[FFmpeg-devel] [PATCH 2/5] x86/vvc_sad: optimize vvc_sad_16

[James Almer]

Signed-off-by: James Almer <jamrial at gmail.com>
---
 libavcodec/x86/vvc/vvc_sad.asm | 27 ++++++++++++++-------------
 1 file changed, 14 insertions(+), 13 deletions(-)

diff --git a/libavcodec/x86/vvc/vvc_sad.asm b/libavcodec/x86/vvc/vvc_sad.asm
index a20818530f..829dbce489 100644
--- a/libavcodec/x86/vvc/vvc_sad.asm
+++ b/libavcodec/x86/vvc/vvc_sad.asm
@@ -96,7 +96,7 @@ cglobal vvc_sad_8, 6, 9, 5, src1, src2, dx, dy, block_w, block_h, off1, off2, ro
         movd          eax, xm0
     RET
 
-cglobal vvc_sad_16, 6, 9, 5, src1, src2, dx, dy, block_w, block_h, off1, off2, row_idx
+cglobal vvc_sad_16, 6, 8, 5, src1, src2, dx, dy, block_w, block_h, off1, off2
     movsxdifnidn    dxq, dxd
     movsxdifnidn    dyq, dyd
 
@@ -121,26 +121,27 @@ cglobal vvc_sad_16, 6, 9, 5, src1, src2, dx, dy, block_w, block_h, off1, off2, r
     pxor               m3, m3
     vpbroadcastd       m4, [pw_1]
 
-        sar      block_wd, 4
+        shl      block_wd, 1
+        add         src1q, block_wq
+        add         src2q, block_wq
+        neg      block_wq
+
+DEFINE_ARGS src1, src2, dx, dy, block_w, block_h, row_idx
         .loop_height:
-        mov         off1q, src1q
-        mov         off2q, src2q
-        mov      row_idxd, block_wd
+        mov      row_idxq, block_wq
 
         .loop_width:
-            movu               m0, [src1q]
-            movu               m1, [src2q]
+            movu               m0, [src1q+row_idxq]
+            movu               m1, [src2q+row_idxq]
             MIN_MAX_SAD        m1, m0, m2
             pmaddwd            m1, m4
             paddd              m3, m1
 
-            add             src1q, 32
-            add             src2q, 32
-            dec          row_idxd
-            jg        .loop_width
+            add          row_idxq, mmsize
+            jl        .loop_width
 
-        lea         src1q, [off1q + ROWS * MAX_PB_SIZE * 2]
-        lea         src2q, [off2q + ROWS * MAX_PB_SIZE * 2]
+        add         src1q, ROWS * MAX_PB_SIZE * 2
+        add         src2q, ROWS * MAX_PB_SIZE * 2
 
         sub      block_hd, 2
         jg   .loop_height
-- 
2.45.1