[vlc-commits] copy: minor speed enhancement to USWC copy

Jean-Yves Avenard git at videolan.org
Mon Jun 16 09:51:17 CEST 2014


vlc | branch: master | Jean-Yves Avenard <jyavenard at mythtv.org> | Sun Jun 15 09:04:37 2014 +0000| [b50a07a3ee8728c602d88432365aae67abc81b28] | committer: Jean-Baptiste Kempf

copy: minor speed enhancement to USWC copy

Adopt suggestions as per https://software.intel.com/en-us/articles/copying-accelerated-video-decode-frame-buffers:
- 64kB aligned memory buffer
- Call to mfence only around copy to cache routine.

Use a single instruction to copy non-aligned memory at startup.
Remove compilation warning on mac

This result in a small 1.3% speed increase on an i7-4650U

Signed-off-by: Jean-Baptiste Kempf <jb at videolan.org>

> http://git.videolan.org/gitweb.cgi/vlc.git/?a=commit;h=b50a07a3ee8728c602d88432365aae67abc81b28
---

 modules/video_chroma/copy.c |   41 +++++++++++++++++++++++------------------
 1 file changed, 23 insertions(+), 18 deletions(-)

diff --git a/modules/video_chroma/copy.c b/modules/video_chroma/copy.c
index d29843c..cc98c92 100644
--- a/modules/video_chroma/copy.c
+++ b/modules/video_chroma/copy.c
@@ -35,8 +35,8 @@
 int CopyInitCache(copy_cache_t *cache, unsigned width)
 {
 #ifdef CAN_COMPILE_SSE2
-    cache->size = __MAX((width + 0x0f) & ~ 0x0f, 4096);
-    cache->buffer = vlc_memalign(16, cache->size);
+    cache->size = __MAX((width + 0x3f) & ~ 0x3f, 4096);
+    cache->buffer = vlc_memalign(64, cache->size);
     if (!cache->buffer)
         return VLC_EGENERIC;
 #else
@@ -57,9 +57,15 @@ void CopyCleanCache(copy_cache_t *cache)
 }
 
 #ifdef CAN_COMPILE_SSE2
-/* Copy 64 bytes from srcp to dstp loading data with the SSE>=2 instruction
+/* Copy 16/64 bytes from srcp to dstp loading data with the SSE>=2 instruction
  * load and storing data with the SSE>=2 instruction store.
  */
+#define COPY16(dstp, srcp, load, store) \
+    asm volatile (                      \
+        load "  0(%[src]), %%xmm1\n"    \
+        store " %%xmm1,    0(%[dst])\n" \
+        : : [dst]"r"(dstp), [src]"r"(srcp) : "memory", "xmm1")
+
 #define COPY64(dstp, srcp, load, store) \
     asm volatile (                      \
         load "  0(%[src]), %%xmm1\n"    \
@@ -97,16 +103,16 @@ static void CopyFromUswc(uint8_t *dst, size_t dst_pitch,
                          unsigned width, unsigned height,
                          unsigned cpu)
 {
+#ifndef CAN_COMPILE_SSSE3
+    VLC_UNUSED(cpu);
+#endif
     assert(((intptr_t)dst & 0x0f) == 0 && (dst_pitch & 0x0f) == 0);
 
     asm volatile ("mfence");
 
     for (unsigned y = 0; y < height; y++) {
         const unsigned unaligned = (-(uintptr_t)src) & 0x0f;
-        unsigned x = 0;
-
-        for (; x < unaligned; x++)
-            dst[x] = src[x];
+        unsigned x = unaligned;
 
 #ifdef CAN_COMPILE_SSE4_1
         if (vlc_CPU_SSE4_1()) {
@@ -114,6 +120,7 @@ static void CopyFromUswc(uint8_t *dst, size_t dst_pitch,
                 for (; x+63 < width; x += 64)
                     COPY64(&dst[x], &src[x], "movntdqa", "movdqa");
             } else {
+                COPY16(dst, src, "movdqu", "movdqa");
                 for (; x+63 < width; x += 64)
                     COPY64(&dst[x], &src[x], "movntdqa", "movdqu");
             }
@@ -124,6 +131,7 @@ static void CopyFromUswc(uint8_t *dst, size_t dst_pitch,
                 for (; x+63 < width; x += 64)
                     COPY64(&dst[x], &src[x], "movdqa", "movdqa");
             } else {
+                COPY16(dst, src, "movdqu", "movdqa");
                 for (; x+63 < width; x += 64)
                     COPY64(&dst[x], &src[x], "movdqa", "movdqu");
             }
@@ -135,6 +143,7 @@ static void CopyFromUswc(uint8_t *dst, size_t dst_pitch,
         src += src_pitch;
         dst += dst_pitch;
     }
+    asm volatile ("mfence");
 }
 
 VLC_SSE
@@ -144,8 +153,6 @@ static void Copy2d(uint8_t *dst, size_t dst_pitch,
 {
     assert(((intptr_t)src & 0x0f) == 0 && (src_pitch & 0x0f) == 0);
 
-    asm volatile ("mfence");
-
     for (unsigned y = 0; y < height; y++) {
         unsigned x = 0;
 
@@ -172,15 +179,15 @@ static void SSE_SplitUV(uint8_t *dstu, size_t dstu_pitch,
                         const uint8_t *src, size_t src_pitch,
                         unsigned width, unsigned height, unsigned cpu)
 {
+#ifndef CAN_COMPILE_SSSE3
     VLC_UNUSED(cpu);
+#endif
     const uint8_t shuffle[] = { 0, 2, 4, 6, 8, 10, 12, 14,
                                 1, 3, 5, 7, 9, 11, 13, 15 };
     const uint8_t mask[] = { 0xff, 0x00, 0xff, 0x00, 0xff, 0x00, 0xff, 0x00,
                              0xff, 0x00, 0xff, 0x00, 0xff, 0x00, 0xff, 0x00 };
 
-    assert(((intptr_t)src & 0x0f) == 0 && (src_pitch & 0x0f) == 0);
-
-    asm volatile ("mfence");
+    assert(((intptr_t)src & 0xf) == 0 && (src_pitch & 0x0f) == 0);
 
     for (unsigned y = 0; y < height; y++) {
         unsigned x = 0;
@@ -280,7 +287,6 @@ static void SSE_CopyPlane(uint8_t *dst, size_t dst_pitch,
         src += src_pitch * hblock;
         dst += dst_pitch * hblock;
     }
-    asm volatile ("mfence");
 }
 
 static void SSE_SplitPlanes(uint8_t *dstu, size_t dstu_pitch,
@@ -289,27 +295,26 @@ static void SSE_SplitPlanes(uint8_t *dstu, size_t dstu_pitch,
                             uint8_t *cache, size_t cache_size,
                             unsigned width, unsigned height, unsigned cpu)
 {
-    const unsigned w2_16 = (2*width+15) & ~15;
-    const unsigned hstep = cache_size / w2_16;
+    const unsigned w16 = (2*width+15) & ~15;
+    const unsigned hstep = cache_size / w16;
     assert(hstep > 0);
 
     for (unsigned y = 0; y < height; y += hstep) {
         const unsigned hblock =  __MIN(hstep, height - y);
 
         /* Copy a bunch of line into our cache */
-        CopyFromUswc(cache, w2_16, src, src_pitch,
+        CopyFromUswc(cache, w16, src, src_pitch,
                      2*width, hblock, cpu);
 
         /* Copy from our cache to the destination */
         SSE_SplitUV(dstu, dstu_pitch, dstv, dstv_pitch,
-                    cache, w2_16, width, hblock, cpu);
+                    cache, w16, width, hblock, cpu);
 
         /* */
         src  += src_pitch  * hblock;
         dstu += dstu_pitch * hblock;
         dstv += dstv_pitch * hblock;
     }
-    asm volatile ("mfence");
 }
 
 static void SSE_CopyFromNv12(picture_t *dst,



More information about the vlc-commits mailing list