[FFmpeg-devel] [PATCH 4/4] lavc/x86/hevc_add_res: Fix coeff overflow in ADD_RES_SSE_16_32_8

[Linjie Fu]

Fix overflow for coeff -32768 in function ADD_RES_SSE_16_32_8 with no
performance drop.(SSE2/AVX/AVX2)

./checkasm --test=hevc_add_res --bench

Mainline:
  - hevc_add_res.add_residual [OK]
    hevc_add_res_32x32_8_sse2: 127.5
    hevc_add_res_32x32_8_avx: 127.0
    hevc_add_res_32x32_8_avx2: 86.5

Add overflow test case:
  - hevc_add_res.add_residual [FAILED]

After:
  - hevc_add_res.add_residual [OK]
    hevc_add_res_32x32_8_sse2: 126.8
    hevc_add_res_32x32_8_avx: 128.3
    hevc_add_res_32x32_8_avx2: 86.8

Signed-off-by: Xu Guangxin <guangxin.xu at intel.com>
Signed-off-by: Linjie Fu <linjie.fu at intel.com>
---
 libavcodec/x86/hevc_add_res.asm | 50 +++++++++++++++++++++--------------------
 1 file changed, 26 insertions(+), 24 deletions(-)

diff --git a/libavcodec/x86/hevc_add_res.asm b/libavcodec/x86/hevc_add_res.asm
index e5e9f24..4d3aefc 100644
--- a/libavcodec/x86/hevc_add_res.asm
+++ b/libavcodec/x86/hevc_add_res.asm
@@ -84,34 +84,36 @@ cglobal hevc_add_residual_4_8, 3, 3, 6
 %endmacro
 
 %macro ADD_RES_SSE_16_32_8 3
-    mova             xm2, [r1+%1]
-    mova             xm6, [r1+%1+16]
+    mova            m1, [%2]
+    mova            m2, m1
+    punpcklbw       m1, m0
+    punpckhbw       m2, m0
+    mova           xm5, [r1+%1]
+    mova           xm6, [r1+%1+16]
 %if cpuflag(avx2)
-    vinserti128       m2, m2, [r1+%1+32], 1
-    vinserti128       m6, m6, [r1+%1+48], 1
+    vinserti128     m5, m5, [r1+%1+32], 1
+    vinserti128     m6, m6, [r1+%1+48], 1
 %endif
-    psubw             m1, m0, m2
-    psubw             m5, m0, m6
-    packuswb          m2, m6
-    packuswb          m1, m5
-
-    mova             xm4, [r1+%1+mmsize*2]
-    mova             xm6, [r1+%1+mmsize*2+16]
+    paddsw          m1, m5
+    paddsw          m2, m6
+
+    mova            m3, [%3]
+    mova            m4, m3
+    punpcklbw       m3, m0
+    punpckhbw       m4, m0
+    mova           xm5, [r1+%1+mmsize*2]
+    mova           xm6, [r1+%1+mmsize*2+16]
 %if cpuflag(avx2)
-    vinserti128       m4, m4, [r1+%1+96 ], 1
-    vinserti128       m6, m6, [r1+%1+112], 1
+    vinserti128     m5, m5, [r1+%1+96], 1
+    vinserti128     m6, m6, [r1+%1+112], 1
 %endif
-    psubw             m3, m0, m4
-    psubw             m5, m0, m6
-    packuswb          m4, m6
-    packuswb          m3, m5
-
-    paddusb           m2, [%2]
-    paddusb           m4, [%3]
-    psubusb           m2, m1
-    psubusb           m4, m3
-    mova            [%2], m2
-    mova            [%3], m4
+    paddsw          m3, m5
+    paddsw          m4, m6
+
+    packuswb        m1, m2
+    packuswb        m3, m4
+    mova          [%2], m1
+    mova          [%3], m3
 %endmacro
 
 
-- 
2.7.4