/*
* VP8 DSP functions x86-optimized
* Copyright (c) 2010 Ronald S. Bultje <rsbultje@gmail.com>
- * Copyright (c) 2010 Jason Garrett-Glaser <darkshikari@gmail.com>
+ * Copyright (c) 2010 Fiona Glaser <fiona@x264.com>
*
* This file is part of Libav.
*
* Foundation, Inc., 51 Franklin Street, Fifth Floor, Boston, MA 02110-1301 USA
*/
+#include "libavutil/attributes.h"
#include "libavutil/cpu.h"
#include "libavutil/mem.h"
#include "libavutil/x86/asm.h"
+#include "libavutil/x86/cpu.h"
#include "libavcodec/vp8dsp.h"
#if HAVE_YASM
/*
* MC functions
*/
-extern void ff_put_vp8_epel4_h4_mmxext(uint8_t *dst, ptrdiff_t dststride,
- uint8_t *src, ptrdiff_t srcstride,
- int height, int mx, int my);
-extern void ff_put_vp8_epel4_h6_mmxext(uint8_t *dst, ptrdiff_t dststride,
- uint8_t *src, ptrdiff_t srcstride,
- int height, int mx, int my);
-extern void ff_put_vp8_epel4_v4_mmxext(uint8_t *dst, ptrdiff_t dststride,
- uint8_t *src, ptrdiff_t srcstride,
- int height, int mx, int my);
-extern void ff_put_vp8_epel4_v6_mmxext(uint8_t *dst, ptrdiff_t dststride,
- uint8_t *src, ptrdiff_t srcstride,
- int height, int mx, int my);
-
-extern void ff_put_vp8_epel8_h4_sse2 (uint8_t *dst, ptrdiff_t dststride,
- uint8_t *src, ptrdiff_t srcstride,
- int height, int mx, int my);
-extern void ff_put_vp8_epel8_h6_sse2 (uint8_t *dst, ptrdiff_t dststride,
- uint8_t *src, ptrdiff_t srcstride,
- int height, int mx, int my);
-extern void ff_put_vp8_epel8_v4_sse2 (uint8_t *dst, ptrdiff_t dststride,
- uint8_t *src, ptrdiff_t srcstride,
- int height, int mx, int my);
-extern void ff_put_vp8_epel8_v6_sse2 (uint8_t *dst, ptrdiff_t dststride,
- uint8_t *src, ptrdiff_t srcstride,
- int height, int mx, int my);
-
-extern void ff_put_vp8_epel4_h4_ssse3 (uint8_t *dst, ptrdiff_t dststride,
- uint8_t *src, ptrdiff_t srcstride,
- int height, int mx, int my);
-extern void ff_put_vp8_epel4_h6_ssse3 (uint8_t *dst, ptrdiff_t dststride,
- uint8_t *src, ptrdiff_t srcstride,
- int height, int mx, int my);
-extern void ff_put_vp8_epel4_v4_ssse3 (uint8_t *dst, ptrdiff_t dststride,
- uint8_t *src, ptrdiff_t srcstride,
- int height, int mx, int my);
-extern void ff_put_vp8_epel4_v6_ssse3 (uint8_t *dst, ptrdiff_t dststride,
- uint8_t *src, ptrdiff_t srcstride,
- int height, int mx, int my);
-extern void ff_put_vp8_epel8_h4_ssse3 (uint8_t *dst, ptrdiff_t dststride,
- uint8_t *src, ptrdiff_t srcstride,
- int height, int mx, int my);
-extern void ff_put_vp8_epel8_h6_ssse3 (uint8_t *dst, ptrdiff_t dststride,
- uint8_t *src, ptrdiff_t srcstride,
- int height, int mx, int my);
-extern void ff_put_vp8_epel8_v4_ssse3 (uint8_t *dst, ptrdiff_t dststride,
- uint8_t *src, ptrdiff_t srcstride,
- int height, int mx, int my);
-extern void ff_put_vp8_epel8_v6_ssse3 (uint8_t *dst, ptrdiff_t dststride,
- uint8_t *src, ptrdiff_t srcstride,
- int height, int mx, int my);
-
-extern void ff_put_vp8_bilinear4_h_mmxext(uint8_t *dst, ptrdiff_t dststride,
- uint8_t *src, ptrdiff_t srcstride,
- int height, int mx, int my);
-extern void ff_put_vp8_bilinear8_h_sse2 (uint8_t *dst, ptrdiff_t dststride,
- uint8_t *src, ptrdiff_t srcstride,
- int height, int mx, int my);
-extern void ff_put_vp8_bilinear4_h_ssse3 (uint8_t *dst, ptrdiff_t dststride,
- uint8_t *src, ptrdiff_t srcstride,
- int height, int mx, int my);
-extern void ff_put_vp8_bilinear8_h_ssse3 (uint8_t *dst, ptrdiff_t dststride,
- uint8_t *src, ptrdiff_t srcstride,
- int height, int mx, int my);
-
-extern void ff_put_vp8_bilinear4_v_mmxext(uint8_t *dst, ptrdiff_t dststride,
- uint8_t *src, ptrdiff_t srcstride,
- int height, int mx, int my);
-extern void ff_put_vp8_bilinear8_v_sse2 (uint8_t *dst, ptrdiff_t dststride,
- uint8_t *src, ptrdiff_t srcstride,
- int height, int mx, int my);
-extern void ff_put_vp8_bilinear4_v_ssse3 (uint8_t *dst, ptrdiff_t dststride,
- uint8_t *src, ptrdiff_t srcstride,
- int height, int mx, int my);
-extern void ff_put_vp8_bilinear8_v_ssse3 (uint8_t *dst, ptrdiff_t dststride,
- uint8_t *src, ptrdiff_t srcstride,
- int height, int mx, int my);
-
-
-extern void ff_put_vp8_pixels8_mmx (uint8_t *dst, ptrdiff_t dststride,
- uint8_t *src, ptrdiff_t srcstride,
- int height, int mx, int my);
-extern void ff_put_vp8_pixels16_mmx(uint8_t *dst, ptrdiff_t dststride,
- uint8_t *src, ptrdiff_t srcstride,
- int height, int mx, int my);
-extern void ff_put_vp8_pixels16_sse(uint8_t *dst, ptrdiff_t dststride,
- uint8_t *src, ptrdiff_t srcstride,
- int height, int mx, int my);
+void ff_put_vp8_epel4_h4_mmxext(uint8_t *dst, ptrdiff_t dststride,
+ uint8_t *src, ptrdiff_t srcstride,
+ int height, int mx, int my);
+void ff_put_vp8_epel4_h6_mmxext(uint8_t *dst, ptrdiff_t dststride,
+ uint8_t *src, ptrdiff_t srcstride,
+ int height, int mx, int my);
+void ff_put_vp8_epel4_v4_mmxext(uint8_t *dst, ptrdiff_t dststride,
+ uint8_t *src, ptrdiff_t srcstride,
+ int height, int mx, int my);
+void ff_put_vp8_epel4_v6_mmxext(uint8_t *dst, ptrdiff_t dststride,
+ uint8_t *src, ptrdiff_t srcstride,
+ int height, int mx, int my);
+
+void ff_put_vp8_epel8_h4_sse2 (uint8_t *dst, ptrdiff_t dststride,
+ uint8_t *src, ptrdiff_t srcstride,
+ int height, int mx, int my);
+void ff_put_vp8_epel8_h6_sse2 (uint8_t *dst, ptrdiff_t dststride,
+ uint8_t *src, ptrdiff_t srcstride,
+ int height, int mx, int my);
+void ff_put_vp8_epel8_v4_sse2 (uint8_t *dst, ptrdiff_t dststride,
+ uint8_t *src, ptrdiff_t srcstride,
+ int height, int mx, int my);
+void ff_put_vp8_epel8_v6_sse2 (uint8_t *dst, ptrdiff_t dststride,
+ uint8_t *src, ptrdiff_t srcstride,
+ int height, int mx, int my);
+
+void ff_put_vp8_epel4_h4_ssse3 (uint8_t *dst, ptrdiff_t dststride,
+ uint8_t *src, ptrdiff_t srcstride,
+ int height, int mx, int my);
+void ff_put_vp8_epel4_h6_ssse3 (uint8_t *dst, ptrdiff_t dststride,
+ uint8_t *src, ptrdiff_t srcstride,
+ int height, int mx, int my);
+void ff_put_vp8_epel4_v4_ssse3 (uint8_t *dst, ptrdiff_t dststride,
+ uint8_t *src, ptrdiff_t srcstride,
+ int height, int mx, int my);
+void ff_put_vp8_epel4_v6_ssse3 (uint8_t *dst, ptrdiff_t dststride,
+ uint8_t *src, ptrdiff_t srcstride,
+ int height, int mx, int my);
+void ff_put_vp8_epel8_h4_ssse3 (uint8_t *dst, ptrdiff_t dststride,
+ uint8_t *src, ptrdiff_t srcstride,
+ int height, int mx, int my);
+void ff_put_vp8_epel8_h6_ssse3 (uint8_t *dst, ptrdiff_t dststride,
+ uint8_t *src, ptrdiff_t srcstride,
+ int height, int mx, int my);
+void ff_put_vp8_epel8_v4_ssse3 (uint8_t *dst, ptrdiff_t dststride,
+ uint8_t *src, ptrdiff_t srcstride,
+ int height, int mx, int my);
+void ff_put_vp8_epel8_v6_ssse3 (uint8_t *dst, ptrdiff_t dststride,
+ uint8_t *src, ptrdiff_t srcstride,
+ int height, int mx, int my);
+
+void ff_put_vp8_bilinear4_h_mmxext(uint8_t *dst, ptrdiff_t dststride,
+ uint8_t *src, ptrdiff_t srcstride,
+ int height, int mx, int my);
+void ff_put_vp8_bilinear8_h_sse2 (uint8_t *dst, ptrdiff_t dststride,
+ uint8_t *src, ptrdiff_t srcstride,
+ int height, int mx, int my);
+void ff_put_vp8_bilinear4_h_ssse3 (uint8_t *dst, ptrdiff_t dststride,
+ uint8_t *src, ptrdiff_t srcstride,
+ int height, int mx, int my);
+void ff_put_vp8_bilinear8_h_ssse3 (uint8_t *dst, ptrdiff_t dststride,
+ uint8_t *src, ptrdiff_t srcstride,
+ int height, int mx, int my);
+
+void ff_put_vp8_bilinear4_v_mmxext(uint8_t *dst, ptrdiff_t dststride,
+ uint8_t *src, ptrdiff_t srcstride,
+ int height, int mx, int my);
+void ff_put_vp8_bilinear8_v_sse2 (uint8_t *dst, ptrdiff_t dststride,
+ uint8_t *src, ptrdiff_t srcstride,
+ int height, int mx, int my);
+void ff_put_vp8_bilinear4_v_ssse3 (uint8_t *dst, ptrdiff_t dststride,
+ uint8_t *src, ptrdiff_t srcstride,
+ int height, int mx, int my);
+void ff_put_vp8_bilinear8_v_ssse3 (uint8_t *dst, ptrdiff_t dststride,
+ uint8_t *src, ptrdiff_t srcstride,
+ int height, int mx, int my);
+
+
+void ff_put_vp8_pixels8_mmx (uint8_t *dst, ptrdiff_t dststride,
+ uint8_t *src, ptrdiff_t srcstride,
+ int height, int mx, int my);
+void ff_put_vp8_pixels16_mmx(uint8_t *dst, ptrdiff_t dststride,
+ uint8_t *src, ptrdiff_t srcstride,
+ int height, int mx, int my);
+void ff_put_vp8_pixels16_sse(uint8_t *dst, ptrdiff_t dststride,
+ uint8_t *src, ptrdiff_t srcstride,
+ int height, int mx, int my);
#define TAP_W16(OPT, FILTERTYPE, TAPTYPE) \
static void ff_put_vp8_ ## FILTERTYPE ## 16_ ## TAPTYPE ## _ ## OPT( \
HVBILIN(ssse3, 8, 8, 16)
HVBILIN(ssse3, 8, 16, 16)
-extern void ff_vp8_idct_dc_add_mmx(uint8_t *dst, DCTELEM block[16],
- ptrdiff_t stride);
-extern void ff_vp8_idct_dc_add_sse4(uint8_t *dst, DCTELEM block[16],
- ptrdiff_t stride);
-extern void ff_vp8_idct_dc_add4y_mmx(uint8_t *dst, DCTELEM block[4][16],
- ptrdiff_t stride);
-extern void ff_vp8_idct_dc_add4y_sse2(uint8_t *dst, DCTELEM block[4][16],
- ptrdiff_t stride);
-extern void ff_vp8_idct_dc_add4uv_mmx(uint8_t *dst, DCTELEM block[2][16],
- ptrdiff_t stride);
-extern void ff_vp8_luma_dc_wht_mmx(DCTELEM block[4][4][16], DCTELEM dc[16]);
-extern void ff_vp8_luma_dc_wht_sse(DCTELEM block[4][4][16], DCTELEM dc[16]);
-extern void ff_vp8_idct_add_mmx(uint8_t *dst, DCTELEM block[16],
- ptrdiff_t stride);
-extern void ff_vp8_idct_add_sse(uint8_t *dst, DCTELEM block[16],
- ptrdiff_t stride);
-
-#define DECLARE_LOOP_FILTER(NAME)\
-extern void ff_vp8_v_loop_filter_simple_ ## NAME(uint8_t *dst, \
- ptrdiff_t stride, \
- int flim);\
-extern void ff_vp8_h_loop_filter_simple_ ## NAME(uint8_t *dst, \
- ptrdiff_t stride, \
- int flim);\
-extern void ff_vp8_v_loop_filter16y_inner_ ## NAME (uint8_t *dst, \
- ptrdiff_t stride,\
- int e, int i, int hvt);\
-extern void ff_vp8_h_loop_filter16y_inner_ ## NAME (uint8_t *dst, \
- ptrdiff_t stride,\
- int e, int i, int hvt);\
-extern void ff_vp8_v_loop_filter8uv_inner_ ## NAME (uint8_t *dstU, \
- uint8_t *dstV,\
- ptrdiff_t s, \
- int e, int i, int hvt);\
-extern void ff_vp8_h_loop_filter8uv_inner_ ## NAME (uint8_t *dstU, \
- uint8_t *dstV,\
- ptrdiff_t s, \
- int e, int i, int hvt);\
-extern void ff_vp8_v_loop_filter16y_mbedge_ ## NAME(uint8_t *dst, \
- ptrdiff_t stride,\
- int e, int i, int hvt);\
-extern void ff_vp8_h_loop_filter16y_mbedge_ ## NAME(uint8_t *dst, \
- ptrdiff_t stride,\
- int e, int i, int hvt);\
-extern void ff_vp8_v_loop_filter8uv_mbedge_ ## NAME(uint8_t *dstU, \
- uint8_t *dstV,\
- ptrdiff_t s, \
- int e, int i, int hvt);\
-extern void ff_vp8_h_loop_filter8uv_mbedge_ ## NAME(uint8_t *dstU, \
- uint8_t *dstV,\
- ptrdiff_t s, \
- int e, int i, int hvt);
+void ff_vp8_idct_dc_add_mmx(uint8_t *dst, int16_t block[16],
+ ptrdiff_t stride);
+void ff_vp8_idct_dc_add_sse4(uint8_t *dst, int16_t block[16],
+ ptrdiff_t stride);
+void ff_vp8_idct_dc_add4y_mmx(uint8_t *dst, int16_t block[4][16],
+ ptrdiff_t stride);
+void ff_vp8_idct_dc_add4y_sse2(uint8_t *dst, int16_t block[4][16],
+ ptrdiff_t stride);
+void ff_vp8_idct_dc_add4uv_mmx(uint8_t *dst, int16_t block[2][16],
+ ptrdiff_t stride);
+void ff_vp8_luma_dc_wht_mmx(int16_t block[4][4][16], int16_t dc[16]);
+void ff_vp8_luma_dc_wht_sse(int16_t block[4][4][16], int16_t dc[16]);
+void ff_vp8_idct_add_mmx(uint8_t *dst, int16_t block[16], ptrdiff_t stride);
+void ff_vp8_idct_add_sse(uint8_t *dst, int16_t block[16], ptrdiff_t stride);
+
+#define DECLARE_LOOP_FILTER(NAME) \
+void ff_vp8_v_loop_filter_simple_ ## NAME(uint8_t *dst, \
+ ptrdiff_t stride, \
+ int flim); \
+void ff_vp8_h_loop_filter_simple_ ## NAME(uint8_t *dst, \
+ ptrdiff_t stride, \
+ int flim); \
+void ff_vp8_v_loop_filter16y_inner_ ## NAME (uint8_t *dst, \
+ ptrdiff_t stride, \
+ int e, int i, int hvt); \
+void ff_vp8_h_loop_filter16y_inner_ ## NAME (uint8_t *dst, \
+ ptrdiff_t stride, \
+ int e, int i, int hvt); \
+void ff_vp8_v_loop_filter8uv_inner_ ## NAME (uint8_t *dstU, \
+ uint8_t *dstV, \
+ ptrdiff_t s, \
+ int e, int i, int hvt); \
+void ff_vp8_h_loop_filter8uv_inner_ ## NAME (uint8_t *dstU, \
+ uint8_t *dstV, \
+ ptrdiff_t s, \
+ int e, int i, int hvt); \
+void ff_vp8_v_loop_filter16y_mbedge_ ## NAME(uint8_t *dst, \
+ ptrdiff_t stride, \
+ int e, int i, int hvt); \
+void ff_vp8_h_loop_filter16y_mbedge_ ## NAME(uint8_t *dst, \
+ ptrdiff_t stride, \
+ int e, int i, int hvt); \
+void ff_vp8_v_loop_filter8uv_mbedge_ ## NAME(uint8_t *dstU, \
+ uint8_t *dstV, \
+ ptrdiff_t s, \
+ int e, int i, int hvt); \
+void ff_vp8_h_loop_filter8uv_mbedge_ ## NAME(uint8_t *dstU, \
+ uint8_t *dstV, \
+ ptrdiff_t s, \
+ int e, int i, int hvt);
DECLARE_LOOP_FILTER(mmx)
DECLARE_LOOP_FILTER(mmxext)
c->put_vp8_bilinear_pixels_tab[IDX][2][2] = ff_put_vp8_bilinear ## SIZE ## _hv_ ## OPT
-av_cold void ff_vp8dsp_init_x86(VP8DSPContext* c)
+av_cold void ff_vp78dsp_init_x86(VP8DSPContext *c)
{
#if HAVE_YASM
- int mm_flags = av_get_cpu_flags();
+ int cpu_flags = av_get_cpu_flags();
- if (mm_flags & AV_CPU_FLAG_MMX) {
- c->vp8_idct_dc_add = ff_vp8_idct_dc_add_mmx;
- c->vp8_idct_dc_add4uv = ff_vp8_idct_dc_add4uv_mmx;
+ if (EXTERNAL_MMX(cpu_flags)) {
#if ARCH_X86_32
- c->vp8_idct_dc_add4y = ff_vp8_idct_dc_add4y_mmx;
- c->vp8_idct_add = ff_vp8_idct_add_mmx;
- c->vp8_luma_dc_wht = ff_vp8_luma_dc_wht_mmx;
c->put_vp8_epel_pixels_tab[0][0][0] =
c->put_vp8_bilinear_pixels_tab[0][0][0] = ff_put_vp8_pixels16_mmx;
#endif
c->put_vp8_epel_pixels_tab[1][0][0] =
c->put_vp8_bilinear_pixels_tab[1][0][0] = ff_put_vp8_pixels8_mmx;
+ }
+
+ /* note that 4-tap width=16 functions are missing because w=16
+ * is only used for luma, and luma is always a copy or sixtap. */
+ if (EXTERNAL_MMXEXT(cpu_flags)) {
+ VP8_MC_FUNC(2, 4, mmxext);
+ VP8_BILINEAR_MC_FUNC(2, 4, mmxext);
+#if ARCH_X86_32
+ VP8_LUMA_MC_FUNC(0, 16, mmxext);
+ VP8_MC_FUNC(1, 8, mmxext);
+ VP8_BILINEAR_MC_FUNC(0, 16, mmxext);
+ VP8_BILINEAR_MC_FUNC(1, 8, mmxext);
+#endif
+ }
+
+ if (EXTERNAL_SSE(cpu_flags)) {
+ c->put_vp8_epel_pixels_tab[0][0][0] =
+ c->put_vp8_bilinear_pixels_tab[0][0][0] = ff_put_vp8_pixels16_sse;
+ }
+
+ if (EXTERNAL_SSE2(cpu_flags) && (cpu_flags & AV_CPU_FLAG_SSE2SLOW)) {
+ VP8_LUMA_MC_FUNC(0, 16, sse2);
+ VP8_MC_FUNC(1, 8, sse2);
+ VP8_BILINEAR_MC_FUNC(0, 16, sse2);
+ VP8_BILINEAR_MC_FUNC(1, 8, sse2);
+ }
+
+ if (EXTERNAL_SSSE3(cpu_flags)) {
+ VP8_LUMA_MC_FUNC(0, 16, ssse3);
+ VP8_MC_FUNC(1, 8, ssse3);
+ VP8_MC_FUNC(2, 4, ssse3);
+ VP8_BILINEAR_MC_FUNC(0, 16, ssse3);
+ VP8_BILINEAR_MC_FUNC(1, 8, ssse3);
+ VP8_BILINEAR_MC_FUNC(2, 4, ssse3);
+ }
+#endif /* HAVE_YASM */
+}
+
+av_cold void ff_vp8dsp_init_x86(VP8DSPContext *c)
+{
+#if HAVE_YASM
+ int cpu_flags = av_get_cpu_flags();
+ if (EXTERNAL_MMX(cpu_flags)) {
+ c->vp8_idct_dc_add = ff_vp8_idct_dc_add_mmx;
+ c->vp8_idct_dc_add4uv = ff_vp8_idct_dc_add4uv_mmx;
#if ARCH_X86_32
+ c->vp8_idct_dc_add4y = ff_vp8_idct_dc_add4y_mmx;
+ c->vp8_idct_add = ff_vp8_idct_add_mmx;
+ c->vp8_luma_dc_wht = ff_vp8_luma_dc_wht_mmx;
+
c->vp8_v_loop_filter_simple = ff_vp8_v_loop_filter_simple_mmx;
c->vp8_h_loop_filter_simple = ff_vp8_h_loop_filter_simple_mmx;
/* note that 4-tap width=16 functions are missing because w=16
* is only used for luma, and luma is always a copy or sixtap. */
- if (mm_flags & AV_CPU_FLAG_MMXEXT) {
- VP8_MC_FUNC(2, 4, mmxext);
- VP8_BILINEAR_MC_FUNC(2, 4, mmxext);
+ if (EXTERNAL_MMXEXT(cpu_flags)) {
#if ARCH_X86_32
- VP8_LUMA_MC_FUNC(0, 16, mmxext);
- VP8_MC_FUNC(1, 8, mmxext);
- VP8_BILINEAR_MC_FUNC(0, 16, mmxext);
- VP8_BILINEAR_MC_FUNC(1, 8, mmxext);
-
c->vp8_v_loop_filter_simple = ff_vp8_v_loop_filter_simple_mmxext;
c->vp8_h_loop_filter_simple = ff_vp8_h_loop_filter_simple_mmxext;
#endif
}
- if (mm_flags & AV_CPU_FLAG_SSE) {
+ if (EXTERNAL_SSE(cpu_flags)) {
c->vp8_idct_add = ff_vp8_idct_add_sse;
c->vp8_luma_dc_wht = ff_vp8_luma_dc_wht_sse;
- c->put_vp8_epel_pixels_tab[0][0][0] =
- c->put_vp8_bilinear_pixels_tab[0][0][0] = ff_put_vp8_pixels16_sse;
}
- if (mm_flags & (AV_CPU_FLAG_SSE2|AV_CPU_FLAG_SSE2SLOW)) {
- VP8_LUMA_MC_FUNC(0, 16, sse2);
- VP8_MC_FUNC(1, 8, sse2);
- VP8_BILINEAR_MC_FUNC(0, 16, sse2);
- VP8_BILINEAR_MC_FUNC(1, 8, sse2);
-
+ if (EXTERNAL_SSE2(cpu_flags) && (cpu_flags & AV_CPU_FLAG_SSE2SLOW)) {
c->vp8_v_loop_filter_simple = ff_vp8_v_loop_filter_simple_sse2;
-#if ARCH_X86_64 || HAVE_ALIGNED_STACK
c->vp8_v_loop_filter16y_inner = ff_vp8_v_loop_filter16y_inner_sse2;
c->vp8_v_loop_filter8uv_inner = ff_vp8_v_loop_filter8uv_inner_sse2;
c->vp8_v_loop_filter16y = ff_vp8_v_loop_filter16y_mbedge_sse2;
c->vp8_v_loop_filter8uv = ff_vp8_v_loop_filter8uv_mbedge_sse2;
-#endif
}
- if (mm_flags & AV_CPU_FLAG_SSE2) {
+ if (EXTERNAL_SSE2(cpu_flags)) {
c->vp8_idct_dc_add4y = ff_vp8_idct_dc_add4y_sse2;
c->vp8_h_loop_filter_simple = ff_vp8_h_loop_filter_simple_sse2;
-#if ARCH_X86_64 || HAVE_ALIGNED_STACK
c->vp8_h_loop_filter16y_inner = ff_vp8_h_loop_filter16y_inner_sse2;
c->vp8_h_loop_filter8uv_inner = ff_vp8_h_loop_filter8uv_inner_sse2;
c->vp8_h_loop_filter16y = ff_vp8_h_loop_filter16y_mbedge_sse2;
c->vp8_h_loop_filter8uv = ff_vp8_h_loop_filter8uv_mbedge_sse2;
-#endif
}
- if (mm_flags & AV_CPU_FLAG_SSSE3) {
- VP8_LUMA_MC_FUNC(0, 16, ssse3);
- VP8_MC_FUNC(1, 8, ssse3);
- VP8_MC_FUNC(2, 4, ssse3);
- VP8_BILINEAR_MC_FUNC(0, 16, ssse3);
- VP8_BILINEAR_MC_FUNC(1, 8, ssse3);
- VP8_BILINEAR_MC_FUNC(2, 4, ssse3);
-
+ if (EXTERNAL_SSSE3(cpu_flags)) {
c->vp8_v_loop_filter_simple = ff_vp8_v_loop_filter_simple_ssse3;
c->vp8_h_loop_filter_simple = ff_vp8_h_loop_filter_simple_ssse3;
-#if ARCH_X86_64 || HAVE_ALIGNED_STACK
c->vp8_v_loop_filter16y_inner = ff_vp8_v_loop_filter16y_inner_ssse3;
c->vp8_h_loop_filter16y_inner = ff_vp8_h_loop_filter16y_inner_ssse3;
c->vp8_v_loop_filter8uv_inner = ff_vp8_v_loop_filter8uv_inner_ssse3;
c->vp8_h_loop_filter16y = ff_vp8_h_loop_filter16y_mbedge_ssse3;
c->vp8_v_loop_filter8uv = ff_vp8_v_loop_filter8uv_mbedge_ssse3;
c->vp8_h_loop_filter8uv = ff_vp8_h_loop_filter8uv_mbedge_ssse3;
-#endif
}
- if (mm_flags & AV_CPU_FLAG_SSE4) {
+ if (EXTERNAL_SSE4(cpu_flags)) {
c->vp8_idct_dc_add = ff_vp8_idct_dc_add_sse4;
c->vp8_h_loop_filter_simple = ff_vp8_h_loop_filter_simple_sse4;
-#if ARCH_X86_64 || HAVE_ALIGNED_STACK
c->vp8_h_loop_filter16y = ff_vp8_h_loop_filter16y_mbedge_sse4;
c->vp8_h_loop_filter8uv = ff_vp8_h_loop_filter8uv_mbedge_sse4;
-#endif
}
#endif /* HAVE_YASM */
}