[FFmpeg-devel] [PATCH 6/7] avcodec/x86/mpegvideoencdsp: speed up draw_edges_mmx by using memcpy()

[Ramiro Polla]

The mmx memory copy code is not nearly as efficient as memcpy(), which
would make draw_edges_mmx much slower than draw_edges_8_c.

Intel(R) Core(TM) i5-5300U CPU @ 2.30GHz:
                             before    after
draw_edges_8_1724_4_mmx:     8697.2   8739.6  ( 1.00x)
draw_edges_8_1724_8_mmx:    10439.0  10548.4  ( 0.99x)
draw_edges_8_1724_16_mmx:   10687.5  10876.1  ( 0.98x)
draw_edges_128_407_4_mmx:    4252.5   3562.4  ( 1.19x)
draw_edges_128_407_8_mmx:    4561.7   3868.1  ( 1.18x)
draw_edges_128_407_16_mmx:   5505.7   4533.4  ( 1.21x)
draw_edges_1080_31_4_mmx:    1667.7    560.4  ( 2.98x)
draw_edges_1080_31_8_mmx:    2804.5   1232.1  ( 2.28x)
draw_edges_1080_31_16_mmx:  12478.5   3829.6  ( 3.26x)
draw_edges_1920_4_4_mmx:     2596.2    812.4  ( 3.20x)
draw_edges_1920_4_8_mmx:     8056.0   2964.4  ( 2.72x)
draw_edges_1920_4_16_mmx:   24628.7   6387.9  ( 3.86x)
---
 libavcodec/x86/mpegvideoencdsp_init.c | 51 ++++++---------------------
 1 file changed, 11 insertions(+), 40 deletions(-)

diff --git a/libavcodec/x86/mpegvideoencdsp_init.c b/libavcodec/x86/mpegvideoencdsp_init.c
index 503548e668..c30fa91175 100644
--- a/libavcodec/x86/mpegvideoencdsp_init.c
+++ b/libavcodec/x86/mpegvideoencdsp_init.c
@@ -166,46 +166,17 @@ static void draw_edges_mmx(uint8_t *buf, int wrap, int width, int height,
               "r" (ptr + wrap * height));
     }
 
-    /* top and bottom (and hopefully also the corners) */
-    if (sides & EDGE_TOP) {
-        for (i = 0; i < h; i += 4) {
-            ptr = buf - (i + 1) * wrap - w;
-            __asm__ volatile (
-                "1:                             \n\t"
-                "movq (%1, %0), %%mm0           \n\t"
-                "movq    %%mm0, (%0)            \n\t"
-                "movq    %%mm0, (%0, %2)        \n\t"
-                "movq    %%mm0, (%0, %2, 2)     \n\t"
-                "movq    %%mm0, (%0, %3)        \n\t"
-                "add        $8, %0              \n\t"
-                "cmp        %4, %0              \n\t"
-                "jb         1b                  \n\t"
-                : "+r" (ptr)
-                : "r" ((x86_reg) buf - (x86_reg) ptr - w),
-                  "r" ((x86_reg) - wrap), "r" ((x86_reg) - wrap * 3),
-                  "r" (ptr + width + 2 * w));
-        }
-    }
-
-    if (sides & EDGE_BOTTOM) {
-        for (i = 0; i < h; i += 4) {
-            ptr = last_line + (i + 1) * wrap - w;
-            __asm__ volatile (
-                "1:                             \n\t"
-                "movq (%1, %0), %%mm0           \n\t"
-                "movq    %%mm0, (%0)            \n\t"
-                "movq    %%mm0, (%0, %2)        \n\t"
-                "movq    %%mm0, (%0, %2, 2)     \n\t"
-                "movq    %%mm0, (%0, %3)        \n\t"
-                "add        $8, %0              \n\t"
-                "cmp        %4, %0              \n\t"
-                "jb         1b                  \n\t"
-                : "+r" (ptr)
-                : "r" ((x86_reg) last_line - (x86_reg) ptr - w),
-                  "r" ((x86_reg) wrap), "r" ((x86_reg) wrap * 3),
-                  "r" (ptr + width + 2 * w));
-        }
-    }
+    /* top and bottom + corners */
+    buf -= w;
+    last_line = buf + (height - 1) * wrap;
+    if (sides & EDGE_TOP)
+        for (i = 0; i < h; i++)
+            // top
+            memcpy(buf - (i + 1) * wrap, buf, width + w + w);
+    if (sides & EDGE_BOTTOM)
+        for (i = 0; i < h; i++)
+            // bottom
+            memcpy(last_line + (i + 1) * wrap, last_line, width + w + w);
 }
 
 #endif /* HAVE_INLINE_ASM */
-- 
2.30.2