From 8d150ce802c40c1fab43d0e165f82de17fb781d0 Mon Sep 17 00:00:00 2001 From: raster Date: Mon, 3 Oct 2011 04:44:23 +0000 Subject: [PATCH] isolate sse3 to only its own file, so it only uses -msse3 for that git-svn-id: http://svn.enlightenment.org/svn/e/trunk/evas@63762 7cbeb6ba-43b4-40fd-8cce-4c39aea84d33 --- configure.ac | 8 ++- src/lib/engines/common/Makefile.am | 3 + src/lib/engines/common/evas_cpu.c | 6 +- src/lib/engines/common/evas_op_blend/Makefile.am | 27 ++++++++ .../common/evas_op_blend/op_blend_master_sse3.c | 72 ++++++++++++++++++++++ src/lib/engines/common/evas_op_blend_main_.c | 53 ++++------------ src/lib/include/evas_blend_ops.h | 12 ++-- 7 files changed, 129 insertions(+), 52 deletions(-) create mode 100644 src/lib/engines/common/evas_op_blend/op_blend_master_sse3.c diff --git a/configure.ac b/configure.ac index 10496b2..f26cc27 100644 --- a/configure.ac +++ b/configure.ac @@ -1225,7 +1225,7 @@ AC_ARG_ENABLE(cpu-sse, build_cpu_sse3="no" case $host_cpu in i*86) - build_cpu_sse3="no" + build_cpu_sse3="yes" ;; x86_64) build_cpu_sse3="yes" @@ -1269,12 +1269,14 @@ AC_ARG_ENABLE(cpu-sse3, # you compile with -msse3. this ALSO tries to optimize REGULAR c code # with sse3 asm.. and this breaks things so badly. so... sse3 will be # off until further notice for 32bit x86. -EVAS_SSE3_CFLAGS="-msse3 " +EVAS_SSE3_CFLAGS=" " if test "x$build_cpu_sse3" = "xyes" ; then - CFLAGS="${CFLAGS} ${EVAS_SSE3_CFLAGS}" + EVAS_SSE3_CFLAGS="-msse3 " +# CFLAGS="${CFLAGS} ${EVAS_SSE3_CFLAGS}" fi AC_SUBST(CFLAGS) +AC_SUBST(EVAS_SSE3_CFLAGS) ####################################### ## ALTIVEC diff --git a/src/lib/engines/common/Makefile.am b/src/lib/engines/common/Makefile.am index 5a059b8..9c9918d 100644 --- a/src/lib/engines/common/Makefile.am +++ b/src/lib/engines/common/Makefile.am @@ -113,3 +113,6 @@ evas_map_image_loop.c libevas_engine_common_la_DEPENDENCIES = \ $(top_builddir)/config.h + +libevas_engine_common_la_LIBADD = \ +evas_op_blend/libevas_engine_common_op_blend_master_sse3.la diff --git a/src/lib/engines/common/evas_cpu.c b/src/lib/engines/common/evas_cpu.c index 4c84de5..cff8b2b 100644 --- a/src/lib/engines/common/evas_cpu.c +++ b/src/lib/engines/common/evas_cpu.c @@ -64,13 +64,13 @@ evas_common_cpu_sse_test(void) #endif } +void evas_common_op_sse3_test(void); + void evas_common_cpu_sse3_test(void) { #ifdef BUILD_SSE3 - int data[4]; - - _mm_lddqu_si128((__m128i *)data); + evas_common_op_sse3_test(); #endif } diff --git a/src/lib/engines/common/evas_op_blend/Makefile.am b/src/lib/engines/common/evas_op_blend/Makefile.am index e3ed139..84add38 100644 --- a/src/lib/engines/common/evas_op_blend/Makefile.am +++ b/src/lib/engines/common/evas_op_blend/Makefile.am @@ -21,3 +21,30 @@ op_blend_pixel_mask_neon.c \ op_blend_pixel_mask_sse3.c \ op_blend_pixel_neon.c \ op_blend_pixel_sse3.c + +noinst_LTLIBRARIES = libevas_engine_common_op_blend_master_sse3.la + +libevas_engine_common_op_blend_master_sse3_la_SOURCES = \ +op_blend_master_sse3.c + +libevas_engine_common_op_blend_master_sse3_la_CFLAGS = \ +-I. \ +-I$(top_srcdir)/src/lib \ +-I$(top_srcdir)/src/lib/engines/common \ +-I$(top_srcdir)/src/lib/engines/common/evas_op_blend \ +-I$(top_srcdir)/src/lib/cserve \ +-I$(top_srcdir)/src/lib/include \ +-DPACKAGE_BIN_DIR=\"$(bindir)\" \ +-DPACKAGE_LIB_DIR=\"$(libdir)\" \ +-DPACKAGE_DATA_DIR=\"$(datadir)/$(PACKAGE)\" \ +@FREETYPE_CFLAGS@ @VALGRIND_CFLAGS@ \ +@PIXMAN_CFLAGS@ \ +@EET_CFLAGS@ @pthread_cflags@ \ +@WIN32_CFLAGS@ @EINA_CFLAGS@ \ +@FRIBIDI_CFLAGS@ @HARFBUZZ_CFLAGS@ \ +@PIXMAN_CFLAGS@ \ +@EVAS_SSE3_CFLAGS@ + +libevas_engine_common_op_blend_master_sse3_la_DEPENENCIES = \ +$(top_builddir)/config.h + diff --git a/src/lib/engines/common/evas_op_blend/op_blend_master_sse3.c b/src/lib/engines/common/evas_op_blend/op_blend_master_sse3.c new file mode 100644 index 0000000..0066589 --- /dev/null +++ b/src/lib/engines/common/evas_op_blend/op_blend_master_sse3.c @@ -0,0 +1,72 @@ +#define NEED_SSE3 1 + +#include "evas_common.h" + +#ifdef BUILD_SSE3 +static __m128i A_MASK_SSE3; +#endif + +extern RGBA_Gfx_Func op_blend_span_funcs[SP_LAST][SM_LAST][SC_LAST][DP_LAST][CPU_LAST]; +extern RGBA_Gfx_Pt_Func op_blend_pt_funcs[SP_LAST][SM_LAST][SC_LAST][DP_LAST][CPU_LAST]; + +extern RGBA_Gfx_Func op_blend_rel_span_funcs[SP_LAST][SM_LAST][SC_LAST][DP_LAST][CPU_LAST]; +extern RGBA_Gfx_Pt_Func op_blend_rel_pt_funcs[SP_LAST][SM_LAST][SC_LAST][DP_LAST][CPU_LAST]; + +# include "op_blend_pixel_sse3.c" +# include "op_blend_color_sse3.c" +# include "op_blend_pixel_color_sse3.c" +# include "op_blend_pixel_mask_sse3.c" +# include "op_blend_mask_color_sse3.c" + +void +evas_common_op_blend_init_sse3(void) +{ +#ifdef BUILD_SSE3 + GA_MASK_SSE3 = _mm_set_epi32(0x00FF00FF, 0x00FF00FF, 0x00FF00FF, 0x00FF00FF); + RB_MASK_SSE3 = _mm_set_epi32(0xFF00FF00, 0xFF00FF00, 0xFF00FF00, 0xFF00FF00); + SYM4_MASK_SSE3 = _mm_set_epi32(0x00FF00FF, 0x000000FF, 0x00FF00FF, 0x000000FF); + RGB_MASK_SSE3 = _mm_set_epi32(0xFFFFFF, 0xFFFFFF, 0xFFFFFF, 0xFFFFFF); + A_MASK_SSE3 = _mm_set_epi32(0xFF000000, 0xFF000000, 0xFF000000, 0xFF000000); + ALPHA_SSE3 = _mm_set_epi32(256, 256, 256, 256); + + init_blend_pixel_span_funcs_sse3(); + init_blend_pixel_color_span_funcs_sse3(); + init_blend_pixel_mask_span_funcs_sse3(); + init_blend_color_span_funcs_sse3(); + init_blend_mask_color_span_funcs_sse3(); + + init_blend_pixel_pt_funcs_sse3(); + init_blend_pixel_color_pt_funcs_sse3(); + init_blend_pixel_mask_pt_funcs_sse3(); + init_blend_color_pt_funcs_sse3(); + init_blend_mask_color_pt_funcs_sse3(); +#endif +} + +void +evas_common_op_blend_rel_init_sse3(void) +{ +#ifdef BUILD_SSE3 + init_blend_rel_pixel_span_funcs_sse3(); + init_blend_rel_pixel_color_span_funcs_sse3(); + init_blend_rel_pixel_mask_span_funcs_sse3(); + init_blend_rel_color_span_funcs_sse3(); + init_blend_rel_mask_color_span_funcs_sse3(); + + init_blend_rel_pixel_pt_funcs_sse3(); + init_blend_rel_pixel_color_pt_funcs_sse3(); + init_blend_rel_pixel_mask_pt_funcs_sse3(); + init_blend_rel_color_pt_funcs_sse3(); + init_blend_rel_mask_color_pt_funcs_sse3(); +#endif +} + +void +evas_common_op_sse3_test(void) +{ +#ifdef BUILD_SSE3 + int data[4]; + + _mm_lddqu_si128((__m128i *)data); +#endif +} diff --git a/src/lib/engines/common/evas_op_blend_main_.c b/src/lib/engines/common/evas_op_blend_main_.c index d358d12..a1a5de8 100644 --- a/src/lib/engines/common/evas_op_blend_main_.c +++ b/src/lib/engines/common/evas_op_blend_main_.c @@ -1,7 +1,7 @@ #include "evas_common.h" -static RGBA_Gfx_Func op_blend_span_funcs[SP_LAST][SM_LAST][SC_LAST][DP_LAST][CPU_LAST]; -static RGBA_Gfx_Pt_Func op_blend_pt_funcs[SP_LAST][SM_LAST][SC_LAST][DP_LAST][CPU_LAST]; +RGBA_Gfx_Func op_blend_span_funcs[SP_LAST][SM_LAST][SC_LAST][DP_LAST][CPU_LAST]; +RGBA_Gfx_Pt_Func op_blend_pt_funcs[SP_LAST][SM_LAST][SC_LAST][DP_LAST][CPU_LAST]; static void op_blend_init(void); static void op_blend_shutdown(void); @@ -35,8 +35,8 @@ evas_common_gfx_compositor_blend_get(void) } -static RGBA_Gfx_Func op_blend_rel_span_funcs[SP_LAST][SM_LAST][SC_LAST][DP_LAST][CPU_LAST]; -static RGBA_Gfx_Pt_Func op_blend_rel_pt_funcs[SP_LAST][SM_LAST][SC_LAST][DP_LAST][CPU_LAST]; +RGBA_Gfx_Func op_blend_rel_span_funcs[SP_LAST][SM_LAST][SC_LAST][DP_LAST][CPU_LAST]; +RGBA_Gfx_Pt_Func op_blend_rel_pt_funcs[SP_LAST][SM_LAST][SC_LAST][DP_LAST][CPU_LAST]; static void op_blend_rel_init(void); static void op_blend_rel_shutdown(void); @@ -84,15 +84,6 @@ evas_common_gfx_compositor_blend_rel_get(void) # include "./evas_op_blend/op_blend_mask_color_i386.c" //# include "./evas_op_blend/op_blend_pixel_mask_color_i386.c" -#ifdef BUILD_SSE3 -static __m128i A_MASK_SSE3; -#endif -# include "./evas_op_blend/op_blend_pixel_sse3.c" -# include "./evas_op_blend/op_blend_color_sse3.c" -# include "./evas_op_blend/op_blend_pixel_color_sse3.c" -# include "./evas_op_blend/op_blend_pixel_mask_sse3.c" -# include "./evas_op_blend/op_blend_mask_color_sse3.c" - # include "./evas_op_blend/op_blend_pixel_neon.c" # include "./evas_op_blend/op_blend_color_neon.c" # include "./evas_op_blend/op_blend_pixel_color_neon.c" @@ -100,30 +91,17 @@ static __m128i A_MASK_SSE3; # include "./evas_op_blend/op_blend_mask_color_neon.c" //# include "./evas_op_blend/op_blend_pixel_mask_color_neon.c" +#ifdef BUILD_SSE3 +void evas_common_op_blend_init_sse3(void); +#endif + static void op_blend_init(void) { memset(op_blend_span_funcs, 0, sizeof(op_blend_span_funcs)); memset(op_blend_pt_funcs, 0, sizeof(op_blend_pt_funcs)); #ifdef BUILD_SSE3 - GA_MASK_SSE3 = _mm_set_epi32(0x00FF00FF, 0x00FF00FF, 0x00FF00FF, 0x00FF00FF); - RB_MASK_SSE3 = _mm_set_epi32(0xFF00FF00, 0xFF00FF00, 0xFF00FF00, 0xFF00FF00); - SYM4_MASK_SSE3 = _mm_set_epi32(0x00FF00FF, 0x000000FF, 0x00FF00FF, 0x000000FF); - RGB_MASK_SSE3 = _mm_set_epi32(0xFFFFFF, 0xFFFFFF, 0xFFFFFF, 0xFFFFFF); - A_MASK_SSE3 = _mm_set_epi32(0xFF000000, 0xFF000000, 0xFF000000, 0xFF000000); - ALPHA_SSE3 = _mm_set_epi32(256, 256, 256, 256); - - init_blend_pixel_span_funcs_sse3(); - init_blend_pixel_color_span_funcs_sse3(); - init_blend_pixel_mask_span_funcs_sse3(); - init_blend_color_span_funcs_sse3(); - init_blend_mask_color_span_funcs_sse3(); - - init_blend_pixel_pt_funcs_sse3(); - init_blend_pixel_color_pt_funcs_sse3(); - init_blend_pixel_mask_pt_funcs_sse3(); - init_blend_color_pt_funcs_sse3(); - init_blend_mask_color_pt_funcs_sse3(); + evas_common_op_blend_init_sse3(); #endif #ifdef BUILD_MMX init_blend_pixel_span_funcs_mmx(); @@ -406,6 +384,7 @@ op_blend_pixel_mask_pt_get(Image_Entry_Flags src_flags, RGBA_Image *dst) return blend_gfx_pt_func_cpu(s, m, c, d); } +void evas_common_op_blend_rel_init_sse3(void); static void op_blend_rel_init(void) @@ -413,17 +392,7 @@ op_blend_rel_init(void) memset(op_blend_rel_span_funcs, 0, sizeof(op_blend_rel_span_funcs)); memset(op_blend_rel_pt_funcs, 0, sizeof(op_blend_rel_pt_funcs)); #ifdef BUILD_SSE3 - init_blend_rel_pixel_span_funcs_sse3(); - init_blend_rel_pixel_color_span_funcs_sse3(); - init_blend_rel_pixel_mask_span_funcs_sse3(); - init_blend_rel_color_span_funcs_sse3(); - init_blend_rel_mask_color_span_funcs_sse3(); - - init_blend_rel_pixel_pt_funcs_sse3(); - init_blend_rel_pixel_color_pt_funcs_sse3(); - init_blend_rel_pixel_mask_pt_funcs_sse3(); - init_blend_rel_color_pt_funcs_sse3(); - init_blend_rel_mask_color_pt_funcs_sse3(); + evas_common_op_blend_rel_init_sse3(); #endif #ifdef BUILD_MMX init_blend_rel_pixel_span_funcs_mmx(); diff --git a/src/lib/include/evas_blend_ops.h b/src/lib/include/evas_blend_ops.h index bc4b81c..1429596 100644 --- a/src/lib/include/evas_blend_ops.h +++ b/src/lib/include/evas_blend_ops.h @@ -5,12 +5,14 @@ #include "evas_mmx.h" #endif -#if defined BUILD_SSE3 -#include -#endif - #include "config.h" +#ifdef NEED_SSE3 +# if defined BUILD_SSE3 +# include +# endif +#endif + /* src pixel flags: */ /* pixels none */ @@ -187,6 +189,7 @@ extern const DATA32 ALPHA_256; /* some useful SSE3 inline functions */ +#ifdef NEED_SSE3 #ifdef BUILD_SSE3 static __m128i GA_MASK_SSE3; @@ -370,5 +373,6 @@ mul3_sym_sse3(__m128i x, __m128i y) { #endif +#endif #endif -- 2.7.4