Changeset bae04ee86f451d93afa221ae95a51d0bb424b634

Timestamp:

31/05/07 11:56:25 (1 year ago)

Author:

Damien Fouilleul <damienf@videolan.org>

git-committer:

Damien Fouilleul <damienf@videolan.org> 1180605385 +0000

git-parent:

[9af412916f0472b37e19194b0c23db9638678d77]

git-author:

Damien Fouilleul <damienf@videolan.org> 1180605385 +0000

Message:

- video_chroma: added support for IA-32/64 SSE2 accelaration (128 bit vector integer registers), needs LOADS of testing

Files:

• configure.ac (modified) (5 diffs)
• modules/video_chroma/Modules.am (modified) (1 diff)
• modules/video_chroma/i420_yuy2.c (modified) (17 diffs)
• modules/video_chroma/i420_yuy2.h (modified) (4 diffs)

configure.ac

@@ -1270 +1270 @@
-THREEDNOW_MODULES="memcpy3dn"
-SSE_MODULES=""
+SSE2_MODULES=""
-ALTIVEC_MODULES="memcpyaltivec i420_yuy2_altivec"
-#ALTIVEC_MODULES="${ALTIVEC_MODULES} idctaltivec motionaltivec"
@@ -1276 +1277 @@
-then
-  MMX_MODULES="${MMX_MODULES} i420_yuy2_mmx"
-
-
-
-
-
-
-
-
-
-
+
-fi
-
@@ -1313 +1305 @@
-fi
-
+dnl  Check for fully workin SSE2 intrinsics
+dnl  We need support for -mmmx, we need <emmintrin.h>, and we also need a
+dnl  working compiler (http://gcc.gnu.org/bugzilla/show_bug.cgi?id=23963)
+AC_CACHE_CHECK([if \$CC groks SSE2 intrinsics],
+    [ac_cv_c_sse2_intrinsics],
+    [CFLAGS="${CFLAGS_save} -O -msse2"
+     AC_TRY_COMPILE([#include <emmintrin.h>
+                     #include <stdint.h>
+                     uint64_t frobzor;],
+                    [__m128i a, b, c;
+                     a = b = c = _mm_set1_epi64((__m64)frobzor);
+                     a = _mm_slli_epi16(a, 3);
+                     a = _mm_adds_epi16(a, b);
+                     c = _mm_srli_epi16(c, 8);
+                     c = _mm_slli_epi16(c, 3);
+                     b = _mm_adds_epi16(b, c);
+                     a = _mm_unpacklo_epi8(a, b);
+                     frobzor = (uint64_t)_mm_movepi64_pi64(a);],
+                    [ac_cv_c_sse2_intrinsics=yes],
+                    [ac_cv_c_sse2_intrinsics=no])])
+if test "${ac_cv_c_sse2_intrinsics}" != "no"; then
+  AC_DEFINE(HAVE_SSE2_INTRINSICS, 1, Define if SSE2 intrinsics are available.)
+  dnl VLC_ADD_CFLAGS([i420_rgb_sse2],[-msse2])
+fi
+
+AC_CACHE_CHECK([if \$CC groks MMX inline assembly],
+    [ac_cv_mmx_inline],
+    [CFLAGS="${CFLAGS_save}"
+     AC_TRY_COMPILE(,[void *p;asm volatile("packuswb %%mm1,%%mm2"::"r"(p));],
+                    ac_cv_mmx_inline=yes, ac_cv_mmx_inline=no)])
+if test "${ac_cv_mmx_inline}" != "no"; then
+  AC_DEFINE(CAN_COMPILE_MMX, 1, Define if \$CC groks MMX inline assembly.)
+  ACCEL_MODULES="${ACCEL_MODULES} ${MMX_MODULES}"
+fi
+
-AC_CACHE_CHECK([if \$CC groks MMX EXT inline assembly],
-    [ac_cv_mmxext_inline],
@@ -1341 +1368 @@
-  AC_DEFINE(CAN_COMPILE_SSE, 1, Define if \$CC groks SSE inline assembly.)
-  ACCEL_MODULES="${ACCEL_MODULES} ${SSE_MODULES}"
+fi
+
+AC_CACHE_CHECK([if \$CC groks SSE2 inline assembly],
+    [ac_cv_sse2_inline],
+    [CFLAGS="${CFLAGS_save}"
+     AC_TRY_COMPILE(,[void *p;asm volatile("punpckhqdq %%xmm1,%%xmm2"::"r"(p));],
+                    ac_cv_sse2_inline=yes, ac_cv_sse2_inline=no)])
+if test "${ac_cv_sse2_inline}" != "no" -a "${SYS}" != "solaris"; then
+  AC_DEFINE(CAN_COMPILE_SSE2, 1, Define if \$CC groks SSE2 inline assembly.)
+  ACCEL_MODULES="${ACCEL_MODULES} ${SSE2_MODULES}"
-fi
-
@@ -1493 +1530 @@
-then
-    ARCH="${ARCH} mmx"
+    VLC_ADD_BUILTINS([${ACCEL_MODULES}])
+fi
+if test "${host_cpu}" = "i686" -o "${host_cpu}" = "x86_64"
+then
+    ARCH="${ARCH} sse sse2"
-    VLC_ADD_BUILTINS([${ACCEL_MODULES}])
-fi

modules/video_chroma/Modules.am

@@ -20 +20 @@
-
-SOURCES_i420_yuy2_mmx = \
+    i420_yuy2.c \
+    i420_yuy2.h \
+    $(NULL)
+
+SOURCES_i420_yuy2_sse2 = \
-    i420_yuy2.c \
-    i420_yuy2.h \

modules/video_chroma/i420_yuy2.c

@@ -6 +6 @@
- *
- * Authors: Samuel Hocevar <sam@zoy.org>
+ *          Damien Fouilleul <damien@videolan.org>
- *
- * This program is free software; you can redistribute it and/or modify
@@ -43 +44 @@
-#elif defined (MODULE_NAME_IS_i420_yuy2_mmx)
-#    define DEST_FOURCC "YUY2,YUNV,YVYU,UYVY,UYNV,Y422,IUYV,cyuv"
+#elif defined (MODULE_NAME_IS_i420_yuy2_sse2)
+#    define DEST_FOURCC "YUY2,YUNV,YVYU,UYVY,UYNV,Y422,IUYV,cyuv"
-#elif defined (MODULE_NAME_IS_i420_yuy2_altivec)
-#    define DEST_FOURCC "YUY2,YUNV,YVYU,UYVY,UYNV,Y422"
@@ -64 +67 @@
-
-#ifdef MODULE_NAME_IS_i420_yuy2_mmx
-
-
+
+
+
-#endif
-
@@ -79 +83 @@
-    set_capability( "chroma", 100 );
-    add_requirement( MMX );
-
-
-
+
+
+
+
-#elif defined (MODULE_NAME_IS_i420_yuy2_altivec)
-    set_description(
@@ -126 +131 @@
-                    p_vout->chroma.pf_convert = I420_UYVY;
-                    break;
-
-#if !defined (MODULE_NAME_IS_i420_yuy2_altivec)
-                case VLC_FOURCC('I','U','Y','V'):
@@ -257 +261 @@
-                               - p_dest->p->i_visible_pitch;
-
+#if !defined(MODULE_NAME_IS_i420_yuy2_sse2)
-    for( i_y = p_vout->render.i_height / 2 ; i_y-- ; )
-    {
@@ -266 +271 @@
-
-#if !defined (MODULE_NAME_IS_i420_yuy2_mmx)
-
-
+
+
+
+
+
-            C_YUV420_YUYV( );
-        }
@@ -275 +283 @@
-            MMX_CALL( MMX_YUV420_YUYV );
-        }
+#endif
-        for( i_x = ( p_vout->render.i_width % 8 ) / 2; i_x-- ; )
-        {
-            C_YUV420_YUYV( );
-        }
-#endif
-
-        p_y1 += i_source_margin;
@@ -289 +297 @@
-    }
-
+#if defined (MODULE_NAME_IS_i420_yuy2_mmx)
+    __asm__ __volatile__("emms" :: );
+#endif
+
-#if defined (MODULE_NAME_IS_i420_yuy2_altivec)
-    }
-#endif
+
+#else // defined(MODULE_NAME_IS_i420_yuy2_sse2)
+    /*
+    ** SSE2 128 bytes fetch/store instructions are faster 
+    ** if memory access is 16 bytes aligned
+    */
+    if( 0 == (15 & (p_source->p[Y_PLANE].i_pitch|p_dest->p->i_pitch|
+        ((int)p_line2|(int)p_y2))) )
+    {
+        /* use faster SSE2 aligned fetch and store */
+        for( i_y = p_vout->render.i_height / 2 ; i_y-- ; )
+        {
+            p_line1 = p_line2;
+            p_line2 += p_dest->p->i_pitch;
+
+            p_y1 = p_y2;
+            p_y2 += p_source->p[Y_PLANE].i_pitch;
+
+            for( i_x = p_vout->render.i_width / 16 ; i_x-- ; )
+            {
+                SSE2_CALL( SSE2_YUV420_YUYV_ALIGNED );
+            }
+            for( i_x = ( p_vout->render.i_width % 16 ) / 2; i_x-- ; )
+            {
+                C_YUV420_YUYV( );
+            }
+
+            p_y1 += i_source_margin;
+            p_y2 += i_source_margin;
+            p_u += i_source_margin_c;
+            p_v += i_source_margin_c;
+            p_line1 += i_dest_margin;
+            p_line2 += i_dest_margin;
+        }
+    }
+    else
+    {
+        /* use slower SSE2 unaligned fetch and store */
+        for( i_y = p_vout->render.i_height / 2 ; i_y-- ; )
+        {
+            p_line1 = p_line2;
+            p_line2 += p_dest->p->i_pitch;
+
+            p_y1 = p_y2;
+            p_y2 += p_source->p[Y_PLANE].i_pitch;
+
+            for( i_x = p_vout->render.i_width / 16 ; i_x-- ; )
+            {
+                SSE2_CALL( SSE2_YUV420_YUYV_UNALIGNED );
+            }
+            for( i_x = ( p_vout->render.i_width % 16 ) / 2; i_x-- ; )
+            {
+                C_YUV420_YUYV( );
+            }
+
+            p_y1 += i_source_margin;
+            p_y2 += i_source_margin;
+            p_u += i_source_margin_c;
+            p_v += i_source_margin_c;
+            p_line1 += i_dest_margin;
+            p_line2 += i_dest_margin;
+        }
+    }
+#endif // defined(MODULE_NAME_IS_i420_yuy2_sse2)
-}
-
@@ -394 +470 @@
-                               - p_dest->p->i_visible_pitch;
-
+#if !defined(MODULE_NAME_IS_i420_yuy2_sse2)
-    for( i_y = p_vout->render.i_height / 2 ; i_y-- ; )
-    {
@@ -421 +498 @@
-        p_line2 += i_dest_margin;
-    }
+
+#if defined (MODULE_NAME_IS_i420_yuy2_mmx)
+    __asm__ __volatile__("emms" :: );
+#endif
+
-#if defined (MODULE_NAME_IS_i420_yuy2_altivec)
-    }
-#endif
+
+#else // defined(MODULE_NAME_IS_i420_yuy2_sse2)
+    /*
+    ** SSE2 128 bytes fetch/store instructions are faster 
+    ** if memory access is 16 bytes aligned
+    */
+    if( 0 == (15 & (p_source->p[Y_PLANE].i_pitch|p_dest->p->i_pitch|
+        ((int)p_line2|(int)p_y2))) )
+    {
+        /* use faster SSE2 aligned fetch and store */
+        for( i_y = p_vout->render.i_height / 2 ; i_y-- ; )
+        {
+            p_line1 = p_line2;
+            p_line2 += p_dest->p->i_pitch;
+
+            p_y1 = p_y2;
+            p_y2 += p_source->p[Y_PLANE].i_pitch;
+
+            for( i_x = p_vout->render.i_width / 16 ; i_x-- ; )
+            {
+                SSE2_CALL( SSE2_YUV420_YVYU_ALIGNED );
+            }
+            for( i_x = ( p_vout->render.i_width % 16 ) / 2; i_x-- ; )
+            {
+                C_YUV420_YVYU( );
+            }
+
+            p_y1 += i_source_margin;
+            p_y2 += i_source_margin;
+            p_u += i_source_margin_c;
+            p_v += i_source_margin_c;
+            p_line1 += i_dest_margin;
+            p_line2 += i_dest_margin;
+        }
+    }
+    else
+    {
+        /* use slower SSE2 unaligned fetch and store */
+        for( i_y = p_vout->render.i_height / 2 ; i_y-- ; )
+        {
+            p_line1 = p_line2;
+            p_line2 += p_dest->p->i_pitch;
+
+            p_y1 = p_y2;
+            p_y2 += p_source->p[Y_PLANE].i_pitch;
+
+            for( i_x = p_vout->render.i_width / 16 ; i_x-- ; )
+            {
+                SSE2_CALL( SSE2_YUV420_YVYU_UNALIGNED );
+            }
+            for( i_x = ( p_vout->render.i_width % 16 ) / 2; i_x-- ; )
+            {
+                C_YUV420_YVYU( );
+            }
+
+            p_y1 += i_source_margin;
+            p_y2 += i_source_margin;
+            p_u += i_source_margin_c;
+            p_v += i_source_margin_c;
+            p_line1 += i_dest_margin;
+            p_line2 += i_dest_margin;
+        }
+    }
+#endif // defined(MODULE_NAME_IS_i420_yuy2_sse2)
-}
-
@@ -526 +672 @@
-                               - p_dest->p->i_visible_pitch;
-
+#if !defined(MODULE_NAME_IS_i420_yuy2_sse2)
-    for( i_y = p_vout->render.i_height / 2 ; i_y-- ; )
-    {
@@ -565 +712 @@
-    }
-#endif
+
+#else // defined(MODULE_NAME_IS_i420_yuy2_sse2)
+    /*
+    ** SSE2 128 bytes fetch/store instructions are faster 
+    ** if memory access is 16 bytes aligned
+    */
+    if( 0 == (15 & (p_source->p[Y_PLANE].i_pitch|p_dest->p->i_pitch|
+        ((int)p_line2|(int)p_y2))) )
+    {
+        /* use faster SSE2 aligned fetch and store */
+        for( i_y = p_vout->render.i_height / 2 ; i_y-- ; )
+        {
+            p_line1 = p_line2;
+            p_line2 += p_dest->p->i_pitch;
+
+            p_y1 = p_y2;
+            p_y2 += p_source->p[Y_PLANE].i_pitch;
+
+            for( i_x = p_vout->render.i_width / 16 ; i_x-- ; )
+            {
+                SSE2_CALL( SSE2_YUV420_UYVY_ALIGNED );
+            }
+            for( i_x = ( p_vout->render.i_width % 16 ) / 2; i_x-- ; )
+            {
+                C_YUV420_UYVY( );
+            }
+
+            p_y1 += i_source_margin;
+            p_y2 += i_source_margin;
+            p_u += i_source_margin_c;
+            p_v += i_source_margin_c;
+            p_line1 += i_dest_margin;
+            p_line2 += i_dest_margin;
+        }
+    }
+    else
+    {
+        /* use slower SSE2 unaligned fetch and store */
+        for( i_y = p_vout->render.i_height / 2 ; i_y-- ; )
+        {
+            p_line1 = p_line2;
+            p_line2 += p_dest->p->i_pitch;
+
+            p_y1 = p_y2;
+            p_y2 += p_source->p[Y_PLANE].i_pitch;
+
+            for( i_x = p_vout->render.i_width / 16 ; i_x-- ; )
+            {
+                SSE2_CALL( SSE2_YUV420_UYVY_UNALIGNED );
+            }
+            for( i_x = ( p_vout->render.i_width % 16 ) / 2; i_x-- ; )
+            {
+                C_YUV420_UYVY( );
+            }
+
+            p_y1 += i_source_margin;
+            p_y2 += i_source_margin;
+            p_u += i_source_margin_c;
+            p_v += i_source_margin_c;
+            p_line1 += i_dest_margin;
+            p_line2 += i_dest_margin;
+        }
+    }
+#endif // defined(MODULE_NAME_IS_i420_yuy2_sse2)
-}
-
@@ -602 +813 @@
-                               - p_dest->p->i_visible_pitch;
-
+#if !defined(MODULE_NAME_IS_i420_yuy2_sse2)
-    for( i_y = p_vout->render.i_height / 2 ; i_y-- ; )
-    {
@@ -612 +824 @@
-        for( i_x = p_vout->render.i_width / 8 ; i_x-- ; )
-        {
-defined (MODULE_NAME_IS_i420_yuy2
+!defined (MODULE_NAME_IS_i420_yuy2_mmx
-            C_YUV420_UYVY( );
-            C_YUV420_UYVY( );
@@ -629 +841 @@
-        p_line2 += i_dest_margin;
-    }
+
+#if defined (MODULE_NAME_IS_i420_yuy2_mmx)
+    __asm__ __volatile__("emms" :: );
+#endif
+
+#else // defined(MODULE_NAME_IS_i420_yuy2_sse2)
+    /*
+    ** SSE2 128 bytes fetch/store instructions are faster 
+    ** if memory access is 16 bytes aligned
+    */
+    if( 0 == (15 & (p_source->p[Y_PLANE].i_pitch|p_dest->p->i_pitch|
+        ((int)p_line2|(int)p_y2))) )
+    {
+        /* use faster SSE2 aligned fetch and store */
+        for( i_y = p_vout->render.i_height / 2 ; i_y-- ; )
+        {
+            p_line1 = p_line2;
+            p_line2 += p_dest->p->i_pitch;
+
+            p_y1 = p_y2;
+            p_y2 += p_source->p[Y_PLANE].i_pitch;
+
+            for( i_x = p_vout->render.i_width / 16 ; i_x-- ; )
+            {
+                SSE2_CALL( SSE2_YUV420_UYVY_ALIGNED );
+            }
+            for( i_x = ( p_vout->render.i_width % 16 ) / 2; i_x-- ; )
+            {
+                C_YUV420_UYVY( );
+            }
+
+            p_y1 += i_source_margin;
+            p_y2 += i_source_margin;
+            p_u += i_source_margin_c;
+            p_v += i_source_margin_c;
+            p_line1 += i_dest_margin;
+            p_line2 += i_dest_margin;
+        }
+    }
+    else
+    {
+        /* use slower SSE2 unaligned fetch and store */
+        for( i_y = p_vout->render.i_height / 2 ; i_y-- ; )
+        {
+            p_line1 = p_line2;
+            p_line2 += p_dest->p->i_pitch;
+
+            p_y1 = p_y2;
+            p_y2 += p_source->p[Y_PLANE].i_pitch;
+
+            for( i_x = p_vout->render.i_width / 16 ; i_x-- ; )
+            {
+                SSE2_CALL( SSE2_YUV420_UYVY_UNALIGNED );
+            }
+            for( i_x = ( p_vout->render.i_width % 16 ) / 2; i_x-- ; )
+            {
+                C_YUV420_UYVY( );
+            }
+
+            p_y1 += i_source_margin;
+            p_y2 += i_source_margin;
+            p_u += i_source_margin_c;
+            p_v += i_source_margin_c;
+            p_line1 += i_dest_margin;
+            p_line2 += i_dest_margin;
+        }
+    }
+#endif // defined(MODULE_NAME_IS_i420_yuy2_sse2)
-}
-#endif // !defined (MODULE_NAME_IS_i420_yuy2_altivec)
@@ -676 +956 @@
-}
-#endif
-

modules/video_chroma/i420_yuy2.h

@@ -6 +6 @@
- *
- * Authors: Samuel Hocevar <sam@zoy.org>
+ *          Damien Fouilleul <damien@videolan.org>
- *
- * This program is free software; you can redistribute it and/or modify
@@ -33 +34 @@
-          "r" (p_u), "r" (p_v) );                                           \
-    p_line1 += 16; p_line2 += 16; p_y1 += 8; p_y2 += 8; p_u += 4; p_v += 4; \
-;                                                             \
+
-
-#define MMX_YUV420_YUYV "                                                 \n\
@@ -112 +113 @@
-"
-
-
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
-
-#define C_YUV420_YVYU( )                                                    \
@@ -128 +257 @@
-    *(p_line1)++ = *(p_line2)++ = *(p_v) - 0x80; p_v += 2;                  \
-
-
-
-
+
-#define C_YUV420_YUYV( )                                                    \
-    *(p_line1)++ = *(p_y1)++; *(p_line2)++ = *(p_y2)++;                     \