From: Tim-Philipp Müller Date: Wed, 15 Sep 2010 20:47:09 +0000 (+0100) Subject: [MOVED FROM BAD 33/68] colorspace: add orc-dist files X-Git-Tag: 1.19.3~511^2~7525 X-Git-Url: http://review.tizen.org/git/?a=commitdiff_plain;h=3b8ae9da70130f910659ae2ad677fbac8bff77a3;p=platform%2Fupstream%2Fgstreamer.git [MOVED FROM BAD 33/68] colorspace: add orc-dist files Should fix build without orc or too old orc. --- diff --git a/gst/colorspace/gstcolorspaceorc-dist.c b/gst/colorspace/gstcolorspaceorc-dist.c new file mode 100644 index 0000000..10b5d8e --- /dev/null +++ b/gst/colorspace/gstcolorspaceorc-dist.c @@ -0,0 +1,16264 @@ + +/* autogenerated from gstcolorspaceorc.orc */ + +#ifdef HAVE_CONFIG_H +#include "config.h" +#endif +#ifndef DISABLE_ORC +#include +#endif +#include + +#ifndef _ORC_INTEGER_TYPEDEFS_ +#define _ORC_INTEGER_TYPEDEFS_ +#if defined(__STDC_VERSION__) && __STDC_VERSION__ >= 199901L +#include +typedef int8_t orc_int8; +typedef int16_t orc_int16; +typedef int32_t orc_int32; +typedef int64_t orc_int64; +typedef uint8_t orc_uint8; +typedef uint16_t orc_uint16; +typedef uint32_t orc_uint32; +typedef uint64_t orc_uint64; +#elif defined(_MSC_VER) +typedef signed __int8 orc_int8; +typedef signed __int16 orc_int16; +typedef signed __int32 orc_int32; +typedef signed __int64 orc_int64; +typedef unsigned __int8 orc_uint8; +typedef unsigned __int16 orc_uint16; +typedef unsigned __int32 orc_uint32; +typedef unsigned __int64 orc_uint64; +#else +#include +typedef signed char orc_int8; +typedef short orc_int16; +typedef int orc_int32; +typedef unsigned char orc_uint8; +typedef unsigned short orc_uint16; +typedef unsigned int orc_uint32; +#if INT_MAX == LONG_MAX +typedef long long orc_int64; +typedef unsigned long long orc_uint64; +#else +typedef long orc_int64; +typedef unsigned long orc_uint64; +#endif +#endif +typedef union +{ + orc_int16 i; + orc_int8 x2[2]; +} orc_union16; +typedef union +{ + orc_int32 i; + float f; + orc_int16 x2[2]; + orc_int8 x4[4]; +} orc_union32; +typedef union +{ + orc_int64 i; + double f; + orc_int32 x2[2]; + orc_int16 x4[4]; +} orc_union64; +#endif + +void cogorc_memcpy_2d (guint8 * d1, int d1_stride, const guint8 * s1, + int s1_stride, int n, int m); +void cogorc_downsample_horiz_cosite_1tap (guint8 * d1, const guint8 * s1, + int n); +void cogorc_downsample_horiz_cosite_3tap (guint8 * d1, const guint8 * s1, + const guint8 * s2, int n); +void cogorc_downsample_420_jpeg (guint8 * d1, const guint8 * s1, + const guint8 * s2, int n); +void cogorc_downsample_vert_halfsite_2tap (guint8 * d1, const guint8 * s1, + const guint8 * s2, int n); +void cogorc_downsample_vert_cosite_3tap (guint8 * d1, const guint8 * s1, + const guint8 * s2, const guint8 * s3, int n); +void cogorc_downsample_vert_halfsite_4tap (guint8 * d1, const guint8 * s1, + const guint8 * s2, const guint8 * s3, const guint8 * s4, int n); +void cogorc_upsample_horiz_cosite_1tap (guint8 * d1, const guint8 * s1, int n); +void cogorc_upsample_horiz_cosite (guint8 * d1, const guint8 * s1, + const guint8 * s2, int n); +void cogorc_upsample_vert_avgub (guint8 * d1, const guint8 * s1, + const guint8 * s2, int n); +void orc_unpack_yuyv_y (guint8 * d1, const guint8 * s1, int n); +void orc_unpack_yuyv_u (guint8 * d1, const guint8 * s1, int n); +void orc_unpack_yuyv_v (guint8 * d1, const guint8 * s1, int n); +void orc_pack_yuyv (guint8 * d1, const guint8 * s1, const guint8 * s2, + const guint8 * s3, int n); +void orc_unpack_uyvy_y (guint8 * d1, const guint8 * s1, int n); +void orc_unpack_uyvy_u (guint8 * d1, const guint8 * s1, int n); +void orc_unpack_uyvy_v (guint8 * d1, const guint8 * s1, int n); +void orc_pack_uyvy (guint8 * d1, const guint8 * s1, const guint8 * s2, + const guint8 * s3, int n); +void orc_matrix2_u8 (guint8 * d1, const guint8 * s1, const guint8 * s2, int p1, + int p2, int p3, int n); +void orc_matrix2_11_u8 (guint8 * d1, const guint8 * s1, const guint8 * s2, + int p1, int p2, int n); +void orc_matrix2_12_u8 (guint8 * d1, const guint8 * s1, const guint8 * s2, + int p1, int p2, int n); +void orc_matrix3_u8 (guint8 * d1, const guint8 * s1, const guint8 * s2, + const guint8 * s3, int p1, int p2, int p3, int p4, int n); +void orc_matrix3_100_u8 (guint8 * d1, const guint8 * s1, const guint8 * s2, + const guint8 * s3, int p1, int p2, int p3, int n); +void orc_matrix3_100_offset_u8 (guint8 * d1, const guint8 * s1, + const guint8 * s2, const guint8 * s3, int p1, int p2, int p3, int p4, + int p5, int n); +void orc_matrix3_000_u8 (guint8 * d1, const guint8 * s1, const guint8 * s2, + const guint8 * s3, int p1, int p2, int p3, int p4, int p5, int n); +void orc_pack_123x (guint32 * d1, const guint8 * s1, const guint8 * s2, + const guint8 * s3, int p1, int n); +void orc_pack_x123 (guint32 * d1, const guint8 * s1, const guint8 * s2, + const guint8 * s3, int p1, int n); +void cogorc_combine2_u8 (guint8 * d1, const guint8 * s1, const guint8 * s2, + int p1, int p2, int n); +void cogorc_convert_I420_UYVY (guint8 * d1, guint8 * d2, const guint8 * s1, + const guint8 * s2, const guint8 * s3, const guint8 * s4, int n); +void cogorc_convert_I420_YUY2 (guint8 * d1, guint8 * d2, const guint8 * s1, + const guint8 * s2, const guint8 * s3, const guint8 * s4, int n); +void cogorc_convert_I420_AYUV (guint8 * d1, guint8 * d2, const guint8 * s1, + const guint8 * s2, const guint8 * s3, const guint8 * s4, int n); +void cogorc_convert_YUY2_I420 (guint8 * d1, guint8 * d2, guint8 * d3, + guint8 * d4, const guint8 * s1, const guint8 * s2, int n); +void cogorc_convert_UYVY_YUY2 (guint8 * d1, int d1_stride, const guint8 * s1, + int s1_stride, int n, int m); +void cogorc_planar_chroma_420_422 (guint8 * d1, int d1_stride, guint8 * d2, + int d2_stride, const guint8 * s1, int s1_stride, int n, int m); +void cogorc_planar_chroma_420_444 (guint8 * d1, int d1_stride, guint8 * d2, + int d2_stride, const guint8 * s1, int s1_stride, int n, int m); +void cogorc_planar_chroma_422_444 (guint8 * d1, int d1_stride, + const guint8 * s1, int s1_stride, int n, int m); +void cogorc_planar_chroma_444_422 (guint8 * d1, int d1_stride, + const guint8 * s1, int s1_stride, int n, int m); +void cogorc_planar_chroma_444_420 (guint8 * d1, int d1_stride, + const guint8 * s1, int s1_stride, const guint8 * s2, int s2_stride, int n, + int m); +void cogorc_planar_chroma_422_420 (guint8 * d1, int d1_stride, + const guint8 * s1, int s1_stride, const guint8 * s2, int s2_stride, int n, + int m); +void cogorc_convert_YUY2_AYUV (guint8 * d1, int d1_stride, const guint8 * s1, + int s1_stride, int n, int m); +void cogorc_convert_UYVY_AYUV (guint8 * d1, int d1_stride, const guint8 * s1, + int s1_stride, int n, int m); +void cogorc_convert_YUY2_Y42B (guint8 * d1, int d1_stride, guint8 * d2, + int d2_stride, guint8 * d3, int d3_stride, const guint8 * s1, int s1_stride, + int n, int m); +void cogorc_convert_UYVY_Y42B (guint8 * d1, int d1_stride, guint8 * d2, + int d2_stride, guint8 * d3, int d3_stride, const guint8 * s1, int s1_stride, + int n, int m); +void cogorc_convert_YUY2_Y444 (guint8 * d1, int d1_stride, guint8 * d2, + int d2_stride, guint8 * d3, int d3_stride, const guint8 * s1, int s1_stride, + int n, int m); +void cogorc_convert_UYVY_Y444 (guint8 * d1, int d1_stride, guint8 * d2, + int d2_stride, guint8 * d3, int d3_stride, const guint8 * s1, int s1_stride, + int n, int m); +void cogorc_convert_UYVY_I420 (guint8 * d1, guint8 * d2, guint8 * d3, + guint8 * d4, const guint8 * s1, const guint8 * s2, int n); +void cogorc_convert_AYUV_I420 (guint8 * d1, int d1_stride, guint8 * d2, + int d2_stride, guint8 * d3, int d3_stride, guint8 * d4, int d4_stride, + const guint8 * s1, int s1_stride, const guint8 * s2, int s2_stride, int n, + int m); +void cogorc_convert_AYUV_YUY2 (guint8 * d1, int d1_stride, const guint8 * s1, + int s1_stride, int n, int m); +void cogorc_convert_AYUV_UYVY (guint8 * d1, int d1_stride, const guint8 * s1, + int s1_stride, int n, int m); +void cogorc_convert_AYUV_Y42B (guint8 * d1, int d1_stride, guint8 * d2, + int d2_stride, guint8 * d3, int d3_stride, const guint8 * s1, int s1_stride, + int n, int m); +void cogorc_convert_AYUV_Y444 (guint8 * d1, int d1_stride, guint8 * d2, + int d2_stride, guint8 * d3, int d3_stride, const guint8 * s1, int s1_stride, + int n, int m); +void cogorc_convert_Y42B_YUY2 (guint8 * d1, int d1_stride, const guint8 * s1, + int s1_stride, const guint8 * s2, int s2_stride, const guint8 * s3, + int s3_stride, int n, int m); +void cogorc_convert_Y42B_UYVY (guint8 * d1, int d1_stride, const guint8 * s1, + int s1_stride, const guint8 * s2, int s2_stride, const guint8 * s3, + int s3_stride, int n, int m); +void cogorc_convert_Y42B_AYUV (guint8 * d1, int d1_stride, const guint8 * s1, + int s1_stride, const guint8 * s2, int s2_stride, const guint8 * s3, + int s3_stride, int n, int m); +void cogorc_convert_Y444_YUY2 (guint8 * d1, int d1_stride, const guint8 * s1, + int s1_stride, const guint8 * s2, int s2_stride, const guint8 * s3, + int s3_stride, int n, int m); +void cogorc_convert_Y444_UYVY (guint8 * d1, int d1_stride, const guint8 * s1, + int s1_stride, const guint8 * s2, int s2_stride, const guint8 * s3, + int s3_stride, int n, int m); +void cogorc_convert_Y444_AYUV (guint8 * d1, int d1_stride, const guint8 * s1, + int s1_stride, const guint8 * s2, int s2_stride, const guint8 * s3, + int s3_stride, int n, int m); +void cogorc_convert_AYUV_ARGB (guint8 * d1, int d1_stride, const guint8 * s1, + int s1_stride, int n, int m); +void cogorc_convert_AYUV_BGRA (guint8 * d1, int d1_stride, const guint8 * s1, + int s1_stride, int n, int m); +void cogorc_convert_AYUV_ABGR (guint8 * d1, int d1_stride, const guint8 * s1, + int s1_stride, int n, int m); +void cogorc_convert_AYUV_RGBA (guint8 * d1, int d1_stride, const guint8 * s1, + int s1_stride, int n, int m); +void cogorc_convert_I420_BGRA (guint8 * d1, const guint8 * s1, + const guint8 * s2, const guint8 * s3, int n); +void cogorc_convert_I420_BGRA_avg (guint8 * d1, const guint8 * s1, + const guint8 * s2, const guint8 * s3, const guint8 * s4, const guint8 * s5, + int n); +void cogorc_getline_I420 (guint8 * d1, const guint8 * s1, const guint8 * s2, + const guint8 * s3, int n); +void cogorc_getline_YUY2 (guint8 * d1, const guint8 * s1, int n); +void cogorc_getline_UYVY (guint8 * d1, const guint8 * s1, int n); +void cogorc_getline_YVYU (guint8 * d1, const guint8 * s1, int n); +void cogorc_getline_Y42B (guint8 * d1, const guint8 * s1, const guint8 * s2, + const guint8 * s3, int n); +void cogorc_getline_Y444 (guint8 * d1, const guint8 * s1, const guint8 * s2, + const guint8 * s3, int n); +void cogorc_getline_Y800 (guint8 * d1, const guint8 * s1, int n); +void cogorc_getline_BGRA (guint8 * d1, const guint8 * s1, int n); +void cogorc_getline_ABGR (guint8 * d1, const guint8 * s1, int n); +void cogorc_getline_RGBA (guint8 * d1, const guint8 * s1, int n); +void cogorc_getline_NV12 (guint8 * d1, const guint8 * s1, const guint8 * s2, + int n); +void cogorc_getline_NV21 (guint8 * d1, const guint8 * s1, const guint8 * s2, + int n); +void cogorc_putline_I420 (guint8 * d1, guint8 * d2, guint8 * d3, + const guint8 * s1, int n); +void cogorc_putline_YUY2 (guint8 * d1, const guint8 * s1, int n); +void cogorc_putline_YVYU (guint8 * d1, const guint8 * s1, int n); +void cogorc_putline_UYVY (guint8 * d1, const guint8 * s1, int n); +void cogorc_putline_Y42B (guint8 * d1, guint8 * d2, guint8 * d3, + const guint8 * s1, int n); +void cogorc_putline_Y444 (guint8 * d1, guint8 * d2, guint8 * d3, + const guint8 * s1, int n); +void cogorc_putline_Y800 (guint8 * d1, const guint8 * s1, int n); +void cogorc_putline_BGRA (guint8 * d1, const guint8 * s1, int n); +void cogorc_putline_ABGR (guint8 * d1, const guint8 * s1, int n); +void cogorc_putline_RGBA (guint8 * d1, const guint8 * s1, int n); +void cogorc_putline_NV12 (guint8 * d1, guint8 * d2, const guint8 * s1, int n); +void cogorc_putline_NV21 (guint8 * d1, guint8 * d2, const guint8 * s1, int n); + + +/* begin Orc C target preamble */ +#define ORC_CLAMP(x,a,b) ((x)<(a) ? (a) : ((x)>(b) ? (b) : (x))) +#define ORC_ABS(a) ((a)<0 ? -(a) : (a)) +#define ORC_MIN(a,b) ((a)<(b) ? (a) : (b)) +#define ORC_MAX(a,b) ((a)>(b) ? (a) : (b)) +#define ORC_SB_MAX 127 +#define ORC_SB_MIN (-1-ORC_SB_MAX) +#define ORC_UB_MAX 255 +#define ORC_UB_MIN 0 +#define ORC_SW_MAX 32767 +#define ORC_SW_MIN (-1-ORC_SW_MAX) +#define ORC_UW_MAX 65535 +#define ORC_UW_MIN 0 +#define ORC_SL_MAX 2147483647 +#define ORC_SL_MIN (-1-ORC_SL_MAX) +#define ORC_UL_MAX 4294967295U +#define ORC_UL_MIN 0 +#define ORC_CLAMP_SB(x) ORC_CLAMP(x,ORC_SB_MIN,ORC_SB_MAX) +#define ORC_CLAMP_UB(x) ORC_CLAMP(x,ORC_UB_MIN,ORC_UB_MAX) +#define ORC_CLAMP_SW(x) ORC_CLAMP(x,ORC_SW_MIN,ORC_SW_MAX) +#define ORC_CLAMP_UW(x) ORC_CLAMP(x,ORC_UW_MIN,ORC_UW_MAX) +#define ORC_CLAMP_SL(x) ORC_CLAMP(x,ORC_SL_MIN,ORC_SL_MAX) +#define ORC_CLAMP_UL(x) ORC_CLAMP(x,ORC_UL_MIN,ORC_UL_MAX) +#define ORC_SWAP_W(x) ((((x)&0xff)<<8) | (((x)&0xff00)>>8)) +#define ORC_SWAP_L(x) ((((x)&0xff)<<24) | (((x)&0xff00)<<8) | (((x)&0xff0000)>>8) | (((x)&0xff000000)>>24)) +#define ORC_SWAP_Q(x) ((((x)&0xffULL)<<56) | (((x)&0xff00ULL)<<40) | (((x)&0xff0000ULL)<<24) | (((x)&0xff000000ULL)<<8) | (((x)&0xff00000000ULL)>>8) | (((x)&0xff0000000000ULL)>>24) | (((x)&0xff000000000000ULL)>>40) | (((x)&0xff00000000000000ULL)>>56)) +#define ORC_PTR_OFFSET(ptr,offset) ((void *)(((unsigned char *)(ptr)) + (offset))) +#define ORC_DENORMAL(x) ((x) & ((((x)&0x7f800000) == 0) ? 0xff800000 : 0xffffffff)) +#define ORC_ISNAN(x) ((((x)&0x7f800000) == 0x7f800000) && (((x)&0x007fffff) != 0)) +#define ORC_DENORMAL_DOUBLE(x) ((x) & ((((x)&0x7ff0000000000000ULL) == 0) ? 0xfff0000000000000ULL : 0xffffffffffffffffULL)) +#define ORC_ISNAN_DOUBLE(x) ((((x)&0x7ff0000000000000ULL) == 0x7ff0000000000000ULL) && (((x)&0x000fffffffffffffULL) != 0)) +#if defined(__STDC_VERSION__) && __STDC_VERSION__ >= 199901L +#define ORC_RESTRICT restrict +#elif defined(__GNUC__) && __GNUC__ >= 4 +#define ORC_RESTRICT __restrict__ +#else +#define ORC_RESTRICT +#endif +/* end Orc C target preamble */ + + + +/* cogorc_memcpy_2d */ +#ifdef DISABLE_ORC +void +cogorc_memcpy_2d (guint8 * d1, int d1_stride, const guint8 * s1, int s1_stride, + int n, int m) +{ + int i; + int j; + orc_int8 *ORC_RESTRICT ptr0; + const orc_int8 *ORC_RESTRICT ptr4; + orc_int8 var32; + orc_int8 var33; + + for (j = 0; j < m; j++) { + ptr0 = ORC_PTR_OFFSET (d1, d1_stride * j); + ptr4 = ORC_PTR_OFFSET (s1, s1_stride * j); + + + for (i = 0; i < n; i++) { + /* 0: loadb */ + var32 = ptr4[i]; + /* 1: copyb */ + var33 = var32; + /* 2: storeb */ + ptr0[i] = var33; + } + } + +} + +#else +static void +_backup_cogorc_memcpy_2d (OrcExecutor * ex) +{ + int i; + int j; + int n = ex->n; + int m = ex->params[ORC_VAR_A1]; + orc_int8 *ORC_RESTRICT ptr0; + const orc_int8 *ORC_RESTRICT ptr4; + orc_int8 var32; + orc_int8 var33; + + for (j = 0; j < m; j++) { + ptr0 = ORC_PTR_OFFSET (ex->arrays[0], ex->params[0] * j); + ptr4 = ORC_PTR_OFFSET (ex->arrays[4], ex->params[4] * j); + + + for (i = 0; i < n; i++) { + /* 0: loadb */ + var32 = ptr4[i]; + /* 1: copyb */ + var33 = var32; + /* 2: storeb */ + ptr0[i] = var33; + } + } + +} + +void +cogorc_memcpy_2d (guint8 * d1, int d1_stride, const guint8 * s1, int s1_stride, + int n, int m) +{ + OrcExecutor _ex, *ex = &_ex; + static int p_inited = 0; + static OrcProgram *p = 0; + void (*func) (OrcExecutor *); + + if (!p_inited) { + orc_once_mutex_lock (); + if (!p_inited) { + OrcCompileResult result; + + p = orc_program_new (); + orc_program_set_2d (p); + orc_program_set_name (p, "cogorc_memcpy_2d"); + orc_program_set_backup_function (p, _backup_cogorc_memcpy_2d); + orc_program_add_destination (p, 1, "d1"); + orc_program_add_source (p, 1, "s1"); + orc_program_add_constant (p, 0, 0x00000000, "c1"); + orc_program_add_constant (p, 0, 0x00000000, "c2"); + orc_program_add_constant (p, 0, 0x00000000, "c3"); + orc_program_add_constant (p, 0, 0x00000000, "c4"); + orc_program_add_constant (p, 0, 0x00000000, "c5"); + orc_program_add_constant (p, 0, 0x00000000, "c6"); + orc_program_add_constant (p, 0, 0x00000000, "c7"); + orc_program_add_constant (p, 0, 0x00000000, "c8"); + + orc_program_append_2 (p, "copyb", 0, ORC_VAR_D1, ORC_VAR_S1, ORC_VAR_D1, + ORC_VAR_D1); + + result = orc_program_compile (p); + } + p_inited = TRUE; + orc_once_mutex_unlock (); + } + ex->program = p; + + ex->n = n; + ORC_EXECUTOR_M (ex) = m; + ex->arrays[ORC_VAR_D1] = d1; + ex->params[ORC_VAR_D1] = d1_stride; + ex->arrays[ORC_VAR_S1] = (void *) s1; + ex->params[ORC_VAR_S1] = s1_stride; + + func = p->code_exec; + func (ex); +} +#endif + + +/* cogorc_downsample_horiz_cosite_1tap */ +#ifdef DISABLE_ORC +void +cogorc_downsample_horiz_cosite_1tap (guint8 * d1, const guint8 * s1, int n) +{ + int i; + orc_int8 *ORC_RESTRICT ptr0; + const orc_union16 *ORC_RESTRICT ptr4; + orc_union16 var32; + orc_int8 var33; + + ptr0 = (orc_int8 *) d1; + ptr4 = (orc_union16 *) s1; + + + for (i = 0; i < n; i++) { + /* 0: loadw */ + var32 = ptr4[i]; + /* 1: select0wb */ + var33 = (orc_uint16) var32.i & 0xff; + /* 2: storeb */ + ptr0[i] = var33; + } + +} + +#else +static void +_backup_cogorc_downsample_horiz_cosite_1tap (OrcExecutor * ex) +{ + int i; + int n = ex->n; + orc_int8 *ORC_RESTRICT ptr0; + const orc_union16 *ORC_RESTRICT ptr4; + orc_union16 var32; + orc_int8 var33; + + ptr0 = (orc_int8 *) ex->arrays[0]; + ptr4 = (orc_union16 *) ex->arrays[4]; + + + for (i = 0; i < n; i++) { + /* 0: loadw */ + var32 = ptr4[i]; + /* 1: select0wb */ + var33 = (orc_uint16) var32.i & 0xff; + /* 2: storeb */ + ptr0[i] = var33; + } + +} + +void +cogorc_downsample_horiz_cosite_1tap (guint8 * d1, const guint8 * s1, int n) +{ + OrcExecutor _ex, *ex = &_ex; + static int p_inited = 0; + static OrcProgram *p = 0; + void (*func) (OrcExecutor *); + + if (!p_inited) { + orc_once_mutex_lock (); + if (!p_inited) { + OrcCompileResult result; + + p = orc_program_new (); + orc_program_set_name (p, "cogorc_downsample_horiz_cosite_1tap"); + orc_program_set_backup_function (p, + _backup_cogorc_downsample_horiz_cosite_1tap); + orc_program_add_destination (p, 1, "d1"); + orc_program_add_source (p, 2, "s1"); + orc_program_add_constant (p, 0, 0x00000000, "c1"); + orc_program_add_constant (p, 0, 0x00000000, "c2"); + orc_program_add_constant (p, 0, 0x00000000, "c3"); + orc_program_add_constant (p, 0, 0x00000000, "c4"); + orc_program_add_constant (p, 0, 0x00000000, "c5"); + orc_program_add_constant (p, 0, 0x00000000, "c6"); + orc_program_add_constant (p, 0, 0x00000000, "c7"); + orc_program_add_constant (p, 0, 0x00000000, "c8"); + + orc_program_append_2 (p, "select0wb", 0, ORC_VAR_D1, ORC_VAR_S1, + ORC_VAR_D1, ORC_VAR_D1); + + result = orc_program_compile (p); + } + p_inited = TRUE; + orc_once_mutex_unlock (); + } + ex->program = p; + + ex->n = n; + ex->arrays[ORC_VAR_D1] = d1; + ex->arrays[ORC_VAR_S1] = (void *) s1; + + func = p->code_exec; + func (ex); +} +#endif + + +/* cogorc_downsample_horiz_cosite_3tap */ +#ifdef DISABLE_ORC +void +cogorc_downsample_horiz_cosite_3tap (guint8 * d1, const guint8 * s1, + const guint8 * s2, int n) +{ + int i; + orc_int8 *ORC_RESTRICT ptr0; + const orc_union16 *ORC_RESTRICT ptr4; + const orc_union16 *ORC_RESTRICT ptr5; + orc_union16 var38; + orc_union16 var39; + orc_union16 var40; + orc_union16 var41; + orc_int8 var42; + orc_union16 var43; + orc_int8 var44; + orc_int8 var45; + orc_int8 var46; + orc_union16 var47; + orc_union16 var48; + orc_union16 var49; + orc_union16 var50; + orc_union16 var51; + orc_union16 var52; + orc_union16 var53; + orc_union16 var54; + + ptr0 = (orc_int8 *) d1; + ptr4 = (orc_union16 *) s1; + ptr5 = (orc_union16 *) s2; + + /* 9: loadpw */ + var40.i = 0x00000002; /* 2 or 9.88131e-324f */ + /* 13: loadpw */ + var41.i = 0x00000002; /* 2 or 9.88131e-324f */ + + for (i = 0; i < n; i++) { + /* 0: loadw */ + var38 = ptr4[i]; + /* 1: copyw */ + var43.i = var38.i; + /* 2: select0wb */ + var44 = (orc_uint16) var43.i & 0xff; + /* 3: select1wb */ + var45 = ((orc_uint16) var43.i >> 8) & 0xff; + /* 4: loadw */ + var39 = ptr5[i]; + /* 5: select0wb */ + var46 = (orc_uint16) var39.i & 0xff; + /* 6: convubw */ + var47.i = (orc_uint8) var44; + /* 7: convubw */ + var48.i = (orc_uint8) var45; + /* 8: convubw */ + var49.i = (orc_uint8) var46; + /* 10: mullw */ + var50.i = (var48.i * var40.i) & 0xffff; + /* 11: addw */ + var51.i = var47.i + var49.i; + /* 12: addw */ + var52.i = var51.i + var50.i; + /* 14: addw */ + var53.i = var52.i + var41.i; + /* 15: shrsw */ + var54.i = var53.i >> 2; + /* 16: convsuswb */ + var42 = ORC_CLAMP_UB (var54.i); + /* 17: storeb */ + ptr0[i] = var42; + } + +} + +#else +static void +_backup_cogorc_downsample_horiz_cosite_3tap (OrcExecutor * ex) +{ + int i; + int n = ex->n; + orc_int8 *ORC_RESTRICT ptr0; + const orc_union16 *ORC_RESTRICT ptr4; + const orc_union16 *ORC_RESTRICT ptr5; + orc_union16 var38; + orc_union16 var39; + orc_union16 var40; + orc_union16 var41; + orc_int8 var42; + orc_union16 var43; + orc_int8 var44; + orc_int8 var45; + orc_int8 var46; + orc_union16 var47; + orc_union16 var48; + orc_union16 var49; + orc_union16 var50; + orc_union16 var51; + orc_union16 var52; + orc_union16 var53; + orc_union16 var54; + + ptr0 = (orc_int8 *) ex->arrays[0]; + ptr4 = (orc_union16 *) ex->arrays[4]; + ptr5 = (orc_union16 *) ex->arrays[5]; + + /* 9: loadpw */ + var40.i = 0x00000002; /* 2 or 9.88131e-324f */ + /* 13: loadpw */ + var41.i = 0x00000002; /* 2 or 9.88131e-324f */ + + for (i = 0; i < n; i++) { + /* 0: loadw */ + var38 = ptr4[i]; + /* 1: copyw */ + var43.i = var38.i; + /* 2: select0wb */ + var44 = (orc_uint16) var43.i & 0xff; + /* 3: select1wb */ + var45 = ((orc_uint16) var43.i >> 8) & 0xff; + /* 4: loadw */ + var39 = ptr5[i]; + /* 5: select0wb */ + var46 = (orc_uint16) var39.i & 0xff; + /* 6: convubw */ + var47.i = (orc_uint8) var44; + /* 7: convubw */ + var48.i = (orc_uint8) var45; + /* 8: convubw */ + var49.i = (orc_uint8) var46; + /* 10: mullw */ + var50.i = (var48.i * var40.i) & 0xffff; + /* 11: addw */ + var51.i = var47.i + var49.i; + /* 12: addw */ + var52.i = var51.i + var50.i; + /* 14: addw */ + var53.i = var52.i + var41.i; + /* 15: shrsw */ + var54.i = var53.i >> 2; + /* 16: convsuswb */ + var42 = ORC_CLAMP_UB (var54.i); + /* 17: storeb */ + ptr0[i] = var42; + } + +} + +void +cogorc_downsample_horiz_cosite_3tap (guint8 * d1, const guint8 * s1, + const guint8 * s2, int n) +{ + OrcExecutor _ex, *ex = &_ex; + static int p_inited = 0; + static OrcProgram *p = 0; + void (*func) (OrcExecutor *); + + if (!p_inited) { + orc_once_mutex_lock (); + if (!p_inited) { + OrcCompileResult result; + + p = orc_program_new (); + orc_program_set_name (p, "cogorc_downsample_horiz_cosite_3tap"); + orc_program_set_backup_function (p, + _backup_cogorc_downsample_horiz_cosite_3tap); + orc_program_add_destination (p, 1, "d1"); + orc_program_add_source (p, 2, "s1"); + orc_program_add_source (p, 2, "s2"); + orc_program_add_constant (p, 4, 0x00000002, "c1"); + orc_program_add_constant (p, 0, 0x00000002, "c2"); + orc_program_add_constant (p, 0, 0x00000000, "c3"); + orc_program_add_constant (p, 0, 0x00000000, "c4"); + orc_program_add_constant (p, 0, 0x00000000, "c5"); + orc_program_add_constant (p, 0, 0x00000000, "c6"); + orc_program_add_constant (p, 0, 0x00000000, "c7"); + orc_program_add_constant (p, 0, 0x00000000, "c8"); + orc_program_add_temporary (p, 1, "t1"); + orc_program_add_temporary (p, 1, "t2"); + orc_program_add_temporary (p, 1, "t3"); + orc_program_add_temporary (p, 2, "t4"); + orc_program_add_temporary (p, 2, "t5"); + orc_program_add_temporary (p, 2, "t6"); + + orc_program_append_2 (p, "copyw", 0, ORC_VAR_T4, ORC_VAR_S1, ORC_VAR_D1, + ORC_VAR_D1); + orc_program_append_2 (p, "select0wb", 0, ORC_VAR_T1, ORC_VAR_T4, + ORC_VAR_D1, ORC_VAR_D1); + orc_program_append_2 (p, "select1wb", 0, ORC_VAR_T2, ORC_VAR_T4, + ORC_VAR_D1, ORC_VAR_D1); + orc_program_append_2 (p, "select0wb", 0, ORC_VAR_T3, ORC_VAR_S2, + ORC_VAR_D1, ORC_VAR_D1); + orc_program_append_2 (p, "convubw", 0, ORC_VAR_T4, ORC_VAR_T1, ORC_VAR_D1, + ORC_VAR_D1); + orc_program_append_2 (p, "convubw", 0, ORC_VAR_T5, ORC_VAR_T2, ORC_VAR_D1, + ORC_VAR_D1); + orc_program_append_2 (p, "convubw", 0, ORC_VAR_T6, ORC_VAR_T3, ORC_VAR_D1, + ORC_VAR_D1); + orc_program_append_2 (p, "mullw", 0, ORC_VAR_T5, ORC_VAR_T5, ORC_VAR_C1, + ORC_VAR_D1); + orc_program_append_2 (p, "addw", 0, ORC_VAR_T4, ORC_VAR_T4, ORC_VAR_T6, + ORC_VAR_D1); + orc_program_append_2 (p, "addw", 0, ORC_VAR_T4, ORC_VAR_T4, ORC_VAR_T5, + ORC_VAR_D1); + orc_program_append_2 (p, "addw", 0, ORC_VAR_T4, ORC_VAR_T4, ORC_VAR_C1, + ORC_VAR_D1); + orc_program_append_2 (p, "shrsw", 0, ORC_VAR_T4, ORC_VAR_T4, ORC_VAR_C1, + ORC_VAR_D1); + orc_program_append_2 (p, "convsuswb", 0, ORC_VAR_D1, ORC_VAR_T4, + ORC_VAR_D1, ORC_VAR_D1); + + result = orc_program_compile (p); + } + p_inited = TRUE; + orc_once_mutex_unlock (); + } + ex->program = p; + + ex->n = n; + ex->arrays[ORC_VAR_D1] = d1; + ex->arrays[ORC_VAR_S1] = (void *) s1; + ex->arrays[ORC_VAR_S2] = (void *) s2; + + func = p->code_exec; + func (ex); +} +#endif + + +/* cogorc_downsample_420_jpeg */ +#ifdef DISABLE_ORC +void +cogorc_downsample_420_jpeg (guint8 * d1, const guint8 * s1, const guint8 * s2, + int n) +{ + int i; + orc_int8 *ORC_RESTRICT ptr0; + const orc_union16 *ORC_RESTRICT ptr4; + const orc_union16 *ORC_RESTRICT ptr5; + orc_union16 var37; + orc_union16 var38; + orc_int8 var39; + orc_union16 var40; + orc_int8 var41; + orc_int8 var42; + orc_int8 var43; + orc_union16 var44; + orc_int8 var45; + orc_int8 var46; + orc_int8 var47; + + ptr0 = (orc_int8 *) d1; + ptr4 = (orc_union16 *) s1; + ptr5 = (orc_union16 *) s2; + + + for (i = 0; i < n; i++) { + /* 0: loadw */ + var37 = ptr4[i]; + /* 1: copyw */ + var40.i = var37.i; + /* 2: select0wb */ + var41 = (orc_uint16) var40.i & 0xff; + /* 3: select1wb */ + var42 = ((orc_uint16) var40.i >> 8) & 0xff; + /* 4: avgub */ + var43 = ((orc_uint8) var41 + (orc_uint8) var42 + 1) >> 1; + /* 5: loadw */ + var38 = ptr5[i]; + /* 6: copyw */ + var44.i = var38.i; + /* 7: select0wb */ + var45 = (orc_uint16) var44.i & 0xff; + /* 8: select1wb */ + var46 = ((orc_uint16) var44.i >> 8) & 0xff; + /* 9: avgub */ + var47 = ((orc_uint8) var45 + (orc_uint8) var46 + 1) >> 1; + /* 10: avgub */ + var39 = ((orc_uint8) var43 + (orc_uint8) var47 + 1) >> 1; + /* 11: storeb */ + ptr0[i] = var39; + } + +} + +#else +static void +_backup_cogorc_downsample_420_jpeg (OrcExecutor * ex) +{ + int i; + int n = ex->n; + orc_int8 *ORC_RESTRICT ptr0; + const orc_union16 *ORC_RESTRICT ptr4; + const orc_union16 *ORC_RESTRICT ptr5; + orc_union16 var37; + orc_union16 var38; + orc_int8 var39; + orc_union16 var40; + orc_int8 var41; + orc_int8 var42; + orc_int8 var43; + orc_union16 var44; + orc_int8 var45; + orc_int8 var46; + orc_int8 var47; + + ptr0 = (orc_int8 *) ex->arrays[0]; + ptr4 = (orc_union16 *) ex->arrays[4]; + ptr5 = (orc_union16 *) ex->arrays[5]; + + + for (i = 0; i < n; i++) { + /* 0: loadw */ + var37 = ptr4[i]; + /* 1: copyw */ + var40.i = var37.i; + /* 2: select0wb */ + var41 = (orc_uint16) var40.i & 0xff; + /* 3: select1wb */ + var42 = ((orc_uint16) var40.i >> 8) & 0xff; + /* 4: avgub */ + var43 = ((orc_uint8) var41 + (orc_uint8) var42 + 1) >> 1; + /* 5: loadw */ + var38 = ptr5[i]; + /* 6: copyw */ + var44.i = var38.i; + /* 7: select0wb */ + var45 = (orc_uint16) var44.i & 0xff; + /* 8: select1wb */ + var46 = ((orc_uint16) var44.i >> 8) & 0xff; + /* 9: avgub */ + var47 = ((orc_uint8) var45 + (orc_uint8) var46 + 1) >> 1; + /* 10: avgub */ + var39 = ((orc_uint8) var43 + (orc_uint8) var47 + 1) >> 1; + /* 11: storeb */ + ptr0[i] = var39; + } + +} + +void +cogorc_downsample_420_jpeg (guint8 * d1, const guint8 * s1, const guint8 * s2, + int n) +{ + OrcExecutor _ex, *ex = &_ex; + static int p_inited = 0; + static OrcProgram *p = 0; + void (*func) (OrcExecutor *); + + if (!p_inited) { + orc_once_mutex_lock (); + if (!p_inited) { + OrcCompileResult result; + + p = orc_program_new (); + orc_program_set_name (p, "cogorc_downsample_420_jpeg"); + orc_program_set_backup_function (p, _backup_cogorc_downsample_420_jpeg); + orc_program_add_destination (p, 1, "d1"); + orc_program_add_source (p, 2, "s1"); + orc_program_add_source (p, 2, "s2"); + orc_program_add_constant (p, 0, 0x00000000, "c1"); + orc_program_add_constant (p, 0, 0x00000000, "c2"); + orc_program_add_constant (p, 0, 0x00000000, "c3"); + orc_program_add_constant (p, 0, 0x00000000, "c4"); + orc_program_add_constant (p, 0, 0x00000000, "c5"); + orc_program_add_constant (p, 0, 0x00000000, "c6"); + orc_program_add_constant (p, 0, 0x00000000, "c7"); + orc_program_add_constant (p, 0, 0x00000000, "c8"); + orc_program_add_temporary (p, 2, "t1"); + orc_program_add_temporary (p, 1, "t2"); + orc_program_add_temporary (p, 1, "t3"); + orc_program_add_temporary (p, 1, "t4"); + orc_program_add_temporary (p, 1, "t5"); + + orc_program_append_2 (p, "copyw", 0, ORC_VAR_T1, ORC_VAR_S1, ORC_VAR_D1, + ORC_VAR_D1); + orc_program_append_2 (p, "select0wb", 0, ORC_VAR_T2, ORC_VAR_T1, + ORC_VAR_D1, ORC_VAR_D1); + orc_program_append_2 (p, "select1wb", 0, ORC_VAR_T3, ORC_VAR_T1, + ORC_VAR_D1, ORC_VAR_D1); + orc_program_append_2 (p, "avgub", 0, ORC_VAR_T2, ORC_VAR_T2, ORC_VAR_T3, + ORC_VAR_D1); + orc_program_append_2 (p, "copyw", 0, ORC_VAR_T1, ORC_VAR_S2, ORC_VAR_D1, + ORC_VAR_D1); + orc_program_append_2 (p, "select0wb", 0, ORC_VAR_T4, ORC_VAR_T1, + ORC_VAR_D1, ORC_VAR_D1); + orc_program_append_2 (p, "select1wb", 0, ORC_VAR_T5, ORC_VAR_T1, + ORC_VAR_D1, ORC_VAR_D1); + orc_program_append_2 (p, "avgub", 0, ORC_VAR_T4, ORC_VAR_T4, ORC_VAR_T5, + ORC_VAR_D1); + orc_program_append_2 (p, "avgub", 0, ORC_VAR_D1, ORC_VAR_T2, ORC_VAR_T4, + ORC_VAR_D1); + + result = orc_program_compile (p); + } + p_inited = TRUE; + orc_once_mutex_unlock (); + } + ex->program = p; + + ex->n = n; + ex->arrays[ORC_VAR_D1] = d1; + ex->arrays[ORC_VAR_S1] = (void *) s1; + ex->arrays[ORC_VAR_S2] = (void *) s2; + + func = p->code_exec; + func (ex); +} +#endif + + +/* cogorc_downsample_vert_halfsite_2tap */ +#ifdef DISABLE_ORC +void +cogorc_downsample_vert_halfsite_2tap (guint8 * d1, const guint8 * s1, + const guint8 * s2, int n) +{ + int i; + orc_int8 *ORC_RESTRICT ptr0; + const orc_int8 *ORC_RESTRICT ptr4; + const orc_int8 *ORC_RESTRICT ptr5; + orc_int8 var32; + orc_int8 var33; + orc_int8 var34; + + ptr0 = (orc_int8 *) d1; + ptr4 = (orc_int8 *) s1; + ptr5 = (orc_int8 *) s2; + + + for (i = 0; i < n; i++) { + /* 0: loadb */ + var32 = ptr4[i]; + /* 1: loadb */ + var33 = ptr5[i]; + /* 2: avgub */ + var34 = ((orc_uint8) var32 + (orc_uint8) var33 + 1) >> 1; + /* 3: storeb */ + ptr0[i] = var34; + } + +} + +#else +static void +_backup_cogorc_downsample_vert_halfsite_2tap (OrcExecutor * ex) +{ + int i; + int n = ex->n; + orc_int8 *ORC_RESTRICT ptr0; + const orc_int8 *ORC_RESTRICT ptr4; + const orc_int8 *ORC_RESTRICT ptr5; + orc_int8 var32; + orc_int8 var33; + orc_int8 var34; + + ptr0 = (orc_int8 *) ex->arrays[0]; + ptr4 = (orc_int8 *) ex->arrays[4]; + ptr5 = (orc_int8 *) ex->arrays[5]; + + + for (i = 0; i < n; i++) { + /* 0: loadb */ + var32 = ptr4[i]; + /* 1: loadb */ + var33 = ptr5[i]; + /* 2: avgub */ + var34 = ((orc_uint8) var32 + (orc_uint8) var33 + 1) >> 1; + /* 3: storeb */ + ptr0[i] = var34; + } + +} + +void +cogorc_downsample_vert_halfsite_2tap (guint8 * d1, const guint8 * s1, + const guint8 * s2, int n) +{ + OrcExecutor _ex, *ex = &_ex; + static int p_inited = 0; + static OrcProgram *p = 0; + void (*func) (OrcExecutor *); + + if (!p_inited) { + orc_once_mutex_lock (); + if (!p_inited) { + OrcCompileResult result; + + p = orc_program_new (); + orc_program_set_name (p, "cogorc_downsample_vert_halfsite_2tap"); + orc_program_set_backup_function (p, + _backup_cogorc_downsample_vert_halfsite_2tap); + orc_program_add_destination (p, 1, "d1"); + orc_program_add_source (p, 1, "s1"); + orc_program_add_source (p, 1, "s2"); + orc_program_add_constant (p, 0, 0x00000000, "c1"); + orc_program_add_constant (p, 0, 0x00000000, "c2"); + orc_program_add_constant (p, 0, 0x00000000, "c3"); + orc_program_add_constant (p, 0, 0x00000000, "c4"); + orc_program_add_constant (p, 0, 0x00000000, "c5"); + orc_program_add_constant (p, 0, 0x00000000, "c6"); + orc_program_add_constant (p, 0, 0x00000000, "c7"); + orc_program_add_constant (p, 0, 0x00000000, "c8"); + + orc_program_append_2 (p, "avgub", 0, ORC_VAR_D1, ORC_VAR_S1, ORC_VAR_S2, + ORC_VAR_D1); + + result = orc_program_compile (p); + } + p_inited = TRUE; + orc_once_mutex_unlock (); + } + ex->program = p; + + ex->n = n; + ex->arrays[ORC_VAR_D1] = d1; + ex->arrays[ORC_VAR_S1] = (void *) s1; + ex->arrays[ORC_VAR_S2] = (void *) s2; + + func = p->code_exec; + func (ex); +} +#endif + + +/* cogorc_downsample_vert_cosite_3tap */ +#ifdef DISABLE_ORC +void +cogorc_downsample_vert_cosite_3tap (guint8 * d1, const guint8 * s1, + const guint8 * s2, const guint8 * s3, int n) +{ + int i; + orc_int8 *ORC_RESTRICT ptr0; + const orc_int8 *ORC_RESTRICT ptr4; + const orc_int8 *ORC_RESTRICT ptr5; + const orc_int8 *ORC_RESTRICT ptr6; + orc_int8 var35; + orc_int8 var36; + orc_int8 var37; + orc_union16 var38; + orc_union16 var39; + orc_int8 var40; + orc_union16 var41; + orc_union16 var42; + orc_union16 var43; + orc_union16 var44; + orc_union16 var45; + orc_union16 var46; + orc_union16 var47; + orc_union16 var48; + + ptr0 = (orc_int8 *) d1; + ptr4 = (orc_int8 *) s1; + ptr5 = (orc_int8 *) s2; + ptr6 = (orc_int8 *) s3; + + /* 6: loadpw */ + var38.i = 0x00000002; /* 2 or 9.88131e-324f */ + /* 10: loadpw */ + var39.i = 0x00000002; /* 2 or 9.88131e-324f */ + + for (i = 0; i < n; i++) { + /* 0: loadb */ + var35 = ptr4[i]; + /* 1: convubw */ + var41.i = (orc_uint8) var35; + /* 2: loadb */ + var36 = ptr5[i]; + /* 3: convubw */ + var42.i = (orc_uint8) var36; + /* 4: loadb */ + var37 = ptr6[i]; + /* 5: convubw */ + var43.i = (orc_uint8) var37; + /* 7: mullw */ + var44.i = (var42.i * var38.i) & 0xffff; + /* 8: addw */ + var45.i = var41.i + var43.i; + /* 9: addw */ + var46.i = var45.i + var44.i; + /* 11: addw */ + var47.i = var46.i + var39.i; + /* 12: shrsw */ + var48.i = var47.i >> 2; + /* 13: convsuswb */ + var40 = ORC_CLAMP_UB (var48.i); + /* 14: storeb */ + ptr0[i] = var40; + } + +} + +#else +static void +_backup_cogorc_downsample_vert_cosite_3tap (OrcExecutor * ex) +{ + int i; + int n = ex->n; + orc_int8 *ORC_RESTRICT ptr0; + const orc_int8 *ORC_RESTRICT ptr4; + const orc_int8 *ORC_RESTRICT ptr5; + const orc_int8 *ORC_RESTRICT ptr6; + orc_int8 var35; + orc_int8 var36; + orc_int8 var37; + orc_union16 var38; + orc_union16 var39; + orc_int8 var40; + orc_union16 var41; + orc_union16 var42; + orc_union16 var43; + orc_union16 var44; + orc_union16 var45; + orc_union16 var46; + orc_union16 var47; + orc_union16 var48; + + ptr0 = (orc_int8 *) ex->arrays[0]; + ptr4 = (orc_int8 *) ex->arrays[4]; + ptr5 = (orc_int8 *) ex->arrays[5]; + ptr6 = (orc_int8 *) ex->arrays[6]; + + /* 6: loadpw */ + var38.i = 0x00000002; /* 2 or 9.88131e-324f */ + /* 10: loadpw */ + var39.i = 0x00000002; /* 2 or 9.88131e-324f */ + + for (i = 0; i < n; i++) { + /* 0: loadb */ + var35 = ptr4[i]; + /* 1: convubw */ + var41.i = (orc_uint8) var35; + /* 2: loadb */ + var36 = ptr5[i]; + /* 3: convubw */ + var42.i = (orc_uint8) var36; + /* 4: loadb */ + var37 = ptr6[i]; + /* 5: convubw */ + var43.i = (orc_uint8) var37; + /* 7: mullw */ + var44.i = (var42.i * var38.i) & 0xffff; + /* 8: addw */ + var45.i = var41.i + var43.i; + /* 9: addw */ + var46.i = var45.i + var44.i; + /* 11: addw */ + var47.i = var46.i + var39.i; + /* 12: shrsw */ + var48.i = var47.i >> 2; + /* 13: convsuswb */ + var40 = ORC_CLAMP_UB (var48.i); + /* 14: storeb */ + ptr0[i] = var40; + } + +} + +void +cogorc_downsample_vert_cosite_3tap (guint8 * d1, const guint8 * s1, + const guint8 * s2, const guint8 * s3, int n) +{ + OrcExecutor _ex, *ex = &_ex; + static int p_inited = 0; + static OrcProgram *p = 0; + void (*func) (OrcExecutor *); + + if (!p_inited) { + orc_once_mutex_lock (); + if (!p_inited) { + OrcCompileResult result; + + p = orc_program_new (); + orc_program_set_name (p, "cogorc_downsample_vert_cosite_3tap"); + orc_program_set_backup_function (p, + _backup_cogorc_downsample_vert_cosite_3tap); + orc_program_add_destination (p, 1, "d1"); + orc_program_add_source (p, 1, "s1"); + orc_program_add_source (p, 1, "s2"); + orc_program_add_source (p, 1, "s3"); + orc_program_add_constant (p, 4, 0x00000002, "c1"); + orc_program_add_constant (p, 0, 0x00000002, "c2"); + orc_program_add_constant (p, 0, 0x00000000, "c3"); + orc_program_add_constant (p, 0, 0x00000000, "c4"); + orc_program_add_constant (p, 0, 0x00000000, "c5"); + orc_program_add_constant (p, 0, 0x00000000, "c6"); + orc_program_add_constant (p, 0, 0x00000000, "c7"); + orc_program_add_constant (p, 0, 0x00000000, "c8"); + orc_program_add_temporary (p, 2, "t1"); + orc_program_add_temporary (p, 2, "t2"); + orc_program_add_temporary (p, 2, "t3"); + + orc_program_append_2 (p, "convubw", 0, ORC_VAR_T1, ORC_VAR_S1, ORC_VAR_D1, + ORC_VAR_D1); + orc_program_append_2 (p, "convubw", 0, ORC_VAR_T2, ORC_VAR_S2, ORC_VAR_D1, + ORC_VAR_D1); + orc_program_append_2 (p, "convubw", 0, ORC_VAR_T3, ORC_VAR_S3, ORC_VAR_D1, + ORC_VAR_D1); + orc_program_append_2 (p, "mullw", 0, ORC_VAR_T2, ORC_VAR_T2, ORC_VAR_C1, + ORC_VAR_D1); + orc_program_append_2 (p, "addw", 0, ORC_VAR_T1, ORC_VAR_T1, ORC_VAR_T3, + ORC_VAR_D1); + orc_program_append_2 (p, "addw", 0, ORC_VAR_T1, ORC_VAR_T1, ORC_VAR_T2, + ORC_VAR_D1); + orc_program_append_2 (p, "addw", 0, ORC_VAR_T1, ORC_VAR_T1, ORC_VAR_C1, + ORC_VAR_D1); + orc_program_append_2 (p, "shrsw", 0, ORC_VAR_T1, ORC_VAR_T1, ORC_VAR_C1, + ORC_VAR_D1); + orc_program_append_2 (p, "convsuswb", 0, ORC_VAR_D1, ORC_VAR_T1, + ORC_VAR_D1, ORC_VAR_D1); + + result = orc_program_compile (p); + } + p_inited = TRUE; + orc_once_mutex_unlock (); + } + ex->program = p; + + ex->n = n; + ex->arrays[ORC_VAR_D1] = d1; + ex->arrays[ORC_VAR_S1] = (void *) s1; + ex->arrays[ORC_VAR_S2] = (void *) s2; + ex->arrays[ORC_VAR_S3] = (void *) s3; + + func = p->code_exec; + func (ex); +} +#endif + + +/* cogorc_downsample_vert_halfsite_4tap */ +#ifdef DISABLE_ORC +void +cogorc_downsample_vert_halfsite_4tap (guint8 * d1, const guint8 * s1, + const guint8 * s2, const guint8 * s3, const guint8 * s4, int n) +{ + int i; + orc_int8 *ORC_RESTRICT ptr0; + const orc_int8 *ORC_RESTRICT ptr4; + const orc_int8 *ORC_RESTRICT ptr5; + const orc_int8 *ORC_RESTRICT ptr6; + const orc_int8 *ORC_RESTRICT ptr7; + orc_int8 var36; + orc_int8 var37; + orc_int8 var38; + orc_int8 var39; + orc_union16 var40; + orc_union16 var41; + orc_union16 var42; + orc_int8 var43; + orc_union16 var44; + orc_union16 var45; + orc_union16 var46; + orc_union16 var47; + orc_union16 var48; + orc_union16 var49; + orc_union16 var50; + orc_union16 var51; + orc_union16 var52; + orc_union16 var53; + orc_union16 var54; + + ptr0 = (orc_int8 *) d1; + ptr4 = (orc_int8 *) s1; + ptr5 = (orc_int8 *) s2; + ptr6 = (orc_int8 *) s3; + ptr7 = (orc_int8 *) s4; + + /* 9: loadpw */ + var40.i = 0x0000001a; /* 26 or 1.28457e-322f */ + /* 12: loadpw */ + var41.i = 0x00000006; /* 6 or 2.96439e-323f */ + /* 15: loadpw */ + var42.i = 0x00000020; /* 32 or 1.58101e-322f */ + + for (i = 0; i < n; i++) { + /* 0: loadb */ + var36 = ptr4[i]; + /* 1: convubw */ + var44.i = (orc_uint8) var36; + /* 2: loadb */ + var37 = ptr5[i]; + /* 3: convubw */ + var45.i = (orc_uint8) var37; + /* 4: loadb */ + var38 = ptr6[i]; + /* 5: convubw */ + var46.i = (orc_uint8) var38; + /* 6: loadb */ + var39 = ptr7[i]; + /* 7: convubw */ + var47.i = (orc_uint8) var39; + /* 8: addw */ + var48.i = var45.i + var46.i; + /* 10: mullw */ + var49.i = (var48.i * var40.i) & 0xffff; + /* 11: addw */ + var50.i = var44.i + var47.i; + /* 13: mullw */ + var51.i = (var50.i * var41.i) & 0xffff; + /* 14: addw */ + var52.i = var49.i + var51.i; + /* 16: addw */ + var53.i = var52.i + var42.i; + /* 17: shrsw */ + var54.i = var53.i >> 6; + /* 18: convsuswb */ + var43 = ORC_CLAMP_UB (var54.i); + /* 19: storeb */ + ptr0[i] = var43; + } + +} + +#else +static void +_backup_cogorc_downsample_vert_halfsite_4tap (OrcExecutor * ex) +{ + int i; + int n = ex->n; + orc_int8 *ORC_RESTRICT ptr0; + const orc_int8 *ORC_RESTRICT ptr4; + const orc_int8 *ORC_RESTRICT ptr5; + const orc_int8 *ORC_RESTRICT ptr6; + const orc_int8 *ORC_RESTRICT ptr7; + orc_int8 var36; + orc_int8 var37; + orc_int8 var38; + orc_int8 var39; + orc_union16 var40; + orc_union16 var41; + orc_union16 var42; + orc_int8 var43; + orc_union16 var44; + orc_union16 var45; + orc_union16 var46; + orc_union16 var47; + orc_union16 var48; + orc_union16 var49; + orc_union16 var50; + orc_union16 var51; + orc_union16 var52; + orc_union16 var53; + orc_union16 var54; + + ptr0 = (orc_int8 *) ex->arrays[0]; + ptr4 = (orc_int8 *) ex->arrays[4]; + ptr5 = (orc_int8 *) ex->arrays[5]; + ptr6 = (orc_int8 *) ex->arrays[6]; + ptr7 = (orc_int8 *) ex->arrays[7]; + + /* 9: loadpw */ + var40.i = 0x0000001a; /* 26 or 1.28457e-322f */ + /* 12: loadpw */ + var41.i = 0x00000006; /* 6 or 2.96439e-323f */ + /* 15: loadpw */ + var42.i = 0x00000020; /* 32 or 1.58101e-322f */ + + for (i = 0; i < n; i++) { + /* 0: loadb */ + var36 = ptr4[i]; + /* 1: convubw */ + var44.i = (orc_uint8) var36; + /* 2: loadb */ + var37 = ptr5[i]; + /* 3: convubw */ + var45.i = (orc_uint8) var37; + /* 4: loadb */ + var38 = ptr6[i]; + /* 5: convubw */ + var46.i = (orc_uint8) var38; + /* 6: loadb */ + var39 = ptr7[i]; + /* 7: convubw */ + var47.i = (orc_uint8) var39; + /* 8: addw */ + var48.i = var45.i + var46.i; + /* 10: mullw */ + var49.i = (var48.i * var40.i) & 0xffff; + /* 11: addw */ + var50.i = var44.i + var47.i; + /* 13: mullw */ + var51.i = (var50.i * var41.i) & 0xffff; + /* 14: addw */ + var52.i = var49.i + var51.i; + /* 16: addw */ + var53.i = var52.i + var42.i; + /* 17: shrsw */ + var54.i = var53.i >> 6; + /* 18: convsuswb */ + var43 = ORC_CLAMP_UB (var54.i); + /* 19: storeb */ + ptr0[i] = var43; + } + +} + +void +cogorc_downsample_vert_halfsite_4tap (guint8 * d1, const guint8 * s1, + const guint8 * s2, const guint8 * s3, const guint8 * s4, int n) +{ + OrcExecutor _ex, *ex = &_ex; + static int p_inited = 0; + static OrcProgram *p = 0; + void (*func) (OrcExecutor *); + + if (!p_inited) { + orc_once_mutex_lock (); + if (!p_inited) { + OrcCompileResult result; + + p = orc_program_new (); + orc_program_set_name (p, "cogorc_downsample_vert_halfsite_4tap"); + orc_program_set_backup_function (p, + _backup_cogorc_downsample_vert_halfsite_4tap); + orc_program_add_destination (p, 1, "d1"); + orc_program_add_source (p, 1, "s1"); + orc_program_add_source (p, 1, "s2"); + orc_program_add_source (p, 1, "s3"); + orc_program_add_source (p, 1, "s4"); + orc_program_add_constant (p, 4, 0x0000001a, "c1"); + orc_program_add_constant (p, 4, 0x00000006, "c2"); + orc_program_add_constant (p, 4, 0x00000020, "c3"); + orc_program_add_constant (p, 0, 0x00000006, "c4"); + orc_program_add_constant (p, 0, 0x00000000, "c5"); + orc_program_add_constant (p, 0, 0x00000000, "c6"); + orc_program_add_constant (p, 0, 0x00000000, "c7"); + orc_program_add_constant (p, 0, 0x00000000, "c8"); + orc_program_add_temporary (p, 2, "t1"); + orc_program_add_temporary (p, 2, "t2"); + orc_program_add_temporary (p, 2, "t3"); + orc_program_add_temporary (p, 2, "t4"); + + orc_program_append_2 (p, "convubw", 0, ORC_VAR_T1, ORC_VAR_S1, ORC_VAR_D1, + ORC_VAR_D1); + orc_program_append_2 (p, "convubw", 0, ORC_VAR_T2, ORC_VAR_S2, ORC_VAR_D1, + ORC_VAR_D1); + orc_program_append_2 (p, "convubw", 0, ORC_VAR_T3, ORC_VAR_S3, ORC_VAR_D1, + ORC_VAR_D1); + orc_program_append_2 (p, "convubw", 0, ORC_VAR_T4, ORC_VAR_S4, ORC_VAR_D1, + ORC_VAR_D1); + orc_program_append_2 (p, "addw", 0, ORC_VAR_T2, ORC_VAR_T2, ORC_VAR_T3, + ORC_VAR_D1); + orc_program_append_2 (p, "mullw", 0, ORC_VAR_T2, ORC_VAR_T2, ORC_VAR_C1, + ORC_VAR_D1); + orc_program_append_2 (p, "addw", 0, ORC_VAR_T1, ORC_VAR_T1, ORC_VAR_T4, + ORC_VAR_D1); + orc_program_append_2 (p, "mullw", 0, ORC_VAR_T1, ORC_VAR_T1, ORC_VAR_C2, + ORC_VAR_D1); + orc_program_append_2 (p, "addw", 0, ORC_VAR_T2, ORC_VAR_T2, ORC_VAR_T1, + ORC_VAR_D1); + orc_program_append_2 (p, "addw", 0, ORC_VAR_T2, ORC_VAR_T2, ORC_VAR_C3, + ORC_VAR_D1); + orc_program_append_2 (p, "shrsw", 0, ORC_VAR_T2, ORC_VAR_T2, ORC_VAR_C2, + ORC_VAR_D1); + orc_program_append_2 (p, "convsuswb", 0, ORC_VAR_D1, ORC_VAR_T2, + ORC_VAR_D1, ORC_VAR_D1); + + result = orc_program_compile (p); + } + p_inited = TRUE; + orc_once_mutex_unlock (); + } + ex->program = p; + + ex->n = n; + ex->arrays[ORC_VAR_D1] = d1; + ex->arrays[ORC_VAR_S1] = (void *) s1; + ex->arrays[ORC_VAR_S2] = (void *) s2; + ex->arrays[ORC_VAR_S3] = (void *) s3; + ex->arrays[ORC_VAR_S4] = (void *) s4; + + func = p->code_exec; + func (ex); +} +#endif + + +/* cogorc_upsample_horiz_cosite_1tap */ +#ifdef DISABLE_ORC +void +cogorc_upsample_horiz_cosite_1tap (guint8 * d1, const guint8 * s1, int n) +{ + int i; + orc_union16 *ORC_RESTRICT ptr0; + const orc_int8 *ORC_RESTRICT ptr4; + orc_int8 var33; + orc_union16 var34; + orc_int8 var35; + + ptr0 = (orc_union16 *) d1; + ptr4 = (orc_int8 *) s1; + + + for (i = 0; i < n; i++) { + /* 0: loadb */ + var33 = ptr4[i]; + /* 1: copyb */ + var35 = var33; + /* 2: mergebw */ + var34.i = ((orc_uint8) var35) | ((orc_uint8) var35 << 8); + /* 3: storew */ + ptr0[i] = var34; + } + +} + +#else +static void +_backup_cogorc_upsample_horiz_cosite_1tap (OrcExecutor * ex) +{ + int i; + int n = ex->n; + orc_union16 *ORC_RESTRICT ptr0; + const orc_int8 *ORC_RESTRICT ptr4; + orc_int8 var33; + orc_union16 var34; + orc_int8 var35; + + ptr0 = (orc_union16 *) ex->arrays[0]; + ptr4 = (orc_int8 *) ex->arrays[4]; + + + for (i = 0; i < n; i++) { + /* 0: loadb */ + var33 = ptr4[i]; + /* 1: copyb */ + var35 = var33; + /* 2: mergebw */ + var34.i = ((orc_uint8) var35) | ((orc_uint8) var35 << 8); + /* 3: storew */ + ptr0[i] = var34; + } + +} + +void +cogorc_upsample_horiz_cosite_1tap (guint8 * d1, const guint8 * s1, int n) +{ + OrcExecutor _ex, *ex = &_ex; + static int p_inited = 0; + static OrcProgram *p = 0; + void (*func) (OrcExecutor *); + + if (!p_inited) { + orc_once_mutex_lock (); + if (!p_inited) { + OrcCompileResult result; + + p = orc_program_new (); + orc_program_set_name (p, "cogorc_upsample_horiz_cosite_1tap"); + orc_program_set_backup_function (p, + _backup_cogorc_upsample_horiz_cosite_1tap); + orc_program_add_destination (p, 2, "d1"); + orc_program_add_source (p, 1, "s1"); + orc_program_add_constant (p, 0, 0x00000000, "c1"); + orc_program_add_constant (p, 0, 0x00000000, "c2"); + orc_program_add_constant (p, 0, 0x00000000, "c3"); + orc_program_add_constant (p, 0, 0x00000000, "c4"); + orc_program_add_constant (p, 0, 0x00000000, "c5"); + orc_program_add_constant (p, 0, 0x00000000, "c6"); + orc_program_add_constant (p, 0, 0x00000000, "c7"); + orc_program_add_constant (p, 0, 0x00000000, "c8"); + orc_program_add_temporary (p, 1, "t1"); + + orc_program_append_2 (p, "copyb", 0, ORC_VAR_T1, ORC_VAR_S1, ORC_VAR_D1, + ORC_VAR_D1); + orc_program_append_2 (p, "mergebw", 0, ORC_VAR_D1, ORC_VAR_T1, ORC_VAR_T1, + ORC_VAR_D1); + + result = orc_program_compile (p); + } + p_inited = TRUE; + orc_once_mutex_unlock (); + } + ex->program = p; + + ex->n = n; + ex->arrays[ORC_VAR_D1] = d1; + ex->arrays[ORC_VAR_S1] = (void *) s1; + + func = p->code_exec; + func (ex); +} +#endif + + +/* cogorc_upsample_horiz_cosite */ +#ifdef DISABLE_ORC +void +cogorc_upsample_horiz_cosite (guint8 * d1, const guint8 * s1, const guint8 * s2, + int n) +{ + int i; + orc_union16 *ORC_RESTRICT ptr0; + const orc_int8 *ORC_RESTRICT ptr4; + const orc_int8 *ORC_RESTRICT ptr5; + orc_int8 var34; + orc_int8 var35; + orc_union16 var36; + orc_int8 var37; + orc_int8 var38; + + ptr0 = (orc_union16 *) d1; + ptr4 = (orc_int8 *) s1; + ptr5 = (orc_int8 *) s2; + + + for (i = 0; i < n; i++) { + /* 0: loadb */ + var34 = ptr4[i]; + /* 1: copyb */ + var37 = var34; + /* 2: loadb */ + var35 = ptr5[i]; + /* 3: avgub */ + var38 = ((orc_uint8) var37 + (orc_uint8) var35 + 1) >> 1; + /* 4: mergebw */ + var36.i = ((orc_uint8) var37) | ((orc_uint8) var38 << 8); + /* 5: storew */ + ptr0[i] = var36; + } + +} + +#else +static void +_backup_cogorc_upsample_horiz_cosite (OrcExecutor * ex) +{ + int i; + int n = ex->n; + orc_union16 *ORC_RESTRICT ptr0; + const orc_int8 *ORC_RESTRICT ptr4; + const orc_int8 *ORC_RESTRICT ptr5; + orc_int8 var34; + orc_int8 var35; + orc_union16 var36; + orc_int8 var37; + orc_int8 var38; + + ptr0 = (orc_union16 *) ex->arrays[0]; + ptr4 = (orc_int8 *) ex->arrays[4]; + ptr5 = (orc_int8 *) ex->arrays[5]; + + + for (i = 0; i < n; i++) { + /* 0: loadb */ + var34 = ptr4[i]; + /* 1: copyb */ + var37 = var34; + /* 2: loadb */ + var35 = ptr5[i]; + /* 3: avgub */ + var38 = ((orc_uint8) var37 + (orc_uint8) var35 + 1) >> 1; + /* 4: mergebw */ + var36.i = ((orc_uint8) var37) | ((orc_uint8) var38 << 8); + /* 5: storew */ + ptr0[i] = var36; + } + +} + +void +cogorc_upsample_horiz_cosite (guint8 * d1, const guint8 * s1, const guint8 * s2, + int n) +{ + OrcExecutor _ex, *ex = &_ex; + static int p_inited = 0; + static OrcProgram *p = 0; + void (*func) (OrcExecutor *); + + if (!p_inited) { + orc_once_mutex_lock (); + if (!p_inited) { + OrcCompileResult result; + + p = orc_program_new (); + orc_program_set_name (p, "cogorc_upsample_horiz_cosite"); + orc_program_set_backup_function (p, _backup_cogorc_upsample_horiz_cosite); + orc_program_add_destination (p, 2, "d1"); + orc_program_add_source (p, 1, "s1"); + orc_program_add_source (p, 1, "s2"); + orc_program_add_constant (p, 0, 0x00000000, "c1"); + orc_program_add_constant (p, 0, 0x00000000, "c2"); + orc_program_add_constant (p, 0, 0x00000000, "c3"); + orc_program_add_constant (p, 0, 0x00000000, "c4"); + orc_program_add_constant (p, 0, 0x00000000, "c5"); + orc_program_add_constant (p, 0, 0x00000000, "c6"); + orc_program_add_constant (p, 0, 0x00000000, "c7"); + orc_program_add_constant (p, 0, 0x00000000, "c8"); + orc_program_add_temporary (p, 1, "t1"); + orc_program_add_temporary (p, 1, "t2"); + + orc_program_append_2 (p, "copyb", 0, ORC_VAR_T1, ORC_VAR_S1, ORC_VAR_D1, + ORC_VAR_D1); + orc_program_append_2 (p, "avgub", 0, ORC_VAR_T2, ORC_VAR_T1, ORC_VAR_S2, + ORC_VAR_D1); + orc_program_append_2 (p, "mergebw", 0, ORC_VAR_D1, ORC_VAR_T1, ORC_VAR_T2, + ORC_VAR_D1); + + result = orc_program_compile (p); + } + p_inited = TRUE; + orc_once_mutex_unlock (); + } + ex->program = p; + + ex->n = n; + ex->arrays[ORC_VAR_D1] = d1; + ex->arrays[ORC_VAR_S1] = (void *) s1; + ex->arrays[ORC_VAR_S2] = (void *) s2; + + func = p->code_exec; + func (ex); +} +#endif + + +/* cogorc_upsample_vert_avgub */ +#ifdef DISABLE_ORC +void +cogorc_upsample_vert_avgub (guint8 * d1, const guint8 * s1, const guint8 * s2, + int n) +{ + int i; + orc_int8 *ORC_RESTRICT ptr0; + const orc_int8 *ORC_RESTRICT ptr4; + const orc_int8 *ORC_RESTRICT ptr5; + orc_int8 var32; + orc_int8 var33; + orc_int8 var34; + + ptr0 = (orc_int8 *) d1; + ptr4 = (orc_int8 *) s1; + ptr5 = (orc_int8 *) s2; + + + for (i = 0; i < n; i++) { + /* 0: loadb */ + var32 = ptr4[i]; + /* 1: loadb */ + var33 = ptr5[i]; + /* 2: avgub */ + var34 = ((orc_uint8) var32 + (orc_uint8) var33 + 1) >> 1; + /* 3: storeb */ + ptr0[i] = var34; + } + +} + +#else +static void +_backup_cogorc_upsample_vert_avgub (OrcExecutor * ex) +{ + int i; + int n = ex->n; + orc_int8 *ORC_RESTRICT ptr0; + const orc_int8 *ORC_RESTRICT ptr4; + const orc_int8 *ORC_RESTRICT ptr5; + orc_int8 var32; + orc_int8 var33; + orc_int8 var34; + + ptr0 = (orc_int8 *) ex->arrays[0]; + ptr4 = (orc_int8 *) ex->arrays[4]; + ptr5 = (orc_int8 *) ex->arrays[5]; + + + for (i = 0; i < n; i++) { + /* 0: loadb */ + var32 = ptr4[i]; + /* 1: loadb */ + var33 = ptr5[i]; + /* 2: avgub */ + var34 = ((orc_uint8) var32 + (orc_uint8) var33 + 1) >> 1; + /* 3: storeb */ + ptr0[i] = var34; + } + +} + +void +cogorc_upsample_vert_avgub (guint8 * d1, const guint8 * s1, const guint8 * s2, + int n) +{ + OrcExecutor _ex, *ex = &_ex; + static int p_inited = 0; + static OrcProgram *p = 0; + void (*func) (OrcExecutor *); + + if (!p_inited) { + orc_once_mutex_lock (); + if (!p_inited) { + OrcCompileResult result; + + p = orc_program_new (); + orc_program_set_name (p, "cogorc_upsample_vert_avgub"); + orc_program_set_backup_function (p, _backup_cogorc_upsample_vert_avgub); + orc_program_add_destination (p, 1, "d1"); + orc_program_add_source (p, 1, "s1"); + orc_program_add_source (p, 1, "s2"); + orc_program_add_constant (p, 0, 0x00000000, "c1"); + orc_program_add_constant (p, 0, 0x00000000, "c2"); + orc_program_add_constant (p, 0, 0x00000000, "c3"); + orc_program_add_constant (p, 0, 0x00000000, "c4"); + orc_program_add_constant (p, 0, 0x00000000, "c5"); + orc_program_add_constant (p, 0, 0x00000000, "c6"); + orc_program_add_constant (p, 0, 0x00000000, "c7"); + orc_program_add_constant (p, 0, 0x00000000, "c8"); + + orc_program_append_2 (p, "avgub", 0, ORC_VAR_D1, ORC_VAR_S1, ORC_VAR_S2, + ORC_VAR_D1); + + result = orc_program_compile (p); + } + p_inited = TRUE; + orc_once_mutex_unlock (); + } + ex->program = p; + + ex->n = n; + ex->arrays[ORC_VAR_D1] = d1; + ex->arrays[ORC_VAR_S1] = (void *) s1; + ex->arrays[ORC_VAR_S2] = (void *) s2; + + func = p->code_exec; + func (ex); +} +#endif + + +/* orc_unpack_yuyv_y */ +#ifdef DISABLE_ORC +void +orc_unpack_yuyv_y (guint8 * d1, const guint8 * s1, int n) +{ + int i; + orc_int8 *ORC_RESTRICT ptr0; + const orc_union16 *ORC_RESTRICT ptr4; + orc_union16 var32; + orc_int8 var33; + + ptr0 = (orc_int8 *) d1; + ptr4 = (orc_union16 *) s1; + + + for (i = 0; i < n; i++) { + /* 0: loadw */ + var32 = ptr4[i]; + /* 1: select0wb */ + var33 = (orc_uint16) var32.i & 0xff; + /* 2: storeb */ + ptr0[i] = var33; + } + +} + +#else +static void +_backup_orc_unpack_yuyv_y (OrcExecutor * ex) +{ + int i; + int n = ex->n; + orc_int8 *ORC_RESTRICT ptr0; + const orc_union16 *ORC_RESTRICT ptr4; + orc_union16 var32; + orc_int8 var33; + + ptr0 = (orc_int8 *) ex->arrays[0]; + ptr4 = (orc_union16 *) ex->arrays[4]; + + + for (i = 0; i < n; i++) { + /* 0: loadw */ + var32 = ptr4[i]; + /* 1: select0wb */ + var33 = (orc_uint16) var32.i & 0xff; + /* 2: storeb */ + ptr0[i] = var33; + } + +} + +void +orc_unpack_yuyv_y (guint8 * d1, const guint8 * s1, int n) +{ + OrcExecutor _ex, *ex = &_ex; + static int p_inited = 0; + static OrcProgram *p = 0; + void (*func) (OrcExecutor *); + + if (!p_inited) { + orc_once_mutex_lock (); + if (!p_inited) { + OrcCompileResult result; + + p = orc_program_new (); + orc_program_set_name (p, "orc_unpack_yuyv_y"); + orc_program_set_backup_function (p, _backup_orc_unpack_yuyv_y); + orc_program_add_destination (p, 1, "d1"); + orc_program_add_source (p, 2, "s1"); + orc_program_add_constant (p, 0, 0x00000000, "c1"); + orc_program_add_constant (p, 0, 0x00000000, "c2"); + orc_program_add_constant (p, 0, 0x00000000, "c3"); + orc_program_add_constant (p, 0, 0x00000000, "c4"); + orc_program_add_constant (p, 0, 0x00000000, "c5"); + orc_program_add_constant (p, 0, 0x00000000, "c6"); + orc_program_add_constant (p, 0, 0x00000000, "c7"); + orc_program_add_constant (p, 0, 0x00000000, "c8"); + + orc_program_append_2 (p, "select0wb", 0, ORC_VAR_D1, ORC_VAR_S1, + ORC_VAR_D1, ORC_VAR_D1); + + result = orc_program_compile (p); + } + p_inited = TRUE; + orc_once_mutex_unlock (); + } + ex->program = p; + + ex->n = n; + ex->arrays[ORC_VAR_D1] = d1; + ex->arrays[ORC_VAR_S1] = (void *) s1; + + func = p->code_exec; + func (ex); +} +#endif + + +/* orc_unpack_yuyv_u */ +#ifdef DISABLE_ORC +void +orc_unpack_yuyv_u (guint8 * d1, const guint8 * s1, int n) +{ + int i; + orc_int8 *ORC_RESTRICT ptr0; + const orc_union32 *ORC_RESTRICT ptr4; + orc_union32 var33; + orc_int8 var34; + orc_union16 var35; + + ptr0 = (orc_int8 *) d1; + ptr4 = (orc_union32 *) s1; + + + for (i = 0; i < n; i++) { + /* 0: loadl */ + var33 = ptr4[i]; + /* 1: select0lw */ + var35.i = (orc_uint32) var33.i & 0xffff; + /* 2: select1wb */ + var34 = ((orc_uint16) var35.i >> 8) & 0xff; + /* 3: storeb */ + ptr0[i] = var34; + } + +} + +#else +static void +_backup_orc_unpack_yuyv_u (OrcExecutor * ex) +{ + int i; + int n = ex->n; + orc_int8 *ORC_RESTRICT ptr0; + const orc_union32 *ORC_RESTRICT ptr4; + orc_union32 var33; + orc_int8 var34; + orc_union16 var35; + + ptr0 = (orc_int8 *) ex->arrays[0]; + ptr4 = (orc_union32 *) ex->arrays[4]; + + + for (i = 0; i < n; i++) { + /* 0: loadl */ + var33 = ptr4[i]; + /* 1: select0lw */ + var35.i = (orc_uint32) var33.i & 0xffff; + /* 2: select1wb */ + var34 = ((orc_uint16) var35.i >> 8) & 0xff; + /* 3: storeb */ + ptr0[i] = var34; + } + +} + +void +orc_unpack_yuyv_u (guint8 * d1, const guint8 * s1, int n) +{ + OrcExecutor _ex, *ex = &_ex; + static int p_inited = 0; + static OrcProgram *p = 0; + void (*func) (OrcExecutor *); + + if (!p_inited) { + orc_once_mutex_lock (); + if (!p_inited) { + OrcCompileResult result; + + p = orc_program_new (); + orc_program_set_name (p, "orc_unpack_yuyv_u"); + orc_program_set_backup_function (p, _backup_orc_unpack_yuyv_u); + orc_program_add_destination (p, 1, "d1"); + orc_program_add_source (p, 4, "s1"); + orc_program_add_constant (p, 0, 0x00000000, "c1"); + orc_program_add_constant (p, 0, 0x00000000, "c2"); + orc_program_add_constant (p, 0, 0x00000000, "c3"); + orc_program_add_constant (p, 0, 0x00000000, "c4"); + orc_program_add_constant (p, 0, 0x00000000, "c5"); + orc_program_add_constant (p, 0, 0x00000000, "c6"); + orc_program_add_constant (p, 0, 0x00000000, "c7"); + orc_program_add_constant (p, 0, 0x00000000, "c8"); + orc_program_add_temporary (p, 2, "t1"); + + orc_program_append_2 (p, "select0lw", 0, ORC_VAR_T1, ORC_VAR_S1, + ORC_VAR_D1, ORC_VAR_D1); + orc_program_append_2 (p, "select1wb", 0, ORC_VAR_D1, ORC_VAR_T1, + ORC_VAR_D1, ORC_VAR_D1); + + result = orc_program_compile (p); + } + p_inited = TRUE; + orc_once_mutex_unlock (); + } + ex->program = p; + + ex->n = n; + ex->arrays[ORC_VAR_D1] = d1; + ex->arrays[ORC_VAR_S1] = (void *) s1; + + func = p->code_exec; + func (ex); +} +#endif + + +/* orc_unpack_yuyv_v */ +#ifdef DISABLE_ORC +void +orc_unpack_yuyv_v (guint8 * d1, const guint8 * s1, int n) +{ + int i; + orc_int8 *ORC_RESTRICT ptr0; + const orc_union32 *ORC_RESTRICT ptr4; + orc_union32 var33; + orc_int8 var34; + orc_union16 var35; + + ptr0 = (orc_int8 *) d1; + ptr4 = (orc_union32 *) s1; + + + for (i = 0; i < n; i++) { + /* 0: loadl */ + var33 = ptr4[i]; + /* 1: select1lw */ + var35.i = ((orc_uint32) var33.i >> 16) & 0xffff; + /* 2: select1wb */ + var34 = ((orc_uint16) var35.i >> 8) & 0xff; + /* 3: storeb */ + ptr0[i] = var34; + } + +} + +#else +static void +_backup_orc_unpack_yuyv_v (OrcExecutor * ex) +{ + int i; + int n = ex->n; + orc_int8 *ORC_RESTRICT ptr0; + const orc_union32 *ORC_RESTRICT ptr4; + orc_union32 var33; + orc_int8 var34; + orc_union16 var35; + + ptr0 = (orc_int8 *) ex->arrays[0]; + ptr4 = (orc_union32 *) ex->arrays[4]; + + + for (i = 0; i < n; i++) { + /* 0: loadl */ + var33 = ptr4[i]; + /* 1: select1lw */ + var35.i = ((orc_uint32) var33.i >> 16) & 0xffff; + /* 2: select1wb */ + var34 = ((orc_uint16) var35.i >> 8) & 0xff; + /* 3: storeb */ + ptr0[i] = var34; + } + +} + +void +orc_unpack_yuyv_v (guint8 * d1, const guint8 * s1, int n) +{ + OrcExecutor _ex, *ex = &_ex; + static int p_inited = 0; + static OrcProgram *p = 0; + void (*func) (OrcExecutor *); + + if (!p_inited) { + orc_once_mutex_lock (); + if (!p_inited) { + OrcCompileResult result; + + p = orc_program_new (); + orc_program_set_name (p, "orc_unpack_yuyv_v"); + orc_program_set_backup_function (p, _backup_orc_unpack_yuyv_v); + orc_program_add_destination (p, 1, "d1"); + orc_program_add_source (p, 4, "s1"); + orc_program_add_constant (p, 0, 0x00000000, "c1"); + orc_program_add_constant (p, 0, 0x00000000, "c2"); + orc_program_add_constant (p, 0, 0x00000000, "c3"); + orc_program_add_constant (p, 0, 0x00000000, "c4"); + orc_program_add_constant (p, 0, 0x00000000, "c5"); + orc_program_add_constant (p, 0, 0x00000000, "c6"); + orc_program_add_constant (p, 0, 0x00000000, "c7"); + orc_program_add_constant (p, 0, 0x00000000, "c8"); + orc_program_add_temporary (p, 2, "t1"); + + orc_program_append_2 (p, "select1lw", 0, ORC_VAR_T1, ORC_VAR_S1, + ORC_VAR_D1, ORC_VAR_D1); + orc_program_append_2 (p, "select1wb", 0, ORC_VAR_D1, ORC_VAR_T1, + ORC_VAR_D1, ORC_VAR_D1); + + result = orc_program_compile (p); + } + p_inited = TRUE; + orc_once_mutex_unlock (); + } + ex->program = p; + + ex->n = n; + ex->arrays[ORC_VAR_D1] = d1; + ex->arrays[ORC_VAR_S1] = (void *) s1; + + func = p->code_exec; + func (ex); +} +#endif + + +/* orc_pack_yuyv */ +#ifdef DISABLE_ORC +void +orc_pack_yuyv (guint8 * d1, const guint8 * s1, const guint8 * s2, + const guint8 * s3, int n) +{ + int i; + orc_union32 *ORC_RESTRICT ptr0; + const orc_union16 *ORC_RESTRICT ptr4; + const orc_int8 *ORC_RESTRICT ptr5; + const orc_int8 *ORC_RESTRICT ptr6; + orc_union16 var37; + orc_int8 var38; + orc_int8 var39; + orc_union32 var40; + orc_union16 var41; + orc_int8 var42; + orc_int8 var43; + orc_union16 var44; + orc_union16 var45; + + ptr0 = (orc_union32 *) d1; + ptr4 = (orc_union16 *) s1; + ptr5 = (orc_int8 *) s2; + ptr6 = (orc_int8 *) s3; + + + for (i = 0; i < n; i++) { + /* 0: loadw */ + var37 = ptr4[i]; + /* 1: copyw */ + var41.i = var37.i; + /* 2: select0wb */ + var42 = (orc_uint16) var41.i & 0xff; + /* 3: select1wb */ + var43 = ((orc_uint16) var41.i >> 8) & 0xff; + /* 4: loadb */ + var38 = ptr5[i]; + /* 5: mergebw */ + var44.i = ((orc_uint8) var42) | ((orc_uint8) var38 << 8); + /* 6: loadb */ + var39 = ptr6[i]; + /* 7: mergebw */ + var45.i = ((orc_uint8) var43) | ((orc_uint8) var39 << 8); + /* 8: mergewl */ + var40.i = ((orc_uint16) var44.i) | ((orc_uint16) var45.i << 16); + /* 9: storel */ + ptr0[i] = var40; + } + +} + +#else +static void +_backup_orc_pack_yuyv (OrcExecutor * ex) +{ + int i; + int n = ex->n; + orc_union32 *ORC_RESTRICT ptr0; + const orc_union16 *ORC_RESTRICT ptr4; + const orc_int8 *ORC_RESTRICT ptr5; + const orc_int8 *ORC_RESTRICT ptr6; + orc_union16 var37; + orc_int8 var38; + orc_int8 var39; + orc_union32 var40; + orc_union16 var41; + orc_int8 var42; + orc_int8 var43; + orc_union16 var44; + orc_union16 var45; + + ptr0 = (orc_union32 *) ex->arrays[0]; + ptr4 = (orc_union16 *) ex->arrays[4]; + ptr5 = (orc_int8 *) ex->arrays[5]; + ptr6 = (orc_int8 *) ex->arrays[6]; + + + for (i = 0; i < n; i++) { + /* 0: loadw */ + var37 = ptr4[i]; + /* 1: copyw */ + var41.i = var37.i; + /* 2: select0wb */ + var42 = (orc_uint16) var41.i & 0xff; + /* 3: select1wb */ + var43 = ((orc_uint16) var41.i >> 8) & 0xff; + /* 4: loadb */ + var38 = ptr5[i]; + /* 5: mergebw */ + var44.i = ((orc_uint8) var42) | ((orc_uint8) var38 << 8); + /* 6: loadb */ + var39 = ptr6[i]; + /* 7: mergebw */ + var45.i = ((orc_uint8) var43) | ((orc_uint8) var39 << 8); + /* 8: mergewl */ + var40.i = ((orc_uint16) var44.i) | ((orc_uint16) var45.i << 16); + /* 9: storel */ + ptr0[i] = var40; + } + +} + +void +orc_pack_yuyv (guint8 * d1, const guint8 * s1, const guint8 * s2, + const guint8 * s3, int n) +{ + OrcExecutor _ex, *ex = &_ex; + static int p_inited = 0; + static OrcProgram *p = 0; + void (*func) (OrcExecutor *); + + if (!p_inited) { + orc_once_mutex_lock (); + if (!p_inited) { + OrcCompileResult result; + + p = orc_program_new (); + orc_program_set_name (p, "orc_pack_yuyv"); + orc_program_set_backup_function (p, _backup_orc_pack_yuyv); + orc_program_add_destination (p, 4, "d1"); + orc_program_add_source (p, 2, "s1"); + orc_program_add_source (p, 1, "s2"); + orc_program_add_source (p, 1, "s3"); + orc_program_add_constant (p, 0, 0x00000000, "c1"); + orc_program_add_constant (p, 0, 0x00000000, "c2"); + orc_program_add_constant (p, 0, 0x00000000, "c3"); + orc_program_add_constant (p, 0, 0x00000000, "c4"); + orc_program_add_constant (p, 0, 0x00000000, "c5"); + orc_program_add_constant (p, 0, 0x00000000, "c6"); + orc_program_add_constant (p, 0, 0x00000000, "c7"); + orc_program_add_constant (p, 0, 0x00000000, "c8"); + orc_program_add_temporary (p, 1, "t1"); + orc_program_add_temporary (p, 1, "t2"); + orc_program_add_temporary (p, 2, "t3"); + orc_program_add_temporary (p, 2, "t4"); + orc_program_add_temporary (p, 2, "t5"); + + orc_program_append_2 (p, "copyw", 0, ORC_VAR_T5, ORC_VAR_S1, ORC_VAR_D1, + ORC_VAR_D1); + orc_program_append_2 (p, "select0wb", 0, ORC_VAR_T1, ORC_VAR_T5, + ORC_VAR_D1, ORC_VAR_D1); + orc_program_append_2 (p, "select1wb", 0, ORC_VAR_T2, ORC_VAR_T5, + ORC_VAR_D1, ORC_VAR_D1); + orc_program_append_2 (p, "mergebw", 0, ORC_VAR_T3, ORC_VAR_T1, ORC_VAR_S2, + ORC_VAR_D1); + orc_program_append_2 (p, "mergebw", 0, ORC_VAR_T4, ORC_VAR_T2, ORC_VAR_S3, + ORC_VAR_D1); + orc_program_append_2 (p, "mergewl", 0, ORC_VAR_D1, ORC_VAR_T3, ORC_VAR_T4, + ORC_VAR_D1); + + result = orc_program_compile (p); + } + p_inited = TRUE; + orc_once_mutex_unlock (); + } + ex->program = p; + + ex->n = n; + ex->arrays[ORC_VAR_D1] = d1; + ex->arrays[ORC_VAR_S1] = (void *) s1; + ex->arrays[ORC_VAR_S2] = (void *) s2; + ex->arrays[ORC_VAR_S3] = (void *) s3; + + func = p->code_exec; + func (ex); +} +#endif + + +/* orc_unpack_uyvy_y */ +#ifdef DISABLE_ORC +void +orc_unpack_uyvy_y (guint8 * d1, const guint8 * s1, int n) +{ + int i; + orc_int8 *ORC_RESTRICT ptr0; + const orc_union16 *ORC_RESTRICT ptr4; + orc_union16 var32; + orc_int8 var33; + + ptr0 = (orc_int8 *) d1; + ptr4 = (orc_union16 *) s1; + + + for (i = 0; i < n; i++) { + /* 0: loadw */ + var32 = ptr4[i]; + /* 1: select1wb */ + var33 = ((orc_uint16) var32.i >> 8) & 0xff; + /* 2: storeb */ + ptr0[i] = var33; + } + +} + +#else +static void +_backup_orc_unpack_uyvy_y (OrcExecutor * ex) +{ + int i; + int n = ex->n; + orc_int8 *ORC_RESTRICT ptr0; + const orc_union16 *ORC_RESTRICT ptr4; + orc_union16 var32; + orc_int8 var33; + + ptr0 = (orc_int8 *) ex->arrays[0]; + ptr4 = (orc_union16 *) ex->arrays[4]; + + + for (i = 0; i < n; i++) { + /* 0: loadw */ + var32 = ptr4[i]; + /* 1: select1wb */ + var33 = ((orc_uint16) var32.i >> 8) & 0xff; + /* 2: storeb */ + ptr0[i] = var33; + } + +} + +void +orc_unpack_uyvy_y (guint8 * d1, const guint8 * s1, int n) +{ + OrcExecutor _ex, *ex = &_ex; + static int p_inited = 0; + static OrcProgram *p = 0; + void (*func) (OrcExecutor *); + + if (!p_inited) { + orc_once_mutex_lock (); + if (!p_inited) { + OrcCompileResult result; + + p = orc_program_new (); + orc_program_set_name (p, "orc_unpack_uyvy_y"); + orc_program_set_backup_function (p, _backup_orc_unpack_uyvy_y); + orc_program_add_destination (p, 1, "d1"); + orc_program_add_source (p, 2, "s1"); + orc_program_add_constant (p, 0, 0x00000000, "c1"); + orc_program_add_constant (p, 0, 0x00000000, "c2"); + orc_program_add_constant (p, 0, 0x00000000, "c3"); + orc_program_add_constant (p, 0, 0x00000000, "c4"); + orc_program_add_constant (p, 0, 0x00000000, "c5"); + orc_program_add_constant (p, 0, 0x00000000, "c6"); + orc_program_add_constant (p, 0, 0x00000000, "c7"); + orc_program_add_constant (p, 0, 0x00000000, "c8"); + + orc_program_append_2 (p, "select1wb", 0, ORC_VAR_D1, ORC_VAR_S1, + ORC_VAR_D1, ORC_VAR_D1); + + result = orc_program_compile (p); + } + p_inited = TRUE; + orc_once_mutex_unlock (); + } + ex->program = p; + + ex->n = n; + ex->arrays[ORC_VAR_D1] = d1; + ex->arrays[ORC_VAR_S1] = (void *) s1; + + func = p->code_exec; + func (ex); +} +#endif + + +/* orc_unpack_uyvy_u */ +#ifdef DISABLE_ORC +void +orc_unpack_uyvy_u (guint8 * d1, const guint8 * s1, int n) +{ + int i; + orc_int8 *ORC_RESTRICT ptr0; + const orc_union32 *ORC_RESTRICT ptr4; + orc_union32 var33; + orc_int8 var34; + orc_union16 var35; + + ptr0 = (orc_int8 *) d1; + ptr4 = (orc_union32 *) s1; + + + for (i = 0; i < n; i++) { + /* 0: loadl */ + var33 = ptr4[i]; + /* 1: select0lw */ + var35.i = (orc_uint32) var33.i & 0xffff; + /* 2: select0wb */ + var34 = (orc_uint16) var35.i & 0xff; + /* 3: storeb */ + ptr0[i] = var34; + } + +} + +#else +static void +_backup_orc_unpack_uyvy_u (OrcExecutor * ex) +{ + int i; + int n = ex->n; + orc_int8 *ORC_RESTRICT ptr0; + const orc_union32 *ORC_RESTRICT ptr4; + orc_union32 var33; + orc_int8 var34; + orc_union16 var35; + + ptr0 = (orc_int8 *) ex->arrays[0]; + ptr4 = (orc_union32 *) ex->arrays[4]; + + + for (i = 0; i < n; i++) { + /* 0: loadl */ + var33 = ptr4[i]; + /* 1: select0lw */ + var35.i = (orc_uint32) var33.i & 0xffff; + /* 2: select0wb */ + var34 = (orc_uint16) var35.i & 0xff; + /* 3: storeb */ + ptr0[i] = var34; + } + +} + +void +orc_unpack_uyvy_u (guint8 * d1, const guint8 * s1, int n) +{ + OrcExecutor _ex, *ex = &_ex; + static int p_inited = 0; + static OrcProgram *p = 0; + void (*func) (OrcExecutor *); + + if (!p_inited) { + orc_once_mutex_lock (); + if (!p_inited) { + OrcCompileResult result; + + p = orc_program_new (); + orc_program_set_name (p, "orc_unpack_uyvy_u"); + orc_program_set_backup_function (p, _backup_orc_unpack_uyvy_u); + orc_program_add_destination (p, 1, "d1"); + orc_program_add_source (p, 4, "s1"); + orc_program_add_constant (p, 0, 0x00000000, "c1"); + orc_program_add_constant (p, 0, 0x00000000, "c2"); + orc_program_add_constant (p, 0, 0x00000000, "c3"); + orc_program_add_constant (p, 0, 0x00000000, "c4"); + orc_program_add_constant (p, 0, 0x00000000, "c5"); + orc_program_add_constant (p, 0, 0x00000000, "c6"); + orc_program_add_constant (p, 0, 0x00000000, "c7"); + orc_program_add_constant (p, 0, 0x00000000, "c8"); + orc_program_add_temporary (p, 2, "t1"); + + orc_program_append_2 (p, "select0lw", 0, ORC_VAR_T1, ORC_VAR_S1, + ORC_VAR_D1, ORC_VAR_D1); + orc_program_append_2 (p, "select0wb", 0, ORC_VAR_D1, ORC_VAR_T1, + ORC_VAR_D1, ORC_VAR_D1); + + result = orc_program_compile (p); + } + p_inited = TRUE; + orc_once_mutex_unlock (); + } + ex->program = p; + + ex->n = n; + ex->arrays[ORC_VAR_D1] = d1; + ex->arrays[ORC_VAR_S1] = (void *) s1; + + func = p->code_exec; + func (ex); +} +#endif + + +/* orc_unpack_uyvy_v */ +#ifdef DISABLE_ORC +void +orc_unpack_uyvy_v (guint8 * d1, const guint8 * s1, int n) +{ + int i; + orc_int8 *ORC_RESTRICT ptr0; + const orc_union32 *ORC_RESTRICT ptr4; + orc_union32 var33; + orc_int8 var34; + orc_union16 var35; + + ptr0 = (orc_int8 *) d1; + ptr4 = (orc_union32 *) s1; + + + for (i = 0; i < n; i++) { + /* 0: loadl */ + var33 = ptr4[i]; + /* 1: select1lw */ + var35.i = ((orc_uint32) var33.i >> 16) & 0xffff; + /* 2: select0wb */ + var34 = (orc_uint16) var35.i & 0xff; + /* 3: storeb */ + ptr0[i] = var34; + } + +} + +#else +static void +_backup_orc_unpack_uyvy_v (OrcExecutor * ex) +{ + int i; + int n = ex->n; + orc_int8 *ORC_RESTRICT ptr0; + const orc_union32 *ORC_RESTRICT ptr4; + orc_union32 var33; + orc_int8 var34; + orc_union16 var35; + + ptr0 = (orc_int8 *) ex->arrays[0]; + ptr4 = (orc_union32 *) ex->arrays[4]; + + + for (i = 0; i < n; i++) { + /* 0: loadl */ + var33 = ptr4[i]; + /* 1: select1lw */ + var35.i = ((orc_uint32) var33.i >> 16) & 0xffff; + /* 2: select0wb */ + var34 = (orc_uint16) var35.i & 0xff; + /* 3: storeb */ + ptr0[i] = var34; + } + +} + +void +orc_unpack_uyvy_v (guint8 * d1, const guint8 * s1, int n) +{ + OrcExecutor _ex, *ex = &_ex; + static int p_inited = 0; + static OrcProgram *p = 0; + void (*func) (OrcExecutor *); + + if (!p_inited) { + orc_once_mutex_lock (); + if (!p_inited) { + OrcCompileResult result; + + p = orc_program_new (); + orc_program_set_name (p, "orc_unpack_uyvy_v"); + orc_program_set_backup_function (p, _backup_orc_unpack_uyvy_v); + orc_program_add_destination (p, 1, "d1"); + orc_program_add_source (p, 4, "s1"); + orc_program_add_constant (p, 0, 0x00000000, "c1"); + orc_program_add_constant (p, 0, 0x00000000, "c2"); + orc_program_add_constant (p, 0, 0x00000000, "c3"); + orc_program_add_constant (p, 0, 0x00000000, "c4"); + orc_program_add_constant (p, 0, 0x00000000, "c5"); + orc_program_add_constant (p, 0, 0x00000000, "c6"); + orc_program_add_constant (p, 0, 0x00000000, "c7"); + orc_program_add_constant (p, 0, 0x00000000, "c8"); + orc_program_add_temporary (p, 2, "t1"); + + orc_program_append_2 (p, "select1lw", 0, ORC_VAR_T1, ORC_VAR_S1, + ORC_VAR_D1, ORC_VAR_D1); + orc_program_append_2 (p, "select0wb", 0, ORC_VAR_D1, ORC_VAR_T1, + ORC_VAR_D1, ORC_VAR_D1); + + result = orc_program_compile (p); + } + p_inited = TRUE; + orc_once_mutex_unlock (); + } + ex->program = p; + + ex->n = n; + ex->arrays[ORC_VAR_D1] = d1; + ex->arrays[ORC_VAR_S1] = (void *) s1; + + func = p->code_exec; + func (ex); +} +#endif + + +/* orc_pack_uyvy */ +#ifdef DISABLE_ORC +void +orc_pack_uyvy (guint8 * d1, const guint8 * s1, const guint8 * s2, + const guint8 * s3, int n) +{ + int i; + orc_union32 *ORC_RESTRICT ptr0; + const orc_union16 *ORC_RESTRICT ptr4; + const orc_int8 *ORC_RESTRICT ptr5; + const orc_int8 *ORC_RESTRICT ptr6; + orc_union16 var37; + orc_int8 var38; + orc_int8 var39; + orc_union32 var40; + orc_union16 var41; + orc_int8 var42; + orc_int8 var43; + orc_union16 var44; + orc_union16 var45; + + ptr0 = (orc_union32 *) d1; + ptr4 = (orc_union16 *) s1; + ptr5 = (orc_int8 *) s2; + ptr6 = (orc_int8 *) s3; + + + for (i = 0; i < n; i++) { + /* 0: loadw */ + var37 = ptr4[i]; + /* 1: copyw */ + var41.i = var37.i; + /* 2: select0wb */ + var42 = (orc_uint16) var41.i & 0xff; + /* 3: select1wb */ + var43 = ((orc_uint16) var41.i >> 8) & 0xff; + /* 4: loadb */ + var38 = ptr5[i]; + /* 5: mergebw */ + var44.i = ((orc_uint8) var38) | ((orc_uint8) var42 << 8); + /* 6: loadb */ + var39 = ptr6[i]; + /* 7: mergebw */ + var45.i = ((orc_uint8) var39) | ((orc_uint8) var43 << 8); + /* 8: mergewl */ + var40.i = ((orc_uint16) var44.i) | ((orc_uint16) var45.i << 16); + /* 9: storel */ + ptr0[i] = var40; + } + +} + +#else +static void +_backup_orc_pack_uyvy (OrcExecutor * ex) +{ + int i; + int n = ex->n; + orc_union32 *ORC_RESTRICT ptr0; + const orc_union16 *ORC_RESTRICT ptr4; + const orc_int8 *ORC_RESTRICT ptr5; + const orc_int8 *ORC_RESTRICT ptr6; + orc_union16 var37; + orc_int8 var38; + orc_int8 var39; + orc_union32 var40; + orc_union16 var41; + orc_int8 var42; + orc_int8 var43; + orc_union16 var44; + orc_union16 var45; + + ptr0 = (orc_union32 *) ex->arrays[0]; + ptr4 = (orc_union16 *) ex->arrays[4]; + ptr5 = (orc_int8 *) ex->arrays[5]; + ptr6 = (orc_int8 *) ex->arrays[6]; + + + for (i = 0; i < n; i++) { + /* 0: loadw */ + var37 = ptr4[i]; + /* 1: copyw */ + var41.i = var37.i; + /* 2: select0wb */ + var42 = (orc_uint16) var41.i & 0xff; + /* 3: select1wb */ + var43 = ((orc_uint16) var41.i >> 8) & 0xff; + /* 4: loadb */ + var38 = ptr5[i]; + /* 5: mergebw */ + var44.i = ((orc_uint8) var38) | ((orc_uint8) var42 << 8); + /* 6: loadb */ + var39 = ptr6[i]; + /* 7: mergebw */ + var45.i = ((orc_uint8) var39) | ((orc_uint8) var43 << 8); + /* 8: mergewl */ + var40.i = ((orc_uint16) var44.i) | ((orc_uint16) var45.i << 16); + /* 9: storel */ + ptr0[i] = var40; + } + +} + +void +orc_pack_uyvy (guint8 * d1, const guint8 * s1, const guint8 * s2, + const guint8 * s3, int n) +{ + OrcExecutor _ex, *ex = &_ex; + static int p_inited = 0; + static OrcProgram *p = 0; + void (*func) (OrcExecutor *); + + if (!p_inited) { + orc_once_mutex_lock (); + if (!p_inited) { + OrcCompileResult result; + + p = orc_program_new (); + orc_program_set_name (p, "orc_pack_uyvy"); + orc_program_set_backup_function (p, _backup_orc_pack_uyvy); + orc_program_add_destination (p, 4, "d1"); + orc_program_add_source (p, 2, "s1"); + orc_program_add_source (p, 1, "s2"); + orc_program_add_source (p, 1, "s3"); + orc_program_add_constant (p, 0, 0x00000000, "c1"); + orc_program_add_constant (p, 0, 0x00000000, "c2"); + orc_program_add_constant (p, 0, 0x00000000, "c3"); + orc_program_add_constant (p, 0, 0x00000000, "c4"); + orc_program_add_constant (p, 0, 0x00000000, "c5"); + orc_program_add_constant (p, 0, 0x00000000, "c6"); + orc_program_add_constant (p, 0, 0x00000000, "c7"); + orc_program_add_constant (p, 0, 0x00000000, "c8"); + orc_program_add_temporary (p, 1, "t1"); + orc_program_add_temporary (p, 1, "t2"); + orc_program_add_temporary (p, 2, "t3"); + orc_program_add_temporary (p, 2, "t4"); + orc_program_add_temporary (p, 2, "t5"); + + orc_program_append_2 (p, "copyw", 0, ORC_VAR_T5, ORC_VAR_S1, ORC_VAR_D1, + ORC_VAR_D1); + orc_program_append_2 (p, "select0wb", 0, ORC_VAR_T1, ORC_VAR_T5, + ORC_VAR_D1, ORC_VAR_D1); + orc_program_append_2 (p, "select1wb", 0, ORC_VAR_T2, ORC_VAR_T5, + ORC_VAR_D1, ORC_VAR_D1); + orc_program_append_2 (p, "mergebw", 0, ORC_VAR_T3, ORC_VAR_S2, ORC_VAR_T1, + ORC_VAR_D1); + orc_program_append_2 (p, "mergebw", 0, ORC_VAR_T4, ORC_VAR_S3, ORC_VAR_T2, + ORC_VAR_D1); + orc_program_append_2 (p, "mergewl", 0, ORC_VAR_D1, ORC_VAR_T3, ORC_VAR_T4, + ORC_VAR_D1); + + result = orc_program_compile (p); + } + p_inited = TRUE; + orc_once_mutex_unlock (); + } + ex->program = p; + + ex->n = n; + ex->arrays[ORC_VAR_D1] = d1; + ex->arrays[ORC_VAR_S1] = (void *) s1; + ex->arrays[ORC_VAR_S2] = (void *) s2; + ex->arrays[ORC_VAR_S3] = (void *) s3; + + func = p->code_exec; + func (ex); +} +#endif + + +/* orc_matrix2_u8 */ +#ifdef DISABLE_ORC +void +orc_matrix2_u8 (guint8 * d1, const guint8 * s1, const guint8 * s2, int p1, + int p2, int p3, int n) +{ + int i; + orc_int8 *ORC_RESTRICT ptr0; + const orc_int8 *ORC_RESTRICT ptr4; + const orc_int8 *ORC_RESTRICT ptr5; + orc_int8 var34; + orc_union16 var35; + orc_int8 var36; + orc_union16 var37; + orc_union16 var38; + orc_int8 var39; + orc_union16 var40; + orc_union16 var41; + orc_union16 var42; + orc_union16 var43; + orc_union16 var44; + orc_union16 var45; + orc_union16 var46; + + ptr0 = (orc_int8 *) d1; + ptr4 = (orc_int8 *) s1; + ptr5 = (orc_int8 *) s2; + + /* 2: loadpw */ + var35.i = p1; + /* 6: loadpw */ + var37.i = p2; + /* 9: loadpw */ + var38.i = p3; + + for (i = 0; i < n; i++) { + /* 0: loadb */ + var34 = ptr4[i]; + /* 1: convubw */ + var40.i = (orc_uint8) var34; + /* 3: mullw */ + var41.i = (var40.i * var35.i) & 0xffff; + /* 4: loadb */ + var36 = ptr5[i]; + /* 5: convubw */ + var42.i = (orc_uint8) var36; + /* 7: mullw */ + var43.i = (var42.i * var37.i) & 0xffff; + /* 8: addw */ + var44.i = var41.i + var43.i; + /* 10: addw */ + var45.i = var44.i + var38.i; + /* 11: shrsw */ + var46.i = var45.i >> 6; + /* 12: convsuswb */ + var39 = ORC_CLAMP_UB (var46.i); + /* 13: storeb */ + ptr0[i] = var39; + } + +} + +#else +static void +_backup_orc_matrix2_u8 (OrcExecutor * ex) +{ + int i; + int n = ex->n; + orc_int8 *ORC_RESTRICT ptr0; + const orc_int8 *ORC_RESTRICT ptr4; + const orc_int8 *ORC_RESTRICT ptr5; + orc_int8 var34; + orc_union16 var35; + orc_int8 var36; + orc_union16 var37; + orc_union16 var38; + orc_int8 var39; + orc_union16 var40; + orc_union16 var41; + orc_union16 var42; + orc_union16 var43; + orc_union16 var44; + orc_union16 var45; + orc_union16 var46; + + ptr0 = (orc_int8 *) ex->arrays[0]; + ptr4 = (orc_int8 *) ex->arrays[4]; + ptr5 = (orc_int8 *) ex->arrays[5]; + + /* 2: loadpw */ + var35.i = ex->params[24]; + /* 6: loadpw */ + var37.i = ex->params[25]; + /* 9: loadpw */ + var38.i = ex->params[26]; + + for (i = 0; i < n; i++) { + /* 0: loadb */ + var34 = ptr4[i]; + /* 1: convubw */ + var40.i = (orc_uint8) var34; + /* 3: mullw */ + var41.i = (var40.i * var35.i) & 0xffff; + /* 4: loadb */ + var36 = ptr5[i]; + /* 5: convubw */ + var42.i = (orc_uint8) var36; + /* 7: mullw */ + var43.i = (var42.i * var37.i) & 0xffff; + /* 8: addw */ + var44.i = var41.i + var43.i; + /* 10: addw */ + var45.i = var44.i + var38.i; + /* 11: shrsw */ + var46.i = var45.i >> 6; + /* 12: convsuswb */ + var39 = ORC_CLAMP_UB (var46.i); + /* 13: storeb */ + ptr0[i] = var39; + } + +} + +void +orc_matrix2_u8 (guint8 * d1, const guint8 * s1, const guint8 * s2, int p1, + int p2, int p3, int n) +{ + OrcExecutor _ex, *ex = &_ex; + static int p_inited = 0; + static OrcProgram *p = 0; + void (*func) (OrcExecutor *); + + if (!p_inited) { + orc_once_mutex_lock (); + if (!p_inited) { + OrcCompileResult result; + + p = orc_program_new (); + orc_program_set_name (p, "orc_matrix2_u8"); + orc_program_set_backup_function (p, _backup_orc_matrix2_u8); + orc_program_add_destination (p, 1, "d1"); + orc_program_add_source (p, 1, "s1"); + orc_program_add_source (p, 1, "s2"); + orc_program_add_constant (p, 4, 0x00000006, "c1"); + orc_program_add_constant (p, 0, 0x00000000, "c2"); + orc_program_add_constant (p, 0, 0x00000000, "c3"); + orc_program_add_constant (p, 0, 0x00000000, "c4"); + orc_program_add_constant (p, 0, 0x00000000, "c5"); + orc_program_add_constant (p, 0, 0x00000000, "c6"); + orc_program_add_constant (p, 0, 0x00000000, "c7"); + orc_program_add_constant (p, 0, 0x00000000, "c8"); + orc_program_add_parameter (p, 2, "p1"); + orc_program_add_parameter (p, 2, "p2"); + orc_program_add_parameter (p, 2, "p3"); + orc_program_add_temporary (p, 2, "t1"); + orc_program_add_temporary (p, 2, "t2"); + + orc_program_append_2 (p, "convubw", 0, ORC_VAR_T1, ORC_VAR_S1, ORC_VAR_D1, + ORC_VAR_D1); + orc_program_append_2 (p, "mullw", 0, ORC_VAR_T1, ORC_VAR_T1, ORC_VAR_P1, + ORC_VAR_D1); + orc_program_append_2 (p, "convubw", 0, ORC_VAR_T2, ORC_VAR_S2, ORC_VAR_D1, + ORC_VAR_D1); + orc_program_append_2 (p, "mullw", 0, ORC_VAR_T2, ORC_VAR_T2, ORC_VAR_P2, + ORC_VAR_D1); + orc_program_append_2 (p, "addw", 0, ORC_VAR_T1, ORC_VAR_T1, ORC_VAR_T2, + ORC_VAR_D1); + orc_program_append_2 (p, "addw", 0, ORC_VAR_T1, ORC_VAR_T1, ORC_VAR_P3, + ORC_VAR_D1); + orc_program_append_2 (p, "shrsw", 0, ORC_VAR_T1, ORC_VAR_T1, ORC_VAR_C1, + ORC_VAR_D1); + orc_program_append_2 (p, "convsuswb", 0, ORC_VAR_D1, ORC_VAR_T1, + ORC_VAR_D1, ORC_VAR_D1); + + result = orc_program_compile (p); + } + p_inited = TRUE; + orc_once_mutex_unlock (); + } + ex->program = p; + + ex->n = n; + ex->arrays[ORC_VAR_D1] = d1; + ex->arrays[ORC_VAR_S1] = (void *) s1; + ex->arrays[ORC_VAR_S2] = (void *) s2; + ex->params[ORC_VAR_P1] = p1; + ex->params[ORC_VAR_P2] = p2; + ex->params[ORC_VAR_P3] = p3; + + func = p->code_exec; + func (ex); +} +#endif + + +/* orc_matrix2_11_u8 */ +#ifdef DISABLE_ORC +void +orc_matrix2_11_u8 (guint8 * d1, const guint8 * s1, const guint8 * s2, int p1, + int p2, int n) +{ + int i; + orc_int8 *ORC_RESTRICT ptr0; + const orc_int8 *ORC_RESTRICT ptr4; + const orc_int8 *ORC_RESTRICT ptr5; + orc_int8 var36; + orc_union16 var37; + orc_union16 var38; + orc_int8 var39; + orc_union16 var40; + orc_union16 var41; + orc_union16 var42; + orc_int8 var43; + orc_union16 var44; + orc_union16 var45; + orc_union16 var46; + orc_union16 var47; + orc_union16 var48; + orc_union16 var49; + orc_union16 var50; + orc_union16 var51; + orc_union16 var52; + orc_union16 var53; + orc_union16 var54; + + ptr0 = (orc_int8 *) d1; + ptr4 = (orc_int8 *) s1; + ptr5 = (orc_int8 *) s2; + + /* 2: loadpw */ + var37.i = 0x00000010; /* 16 or 7.90505e-323f */ + /* 4: loadpw */ + var38.i = p1; + /* 8: loadpw */ + var40.i = 0x00000080; /* 128 or 6.32404e-322f */ + /* 10: loadpw */ + var41.i = p2; + /* 13: loadpw */ + var42.i = 0x00000080; /* 128 or 6.32404e-322f */ + + for (i = 0; i < n; i++) { + /* 0: loadb */ + var36 = ptr4[i]; + /* 1: convubw */ + var44.i = (orc_uint8) var36; + /* 3: subw */ + var45.i = var44.i - var37.i; + /* 5: mullw */ + var46.i = (var45.i * var38.i) & 0xffff; + /* 6: loadb */ + var39 = ptr5[i]; + /* 7: convubw */ + var47.i = (orc_uint8) var39; + /* 9: subw */ + var48.i = var47.i - var40.i; + /* 11: mullw */ + var49.i = (var48.i * var41.i) & 0xffff; + /* 12: addw */ + var50.i = var46.i + var49.i; + /* 14: addw */ + var51.i = var50.i + var42.i; + /* 15: shrsw */ + var52.i = var51.i >> 8; + /* 16: addw */ + var53.i = var52.i + var45.i; + /* 17: addw */ + var54.i = var53.i + var48.i; + /* 18: convsuswb */ + var43 = ORC_CLAMP_UB (var54.i); + /* 19: storeb */ + ptr0[i] = var43; + } + +} + +#else +static void +_backup_orc_matrix2_11_u8 (OrcExecutor * ex) +{ + int i; + int n = ex->n; + orc_int8 *ORC_RESTRICT ptr0; + const orc_int8 *ORC_RESTRICT ptr4; + const orc_int8 *ORC_RESTRICT ptr5; + orc_int8 var36; + orc_union16 var37; + orc_union16 var38; + orc_int8 var39; + orc_union16 var40; + orc_union16 var41; + orc_union16 var42; + orc_int8 var43; + orc_union16 var44; + orc_union16 var45; + orc_union16 var46; + orc_union16 var47; + orc_union16 var48; + orc_union16 var49; + orc_union16 var50; + orc_union16 var51; + orc_union16 var52; + orc_union16 var53; + orc_union16 var54; + + ptr0 = (orc_int8 *) ex->arrays[0]; + ptr4 = (orc_int8 *) ex->arrays[4]; + ptr5 = (orc_int8 *) ex->arrays[5]; + + /* 2: loadpw */ + var37.i = 0x00000010; /* 16 or 7.90505e-323f */ + /* 4: loadpw */ + var38.i = ex->params[24]; + /* 8: loadpw */ + var40.i = 0x00000080; /* 128 or 6.32404e-322f */ + /* 10: loadpw */ + var41.i = ex->params[25]; + /* 13: loadpw */ + var42.i = 0x00000080; /* 128 or 6.32404e-322f */ + + for (i = 0; i < n; i++) { + /* 0: loadb */ + var36 = ptr4[i]; + /* 1: convubw */ + var44.i = (orc_uint8) var36; + /* 3: subw */ + var45.i = var44.i - var37.i; + /* 5: mullw */ + var46.i = (var45.i * var38.i) & 0xffff; + /* 6: loadb */ + var39 = ptr5[i]; + /* 7: convubw */ + var47.i = (orc_uint8) var39; + /* 9: subw */ + var48.i = var47.i - var40.i; + /* 11: mullw */ + var49.i = (var48.i * var41.i) & 0xffff; + /* 12: addw */ + var50.i = var46.i + var49.i; + /* 14: addw */ + var51.i = var50.i + var42.i; + /* 15: shrsw */ + var52.i = var51.i >> 8; + /* 16: addw */ + var53.i = var52.i + var45.i; + /* 17: addw */ + var54.i = var53.i + var48.i; + /* 18: convsuswb */ + var43 = ORC_CLAMP_UB (var54.i); + /* 19: storeb */ + ptr0[i] = var43; + } + +} + +void +orc_matrix2_11_u8 (guint8 * d1, const guint8 * s1, const guint8 * s2, int p1, + int p2, int n) +{ + OrcExecutor _ex, *ex = &_ex; + static int p_inited = 0; + static OrcProgram *p = 0; + void (*func) (OrcExecutor *); + + if (!p_inited) { + orc_once_mutex_lock (); + if (!p_inited) { + OrcCompileResult result; + + p = orc_program_new (); + orc_program_set_name (p, "orc_matrix2_11_u8"); + orc_program_set_backup_function (p, _backup_orc_matrix2_11_u8); + orc_program_add_destination (p, 1, "d1"); + orc_program_add_source (p, 1, "s1"); + orc_program_add_source (p, 1, "s2"); + orc_program_add_constant (p, 4, 0x00000010, "c1"); + orc_program_add_constant (p, 4, 0x00000080, "c2"); + orc_program_add_constant (p, 4, 0x00000008, "c3"); + orc_program_add_constant (p, 0, 0x00000000, "c4"); + orc_program_add_constant (p, 0, 0x00000000, "c5"); + orc_program_add_constant (p, 0, 0x00000000, "c6"); + orc_program_add_constant (p, 0, 0x00000000, "c7"); + orc_program_add_constant (p, 0, 0x00000000, "c8"); + orc_program_add_parameter (p, 2, "p1"); + orc_program_add_parameter (p, 2, "p2"); + orc_program_add_temporary (p, 2, "t1"); + orc_program_add_temporary (p, 2, "t2"); + orc_program_add_temporary (p, 2, "t3"); + orc_program_add_temporary (p, 2, "t4"); + + orc_program_append_2 (p, "convubw", 0, ORC_VAR_T1, ORC_VAR_S1, ORC_VAR_D1, + ORC_VAR_D1); + orc_program_append_2 (p, "subw", 0, ORC_VAR_T1, ORC_VAR_T1, ORC_VAR_C1, + ORC_VAR_D1); + orc_program_append_2 (p, "mullw", 0, ORC_VAR_T3, ORC_VAR_T1, ORC_VAR_P1, + ORC_VAR_D1); + orc_program_append_2 (p, "convubw", 0, ORC_VAR_T2, ORC_VAR_S2, ORC_VAR_D1, + ORC_VAR_D1); + orc_program_append_2 (p, "subw", 0, ORC_VAR_T2, ORC_VAR_T2, ORC_VAR_C2, + ORC_VAR_D1); + orc_program_append_2 (p, "mullw", 0, ORC_VAR_T4, ORC_VAR_T2, ORC_VAR_P2, + ORC_VAR_D1); + orc_program_append_2 (p, "addw", 0, ORC_VAR_T3, ORC_VAR_T3, ORC_VAR_T4, + ORC_VAR_D1); + orc_program_append_2 (p, "addw", 0, ORC_VAR_T3, ORC_VAR_T3, ORC_VAR_C2, + ORC_VAR_D1); + orc_program_append_2 (p, "shrsw", 0, ORC_VAR_T3, ORC_VAR_T3, ORC_VAR_C3, + ORC_VAR_D1); + orc_program_append_2 (p, "addw", 0, ORC_VAR_T3, ORC_VAR_T3, ORC_VAR_T1, + ORC_VAR_D1); + orc_program_append_2 (p, "addw", 0, ORC_VAR_T3, ORC_VAR_T3, ORC_VAR_T2, + ORC_VAR_D1); + orc_program_append_2 (p, "convsuswb", 0, ORC_VAR_D1, ORC_VAR_T3, + ORC_VAR_D1, ORC_VAR_D1); + + result = orc_program_compile (p); + } + p_inited = TRUE; + orc_once_mutex_unlock (); + } + ex->program = p; + + ex->n = n; + ex->arrays[ORC_VAR_D1] = d1; + ex->arrays[ORC_VAR_S1] = (void *) s1; + ex->arrays[ORC_VAR_S2] = (void *) s2; + ex->params[ORC_VAR_P1] = p1; + ex->params[ORC_VAR_P2] = p2; + + func = p->code_exec; + func (ex); +} +#endif + + +/* orc_matrix2_12_u8 */ +#ifdef DISABLE_ORC +void +orc_matrix2_12_u8 (guint8 * d1, const guint8 * s1, const guint8 * s2, int p1, + int p2, int n) +{ + int i; + orc_int8 *ORC_RESTRICT ptr0; + const orc_int8 *ORC_RESTRICT ptr4; + const orc_int8 *ORC_RESTRICT ptr5; + orc_int8 var36; + orc_union16 var37; + orc_union16 var38; + orc_int8 var39; + orc_union16 var40; + orc_union16 var41; + orc_union16 var42; + orc_int8 var43; + orc_union16 var44; + orc_union16 var45; + orc_union16 var46; + orc_union16 var47; + orc_union16 var48; + orc_union16 var49; + orc_union16 var50; + orc_union16 var51; + orc_union16 var52; + orc_union16 var53; + orc_union16 var54; + orc_union16 var55; + + ptr0 = (orc_int8 *) d1; + ptr4 = (orc_int8 *) s1; + ptr5 = (orc_int8 *) s2; + + /* 2: loadpw */ + var37.i = 0x00000010; /* 16 or 7.90505e-323f */ + /* 4: loadpw */ + var38.i = p1; + /* 8: loadpw */ + var40.i = 0x00000080; /* 128 or 6.32404e-322f */ + /* 10: loadpw */ + var41.i = p2; + /* 13: loadpw */ + var42.i = 0x00000080; /* 128 or 6.32404e-322f */ + + for (i = 0; i < n; i++) { + /* 0: loadb */ + var36 = ptr4[i]; + /* 1: convubw */ + var44.i = (orc_uint8) var36; + /* 3: subw */ + var45.i = var44.i - var37.i; + /* 5: mullw */ + var46.i = (var45.i * var38.i) & 0xffff; + /* 6: loadb */ + var39 = ptr5[i]; + /* 7: convubw */ + var47.i = (orc_uint8) var39; + /* 9: subw */ + var48.i = var47.i - var40.i; + /* 11: mullw */ + var49.i = (var48.i * var41.i) & 0xffff; + /* 12: addw */ + var50.i = var46.i + var49.i; + /* 14: addw */ + var51.i = var50.i + var42.i; + /* 15: shrsw */ + var52.i = var51.i >> 8; + /* 16: addw */ + var53.i = var52.i + var45.i; + /* 17: addw */ + var54.i = var53.i + var48.i; + /* 18: addw */ + var55.i = var54.i + var48.i; + /* 19: convsuswb */ + var43 = ORC_CLAMP_UB (var55.i); + /* 20: storeb */ + ptr0[i] = var43; + } + +} + +#else +static void +_backup_orc_matrix2_12_u8 (OrcExecutor * ex) +{ + int i; + int n = ex->n; + orc_int8 *ORC_RESTRICT ptr0; + const orc_int8 *ORC_RESTRICT ptr4; + const orc_int8 *ORC_RESTRICT ptr5; + orc_int8 var36; + orc_union16 var37; + orc_union16 var38; + orc_int8 var39; + orc_union16 var40; + orc_union16 var41; + orc_union16 var42; + orc_int8 var43; + orc_union16 var44; + orc_union16 var45; + orc_union16 var46; + orc_union16 var47; + orc_union16 var48; + orc_union16 var49; + orc_union16 var50; + orc_union16 var51; + orc_union16 var52; + orc_union16 var53; + orc_union16 var54; + orc_union16 var55; + + ptr0 = (orc_int8 *) ex->arrays[0]; + ptr4 = (orc_int8 *) ex->arrays[4]; + ptr5 = (orc_int8 *) ex->arrays[5]; + + /* 2: loadpw */ + var37.i = 0x00000010; /* 16 or 7.90505e-323f */ + /* 4: loadpw */ + var38.i = ex->params[24]; + /* 8: loadpw */ + var40.i = 0x00000080; /* 128 or 6.32404e-322f */ + /* 10: loadpw */ + var41.i = ex->params[25]; + /* 13: loadpw */ + var42.i = 0x00000080; /* 128 or 6.32404e-322f */ + + for (i = 0; i < n; i++) { + /* 0: loadb */ + var36 = ptr4[i]; + /* 1: convubw */ + var44.i = (orc_uint8) var36; + /* 3: subw */ + var45.i = var44.i - var37.i; + /* 5: mullw */ + var46.i = (var45.i * var38.i) & 0xffff; + /* 6: loadb */ + var39 = ptr5[i]; + /* 7: convubw */ + var47.i = (orc_uint8) var39; + /* 9: subw */ + var48.i = var47.i - var40.i; + /* 11: mullw */ + var49.i = (var48.i * var41.i) & 0xffff; + /* 12: addw */ + var50.i = var46.i + var49.i; + /* 14: addw */ + var51.i = var50.i + var42.i; + /* 15: shrsw */ + var52.i = var51.i >> 8; + /* 16: addw */ + var53.i = var52.i + var45.i; + /* 17: addw */ + var54.i = var53.i + var48.i; + /* 18: addw */ + var55.i = var54.i + var48.i; + /* 19: convsuswb */ + var43 = ORC_CLAMP_UB (var55.i); + /* 20: storeb */ + ptr0[i] = var43; + } + +} + +void +orc_matrix2_12_u8 (guint8 * d1, const guint8 * s1, const guint8 * s2, int p1, + int p2, int n) +{ + OrcExecutor _ex, *ex = &_ex; + static int p_inited = 0; + static OrcProgram *p = 0; + void (*func) (OrcExecutor *); + + if (!p_inited) { + orc_once_mutex_lock (); + if (!p_inited) { + OrcCompileResult result; + + p = orc_program_new (); + orc_program_set_name (p, "orc_matrix2_12_u8"); + orc_program_set_backup_function (p, _backup_orc_matrix2_12_u8); + orc_program_add_destination (p, 1, "d1"); + orc_program_add_source (p, 1, "s1"); + orc_program_add_source (p, 1, "s2"); + orc_program_add_constant (p, 4, 0x00000010, "c1"); + orc_program_add_constant (p, 4, 0x00000080, "c2"); + orc_program_add_constant (p, 4, 0x00000008, "c3"); + orc_program_add_constant (p, 0, 0x00000000, "c4"); + orc_program_add_constant (p, 0, 0x00000000, "c5"); + orc_program_add_constant (p, 0, 0x00000000, "c6"); + orc_program_add_constant (p, 0, 0x00000000, "c7"); + orc_program_add_constant (p, 0, 0x00000000, "c8"); + orc_program_add_parameter (p, 2, "p1"); + orc_program_add_parameter (p, 2, "p2"); + orc_program_add_temporary (p, 2, "t1"); + orc_program_add_temporary (p, 2, "t2"); + orc_program_add_temporary (p, 2, "t3"); + orc_program_add_temporary (p, 2, "t4"); + + orc_program_append_2 (p, "convubw", 0, ORC_VAR_T1, ORC_VAR_S1, ORC_VAR_D1, + ORC_VAR_D1); + orc_program_append_2 (p, "subw", 0, ORC_VAR_T1, ORC_VAR_T1, ORC_VAR_C1, + ORC_VAR_D1); + orc_program_append_2 (p, "mullw", 0, ORC_VAR_T3, ORC_VAR_T1, ORC_VAR_P1, + ORC_VAR_D1); + orc_program_append_2 (p, "convubw", 0, ORC_VAR_T2, ORC_VAR_S2, ORC_VAR_D1, + ORC_VAR_D1); + orc_program_append_2 (p, "subw", 0, ORC_VAR_T2, ORC_VAR_T2, ORC_VAR_C2, + ORC_VAR_D1); + orc_program_append_2 (p, "mullw", 0, ORC_VAR_T4, ORC_VAR_T2, ORC_VAR_P2, + ORC_VAR_D1); + orc_program_append_2 (p, "addw", 0, ORC_VAR_T3, ORC_VAR_T3, ORC_VAR_T4, + ORC_VAR_D1); + orc_program_append_2 (p, "addw", 0, ORC_VAR_T3, ORC_VAR_T3, ORC_VAR_C2, + ORC_VAR_D1); + orc_program_append_2 (p, "shrsw", 0, ORC_VAR_T3, ORC_VAR_T3, ORC_VAR_C3, + ORC_VAR_D1); + orc_program_append_2 (p, "addw", 0, ORC_VAR_T3, ORC_VAR_T3, ORC_VAR_T1, + ORC_VAR_D1); + orc_program_append_2 (p, "addw", 0, ORC_VAR_T3, ORC_VAR_T3, ORC_VAR_T2, + ORC_VAR_D1); + orc_program_append_2 (p, "addw", 0, ORC_VAR_T3, ORC_VAR_T3, ORC_VAR_T2, + ORC_VAR_D1); + orc_program_append_2 (p, "convsuswb", 0, ORC_VAR_D1, ORC_VAR_T3, + ORC_VAR_D1, ORC_VAR_D1); + + result = orc_program_compile (p); + } + p_inited = TRUE; + orc_once_mutex_unlock (); + } + ex->program = p; + + ex->n = n; + ex->arrays[ORC_VAR_D1] = d1; + ex->arrays[ORC_VAR_S1] = (void *) s1; + ex->arrays[ORC_VAR_S2] = (void *) s2; + ex->params[ORC_VAR_P1] = p1; + ex->params[ORC_VAR_P2] = p2; + + func = p->code_exec; + func (ex); +} +#endif + + +/* orc_matrix3_u8 */ +#ifdef DISABLE_ORC +void +orc_matrix3_u8 (guint8 * d1, const guint8 * s1, const guint8 * s2, + const guint8 * s3, int p1, int p2, int p3, int p4, int n) +{ + int i; + orc_int8 *ORC_RESTRICT ptr0; + const orc_int8 *ORC_RESTRICT ptr4; + const orc_int8 *ORC_RESTRICT ptr5; + const orc_int8 *ORC_RESTRICT ptr6; + orc_int8 var34; + orc_union16 var35; + orc_int8 var36; + orc_union16 var37; + orc_int8 var38; + orc_union16 var39; + orc_union16 var40; + orc_int8 var41; + orc_union16 var42; + orc_union16 var43; + orc_union16 var44; + orc_union16 var45; + orc_union16 var46; + orc_union16 var47; + orc_union16 var48; + orc_union16 var49; + orc_union16 var50; + orc_union16 var51; + + ptr0 = (orc_int8 *) d1; + ptr4 = (orc_int8 *) s1; + ptr5 = (orc_int8 *) s2; + ptr6 = (orc_int8 *) s3; + + /* 2: loadpw */ + var35.i = p1; + /* 6: loadpw */ + var37.i = p2; + /* 11: loadpw */ + var39.i = p3; + /* 14: loadpw */ + var40.i = p4; + + for (i = 0; i < n; i++) { + /* 0: loadb */ + var34 = ptr4[i]; + /* 1: convubw */ + var42.i = (orc_uint8) var34; + /* 3: mullw */ + var43.i = (var42.i * var35.i) & 0xffff; + /* 4: loadb */ + var36 = ptr5[i]; + /* 5: convubw */ + var44.i = (orc_uint8) var36; + /* 7: mullw */ + var45.i = (var44.i * var37.i) & 0xffff; + /* 8: addw */ + var46.i = var43.i + var45.i; + /* 9: loadb */ + var38 = ptr6[i]; + /* 10: convubw */ + var47.i = (orc_uint8) var38; + /* 12: mullw */ + var48.i = (var47.i * var39.i) & 0xffff; + /* 13: addw */ + var49.i = var46.i + var48.i; + /* 15: addw */ + var50.i = var49.i + var40.i; + /* 16: shrsw */ + var51.i = var50.i >> 6; + /* 17: convsuswb */ + var41 = ORC_CLAMP_UB (var51.i); + /* 18: storeb */ + ptr0[i] = var41; + } + +} + +#else +static void +_backup_orc_matrix3_u8 (OrcExecutor * ex) +{ + int i; + int n = ex->n; + orc_int8 *ORC_RESTRICT ptr0; + const orc_int8 *ORC_RESTRICT ptr4; + const orc_int8 *ORC_RESTRICT ptr5; + const orc_int8 *ORC_RESTRICT ptr6; + orc_int8 var34; + orc_union16 var35; + orc_int8 var36; + orc_union16 var37; + orc_int8 var38; + orc_union16 var39; + orc_union16 var40; + orc_int8 var41; + orc_union16 var42; + orc_union16 var43; + orc_union16 var44; + orc_union16 var45; + orc_union16 var46; + orc_union16 var47; + orc_union16 var48; + orc_union16 var49; + orc_union16 var50; + orc_union16 var51; + + ptr0 = (orc_int8 *) ex->arrays[0]; + ptr4 = (orc_int8 *) ex->arrays[4]; + ptr5 = (orc_int8 *) ex->arrays[5]; + ptr6 = (orc_int8 *) ex->arrays[6]; + + /* 2: loadpw */ + var35.i = ex->params[24]; + /* 6: loadpw */ + var37.i = ex->params[25]; + /* 11: loadpw */ + var39.i = ex->params[26]; + /* 14: loadpw */ + var40.i = ex->params[27]; + + for (i = 0; i < n; i++) { + /* 0: loadb */ + var34 = ptr4[i]; + /* 1: convubw */ + var42.i = (orc_uint8) var34; + /* 3: mullw */ + var43.i = (var42.i * var35.i) & 0xffff; + /* 4: loadb */ + var36 = ptr5[i]; + /* 5: convubw */ + var44.i = (orc_uint8) var36; + /* 7: mullw */ + var45.i = (var44.i * var37.i) & 0xffff; + /* 8: addw */ + var46.i = var43.i + var45.i; + /* 9: loadb */ + var38 = ptr6[i]; + /* 10: convubw */ + var47.i = (orc_uint8) var38; + /* 12: mullw */ + var48.i = (var47.i * var39.i) & 0xffff; + /* 13: addw */ + var49.i = var46.i + var48.i; + /* 15: addw */ + var50.i = var49.i + var40.i; + /* 16: shrsw */ + var51.i = var50.i >> 6; + /* 17: convsuswb */ + var41 = ORC_CLAMP_UB (var51.i); + /* 18: storeb */ + ptr0[i] = var41; + } + +} + +void +orc_matrix3_u8 (guint8 * d1, const guint8 * s1, const guint8 * s2, + const guint8 * s3, int p1, int p2, int p3, int p4, int n) +{ + OrcExecutor _ex, *ex = &_ex; + static int p_inited = 0; + static OrcProgram *p = 0; + void (*func) (OrcExecutor *); + + if (!p_inited) { + orc_once_mutex_lock (); + if (!p_inited) { + OrcCompileResult result; + + p = orc_program_new (); + orc_program_set_name (p, "orc_matrix3_u8"); + orc_program_set_backup_function (p, _backup_orc_matrix3_u8); + orc_program_add_destination (p, 1, "d1"); + orc_program_add_source (p, 1, "s1"); + orc_program_add_source (p, 1, "s2"); + orc_program_add_source (p, 1, "s3"); + orc_program_add_constant (p, 4, 0x00000006, "c1"); + orc_program_add_constant (p, 0, 0x00000000, "c2"); + orc_program_add_constant (p, 0, 0x00000000, "c3"); + orc_program_add_constant (p, 0, 0x00000000, "c4"); + orc_program_add_constant (p, 0, 0x00000000, "c5"); + orc_program_add_constant (p, 0, 0x00000000, "c6"); + orc_program_add_constant (p, 0, 0x00000000, "c7"); + orc_program_add_constant (p, 0, 0x00000000, "c8"); + orc_program_add_parameter (p, 2, "p1"); + orc_program_add_parameter (p, 2, "p2"); + orc_program_add_parameter (p, 2, "p3"); + orc_program_add_parameter (p, 2, "p4"); + orc_program_add_temporary (p, 2, "t1"); + orc_program_add_temporary (p, 2, "t2"); + + orc_program_append_2 (p, "convubw", 0, ORC_VAR_T1, ORC_VAR_S1, ORC_VAR_D1, + ORC_VAR_D1); + orc_program_append_2 (p, "mullw", 0, ORC_VAR_T1, ORC_VAR_T1, ORC_VAR_P1, + ORC_VAR_D1); + orc_program_append_2 (p, "convubw", 0, ORC_VAR_T2, ORC_VAR_S2, ORC_VAR_D1, + ORC_VAR_D1); + orc_program_append_2 (p, "mullw", 0, ORC_VAR_T2, ORC_VAR_T2, ORC_VAR_P2, + ORC_VAR_D1); + orc_program_append_2 (p, "addw", 0, ORC_VAR_T1, ORC_VAR_T1, ORC_VAR_T2, + ORC_VAR_D1); + orc_program_append_2 (p, "convubw", 0, ORC_VAR_T2, ORC_VAR_S3, ORC_VAR_D1, + ORC_VAR_D1); + orc_program_append_2 (p, "mullw", 0, ORC_VAR_T2, ORC_VAR_T2, ORC_VAR_P3, + ORC_VAR_D1); + orc_program_append_2 (p, "addw", 0, ORC_VAR_T1, ORC_VAR_T1, ORC_VAR_T2, + ORC_VAR_D1); + orc_program_append_2 (p, "addw", 0, ORC_VAR_T1, ORC_VAR_T1, ORC_VAR_P4, + ORC_VAR_D1); + orc_program_append_2 (p, "shrsw", 0, ORC_VAR_T1, ORC_VAR_T1, ORC_VAR_C1, + ORC_VAR_D1); + orc_program_append_2 (p, "convsuswb", 0, ORC_VAR_D1, ORC_VAR_T1, + ORC_VAR_D1, ORC_VAR_D1); + + result = orc_program_compile (p); + } + p_inited = TRUE; + orc_once_mutex_unlock (); + } + ex->program = p; + + ex->n = n; + ex->arrays[ORC_VAR_D1] = d1; + ex->arrays[ORC_VAR_S1] = (void *) s1; + ex->arrays[ORC_VAR_S2] = (void *) s2; + ex->arrays[ORC_VAR_S3] = (void *) s3; + ex->params[ORC_VAR_P1] = p1; + ex->params[ORC_VAR_P2] = p2; + ex->params[ORC_VAR_P3] = p3; + ex->params[ORC_VAR_P4] = p4; + + func = p->code_exec; + func (ex); +} +#endif + + +/* orc_matrix3_100_u8 */ +#ifdef DISABLE_ORC +void +orc_matrix3_100_u8 (guint8 * d1, const guint8 * s1, const guint8 * s2, + const guint8 * s3, int p1, int p2, int p3, int n) +{ + int i; + orc_int8 *ORC_RESTRICT ptr0; + const orc_int8 *ORC_RESTRICT ptr4; + const orc_int8 *ORC_RESTRICT ptr5; + const orc_int8 *ORC_RESTRICT ptr6; + orc_int8 var35; + orc_union16 var36; + orc_union16 var37; + orc_int8 var38; + orc_union16 var39; + orc_union16 var40; + orc_int8 var41; + orc_union16 var42; + orc_union16 var43; + orc_union16 var44; + orc_int8 var45; + orc_union16 var46; + orc_union16 var47; + orc_union16 var48; + orc_union16 var49; + orc_union16 var50; + orc_union16 var51; + orc_union16 var52; + orc_union16 var53; + orc_union16 var54; + orc_union16 var55; + orc_union16 var56; + orc_union16 var57; + orc_union16 var58; + orc_union16 var59; + + ptr0 = (orc_int8 *) d1; + ptr4 = (orc_int8 *) s1; + ptr5 = (orc_int8 *) s2; + ptr6 = (orc_int8 *) s3; + + /* 2: loadpw */ + var36.i = 0x00000010; /* 16 or 7.90505e-323f */ + /* 4: loadpw */ + var37.i = p1; + /* 8: loadpw */ + var39.i = 0x00000080; /* 128 or 6.32404e-322f */ + /* 10: loadpw */ + var40.i = p2; + /* 15: loadpw */ + var42.i = 0x00000080; /* 128 or 6.32404e-322f */ + /* 17: loadpw */ + var43.i = p3; + /* 20: loadpw */ + var44.i = 0x00000080; /* 128 or 6.32404e-322f */ + + for (i = 0; i < n; i++) { + /* 0: loadb */ + var35 = ptr4[i]; + /* 1: convubw */ + var46.i = (orc_uint8) var35; + /* 3: subw */ + var47.i = var46.i - var36.i; + /* 5: mullw */ + var48.i = (var47.i * var37.i) & 0xffff; + /* 6: loadb */ + var38 = ptr5[i]; + /* 7: convubw */ + var49.i = (orc_uint8) var38; + /* 9: subw */ + var50.i = var49.i - var39.i; + /* 11: mullw */ + var51.i = (var50.i * var40.i) & 0xffff; + /* 12: addw */ + var52.i = var48.i + var51.i; + /* 13: loadb */ + var41 = ptr6[i]; + /* 14: convubw */ + var53.i = (orc_uint8) var41; + /* 16: subw */ + var54.i = var53.i - var42.i; + /* 18: mullw */ + var55.i = (var54.i * var43.i) & 0xffff; + /* 19: addw */ + var56.i = var52.i + var55.i; + /* 21: addw */ + var57.i = var56.i + var44.i; + /* 22: shrsw */ + var58.i = var57.i >> 8; + /* 23: addw */ + var59.i = var58.i + var47.i; + /* 24: convsuswb */ + var45 = ORC_CLAMP_UB (var59.i); + /* 25: storeb */ + ptr0[i] = var45; + } + +} + +#else +static void +_backup_orc_matrix3_100_u8 (OrcExecutor * ex) +{ + int i; + int n = ex->n; + orc_int8 *ORC_RESTRICT ptr0; + const orc_int8 *ORC_RESTRICT ptr4; + const orc_int8 *ORC_RESTRICT ptr5; + const orc_int8 *ORC_RESTRICT ptr6; + orc_int8 var35; + orc_union16 var36; + orc_union16 var37; + orc_int8 var38; + orc_union16 var39; + orc_union16 var40; + orc_int8 var41; + orc_union16 var42; + orc_union16 var43; + orc_union16 var44; + orc_int8 var45; + orc_union16 var46; + orc_union16 var47; + orc_union16 var48; + orc_union16 var49; + orc_union16 var50; + orc_union16 var51; + orc_union16 var52; + orc_union16 var53; + orc_union16 var54; + orc_union16 var55; + orc_union16 var56; + orc_union16 var57; + orc_union16 var58; + orc_union16 var59; + + ptr0 = (orc_int8 *) ex->arrays[0]; + ptr4 = (orc_int8 *) ex->arrays[4]; + ptr5 = (orc_int8 *) ex->arrays[5]; + ptr6 = (orc_int8 *) ex->arrays[6]; + + /* 2: loadpw */ + var36.i = 0x00000010; /* 16 or 7.90505e-323f */ + /* 4: loadpw */ + var37.i = ex->params[24]; + /* 8: loadpw */ + var39.i = 0x00000080; /* 128 or 6.32404e-322f */ + /* 10: loadpw */ + var40.i = ex->params[25]; + /* 15: loadpw */ + var42.i = 0x00000080; /* 128 or 6.32404e-322f */ + /* 17: loadpw */ + var43.i = ex->params[26]; + /* 20: loadpw */ + var44.i = 0x00000080; /* 128 or 6.32404e-322f */ + + for (i = 0; i < n; i++) { + /* 0: loadb */ + var35 = ptr4[i]; + /* 1: convubw */ + var46.i = (orc_uint8) var35; + /* 3: subw */ + var47.i = var46.i - var36.i; + /* 5: mullw */ + var48.i = (var47.i * var37.i) & 0xffff; + /* 6: loadb */ + var38 = ptr5[i]; + /* 7: convubw */ + var49.i = (orc_uint8) var38; + /* 9: subw */ + var50.i = var49.i - var39.i; + /* 11: mullw */ + var51.i = (var50.i * var40.i) & 0xffff; + /* 12: addw */ + var52.i = var48.i + var51.i; + /* 13: loadb */ + var41 = ptr6[i]; + /* 14: convubw */ + var53.i = (orc_uint8) var41; + /* 16: subw */ + var54.i = var53.i - var42.i; + /* 18: mullw */ + var55.i = (var54.i * var43.i) & 0xffff; + /* 19: addw */ + var56.i = var52.i + var55.i; + /* 21: addw */ + var57.i = var56.i + var44.i; + /* 22: shrsw */ + var58.i = var57.i >> 8; + /* 23: addw */ + var59.i = var58.i + var47.i; + /* 24: convsuswb */ + var45 = ORC_CLAMP_UB (var59.i); + /* 25: storeb */ + ptr0[i] = var45; + } + +} + +void +orc_matrix3_100_u8 (guint8 * d1, const guint8 * s1, const guint8 * s2, + const guint8 * s3, int p1, int p2, int p3, int n) +{ + OrcExecutor _ex, *ex = &_ex; + static int p_inited = 0; + static OrcProgram *p = 0; + void (*func) (OrcExecutor *); + + if (!p_inited) { + orc_once_mutex_lock (); + if (!p_inited) { + OrcCompileResult result; + + p = orc_program_new (); + orc_program_set_name (p, "orc_matrix3_100_u8"); + orc_program_set_backup_function (p, _backup_orc_matrix3_100_u8); + orc_program_add_destination (p, 1, "d1"); + orc_program_add_source (p, 1, "s1"); + orc_program_add_source (p, 1, "s2"); + orc_program_add_source (p, 1, "s3"); + orc_program_add_constant (p, 4, 0x00000010, "c1"); + orc_program_add_constant (p, 4, 0x00000080, "c2"); + orc_program_add_constant (p, 4, 0x00000008, "c3"); + orc_program_add_constant (p, 0, 0x00000000, "c4"); + orc_program_add_constant (p, 0, 0x00000000, "c5"); + orc_program_add_constant (p, 0, 0x00000000, "c6"); + orc_program_add_constant (p, 0, 0x00000000, "c7"); + orc_program_add_constant (p, 0, 0x00000000, "c8"); + orc_program_add_parameter (p, 2, "p1"); + orc_program_add_parameter (p, 2, "p2"); + orc_program_add_parameter (p, 2, "p3"); + orc_program_add_temporary (p, 2, "t1"); + orc_program_add_temporary (p, 2, "t2"); + orc_program_add_temporary (p, 2, "t3"); + + orc_program_append_2 (p, "convubw", 0, ORC_VAR_T1, ORC_VAR_S1, ORC_VAR_D1, + ORC_VAR_D1); + orc_program_append_2 (p, "subw", 0, ORC_VAR_T1, ORC_VAR_T1, ORC_VAR_C1, + ORC_VAR_D1); + orc_program_append_2 (p, "mullw", 0, ORC_VAR_T3, ORC_VAR_T1, ORC_VAR_P1, + ORC_VAR_D1); + orc_program_append_2 (p, "convubw", 0, ORC_VAR_T2, ORC_VAR_S2, ORC_VAR_D1, + ORC_VAR_D1); + orc_program_append_2 (p, "subw", 0, ORC_VAR_T2, ORC_VAR_T2, ORC_VAR_C2, + ORC_VAR_D1); + orc_program_append_2 (p, "mullw", 0, ORC_VAR_T2, ORC_VAR_T2, ORC_VAR_P2, + ORC_VAR_D1); + orc_program_append_2 (p, "addw", 0, ORC_VAR_T3, ORC_VAR_T3, ORC_VAR_T2, + ORC_VAR_D1); + orc_program_append_2 (p, "convubw", 0, ORC_VAR_T2, ORC_VAR_S3, ORC_VAR_D1, + ORC_VAR_D1); + orc_program_append_2 (p, "subw", 0, ORC_VAR_T2, ORC_VAR_T2, ORC_VAR_C2, + ORC_VAR_D1); + orc_program_append_2 (p, "mullw", 0, ORC_VAR_T2, ORC_VAR_T2, ORC_VAR_P3, + ORC_VAR_D1); + orc_program_append_2 (p, "addw", 0, ORC_VAR_T3, ORC_VAR_T3, ORC_VAR_T2, + ORC_VAR_D1); + orc_program_append_2 (p, "addw", 0, ORC_VAR_T3, ORC_VAR_T3, ORC_VAR_C2, + ORC_VAR_D1); + orc_program_append_2 (p, "shrsw", 0, ORC_VAR_T3, ORC_VAR_T3, ORC_VAR_C3, + ORC_VAR_D1); + orc_program_append_2 (p, "addw", 0, ORC_VAR_T3, ORC_VAR_T3, ORC_VAR_T1, + ORC_VAR_D1); + orc_program_append_2 (p, "convsuswb", 0, ORC_VAR_D1, ORC_VAR_T3, + ORC_VAR_D1, ORC_VAR_D1); + + result = orc_program_compile (p); + } + p_inited = TRUE; + orc_once_mutex_unlock (); + } + ex->program = p; + + ex->n = n; + ex->arrays[ORC_VAR_D1] = d1; + ex->arrays[ORC_VAR_S1] = (void *) s1; + ex->arrays[ORC_VAR_S2] = (void *) s2; + ex->arrays[ORC_VAR_S3] = (void *) s3; + ex->params[ORC_VAR_P1] = p1; + ex->params[ORC_VAR_P2] = p2; + ex->params[ORC_VAR_P3] = p3; + + func = p->code_exec; + func (ex); +} +#endif + + +/* orc_matrix3_100_offset_u8 */ +#ifdef DISABLE_ORC +void +orc_matrix3_100_offset_u8 (guint8 * d1, const guint8 * s1, const guint8 * s2, + const guint8 * s3, int p1, int p2, int p3, int p4, int p5, int n) +{ + int i; + orc_int8 *ORC_RESTRICT ptr0; + const orc_int8 *ORC_RESTRICT ptr4; + const orc_int8 *ORC_RESTRICT ptr5; + const orc_int8 *ORC_RESTRICT ptr6; + orc_int8 var35; + orc_union16 var36; + orc_int8 var37; + orc_union16 var38; + orc_int8 var39; + orc_union16 var40; + orc_union16 var41; + orc_int8 var42; + orc_union16 var43; + orc_union16 var44; + orc_union16 var45; + orc_union16 var46; + orc_union16 var47; + orc_union16 var48; + orc_union16 var49; + orc_union16 var50; + orc_union16 var51; + orc_union16 var52; + orc_union16 var53; + + ptr0 = (orc_int8 *) d1; + ptr4 = (orc_int8 *) s1; + ptr5 = (orc_int8 *) s2; + ptr6 = (orc_int8 *) s3; + + /* 2: loadpw */ + var36.i = p1; + /* 6: loadpw */ + var38.i = p2; + /* 11: loadpw */ + var40.i = p3; + /* 14: loadpw */ + var41.i = p4; + + for (i = 0; i < n; i++) { + /* 0: loadb */ + var35 = ptr4[i]; + /* 1: convubw */ + var43.i = (orc_uint8) var35; + /* 3: mullw */ + var44.i = (var43.i * var36.i) & 0xffff; + /* 4: loadb */ + var37 = ptr5[i]; + /* 5: convubw */ + var45.i = (orc_uint8) var37; + /* 7: mullw */ + var46.i = (var45.i * var38.i) & 0xffff; + /* 8: addw */ + var47.i = var44.i + var46.i; + /* 9: loadb */ + var39 = ptr6[i]; + /* 10: convubw */ + var48.i = (orc_uint8) var39; + /* 12: mullw */ + var49.i = (var48.i * var40.i) & 0xffff; + /* 13: addw */ + var50.i = var47.i + var49.i; + /* 15: addw */ + var51.i = var50.i + var41.i; + /* 16: shrsw */ + var52.i = var51.i >> p5; + /* 17: addw */ + var53.i = var52.i + var43.i; + /* 18: convsuswb */ + var42 = ORC_CLAMP_UB (var53.i); + /* 19: storeb */ + ptr0[i] = var42; + } + +} + +#else +static void +_backup_orc_matrix3_100_offset_u8 (OrcExecutor * ex) +{ + int i; + int n = ex->n; + orc_int8 *ORC_RESTRICT ptr0; + const orc_int8 *ORC_RESTRICT ptr4; + const orc_int8 *ORC_RESTRICT ptr5; + const orc_int8 *ORC_RESTRICT ptr6; + orc_int8 var35; + orc_union16 var36; + orc_int8 var37; + orc_union16 var38; + orc_int8 var39; + orc_union16 var40; + orc_union16 var41; + orc_int8 var42; + orc_union16 var43; + orc_union16 var44; + orc_union16 var45; + orc_union16 var46; + orc_union16 var47; + orc_union16 var48; + orc_union16 var49; + orc_union16 var50; + orc_union16 var51; + orc_union16 var52; + orc_union16 var53; + + ptr0 = (orc_int8 *) ex->arrays[0]; + ptr4 = (orc_int8 *) ex->arrays[4]; + ptr5 = (orc_int8 *) ex->arrays[5]; + ptr6 = (orc_int8 *) ex->arrays[6]; + + /* 2: loadpw */ + var36.i = ex->params[24]; + /* 6: loadpw */ + var38.i = ex->params[25]; + /* 11: loadpw */ + var40.i = ex->params[26]; + /* 14: loadpw */ + var41.i = ex->params[27]; + + for (i = 0; i < n; i++) { + /* 0: loadb */ + var35 = ptr4[i]; + /* 1: convubw */ + var43.i = (orc_uint8) var35; + /* 3: mullw */ + var44.i = (var43.i * var36.i) & 0xffff; + /* 4: loadb */ + var37 = ptr5[i]; + /* 5: convubw */ + var45.i = (orc_uint8) var37; + /* 7: mullw */ + var46.i = (var45.i * var38.i) & 0xffff; + /* 8: addw */ + var47.i = var44.i + var46.i; + /* 9: loadb */ + var39 = ptr6[i]; + /* 10: convubw */ + var48.i = (orc_uint8) var39; + /* 12: mullw */ + var49.i = (var48.i * var40.i) & 0xffff; + /* 13: addw */ + var50.i = var47.i + var49.i; + /* 15: addw */ + var51.i = var50.i + var41.i; + /* 16: shrsw */ + var52.i = var51.i >> ex->params[28]; + /* 17: addw */ + var53.i = var52.i + var43.i; + /* 18: convsuswb */ + var42 = ORC_CLAMP_UB (var53.i); + /* 19: storeb */ + ptr0[i] = var42; + } + +} + +void +orc_matrix3_100_offset_u8 (guint8 * d1, const guint8 * s1, const guint8 * s2, + const guint8 * s3, int p1, int p2, int p3, int p4, int p5, int n) +{ + OrcExecutor _ex, *ex = &_ex; + static int p_inited = 0; + static OrcProgram *p = 0; + void (*func) (OrcExecutor *); + + if (!p_inited) { + orc_once_mutex_lock (); + if (!p_inited) { + OrcCompileResult result; + + p = orc_program_new (); + orc_program_set_name (p, "orc_matrix3_100_offset_u8"); + orc_program_set_backup_function (p, _backup_orc_matrix3_100_offset_u8); + orc_program_add_destination (p, 1, "d1"); + orc_program_add_source (p, 1, "s1"); + orc_program_add_source (p, 1, "s2"); + orc_program_add_source (p, 1, "s3"); + orc_program_add_constant (p, 0, 0x00000000, "c1"); + orc_program_add_constant (p, 0, 0x00000000, "c2"); + orc_program_add_constant (p, 0, 0x00000000, "c3"); + orc_program_add_constant (p, 0, 0x00000000, "c4"); + orc_program_add_constant (p, 0, 0x00000000, "c5"); + orc_program_add_constant (p, 0, 0x00000000, "c6"); + orc_program_add_constant (p, 0, 0x00000000, "c7"); + orc_program_add_constant (p, 0, 0x00000000, "c8"); + orc_program_add_parameter (p, 2, "p1"); + orc_program_add_parameter (p, 2, "p2"); + orc_program_add_parameter (p, 2, "p3"); + orc_program_add_parameter (p, 2, "p4"); + orc_program_add_parameter (p, 2, "p5"); + orc_program_add_temporary (p, 2, "t1"); + orc_program_add_temporary (p, 2, "t2"); + orc_program_add_temporary (p, 2, "t3"); + + orc_program_append_2 (p, "convubw", 0, ORC_VAR_T3, ORC_VAR_S1, ORC_VAR_D1, + ORC_VAR_D1); + orc_program_append_2 (p, "mullw", 0, ORC_VAR_T1, ORC_VAR_T3, ORC_VAR_P1, + ORC_VAR_D1); + orc_program_append_2 (p, "convubw", 0, ORC_VAR_T2, ORC_VAR_S2, ORC_VAR_D1, + ORC_VAR_D1); + orc_program_append_2 (p, "mullw", 0, ORC_VAR_T2, ORC_VAR_T2, ORC_VAR_P2, + ORC_VAR_D1); + orc_program_append_2 (p, "addw", 0, ORC_VAR_T1, ORC_VAR_T1, ORC_VAR_T2, + ORC_VAR_D1); + orc_program_append_2 (p, "convubw", 0, ORC_VAR_T2, ORC_VAR_S3, ORC_VAR_D1, + ORC_VAR_D1); + orc_program_append_2 (p, "mullw", 0, ORC_VAR_T2, ORC_VAR_T2, ORC_VAR_P3, + ORC_VAR_D1); + orc_program_append_2 (p, "addw", 0, ORC_VAR_T1, ORC_VAR_T1, ORC_VAR_T2, + ORC_VAR_D1); + orc_program_append_2 (p, "addw", 0, ORC_VAR_T1, ORC_VAR_T1, ORC_VAR_P4, + ORC_VAR_D1); + orc_program_append_2 (p, "shrsw", 0, ORC_VAR_T1, ORC_VAR_T1, ORC_VAR_P5, + ORC_VAR_D1); + orc_program_append_2 (p, "addw", 0, ORC_VAR_T1, ORC_VAR_T1, ORC_VAR_T3, + ORC_VAR_D1); + orc_program_append_2 (p, "convsuswb", 0, ORC_VAR_D1, ORC_VAR_T1, + ORC_VAR_D1, ORC_VAR_D1); + + result = orc_program_compile (p); + } + p_inited = TRUE; + orc_once_mutex_unlock (); + } + ex->program = p; + + ex->n = n; + ex->arrays[ORC_VAR_D1] = d1; + ex->arrays[ORC_VAR_S1] = (void *) s1; + ex->arrays[ORC_VAR_S2] = (void *) s2; + ex->arrays[ORC_VAR_S3] = (void *) s3; + ex->params[ORC_VAR_P1] = p1; + ex->params[ORC_VAR_P2] = p2; + ex->params[ORC_VAR_P3] = p3; + ex->params[ORC_VAR_P4] = p4; + ex->params[ORC_VAR_P5] = p5; + + func = p->code_exec; + func (ex); +} +#endif + + +/* orc_matrix3_000_u8 */ +#ifdef DISABLE_ORC +void +orc_matrix3_000_u8 (guint8 * d1, const guint8 * s1, const guint8 * s2, + const guint8 * s3, int p1, int p2, int p3, int p4, int p5, int n) +{ + int i; + orc_int8 *ORC_RESTRICT ptr0; + const orc_int8 *ORC_RESTRICT ptr4; + const orc_int8 *ORC_RESTRICT ptr5; + const orc_int8 *ORC_RESTRICT ptr6; + orc_int8 var34; + orc_union16 var35; + orc_int8 var36; + orc_union16 var37; + orc_int8 var38; + orc_union16 var39; + orc_union16 var40; + orc_int8 var41; + orc_union16 var42; + orc_union16 var43; + orc_union16 var44; + orc_union16 var45; + orc_union16 var46; + orc_union16 var47; + orc_union16 var48; + orc_union16 var49; + orc_union16 var50; + orc_union16 var51; + + ptr0 = (orc_int8 *) d1; + ptr4 = (orc_int8 *) s1; + ptr5 = (orc_int8 *) s2; + ptr6 = (orc_int8 *) s3; + + /* 2: loadpw */ + var35.i = p1; + /* 6: loadpw */ + var37.i = p2; + /* 11: loadpw */ + var39.i = p3; + /* 14: loadpw */ + var40.i = p4; + + for (i = 0; i < n; i++) { + /* 0: loadb */ + var34 = ptr4[i]; + /* 1: convubw */ + var42.i = (orc_uint8) var34; + /* 3: mullw */ + var43.i = (var42.i * var35.i) & 0xffff; + /* 4: loadb */ + var36 = ptr5[i]; + /* 5: convubw */ + var44.i = (orc_uint8) var36; + /* 7: mullw */ + var45.i = (var44.i * var37.i) & 0xffff; + /* 8: addw */ + var46.i = var43.i + var45.i; + /* 9: loadb */ + var38 = ptr6[i]; + /* 10: convubw */ + var47.i = (orc_uint8) var38; + /* 12: mullw */ + var48.i = (var47.i * var39.i) & 0xffff; + /* 13: addw */ + var49.i = var46.i + var48.i; + /* 15: addw */ + var50.i = var49.i + var40.i; + /* 16: shrsw */ + var51.i = var50.i >> p5; + /* 17: convwb */ + var41 = var51.i; + /* 18: storeb */ + ptr0[i] = var41; + } + +} + +#else +static void +_backup_orc_matrix3_000_u8 (OrcExecutor * ex) +{ + int i; + int n = ex->n; + orc_int8 *ORC_RESTRICT ptr0; + const orc_int8 *ORC_RESTRICT ptr4; + const orc_int8 *ORC_RESTRICT ptr5; + const orc_int8 *ORC_RESTRICT ptr6; + orc_int8 var34; + orc_union16 var35; + orc_int8 var36; + orc_union16 var37; + orc_int8 var38; + orc_union16 var39; + orc_union16 var40; + orc_int8 var41; + orc_union16 var42; + orc_union16 var43; + orc_union16 var44; + orc_union16 var45; + orc_union16 var46; + orc_union16 var47; + orc_union16 var48; + orc_union16 var49; + orc_union16 var50; + orc_union16 var51; + + ptr0 = (orc_int8 *) ex->arrays[0]; + ptr4 = (orc_int8 *) ex->arrays[4]; + ptr5 = (orc_int8 *) ex->arrays[5]; + ptr6 = (orc_int8 *) ex->arrays[6]; + + /* 2: loadpw */ + var35.i = ex->params[24]; + /* 6: loadpw */ + var37.i = ex->params[25]; + /* 11: loadpw */ + var39.i = ex->params[26]; + /* 14: loadpw */ + var40.i = ex->params[27]; + + for (i = 0; i < n; i++) { + /* 0: loadb */ + var34 = ptr4[i]; + /* 1: convubw */ + var42.i = (orc_uint8) var34; + /* 3: mullw */ + var43.i = (var42.i * var35.i) & 0xffff; + /* 4: loadb */ + var36 = ptr5[i]; + /* 5: convubw */ + var44.i = (orc_uint8) var36; + /* 7: mullw */ + var45.i = (var44.i * var37.i) & 0xffff; + /* 8: addw */ + var46.i = var43.i + var45.i; + /* 9: loadb */ + var38 = ptr6[i]; + /* 10: convubw */ + var47.i = (orc_uint8) var38; + /* 12: mullw */ + var48.i = (var47.i * var39.i) & 0xffff; + /* 13: addw */ + var49.i = var46.i + var48.i; + /* 15: addw */ + var50.i = var49.i + var40.i; + /* 16: shrsw */ + var51.i = var50.i >> ex->params[28]; + /* 17: convwb */ + var41 = var51.i; + /* 18: storeb */ + ptr0[i] = var41; + } + +} + +void +orc_matrix3_000_u8 (guint8 * d1, const guint8 * s1, const guint8 * s2, + const guint8 * s3, int p1, int p2, int p3, int p4, int p5, int n) +{ + OrcExecutor _ex, *ex = &_ex; + static int p_inited = 0; + static OrcProgram *p = 0; + void (*func) (OrcExecutor *); + + if (!p_inited) { + orc_once_mutex_lock (); + if (!p_inited) { + OrcCompileResult result; + + p = orc_program_new (); + orc_program_set_name (p, "orc_matrix3_000_u8"); + orc_program_set_backup_function (p, _backup_orc_matrix3_000_u8); + orc_program_add_destination (p, 1, "d1"); + orc_program_add_source (p, 1, "s1"); + orc_program_add_source (p, 1, "s2"); + orc_program_add_source (p, 1, "s3"); + orc_program_add_constant (p, 0, 0x00000000, "c1"); + orc_program_add_constant (p, 0, 0x00000000, "c2"); + orc_program_add_constant (p, 0, 0x00000000, "c3"); + orc_program_add_constant (p, 0, 0x00000000, "c4"); + orc_program_add_constant (p, 0, 0x00000000, "c5"); + orc_program_add_constant (p, 0, 0x00000000, "c6"); + orc_program_add_constant (p, 0, 0x00000000, "c7"); + orc_program_add_constant (p, 0, 0x00000000, "c8"); + orc_program_add_parameter (p, 2, "p1"); + orc_program_add_parameter (p, 2, "p2"); + orc_program_add_parameter (p, 2, "p3"); + orc_program_add_parameter (p, 2, "p4"); + orc_program_add_parameter (p, 2, "p5"); + orc_program_add_temporary (p, 2, "t1"); + orc_program_add_temporary (p, 2, "t2"); + + orc_program_append_2 (p, "convubw", 0, ORC_VAR_T1, ORC_VAR_S1, ORC_VAR_D1, + ORC_VAR_D1); + orc_program_append_2 (p, "mullw", 0, ORC_VAR_T1, ORC_VAR_T1, ORC_VAR_P1, + ORC_VAR_D1); + orc_program_append_2 (p, "convubw", 0, ORC_VAR_T2, ORC_VAR_S2, ORC_VAR_D1, + ORC_VAR_D1); + orc_program_append_2 (p, "mullw", 0, ORC_VAR_T2, ORC_VAR_T2, ORC_VAR_P2, + ORC_VAR_D1); + orc_program_append_2 (p, "addw", 0, ORC_VAR_T1, ORC_VAR_T1, ORC_VAR_T2, + ORC_VAR_D1); + orc_program_append_2 (p, "convubw", 0, ORC_VAR_T2, ORC_VAR_S3, ORC_VAR_D1, + ORC_VAR_D1); + orc_program_append_2 (p, "mullw", 0, ORC_VAR_T2, ORC_VAR_T2, ORC_VAR_P3, + ORC_VAR_D1); + orc_program_append_2 (p, "addw", 0, ORC_VAR_T1, ORC_VAR_T1, ORC_VAR_T2, + ORC_VAR_D1); + orc_program_append_2 (p, "addw", 0, ORC_VAR_T1, ORC_VAR_T1, ORC_VAR_P4, + ORC_VAR_D1); + orc_program_append_2 (p, "shrsw", 0, ORC_VAR_T1, ORC_VAR_T1, ORC_VAR_P5, + ORC_VAR_D1); + orc_program_append_2 (p, "convwb", 0, ORC_VAR_D1, ORC_VAR_T1, ORC_VAR_D1, + ORC_VAR_D1); + + result = orc_program_compile (p); + } + p_inited = TRUE; + orc_once_mutex_unlock (); + } + ex->program = p; + + ex->n = n; + ex->arrays[ORC_VAR_D1] = d1; + ex->arrays[ORC_VAR_S1] = (void *) s1; + ex->arrays[ORC_VAR_S2] = (void *) s2; + ex->arrays[ORC_VAR_S3] = (void *) s3; + ex->params[ORC_VAR_P1] = p1; + ex->params[ORC_VAR_P2] = p2; + ex->params[ORC_VAR_P3] = p3; + ex->params[ORC_VAR_P4] = p4; + ex->params[ORC_VAR_P5] = p5; + + func = p->code_exec; + func (ex); +} +#endif + + +/* orc_pack_123x */ +#ifdef DISABLE_ORC +void +orc_pack_123x (guint32 * d1, const guint8 * s1, const guint8 * s2, + const guint8 * s3, int p1, int n) +{ + int i; + orc_union32 *ORC_RESTRICT ptr0; + const orc_int8 *ORC_RESTRICT ptr4; + const orc_int8 *ORC_RESTRICT ptr5; + const orc_int8 *ORC_RESTRICT ptr6; + orc_int8 var34; + orc_int8 var35; + orc_int8 var36; + orc_int8 var37; + orc_union32 var38; + orc_union16 var39; + orc_union16 var40; + + ptr0 = (orc_union32 *) d1; + ptr4 = (orc_int8 *) s1; + ptr5 = (orc_int8 *) s2; + ptr6 = (orc_int8 *) s3; + + /* 4: loadpb */ + var37 = p1; + + for (i = 0; i < n; i++) { + /* 0: loadb */ + var34 = ptr4[i]; + /* 1: loadb */ + var35 = ptr5[i]; + /* 2: mergebw */ + var39.i = ((orc_uint8) var34) | ((orc_uint8) var35 << 8); + /* 3: loadb */ + var36 = ptr6[i]; + /* 5: mergebw */ + var40.i = ((orc_uint8) var36) | ((orc_uint8) var37 << 8); + /* 6: mergewl */ + var38.i = ((orc_uint16) var39.i) | ((orc_uint16) var40.i << 16); + /* 7: storel */ + ptr0[i] = var38; + } + +} + +#else +static void +_backup_orc_pack_123x (OrcExecutor * ex) +{ + int i; + int n = ex->n; + orc_union32 *ORC_RESTRICT ptr0; + const orc_int8 *ORC_RESTRICT ptr4; + const orc_int8 *ORC_RESTRICT ptr5; + const orc_int8 *ORC_RESTRICT ptr6; + orc_int8 var34; + orc_int8 var35; + orc_int8 var36; + orc_int8 var37; + orc_union32 var38; + orc_union16 var39; + orc_union16 var40; + + ptr0 = (orc_union32 *) ex->arrays[0]; + ptr4 = (orc_int8 *) ex->arrays[4]; + ptr5 = (orc_int8 *) ex->arrays[5]; + ptr6 = (orc_int8 *) ex->arrays[6]; + + /* 4: loadpb */ + var37 = ex->params[24]; + + for (i = 0; i < n; i++) { + /* 0: loadb */ + var34 = ptr4[i]; + /* 1: loadb */ + var35 = ptr5[i]; + /* 2: mergebw */ + var39.i = ((orc_uint8) var34) | ((orc_uint8) var35 << 8); + /* 3: loadb */ + var36 = ptr6[i]; + /* 5: mergebw */ + var40.i = ((orc_uint8) var36) | ((orc_uint8) var37 << 8); + /* 6: mergewl */ + var38.i = ((orc_uint16) var39.i) | ((orc_uint16) var40.i << 16); + /* 7: storel */ + ptr0[i] = var38; + } + +} + +void +orc_pack_123x (guint32 * d1, const guint8 * s1, const guint8 * s2, + const guint8 * s3, int p1, int n) +{ + OrcExecutor _ex, *ex = &_ex; + static int p_inited = 0; + static OrcProgram *p = 0; + void (*func) (OrcExecutor *); + + if (!p_inited) { + orc_once_mutex_lock (); + if (!p_inited) { + OrcCompileResult result; + + p = orc_program_new (); + orc_program_set_name (p, "orc_pack_123x"); + orc_program_set_backup_function (p, _backup_orc_pack_123x); + orc_program_add_destination (p, 4, "d1"); + orc_program_add_source (p, 1, "s1"); + orc_program_add_source (p, 1, "s2"); + orc_program_add_source (p, 1, "s3"); + orc_program_add_constant (p, 0, 0x00000000, "c1"); + orc_program_add_constant (p, 0, 0x00000000, "c2"); + orc_program_add_constant (p, 0, 0x00000000, "c3"); + orc_program_add_constant (p, 0, 0x00000000, "c4"); + orc_program_add_constant (p, 0, 0x00000000, "c5"); + orc_program_add_constant (p, 0, 0x00000000, "c6"); + orc_program_add_constant (p, 0, 0x00000000, "c7"); + orc_program_add_constant (p, 0, 0x00000000, "c8"); + orc_program_add_parameter (p, 1, "p1"); + orc_program_add_temporary (p, 2, "t1"); + orc_program_add_temporary (p, 2, "t2"); + + orc_program_append_2 (p, "mergebw", 0, ORC_VAR_T1, ORC_VAR_S1, ORC_VAR_S2, + ORC_VAR_D1); + orc_program_append_2 (p, "mergebw", 0, ORC_VAR_T2, ORC_VAR_S3, ORC_VAR_P1, + ORC_VAR_D1); + orc_program_append_2 (p, "mergewl", 0, ORC_VAR_D1, ORC_VAR_T1, ORC_VAR_T2, + ORC_VAR_D1); + + result = orc_program_compile (p); + } + p_inited = TRUE; + orc_once_mutex_unlock (); + } + ex->program = p; + + ex->n = n; + ex->arrays[ORC_VAR_D1] = d1; + ex->arrays[ORC_VAR_S1] = (void *) s1; + ex->arrays[ORC_VAR_S2] = (void *) s2; + ex->arrays[ORC_VAR_S3] = (void *) s3; + ex->params[ORC_VAR_P1] = p1; + + func = p->code_exec; + func (ex); +} +#endif + + +/* orc_pack_x123 */ +#ifdef DISABLE_ORC +void +orc_pack_x123 (guint32 * d1, const guint8 * s1, const guint8 * s2, + const guint8 * s3, int p1, int n) +{ + int i; + orc_union32 *ORC_RESTRICT ptr0; + const orc_int8 *ORC_RESTRICT ptr4; + const orc_int8 *ORC_RESTRICT ptr5; + const orc_int8 *ORC_RESTRICT ptr6; + orc_int8 var34; + orc_int8 var35; + orc_int8 var36; + orc_int8 var37; + orc_union32 var38; + orc_union16 var39; + orc_union16 var40; + + ptr0 = (orc_union32 *) d1; + ptr4 = (orc_int8 *) s1; + ptr5 = (orc_int8 *) s2; + ptr6 = (orc_int8 *) s3; + + /* 0: loadpb */ + var34 = p1; + + for (i = 0; i < n; i++) { + /* 1: loadb */ + var35 = ptr4[i]; + /* 2: mergebw */ + var39.i = ((orc_uint8) var34) | ((orc_uint8) var35 << 8); + /* 3: loadb */ + var36 = ptr5[i]; + /* 4: loadb */ + var37 = ptr6[i]; + /* 5: mergebw */ + var40.i = ((orc_uint8) var36) | ((orc_uint8) var37 << 8); + /* 6: mergewl */ + var38.i = ((orc_uint16) var39.i) | ((orc_uint16) var40.i << 16); + /* 7: storel */ + ptr0[i] = var38; + } + +} + +#else +static void +_backup_orc_pack_x123 (OrcExecutor * ex) +{ + int i; + int n = ex->n; + orc_union32 *ORC_RESTRICT ptr0; + const orc_int8 *ORC_RESTRICT ptr4; + const orc_int8 *ORC_RESTRICT ptr5; + const orc_int8 *ORC_RESTRICT ptr6; + orc_int8 var34; + orc_int8 var35; + orc_int8 var36; + orc_int8 var37; + orc_union32 var38; + orc_union16 var39; + orc_union16 var40; + + ptr0 = (orc_union32 *) ex->arrays[0]; + ptr4 = (orc_int8 *) ex->arrays[4]; + ptr5 = (orc_int8 *) ex->arrays[5]; + ptr6 = (orc_int8 *) ex->arrays[6]; + + /* 0: loadpb */ + var34 = ex->params[24]; + + for (i = 0; i < n; i++) { + /* 1: loadb */ + var35 = ptr4[i]; + /* 2: mergebw */ + var39.i = ((orc_uint8) var34) | ((orc_uint8) var35 << 8); + /* 3: loadb */ + var36 = ptr5[i]; + /* 4: loadb */ + var37 = ptr6[i]; + /* 5: mergebw */ + var40.i = ((orc_uint8) var36) | ((orc_uint8) var37 << 8); + /* 6: mergewl */ + var38.i = ((orc_uint16) var39.i) | ((orc_uint16) var40.i << 16); + /* 7: storel */ + ptr0[i] = var38; + } + +} + +void +orc_pack_x123 (guint32 * d1, const guint8 * s1, const guint8 * s2, + const guint8 * s3, int p1, int n) +{ + OrcExecutor _ex, *ex = &_ex; + static int p_inited = 0; + static OrcProgram *p = 0; + void (*func) (OrcExecutor *); + + if (!p_inited) { + orc_once_mutex_lock (); + if (!p_inited) { + OrcCompileResult result; + + p = orc_program_new (); + orc_program_set_name (p, "orc_pack_x123"); + orc_program_set_backup_function (p, _backup_orc_pack_x123); + orc_program_add_destination (p, 4, "d1"); + orc_program_add_source (p, 1, "s1"); + orc_program_add_source (p, 1, "s2"); + orc_program_add_source (p, 1, "s3"); + orc_program_add_constant (p, 0, 0x00000000, "c1"); + orc_program_add_constant (p, 0, 0x00000000, "c2"); + orc_program_add_constant (p, 0, 0x00000000, "c3"); + orc_program_add_constant (p, 0, 0x00000000, "c4"); + orc_program_add_constant (p, 0, 0x00000000, "c5"); + orc_program_add_constant (p, 0, 0x00000000, "c6"); + orc_program_add_constant (p, 0, 0x00000000, "c7"); + orc_program_add_constant (p, 0, 0x00000000, "c8"); + orc_program_add_parameter (p, 1, "p1"); + orc_program_add_temporary (p, 2, "t1"); + orc_program_add_temporary (p, 2, "t2"); + + orc_program_append_2 (p, "mergebw", 0, ORC_VAR_T1, ORC_VAR_P1, ORC_VAR_S1, + ORC_VAR_D1); + orc_program_append_2 (p, "mergebw", 0, ORC_VAR_T2, ORC_VAR_S2, ORC_VAR_S3, + ORC_VAR_D1); + orc_program_append_2 (p, "mergewl", 0, ORC_VAR_D1, ORC_VAR_T1, ORC_VAR_T2, + ORC_VAR_D1); + + result = orc_program_compile (p); + } + p_inited = TRUE; + orc_once_mutex_unlock (); + } + ex->program = p; + + ex->n = n; + ex->arrays[ORC_VAR_D1] = d1; + ex->arrays[ORC_VAR_S1] = (void *) s1; + ex->arrays[ORC_VAR_S2] = (void *) s2; + ex->arrays[ORC_VAR_S3] = (void *) s3; + ex->params[ORC_VAR_P1] = p1; + + func = p->code_exec; + func (ex); +} +#endif + + +/* cogorc_combine2_u8 */ +#ifdef DISABLE_ORC +void +cogorc_combine2_u8 (guint8 * d1, const guint8 * s1, const guint8 * s2, int p1, + int p2, int n) +{ + int i; + orc_int8 *ORC_RESTRICT ptr0; + const orc_int8 *ORC_RESTRICT ptr4; + const orc_int8 *ORC_RESTRICT ptr5; + orc_int8 var34; + orc_union16 var35; + orc_int8 var36; + orc_union16 var37; + orc_int8 var38; + orc_union16 var39; + orc_union16 var40; + orc_union16 var41; + orc_union16 var42; + orc_union16 var43; + orc_union16 var44; + + ptr0 = (orc_int8 *) d1; + ptr4 = (orc_int8 *) s1; + ptr5 = (orc_int8 *) s2; + + /* 2: loadpw */ + var35.i = p1; + /* 6: loadpw */ + var37.i = p2; + + for (i = 0; i < n; i++) { + /* 0: loadb */ + var34 = ptr4[i]; + /* 1: convubw */ + var39.i = (orc_uint8) var34; + /* 3: mullw */ + var40.i = (var39.i * var35.i) & 0xffff; + /* 4: loadb */ + var36 = ptr5[i]; + /* 5: convubw */ + var41.i = (orc_uint8) var36; + /* 7: mullw */ + var42.i = (var41.i * var37.i) & 0xffff; + /* 8: addw */ + var43.i = var40.i + var42.i; + /* 9: shruw */ + var44.i = ((orc_uint16) var43.i) >> 8; + /* 10: convsuswb */ + var38 = ORC_CLAMP_UB (var44.i); + /* 11: storeb */ + ptr0[i] = var38; + } + +} + +#else +static void +_backup_cogorc_combine2_u8 (OrcExecutor * ex) +{ + int i; + int n = ex->n; + orc_int8 *ORC_RESTRICT ptr0; + const orc_int8 *ORC_RESTRICT ptr4; + const orc_int8 *ORC_RESTRICT ptr5; + orc_int8 var34; + orc_union16 var35; + orc_int8 var36; + orc_union16 var37; + orc_int8 var38; + orc_union16 var39; + orc_union16 var40; + orc_union16 var41; + orc_union16 var42; + orc_union16 var43; + orc_union16 var44; + + ptr0 = (orc_int8 *) ex->arrays[0]; + ptr4 = (orc_int8 *) ex->arrays[4]; + ptr5 = (orc_int8 *) ex->arrays[5]; + + /* 2: loadpw */ + var35.i = ex->params[24]; + /* 6: loadpw */ + var37.i = ex->params[25]; + + for (i = 0; i < n; i++) { + /* 0: loadb */ + var34 = ptr4[i]; + /* 1: convubw */ + var39.i = (orc_uint8) var34; + /* 3: mullw */ + var40.i = (var39.i * var35.i) & 0xffff; + /* 4: loadb */ + var36 = ptr5[i]; + /* 5: convubw */ + var41.i = (orc_uint8) var36; + /* 7: mullw */ + var42.i = (var41.i * var37.i) & 0xffff; + /* 8: addw */ + var43.i = var40.i + var42.i; + /* 9: shruw */ + var44.i = ((orc_uint16) var43.i) >> 8; + /* 10: convsuswb */ + var38 = ORC_CLAMP_UB (var44.i); + /* 11: storeb */ + ptr0[i] = var38; + } + +} + +void +cogorc_combine2_u8 (guint8 * d1, const guint8 * s1, const guint8 * s2, int p1, + int p2, int n) +{ + OrcExecutor _ex, *ex = &_ex; + static int p_inited = 0; + static OrcProgram *p = 0; + void (*func) (OrcExecutor *); + + if (!p_inited) { + orc_once_mutex_lock (); + if (!p_inited) { + OrcCompileResult result; + + p = orc_program_new (); + orc_program_set_name (p, "cogorc_combine2_u8"); + orc_program_set_backup_function (p, _backup_cogorc_combine2_u8); + orc_program_add_destination (p, 1, "d1"); + orc_program_add_source (p, 1, "s1"); + orc_program_add_source (p, 1, "s2"); + orc_program_add_constant (p, 4, 0x00000008, "c1"); + orc_program_add_constant (p, 0, 0x00000000, "c2"); + orc_program_add_constant (p, 0, 0x00000000, "c3"); + orc_program_add_constant (p, 0, 0x00000000, "c4"); + orc_program_add_constant (p, 0, 0x00000000, "c5"); + orc_program_add_constant (p, 0, 0x00000000, "c6"); + orc_program_add_constant (p, 0, 0x00000000, "c7"); + orc_program_add_constant (p, 0, 0x00000000, "c8"); + orc_program_add_parameter (p, 2, "p1"); + orc_program_add_parameter (p, 2, "p2"); + orc_program_add_temporary (p, 2, "t1"); + orc_program_add_temporary (p, 2, "t2"); + + orc_program_append_2 (p, "convubw", 0, ORC_VAR_T1, ORC_VAR_S1, ORC_VAR_D1, + ORC_VAR_D1); + orc_program_append_2 (p, "mullw", 0, ORC_VAR_T1, ORC_VAR_T1, ORC_VAR_P1, + ORC_VAR_D1); + orc_program_append_2 (p, "convubw", 0, ORC_VAR_T2, ORC_VAR_S2, ORC_VAR_D1, + ORC_VAR_D1); + orc_program_append_2 (p, "mullw", 0, ORC_VAR_T2, ORC_VAR_T2, ORC_VAR_P2, + ORC_VAR_D1); + orc_program_append_2 (p, "addw", 0, ORC_VAR_T1, ORC_VAR_T1, ORC_VAR_T2, + ORC_VAR_D1); + orc_program_append_2 (p, "shruw", 0, ORC_VAR_T1, ORC_VAR_T1, ORC_VAR_C1, + ORC_VAR_D1); + orc_program_append_2 (p, "convsuswb", 0, ORC_VAR_D1, ORC_VAR_T1, + ORC_VAR_D1, ORC_VAR_D1); + + result = orc_program_compile (p); + } + p_inited = TRUE; + orc_once_mutex_unlock (); + } + ex->program = p; + + ex->n = n; + ex->arrays[ORC_VAR_D1] = d1; + ex->arrays[ORC_VAR_S1] = (void *) s1; + ex->arrays[ORC_VAR_S2] = (void *) s2; + ex->params[ORC_VAR_P1] = p1; + ex->params[ORC_VAR_P2] = p2; + + func = p->code_exec; + func (ex); +} +#endif + + +/* cogorc_convert_I420_UYVY */ +#ifdef DISABLE_ORC +void +cogorc_convert_I420_UYVY (guint8 * d1, guint8 * d2, const guint8 * s1, + const guint8 * s2, const guint8 * s3, const guint8 * s4, int n) +{ + int i; + orc_union32 *ORC_RESTRICT ptr0; + orc_union32 *ORC_RESTRICT ptr1; + const orc_union16 *ORC_RESTRICT ptr4; + const orc_union16 *ORC_RESTRICT ptr5; + const orc_int8 *ORC_RESTRICT ptr6; + const orc_int8 *ORC_RESTRICT ptr7; + orc_int8 var33; + orc_int8 var34; + orc_union16 var35; + orc_union32 var36; + orc_union16 var37; + orc_union32 var38; + orc_union16 var39; + + ptr0 = (orc_union32 *) d1; + ptr1 = (orc_union32 *) d2; + ptr4 = (orc_union16 *) s1; + ptr5 = (orc_union16 *) s2; + ptr6 = (orc_int8 *) s3; + ptr7 = (orc_int8 *) s4; + + + for (i = 0; i < n; i++) { + /* 0: loadb */ + var33 = ptr6[i]; + /* 1: loadb */ + var34 = ptr7[i]; + /* 2: mergebw */ + var39.i = ((orc_uint8) var33) | ((orc_uint8) var34 << 8); + /* 3: loadw */ + var35 = ptr4[i]; + /* 4: mergebw */ + var36.x2[0] = ((orc_uint8) var39.x2[0]) | ((orc_uint8) var35.x2[0] << 8); + var36.x2[1] = ((orc_uint8) var39.x2[1]) | ((orc_uint8) var35.x2[1] << 8); + /* 5: storel */ + ptr0[i] = var36; + /* 6: loadw */ + var37 = ptr5[i]; + /* 7: mergebw */ + var38.x2[0] = ((orc_uint8) var39.x2[0]) | ((orc_uint8) var37.x2[0] << 8); + var38.x2[1] = ((orc_uint8) var39.x2[1]) | ((orc_uint8) var37.x2[1] << 8); + /* 8: storel */ + ptr1[i] = var38; + } + +} + +#else +static void +_backup_cogorc_convert_I420_UYVY (OrcExecutor * ex) +{ + int i; + int n = ex->n; + orc_union32 *ORC_RESTRICT ptr0; + orc_union32 *ORC_RESTRICT ptr1; + const orc_union16 *ORC_RESTRICT ptr4; + const orc_union16 *ORC_RESTRICT ptr5; + const orc_int8 *ORC_RESTRICT ptr6; + const orc_int8 *ORC_RESTRICT ptr7; + orc_int8 var33; + orc_int8 var34; + orc_union16 var35; + orc_union32 var36; + orc_union16 var37; + orc_union32 var38; + orc_union16 var39; + + ptr0 = (orc_union32 *) ex->arrays[0]; + ptr1 = (orc_union32 *) ex->arrays[1]; + ptr4 = (orc_union16 *) ex->arrays[4]; + ptr5 = (orc_union16 *) ex->arrays[5]; + ptr6 = (orc_int8 *) ex->arrays[6]; + ptr7 = (orc_int8 *) ex->arrays[7]; + + + for (i = 0; i < n; i++) { + /* 0: loadb */ + var33 = ptr6[i]; + /* 1: loadb */ + var34 = ptr7[i]; + /* 2: mergebw */ + var39.i = ((orc_uint8) var33) | ((orc_uint8) var34 << 8); + /* 3: loadw */ + var35 = ptr4[i]; + /* 4: mergebw */ + var36.x2[0] = ((orc_uint8) var39.x2[0]) | ((orc_uint8) var35.x2[0] << 8); + var36.x2[1] = ((orc_uint8) var39.x2[1]) | ((orc_uint8) var35.x2[1] << 8); + /* 5: storel */ + ptr0[i] = var36; + /* 6: loadw */ + var37 = ptr5[i]; + /* 7: mergebw */ + var38.x2[0] = ((orc_uint8) var39.x2[0]) | ((orc_uint8) var37.x2[0] << 8); + var38.x2[1] = ((orc_uint8) var39.x2[1]) | ((orc_uint8) var37.x2[1] << 8); + /* 8: storel */ + ptr1[i] = var38; + } + +} + +void +cogorc_convert_I420_UYVY (guint8 * d1, guint8 * d2, const guint8 * s1, + const guint8 * s2, const guint8 * s3, const guint8 * s4, int n) +{ + OrcExecutor _ex, *ex = &_ex; + static int p_inited = 0; + static OrcProgram *p = 0; + void (*func) (OrcExecutor *); + + if (!p_inited) { + orc_once_mutex_lock (); + if (!p_inited) { + OrcCompileResult result; + + p = orc_program_new (); + orc_program_set_name (p, "cogorc_convert_I420_UYVY"); + orc_program_set_backup_function (p, _backup_cogorc_convert_I420_UYVY); + orc_program_add_destination (p, 4, "d1"); + orc_program_add_destination (p, 4, "d2"); + orc_program_add_source (p, 2, "s1"); + orc_program_add_source (p, 2, "s2"); + orc_program_add_source (p, 1, "s3"); + orc_program_add_source (p, 1, "s4"); + orc_program_add_constant (p, 0, 0x00000000, "c1"); + orc_program_add_constant (p, 0, 0x00000000, "c2"); + orc_program_add_constant (p, 0, 0x00000000, "c3"); + orc_program_add_constant (p, 0, 0x00000000, "c4"); + orc_program_add_constant (p, 0, 0x00000000, "c5"); + orc_program_add_constant (p, 0, 0x00000000, "c6"); + orc_program_add_constant (p, 0, 0x00000000, "c7"); + orc_program_add_constant (p, 0, 0x00000000, "c8"); + orc_program_add_temporary (p, 2, "t1"); + + orc_program_append_2 (p, "mergebw", 0, ORC_VAR_T1, ORC_VAR_S3, ORC_VAR_S4, + ORC_VAR_D1); + orc_program_append_2 (p, "mergebw", 1, ORC_VAR_D1, ORC_VAR_T1, ORC_VAR_S1, + ORC_VAR_D1); + orc_program_append_2 (p, "mergebw", 1, ORC_VAR_D2, ORC_VAR_T1, ORC_VAR_S2, + ORC_VAR_D1); + + result = orc_program_compile (p); + } + p_inited = TRUE; + orc_once_mutex_unlock (); + } + ex->program = p; + + ex->n = n; + ex->arrays[ORC_VAR_D1] = d1; + ex->arrays[ORC_VAR_D2] = d2; + ex->arrays[ORC_VAR_S1] = (void *) s1; + ex->arrays[ORC_VAR_S2] = (void *) s2; + ex->arrays[ORC_VAR_S3] = (void *) s3; + ex->arrays[ORC_VAR_S4] = (void *) s4; + + func = p->code_exec; + func (ex); +} +#endif + + +/* cogorc_convert_I420_YUY2 */ +#ifdef DISABLE_ORC +void +cogorc_convert_I420_YUY2 (guint8 * d1, guint8 * d2, const guint8 * s1, + const guint8 * s2, const guint8 * s3, const guint8 * s4, int n) +{ + int i; + orc_union32 *ORC_RESTRICT ptr0; + orc_union32 *ORC_RESTRICT ptr1; + const orc_union16 *ORC_RESTRICT ptr4; + const orc_union16 *ORC_RESTRICT ptr5; + const orc_int8 *ORC_RESTRICT ptr6; + const orc_int8 *ORC_RESTRICT ptr7; + orc_int8 var33; + orc_int8 var34; + orc_union16 var35; + orc_union32 var36; + orc_union16 var37; + orc_union32 var38; + orc_union16 var39; + + ptr0 = (orc_union32 *) d1; + ptr1 = (orc_union32 *) d2; + ptr4 = (orc_union16 *) s1; + ptr5 = (orc_union16 *) s2; + ptr6 = (orc_int8 *) s3; + ptr7 = (orc_int8 *) s4; + + + for (i = 0; i < n; i++) { + /* 0: loadb */ + var33 = ptr6[i]; + /* 1: loadb */ + var34 = ptr7[i]; + /* 2: mergebw */ + var39.i = ((orc_uint8) var33) | ((orc_uint8) var34 << 8); + /* 3: loadw */ + var35 = ptr4[i]; + /* 4: mergebw */ + var36.x2[0] = ((orc_uint8) var35.x2[0]) | ((orc_uint8) var39.x2[0] << 8); + var36.x2[1] = ((orc_uint8) var35.x2[1]) | ((orc_uint8) var39.x2[1] << 8); + /* 5: storel */ + ptr0[i] = var36; + /* 6: loadw */ + var37 = ptr5[i]; + /* 7: mergebw */ + var38.x2[0] = ((orc_uint8) var37.x2[0]) | ((orc_uint8) var39.x2[0] << 8); + var38.x2[1] = ((orc_uint8) var37.x2[1]) | ((orc_uint8) var39.x2[1] << 8); + /* 8: storel */ + ptr1[i] = var38; + } + +} + +#else +static void +_backup_cogorc_convert_I420_YUY2 (OrcExecutor * ex) +{ + int i; + int n = ex->n; + orc_union32 *ORC_RESTRICT ptr0; + orc_union32 *ORC_RESTRICT ptr1; + const orc_union16 *ORC_RESTRICT ptr4; + const orc_union16 *ORC_RESTRICT ptr5; + const orc_int8 *ORC_RESTRICT ptr6; + const orc_int8 *ORC_RESTRICT ptr7; + orc_int8 var33; + orc_int8 var34; + orc_union16 var35; + orc_union32 var36; + orc_union16 var37; + orc_union32 var38; + orc_union16 var39; + + ptr0 = (orc_union32 *) ex->arrays[0]; + ptr1 = (orc_union32 *) ex->arrays[1]; + ptr4 = (orc_union16 *) ex->arrays[4]; + ptr5 = (orc_union16 *) ex->arrays[5]; + ptr6 = (orc_int8 *) ex->arrays[6]; + ptr7 = (orc_int8 *) ex->arrays[7]; + + + for (i = 0; i < n; i++) { + /* 0: loadb */ + var33 = ptr6[i]; + /* 1: loadb */ + var34 = ptr7[i]; + /* 2: mergebw */ + var39.i = ((orc_uint8) var33) | ((orc_uint8) var34 << 8); + /* 3: loadw */ + var35 = ptr4[i]; + /* 4: mergebw */ + var36.x2[0] = ((orc_uint8) var35.x2[0]) | ((orc_uint8) var39.x2[0] << 8); + var36.x2[1] = ((orc_uint8) var35.x2[1]) | ((orc_uint8) var39.x2[1] << 8); + /* 5: storel */ + ptr0[i] = var36; + /* 6: loadw */ + var37 = ptr5[i]; + /* 7: mergebw */ + var38.x2[0] = ((orc_uint8) var37.x2[0]) | ((orc_uint8) var39.x2[0] << 8); + var38.x2[1] = ((orc_uint8) var37.x2[1]) | ((orc_uint8) var39.x2[1] << 8); + /* 8: storel */ + ptr1[i] = var38; + } + +} + +void +cogorc_convert_I420_YUY2 (guint8 * d1, guint8 * d2, const guint8 * s1, + const guint8 * s2, const guint8 * s3, const guint8 * s4, int n) +{ + OrcExecutor _ex, *ex = &_ex; + static int p_inited = 0; + static OrcProgram *p = 0; + void (*func) (OrcExecutor *); + + if (!p_inited) { + orc_once_mutex_lock (); + if (!p_inited) { + OrcCompileResult result; + + p = orc_program_new (); + orc_program_set_name (p, "cogorc_convert_I420_YUY2"); + orc_program_set_backup_function (p, _backup_cogorc_convert_I420_YUY2); + orc_program_add_destination (p, 4, "d1"); + orc_program_add_destination (p, 4, "d2"); + orc_program_add_source (p, 2, "s1"); + orc_program_add_source (p, 2, "s2"); + orc_program_add_source (p, 1, "s3"); + orc_program_add_source (p, 1, "s4"); + orc_program_add_constant (p, 0, 0x00000000, "c1"); + orc_program_add_constant (p, 0, 0x00000000, "c2"); + orc_program_add_constant (p, 0, 0x00000000, "c3"); + orc_program_add_constant (p, 0, 0x00000000, "c4"); + orc_program_add_constant (p, 0, 0x00000000, "c5"); + orc_program_add_constant (p, 0, 0x00000000, "c6"); + orc_program_add_constant (p, 0, 0x00000000, "c7"); + orc_program_add_constant (p, 0, 0x00000000, "c8"); + orc_program_add_temporary (p, 2, "t1"); + + orc_program_append_2 (p, "mergebw", 0, ORC_VAR_T1, ORC_VAR_S3, ORC_VAR_S4, + ORC_VAR_D1); + orc_program_append_2 (p, "mergebw", 1, ORC_VAR_D1, ORC_VAR_S1, ORC_VAR_T1, + ORC_VAR_D1); + orc_program_append_2 (p, "mergebw", 1, ORC_VAR_D2, ORC_VAR_S2, ORC_VAR_T1, + ORC_VAR_D1); + + result = orc_program_compile (p); + } + p_inited = TRUE; + orc_once_mutex_unlock (); + } + ex->program = p; + + ex->n = n; + ex->arrays[ORC_VAR_D1] = d1; + ex->arrays[ORC_VAR_D2] = d2; + ex->arrays[ORC_VAR_S1] = (void *) s1; + ex->arrays[ORC_VAR_S2] = (void *) s2; + ex->arrays[ORC_VAR_S3] = (void *) s3; + ex->arrays[ORC_VAR_S4] = (void *) s4; + + func = p->code_exec; + func (ex); +} +#endif + + +/* cogorc_convert_I420_AYUV */ +#ifdef DISABLE_ORC +void +cogorc_convert_I420_AYUV (guint8 * d1, guint8 * d2, const guint8 * s1, + const guint8 * s2, const guint8 * s3, const guint8 * s4, int n) +{ + int i; + orc_union32 *ORC_RESTRICT ptr0; + orc_union32 *ORC_RESTRICT ptr1; + const orc_int8 *ORC_RESTRICT ptr4; + const orc_int8 *ORC_RESTRICT ptr5; + const orc_int8 *ORC_RESTRICT ptr6; + const orc_int8 *ORC_RESTRICT ptr7; + orc_int8 var36; + orc_int8 var37; + orc_union32 var38; + orc_int8 var39; + orc_int8 var40; + orc_union32 var41; + orc_int8 var42; + orc_int8 var43; + orc_union16 var44; + orc_union16 var45; + orc_union16 var46; + + ptr0 = (orc_union32 *) d1; + ptr1 = (orc_union32 *) d2; + ptr4 = (orc_int8 *) s1; + ptr5 = (orc_int8 *) s2; + ptr6 = (orc_int8 *) s3; + ptr7 = (orc_int8 *) s4; + + /* 3: loadpb */ + var36 = 0x000000ff; /* 255 or 1.25987e-321f */ + /* 8: loadpb */ + var39 = 0x000000ff; /* 255 or 1.25987e-321f */ + + for (i = 0; i < n; i++) { + /* 0: loadupdb */ + var42 = ptr6[i >> 1]; + /* 1: loadupdb */ + var43 = ptr7[i >> 1]; + /* 2: mergebw */ + var44.i = ((orc_uint8) var42) | ((orc_uint8) var43 << 8); + /* 4: loadb */ + var37 = ptr4[i]; + /* 5: mergebw */ + var45.i = ((orc_uint8) var36) | ((orc_uint8) var37 << 8); + /* 6: mergewl */ + var38.i = ((orc_uint16) var45.i) | ((orc_uint16) var44.i << 16); + /* 7: storel */ + ptr0[i] = var38; + /* 9: loadb */ + var40 = ptr5[i]; + /* 10: mergebw */ + var46.i = ((orc_uint8) var39) | ((orc_uint8) var40 << 8); + /* 11: mergewl */ + var41.i = ((orc_uint16) var46.i) | ((orc_uint16) var44.i << 16); + /* 12: storel */ + ptr1[i] = var41; + } + +} + +#else +static void +_backup_cogorc_convert_I420_AYUV (OrcExecutor * ex) +{ + int i; + int n = ex->n; + orc_union32 *ORC_RESTRICT ptr0; + orc_union32 *ORC_RESTRICT ptr1; + const orc_int8 *ORC_RESTRICT ptr4; + const orc_int8 *ORC_RESTRICT ptr5; + const orc_int8 *ORC_RESTRICT ptr6; + const orc_int8 *ORC_RESTRICT ptr7; + orc_int8 var36; + orc_int8 var37; + orc_union32 var38; + orc_int8 var39; + orc_int8 var40; + orc_union32 var41; + orc_int8 var42; + orc_int8 var43; + orc_union16 var44; + orc_union16 var45; + orc_union16 var46; + + ptr0 = (orc_union32 *) ex->arrays[0]; + ptr1 = (orc_union32 *) ex->arrays[1]; + ptr4 = (orc_int8 *) ex->arrays[4]; + ptr5 = (orc_int8 *) ex->arrays[5]; + ptr6 = (orc_int8 *) ex->arrays[6]; + ptr7 = (orc_int8 *) ex->arrays[7]; + + /* 3: loadpb */ + var36 = 0x000000ff; /* 255 or 1.25987e-321f */ + /* 8: loadpb */ + var39 = 0x000000ff; /* 255 or 1.25987e-321f */ + + for (i = 0; i < n; i++) { + /* 0: loadupdb */ + var42 = ptr6[i >> 1]; + /* 1: loadupdb */ + var43 = ptr7[i >> 1]; + /* 2: mergebw */ + var44.i = ((orc_uint8) var42) | ((orc_uint8) var43 << 8); + /* 4: loadb */ + var37 = ptr4[i]; + /* 5: mergebw */ + var45.i = ((orc_uint8) var36) | ((orc_uint8) var37 << 8); + /* 6: mergewl */ + var38.i = ((orc_uint16) var45.i) | ((orc_uint16) var44.i << 16); + /* 7: storel */ + ptr0[i] = var38; + /* 9: loadb */ + var40 = ptr5[i]; + /* 10: mergebw */ + var46.i = ((orc_uint8) var39) | ((orc_uint8) var40 << 8); + /* 11: mergewl */ + var41.i = ((orc_uint16) var46.i) | ((orc_uint16) var44.i << 16); + /* 12: storel */ + ptr1[i] = var41; + } + +} + +void +cogorc_convert_I420_AYUV (guint8 * d1, guint8 * d2, const guint8 * s1, + const guint8 * s2, const guint8 * s3, const guint8 * s4, int n) +{ + OrcExecutor _ex, *ex = &_ex; + static int p_inited = 0; + static OrcProgram *p = 0; + void (*func) (OrcExecutor *); + + if (!p_inited) { + orc_once_mutex_lock (); + if (!p_inited) { + OrcCompileResult result; + + p = orc_program_new (); + orc_program_set_name (p, "cogorc_convert_I420_AYUV"); + orc_program_set_backup_function (p, _backup_cogorc_convert_I420_AYUV); + orc_program_add_destination (p, 4, "d1"); + orc_program_add_destination (p, 4, "d2"); + orc_program_add_source (p, 1, "s1"); + orc_program_add_source (p, 1, "s2"); + orc_program_add_source (p, 1, "s3"); + orc_program_add_source (p, 1, "s4"); + orc_program_add_constant (p, 1, 0x000000ff, "c1"); + orc_program_add_constant (p, 0, 0x00000000, "c2"); + orc_program_add_constant (p, 0, 0x00000000, "c3"); + orc_program_add_constant (p, 0, 0x00000000, "c4"); + orc_program_add_constant (p, 0, 0x00000000, "c5"); + orc_program_add_constant (p, 0, 0x00000000, "c6"); + orc_program_add_constant (p, 0, 0x00000000, "c7"); + orc_program_add_constant (p, 0, 0x00000000, "c8"); + orc_program_add_temporary (p, 2, "t1"); + orc_program_add_temporary (p, 2, "t2"); + orc_program_add_temporary (p, 1, "t3"); + orc_program_add_temporary (p, 1, "t4"); + + orc_program_append_2 (p, "loadupdb", 0, ORC_VAR_T3, ORC_VAR_S3, + ORC_VAR_D1, ORC_VAR_D1); + orc_program_append_2 (p, "loadupdb", 0, ORC_VAR_T4, ORC_VAR_S4, + ORC_VAR_D1, ORC_VAR_D1); + orc_program_append_2 (p, "mergebw", 0, ORC_VAR_T1, ORC_VAR_T3, ORC_VAR_T4, + ORC_VAR_D1); + orc_program_append_2 (p, "mergebw", 0, ORC_VAR_T2, ORC_VAR_C1, ORC_VAR_S1, + ORC_VAR_D1); + orc_program_append_2 (p, "mergewl", 0, ORC_VAR_D1, ORC_VAR_T2, ORC_VAR_T1, + ORC_VAR_D1); + orc_program_append_2 (p, "mergebw", 0, ORC_VAR_T2, ORC_VAR_C1, ORC_VAR_S2, + ORC_VAR_D1); + orc_program_append_2 (p, "mergewl", 0, ORC_VAR_D2, ORC_VAR_T2, ORC_VAR_T1, + ORC_VAR_D1); + + result = orc_program_compile (p); + } + p_inited = TRUE; + orc_once_mutex_unlock (); + } + ex->program = p; + + ex->n = n; + ex->arrays[ORC_VAR_D1] = d1; + ex->arrays[ORC_VAR_D2] = d2; + ex->arrays[ORC_VAR_S1] = (void *) s1; + ex->arrays[ORC_VAR_S2] = (void *) s2; + ex->arrays[ORC_VAR_S3] = (void *) s3; + ex->arrays[ORC_VAR_S4] = (void *) s4; + + func = p->code_exec; + func (ex); +} +#endif + + +/* cogorc_convert_YUY2_I420 */ +#ifdef DISABLE_ORC +void +cogorc_convert_YUY2_I420 (guint8 * d1, guint8 * d2, guint8 * d3, guint8 * d4, + const guint8 * s1, const guint8 * s2, int n) +{ + int i; + orc_union16 *ORC_RESTRICT ptr0; + orc_union16 *ORC_RESTRICT ptr1; + orc_int8 *ORC_RESTRICT ptr2; + orc_int8 *ORC_RESTRICT ptr3; + const orc_union32 *ORC_RESTRICT ptr4; + const orc_union32 *ORC_RESTRICT ptr5; + orc_union32 var35; + orc_union32 var36; + orc_int8 var37; + orc_int8 var38; + orc_union16 var39; + orc_union16 var40; + orc_union16 var41; + orc_union16 var42; + orc_union16 var43; + + ptr0 = (orc_union16 *) d1; + ptr1 = (orc_union16 *) d2; + ptr2 = (orc_int8 *) d3; + ptr3 = (orc_int8 *) d4; + ptr4 = (orc_union32 *) s1; + ptr5 = (orc_union32 *) s2; + + + for (i = 0; i < n; i++) { + /* 0: loadl */ + var35 = ptr4[i]; + /* 1: splitwb */ + var39.x2[0] = (var35.x2[0] >> 8) & 0xff; + var40.x2[0] = var35.x2[0] & 0xff; + var39.x2[1] = (var35.x2[1] >> 8) & 0xff; + var40.x2[1] = var35.x2[1] & 0xff; + /* 2: storew */ + ptr0[i] = var40; + /* 3: loadl */ + var36 = ptr5[i]; + /* 4: splitwb */ + var41.x2[0] = (var36.x2[0] >> 8) & 0xff; + var42.x2[0] = var36.x2[0] & 0xff; + var41.x2[1] = (var36.x2[1] >> 8) & 0xff; + var42.x2[1] = var36.x2[1] & 0xff; + /* 5: storew */ + ptr1[i] = var42; + /* 6: avgub */ + var43.x2[0] = ((orc_uint8) var39.x2[0] + (orc_uint8) var41.x2[0] + 1) >> 1; + var43.x2[1] = ((orc_uint8) var39.x2[1] + (orc_uint8) var41.x2[1] + 1) >> 1; + /* 7: splitwb */ + var37 = (var43.i >> 8) & 0xff; + var38 = var43.i & 0xff; + /* 8: storeb */ + ptr3[i] = var37; + /* 9: storeb */ + ptr2[i] = var38; + } + +} + +#else +static void +_backup_cogorc_convert_YUY2_I420 (OrcExecutor * ex) +{ + int i; + int n = ex->n; + orc_union16 *ORC_RESTRICT ptr0; + orc_union16 *ORC_RESTRICT ptr1; + orc_int8 *ORC_RESTRICT ptr2; + orc_int8 *ORC_RESTRICT ptr3; + const orc_union32 *ORC_RESTRICT ptr4; + const orc_union32 *ORC_RESTRICT ptr5; + orc_union32 var35; + orc_union32 var36; + orc_int8 var37; + orc_int8 var38; + orc_union16 var39; + orc_union16 var40; + orc_union16 var41; + orc_union16 var42; + orc_union16 var43; + + ptr0 = (orc_union16 *) ex->arrays[0]; + ptr1 = (orc_union16 *) ex->arrays[1]; + ptr2 = (orc_int8 *) ex->arrays[2]; + ptr3 = (orc_int8 *) ex->arrays[3]; + ptr4 = (orc_union32 *) ex->arrays[4]; + ptr5 = (orc_union32 *) ex->arrays[5]; + + + for (i = 0; i < n; i++) { + /* 0: loadl */ + var35 = ptr4[i]; + /* 1: splitwb */ + var39.x2[0] = (var35.x2[0] >> 8) & 0xff; + var40.x2[0] = var35.x2[0] & 0xff; + var39.x2[1] = (var35.x2[1] >> 8) & 0xff; + var40.x2[1] = var35.x2[1] & 0xff; + /* 2: storew */ + ptr0[i] = var40; + /* 3: loadl */ + var36 = ptr5[i]; + /* 4: splitwb */ + var41.x2[0] = (var36.x2[0] >> 8) & 0xff; + var42.x2[0] = var36.x2[0] & 0xff; + var41.x2[1] = (var36.x2[1] >> 8) & 0xff; + var42.x2[1] = var36.x2[1] & 0xff; + /* 5: storew */ + ptr1[i] = var42; + /* 6: avgub */ + var43.x2[0] = ((orc_uint8) var39.x2[0] + (orc_uint8) var41.x2[0] + 1) >> 1; + var43.x2[1] = ((orc_uint8) var39.x2[1] + (orc_uint8) var41.x2[1] + 1) >> 1; + /* 7: splitwb */ + var37 = (var43.i >> 8) & 0xff; + var38 = var43.i & 0xff; + /* 8: storeb */ + ptr3[i] = var37; + /* 9: storeb */ + ptr2[i] = var38; + } + +} + +void +cogorc_convert_YUY2_I420 (guint8 * d1, guint8 * d2, guint8 * d3, guint8 * d4, + const guint8 * s1, const guint8 * s2, int n) +{ + OrcExecutor _ex, *ex = &_ex; + static int p_inited = 0; + static OrcProgram *p = 0; + void (*func) (OrcExecutor *); + + if (!p_inited) { + orc_once_mutex_lock (); + if (!p_inited) { + OrcCompileResult result; + + p = orc_program_new (); + orc_program_set_name (p, "cogorc_convert_YUY2_I420"); + orc_program_set_backup_function (p, _backup_cogorc_convert_YUY2_I420); + orc_program_add_destination (p, 2, "d1"); + orc_program_add_destination (p, 2, "d2"); + orc_program_add_destination (p, 1, "d3"); + orc_program_add_destination (p, 1, "d4"); + orc_program_add_source (p, 4, "s1"); + orc_program_add_source (p, 4, "s2"); + orc_program_add_constant (p, 0, 0x00000000, "c1"); + orc_program_add_constant (p, 0, 0x00000000, "c2"); + orc_program_add_constant (p, 0, 0x00000000, "c3"); + orc_program_add_constant (p, 0, 0x00000000, "c4"); + orc_program_add_constant (p, 0, 0x00000000, "c5"); + orc_program_add_constant (p, 0, 0x00000000, "c6"); + orc_program_add_constant (p, 0, 0x00000000, "c7"); + orc_program_add_constant (p, 0, 0x00000000, "c8"); + orc_program_add_temporary (p, 2, "t1"); + orc_program_add_temporary (p, 2, "t2"); + orc_program_add_temporary (p, 2, "t3"); + + orc_program_append_2 (p, "splitwb", 1, ORC_VAR_T1, ORC_VAR_T3, ORC_VAR_S1, + ORC_VAR_D1); + orc_program_append_2 (p, "storew", 0, ORC_VAR_D1, ORC_VAR_T3, ORC_VAR_D1, + ORC_VAR_D1); + orc_program_append_2 (p, "splitwb", 1, ORC_VAR_T2, ORC_VAR_T3, ORC_VAR_S2, + ORC_VAR_D1); + orc_program_append_2 (p, "storew", 0, ORC_VAR_D2, ORC_VAR_T3, ORC_VAR_D1, + ORC_VAR_D1); + orc_program_append_2 (p, "avgub", 1, ORC_VAR_T1, ORC_VAR_T1, ORC_VAR_T2, + ORC_VAR_D1); + orc_program_append_2 (p, "splitwb", 0, ORC_VAR_D4, ORC_VAR_D3, ORC_VAR_T1, + ORC_VAR_D1); + + result = orc_program_compile (p); + } + p_inited = TRUE; + orc_once_mutex_unlock (); + } + ex->program = p; + + ex->n = n; + ex->arrays[ORC_VAR_D1] = d1; + ex->arrays[ORC_VAR_D2] = d2; + ex->arrays[ORC_VAR_D3] = d3; + ex->arrays[ORC_VAR_D4] = d4; + ex->arrays[ORC_VAR_S1] = (void *) s1; + ex->arrays[ORC_VAR_S2] = (void *) s2; + + func = p->code_exec; + func (ex); +} +#endif + + +/* cogorc_convert_UYVY_YUY2 */ +#ifdef DISABLE_ORC +void +cogorc_convert_UYVY_YUY2 (guint8 * d1, int d1_stride, const guint8 * s1, + int s1_stride, int n, int m) +{ + int i; + int j; + orc_union32 *ORC_RESTRICT ptr0; + const orc_union32 *ORC_RESTRICT ptr4; + orc_union32 var32; + orc_union32 var33; + + for (j = 0; j < m; j++) { + ptr0 = ORC_PTR_OFFSET (d1, d1_stride * j); + ptr4 = ORC_PTR_OFFSET (s1, s1_stride * j); + + + for (i = 0; i < n; i++) { + /* 0: loadl */ + var32 = ptr4[i]; + /* 1: swapw */ + var33.x2[0] = ORC_SWAP_W (var32.x2[0]); + var33.x2[1] = ORC_SWAP_W (var32.x2[1]); + /* 2: storel */ + ptr0[i] = var33; + } + } + +} + +#else +static void +_backup_cogorc_convert_UYVY_YUY2 (OrcExecutor * ex) +{ + int i; + int j; + int n = ex->n; + int m = ex->params[ORC_VAR_A1]; + orc_union32 *ORC_RESTRICT ptr0; + const orc_union32 *ORC_RESTRICT ptr4; + orc_union32 var32; + orc_union32 var33; + + for (j = 0; j < m; j++) { + ptr0 = ORC_PTR_OFFSET (ex->arrays[0], ex->params[0] * j); + ptr4 = ORC_PTR_OFFSET (ex->arrays[4], ex->params[4] * j); + + + for (i = 0; i < n; i++) { + /* 0: loadl */ + var32 = ptr4[i]; + /* 1: swapw */ + var33.x2[0] = ORC_SWAP_W (var32.x2[0]); + var33.x2[1] = ORC_SWAP_W (var32.x2[1]); + /* 2: storel */ + ptr0[i] = var33; + } + } + +} + +void +cogorc_convert_UYVY_YUY2 (guint8 * d1, int d1_stride, const guint8 * s1, + int s1_stride, int n, int m) +{ + OrcExecutor _ex, *ex = &_ex; + static int p_inited = 0; + static OrcProgram *p = 0; + void (*func) (OrcExecutor *); + + if (!p_inited) { + orc_once_mutex_lock (); + if (!p_inited) { + OrcCompileResult result; + + p = orc_program_new (); + orc_program_set_2d (p); + orc_program_set_name (p, "cogorc_convert_UYVY_YUY2"); + orc_program_set_backup_function (p, _backup_cogorc_convert_UYVY_YUY2); + orc_program_add_destination (p, 4, "d1"); + orc_program_add_source (p, 4, "s1"); + orc_program_add_constant (p, 0, 0x00000000, "c1"); + orc_program_add_constant (p, 0, 0x00000000, "c2"); + orc_program_add_constant (p, 0, 0x00000000, "c3"); + orc_program_add_constant (p, 0, 0x00000000, "c4"); + orc_program_add_constant (p, 0, 0x00000000, "c5"); + orc_program_add_constant (p, 0, 0x00000000, "c6"); + orc_program_add_constant (p, 0, 0x00000000, "c7"); + orc_program_add_constant (p, 0, 0x00000000, "c8"); + + orc_program_append_2 (p, "swapw", 1, ORC_VAR_D1, ORC_VAR_S1, ORC_VAR_D1, + ORC_VAR_D1); + + result = orc_program_compile (p); + } + p_inited = TRUE; + orc_once_mutex_unlock (); + } + ex->program = p; + + ex->n = n; + ORC_EXECUTOR_M (ex) = m; + ex->arrays[ORC_VAR_D1] = d1; + ex->params[ORC_VAR_D1] = d1_stride; + ex->arrays[ORC_VAR_S1] = (void *) s1; + ex->params[ORC_VAR_S1] = s1_stride; + + func = p->code_exec; + func (ex); +} +#endif + + +/* cogorc_planar_chroma_420_422 */ +#ifdef DISABLE_ORC +void +cogorc_planar_chroma_420_422 (guint8 * d1, int d1_stride, guint8 * d2, + int d2_stride, const guint8 * s1, int s1_stride, int n, int m) +{ + int i; + int j; + orc_int8 *ORC_RESTRICT ptr0; + orc_int8 *ORC_RESTRICT ptr1; + const orc_int8 *ORC_RESTRICT ptr4; + orc_int8 var32; + orc_int8 var33; + orc_int8 var34; + orc_int8 var35; + + for (j = 0; j < m; j++) { + ptr0 = ORC_PTR_OFFSET (d1, d1_stride * j); + ptr1 = ORC_PTR_OFFSET (d2, d2_stride * j); + ptr4 = ORC_PTR_OFFSET (s1, s1_stride * j); + + + for (i = 0; i < n; i++) { + /* 0: loadb */ + var32 = ptr4[i]; + /* 1: copyb */ + var33 = var32; + /* 2: storeb */ + ptr0[i] = var33; + /* 3: loadb */ + var34 = ptr4[i]; + /* 4: copyb */ + var35 = var34; + /* 5: storeb */ + ptr1[i] = var35; + } + } + +} + +#else +static void +_backup_cogorc_planar_chroma_420_422 (OrcExecutor * ex) +{ + int i; + int j; + int n = ex->n; + int m = ex->params[ORC_VAR_A1]; + orc_int8 *ORC_RESTRICT ptr0; + orc_int8 *ORC_RESTRICT ptr1; + const orc_int8 *ORC_RESTRICT ptr4; + orc_int8 var32; + orc_int8 var33; + orc_int8 var34; + orc_int8 var35; + + for (j = 0; j < m; j++) { + ptr0 = ORC_PTR_OFFSET (ex->arrays[0], ex->params[0] * j); + ptr1 = ORC_PTR_OFFSET (ex->arrays[1], ex->params[1] * j); + ptr4 = ORC_PTR_OFFSET (ex->arrays[4], ex->params[4] * j); + + + for (i = 0; i < n; i++) { + /* 0: loadb */ + var32 = ptr4[i]; + /* 1: copyb */ + var33 = var32; + /* 2: storeb */ + ptr0[i] = var33; + /* 3: loadb */ + var34 = ptr4[i]; + /* 4: copyb */ + var35 = var34; + /* 5: storeb */ + ptr1[i] = var35; + } + } + +} + +void +cogorc_planar_chroma_420_422 (guint8 * d1, int d1_stride, guint8 * d2, + int d2_stride, const guint8 * s1, int s1_stride, int n, int m) +{ + OrcExecutor _ex, *ex = &_ex; + static int p_inited = 0; + static OrcProgram *p = 0; + void (*func) (OrcExecutor *); + + if (!p_inited) { + orc_once_mutex_lock (); + if (!p_inited) { + OrcCompileResult result; + + p = orc_program_new (); + orc_program_set_2d (p); + orc_program_set_name (p, "cogorc_planar_chroma_420_422"); + orc_program_set_backup_function (p, _backup_cogorc_planar_chroma_420_422); + orc_program_add_destination (p, 1, "d1"); + orc_program_add_destination (p, 1, "d2"); + orc_program_add_source (p, 1, "s1"); + orc_program_add_constant (p, 0, 0x00000000, "c1"); + orc_program_add_constant (p, 0, 0x00000000, "c2"); + orc_program_add_constant (p, 0, 0x00000000, "c3"); + orc_program_add_constant (p, 0, 0x00000000, "c4"); + orc_program_add_constant (p, 0, 0x00000000, "c5"); + orc_program_add_constant (p, 0, 0x00000000, "c6"); + orc_program_add_constant (p, 0, 0x00000000, "c7"); + orc_program_add_constant (p, 0, 0x00000000, "c8"); + + orc_program_append_2 (p, "copyb", 0, ORC_VAR_D1, ORC_VAR_S1, ORC_VAR_D1, + ORC_VAR_D1); + orc_program_append_2 (p, "copyb", 0, ORC_VAR_D2, ORC_VAR_S1, ORC_VAR_D1, + ORC_VAR_D1); + + result = orc_program_compile (p); + } + p_inited = TRUE; + orc_once_mutex_unlock (); + } + ex->program = p; + + ex->n = n; + ORC_EXECUTOR_M (ex) = m; + ex->arrays[ORC_VAR_D1] = d1; + ex->params[ORC_VAR_D1] = d1_stride; + ex->arrays[ORC_VAR_D2] = d2; + ex->params[ORC_VAR_D2] = d2_stride; + ex->arrays[ORC_VAR_S1] = (void *) s1; + ex->params[ORC_VAR_S1] = s1_stride; + + func = p->code_exec; + func (ex); +} +#endif + + +/* cogorc_planar_chroma_420_444 */ +#ifdef DISABLE_ORC +void +cogorc_planar_chroma_420_444 (guint8 * d1, int d1_stride, guint8 * d2, + int d2_stride, const guint8 * s1, int s1_stride, int n, int m) +{ + int i; + int j; + orc_union16 *ORC_RESTRICT ptr0; + orc_union16 *ORC_RESTRICT ptr1; + const orc_int8 *ORC_RESTRICT ptr4; + orc_int8 var33; + orc_union16 var34; + + for (j = 0; j < m; j++) { + ptr0 = ORC_PTR_OFFSET (d1, d1_stride * j); + ptr1 = ORC_PTR_OFFSET (d2, d2_stride * j); + ptr4 = ORC_PTR_OFFSET (s1, s1_stride * j); + + + for (i = 0; i < n; i++) { + /* 0: loadb */ + var33 = ptr4[i]; + /* 1: splatbw */ + var34.i = ((var33 & 0xff) << 8) | (var33 & 0xff); + /* 2: storew */ + ptr0[i] = var34; + /* 3: storew */ + ptr1[i] = var34; + } + } + +} + +#else +static void +_backup_cogorc_planar_chroma_420_444 (OrcExecutor * ex) +{ + int i; + int j; + int n = ex->n; + int m = ex->params[ORC_VAR_A1]; + orc_union16 *ORC_RESTRICT ptr0; + orc_union16 *ORC_RESTRICT ptr1; + const orc_int8 *ORC_RESTRICT ptr4; + orc_int8 var33; + orc_union16 var34; + + for (j = 0; j < m; j++) { + ptr0 = ORC_PTR_OFFSET (ex->arrays[0], ex->params[0] * j); + ptr1 = ORC_PTR_OFFSET (ex->arrays[1], ex->params[1] * j); + ptr4 = ORC_PTR_OFFSET (ex->arrays[4], ex->params[4] * j); + + + for (i = 0; i < n; i++) { + /* 0: loadb */ + var33 = ptr4[i]; + /* 1: splatbw */ + var34.i = ((var33 & 0xff) << 8) | (var33 & 0xff); + /* 2: storew */ + ptr0[i] = var34; + /* 3: storew */ + ptr1[i] = var34; + } + } + +} + +void +cogorc_planar_chroma_420_444 (guint8 * d1, int d1_stride, guint8 * d2, + int d2_stride, const guint8 * s1, int s1_stride, int n, int m) +{ + OrcExecutor _ex, *ex = &_ex; + static int p_inited = 0; + static OrcProgram *p = 0; + void (*func) (OrcExecutor *); + + if (!p_inited) { + orc_once_mutex_lock (); + if (!p_inited) { + OrcCompileResult result; + + p = orc_program_new (); + orc_program_set_2d (p); + orc_program_set_name (p, "cogorc_planar_chroma_420_444"); + orc_program_set_backup_function (p, _backup_cogorc_planar_chroma_420_444); + orc_program_add_destination (p, 2, "d1"); + orc_program_add_destination (p, 2, "d2"); + orc_program_add_source (p, 1, "s1"); + orc_program_add_constant (p, 0, 0x00000000, "c1"); + orc_program_add_constant (p, 0, 0x00000000, "c2"); + orc_program_add_constant (p, 0, 0x00000000, "c3"); + orc_program_add_constant (p, 0, 0x00000000, "c4"); + orc_program_add_constant (p, 0, 0x00000000, "c5"); + orc_program_add_constant (p, 0, 0x00000000, "c6"); + orc_program_add_constant (p, 0, 0x00000000, "c7"); + orc_program_add_constant (p, 0, 0x00000000, "c8"); + orc_program_add_temporary (p, 2, "t1"); + + orc_program_append_2 (p, "splatbw", 0, ORC_VAR_T1, ORC_VAR_S1, ORC_VAR_D1, + ORC_VAR_D1); + orc_program_append_2 (p, "storew", 0, ORC_VAR_D1, ORC_VAR_T1, ORC_VAR_D1, + ORC_VAR_D1); + orc_program_append_2 (p, "storew", 0, ORC_VAR_D2, ORC_VAR_T1, ORC_VAR_D1, + ORC_VAR_D1); + + result = orc_program_compile (p); + } + p_inited = TRUE; + orc_once_mutex_unlock (); + } + ex->program = p; + + ex->n = n; + ORC_EXECUTOR_M (ex) = m; + ex->arrays[ORC_VAR_D1] = d1; + ex->params[ORC_VAR_D1] = d1_stride; + ex->arrays[ORC_VAR_D2] = d2; + ex->params[ORC_VAR_D2] = d2_stride; + ex->arrays[ORC_VAR_S1] = (void *) s1; + ex->params[ORC_VAR_S1] = s1_stride; + + func = p->code_exec; + func (ex); +} +#endif + + +/* cogorc_planar_chroma_422_444 */ +#ifdef DISABLE_ORC +void +cogorc_planar_chroma_422_444 (guint8 * d1, int d1_stride, const guint8 * s1, + int s1_stride, int n, int m) +{ + int i; + int j; + orc_union16 *ORC_RESTRICT ptr0; + const orc_int8 *ORC_RESTRICT ptr4; + orc_int8 var33; + orc_union16 var34; + + for (j = 0; j < m; j++) { + ptr0 = ORC_PTR_OFFSET (d1, d1_stride * j); + ptr4 = ORC_PTR_OFFSET (s1, s1_stride * j); + + + for (i = 0; i < n; i++) { + /* 0: loadb */ + var33 = ptr4[i]; + /* 1: splatbw */ + var34.i = ((var33 & 0xff) << 8) | (var33 & 0xff); + /* 2: storew */ + ptr0[i] = var34; + } + } + +} + +#else +static void +_backup_cogorc_planar_chroma_422_444 (OrcExecutor * ex) +{ + int i; + int j; + int n = ex->n; + int m = ex->params[ORC_VAR_A1]; + orc_union16 *ORC_RESTRICT ptr0; + const orc_int8 *ORC_RESTRICT ptr4; + orc_int8 var33; + orc_union16 var34; + + for (j = 0; j < m; j++) { + ptr0 = ORC_PTR_OFFSET (ex->arrays[0], ex->params[0] * j); + ptr4 = ORC_PTR_OFFSET (ex->arrays[4], ex->params[4] * j); + + + for (i = 0; i < n; i++) { + /* 0: loadb */ + var33 = ptr4[i]; + /* 1: splatbw */ + var34.i = ((var33 & 0xff) << 8) | (var33 & 0xff); + /* 2: storew */ + ptr0[i] = var34; + } + } + +} + +void +cogorc_planar_chroma_422_444 (guint8 * d1, int d1_stride, const guint8 * s1, + int s1_stride, int n, int m) +{ + OrcExecutor _ex, *ex = &_ex; + static int p_inited = 0; + static OrcProgram *p = 0; + void (*func) (OrcExecutor *); + + if (!p_inited) { + orc_once_mutex_lock (); + if (!p_inited) { + OrcCompileResult result; + + p = orc_program_new (); + orc_program_set_2d (p); + orc_program_set_name (p, "cogorc_planar_chroma_422_444"); + orc_program_set_backup_function (p, _backup_cogorc_planar_chroma_422_444); + orc_program_add_destination (p, 2, "d1"); + orc_program_add_source (p, 1, "s1"); + orc_program_add_constant (p, 0, 0x00000000, "c1"); + orc_program_add_constant (p, 0, 0x00000000, "c2"); + orc_program_add_constant (p, 0, 0x00000000, "c3"); + orc_program_add_constant (p, 0, 0x00000000, "c4"); + orc_program_add_constant (p, 0, 0x00000000, "c5"); + orc_program_add_constant (p, 0, 0x00000000, "c6"); + orc_program_add_constant (p, 0, 0x00000000, "c7"); + orc_program_add_constant (p, 0, 0x00000000, "c8"); + orc_program_add_temporary (p, 2, "t1"); + + orc_program_append_2 (p, "splatbw", 0, ORC_VAR_T1, ORC_VAR_S1, ORC_VAR_D1, + ORC_VAR_D1); + orc_program_append_2 (p, "storew", 0, ORC_VAR_D1, ORC_VAR_T1, ORC_VAR_D1, + ORC_VAR_D1); + + result = orc_program_compile (p); + } + p_inited = TRUE; + orc_once_mutex_unlock (); + } + ex->program = p; + + ex->n = n; + ORC_EXECUTOR_M (ex) = m; + ex->arrays[ORC_VAR_D1] = d1; + ex->params[ORC_VAR_D1] = d1_stride; + ex->arrays[ORC_VAR_S1] = (void *) s1; + ex->params[ORC_VAR_S1] = s1_stride; + + func = p->code_exec; + func (ex); +} +#endif + + +/* cogorc_planar_chroma_444_422 */ +#ifdef DISABLE_ORC +void +cogorc_planar_chroma_444_422 (guint8 * d1, int d1_stride, const guint8 * s1, + int s1_stride, int n, int m) +{ + int i; + int j; + orc_int8 *ORC_RESTRICT ptr0; + const orc_union16 *ORC_RESTRICT ptr4; + orc_union16 var34; + orc_int8 var35; + orc_int8 var36; + orc_int8 var37; + + for (j = 0; j < m; j++) { + ptr0 = ORC_PTR_OFFSET (d1, d1_stride * j); + ptr4 = ORC_PTR_OFFSET (s1, s1_stride * j); + + + for (i = 0; i < n; i++) { + /* 0: loadw */ + var34 = ptr4[i]; + /* 1: splitwb */ + var36 = (var34.i >> 8) & 0xff; + var37 = var34.i & 0xff; + /* 2: avgub */ + var35 = ((orc_uint8) var36 + (orc_uint8) var37 + 1) >> 1; + /* 3: storeb */ + ptr0[i] = var35; + } + } + +} + +#else +static void +_backup_cogorc_planar_chroma_444_422 (OrcExecutor * ex) +{ + int i; + int j; + int n = ex->n; + int m = ex->params[ORC_VAR_A1]; + orc_int8 *ORC_RESTRICT ptr0; + const orc_union16 *ORC_RESTRICT ptr4; + orc_union16 var34; + orc_int8 var35; + orc_int8 var36; + orc_int8 var37; + + for (j = 0; j < m; j++) { + ptr0 = ORC_PTR_OFFSET (ex->arrays[0], ex->params[0] * j); + ptr4 = ORC_PTR_OFFSET (ex->arrays[4], ex->params[4] * j); + + + for (i = 0; i < n; i++) { + /* 0: loadw */ + var34 = ptr4[i]; + /* 1: splitwb */ + var36 = (var34.i >> 8) & 0xff; + var37 = var34.i & 0xff; + /* 2: avgub */ + var35 = ((orc_uint8) var36 + (orc_uint8) var37 + 1) >> 1; + /* 3: storeb */ + ptr0[i] = var35; + } + } + +} + +void +cogorc_planar_chroma_444_422 (guint8 * d1, int d1_stride, const guint8 * s1, + int s1_stride, int n, int m) +{ + OrcExecutor _ex, *ex = &_ex; + static int p_inited = 0; + static OrcProgram *p = 0; + void (*func) (OrcExecutor *); + + if (!p_inited) { + orc_once_mutex_lock (); + if (!p_inited) { + OrcCompileResult result; + + p = orc_program_new (); + orc_program_set_2d (p); + orc_program_set_name (p, "cogorc_planar_chroma_444_422"); + orc_program_set_backup_function (p, _backup_cogorc_planar_chroma_444_422); + orc_program_add_destination (p, 1, "d1"); + orc_program_add_source (p, 2, "s1"); + orc_program_add_constant (p, 0, 0x00000000, "c1"); + orc_program_add_constant (p, 0, 0x00000000, "c2"); + orc_program_add_constant (p, 0, 0x00000000, "c3"); + orc_program_add_constant (p, 0, 0x00000000, "c4"); + orc_program_add_constant (p, 0, 0x00000000, "c5"); + orc_program_add_constant (p, 0, 0x00000000, "c6"); + orc_program_add_constant (p, 0, 0x00000000, "c7"); + orc_program_add_constant (p, 0, 0x00000000, "c8"); + orc_program_add_temporary (p, 1, "t1"); + orc_program_add_temporary (p, 1, "t2"); + + orc_program_append_2 (p, "splitwb", 0, ORC_VAR_T1, ORC_VAR_T2, ORC_VAR_S1, + ORC_VAR_D1); + orc_program_append_2 (p, "avgub", 0, ORC_VAR_D1, ORC_VAR_T1, ORC_VAR_T2, + ORC_VAR_D1); + + result = orc_program_compile (p); + } + p_inited = TRUE; + orc_once_mutex_unlock (); + } + ex->program = p; + + ex->n = n; + ORC_EXECUTOR_M (ex) = m; + ex->arrays[ORC_VAR_D1] = d1; + ex->params[ORC_VAR_D1] = d1_stride; + ex->arrays[ORC_VAR_S1] = (void *) s1; + ex->params[ORC_VAR_S1] = s1_stride; + + func = p->code_exec; + func (ex); +} +#endif + + +/* cogorc_planar_chroma_444_420 */ +#ifdef DISABLE_ORC +void +cogorc_planar_chroma_444_420 (guint8 * d1, int d1_stride, const guint8 * s1, + int s1_stride, const guint8 * s2, int s2_stride, int n, int m) +{ + int i; + int j; + orc_int8 *ORC_RESTRICT ptr0; + const orc_union16 *ORC_RESTRICT ptr4; + const orc_union16 *ORC_RESTRICT ptr5; + orc_union16 var35; + orc_union16 var36; + orc_int8 var37; + orc_union16 var38; + orc_int8 var39; + orc_int8 var40; + + for (j = 0; j < m; j++) { + ptr0 = ORC_PTR_OFFSET (d1, d1_stride * j); + ptr4 = ORC_PTR_OFFSET (s1, s1_stride * j); + ptr5 = ORC_PTR_OFFSET (s2, s2_stride * j); + + + for (i = 0; i < n; i++) { + /* 0: loadw */ + var35 = ptr4[i]; + /* 1: loadw */ + var36 = ptr5[i]; + /* 2: avgub */ + var38.x2[0] = + ((orc_uint8) var35.x2[0] + (orc_uint8) var36.x2[0] + 1) >> 1; + var38.x2[1] = + ((orc_uint8) var35.x2[1] + (orc_uint8) var36.x2[1] + 1) >> 1; + /* 3: splitwb */ + var39 = (var38.i >> 8) & 0xff; + var40 = var38.i & 0xff; + /* 4: avgub */ + var37 = ((orc_uint8) var39 + (orc_uint8) var40 + 1) >> 1; + /* 5: storeb */ + ptr0[i] = var37; + } + } + +} + +#else +static void +_backup_cogorc_planar_chroma_444_420 (OrcExecutor * ex) +{ + int i; + int j; + int n = ex->n; + int m = ex->params[ORC_VAR_A1]; + orc_int8 *ORC_RESTRICT ptr0; + const orc_union16 *ORC_RESTRICT ptr4; + const orc_union16 *ORC_RESTRICT ptr5; + orc_union16 var35; + orc_union16 var36; + orc_int8 var37; + orc_union16 var38; + orc_int8 var39; + orc_int8 var40; + + for (j = 0; j < m; j++) { + ptr0 = ORC_PTR_OFFSET (ex->arrays[0], ex->params[0] * j); + ptr4 = ORC_PTR_OFFSET (ex->arrays[4], ex->params[4] * j); + ptr5 = ORC_PTR_OFFSET (ex->arrays[5], ex->params[5] * j); + + + for (i = 0; i < n; i++) { + /* 0: loadw */ + var35 = ptr4[i]; + /* 1: loadw */ + var36 = ptr5[i]; + /* 2: avgub */ + var38.x2[0] = + ((orc_uint8) var35.x2[0] + (orc_uint8) var36.x2[0] + 1) >> 1; + var38.x2[1] = + ((orc_uint8) var35.x2[1] + (orc_uint8) var36.x2[1] + 1) >> 1; + /* 3: splitwb */ + var39 = (var38.i >> 8) & 0xff; + var40 = var38.i & 0xff; + /* 4: avgub */ + var37 = ((orc_uint8) var39 + (orc_uint8) var40 + 1) >> 1; + /* 5: storeb */ + ptr0[i] = var37; + } + } + +} + +void +cogorc_planar_chroma_444_420 (guint8 * d1, int d1_stride, const guint8 * s1, + int s1_stride, const guint8 * s2, int s2_stride, int n, int m) +{ + OrcExecutor _ex, *ex = &_ex; + static int p_inited = 0; + static OrcProgram *p = 0; + void (*func) (OrcExecutor *); + + if (!p_inited) { + orc_once_mutex_lock (); + if (!p_inited) { + OrcCompileResult result; + + p = orc_program_new (); + orc_program_set_2d (p); + orc_program_set_name (p, "cogorc_planar_chroma_444_420"); + orc_program_set_backup_function (p, _backup_cogorc_planar_chroma_444_420); + orc_program_add_destination (p, 1, "d1"); + orc_program_add_source (p, 2, "s1"); + orc_program_add_source (p, 2, "s2"); + orc_program_add_constant (p, 0, 0x00000000, "c1"); + orc_program_add_constant (p, 0, 0x00000000, "c2"); + orc_program_add_constant (p, 0, 0x00000000, "c3"); + orc_program_add_constant (p, 0, 0x00000000, "c4"); + orc_program_add_constant (p, 0, 0x00000000, "c5"); + orc_program_add_constant (p, 0, 0x00000000, "c6"); + orc_program_add_constant (p, 0, 0x00000000, "c7"); + orc_program_add_constant (p, 0, 0x00000000, "c8"); + orc_program_add_temporary (p, 2, "t1"); + orc_program_add_temporary (p, 1, "t2"); + orc_program_add_temporary (p, 1, "t3"); + + orc_program_append_2 (p, "avgub", 1, ORC_VAR_T1, ORC_VAR_S1, ORC_VAR_S2, + ORC_VAR_D1); + orc_program_append_2 (p, "splitwb", 0, ORC_VAR_T2, ORC_VAR_T3, ORC_VAR_T1, + ORC_VAR_D1); + orc_program_append_2 (p, "avgub", 0, ORC_VAR_D1, ORC_VAR_T2, ORC_VAR_T3, + ORC_VAR_D1); + + result = orc_program_compile (p); + } + p_inited = TRUE; + orc_once_mutex_unlock (); + } + ex->program = p; + + ex->n = n; + ORC_EXECUTOR_M (ex) = m; + ex->arrays[ORC_VAR_D1] = d1; + ex->params[ORC_VAR_D1] = d1_stride; + ex->arrays[ORC_VAR_S1] = (void *) s1; + ex->params[ORC_VAR_S1] = s1_stride; + ex->arrays[ORC_VAR_S2] = (void *) s2; + ex->params[ORC_VAR_S2] = s2_stride; + + func = p->code_exec; + func (ex); +} +#endif + + +/* cogorc_planar_chroma_422_420 */ +#ifdef DISABLE_ORC +void +cogorc_planar_chroma_422_420 (guint8 * d1, int d1_stride, const guint8 * s1, + int s1_stride, const guint8 * s2, int s2_stride, int n, int m) +{ + int i; + int j; + orc_int8 *ORC_RESTRICT ptr0; + const orc_int8 *ORC_RESTRICT ptr4; + const orc_int8 *ORC_RESTRICT ptr5; + orc_int8 var32; + orc_int8 var33; + orc_int8 var34; + + for (j = 0; j < m; j++) { + ptr0 = ORC_PTR_OFFSET (d1, d1_stride * j); + ptr4 = ORC_PTR_OFFSET (s1, s1_stride * j); + ptr5 = ORC_PTR_OFFSET (s2, s2_stride * j); + + + for (i = 0; i < n; i++) { + /* 0: loadb */ + var32 = ptr4[i]; + /* 1: loadb */ + var33 = ptr5[i]; + /* 2: avgub */ + var34 = ((orc_uint8) var32 + (orc_uint8) var33 + 1) >> 1; + /* 3: storeb */ + ptr0[i] = var34; + } + } + +} + +#else +static void +_backup_cogorc_planar_chroma_422_420 (OrcExecutor * ex) +{ + int i; + int j; + int n = ex->n; + int m = ex->params[ORC_VAR_A1]; + orc_int8 *ORC_RESTRICT ptr0; + const orc_int8 *ORC_RESTRICT ptr4; + const orc_int8 *ORC_RESTRICT ptr5; + orc_int8 var32; + orc_int8 var33; + orc_int8 var34; + + for (j = 0; j < m; j++) { + ptr0 = ORC_PTR_OFFSET (ex->arrays[0], ex->params[0] * j); + ptr4 = ORC_PTR_OFFSET (ex->arrays[4], ex->params[4] * j); + ptr5 = ORC_PTR_OFFSET (ex->arrays[5], ex->params[5] * j); + + + for (i = 0; i < n; i++) { + /* 0: loadb */ + var32 = ptr4[i]; + /* 1: loadb */ + var33 = ptr5[i]; + /* 2: avgub */ + var34 = ((orc_uint8) var32 + (orc_uint8) var33 + 1) >> 1; + /* 3: storeb */ + ptr0[i] = var34; + } + } + +} + +void +cogorc_planar_chroma_422_420 (guint8 * d1, int d1_stride, const guint8 * s1, + int s1_stride, const guint8 * s2, int s2_stride, int n, int m) +{ + OrcExecutor _ex, *ex = &_ex; + static int p_inited = 0; + static OrcProgram *p = 0; + void (*func) (OrcExecutor *); + + if (!p_inited) { + orc_once_mutex_lock (); + if (!p_inited) { + OrcCompileResult result; + + p = orc_program_new (); + orc_program_set_2d (p); + orc_program_set_name (p, "cogorc_planar_chroma_422_420"); + orc_program_set_backup_function (p, _backup_cogorc_planar_chroma_422_420); + orc_program_add_destination (p, 1, "d1"); + orc_program_add_source (p, 1, "s1"); + orc_program_add_source (p, 1, "s2"); + orc_program_add_constant (p, 0, 0x00000000, "c1"); + orc_program_add_constant (p, 0, 0x00000000, "c2"); + orc_program_add_constant (p, 0, 0x00000000, "c3"); + orc_program_add_constant (p, 0, 0x00000000, "c4"); + orc_program_add_constant (p, 0, 0x00000000, "c5"); + orc_program_add_constant (p, 0, 0x00000000, "c6"); + orc_program_add_constant (p, 0, 0x00000000, "c7"); + orc_program_add_constant (p, 0, 0x00000000, "c8"); + + orc_program_append_2 (p, "avgub", 0, ORC_VAR_D1, ORC_VAR_S1, ORC_VAR_S2, + ORC_VAR_D1); + + result = orc_program_compile (p); + } + p_inited = TRUE; + orc_once_mutex_unlock (); + } + ex->program = p; + + ex->n = n; + ORC_EXECUTOR_M (ex) = m; + ex->arrays[ORC_VAR_D1] = d1; + ex->params[ORC_VAR_D1] = d1_stride; + ex->arrays[ORC_VAR_S1] = (void *) s1; + ex->params[ORC_VAR_S1] = s1_stride; + ex->arrays[ORC_VAR_S2] = (void *) s2; + ex->params[ORC_VAR_S2] = s2_stride; + + func = p->code_exec; + func (ex); +} +#endif + + +/* cogorc_convert_YUY2_AYUV */ +#ifdef DISABLE_ORC +void +cogorc_convert_YUY2_AYUV (guint8 * d1, int d1_stride, const guint8 * s1, + int s1_stride, int n, int m) +{ + int i; + int j; + orc_union64 *ORC_RESTRICT ptr0; + const orc_union32 *ORC_RESTRICT ptr4; + orc_union32 var36; + orc_union16 var37; + orc_union64 var38; + orc_union16 var39; + orc_union16 var40; + orc_union32 var41; + orc_union32 var42; + + for (j = 0; j < m; j++) { + ptr0 = ORC_PTR_OFFSET (d1, d1_stride * j); + ptr4 = ORC_PTR_OFFSET (s1, s1_stride * j); + + /* 2: loadpb */ + var37.x2[0] = 0x000000ff; /* 255 or 1.25987e-321f */ + var37.x2[1] = 0x000000ff; /* 255 or 1.25987e-321f */ + + for (i = 0; i < n; i++) { + /* 0: loadl */ + var36 = ptr4[i]; + /* 1: splitwb */ + var39.x2[0] = (var36.x2[0] >> 8) & 0xff; + var40.x2[0] = var36.x2[0] & 0xff; + var39.x2[1] = (var36.x2[1] >> 8) & 0xff; + var40.x2[1] = var36.x2[1] & 0xff; + /* 3: mergebw */ + var41.x2[0] = ((orc_uint8) var37.x2[0]) | ((orc_uint8) var40.x2[0] << 8); + var41.x2[1] = ((orc_uint8) var37.x2[1]) | ((orc_uint8) var40.x2[1] << 8); + /* 4: mergewl */ + var42.i = ((orc_uint16) var39.i) | ((orc_uint16) var39.i << 16); + /* 5: mergewl */ + var38.x2[0] = + ((orc_uint16) var41.x2[0]) | ((orc_uint16) var42.x2[0] << 16); + var38.x2[1] = + ((orc_uint16) var41.x2[1]) | ((orc_uint16) var42.x2[1] << 16); + /* 6: storeq */ + ptr0[i] = var38; + } + } + +} + +#else +static void +_backup_cogorc_convert_YUY2_AYUV (OrcExecutor * ex) +{ + int i; + int j; + int n = ex->n; + int m = ex->params[ORC_VAR_A1]; + orc_union64 *ORC_RESTRICT ptr0; + const orc_union32 *ORC_RESTRICT ptr4; + orc_union32 var36; + orc_union16 var37; + orc_union64 var38; + orc_union16 var39; + orc_union16 var40; + orc_union32 var41; + orc_union32 var42; + + for (j = 0; j < m; j++) { + ptr0 = ORC_PTR_OFFSET (ex->arrays[0], ex->params[0] * j); + ptr4 = ORC_PTR_OFFSET (ex->arrays[4], ex->params[4] * j); + + /* 2: loadpb */ + var37.x2[0] = 0x000000ff; /* 255 or 1.25987e-321f */ + var37.x2[1] = 0x000000ff; /* 255 or 1.25987e-321f */ + + for (i = 0; i < n; i++) { + /* 0: loadl */ + var36 = ptr4[i]; + /* 1: splitwb */ + var39.x2[0] = (var36.x2[0] >> 8) & 0xff; + var40.x2[0] = var36.x2[0] & 0xff; + var39.x2[1] = (var36.x2[1] >> 8) & 0xff; + var40.x2[1] = var36.x2[1] & 0xff; + /* 3: mergebw */ + var41.x2[0] = ((orc_uint8) var37.x2[0]) | ((orc_uint8) var40.x2[0] << 8); + var41.x2[1] = ((orc_uint8) var37.x2[1]) | ((orc_uint8) var40.x2[1] << 8); + /* 4: mergewl */ + var42.i = ((orc_uint16) var39.i) | ((orc_uint16) var39.i << 16); + /* 5: mergewl */ + var38.x2[0] = + ((orc_uint16) var41.x2[0]) | ((orc_uint16) var42.x2[0] << 16); + var38.x2[1] = + ((orc_uint16) var41.x2[1]) | ((orc_uint16) var42.x2[1] << 16); + /* 6: storeq */ + ptr0[i] = var38; + } + } + +} + +void +cogorc_convert_YUY2_AYUV (guint8 * d1, int d1_stride, const guint8 * s1, + int s1_stride, int n, int m) +{ + OrcExecutor _ex, *ex = &_ex; + static int p_inited = 0; + static OrcProgram *p = 0; + void (*func) (OrcExecutor *); + + if (!p_inited) { + orc_once_mutex_lock (); + if (!p_inited) { + OrcCompileResult result; + + p = orc_program_new (); + orc_program_set_2d (p); + orc_program_set_name (p, "cogorc_convert_YUY2_AYUV"); + orc_program_set_backup_function (p, _backup_cogorc_convert_YUY2_AYUV); + orc_program_add_destination (p, 8, "d1"); + orc_program_add_source (p, 4, "s1"); + orc_program_add_constant (p, 2, 0x000000ff, "c1"); + orc_program_add_constant (p, 0, 0x00000000, "c2"); + orc_program_add_constant (p, 0, 0x00000000, "c3"); + orc_program_add_constant (p, 0, 0x00000000, "c4"); + orc_program_add_constant (p, 0, 0x00000000, "c5"); + orc_program_add_constant (p, 0, 0x00000000, "c6"); + orc_program_add_constant (p, 0, 0x00000000, "c7"); + orc_program_add_constant (p, 0, 0x00000000, "c8"); + orc_program_add_temporary (p, 2, "t1"); + orc_program_add_temporary (p, 2, "t2"); + orc_program_add_temporary (p, 4, "t3"); + orc_program_add_temporary (p, 4, "t4"); + + orc_program_append_2 (p, "splitwb", 1, ORC_VAR_T2, ORC_VAR_T1, ORC_VAR_S1, + ORC_VAR_D1); + orc_program_append_2 (p, "mergebw", 1, ORC_VAR_T3, ORC_VAR_C1, ORC_VAR_T1, + ORC_VAR_D1); + orc_program_append_2 (p, "mergewl", 0, ORC_VAR_T4, ORC_VAR_T2, ORC_VAR_T2, + ORC_VAR_D1); + orc_program_append_2 (p, "mergewl", 1, ORC_VAR_D1, ORC_VAR_T3, ORC_VAR_T4, + ORC_VAR_D1); + + result = orc_program_compile (p); + } + p_inited = TRUE; + orc_once_mutex_unlock (); + } + ex->program = p; + + ex->n = n; + ORC_EXECUTOR_M (ex) = m; + ex->arrays[ORC_VAR_D1] = d1; + ex->params[ORC_VAR_D1] = d1_stride; + ex->arrays[ORC_VAR_S1] = (void *) s1; + ex->params[ORC_VAR_S1] = s1_stride; + + func = p->code_exec; + func (ex); +} +#endif + + +/* cogorc_convert_UYVY_AYUV */ +#ifdef DISABLE_ORC +void +cogorc_convert_UYVY_AYUV (guint8 * d1, int d1_stride, const guint8 * s1, + int s1_stride, int n, int m) +{ + int i; + int j; + orc_union64 *ORC_RESTRICT ptr0; + const orc_union32 *ORC_RESTRICT ptr4; + orc_union32 var36; + orc_union16 var37; + orc_union64 var38; + orc_union16 var39; + orc_union16 var40; + orc_union32 var41; + orc_union32 var42; + + for (j = 0; j < m; j++) { + ptr0 = ORC_PTR_OFFSET (d1, d1_stride * j); + ptr4 = ORC_PTR_OFFSET (s1, s1_stride * j); + + /* 2: loadpb */ + var37.x2[0] = 0x000000ff; /* 255 or 1.25987e-321f */ + var37.x2[1] = 0x000000ff; /* 255 or 1.25987e-321f */ + + for (i = 0; i < n; i++) { + /* 0: loadl */ + var36 = ptr4[i]; + /* 1: splitwb */ + var39.x2[0] = (var36.x2[0] >> 8) & 0xff; + var40.x2[0] = var36.x2[0] & 0xff; + var39.x2[1] = (var36.x2[1] >> 8) & 0xff; + var40.x2[1] = var36.x2[1] & 0xff; + /* 3: mergebw */ + var41.x2[0] = ((orc_uint8) var37.x2[0]) | ((orc_uint8) var39.x2[0] << 8); + var41.x2[1] = ((orc_uint8) var37.x2[1]) | ((orc_uint8) var39.x2[1] << 8); + /* 4: mergewl */ + var42.i = ((orc_uint16) var40.i) | ((orc_uint16) var40.i << 16); + /* 5: mergewl */ + var38.x2[0] = + ((orc_uint16) var41.x2[0]) | ((orc_uint16) var42.x2[0] << 16); + var38.x2[1] = + ((orc_uint16) var41.x2[1]) | ((orc_uint16) var42.x2[1] << 16); + /* 6: storeq */ + ptr0[i] = var38; + } + } + +} + +#else +static void +_backup_cogorc_convert_UYVY_AYUV (OrcExecutor * ex) +{ + int i; + int j; + int n = ex->n; + int m = ex->params[ORC_VAR_A1]; + orc_union64 *ORC_RESTRICT ptr0; + const orc_union32 *ORC_RESTRICT ptr4; + orc_union32 var36; + orc_union16 var37; + orc_union64 var38; + orc_union16 var39; + orc_union16 var40; + orc_union32 var41; + orc_union32 var42; + + for (j = 0; j < m; j++) { + ptr0 = ORC_PTR_OFFSET (ex->arrays[0], ex->params[0] * j); + ptr4 = ORC_PTR_OFFSET (ex->arrays[4], ex->params[4] * j); + + /* 2: loadpb */ + var37.x2[0] = 0x000000ff; /* 255 or 1.25987e-321f */ + var37.x2[1] = 0x000000ff; /* 255 or 1.25987e-321f */ + + for (i = 0; i < n; i++) { + /* 0: loadl */ + var36 = ptr4[i]; + /* 1: splitwb */ + var39.x2[0] = (var36.x2[0] >> 8) & 0xff; + var40.x2[0] = var36.x2[0] & 0xff; + var39.x2[1] = (var36.x2[1] >> 8) & 0xff; + var40.x2[1] = var36.x2[1] & 0xff; + /* 3: mergebw */ + var41.x2[0] = ((orc_uint8) var37.x2[0]) | ((orc_uint8) var39.x2[0] << 8); + var41.x2[1] = ((orc_uint8) var37.x2[1]) | ((orc_uint8) var39.x2[1] << 8); + /* 4: mergewl */ + var42.i = ((orc_uint16) var40.i) | ((orc_uint16) var40.i << 16); + /* 5: mergewl */ + var38.x2[0] = + ((orc_uint16) var41.x2[0]) | ((orc_uint16) var42.x2[0] << 16); + var38.x2[1] = + ((orc_uint16) var41.x2[1]) | ((orc_uint16) var42.x2[1] << 16); + /* 6: storeq */ + ptr0[i] = var38; + } + } + +} + +void +cogorc_convert_UYVY_AYUV (guint8 * d1, int d1_stride, const guint8 * s1, + int s1_stride, int n, int m) +{ + OrcExecutor _ex, *ex = &_ex; + static int p_inited = 0; + static OrcProgram *p = 0; + void (*func) (OrcExecutor *); + + if (!p_inited) { + orc_once_mutex_lock (); + if (!p_inited) { + OrcCompileResult result; + + p = orc_program_new (); + orc_program_set_2d (p); + orc_program_set_name (p, "cogorc_convert_UYVY_AYUV"); + orc_program_set_backup_function (p, _backup_cogorc_convert_UYVY_AYUV); + orc_program_add_destination (p, 8, "d1"); + orc_program_add_source (p, 4, "s1"); + orc_program_add_constant (p, 2, 0x000000ff, "c1"); + orc_program_add_constant (p, 0, 0x00000000, "c2"); + orc_program_add_constant (p, 0, 0x00000000, "c3"); + orc_program_add_constant (p, 0, 0x00000000, "c4"); + orc_program_add_constant (p, 0, 0x00000000, "c5"); + orc_program_add_constant (p, 0, 0x00000000, "c6"); + orc_program_add_constant (p, 0, 0x00000000, "c7"); + orc_program_add_constant (p, 0, 0x00000000, "c8"); + orc_program_add_temporary (p, 2, "t1"); + orc_program_add_temporary (p, 2, "t2"); + orc_program_add_temporary (p, 4, "t3"); + orc_program_add_temporary (p, 4, "t4"); + + orc_program_append_2 (p, "splitwb", 1, ORC_VAR_T1, ORC_VAR_T2, ORC_VAR_S1, + ORC_VAR_D1); + orc_program_append_2 (p, "mergebw", 1, ORC_VAR_T3, ORC_VAR_C1, ORC_VAR_T1, + ORC_VAR_D1); + orc_program_append_2 (p, "mergewl", 0, ORC_VAR_T4, ORC_VAR_T2, ORC_VAR_T2, + ORC_VAR_D1); + orc_program_append_2 (p, "mergewl", 1, ORC_VAR_D1, ORC_VAR_T3, ORC_VAR_T4, + ORC_VAR_D1); + + result = orc_program_compile (p); + } + p_inited = TRUE; + orc_once_mutex_unlock (); + } + ex->program = p; + + ex->n = n; + ORC_EXECUTOR_M (ex) = m; + ex->arrays[ORC_VAR_D1] = d1; + ex->params[ORC_VAR_D1] = d1_stride; + ex->arrays[ORC_VAR_S1] = (void *) s1; + ex->params[ORC_VAR_S1] = s1_stride; + + func = p->code_exec; + func (ex); +} +#endif + + +/* cogorc_convert_YUY2_Y42B */ +#ifdef DISABLE_ORC +void +cogorc_convert_YUY2_Y42B (guint8 * d1, int d1_stride, guint8 * d2, + int d2_stride, guint8 * d3, int d3_stride, const guint8 * s1, int s1_stride, + int n, int m) +{ + int i; + int j; + orc_union16 *ORC_RESTRICT ptr0; + orc_int8 *ORC_RESTRICT ptr1; + orc_int8 *ORC_RESTRICT ptr2; + const orc_union32 *ORC_RESTRICT ptr4; + orc_union32 var33; + orc_union16 var34; + orc_int8 var35; + orc_int8 var36; + orc_union16 var37; + + for (j = 0; j < m; j++) { + ptr0 = ORC_PTR_OFFSET (d1, d1_stride * j); + ptr1 = ORC_PTR_OFFSET (d2, d2_stride * j); + ptr2 = ORC_PTR_OFFSET (d3, d3_stride * j); + ptr4 = ORC_PTR_OFFSET (s1, s1_stride * j); + + + for (i = 0; i < n; i++) { + /* 0: loadl */ + var33 = ptr4[i]; + /* 1: splitwb */ + var37.x2[0] = (var33.x2[0] >> 8) & 0xff; + var34.x2[0] = var33.x2[0] & 0xff; + var37.x2[1] = (var33.x2[1] >> 8) & 0xff; + var34.x2[1] = var33.x2[1] & 0xff; + /* 2: storew */ + ptr0[i] = var34; + /* 3: splitwb */ + var35 = (var37.i >> 8) & 0xff; + var36 = var37.i & 0xff; + /* 4: storeb */ + ptr2[i] = var35; + /* 5: storeb */ + ptr1[i] = var36; + } + } + +} + +#else +static void +_backup_cogorc_convert_YUY2_Y42B (OrcExecutor * ex) +{ + int i; + int j; + int n = ex->n; + int m = ex->params[ORC_VAR_A1]; + orc_union16 *ORC_RESTRICT ptr0; + orc_int8 *ORC_RESTRICT ptr1; + orc_int8 *ORC_RESTRICT ptr2; + const orc_union32 *ORC_RESTRICT ptr4; + orc_union32 var33; + orc_union16 var34; + orc_int8 var35; + orc_int8 var36; + orc_union16 var37; + + for (j = 0; j < m; j++) { + ptr0 = ORC_PTR_OFFSET (ex->arrays[0], ex->params[0] * j); + ptr1 = ORC_PTR_OFFSET (ex->arrays[1], ex->params[1] * j); + ptr2 = ORC_PTR_OFFSET (ex->arrays[2], ex->params[2] * j); + ptr4 = ORC_PTR_OFFSET (ex->arrays[4], ex->params[4] * j); + + + for (i = 0; i < n; i++) { + /* 0: loadl */ + var33 = ptr4[i]; + /* 1: splitwb */ + var37.x2[0] = (var33.x2[0] >> 8) & 0xff; + var34.x2[0] = var33.x2[0] & 0xff; + var37.x2[1] = (var33.x2[1] >> 8) & 0xff; + var34.x2[1] = var33.x2[1] & 0xff; + /* 2: storew */ + ptr0[i] = var34; + /* 3: splitwb */ + var35 = (var37.i >> 8) & 0xff; + var36 = var37.i & 0xff; + /* 4: storeb */ + ptr2[i] = var35; + /* 5: storeb */ + ptr1[i] = var36; + } + } + +} + +void +cogorc_convert_YUY2_Y42B (guint8 * d1, int d1_stride, guint8 * d2, + int d2_stride, guint8 * d3, int d3_stride, const guint8 * s1, int s1_stride, + int n, int m) +{ + OrcExecutor _ex, *ex = &_ex; + static int p_inited = 0; + static OrcProgram *p = 0; + void (*func) (OrcExecutor *); + + if (!p_inited) { + orc_once_mutex_lock (); + if (!p_inited) { + OrcCompileResult result; + + p = orc_program_new (); + orc_program_set_2d (p); + orc_program_set_name (p, "cogorc_convert_YUY2_Y42B"); + orc_program_set_backup_function (p, _backup_cogorc_convert_YUY2_Y42B); + orc_program_add_destination (p, 2, "d1"); + orc_program_add_destination (p, 1, "d2"); + orc_program_add_destination (p, 1, "d3"); + orc_program_add_source (p, 4, "s1"); + orc_program_add_constant (p, 0, 0x00000000, "c1"); + orc_program_add_constant (p, 0, 0x00000000, "c2"); + orc_program_add_constant (p, 0, 0x00000000, "c3"); + orc_program_add_constant (p, 0, 0x00000000, "c4"); + orc_program_add_constant (p, 0, 0x00000000, "c5"); + orc_program_add_constant (p, 0, 0x00000000, "c6"); + orc_program_add_constant (p, 0, 0x00000000, "c7"); + orc_program_add_constant (p, 0, 0x00000000, "c8"); + orc_program_add_temporary (p, 2, "t1"); + + orc_program_append_2 (p, "splitwb", 1, ORC_VAR_T1, ORC_VAR_D1, ORC_VAR_S1, + ORC_VAR_D1); + orc_program_append_2 (p, "splitwb", 0, ORC_VAR_D3, ORC_VAR_D2, ORC_VAR_T1, + ORC_VAR_D1); + + result = orc_program_compile (p); + } + p_inited = TRUE; + orc_once_mutex_unlock (); + } + ex->program = p; + + ex->n = n; + ORC_EXECUTOR_M (ex) = m; + ex->arrays[ORC_VAR_D1] = d1; + ex->params[ORC_VAR_D1] = d1_stride; + ex->arrays[ORC_VAR_D2] = d2; + ex->params[ORC_VAR_D2] = d2_stride; + ex->arrays[ORC_VAR_D3] = d3; + ex->params[ORC_VAR_D3] = d3_stride; + ex->arrays[ORC_VAR_S1] = (void *) s1; + ex->params[ORC_VAR_S1] = s1_stride; + + func = p->code_exec; + func (ex); +} +#endif + + +/* cogorc_convert_UYVY_Y42B */ +#ifdef DISABLE_ORC +void +cogorc_convert_UYVY_Y42B (guint8 * d1, int d1_stride, guint8 * d2, + int d2_stride, guint8 * d3, int d3_stride, const guint8 * s1, int s1_stride, + int n, int m) +{ + int i; + int j; + orc_union16 *ORC_RESTRICT ptr0; + orc_int8 *ORC_RESTRICT ptr1; + orc_int8 *ORC_RESTRICT ptr2; + const orc_union32 *ORC_RESTRICT ptr4; + orc_union32 var33; + orc_union16 var34; + orc_int8 var35; + orc_int8 var36; + orc_union16 var37; + + for (j = 0; j < m; j++) { + ptr0 = ORC_PTR_OFFSET (d1, d1_stride * j); + ptr1 = ORC_PTR_OFFSET (d2, d2_stride * j); + ptr2 = ORC_PTR_OFFSET (d3, d3_stride * j); + ptr4 = ORC_PTR_OFFSET (s1, s1_stride * j); + + + for (i = 0; i < n; i++) { + /* 0: loadl */ + var33 = ptr4[i]; + /* 1: splitwb */ + var34.x2[0] = (var33.x2[0] >> 8) & 0xff; + var37.x2[0] = var33.x2[0] & 0xff; + var34.x2[1] = (var33.x2[1] >> 8) & 0xff; + var37.x2[1] = var33.x2[1] & 0xff; + /* 2: storew */ + ptr0[i] = var34; + /* 3: splitwb */ + var35 = (var37.i >> 8) & 0xff; + var36 = var37.i & 0xff; + /* 4: storeb */ + ptr2[i] = var35; + /* 5: storeb */ + ptr1[i] = var36; + } + } + +} + +#else +static void +_backup_cogorc_convert_UYVY_Y42B (OrcExecutor * ex) +{ + int i; + int j; + int n = ex->n; + int m = ex->params[ORC_VAR_A1]; + orc_union16 *ORC_RESTRICT ptr0; + orc_int8 *ORC_RESTRICT ptr1; + orc_int8 *ORC_RESTRICT ptr2; + const orc_union32 *ORC_RESTRICT ptr4; + orc_union32 var33; + orc_union16 var34; + orc_int8 var35; + orc_int8 var36; + orc_union16 var37; + + for (j = 0; j < m; j++) { + ptr0 = ORC_PTR_OFFSET (ex->arrays[0], ex->params[0] * j); + ptr1 = ORC_PTR_OFFSET (ex->arrays[1], ex->params[1] * j); + ptr2 = ORC_PTR_OFFSET (ex->arrays[2], ex->params[2] * j); + ptr4 = ORC_PTR_OFFSET (ex->arrays[4], ex->params[4] * j); + + + for (i = 0; i < n; i++) { + /* 0: loadl */ + var33 = ptr4[i]; + /* 1: splitwb */ + var34.x2[0] = (var33.x2[0] >> 8) & 0xff; + var37.x2[0] = var33.x2[0] & 0xff; + var34.x2[1] = (var33.x2[1] >> 8) & 0xff; + var37.x2[1] = var33.x2[1] & 0xff; + /* 2: storew */ + ptr0[i] = var34; + /* 3: splitwb */ + var35 = (var37.i >> 8) & 0xff; + var36 = var37.i & 0xff; + /* 4: storeb */ + ptr2[i] = var35; + /* 5: storeb */ + ptr1[i] = var36; + } + } + +} + +void +cogorc_convert_UYVY_Y42B (guint8 * d1, int d1_stride, guint8 * d2, + int d2_stride, guint8 * d3, int d3_stride, const guint8 * s1, int s1_stride, + int n, int m) +{ + OrcExecutor _ex, *ex = &_ex; + static int p_inited = 0; + static OrcProgram *p = 0; + void (*func) (OrcExecutor *); + + if (!p_inited) { + orc_once_mutex_lock (); + if (!p_inited) { + OrcCompileResult result; + + p = orc_program_new (); + orc_program_set_2d (p); + orc_program_set_name (p, "cogorc_convert_UYVY_Y42B"); + orc_program_set_backup_function (p, _backup_cogorc_convert_UYVY_Y42B); + orc_program_add_destination (p, 2, "d1"); + orc_program_add_destination (p, 1, "d2"); + orc_program_add_destination (p, 1, "d3"); + orc_program_add_source (p, 4, "s1"); + orc_program_add_constant (p, 0, 0x00000000, "c1"); + orc_program_add_constant (p, 0, 0x00000000, "c2"); + orc_program_add_constant (p, 0, 0x00000000, "c3"); + orc_program_add_constant (p, 0, 0x00000000, "c4"); + orc_program_add_constant (p, 0, 0x00000000, "c5"); + orc_program_add_constant (p, 0, 0x00000000, "c6"); + orc_program_add_constant (p, 0, 0x00000000, "c7"); + orc_program_add_constant (p, 0, 0x00000000, "c8"); + orc_program_add_temporary (p, 2, "t1"); + + orc_program_append_2 (p, "splitwb", 1, ORC_VAR_D1, ORC_VAR_T1, ORC_VAR_S1, + ORC_VAR_D1); + orc_program_append_2 (p, "splitwb", 0, ORC_VAR_D3, ORC_VAR_D2, ORC_VAR_T1, + ORC_VAR_D1); + + result = orc_program_compile (p); + } + p_inited = TRUE; + orc_once_mutex_unlock (); + } + ex->program = p; + + ex->n = n; + ORC_EXECUTOR_M (ex) = m; + ex->arrays[ORC_VAR_D1] = d1; + ex->params[ORC_VAR_D1] = d1_stride; + ex->arrays[ORC_VAR_D2] = d2; + ex->params[ORC_VAR_D2] = d2_stride; + ex->arrays[ORC_VAR_D3] = d3; + ex->params[ORC_VAR_D3] = d3_stride; + ex->arrays[ORC_VAR_S1] = (void *) s1; + ex->params[ORC_VAR_S1] = s1_stride; + + func = p->code_exec; + func (ex); +} +#endif + + +/* cogorc_convert_YUY2_Y444 */ +#ifdef DISABLE_ORC +void +cogorc_convert_YUY2_Y444 (guint8 * d1, int d1_stride, guint8 * d2, + int d2_stride, guint8 * d3, int d3_stride, const guint8 * s1, int s1_stride, + int n, int m) +{ + int i; + int j; + orc_union16 *ORC_RESTRICT ptr0; + orc_union16 *ORC_RESTRICT ptr1; + orc_union16 *ORC_RESTRICT ptr2; + const orc_union32 *ORC_RESTRICT ptr4; + orc_union32 var35; + orc_union16 var36; + orc_union16 var37; + orc_union16 var38; + orc_union16 var39; + orc_int8 var40; + orc_int8 var41; + + for (j = 0; j < m; j++) { + ptr0 = ORC_PTR_OFFSET (d1, d1_stride * j); + ptr1 = ORC_PTR_OFFSET (d2, d2_stride * j); + ptr2 = ORC_PTR_OFFSET (d3, d3_stride * j); + ptr4 = ORC_PTR_OFFSET (s1, s1_stride * j); + + + for (i = 0; i < n; i++) { + /* 0: loadl */ + var35 = ptr4[i]; + /* 1: splitwb */ + var39.x2[0] = (var35.x2[0] >> 8) & 0xff; + var36.x2[0] = var35.x2[0] & 0xff; + var39.x2[1] = (var35.x2[1] >> 8) & 0xff; + var36.x2[1] = var35.x2[1] & 0xff; + /* 2: storew */ + ptr0[i] = var36; + /* 3: splitwb */ + var40 = (var39.i >> 8) & 0xff; + var41 = var39.i & 0xff; + /* 4: splatbw */ + var37.i = ((var41 & 0xff) << 8) | (var41 & 0xff); + /* 5: storew */ + ptr1[i] = var37; + /* 6: splatbw */ + var38.i = ((var40 & 0xff) << 8) | (var40 & 0xff); + /* 7: storew */ + ptr2[i] = var38; + } + } + +} + +#else +static void +_backup_cogorc_convert_YUY2_Y444 (OrcExecutor * ex) +{ + int i; + int j; + int n = ex->n; + int m = ex->params[ORC_VAR_A1]; + orc_union16 *ORC_RESTRICT ptr0; + orc_union16 *ORC_RESTRICT ptr1; + orc_union16 *ORC_RESTRICT ptr2; + const orc_union32 *ORC_RESTRICT ptr4; + orc_union32 var35; + orc_union16 var36; + orc_union16 var37; + orc_union16 var38; + orc_union16 var39; + orc_int8 var40; + orc_int8 var41; + + for (j = 0; j < m; j++) { + ptr0 = ORC_PTR_OFFSET (ex->arrays[0], ex->params[0] * j); + ptr1 = ORC_PTR_OFFSET (ex->arrays[1], ex->params[1] * j); + ptr2 = ORC_PTR_OFFSET (ex->arrays[2], ex->params[2] * j); + ptr4 = ORC_PTR_OFFSET (ex->arrays[4], ex->params[4] * j); + + + for (i = 0; i < n; i++) { + /* 0: loadl */ + var35 = ptr4[i]; + /* 1: splitwb */ + var39.x2[0] = (var35.x2[0] >> 8) & 0xff; + var36.x2[0] = var35.x2[0] & 0xff; + var39.x2[1] = (var35.x2[1] >> 8) & 0xff; + var36.x2[1] = var35.x2[1] & 0xff; + /* 2: storew */ + ptr0[i] = var36; + /* 3: splitwb */ + var40 = (var39.i >> 8) & 0xff; + var41 = var39.i & 0xff; + /* 4: splatbw */ + var37.i = ((var41 & 0xff) << 8) | (var41 & 0xff); + /* 5: storew */ + ptr1[i] = var37; + /* 6: splatbw */ + var38.i = ((var40 & 0xff) << 8) | (var40 & 0xff); + /* 7: storew */ + ptr2[i] = var38; + } + } + +} + +void +cogorc_convert_YUY2_Y444 (guint8 * d1, int d1_stride, guint8 * d2, + int d2_stride, guint8 * d3, int d3_stride, const guint8 * s1, int s1_stride, + int n, int m) +{ + OrcExecutor _ex, *ex = &_ex; + static int p_inited = 0; + static OrcProgram *p = 0; + void (*func) (OrcExecutor *); + + if (!p_inited) { + orc_once_mutex_lock (); + if (!p_inited) { + OrcCompileResult result; + + p = orc_program_new (); + orc_program_set_2d (p); + orc_program_set_name (p, "cogorc_convert_YUY2_Y444"); + orc_program_set_backup_function (p, _backup_cogorc_convert_YUY2_Y444); + orc_program_add_destination (p, 2, "d1"); + orc_program_add_destination (p, 2, "d2"); + orc_program_add_destination (p, 2, "d3"); + orc_program_add_source (p, 4, "s1"); + orc_program_add_constant (p, 0, 0x00000000, "c1"); + orc_program_add_constant (p, 0, 0x00000000, "c2"); + orc_program_add_constant (p, 0, 0x00000000, "c3"); + orc_program_add_constant (p, 0, 0x00000000, "c4"); + orc_program_add_constant (p, 0, 0x00000000, "c5"); + orc_program_add_constant (p, 0, 0x00000000, "c6"); + orc_program_add_constant (p, 0, 0x00000000, "c7"); + orc_program_add_constant (p, 0, 0x00000000, "c8"); + orc_program_add_temporary (p, 2, "t1"); + orc_program_add_temporary (p, 1, "t2"); + orc_program_add_temporary (p, 1, "t3"); + + orc_program_append_2 (p, "splitwb", 1, ORC_VAR_T1, ORC_VAR_D1, ORC_VAR_S1, + ORC_VAR_D1); + orc_program_append_2 (p, "splitwb", 0, ORC_VAR_T3, ORC_VAR_T2, ORC_VAR_T1, + ORC_VAR_D1); + orc_program_append_2 (p, "splatbw", 0, ORC_VAR_D2, ORC_VAR_T2, ORC_VAR_D1, + ORC_VAR_D1); + orc_program_append_2 (p, "splatbw", 0, ORC_VAR_D3, ORC_VAR_T3, ORC_VAR_D1, + ORC_VAR_D1); + + result = orc_program_compile (p); + } + p_inited = TRUE; + orc_once_mutex_unlock (); + } + ex->program = p; + + ex->n = n; + ORC_EXECUTOR_M (ex) = m; + ex->arrays[ORC_VAR_D1] = d1; + ex->params[ORC_VAR_D1] = d1_stride; + ex->arrays[ORC_VAR_D2] = d2; + ex->params[ORC_VAR_D2] = d2_stride; + ex->arrays[ORC_VAR_D3] = d3; + ex->params[ORC_VAR_D3] = d3_stride; + ex->arrays[ORC_VAR_S1] = (void *) s1; + ex->params[ORC_VAR_S1] = s1_stride; + + func = p->code_exec; + func (ex); +} +#endif + + +/* cogorc_convert_UYVY_Y444 */ +#ifdef DISABLE_ORC +void +cogorc_convert_UYVY_Y444 (guint8 * d1, int d1_stride, guint8 * d2, + int d2_stride, guint8 * d3, int d3_stride, const guint8 * s1, int s1_stride, + int n, int m) +{ + int i; + int j; + orc_union16 *ORC_RESTRICT ptr0; + orc_union16 *ORC_RESTRICT ptr1; + orc_union16 *ORC_RESTRICT ptr2; + const orc_union32 *ORC_RESTRICT ptr4; + orc_union32 var35; + orc_union16 var36; + orc_union16 var37; + orc_union16 var38; + orc_union16 var39; + orc_int8 var40; + orc_int8 var41; + + for (j = 0; j < m; j++) { + ptr0 = ORC_PTR_OFFSET (d1, d1_stride * j); + ptr1 = ORC_PTR_OFFSET (d2, d2_stride * j); + ptr2 = ORC_PTR_OFFSET (d3, d3_stride * j); + ptr4 = ORC_PTR_OFFSET (s1, s1_stride * j); + + + for (i = 0; i < n; i++) { + /* 0: loadl */ + var35 = ptr4[i]; + /* 1: splitwb */ + var36.x2[0] = (var35.x2[0] >> 8) & 0xff; + var39.x2[0] = var35.x2[0] & 0xff; + var36.x2[1] = (var35.x2[1] >> 8) & 0xff; + var39.x2[1] = var35.x2[1] & 0xff; + /* 2: storew */ + ptr0[i] = var36; + /* 3: splitwb */ + var40 = (var39.i >> 8) & 0xff; + var41 = var39.i & 0xff; + /* 4: splatbw */ + var37.i = ((var41 & 0xff) << 8) | (var41 & 0xff); + /* 5: storew */ + ptr1[i] = var37; + /* 6: splatbw */ + var38.i = ((var40 & 0xff) << 8) | (var40 & 0xff); + /* 7: storew */ + ptr2[i] = var38; + } + } + +} + +#else +static void +_backup_cogorc_convert_UYVY_Y444 (OrcExecutor * ex) +{ + int i; + int j; + int n = ex->n; + int m = ex->params[ORC_VAR_A1]; + orc_union16 *ORC_RESTRICT ptr0; + orc_union16 *ORC_RESTRICT ptr1; + orc_union16 *ORC_RESTRICT ptr2; + const orc_union32 *ORC_RESTRICT ptr4; + orc_union32 var35; + orc_union16 var36; + orc_union16 var37; + orc_union16 var38; + orc_union16 var39; + orc_int8 var40; + orc_int8 var41; + + for (j = 0; j < m; j++) { + ptr0 = ORC_PTR_OFFSET (ex->arrays[0], ex->params[0] * j); + ptr1 = ORC_PTR_OFFSET (ex->arrays[1], ex->params[1] * j); + ptr2 = ORC_PTR_OFFSET (ex->arrays[2], ex->params[2] * j); + ptr4 = ORC_PTR_OFFSET (ex->arrays[4], ex->params[4] * j); + + + for (i = 0; i < n; i++) { + /* 0: loadl */ + var35 = ptr4[i]; + /* 1: splitwb */ + var36.x2[0] = (var35.x2[0] >> 8) & 0xff; + var39.x2[0] = var35.x2[0] & 0xff; + var36.x2[1] = (var35.x2[1] >> 8) & 0xff; + var39.x2[1] = var35.x2[1] & 0xff; + /* 2: storew */ + ptr0[i] = var36; + /* 3: splitwb */ + var40 = (var39.i >> 8) & 0xff; + var41 = var39.i & 0xff; + /* 4: splatbw */ + var37.i = ((var41 & 0xff) << 8) | (var41 & 0xff); + /* 5: storew */ + ptr1[i] = var37; + /* 6: splatbw */ + var38.i = ((var40 & 0xff) << 8) | (var40 & 0xff); + /* 7: storew */ + ptr2[i] = var38; + } + } + +} + +void +cogorc_convert_UYVY_Y444 (guint8 * d1, int d1_stride, guint8 * d2, + int d2_stride, guint8 * d3, int d3_stride, const guint8 * s1, int s1_stride, + int n, int m) +{ + OrcExecutor _ex, *ex = &_ex; + static int p_inited = 0; + static OrcProgram *p = 0; + void (*func) (OrcExecutor *); + + if (!p_inited) { + orc_once_mutex_lock (); + if (!p_inited) { + OrcCompileResult result; + + p = orc_program_new (); + orc_program_set_2d (p); + orc_program_set_name (p, "cogorc_convert_UYVY_Y444"); + orc_program_set_backup_function (p, _backup_cogorc_convert_UYVY_Y444); + orc_program_add_destination (p, 2, "d1"); + orc_program_add_destination (p, 2, "d2"); + orc_program_add_destination (p, 2, "d3"); + orc_program_add_source (p, 4, "s1"); + orc_program_add_constant (p, 0, 0x00000000, "c1"); + orc_program_add_constant (p, 0, 0x00000000, "c2"); + orc_program_add_constant (p, 0, 0x00000000, "c3"); + orc_program_add_constant (p, 0, 0x00000000, "c4"); + orc_program_add_constant (p, 0, 0x00000000, "c5"); + orc_program_add_constant (p, 0, 0x00000000, "c6"); + orc_program_add_constant (p, 0, 0x00000000, "c7"); + orc_program_add_constant (p, 0, 0x00000000, "c8"); + orc_program_add_temporary (p, 2, "t1"); + orc_program_add_temporary (p, 1, "t2"); + orc_program_add_temporary (p, 1, "t3"); + + orc_program_append_2 (p, "splitwb", 1, ORC_VAR_D1, ORC_VAR_T1, ORC_VAR_S1, + ORC_VAR_D1); + orc_program_append_2 (p, "splitwb", 0, ORC_VAR_T3, ORC_VAR_T2, ORC_VAR_T1, + ORC_VAR_D1); + orc_program_append_2 (p, "splatbw", 0, ORC_VAR_D2, ORC_VAR_T2, ORC_VAR_D1, + ORC_VAR_D1); + orc_program_append_2 (p, "splatbw", 0, ORC_VAR_D3, ORC_VAR_T3, ORC_VAR_D1, + ORC_VAR_D1); + + result = orc_program_compile (p); + } + p_inited = TRUE; + orc_once_mutex_unlock (); + } + ex->program = p; + + ex->n = n; + ORC_EXECUTOR_M (ex) = m; + ex->arrays[ORC_VAR_D1] = d1; + ex->params[ORC_VAR_D1] = d1_stride; + ex->arrays[ORC_VAR_D2] = d2; + ex->params[ORC_VAR_D2] = d2_stride; + ex->arrays[ORC_VAR_D3] = d3; + ex->params[ORC_VAR_D3] = d3_stride; + ex->arrays[ORC_VAR_S1] = (void *) s1; + ex->params[ORC_VAR_S1] = s1_stride; + + func = p->code_exec; + func (ex); +} +#endif + + +/* cogorc_convert_UYVY_I420 */ +#ifdef DISABLE_ORC +void +cogorc_convert_UYVY_I420 (guint8 * d1, guint8 * d2, guint8 * d3, guint8 * d4, + const guint8 * s1, const guint8 * s2, int n) +{ + int i; + orc_union16 *ORC_RESTRICT ptr0; + orc_union16 *ORC_RESTRICT ptr1; + orc_int8 *ORC_RESTRICT ptr2; + orc_int8 *ORC_RESTRICT ptr3; + const orc_union32 *ORC_RESTRICT ptr4; + const orc_union32 *ORC_RESTRICT ptr5; + orc_union32 var35; + orc_union32 var36; + orc_int8 var37; + orc_int8 var38; + orc_union16 var39; + orc_union16 var40; + orc_union16 var41; + orc_union16 var42; + orc_union16 var43; + + ptr0 = (orc_union16 *) d1; + ptr1 = (orc_union16 *) d2; + ptr2 = (orc_int8 *) d3; + ptr3 = (orc_int8 *) d4; + ptr4 = (orc_union32 *) s1; + ptr5 = (orc_union32 *) s2; + + + for (i = 0; i < n; i++) { + /* 0: loadl */ + var35 = ptr4[i]; + /* 1: splitwb */ + var39.x2[0] = (var35.x2[0] >> 8) & 0xff; + var40.x2[0] = var35.x2[0] & 0xff; + var39.x2[1] = (var35.x2[1] >> 8) & 0xff; + var40.x2[1] = var35.x2[1] & 0xff; + /* 2: storew */ + ptr0[i] = var39; + /* 3: loadl */ + var36 = ptr5[i]; + /* 4: splitwb */ + var41.x2[0] = (var36.x2[0] >> 8) & 0xff; + var42.x2[0] = var36.x2[0] & 0xff; + var41.x2[1] = (var36.x2[1] >> 8) & 0xff; + var42.x2[1] = var36.x2[1] & 0xff; + /* 5: storew */ + ptr1[i] = var41; + /* 6: avgub */ + var43.x2[0] = ((orc_uint8) var40.x2[0] + (orc_uint8) var42.x2[0] + 1) >> 1; + var43.x2[1] = ((orc_uint8) var40.x2[1] + (orc_uint8) var42.x2[1] + 1) >> 1; + /* 7: splitwb */ + var37 = (var43.i >> 8) & 0xff; + var38 = var43.i & 0xff; + /* 8: storeb */ + ptr3[i] = var37; + /* 9: storeb */ + ptr2[i] = var38; + } + +} + +#else +static void +_backup_cogorc_convert_UYVY_I420 (OrcExecutor * ex) +{ + int i; + int n = ex->n; + orc_union16 *ORC_RESTRICT ptr0; + orc_union16 *ORC_RESTRICT ptr1; + orc_int8 *ORC_RESTRICT ptr2; + orc_int8 *ORC_RESTRICT ptr3; + const orc_union32 *ORC_RESTRICT ptr4; + const orc_union32 *ORC_RESTRICT ptr5; + orc_union32 var35; + orc_union32 var36; + orc_int8 var37; + orc_int8 var38; + orc_union16 var39; + orc_union16 var40; + orc_union16 var41; + orc_union16 var42; + orc_union16 var43; + + ptr0 = (orc_union16 *) ex->arrays[0]; + ptr1 = (orc_union16 *) ex->arrays[1]; + ptr2 = (orc_int8 *) ex->arrays[2]; + ptr3 = (orc_int8 *) ex->arrays[3]; + ptr4 = (orc_union32 *) ex->arrays[4]; + ptr5 = (orc_union32 *) ex->arrays[5]; + + + for (i = 0; i < n; i++) { + /* 0: loadl */ + var35 = ptr4[i]; + /* 1: splitwb */ + var39.x2[0] = (var35.x2[0] >> 8) & 0xff; + var40.x2[0] = var35.x2[0] & 0xff; + var39.x2[1] = (var35.x2[1] >> 8) & 0xff; + var40.x2[1] = var35.x2[1] & 0xff; + /* 2: storew */ + ptr0[i] = var39; + /* 3: loadl */ + var36 = ptr5[i]; + /* 4: splitwb */ + var41.x2[0] = (var36.x2[0] >> 8) & 0xff; + var42.x2[0] = var36.x2[0] & 0xff; + var41.x2[1] = (var36.x2[1] >> 8) & 0xff; + var42.x2[1] = var36.x2[1] & 0xff; + /* 5: storew */ + ptr1[i] = var41; + /* 6: avgub */ + var43.x2[0] = ((orc_uint8) var40.x2[0] + (orc_uint8) var42.x2[0] + 1) >> 1; + var43.x2[1] = ((orc_uint8) var40.x2[1] + (orc_uint8) var42.x2[1] + 1) >> 1; + /* 7: splitwb */ + var37 = (var43.i >> 8) & 0xff; + var38 = var43.i & 0xff; + /* 8: storeb */ + ptr3[i] = var37; + /* 9: storeb */ + ptr2[i] = var38; + } + +} + +void +cogorc_convert_UYVY_I420 (guint8 * d1, guint8 * d2, guint8 * d3, guint8 * d4, + const guint8 * s1, const guint8 * s2, int n) +{ + OrcExecutor _ex, *ex = &_ex; + static int p_inited = 0; + static OrcProgram *p = 0; + void (*func) (OrcExecutor *); + + if (!p_inited) { + orc_once_mutex_lock (); + if (!p_inited) { + OrcCompileResult result; + + p = orc_program_new (); + orc_program_set_name (p, "cogorc_convert_UYVY_I420"); + orc_program_set_backup_function (p, _backup_cogorc_convert_UYVY_I420); + orc_program_add_destination (p, 2, "d1"); + orc_program_add_destination (p, 2, "d2"); + orc_program_add_destination (p, 1, "d3"); + orc_program_add_destination (p, 1, "d4"); + orc_program_add_source (p, 4, "s1"); + orc_program_add_source (p, 4, "s2"); + orc_program_add_constant (p, 0, 0x00000000, "c1"); + orc_program_add_constant (p, 0, 0x00000000, "c2"); + orc_program_add_constant (p, 0, 0x00000000, "c3"); + orc_program_add_constant (p, 0, 0x00000000, "c4"); + orc_program_add_constant (p, 0, 0x00000000, "c5"); + orc_program_add_constant (p, 0, 0x00000000, "c6"); + orc_program_add_constant (p, 0, 0x00000000, "c7"); + orc_program_add_constant (p, 0, 0x00000000, "c8"); + orc_program_add_temporary (p, 2, "t1"); + orc_program_add_temporary (p, 2, "t2"); + orc_program_add_temporary (p, 2, "t3"); + + orc_program_append_2 (p, "splitwb", 1, ORC_VAR_T3, ORC_VAR_T1, ORC_VAR_S1, + ORC_VAR_D1); + orc_program_append_2 (p, "storew", 0, ORC_VAR_D1, ORC_VAR_T3, ORC_VAR_D1, + ORC_VAR_D1); + orc_program_append_2 (p, "splitwb", 1, ORC_VAR_T3, ORC_VAR_T2, ORC_VAR_S2, + ORC_VAR_D1); + orc_program_append_2 (p, "storew", 0, ORC_VAR_D2, ORC_VAR_T3, ORC_VAR_D1, + ORC_VAR_D1); + orc_program_append_2 (p, "avgub", 1, ORC_VAR_T1, ORC_VAR_T1, ORC_VAR_T2, + ORC_VAR_D1); + orc_program_append_2 (p, "splitwb", 0, ORC_VAR_D4, ORC_VAR_D3, ORC_VAR_T1, + ORC_VAR_D1); + + result = orc_program_compile (p); + } + p_inited = TRUE; + orc_once_mutex_unlock (); + } + ex->program = p; + + ex->n = n; + ex->arrays[ORC_VAR_D1] = d1; + ex->arrays[ORC_VAR_D2] = d2; + ex->arrays[ORC_VAR_D3] = d3; + ex->arrays[ORC_VAR_D4] = d4; + ex->arrays[ORC_VAR_S1] = (void *) s1; + ex->arrays[ORC_VAR_S2] = (void *) s2; + + func = p->code_exec; + func (ex); +} +#endif + + +/* cogorc_convert_AYUV_I420 */ +#ifdef DISABLE_ORC +void +cogorc_convert_AYUV_I420 (guint8 * d1, int d1_stride, guint8 * d2, + int d2_stride, guint8 * d3, int d3_stride, guint8 * d4, int d4_stride, + const guint8 * s1, int s1_stride, const guint8 * s2, int s2_stride, int n, + int m) +{ + int i; + int j; + orc_union16 *ORC_RESTRICT ptr0; + orc_union16 *ORC_RESTRICT ptr1; + orc_int8 *ORC_RESTRICT ptr2; + orc_int8 *ORC_RESTRICT ptr3; + const orc_union64 *ORC_RESTRICT ptr4; + const orc_union64 *ORC_RESTRICT ptr5; + orc_union64 var40; + orc_union16 var41; + orc_union64 var42; + orc_union16 var43; + orc_int8 var44; + orc_int8 var45; + orc_union32 var46; + orc_union32 var47; + orc_union32 var48; + orc_union32 var49; + orc_union32 var50; + orc_union16 var51; + orc_union16 var52; + orc_int8 var53; + orc_int8 var54; + orc_int8 var55; + orc_int8 var56; + + for (j = 0; j < m; j++) { + ptr0 = ORC_PTR_OFFSET (d1, d1_stride * j); + ptr1 = ORC_PTR_OFFSET (d2, d2_stride * j); + ptr2 = ORC_PTR_OFFSET (d3, d3_stride * j); + ptr3 = ORC_PTR_OFFSET (d4, d4_stride * j); + ptr4 = ORC_PTR_OFFSET (s1, s1_stride * j); + ptr5 = ORC_PTR_OFFSET (s2, s2_stride * j); + + + for (i = 0; i < n; i++) { + /* 0: loadq */ + var40 = ptr4[i]; + /* 1: splitlw */ + var46.x2[0] = (var40.x2[0] >> 16) & 0xffff; + var47.x2[0] = var40.x2[0] & 0xffff; + var46.x2[1] = (var40.x2[1] >> 16) & 0xffff; + var47.x2[1] = var40.x2[1] & 0xffff; + /* 2: select1wb */ + var41.x2[0] = ((orc_uint16) var47.x2[0] >> 8) & 0xff; + var41.x2[1] = ((orc_uint16) var47.x2[1] >> 8) & 0xff; + /* 3: storew */ + ptr0[i] = var41; + /* 4: loadq */ + var42 = ptr5[i]; + /* 5: splitlw */ + var48.x2[0] = (var42.x2[0] >> 16) & 0xffff; + var49.x2[0] = var42.x2[0] & 0xffff; + var48.x2[1] = (var42.x2[1] >> 16) & 0xffff; + var49.x2[1] = var42.x2[1] & 0xffff; + /* 6: select1wb */ + var43.x2[0] = ((orc_uint16) var49.x2[0] >> 8) & 0xff; + var43.x2[1] = ((orc_uint16) var49.x2[1] >> 8) & 0xff; + /* 7: storew */ + ptr1[i] = var43; + /* 8: avgub */ + var50.x4[0] = + ((orc_uint8) var46.x4[0] + (orc_uint8) var48.x4[0] + 1) >> 1; + var50.x4[1] = + ((orc_uint8) var46.x4[1] + (orc_uint8) var48.x4[1] + 1) >> 1; + var50.x4[2] = + ((orc_uint8) var46.x4[2] + (orc_uint8) var48.x4[2] + 1) >> 1; + var50.x4[3] = + ((orc_uint8) var46.x4[3] + (orc_uint8) var48.x4[3] + 1) >> 1; + /* 9: splitwb */ + var51.x2[0] = (var50.x2[0] >> 8) & 0xff; + var52.x2[0] = var50.x2[0] & 0xff; + var51.x2[1] = (var50.x2[1] >> 8) & 0xff; + var52.x2[1] = var50.x2[1] & 0xff; + /* 10: splitwb */ + var53 = (var52.i >> 8) & 0xff; + var54 = var52.i & 0xff; + /* 11: avgub */ + var44 = ((orc_uint8) var53 + (orc_uint8) var54 + 1) >> 1; + /* 12: storeb */ + ptr2[i] = var44; + /* 13: splitwb */ + var55 = (var51.i >> 8) & 0xff; + var56 = var51.i & 0xff; + /* 14: avgub */ + var45 = ((orc_uint8) var55 + (orc_uint8) var56 + 1) >> 1; + /* 15: storeb */ + ptr3[i] = var45; + } + } + +} + +#else +static void +_backup_cogorc_convert_AYUV_I420 (OrcExecutor * ex) +{ + int i; + int j; + int n = ex->n; + int m = ex->params[ORC_VAR_A1]; + orc_union16 *ORC_RESTRICT ptr0; + orc_union16 *ORC_RESTRICT ptr1; + orc_int8 *ORC_RESTRICT ptr2; + orc_int8 *ORC_RESTRICT ptr3; + const orc_union64 *ORC_RESTRICT ptr4; + const orc_union64 *ORC_RESTRICT ptr5; + orc_union64 var40; + orc_union16 var41; + orc_union64 var42; + orc_union16 var43; + orc_int8 var44; + orc_int8 var45; + orc_union32 var46; + orc_union32 var47; + orc_union32 var48; + orc_union32 var49; + orc_union32 var50; + orc_union16 var51; + orc_union16 var52; + orc_int8 var53; + orc_int8 var54; + orc_int8 var55; + orc_int8 var56; + + for (j = 0; j < m; j++) { + ptr0 = ORC_PTR_OFFSET (ex->arrays[0], ex->params[0] * j); + ptr1 = ORC_PTR_OFFSET (ex->arrays[1], ex->params[1] * j); + ptr2 = ORC_PTR_OFFSET (ex->arrays[2], ex->params[2] * j); + ptr3 = ORC_PTR_OFFSET (ex->arrays[3], ex->params[3] * j); + ptr4 = ORC_PTR_OFFSET (ex->arrays[4], ex->params[4] * j); + ptr5 = ORC_PTR_OFFSET (ex->arrays[5], ex->params[5] * j); + + + for (i = 0; i < n; i++) { + /* 0: loadq */ + var40 = ptr4[i]; + /* 1: splitlw */ + var46.x2[0] = (var40.x2[0] >> 16) & 0xffff; + var47.x2[0] = var40.x2[0] & 0xffff; + var46.x2[1] = (var40.x2[1] >> 16) & 0xffff; + var47.x2[1] = var40.x2[1] & 0xffff; + /* 2: select1wb */ + var41.x2[0] = ((orc_uint16) var47.x2[0] >> 8) & 0xff; + var41.x2[1] = ((orc_uint16) var47.x2[1] >> 8) & 0xff; + /* 3: storew */ + ptr0[i] = var41; + /* 4: loadq */ + var42 = ptr5[i]; + /* 5: splitlw */ + var48.x2[0] = (var42.x2[0] >> 16) & 0xffff; + var49.x2[0] = var42.x2[0] & 0xffff; + var48.x2[1] = (var42.x2[1] >> 16) & 0xffff; + var49.x2[1] = var42.x2[1] & 0xffff; + /* 6: select1wb */ + var43.x2[0] = ((orc_uint16) var49.x2[0] >> 8) & 0xff; + var43.x2[1] = ((orc_uint16) var49.x2[1] >> 8) & 0xff; + /* 7: storew */ + ptr1[i] = var43; + /* 8: avgub */ + var50.x4[0] = + ((orc_uint8) var46.x4[0] + (orc_uint8) var48.x4[0] + 1) >> 1; + var50.x4[1] = + ((orc_uint8) var46.x4[1] + (orc_uint8) var48.x4[1] + 1) >> 1; + var50.x4[2] = + ((orc_uint8) var46.x4[2] + (orc_uint8) var48.x4[2] + 1) >> 1; + var50.x4[3] = + ((orc_uint8) var46.x4[3] + (orc_uint8) var48.x4[3] + 1) >> 1; + /* 9: splitwb */ + var51.x2[0] = (var50.x2[0] >> 8) & 0xff; + var52.x2[0] = var50.x2[0] & 0xff; + var51.x2[1] = (var50.x2[1] >> 8) & 0xff; + var52.x2[1] = var50.x2[1] & 0xff; + /* 10: splitwb */ + var53 = (var52.i >> 8) & 0xff; + var54 = var52.i & 0xff; + /* 11: avgub */ + var44 = ((orc_uint8) var53 + (orc_uint8) var54 + 1) >> 1; + /* 12: storeb */ + ptr2[i] = var44; + /* 13: splitwb */ + var55 = (var51.i >> 8) & 0xff; + var56 = var51.i & 0xff; + /* 14: avgub */ + var45 = ((orc_uint8) var55 + (orc_uint8) var56 + 1) >> 1; + /* 15: storeb */ + ptr3[i] = var45; + } + } + +} + +void +cogorc_convert_AYUV_I420 (guint8 * d1, int d1_stride, guint8 * d2, + int d2_stride, guint8 * d3, int d3_stride, guint8 * d4, int d4_stride, + const guint8 * s1, int s1_stride, const guint8 * s2, int s2_stride, int n, + int m) +{ + OrcExecutor _ex, *ex = &_ex; + static int p_inited = 0; + static OrcProgram *p = 0; + void (*func) (OrcExecutor *); + + if (!p_inited) { + orc_once_mutex_lock (); + if (!p_inited) { + OrcCompileResult result; + + p = orc_program_new (); + orc_program_set_2d (p); + orc_program_set_name (p, "cogorc_convert_AYUV_I420"); + orc_program_set_backup_function (p, _backup_cogorc_convert_AYUV_I420); + orc_program_add_destination (p, 2, "d1"); + orc_program_add_destination (p, 2, "d2"); + orc_program_add_destination (p, 1, "d3"); + orc_program_add_destination (p, 1, "d4"); + orc_program_add_source (p, 8, "s1"); + orc_program_add_source (p, 8, "s2"); + orc_program_add_constant (p, 0, 0x00000000, "c1"); + orc_program_add_constant (p, 0, 0x00000000, "c2"); + orc_program_add_constant (p, 0, 0x00000000, "c3"); + orc_program_add_constant (p, 0, 0x00000000, "c4"); + orc_program_add_constant (p, 0, 0x00000000, "c5"); + orc_program_add_constant (p, 0, 0x00000000, "c6"); + orc_program_add_constant (p, 0, 0x00000000, "c7"); + orc_program_add_constant (p, 0, 0x00000000, "c8"); + orc_program_add_temporary (p, 4, "t1"); + orc_program_add_temporary (p, 4, "t2"); + orc_program_add_temporary (p, 4, "t3"); + orc_program_add_temporary (p, 4, "t4"); + orc_program_add_temporary (p, 2, "t5"); + orc_program_add_temporary (p, 2, "t6"); + orc_program_add_temporary (p, 1, "t7"); + orc_program_add_temporary (p, 1, "t8"); + + orc_program_append_2 (p, "splitlw", 1, ORC_VAR_T2, ORC_VAR_T1, ORC_VAR_S1, + ORC_VAR_D1); + orc_program_append_2 (p, "select1wb", 1, ORC_VAR_D1, ORC_VAR_T1, + ORC_VAR_D1, ORC_VAR_D1); + orc_program_append_2 (p, "splitlw", 1, ORC_VAR_T3, ORC_VAR_T1, ORC_VAR_S2, + ORC_VAR_D1); + orc_program_append_2 (p, "select1wb", 1, ORC_VAR_D2, ORC_VAR_T1, + ORC_VAR_D1, ORC_VAR_D1); + orc_program_append_2 (p, "avgub", 2, ORC_VAR_T4, ORC_VAR_T2, ORC_VAR_T3, + ORC_VAR_D1); + orc_program_append_2 (p, "splitwb", 1, ORC_VAR_T6, ORC_VAR_T5, ORC_VAR_T4, + ORC_VAR_D1); + orc_program_append_2 (p, "splitwb", 0, ORC_VAR_T7, ORC_VAR_T8, ORC_VAR_T5, + ORC_VAR_D1); + orc_program_append_2 (p, "avgub", 0, ORC_VAR_D3, ORC_VAR_T7, ORC_VAR_T8, + ORC_VAR_D1); + orc_program_append_2 (p, "splitwb", 0, ORC_VAR_T7, ORC_VAR_T8, ORC_VAR_T6, + ORC_VAR_D1); + orc_program_append_2 (p, "avgub", 0, ORC_VAR_D4, ORC_VAR_T7, ORC_VAR_T8, + ORC_VAR_D1); + + result = orc_program_compile (p); + } + p_inited = TRUE; + orc_once_mutex_unlock (); + } + ex->program = p; + + ex->n = n; + ORC_EXECUTOR_M (ex) = m; + ex->arrays[ORC_VAR_D1] = d1; + ex->params[ORC_VAR_D1] = d1_stride; + ex->arrays[ORC_VAR_D2] = d2; + ex->params[ORC_VAR_D2] = d2_stride; + ex->arrays[ORC_VAR_D3] = d3; + ex->params[ORC_VAR_D3] = d3_stride; + ex->arrays[ORC_VAR_D4] = d4; + ex->params[ORC_VAR_D4] = d4_stride; + ex->arrays[ORC_VAR_S1] = (void *) s1; + ex->params[ORC_VAR_S1] = s1_stride; + ex->arrays[ORC_VAR_S2] = (void *) s2; + ex->params[ORC_VAR_S2] = s2_stride; + + func = p->code_exec; + func (ex); +} +#endif + + +/* cogorc_convert_AYUV_YUY2 */ +#ifdef DISABLE_ORC +void +cogorc_convert_AYUV_YUY2 (guint8 * d1, int d1_stride, const guint8 * s1, + int s1_stride, int n, int m) +{ + int i; + int j; + orc_union32 *ORC_RESTRICT ptr0; + const orc_union64 *ORC_RESTRICT ptr4; + orc_union64 var37; + orc_union32 var38; + orc_union32 var39; + orc_union32 var40; + orc_union16 var41; + orc_union16 var42; + orc_union16 var43; + orc_union16 var44; + + for (j = 0; j < m; j++) { + ptr0 = ORC_PTR_OFFSET (d1, d1_stride * j); + ptr4 = ORC_PTR_OFFSET (s1, s1_stride * j); + + + for (i = 0; i < n; i++) { + /* 0: loadq */ + var37 = ptr4[i]; + /* 1: splitlw */ + var39.x2[0] = (var37.x2[0] >> 16) & 0xffff; + var40.x2[0] = var37.x2[0] & 0xffff; + var39.x2[1] = (var37.x2[1] >> 16) & 0xffff; + var40.x2[1] = var37.x2[1] & 0xffff; + /* 2: splitlw */ + var41.i = (var39.i >> 16) & 0xffff; + var42.i = var39.i & 0xffff; + /* 3: avgub */ + var43.x2[0] = + ((orc_uint8) var41.x2[0] + (orc_uint8) var42.x2[0] + 1) >> 1; + var43.x2[1] = + ((orc_uint8) var41.x2[1] + (orc_uint8) var42.x2[1] + 1) >> 1; + /* 4: select1wb */ + var44.x2[0] = ((orc_uint16) var40.x2[0] >> 8) & 0xff; + var44.x2[1] = ((orc_uint16) var40.x2[1] >> 8) & 0xff; + /* 5: mergebw */ + var38.x2[0] = ((orc_uint8) var44.x2[0]) | ((orc_uint8) var43.x2[0] << 8); + var38.x2[1] = ((orc_uint8) var44.x2[1]) | ((orc_uint8) var43.x2[1] << 8); + /* 6: storel */ + ptr0[i] = var38; + } + } + +} + +#else +static void +_backup_cogorc_convert_AYUV_YUY2 (OrcExecutor * ex) +{ + int i; + int j; + int n = ex->n; + int m = ex->params[ORC_VAR_A1]; + orc_union32 *ORC_RESTRICT ptr0; + const orc_union64 *ORC_RESTRICT ptr4; + orc_union64 var37; + orc_union32 var38; + orc_union32 var39; + orc_union32 var40; + orc_union16 var41; + orc_union16 var42; + orc_union16 var43; + orc_union16 var44; + + for (j = 0; j < m; j++) { + ptr0 = ORC_PTR_OFFSET (ex->arrays[0], ex->params[0] * j); + ptr4 = ORC_PTR_OFFSET (ex->arrays[4], ex->params[4] * j); + + + for (i = 0; i < n; i++) { + /* 0: loadq */ + var37 = ptr4[i]; + /* 1: splitlw */ + var39.x2[0] = (var37.x2[0] >> 16) & 0xffff; + var40.x2[0] = var37.x2[0] & 0xffff; + var39.x2[1] = (var37.x2[1] >> 16) & 0xffff; + var40.x2[1] = var37.x2[1] & 0xffff; + /* 2: splitlw */ + var41.i = (var39.i >> 16) & 0xffff; + var42.i = var39.i & 0xffff; + /* 3: avgub */ + var43.x2[0] = + ((orc_uint8) var41.x2[0] + (orc_uint8) var42.x2[0] + 1) >> 1; + var43.x2[1] = + ((orc_uint8) var41.x2[1] + (orc_uint8) var42.x2[1] + 1) >> 1; + /* 4: select1wb */ + var44.x2[0] = ((orc_uint16) var40.x2[0] >> 8) & 0xff; + var44.x2[1] = ((orc_uint16) var40.x2[1] >> 8) & 0xff; + /* 5: mergebw */ + var38.x2[0] = ((orc_uint8) var44.x2[0]) | ((orc_uint8) var43.x2[0] << 8); + var38.x2[1] = ((orc_uint8) var44.x2[1]) | ((orc_uint8) var43.x2[1] << 8); + /* 6: storel */ + ptr0[i] = var38; + } + } + +} + +void +cogorc_convert_AYUV_YUY2 (guint8 * d1, int d1_stride, const guint8 * s1, + int s1_stride, int n, int m) +{ + OrcExecutor _ex, *ex = &_ex; + static int p_inited = 0; + static OrcProgram *p = 0; + void (*func) (OrcExecutor *); + + if (!p_inited) { + orc_once_mutex_lock (); + if (!p_inited) { + OrcCompileResult result; + + p = orc_program_new (); + orc_program_set_2d (p); + orc_program_set_name (p, "cogorc_convert_AYUV_YUY2"); + orc_program_set_backup_function (p, _backup_cogorc_convert_AYUV_YUY2); + orc_program_add_destination (p, 4, "d1"); + orc_program_add_source (p, 8, "s1"); + orc_program_add_constant (p, 0, 0x00000000, "c1"); + orc_program_add_constant (p, 0, 0x00000000, "c2"); + orc_program_add_constant (p, 0, 0x00000000, "c3"); + orc_program_add_constant (p, 0, 0x00000000, "c4"); + orc_program_add_constant (p, 0, 0x00000000, "c5"); + orc_program_add_constant (p, 0, 0x00000000, "c6"); + orc_program_add_constant (p, 0, 0x00000000, "c7"); + orc_program_add_constant (p, 0, 0x00000000, "c8"); + orc_program_add_temporary (p, 2, "t1"); + orc_program_add_temporary (p, 2, "t2"); + orc_program_add_temporary (p, 2, "t3"); + orc_program_add_temporary (p, 4, "t4"); + orc_program_add_temporary (p, 4, "t5"); + + orc_program_append_2 (p, "splitlw", 1, ORC_VAR_T5, ORC_VAR_T4, ORC_VAR_S1, + ORC_VAR_D1); + orc_program_append_2 (p, "splitlw", 0, ORC_VAR_T2, ORC_VAR_T3, ORC_VAR_T5, + ORC_VAR_D1); + orc_program_append_2 (p, "avgub", 1, ORC_VAR_T2, ORC_VAR_T2, ORC_VAR_T3, + ORC_VAR_D1); + orc_program_append_2 (p, "select1wb", 1, ORC_VAR_T1, ORC_VAR_T4, + ORC_VAR_D1, ORC_VAR_D1); + orc_program_append_2 (p, "mergebw", 1, ORC_VAR_D1, ORC_VAR_T1, ORC_VAR_T2, + ORC_VAR_D1); + + result = orc_program_compile (p); + } + p_inited = TRUE; + orc_once_mutex_unlock (); + } + ex->program = p; + + ex->n = n; + ORC_EXECUTOR_M (ex) = m; + ex->arrays[ORC_VAR_D1] = d1; + ex->params[ORC_VAR_D1] = d1_stride; + ex->arrays[ORC_VAR_S1] = (void *) s1; + ex->params[ORC_VAR_S1] = s1_stride; + + func = p->code_exec; + func (ex); +} +#endif + + +/* cogorc_convert_AYUV_UYVY */ +#ifdef DISABLE_ORC +void +cogorc_convert_AYUV_UYVY (guint8 * d1, int d1_stride, const guint8 * s1, + int s1_stride, int n, int m) +{ + int i; + int j; + orc_union32 *ORC_RESTRICT ptr0; + const orc_union64 *ORC_RESTRICT ptr4; + orc_union64 var37; + orc_union32 var38; + orc_union32 var39; + orc_union32 var40; + orc_union16 var41; + orc_union16 var42; + orc_union16 var43; + orc_union16 var44; + + for (j = 0; j < m; j++) { + ptr0 = ORC_PTR_OFFSET (d1, d1_stride * j); + ptr4 = ORC_PTR_OFFSET (s1, s1_stride * j); + + + for (i = 0; i < n; i++) { + /* 0: loadq */ + var37 = ptr4[i]; + /* 1: splitlw */ + var39.x2[0] = (var37.x2[0] >> 16) & 0xffff; + var40.x2[0] = var37.x2[0] & 0xffff; + var39.x2[1] = (var37.x2[1] >> 16) & 0xffff; + var40.x2[1] = var37.x2[1] & 0xffff; + /* 2: splitlw */ + var41.i = (var39.i >> 16) & 0xffff; + var42.i = var39.i & 0xffff; + /* 3: avgub */ + var43.x2[0] = + ((orc_uint8) var41.x2[0] + (orc_uint8) var42.x2[0] + 1) >> 1; + var43.x2[1] = + ((orc_uint8) var41.x2[1] + (orc_uint8) var42.x2[1] + 1) >> 1; + /* 4: select1wb */ + var44.x2[0] = ((orc_uint16) var40.x2[0] >> 8) & 0xff; + var44.x2[1] = ((orc_uint16) var40.x2[1] >> 8) & 0xff; + /* 5: mergebw */ + var38.x2[0] = ((orc_uint8) var43.x2[0]) | ((orc_uint8) var44.x2[0] << 8); + var38.x2[1] = ((orc_uint8) var43.x2[1]) | ((orc_uint8) var44.x2[1] << 8); + /* 6: storel */ + ptr0[i] = var38; + } + } + +} + +#else +static void +_backup_cogorc_convert_AYUV_UYVY (OrcExecutor * ex) +{ + int i; + int j; + int n = ex->n; + int m = ex->params[ORC_VAR_A1]; + orc_union32 *ORC_RESTRICT ptr0; + const orc_union64 *ORC_RESTRICT ptr4; + orc_union64 var37; + orc_union32 var38; + orc_union32 var39; + orc_union32 var40; + orc_union16 var41; + orc_union16 var42; + orc_union16 var43; + orc_union16 var44; + + for (j = 0; j < m; j++) { + ptr0 = ORC_PTR_OFFSET (ex->arrays[0], ex->params[0] * j); + ptr4 = ORC_PTR_OFFSET (ex->arrays[4], ex->params[4] * j); + + + for (i = 0; i < n; i++) { + /* 0: loadq */ + var37 = ptr4[i]; + /* 1: splitlw */ + var39.x2[0] = (var37.x2[0] >> 16) & 0xffff; + var40.x2[0] = var37.x2[0] & 0xffff; + var39.x2[1] = (var37.x2[1] >> 16) & 0xffff; + var40.x2[1] = var37.x2[1] & 0xffff; + /* 2: splitlw */ + var41.i = (var39.i >> 16) & 0xffff; + var42.i = var39.i & 0xffff; + /* 3: avgub */ + var43.x2[0] = + ((orc_uint8) var41.x2[0] + (orc_uint8) var42.x2[0] + 1) >> 1; + var43.x2[1] = + ((orc_uint8) var41.x2[1] + (orc_uint8) var42.x2[1] + 1) >> 1; + /* 4: select1wb */ + var44.x2[0] = ((orc_uint16) var40.x2[0] >> 8) & 0xff; + var44.x2[1] = ((orc_uint16) var40.x2[1] >> 8) & 0xff; + /* 5: mergebw */ + var38.x2[0] = ((orc_uint8) var43.x2[0]) | ((orc_uint8) var44.x2[0] << 8); + var38.x2[1] = ((orc_uint8) var43.x2[1]) | ((orc_uint8) var44.x2[1] << 8); + /* 6: storel */ + ptr0[i] = var38; + } + } + +} + +void +cogorc_convert_AYUV_UYVY (guint8 * d1, int d1_stride, const guint8 * s1, + int s1_stride, int n, int m) +{ + OrcExecutor _ex, *ex = &_ex; + static int p_inited = 0; + static OrcProgram *p = 0; + void (*func) (OrcExecutor *); + + if (!p_inited) { + orc_once_mutex_lock (); + if (!p_inited) { + OrcCompileResult result; + + p = orc_program_new (); + orc_program_set_2d (p); + orc_program_set_name (p, "cogorc_convert_AYUV_UYVY"); + orc_program_set_backup_function (p, _backup_cogorc_convert_AYUV_UYVY); + orc_program_add_destination (p, 4, "d1"); + orc_program_add_source (p, 8, "s1"); + orc_program_add_constant (p, 0, 0x00000000, "c1"); + orc_program_add_constant (p, 0, 0x00000000, "c2"); + orc_program_add_constant (p, 0, 0x00000000, "c3"); + orc_program_add_constant (p, 0, 0x00000000, "c4"); + orc_program_add_constant (p, 0, 0x00000000, "c5"); + orc_program_add_constant (p, 0, 0x00000000, "c6"); + orc_program_add_constant (p, 0, 0x00000000, "c7"); + orc_program_add_constant (p, 0, 0x00000000, "c8"); + orc_program_add_temporary (p, 2, "t1"); + orc_program_add_temporary (p, 2, "t2"); + orc_program_add_temporary (p, 2, "t3"); + orc_program_add_temporary (p, 4, "t4"); + orc_program_add_temporary (p, 4, "t5"); + + orc_program_append_2 (p, "splitlw", 1, ORC_VAR_T5, ORC_VAR_T4, ORC_VAR_S1, + ORC_VAR_D1); + orc_program_append_2 (p, "splitlw", 0, ORC_VAR_T2, ORC_VAR_T3, ORC_VAR_T5, + ORC_VAR_D1); + orc_program_append_2 (p, "avgub", 1, ORC_VAR_T2, ORC_VAR_T2, ORC_VAR_T3, + ORC_VAR_D1); + orc_program_append_2 (p, "select1wb", 1, ORC_VAR_T1, ORC_VAR_T4, + ORC_VAR_D1, ORC_VAR_D1); + orc_program_append_2 (p, "mergebw", 1, ORC_VAR_D1, ORC_VAR_T2, ORC_VAR_T1, + ORC_VAR_D1); + + result = orc_program_compile (p); + } + p_inited = TRUE; + orc_once_mutex_unlock (); + } + ex->program = p; + + ex->n = n; + ORC_EXECUTOR_M (ex) = m; + ex->arrays[ORC_VAR_D1] = d1; + ex->params[ORC_VAR_D1] = d1_stride; + ex->arrays[ORC_VAR_S1] = (void *) s1; + ex->params[ORC_VAR_S1] = s1_stride; + + func = p->code_exec; + func (ex); +} +#endif + + +/* cogorc_convert_AYUV_Y42B */ +#ifdef DISABLE_ORC +void +cogorc_convert_AYUV_Y42B (guint8 * d1, int d1_stride, guint8 * d2, + int d2_stride, guint8 * d3, int d3_stride, const guint8 * s1, int s1_stride, + int n, int m) +{ + int i; + int j; + orc_union16 *ORC_RESTRICT ptr0; + orc_int8 *ORC_RESTRICT ptr1; + orc_int8 *ORC_RESTRICT ptr2; + const orc_union64 *ORC_RESTRICT ptr4; + orc_union64 var36; + orc_int8 var37; + orc_int8 var38; + orc_union16 var39; + orc_union32 var40; + orc_union32 var41; + orc_union16 var42; + orc_union16 var43; + orc_union16 var44; + + for (j = 0; j < m; j++) { + ptr0 = ORC_PTR_OFFSET (d1, d1_stride * j); + ptr1 = ORC_PTR_OFFSET (d2, d2_stride * j); + ptr2 = ORC_PTR_OFFSET (d3, d3_stride * j); + ptr4 = ORC_PTR_OFFSET (s1, s1_stride * j); + + + for (i = 0; i < n; i++) { + /* 0: loadq */ + var36 = ptr4[i]; + /* 1: splitlw */ + var40.x2[0] = (var36.x2[0] >> 16) & 0xffff; + var41.x2[0] = var36.x2[0] & 0xffff; + var40.x2[1] = (var36.x2[1] >> 16) & 0xffff; + var41.x2[1] = var36.x2[1] & 0xffff; + /* 2: splitlw */ + var42.i = (var40.i >> 16) & 0xffff; + var43.i = var40.i & 0xffff; + /* 3: avgub */ + var44.x2[0] = + ((orc_uint8) var42.x2[0] + (orc_uint8) var43.x2[0] + 1) >> 1; + var44.x2[1] = + ((orc_uint8) var42.x2[1] + (orc_uint8) var43.x2[1] + 1) >> 1; + /* 4: splitwb */ + var37 = (var44.i >> 8) & 0xff; + var38 = var44.i & 0xff; + /* 5: storeb */ + ptr2[i] = var37; + /* 6: storeb */ + ptr1[i] = var38; + /* 7: select1wb */ + var39.x2[0] = ((orc_uint16) var41.x2[0] >> 8) & 0xff; + var39.x2[1] = ((orc_uint16) var41.x2[1] >> 8) & 0xff; + /* 8: storew */ + ptr0[i] = var39; + } + } + +} + +#else +static void +_backup_cogorc_convert_AYUV_Y42B (OrcExecutor * ex) +{ + int i; + int j; + int n = ex->n; + int m = ex->params[ORC_VAR_A1]; + orc_union16 *ORC_RESTRICT ptr0; + orc_int8 *ORC_RESTRICT ptr1; + orc_int8 *ORC_RESTRICT ptr2; + const orc_union64 *ORC_RESTRICT ptr4; + orc_union64 var36; + orc_int8 var37; + orc_int8 var38; + orc_union16 var39; + orc_union32 var40; + orc_union32 var41; + orc_union16 var42; + orc_union16 var43; + orc_union16 var44; + + for (j = 0; j < m; j++) { + ptr0 = ORC_PTR_OFFSET (ex->arrays[0], ex->params[0] * j); + ptr1 = ORC_PTR_OFFSET (ex->arrays[1], ex->params[1] * j); + ptr2 = ORC_PTR_OFFSET (ex->arrays[2], ex->params[2] * j); + ptr4 = ORC_PTR_OFFSET (ex->arrays[4], ex->params[4] * j); + + + for (i = 0; i < n; i++) { + /* 0: loadq */ + var36 = ptr4[i]; + /* 1: splitlw */ + var40.x2[0] = (var36.x2[0] >> 16) & 0xffff; + var41.x2[0] = var36.x2[0] & 0xffff; + var40.x2[1] = (var36.x2[1] >> 16) & 0xffff; + var41.x2[1] = var36.x2[1] & 0xffff; + /* 2: splitlw */ + var42.i = (var40.i >> 16) & 0xffff; + var43.i = var40.i & 0xffff; + /* 3: avgub */ + var44.x2[0] = + ((orc_uint8) var42.x2[0] + (orc_uint8) var43.x2[0] + 1) >> 1; + var44.x2[1] = + ((orc_uint8) var42.x2[1] + (orc_uint8) var43.x2[1] + 1) >> 1; + /* 4: splitwb */ + var37 = (var44.i >> 8) & 0xff; + var38 = var44.i & 0xff; + /* 5: storeb */ + ptr2[i] = var37; + /* 6: storeb */ + ptr1[i] = var38; + /* 7: select1wb */ + var39.x2[0] = ((orc_uint16) var41.x2[0] >> 8) & 0xff; + var39.x2[1] = ((orc_uint16) var41.x2[1] >> 8) & 0xff; + /* 8: storew */ + ptr0[i] = var39; + } + } + +} + +void +cogorc_convert_AYUV_Y42B (guint8 * d1, int d1_stride, guint8 * d2, + int d2_stride, guint8 * d3, int d3_stride, const guint8 * s1, int s1_stride, + int n, int m) +{ + OrcExecutor _ex, *ex = &_ex; + static int p_inited = 0; + static OrcProgram *p = 0; + void (*func) (OrcExecutor *); + + if (!p_inited) { + orc_once_mutex_lock (); + if (!p_inited) { + OrcCompileResult result; + + p = orc_program_new (); + orc_program_set_2d (p); + orc_program_set_name (p, "cogorc_convert_AYUV_Y42B"); + orc_program_set_backup_function (p, _backup_cogorc_convert_AYUV_Y42B); + orc_program_add_destination (p, 2, "d1"); + orc_program_add_destination (p, 1, "d2"); + orc_program_add_destination (p, 1, "d3"); + orc_program_add_source (p, 8, "s1"); + orc_program_add_constant (p, 0, 0x00000000, "c1"); + orc_program_add_constant (p, 0, 0x00000000, "c2"); + orc_program_add_constant (p, 0, 0x00000000, "c3"); + orc_program_add_constant (p, 0, 0x00000000, "c4"); + orc_program_add_constant (p, 0, 0x00000000, "c5"); + orc_program_add_constant (p, 0, 0x00000000, "c6"); + orc_program_add_constant (p, 0, 0x00000000, "c7"); + orc_program_add_constant (p, 0, 0x00000000, "c8"); + orc_program_add_temporary (p, 4, "t1"); + orc_program_add_temporary (p, 4, "t2"); + orc_program_add_temporary (p, 2, "t3"); + orc_program_add_temporary (p, 2, "t4"); + + orc_program_append_2 (p, "splitlw", 1, ORC_VAR_T2, ORC_VAR_T1, ORC_VAR_S1, + ORC_VAR_D1); + orc_program_append_2 (p, "splitlw", 0, ORC_VAR_T3, ORC_VAR_T4, ORC_VAR_T2, + ORC_VAR_D1); + orc_program_append_2 (p, "avgub", 1, ORC_VAR_T3, ORC_VAR_T3, ORC_VAR_T4, + ORC_VAR_D1); + orc_program_append_2 (p, "splitwb", 0, ORC_VAR_D3, ORC_VAR_D2, ORC_VAR_T3, + ORC_VAR_D1); + orc_program_append_2 (p, "select1wb", 1, ORC_VAR_D1, ORC_VAR_T1, + ORC_VAR_D1, ORC_VAR_D1); + + result = orc_program_compile (p); + } + p_inited = TRUE; + orc_once_mutex_unlock (); + } + ex->program = p; + + ex->n = n; + ORC_EXECUTOR_M (ex) = m; + ex->arrays[ORC_VAR_D1] = d1; + ex->params[ORC_VAR_D1] = d1_stride; + ex->arrays[ORC_VAR_D2] = d2; + ex->params[ORC_VAR_D2] = d2_stride; + ex->arrays[ORC_VAR_D3] = d3; + ex->params[ORC_VAR_D3] = d3_stride; + ex->arrays[ORC_VAR_S1] = (void *) s1; + ex->params[ORC_VAR_S1] = s1_stride; + + func = p->code_exec; + func (ex); +} +#endif + + +/* cogorc_convert_AYUV_Y444 */ +#ifdef DISABLE_ORC +void +cogorc_convert_AYUV_Y444 (guint8 * d1, int d1_stride, guint8 * d2, + int d2_stride, guint8 * d3, int d3_stride, const guint8 * s1, int s1_stride, + int n, int m) +{ + int i; + int j; + orc_int8 *ORC_RESTRICT ptr0; + orc_int8 *ORC_RESTRICT ptr1; + orc_int8 *ORC_RESTRICT ptr2; + const orc_union32 *ORC_RESTRICT ptr4; + orc_union32 var34; + orc_int8 var35; + orc_int8 var36; + orc_int8 var37; + orc_union16 var38; + orc_union16 var39; + + for (j = 0; j < m; j++) { + ptr0 = ORC_PTR_OFFSET (d1, d1_stride * j); + ptr1 = ORC_PTR_OFFSET (d2, d2_stride * j); + ptr2 = ORC_PTR_OFFSET (d3, d3_stride * j); + ptr4 = ORC_PTR_OFFSET (s1, s1_stride * j); + + + for (i = 0; i < n; i++) { + /* 0: loadl */ + var34 = ptr4[i]; + /* 1: splitlw */ + var38.i = (var34.i >> 16) & 0xffff; + var39.i = var34.i & 0xffff; + /* 2: splitwb */ + var35 = (var38.i >> 8) & 0xff; + var36 = var38.i & 0xff; + /* 3: storeb */ + ptr2[i] = var35; + /* 4: storeb */ + ptr1[i] = var36; + /* 5: select1wb */ + var37 = ((orc_uint16) var39.i >> 8) & 0xff; + /* 6: storeb */ + ptr0[i] = var37; + } + } + +} + +#else +static void +_backup_cogorc_convert_AYUV_Y444 (OrcExecutor * ex) +{ + int i; + int j; + int n = ex->n; + int m = ex->params[ORC_VAR_A1]; + orc_int8 *ORC_RESTRICT ptr0; + orc_int8 *ORC_RESTRICT ptr1; + orc_int8 *ORC_RESTRICT ptr2; + const orc_union32 *ORC_RESTRICT ptr4; + orc_union32 var34; + orc_int8 var35; + orc_int8 var36; + orc_int8 var37; + orc_union16 var38; + orc_union16 var39; + + for (j = 0; j < m; j++) { + ptr0 = ORC_PTR_OFFSET (ex->arrays[0], ex->params[0] * j); + ptr1 = ORC_PTR_OFFSET (ex->arrays[1], ex->params[1] * j); + ptr2 = ORC_PTR_OFFSET (ex->arrays[2], ex->params[2] * j); + ptr4 = ORC_PTR_OFFSET (ex->arrays[4], ex->params[4] * j); + + + for (i = 0; i < n; i++) { + /* 0: loadl */ + var34 = ptr4[i]; + /* 1: splitlw */ + var38.i = (var34.i >> 16) & 0xffff; + var39.i = var34.i & 0xffff; + /* 2: splitwb */ + var35 = (var38.i >> 8) & 0xff; + var36 = var38.i & 0xff; + /* 3: storeb */ + ptr2[i] = var35; + /* 4: storeb */ + ptr1[i] = var36; + /* 5: select1wb */ + var37 = ((orc_uint16) var39.i >> 8) & 0xff; + /* 6: storeb */ + ptr0[i] = var37; + } + } + +} + +void +cogorc_convert_AYUV_Y444 (guint8 * d1, int d1_stride, guint8 * d2, + int d2_stride, guint8 * d3, int d3_stride, const guint8 * s1, int s1_stride, + int n, int m) +{ + OrcExecutor _ex, *ex = &_ex; + static int p_inited = 0; + static OrcProgram *p = 0; + void (*func) (OrcExecutor *); + + if (!p_inited) { + orc_once_mutex_lock (); + if (!p_inited) { + OrcCompileResult result; + + p = orc_program_new (); + orc_program_set_2d (p); + orc_program_set_name (p, "cogorc_convert_AYUV_Y444"); + orc_program_set_backup_function (p, _backup_cogorc_convert_AYUV_Y444); + orc_program_add_destination (p, 1, "d1"); + orc_program_add_destination (p, 1, "d2"); + orc_program_add_destination (p, 1, "d3"); + orc_program_add_source (p, 4, "s1"); + orc_program_add_constant (p, 0, 0x00000000, "c1"); + orc_program_add_constant (p, 0, 0x00000000, "c2"); + orc_program_add_constant (p, 0, 0x00000000, "c3"); + orc_program_add_constant (p, 0, 0x00000000, "c4"); + orc_program_add_constant (p, 0, 0x00000000, "c5"); + orc_program_add_constant (p, 0, 0x00000000, "c6"); + orc_program_add_constant (p, 0, 0x00000000, "c7"); + orc_program_add_constant (p, 0, 0x00000000, "c8"); + orc_program_add_temporary (p, 2, "t1"); + orc_program_add_temporary (p, 2, "t2"); + + orc_program_append_2 (p, "splitlw", 0, ORC_VAR_T2, ORC_VAR_T1, ORC_VAR_S1, + ORC_VAR_D1); + orc_program_append_2 (p, "splitwb", 0, ORC_VAR_D3, ORC_VAR_D2, ORC_VAR_T2, + ORC_VAR_D1); + orc_program_append_2 (p, "select1wb", 0, ORC_VAR_D1, ORC_VAR_T1, + ORC_VAR_D1, ORC_VAR_D1); + + result = orc_program_compile (p); + } + p_inited = TRUE; + orc_once_mutex_unlock (); + } + ex->program = p; + + ex->n = n; + ORC_EXECUTOR_M (ex) = m; + ex->arrays[ORC_VAR_D1] = d1; + ex->params[ORC_VAR_D1] = d1_stride; + ex->arrays[ORC_VAR_D2] = d2; + ex->params[ORC_VAR_D2] = d2_stride; + ex->arrays[ORC_VAR_D3] = d3; + ex->params[ORC_VAR_D3] = d3_stride; + ex->arrays[ORC_VAR_S1] = (void *) s1; + ex->params[ORC_VAR_S1] = s1_stride; + + func = p->code_exec; + func (ex); +} +#endif + + +/* cogorc_convert_Y42B_YUY2 */ +#ifdef DISABLE_ORC +void +cogorc_convert_Y42B_YUY2 (guint8 * d1, int d1_stride, const guint8 * s1, + int s1_stride, const guint8 * s2, int s2_stride, const guint8 * s3, + int s3_stride, int n, int m) +{ + int i; + int j; + orc_union32 *ORC_RESTRICT ptr0; + const orc_union16 *ORC_RESTRICT ptr4; + const orc_int8 *ORC_RESTRICT ptr5; + const orc_int8 *ORC_RESTRICT ptr6; + orc_int8 var33; + orc_int8 var34; + orc_union16 var35; + orc_union32 var36; + orc_union16 var37; + + for (j = 0; j < m; j++) { + ptr0 = ORC_PTR_OFFSET (d1, d1_stride * j); + ptr4 = ORC_PTR_OFFSET (s1, s1_stride * j); + ptr5 = ORC_PTR_OFFSET (s2, s2_stride * j); + ptr6 = ORC_PTR_OFFSET (s3, s3_stride * j); + + + for (i = 0; i < n; i++) { + /* 0: loadb */ + var33 = ptr5[i]; + /* 1: loadb */ + var34 = ptr6[i]; + /* 2: mergebw */ + var37.i = ((orc_uint8) var33) | ((orc_uint8) var34 << 8); + /* 3: loadw */ + var35 = ptr4[i]; + /* 4: mergebw */ + var36.x2[0] = ((orc_uint8) var35.x2[0]) | ((orc_uint8) var37.x2[0] << 8); + var36.x2[1] = ((orc_uint8) var35.x2[1]) | ((orc_uint8) var37.x2[1] << 8); + /* 5: storel */ + ptr0[i] = var36; + } + } + +} + +#else +static void +_backup_cogorc_convert_Y42B_YUY2 (OrcExecutor * ex) +{ + int i; + int j; + int n = ex->n; + int m = ex->params[ORC_VAR_A1]; + orc_union32 *ORC_RESTRICT ptr0; + const orc_union16 *ORC_RESTRICT ptr4; + const orc_int8 *ORC_RESTRICT ptr5; + const orc_int8 *ORC_RESTRICT ptr6; + orc_int8 var33; + orc_int8 var34; + orc_union16 var35; + orc_union32 var36; + orc_union16 var37; + + for (j = 0; j < m; j++) { + ptr0 = ORC_PTR_OFFSET (ex->arrays[0], ex->params[0] * j); + ptr4 = ORC_PTR_OFFSET (ex->arrays[4], ex->params[4] * j); + ptr5 = ORC_PTR_OFFSET (ex->arrays[5], ex->params[5] * j); + ptr6 = ORC_PTR_OFFSET (ex->arrays[6], ex->params[6] * j); + + + for (i = 0; i < n; i++) { + /* 0: loadb */ + var33 = ptr5[i]; + /* 1: loadb */ + var34 = ptr6[i]; + /* 2: mergebw */ + var37.i = ((orc_uint8) var33) | ((orc_uint8) var34 << 8); + /* 3: loadw */ + var35 = ptr4[i]; + /* 4: mergebw */ + var36.x2[0] = ((orc_uint8) var35.x2[0]) | ((orc_uint8) var37.x2[0] << 8); + var36.x2[1] = ((orc_uint8) var35.x2[1]) | ((orc_uint8) var37.x2[1] << 8); + /* 5: storel */ + ptr0[i] = var36; + } + } + +} + +void +cogorc_convert_Y42B_YUY2 (guint8 * d1, int d1_stride, const guint8 * s1, + int s1_stride, const guint8 * s2, int s2_stride, const guint8 * s3, + int s3_stride, int n, int m) +{ + OrcExecutor _ex, *ex = &_ex; + static int p_inited = 0; + static OrcProgram *p = 0; + void (*func) (OrcExecutor *); + + if (!p_inited) { + orc_once_mutex_lock (); + if (!p_inited) { + OrcCompileResult result; + + p = orc_program_new (); + orc_program_set_2d (p); + orc_program_set_name (p, "cogorc_convert_Y42B_YUY2"); + orc_program_set_backup_function (p, _backup_cogorc_convert_Y42B_YUY2); + orc_program_add_destination (p, 4, "d1"); + orc_program_add_source (p, 2, "s1"); + orc_program_add_source (p, 1, "s2"); + orc_program_add_source (p, 1, "s3"); + orc_program_add_constant (p, 0, 0x00000000, "c1"); + orc_program_add_constant (p, 0, 0x00000000, "c2"); + orc_program_add_constant (p, 0, 0x00000000, "c3"); + orc_program_add_constant (p, 0, 0x00000000, "c4"); + orc_program_add_constant (p, 0, 0x00000000, "c5"); + orc_program_add_constant (p, 0, 0x00000000, "c6"); + orc_program_add_constant (p, 0, 0x00000000, "c7"); + orc_program_add_constant (p, 0, 0x00000000, "c8"); + orc_program_add_temporary (p, 2, "t1"); + + orc_program_append_2 (p, "mergebw", 0, ORC_VAR_T1, ORC_VAR_S2, ORC_VAR_S3, + ORC_VAR_D1); + orc_program_append_2 (p, "mergebw", 1, ORC_VAR_D1, ORC_VAR_S1, ORC_VAR_T1, + ORC_VAR_D1); + + result = orc_program_compile (p); + } + p_inited = TRUE; + orc_once_mutex_unlock (); + } + ex->program = p; + + ex->n = n; + ORC_EXECUTOR_M (ex) = m; + ex->arrays[ORC_VAR_D1] = d1; + ex->params[ORC_VAR_D1] = d1_stride; + ex->arrays[ORC_VAR_S1] = (void *) s1; + ex->params[ORC_VAR_S1] = s1_stride; + ex->arrays[ORC_VAR_S2] = (void *) s2; + ex->params[ORC_VAR_S2] = s2_stride; + ex->arrays[ORC_VAR_S3] = (void *) s3; + ex->params[ORC_VAR_S3] = s3_stride; + + func = p->code_exec; + func (ex); +} +#endif + + +/* cogorc_convert_Y42B_UYVY */ +#ifdef DISABLE_ORC +void +cogorc_convert_Y42B_UYVY (guint8 * d1, int d1_stride, const guint8 * s1, + int s1_stride, const guint8 * s2, int s2_stride, const guint8 * s3, + int s3_stride, int n, int m) +{ + int i; + int j; + orc_union32 *ORC_RESTRICT ptr0; + const orc_union16 *ORC_RESTRICT ptr4; + const orc_int8 *ORC_RESTRICT ptr5; + const orc_int8 *ORC_RESTRICT ptr6; + orc_int8 var33; + orc_int8 var34; + orc_union16 var35; + orc_union32 var36; + orc_union16 var37; + + for (j = 0; j < m; j++) { + ptr0 = ORC_PTR_OFFSET (d1, d1_stride * j); + ptr4 = ORC_PTR_OFFSET (s1, s1_stride * j); + ptr5 = ORC_PTR_OFFSET (s2, s2_stride * j); + ptr6 = ORC_PTR_OFFSET (s3, s3_stride * j); + + + for (i = 0; i < n; i++) { + /* 0: loadb */ + var33 = ptr5[i]; + /* 1: loadb */ + var34 = ptr6[i]; + /* 2: mergebw */ + var37.i = ((orc_uint8) var33) | ((orc_uint8) var34 << 8); + /* 3: loadw */ + var35 = ptr4[i]; + /* 4: mergebw */ + var36.x2[0] = ((orc_uint8) var37.x2[0]) | ((orc_uint8) var35.x2[0] << 8); + var36.x2[1] = ((orc_uint8) var37.x2[1]) | ((orc_uint8) var35.x2[1] << 8); + /* 5: storel */ + ptr0[i] = var36; + } + } + +} + +#else +static void +_backup_cogorc_convert_Y42B_UYVY (OrcExecutor * ex) +{ + int i; + int j; + int n = ex->n; + int m = ex->params[ORC_VAR_A1]; + orc_union32 *ORC_RESTRICT ptr0; + const orc_union16 *ORC_RESTRICT ptr4; + const orc_int8 *ORC_RESTRICT ptr5; + const orc_int8 *ORC_RESTRICT ptr6; + orc_int8 var33; + orc_int8 var34; + orc_union16 var35; + orc_union32 var36; + orc_union16 var37; + + for (j = 0; j < m; j++) { + ptr0 = ORC_PTR_OFFSET (ex->arrays[0], ex->params[0] * j); + ptr4 = ORC_PTR_OFFSET (ex->arrays[4], ex->params[4] * j); + ptr5 = ORC_PTR_OFFSET (ex->arrays[5], ex->params[5] * j); + ptr6 = ORC_PTR_OFFSET (ex->arrays[6], ex->params[6] * j); + + + for (i = 0; i < n; i++) { + /* 0: loadb */ + var33 = ptr5[i]; + /* 1: loadb */ + var34 = ptr6[i]; + /* 2: mergebw */ + var37.i = ((orc_uint8) var33) | ((orc_uint8) var34 << 8); + /* 3: loadw */ + var35 = ptr4[i]; + /* 4: mergebw */ + var36.x2[0] = ((orc_uint8) var37.x2[0]) | ((orc_uint8) var35.x2[0] << 8); + var36.x2[1] = ((orc_uint8) var37.x2[1]) | ((orc_uint8) var35.x2[1] << 8); + /* 5: storel */ + ptr0[i] = var36; + } + } + +} + +void +cogorc_convert_Y42B_UYVY (guint8 * d1, int d1_stride, const guint8 * s1, + int s1_stride, const guint8 * s2, int s2_stride, const guint8 * s3, + int s3_stride, int n, int m) +{ + OrcExecutor _ex, *ex = &_ex; + static int p_inited = 0; + static OrcProgram *p = 0; + void (*func) (OrcExecutor *); + + if (!p_inited) { + orc_once_mutex_lock (); + if (!p_inited) { + OrcCompileResult result; + + p = orc_program_new (); + orc_program_set_2d (p); + orc_program_set_name (p, "cogorc_convert_Y42B_UYVY"); + orc_program_set_backup_function (p, _backup_cogorc_convert_Y42B_UYVY); + orc_program_add_destination (p, 4, "d1"); + orc_program_add_source (p, 2, "s1"); + orc_program_add_source (p, 1, "s2"); + orc_program_add_source (p, 1, "s3"); + orc_program_add_constant (p, 0, 0x00000000, "c1"); + orc_program_add_constant (p, 0, 0x00000000, "c2"); + orc_program_add_constant (p, 0, 0x00000000, "c3"); + orc_program_add_constant (p, 0, 0x00000000, "c4"); + orc_program_add_constant (p, 0, 0x00000000, "c5"); + orc_program_add_constant (p, 0, 0x00000000, "c6"); + orc_program_add_constant (p, 0, 0x00000000, "c7"); + orc_program_add_constant (p, 0, 0x00000000, "c8"); + orc_program_add_temporary (p, 2, "t1"); + + orc_program_append_2 (p, "mergebw", 0, ORC_VAR_T1, ORC_VAR_S2, ORC_VAR_S3, + ORC_VAR_D1); + orc_program_append_2 (p, "mergebw", 1, ORC_VAR_D1, ORC_VAR_T1, ORC_VAR_S1, + ORC_VAR_D1); + + result = orc_program_compile (p); + } + p_inited = TRUE; + orc_once_mutex_unlock (); + } + ex->program = p; + + ex->n = n; + ORC_EXECUTOR_M (ex) = m; + ex->arrays[ORC_VAR_D1] = d1; + ex->params[ORC_VAR_D1] = d1_stride; + ex->arrays[ORC_VAR_S1] = (void *) s1; + ex->params[ORC_VAR_S1] = s1_stride; + ex->arrays[ORC_VAR_S2] = (void *) s2; + ex->params[ORC_VAR_S2] = s2_stride; + ex->arrays[ORC_VAR_S3] = (void *) s3; + ex->params[ORC_VAR_S3] = s3_stride; + + func = p->code_exec; + func (ex); +} +#endif + + +/* cogorc_convert_Y42B_AYUV */ +#ifdef DISABLE_ORC +void +cogorc_convert_Y42B_AYUV (guint8 * d1, int d1_stride, const guint8 * s1, + int s1_stride, const guint8 * s2, int s2_stride, const guint8 * s3, + int s3_stride, int n, int m) +{ + int i; + int j; + orc_union64 *ORC_RESTRICT ptr0; + const orc_union16 *ORC_RESTRICT ptr4; + const orc_int8 *ORC_RESTRICT ptr5; + const orc_int8 *ORC_RESTRICT ptr6; + orc_int8 var36; + orc_int8 var37; + orc_union16 var38; + orc_union16 var39; + orc_union64 var40; + orc_union16 var41; + orc_union32 var42; + orc_union32 var43; + + for (j = 0; j < m; j++) { + ptr0 = ORC_PTR_OFFSET (d1, d1_stride * j); + ptr4 = ORC_PTR_OFFSET (s1, s1_stride * j); + ptr5 = ORC_PTR_OFFSET (s2, s2_stride * j); + ptr6 = ORC_PTR_OFFSET (s3, s3_stride * j); + + /* 3: loadpb */ + var38.x2[0] = 0x000000ff; /* 255 or 1.25987e-321f */ + var38.x2[1] = 0x000000ff; /* 255 or 1.25987e-321f */ + + for (i = 0; i < n; i++) { + /* 0: loadb */ + var36 = ptr5[i]; + /* 1: loadb */ + var37 = ptr6[i]; + /* 2: mergebw */ + var41.i = ((orc_uint8) var36) | ((orc_uint8) var37 << 8); + /* 4: loadw */ + var39 = ptr4[i]; + /* 5: mergebw */ + var42.x2[0] = ((orc_uint8) var38.x2[0]) | ((orc_uint8) var39.x2[0] << 8); + var42.x2[1] = ((orc_uint8) var38.x2[1]) | ((orc_uint8) var39.x2[1] << 8); + /* 6: mergewl */ + var43.i = ((orc_uint16) var41.i) | ((orc_uint16) var41.i << 16); + /* 7: mergewl */ + var40.x2[0] = + ((orc_uint16) var42.x2[0]) | ((orc_uint16) var43.x2[0] << 16); + var40.x2[1] = + ((orc_uint16) var42.x2[1]) | ((orc_uint16) var43.x2[1] << 16); + /* 8: storeq */ + ptr0[i] = var40; + } + } + +} + +#else +static void +_backup_cogorc_convert_Y42B_AYUV (OrcExecutor * ex) +{ + int i; + int j; + int n = ex->n; + int m = ex->params[ORC_VAR_A1]; + orc_union64 *ORC_RESTRICT ptr0; + const orc_union16 *ORC_RESTRICT ptr4; + const orc_int8 *ORC_RESTRICT ptr5; + const orc_int8 *ORC_RESTRICT ptr6; + orc_int8 var36; + orc_int8 var37; + orc_union16 var38; + orc_union16 var39; + orc_union64 var40; + orc_union16 var41; + orc_union32 var42; + orc_union32 var43; + + for (j = 0; j < m; j++) { + ptr0 = ORC_PTR_OFFSET (ex->arrays[0], ex->params[0] * j); + ptr4 = ORC_PTR_OFFSET (ex->arrays[4], ex->params[4] * j); + ptr5 = ORC_PTR_OFFSET (ex->arrays[5], ex->params[5] * j); + ptr6 = ORC_PTR_OFFSET (ex->arrays[6], ex->params[6] * j); + + /* 3: loadpb */ + var38.x2[0] = 0x000000ff; /* 255 or 1.25987e-321f */ + var38.x2[1] = 0x000000ff; /* 255 or 1.25987e-321f */ + + for (i = 0; i < n; i++) { + /* 0: loadb */ + var36 = ptr5[i]; + /* 1: loadb */ + var37 = ptr6[i]; + /* 2: mergebw */ + var41.i = ((orc_uint8) var36) | ((orc_uint8) var37 << 8); + /* 4: loadw */ + var39 = ptr4[i]; + /* 5: mergebw */ + var42.x2[0] = ((orc_uint8) var38.x2[0]) | ((orc_uint8) var39.x2[0] << 8); + var42.x2[1] = ((orc_uint8) var38.x2[1]) | ((orc_uint8) var39.x2[1] << 8); + /* 6: mergewl */ + var43.i = ((orc_uint16) var41.i) | ((orc_uint16) var41.i << 16); + /* 7: mergewl */ + var40.x2[0] = + ((orc_uint16) var42.x2[0]) | ((orc_uint16) var43.x2[0] << 16); + var40.x2[1] = + ((orc_uint16) var42.x2[1]) | ((orc_uint16) var43.x2[1] << 16); + /* 8: storeq */ + ptr0[i] = var40; + } + } + +} + +void +cogorc_convert_Y42B_AYUV (guint8 * d1, int d1_stride, const guint8 * s1, + int s1_stride, const guint8 * s2, int s2_stride, const guint8 * s3, + int s3_stride, int n, int m) +{ + OrcExecutor _ex, *ex = &_ex; + static int p_inited = 0; + static OrcProgram *p = 0; + void (*func) (OrcExecutor *); + + if (!p_inited) { + orc_once_mutex_lock (); + if (!p_inited) { + OrcCompileResult result; + + p = orc_program_new (); + orc_program_set_2d (p); + orc_program_set_name (p, "cogorc_convert_Y42B_AYUV"); + orc_program_set_backup_function (p, _backup_cogorc_convert_Y42B_AYUV); + orc_program_add_destination (p, 8, "d1"); + orc_program_add_source (p, 2, "s1"); + orc_program_add_source (p, 1, "s2"); + orc_program_add_source (p, 1, "s3"); + orc_program_add_constant (p, 1, 0x000000ff, "c1"); + orc_program_add_constant (p, 0, 0x00000000, "c2"); + orc_program_add_constant (p, 0, 0x00000000, "c3"); + orc_program_add_constant (p, 0, 0x00000000, "c4"); + orc_program_add_constant (p, 0, 0x00000000, "c5"); + orc_program_add_constant (p, 0, 0x00000000, "c6"); + orc_program_add_constant (p, 0, 0x00000000, "c7"); + orc_program_add_constant (p, 0, 0x00000000, "c8"); + orc_program_add_temporary (p, 2, "t1"); + orc_program_add_temporary (p, 2, "t2"); + orc_program_add_temporary (p, 4, "t3"); + orc_program_add_temporary (p, 4, "t4"); + + orc_program_append_2 (p, "mergebw", 0, ORC_VAR_T1, ORC_VAR_S2, ORC_VAR_S3, + ORC_VAR_D1); + orc_program_append_2 (p, "mergebw", 1, ORC_VAR_T4, ORC_VAR_C1, ORC_VAR_S1, + ORC_VAR_D1); + orc_program_append_2 (p, "mergewl", 0, ORC_VAR_T3, ORC_VAR_T1, ORC_VAR_T1, + ORC_VAR_D1); + orc_program_append_2 (p, "mergewl", 1, ORC_VAR_D1, ORC_VAR_T4, ORC_VAR_T3, + ORC_VAR_D1); + + result = orc_program_compile (p); + } + p_inited = TRUE; + orc_once_mutex_unlock (); + } + ex->program = p; + + ex->n = n; + ORC_EXECUTOR_M (ex) = m; + ex->arrays[ORC_VAR_D1] = d1; + ex->params[ORC_VAR_D1] = d1_stride; + ex->arrays[ORC_VAR_S1] = (void *) s1; + ex->params[ORC_VAR_S1] = s1_stride; + ex->arrays[ORC_VAR_S2] = (void *) s2; + ex->params[ORC_VAR_S2] = s2_stride; + ex->arrays[ORC_VAR_S3] = (void *) s3; + ex->params[ORC_VAR_S3] = s3_stride; + + func = p->code_exec; + func (ex); +} +#endif + + +/* cogorc_convert_Y444_YUY2 */ +#ifdef DISABLE_ORC +void +cogorc_convert_Y444_YUY2 (guint8 * d1, int d1_stride, const guint8 * s1, + int s1_stride, const guint8 * s2, int s2_stride, const guint8 * s3, + int s3_stride, int n, int m) +{ + int i; + int j; + orc_union32 *ORC_RESTRICT ptr0; + const orc_union16 *ORC_RESTRICT ptr4; + const orc_union16 *ORC_RESTRICT ptr5; + const orc_union16 *ORC_RESTRICT ptr6; + orc_union16 var36; + orc_union16 var37; + orc_union16 var38; + orc_union32 var39; + orc_union32 var40; + orc_union16 var41; + orc_union16 var42; + orc_union16 var43; + + for (j = 0; j < m; j++) { + ptr0 = ORC_PTR_OFFSET (d1, d1_stride * j); + ptr4 = ORC_PTR_OFFSET (s1, s1_stride * j); + ptr5 = ORC_PTR_OFFSET (s2, s2_stride * j); + ptr6 = ORC_PTR_OFFSET (s3, s3_stride * j); + + + for (i = 0; i < n; i++) { + /* 0: loadw */ + var36 = ptr5[i]; + /* 1: loadw */ + var37 = ptr6[i]; + /* 2: mergebw */ + var40.x2[0] = ((orc_uint8) var36.x2[0]) | ((orc_uint8) var37.x2[0] << 8); + var40.x2[1] = ((orc_uint8) var36.x2[1]) | ((orc_uint8) var37.x2[1] << 8); + /* 3: splitlw */ + var41.i = (var40.i >> 16) & 0xffff; + var42.i = var40.i & 0xffff; + /* 4: avgub */ + var43.x2[0] = + ((orc_uint8) var41.x2[0] + (orc_uint8) var42.x2[0] + 1) >> 1; + var43.x2[1] = + ((orc_uint8) var41.x2[1] + (orc_uint8) var42.x2[1] + 1) >> 1; + /* 5: loadw */ + var38 = ptr4[i]; + /* 6: mergebw */ + var39.x2[0] = ((orc_uint8) var38.x2[0]) | ((orc_uint8) var43.x2[0] << 8); + var39.x2[1] = ((orc_uint8) var38.x2[1]) | ((orc_uint8) var43.x2[1] << 8); + /* 7: storel */ + ptr0[i] = var39; + } + } + +} + +#else +static void +_backup_cogorc_convert_Y444_YUY2 (OrcExecutor * ex) +{ + int i; + int j; + int n = ex->n; + int m = ex->params[ORC_VAR_A1]; + orc_union32 *ORC_RESTRICT ptr0; + const orc_union16 *ORC_RESTRICT ptr4; + const orc_union16 *ORC_RESTRICT ptr5; + const orc_union16 *ORC_RESTRICT ptr6; + orc_union16 var36; + orc_union16 var37; + orc_union16 var38; + orc_union32 var39; + orc_union32 var40; + orc_union16 var41; + orc_union16 var42; + orc_union16 var43; + + for (j = 0; j < m; j++) { + ptr0 = ORC_PTR_OFFSET (ex->arrays[0], ex->params[0] * j); + ptr4 = ORC_PTR_OFFSET (ex->arrays[4], ex->params[4] * j); + ptr5 = ORC_PTR_OFFSET (ex->arrays[5], ex->params[5] * j); + ptr6 = ORC_PTR_OFFSET (ex->arrays[6], ex->params[6] * j); + + + for (i = 0; i < n; i++) { + /* 0: loadw */ + var36 = ptr5[i]; + /* 1: loadw */ + var37 = ptr6[i]; + /* 2: mergebw */ + var40.x2[0] = ((orc_uint8) var36.x2[0]) | ((orc_uint8) var37.x2[0] << 8); + var40.x2[1] = ((orc_uint8) var36.x2[1]) | ((orc_uint8) var37.x2[1] << 8); + /* 3: splitlw */ + var41.i = (var40.i >> 16) & 0xffff; + var42.i = var40.i & 0xffff; + /* 4: avgub */ + var43.x2[0] = + ((orc_uint8) var41.x2[0] + (orc_uint8) var42.x2[0] + 1) >> 1; + var43.x2[1] = + ((orc_uint8) var41.x2[1] + (orc_uint8) var42.x2[1] + 1) >> 1; + /* 5: loadw */ + var38 = ptr4[i]; + /* 6: mergebw */ + var39.x2[0] = ((orc_uint8) var38.x2[0]) | ((orc_uint8) var43.x2[0] << 8); + var39.x2[1] = ((orc_uint8) var38.x2[1]) | ((orc_uint8) var43.x2[1] << 8); + /* 7: storel */ + ptr0[i] = var39; + } + } + +} + +void +cogorc_convert_Y444_YUY2 (guint8 * d1, int d1_stride, const guint8 * s1, + int s1_stride, const guint8 * s2, int s2_stride, const guint8 * s3, + int s3_stride, int n, int m) +{ + OrcExecutor _ex, *ex = &_ex; + static int p_inited = 0; + static OrcProgram *p = 0; + void (*func) (OrcExecutor *); + + if (!p_inited) { + orc_once_mutex_lock (); + if (!p_inited) { + OrcCompileResult result; + + p = orc_program_new (); + orc_program_set_2d (p); + orc_program_set_name (p, "cogorc_convert_Y444_YUY2"); + orc_program_set_backup_function (p, _backup_cogorc_convert_Y444_YUY2); + orc_program_add_destination (p, 4, "d1"); + orc_program_add_source (p, 2, "s1"); + orc_program_add_source (p, 2, "s2"); + orc_program_add_source (p, 2, "s3"); + orc_program_add_constant (p, 0, 0x00000000, "c1"); + orc_program_add_constant (p, 0, 0x00000000, "c2"); + orc_program_add_constant (p, 0, 0x00000000, "c3"); + orc_program_add_constant (p, 0, 0x00000000, "c4"); + orc_program_add_constant (p, 0, 0x00000000, "c5"); + orc_program_add_constant (p, 0, 0x00000000, "c6"); + orc_program_add_constant (p, 0, 0x00000000, "c7"); + orc_program_add_constant (p, 0, 0x00000000, "c8"); + orc_program_add_temporary (p, 2, "t1"); + orc_program_add_temporary (p, 4, "t2"); + orc_program_add_temporary (p, 2, "t3"); + orc_program_add_temporary (p, 2, "t4"); + + orc_program_append_2 (p, "mergebw", 1, ORC_VAR_T2, ORC_VAR_S2, ORC_VAR_S3, + ORC_VAR_D1); + orc_program_append_2 (p, "splitlw", 0, ORC_VAR_T3, ORC_VAR_T4, ORC_VAR_T2, + ORC_VAR_D1); + orc_program_append_2 (p, "avgub", 1, ORC_VAR_T1, ORC_VAR_T3, ORC_VAR_T4, + ORC_VAR_D1); + orc_program_append_2 (p, "mergebw", 1, ORC_VAR_D1, ORC_VAR_S1, ORC_VAR_T1, + ORC_VAR_D1); + + result = orc_program_compile (p); + } + p_inited = TRUE; + orc_once_mutex_unlock (); + } + ex->program = p; + + ex->n = n; + ORC_EXECUTOR_M (ex) = m; + ex->arrays[ORC_VAR_D1] = d1; + ex->params[ORC_VAR_D1] = d1_stride; + ex->arrays[ORC_VAR_S1] = (void *) s1; + ex->params[ORC_VAR_S1] = s1_stride; + ex->arrays[ORC_VAR_S2] = (void *) s2; + ex->params[ORC_VAR_S2] = s2_stride; + ex->arrays[ORC_VAR_S3] = (void *) s3; + ex->params[ORC_VAR_S3] = s3_stride; + + func = p->code_exec; + func (ex); +} +#endif + + +/* cogorc_convert_Y444_UYVY */ +#ifdef DISABLE_ORC +void +cogorc_convert_Y444_UYVY (guint8 * d1, int d1_stride, const guint8 * s1, + int s1_stride, const guint8 * s2, int s2_stride, const guint8 * s3, + int s3_stride, int n, int m) +{ + int i; + int j; + orc_union32 *ORC_RESTRICT ptr0; + const orc_union16 *ORC_RESTRICT ptr4; + const orc_union16 *ORC_RESTRICT ptr5; + const orc_union16 *ORC_RESTRICT ptr6; + orc_union16 var36; + orc_union16 var37; + orc_union16 var38; + orc_union32 var39; + orc_union32 var40; + orc_union16 var41; + orc_union16 var42; + orc_union16 var43; + + for (j = 0; j < m; j++) { + ptr0 = ORC_PTR_OFFSET (d1, d1_stride * j); + ptr4 = ORC_PTR_OFFSET (s1, s1_stride * j); + ptr5 = ORC_PTR_OFFSET (s2, s2_stride * j); + ptr6 = ORC_PTR_OFFSET (s3, s3_stride * j); + + + for (i = 0; i < n; i++) { + /* 0: loadw */ + var36 = ptr5[i]; + /* 1: loadw */ + var37 = ptr6[i]; + /* 2: mergebw */ + var40.x2[0] = ((orc_uint8) var36.x2[0]) | ((orc_uint8) var37.x2[0] << 8); + var40.x2[1] = ((orc_uint8) var36.x2[1]) | ((orc_uint8) var37.x2[1] << 8); + /* 3: splitlw */ + var41.i = (var40.i >> 16) & 0xffff; + var42.i = var40.i & 0xffff; + /* 4: avgub */ + var43.x2[0] = + ((orc_uint8) var41.x2[0] + (orc_uint8) var42.x2[0] + 1) >> 1; + var43.x2[1] = + ((orc_uint8) var41.x2[1] + (orc_uint8) var42.x2[1] + 1) >> 1; + /* 5: loadw */ + var38 = ptr4[i]; + /* 6: mergebw */ + var39.x2[0] = ((orc_uint8) var43.x2[0]) | ((orc_uint8) var38.x2[0] << 8); + var39.x2[1] = ((orc_uint8) var43.x2[1]) | ((orc_uint8) var38.x2[1] << 8); + /* 7: storel */ + ptr0[i] = var39; + } + } + +} + +#else +static void +_backup_cogorc_convert_Y444_UYVY (OrcExecutor * ex) +{ + int i; + int j; + int n = ex->n; + int m = ex->params[ORC_VAR_A1]; + orc_union32 *ORC_RESTRICT ptr0; + const orc_union16 *ORC_RESTRICT ptr4; + const orc_union16 *ORC_RESTRICT ptr5; + const orc_union16 *ORC_RESTRICT ptr6; + orc_union16 var36; + orc_union16 var37; + orc_union16 var38; + orc_union32 var39; + orc_union32 var40; + orc_union16 var41; + orc_union16 var42; + orc_union16 var43; + + for (j = 0; j < m; j++) { + ptr0 = ORC_PTR_OFFSET (ex->arrays[0], ex->params[0] * j); + ptr4 = ORC_PTR_OFFSET (ex->arrays[4], ex->params[4] * j); + ptr5 = ORC_PTR_OFFSET (ex->arrays[5], ex->params[5] * j); + ptr6 = ORC_PTR_OFFSET (ex->arrays[6], ex->params[6] * j); + + + for (i = 0; i < n; i++) { + /* 0: loadw */ + var36 = ptr5[i]; + /* 1: loadw */ + var37 = ptr6[i]; + /* 2: mergebw */ + var40.x2[0] = ((orc_uint8) var36.x2[0]) | ((orc_uint8) var37.x2[0] << 8); + var40.x2[1] = ((orc_uint8) var36.x2[1]) | ((orc_uint8) var37.x2[1] << 8); + /* 3: splitlw */ + var41.i = (var40.i >> 16) & 0xffff; + var42.i = var40.i & 0xffff; + /* 4: avgub */ + var43.x2[0] = + ((orc_uint8) var41.x2[0] + (orc_uint8) var42.x2[0] + 1) >> 1; + var43.x2[1] = + ((orc_uint8) var41.x2[1] + (orc_uint8) var42.x2[1] + 1) >> 1; + /* 5: loadw */ + var38 = ptr4[i]; + /* 6: mergebw */ + var39.x2[0] = ((orc_uint8) var43.x2[0]) | ((orc_uint8) var38.x2[0] << 8); + var39.x2[1] = ((orc_uint8) var43.x2[1]) | ((orc_uint8) var38.x2[1] << 8); + /* 7: storel */ + ptr0[i] = var39; + } + } + +} + +void +cogorc_convert_Y444_UYVY (guint8 * d1, int d1_stride, const guint8 * s1, + int s1_stride, const guint8 * s2, int s2_stride, const guint8 * s3, + int s3_stride, int n, int m) +{ + OrcExecutor _ex, *ex = &_ex; + static int p_inited = 0; + static OrcProgram *p = 0; + void (*func) (OrcExecutor *); + + if (!p_inited) { + orc_once_mutex_lock (); + if (!p_inited) { + OrcCompileResult result; + + p = orc_program_new (); + orc_program_set_2d (p); + orc_program_set_name (p, "cogorc_convert_Y444_UYVY"); + orc_program_set_backup_function (p, _backup_cogorc_convert_Y444_UYVY); + orc_program_add_destination (p, 4, "d1"); + orc_program_add_source (p, 2, "s1"); + orc_program_add_source (p, 2, "s2"); + orc_program_add_source (p, 2, "s3"); + orc_program_add_constant (p, 0, 0x00000000, "c1"); + orc_program_add_constant (p, 0, 0x00000000, "c2"); + orc_program_add_constant (p, 0, 0x00000000, "c3"); + orc_program_add_constant (p, 0, 0x00000000, "c4"); + orc_program_add_constant (p, 0, 0x00000000, "c5"); + orc_program_add_constant (p, 0, 0x00000000, "c6"); + orc_program_add_constant (p, 0, 0x00000000, "c7"); + orc_program_add_constant (p, 0, 0x00000000, "c8"); + orc_program_add_temporary (p, 2, "t1"); + orc_program_add_temporary (p, 4, "t2"); + orc_program_add_temporary (p, 2, "t3"); + orc_program_add_temporary (p, 2, "t4"); + + orc_program_append_2 (p, "mergebw", 1, ORC_VAR_T2, ORC_VAR_S2, ORC_VAR_S3, + ORC_VAR_D1); + orc_program_append_2 (p, "splitlw", 0, ORC_VAR_T3, ORC_VAR_T4, ORC_VAR_T2, + ORC_VAR_D1); + orc_program_append_2 (p, "avgub", 1, ORC_VAR_T1, ORC_VAR_T3, ORC_VAR_T4, + ORC_VAR_D1); + orc_program_append_2 (p, "mergebw", 1, ORC_VAR_D1, ORC_VAR_T1, ORC_VAR_S1, + ORC_VAR_D1); + + result = orc_program_compile (p); + } + p_inited = TRUE; + orc_once_mutex_unlock (); + } + ex->program = p; + + ex->n = n; + ORC_EXECUTOR_M (ex) = m; + ex->arrays[ORC_VAR_D1] = d1; + ex->params[ORC_VAR_D1] = d1_stride; + ex->arrays[ORC_VAR_S1] = (void *) s1; + ex->params[ORC_VAR_S1] = s1_stride; + ex->arrays[ORC_VAR_S2] = (void *) s2; + ex->params[ORC_VAR_S2] = s2_stride; + ex->arrays[ORC_VAR_S3] = (void *) s3; + ex->params[ORC_VAR_S3] = s3_stride; + + func = p->code_exec; + func (ex); +} +#endif + + +/* cogorc_convert_Y444_AYUV */ +#ifdef DISABLE_ORC +void +cogorc_convert_Y444_AYUV (guint8 * d1, int d1_stride, const guint8 * s1, + int s1_stride, const guint8 * s2, int s2_stride, const guint8 * s3, + int s3_stride, int n, int m) +{ + int i; + int j; + orc_union32 *ORC_RESTRICT ptr0; + const orc_int8 *ORC_RESTRICT ptr4; + const orc_int8 *ORC_RESTRICT ptr5; + const orc_int8 *ORC_RESTRICT ptr6; + orc_int8 var34; + orc_int8 var35; + orc_int8 var36; + orc_int8 var37; + orc_union32 var38; + orc_union16 var39; + orc_union16 var40; + + for (j = 0; j < m; j++) { + ptr0 = ORC_PTR_OFFSET (d1, d1_stride * j); + ptr4 = ORC_PTR_OFFSET (s1, s1_stride * j); + ptr5 = ORC_PTR_OFFSET (s2, s2_stride * j); + ptr6 = ORC_PTR_OFFSET (s3, s3_stride * j); + + /* 3: loadpb */ + var36 = 0x000000ff; /* 255 or 1.25987e-321f */ + + for (i = 0; i < n; i++) { + /* 0: loadb */ + var34 = ptr5[i]; + /* 1: loadb */ + var35 = ptr6[i]; + /* 2: mergebw */ + var39.i = ((orc_uint8) var34) | ((orc_uint8) var35 << 8); + /* 4: loadb */ + var37 = ptr4[i]; + /* 5: mergebw */ + var40.i = ((orc_uint8) var36) | ((orc_uint8) var37 << 8); + /* 6: mergewl */ + var38.i = ((orc_uint16) var40.i) | ((orc_uint16) var39.i << 16); + /* 7: storel */ + ptr0[i] = var38; + } + } + +} + +#else +static void +_backup_cogorc_convert_Y444_AYUV (OrcExecutor * ex) +{ + int i; + int j; + int n = ex->n; + int m = ex->params[ORC_VAR_A1]; + orc_union32 *ORC_RESTRICT ptr0; + const orc_int8 *ORC_RESTRICT ptr4; + const orc_int8 *ORC_RESTRICT ptr5; + const orc_int8 *ORC_RESTRICT ptr6; + orc_int8 var34; + orc_int8 var35; + orc_int8 var36; + orc_int8 var37; + orc_union32 var38; + orc_union16 var39; + orc_union16 var40; + + for (j = 0; j < m; j++) { + ptr0 = ORC_PTR_OFFSET (ex->arrays[0], ex->params[0] * j); + ptr4 = ORC_PTR_OFFSET (ex->arrays[4], ex->params[4] * j); + ptr5 = ORC_PTR_OFFSET (ex->arrays[5], ex->params[5] * j); + ptr6 = ORC_PTR_OFFSET (ex->arrays[6], ex->params[6] * j); + + /* 3: loadpb */ + var36 = 0x000000ff; /* 255 or 1.25987e-321f */ + + for (i = 0; i < n; i++) { + /* 0: loadb */ + var34 = ptr5[i]; + /* 1: loadb */ + var35 = ptr6[i]; + /* 2: mergebw */ + var39.i = ((orc_uint8) var34) | ((orc_uint8) var35 << 8); + /* 4: loadb */ + var37 = ptr4[i]; + /* 5: mergebw */ + var40.i = ((orc_uint8) var36) | ((orc_uint8) var37 << 8); + /* 6: mergewl */ + var38.i = ((orc_uint16) var40.i) | ((orc_uint16) var39.i << 16); + /* 7: storel */ + ptr0[i] = var38; + } + } + +} + +void +cogorc_convert_Y444_AYUV (guint8 * d1, int d1_stride, const guint8 * s1, + int s1_stride, const guint8 * s2, int s2_stride, const guint8 * s3, + int s3_stride, int n, int m) +{ + OrcExecutor _ex, *ex = &_ex; + static int p_inited = 0; + static OrcProgram *p = 0; + void (*func) (OrcExecutor *); + + if (!p_inited) { + orc_once_mutex_lock (); + if (!p_inited) { + OrcCompileResult result; + + p = orc_program_new (); + orc_program_set_2d (p); + orc_program_set_name (p, "cogorc_convert_Y444_AYUV"); + orc_program_set_backup_function (p, _backup_cogorc_convert_Y444_AYUV); + orc_program_add_destination (p, 4, "d1"); + orc_program_add_source (p, 1, "s1"); + orc_program_add_source (p, 1, "s2"); + orc_program_add_source (p, 1, "s3"); + orc_program_add_constant (p, 1, 0x000000ff, "c1"); + orc_program_add_constant (p, 0, 0x00000000, "c2"); + orc_program_add_constant (p, 0, 0x00000000, "c3"); + orc_program_add_constant (p, 0, 0x00000000, "c4"); + orc_program_add_constant (p, 0, 0x00000000, "c5"); + orc_program_add_constant (p, 0, 0x00000000, "c6"); + orc_program_add_constant (p, 0, 0x00000000, "c7"); + orc_program_add_constant (p, 0, 0x00000000, "c8"); + orc_program_add_temporary (p, 2, "t1"); + orc_program_add_temporary (p, 2, "t2"); + + orc_program_append_2 (p, "mergebw", 0, ORC_VAR_T1, ORC_VAR_S2, ORC_VAR_S3, + ORC_VAR_D1); + orc_program_append_2 (p, "mergebw", 0, ORC_VAR_T2, ORC_VAR_C1, ORC_VAR_S1, + ORC_VAR_D1); + orc_program_append_2 (p, "mergewl", 0, ORC_VAR_D1, ORC_VAR_T2, ORC_VAR_T1, + ORC_VAR_D1); + + result = orc_program_compile (p); + } + p_inited = TRUE; + orc_once_mutex_unlock (); + } + ex->program = p; + + ex->n = n; + ORC_EXECUTOR_M (ex) = m; + ex->arrays[ORC_VAR_D1] = d1; + ex->params[ORC_VAR_D1] = d1_stride; + ex->arrays[ORC_VAR_S1] = (void *) s1; + ex->params[ORC_VAR_S1] = s1_stride; + ex->arrays[ORC_VAR_S2] = (void *) s2; + ex->params[ORC_VAR_S2] = s2_stride; + ex->arrays[ORC_VAR_S3] = (void *) s3; + ex->params[ORC_VAR_S3] = s3_stride; + + func = p->code_exec; + func (ex); +} +#endif + + +/* cogorc_convert_AYUV_ARGB */ +#ifdef DISABLE_ORC +void +cogorc_convert_AYUV_ARGB (guint8 * d1, int d1_stride, const guint8 * s1, + int s1_stride, int n, int m) +{ + int i; + int j; + orc_union32 *ORC_RESTRICT ptr0; + const orc_union32 *ORC_RESTRICT ptr4; + orc_union32 var48; + orc_union32 var49; + orc_union16 var50; + orc_union16 var51; + orc_union16 var52; + orc_union16 var53; + orc_union16 var54; + orc_union32 var55; + orc_union32 var56; + orc_union32 var57; + orc_union16 var58; + orc_union16 var59; + orc_int8 var60; + orc_int8 var61; + orc_int8 var62; + orc_int8 var63; + orc_union16 var64; + orc_union16 var65; + orc_union16 var66; + orc_union16 var67; + orc_union16 var68; + orc_union16 var69; + orc_union16 var70; + orc_union16 var71; + orc_union16 var72; + orc_union16 var73; + orc_union16 var74; + orc_union16 var75; + orc_union16 var76; + orc_union16 var77; + orc_union16 var78; + orc_union16 var79; + orc_union16 var80; + orc_union16 var81; + orc_union16 var82; + orc_union16 var83; + orc_union16 var84; + orc_union16 var85; + orc_union16 var86; + orc_int8 var87; + orc_int8 var88; + orc_int8 var89; + orc_union16 var90; + orc_union16 var91; + orc_union32 var92; + + for (j = 0; j < m; j++) { + ptr0 = ORC_PTR_OFFSET (d1, d1_stride * j); + ptr4 = ORC_PTR_OFFSET (s1, s1_stride * j); + + /* 1: loadpb */ + var49.x4[0] = 0x00000080; /* 128 or 6.32404e-322f */ + var49.x4[1] = 0x00000080; /* 128 or 6.32404e-322f */ + var49.x4[2] = 0x00000080; /* 128 or 6.32404e-322f */ + var49.x4[3] = 0x00000080; /* 128 or 6.32404e-322f */ + /* 9: loadpw */ + var50.i = 0x0000002a; /* 42 or 2.07508e-322f */ + /* 14: loadpw */ + var51.i = 0x00000067; /* 103 or 5.08888e-322f */ + /* 21: loadpw */ + var52.i = 0x00000004; /* 4 or 1.97626e-323f */ + /* 25: loadpw */ + var53.i = 0x00000064; /* 100 or 4.94066e-322f */ + /* 29: loadpw */ + var54.i = 0x00000068; /* 104 or 5.13828e-322f */ + /* 40: loadpb */ + var55.x4[0] = 0x00000080; /* 128 or 6.32404e-322f */ + var55.x4[1] = 0x00000080; /* 128 or 6.32404e-322f */ + var55.x4[2] = 0x00000080; /* 128 or 6.32404e-322f */ + var55.x4[3] = 0x00000080; /* 128 or 6.32404e-322f */ + + for (i = 0; i < n; i++) { + /* 0: loadl */ + var48 = ptr4[i]; + /* 2: subb */ + var57.x4[0] = var48.x4[0] - var49.x4[0]; + var57.x4[1] = var48.x4[1] - var49.x4[1]; + var57.x4[2] = var48.x4[2] - var49.x4[2]; + var57.x4[3] = var48.x4[3] - var49.x4[3]; + /* 3: splitlw */ + var58.i = (var57.i >> 16) & 0xffff; + var59.i = var57.i & 0xffff; + /* 4: splitwb */ + var60 = (var59.i >> 8) & 0xff; + var61 = var59.i & 0xff; + /* 5: splitwb */ + var62 = (var58.i >> 8) & 0xff; + var63 = var58.i & 0xff; + /* 6: convsbw */ + var64.i = var60; + /* 7: convsbw */ + var65.i = var63; + /* 8: convsbw */ + var66.i = var62; + /* 10: mullw */ + var67.i = (var64.i * var50.i) & 0xffff; + /* 11: shrsw */ + var68.i = var67.i >> 8; + /* 12: addssw */ + var69.i = ORC_CLAMP_SW (var64.i + var68.i); + /* 13: addssw */ + var70.i = ORC_CLAMP_SW (var69.i + var66.i); + /* 15: mullw */ + var71.i = (var66.i * var51.i) & 0xffff; + /* 16: shrsw */ + var72.i = var71.i >> 8; + /* 17: subssw */ + var73.i = ORC_CLAMP_SW (var70.i - var72.i); + /* 18: addssw */ + var74.i = ORC_CLAMP_SW (var73.i + var66.i); + /* 19: addssw */ + var75.i = ORC_CLAMP_SW (var69.i + var65.i); + /* 20: addssw */ + var76.i = ORC_CLAMP_SW (var75.i + var65.i); + /* 22: mullw */ + var77.i = (var65.i * var52.i) & 0xffff; + /* 23: shrsw */ + var78.i = var77.i >> 8; + /* 24: addssw */ + var79.i = ORC_CLAMP_SW (var76.i + var78.i); + /* 26: mullw */ + var80.i = (var65.i * var53.i) & 0xffff; + /* 27: shrsw */ + var81.i = var80.i >> 8; + /* 28: subssw */ + var82.i = ORC_CLAMP_SW (var69.i - var81.i); + /* 30: mullw */ + var83.i = (var66.i * var54.i) & 0xffff; + /* 31: shrsw */ + var84.i = var83.i >> 8; + /* 32: subssw */ + var85.i = ORC_CLAMP_SW (var82.i - var84.i); + /* 33: subssw */ + var86.i = ORC_CLAMP_SW (var85.i - var84.i); + /* 34: convssswb */ + var87 = ORC_CLAMP_SB (var74.i); + /* 35: convssswb */ + var88 = ORC_CLAMP_SB (var86.i); + /* 36: convssswb */ + var89 = ORC_CLAMP_SB (var79.i); + /* 37: mergebw */ + var90.i = ((orc_uint8) var61) | ((orc_uint8) var87 << 8); + /* 38: mergebw */ + var91.i = ((orc_uint8) var88) | ((orc_uint8) var89 << 8); + /* 39: mergewl */ + var92.i = ((orc_uint16) var90.i) | ((orc_uint16) var91.i << 16); + /* 41: addb */ + var56.x4[0] = var92.x4[0] + var55.x4[0]; + var56.x4[1] = var92.x4[1] + var55.x4[1]; + var56.x4[2] = var92.x4[2] + var55.x4[2]; + var56.x4[3] = var92.x4[3] + var55.x4[3]; + /* 42: storel */ + ptr0[i] = var56; + } + } + +} + +#else +static void +_backup_cogorc_convert_AYUV_ARGB (OrcExecutor * ex) +{ + int i; + int j; + int n = ex->n; + int m = ex->params[ORC_VAR_A1]; + orc_union32 *ORC_RESTRICT ptr0; + const orc_union32 *ORC_RESTRICT ptr4; + orc_union32 var48; + orc_union32 var49; + orc_union16 var50; + orc_union16 var51; + orc_union16 var52; + orc_union16 var53; + orc_union16 var54; + orc_union32 var55; + orc_union32 var56; + orc_union32 var57; + orc_union16 var58; + orc_union16 var59; + orc_int8 var60; + orc_int8 var61; + orc_int8 var62; + orc_int8 var63; + orc_union16 var64; + orc_union16 var65; + orc_union16 var66; + orc_union16 var67; + orc_union16 var68; + orc_union16 var69; + orc_union16 var70; + orc_union16 var71; + orc_union16 var72; + orc_union16 var73; + orc_union16 var74; + orc_union16 var75; + orc_union16 var76; + orc_union16 var77; + orc_union16 var78; + orc_union16 var79; + orc_union16 var80; + orc_union16 var81; + orc_union16 var82; + orc_union16 var83; + orc_union16 var84; + orc_union16 var85; + orc_union16 var86; + orc_int8 var87; + orc_int8 var88; + orc_int8 var89; + orc_union16 var90; + orc_union16 var91; + orc_union32 var92; + + for (j = 0; j < m; j++) { + ptr0 = ORC_PTR_OFFSET (ex->arrays[0], ex->params[0] * j); + ptr4 = ORC_PTR_OFFSET (ex->arrays[4], ex->params[4] * j); + + /* 1: loadpb */ + var49.x4[0] = 0x00000080; /* 128 or 6.32404e-322f */ + var49.x4[1] = 0x00000080; /* 128 or 6.32404e-322f */ + var49.x4[2] = 0x00000080; /* 128 or 6.32404e-322f */ + var49.x4[3] = 0x00000080; /* 128 or 6.32404e-322f */ + /* 9: loadpw */ + var50.i = 0x0000002a; /* 42 or 2.07508e-322f */ + /* 14: loadpw */ + var51.i = 0x00000067; /* 103 or 5.08888e-322f */ + /* 21: loadpw */ + var52.i = 0x00000004; /* 4 or 1.97626e-323f */ + /* 25: loadpw */ + var53.i = 0x00000064; /* 100 or 4.94066e-322f */ + /* 29: loadpw */ + var54.i = 0x00000068; /* 104 or 5.13828e-322f */ + /* 40: loadpb */ + var55.x4[0] = 0x00000080; /* 128 or 6.32404e-322f */ + var55.x4[1] = 0x00000080; /* 128 or 6.32404e-322f */ + var55.x4[2] = 0x00000080; /* 128 or 6.32404e-322f */ + var55.x4[3] = 0x00000080; /* 128 or 6.32404e-322f */ + + for (i = 0; i < n; i++) { + /* 0: loadl */ + var48 = ptr4[i]; + /* 2: subb */ + var57.x4[0] = var48.x4[0] - var49.x4[0]; + var57.x4[1] = var48.x4[1] - var49.x4[1]; + var57.x4[2] = var48.x4[2] - var49.x4[2]; + var57.x4[3] = var48.x4[3] - var49.x4[3]; + /* 3: splitlw */ + var58.i = (var57.i >> 16) & 0xffff; + var59.i = var57.i & 0xffff; + /* 4: splitwb */ + var60 = (var59.i >> 8) & 0xff; + var61 = var59.i & 0xff; + /* 5: splitwb */ + var62 = (var58.i >> 8) & 0xff; + var63 = var58.i & 0xff; + /* 6: convsbw */ + var64.i = var60; + /* 7: convsbw */ + var65.i = var63; + /* 8: convsbw */ + var66.i = var62; + /* 10: mullw */ + var67.i = (var64.i * var50.i) & 0xffff; + /* 11: shrsw */ + var68.i = var67.i >> 8; + /* 12: addssw */ + var69.i = ORC_CLAMP_SW (var64.i + var68.i); + /* 13: addssw */ + var70.i = ORC_CLAMP_SW (var69.i + var66.i); + /* 15: mullw */ + var71.i = (var66.i * var51.i) & 0xffff; + /* 16: shrsw */ + var72.i = var71.i >> 8; + /* 17: subssw */ + var73.i = ORC_CLAMP_SW (var70.i - var72.i); + /* 18: addssw */ + var74.i = ORC_CLAMP_SW (var73.i + var66.i); + /* 19: addssw */ + var75.i = ORC_CLAMP_SW (var69.i + var65.i); + /* 20: addssw */ + var76.i = ORC_CLAMP_SW (var75.i + var65.i); + /* 22: mullw */ + var77.i = (var65.i * var52.i) & 0xffff; + /* 23: shrsw */ + var78.i = var77.i >> 8; + /* 24: addssw */ + var79.i = ORC_CLAMP_SW (var76.i + var78.i); + /* 26: mullw */ + var80.i = (var65.i * var53.i) & 0xffff; + /* 27: shrsw */ + var81.i = var80.i >> 8; + /* 28: subssw */ + var82.i = ORC_CLAMP_SW (var69.i - var81.i); + /* 30: mullw */ + var83.i = (var66.i * var54.i) & 0xffff; + /* 31: shrsw */ + var84.i = var83.i >> 8; + /* 32: subssw */ + var85.i = ORC_CLAMP_SW (var82.i - var84.i); + /* 33: subssw */ + var86.i = ORC_CLAMP_SW (var85.i - var84.i); + /* 34: convssswb */ + var87 = ORC_CLAMP_SB (var74.i); + /* 35: convssswb */ + var88 = ORC_CLAMP_SB (var86.i); + /* 36: convssswb */ + var89 = ORC_CLAMP_SB (var79.i); + /* 37: mergebw */ + var90.i = ((orc_uint8) var61) | ((orc_uint8) var87 << 8); + /* 38: mergebw */ + var91.i = ((orc_uint8) var88) | ((orc_uint8) var89 << 8); + /* 39: mergewl */ + var92.i = ((orc_uint16) var90.i) | ((orc_uint16) var91.i << 16); + /* 41: addb */ + var56.x4[0] = var92.x4[0] + var55.x4[0]; + var56.x4[1] = var92.x4[1] + var55.x4[1]; + var56.x4[2] = var92.x4[2] + var55.x4[2]; + var56.x4[3] = var92.x4[3] + var55.x4[3]; + /* 42: storel */ + ptr0[i] = var56; + } + } + +} + +void +cogorc_convert_AYUV_ARGB (guint8 * d1, int d1_stride, const guint8 * s1, + int s1_stride, int n, int m) +{ + OrcExecutor _ex, *ex = &_ex; + static int p_inited = 0; + static OrcProgram *p = 0; + void (*func) (OrcExecutor *); + + if (!p_inited) { + orc_once_mutex_lock (); + if (!p_inited) { + OrcCompileResult result; + + p = orc_program_new (); + orc_program_set_2d (p); + orc_program_set_name (p, "cogorc_convert_AYUV_ARGB"); + orc_program_set_backup_function (p, _backup_cogorc_convert_AYUV_ARGB); + orc_program_add_destination (p, 4, "d1"); + orc_program_add_source (p, 4, "s1"); + orc_program_add_constant (p, 1, 0x00000008, "c1"); + orc_program_add_constant (p, 4, 0x00000080, "c2"); + orc_program_add_constant (p, 4, 0x0000002a, "c3"); + orc_program_add_constant (p, 4, 0x00000067, "c4"); + orc_program_add_constant (p, 4, 0x00000004, "c5"); + orc_program_add_constant (p, 4, 0x00000064, "c6"); + orc_program_add_constant (p, 4, 0x00000068, "c7"); + orc_program_add_constant (p, 0, 0x00000080, "c8"); + orc_program_add_temporary (p, 2, "t1"); + orc_program_add_temporary (p, 2, "t2"); + orc_program_add_temporary (p, 1, "t3"); + orc_program_add_temporary (p, 1, "t4"); + orc_program_add_temporary (p, 1, "t5"); + orc_program_add_temporary (p, 1, "t6"); + orc_program_add_temporary (p, 2, "t7"); + orc_program_add_temporary (p, 2, "t8"); + orc_program_add_temporary (p, 2, "t9"); + orc_program_add_temporary (p, 2, "t10"); + orc_program_add_temporary (p, 2, "t11"); + orc_program_add_temporary (p, 2, "t12"); + orc_program_add_temporary (p, 1, "t13"); + orc_program_add_temporary (p, 1, "t14"); + orc_program_add_temporary (p, 1, "t15"); + orc_program_add_temporary (p, 4, "t16"); + + orc_program_append_2 (p, "subb", 2, 47, ORC_VAR_S1, ORC_VAR_C2, + ORC_VAR_D1); + orc_program_append_2 (p, "splitlw", 0, ORC_VAR_T1, ORC_VAR_T2, 47, + ORC_VAR_D1); + orc_program_append_2 (p, "splitwb", 0, ORC_VAR_T4, ORC_VAR_T3, ORC_VAR_T2, + ORC_VAR_D1); + orc_program_append_2 (p, "splitwb", 0, ORC_VAR_T6, ORC_VAR_T5, ORC_VAR_T1, + ORC_VAR_D1); + orc_program_append_2 (p, "convsbw", 0, ORC_VAR_T7, ORC_VAR_T4, ORC_VAR_D1, + ORC_VAR_D1); + orc_program_append_2 (p, "convsbw", 0, ORC_VAR_T8, ORC_VAR_T5, ORC_VAR_D1, + ORC_VAR_D1); + orc_program_append_2 (p, "convsbw", 0, ORC_VAR_T9, ORC_VAR_T6, ORC_VAR_D1, + ORC_VAR_D1); + orc_program_append_2 (p, "mullw", 0, ORC_VAR_T1, ORC_VAR_T7, ORC_VAR_C3, + ORC_VAR_D1); + orc_program_append_2 (p, "shrsw", 0, ORC_VAR_T1, ORC_VAR_T1, ORC_VAR_C1, + ORC_VAR_D1); + orc_program_append_2 (p, "addssw", 0, ORC_VAR_T7, ORC_VAR_T7, ORC_VAR_T1, + ORC_VAR_D1); + orc_program_append_2 (p, "addssw", 0, ORC_VAR_T10, ORC_VAR_T7, ORC_VAR_T9, + ORC_VAR_D1); + orc_program_append_2 (p, "mullw", 0, ORC_VAR_T1, ORC_VAR_T9, ORC_VAR_C4, + ORC_VAR_D1); + orc_program_append_2 (p, "shrsw", 0, ORC_VAR_T1, ORC_VAR_T1, ORC_VAR_C1, + ORC_VAR_D1); + orc_program_append_2 (p, "subssw", 0, ORC_VAR_T10, ORC_VAR_T10, + ORC_VAR_T1, ORC_VAR_D1); + orc_program_append_2 (p, "addssw", 0, ORC_VAR_T10, ORC_VAR_T10, + ORC_VAR_T9, ORC_VAR_D1); + orc_program_append_2 (p, "addssw", 0, ORC_VAR_T12, ORC_VAR_T7, ORC_VAR_T8, + ORC_VAR_D1); + orc_program_append_2 (p, "addssw", 0, ORC_VAR_T12, ORC_VAR_T12, + ORC_VAR_T8, ORC_VAR_D1); + orc_program_append_2 (p, "mullw", 0, ORC_VAR_T1, ORC_VAR_T8, ORC_VAR_C5, + ORC_VAR_D1); + orc_program_append_2 (p, "shrsw", 0, ORC_VAR_T1, ORC_VAR_T1, ORC_VAR_C1, + ORC_VAR_D1); + orc_program_append_2 (p, "addssw", 0, ORC_VAR_T12, ORC_VAR_T12, + ORC_VAR_T1, ORC_VAR_D1); + orc_program_append_2 (p, "mullw", 0, ORC_VAR_T1, ORC_VAR_T8, ORC_VAR_C6, + ORC_VAR_D1); + orc_program_append_2 (p, "shrsw", 0, ORC_VAR_T1, ORC_VAR_T1, ORC_VAR_C1, + ORC_VAR_D1); + orc_program_append_2 (p, "subssw", 0, ORC_VAR_T11, ORC_VAR_T7, ORC_VAR_T1, + ORC_VAR_D1); + orc_program_append_2 (p, "mullw", 0, ORC_VAR_T1, ORC_VAR_T9, ORC_VAR_C7, + ORC_VAR_D1); + orc_program_append_2 (p, "shrsw", 0, ORC_VAR_T1, ORC_VAR_T1, ORC_VAR_C1, + ORC_VAR_D1); + orc_program_append_2 (p, "subssw", 0, ORC_VAR_T11, ORC_VAR_T11, + ORC_VAR_T1, ORC_VAR_D1); + orc_program_append_2 (p, "subssw", 0, ORC_VAR_T11, ORC_VAR_T11, + ORC_VAR_T1, ORC_VAR_D1); + orc_program_append_2 (p, "convssswb", 0, ORC_VAR_T13, ORC_VAR_T10, + ORC_VAR_D1, ORC_VAR_D1); + orc_program_append_2 (p, "convssswb", 0, ORC_VAR_T14, ORC_VAR_T11, + ORC_VAR_D1, ORC_VAR_D1); + orc_program_append_2 (p, "convssswb", 0, ORC_VAR_T15, ORC_VAR_T12, + ORC_VAR_D1, ORC_VAR_D1); + orc_program_append_2 (p, "mergebw", 0, ORC_VAR_T1, ORC_VAR_T3, + ORC_VAR_T13, ORC_VAR_D1); + orc_program_append_2 (p, "mergebw", 0, ORC_VAR_T2, ORC_VAR_T14, + ORC_VAR_T15, ORC_VAR_D1); + orc_program_append_2 (p, "mergewl", 0, 47, ORC_VAR_T1, ORC_VAR_T2, + ORC_VAR_D1); + orc_program_append_2 (p, "addb", 2, ORC_VAR_D1, 47, ORC_VAR_C2, + ORC_VAR_D1); + + result = orc_program_compile (p); + } + p_inited = TRUE; + orc_once_mutex_unlock (); + } + ex->program = p; + + ex->n = n; + ORC_EXECUTOR_M (ex) = m; + ex->arrays[ORC_VAR_D1] = d1; + ex->params[ORC_VAR_D1] = d1_stride; + ex->arrays[ORC_VAR_S1] = (void *) s1; + ex->params[ORC_VAR_S1] = s1_stride; + + func = p->code_exec; + func (ex); +} +#endif + + +/* cogorc_convert_AYUV_BGRA */ +#ifdef DISABLE_ORC +void +cogorc_convert_AYUV_BGRA (guint8 * d1, int d1_stride, const guint8 * s1, + int s1_stride, int n, int m) +{ + int i; + int j; + orc_union32 *ORC_RESTRICT ptr0; + const orc_union32 *ORC_RESTRICT ptr4; + orc_union32 var48; + orc_union32 var49; + orc_union16 var50; + orc_union16 var51; + orc_union16 var52; + orc_union16 var53; + orc_union16 var54; + orc_union32 var55; + orc_union32 var56; + orc_union32 var57; + orc_union16 var58; + orc_union16 var59; + orc_int8 var60; + orc_int8 var61; + orc_int8 var62; + orc_int8 var63; + orc_union16 var64; + orc_union16 var65; + orc_union16 var66; + orc_union16 var67; + orc_union16 var68; + orc_union16 var69; + orc_union16 var70; + orc_union16 var71; + orc_union16 var72; + orc_union16 var73; + orc_union16 var74; + orc_union16 var75; + orc_union16 var76; + orc_union16 var77; + orc_union16 var78; + orc_union16 var79; + orc_union16 var80; + orc_union16 var81; + orc_union16 var82; + orc_union16 var83; + orc_union16 var84; + orc_union16 var85; + orc_union16 var86; + orc_int8 var87; + orc_int8 var88; + orc_int8 var89; + orc_union16 var90; + orc_union16 var91; + orc_union32 var92; + + for (j = 0; j < m; j++) { + ptr0 = ORC_PTR_OFFSET (d1, d1_stride * j); + ptr4 = ORC_PTR_OFFSET (s1, s1_stride * j); + + /* 1: loadpb */ + var49.x4[0] = 0x00000080; /* 128 or 6.32404e-322f */ + var49.x4[1] = 0x00000080; /* 128 or 6.32404e-322f */ + var49.x4[2] = 0x00000080; /* 128 or 6.32404e-322f */ + var49.x4[3] = 0x00000080; /* 128 or 6.32404e-322f */ + /* 9: loadpw */ + var50.i = 0x0000002a; /* 42 or 2.07508e-322f */ + /* 14: loadpw */ + var51.i = 0x00000067; /* 103 or 5.08888e-322f */ + /* 21: loadpw */ + var52.i = 0x00000004; /* 4 or 1.97626e-323f */ + /* 25: loadpw */ + var53.i = 0x00000064; /* 100 or 4.94066e-322f */ + /* 29: loadpw */ + var54.i = 0x00000068; /* 104 or 5.13828e-322f */ + /* 40: loadpb */ + var55.x4[0] = 0x00000080; /* 128 or 6.32404e-322f */ + var55.x4[1] = 0x00000080; /* 128 or 6.32404e-322f */ + var55.x4[2] = 0x00000080; /* 128 or 6.32404e-322f */ + var55.x4[3] = 0x00000080; /* 128 or 6.32404e-322f */ + + for (i = 0; i < n; i++) { + /* 0: loadl */ + var48 = ptr4[i]; + /* 2: subb */ + var57.x4[0] = var48.x4[0] - var49.x4[0]; + var57.x4[1] = var48.x4[1] - var49.x4[1]; + var57.x4[2] = var48.x4[2] - var49.x4[2]; + var57.x4[3] = var48.x4[3] - var49.x4[3]; + /* 3: splitlw */ + var58.i = (var57.i >> 16) & 0xffff; + var59.i = var57.i & 0xffff; + /* 4: splitwb */ + var60 = (var59.i >> 8) & 0xff; + var61 = var59.i & 0xff; + /* 5: splitwb */ + var62 = (var58.i >> 8) & 0xff; + var63 = var58.i & 0xff; + /* 6: convsbw */ + var64.i = var60; + /* 7: convsbw */ + var65.i = var63; + /* 8: convsbw */ + var66.i = var62; + /* 10: mullw */ + var67.i = (var64.i * var50.i) & 0xffff; + /* 11: shrsw */ + var68.i = var67.i >> 8; + /* 12: addssw */ + var69.i = ORC_CLAMP_SW (var64.i + var68.i); + /* 13: addssw */ + var70.i = ORC_CLAMP_SW (var69.i + var66.i); + /* 15: mullw */ + var71.i = (var66.i * var51.i) & 0xffff; + /* 16: shrsw */ + var72.i = var71.i >> 8; + /* 17: subssw */ + var73.i = ORC_CLAMP_SW (var70.i - var72.i); + /* 18: addssw */ + var74.i = ORC_CLAMP_SW (var73.i + var66.i); + /* 19: addssw */ + var75.i = ORC_CLAMP_SW (var69.i + var65.i); + /* 20: addssw */ + var76.i = ORC_CLAMP_SW (var75.i + var65.i); + /* 22: mullw */ + var77.i = (var65.i * var52.i) & 0xffff; + /* 23: shrsw */ + var78.i = var77.i >> 8; + /* 24: addssw */ + var79.i = ORC_CLAMP_SW (var76.i + var78.i); + /* 26: mullw */ + var80.i = (var65.i * var53.i) & 0xffff; + /* 27: shrsw */ + var81.i = var80.i >> 8; + /* 28: subssw */ + var82.i = ORC_CLAMP_SW (var69.i - var81.i); + /* 30: mullw */ + var83.i = (var66.i * var54.i) & 0xffff; + /* 31: shrsw */ + var84.i = var83.i >> 8; + /* 32: subssw */ + var85.i = ORC_CLAMP_SW (var82.i - var84.i); + /* 33: subssw */ + var86.i = ORC_CLAMP_SW (var85.i - var84.i); + /* 34: convssswb */ + var87 = ORC_CLAMP_SB (var74.i); + /* 35: convssswb */ + var88 = ORC_CLAMP_SB (var86.i); + /* 36: convssswb */ + var89 = ORC_CLAMP_SB (var79.i); + /* 37: mergebw */ + var90.i = ((orc_uint8) var89) | ((orc_uint8) var88 << 8); + /* 38: mergebw */ + var91.i = ((orc_uint8) var87) | ((orc_uint8) var61 << 8); + /* 39: mergewl */ + var92.i = ((orc_uint16) var90.i) | ((orc_uint16) var91.i << 16); + /* 41: addb */ + var56.x4[0] = var92.x4[0] + var55.x4[0]; + var56.x4[1] = var92.x4[1] + var55.x4[1]; + var56.x4[2] = var92.x4[2] + var55.x4[2]; + var56.x4[3] = var92.x4[3] + var55.x4[3]; + /* 42: storel */ + ptr0[i] = var56; + } + } + +} + +#else +static void +_backup_cogorc_convert_AYUV_BGRA (OrcExecutor * ex) +{ + int i; + int j; + int n = ex->n; + int m = ex->params[ORC_VAR_A1]; + orc_union32 *ORC_RESTRICT ptr0; + const orc_union32 *ORC_RESTRICT ptr4; + orc_union32 var48; + orc_union32 var49; + orc_union16 var50; + orc_union16 var51; + orc_union16 var52; + orc_union16 var53; + orc_union16 var54; + orc_union32 var55; + orc_union32 var56; + orc_union32 var57; + orc_union16 var58; + orc_union16 var59; + orc_int8 var60; + orc_int8 var61; + orc_int8 var62; + orc_int8 var63; + orc_union16 var64; + orc_union16 var65; + orc_union16 var66; + orc_union16 var67; + orc_union16 var68; + orc_union16 var69; + orc_union16 var70; + orc_union16 var71; + orc_union16 var72; + orc_union16 var73; + orc_union16 var74; + orc_union16 var75; + orc_union16 var76; + orc_union16 var77; + orc_union16 var78; + orc_union16 var79; + orc_union16 var80; + orc_union16 var81; + orc_union16 var82; + orc_union16 var83; + orc_union16 var84; + orc_union16 var85; + orc_union16 var86; + orc_int8 var87; + orc_int8 var88; + orc_int8 var89; + orc_union16 var90; + orc_union16 var91; + orc_union32 var92; + + for (j = 0; j < m; j++) { + ptr0 = ORC_PTR_OFFSET (ex->arrays[0], ex->params[0] * j); + ptr4 = ORC_PTR_OFFSET (ex->arrays[4], ex->params[4] * j); + + /* 1: loadpb */ + var49.x4[0] = 0x00000080; /* 128 or 6.32404e-322f */ + var49.x4[1] = 0x00000080; /* 128 or 6.32404e-322f */ + var49.x4[2] = 0x00000080; /* 128 or 6.32404e-322f */ + var49.x4[3] = 0x00000080; /* 128 or 6.32404e-322f */ + /* 9: loadpw */ + var50.i = 0x0000002a; /* 42 or 2.07508e-322f */ + /* 14: loadpw */ + var51.i = 0x00000067; /* 103 or 5.08888e-322f */ + /* 21: loadpw */ + var52.i = 0x00000004; /* 4 or 1.97626e-323f */ + /* 25: loadpw */ + var53.i = 0x00000064; /* 100 or 4.94066e-322f */ + /* 29: loadpw */ + var54.i = 0x00000068; /* 104 or 5.13828e-322f */ + /* 40: loadpb */ + var55.x4[0] = 0x00000080; /* 128 or 6.32404e-322f */ + var55.x4[1] = 0x00000080; /* 128 or 6.32404e-322f */ + var55.x4[2] = 0x00000080; /* 128 or 6.32404e-322f */ + var55.x4[3] = 0x00000080; /* 128 or 6.32404e-322f */ + + for (i = 0; i < n; i++) { + /* 0: loadl */ + var48 = ptr4[i]; + /* 2: subb */ + var57.x4[0] = var48.x4[0] - var49.x4[0]; + var57.x4[1] = var48.x4[1] - var49.x4[1]; + var57.x4[2] = var48.x4[2] - var49.x4[2]; + var57.x4[3] = var48.x4[3] - var49.x4[3]; + /* 3: splitlw */ + var58.i = (var57.i >> 16) & 0xffff; + var59.i = var57.i & 0xffff; + /* 4: splitwb */ + var60 = (var59.i >> 8) & 0xff; + var61 = var59.i & 0xff; + /* 5: splitwb */ + var62 = (var58.i >> 8) & 0xff; + var63 = var58.i & 0xff; + /* 6: convsbw */ + var64.i = var60; + /* 7: convsbw */ + var65.i = var63; + /* 8: convsbw */ + var66.i = var62; + /* 10: mullw */ + var67.i = (var64.i * var50.i) & 0xffff; + /* 11: shrsw */ + var68.i = var67.i >> 8; + /* 12: addssw */ + var69.i = ORC_CLAMP_SW (var64.i + var68.i); + /* 13: addssw */ + var70.i = ORC_CLAMP_SW (var69.i + var66.i); + /* 15: mullw */ + var71.i = (var66.i * var51.i) & 0xffff; + /* 16: shrsw */ + var72.i = var71.i >> 8; + /* 17: subssw */ + var73.i = ORC_CLAMP_SW (var70.i - var72.i); + /* 18: addssw */ + var74.i = ORC_CLAMP_SW (var73.i + var66.i); + /* 19: addssw */ + var75.i = ORC_CLAMP_SW (var69.i + var65.i); + /* 20: addssw */ + var76.i = ORC_CLAMP_SW (var75.i + var65.i); + /* 22: mullw */ + var77.i = (var65.i * var52.i) & 0xffff; + /* 23: shrsw */ + var78.i = var77.i >> 8; + /* 24: addssw */ + var79.i = ORC_CLAMP_SW (var76.i + var78.i); + /* 26: mullw */ + var80.i = (var65.i * var53.i) & 0xffff; + /* 27: shrsw */ + var81.i = var80.i >> 8; + /* 28: subssw */ + var82.i = ORC_CLAMP_SW (var69.i - var81.i); + /* 30: mullw */ + var83.i = (var66.i * var54.i) & 0xffff; + /* 31: shrsw */ + var84.i = var83.i >> 8; + /* 32: subssw */ + var85.i = ORC_CLAMP_SW (var82.i - var84.i); + /* 33: subssw */ + var86.i = ORC_CLAMP_SW (var85.i - var84.i); + /* 34: convssswb */ + var87 = ORC_CLAMP_SB (var74.i); + /* 35: convssswb */ + var88 = ORC_CLAMP_SB (var86.i); + /* 36: convssswb */ + var89 = ORC_CLAMP_SB (var79.i); + /* 37: mergebw */ + var90.i = ((orc_uint8) var89) | ((orc_uint8) var88 << 8); + /* 38: mergebw */ + var91.i = ((orc_uint8) var87) | ((orc_uint8) var61 << 8); + /* 39: mergewl */ + var92.i = ((orc_uint16) var90.i) | ((orc_uint16) var91.i << 16); + /* 41: addb */ + var56.x4[0] = var92.x4[0] + var55.x4[0]; + var56.x4[1] = var92.x4[1] + var55.x4[1]; + var56.x4[2] = var92.x4[2] + var55.x4[2]; + var56.x4[3] = var92.x4[3] + var55.x4[3]; + /* 42: storel */ + ptr0[i] = var56; + } + } + +} + +void +cogorc_convert_AYUV_BGRA (guint8 * d1, int d1_stride, const guint8 * s1, + int s1_stride, int n, int m) +{ + OrcExecutor _ex, *ex = &_ex; + static int p_inited = 0; + static OrcProgram *p = 0; + void (*func) (OrcExecutor *); + + if (!p_inited) { + orc_once_mutex_lock (); + if (!p_inited) { + OrcCompileResult result; + + p = orc_program_new (); + orc_program_set_2d (p); + orc_program_set_name (p, "cogorc_convert_AYUV_BGRA"); + orc_program_set_backup_function (p, _backup_cogorc_convert_AYUV_BGRA); + orc_program_add_destination (p, 4, "d1"); + orc_program_add_source (p, 4, "s1"); + orc_program_add_constant (p, 1, 0x00000008, "c1"); + orc_program_add_constant (p, 4, 0x00000080, "c2"); + orc_program_add_constant (p, 4, 0x0000002a, "c3"); + orc_program_add_constant (p, 4, 0x00000067, "c4"); + orc_program_add_constant (p, 4, 0x00000004, "c5"); + orc_program_add_constant (p, 4, 0x00000064, "c6"); + orc_program_add_constant (p, 4, 0x00000068, "c7"); + orc_program_add_constant (p, 0, 0x00000080, "c8"); + orc_program_add_temporary (p, 2, "t1"); + orc_program_add_temporary (p, 2, "t2"); + orc_program_add_temporary (p, 1, "t3"); + orc_program_add_temporary (p, 1, "t4"); + orc_program_add_temporary (p, 1, "t5"); + orc_program_add_temporary (p, 1, "t6"); + orc_program_add_temporary (p, 2, "t7"); + orc_program_add_temporary (p, 2, "t8"); + orc_program_add_temporary (p, 2, "t9"); + orc_program_add_temporary (p, 2, "t10"); + orc_program_add_temporary (p, 2, "t11"); + orc_program_add_temporary (p, 2, "t12"); + orc_program_add_temporary (p, 1, "t13"); + orc_program_add_temporary (p, 1, "t14"); + orc_program_add_temporary (p, 1, "t15"); + orc_program_add_temporary (p, 4, "t16"); + + orc_program_append_2 (p, "subb", 2, 47, ORC_VAR_S1, ORC_VAR_C2, + ORC_VAR_D1); + orc_program_append_2 (p, "splitlw", 0, ORC_VAR_T1, ORC_VAR_T2, 47, + ORC_VAR_D1); + orc_program_append_2 (p, "splitwb", 0, ORC_VAR_T4, ORC_VAR_T3, ORC_VAR_T2, + ORC_VAR_D1); + orc_program_append_2 (p, "splitwb", 0, ORC_VAR_T6, ORC_VAR_T5, ORC_VAR_T1, + ORC_VAR_D1); + orc_program_append_2 (p, "convsbw", 0, ORC_VAR_T7, ORC_VAR_T4, ORC_VAR_D1, + ORC_VAR_D1); + orc_program_append_2 (p, "convsbw", 0, ORC_VAR_T8, ORC_VAR_T5, ORC_VAR_D1, + ORC_VAR_D1); + orc_program_append_2 (p, "convsbw", 0, ORC_VAR_T9, ORC_VAR_T6, ORC_VAR_D1, + ORC_VAR_D1); + orc_program_append_2 (p, "mullw", 0, ORC_VAR_T1, ORC_VAR_T7, ORC_VAR_C3, + ORC_VAR_D1); + orc_program_append_2 (p, "shrsw", 0, ORC_VAR_T1, ORC_VAR_T1, ORC_VAR_C1, + ORC_VAR_D1); + orc_program_append_2 (p, "addssw", 0, ORC_VAR_T7, ORC_VAR_T7, ORC_VAR_T1, + ORC_VAR_D1); + orc_program_append_2 (p, "addssw", 0, ORC_VAR_T10, ORC_VAR_T7, ORC_VAR_T9, + ORC_VAR_D1); + orc_program_append_2 (p, "mullw", 0, ORC_VAR_T1, ORC_VAR_T9, ORC_VAR_C4, + ORC_VAR_D1); + orc_program_append_2 (p, "shrsw", 0, ORC_VAR_T1, ORC_VAR_T1, ORC_VAR_C1, + ORC_VAR_D1); + orc_program_append_2 (p, "subssw", 0, ORC_VAR_T10, ORC_VAR_T10, + ORC_VAR_T1, ORC_VAR_D1); + orc_program_append_2 (p, "addssw", 0, ORC_VAR_T10, ORC_VAR_T10, + ORC_VAR_T9, ORC_VAR_D1); + orc_program_append_2 (p, "addssw", 0, ORC_VAR_T12, ORC_VAR_T7, ORC_VAR_T8, + ORC_VAR_D1); + orc_program_append_2 (p, "addssw", 0, ORC_VAR_T12, ORC_VAR_T12, + ORC_VAR_T8, ORC_VAR_D1); + orc_program_append_2 (p, "mullw", 0, ORC_VAR_T1, ORC_VAR_T8, ORC_VAR_C5, + ORC_VAR_D1); + orc_program_append_2 (p, "shrsw", 0, ORC_VAR_T1, ORC_VAR_T1, ORC_VAR_C1, + ORC_VAR_D1); + orc_program_append_2 (p, "addssw", 0, ORC_VAR_T12, ORC_VAR_T12, + ORC_VAR_T1, ORC_VAR_D1); + orc_program_append_2 (p, "mullw", 0, ORC_VAR_T1, ORC_VAR_T8, ORC_VAR_C6, + ORC_VAR_D1); + orc_program_append_2 (p, "shrsw", 0, ORC_VAR_T1, ORC_VAR_T1, ORC_VAR_C1, + ORC_VAR_D1); + orc_program_append_2 (p, "subssw", 0, ORC_VAR_T11, ORC_VAR_T7, ORC_VAR_T1, + ORC_VAR_D1); + orc_program_append_2 (p, "mullw", 0, ORC_VAR_T1, ORC_VAR_T9, ORC_VAR_C7, + ORC_VAR_D1); + orc_program_append_2 (p, "shrsw", 0, ORC_VAR_T1, ORC_VAR_T1, ORC_VAR_C1, + ORC_VAR_D1); + orc_program_append_2 (p, "subssw", 0, ORC_VAR_T11, ORC_VAR_T11, + ORC_VAR_T1, ORC_VAR_D1); + orc_program_append_2 (p, "subssw", 0, ORC_VAR_T11, ORC_VAR_T11, + ORC_VAR_T1, ORC_VAR_D1); + orc_program_append_2 (p, "convssswb", 0, ORC_VAR_T13, ORC_VAR_T10, + ORC_VAR_D1, ORC_VAR_D1); + orc_program_append_2 (p, "convssswb", 0, ORC_VAR_T14, ORC_VAR_T11, + ORC_VAR_D1, ORC_VAR_D1); + orc_program_append_2 (p, "convssswb", 0, ORC_VAR_T15, ORC_VAR_T12, + ORC_VAR_D1, ORC_VAR_D1); + orc_program_append_2 (p, "mergebw", 0, ORC_VAR_T1, ORC_VAR_T15, + ORC_VAR_T14, ORC_VAR_D1); + orc_program_append_2 (p, "mergebw", 0, ORC_VAR_T2, ORC_VAR_T13, + ORC_VAR_T3, ORC_VAR_D1); + orc_program_append_2 (p, "mergewl", 0, 47, ORC_VAR_T1, ORC_VAR_T2, + ORC_VAR_D1); + orc_program_append_2 (p, "addb", 2, ORC_VAR_D1, 47, ORC_VAR_C2, + ORC_VAR_D1); + + result = orc_program_compile (p); + } + p_inited = TRUE; + orc_once_mutex_unlock (); + } + ex->program = p; + + ex->n = n; + ORC_EXECUTOR_M (ex) = m; + ex->arrays[ORC_VAR_D1] = d1; + ex->params[ORC_VAR_D1] = d1_stride; + ex->arrays[ORC_VAR_S1] = (void *) s1; + ex->params[ORC_VAR_S1] = s1_stride; + + func = p->code_exec; + func (ex); +} +#endif + + +/* cogorc_convert_AYUV_ABGR */ +#ifdef DISABLE_ORC +void +cogorc_convert_AYUV_ABGR (guint8 * d1, int d1_stride, const guint8 * s1, + int s1_stride, int n, int m) +{ + int i; + int j; + orc_union32 *ORC_RESTRICT ptr0; + const orc_union32 *ORC_RESTRICT ptr4; + orc_union32 var48; + orc_union32 var49; + orc_union16 var50; + orc_union16 var51; + orc_union16 var52; + orc_union16 var53; + orc_union16 var54; + orc_union32 var55; + orc_union32 var56; + orc_union32 var57; + orc_union16 var58; + orc_union16 var59; + orc_int8 var60; + orc_int8 var61; + orc_int8 var62; + orc_int8 var63; + orc_union16 var64; + orc_union16 var65; + orc_union16 var66; + orc_union16 var67; + orc_union16 var68; + orc_union16 var69; + orc_union16 var70; + orc_union16 var71; + orc_union16 var72; + orc_union16 var73; + orc_union16 var74; + orc_union16 var75; + orc_union16 var76; + orc_union16 var77; + orc_union16 var78; + orc_union16 var79; + orc_union16 var80; + orc_union16 var81; + orc_union16 var82; + orc_union16 var83; + orc_union16 var84; + orc_union16 var85; + orc_union16 var86; + orc_int8 var87; + orc_int8 var88; + orc_int8 var89; + orc_union16 var90; + orc_union16 var91; + orc_union32 var92; + + for (j = 0; j < m; j++) { + ptr0 = ORC_PTR_OFFSET (d1, d1_stride * j); + ptr4 = ORC_PTR_OFFSET (s1, s1_stride * j); + + /* 1: loadpb */ + var49.x4[0] = 0x00000080; /* 128 or 6.32404e-322f */ + var49.x4[1] = 0x00000080; /* 128 or 6.32404e-322f */ + var49.x4[2] = 0x00000080; /* 128 or 6.32404e-322f */ + var49.x4[3] = 0x00000080; /* 128 or 6.32404e-322f */ + /* 9: loadpw */ + var50.i = 0x0000002a; /* 42 or 2.07508e-322f */ + /* 14: loadpw */ + var51.i = 0x00000067; /* 103 or 5.08888e-322f */ + /* 21: loadpw */ + var52.i = 0x00000004; /* 4 or 1.97626e-323f */ + /* 25: loadpw */ + var53.i = 0x00000064; /* 100 or 4.94066e-322f */ + /* 29: loadpw */ + var54.i = 0x00000068; /* 104 or 5.13828e-322f */ + /* 40: loadpb */ + var55.x4[0] = 0x00000080; /* 128 or 6.32404e-322f */ + var55.x4[1] = 0x00000080; /* 128 or 6.32404e-322f */ + var55.x4[2] = 0x00000080; /* 128 or 6.32404e-322f */ + var55.x4[3] = 0x00000080; /* 128 or 6.32404e-322f */ + + for (i = 0; i < n; i++) { + /* 0: loadl */ + var48 = ptr4[i]; + /* 2: subb */ + var57.x4[0] = var48.x4[0] - var49.x4[0]; + var57.x4[1] = var48.x4[1] - var49.x4[1]; + var57.x4[2] = var48.x4[2] - var49.x4[2]; + var57.x4[3] = var48.x4[3] - var49.x4[3]; + /* 3: splitlw */ + var58.i = (var57.i >> 16) & 0xffff; + var59.i = var57.i & 0xffff; + /* 4: splitwb */ + var60 = (var59.i >> 8) & 0xff; + var61 = var59.i & 0xff; + /* 5: splitwb */ + var62 = (var58.i >> 8) & 0xff; + var63 = var58.i & 0xff; + /* 6: convsbw */ + var64.i = var60; + /* 7: convsbw */ + var65.i = var63; + /* 8: convsbw */ + var66.i = var62; + /* 10: mullw */ + var67.i = (var64.i * var50.i) & 0xffff; + /* 11: shrsw */ + var68.i = var67.i >> 8; + /* 12: addssw */ + var69.i = ORC_CLAMP_SW (var64.i + var68.i); + /* 13: addssw */ + var70.i = ORC_CLAMP_SW (var69.i + var66.i); + /* 15: mullw */ + var71.i = (var66.i * var51.i) & 0xffff; + /* 16: shrsw */ + var72.i = var71.i >> 8; + /* 17: subssw */ + var73.i = ORC_CLAMP_SW (var70.i - var72.i); + /* 18: addssw */ + var74.i = ORC_CLAMP_SW (var73.i + var66.i); + /* 19: addssw */ + var75.i = ORC_CLAMP_SW (var69.i + var65.i); + /* 20: addssw */ + var76.i = ORC_CLAMP_SW (var75.i + var65.i); + /* 22: mullw */ + var77.i = (var65.i * var52.i) & 0xffff; + /* 23: shrsw */ + var78.i = var77.i >> 8; + /* 24: addssw */ + var79.i = ORC_CLAMP_SW (var76.i + var78.i); + /* 26: mullw */ + var80.i = (var65.i * var53.i) & 0xffff; + /* 27: shrsw */ + var81.i = var80.i >> 8; + /* 28: subssw */ + var82.i = ORC_CLAMP_SW (var69.i - var81.i); + /* 30: mullw */ + var83.i = (var66.i * var54.i) & 0xffff; + /* 31: shrsw */ + var84.i = var83.i >> 8; + /* 32: subssw */ + var85.i = ORC_CLAMP_SW (var82.i - var84.i); + /* 33: subssw */ + var86.i = ORC_CLAMP_SW (var85.i - var84.i); + /* 34: convssswb */ + var87 = ORC_CLAMP_SB (var74.i); + /* 35: convssswb */ + var88 = ORC_CLAMP_SB (var86.i); + /* 36: convssswb */ + var89 = ORC_CLAMP_SB (var79.i); + /* 37: mergebw */ + var90.i = ((orc_uint8) var61) | ((orc_uint8) var89 << 8); + /* 38: mergebw */ + var91.i = ((orc_uint8) var88) | ((orc_uint8) var87 << 8); + /* 39: mergewl */ + var92.i = ((orc_uint16) var90.i) | ((orc_uint16) var91.i << 16); + /* 41: addb */ + var56.x4[0] = var92.x4[0] + var55.x4[0]; + var56.x4[1] = var92.x4[1] + var55.x4[1]; + var56.x4[2] = var92.x4[2] + var55.x4[2]; + var56.x4[3] = var92.x4[3] + var55.x4[3]; + /* 42: storel */ + ptr0[i] = var56; + } + } + +} + +#else +static void +_backup_cogorc_convert_AYUV_ABGR (OrcExecutor * ex) +{ + int i; + int j; + int n = ex->n; + int m = ex->params[ORC_VAR_A1]; + orc_union32 *ORC_RESTRICT ptr0; + const orc_union32 *ORC_RESTRICT ptr4; + orc_union32 var48; + orc_union32 var49; + orc_union16 var50; + orc_union16 var51; + orc_union16 var52; + orc_union16 var53; + orc_union16 var54; + orc_union32 var55; + orc_union32 var56; + orc_union32 var57; + orc_union16 var58; + orc_union16 var59; + orc_int8 var60; + orc_int8 var61; + orc_int8 var62; + orc_int8 var63; + orc_union16 var64; + orc_union16 var65; + orc_union16 var66; + orc_union16 var67; + orc_union16 var68; + orc_union16 var69; + orc_union16 var70; + orc_union16 var71; + orc_union16 var72; + orc_union16 var73; + orc_union16 var74; + orc_union16 var75; + orc_union16 var76; + orc_union16 var77; + orc_union16 var78; + orc_union16 var79; + orc_union16 var80; + orc_union16 var81; + orc_union16 var82; + orc_union16 var83; + orc_union16 var84; + orc_union16 var85; + orc_union16 var86; + orc_int8 var87; + orc_int8 var88; + orc_int8 var89; + orc_union16 var90; + orc_union16 var91; + orc_union32 var92; + + for (j = 0; j < m; j++) { + ptr0 = ORC_PTR_OFFSET (ex->arrays[0], ex->params[0] * j); + ptr4 = ORC_PTR_OFFSET (ex->arrays[4], ex->params[4] * j); + + /* 1: loadpb */ + var49.x4[0] = 0x00000080; /* 128 or 6.32404e-322f */ + var49.x4[1] = 0x00000080; /* 128 or 6.32404e-322f */ + var49.x4[2] = 0x00000080; /* 128 or 6.32404e-322f */ + var49.x4[3] = 0x00000080; /* 128 or 6.32404e-322f */ + /* 9: loadpw */ + var50.i = 0x0000002a; /* 42 or 2.07508e-322f */ + /* 14: loadpw */ + var51.i = 0x00000067; /* 103 or 5.08888e-322f */ + /* 21: loadpw */ + var52.i = 0x00000004; /* 4 or 1.97626e-323f */ + /* 25: loadpw */ + var53.i = 0x00000064; /* 100 or 4.94066e-322f */ + /* 29: loadpw */ + var54.i = 0x00000068; /* 104 or 5.13828e-322f */ + /* 40: loadpb */ + var55.x4[0] = 0x00000080; /* 128 or 6.32404e-322f */ + var55.x4[1] = 0x00000080; /* 128 or 6.32404e-322f */ + var55.x4[2] = 0x00000080; /* 128 or 6.32404e-322f */ + var55.x4[3] = 0x00000080; /* 128 or 6.32404e-322f */ + + for (i = 0; i < n; i++) { + /* 0: loadl */ + var48 = ptr4[i]; + /* 2: subb */ + var57.x4[0] = var48.x4[0] - var49.x4[0]; + var57.x4[1] = var48.x4[1] - var49.x4[1]; + var57.x4[2] = var48.x4[2] - var49.x4[2]; + var57.x4[3] = var48.x4[3] - var49.x4[3]; + /* 3: splitlw */ + var58.i = (var57.i >> 16) & 0xffff; + var59.i = var57.i & 0xffff; + /* 4: splitwb */ + var60 = (var59.i >> 8) & 0xff; + var61 = var59.i & 0xff; + /* 5: splitwb */ + var62 = (var58.i >> 8) & 0xff; + var63 = var58.i & 0xff; + /* 6: convsbw */ + var64.i = var60; + /* 7: convsbw */ + var65.i = var63; + /* 8: convsbw */ + var66.i = var62; + /* 10: mullw */ + var67.i = (var64.i * var50.i) & 0xffff; + /* 11: shrsw */ + var68.i = var67.i >> 8; + /* 12: addssw */ + var69.i = ORC_CLAMP_SW (var64.i + var68.i); + /* 13: addssw */ + var70.i = ORC_CLAMP_SW (var69.i + var66.i); + /* 15: mullw */ + var71.i = (var66.i * var51.i) & 0xffff; + /* 16: shrsw */ + var72.i = var71.i >> 8; + /* 17: subssw */ + var73.i = ORC_CLAMP_SW (var70.i - var72.i); + /* 18: addssw */ + var74.i = ORC_CLAMP_SW (var73.i + var66.i); + /* 19: addssw */ + var75.i = ORC_CLAMP_SW (var69.i + var65.i); + /* 20: addssw */ + var76.i = ORC_CLAMP_SW (var75.i + var65.i); + /* 22: mullw */ + var77.i = (var65.i * var52.i) & 0xffff; + /* 23: shrsw */ + var78.i = var77.i >> 8; + /* 24: addssw */ + var79.i = ORC_CLAMP_SW (var76.i + var78.i); + /* 26: mullw */ + var80.i = (var65.i * var53.i) & 0xffff; + /* 27: shrsw */ + var81.i = var80.i >> 8; + /* 28: subssw */ + var82.i = ORC_CLAMP_SW (var69.i - var81.i); + /* 30: mullw */ + var83.i = (var66.i * var54.i) & 0xffff; + /* 31: shrsw */ + var84.i = var83.i >> 8; + /* 32: subssw */ + var85.i = ORC_CLAMP_SW (var82.i - var84.i); + /* 33: subssw */ + var86.i = ORC_CLAMP_SW (var85.i - var84.i); + /* 34: convssswb */ + var87 = ORC_CLAMP_SB (var74.i); + /* 35: convssswb */ + var88 = ORC_CLAMP_SB (var86.i); + /* 36: convssswb */ + var89 = ORC_CLAMP_SB (var79.i); + /* 37: mergebw */ + var90.i = ((orc_uint8) var61) | ((orc_uint8) var89 << 8); + /* 38: mergebw */ + var91.i = ((orc_uint8) var88) | ((orc_uint8) var87 << 8); + /* 39: mergewl */ + var92.i = ((orc_uint16) var90.i) | ((orc_uint16) var91.i << 16); + /* 41: addb */ + var56.x4[0] = var92.x4[0] + var55.x4[0]; + var56.x4[1] = var92.x4[1] + var55.x4[1]; + var56.x4[2] = var92.x4[2] + var55.x4[2]; + var56.x4[3] = var92.x4[3] + var55.x4[3]; + /* 42: storel */ + ptr0[i] = var56; + } + } + +} + +void +cogorc_convert_AYUV_ABGR (guint8 * d1, int d1_stride, const guint8 * s1, + int s1_stride, int n, int m) +{ + OrcExecutor _ex, *ex = &_ex; + static int p_inited = 0; + static OrcProgram *p = 0; + void (*func) (OrcExecutor *); + + if (!p_inited) { + orc_once_mutex_lock (); + if (!p_inited) { + OrcCompileResult result; + + p = orc_program_new (); + orc_program_set_2d (p); + orc_program_set_name (p, "cogorc_convert_AYUV_ABGR"); + orc_program_set_backup_function (p, _backup_cogorc_convert_AYUV_ABGR); + orc_program_add_destination (p, 4, "d1"); + orc_program_add_source (p, 4, "s1"); + orc_program_add_constant (p, 1, 0x00000008, "c1"); + orc_program_add_constant (p, 4, 0x00000080, "c2"); + orc_program_add_constant (p, 4, 0x0000002a, "c3"); + orc_program_add_constant (p, 4, 0x00000067, "c4"); + orc_program_add_constant (p, 4, 0x00000004, "c5"); + orc_program_add_constant (p, 4, 0x00000064, "c6"); + orc_program_add_constant (p, 4, 0x00000068, "c7"); + orc_program_add_constant (p, 0, 0x00000080, "c8"); + orc_program_add_temporary (p, 2, "t1"); + orc_program_add_temporary (p, 2, "t2"); + orc_program_add_temporary (p, 1, "t3"); + orc_program_add_temporary (p, 1, "t4"); + orc_program_add_temporary (p, 1, "t5"); + orc_program_add_temporary (p, 1, "t6"); + orc_program_add_temporary (p, 2, "t7"); + orc_program_add_temporary (p, 2, "t8"); + orc_program_add_temporary (p, 2, "t9"); + orc_program_add_temporary (p, 2, "t10"); + orc_program_add_temporary (p, 2, "t11"); + orc_program_add_temporary (p, 2, "t12"); + orc_program_add_temporary (p, 1, "t13"); + orc_program_add_temporary (p, 1, "t14"); + orc_program_add_temporary (p, 1, "t15"); + orc_program_add_temporary (p, 4, "t16"); + + orc_program_append_2 (p, "subb", 2, 47, ORC_VAR_S1, ORC_VAR_C2, + ORC_VAR_D1); + orc_program_append_2 (p, "splitlw", 0, ORC_VAR_T1, ORC_VAR_T2, 47, + ORC_VAR_D1); + orc_program_append_2 (p, "splitwb", 0, ORC_VAR_T4, ORC_VAR_T3, ORC_VAR_T2, + ORC_VAR_D1); + orc_program_append_2 (p, "splitwb", 0, ORC_VAR_T6, ORC_VAR_T5, ORC_VAR_T1, + ORC_VAR_D1); + orc_program_append_2 (p, "convsbw", 0, ORC_VAR_T7, ORC_VAR_T4, ORC_VAR_D1, + ORC_VAR_D1); + orc_program_append_2 (p, "convsbw", 0, ORC_VAR_T8, ORC_VAR_T5, ORC_VAR_D1, + ORC_VAR_D1); + orc_program_append_2 (p, "convsbw", 0, ORC_VAR_T9, ORC_VAR_T6, ORC_VAR_D1, + ORC_VAR_D1); + orc_program_append_2 (p, "mullw", 0, ORC_VAR_T1, ORC_VAR_T7, ORC_VAR_C3, + ORC_VAR_D1); + orc_program_append_2 (p, "shrsw", 0, ORC_VAR_T1, ORC_VAR_T1, ORC_VAR_C1, + ORC_VAR_D1); + orc_program_append_2 (p, "addssw", 0, ORC_VAR_T7, ORC_VAR_T7, ORC_VAR_T1, + ORC_VAR_D1); + orc_program_append_2 (p, "addssw", 0, ORC_VAR_T10, ORC_VAR_T7, ORC_VAR_T9, + ORC_VAR_D1); + orc_program_append_2 (p, "mullw", 0, ORC_VAR_T1, ORC_VAR_T9, ORC_VAR_C4, + ORC_VAR_D1); + orc_program_append_2 (p, "shrsw", 0, ORC_VAR_T1, ORC_VAR_T1, ORC_VAR_C1, + ORC_VAR_D1); + orc_program_append_2 (p, "subssw", 0, ORC_VAR_T10, ORC_VAR_T10, + ORC_VAR_T1, ORC_VAR_D1); + orc_program_append_2 (p, "addssw", 0, ORC_VAR_T10, ORC_VAR_T10, + ORC_VAR_T9, ORC_VAR_D1); + orc_program_append_2 (p, "addssw", 0, ORC_VAR_T12, ORC_VAR_T7, ORC_VAR_T8, + ORC_VAR_D1); + orc_program_append_2 (p, "addssw", 0, ORC_VAR_T12, ORC_VAR_T12, + ORC_VAR_T8, ORC_VAR_D1); + orc_program_append_2 (p, "mullw", 0, ORC_VAR_T1, ORC_VAR_T8, ORC_VAR_C5, + ORC_VAR_D1); + orc_program_append_2 (p, "shrsw", 0, ORC_VAR_T1, ORC_VAR_T1, ORC_VAR_C1, + ORC_VAR_D1); + orc_program_append_2 (p, "addssw", 0, ORC_VAR_T12, ORC_VAR_T12, + ORC_VAR_T1, ORC_VAR_D1); + orc_program_append_2 (p, "mullw", 0, ORC_VAR_T1, ORC_VAR_T8, ORC_VAR_C6, + ORC_VAR_D1); + orc_program_append_2 (p, "shrsw", 0, ORC_VAR_T1, ORC_VAR_T1, ORC_VAR_C1, + ORC_VAR_D1); + orc_program_append_2 (p, "subssw", 0, ORC_VAR_T11, ORC_VAR_T7, ORC_VAR_T1, + ORC_VAR_D1); + orc_program_append_2 (p, "mullw", 0, ORC_VAR_T1, ORC_VAR_T9, ORC_VAR_C7, + ORC_VAR_D1); + orc_program_append_2 (p, "shrsw", 0, ORC_VAR_T1, ORC_VAR_T1, ORC_VAR_C1, + ORC_VAR_D1); + orc_program_append_2 (p, "subssw", 0, ORC_VAR_T11, ORC_VAR_T11, + ORC_VAR_T1, ORC_VAR_D1); + orc_program_append_2 (p, "subssw", 0, ORC_VAR_T11, ORC_VAR_T11, + ORC_VAR_T1, ORC_VAR_D1); + orc_program_append_2 (p, "convssswb", 0, ORC_VAR_T13, ORC_VAR_T10, + ORC_VAR_D1, ORC_VAR_D1); + orc_program_append_2 (p, "convssswb", 0, ORC_VAR_T14, ORC_VAR_T11, + ORC_VAR_D1, ORC_VAR_D1); + orc_program_append_2 (p, "convssswb", 0, ORC_VAR_T15, ORC_VAR_T12, + ORC_VAR_D1, ORC_VAR_D1); + orc_program_append_2 (p, "mergebw", 0, ORC_VAR_T1, ORC_VAR_T3, + ORC_VAR_T15, ORC_VAR_D1); + orc_program_append_2 (p, "mergebw", 0, ORC_VAR_T2, ORC_VAR_T14, + ORC_VAR_T13, ORC_VAR_D1); + orc_program_append_2 (p, "mergewl", 0, 47, ORC_VAR_T1, ORC_VAR_T2, + ORC_VAR_D1); + orc_program_append_2 (p, "addb", 2, ORC_VAR_D1, 47, ORC_VAR_C2, + ORC_VAR_D1); + + result = orc_program_compile (p); + } + p_inited = TRUE; + orc_once_mutex_unlock (); + } + ex->program = p; + + ex->n = n; + ORC_EXECUTOR_M (ex) = m; + ex->arrays[ORC_VAR_D1] = d1; + ex->params[ORC_VAR_D1] = d1_stride; + ex->arrays[ORC_VAR_S1] = (void *) s1; + ex->params[ORC_VAR_S1] = s1_stride; + + func = p->code_exec; + func (ex); +} +#endif + + +/* cogorc_convert_AYUV_RGBA */ +#ifdef DISABLE_ORC +void +cogorc_convert_AYUV_RGBA (guint8 * d1, int d1_stride, const guint8 * s1, + int s1_stride, int n, int m) +{ + int i; + int j; + orc_union32 *ORC_RESTRICT ptr0; + const orc_union32 *ORC_RESTRICT ptr4; + orc_union32 var48; + orc_union32 var49; + orc_union16 var50; + orc_union16 var51; + orc_union16 var52; + orc_union16 var53; + orc_union16 var54; + orc_union32 var55; + orc_union32 var56; + orc_union32 var57; + orc_union16 var58; + orc_union16 var59; + orc_int8 var60; + orc_int8 var61; + orc_int8 var62; + orc_int8 var63; + orc_union16 var64; + orc_union16 var65; + orc_union16 var66; + orc_union16 var67; + orc_union16 var68; + orc_union16 var69; + orc_union16 var70; + orc_union16 var71; + orc_union16 var72; + orc_union16 var73; + orc_union16 var74; + orc_union16 var75; + orc_union16 var76; + orc_union16 var77; + orc_union16 var78; + orc_union16 var79; + orc_union16 var80; + orc_union16 var81; + orc_union16 var82; + orc_union16 var83; + orc_union16 var84; + orc_union16 var85; + orc_union16 var86; + orc_int8 var87; + orc_int8 var88; + orc_int8 var89; + orc_union16 var90; + orc_union16 var91; + orc_union32 var92; + + for (j = 0; j < m; j++) { + ptr0 = ORC_PTR_OFFSET (d1, d1_stride * j); + ptr4 = ORC_PTR_OFFSET (s1, s1_stride * j); + + /* 1: loadpb */ + var49.x4[0] = 0x00000080; /* 128 or 6.32404e-322f */ + var49.x4[1] = 0x00000080; /* 128 or 6.32404e-322f */ + var49.x4[2] = 0x00000080; /* 128 or 6.32404e-322f */ + var49.x4[3] = 0x00000080; /* 128 or 6.32404e-322f */ + /* 9: loadpw */ + var50.i = 0x0000002a; /* 42 or 2.07508e-322f */ + /* 14: loadpw */ + var51.i = 0x00000067; /* 103 or 5.08888e-322f */ + /* 21: loadpw */ + var52.i = 0x00000004; /* 4 or 1.97626e-323f */ + /* 25: loadpw */ + var53.i = 0x00000064; /* 100 or 4.94066e-322f */ + /* 29: loadpw */ + var54.i = 0x00000068; /* 104 or 5.13828e-322f */ + /* 40: loadpb */ + var55.x4[0] = 0x00000080; /* 128 or 6.32404e-322f */ + var55.x4[1] = 0x00000080; /* 128 or 6.32404e-322f */ + var55.x4[2] = 0x00000080; /* 128 or 6.32404e-322f */ + var55.x4[3] = 0x00000080; /* 128 or 6.32404e-322f */ + + for (i = 0; i < n; i++) { + /* 0: loadl */ + var48 = ptr4[i]; + /* 2: subb */ + var57.x4[0] = var48.x4[0] - var49.x4[0]; + var57.x4[1] = var48.x4[1] - var49.x4[1]; + var57.x4[2] = var48.x4[2] - var49.x4[2]; + var57.x4[3] = var48.x4[3] - var49.x4[3]; + /* 3: splitlw */ + var58.i = (var57.i >> 16) & 0xffff; + var59.i = var57.i & 0xffff; + /* 4: splitwb */ + var60 = (var59.i >> 8) & 0xff; + var61 = var59.i & 0xff; + /* 5: splitwb */ + var62 = (var58.i >> 8) & 0xff; + var63 = var58.i & 0xff; + /* 6: convsbw */ + var64.i = var60; + /* 7: convsbw */ + var65.i = var63; + /* 8: convsbw */ + var66.i = var62; + /* 10: mullw */ + var67.i = (var64.i * var50.i) & 0xffff; + /* 11: shrsw */ + var68.i = var67.i >> 8; + /* 12: addssw */ + var69.i = ORC_CLAMP_SW (var64.i + var68.i); + /* 13: addssw */ + var70.i = ORC_CLAMP_SW (var69.i + var66.i); + /* 15: mullw */ + var71.i = (var66.i * var51.i) & 0xffff; + /* 16: shrsw */ + var72.i = var71.i >> 8; + /* 17: subssw */ + var73.i = ORC_CLAMP_SW (var70.i - var72.i); + /* 18: addssw */ + var74.i = ORC_CLAMP_SW (var73.i + var66.i); + /* 19: addssw */ + var75.i = ORC_CLAMP_SW (var69.i + var65.i); + /* 20: addssw */ + var76.i = ORC_CLAMP_SW (var75.i + var65.i); + /* 22: mullw */ + var77.i = (var65.i * var52.i) & 0xffff; + /* 23: shrsw */ + var78.i = var77.i >> 8; + /* 24: addssw */ + var79.i = ORC_CLAMP_SW (var76.i + var78.i); + /* 26: mullw */ + var80.i = (var65.i * var53.i) & 0xffff; + /* 27: shrsw */ + var81.i = var80.i >> 8; + /* 28: subssw */ + var82.i = ORC_CLAMP_SW (var69.i - var81.i); + /* 30: mullw */ + var83.i = (var66.i * var54.i) & 0xffff; + /* 31: shrsw */ + var84.i = var83.i >> 8; + /* 32: subssw */ + var85.i = ORC_CLAMP_SW (var82.i - var84.i); + /* 33: subssw */ + var86.i = ORC_CLAMP_SW (var85.i - var84.i); + /* 34: convssswb */ + var87 = ORC_CLAMP_SB (var74.i); + /* 35: convssswb */ + var88 = ORC_CLAMP_SB (var86.i); + /* 36: convssswb */ + var89 = ORC_CLAMP_SB (var79.i); + /* 37: mergebw */ + var90.i = ((orc_uint8) var87) | ((orc_uint8) var88 << 8); + /* 38: mergebw */ + var91.i = ((orc_uint8) var89) | ((orc_uint8) var61 << 8); + /* 39: mergewl */ + var92.i = ((orc_uint16) var90.i) | ((orc_uint16) var91.i << 16); + /* 41: addb */ + var56.x4[0] = var92.x4[0] + var55.x4[0]; + var56.x4[1] = var92.x4[1] + var55.x4[1]; + var56.x4[2] = var92.x4[2] + var55.x4[2]; + var56.x4[3] = var92.x4[3] + var55.x4[3]; + /* 42: storel */ + ptr0[i] = var56; + } + } + +} + +#else +static void +_backup_cogorc_convert_AYUV_RGBA (OrcExecutor * ex) +{ + int i; + int j; + int n = ex->n; + int m = ex->params[ORC_VAR_A1]; + orc_union32 *ORC_RESTRICT ptr0; + const orc_union32 *ORC_RESTRICT ptr4; + orc_union32 var48; + orc_union32 var49; + orc_union16 var50; + orc_union16 var51; + orc_union16 var52; + orc_union16 var53; + orc_union16 var54; + orc_union32 var55; + orc_union32 var56; + orc_union32 var57; + orc_union16 var58; + orc_union16 var59; + orc_int8 var60; + orc_int8 var61; + orc_int8 var62; + orc_int8 var63; + orc_union16 var64; + orc_union16 var65; + orc_union16 var66; + orc_union16 var67; + orc_union16 var68; + orc_union16 var69; + orc_union16 var70; + orc_union16 var71; + orc_union16 var72; + orc_union16 var73; + orc_union16 var74; + orc_union16 var75; + orc_union16 var76; + orc_union16 var77; + orc_union16 var78; + orc_union16 var79; + orc_union16 var80; + orc_union16 var81; + orc_union16 var82; + orc_union16 var83; + orc_union16 var84; + orc_union16 var85; + orc_union16 var86; + orc_int8 var87; + orc_int8 var88; + orc_int8 var89; + orc_union16 var90; + orc_union16 var91; + orc_union32 var92; + + for (j = 0; j < m; j++) { + ptr0 = ORC_PTR_OFFSET (ex->arrays[0], ex->params[0] * j); + ptr4 = ORC_PTR_OFFSET (ex->arrays[4], ex->params[4] * j); + + /* 1: loadpb */ + var49.x4[0] = 0x00000080; /* 128 or 6.32404e-322f */ + var49.x4[1] = 0x00000080; /* 128 or 6.32404e-322f */ + var49.x4[2] = 0x00000080; /* 128 or 6.32404e-322f */ + var49.x4[3] = 0x00000080; /* 128 or 6.32404e-322f */ + /* 9: loadpw */ + var50.i = 0x0000002a; /* 42 or 2.07508e-322f */ + /* 14: loadpw */ + var51.i = 0x00000067; /* 103 or 5.08888e-322f */ + /* 21: loadpw */ + var52.i = 0x00000004; /* 4 or 1.97626e-323f */ + /* 25: loadpw */ + var53.i = 0x00000064; /* 100 or 4.94066e-322f */ + /* 29: loadpw */ + var54.i = 0x00000068; /* 104 or 5.13828e-322f */ + /* 40: loadpb */ + var55.x4[0] = 0x00000080; /* 128 or 6.32404e-322f */ + var55.x4[1] = 0x00000080; /* 128 or 6.32404e-322f */ + var55.x4[2] = 0x00000080; /* 128 or 6.32404e-322f */ + var55.x4[3] = 0x00000080; /* 128 or 6.32404e-322f */ + + for (i = 0; i < n; i++) { + /* 0: loadl */ + var48 = ptr4[i]; + /* 2: subb */ + var57.x4[0] = var48.x4[0] - var49.x4[0]; + var57.x4[1] = var48.x4[1] - var49.x4[1]; + var57.x4[2] = var48.x4[2] - var49.x4[2]; + var57.x4[3] = var48.x4[3] - var49.x4[3]; + /* 3: splitlw */ + var58.i = (var57.i >> 16) & 0xffff; + var59.i = var57.i & 0xffff; + /* 4: splitwb */ + var60 = (var59.i >> 8) & 0xff; + var61 = var59.i & 0xff; + /* 5: splitwb */ + var62 = (var58.i >> 8) & 0xff; + var63 = var58.i & 0xff; + /* 6: convsbw */ + var64.i = var60; + /* 7: convsbw */ + var65.i = var63; + /* 8: convsbw */ + var66.i = var62; + /* 10: mullw */ + var67.i = (var64.i * var50.i) & 0xffff; + /* 11: shrsw */ + var68.i = var67.i >> 8; + /* 12: addssw */ + var69.i = ORC_CLAMP_SW (var64.i + var68.i); + /* 13: addssw */ + var70.i = ORC_CLAMP_SW (var69.i + var66.i); + /* 15: mullw */ + var71.i = (var66.i * var51.i) & 0xffff; + /* 16: shrsw */ + var72.i = var71.i >> 8; + /* 17: subssw */ + var73.i = ORC_CLAMP_SW (var70.i - var72.i); + /* 18: addssw */ + var74.i = ORC_CLAMP_SW (var73.i + var66.i); + /* 19: addssw */ + var75.i = ORC_CLAMP_SW (var69.i + var65.i); + /* 20: addssw */ + var76.i = ORC_CLAMP_SW (var75.i + var65.i); + /* 22: mullw */ + var77.i = (var65.i * var52.i) & 0xffff; + /* 23: shrsw */ + var78.i = var77.i >> 8; + /* 24: addssw */ + var79.i = ORC_CLAMP_SW (var76.i + var78.i); + /* 26: mullw */ + var80.i = (var65.i * var53.i) & 0xffff; + /* 27: shrsw */ + var81.i = var80.i >> 8; + /* 28: subssw */ + var82.i = ORC_CLAMP_SW (var69.i - var81.i); + /* 30: mullw */ + var83.i = (var66.i * var54.i) & 0xffff; + /* 31: shrsw */ + var84.i = var83.i >> 8; + /* 32: subssw */ + var85.i = ORC_CLAMP_SW (var82.i - var84.i); + /* 33: subssw */ + var86.i = ORC_CLAMP_SW (var85.i - var84.i); + /* 34: convssswb */ + var87 = ORC_CLAMP_SB (var74.i); + /* 35: convssswb */ + var88 = ORC_CLAMP_SB (var86.i); + /* 36: convssswb */ + var89 = ORC_CLAMP_SB (var79.i); + /* 37: mergebw */ + var90.i = ((orc_uint8) var87) | ((orc_uint8) var88 << 8); + /* 38: mergebw */ + var91.i = ((orc_uint8) var89) | ((orc_uint8) var61 << 8); + /* 39: mergewl */ + var92.i = ((orc_uint16) var90.i) | ((orc_uint16) var91.i << 16); + /* 41: addb */ + var56.x4[0] = var92.x4[0] + var55.x4[0]; + var56.x4[1] = var92.x4[1] + var55.x4[1]; + var56.x4[2] = var92.x4[2] + var55.x4[2]; + var56.x4[3] = var92.x4[3] + var55.x4[3]; + /* 42: storel */ + ptr0[i] = var56; + } + } + +} + +void +cogorc_convert_AYUV_RGBA (guint8 * d1, int d1_stride, const guint8 * s1, + int s1_stride, int n, int m) +{ + OrcExecutor _ex, *ex = &_ex; + static int p_inited = 0; + static OrcProgram *p = 0; + void (*func) (OrcExecutor *); + + if (!p_inited) { + orc_once_mutex_lock (); + if (!p_inited) { + OrcCompileResult result; + + p = orc_program_new (); + orc_program_set_2d (p); + orc_program_set_name (p, "cogorc_convert_AYUV_RGBA"); + orc_program_set_backup_function (p, _backup_cogorc_convert_AYUV_RGBA); + orc_program_add_destination (p, 4, "d1"); + orc_program_add_source (p, 4, "s1"); + orc_program_add_constant (p, 1, 0x00000008, "c1"); + orc_program_add_constant (p, 4, 0x00000080, "c2"); + orc_program_add_constant (p, 4, 0x0000002a, "c3"); + orc_program_add_constant (p, 4, 0x00000067, "c4"); + orc_program_add_constant (p, 4, 0x00000004, "c5"); + orc_program_add_constant (p, 4, 0x00000064, "c6"); + orc_program_add_constant (p, 4, 0x00000068, "c7"); + orc_program_add_constant (p, 0, 0x00000080, "c8"); + orc_program_add_temporary (p, 2, "t1"); + orc_program_add_temporary (p, 2, "t2"); + orc_program_add_temporary (p, 1, "t3"); + orc_program_add_temporary (p, 1, "t4"); + orc_program_add_temporary (p, 1, "t5"); + orc_program_add_temporary (p, 1, "t6"); + orc_program_add_temporary (p, 2, "t7"); + orc_program_add_temporary (p, 2, "t8"); + orc_program_add_temporary (p, 2, "t9"); + orc_program_add_temporary (p, 2, "t10"); + orc_program_add_temporary (p, 2, "t11"); + orc_program_add_temporary (p, 2, "t12"); + orc_program_add_temporary (p, 1, "t13"); + orc_program_add_temporary (p, 1, "t14"); + orc_program_add_temporary (p, 1, "t15"); + orc_program_add_temporary (p, 4, "t16"); + + orc_program_append_2 (p, "subb", 2, 47, ORC_VAR_S1, ORC_VAR_C2, + ORC_VAR_D1); + orc_program_append_2 (p, "splitlw", 0, ORC_VAR_T1, ORC_VAR_T2, 47, + ORC_VAR_D1); + orc_program_append_2 (p, "splitwb", 0, ORC_VAR_T4, ORC_VAR_T3, ORC_VAR_T2, + ORC_VAR_D1); + orc_program_append_2 (p, "splitwb", 0, ORC_VAR_T6, ORC_VAR_T5, ORC_VAR_T1, + ORC_VAR_D1); + orc_program_append_2 (p, "convsbw", 0, ORC_VAR_T7, ORC_VAR_T4, ORC_VAR_D1, + ORC_VAR_D1); + orc_program_append_2 (p, "convsbw", 0, ORC_VAR_T8, ORC_VAR_T5, ORC_VAR_D1, + ORC_VAR_D1); + orc_program_append_2 (p, "convsbw", 0, ORC_VAR_T9, ORC_VAR_T6, ORC_VAR_D1, + ORC_VAR_D1); + orc_program_append_2 (p, "mullw", 0, ORC_VAR_T1, ORC_VAR_T7, ORC_VAR_C3, + ORC_VAR_D1); + orc_program_append_2 (p, "shrsw", 0, ORC_VAR_T1, ORC_VAR_T1, ORC_VAR_C1, + ORC_VAR_D1); + orc_program_append_2 (p, "addssw", 0, ORC_VAR_T7, ORC_VAR_T7, ORC_VAR_T1, + ORC_VAR_D1); + orc_program_append_2 (p, "addssw", 0, ORC_VAR_T10, ORC_VAR_T7, ORC_VAR_T9, + ORC_VAR_D1); + orc_program_append_2 (p, "mullw", 0, ORC_VAR_T1, ORC_VAR_T9, ORC_VAR_C4, + ORC_VAR_D1); + orc_program_append_2 (p, "shrsw", 0, ORC_VAR_T1, ORC_VAR_T1, ORC_VAR_C1, + ORC_VAR_D1); + orc_program_append_2 (p, "subssw", 0, ORC_VAR_T10, ORC_VAR_T10, + ORC_VAR_T1, ORC_VAR_D1); + orc_program_append_2 (p, "addssw", 0, ORC_VAR_T10, ORC_VAR_T10, + ORC_VAR_T9, ORC_VAR_D1); + orc_program_append_2 (p, "addssw", 0, ORC_VAR_T12, ORC_VAR_T7, ORC_VAR_T8, + ORC_VAR_D1); + orc_program_append_2 (p, "addssw", 0, ORC_VAR_T12, ORC_VAR_T12, + ORC_VAR_T8, ORC_VAR_D1); + orc_program_append_2 (p, "mullw", 0, ORC_VAR_T1, ORC_VAR_T8, ORC_VAR_C5, + ORC_VAR_D1); + orc_program_append_2 (p, "shrsw", 0, ORC_VAR_T1, ORC_VAR_T1, ORC_VAR_C1, + ORC_VAR_D1); + orc_program_append_2 (p, "addssw", 0, ORC_VAR_T12, ORC_VAR_T12, + ORC_VAR_T1, ORC_VAR_D1); + orc_program_append_2 (p, "mullw", 0, ORC_VAR_T1, ORC_VAR_T8, ORC_VAR_C6, + ORC_VAR_D1); + orc_program_append_2 (p, "shrsw", 0, ORC_VAR_T1, ORC_VAR_T1, ORC_VAR_C1, + ORC_VAR_D1); + orc_program_append_2 (p, "subssw", 0, ORC_VAR_T11, ORC_VAR_T7, ORC_VAR_T1, + ORC_VAR_D1); + orc_program_append_2 (p, "mullw", 0, ORC_VAR_T1, ORC_VAR_T9, ORC_VAR_C7, + ORC_VAR_D1); + orc_program_append_2 (p, "shrsw", 0, ORC_VAR_T1, ORC_VAR_T1, ORC_VAR_C1, + ORC_VAR_D1); + orc_program_append_2 (p, "subssw", 0, ORC_VAR_T11, ORC_VAR_T11, + ORC_VAR_T1, ORC_VAR_D1); + orc_program_append_2 (p, "subssw", 0, ORC_VAR_T11, ORC_VAR_T11, + ORC_VAR_T1, ORC_VAR_D1); + orc_program_append_2 (p, "convssswb", 0, ORC_VAR_T13, ORC_VAR_T10, + ORC_VAR_D1, ORC_VAR_D1); + orc_program_append_2 (p, "convssswb", 0, ORC_VAR_T14, ORC_VAR_T11, + ORC_VAR_D1, ORC_VAR_D1); + orc_program_append_2 (p, "convssswb", 0, ORC_VAR_T15, ORC_VAR_T12, + ORC_VAR_D1, ORC_VAR_D1); + orc_program_append_2 (p, "mergebw", 0, ORC_VAR_T1, ORC_VAR_T13, + ORC_VAR_T14, ORC_VAR_D1); + orc_program_append_2 (p, "mergebw", 0, ORC_VAR_T2, ORC_VAR_T15, + ORC_VAR_T3, ORC_VAR_D1); + orc_program_append_2 (p, "mergewl", 0, 47, ORC_VAR_T1, ORC_VAR_T2, + ORC_VAR_D1); + orc_program_append_2 (p, "addb", 2, ORC_VAR_D1, 47, ORC_VAR_C2, + ORC_VAR_D1); + + result = orc_program_compile (p); + } + p_inited = TRUE; + orc_once_mutex_unlock (); + } + ex->program = p; + + ex->n = n; + ORC_EXECUTOR_M (ex) = m; + ex->arrays[ORC_VAR_D1] = d1; + ex->params[ORC_VAR_D1] = d1_stride; + ex->arrays[ORC_VAR_S1] = (void *) s1; + ex->params[ORC_VAR_S1] = s1_stride; + + func = p->code_exec; + func (ex); +} +#endif + + +/* cogorc_convert_I420_BGRA */ +#ifdef DISABLE_ORC +void +cogorc_convert_I420_BGRA (guint8 * d1, const guint8 * s1, const guint8 * s2, + const guint8 * s3, int n) +{ + int i; + orc_union32 *ORC_RESTRICT ptr0; + const orc_int8 *ORC_RESTRICT ptr4; + const orc_int8 *ORC_RESTRICT ptr5; + const orc_int8 *ORC_RESTRICT ptr6; + orc_int8 var45; + orc_int8 var46; + orc_int8 var47; + orc_int8 var48; + orc_union16 var49; + orc_union16 var50; + orc_union16 var51; + orc_union16 var52; + orc_union16 var53; + orc_int8 var54; + orc_union32 var55; + orc_union32 var56; + orc_int8 var57; + orc_union16 var58; + orc_int8 var59; + orc_int8 var60; + orc_union16 var61; + orc_int8 var62; + orc_int8 var63; + orc_union16 var64; + orc_union16 var65; + orc_union16 var66; + orc_union16 var67; + orc_union16 var68; + orc_union16 var69; + orc_union16 var70; + orc_union16 var71; + orc_union16 var72; + orc_union16 var73; + orc_union16 var74; + orc_union16 var75; + orc_union16 var76; + orc_union16 var77; + orc_union16 var78; + orc_union16 var79; + orc_union16 var80; + orc_union16 var81; + orc_union16 var82; + orc_union16 var83; + orc_union16 var84; + orc_int8 var85; + orc_int8 var86; + orc_int8 var87; + orc_union16 var88; + orc_union16 var89; + orc_union32 var90; + + ptr0 = (orc_union32 *) d1; + ptr4 = (orc_int8 *) s1; + ptr5 = (orc_int8 *) s2; + ptr6 = (orc_int8 *) s3; + + /* 1: loadpb */ + var46 = 0x00000080; /* 128 or 6.32404e-322f */ + /* 5: loadpb */ + var47 = 0x00000080; /* 128 or 6.32404e-322f */ + /* 9: loadpb */ + var48 = 0x00000080; /* 128 or 6.32404e-322f */ + /* 12: loadpw */ + var49.i = 0x0000002a; /* 42 or 2.07508e-322f */ + /* 17: loadpw */ + var50.i = 0x00000067; /* 103 or 5.08888e-322f */ + /* 24: loadpw */ + var51.i = 0x00000004; /* 4 or 1.97626e-323f */ + /* 28: loadpw */ + var52.i = 0x00000064; /* 100 or 4.94066e-322f */ + /* 32: loadpw */ + var53.i = 0x00000068; /* 104 or 5.13828e-322f */ + /* 41: loadpb */ + var54 = 0x000000ff; /* 255 or 1.25987e-321f */ + /* 44: loadpb */ + var55.x4[0] = 0x00000080; /* 128 or 6.32404e-322f */ + var55.x4[1] = 0x00000080; /* 128 or 6.32404e-322f */ + var55.x4[2] = 0x00000080; /* 128 or 6.32404e-322f */ + var55.x4[3] = 0x00000080; /* 128 or 6.32404e-322f */ + + for (i = 0; i < n; i++) { + /* 0: loadb */ + var45 = ptr4[i]; + /* 2: subb */ + var57 = var45 - var46; + /* 3: convsbw */ + var58.i = var57; + /* 4: loadupib */ + var59 = + (i & 1) ? ((orc_uint8) ptr5[i >> 1] + (orc_uint8) ptr5[(i >> 1) + 1] + + 1) >> 1 : ptr5[i >> 1]; + /* 6: subb */ + var60 = var59 - var47; + /* 7: convsbw */ + var61.i = var60; + /* 8: loadupib */ + var62 = + (i & 1) ? ((orc_uint8) ptr6[i >> 1] + (orc_uint8) ptr6[(i >> 1) + 1] + + 1) >> 1 : ptr6[i >> 1]; + /* 10: subb */ + var63 = var62 - var48; + /* 11: convsbw */ + var64.i = var63; + /* 13: mullw */ + var65.i = (var58.i * var49.i) & 0xffff; + /* 14: shrsw */ + var66.i = var65.i >> 8; + /* 15: addssw */ + var67.i = ORC_CLAMP_SW (var58.i + var66.i); + /* 16: addssw */ + var68.i = ORC_CLAMP_SW (var67.i + var64.i); + /* 18: mullw */ + var69.i = (var64.i * var50.i) & 0xffff; + /* 19: shrsw */ + var70.i = var69.i >> 8; + /* 20: subssw */ + var71.i = ORC_CLAMP_SW (var68.i - var70.i); + /* 21: addssw */ + var72.i = ORC_CLAMP_SW (var71.i + var64.i); + /* 22: addssw */ + var73.i = ORC_CLAMP_SW (var67.i + var61.i); + /* 23: addssw */ + var74.i = ORC_CLAMP_SW (var73.i + var61.i); + /* 25: mullw */ + var75.i = (var61.i * var51.i) & 0xffff; + /* 26: shrsw */ + var76.i = var75.i >> 8; + /* 27: addssw */ + var77.i = ORC_CLAMP_SW (var74.i + var76.i); + /* 29: mullw */ + var78.i = (var61.i * var52.i) & 0xffff; + /* 30: shrsw */ + var79.i = var78.i >> 8; + /* 31: subssw */ + var80.i = ORC_CLAMP_SW (var67.i - var79.i); + /* 33: mullw */ + var81.i = (var64.i * var53.i) & 0xffff; + /* 34: shrsw */ + var82.i = var81.i >> 8; + /* 35: subssw */ + var83.i = ORC_CLAMP_SW (var80.i - var82.i); + /* 36: subssw */ + var84.i = ORC_CLAMP_SW (var83.i - var82.i); + /* 37: convssswb */ + var85 = ORC_CLAMP_SB (var72.i); + /* 38: convssswb */ + var86 = ORC_CLAMP_SB (var84.i); + /* 39: convssswb */ + var87 = ORC_CLAMP_SB (var77.i); + /* 40: mergebw */ + var88.i = ((orc_uint8) var87) | ((orc_uint8) var86 << 8); + /* 42: mergebw */ + var89.i = ((orc_uint8) var85) | ((orc_uint8) var54 << 8); + /* 43: mergewl */ + var90.i = ((orc_uint16) var88.i) | ((orc_uint16) var89.i << 16); + /* 45: addb */ + var56.x4[0] = var90.x4[0] + var55.x4[0]; + var56.x4[1] = var90.x4[1] + var55.x4[1]; + var56.x4[2] = var90.x4[2] + var55.x4[2]; + var56.x4[3] = var90.x4[3] + var55.x4[3]; + /* 46: storel */ + ptr0[i] = var56; + } + +} + +#else +static void +_backup_cogorc_convert_I420_BGRA (OrcExecutor * ex) +{ + int i; + int n = ex->n; + orc_union32 *ORC_RESTRICT ptr0; + const orc_int8 *ORC_RESTRICT ptr4; + const orc_int8 *ORC_RESTRICT ptr5; + const orc_int8 *ORC_RESTRICT ptr6; + orc_int8 var45; + orc_int8 var46; + orc_int8 var47; + orc_int8 var48; + orc_union16 var49; + orc_union16 var50; + orc_union16 var51; + orc_union16 var52; + orc_union16 var53; + orc_int8 var54; + orc_union32 var55; + orc_union32 var56; + orc_int8 var57; + orc_union16 var58; + orc_int8 var59; + orc_int8 var60; + orc_union16 var61; + orc_int8 var62; + orc_int8 var63; + orc_union16 var64; + orc_union16 var65; + orc_union16 var66; + orc_union16 var67; + orc_union16 var68; + orc_union16 var69; + orc_union16 var70; + orc_union16 var71; + orc_union16 var72; + orc_union16 var73; + orc_union16 var74; + orc_union16 var75; + orc_union16 var76; + orc_union16 var77; + orc_union16 var78; + orc_union16 var79; + orc_union16 var80; + orc_union16 var81; + orc_union16 var82; + orc_union16 var83; + orc_union16 var84; + orc_int8 var85; + orc_int8 var86; + orc_int8 var87; + orc_union16 var88; + orc_union16 var89; + orc_union32 var90; + + ptr0 = (orc_union32 *) ex->arrays[0]; + ptr4 = (orc_int8 *) ex->arrays[4]; + ptr5 = (orc_int8 *) ex->arrays[5]; + ptr6 = (orc_int8 *) ex->arrays[6]; + + /* 1: loadpb */ + var46 = 0x00000080; /* 128 or 6.32404e-322f */ + /* 5: loadpb */ + var47 = 0x00000080; /* 128 or 6.32404e-322f */ + /* 9: loadpb */ + var48 = 0x00000080; /* 128 or 6.32404e-322f */ + /* 12: loadpw */ + var49.i = 0x0000002a; /* 42 or 2.07508e-322f */ + /* 17: loadpw */ + var50.i = 0x00000067; /* 103 or 5.08888e-322f */ + /* 24: loadpw */ + var51.i = 0x00000004; /* 4 or 1.97626e-323f */ + /* 28: loadpw */ + var52.i = 0x00000064; /* 100 or 4.94066e-322f */ + /* 32: loadpw */ + var53.i = 0x00000068; /* 104 or 5.13828e-322f */ + /* 41: loadpb */ + var54 = 0x000000ff; /* 255 or 1.25987e-321f */ + /* 44: loadpb */ + var55.x4[0] = 0x00000080; /* 128 or 6.32404e-322f */ + var55.x4[1] = 0x00000080; /* 128 or 6.32404e-322f */ + var55.x4[2] = 0x00000080; /* 128 or 6.32404e-322f */ + var55.x4[3] = 0x00000080; /* 128 or 6.32404e-322f */ + + for (i = 0; i < n; i++) { + /* 0: loadb */ + var45 = ptr4[i]; + /* 2: subb */ + var57 = var45 - var46; + /* 3: convsbw */ + var58.i = var57; + /* 4: loadupib */ + var59 = + (i & 1) ? ((orc_uint8) ptr5[i >> 1] + (orc_uint8) ptr5[(i >> 1) + 1] + + 1) >> 1 : ptr5[i >> 1]; + /* 6: subb */ + var60 = var59 - var47; + /* 7: convsbw */ + var61.i = var60; + /* 8: loadupib */ + var62 = + (i & 1) ? ((orc_uint8) ptr6[i >> 1] + (orc_uint8) ptr6[(i >> 1) + 1] + + 1) >> 1 : ptr6[i >> 1]; + /* 10: subb */ + var63 = var62 - var48; + /* 11: convsbw */ + var64.i = var63; + /* 13: mullw */ + var65.i = (var58.i * var49.i) & 0xffff; + /* 14: shrsw */ + var66.i = var65.i >> 8; + /* 15: addssw */ + var67.i = ORC_CLAMP_SW (var58.i + var66.i); + /* 16: addssw */ + var68.i = ORC_CLAMP_SW (var67.i + var64.i); + /* 18: mullw */ + var69.i = (var64.i * var50.i) & 0xffff; + /* 19: shrsw */ + var70.i = var69.i >> 8; + /* 20: subssw */ + var71.i = ORC_CLAMP_SW (var68.i - var70.i); + /* 21: addssw */ + var72.i = ORC_CLAMP_SW (var71.i + var64.i); + /* 22: addssw */ + var73.i = ORC_CLAMP_SW (var67.i + var61.i); + /* 23: addssw */ + var74.i = ORC_CLAMP_SW (var73.i + var61.i); + /* 25: mullw */ + var75.i = (var61.i * var51.i) & 0xffff; + /* 26: shrsw */ + var76.i = var75.i >> 8; + /* 27: addssw */ + var77.i = ORC_CLAMP_SW (var74.i + var76.i); + /* 29: mullw */ + var78.i = (var61.i * var52.i) & 0xffff; + /* 30: shrsw */ + var79.i = var78.i >> 8; + /* 31: subssw */ + var80.i = ORC_CLAMP_SW (var67.i - var79.i); + /* 33: mullw */ + var81.i = (var64.i * var53.i) & 0xffff; + /* 34: shrsw */ + var82.i = var81.i >> 8; + /* 35: subssw */ + var83.i = ORC_CLAMP_SW (var80.i - var82.i); + /* 36: subssw */ + var84.i = ORC_CLAMP_SW (var83.i - var82.i); + /* 37: convssswb */ + var85 = ORC_CLAMP_SB (var72.i); + /* 38: convssswb */ + var86 = ORC_CLAMP_SB (var84.i); + /* 39: convssswb */ + var87 = ORC_CLAMP_SB (var77.i); + /* 40: mergebw */ + var88.i = ((orc_uint8) var87) | ((orc_uint8) var86 << 8); + /* 42: mergebw */ + var89.i = ((orc_uint8) var85) | ((orc_uint8) var54 << 8); + /* 43: mergewl */ + var90.i = ((orc_uint16) var88.i) | ((orc_uint16) var89.i << 16); + /* 45: addb */ + var56.x4[0] = var90.x4[0] + var55.x4[0]; + var56.x4[1] = var90.x4[1] + var55.x4[1]; + var56.x4[2] = var90.x4[2] + var55.x4[2]; + var56.x4[3] = var90.x4[3] + var55.x4[3]; + /* 46: storel */ + ptr0[i] = var56; + } + +} + +void +cogorc_convert_I420_BGRA (guint8 * d1, const guint8 * s1, const guint8 * s2, + const guint8 * s3, int n) +{ + OrcExecutor _ex, *ex = &_ex; + static int p_inited = 0; + static OrcProgram *p = 0; + void (*func) (OrcExecutor *); + + if (!p_inited) { + orc_once_mutex_lock (); + if (!p_inited) { + OrcCompileResult result; + + p = orc_program_new (); + orc_program_set_name (p, "cogorc_convert_I420_BGRA"); + orc_program_set_backup_function (p, _backup_cogorc_convert_I420_BGRA); + orc_program_add_destination (p, 4, "d1"); + orc_program_add_source (p, 1, "s1"); + orc_program_add_source (p, 1, "s2"); + orc_program_add_source (p, 1, "s3"); + orc_program_add_constant (p, 1, 0x00000008, "c1"); + orc_program_add_constant (p, 1, 0x00000080, "c2"); + orc_program_add_constant (p, 4, 0x0000002a, "c3"); + orc_program_add_constant (p, 4, 0x00000067, "c4"); + orc_program_add_constant (p, 4, 0x00000004, "c5"); + orc_program_add_constant (p, 4, 0x00000064, "c6"); + orc_program_add_constant (p, 4, 0x00000068, "c7"); + orc_program_add_constant (p, 4, 0x000000ff, "c8"); + orc_program_add_temporary (p, 2, "t1"); + orc_program_add_temporary (p, 2, "t2"); + orc_program_add_temporary (p, 1, "t3"); + orc_program_add_temporary (p, 2, "t4"); + orc_program_add_temporary (p, 2, "t5"); + orc_program_add_temporary (p, 2, "t6"); + orc_program_add_temporary (p, 2, "t7"); + orc_program_add_temporary (p, 2, "t8"); + orc_program_add_temporary (p, 2, "t9"); + orc_program_add_temporary (p, 1, "t10"); + orc_program_add_temporary (p, 1, "t11"); + orc_program_add_temporary (p, 1, "t12"); + orc_program_add_temporary (p, 4, "t13"); + + orc_program_append_2 (p, "subb", 0, ORC_VAR_T3, ORC_VAR_S1, ORC_VAR_C2, + ORC_VAR_D1); + orc_program_append_2 (p, "convsbw", 0, ORC_VAR_T4, ORC_VAR_T3, ORC_VAR_D1, + ORC_VAR_D1); + orc_program_append_2 (p, "loadupib", 0, ORC_VAR_T3, ORC_VAR_S2, + ORC_VAR_D1, ORC_VAR_D1); + orc_program_append_2 (p, "subb", 0, ORC_VAR_T3, ORC_VAR_T3, ORC_VAR_C2, + ORC_VAR_D1); + orc_program_append_2 (p, "convsbw", 0, ORC_VAR_T5, ORC_VAR_T3, ORC_VAR_D1, + ORC_VAR_D1); + orc_program_append_2 (p, "loadupib", 0, ORC_VAR_T3, ORC_VAR_S3, + ORC_VAR_D1, ORC_VAR_D1); + orc_program_append_2 (p, "subb", 0, ORC_VAR_T3, ORC_VAR_T3, ORC_VAR_C2, + ORC_VAR_D1); + orc_program_append_2 (p, "convsbw", 0, ORC_VAR_T6, ORC_VAR_T3, ORC_VAR_D1, + ORC_VAR_D1); + orc_program_append_2 (p, "mullw", 0, ORC_VAR_T1, ORC_VAR_T4, ORC_VAR_C3, + ORC_VAR_D1); + orc_program_append_2 (p, "shrsw", 0, ORC_VAR_T1, ORC_VAR_T1, ORC_VAR_C1, + ORC_VAR_D1); + orc_program_append_2 (p, "addssw", 0, ORC_VAR_T4, ORC_VAR_T4, ORC_VAR_T1, + ORC_VAR_D1); + orc_program_append_2 (p, "addssw", 0, ORC_VAR_T7, ORC_VAR_T4, ORC_VAR_T6, + ORC_VAR_D1); + orc_program_append_2 (p, "mullw", 0, ORC_VAR_T1, ORC_VAR_T6, ORC_VAR_C4, + ORC_VAR_D1); + orc_program_append_2 (p, "shrsw", 0, ORC_VAR_T1, ORC_VAR_T1, ORC_VAR_C1, + ORC_VAR_D1); + orc_program_append_2 (p, "subssw", 0, ORC_VAR_T7, ORC_VAR_T7, ORC_VAR_T1, + ORC_VAR_D1); + orc_program_append_2 (p, "addssw", 0, ORC_VAR_T7, ORC_VAR_T7, ORC_VAR_T6, + ORC_VAR_D1); + orc_program_append_2 (p, "addssw", 0, ORC_VAR_T9, ORC_VAR_T4, ORC_VAR_T5, + ORC_VAR_D1); + orc_program_append_2 (p, "addssw", 0, ORC_VAR_T9, ORC_VAR_T9, ORC_VAR_T5, + ORC_VAR_D1); + orc_program_append_2 (p, "mullw", 0, ORC_VAR_T1, ORC_VAR_T5, ORC_VAR_C5, + ORC_VAR_D1); + orc_program_append_2 (p, "shrsw", 0, ORC_VAR_T1, ORC_VAR_T1, ORC_VAR_C1, + ORC_VAR_D1); + orc_program_append_2 (p, "addssw", 0, ORC_VAR_T9, ORC_VAR_T9, ORC_VAR_T1, + ORC_VAR_D1); + orc_program_append_2 (p, "mullw", 0, ORC_VAR_T1, ORC_VAR_T5, ORC_VAR_C6, + ORC_VAR_D1); + orc_program_append_2 (p, "shrsw", 0, ORC_VAR_T1, ORC_VAR_T1, ORC_VAR_C1, + ORC_VAR_D1); + orc_program_append_2 (p, "subssw", 0, ORC_VAR_T8, ORC_VAR_T4, ORC_VAR_T1, + ORC_VAR_D1); + orc_program_append_2 (p, "mullw", 0, ORC_VAR_T1, ORC_VAR_T6, ORC_VAR_C7, + ORC_VAR_D1); + orc_program_append_2 (p, "shrsw", 0, ORC_VAR_T1, ORC_VAR_T1, ORC_VAR_C1, + ORC_VAR_D1); + orc_program_append_2 (p, "subssw", 0, ORC_VAR_T8, ORC_VAR_T8, ORC_VAR_T1, + ORC_VAR_D1); + orc_program_append_2 (p, "subssw", 0, ORC_VAR_T8, ORC_VAR_T8, ORC_VAR_T1, + ORC_VAR_D1); + orc_program_append_2 (p, "convssswb", 0, ORC_VAR_T10, ORC_VAR_T7, + ORC_VAR_D1, ORC_VAR_D1); + orc_program_append_2 (p, "convssswb", 0, ORC_VAR_T11, ORC_VAR_T8, + ORC_VAR_D1, ORC_VAR_D1); + orc_program_append_2 (p, "convssswb", 0, ORC_VAR_T12, ORC_VAR_T9, + ORC_VAR_D1, ORC_VAR_D1); + orc_program_append_2 (p, "mergebw", 0, ORC_VAR_T1, ORC_VAR_T12, + ORC_VAR_T11, ORC_VAR_D1); + orc_program_append_2 (p, "mergebw", 0, ORC_VAR_T2, ORC_VAR_T10, + ORC_VAR_C8, ORC_VAR_D1); + orc_program_append_2 (p, "mergewl", 0, ORC_VAR_T13, ORC_VAR_T1, + ORC_VAR_T2, ORC_VAR_D1); + orc_program_append_2 (p, "addb", 2, ORC_VAR_D1, ORC_VAR_T13, ORC_VAR_C2, + ORC_VAR_D1); + + result = orc_program_compile (p); + } + p_inited = TRUE; + orc_once_mutex_unlock (); + } + ex->program = p; + + ex->n = n; + ex->arrays[ORC_VAR_D1] = d1; + ex->arrays[ORC_VAR_S1] = (void *) s1; + ex->arrays[ORC_VAR_S2] = (void *) s2; + ex->arrays[ORC_VAR_S3] = (void *) s3; + + func = p->code_exec; + func (ex); +} +#endif + + +/* cogorc_convert_I420_BGRA_avg */ +#ifdef DISABLE_ORC +void +cogorc_convert_I420_BGRA_avg (guint8 * d1, const guint8 * s1, const guint8 * s2, + const guint8 * s3, const guint8 * s4, const guint8 * s5, int n) +{ + int i; + orc_union32 *ORC_RESTRICT ptr0; + const orc_int8 *ORC_RESTRICT ptr4; + const orc_int8 *ORC_RESTRICT ptr5; + const orc_int8 *ORC_RESTRICT ptr6; + const orc_int8 *ORC_RESTRICT ptr7; + const orc_int8 *ORC_RESTRICT ptr8; + orc_int8 var46; + orc_int8 var47; + orc_int8 var48; + orc_int8 var49; + orc_union16 var50; + orc_union16 var51; + orc_union16 var52; + orc_union16 var53; + orc_union16 var54; + orc_int8 var55; + orc_union32 var56; + orc_union32 var57; + orc_int8 var58; + orc_union16 var59; + orc_int8 var60; + orc_int8 var61; + orc_int8 var62; + orc_int8 var63; + orc_union16 var64; + orc_int8 var65; + orc_int8 var66; + orc_int8 var67; + orc_int8 var68; + orc_union16 var69; + orc_union16 var70; + orc_union16 var71; + orc_union16 var72; + orc_union16 var73; + orc_union16 var74; + orc_union16 var75; + orc_union16 var76; + orc_union16 var77; + orc_union16 var78; + orc_union16 var79; + orc_union16 var80; + orc_union16 var81; + orc_union16 var82; + orc_union16 var83; + orc_union16 var84; + orc_union16 var85; + orc_union16 var86; + orc_union16 var87; + orc_union16 var88; + orc_union16 var89; + orc_int8 var90; + orc_int8 var91; + orc_int8 var92; + orc_union16 var93; + orc_union16 var94; + orc_union32 var95; + + ptr0 = (orc_union32 *) d1; + ptr4 = (orc_int8 *) s1; + ptr5 = (orc_int8 *) s2; + ptr6 = (orc_int8 *) s3; + ptr7 = (orc_int8 *) s4; + ptr8 = (orc_int8 *) s5; + + /* 1: loadpb */ + var47 = 0x00000080; /* 128 or 6.32404e-322f */ + /* 7: loadpb */ + var48 = 0x00000080; /* 128 or 6.32404e-322f */ + /* 13: loadpb */ + var49 = 0x00000080; /* 128 or 6.32404e-322f */ + /* 16: loadpw */ + var50.i = 0x0000002a; /* 42 or 2.07508e-322f */ + /* 21: loadpw */ + var51.i = 0x00000067; /* 103 or 5.08888e-322f */ + /* 28: loadpw */ + var52.i = 0x00000004; /* 4 or 1.97626e-323f */ + /* 32: loadpw */ + var53.i = 0x00000064; /* 100 or 4.94066e-322f */ + /* 36: loadpw */ + var54.i = 0x00000068; /* 104 or 5.13828e-322f */ + /* 45: loadpb */ + var55 = 0x000000ff; /* 255 or 1.25987e-321f */ + /* 48: loadpb */ + var56.x4[0] = 0x00000080; /* 128 or 6.32404e-322f */ + var56.x4[1] = 0x00000080; /* 128 or 6.32404e-322f */ + var56.x4[2] = 0x00000080; /* 128 or 6.32404e-322f */ + var56.x4[3] = 0x00000080; /* 128 or 6.32404e-322f */ + + for (i = 0; i < n; i++) { + /* 0: loadb */ + var46 = ptr4[i]; + /* 2: subb */ + var58 = var46 - var47; + /* 3: convsbw */ + var59.i = var58; + /* 4: loadupib */ + var60 = + (i & 1) ? ((orc_uint8) ptr5[i >> 1] + (orc_uint8) ptr5[(i >> 1) + 1] + + 1) >> 1 : ptr5[i >> 1]; + /* 5: loadupib */ + var61 = + (i & 1) ? ((orc_uint8) ptr6[i >> 1] + (orc_uint8) ptr6[(i >> 1) + 1] + + 1) >> 1 : ptr6[i >> 1]; + /* 6: avgub */ + var62 = ((orc_uint8) var60 + (orc_uint8) var61 + 1) >> 1; + /* 8: subb */ + var63 = var62 - var48; + /* 9: convsbw */ + var64.i = var63; + /* 10: loadupib */ + var65 = + (i & 1) ? ((orc_uint8) ptr7[i >> 1] + (orc_uint8) ptr7[(i >> 1) + 1] + + 1) >> 1 : ptr7[i >> 1]; + /* 11: loadupib */ + var66 = + (i & 1) ? ((orc_uint8) ptr8[i >> 1] + (orc_uint8) ptr8[(i >> 1) + 1] + + 1) >> 1 : ptr8[i >> 1]; + /* 12: avgub */ + var67 = ((orc_uint8) var65 + (orc_uint8) var66 + 1) >> 1; + /* 14: subb */ + var68 = var67 - var49; + /* 15: convsbw */ + var69.i = var68; + /* 17: mullw */ + var70.i = (var59.i * var50.i) & 0xffff; + /* 18: shrsw */ + var71.i = var70.i >> 8; + /* 19: addssw */ + var72.i = ORC_CLAMP_SW (var59.i + var71.i); + /* 20: addssw */ + var73.i = ORC_CLAMP_SW (var72.i + var69.i); + /* 22: mullw */ + var74.i = (var69.i * var51.i) & 0xffff; + /* 23: shrsw */ + var75.i = var74.i >> 8; + /* 24: subssw */ + var76.i = ORC_CLAMP_SW (var73.i - var75.i); + /* 25: addssw */ + var77.i = ORC_CLAMP_SW (var76.i + var69.i); + /* 26: addssw */ + var78.i = ORC_CLAMP_SW (var72.i + var64.i); + /* 27: addssw */ + var79.i = ORC_CLAMP_SW (var78.i + var64.i); + /* 29: mullw */ + var80.i = (var64.i * var52.i) & 0xffff; + /* 30: shrsw */ + var81.i = var80.i >> 8; + /* 31: addssw */ + var82.i = ORC_CLAMP_SW (var79.i + var81.i); + /* 33: mullw */ + var83.i = (var64.i * var53.i) & 0xffff; + /* 34: shrsw */ + var84.i = var83.i >> 8; + /* 35: subssw */ + var85.i = ORC_CLAMP_SW (var72.i - var84.i); + /* 37: mullw */ + var86.i = (var69.i * var54.i) & 0xffff; + /* 38: shrsw */ + var87.i = var86.i >> 8; + /* 39: subssw */ + var88.i = ORC_CLAMP_SW (var85.i - var87.i); + /* 40: subssw */ + var89.i = ORC_CLAMP_SW (var88.i - var87.i); + /* 41: convssswb */ + var90 = ORC_CLAMP_SB (var77.i); + /* 42: convssswb */ + var91 = ORC_CLAMP_SB (var89.i); + /* 43: convssswb */ + var92 = ORC_CLAMP_SB (var82.i); + /* 44: mergebw */ + var93.i = ((orc_uint8) var92) | ((orc_uint8) var91 << 8); + /* 46: mergebw */ + var94.i = ((orc_uint8) var90) | ((orc_uint8) var55 << 8); + /* 47: mergewl */ + var95.i = ((orc_uint16) var93.i) | ((orc_uint16) var94.i << 16); + /* 49: addb */ + var57.x4[0] = var95.x4[0] + var56.x4[0]; + var57.x4[1] = var95.x4[1] + var56.x4[1]; + var57.x4[2] = var95.x4[2] + var56.x4[2]; + var57.x4[3] = var95.x4[3] + var56.x4[3]; + /* 50: storel */ + ptr0[i] = var57; + } + +} + +#else +static void +_backup_cogorc_convert_I420_BGRA_avg (OrcExecutor * ex) +{ + int i; + int n = ex->n; + orc_union32 *ORC_RESTRICT ptr0; + const orc_int8 *ORC_RESTRICT ptr4; + const orc_int8 *ORC_RESTRICT ptr5; + const orc_int8 *ORC_RESTRICT ptr6; + const orc_int8 *ORC_RESTRICT ptr7; + const orc_int8 *ORC_RESTRICT ptr8; + orc_int8 var46; + orc_int8 var47; + orc_int8 var48; + orc_int8 var49; + orc_union16 var50; + orc_union16 var51; + orc_union16 var52; + orc_union16 var53; + orc_union16 var54; + orc_int8 var55; + orc_union32 var56; + orc_union32 var57; + orc_int8 var58; + orc_union16 var59; + orc_int8 var60; + orc_int8 var61; + orc_int8 var62; + orc_int8 var63; + orc_union16 var64; + orc_int8 var65; + orc_int8 var66; + orc_int8 var67; + orc_int8 var68; + orc_union16 var69; + orc_union16 var70; + orc_union16 var71; + orc_union16 var72; + orc_union16 var73; + orc_union16 var74; + orc_union16 var75; + orc_union16 var76; + orc_union16 var77; + orc_union16 var78; + orc_union16 var79; + orc_union16 var80; + orc_union16 var81; + orc_union16 var82; + orc_union16 var83; + orc_union16 var84; + orc_union16 var85; + orc_union16 var86; + orc_union16 var87; + orc_union16 var88; + orc_union16 var89; + orc_int8 var90; + orc_int8 var91; + orc_int8 var92; + orc_union16 var93; + orc_union16 var94; + orc_union32 var95; + + ptr0 = (orc_union32 *) ex->arrays[0]; + ptr4 = (orc_int8 *) ex->arrays[4]; + ptr5 = (orc_int8 *) ex->arrays[5]; + ptr6 = (orc_int8 *) ex->arrays[6]; + ptr7 = (orc_int8 *) ex->arrays[7]; + ptr8 = (orc_int8 *) ex->arrays[8]; + + /* 1: loadpb */ + var47 = 0x00000080; /* 128 or 6.32404e-322f */ + /* 7: loadpb */ + var48 = 0x00000080; /* 128 or 6.32404e-322f */ + /* 13: loadpb */ + var49 = 0x00000080; /* 128 or 6.32404e-322f */ + /* 16: loadpw */ + var50.i = 0x0000002a; /* 42 or 2.07508e-322f */ + /* 21: loadpw */ + var51.i = 0x00000067; /* 103 or 5.08888e-322f */ + /* 28: loadpw */ + var52.i = 0x00000004; /* 4 or 1.97626e-323f */ + /* 32: loadpw */ + var53.i = 0x00000064; /* 100 or 4.94066e-322f */ + /* 36: loadpw */ + var54.i = 0x00000068; /* 104 or 5.13828e-322f */ + /* 45: loadpb */ + var55 = 0x000000ff; /* 255 or 1.25987e-321f */ + /* 48: loadpb */ + var56.x4[0] = 0x00000080; /* 128 or 6.32404e-322f */ + var56.x4[1] = 0x00000080; /* 128 or 6.32404e-322f */ + var56.x4[2] = 0x00000080; /* 128 or 6.32404e-322f */ + var56.x4[3] = 0x00000080; /* 128 or 6.32404e-322f */ + + for (i = 0; i < n; i++) { + /* 0: loadb */ + var46 = ptr4[i]; + /* 2: subb */ + var58 = var46 - var47; + /* 3: convsbw */ + var59.i = var58; + /* 4: loadupib */ + var60 = + (i & 1) ? ((orc_uint8) ptr5[i >> 1] + (orc_uint8) ptr5[(i >> 1) + 1] + + 1) >> 1 : ptr5[i >> 1]; + /* 5: loadupib */ + var61 = + (i & 1) ? ((orc_uint8) ptr6[i >> 1] + (orc_uint8) ptr6[(i >> 1) + 1] + + 1) >> 1 : ptr6[i >> 1]; + /* 6: avgub */ + var62 = ((orc_uint8) var60 + (orc_uint8) var61 + 1) >> 1; + /* 8: subb */ + var63 = var62 - var48; + /* 9: convsbw */ + var64.i = var63; + /* 10: loadupib */ + var65 = + (i & 1) ? ((orc_uint8) ptr7[i >> 1] + (orc_uint8) ptr7[(i >> 1) + 1] + + 1) >> 1 : ptr7[i >> 1]; + /* 11: loadupib */ + var66 = + (i & 1) ? ((orc_uint8) ptr8[i >> 1] + (orc_uint8) ptr8[(i >> 1) + 1] + + 1) >> 1 : ptr8[i >> 1]; + /* 12: avgub */ + var67 = ((orc_uint8) var65 + (orc_uint8) var66 + 1) >> 1; + /* 14: subb */ + var68 = var67 - var49; + /* 15: convsbw */ + var69.i = var68; + /* 17: mullw */ + var70.i = (var59.i * var50.i) & 0xffff; + /* 18: shrsw */ + var71.i = var70.i >> 8; + /* 19: addssw */ + var72.i = ORC_CLAMP_SW (var59.i + var71.i); + /* 20: addssw */ + var73.i = ORC_CLAMP_SW (var72.i + var69.i); + /* 22: mullw */ + var74.i = (var69.i * var51.i) & 0xffff; + /* 23: shrsw */ + var75.i = var74.i >> 8; + /* 24: subssw */ + var76.i = ORC_CLAMP_SW (var73.i - var75.i); + /* 25: addssw */ + var77.i = ORC_CLAMP_SW (var76.i + var69.i); + /* 26: addssw */ + var78.i = ORC_CLAMP_SW (var72.i + var64.i); + /* 27: addssw */ + var79.i = ORC_CLAMP_SW (var78.i + var64.i); + /* 29: mullw */ + var80.i = (var64.i * var52.i) & 0xffff; + /* 30: shrsw */ + var81.i = var80.i >> 8; + /* 31: addssw */ + var82.i = ORC_CLAMP_SW (var79.i + var81.i); + /* 33: mullw */ + var83.i = (var64.i * var53.i) & 0xffff; + /* 34: shrsw */ + var84.i = var83.i >> 8; + /* 35: subssw */ + var85.i = ORC_CLAMP_SW (var72.i - var84.i); + /* 37: mullw */ + var86.i = (var69.i * var54.i) & 0xffff; + /* 38: shrsw */ + var87.i = var86.i >> 8; + /* 39: subssw */ + var88.i = ORC_CLAMP_SW (var85.i - var87.i); + /* 40: subssw */ + var89.i = ORC_CLAMP_SW (var88.i - var87.i); + /* 41: convssswb */ + var90 = ORC_CLAMP_SB (var77.i); + /* 42: convssswb */ + var91 = ORC_CLAMP_SB (var89.i); + /* 43: convssswb */ + var92 = ORC_CLAMP_SB (var82.i); + /* 44: mergebw */ + var93.i = ((orc_uint8) var92) | ((orc_uint8) var91 << 8); + /* 46: mergebw */ + var94.i = ((orc_uint8) var90) | ((orc_uint8) var55 << 8); + /* 47: mergewl */ + var95.i = ((orc_uint16) var93.i) | ((orc_uint16) var94.i << 16); + /* 49: addb */ + var57.x4[0] = var95.x4[0] + var56.x4[0]; + var57.x4[1] = var95.x4[1] + var56.x4[1]; + var57.x4[2] = var95.x4[2] + var56.x4[2]; + var57.x4[3] = var95.x4[3] + var56.x4[3]; + /* 50: storel */ + ptr0[i] = var57; + } + +} + +void +cogorc_convert_I420_BGRA_avg (guint8 * d1, const guint8 * s1, const guint8 * s2, + const guint8 * s3, const guint8 * s4, const guint8 * s5, int n) +{ + OrcExecutor _ex, *ex = &_ex; + static int p_inited = 0; + static OrcProgram *p = 0; + void (*func) (OrcExecutor *); + + if (!p_inited) { + orc_once_mutex_lock (); + if (!p_inited) { + OrcCompileResult result; + + p = orc_program_new (); + orc_program_set_name (p, "cogorc_convert_I420_BGRA_avg"); + orc_program_set_backup_function (p, _backup_cogorc_convert_I420_BGRA_avg); + orc_program_add_destination (p, 4, "d1"); + orc_program_add_source (p, 1, "s1"); + orc_program_add_source (p, 1, "s2"); + orc_program_add_source (p, 1, "s3"); + orc_program_add_source (p, 1, "s4"); + orc_program_add_source (p, 1, "s5"); + orc_program_add_constant (p, 1, 0x00000008, "c1"); + orc_program_add_constant (p, 1, 0x00000080, "c2"); + orc_program_add_constant (p, 4, 0x0000002a, "c3"); + orc_program_add_constant (p, 4, 0x00000067, "c4"); + orc_program_add_constant (p, 4, 0x00000004, "c5"); + orc_program_add_constant (p, 4, 0x00000064, "c6"); + orc_program_add_constant (p, 4, 0x00000068, "c7"); + orc_program_add_constant (p, 4, 0x000000ff, "c8"); + orc_program_add_temporary (p, 2, "t1"); + orc_program_add_temporary (p, 2, "t2"); + orc_program_add_temporary (p, 1, "t3"); + orc_program_add_temporary (p, 1, "t4"); + orc_program_add_temporary (p, 2, "t5"); + orc_program_add_temporary (p, 2, "t6"); + orc_program_add_temporary (p, 2, "t7"); + orc_program_add_temporary (p, 2, "t8"); + orc_program_add_temporary (p, 2, "t9"); + orc_program_add_temporary (p, 2, "t10"); + orc_program_add_temporary (p, 1, "t11"); + orc_program_add_temporary (p, 1, "t12"); + orc_program_add_temporary (p, 1, "t13"); + orc_program_add_temporary (p, 4, "t14"); + + orc_program_append_2 (p, "subb", 0, ORC_VAR_T3, ORC_VAR_S1, ORC_VAR_C2, + ORC_VAR_D1); + orc_program_append_2 (p, "convsbw", 0, ORC_VAR_T5, ORC_VAR_T3, ORC_VAR_D1, + ORC_VAR_D1); + orc_program_append_2 (p, "loadupib", 0, ORC_VAR_T3, ORC_VAR_S2, + ORC_VAR_D1, ORC_VAR_D1); + orc_program_append_2 (p, "loadupib", 0, ORC_VAR_T4, ORC_VAR_S3, + ORC_VAR_D1, ORC_VAR_D1); + orc_program_append_2 (p, "avgub", 0, ORC_VAR_T3, ORC_VAR_T3, ORC_VAR_T4, + ORC_VAR_D1); + orc_program_append_2 (p, "subb", 0, ORC_VAR_T3, ORC_VAR_T3, ORC_VAR_C2, + ORC_VAR_D1); + orc_program_append_2 (p, "convsbw", 0, ORC_VAR_T6, ORC_VAR_T3, ORC_VAR_D1, + ORC_VAR_D1); + orc_program_append_2 (p, "loadupib", 0, ORC_VAR_T3, ORC_VAR_S4, + ORC_VAR_D1, ORC_VAR_D1); + orc_program_append_2 (p, "loadupib", 0, ORC_VAR_T4, ORC_VAR_S5, + ORC_VAR_D1, ORC_VAR_D1); + orc_program_append_2 (p, "avgub", 0, ORC_VAR_T3, ORC_VAR_T3, ORC_VAR_T4, + ORC_VAR_D1); + orc_program_append_2 (p, "subb", 0, ORC_VAR_T3, ORC_VAR_T3, ORC_VAR_C2, + ORC_VAR_D1); + orc_program_append_2 (p, "convsbw", 0, ORC_VAR_T7, ORC_VAR_T3, ORC_VAR_D1, + ORC_VAR_D1); + orc_program_append_2 (p, "mullw", 0, ORC_VAR_T1, ORC_VAR_T5, ORC_VAR_C3, + ORC_VAR_D1); + orc_program_append_2 (p, "shrsw", 0, ORC_VAR_T1, ORC_VAR_T1, ORC_VAR_C1, + ORC_VAR_D1); + orc_program_append_2 (p, "addssw", 0, ORC_VAR_T5, ORC_VAR_T5, ORC_VAR_T1, + ORC_VAR_D1); + orc_program_append_2 (p, "addssw", 0, ORC_VAR_T8, ORC_VAR_T5, ORC_VAR_T7, + ORC_VAR_D1); + orc_program_append_2 (p, "mullw", 0, ORC_VAR_T1, ORC_VAR_T7, ORC_VAR_C4, + ORC_VAR_D1); + orc_program_append_2 (p, "shrsw", 0, ORC_VAR_T1, ORC_VAR_T1, ORC_VAR_C1, + ORC_VAR_D1); + orc_program_append_2 (p, "subssw", 0, ORC_VAR_T8, ORC_VAR_T8, ORC_VAR_T1, + ORC_VAR_D1); + orc_program_append_2 (p, "addssw", 0, ORC_VAR_T8, ORC_VAR_T8, ORC_VAR_T7, + ORC_VAR_D1); + orc_program_append_2 (p, "addssw", 0, ORC_VAR_T10, ORC_VAR_T5, ORC_VAR_T6, + ORC_VAR_D1); + orc_program_append_2 (p, "addssw", 0, ORC_VAR_T10, ORC_VAR_T10, + ORC_VAR_T6, ORC_VAR_D1); + orc_program_append_2 (p, "mullw", 0, ORC_VAR_T1, ORC_VAR_T6, ORC_VAR_C5, + ORC_VAR_D1); + orc_program_append_2 (p, "shrsw", 0, ORC_VAR_T1, ORC_VAR_T1, ORC_VAR_C1, + ORC_VAR_D1); + orc_program_append_2 (p, "addssw", 0, ORC_VAR_T10, ORC_VAR_T10, + ORC_VAR_T1, ORC_VAR_D1); + orc_program_append_2 (p, "mullw", 0, ORC_VAR_T1, ORC_VAR_T6, ORC_VAR_C6, + ORC_VAR_D1); + orc_program_append_2 (p, "shrsw", 0, ORC_VAR_T1, ORC_VAR_T1, ORC_VAR_C1, + ORC_VAR_D1); + orc_program_append_2 (p, "subssw", 0, ORC_VAR_T9, ORC_VAR_T5, ORC_VAR_T1, + ORC_VAR_D1); + orc_program_append_2 (p, "mullw", 0, ORC_VAR_T1, ORC_VAR_T7, ORC_VAR_C7, + ORC_VAR_D1); + orc_program_append_2 (p, "shrsw", 0, ORC_VAR_T1, ORC_VAR_T1, ORC_VAR_C1, + ORC_VAR_D1); + orc_program_append_2 (p, "subssw", 0, ORC_VAR_T9, ORC_VAR_T9, ORC_VAR_T1, + ORC_VAR_D1); + orc_program_append_2 (p, "subssw", 0, ORC_VAR_T9, ORC_VAR_T9, ORC_VAR_T1, + ORC_VAR_D1); + orc_program_append_2 (p, "convssswb", 0, ORC_VAR_T11, ORC_VAR_T8, + ORC_VAR_D1, ORC_VAR_D1); + orc_program_append_2 (p, "convssswb", 0, ORC_VAR_T12, ORC_VAR_T9, + ORC_VAR_D1, ORC_VAR_D1); + orc_program_append_2 (p, "convssswb", 0, ORC_VAR_T13, ORC_VAR_T10, + ORC_VAR_D1, ORC_VAR_D1); + orc_program_append_2 (p, "mergebw", 0, ORC_VAR_T1, ORC_VAR_T13, + ORC_VAR_T12, ORC_VAR_D1); + orc_program_append_2 (p, "mergebw", 0, ORC_VAR_T2, ORC_VAR_T11, + ORC_VAR_C8, ORC_VAR_D1); + orc_program_append_2 (p, "mergewl", 0, ORC_VAR_T14, ORC_VAR_T1, + ORC_VAR_T2, ORC_VAR_D1); + orc_program_append_2 (p, "addb", 2, ORC_VAR_D1, ORC_VAR_T14, ORC_VAR_C2, + ORC_VAR_D1); + + result = orc_program_compile (p); + } + p_inited = TRUE; + orc_once_mutex_unlock (); + } + ex->program = p; + + ex->n = n; + ex->arrays[ORC_VAR_D1] = d1; + ex->arrays[ORC_VAR_S1] = (void *) s1; + ex->arrays[ORC_VAR_S2] = (void *) s2; + ex->arrays[ORC_VAR_S3] = (void *) s3; + ex->arrays[ORC_VAR_S4] = (void *) s4; + ex->arrays[ORC_VAR_S5] = (void *) s5; + + func = p->code_exec; + func (ex); +} +#endif + + +/* cogorc_getline_I420 */ +#ifdef DISABLE_ORC +void +cogorc_getline_I420 (guint8 * d1, const guint8 * s1, const guint8 * s2, + const guint8 * s3, int n) +{ + int i; + orc_union32 *ORC_RESTRICT ptr0; + const orc_int8 *ORC_RESTRICT ptr4; + const orc_int8 *ORC_RESTRICT ptr5; + const orc_int8 *ORC_RESTRICT ptr6; + orc_int8 var36; + orc_int8 var37; + orc_union32 var38; + orc_int8 var39; + orc_int8 var40; + orc_union16 var41; + orc_union16 var42; + + ptr0 = (orc_union32 *) d1; + ptr4 = (orc_int8 *) s1; + ptr5 = (orc_int8 *) s2; + ptr6 = (orc_int8 *) s3; + + /* 3: loadpb */ + var36 = 0x000000ff; /* 255 or 1.25987e-321f */ + + for (i = 0; i < n; i++) { + /* 0: loadupdb */ + var39 = ptr5[i >> 1]; + /* 1: loadupdb */ + var40 = ptr6[i >> 1]; + /* 2: mergebw */ + var41.i = ((orc_uint8) var39) | ((orc_uint8) var40 << 8); + /* 4: loadb */ + var37 = ptr4[i]; + /* 5: mergebw */ + var42.i = ((orc_uint8) var36) | ((orc_uint8) var37 << 8); + /* 6: mergewl */ + var38.i = ((orc_uint16) var42.i) | ((orc_uint16) var41.i << 16); + /* 7: storel */ + ptr0[i] = var38; + } + +} + +#else +static void +_backup_cogorc_getline_I420 (OrcExecutor * ex) +{ + int i; + int n = ex->n; + orc_union32 *ORC_RESTRICT ptr0; + const orc_int8 *ORC_RESTRICT ptr4; + const orc_int8 *ORC_RESTRICT ptr5; + const orc_int8 *ORC_RESTRICT ptr6; + orc_int8 var36; + orc_int8 var37; + orc_union32 var38; + orc_int8 var39; + orc_int8 var40; + orc_union16 var41; + orc_union16 var42; + + ptr0 = (orc_union32 *) ex->arrays[0]; + ptr4 = (orc_int8 *) ex->arrays[4]; + ptr5 = (orc_int8 *) ex->arrays[5]; + ptr6 = (orc_int8 *) ex->arrays[6]; + + /* 3: loadpb */ + var36 = 0x000000ff; /* 255 or 1.25987e-321f */ + + for (i = 0; i < n; i++) { + /* 0: loadupdb */ + var39 = ptr5[i >> 1]; + /* 1: loadupdb */ + var40 = ptr6[i >> 1]; + /* 2: mergebw */ + var41.i = ((orc_uint8) var39) | ((orc_uint8) var40 << 8); + /* 4: loadb */ + var37 = ptr4[i]; + /* 5: mergebw */ + var42.i = ((orc_uint8) var36) | ((orc_uint8) var37 << 8); + /* 6: mergewl */ + var38.i = ((orc_uint16) var42.i) | ((orc_uint16) var41.i << 16); + /* 7: storel */ + ptr0[i] = var38; + } + +} + +void +cogorc_getline_I420 (guint8 * d1, const guint8 * s1, const guint8 * s2, + const guint8 * s3, int n) +{ + OrcExecutor _ex, *ex = &_ex; + static int p_inited = 0; + static OrcProgram *p = 0; + void (*func) (OrcExecutor *); + + if (!p_inited) { + orc_once_mutex_lock (); + if (!p_inited) { + OrcCompileResult result; + + p = orc_program_new (); + orc_program_set_name (p, "cogorc_getline_I420"); + orc_program_set_backup_function (p, _backup_cogorc_getline_I420); + orc_program_add_destination (p, 4, "d1"); + orc_program_add_source (p, 1, "s1"); + orc_program_add_source (p, 1, "s2"); + orc_program_add_source (p, 1, "s3"); + orc_program_add_constant (p, 1, 0x000000ff, "c1"); + orc_program_add_constant (p, 0, 0x00000000, "c2"); + orc_program_add_constant (p, 0, 0x00000000, "c3"); + orc_program_add_constant (p, 0, 0x00000000, "c4"); + orc_program_add_constant (p, 0, 0x00000000, "c5"); + orc_program_add_constant (p, 0, 0x00000000, "c6"); + orc_program_add_constant (p, 0, 0x00000000, "c7"); + orc_program_add_constant (p, 0, 0x00000000, "c8"); + orc_program_add_temporary (p, 2, "t1"); + orc_program_add_temporary (p, 2, "t2"); + orc_program_add_temporary (p, 1, "t3"); + orc_program_add_temporary (p, 1, "t4"); + + orc_program_append_2 (p, "loadupdb", 0, ORC_VAR_T3, ORC_VAR_S2, + ORC_VAR_D1, ORC_VAR_D1); + orc_program_append_2 (p, "loadupdb", 0, ORC_VAR_T4, ORC_VAR_S3, + ORC_VAR_D1, ORC_VAR_D1); + orc_program_append_2 (p, "mergebw", 0, ORC_VAR_T1, ORC_VAR_T3, ORC_VAR_T4, + ORC_VAR_D1); + orc_program_append_2 (p, "mergebw", 0, ORC_VAR_T2, ORC_VAR_C1, ORC_VAR_S1, + ORC_VAR_D1); + orc_program_append_2 (p, "mergewl", 0, ORC_VAR_D1, ORC_VAR_T2, ORC_VAR_T1, + ORC_VAR_D1); + + result = orc_program_compile (p); + } + p_inited = TRUE; + orc_once_mutex_unlock (); + } + ex->program = p; + + ex->n = n; + ex->arrays[ORC_VAR_D1] = d1; + ex->arrays[ORC_VAR_S1] = (void *) s1; + ex->arrays[ORC_VAR_S2] = (void *) s2; + ex->arrays[ORC_VAR_S3] = (void *) s3; + + func = p->code_exec; + func (ex); +} +#endif + + +/* cogorc_getline_YUY2 */ +#ifdef DISABLE_ORC +void +cogorc_getline_YUY2 (guint8 * d1, const guint8 * s1, int n) +{ + int i; + orc_union64 *ORC_RESTRICT ptr0; + const orc_union32 *ORC_RESTRICT ptr4; + orc_union32 var36; + orc_union16 var37; + orc_union64 var38; + orc_union16 var39; + orc_union16 var40; + orc_union32 var41; + orc_union32 var42; + + ptr0 = (orc_union64 *) d1; + ptr4 = (orc_union32 *) s1; + + /* 2: loadpb */ + var37.x2[0] = 0x000000ff; /* 255 or 1.25987e-321f */ + var37.x2[1] = 0x000000ff; /* 255 or 1.25987e-321f */ + + for (i = 0; i < n; i++) { + /* 0: loadl */ + var36 = ptr4[i]; + /* 1: splitwb */ + var39.x2[0] = (var36.x2[0] >> 8) & 0xff; + var40.x2[0] = var36.x2[0] & 0xff; + var39.x2[1] = (var36.x2[1] >> 8) & 0xff; + var40.x2[1] = var36.x2[1] & 0xff; + /* 3: mergebw */ + var41.x2[0] = ((orc_uint8) var37.x2[0]) | ((orc_uint8) var40.x2[0] << 8); + var41.x2[1] = ((orc_uint8) var37.x2[1]) | ((orc_uint8) var40.x2[1] << 8); + /* 4: mergewl */ + var42.i = ((orc_uint16) var39.i) | ((orc_uint16) var39.i << 16); + /* 5: mergewl */ + var38.x2[0] = ((orc_uint16) var41.x2[0]) | ((orc_uint16) var42.x2[0] << 16); + var38.x2[1] = ((orc_uint16) var41.x2[1]) | ((orc_uint16) var42.x2[1] << 16); + /* 6: storeq */ + ptr0[i] = var38; + } + +} + +#else +static void +_backup_cogorc_getline_YUY2 (OrcExecutor * ex) +{ + int i; + int n = ex->n; + orc_union64 *ORC_RESTRICT ptr0; + const orc_union32 *ORC_RESTRICT ptr4; + orc_union32 var36; + orc_union16 var37; + orc_union64 var38; + orc_union16 var39; + orc_union16 var40; + orc_union32 var41; + orc_union32 var42; + + ptr0 = (orc_union64 *) ex->arrays[0]; + ptr4 = (orc_union32 *) ex->arrays[4]; + + /* 2: loadpb */ + var37.x2[0] = 0x000000ff; /* 255 or 1.25987e-321f */ + var37.x2[1] = 0x000000ff; /* 255 or 1.25987e-321f */ + + for (i = 0; i < n; i++) { + /* 0: loadl */ + var36 = ptr4[i]; + /* 1: splitwb */ + var39.x2[0] = (var36.x2[0] >> 8) & 0xff; + var40.x2[0] = var36.x2[0] & 0xff; + var39.x2[1] = (var36.x2[1] >> 8) & 0xff; + var40.x2[1] = var36.x2[1] & 0xff; + /* 3: mergebw */ + var41.x2[0] = ((orc_uint8) var37.x2[0]) | ((orc_uint8) var40.x2[0] << 8); + var41.x2[1] = ((orc_uint8) var37.x2[1]) | ((orc_uint8) var40.x2[1] << 8); + /* 4: mergewl */ + var42.i = ((orc_uint16) var39.i) | ((orc_uint16) var39.i << 16); + /* 5: mergewl */ + var38.x2[0] = ((orc_uint16) var41.x2[0]) | ((orc_uint16) var42.x2[0] << 16); + var38.x2[1] = ((orc_uint16) var41.x2[1]) | ((orc_uint16) var42.x2[1] << 16); + /* 6: storeq */ + ptr0[i] = var38; + } + +} + +void +cogorc_getline_YUY2 (guint8 * d1, const guint8 * s1, int n) +{ + OrcExecutor _ex, *ex = &_ex; + static int p_inited = 0; + static OrcProgram *p = 0; + void (*func) (OrcExecutor *); + + if (!p_inited) { + orc_once_mutex_lock (); + if (!p_inited) { + OrcCompileResult result; + + p = orc_program_new (); + orc_program_set_name (p, "cogorc_getline_YUY2"); + orc_program_set_backup_function (p, _backup_cogorc_getline_YUY2); + orc_program_add_destination (p, 8, "d1"); + orc_program_add_source (p, 4, "s1"); + orc_program_add_constant (p, 2, 0x000000ff, "c1"); + orc_program_add_constant (p, 0, 0x00000000, "c2"); + orc_program_add_constant (p, 0, 0x00000000, "c3"); + orc_program_add_constant (p, 0, 0x00000000, "c4"); + orc_program_add_constant (p, 0, 0x00000000, "c5"); + orc_program_add_constant (p, 0, 0x00000000, "c6"); + orc_program_add_constant (p, 0, 0x00000000, "c7"); + orc_program_add_constant (p, 0, 0x00000000, "c8"); + orc_program_add_temporary (p, 2, "t1"); + orc_program_add_temporary (p, 2, "t2"); + orc_program_add_temporary (p, 4, "t3"); + orc_program_add_temporary (p, 4, "t4"); + + orc_program_append_2 (p, "splitwb", 1, ORC_VAR_T2, ORC_VAR_T1, ORC_VAR_S1, + ORC_VAR_D1); + orc_program_append_2 (p, "mergebw", 1, ORC_VAR_T3, ORC_VAR_C1, ORC_VAR_T1, + ORC_VAR_D1); + orc_program_append_2 (p, "mergewl", 0, ORC_VAR_T4, ORC_VAR_T2, ORC_VAR_T2, + ORC_VAR_D1); + orc_program_append_2 (p, "mergewl", 1, ORC_VAR_D1, ORC_VAR_T3, ORC_VAR_T4, + ORC_VAR_D1); + + result = orc_program_compile (p); + } + p_inited = TRUE; + orc_once_mutex_unlock (); + } + ex->program = p; + + ex->n = n; + ex->arrays[ORC_VAR_D1] = d1; + ex->arrays[ORC_VAR_S1] = (void *) s1; + + func = p->code_exec; + func (ex); +} +#endif + + +/* cogorc_getline_UYVY */ +#ifdef DISABLE_ORC +void +cogorc_getline_UYVY (guint8 * d1, const guint8 * s1, int n) +{ + int i; + orc_union64 *ORC_RESTRICT ptr0; + const orc_union32 *ORC_RESTRICT ptr4; + orc_union32 var36; + orc_union16 var37; + orc_union64 var38; + orc_union16 var39; + orc_union16 var40; + orc_union32 var41; + orc_union32 var42; + + ptr0 = (orc_union64 *) d1; + ptr4 = (orc_union32 *) s1; + + /* 2: loadpb */ + var37.x2[0] = 0x000000ff; /* 255 or 1.25987e-321f */ + var37.x2[1] = 0x000000ff; /* 255 or 1.25987e-321f */ + + for (i = 0; i < n; i++) { + /* 0: loadl */ + var36 = ptr4[i]; + /* 1: splitwb */ + var39.x2[0] = (var36.x2[0] >> 8) & 0xff; + var40.x2[0] = var36.x2[0] & 0xff; + var39.x2[1] = (var36.x2[1] >> 8) & 0xff; + var40.x2[1] = var36.x2[1] & 0xff; + /* 3: mergebw */ + var41.x2[0] = ((orc_uint8) var37.x2[0]) | ((orc_uint8) var39.x2[0] << 8); + var41.x2[1] = ((orc_uint8) var37.x2[1]) | ((orc_uint8) var39.x2[1] << 8); + /* 4: mergewl */ + var42.i = ((orc_uint16) var40.i) | ((orc_uint16) var40.i << 16); + /* 5: mergewl */ + var38.x2[0] = ((orc_uint16) var41.x2[0]) | ((orc_uint16) var42.x2[0] << 16); + var38.x2[1] = ((orc_uint16) var41.x2[1]) | ((orc_uint16) var42.x2[1] << 16); + /* 6: storeq */ + ptr0[i] = var38; + } + +} + +#else +static void +_backup_cogorc_getline_UYVY (OrcExecutor * ex) +{ + int i; + int n = ex->n; + orc_union64 *ORC_RESTRICT ptr0; + const orc_union32 *ORC_RESTRICT ptr4; + orc_union32 var36; + orc_union16 var37; + orc_union64 var38; + orc_union16 var39; + orc_union16 var40; + orc_union32 var41; + orc_union32 var42; + + ptr0 = (orc_union64 *) ex->arrays[0]; + ptr4 = (orc_union32 *) ex->arrays[4]; + + /* 2: loadpb */ + var37.x2[0] = 0x000000ff; /* 255 or 1.25987e-321f */ + var37.x2[1] = 0x000000ff; /* 255 or 1.25987e-321f */ + + for (i = 0; i < n; i++) { + /* 0: loadl */ + var36 = ptr4[i]; + /* 1: splitwb */ + var39.x2[0] = (var36.x2[0] >> 8) & 0xff; + var40.x2[0] = var36.x2[0] & 0xff; + var39.x2[1] = (var36.x2[1] >> 8) & 0xff; + var40.x2[1] = var36.x2[1] & 0xff; + /* 3: mergebw */ + var41.x2[0] = ((orc_uint8) var37.x2[0]) | ((orc_uint8) var39.x2[0] << 8); + var41.x2[1] = ((orc_uint8) var37.x2[1]) | ((orc_uint8) var39.x2[1] << 8); + /* 4: mergewl */ + var42.i = ((orc_uint16) var40.i) | ((orc_uint16) var40.i << 16); + /* 5: mergewl */ + var38.x2[0] = ((orc_uint16) var41.x2[0]) | ((orc_uint16) var42.x2[0] << 16); + var38.x2[1] = ((orc_uint16) var41.x2[1]) | ((orc_uint16) var42.x2[1] << 16); + /* 6: storeq */ + ptr0[i] = var38; + } + +} + +void +cogorc_getline_UYVY (guint8 * d1, const guint8 * s1, int n) +{ + OrcExecutor _ex, *ex = &_ex; + static int p_inited = 0; + static OrcProgram *p = 0; + void (*func) (OrcExecutor *); + + if (!p_inited) { + orc_once_mutex_lock (); + if (!p_inited) { + OrcCompileResult result; + + p = orc_program_new (); + orc_program_set_name (p, "cogorc_getline_UYVY"); + orc_program_set_backup_function (p, _backup_cogorc_getline_UYVY); + orc_program_add_destination (p, 8, "d1"); + orc_program_add_source (p, 4, "s1"); + orc_program_add_constant (p, 2, 0x000000ff, "c1"); + orc_program_add_constant (p, 0, 0x00000000, "c2"); + orc_program_add_constant (p, 0, 0x00000000, "c3"); + orc_program_add_constant (p, 0, 0x00000000, "c4"); + orc_program_add_constant (p, 0, 0x00000000, "c5"); + orc_program_add_constant (p, 0, 0x00000000, "c6"); + orc_program_add_constant (p, 0, 0x00000000, "c7"); + orc_program_add_constant (p, 0, 0x00000000, "c8"); + orc_program_add_temporary (p, 2, "t1"); + orc_program_add_temporary (p, 2, "t2"); + orc_program_add_temporary (p, 4, "t3"); + orc_program_add_temporary (p, 4, "t4"); + + orc_program_append_2 (p, "splitwb", 1, ORC_VAR_T1, ORC_VAR_T2, ORC_VAR_S1, + ORC_VAR_D1); + orc_program_append_2 (p, "mergebw", 1, ORC_VAR_T3, ORC_VAR_C1, ORC_VAR_T1, + ORC_VAR_D1); + orc_program_append_2 (p, "mergewl", 0, ORC_VAR_T4, ORC_VAR_T2, ORC_VAR_T2, + ORC_VAR_D1); + orc_program_append_2 (p, "mergewl", 1, ORC_VAR_D1, ORC_VAR_T3, ORC_VAR_T4, + ORC_VAR_D1); + + result = orc_program_compile (p); + } + p_inited = TRUE; + orc_once_mutex_unlock (); + } + ex->program = p; + + ex->n = n; + ex->arrays[ORC_VAR_D1] = d1; + ex->arrays[ORC_VAR_S1] = (void *) s1; + + func = p->code_exec; + func (ex); +} +#endif + + +/* cogorc_getline_YVYU */ +#ifdef DISABLE_ORC +void +cogorc_getline_YVYU (guint8 * d1, const guint8 * s1, int n) +{ + int i; + orc_union64 *ORC_RESTRICT ptr0; + const orc_union32 *ORC_RESTRICT ptr4; + orc_union32 var36; + orc_union16 var37; + orc_union64 var38; + orc_union16 var39; + orc_union16 var40; + orc_union32 var41; + orc_union32 var42; + + ptr0 = (orc_union64 *) d1; + ptr4 = (orc_union32 *) s1; + + /* 2: loadpb */ + var37.x2[0] = 0x000000ff; /* 255 or 1.25987e-321f */ + var37.x2[1] = 0x000000ff; /* 255 or 1.25987e-321f */ + + for (i = 0; i < n; i++) { + /* 0: loadl */ + var36 = ptr4[i]; + /* 1: splitwb */ + var39.x2[0] = (var36.x2[0] >> 8) & 0xff; + var40.x2[0] = var36.x2[0] & 0xff; + var39.x2[1] = (var36.x2[1] >> 8) & 0xff; + var40.x2[1] = var36.x2[1] & 0xff; + /* 3: mergebw */ + var41.x2[0] = ((orc_uint8) var37.x2[0]) | ((orc_uint8) var39.x2[0] << 8); + var41.x2[1] = ((orc_uint8) var37.x2[1]) | ((orc_uint8) var39.x2[1] << 8); + /* 4: mergewl */ + var42.i = ((orc_uint16) var40.i) | ((orc_uint16) var40.i << 16); + /* 5: mergewl */ + var38.x2[0] = ((orc_uint16) var41.x2[0]) | ((orc_uint16) var42.x2[0] << 16); + var38.x2[1] = ((orc_uint16) var41.x2[1]) | ((orc_uint16) var42.x2[1] << 16); + /* 6: storeq */ + ptr0[i] = var38; + } + +} + +#else +static void +_backup_cogorc_getline_YVYU (OrcExecutor * ex) +{ + int i; + int n = ex->n; + orc_union64 *ORC_RESTRICT ptr0; + const orc_union32 *ORC_RESTRICT ptr4; + orc_union32 var36; + orc_union16 var37; + orc_union64 var38; + orc_union16 var39; + orc_union16 var40; + orc_union32 var41; + orc_union32 var42; + + ptr0 = (orc_union64 *) ex->arrays[0]; + ptr4 = (orc_union32 *) ex->arrays[4]; + + /* 2: loadpb */ + var37.x2[0] = 0x000000ff; /* 255 or 1.25987e-321f */ + var37.x2[1] = 0x000000ff; /* 255 or 1.25987e-321f */ + + for (i = 0; i < n; i++) { + /* 0: loadl */ + var36 = ptr4[i]; + /* 1: splitwb */ + var39.x2[0] = (var36.x2[0] >> 8) & 0xff; + var40.x2[0] = var36.x2[0] & 0xff; + var39.x2[1] = (var36.x2[1] >> 8) & 0xff; + var40.x2[1] = var36.x2[1] & 0xff; + /* 3: mergebw */ + var41.x2[0] = ((orc_uint8) var37.x2[0]) | ((orc_uint8) var39.x2[0] << 8); + var41.x2[1] = ((orc_uint8) var37.x2[1]) | ((orc_uint8) var39.x2[1] << 8); + /* 4: mergewl */ + var42.i = ((orc_uint16) var40.i) | ((orc_uint16) var40.i << 16); + /* 5: mergewl */ + var38.x2[0] = ((orc_uint16) var41.x2[0]) | ((orc_uint16) var42.x2[0] << 16); + var38.x2[1] = ((orc_uint16) var41.x2[1]) | ((orc_uint16) var42.x2[1] << 16); + /* 6: storeq */ + ptr0[i] = var38; + } + +} + +void +cogorc_getline_YVYU (guint8 * d1, const guint8 * s1, int n) +{ + OrcExecutor _ex, *ex = &_ex; + static int p_inited = 0; + static OrcProgram *p = 0; + void (*func) (OrcExecutor *); + + if (!p_inited) { + orc_once_mutex_lock (); + if (!p_inited) { + OrcCompileResult result; + + p = orc_program_new (); + orc_program_set_name (p, "cogorc_getline_YVYU"); + orc_program_set_backup_function (p, _backup_cogorc_getline_YVYU); + orc_program_add_destination (p, 8, "d1"); + orc_program_add_source (p, 4, "s1"); + orc_program_add_constant (p, 2, 0x000000ff, "c1"); + orc_program_add_constant (p, 0, 0x00000000, "c2"); + orc_program_add_constant (p, 0, 0x00000000, "c3"); + orc_program_add_constant (p, 0, 0x00000000, "c4"); + orc_program_add_constant (p, 0, 0x00000000, "c5"); + orc_program_add_constant (p, 0, 0x00000000, "c6"); + orc_program_add_constant (p, 0, 0x00000000, "c7"); + orc_program_add_constant (p, 0, 0x00000000, "c8"); + orc_program_add_temporary (p, 2, "t1"); + orc_program_add_temporary (p, 2, "t2"); + orc_program_add_temporary (p, 4, "t3"); + orc_program_add_temporary (p, 4, "t4"); + + orc_program_append_2 (p, "splitwb", 1, ORC_VAR_T1, ORC_VAR_T2, ORC_VAR_S1, + ORC_VAR_D1); + orc_program_append_2 (p, "mergebw", 1, ORC_VAR_T3, ORC_VAR_C1, ORC_VAR_T1, + ORC_VAR_D1); + orc_program_append_2 (p, "mergewl", 0, ORC_VAR_T4, ORC_VAR_T2, ORC_VAR_T2, + ORC_VAR_D1); + orc_program_append_2 (p, "mergewl", 1, ORC_VAR_D1, ORC_VAR_T3, ORC_VAR_T4, + ORC_VAR_D1); + + result = orc_program_compile (p); + } + p_inited = TRUE; + orc_once_mutex_unlock (); + } + ex->program = p; + + ex->n = n; + ex->arrays[ORC_VAR_D1] = d1; + ex->arrays[ORC_VAR_S1] = (void *) s1; + + func = p->code_exec; + func (ex); +} +#endif + + +/* cogorc_getline_Y42B */ +#ifdef DISABLE_ORC +void +cogorc_getline_Y42B (guint8 * d1, const guint8 * s1, const guint8 * s2, + const guint8 * s3, int n) +{ + int i; + orc_union64 *ORC_RESTRICT ptr0; + const orc_union16 *ORC_RESTRICT ptr4; + const orc_int8 *ORC_RESTRICT ptr5; + const orc_int8 *ORC_RESTRICT ptr6; + orc_int8 var36; + orc_int8 var37; + orc_union16 var38; + orc_union16 var39; + orc_union64 var40; + orc_union16 var41; + orc_union32 var42; + orc_union32 var43; + + ptr0 = (orc_union64 *) d1; + ptr4 = (orc_union16 *) s1; + ptr5 = (orc_int8 *) s2; + ptr6 = (orc_int8 *) s3; + + /* 3: loadpb */ + var38.x2[0] = 0x000000ff; /* 255 or 1.25987e-321f */ + var38.x2[1] = 0x000000ff; /* 255 or 1.25987e-321f */ + + for (i = 0; i < n; i++) { + /* 0: loadb */ + var36 = ptr5[i]; + /* 1: loadb */ + var37 = ptr6[i]; + /* 2: mergebw */ + var41.i = ((orc_uint8) var36) | ((orc_uint8) var37 << 8); + /* 4: loadw */ + var39 = ptr4[i]; + /* 5: mergebw */ + var42.x2[0] = ((orc_uint8) var38.x2[0]) | ((orc_uint8) var39.x2[0] << 8); + var42.x2[1] = ((orc_uint8) var38.x2[1]) | ((orc_uint8) var39.x2[1] << 8); + /* 6: mergewl */ + var43.i = ((orc_uint16) var41.i) | ((orc_uint16) var41.i << 16); + /* 7: mergewl */ + var40.x2[0] = ((orc_uint16) var42.x2[0]) | ((orc_uint16) var43.x2[0] << 16); + var40.x2[1] = ((orc_uint16) var42.x2[1]) | ((orc_uint16) var43.x2[1] << 16); + /* 8: storeq */ + ptr0[i] = var40; + } + +} + +#else +static void +_backup_cogorc_getline_Y42B (OrcExecutor * ex) +{ + int i; + int n = ex->n; + orc_union64 *ORC_RESTRICT ptr0; + const orc_union16 *ORC_RESTRICT ptr4; + const orc_int8 *ORC_RESTRICT ptr5; + const orc_int8 *ORC_RESTRICT ptr6; + orc_int8 var36; + orc_int8 var37; + orc_union16 var38; + orc_union16 var39; + orc_union64 var40; + orc_union16 var41; + orc_union32 var42; + orc_union32 var43; + + ptr0 = (orc_union64 *) ex->arrays[0]; + ptr4 = (orc_union16 *) ex->arrays[4]; + ptr5 = (orc_int8 *) ex->arrays[5]; + ptr6 = (orc_int8 *) ex->arrays[6]; + + /* 3: loadpb */ + var38.x2[0] = 0x000000ff; /* 255 or 1.25987e-321f */ + var38.x2[1] = 0x000000ff; /* 255 or 1.25987e-321f */ + + for (i = 0; i < n; i++) { + /* 0: loadb */ + var36 = ptr5[i]; + /* 1: loadb */ + var37 = ptr6[i]; + /* 2: mergebw */ + var41.i = ((orc_uint8) var36) | ((orc_uint8) var37 << 8); + /* 4: loadw */ + var39 = ptr4[i]; + /* 5: mergebw */ + var42.x2[0] = ((orc_uint8) var38.x2[0]) | ((orc_uint8) var39.x2[0] << 8); + var42.x2[1] = ((orc_uint8) var38.x2[1]) | ((orc_uint8) var39.x2[1] << 8); + /* 6: mergewl */ + var43.i = ((orc_uint16) var41.i) | ((orc_uint16) var41.i << 16); + /* 7: mergewl */ + var40.x2[0] = ((orc_uint16) var42.x2[0]) | ((orc_uint16) var43.x2[0] << 16); + var40.x2[1] = ((orc_uint16) var42.x2[1]) | ((orc_uint16) var43.x2[1] << 16); + /* 8: storeq */ + ptr0[i] = var40; + } + +} + +void +cogorc_getline_Y42B (guint8 * d1, const guint8 * s1, const guint8 * s2, + const guint8 * s3, int n) +{ + OrcExecutor _ex, *ex = &_ex; + static int p_inited = 0; + static OrcProgram *p = 0; + void (*func) (OrcExecutor *); + + if (!p_inited) { + orc_once_mutex_lock (); + if (!p_inited) { + OrcCompileResult result; + + p = orc_program_new (); + orc_program_set_name (p, "cogorc_getline_Y42B"); + orc_program_set_backup_function (p, _backup_cogorc_getline_Y42B); + orc_program_add_destination (p, 8, "d1"); + orc_program_add_source (p, 2, "s1"); + orc_program_add_source (p, 1, "s2"); + orc_program_add_source (p, 1, "s3"); + orc_program_add_constant (p, 1, 0x000000ff, "c1"); + orc_program_add_constant (p, 0, 0x00000000, "c2"); + orc_program_add_constant (p, 0, 0x00000000, "c3"); + orc_program_add_constant (p, 0, 0x00000000, "c4"); + orc_program_add_constant (p, 0, 0x00000000, "c5"); + orc_program_add_constant (p, 0, 0x00000000, "c6"); + orc_program_add_constant (p, 0, 0x00000000, "c7"); + orc_program_add_constant (p, 0, 0x00000000, "c8"); + orc_program_add_temporary (p, 2, "t1"); + orc_program_add_temporary (p, 2, "t2"); + orc_program_add_temporary (p, 4, "t3"); + orc_program_add_temporary (p, 4, "t4"); + + orc_program_append_2 (p, "mergebw", 0, ORC_VAR_T1, ORC_VAR_S2, ORC_VAR_S3, + ORC_VAR_D1); + orc_program_append_2 (p, "mergebw", 1, ORC_VAR_T4, ORC_VAR_C1, ORC_VAR_S1, + ORC_VAR_D1); + orc_program_append_2 (p, "mergewl", 0, ORC_VAR_T3, ORC_VAR_T1, ORC_VAR_T1, + ORC_VAR_D1); + orc_program_append_2 (p, "mergewl", 1, ORC_VAR_D1, ORC_VAR_T4, ORC_VAR_T3, + ORC_VAR_D1); + + result = orc_program_compile (p); + } + p_inited = TRUE; + orc_once_mutex_unlock (); + } + ex->program = p; + + ex->n = n; + ex->arrays[ORC_VAR_D1] = d1; + ex->arrays[ORC_VAR_S1] = (void *) s1; + ex->arrays[ORC_VAR_S2] = (void *) s2; + ex->arrays[ORC_VAR_S3] = (void *) s3; + + func = p->code_exec; + func (ex); +} +#endif + + +/* cogorc_getline_Y444 */ +#ifdef DISABLE_ORC +void +cogorc_getline_Y444 (guint8 * d1, const guint8 * s1, const guint8 * s2, + const guint8 * s3, int n) +{ + int i; + orc_union32 *ORC_RESTRICT ptr0; + const orc_int8 *ORC_RESTRICT ptr4; + const orc_int8 *ORC_RESTRICT ptr5; + const orc_int8 *ORC_RESTRICT ptr6; + orc_int8 var34; + orc_int8 var35; + orc_int8 var36; + orc_int8 var37; + orc_union32 var38; + orc_union16 var39; + orc_union16 var40; + + ptr0 = (orc_union32 *) d1; + ptr4 = (orc_int8 *) s1; + ptr5 = (orc_int8 *) s2; + ptr6 = (orc_int8 *) s3; + + /* 3: loadpb */ + var36 = 0x000000ff; /* 255 or 1.25987e-321f */ + + for (i = 0; i < n; i++) { + /* 0: loadb */ + var34 = ptr5[i]; + /* 1: loadb */ + var35 = ptr6[i]; + /* 2: mergebw */ + var39.i = ((orc_uint8) var34) | ((orc_uint8) var35 << 8); + /* 4: loadb */ + var37 = ptr4[i]; + /* 5: mergebw */ + var40.i = ((orc_uint8) var36) | ((orc_uint8) var37 << 8); + /* 6: mergewl */ + var38.i = ((orc_uint16) var40.i) | ((orc_uint16) var39.i << 16); + /* 7: storel */ + ptr0[i] = var38; + } + +} + +#else +static void +_backup_cogorc_getline_Y444 (OrcExecutor * ex) +{ + int i; + int n = ex->n; + orc_union32 *ORC_RESTRICT ptr0; + const orc_int8 *ORC_RESTRICT ptr4; + const orc_int8 *ORC_RESTRICT ptr5; + const orc_int8 *ORC_RESTRICT ptr6; + orc_int8 var34; + orc_int8 var35; + orc_int8 var36; + orc_int8 var37; + orc_union32 var38; + orc_union16 var39; + orc_union16 var40; + + ptr0 = (orc_union32 *) ex->arrays[0]; + ptr4 = (orc_int8 *) ex->arrays[4]; + ptr5 = (orc_int8 *) ex->arrays[5]; + ptr6 = (orc_int8 *) ex->arrays[6]; + + /* 3: loadpb */ + var36 = 0x000000ff; /* 255 or 1.25987e-321f */ + + for (i = 0; i < n; i++) { + /* 0: loadb */ + var34 = ptr5[i]; + /* 1: loadb */ + var35 = ptr6[i]; + /* 2: mergebw */ + var39.i = ((orc_uint8) var34) | ((orc_uint8) var35 << 8); + /* 4: loadb */ + var37 = ptr4[i]; + /* 5: mergebw */ + var40.i = ((orc_uint8) var36) | ((orc_uint8) var37 << 8); + /* 6: mergewl */ + var38.i = ((orc_uint16) var40.i) | ((orc_uint16) var39.i << 16); + /* 7: storel */ + ptr0[i] = var38; + } + +} + +void +cogorc_getline_Y444 (guint8 * d1, const guint8 * s1, const guint8 * s2, + const guint8 * s3, int n) +{ + OrcExecutor _ex, *ex = &_ex; + static int p_inited = 0; + static OrcProgram *p = 0; + void (*func) (OrcExecutor *); + + if (!p_inited) { + orc_once_mutex_lock (); + if (!p_inited) { + OrcCompileResult result; + + p = orc_program_new (); + orc_program_set_name (p, "cogorc_getline_Y444"); + orc_program_set_backup_function (p, _backup_cogorc_getline_Y444); + orc_program_add_destination (p, 4, "d1"); + orc_program_add_source (p, 1, "s1"); + orc_program_add_source (p, 1, "s2"); + orc_program_add_source (p, 1, "s3"); + orc_program_add_constant (p, 1, 0x000000ff, "c1"); + orc_program_add_constant (p, 0, 0x00000000, "c2"); + orc_program_add_constant (p, 0, 0x00000000, "c3"); + orc_program_add_constant (p, 0, 0x00000000, "c4"); + orc_program_add_constant (p, 0, 0x00000000, "c5"); + orc_program_add_constant (p, 0, 0x00000000, "c6"); + orc_program_add_constant (p, 0, 0x00000000, "c7"); + orc_program_add_constant (p, 0, 0x00000000, "c8"); + orc_program_add_temporary (p, 2, "t1"); + orc_program_add_temporary (p, 2, "t2"); + + orc_program_append_2 (p, "mergebw", 0, ORC_VAR_T1, ORC_VAR_S2, ORC_VAR_S3, + ORC_VAR_D1); + orc_program_append_2 (p, "mergebw", 0, ORC_VAR_T2, ORC_VAR_C1, ORC_VAR_S1, + ORC_VAR_D1); + orc_program_append_2 (p, "mergewl", 0, ORC_VAR_D1, ORC_VAR_T2, ORC_VAR_T1, + ORC_VAR_D1); + + result = orc_program_compile (p); + } + p_inited = TRUE; + orc_once_mutex_unlock (); + } + ex->program = p; + + ex->n = n; + ex->arrays[ORC_VAR_D1] = d1; + ex->arrays[ORC_VAR_S1] = (void *) s1; + ex->arrays[ORC_VAR_S2] = (void *) s2; + ex->arrays[ORC_VAR_S3] = (void *) s3; + + func = p->code_exec; + func (ex); +} +#endif + + +/* cogorc_getline_Y800 */ +#ifdef DISABLE_ORC +void +cogorc_getline_Y800 (guint8 * d1, const guint8 * s1, int n) +{ + int i; + orc_union32 *ORC_RESTRICT ptr0; + const orc_int8 *ORC_RESTRICT ptr4; + orc_int8 var33; + orc_int8 var34; + orc_union16 var35; + orc_union32 var36; + orc_union16 var37; + + ptr0 = (orc_union32 *) d1; + ptr4 = (orc_int8 *) s1; + + /* 0: loadpb */ + var33 = 0x000000ff; /* 255 or 1.25987e-321f */ + /* 3: loadpw */ + var35.i = 0x0000ffff; /* 65535 or 3.23786e-319f */ + + for (i = 0; i < n; i++) { + /* 1: loadb */ + var34 = ptr4[i]; + /* 2: mergebw */ + var37.i = ((orc_uint8) var33) | ((orc_uint8) var34 << 8); + /* 4: mergewl */ + var36.i = ((orc_uint16) var37.i) | ((orc_uint16) var35.i << 16); + /* 5: storel */ + ptr0[i] = var36; + } + +} + +#else +static void +_backup_cogorc_getline_Y800 (OrcExecutor * ex) +{ + int i; + int n = ex->n; + orc_union32 *ORC_RESTRICT ptr0; + const orc_int8 *ORC_RESTRICT ptr4; + orc_int8 var33; + orc_int8 var34; + orc_union16 var35; + orc_union32 var36; + orc_union16 var37; + + ptr0 = (orc_union32 *) ex->arrays[0]; + ptr4 = (orc_int8 *) ex->arrays[4]; + + /* 0: loadpb */ + var33 = 0x000000ff; /* 255 or 1.25987e-321f */ + /* 3: loadpw */ + var35.i = 0x0000ffff; /* 65535 or 3.23786e-319f */ + + for (i = 0; i < n; i++) { + /* 1: loadb */ + var34 = ptr4[i]; + /* 2: mergebw */ + var37.i = ((orc_uint8) var33) | ((orc_uint8) var34 << 8); + /* 4: mergewl */ + var36.i = ((orc_uint16) var37.i) | ((orc_uint16) var35.i << 16); + /* 5: storel */ + ptr0[i] = var36; + } + +} + +void +cogorc_getline_Y800 (guint8 * d1, const guint8 * s1, int n) +{ + OrcExecutor _ex, *ex = &_ex; + static int p_inited = 0; + static OrcProgram *p = 0; + void (*func) (OrcExecutor *); + + if (!p_inited) { + orc_once_mutex_lock (); + if (!p_inited) { + OrcCompileResult result; + + p = orc_program_new (); + orc_program_set_name (p, "cogorc_getline_Y800"); + orc_program_set_backup_function (p, _backup_cogorc_getline_Y800); + orc_program_add_destination (p, 4, "d1"); + orc_program_add_source (p, 1, "s1"); + orc_program_add_constant (p, 1, 0x000000ff, "c1"); + orc_program_add_constant (p, 2, 0x0000ffff, "c2"); + orc_program_add_constant (p, 0, 0x00000000, "c3"); + orc_program_add_constant (p, 0, 0x00000000, "c4"); + orc_program_add_constant (p, 0, 0x00000000, "c5"); + orc_program_add_constant (p, 0, 0x00000000, "c6"); + orc_program_add_constant (p, 0, 0x00000000, "c7"); + orc_program_add_constant (p, 0, 0x00000000, "c8"); + orc_program_add_temporary (p, 2, "t1"); + + orc_program_append_2 (p, "mergebw", 0, ORC_VAR_T1, ORC_VAR_C1, ORC_VAR_S1, + ORC_VAR_D1); + orc_program_append_2 (p, "mergewl", 0, ORC_VAR_D1, ORC_VAR_T1, ORC_VAR_C2, + ORC_VAR_D1); + + result = orc_program_compile (p); + } + p_inited = TRUE; + orc_once_mutex_unlock (); + } + ex->program = p; + + ex->n = n; + ex->arrays[ORC_VAR_D1] = d1; + ex->arrays[ORC_VAR_S1] = (void *) s1; + + func = p->code_exec; + func (ex); +} +#endif + + +/* cogorc_getline_BGRA */ +#ifdef DISABLE_ORC +void +cogorc_getline_BGRA (guint8 * d1, const guint8 * s1, int n) +{ + int i; + orc_union32 *ORC_RESTRICT ptr0; + const orc_union32 *ORC_RESTRICT ptr4; + orc_union32 var32; + orc_union32 var33; + + ptr0 = (orc_union32 *) d1; + ptr4 = (orc_union32 *) s1; + + + for (i = 0; i < n; i++) { + /* 0: loadl */ + var32 = ptr4[i]; + /* 1: swapl */ + var33.i = ORC_SWAP_L (var32.i); + /* 2: storel */ + ptr0[i] = var33; + } + +} + +#else +static void +_backup_cogorc_getline_BGRA (OrcExecutor * ex) +{ + int i; + int n = ex->n; + orc_union32 *ORC_RESTRICT ptr0; + const orc_union32 *ORC_RESTRICT ptr4; + orc_union32 var32; + orc_union32 var33; + + ptr0 = (orc_union32 *) ex->arrays[0]; + ptr4 = (orc_union32 *) ex->arrays[4]; + + + for (i = 0; i < n; i++) { + /* 0: loadl */ + var32 = ptr4[i]; + /* 1: swapl */ + var33.i = ORC_SWAP_L (var32.i); + /* 2: storel */ + ptr0[i] = var33; + } + +} + +void +cogorc_getline_BGRA (guint8 * d1, const guint8 * s1, int n) +{ + OrcExecutor _ex, *ex = &_ex; + static int p_inited = 0; + static OrcProgram *p = 0; + void (*func) (OrcExecutor *); + + if (!p_inited) { + orc_once_mutex_lock (); + if (!p_inited) { + OrcCompileResult result; + + p = orc_program_new (); + orc_program_set_name (p, "cogorc_getline_BGRA"); + orc_program_set_backup_function (p, _backup_cogorc_getline_BGRA); + orc_program_add_destination (p, 4, "d1"); + orc_program_add_source (p, 4, "s1"); + orc_program_add_constant (p, 0, 0x00000000, "c1"); + orc_program_add_constant (p, 0, 0x00000000, "c2"); + orc_program_add_constant (p, 0, 0x00000000, "c3"); + orc_program_add_constant (p, 0, 0x00000000, "c4"); + orc_program_add_constant (p, 0, 0x00000000, "c5"); + orc_program_add_constant (p, 0, 0x00000000, "c6"); + orc_program_add_constant (p, 0, 0x00000000, "c7"); + orc_program_add_constant (p, 0, 0x00000000, "c8"); + + orc_program_append_2 (p, "swapl", 0, ORC_VAR_D1, ORC_VAR_S1, ORC_VAR_D1, + ORC_VAR_D1); + + result = orc_program_compile (p); + } + p_inited = TRUE; + orc_once_mutex_unlock (); + } + ex->program = p; + + ex->n = n; + ex->arrays[ORC_VAR_D1] = d1; + ex->arrays[ORC_VAR_S1] = (void *) s1; + + func = p->code_exec; + func (ex); +} +#endif + + +/* cogorc_getline_ABGR */ +#ifdef DISABLE_ORC +void +cogorc_getline_ABGR (guint8 * d1, const guint8 * s1, int n) +{ + int i; + orc_union32 *ORC_RESTRICT ptr0; + const orc_union32 *ORC_RESTRICT ptr4; + orc_union32 var40; + orc_union32 var41; + orc_union16 var42; + orc_union16 var43; + orc_int8 var44; + orc_int8 var45; + orc_int8 var46; + orc_int8 var47; + orc_union16 var48; + orc_union16 var49; + + ptr0 = (orc_union32 *) d1; + ptr4 = (orc_union32 *) s1; + + + for (i = 0; i < n; i++) { + /* 0: loadl */ + var40 = ptr4[i]; + /* 1: splitlw */ + var42.i = (var40.i >> 16) & 0xffff; + var43.i = var40.i & 0xffff; + /* 2: splitwb */ + var44 = (var42.i >> 8) & 0xff; + var45 = var42.i & 0xff; + /* 3: splitwb */ + var46 = (var43.i >> 8) & 0xff; + var47 = var43.i & 0xff; + /* 4: mergebw */ + var48.i = ((orc_uint8) var47) | ((orc_uint8) var44 << 8); + /* 5: mergebw */ + var49.i = ((orc_uint8) var45) | ((orc_uint8) var46 << 8); + /* 6: mergewl */ + var41.i = ((orc_uint16) var48.i) | ((orc_uint16) var49.i << 16); + /* 7: storel */ + ptr0[i] = var41; + } + +} + +#else +static void +_backup_cogorc_getline_ABGR (OrcExecutor * ex) +{ + int i; + int n = ex->n; + orc_union32 *ORC_RESTRICT ptr0; + const orc_union32 *ORC_RESTRICT ptr4; + orc_union32 var40; + orc_union32 var41; + orc_union16 var42; + orc_union16 var43; + orc_int8 var44; + orc_int8 var45; + orc_int8 var46; + orc_int8 var47; + orc_union16 var48; + orc_union16 var49; + + ptr0 = (orc_union32 *) ex->arrays[0]; + ptr4 = (orc_union32 *) ex->arrays[4]; + + + for (i = 0; i < n; i++) { + /* 0: loadl */ + var40 = ptr4[i]; + /* 1: splitlw */ + var42.i = (var40.i >> 16) & 0xffff; + var43.i = var40.i & 0xffff; + /* 2: splitwb */ + var44 = (var42.i >> 8) & 0xff; + var45 = var42.i & 0xff; + /* 3: splitwb */ + var46 = (var43.i >> 8) & 0xff; + var47 = var43.i & 0xff; + /* 4: mergebw */ + var48.i = ((orc_uint8) var47) | ((orc_uint8) var44 << 8); + /* 5: mergebw */ + var49.i = ((orc_uint8) var45) | ((orc_uint8) var46 << 8); + /* 6: mergewl */ + var41.i = ((orc_uint16) var48.i) | ((orc_uint16) var49.i << 16); + /* 7: storel */ + ptr0[i] = var41; + } + +} + +void +cogorc_getline_ABGR (guint8 * d1, const guint8 * s1, int n) +{ + OrcExecutor _ex, *ex = &_ex; + static int p_inited = 0; + static OrcProgram *p = 0; + void (*func) (OrcExecutor *); + + if (!p_inited) { + orc_once_mutex_lock (); + if (!p_inited) { + OrcCompileResult result; + + p = orc_program_new (); + orc_program_set_name (p, "cogorc_getline_ABGR"); + orc_program_set_backup_function (p, _backup_cogorc_getline_ABGR); + orc_program_add_destination (p, 4, "d1"); + orc_program_add_source (p, 4, "s1"); + orc_program_add_constant (p, 0, 0x00000000, "c1"); + orc_program_add_constant (p, 0, 0x00000000, "c2"); + orc_program_add_constant (p, 0, 0x00000000, "c3"); + orc_program_add_constant (p, 0, 0x00000000, "c4"); + orc_program_add_constant (p, 0, 0x00000000, "c5"); + orc_program_add_constant (p, 0, 0x00000000, "c6"); + orc_program_add_constant (p, 0, 0x00000000, "c7"); + orc_program_add_constant (p, 0, 0x00000000, "c8"); + orc_program_add_temporary (p, 1, "t1"); + orc_program_add_temporary (p, 1, "t2"); + orc_program_add_temporary (p, 1, "t3"); + orc_program_add_temporary (p, 1, "t4"); + orc_program_add_temporary (p, 2, "t5"); + orc_program_add_temporary (p, 2, "t6"); + orc_program_add_temporary (p, 2, "t7"); + orc_program_add_temporary (p, 2, "t8"); + + orc_program_append_2 (p, "splitlw", 0, ORC_VAR_T5, ORC_VAR_T6, ORC_VAR_S1, + ORC_VAR_D1); + orc_program_append_2 (p, "splitwb", 0, ORC_VAR_T2, ORC_VAR_T3, ORC_VAR_T5, + ORC_VAR_D1); + orc_program_append_2 (p, "splitwb", 0, ORC_VAR_T4, ORC_VAR_T1, ORC_VAR_T6, + ORC_VAR_D1); + orc_program_append_2 (p, "mergebw", 0, ORC_VAR_T7, ORC_VAR_T1, ORC_VAR_T2, + ORC_VAR_D1); + orc_program_append_2 (p, "mergebw", 0, ORC_VAR_T8, ORC_VAR_T3, ORC_VAR_T4, + ORC_VAR_D1); + orc_program_append_2 (p, "mergewl", 0, ORC_VAR_D1, ORC_VAR_T7, ORC_VAR_T8, + ORC_VAR_D1); + + result = orc_program_compile (p); + } + p_inited = TRUE; + orc_once_mutex_unlock (); + } + ex->program = p; + + ex->n = n; + ex->arrays[ORC_VAR_D1] = d1; + ex->arrays[ORC_VAR_S1] = (void *) s1; + + func = p->code_exec; + func (ex); +} +#endif + + +/* cogorc_getline_RGBA */ +#ifdef DISABLE_ORC +void +cogorc_getline_RGBA (guint8 * d1, const guint8 * s1, int n) +{ + int i; + orc_union32 *ORC_RESTRICT ptr0; + const orc_union32 *ORC_RESTRICT ptr4; + orc_union32 var40; + orc_union32 var41; + orc_union16 var42; + orc_union16 var43; + orc_int8 var44; + orc_int8 var45; + orc_int8 var46; + orc_int8 var47; + orc_union16 var48; + orc_union16 var49; + + ptr0 = (orc_union32 *) d1; + ptr4 = (orc_union32 *) s1; + + + for (i = 0; i < n; i++) { + /* 0: loadl */ + var40 = ptr4[i]; + /* 1: splitlw */ + var42.i = (var40.i >> 16) & 0xffff; + var43.i = var40.i & 0xffff; + /* 2: splitwb */ + var44 = (var43.i >> 8) & 0xff; + var45 = var43.i & 0xff; + /* 3: splitwb */ + var46 = (var42.i >> 8) & 0xff; + var47 = var42.i & 0xff; + /* 4: mergebw */ + var48.i = ((orc_uint8) var46) | ((orc_uint8) var45 << 8); + /* 5: mergebw */ + var49.i = ((orc_uint8) var44) | ((orc_uint8) var47 << 8); + /* 6: mergewl */ + var41.i = ((orc_uint16) var48.i) | ((orc_uint16) var49.i << 16); + /* 7: storel */ + ptr0[i] = var41; + } + +} + +#else +static void +_backup_cogorc_getline_RGBA (OrcExecutor * ex) +{ + int i; + int n = ex->n; + orc_union32 *ORC_RESTRICT ptr0; + const orc_union32 *ORC_RESTRICT ptr4; + orc_union32 var40; + orc_union32 var41; + orc_union16 var42; + orc_union16 var43; + orc_int8 var44; + orc_int8 var45; + orc_int8 var46; + orc_int8 var47; + orc_union16 var48; + orc_union16 var49; + + ptr0 = (orc_union32 *) ex->arrays[0]; + ptr4 = (orc_union32 *) ex->arrays[4]; + + + for (i = 0; i < n; i++) { + /* 0: loadl */ + var40 = ptr4[i]; + /* 1: splitlw */ + var42.i = (var40.i >> 16) & 0xffff; + var43.i = var40.i & 0xffff; + /* 2: splitwb */ + var44 = (var43.i >> 8) & 0xff; + var45 = var43.i & 0xff; + /* 3: splitwb */ + var46 = (var42.i >> 8) & 0xff; + var47 = var42.i & 0xff; + /* 4: mergebw */ + var48.i = ((orc_uint8) var46) | ((orc_uint8) var45 << 8); + /* 5: mergebw */ + var49.i = ((orc_uint8) var44) | ((orc_uint8) var47 << 8); + /* 6: mergewl */ + var41.i = ((orc_uint16) var48.i) | ((orc_uint16) var49.i << 16); + /* 7: storel */ + ptr0[i] = var41; + } + +} + +void +cogorc_getline_RGBA (guint8 * d1, const guint8 * s1, int n) +{ + OrcExecutor _ex, *ex = &_ex; + static int p_inited = 0; + static OrcProgram *p = 0; + void (*func) (OrcExecutor *); + + if (!p_inited) { + orc_once_mutex_lock (); + if (!p_inited) { + OrcCompileResult result; + + p = orc_program_new (); + orc_program_set_name (p, "cogorc_getline_RGBA"); + orc_program_set_backup_function (p, _backup_cogorc_getline_RGBA); + orc_program_add_destination (p, 4, "d1"); + orc_program_add_source (p, 4, "s1"); + orc_program_add_constant (p, 0, 0x00000000, "c1"); + orc_program_add_constant (p, 0, 0x00000000, "c2"); + orc_program_add_constant (p, 0, 0x00000000, "c3"); + orc_program_add_constant (p, 0, 0x00000000, "c4"); + orc_program_add_constant (p, 0, 0x00000000, "c5"); + orc_program_add_constant (p, 0, 0x00000000, "c6"); + orc_program_add_constant (p, 0, 0x00000000, "c7"); + orc_program_add_constant (p, 0, 0x00000000, "c8"); + orc_program_add_temporary (p, 1, "t1"); + orc_program_add_temporary (p, 1, "t2"); + orc_program_add_temporary (p, 1, "t3"); + orc_program_add_temporary (p, 1, "t4"); + orc_program_add_temporary (p, 2, "t5"); + orc_program_add_temporary (p, 2, "t6"); + orc_program_add_temporary (p, 2, "t7"); + orc_program_add_temporary (p, 2, "t8"); + + orc_program_append_2 (p, "splitlw", 0, ORC_VAR_T6, ORC_VAR_T5, ORC_VAR_S1, + ORC_VAR_D1); + orc_program_append_2 (p, "splitwb", 0, ORC_VAR_T3, ORC_VAR_T2, ORC_VAR_T5, + ORC_VAR_D1); + orc_program_append_2 (p, "splitwb", 0, ORC_VAR_T1, ORC_VAR_T4, ORC_VAR_T6, + ORC_VAR_D1); + orc_program_append_2 (p, "mergebw", 0, ORC_VAR_T7, ORC_VAR_T1, ORC_VAR_T2, + ORC_VAR_D1); + orc_program_append_2 (p, "mergebw", 0, ORC_VAR_T8, ORC_VAR_T3, ORC_VAR_T4, + ORC_VAR_D1); + orc_program_append_2 (p, "mergewl", 0, ORC_VAR_D1, ORC_VAR_T7, ORC_VAR_T8, + ORC_VAR_D1); + + result = orc_program_compile (p); + } + p_inited = TRUE; + orc_once_mutex_unlock (); + } + ex->program = p; + + ex->n = n; + ex->arrays[ORC_VAR_D1] = d1; + ex->arrays[ORC_VAR_S1] = (void *) s1; + + func = p->code_exec; + func (ex); +} +#endif + + +/* cogorc_getline_NV12 */ +#ifdef DISABLE_ORC +void +cogorc_getline_NV12 (guint8 * d1, const guint8 * s1, const guint8 * s2, int n) +{ + int i; + orc_union64 *ORC_RESTRICT ptr0; + const orc_union16 *ORC_RESTRICT ptr4; + const orc_union16 *ORC_RESTRICT ptr5; + orc_union16 var34; + orc_union16 var35; + orc_union16 var36; + orc_union16 var37; + orc_union64 var38; + orc_union32 var39; + orc_union32 var40; + + ptr0 = (orc_union64 *) d1; + ptr4 = (orc_union16 *) s1; + ptr5 = (orc_union16 *) s2; + + /* 3: loadpb */ + var36.x2[0] = 0x000000ff; /* 255 or 1.25987e-321f */ + var36.x2[1] = 0x000000ff; /* 255 or 1.25987e-321f */ + + for (i = 0; i < n; i++) { + /* 0: loadw */ + var34 = ptr5[i]; + /* 1: loadw */ + var35 = ptr5[i]; + /* 2: mergewl */ + var39.i = ((orc_uint16) var34.i) | ((orc_uint16) var35.i << 16); + /* 4: loadw */ + var37 = ptr4[i]; + /* 5: mergebw */ + var40.x2[0] = ((orc_uint8) var36.x2[0]) | ((orc_uint8) var37.x2[0] << 8); + var40.x2[1] = ((orc_uint8) var36.x2[1]) | ((orc_uint8) var37.x2[1] << 8); + /* 6: mergewl */ + var38.x2[0] = ((orc_uint16) var40.x2[0]) | ((orc_uint16) var39.x2[0] << 16); + var38.x2[1] = ((orc_uint16) var40.x2[1]) | ((orc_uint16) var39.x2[1] << 16); + /* 7: storeq */ + ptr0[i] = var38; + } + +} + +#else +static void +_backup_cogorc_getline_NV12 (OrcExecutor * ex) +{ + int i; + int n = ex->n; + orc_union64 *ORC_RESTRICT ptr0; + const orc_union16 *ORC_RESTRICT ptr4; + const orc_union16 *ORC_RESTRICT ptr5; + orc_union16 var34; + orc_union16 var35; + orc_union16 var36; + orc_union16 var37; + orc_union64 var38; + orc_union32 var39; + orc_union32 var40; + + ptr0 = (orc_union64 *) ex->arrays[0]; + ptr4 = (orc_union16 *) ex->arrays[4]; + ptr5 = (orc_union16 *) ex->arrays[5]; + + /* 3: loadpb */ + var36.x2[0] = 0x000000ff; /* 255 or 1.25987e-321f */ + var36.x2[1] = 0x000000ff; /* 255 or 1.25987e-321f */ + + for (i = 0; i < n; i++) { + /* 0: loadw */ + var34 = ptr5[i]; + /* 1: loadw */ + var35 = ptr5[i]; + /* 2: mergewl */ + var39.i = ((orc_uint16) var34.i) | ((orc_uint16) var35.i << 16); + /* 4: loadw */ + var37 = ptr4[i]; + /* 5: mergebw */ + var40.x2[0] = ((orc_uint8) var36.x2[0]) | ((orc_uint8) var37.x2[0] << 8); + var40.x2[1] = ((orc_uint8) var36.x2[1]) | ((orc_uint8) var37.x2[1] << 8); + /* 6: mergewl */ + var38.x2[0] = ((orc_uint16) var40.x2[0]) | ((orc_uint16) var39.x2[0] << 16); + var38.x2[1] = ((orc_uint16) var40.x2[1]) | ((orc_uint16) var39.x2[1] << 16); + /* 7: storeq */ + ptr0[i] = var38; + } + +} + +void +cogorc_getline_NV12 (guint8 * d1, const guint8 * s1, const guint8 * s2, int n) +{ + OrcExecutor _ex, *ex = &_ex; + static int p_inited = 0; + static OrcProgram *p = 0; + void (*func) (OrcExecutor *); + + if (!p_inited) { + orc_once_mutex_lock (); + if (!p_inited) { + OrcCompileResult result; + + p = orc_program_new (); + orc_program_set_name (p, "cogorc_getline_NV12"); + orc_program_set_backup_function (p, _backup_cogorc_getline_NV12); + orc_program_add_destination (p, 8, "d1"); + orc_program_add_source (p, 2, "s1"); + orc_program_add_source (p, 2, "s2"); + orc_program_add_constant (p, 1, 0x000000ff, "c1"); + orc_program_add_constant (p, 0, 0x00000000, "c2"); + orc_program_add_constant (p, 0, 0x00000000, "c3"); + orc_program_add_constant (p, 0, 0x00000000, "c4"); + orc_program_add_constant (p, 0, 0x00000000, "c5"); + orc_program_add_constant (p, 0, 0x00000000, "c6"); + orc_program_add_constant (p, 0, 0x00000000, "c7"); + orc_program_add_constant (p, 0, 0x00000000, "c8"); + orc_program_add_temporary (p, 4, "t1"); + orc_program_add_temporary (p, 4, "t2"); + + orc_program_append_2 (p, "mergewl", 0, ORC_VAR_T2, ORC_VAR_S2, ORC_VAR_S2, + ORC_VAR_D1); + orc_program_append_2 (p, "mergebw", 1, ORC_VAR_T1, ORC_VAR_C1, ORC_VAR_S1, + ORC_VAR_D1); + orc_program_append_2 (p, "mergewl", 1, ORC_VAR_D1, ORC_VAR_T1, ORC_VAR_T2, + ORC_VAR_D1); + + result = orc_program_compile (p); + } + p_inited = TRUE; + orc_once_mutex_unlock (); + } + ex->program = p; + + ex->n = n; + ex->arrays[ORC_VAR_D1] = d1; + ex->arrays[ORC_VAR_S1] = (void *) s1; + ex->arrays[ORC_VAR_S2] = (void *) s2; + + func = p->code_exec; + func (ex); +} +#endif + + +/* cogorc_getline_NV21 */ +#ifdef DISABLE_ORC +void +cogorc_getline_NV21 (guint8 * d1, const guint8 * s1, const guint8 * s2, int n) +{ + int i; + orc_union64 *ORC_RESTRICT ptr0; + const orc_union16 *ORC_RESTRICT ptr4; + const orc_union16 *ORC_RESTRICT ptr5; + orc_union16 var35; + orc_union16 var36; + orc_union16 var37; + orc_union64 var38; + orc_union16 var39; + orc_union32 var40; + orc_union32 var41; + + ptr0 = (orc_union64 *) d1; + ptr4 = (orc_union16 *) s1; + ptr5 = (orc_union16 *) s2; + + /* 3: loadpb */ + var36.x2[0] = 0x000000ff; /* 255 or 1.25987e-321f */ + var36.x2[1] = 0x000000ff; /* 255 or 1.25987e-321f */ + + for (i = 0; i < n; i++) { + /* 0: loadw */ + var35 = ptr5[i]; + /* 1: swapw */ + var39.i = ORC_SWAP_W (var35.i); + /* 2: mergewl */ + var40.i = ((orc_uint16) var39.i) | ((orc_uint16) var39.i << 16); + /* 4: loadw */ + var37 = ptr4[i]; + /* 5: mergebw */ + var41.x2[0] = ((orc_uint8) var36.x2[0]) | ((orc_uint8) var37.x2[0] << 8); + var41.x2[1] = ((orc_uint8) var36.x2[1]) | ((orc_uint8) var37.x2[1] << 8); + /* 6: mergewl */ + var38.x2[0] = ((orc_uint16) var41.x2[0]) | ((orc_uint16) var40.x2[0] << 16); + var38.x2[1] = ((orc_uint16) var41.x2[1]) | ((orc_uint16) var40.x2[1] << 16); + /* 7: storeq */ + ptr0[i] = var38; + } + +} + +#else +static void +_backup_cogorc_getline_NV21 (OrcExecutor * ex) +{ + int i; + int n = ex->n; + orc_union64 *ORC_RESTRICT ptr0; + const orc_union16 *ORC_RESTRICT ptr4; + const orc_union16 *ORC_RESTRICT ptr5; + orc_union16 var35; + orc_union16 var36; + orc_union16 var37; + orc_union64 var38; + orc_union16 var39; + orc_union32 var40; + orc_union32 var41; + + ptr0 = (orc_union64 *) ex->arrays[0]; + ptr4 = (orc_union16 *) ex->arrays[4]; + ptr5 = (orc_union16 *) ex->arrays[5]; + + /* 3: loadpb */ + var36.x2[0] = 0x000000ff; /* 255 or 1.25987e-321f */ + var36.x2[1] = 0x000000ff; /* 255 or 1.25987e-321f */ + + for (i = 0; i < n; i++) { + /* 0: loadw */ + var35 = ptr5[i]; + /* 1: swapw */ + var39.i = ORC_SWAP_W (var35.i); + /* 2: mergewl */ + var40.i = ((orc_uint16) var39.i) | ((orc_uint16) var39.i << 16); + /* 4: loadw */ + var37 = ptr4[i]; + /* 5: mergebw */ + var41.x2[0] = ((orc_uint8) var36.x2[0]) | ((orc_uint8) var37.x2[0] << 8); + var41.x2[1] = ((orc_uint8) var36.x2[1]) | ((orc_uint8) var37.x2[1] << 8); + /* 6: mergewl */ + var38.x2[0] = ((orc_uint16) var41.x2[0]) | ((orc_uint16) var40.x2[0] << 16); + var38.x2[1] = ((orc_uint16) var41.x2[1]) | ((orc_uint16) var40.x2[1] << 16); + /* 7: storeq */ + ptr0[i] = var38; + } + +} + +void +cogorc_getline_NV21 (guint8 * d1, const guint8 * s1, const guint8 * s2, int n) +{ + OrcExecutor _ex, *ex = &_ex; + static int p_inited = 0; + static OrcProgram *p = 0; + void (*func) (OrcExecutor *); + + if (!p_inited) { + orc_once_mutex_lock (); + if (!p_inited) { + OrcCompileResult result; + + p = orc_program_new (); + orc_program_set_name (p, "cogorc_getline_NV21"); + orc_program_set_backup_function (p, _backup_cogorc_getline_NV21); + orc_program_add_destination (p, 8, "d1"); + orc_program_add_source (p, 2, "s1"); + orc_program_add_source (p, 2, "s2"); + orc_program_add_constant (p, 1, 0x000000ff, "c1"); + orc_program_add_constant (p, 0, 0x00000000, "c2"); + orc_program_add_constant (p, 0, 0x00000000, "c3"); + orc_program_add_constant (p, 0, 0x00000000, "c4"); + orc_program_add_constant (p, 0, 0x00000000, "c5"); + orc_program_add_constant (p, 0, 0x00000000, "c6"); + orc_program_add_constant (p, 0, 0x00000000, "c7"); + orc_program_add_constant (p, 0, 0x00000000, "c8"); + orc_program_add_temporary (p, 2, "t1"); + orc_program_add_temporary (p, 4, "t2"); + orc_program_add_temporary (p, 4, "t3"); + + orc_program_append_2 (p, "swapw", 0, ORC_VAR_T1, ORC_VAR_S2, ORC_VAR_D1, + ORC_VAR_D1); + orc_program_append_2 (p, "mergewl", 0, ORC_VAR_T3, ORC_VAR_T1, ORC_VAR_T1, + ORC_VAR_D1); + orc_program_append_2 (p, "mergebw", 1, ORC_VAR_T2, ORC_VAR_C1, ORC_VAR_S1, + ORC_VAR_D1); + orc_program_append_2 (p, "mergewl", 1, ORC_VAR_D1, ORC_VAR_T2, ORC_VAR_T3, + ORC_VAR_D1); + + result = orc_program_compile (p); + } + p_inited = TRUE; + orc_once_mutex_unlock (); + } + ex->program = p; + + ex->n = n; + ex->arrays[ORC_VAR_D1] = d1; + ex->arrays[ORC_VAR_S1] = (void *) s1; + ex->arrays[ORC_VAR_S2] = (void *) s2; + + func = p->code_exec; + func (ex); +} +#endif + + +/* cogorc_putline_I420 */ +#ifdef DISABLE_ORC +void +cogorc_putline_I420 (guint8 * d1, guint8 * d2, guint8 * d3, const guint8 * s1, + int n) +{ + int i; + orc_union16 *ORC_RESTRICT ptr0; + orc_int8 *ORC_RESTRICT ptr1; + orc_int8 *ORC_RESTRICT ptr2; + const orc_union64 *ORC_RESTRICT ptr4; + orc_union64 var38; + orc_union16 var39; + orc_int8 var40; + orc_int8 var41; + orc_union32 var42; + orc_union32 var43; + orc_union16 var44; + orc_union16 var45; + orc_int8 var46; + orc_int8 var47; + orc_int8 var48; + orc_int8 var49; + + ptr0 = (orc_union16 *) d1; + ptr1 = (orc_int8 *) d2; + ptr2 = (orc_int8 *) d3; + ptr4 = (orc_union64 *) s1; + + + for (i = 0; i < n; i++) { + /* 0: loadq */ + var38 = ptr4[i]; + /* 1: splitlw */ + var42.x2[0] = (var38.x2[0] >> 16) & 0xffff; + var43.x2[0] = var38.x2[0] & 0xffff; + var42.x2[1] = (var38.x2[1] >> 16) & 0xffff; + var43.x2[1] = var38.x2[1] & 0xffff; + /* 2: select1wb */ + var39.x2[0] = ((orc_uint16) var43.x2[0] >> 8) & 0xff; + var39.x2[1] = ((orc_uint16) var43.x2[1] >> 8) & 0xff; + /* 3: storew */ + ptr0[i] = var39; + /* 4: splitwb */ + var44.x2[0] = (var42.x2[0] >> 8) & 0xff; + var45.x2[0] = var42.x2[0] & 0xff; + var44.x2[1] = (var42.x2[1] >> 8) & 0xff; + var45.x2[1] = var42.x2[1] & 0xff; + /* 5: splitwb */ + var46 = (var45.i >> 8) & 0xff; + var47 = var45.i & 0xff; + /* 6: avgub */ + var40 = ((orc_uint8) var46 + (orc_uint8) var47 + 1) >> 1; + /* 7: storeb */ + ptr1[i] = var40; + /* 8: splitwb */ + var48 = (var44.i >> 8) & 0xff; + var49 = var44.i & 0xff; + /* 9: avgub */ + var41 = ((orc_uint8) var48 + (orc_uint8) var49 + 1) >> 1; + /* 10: storeb */ + ptr2[i] = var41; + } + +} + +#else +static void +_backup_cogorc_putline_I420 (OrcExecutor * ex) +{ + int i; + int n = ex->n; + orc_union16 *ORC_RESTRICT ptr0; + orc_int8 *ORC_RESTRICT ptr1; + orc_int8 *ORC_RESTRICT ptr2; + const orc_union64 *ORC_RESTRICT ptr4; + orc_union64 var38; + orc_union16 var39; + orc_int8 var40; + orc_int8 var41; + orc_union32 var42; + orc_union32 var43; + orc_union16 var44; + orc_union16 var45; + orc_int8 var46; + orc_int8 var47; + orc_int8 var48; + orc_int8 var49; + + ptr0 = (orc_union16 *) ex->arrays[0]; + ptr1 = (orc_int8 *) ex->arrays[1]; + ptr2 = (orc_int8 *) ex->arrays[2]; + ptr4 = (orc_union64 *) ex->arrays[4]; + + + for (i = 0; i < n; i++) { + /* 0: loadq */ + var38 = ptr4[i]; + /* 1: splitlw */ + var42.x2[0] = (var38.x2[0] >> 16) & 0xffff; + var43.x2[0] = var38.x2[0] & 0xffff; + var42.x2[1] = (var38.x2[1] >> 16) & 0xffff; + var43.x2[1] = var38.x2[1] & 0xffff; + /* 2: select1wb */ + var39.x2[0] = ((orc_uint16) var43.x2[0] >> 8) & 0xff; + var39.x2[1] = ((orc_uint16) var43.x2[1] >> 8) & 0xff; + /* 3: storew */ + ptr0[i] = var39; + /* 4: splitwb */ + var44.x2[0] = (var42.x2[0] >> 8) & 0xff; + var45.x2[0] = var42.x2[0] & 0xff; + var44.x2[1] = (var42.x2[1] >> 8) & 0xff; + var45.x2[1] = var42.x2[1] & 0xff; + /* 5: splitwb */ + var46 = (var45.i >> 8) & 0xff; + var47 = var45.i & 0xff; + /* 6: avgub */ + var40 = ((orc_uint8) var46 + (orc_uint8) var47 + 1) >> 1; + /* 7: storeb */ + ptr1[i] = var40; + /* 8: splitwb */ + var48 = (var44.i >> 8) & 0xff; + var49 = var44.i & 0xff; + /* 9: avgub */ + var41 = ((orc_uint8) var48 + (orc_uint8) var49 + 1) >> 1; + /* 10: storeb */ + ptr2[i] = var41; + } + +} + +void +cogorc_putline_I420 (guint8 * d1, guint8 * d2, guint8 * d3, const guint8 * s1, + int n) +{ + OrcExecutor _ex, *ex = &_ex; + static int p_inited = 0; + static OrcProgram *p = 0; + void (*func) (OrcExecutor *); + + if (!p_inited) { + orc_once_mutex_lock (); + if (!p_inited) { + OrcCompileResult result; + + p = orc_program_new (); + orc_program_set_name (p, "cogorc_putline_I420"); + orc_program_set_backup_function (p, _backup_cogorc_putline_I420); + orc_program_add_destination (p, 2, "d1"); + orc_program_add_destination (p, 1, "d2"); + orc_program_add_destination (p, 1, "d3"); + orc_program_add_source (p, 8, "s1"); + orc_program_add_constant (p, 0, 0x00000000, "c1"); + orc_program_add_constant (p, 0, 0x00000000, "c2"); + orc_program_add_constant (p, 0, 0x00000000, "c3"); + orc_program_add_constant (p, 0, 0x00000000, "c4"); + orc_program_add_constant (p, 0, 0x00000000, "c5"); + orc_program_add_constant (p, 0, 0x00000000, "c6"); + orc_program_add_constant (p, 0, 0x00000000, "c7"); + orc_program_add_constant (p, 0, 0x00000000, "c8"); + orc_program_add_temporary (p, 4, "t1"); + orc_program_add_temporary (p, 4, "t2"); + orc_program_add_temporary (p, 2, "t3"); + orc_program_add_temporary (p, 2, "t4"); + orc_program_add_temporary (p, 1, "t5"); + orc_program_add_temporary (p, 1, "t6"); + + orc_program_append_2 (p, "splitlw", 1, ORC_VAR_T2, ORC_VAR_T1, ORC_VAR_S1, + ORC_VAR_D1); + orc_program_append_2 (p, "select1wb", 1, ORC_VAR_D1, ORC_VAR_T1, + ORC_VAR_D1, ORC_VAR_D1); + orc_program_append_2 (p, "splitwb", 1, ORC_VAR_T4, ORC_VAR_T3, ORC_VAR_T2, + ORC_VAR_D1); + orc_program_append_2 (p, "splitwb", 0, ORC_VAR_T5, ORC_VAR_T6, ORC_VAR_T3, + ORC_VAR_D1); + orc_program_append_2 (p, "avgub", 0, ORC_VAR_D2, ORC_VAR_T5, ORC_VAR_T6, + ORC_VAR_D1); + orc_program_append_2 (p, "splitwb", 0, ORC_VAR_T5, ORC_VAR_T6, ORC_VAR_T4, + ORC_VAR_D1); + orc_program_append_2 (p, "avgub", 0, ORC_VAR_D3, ORC_VAR_T5, ORC_VAR_T6, + ORC_VAR_D1); + + result = orc_program_compile (p); + } + p_inited = TRUE; + orc_once_mutex_unlock (); + } + ex->program = p; + + ex->n = n; + ex->arrays[ORC_VAR_D1] = d1; + ex->arrays[ORC_VAR_D2] = d2; + ex->arrays[ORC_VAR_D3] = d3; + ex->arrays[ORC_VAR_S1] = (void *) s1; + + func = p->code_exec; + func (ex); +} +#endif + + +/* cogorc_putline_YUY2 */ +#ifdef DISABLE_ORC +void +cogorc_putline_YUY2 (guint8 * d1, const guint8 * s1, int n) +{ + int i; + orc_union32 *ORC_RESTRICT ptr0; + const orc_union64 *ORC_RESTRICT ptr4; + orc_union64 var37; + orc_union32 var38; + orc_union32 var39; + orc_union32 var40; + orc_union16 var41; + orc_union16 var42; + orc_union16 var43; + orc_union16 var44; + + ptr0 = (orc_union32 *) d1; + ptr4 = (orc_union64 *) s1; + + + for (i = 0; i < n; i++) { + /* 0: loadq */ + var37 = ptr4[i]; + /* 1: splitlw */ + var39.x2[0] = (var37.x2[0] >> 16) & 0xffff; + var40.x2[0] = var37.x2[0] & 0xffff; + var39.x2[1] = (var37.x2[1] >> 16) & 0xffff; + var40.x2[1] = var37.x2[1] & 0xffff; + /* 2: splitlw */ + var41.i = (var39.i >> 16) & 0xffff; + var42.i = var39.i & 0xffff; + /* 3: avgub */ + var43.x2[0] = ((orc_uint8) var41.x2[0] + (orc_uint8) var42.x2[0] + 1) >> 1; + var43.x2[1] = ((orc_uint8) var41.x2[1] + (orc_uint8) var42.x2[1] + 1) >> 1; + /* 4: select1wb */ + var44.x2[0] = ((orc_uint16) var40.x2[0] >> 8) & 0xff; + var44.x2[1] = ((orc_uint16) var40.x2[1] >> 8) & 0xff; + /* 5: mergebw */ + var38.x2[0] = ((orc_uint8) var44.x2[0]) | ((orc_uint8) var43.x2[0] << 8); + var38.x2[1] = ((orc_uint8) var44.x2[1]) | ((orc_uint8) var43.x2[1] << 8); + /* 6: storel */ + ptr0[i] = var38; + } + +} + +#else +static void +_backup_cogorc_putline_YUY2 (OrcExecutor * ex) +{ + int i; + int n = ex->n; + orc_union32 *ORC_RESTRICT ptr0; + const orc_union64 *ORC_RESTRICT ptr4; + orc_union64 var37; + orc_union32 var38; + orc_union32 var39; + orc_union32 var40; + orc_union16 var41; + orc_union16 var42; + orc_union16 var43; + orc_union16 var44; + + ptr0 = (orc_union32 *) ex->arrays[0]; + ptr4 = (orc_union64 *) ex->arrays[4]; + + + for (i = 0; i < n; i++) { + /* 0: loadq */ + var37 = ptr4[i]; + /* 1: splitlw */ + var39.x2[0] = (var37.x2[0] >> 16) & 0xffff; + var40.x2[0] = var37.x2[0] & 0xffff; + var39.x2[1] = (var37.x2[1] >> 16) & 0xffff; + var40.x2[1] = var37.x2[1] & 0xffff; + /* 2: splitlw */ + var41.i = (var39.i >> 16) & 0xffff; + var42.i = var39.i & 0xffff; + /* 3: avgub */ + var43.x2[0] = ((orc_uint8) var41.x2[0] + (orc_uint8) var42.x2[0] + 1) >> 1; + var43.x2[1] = ((orc_uint8) var41.x2[1] + (orc_uint8) var42.x2[1] + 1) >> 1; + /* 4: select1wb */ + var44.x2[0] = ((orc_uint16) var40.x2[0] >> 8) & 0xff; + var44.x2[1] = ((orc_uint16) var40.x2[1] >> 8) & 0xff; + /* 5: mergebw */ + var38.x2[0] = ((orc_uint8) var44.x2[0]) | ((orc_uint8) var43.x2[0] << 8); + var38.x2[1] = ((orc_uint8) var44.x2[1]) | ((orc_uint8) var43.x2[1] << 8); + /* 6: storel */ + ptr0[i] = var38; + } + +} + +void +cogorc_putline_YUY2 (guint8 * d1, const guint8 * s1, int n) +{ + OrcExecutor _ex, *ex = &_ex; + static int p_inited = 0; + static OrcProgram *p = 0; + void (*func) (OrcExecutor *); + + if (!p_inited) { + orc_once_mutex_lock (); + if (!p_inited) { + OrcCompileResult result; + + p = orc_program_new (); + orc_program_set_name (p, "cogorc_putline_YUY2"); + orc_program_set_backup_function (p, _backup_cogorc_putline_YUY2); + orc_program_add_destination (p, 4, "d1"); + orc_program_add_source (p, 8, "s1"); + orc_program_add_constant (p, 0, 0x00000000, "c1"); + orc_program_add_constant (p, 0, 0x00000000, "c2"); + orc_program_add_constant (p, 0, 0x00000000, "c3"); + orc_program_add_constant (p, 0, 0x00000000, "c4"); + orc_program_add_constant (p, 0, 0x00000000, "c5"); + orc_program_add_constant (p, 0, 0x00000000, "c6"); + orc_program_add_constant (p, 0, 0x00000000, "c7"); + orc_program_add_constant (p, 0, 0x00000000, "c8"); + orc_program_add_temporary (p, 2, "t1"); + orc_program_add_temporary (p, 2, "t2"); + orc_program_add_temporary (p, 2, "t3"); + orc_program_add_temporary (p, 4, "t4"); + orc_program_add_temporary (p, 4, "t5"); + + orc_program_append_2 (p, "splitlw", 1, ORC_VAR_T5, ORC_VAR_T4, ORC_VAR_S1, + ORC_VAR_D1); + orc_program_append_2 (p, "splitlw", 0, ORC_VAR_T2, ORC_VAR_T3, ORC_VAR_T5, + ORC_VAR_D1); + orc_program_append_2 (p, "avgub", 1, ORC_VAR_T2, ORC_VAR_T2, ORC_VAR_T3, + ORC_VAR_D1); + orc_program_append_2 (p, "select1wb", 1, ORC_VAR_T1, ORC_VAR_T4, + ORC_VAR_D1, ORC_VAR_D1); + orc_program_append_2 (p, "mergebw", 1, ORC_VAR_D1, ORC_VAR_T1, ORC_VAR_T2, + ORC_VAR_D1); + + result = orc_program_compile (p); + } + p_inited = TRUE; + orc_once_mutex_unlock (); + } + ex->program = p; + + ex->n = n; + ex->arrays[ORC_VAR_D1] = d1; + ex->arrays[ORC_VAR_S1] = (void *) s1; + + func = p->code_exec; + func (ex); +} +#endif + + +/* cogorc_putline_YVYU */ +#ifdef DISABLE_ORC +void +cogorc_putline_YVYU (guint8 * d1, const guint8 * s1, int n) +{ + int i; + orc_union32 *ORC_RESTRICT ptr0; + const orc_union64 *ORC_RESTRICT ptr4; + orc_union64 var37; + orc_union32 var38; + orc_union32 var39; + orc_union32 var40; + orc_union16 var41; + orc_union16 var42; + orc_union16 var43; + orc_union16 var44; + orc_union16 var45; + + ptr0 = (orc_union32 *) d1; + ptr4 = (orc_union64 *) s1; + + + for (i = 0; i < n; i++) { + /* 0: loadq */ + var37 = ptr4[i]; + /* 1: splitlw */ + var39.x2[0] = (var37.x2[0] >> 16) & 0xffff; + var40.x2[0] = var37.x2[0] & 0xffff; + var39.x2[1] = (var37.x2[1] >> 16) & 0xffff; + var40.x2[1] = var37.x2[1] & 0xffff; + /* 2: splitlw */ + var41.i = (var39.i >> 16) & 0xffff; + var42.i = var39.i & 0xffff; + /* 3: avgub */ + var43.x2[0] = ((orc_uint8) var41.x2[0] + (orc_uint8) var42.x2[0] + 1) >> 1; + var43.x2[1] = ((orc_uint8) var41.x2[1] + (orc_uint8) var42.x2[1] + 1) >> 1; + /* 4: select1wb */ + var44.x2[0] = ((orc_uint16) var40.x2[0] >> 8) & 0xff; + var44.x2[1] = ((orc_uint16) var40.x2[1] >> 8) & 0xff; + /* 5: swapw */ + var45.i = ORC_SWAP_W (var43.i); + /* 6: mergebw */ + var38.x2[0] = ((orc_uint8) var44.x2[0]) | ((orc_uint8) var45.x2[0] << 8); + var38.x2[1] = ((orc_uint8) var44.x2[1]) | ((orc_uint8) var45.x2[1] << 8); + /* 7: storel */ + ptr0[i] = var38; + } + +} + +#else +static void +_backup_cogorc_putline_YVYU (OrcExecutor * ex) +{ + int i; + int n = ex->n; + orc_union32 *ORC_RESTRICT ptr0; + const orc_union64 *ORC_RESTRICT ptr4; + orc_union64 var37; + orc_union32 var38; + orc_union32 var39; + orc_union32 var40; + orc_union16 var41; + orc_union16 var42; + orc_union16 var43; + orc_union16 var44; + orc_union16 var45; + + ptr0 = (orc_union32 *) ex->arrays[0]; + ptr4 = (orc_union64 *) ex->arrays[4]; + + + for (i = 0; i < n; i++) { + /* 0: loadq */ + var37 = ptr4[i]; + /* 1: splitlw */ + var39.x2[0] = (var37.x2[0] >> 16) & 0xffff; + var40.x2[0] = var37.x2[0] & 0xffff; + var39.x2[1] = (var37.x2[1] >> 16) & 0xffff; + var40.x2[1] = var37.x2[1] & 0xffff; + /* 2: splitlw */ + var41.i = (var39.i >> 16) & 0xffff; + var42.i = var39.i & 0xffff; + /* 3: avgub */ + var43.x2[0] = ((orc_uint8) var41.x2[0] + (orc_uint8) var42.x2[0] + 1) >> 1; + var43.x2[1] = ((orc_uint8) var41.x2[1] + (orc_uint8) var42.x2[1] + 1) >> 1; + /* 4: select1wb */ + var44.x2[0] = ((orc_uint16) var40.x2[0] >> 8) & 0xff; + var44.x2[1] = ((orc_uint16) var40.x2[1] >> 8) & 0xff; + /* 5: swapw */ + var45.i = ORC_SWAP_W (var43.i); + /* 6: mergebw */ + var38.x2[0] = ((orc_uint8) var44.x2[0]) | ((orc_uint8) var45.x2[0] << 8); + var38.x2[1] = ((orc_uint8) var44.x2[1]) | ((orc_uint8) var45.x2[1] << 8); + /* 7: storel */ + ptr0[i] = var38; + } + +} + +void +cogorc_putline_YVYU (guint8 * d1, const guint8 * s1, int n) +{ + OrcExecutor _ex, *ex = &_ex; + static int p_inited = 0; + static OrcProgram *p = 0; + void (*func) (OrcExecutor *); + + if (!p_inited) { + orc_once_mutex_lock (); + if (!p_inited) { + OrcCompileResult result; + + p = orc_program_new (); + orc_program_set_name (p, "cogorc_putline_YVYU"); + orc_program_set_backup_function (p, _backup_cogorc_putline_YVYU); + orc_program_add_destination (p, 4, "d1"); + orc_program_add_source (p, 8, "s1"); + orc_program_add_constant (p, 0, 0x00000000, "c1"); + orc_program_add_constant (p, 0, 0x00000000, "c2"); + orc_program_add_constant (p, 0, 0x00000000, "c3"); + orc_program_add_constant (p, 0, 0x00000000, "c4"); + orc_program_add_constant (p, 0, 0x00000000, "c5"); + orc_program_add_constant (p, 0, 0x00000000, "c6"); + orc_program_add_constant (p, 0, 0x00000000, "c7"); + orc_program_add_constant (p, 0, 0x00000000, "c8"); + orc_program_add_temporary (p, 2, "t1"); + orc_program_add_temporary (p, 2, "t2"); + orc_program_add_temporary (p, 2, "t3"); + orc_program_add_temporary (p, 4, "t4"); + orc_program_add_temporary (p, 4, "t5"); + + orc_program_append_2 (p, "splitlw", 1, ORC_VAR_T5, ORC_VAR_T4, ORC_VAR_S1, + ORC_VAR_D1); + orc_program_append_2 (p, "splitlw", 0, ORC_VAR_T2, ORC_VAR_T3, ORC_VAR_T5, + ORC_VAR_D1); + orc_program_append_2 (p, "avgub", 1, ORC_VAR_T2, ORC_VAR_T2, ORC_VAR_T3, + ORC_VAR_D1); + orc_program_append_2 (p, "select1wb", 1, ORC_VAR_T1, ORC_VAR_T4, + ORC_VAR_D1, ORC_VAR_D1); + orc_program_append_2 (p, "swapw", 0, ORC_VAR_T2, ORC_VAR_T2, ORC_VAR_D1, + ORC_VAR_D1); + orc_program_append_2 (p, "mergebw", 1, ORC_VAR_D1, ORC_VAR_T1, ORC_VAR_T2, + ORC_VAR_D1); + + result = orc_program_compile (p); + } + p_inited = TRUE; + orc_once_mutex_unlock (); + } + ex->program = p; + + ex->n = n; + ex->arrays[ORC_VAR_D1] = d1; + ex->arrays[ORC_VAR_S1] = (void *) s1; + + func = p->code_exec; + func (ex); +} +#endif + + +/* cogorc_putline_UYVY */ +#ifdef DISABLE_ORC +void +cogorc_putline_UYVY (guint8 * d1, const guint8 * s1, int n) +{ + int i; + orc_union32 *ORC_RESTRICT ptr0; + const orc_union64 *ORC_RESTRICT ptr4; + orc_union64 var37; + orc_union32 var38; + orc_union32 var39; + orc_union32 var40; + orc_union16 var41; + orc_union16 var42; + orc_union16 var43; + orc_union16 var44; + + ptr0 = (orc_union32 *) d1; + ptr4 = (orc_union64 *) s1; + + + for (i = 0; i < n; i++) { + /* 0: loadq */ + var37 = ptr4[i]; + /* 1: splitlw */ + var39.x2[0] = (var37.x2[0] >> 16) & 0xffff; + var40.x2[0] = var37.x2[0] & 0xffff; + var39.x2[1] = (var37.x2[1] >> 16) & 0xffff; + var40.x2[1] = var37.x2[1] & 0xffff; + /* 2: splitlw */ + var41.i = (var39.i >> 16) & 0xffff; + var42.i = var39.i & 0xffff; + /* 3: avgub */ + var43.x2[0] = ((orc_uint8) var41.x2[0] + (orc_uint8) var42.x2[0] + 1) >> 1; + var43.x2[1] = ((orc_uint8) var41.x2[1] + (orc_uint8) var42.x2[1] + 1) >> 1; + /* 4: select1wb */ + var44.x2[0] = ((orc_uint16) var40.x2[0] >> 8) & 0xff; + var44.x2[1] = ((orc_uint16) var40.x2[1] >> 8) & 0xff; + /* 5: mergebw */ + var38.x2[0] = ((orc_uint8) var43.x2[0]) | ((orc_uint8) var44.x2[0] << 8); + var38.x2[1] = ((orc_uint8) var43.x2[1]) | ((orc_uint8) var44.x2[1] << 8); + /* 6: storel */ + ptr0[i] = var38; + } + +} + +#else +static void +_backup_cogorc_putline_UYVY (OrcExecutor * ex) +{ + int i; + int n = ex->n; + orc_union32 *ORC_RESTRICT ptr0; + const orc_union64 *ORC_RESTRICT ptr4; + orc_union64 var37; + orc_union32 var38; + orc_union32 var39; + orc_union32 var40; + orc_union16 var41; + orc_union16 var42; + orc_union16 var43; + orc_union16 var44; + + ptr0 = (orc_union32 *) ex->arrays[0]; + ptr4 = (orc_union64 *) ex->arrays[4]; + + + for (i = 0; i < n; i++) { + /* 0: loadq */ + var37 = ptr4[i]; + /* 1: splitlw */ + var39.x2[0] = (var37.x2[0] >> 16) & 0xffff; + var40.x2[0] = var37.x2[0] & 0xffff; + var39.x2[1] = (var37.x2[1] >> 16) & 0xffff; + var40.x2[1] = var37.x2[1] & 0xffff; + /* 2: splitlw */ + var41.i = (var39.i >> 16) & 0xffff; + var42.i = var39.i & 0xffff; + /* 3: avgub */ + var43.x2[0] = ((orc_uint8) var41.x2[0] + (orc_uint8) var42.x2[0] + 1) >> 1; + var43.x2[1] = ((orc_uint8) var41.x2[1] + (orc_uint8) var42.x2[1] + 1) >> 1; + /* 4: select1wb */ + var44.x2[0] = ((orc_uint16) var40.x2[0] >> 8) & 0xff; + var44.x2[1] = ((orc_uint16) var40.x2[1] >> 8) & 0xff; + /* 5: mergebw */ + var38.x2[0] = ((orc_uint8) var43.x2[0]) | ((orc_uint8) var44.x2[0] << 8); + var38.x2[1] = ((orc_uint8) var43.x2[1]) | ((orc_uint8) var44.x2[1] << 8); + /* 6: storel */ + ptr0[i] = var38; + } + +} + +void +cogorc_putline_UYVY (guint8 * d1, const guint8 * s1, int n) +{ + OrcExecutor _ex, *ex = &_ex; + static int p_inited = 0; + static OrcProgram *p = 0; + void (*func) (OrcExecutor *); + + if (!p_inited) { + orc_once_mutex_lock (); + if (!p_inited) { + OrcCompileResult result; + + p = orc_program_new (); + orc_program_set_name (p, "cogorc_putline_UYVY"); + orc_program_set_backup_function (p, _backup_cogorc_putline_UYVY); + orc_program_add_destination (p, 4, "d1"); + orc_program_add_source (p, 8, "s1"); + orc_program_add_constant (p, 0, 0x00000000, "c1"); + orc_program_add_constant (p, 0, 0x00000000, "c2"); + orc_program_add_constant (p, 0, 0x00000000, "c3"); + orc_program_add_constant (p, 0, 0x00000000, "c4"); + orc_program_add_constant (p, 0, 0x00000000, "c5"); + orc_program_add_constant (p, 0, 0x00000000, "c6"); + orc_program_add_constant (p, 0, 0x00000000, "c7"); + orc_program_add_constant (p, 0, 0x00000000, "c8"); + orc_program_add_temporary (p, 2, "t1"); + orc_program_add_temporary (p, 2, "t2"); + orc_program_add_temporary (p, 2, "t3"); + orc_program_add_temporary (p, 4, "t4"); + orc_program_add_temporary (p, 4, "t5"); + + orc_program_append_2 (p, "splitlw", 1, ORC_VAR_T5, ORC_VAR_T4, ORC_VAR_S1, + ORC_VAR_D1); + orc_program_append_2 (p, "splitlw", 0, ORC_VAR_T2, ORC_VAR_T3, ORC_VAR_T5, + ORC_VAR_D1); + orc_program_append_2 (p, "avgub", 1, ORC_VAR_T2, ORC_VAR_T2, ORC_VAR_T3, + ORC_VAR_D1); + orc_program_append_2 (p, "select1wb", 1, ORC_VAR_T1, ORC_VAR_T4, + ORC_VAR_D1, ORC_VAR_D1); + orc_program_append_2 (p, "mergebw", 1, ORC_VAR_D1, ORC_VAR_T2, ORC_VAR_T1, + ORC_VAR_D1); + + result = orc_program_compile (p); + } + p_inited = TRUE; + orc_once_mutex_unlock (); + } + ex->program = p; + + ex->n = n; + ex->arrays[ORC_VAR_D1] = d1; + ex->arrays[ORC_VAR_S1] = (void *) s1; + + func = p->code_exec; + func (ex); +} +#endif + + +/* cogorc_putline_Y42B */ +#ifdef DISABLE_ORC +void +cogorc_putline_Y42B (guint8 * d1, guint8 * d2, guint8 * d3, const guint8 * s1, + int n) +{ + int i; + orc_union16 *ORC_RESTRICT ptr0; + orc_int8 *ORC_RESTRICT ptr1; + orc_int8 *ORC_RESTRICT ptr2; + const orc_union64 *ORC_RESTRICT ptr4; + orc_union64 var36; + orc_int8 var37; + orc_int8 var38; + orc_union16 var39; + orc_union32 var40; + orc_union32 var41; + orc_union16 var42; + orc_union16 var43; + orc_union16 var44; + + ptr0 = (orc_union16 *) d1; + ptr1 = (orc_int8 *) d2; + ptr2 = (orc_int8 *) d3; + ptr4 = (orc_union64 *) s1; + + + for (i = 0; i < n; i++) { + /* 0: loadq */ + var36 = ptr4[i]; + /* 1: splitlw */ + var40.x2[0] = (var36.x2[0] >> 16) & 0xffff; + var41.x2[0] = var36.x2[0] & 0xffff; + var40.x2[1] = (var36.x2[1] >> 16) & 0xffff; + var41.x2[1] = var36.x2[1] & 0xffff; + /* 2: splitlw */ + var42.i = (var40.i >> 16) & 0xffff; + var43.i = var40.i & 0xffff; + /* 3: avgub */ + var44.x2[0] = ((orc_uint8) var42.x2[0] + (orc_uint8) var43.x2[0] + 1) >> 1; + var44.x2[1] = ((orc_uint8) var42.x2[1] + (orc_uint8) var43.x2[1] + 1) >> 1; + /* 4: splitwb */ + var37 = (var44.i >> 8) & 0xff; + var38 = var44.i & 0xff; + /* 5: storeb */ + ptr2[i] = var37; + /* 6: storeb */ + ptr1[i] = var38; + /* 7: select1wb */ + var39.x2[0] = ((orc_uint16) var41.x2[0] >> 8) & 0xff; + var39.x2[1] = ((orc_uint16) var41.x2[1] >> 8) & 0xff; + /* 8: storew */ + ptr0[i] = var39; + } + +} + +#else +static void +_backup_cogorc_putline_Y42B (OrcExecutor * ex) +{ + int i; + int n = ex->n; + orc_union16 *ORC_RESTRICT ptr0; + orc_int8 *ORC_RESTRICT ptr1; + orc_int8 *ORC_RESTRICT ptr2; + const orc_union64 *ORC_RESTRICT ptr4; + orc_union64 var36; + orc_int8 var37; + orc_int8 var38; + orc_union16 var39; + orc_union32 var40; + orc_union32 var41; + orc_union16 var42; + orc_union16 var43; + orc_union16 var44; + + ptr0 = (orc_union16 *) ex->arrays[0]; + ptr1 = (orc_int8 *) ex->arrays[1]; + ptr2 = (orc_int8 *) ex->arrays[2]; + ptr4 = (orc_union64 *) ex->arrays[4]; + + + for (i = 0; i < n; i++) { + /* 0: loadq */ + var36 = ptr4[i]; + /* 1: splitlw */ + var40.x2[0] = (var36.x2[0] >> 16) & 0xffff; + var41.x2[0] = var36.x2[0] & 0xffff; + var40.x2[1] = (var36.x2[1] >> 16) & 0xffff; + var41.x2[1] = var36.x2[1] & 0xffff; + /* 2: splitlw */ + var42.i = (var40.i >> 16) & 0xffff; + var43.i = var40.i & 0xffff; + /* 3: avgub */ + var44.x2[0] = ((orc_uint8) var42.x2[0] + (orc_uint8) var43.x2[0] + 1) >> 1; + var44.x2[1] = ((orc_uint8) var42.x2[1] + (orc_uint8) var43.x2[1] + 1) >> 1; + /* 4: splitwb */ + var37 = (var44.i >> 8) & 0xff; + var38 = var44.i & 0xff; + /* 5: storeb */ + ptr2[i] = var37; + /* 6: storeb */ + ptr1[i] = var38; + /* 7: select1wb */ + var39.x2[0] = ((orc_uint16) var41.x2[0] >> 8) & 0xff; + var39.x2[1] = ((orc_uint16) var41.x2[1] >> 8) & 0xff; + /* 8: storew */ + ptr0[i] = var39; + } + +} + +void +cogorc_putline_Y42B (guint8 * d1, guint8 * d2, guint8 * d3, const guint8 * s1, + int n) +{ + OrcExecutor _ex, *ex = &_ex; + static int p_inited = 0; + static OrcProgram *p = 0; + void (*func) (OrcExecutor *); + + if (!p_inited) { + orc_once_mutex_lock (); + if (!p_inited) { + OrcCompileResult result; + + p = orc_program_new (); + orc_program_set_name (p, "cogorc_putline_Y42B"); + orc_program_set_backup_function (p, _backup_cogorc_putline_Y42B); + orc_program_add_destination (p, 2, "d1"); + orc_program_add_destination (p, 1, "d2"); + orc_program_add_destination (p, 1, "d3"); + orc_program_add_source (p, 8, "s1"); + orc_program_add_constant (p, 0, 0x00000000, "c1"); + orc_program_add_constant (p, 0, 0x00000000, "c2"); + orc_program_add_constant (p, 0, 0x00000000, "c3"); + orc_program_add_constant (p, 0, 0x00000000, "c4"); + orc_program_add_constant (p, 0, 0x00000000, "c5"); + orc_program_add_constant (p, 0, 0x00000000, "c6"); + orc_program_add_constant (p, 0, 0x00000000, "c7"); + orc_program_add_constant (p, 0, 0x00000000, "c8"); + orc_program_add_temporary (p, 4, "t1"); + orc_program_add_temporary (p, 4, "t2"); + orc_program_add_temporary (p, 2, "t3"); + orc_program_add_temporary (p, 2, "t4"); + + orc_program_append_2 (p, "splitlw", 1, ORC_VAR_T2, ORC_VAR_T1, ORC_VAR_S1, + ORC_VAR_D1); + orc_program_append_2 (p, "splitlw", 0, ORC_VAR_T3, ORC_VAR_T4, ORC_VAR_T2, + ORC_VAR_D1); + orc_program_append_2 (p, "avgub", 1, ORC_VAR_T3, ORC_VAR_T3, ORC_VAR_T4, + ORC_VAR_D1); + orc_program_append_2 (p, "splitwb", 0, ORC_VAR_D3, ORC_VAR_D2, ORC_VAR_T3, + ORC_VAR_D1); + orc_program_append_2 (p, "select1wb", 1, ORC_VAR_D1, ORC_VAR_T1, + ORC_VAR_D1, ORC_VAR_D1); + + result = orc_program_compile (p); + } + p_inited = TRUE; + orc_once_mutex_unlock (); + } + ex->program = p; + + ex->n = n; + ex->arrays[ORC_VAR_D1] = d1; + ex->arrays[ORC_VAR_D2] = d2; + ex->arrays[ORC_VAR_D3] = d3; + ex->arrays[ORC_VAR_S1] = (void *) s1; + + func = p->code_exec; + func (ex); +} +#endif + + +/* cogorc_putline_Y444 */ +#ifdef DISABLE_ORC +void +cogorc_putline_Y444 (guint8 * d1, guint8 * d2, guint8 * d3, const guint8 * s1, + int n) +{ + int i; + orc_int8 *ORC_RESTRICT ptr0; + orc_int8 *ORC_RESTRICT ptr1; + orc_int8 *ORC_RESTRICT ptr2; + const orc_union32 *ORC_RESTRICT ptr4; + orc_union32 var34; + orc_int8 var35; + orc_int8 var36; + orc_int8 var37; + orc_union16 var38; + orc_union16 var39; + + ptr0 = (orc_int8 *) d1; + ptr1 = (orc_int8 *) d2; + ptr2 = (orc_int8 *) d3; + ptr4 = (orc_union32 *) s1; + + + for (i = 0; i < n; i++) { + /* 0: loadl */ + var34 = ptr4[i]; + /* 1: splitlw */ + var38.i = (var34.i >> 16) & 0xffff; + var39.i = var34.i & 0xffff; + /* 2: splitwb */ + var35 = (var38.i >> 8) & 0xff; + var36 = var38.i & 0xff; + /* 3: storeb */ + ptr2[i] = var35; + /* 4: storeb */ + ptr1[i] = var36; + /* 5: select1wb */ + var37 = ((orc_uint16) var39.i >> 8) & 0xff; + /* 6: storeb */ + ptr0[i] = var37; + } + +} + +#else +static void +_backup_cogorc_putline_Y444 (OrcExecutor * ex) +{ + int i; + int n = ex->n; + orc_int8 *ORC_RESTRICT ptr0; + orc_int8 *ORC_RESTRICT ptr1; + orc_int8 *ORC_RESTRICT ptr2; + const orc_union32 *ORC_RESTRICT ptr4; + orc_union32 var34; + orc_int8 var35; + orc_int8 var36; + orc_int8 var37; + orc_union16 var38; + orc_union16 var39; + + ptr0 = (orc_int8 *) ex->arrays[0]; + ptr1 = (orc_int8 *) ex->arrays[1]; + ptr2 = (orc_int8 *) ex->arrays[2]; + ptr4 = (orc_union32 *) ex->arrays[4]; + + + for (i = 0; i < n; i++) { + /* 0: loadl */ + var34 = ptr4[i]; + /* 1: splitlw */ + var38.i = (var34.i >> 16) & 0xffff; + var39.i = var34.i & 0xffff; + /* 2: splitwb */ + var35 = (var38.i >> 8) & 0xff; + var36 = var38.i & 0xff; + /* 3: storeb */ + ptr2[i] = var35; + /* 4: storeb */ + ptr1[i] = var36; + /* 5: select1wb */ + var37 = ((orc_uint16) var39.i >> 8) & 0xff; + /* 6: storeb */ + ptr0[i] = var37; + } + +} + +void +cogorc_putline_Y444 (guint8 * d1, guint8 * d2, guint8 * d3, const guint8 * s1, + int n) +{ + OrcExecutor _ex, *ex = &_ex; + static int p_inited = 0; + static OrcProgram *p = 0; + void (*func) (OrcExecutor *); + + if (!p_inited) { + orc_once_mutex_lock (); + if (!p_inited) { + OrcCompileResult result; + + p = orc_program_new (); + orc_program_set_name (p, "cogorc_putline_Y444"); + orc_program_set_backup_function (p, _backup_cogorc_putline_Y444); + orc_program_add_destination (p, 1, "d1"); + orc_program_add_destination (p, 1, "d2"); + orc_program_add_destination (p, 1, "d3"); + orc_program_add_source (p, 4, "s1"); + orc_program_add_constant (p, 0, 0x00000000, "c1"); + orc_program_add_constant (p, 0, 0x00000000, "c2"); + orc_program_add_constant (p, 0, 0x00000000, "c3"); + orc_program_add_constant (p, 0, 0x00000000, "c4"); + orc_program_add_constant (p, 0, 0x00000000, "c5"); + orc_program_add_constant (p, 0, 0x00000000, "c6"); + orc_program_add_constant (p, 0, 0x00000000, "c7"); + orc_program_add_constant (p, 0, 0x00000000, "c8"); + orc_program_add_temporary (p, 2, "t1"); + orc_program_add_temporary (p, 2, "t2"); + + orc_program_append_2 (p, "splitlw", 0, ORC_VAR_T2, ORC_VAR_T1, ORC_VAR_S1, + ORC_VAR_D1); + orc_program_append_2 (p, "splitwb", 0, ORC_VAR_D3, ORC_VAR_D2, ORC_VAR_T2, + ORC_VAR_D1); + orc_program_append_2 (p, "select1wb", 0, ORC_VAR_D1, ORC_VAR_T1, + ORC_VAR_D1, ORC_VAR_D1); + + result = orc_program_compile (p); + } + p_inited = TRUE; + orc_once_mutex_unlock (); + } + ex->program = p; + + ex->n = n; + ex->arrays[ORC_VAR_D1] = d1; + ex->arrays[ORC_VAR_D2] = d2; + ex->arrays[ORC_VAR_D3] = d3; + ex->arrays[ORC_VAR_S1] = (void *) s1; + + func = p->code_exec; + func (ex); +} +#endif + + +/* cogorc_putline_Y800 */ +#ifdef DISABLE_ORC +void +cogorc_putline_Y800 (guint8 * d1, const guint8 * s1, int n) +{ + int i; + orc_int8 *ORC_RESTRICT ptr0; + const orc_union32 *ORC_RESTRICT ptr4; + orc_union32 var33; + orc_int8 var34; + orc_union16 var35; + + ptr0 = (orc_int8 *) d1; + ptr4 = (orc_union32 *) s1; + + + for (i = 0; i < n; i++) { + /* 0: loadl */ + var33 = ptr4[i]; + /* 1: select0lw */ + var35.i = (orc_uint32) var33.i & 0xffff; + /* 2: select1wb */ + var34 = ((orc_uint16) var35.i >> 8) & 0xff; + /* 3: storeb */ + ptr0[i] = var34; + } + +} + +#else +static void +_backup_cogorc_putline_Y800 (OrcExecutor * ex) +{ + int i; + int n = ex->n; + orc_int8 *ORC_RESTRICT ptr0; + const orc_union32 *ORC_RESTRICT ptr4; + orc_union32 var33; + orc_int8 var34; + orc_union16 var35; + + ptr0 = (orc_int8 *) ex->arrays[0]; + ptr4 = (orc_union32 *) ex->arrays[4]; + + + for (i = 0; i < n; i++) { + /* 0: loadl */ + var33 = ptr4[i]; + /* 1: select0lw */ + var35.i = (orc_uint32) var33.i & 0xffff; + /* 2: select1wb */ + var34 = ((orc_uint16) var35.i >> 8) & 0xff; + /* 3: storeb */ + ptr0[i] = var34; + } + +} + +void +cogorc_putline_Y800 (guint8 * d1, const guint8 * s1, int n) +{ + OrcExecutor _ex, *ex = &_ex; + static int p_inited = 0; + static OrcProgram *p = 0; + void (*func) (OrcExecutor *); + + if (!p_inited) { + orc_once_mutex_lock (); + if (!p_inited) { + OrcCompileResult result; + + p = orc_program_new (); + orc_program_set_name (p, "cogorc_putline_Y800"); + orc_program_set_backup_function (p, _backup_cogorc_putline_Y800); + orc_program_add_destination (p, 1, "d1"); + orc_program_add_source (p, 4, "s1"); + orc_program_add_constant (p, 0, 0x00000000, "c1"); + orc_program_add_constant (p, 0, 0x00000000, "c2"); + orc_program_add_constant (p, 0, 0x00000000, "c3"); + orc_program_add_constant (p, 0, 0x00000000, "c4"); + orc_program_add_constant (p, 0, 0x00000000, "c5"); + orc_program_add_constant (p, 0, 0x00000000, "c6"); + orc_program_add_constant (p, 0, 0x00000000, "c7"); + orc_program_add_constant (p, 0, 0x00000000, "c8"); + orc_program_add_temporary (p, 2, "t1"); + + orc_program_append_2 (p, "select0lw", 0, ORC_VAR_T1, ORC_VAR_S1, + ORC_VAR_D1, ORC_VAR_D1); + orc_program_append_2 (p, "select1wb", 0, ORC_VAR_D1, ORC_VAR_T1, + ORC_VAR_D1, ORC_VAR_D1); + + result = orc_program_compile (p); + } + p_inited = TRUE; + orc_once_mutex_unlock (); + } + ex->program = p; + + ex->n = n; + ex->arrays[ORC_VAR_D1] = d1; + ex->arrays[ORC_VAR_S1] = (void *) s1; + + func = p->code_exec; + func (ex); +} +#endif + + +/* cogorc_putline_BGRA */ +#ifdef DISABLE_ORC +void +cogorc_putline_BGRA (guint8 * d1, const guint8 * s1, int n) +{ + int i; + orc_union32 *ORC_RESTRICT ptr0; + const orc_union32 *ORC_RESTRICT ptr4; + orc_union32 var32; + orc_union32 var33; + + ptr0 = (orc_union32 *) d1; + ptr4 = (orc_union32 *) s1; + + + for (i = 0; i < n; i++) { + /* 0: loadl */ + var32 = ptr4[i]; + /* 1: swapl */ + var33.i = ORC_SWAP_L (var32.i); + /* 2: storel */ + ptr0[i] = var33; + } + +} + +#else +static void +_backup_cogorc_putline_BGRA (OrcExecutor * ex) +{ + int i; + int n = ex->n; + orc_union32 *ORC_RESTRICT ptr0; + const orc_union32 *ORC_RESTRICT ptr4; + orc_union32 var32; + orc_union32 var33; + + ptr0 = (orc_union32 *) ex->arrays[0]; + ptr4 = (orc_union32 *) ex->arrays[4]; + + + for (i = 0; i < n; i++) { + /* 0: loadl */ + var32 = ptr4[i]; + /* 1: swapl */ + var33.i = ORC_SWAP_L (var32.i); + /* 2: storel */ + ptr0[i] = var33; + } + +} + +void +cogorc_putline_BGRA (guint8 * d1, const guint8 * s1, int n) +{ + OrcExecutor _ex, *ex = &_ex; + static int p_inited = 0; + static OrcProgram *p = 0; + void (*func) (OrcExecutor *); + + if (!p_inited) { + orc_once_mutex_lock (); + if (!p_inited) { + OrcCompileResult result; + + p = orc_program_new (); + orc_program_set_name (p, "cogorc_putline_BGRA"); + orc_program_set_backup_function (p, _backup_cogorc_putline_BGRA); + orc_program_add_destination (p, 4, "d1"); + orc_program_add_source (p, 4, "s1"); + orc_program_add_constant (p, 0, 0x00000000, "c1"); + orc_program_add_constant (p, 0, 0x00000000, "c2"); + orc_program_add_constant (p, 0, 0x00000000, "c3"); + orc_program_add_constant (p, 0, 0x00000000, "c4"); + orc_program_add_constant (p, 0, 0x00000000, "c5"); + orc_program_add_constant (p, 0, 0x00000000, "c6"); + orc_program_add_constant (p, 0, 0x00000000, "c7"); + orc_program_add_constant (p, 0, 0x00000000, "c8"); + + orc_program_append_2 (p, "swapl", 0, ORC_VAR_D1, ORC_VAR_S1, ORC_VAR_D1, + ORC_VAR_D1); + + result = orc_program_compile (p); + } + p_inited = TRUE; + orc_once_mutex_unlock (); + } + ex->program = p; + + ex->n = n; + ex->arrays[ORC_VAR_D1] = d1; + ex->arrays[ORC_VAR_S1] = (void *) s1; + + func = p->code_exec; + func (ex); +} +#endif + + +/* cogorc_putline_ABGR */ +#ifdef DISABLE_ORC +void +cogorc_putline_ABGR (guint8 * d1, const guint8 * s1, int n) +{ + int i; + orc_union32 *ORC_RESTRICT ptr0; + const orc_union32 *ORC_RESTRICT ptr4; + orc_union32 var40; + orc_union32 var41; + orc_union16 var42; + orc_union16 var43; + orc_int8 var44; + orc_int8 var45; + orc_int8 var46; + orc_int8 var47; + orc_union16 var48; + orc_union16 var49; + + ptr0 = (orc_union32 *) d1; + ptr4 = (orc_union32 *) s1; + + + for (i = 0; i < n; i++) { + /* 0: loadl */ + var40 = ptr4[i]; + /* 1: splitlw */ + var42.i = (var40.i >> 16) & 0xffff; + var43.i = var40.i & 0xffff; + /* 2: splitwb */ + var44 = (var42.i >> 8) & 0xff; + var45 = var42.i & 0xff; + /* 3: splitwb */ + var46 = (var43.i >> 8) & 0xff; + var47 = var43.i & 0xff; + /* 4: mergebw */ + var48.i = ((orc_uint8) var47) | ((orc_uint8) var44 << 8); + /* 5: mergebw */ + var49.i = ((orc_uint8) var45) | ((orc_uint8) var46 << 8); + /* 6: mergewl */ + var41.i = ((orc_uint16) var48.i) | ((orc_uint16) var49.i << 16); + /* 7: storel */ + ptr0[i] = var41; + } + +} + +#else +static void +_backup_cogorc_putline_ABGR (OrcExecutor * ex) +{ + int i; + int n = ex->n; + orc_union32 *ORC_RESTRICT ptr0; + const orc_union32 *ORC_RESTRICT ptr4; + orc_union32 var40; + orc_union32 var41; + orc_union16 var42; + orc_union16 var43; + orc_int8 var44; + orc_int8 var45; + orc_int8 var46; + orc_int8 var47; + orc_union16 var48; + orc_union16 var49; + + ptr0 = (orc_union32 *) ex->arrays[0]; + ptr4 = (orc_union32 *) ex->arrays[4]; + + + for (i = 0; i < n; i++) { + /* 0: loadl */ + var40 = ptr4[i]; + /* 1: splitlw */ + var42.i = (var40.i >> 16) & 0xffff; + var43.i = var40.i & 0xffff; + /* 2: splitwb */ + var44 = (var42.i >> 8) & 0xff; + var45 = var42.i & 0xff; + /* 3: splitwb */ + var46 = (var43.i >> 8) & 0xff; + var47 = var43.i & 0xff; + /* 4: mergebw */ + var48.i = ((orc_uint8) var47) | ((orc_uint8) var44 << 8); + /* 5: mergebw */ + var49.i = ((orc_uint8) var45) | ((orc_uint8) var46 << 8); + /* 6: mergewl */ + var41.i = ((orc_uint16) var48.i) | ((orc_uint16) var49.i << 16); + /* 7: storel */ + ptr0[i] = var41; + } + +} + +void +cogorc_putline_ABGR (guint8 * d1, const guint8 * s1, int n) +{ + OrcExecutor _ex, *ex = &_ex; + static int p_inited = 0; + static OrcProgram *p = 0; + void (*func) (OrcExecutor *); + + if (!p_inited) { + orc_once_mutex_lock (); + if (!p_inited) { + OrcCompileResult result; + + p = orc_program_new (); + orc_program_set_name (p, "cogorc_putline_ABGR"); + orc_program_set_backup_function (p, _backup_cogorc_putline_ABGR); + orc_program_add_destination (p, 4, "d1"); + orc_program_add_source (p, 4, "s1"); + orc_program_add_constant (p, 0, 0x00000000, "c1"); + orc_program_add_constant (p, 0, 0x00000000, "c2"); + orc_program_add_constant (p, 0, 0x00000000, "c3"); + orc_program_add_constant (p, 0, 0x00000000, "c4"); + orc_program_add_constant (p, 0, 0x00000000, "c5"); + orc_program_add_constant (p, 0, 0x00000000, "c6"); + orc_program_add_constant (p, 0, 0x00000000, "c7"); + orc_program_add_constant (p, 0, 0x00000000, "c8"); + orc_program_add_temporary (p, 1, "t1"); + orc_program_add_temporary (p, 1, "t2"); + orc_program_add_temporary (p, 1, "t3"); + orc_program_add_temporary (p, 1, "t4"); + orc_program_add_temporary (p, 2, "t5"); + orc_program_add_temporary (p, 2, "t6"); + orc_program_add_temporary (p, 2, "t7"); + orc_program_add_temporary (p, 2, "t8"); + + orc_program_append_2 (p, "splitlw", 0, ORC_VAR_T8, ORC_VAR_T7, ORC_VAR_S1, + ORC_VAR_D1); + orc_program_append_2 (p, "splitwb", 0, ORC_VAR_T4, ORC_VAR_T3, ORC_VAR_T8, + ORC_VAR_D1); + orc_program_append_2 (p, "splitwb", 0, ORC_VAR_T2, ORC_VAR_T1, ORC_VAR_T7, + ORC_VAR_D1); + orc_program_append_2 (p, "mergebw", 0, ORC_VAR_T6, ORC_VAR_T1, ORC_VAR_T4, + ORC_VAR_D1); + orc_program_append_2 (p, "mergebw", 0, ORC_VAR_T5, ORC_VAR_T3, ORC_VAR_T2, + ORC_VAR_D1); + orc_program_append_2 (p, "mergewl", 0, ORC_VAR_D1, ORC_VAR_T6, ORC_VAR_T5, + ORC_VAR_D1); + + result = orc_program_compile (p); + } + p_inited = TRUE; + orc_once_mutex_unlock (); + } + ex->program = p; + + ex->n = n; + ex->arrays[ORC_VAR_D1] = d1; + ex->arrays[ORC_VAR_S1] = (void *) s1; + + func = p->code_exec; + func (ex); +} +#endif + + +/* cogorc_putline_RGBA */ +#ifdef DISABLE_ORC +void +cogorc_putline_RGBA (guint8 * d1, const guint8 * s1, int n) +{ + int i; + orc_union32 *ORC_RESTRICT ptr0; + const orc_union32 *ORC_RESTRICT ptr4; + orc_union32 var40; + orc_union32 var41; + orc_union16 var42; + orc_union16 var43; + orc_int8 var44; + orc_int8 var45; + orc_int8 var46; + orc_int8 var47; + orc_union16 var48; + orc_union16 var49; + + ptr0 = (orc_union32 *) d1; + ptr4 = (orc_union32 *) s1; + + + for (i = 0; i < n; i++) { + /* 0: loadl */ + var40 = ptr4[i]; + /* 1: splitlw */ + var42.i = (var40.i >> 16) & 0xffff; + var43.i = var40.i & 0xffff; + /* 2: splitwb */ + var44 = (var42.i >> 8) & 0xff; + var45 = var42.i & 0xff; + /* 3: splitwb */ + var46 = (var43.i >> 8) & 0xff; + var47 = var43.i & 0xff; + /* 4: mergebw */ + var48.i = ((orc_uint8) var44) | ((orc_uint8) var47 << 8); + /* 5: mergebw */ + var49.i = ((orc_uint8) var46) | ((orc_uint8) var45 << 8); + /* 6: mergewl */ + var41.i = ((orc_uint16) var49.i) | ((orc_uint16) var48.i << 16); + /* 7: storel */ + ptr0[i] = var41; + } + +} + +#else +static void +_backup_cogorc_putline_RGBA (OrcExecutor * ex) +{ + int i; + int n = ex->n; + orc_union32 *ORC_RESTRICT ptr0; + const orc_union32 *ORC_RESTRICT ptr4; + orc_union32 var40; + orc_union32 var41; + orc_union16 var42; + orc_union16 var43; + orc_int8 var44; + orc_int8 var45; + orc_int8 var46; + orc_int8 var47; + orc_union16 var48; + orc_union16 var49; + + ptr0 = (orc_union32 *) ex->arrays[0]; + ptr4 = (orc_union32 *) ex->arrays[4]; + + + for (i = 0; i < n; i++) { + /* 0: loadl */ + var40 = ptr4[i]; + /* 1: splitlw */ + var42.i = (var40.i >> 16) & 0xffff; + var43.i = var40.i & 0xffff; + /* 2: splitwb */ + var44 = (var42.i >> 8) & 0xff; + var45 = var42.i & 0xff; + /* 3: splitwb */ + var46 = (var43.i >> 8) & 0xff; + var47 = var43.i & 0xff; + /* 4: mergebw */ + var48.i = ((orc_uint8) var44) | ((orc_uint8) var47 << 8); + /* 5: mergebw */ + var49.i = ((orc_uint8) var46) | ((orc_uint8) var45 << 8); + /* 6: mergewl */ + var41.i = ((orc_uint16) var49.i) | ((orc_uint16) var48.i << 16); + /* 7: storel */ + ptr0[i] = var41; + } + +} + +void +cogorc_putline_RGBA (guint8 * d1, const guint8 * s1, int n) +{ + OrcExecutor _ex, *ex = &_ex; + static int p_inited = 0; + static OrcProgram *p = 0; + void (*func) (OrcExecutor *); + + if (!p_inited) { + orc_once_mutex_lock (); + if (!p_inited) { + OrcCompileResult result; + + p = orc_program_new (); + orc_program_set_name (p, "cogorc_putline_RGBA"); + orc_program_set_backup_function (p, _backup_cogorc_putline_RGBA); + orc_program_add_destination (p, 4, "d1"); + orc_program_add_source (p, 4, "s1"); + orc_program_add_constant (p, 0, 0x00000000, "c1"); + orc_program_add_constant (p, 0, 0x00000000, "c2"); + orc_program_add_constant (p, 0, 0x00000000, "c3"); + orc_program_add_constant (p, 0, 0x00000000, "c4"); + orc_program_add_constant (p, 0, 0x00000000, "c5"); + orc_program_add_constant (p, 0, 0x00000000, "c6"); + orc_program_add_constant (p, 0, 0x00000000, "c7"); + orc_program_add_constant (p, 0, 0x00000000, "c8"); + orc_program_add_temporary (p, 1, "t1"); + orc_program_add_temporary (p, 1, "t2"); + orc_program_add_temporary (p, 1, "t3"); + orc_program_add_temporary (p, 1, "t4"); + orc_program_add_temporary (p, 2, "t5"); + orc_program_add_temporary (p, 2, "t6"); + orc_program_add_temporary (p, 2, "t7"); + orc_program_add_temporary (p, 2, "t8"); + + orc_program_append_2 (p, "splitlw", 0, ORC_VAR_T8, ORC_VAR_T7, ORC_VAR_S1, + ORC_VAR_D1); + orc_program_append_2 (p, "splitwb", 0, ORC_VAR_T4, ORC_VAR_T3, ORC_VAR_T8, + ORC_VAR_D1); + orc_program_append_2 (p, "splitwb", 0, ORC_VAR_T2, ORC_VAR_T1, ORC_VAR_T7, + ORC_VAR_D1); + orc_program_append_2 (p, "mergebw", 0, ORC_VAR_T6, ORC_VAR_T4, ORC_VAR_T1, + ORC_VAR_D1); + orc_program_append_2 (p, "mergebw", 0, ORC_VAR_T5, ORC_VAR_T2, ORC_VAR_T3, + ORC_VAR_D1); + orc_program_append_2 (p, "mergewl", 0, ORC_VAR_D1, ORC_VAR_T5, ORC_VAR_T6, + ORC_VAR_D1); + + result = orc_program_compile (p); + } + p_inited = TRUE; + orc_once_mutex_unlock (); + } + ex->program = p; + + ex->n = n; + ex->arrays[ORC_VAR_D1] = d1; + ex->arrays[ORC_VAR_S1] = (void *) s1; + + func = p->code_exec; + func (ex); +} +#endif + + +/* cogorc_putline_NV12 */ +#ifdef DISABLE_ORC +void +cogorc_putline_NV12 (guint8 * d1, guint8 * d2, const guint8 * s1, int n) +{ + int i; + orc_union16 *ORC_RESTRICT ptr0; + orc_union16 *ORC_RESTRICT ptr1; + const orc_union64 *ORC_RESTRICT ptr4; + orc_union64 var36; + orc_union16 var37; + orc_union16 var38; + orc_union32 var39; + orc_union32 var40; + orc_union16 var41; + orc_union16 var42; + + ptr0 = (orc_union16 *) d1; + ptr1 = (orc_union16 *) d2; + ptr4 = (orc_union64 *) s1; + + + for (i = 0; i < n; i++) { + /* 0: loadq */ + var36 = ptr4[i]; + /* 1: splitlw */ + var39.x2[0] = (var36.x2[0] >> 16) & 0xffff; + var40.x2[0] = var36.x2[0] & 0xffff; + var39.x2[1] = (var36.x2[1] >> 16) & 0xffff; + var40.x2[1] = var36.x2[1] & 0xffff; + /* 2: select1wb */ + var37.x2[0] = ((orc_uint16) var40.x2[0] >> 8) & 0xff; + var37.x2[1] = ((orc_uint16) var40.x2[1] >> 8) & 0xff; + /* 3: storew */ + ptr0[i] = var37; + /* 4: splitlw */ + var41.i = (var39.i >> 16) & 0xffff; + var42.i = var39.i & 0xffff; + /* 5: avgub */ + var38.x2[0] = ((orc_uint8) var41.x2[0] + (orc_uint8) var42.x2[0] + 1) >> 1; + var38.x2[1] = ((orc_uint8) var41.x2[1] + (orc_uint8) var42.x2[1] + 1) >> 1; + /* 6: storew */ + ptr1[i] = var38; + } + +} + +#else +static void +_backup_cogorc_putline_NV12 (OrcExecutor * ex) +{ + int i; + int n = ex->n; + orc_union16 *ORC_RESTRICT ptr0; + orc_union16 *ORC_RESTRICT ptr1; + const orc_union64 *ORC_RESTRICT ptr4; + orc_union64 var36; + orc_union16 var37; + orc_union16 var38; + orc_union32 var39; + orc_union32 var40; + orc_union16 var41; + orc_union16 var42; + + ptr0 = (orc_union16 *) ex->arrays[0]; + ptr1 = (orc_union16 *) ex->arrays[1]; + ptr4 = (orc_union64 *) ex->arrays[4]; + + + for (i = 0; i < n; i++) { + /* 0: loadq */ + var36 = ptr4[i]; + /* 1: splitlw */ + var39.x2[0] = (var36.x2[0] >> 16) & 0xffff; + var40.x2[0] = var36.x2[0] & 0xffff; + var39.x2[1] = (var36.x2[1] >> 16) & 0xffff; + var40.x2[1] = var36.x2[1] & 0xffff; + /* 2: select1wb */ + var37.x2[0] = ((orc_uint16) var40.x2[0] >> 8) & 0xff; + var37.x2[1] = ((orc_uint16) var40.x2[1] >> 8) & 0xff; + /* 3: storew */ + ptr0[i] = var37; + /* 4: splitlw */ + var41.i = (var39.i >> 16) & 0xffff; + var42.i = var39.i & 0xffff; + /* 5: avgub */ + var38.x2[0] = ((orc_uint8) var41.x2[0] + (orc_uint8) var42.x2[0] + 1) >> 1; + var38.x2[1] = ((orc_uint8) var41.x2[1] + (orc_uint8) var42.x2[1] + 1) >> 1; + /* 6: storew */ + ptr1[i] = var38; + } + +} + +void +cogorc_putline_NV12 (guint8 * d1, guint8 * d2, const guint8 * s1, int n) +{ + OrcExecutor _ex, *ex = &_ex; + static int p_inited = 0; + static OrcProgram *p = 0; + void (*func) (OrcExecutor *); + + if (!p_inited) { + orc_once_mutex_lock (); + if (!p_inited) { + OrcCompileResult result; + + p = orc_program_new (); + orc_program_set_name (p, "cogorc_putline_NV12"); + orc_program_set_backup_function (p, _backup_cogorc_putline_NV12); + orc_program_add_destination (p, 2, "d1"); + orc_program_add_destination (p, 2, "d2"); + orc_program_add_source (p, 8, "s1"); + orc_program_add_constant (p, 0, 0x00000000, "c1"); + orc_program_add_constant (p, 0, 0x00000000, "c2"); + orc_program_add_constant (p, 0, 0x00000000, "c3"); + orc_program_add_constant (p, 0, 0x00000000, "c4"); + orc_program_add_constant (p, 0, 0x00000000, "c5"); + orc_program_add_constant (p, 0, 0x00000000, "c6"); + orc_program_add_constant (p, 0, 0x00000000, "c7"); + orc_program_add_constant (p, 0, 0x00000000, "c8"); + orc_program_add_temporary (p, 4, "t1"); + orc_program_add_temporary (p, 4, "t2"); + orc_program_add_temporary (p, 2, "t3"); + orc_program_add_temporary (p, 2, "t4"); + + orc_program_append_2 (p, "splitlw", 1, ORC_VAR_T2, ORC_VAR_T1, ORC_VAR_S1, + ORC_VAR_D1); + orc_program_append_2 (p, "select1wb", 1, ORC_VAR_D1, ORC_VAR_T1, + ORC_VAR_D1, ORC_VAR_D1); + orc_program_append_2 (p, "splitlw", 0, ORC_VAR_T3, ORC_VAR_T4, ORC_VAR_T2, + ORC_VAR_D1); + orc_program_append_2 (p, "avgub", 1, ORC_VAR_D2, ORC_VAR_T3, ORC_VAR_T4, + ORC_VAR_D1); + + result = orc_program_compile (p); + } + p_inited = TRUE; + orc_once_mutex_unlock (); + } + ex->program = p; + + ex->n = n; + ex->arrays[ORC_VAR_D1] = d1; + ex->arrays[ORC_VAR_D2] = d2; + ex->arrays[ORC_VAR_S1] = (void *) s1; + + func = p->code_exec; + func (ex); +} +#endif + + +/* cogorc_putline_NV21 */ +#ifdef DISABLE_ORC +void +cogorc_putline_NV21 (guint8 * d1, guint8 * d2, const guint8 * s1, int n) +{ + int i; + orc_union16 *ORC_RESTRICT ptr0; + orc_union16 *ORC_RESTRICT ptr1; + const orc_union64 *ORC_RESTRICT ptr4; + orc_union64 var37; + orc_union16 var38; + orc_union16 var39; + orc_union32 var40; + orc_union32 var41; + orc_union16 var42; + orc_union16 var43; + orc_union16 var44; + + ptr0 = (orc_union16 *) d1; + ptr1 = (orc_union16 *) d2; + ptr4 = (orc_union64 *) s1; + + + for (i = 0; i < n; i++) { + /* 0: loadq */ + var37 = ptr4[i]; + /* 1: splitlw */ + var40.x2[0] = (var37.x2[0] >> 16) & 0xffff; + var41.x2[0] = var37.x2[0] & 0xffff; + var40.x2[1] = (var37.x2[1] >> 16) & 0xffff; + var41.x2[1] = var37.x2[1] & 0xffff; + /* 2: select1wb */ + var38.x2[0] = ((orc_uint16) var41.x2[0] >> 8) & 0xff; + var38.x2[1] = ((orc_uint16) var41.x2[1] >> 8) & 0xff; + /* 3: storew */ + ptr0[i] = var38; + /* 4: splitlw */ + var42.i = (var40.i >> 16) & 0xffff; + var43.i = var40.i & 0xffff; + /* 5: avgub */ + var44.x2[0] = ((orc_uint8) var42.x2[0] + (orc_uint8) var43.x2[0] + 1) >> 1; + var44.x2[1] = ((orc_uint8) var42.x2[1] + (orc_uint8) var43.x2[1] + 1) >> 1; + /* 6: swapw */ + var39.i = ORC_SWAP_W (var44.i); + /* 7: storew */ + ptr1[i] = var39; + } + +} + +#else +static void +_backup_cogorc_putline_NV21 (OrcExecutor * ex) +{ + int i; + int n = ex->n; + orc_union16 *ORC_RESTRICT ptr0; + orc_union16 *ORC_RESTRICT ptr1; + const orc_union64 *ORC_RESTRICT ptr4; + orc_union64 var37; + orc_union16 var38; + orc_union16 var39; + orc_union32 var40; + orc_union32 var41; + orc_union16 var42; + orc_union16 var43; + orc_union16 var44; + + ptr0 = (orc_union16 *) ex->arrays[0]; + ptr1 = (orc_union16 *) ex->arrays[1]; + ptr4 = (orc_union64 *) ex->arrays[4]; + + + for (i = 0; i < n; i++) { + /* 0: loadq */ + var37 = ptr4[i]; + /* 1: splitlw */ + var40.x2[0] = (var37.x2[0] >> 16) & 0xffff; + var41.x2[0] = var37.x2[0] & 0xffff; + var40.x2[1] = (var37.x2[1] >> 16) & 0xffff; + var41.x2[1] = var37.x2[1] & 0xffff; + /* 2: select1wb */ + var38.x2[0] = ((orc_uint16) var41.x2[0] >> 8) & 0xff; + var38.x2[1] = ((orc_uint16) var41.x2[1] >> 8) & 0xff; + /* 3: storew */ + ptr0[i] = var38; + /* 4: splitlw */ + var42.i = (var40.i >> 16) & 0xffff; + var43.i = var40.i & 0xffff; + /* 5: avgub */ + var44.x2[0] = ((orc_uint8) var42.x2[0] + (orc_uint8) var43.x2[0] + 1) >> 1; + var44.x2[1] = ((orc_uint8) var42.x2[1] + (orc_uint8) var43.x2[1] + 1) >> 1; + /* 6: swapw */ + var39.i = ORC_SWAP_W (var44.i); + /* 7: storew */ + ptr1[i] = var39; + } + +} + +void +cogorc_putline_NV21 (guint8 * d1, guint8 * d2, const guint8 * s1, int n) +{ + OrcExecutor _ex, *ex = &_ex; + static int p_inited = 0; + static OrcProgram *p = 0; + void (*func) (OrcExecutor *); + + if (!p_inited) { + orc_once_mutex_lock (); + if (!p_inited) { + OrcCompileResult result; + + p = orc_program_new (); + orc_program_set_name (p, "cogorc_putline_NV21"); + orc_program_set_backup_function (p, _backup_cogorc_putline_NV21); + orc_program_add_destination (p, 2, "d1"); + orc_program_add_destination (p, 2, "d2"); + orc_program_add_source (p, 8, "s1"); + orc_program_add_constant (p, 0, 0x00000000, "c1"); + orc_program_add_constant (p, 0, 0x00000000, "c2"); + orc_program_add_constant (p, 0, 0x00000000, "c3"); + orc_program_add_constant (p, 0, 0x00000000, "c4"); + orc_program_add_constant (p, 0, 0x00000000, "c5"); + orc_program_add_constant (p, 0, 0x00000000, "c6"); + orc_program_add_constant (p, 0, 0x00000000, "c7"); + orc_program_add_constant (p, 0, 0x00000000, "c8"); + orc_program_add_temporary (p, 4, "t1"); + orc_program_add_temporary (p, 4, "t2"); + orc_program_add_temporary (p, 2, "t3"); + orc_program_add_temporary (p, 2, "t4"); + orc_program_add_temporary (p, 2, "t5"); + + orc_program_append_2 (p, "splitlw", 1, ORC_VAR_T2, ORC_VAR_T1, ORC_VAR_S1, + ORC_VAR_D1); + orc_program_append_2 (p, "select1wb", 1, ORC_VAR_D1, ORC_VAR_T1, + ORC_VAR_D1, ORC_VAR_D1); + orc_program_append_2 (p, "splitlw", 0, ORC_VAR_T3, ORC_VAR_T4, ORC_VAR_T2, + ORC_VAR_D1); + orc_program_append_2 (p, "avgub", 1, ORC_VAR_T5, ORC_VAR_T3, ORC_VAR_T4, + ORC_VAR_D1); + orc_program_append_2 (p, "swapw", 0, ORC_VAR_D2, ORC_VAR_T5, ORC_VAR_D1, + ORC_VAR_D1); + + result = orc_program_compile (p); + } + p_inited = TRUE; + orc_once_mutex_unlock (); + } + ex->program = p; + + ex->n = n; + ex->arrays[ORC_VAR_D1] = d1; + ex->arrays[ORC_VAR_D2] = d2; + ex->arrays[ORC_VAR_S1] = (void *) s1; + + func = p->code_exec; + func (ex); +} +#endif diff --git a/gst/colorspace/gstcolorspaceorc-dist.h b/gst/colorspace/gstcolorspaceorc-dist.h new file mode 100644 index 0000000..3e29282 --- /dev/null +++ b/gst/colorspace/gstcolorspaceorc-dist.h @@ -0,0 +1,149 @@ + +/* autogenerated from gstcolorspaceorc.orc */ + +#ifndef _GSTCOLORSPACEORC_H_ +#define _GSTCOLORSPACEORC_H_ + +#include + +#ifdef __cplusplus +extern "C" { +#endif + + + +#ifndef _ORC_INTEGER_TYPEDEFS_ +#define _ORC_INTEGER_TYPEDEFS_ +#if defined(__STDC_VERSION__) && __STDC_VERSION__ >= 199901L +#include +typedef int8_t orc_int8; +typedef int16_t orc_int16; +typedef int32_t orc_int32; +typedef int64_t orc_int64; +typedef uint8_t orc_uint8; +typedef uint16_t orc_uint16; +typedef uint32_t orc_uint32; +typedef uint64_t orc_uint64; +#elif defined(_MSC_VER) +typedef signed __int8 orc_int8; +typedef signed __int16 orc_int16; +typedef signed __int32 orc_int32; +typedef signed __int64 orc_int64; +typedef unsigned __int8 orc_uint8; +typedef unsigned __int16 orc_uint16; +typedef unsigned __int32 orc_uint32; +typedef unsigned __int64 orc_uint64; +#else +#include +typedef signed char orc_int8; +typedef short orc_int16; +typedef int orc_int32; +typedef unsigned char orc_uint8; +typedef unsigned short orc_uint16; +typedef unsigned int orc_uint32; +#if INT_MAX == LONG_MAX +typedef long long orc_int64; +typedef unsigned long long orc_uint64; +#else +typedef long orc_int64; +typedef unsigned long orc_uint64; +#endif +#endif +typedef union { orc_int16 i; orc_int8 x2[2]; } orc_union16; +typedef union { orc_int32 i; float f; orc_int16 x2[2]; orc_int8 x4[4]; } orc_union32; +typedef union { orc_int64 i; double f; orc_int32 x2[2]; orc_int16 x4[4]; } orc_union64; +#endif +void cogorc_memcpy_2d (guint8 * d1, int d1_stride, const guint8 * s1, int s1_stride, int n, int m); +void cogorc_downsample_horiz_cosite_1tap (guint8 * d1, const guint8 * s1, int n); +void cogorc_downsample_horiz_cosite_3tap (guint8 * d1, const guint8 * s1, const guint8 * s2, int n); +void cogorc_downsample_420_jpeg (guint8 * d1, const guint8 * s1, const guint8 * s2, int n); +void cogorc_downsample_vert_halfsite_2tap (guint8 * d1, const guint8 * s1, const guint8 * s2, int n); +void cogorc_downsample_vert_cosite_3tap (guint8 * d1, const guint8 * s1, const guint8 * s2, const guint8 * s3, int n); +void cogorc_downsample_vert_halfsite_4tap (guint8 * d1, const guint8 * s1, const guint8 * s2, const guint8 * s3, const guint8 * s4, int n); +void cogorc_upsample_horiz_cosite_1tap (guint8 * d1, const guint8 * s1, int n); +void cogorc_upsample_horiz_cosite (guint8 * d1, const guint8 * s1, const guint8 * s2, int n); +void cogorc_upsample_vert_avgub (guint8 * d1, const guint8 * s1, const guint8 * s2, int n); +void orc_unpack_yuyv_y (guint8 * d1, const guint8 * s1, int n); +void orc_unpack_yuyv_u (guint8 * d1, const guint8 * s1, int n); +void orc_unpack_yuyv_v (guint8 * d1, const guint8 * s1, int n); +void orc_pack_yuyv (guint8 * d1, const guint8 * s1, const guint8 * s2, const guint8 * s3, int n); +void orc_unpack_uyvy_y (guint8 * d1, const guint8 * s1, int n); +void orc_unpack_uyvy_u (guint8 * d1, const guint8 * s1, int n); +void orc_unpack_uyvy_v (guint8 * d1, const guint8 * s1, int n); +void orc_pack_uyvy (guint8 * d1, const guint8 * s1, const guint8 * s2, const guint8 * s3, int n); +void orc_matrix2_u8 (guint8 * d1, const guint8 * s1, const guint8 * s2, int p1, int p2, int p3, int n); +void orc_matrix2_11_u8 (guint8 * d1, const guint8 * s1, const guint8 * s2, int p1, int p2, int n); +void orc_matrix2_12_u8 (guint8 * d1, const guint8 * s1, const guint8 * s2, int p1, int p2, int n); +void orc_matrix3_u8 (guint8 * d1, const guint8 * s1, const guint8 * s2, const guint8 * s3, int p1, int p2, int p3, int p4, int n); +void orc_matrix3_100_u8 (guint8 * d1, const guint8 * s1, const guint8 * s2, const guint8 * s3, int p1, int p2, int p3, int n); +void orc_matrix3_100_offset_u8 (guint8 * d1, const guint8 * s1, const guint8 * s2, const guint8 * s3, int p1, int p2, int p3, int p4, int p5, int n); +void orc_matrix3_000_u8 (guint8 * d1, const guint8 * s1, const guint8 * s2, const guint8 * s3, int p1, int p2, int p3, int p4, int p5, int n); +void orc_pack_123x (guint32 * d1, const guint8 * s1, const guint8 * s2, const guint8 * s3, int p1, int n); +void orc_pack_x123 (guint32 * d1, const guint8 * s1, const guint8 * s2, const guint8 * s3, int p1, int n); +void cogorc_combine2_u8 (guint8 * d1, const guint8 * s1, const guint8 * s2, int p1, int p2, int n); +void cogorc_convert_I420_UYVY (guint8 * d1, guint8 * d2, const guint8 * s1, const guint8 * s2, const guint8 * s3, const guint8 * s4, int n); +void cogorc_convert_I420_YUY2 (guint8 * d1, guint8 * d2, const guint8 * s1, const guint8 * s2, const guint8 * s3, const guint8 * s4, int n); +void cogorc_convert_I420_AYUV (guint8 * d1, guint8 * d2, const guint8 * s1, const guint8 * s2, const guint8 * s3, const guint8 * s4, int n); +void cogorc_convert_YUY2_I420 (guint8 * d1, guint8 * d2, guint8 * d3, guint8 * d4, const guint8 * s1, const guint8 * s2, int n); +void cogorc_convert_UYVY_YUY2 (guint8 * d1, int d1_stride, const guint8 * s1, int s1_stride, int n, int m); +void cogorc_planar_chroma_420_422 (guint8 * d1, int d1_stride, guint8 * d2, int d2_stride, const guint8 * s1, int s1_stride, int n, int m); +void cogorc_planar_chroma_420_444 (guint8 * d1, int d1_stride, guint8 * d2, int d2_stride, const guint8 * s1, int s1_stride, int n, int m); +void cogorc_planar_chroma_422_444 (guint8 * d1, int d1_stride, const guint8 * s1, int s1_stride, int n, int m); +void cogorc_planar_chroma_444_422 (guint8 * d1, int d1_stride, const guint8 * s1, int s1_stride, int n, int m); +void cogorc_planar_chroma_444_420 (guint8 * d1, int d1_stride, const guint8 * s1, int s1_stride, const guint8 * s2, int s2_stride, int n, int m); +void cogorc_planar_chroma_422_420 (guint8 * d1, int d1_stride, const guint8 * s1, int s1_stride, const guint8 * s2, int s2_stride, int n, int m); +void cogorc_convert_YUY2_AYUV (guint8 * d1, int d1_stride, const guint8 * s1, int s1_stride, int n, int m); +void cogorc_convert_UYVY_AYUV (guint8 * d1, int d1_stride, const guint8 * s1, int s1_stride, int n, int m); +void cogorc_convert_YUY2_Y42B (guint8 * d1, int d1_stride, guint8 * d2, int d2_stride, guint8 * d3, int d3_stride, const guint8 * s1, int s1_stride, int n, int m); +void cogorc_convert_UYVY_Y42B (guint8 * d1, int d1_stride, guint8 * d2, int d2_stride, guint8 * d3, int d3_stride, const guint8 * s1, int s1_stride, int n, int m); +void cogorc_convert_YUY2_Y444 (guint8 * d1, int d1_stride, guint8 * d2, int d2_stride, guint8 * d3, int d3_stride, const guint8 * s1, int s1_stride, int n, int m); +void cogorc_convert_UYVY_Y444 (guint8 * d1, int d1_stride, guint8 * d2, int d2_stride, guint8 * d3, int d3_stride, const guint8 * s1, int s1_stride, int n, int m); +void cogorc_convert_UYVY_I420 (guint8 * d1, guint8 * d2, guint8 * d3, guint8 * d4, const guint8 * s1, const guint8 * s2, int n); +void cogorc_convert_AYUV_I420 (guint8 * d1, int d1_stride, guint8 * d2, int d2_stride, guint8 * d3, int d3_stride, guint8 * d4, int d4_stride, const guint8 * s1, int s1_stride, const guint8 * s2, int s2_stride, int n, int m); +void cogorc_convert_AYUV_YUY2 (guint8 * d1, int d1_stride, const guint8 * s1, int s1_stride, int n, int m); +void cogorc_convert_AYUV_UYVY (guint8 * d1, int d1_stride, const guint8 * s1, int s1_stride, int n, int m); +void cogorc_convert_AYUV_Y42B (guint8 * d1, int d1_stride, guint8 * d2, int d2_stride, guint8 * d3, int d3_stride, const guint8 * s1, int s1_stride, int n, int m); +void cogorc_convert_AYUV_Y444 (guint8 * d1, int d1_stride, guint8 * d2, int d2_stride, guint8 * d3, int d3_stride, const guint8 * s1, int s1_stride, int n, int m); +void cogorc_convert_Y42B_YUY2 (guint8 * d1, int d1_stride, const guint8 * s1, int s1_stride, const guint8 * s2, int s2_stride, const guint8 * s3, int s3_stride, int n, int m); +void cogorc_convert_Y42B_UYVY (guint8 * d1, int d1_stride, const guint8 * s1, int s1_stride, const guint8 * s2, int s2_stride, const guint8 * s3, int s3_stride, int n, int m); +void cogorc_convert_Y42B_AYUV (guint8 * d1, int d1_stride, const guint8 * s1, int s1_stride, const guint8 * s2, int s2_stride, const guint8 * s3, int s3_stride, int n, int m); +void cogorc_convert_Y444_YUY2 (guint8 * d1, int d1_stride, const guint8 * s1, int s1_stride, const guint8 * s2, int s2_stride, const guint8 * s3, int s3_stride, int n, int m); +void cogorc_convert_Y444_UYVY (guint8 * d1, int d1_stride, const guint8 * s1, int s1_stride, const guint8 * s2, int s2_stride, const guint8 * s3, int s3_stride, int n, int m); +void cogorc_convert_Y444_AYUV (guint8 * d1, int d1_stride, const guint8 * s1, int s1_stride, const guint8 * s2, int s2_stride, const guint8 * s3, int s3_stride, int n, int m); +void cogorc_convert_AYUV_ARGB (guint8 * d1, int d1_stride, const guint8 * s1, int s1_stride, int n, int m); +void cogorc_convert_AYUV_BGRA (guint8 * d1, int d1_stride, const guint8 * s1, int s1_stride, int n, int m); +void cogorc_convert_AYUV_ABGR (guint8 * d1, int d1_stride, const guint8 * s1, int s1_stride, int n, int m); +void cogorc_convert_AYUV_RGBA (guint8 * d1, int d1_stride, const guint8 * s1, int s1_stride, int n, int m); +void cogorc_convert_I420_BGRA (guint8 * d1, const guint8 * s1, const guint8 * s2, const guint8 * s3, int n); +void cogorc_convert_I420_BGRA_avg (guint8 * d1, const guint8 * s1, const guint8 * s2, const guint8 * s3, const guint8 * s4, const guint8 * s5, int n); +void cogorc_getline_I420 (guint8 * d1, const guint8 * s1, const guint8 * s2, const guint8 * s3, int n); +void cogorc_getline_YUY2 (guint8 * d1, const guint8 * s1, int n); +void cogorc_getline_UYVY (guint8 * d1, const guint8 * s1, int n); +void cogorc_getline_YVYU (guint8 * d1, const guint8 * s1, int n); +void cogorc_getline_Y42B (guint8 * d1, const guint8 * s1, const guint8 * s2, const guint8 * s3, int n); +void cogorc_getline_Y444 (guint8 * d1, const guint8 * s1, const guint8 * s2, const guint8 * s3, int n); +void cogorc_getline_Y800 (guint8 * d1, const guint8 * s1, int n); +void cogorc_getline_BGRA (guint8 * d1, const guint8 * s1, int n); +void cogorc_getline_ABGR (guint8 * d1, const guint8 * s1, int n); +void cogorc_getline_RGBA (guint8 * d1, const guint8 * s1, int n); +void cogorc_getline_NV12 (guint8 * d1, const guint8 * s1, const guint8 * s2, int n); +void cogorc_getline_NV21 (guint8 * d1, const guint8 * s1, const guint8 * s2, int n); +void cogorc_putline_I420 (guint8 * d1, guint8 * d2, guint8 * d3, const guint8 * s1, int n); +void cogorc_putline_YUY2 (guint8 * d1, const guint8 * s1, int n); +void cogorc_putline_YVYU (guint8 * d1, const guint8 * s1, int n); +void cogorc_putline_UYVY (guint8 * d1, const guint8 * s1, int n); +void cogorc_putline_Y42B (guint8 * d1, guint8 * d2, guint8 * d3, const guint8 * s1, int n); +void cogorc_putline_Y444 (guint8 * d1, guint8 * d2, guint8 * d3, const guint8 * s1, int n); +void cogorc_putline_Y800 (guint8 * d1, const guint8 * s1, int n); +void cogorc_putline_BGRA (guint8 * d1, const guint8 * s1, int n); +void cogorc_putline_ABGR (guint8 * d1, const guint8 * s1, int n); +void cogorc_putline_RGBA (guint8 * d1, const guint8 * s1, int n); +void cogorc_putline_NV12 (guint8 * d1, guint8 * d2, const guint8 * s1, int n); +void cogorc_putline_NV21 (guint8 * d1, guint8 * d2, const guint8 * s1, int n); + +#ifdef __cplusplus +} +#endif + +#endif +