Changeset 262b177b0890b6b1943cb9c4838bcffc5152f290

Timestamp:

08/10/07 20:28:49 (1 year ago)

Author:

Damien Fouilleul <damienf@videolan.org>

git-committer:

Damien Fouilleul <damienf@videolan.org> 1186770529 +0000

git-parent:

[43bd7bed43a7b997985d2ed18b32d156a25164b6]

git-author:

Damien Fouilleul <damienf@videolan.org> 1186770529 +0000

Message:

i422_yuy2: SSE2 improvements

Files:

• configure.ac (modified) (3 diffs)
• modules/video_chroma/Modules.am (modified) (1 diff)
• modules/video_chroma/i422_yuy2.c (modified) (10 diffs)
• modules/video_chroma/i422_yuy2.h (modified) (4 diffs)

configure.ac

@@ -1258 +1258 @@
-THREEDNOW_MODULES="memcpy3dn"
-SSE_MODULES=""
-
+ i422_yuy2_sse2
-ALTIVEC_MODULES="memcpyaltivec i420_yuy2_altivec"
-#ALTIVEC_MODULES="${ALTIVEC_MODULES} idctaltivec motionaltivec"
@@ -1284 +1284 @@
-if test "${ac_cv_c_mmx_intrinsics}" != "no"; then
-  AC_DEFINE(HAVE_MMX_INTRINSICS, 1, Define if MMX intrinsics are available.)
-i420_rgb_mmx
+${MMX_MODULES}
-fi
-
@@ -1309 +1309 @@
-if test "${ac_cv_c_sse2_intrinsics}" != "no"; then
-  AC_DEFINE(HAVE_SSE2_INTRINSICS, 1, Define if SSE2 intrinsics are available.)
-i420_rgb_sse2
+${SSE2_MODULES}
-fi
-

modules/video_chroma/Modules.am

@@ -51 +51 @@
-    $(NULL)
-
+SOURCES_i422_yuy2_sse2 = \
+    i422_yuy2.c \
+    i422_yuy2.h \
+    $(NULL)
+
-SOURCES_i420_ymga = \
-    i420_ymga.c \

modules/video_chroma/i422_yuy2.c

@@ -68 +68 @@
-    set_capability( "chroma", 100 );
-    add_requirement( MMX );
+#elif defined (MODULE_NAME_IS_i422_yuy2_sse2)
+    set_description( _("SSE2 conversions from " SRC_FOURCC " to " DEST_FOURCC) );
+    set_capability( "chroma", 120 );
+    add_requirement( MMX );
-#endif
-    set_callbacks( Activate, NULL );
@@ -144 +148 @@
-                                              picture_t *p_dest )
-{
-
-
+
-    uint8_t *p_y = p_source->Y_PIXELS;
-    uint8_t *p_u = p_source->U_PIXELS;
@@ -152 +155 @@
-    int i_x, i_y;
-
+    const int i_source_margin = p_source->p[0].i_pitch
+                                 - p_source->p[0].i_visible_pitch;
+    const int i_source_margin_c = p_source->p[1].i_pitch
+                                 - p_source->p[1].i_visible_pitch;
+    const int i_dest_margin = p_dest->p->i_pitch
+                               - p_dest->p->i_visible_pitch;
+
+#if defined (MODULE_NAME_IS_i422_yuy2_sse2)
+
+    if( 0 == (15 & (p_source->p[Y_PLANE].i_pitch|p_dest->p->i_pitch|
+        ((int)p_line|(int)p_y))) )
+    {
+        /* use faster SSE2 aligned fetch and store */
+        for( i_y = p_vout->render.i_height ; i_y-- ; )
+        {
+            for( i_x = p_vout->render.i_width / 16 ; i_x-- ; )
+            {
+                SSE2_CALL( SSE2_YUV422_YUYV_ALIGNED );
+            }
+            for( i_x = ( p_vout->render.i_width % 16 ) / 2; i_x-- ; )
+            {
+                C_YUV422_YUYV( p_line, p_y, p_u, p_v );
+            }
+            p_y += i_source_margin;
+            p_u += i_source_margin_c;
+            p_v += i_source_margin_c;
+            p_line += i_dest_margin;
+        }
+    }
+    else {
+        /* use slower SSE2 unaligned fetch and store */
+        for( i_y = p_vout->render.i_height ; i_y-- ; )
+        {
+            for( i_x = p_vout->render.i_width / 16 ; i_x-- ; )
+            {
+                SSE2_CALL( SSE2_YUV422_YUYV_UNALIGNED );
+            }
+            for( i_x = ( p_vout->render.i_width % 16 ) / 2; i_x-- ; )
+            {
+                C_YUV422_YUYV( p_line, p_y, p_u, p_v );
+            }
+            p_y += i_source_margin;
+            p_u += i_source_margin_c;
+            p_v += i_source_margin_c;
+            p_line += i_dest_margin;
+        }
+    }
+    SSE2_END;
+
+#else
+
-    for( i_y = p_vout->render.i_height ; i_y-- ; )
-    {
-        uint8_t *p_line = p_pixels;
-        for( i_x = p_vout->render.i_width / 8 ; i_x-- ; )
-        {
@@ -166 +219 @@
-#endif
-        }
-
+
+
+
+
+
+
+
+
-    }
-#if defined (MODULE_NAME_IS_i422_yuy2_mmx)
-    MMX_END;
-lif defined (MODULE_NAME_IS_i422_yuy2_sse2)
-
+ndif
+
-#endif
-}
@@ -181 +241 @@
-                                              picture_t *p_dest )
-{
-
-
+
-    uint8_t *p_y = p_source->Y_PIXELS;
-    uint8_t *p_u = p_source->U_PIXELS;
@@ -189 +248 @@
-    int i_x, i_y;
-
+    const int i_source_margin = p_source->p[0].i_pitch
+                                 - p_source->p[0].i_visible_pitch;
+    const int i_source_margin_c = p_source->p[1].i_pitch
+                                 - p_source->p[1].i_visible_pitch;
+    const int i_dest_margin = p_dest->p->i_pitch
+                               - p_dest->p->i_visible_pitch;
+
+#if defined (MODULE_NAME_IS_i422_yuy2_sse2)
+
+    if( 0 == (15 & (p_source->p[Y_PLANE].i_pitch|p_dest->p->i_pitch|
+        ((int)p_line|(int)p_y))) )
+    {
+        /* use faster SSE2 aligned fetch and store */
+        for( i_y = p_vout->render.i_height ; i_y-- ; )
+        {
+            for( i_x = p_vout->render.i_width / 16 ; i_x-- ; )
+            {
+                SSE2_CALL( SSE2_YUV422_YVYU_ALIGNED );
+            }
+            for( i_x = ( p_vout->render.i_width % 16 ) / 2; i_x-- ; )
+            {
+                C_YUV422_YVYU( p_line, p_y, p_u, p_v );
+            }
+            p_y += i_source_margin;
+            p_u += i_source_margin_c;
+            p_v += i_source_margin_c;
+            p_line += i_dest_margin;
+        }
+    }
+    else {
+        /* use slower SSE2 unaligned fetch and store */
+        for( i_y = p_vout->render.i_height ; i_y-- ; )
+        {
+            for( i_x = p_vout->render.i_width / 16 ; i_x-- ; )
+            {
+                SSE2_CALL( SSE2_YUV422_YVYU_UNALIGNED );
+            }
+            for( i_x = ( p_vout->render.i_width % 16 ) / 2; i_x-- ; )
+            {
+                C_YUV422_YVYU( p_line, p_y, p_u, p_v );
+            }
+            p_y += i_source_margin;
+            p_u += i_source_margin_c;
+            p_v += i_source_margin_c;
+            p_line += i_dest_margin;
+        }
+    }
+    SSE2_END;
+
+#else
+
-    for( i_y = p_vout->render.i_height ; i_y-- ; )
-    {
-        uint8_t *p_line = p_pixels;
-        for( i_x = p_vout->render.i_width / 8 ; i_x-- ; )
-        {
@@ -203 +312 @@
-#endif
-        }
-
+
+
+
+
+
+
+
+
-    }
-#if defined (MODULE_NAME_IS_i422_yuy2_mmx)
-    MMX_END;
-lif defined (MODULE_NAME_IS_i422_yuy2_sse2)
-
+ndif
+
-#endif
-}
@@ -218 +334 @@
-                                              picture_t *p_dest )
-{
-
-
+
-    uint8_t *p_y = p_source->Y_PIXELS;
-    uint8_t *p_u = p_source->U_PIXELS;
@@ -226 +341 @@
-    int i_x, i_y;
-
+    const int i_source_margin = p_source->p[0].i_pitch
+                                 - p_source->p[0].i_visible_pitch;
+    const int i_source_margin_c = p_source->p[1].i_pitch
+                                 - p_source->p[1].i_visible_pitch;
+    const int i_dest_margin = p_dest->p->i_pitch
+                               - p_dest->p->i_visible_pitch;
+
+#if defined (MODULE_NAME_IS_i422_yuy2_sse2)
+
+    if( 0 == (15 & (p_source->p[Y_PLANE].i_pitch|p_dest->p->i_pitch|
+        ((int)p_line|(int)p_y))) )
+    {
+        /* use faster SSE2 aligned fetch and store */
+        for( i_y = p_vout->render.i_height ; i_y-- ; )
+        {
+            for( i_x = p_vout->render.i_width / 16 ; i_x-- ; )
+            {
+                SSE2_CALL( SSE2_YUV422_UYVY_ALIGNED );
+            }
+            for( i_x = ( p_vout->render.i_width % 16 ) / 2; i_x-- ; )
+            {
+                C_YUV422_UYVY( p_line, p_y, p_u, p_v );
+            }
+            p_y += i_source_margin;
+            p_u += i_source_margin_c;
+            p_v += i_source_margin_c;
+            p_line += i_dest_margin;
+        }
+    }
+    else {
+        /* use slower SSE2 unaligned fetch and store */
+        for( i_y = p_vout->render.i_height ; i_y-- ; )
+        {
+            for( i_x = p_vout->render.i_width / 16 ; i_x-- ; )
+            {
+                SSE2_CALL( SSE2_YUV422_UYVY_UNALIGNED );
+            }
+            for( i_x = ( p_vout->render.i_width % 16 ) / 2; i_x-- ; )
+            {
+                C_YUV422_UYVY( p_line, p_y, p_u, p_v );
+            }
+            p_y += i_source_margin;
+            p_u += i_source_margin_c;
+            p_v += i_source_margin_c;
+            p_line += i_dest_margin;
+        }
+    }
+    SSE2_END;
+
+#else
+
-    for( i_y = p_vout->render.i_height ; i_y-- ; )
-    {
-        uint8_t *p_line = p_pixels;
-        for( i_x = p_vout->render.i_width / 8 ; i_x-- ; )
-        {
@@ -240 +405 @@
-#endif
-        }
-
+
+
+
+
+
+
+
+
-    }
-#if defined (MODULE_NAME_IS_i422_yuy2_mmx)
-    MMX_END;
-lif defined (MODULE_NAME_IS_i422_yuy2_sse2)
-
+ndif
+
-#endif
-}

modules/video_chroma/i422_yuy2.h

@@ -88 +88 @@
-#include <mmintrin.h>
-
+#define MMX_CALL(MMX_INSTRUCTIONS)  \
+    do {                            \
+        __m64 mm0, mm1, mm2;        \
+        MMX_INSTRUCTIONS            \
+        p_line += 16; p_y += 8;     \
+        p_u += 4; p_v += 4;         \
+    } while(0)
+
-#define MMX_END _mm_empty()
+
+#define MMX_YUV422_YUYV                     \
+    mm0 = (__m64)*(uint64_t*)p_y;           \
+    mm1 = _mm_cvtsi32_si64(*(int*)p_u);     \
+    mm2 = _mm_cvtsi32_si64(*(int*)p_v);     \
+    mm1 = _mm_unpacklo_pi8(mm1, mm2);       \
+    mm2 = mm0;                              \
+    mm2 = _mm_unpacklo_pi8(mm2, mm1);       \
+    *(uint64_t*)p_line = (uint64_t)mm2;     \
+    mm0 = _mm_unpackhi_pi8(mm0, mm1);       \
+    *(uint64_t*)(p_line+8) = (uint64_t)mm0;
+    
+#define MMX_YUV422_YVYU                     \
+    mm0 = (__m64)*(uint64_t*)p_y;           \
+    mm2 = _mm_cvtsi32_si64(*(int*)p_u);     \
+    mm1 = _mm_cvtsi32_si64(*(int*)p_v);     \
+    mm1 = _mm_unpacklo_pi8(mm1, mm2);       \
+    mm2 = mm0;                              \
+    mm2 = _mm_unpacklo_pi8(mm2, mm1);       \
+    *(uint64_t*)p_line = (uint64_t)mm2;     \
+    mm0 = _mm_unpackhi_pi8(mm0, mm1);       \
+    *(uint64_t*)(p_line+8) = (uint64_t)mm0;
+
+#define MMX_YUV422_UYVY                     \
+    mm0 = (__m64)*(uint64_t*)p_y;           \
+    mm1 = _mm_cvtsi32_si64(*(int*)p_u);     \
+    mm2 = _mm_cvtsi32_si64(*(int*)p_v);     \
+    mm1 = _mm_unpacklo_pi8(mm1, mm2);       \
+    mm2 = mm1;                              \
+    mm2 = _mm_unpacklo_pi8(mm2, mm0);       \
+    *(uint64_t*)p_line = (uint64_t)mm2;     \
+    mm1 = _mm_unpackhi_pi8(mm1, mm0);       \
+    *(uint64_t*)(p_line+8) = (uint64_t)mm1;
-
-#endif
@@ -98 +139 @@
-/* SSE2 assembly */
-
+#define SSE2_CALL(MMX_INSTRUCTIONS)         \
+    do {                                    \
+    __asm__ __volatile__(                   \
+        ".p2align 3 \n\t"                   \
+        MMX_INSTRUCTIONS                    \
+        :                                   \
+        : "r" (p_line), "r" (p_y),          \
+          "r" (p_u), "r" (p_v) );           \
+        p_line += 32; p_y += 16;            \
+        p_u += 8; p_v += 8;                 \
+    } while(0)
+
-#define SSE2_END  __asm__ __volatile__ ( "sfence" ::: "memory" )
-
+#define SSE2_YUV422_YUYV_ALIGNED "                                        \n\
+movdqa      (%1), %%xmm0  # Load 8 Y            y7 y6 y5 y4 y3 y2 y1 y0   \n\
+movq        (%2), %%xmm1  # Load 4 Cb           00 00 00 00 u3 u2 u1 u0   \n\
+movq        (%3), %%xmm2  # Load 4 Cr           00 00 00 00 v3 v2 v1 v0   \n\
+punpcklbw %%xmm2, %%xmm1  #                     v3 u3 v2 u2 v1 u1 v0 u0   \n\
+movdqa    %%xmm0, %%xmm2  #                     y7 y6 y5 y4 y3 y2 y1 y0   \n\
+punpcklbw %%xmm1, %%xmm2  #                     v1 y3 u1 y2 v0 y1 u0 y0   \n\
+movntdq   %%xmm2, (%0)    # Store low YUYV                                \n\
+punpckhbw %%xmm1, %%xmm0  #                     v3 y7 u3 y6 v2 y5 u2 y4   \n\
+movntdq   %%xmm0, 16(%0)  # Store high YUYV                               \n\
+"
+
+#define SSE2_YUV422_YUYV_UNALIGNED "                                      \n\
+movdqu      (%1), %%xmm0  # Load 8 Y            y7 y6 y5 y4 y3 y2 y1 y0   \n\
+movq        (%2), %%xmm1  # Load 4 Cb           00 00 00 00 u3 u2 u1 u0   \n\
+movq        (%3), %%xmm2  # Load 4 Cr           00 00 00 00 v3 v2 v1 v0   \n\
+prefetchnta (%0)          # Tell CPU not to cache output YUYV data        \n\
+punpcklbw %%xmm2, %%xmm1  #                     v3 u3 v2 u2 v1 u1 v0 u0   \n\
+movdqa    %%xmm0, %%xmm2  #                     y7 y6 y5 y4 y3 y2 y1 y0   \n\
+punpcklbw %%xmm1, %%xmm2  #                     v1 y3 u1 y2 v0 y1 u0 y0   \n\
+movdqu    %%xmm2, (%0)    # Store low YUYV                                \n\
+punpckhbw %%xmm1, %%xmm0  #                     v3 y7 u3 y6 v2 y5 u2 y4   \n\
+movdqu    %%xmm0, 16(%0)  # Store high YUYV                               \n\
+"
+
+#define SSE2_YUV422_YVYU_ALIGNED "                                        \n\
+movdqa      (%1), %%xmm0  # Load 8 Y            y7 y6 y5 y4 y3 y2 y1 y0   \n\
+movq        (%2), %%xmm2  # Load 4 Cb           00 00 00 00 u3 u2 u1 u0   \n\
+movq        (%3), %%xmm1  # Load 4 Cr           00 00 00 00 v3 v2 v1 v0   \n\
+punpcklbw %%xmm2, %%xmm1  #                     u3 v3 u2 v2 u1 v1 u0 v0   \n\
+movdqa    %%xmm0, %%xmm2  #                     y7 y6 y5 y4 y3 y2 y1 y0   \n\
+punpcklbw %%xmm1, %%xmm2  #                     u1 y3 v1 y2 u0 y1 v0 y0   \n\
+movntdq   %%xmm2, (%0)    # Store low YUYV                                \n\
+punpckhbw %%xmm1, %%xmm0  #                     u3 y7 v3 y6 u2 y5 v2 y4   \n\
+movntdq   %%xmm0, 16(%0)  # Store high YUYV                               \n\
+"
+
+#define SSE2_YUV422_YVYU_UNALIGNED "                                      \n\
+movdqu      (%1), %%xmm0  # Load 8 Y            y7 y6 y5 y4 y3 y2 y1 y0   \n\
+movq        (%2), %%xmm2  # Load 4 Cb           00 00 00 00 u3 u2 u1 u0   \n\
+movq        (%3), %%xmm1  # Load 4 Cr           00 00 00 00 v3 v2 v1 v0   \n\
+prefetchnta (%0)          # Tell CPU not to cache output YUYV data        \n\
+punpcklbw %%xmm2, %%xmm1  #                     u3 v3 u2 v2 u1 v1 u0 v0   \n\
+movdqa    %%xmm0, %%xmm2  #                     y7 y6 y5 y4 y3 y2 y1 y0   \n\
+punpcklbw %%xmm1, %%xmm2  #                     u1 y3 v1 y2 u0 y1 v0 y0   \n\
+movdqu    %%xmm2, (%0)    # Store low YUYV                                \n\
+punpckhbw %%xmm1, %%xmm0  #                     u3 y7 v3 y6 u2 y5 v2 y4   \n\
+movdqu    %%xmm0, 16(%0)  # Store high YUYV                               \n\
+"
+
+#define SSE2_YUV422_UYVY_ALIGNED "                                        \n\
+movdqa      (%1), %%xmm0  # Load 8 Y            y7 y6 y5 y4 y3 y2 y1 y0   \n\
+movq        (%2), %%xmm1  # Load 4 Cb           00 00 00 00 u3 u2 u1 u0   \n\
+movq        (%3), %%xmm2  # Load 4 Cr           00 00 00 00 v3 v2 v1 v0   \n\
+punpcklbw %%xmm2, %%xmm1  #                     v3 u3 v2 u2 v1 u1 v0 u0   \n\
+movdqa    %%xmm1, %%xmm2  #                     v3 u3 v2 u2 v1 u1 v0 u0   \n\
+punpcklbw %%xmm0, %%xmm2  #                     y3 v1 y2 u1 y1 v0 y0 u0   \n\
+movntdq   %%xmm2, (%0)    # Store low UYVY                                \n\
+punpckhbw %%xmm0, %%xmm1  #                     y7 v3 y6 u3 y5 v2 y4 u2   \n\
+movntdq   %%xmm1, 16(%0)  # Store high UYVY                               \n\
+"
+
+#define SSE2_YUV422_UYVY_UNALIGNED "                                      \n\
+movdqu      (%1), %%xmm0  # Load 8 Y            y7 y6 y5 y4 y3 y2 y1 y0   \n\
+movq        (%2), %%xmm1  # Load 4 Cb           00 00 00 00 u3 u2 u1 u0   \n\
+movq        (%3), %%xmm2  # Load 4 Cr           00 00 00 00 v3 v2 v1 v0   \n\
+prefetchnta (%0)          # Tell CPU not to cache output YUYV data        \n\
+punpcklbw %%xmm2, %%xmm1  #                     v3 u3 v2 u2 v1 u1 v0 u0   \n\
+movdqa    %%xmm1, %%xmm2  #                     v3 u3 v2 u2 v1 u1 v0 u0   \n\
+punpcklbw %%xmm0, %%xmm2  #                     y3 v1 y2 u1 y1 v0 y0 u0   \n\
+movdqu    %%xmm2, (%0)    # Store low UYVY                                \n\
+punpckhbw %%xmm0, %%xmm1  #                     y7 v3 y6 u3 y5 v2 y4 u2   \n\
+movdqu    %%xmm1, 16(%0)  # Store high UYVY                               \n\
+"
+
-#elif defined(HAVE_SSE2_INTRINSICS)
-
@@ -111 +239 @@
-#endif
-
-lif defined (MODULE_NAME_IS_i422_yuy2)
+ndif
-
-#define C_YUV422_YUYV( p_line, p_y, p_u, p_v )                              \
@@ -137 +265 @@
-    *(p_line)++ = *(p_v) - 0x80; p_v += 2;                                  \
-
-
-
+