Changeset df3b5eec76004d2dcfbabc36ac65f64755bdbc33

Timestamp:

08/15/07 18:15:45 (11 months ago)

Author:

Damien Fouilleul <damienf@videolan.org>

git-committer:

Damien Fouilleul <damienf@videolan.org> 1187194545 +0000

git-parent:

[30900cb3df556d168d425fc7d42c11cb83c5aef7]

git-author:

Damien Fouilleul <damienf@videolan.org> 1187194545 +0000

Message:

video chromas: finalize SSE2 improvements

Files:

• AUTHORS (modified) (1 diff)
• NEWS (modified) (1 diff)
• modules/video_chroma/i420_yuy2.h (modified) (3 diffs)
• modules/video_chroma/i422_yuy2.c (modified) (3 diffs)
• modules/video_chroma/i422_yuy2.h (modified) (1 diff)

AUTHORS

@@ -191 +191 @@
-D: ActiveX control
-D: Safari/Firefox plugin for MacOS X
+D: Direct3D Video output
+D: SSE2 chroma converters
+D: improved MMX chroma converters
-S: Ireland
-

NEWS

@@ -82 +82 @@
- * Fix random characters problem in RSS filter.
- * Add rotate-deciangle for more precision on rotate filter
+ * Support for Intel SSE2 intruction set in chroma converters
+ * Improved use of Intel MMX intruction set in chroma converters
-
-Audio output

modules/video_chroma/i420_yuy2.h

@@ -367 +367 @@
-    xmm1 = _mm_loadl_epi64((__m128i *)p_u);         \
-    xmm2 = _mm_loadl_epi64((__m128i *)p_v);         \
-_si128((__m128i *)p_y1); 
-_si128((__m128i *)p_y2); 
+u_si128((__m128i *)p_y1);
+u_si128((__m128i *)p_y2);
-    _mm_prefetch(p_line1, _MM_HINT_NTA);            \
-    _mm_prefetch(p_line2, _MM_HINT_NTA);            \
@@ -403 +403 @@
-    xmm1 = _mm_loadl_epi64((__m128i *)p_v);         \
-    xmm2 = _mm_loadl_epi64((__m128i *)p_u);         \
-_si128((__m128i *)p_y1); 
-_si128((__m128i *)p_y2); 
+u_si128((__m128i *)p_y1);
+u_si128((__m128i *)p_y2);
-    _mm_prefetch(p_line1, _MM_HINT_NTA);            \
-    _mm_prefetch(p_line2, _MM_HINT_NTA);            \
@@ -440 +440 @@
-    xmm1 = _mm_loadl_epi64((__m128i *)p_u);         \
-    xmm2 = _mm_loadl_epi64((__m128i *)p_v);         \
-_si128((__m128i *)p_y1); 
-_si128((__m128i *)p_y2); 
+u_si128((__m128i *)p_y1);
+u_si128((__m128i *)p_y2);
-    _mm_prefetch(p_line1, _MM_HINT_NTA);            \
-    _mm_prefetch(p_line2, _MM_HINT_NTA);            \

modules/video_chroma/i422_yuy2.c

@@ -443 +443 @@
-    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-- ; )
+        {
+            p_line -= 2 * p_dest->p->i_pitch;
+
+            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-- ; )
+        {
+            p_line -= 2 * p_dest->p->i_pitch;
+
+            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-- ; )
-    {
@@ -458 +513 @@
-#endif
-        }
+        p_y += i_source_margin;
+        p_u += i_source_margin_c;
+        p_v += i_source_margin_c;
+        p_line += i_dest_margin;
-    }
-#if defined (MODULE_NAME_IS_i422_yuy2_mmx)
@@ -463 +522 @@
-#elif defined (MODULE_NAME_IS_i422_yuy2_sse2)
-    SSE2_END;
+#endif
+
-#endif
-}

modules/video_chroma/i422_yuy2.h

@@ -234 +234 @@
-#include <emmintrin.h>
-
+#define SSE2_CALL(SSE2_INSTRUCTIONS)    \
+    do {                                \
+        __m128i xmm0, xmm1, xmm2;        \
+        SSE2_INSTRUCTIONS               \
+        p_line += 32; p_y += 16;        \
+        p_u += 8; p_v += 8;             \
+    } while(0)
-
-#define SSE2_END  _mm_sfence()
+
+#define SSE2_YUV422_YUYV_ALIGNED                \
+    xmm0 = _mm_load_si128((__m128i *)p_y);      \
+    xmm1 = _mm_loadl_epi64((__m128i *)p_u);     \
+    xmm2 = _mm_loadl_epi64((__m128i *)p_v);     \
+    xmm1 = _mm_unpacklo_epi8(xmm1, xmm2);       \
+    xmm2 = xmm0;                                \
+    xmm2 = _mm_unpacklo_epi8(xmm2, xmm1);       \
+    _mm_stream_si128((__m128i*)(p_line), xmm2); \
+    xmm0 = _mm_unpackhi_epi8(xmm0, xmm1);       \
+    _mm_stream_si128((__m128i*)(p_line+16), xmm0);
+ 
+#define SSE2_YUV422_YUYV_UNALIGNED              \
+    xmm0 = _mm_loadu_si128((__m128i *)p_y);     \
+    xmm1 = _mm_loadl_epi64((__m128i *)p_u);     \
+    xmm2 = _mm_loadl_epi64((__m128i *)p_v);     \
+    xmm1 = _mm_unpacklo_epi8(xmm1, xmm2);       \
+    xmm2 = xmm0;                                \
+    xmm2 = _mm_unpacklo_epi8(xmm2, xmm1);       \
+    _mm_storeu_si128((__m128i*)(p_line), xmm2); \
+    xmm0 = _mm_unpackhi_epi8(xmm0, xmm1);       \
+    _mm_storeu_si128((__m128i*)(p_line+16), xmm0);
+ 
+#define SSE2_YUV422_YVYU_ALIGNED                \
+    xmm0 = _mm_load_si128((__m128i *)p_y);      \
+    xmm2 = _mm_loadl_epi64((__m128i *)p_u);     \
+    xmm1 = _mm_loadl_epi64((__m128i *)p_v);     \
+    xmm1 = _mm_unpacklo_epi8(xmm1, xmm2);       \
+    xmm2 = xmm0;                                \
+    xmm2 = _mm_unpacklo_epi8(xmm2, xmm1);       \
+    _mm_stream_si128((__m128i*)(p_line), xmm2); \
+    xmm0 = _mm_unpackhi_epi8(xmm0, xmm1);       \
+    _mm_stream_si128((__m128i*)(p_line+16), xmm0);
+
+#define SSE2_YUV422_YVYU_UNALIGNED              \
+    xmm0 = _mm_loadu_si128((__m128i *)p_y);     \
+    xmm2 = _mm_loadl_epi64((__m128i *)p_u);     \
+    xmm1 = _mm_loadl_epi64((__m128i *)p_v);     \
+    xmm1 = _mm_unpacklo_epi8(xmm1, xmm2);       \
+    xmm2 = xmm0;                                \
+    xmm2 = _mm_unpacklo_epi8(xmm2, xmm1);       \
+    _mm_storeu_si128((__m128i*)(p_line), xmm2); \
+    xmm0 = _mm_unpackhi_epi8(xmm0, xmm1);       \
+    _mm_storeu_si128((__m128i*)(p_line+16), xmm0);
+
+#define SSE2_YUV422_UYVY_ALIGNED                \
+    xmm0 = _mm_load_si128((__m128i *)p_y);      \
+    xmm1 = _mm_loadl_epi64((__m128i *)p_u);     \
+    xmm2 = _mm_loadl_epi64((__m128i *)p_v);     \
+    xmm1 = _mm_unpacklo_epi8(xmm1, xmm2);       \
+    xmm2 = xmm1;                                \
+    xmm2 = _mm_unpacklo_epi8(xmm2, xmm0);       \
+    _mm_stream_si128((__m128i*)(p_line), xmm2); \
+    xmm1 = _mm_unpackhi_epi8(xmm1, xmm0);       \
+    _mm_stream_si128((__m128i*)(p_line+16), xmm1);
+
+#define SSE2_YUV422_UYVY_UNALIGNED              \
+    xmm0 = _mm_loadu_si128((__m128i *)p_y);     \
+    xmm1 = _mm_loadl_epi64((__m128i *)p_u);     \
+    xmm2 = _mm_loadl_epi64((__m128i *)p_v);     \
+    xmm1 = _mm_unpacklo_epi8(xmm1, xmm2);       \
+    xmm2 = xmm1;                                \
+    xmm2 = _mm_unpacklo_epi8(xmm2, xmm0);       \
+    _mm_storeu_si128((__m128i*)(p_line), xmm2); \
+    xmm1 = _mm_unpackhi_epi8(xmm1, xmm0);       \
+    _mm_storeu_si128((__m128i*)(p_line+16), xmm1);
-
-#endif