From 30466f26b48d7d0d7eb29a290053a2ccd4dc71ea Mon Sep 17 00:00:00 2001 From: James Zern Date: Fri, 6 Nov 2015 13:03:13 -0800 Subject: [PATCH] Revert "Add AVX vectorized vp9_diamond_search_sad" This reverts commit f1342a7b070ef61b9fbdf03e899ac2107cfcb6bd. This breaks 32-bit builds: runtime error: load of misaligned address 0xf72fdd48 for type 'const __m128i' (vector of 2 'long long' values), which requires 16 byte alignment + _mm_set1_epi64x is incompatible with some versions of visual studio Change-Id: I6f6fc3c11403344cef78d1c432cdc9147e5c1673 --- vp9/common/vp9_rtcd_defs.pl | 2 +- vp9/encoder/vp9_encoder.c | 30 --- vp9/encoder/vp9_mcomp.c | 26 ++- vp9/encoder/vp9_mcomp.h | 6 +- vp9/encoder/x86/vp9_diamond_search_sad_avx.c | 308 --------------------------- vp9/vp9cx.mk | 1 - 6 files changed, 20 insertions(+), 353 deletions(-) delete mode 100644 vp9/encoder/x86/vp9_diamond_search_sad_avx.c diff --git a/vp9/common/vp9_rtcd_defs.pl b/vp9/common/vp9_rtcd_defs.pl index f41ee09..5bf71ef 100644 --- a/vp9/common/vp9_rtcd_defs.pl +++ b/vp9/common/vp9_rtcd_defs.pl @@ -312,7 +312,7 @@ $vp9_full_search_sad_sse3=vp9_full_search_sadx3; $vp9_full_search_sad_sse4_1=vp9_full_search_sadx8; add_proto qw/int vp9_diamond_search_sad/, "const struct macroblock *x, const struct search_site_config *cfg, struct mv *ref_mv, struct mv *best_mv, int search_param, int sad_per_bit, int *num00, const struct vp9_variance_vtable *fn_ptr, const struct mv *center_mv"; -specialize qw/vp9_diamond_search_sad avx/; +specialize qw/vp9_diamond_search_sad/; add_proto qw/int vp9_full_range_search/, "const struct macroblock *x, const struct search_site_config *cfg, struct mv *ref_mv, struct mv *best_mv, int search_param, int sad_per_bit, int *num00, const struct vp9_variance_vtable *fn_ptr, const struct mv *center_mv"; specialize qw/vp9_full_range_search/; diff --git a/vp9/encoder/vp9_encoder.c b/vp9/encoder/vp9_encoder.c index eebd7c5..d86a7a7 100644 --- a/vp9/encoder/vp9_encoder.c +++ b/vp9/encoder/vp9_encoder.c @@ -1570,30 +1570,7 @@ void vp9_change_config(struct VP9_COMP *cpi, const VP9EncoderConfig *oxcf) { #endif #define log2f(x) (log (x) / (float) M_LOG2_E) -/*********************************************************************** - * Read before modifying 'cal_nmvjointsadcost' or 'cal_nmvsadcosts' * - *********************************************************************** - * The following 2 functions ('cal_nmvjointsadcost' and * - * 'cal_nmvsadcosts') are used to calculate cost lookup tables * - * used by 'vp9_diamond_search_sad'. The C implementation of the * - * function is generic, but the AVX intrinsics optimised version * - * relies on the following properties of the computed tables: * - * For cal_nmvjointsadcost: * - * - mvjointsadcost[1] == mvjointsadcost[2] == mvjointsadcost[3] * - * For cal_nmvsadcosts: * - * - For all i: mvsadcost[0][i] == mvsadcost[1][i] * - * (Equal costs for both components) * - * - For all i: mvsadcost[0][i] == mvsadcost[0][-i] * - * (Cost function is even) * - * If these do not hold, then the AVX optimised version of the * - * 'vp9_diamond_search_sad' function cannot be used as it is, in which * - * case you can revert to using the C function instead. * - ***********************************************************************/ - static void cal_nmvjointsadcost(int *mvjointsadcost) { - /********************************************************************* - * Warning: Read the comments above before modifying this function * - *********************************************************************/ mvjointsadcost[0] = 600; mvjointsadcost[1] = 300; mvjointsadcost[2] = 300; @@ -1601,9 +1578,6 @@ static void cal_nmvjointsadcost(int *mvjointsadcost) { } static void cal_nmvsadcosts(int *mvsadcost[2]) { - /********************************************************************* - * Warning: Read the comments above before modifying this function * - *********************************************************************/ int i = 1; mvsadcost[0][0] = 0; @@ -1765,10 +1739,6 @@ VP9_COMP *vp9_create_compressor(VP9EncoderConfig *oxcf, cpi->first_time_stamp_ever = INT64_MAX; - /********************************************************************* - * Warning: Read the comments around 'cal_nmvjointsadcost' and * - * 'cal_nmvsadcosts' before modifying how these tables are computed. * - *********************************************************************/ cal_nmvjointsadcost(cpi->td.mb.nmvjointsadcost); cpi->td.mb.nmvcost[0] = &cpi->nmvcosts[0][MV_MAX]; cpi->td.mb.nmvcost[1] = &cpi->nmvcosts[1][MV_MAX]; diff --git a/vp9/encoder/vp9_mcomp.c b/vp9/encoder/vp9_mcomp.c index 4071755..9744e43 100644 --- a/vp9/encoder/vp9_mcomp.c +++ b/vp9/encoder/vp9_mcomp.c @@ -101,8 +101,11 @@ static int mvsad_err_cost(const MACROBLOCK *x, const MV *mv, const MV *ref, } void vp9_init_dsmotion_compensation(search_site_config *cfg, int stride) { - int len; - int ss_count = 0; + int len, ss_count = 1; + + cfg->ss_mv[0].col = 0; + cfg->ss_mv[0].row = 0; + cfg->ss_os[0] = 0; for (len = MAX_FIRST_STEP; len > 0; len /= 2) { // Generate offsets for 4 search sites per step. @@ -114,13 +117,16 @@ void vp9_init_dsmotion_compensation(search_site_config *cfg, int stride) { } } + cfg->ss_count = ss_count; cfg->searches_per_step = 4; - cfg->total_steps = ss_count / cfg->searches_per_step; } void vp9_init3smotion_compensation(search_site_config *cfg, int stride) { - int len; - int ss_count = 0; + int len, ss_count = 1; + + cfg->ss_mv[0].col = 0; + cfg->ss_mv[0].row = 0; + cfg->ss_os[0] = 0; for (len = MAX_FIRST_STEP; len > 0; len /= 2) { // Generate offsets for 8 search sites per step. @@ -135,8 +141,8 @@ void vp9_init3smotion_compensation(search_site_config *cfg, int stride) { } } + cfg->ss_count = ss_count; cfg->searches_per_step = 8; - cfg->total_steps = ss_count / cfg->searches_per_step; } /* @@ -1606,8 +1612,8 @@ int vp9_diamond_search_sad_c(const MACROBLOCK *x, const uint8_t *best_address; unsigned int bestsad = INT_MAX; - int best_site = -1; - int last_site = -1; + int best_site = 0; + int last_site = 0; int ref_row; int ref_col; @@ -1620,7 +1626,7 @@ int vp9_diamond_search_sad_c(const MACROBLOCK *x, // const search_site *ss = &cfg->ss[search_param * cfg->searches_per_step]; const MV *ss_mv = &cfg->ss_mv[search_param * cfg->searches_per_step]; const intptr_t *ss_os = &cfg->ss_os[search_param * cfg->searches_per_step]; - const int tot_steps = (cfg->total_steps) - search_param; + const int tot_steps = (cfg->ss_count / cfg->searches_per_step) - search_param; const MV fcenter_mv = {center_mv->row >> 3, center_mv->col >> 3}; clamp_mv(ref_mv, x->mv_col_min, x->mv_col_max, x->mv_row_min, x->mv_row_max); @@ -1638,7 +1644,7 @@ int vp9_diamond_search_sad_c(const MACROBLOCK *x, bestsad = fn_ptr->sdf(what, what_stride, in_what, in_what_stride) + mvsad_err_cost(x, best_mv, &fcenter_mv, sad_per_bit); - i = 0; + i = 1; for (step = 0; step < tot_steps; step++) { int all_in = 1, t; diff --git a/vp9/encoder/vp9_mcomp.h b/vp9/encoder/vp9_mcomp.h index 1c101f2..e48259f 100644 --- a/vp9/encoder/vp9_mcomp.h +++ b/vp9/encoder/vp9_mcomp.h @@ -33,10 +33,10 @@ extern "C" { typedef struct search_site_config { // motion search sites - MV ss_mv[8 * MAX_MVSEARCH_STEPS]; // Motion vector - intptr_t ss_os[8 * MAX_MVSEARCH_STEPS]; // Offset + MV ss_mv[8 * MAX_MVSEARCH_STEPS + 1]; // Motion vector + intptr_t ss_os[8 * MAX_MVSEARCH_STEPS + 1]; // Offset + int ss_count; int searches_per_step; - int total_steps; } search_site_config; void vp9_init_dsmotion_compensation(search_site_config *cfg, int stride); diff --git a/vp9/encoder/x86/vp9_diamond_search_sad_avx.c b/vp9/encoder/x86/vp9_diamond_search_sad_avx.c deleted file mode 100644 index f216437..0000000 --- a/vp9/encoder/x86/vp9_diamond_search_sad_avx.c +++ /dev/null @@ -1,308 +0,0 @@ -/* - * Copyright (c) 2015 The WebM project authors. All Rights Reserved. - * - * Use of this source code is governed by a BSD-style license - * that can be found in the LICENSE file in the root of the source - * tree. An additional intellectual property rights grant can be found - * in the file PATENTS. All contributing project authors may - * be found in the AUTHORS file in the root of the source tree. - */ - -#include -#include - -#include "vpx_dsp/vpx_dsp_common.h" -#include "vp9/encoder/vp9_encoder.h" -#include "vpx_ports/mem.h" - -#ifdef __GNUC__ -# define __likely__(v) __builtin_expect(v, 1) -# define __unlikely__(v) __builtin_expect(v, 0) -#else -# define __likely__(v) (v) -# define __unlikely__(v) (v) -#endif - -static INLINE MV_JOINT_TYPE get_mv_joint(const int_mv mv) { - // This is simplified from the C implementation to utilise that - // x->nmvjointsadcost[1] == x->nmvjointsadcost[2] and - // x->nmvjointsadcost[1] == x->nmvjointsadcost[3] - return mv.as_int == 0 ? 0 : 1; -} - -static INLINE int mv_cost(const int_mv mv, - const int *joint_cost, int *const comp_cost[2]) { - return joint_cost[get_mv_joint(mv)] + - comp_cost[0][mv.as_mv.row] + comp_cost[1][mv.as_mv.col]; -} - -static int mvsad_err_cost(const MACROBLOCK *x, const int_mv mv, const MV *ref, - int error_per_bit) { - const int_mv diff = { .as_mv={ mv.as_mv.row - ref->row, - mv.as_mv.col - ref->col } }; - return ROUND_POWER_OF_TWO(mv_cost(diff, x->nmvjointsadcost, - x->nmvsadcost) * error_per_bit, 8); -} - -/***************************************************************************** - * This function utilises 3 properties of the cost function lookup tables, * - * constructed in using 'cal_nmvjointsadcost' and 'cal_nmvsadcosts' in * - * vp9_encoder.c. * - * For the joint cost: * - * - mvjointsadcost[1] == mvjointsadcost[2] == mvjointsadcost[3] * - * For the component costs: * - * - For all i: mvsadcost[0][i] == mvsadcost[1][i] * - * (Equal costs for both components) * - * - For all i: mvsadcost[0][i] == mvsadcost[0][-i] * - * (Cost function is even) * - * If these do not hold, then this function cannot be used without * - * modification, in which case you can revert to using the C implementation, * - * which does not rely on these properties. * - *****************************************************************************/ -int vp9_diamond_search_sad_avx(const MACROBLOCK *x, - const search_site_config *cfg, - MV *ref_mv, MV *best_mv, int search_param, - int sad_per_bit, int *num00, - const vp9_variance_fn_ptr_t *fn_ptr, - const MV *center_mv) { - const int_mv maxmv = { .as_mv={x->mv_row_max, x->mv_col_max} }; - const __m128i v_max_mv_w = _mm_set1_epi32(maxmv.as_int); - const int_mv minmv = { .as_mv={x->mv_row_min, x->mv_col_min} }; - const __m128i v_min_mv_w = _mm_set1_epi32(minmv.as_int); - - const __m128i v_spb_d = _mm_set1_epi32(sad_per_bit); - - const __m128i v_joint_cost_0_d = _mm_set1_epi32(x->nmvjointsadcost[0]); - const __m128i v_joint_cost_1_d = _mm_set1_epi32(x->nmvjointsadcost[1]); - - // search_param determines the length of the initial step and hence the number - // of iterations. - // 0 = initial step (MAX_FIRST_STEP) pel - // 1 = (MAX_FIRST_STEP/2) pel, - // 2 = (MAX_FIRST_STEP/4) pel... - const MV *ss_mv = &cfg->ss_mv[cfg->searches_per_step * search_param]; - const intptr_t *ss_os = &cfg->ss_os[cfg->searches_per_step * search_param]; - const int tot_steps = cfg->total_steps - search_param; - - const int_mv fcenter_mv = { .as_mv={center_mv->row >> 3, - center_mv->col >> 3} }; - const __m128i vfcmv = _mm_set1_epi32(fcenter_mv.as_int); - - const int ref_row = clamp(ref_mv->row, minmv.as_mv.row, maxmv.as_mv.row); - const int ref_col = clamp(ref_mv->col, minmv.as_mv.col, maxmv.as_mv.col); - - int_mv bmv = { .as_mv={ ref_row, ref_col} }; - int_mv new_bmv = bmv; - __m128i v_bmv_w = _mm_set1_epi32(bmv.as_int); - - const int what_stride = x->plane[0].src.stride; - const uint8_t * const what = x->plane[0].src.buf; - const int in_what_stride = x->e_mbd.plane[0].pre[0].stride; - const uint8_t * const in_what = x->e_mbd.plane[0].pre[0].buf + - ref_row * in_what_stride + ref_col; - - // Work out the start point for the search - const uint8_t *best_address = in_what; - const uint8_t *new_best_address = best_address; -#if ARCH_X86_64 - __m128i v_ba_q = _mm_set1_epi64x((intptr_t)best_address); -#else - __m128i v_ba_d = _mm_set1_epi32((intptr_t)best_address); -#endif - - unsigned int best_sad; - - int i; - int j; - int step; - - // Check the starting position - best_sad = fn_ptr->sdf(what, what_stride, in_what, in_what_stride); - best_sad += mvsad_err_cost(x, bmv, &fcenter_mv.as_mv, sad_per_bit); - - *num00 = 0; - - for (i = 0, step = 0; step < tot_steps; step++) { - for (j = 0; j < cfg->searches_per_step; j += 4, i += 4) { - __m128i v_sad_d; - __m128i v_cost_d; - __m128i v_outside_d; - __m128i v_inside_d; - __m128i v_diff_mv_w; -#if ARCH_X86_64 - __m128i v_blocka[2]; -#else - __m128i v_blocka[1]; -#endif - - // Compute the candidate motion vectors - const __m128i v_ss_mv_w = _mm_load_si128((const __m128i*)&ss_mv[i]); - const __m128i v_these_mv_w = _mm_add_epi16(v_bmv_w, v_ss_mv_w); - // Clamp them to the search bounds - __m128i v_these_mv_clamp_w = v_these_mv_w; - v_these_mv_clamp_w = _mm_min_epi16(v_these_mv_clamp_w, v_max_mv_w); - v_these_mv_clamp_w = _mm_max_epi16(v_these_mv_clamp_w, v_min_mv_w); - // The ones that did not change are inside the search area - v_inside_d = _mm_cmpeq_epi32(v_these_mv_clamp_w, v_these_mv_w); - - // If none of them are inside, then move on - if (__likely__(_mm_test_all_zeros(v_inside_d, v_inside_d))) { - continue; - } - - // The inverse mask indicates which of the MVs are outside - v_outside_d = _mm_xor_si128(v_inside_d, _mm_set1_epi8(0xff)); - // Shift right to keep the sign bit clear, we will use this later - // to set the cost to the maximum value. - v_outside_d = _mm_srli_epi32(v_outside_d, 1); - - // Compute the difference MV - v_diff_mv_w = _mm_sub_epi16(v_these_mv_clamp_w, vfcmv); - // We utilise the fact that the cost function is even, and use the - // absolute difference. This allows us to use unsigned indexes later - // and reduces cache pressure somewhat as only a half of the table - // is ever referenced. - v_diff_mv_w = _mm_abs_epi16(v_diff_mv_w); - - // Compute the SIMD pointer offsets. - { -#if ARCH_X86_64 // sizeof(intptr_t) == 8 - // Load the offsets (could use _mm_maskload_ps here, instead of the - // extra 'and' but it's slower that way) - __m128i v_bo10_q = _mm_load_si128((const __m128i*)&ss_os[i+0]); - __m128i v_bo32_q = _mm_load_si128((const __m128i*)&ss_os[i+2]); - // Set the ones falling outside to zero - v_bo10_q = _mm_and_si128(v_bo10_q, - _mm_cvtepi32_epi64(v_inside_d)); - v_bo32_q = _mm_and_si128(v_bo32_q, - _mm_unpackhi_epi32(v_inside_d, v_inside_d)); - // Compute the candidate addresses - v_blocka[0] = _mm_add_epi64(v_ba_q, v_bo10_q); - v_blocka[1] = _mm_add_epi64(v_ba_q, v_bo32_q); -#else // ARCH_X86 // sizeof(intptr_t) == 4 - __m128i v_bo_d = _mm_load_si128((const __m128i*)&ss_os[i]); - v_bo_d = _mm_and_si128(v_bo_d, v_inside_d); - v_blocka[0] = _mm_add_epi32(v_ba_d, v_bo_d); -#endif - } - - fn_ptr->sdx4df(what, what_stride, - (const uint8_t **)&v_blocka[0], in_what_stride, - (uint32_t*)&v_sad_d); - - // Look up the component cost of the residual motion vector - { - const int32_t row0 = _mm_extract_epi16(v_diff_mv_w, 0); - const int32_t col0 = _mm_extract_epi16(v_diff_mv_w, 1); - const int32_t row1 = _mm_extract_epi16(v_diff_mv_w, 2); - const int32_t col1 = _mm_extract_epi16(v_diff_mv_w, 3); - const int32_t row2 = _mm_extract_epi16(v_diff_mv_w, 4); - const int32_t col2 = _mm_extract_epi16(v_diff_mv_w, 5); - const int32_t row3 = _mm_extract_epi16(v_diff_mv_w, 6); - const int32_t col3 = _mm_extract_epi16(v_diff_mv_w, 7); - - // Note: This is a use case for vpgather in AVX2 - const uint32_t cost0 = x->nmvsadcost[0][row0] + x->nmvsadcost[0][col0]; - const uint32_t cost1 = x->nmvsadcost[0][row1] + x->nmvsadcost[0][col1]; - const uint32_t cost2 = x->nmvsadcost[0][row2] + x->nmvsadcost[0][col2]; - const uint32_t cost3 = x->nmvsadcost[0][row3] + x->nmvsadcost[0][col3]; - - __m128i v_cost_10_d; - __m128i v_cost_32_d; - - v_cost_10_d = _mm_cvtsi32_si128(cost0); - v_cost_10_d = _mm_insert_epi32(v_cost_10_d, cost1, 1); - - v_cost_32_d = _mm_cvtsi32_si128(cost2); - v_cost_32_d = _mm_insert_epi32(v_cost_32_d, cost3, 1); - - v_cost_d = _mm_unpacklo_epi64(v_cost_10_d, v_cost_32_d); - } - - // Now add in the joint cost - { - const __m128i v_sel_d = _mm_cmpeq_epi32(v_diff_mv_w, - _mm_setzero_si128()); - const __m128i v_joint_cost_d = _mm_blendv_epi8(v_joint_cost_1_d, - v_joint_cost_0_d, - v_sel_d); - v_cost_d = _mm_add_epi32(v_cost_d, v_joint_cost_d); - } - - // Multiply by sad_per_bit - v_cost_d = _mm_mullo_epi32(v_cost_d, v_spb_d); - // ROUND_POWER_OF_TWO(v_cost_d, 8) - v_cost_d = _mm_add_epi32(v_cost_d, _mm_set1_epi32(0x80)); - v_cost_d = _mm_srai_epi32(v_cost_d, 8); - // Add the cost to the sad - v_sad_d = _mm_add_epi32(v_sad_d, v_cost_d); - - // Make the motion vectors outside the search area have max cost - // by or'ing in the comparison mask, this way the minimum search won't - // pick them. - v_sad_d = _mm_or_si128(v_sad_d, v_outside_d); - - // Find the minimum value and index horizontally in v_sad_d - { - // Try speculatively on 16 bits, so we can use the minpos intrinsic - const __m128i v_sad_w = _mm_packus_epi32(v_sad_d, v_sad_d); - const __m128i v_minp_w = _mm_minpos_epu16(v_sad_w); - - uint32_t local_best_sad = _mm_extract_epi16(v_minp_w, 0); - uint32_t local_best_idx = _mm_extract_epi16(v_minp_w, 1); - - // If the local best value is not saturated, just use it, otherwise - // find the horizontal minimum again the hard way on 32 bits. - // This is executed rarely. - if (__unlikely__(local_best_sad == 0xffff)) { - __m128i v_loval_d, v_hival_d, v_loidx_d, v_hiidx_d, v_sel_d; - - v_loval_d = v_sad_d; - v_loidx_d = _mm_set_epi32(3, 2, 1, 0); - v_hival_d = _mm_srli_si128(v_loval_d, 8); - v_hiidx_d = _mm_srli_si128(v_loidx_d, 8); - - v_sel_d = _mm_cmplt_epi32(v_hival_d, v_loval_d); - - v_loval_d = _mm_blendv_epi8(v_loval_d, v_hival_d, v_sel_d); - v_loidx_d = _mm_blendv_epi8(v_loidx_d, v_hiidx_d, v_sel_d); - v_hival_d = _mm_srli_si128(v_loval_d, 4); - v_hiidx_d = _mm_srli_si128(v_loidx_d, 4); - - v_sel_d = _mm_cmplt_epi32(v_hival_d, v_loval_d); - - v_loval_d = _mm_blendv_epi8(v_loval_d, v_hival_d, v_sel_d); - v_loidx_d = _mm_blendv_epi8(v_loidx_d, v_hiidx_d, v_sel_d); - - local_best_sad = _mm_extract_epi32(v_loval_d, 0); - local_best_idx = _mm_extract_epi32(v_loidx_d, 0); - } - - // Update the global minimum if the local minimum is smaller - if (__likely__(local_best_sad < best_sad)) { - new_bmv = ((const int_mv *)&v_these_mv_w)[local_best_idx]; - new_best_address = ((const uint8_t **)v_blocka)[local_best_idx]; - - best_sad = local_best_sad; - } - } - } - - bmv = new_bmv; - best_address = new_best_address; - - v_bmv_w = _mm_set1_epi32(bmv.as_int); -#if ARCH_X86_64 - v_ba_q = _mm_set1_epi64x((intptr_t)best_address); -#else - v_ba_d = _mm_set1_epi32((intptr_t)best_address); -#endif - - if (__unlikely__(best_address == in_what)) { - (*num00)++; - } - } - - *best_mv = bmv.as_mv; - return best_sad; -} diff --git a/vp9/vp9cx.mk b/vp9/vp9cx.mk index 5918240..3f3bdef 100644 --- a/vp9/vp9cx.mk +++ b/vp9/vp9cx.mk @@ -96,7 +96,6 @@ VP9_CX_SRCS-yes += encoder/vp9_mbgraph.h VP9_CX_SRCS-$(HAVE_SSE2) += encoder/x86/vp9_avg_intrin_sse2.c VP9_CX_SRCS-$(HAVE_SSE2) += encoder/x86/vp9_temporal_filter_apply_sse2.asm VP9_CX_SRCS-$(HAVE_SSE2) += encoder/x86/vp9_quantize_sse2.c -VP9_CX_SRCS-$(HAVE_AVX) += encoder/x86/vp9_diamond_search_sad_avx.c ifeq ($(CONFIG_VP9_HIGHBITDEPTH),yes) VP9_CX_SRCS-$(HAVE_SSE2) += encoder/x86/vp9_highbd_block_error_intrin_sse2.c endif -- 2.7.4