[x265] [PATCH] denoise_dct: avx2 asm code

praveen at multicorewareinc.com praveen at multicorewareinc.com
Thu Sep 18 12:00:39 CEST 2014


# HG changeset patch
# User Praveen Tiwari
# Date 1411034418 -19800
# Node ID e723ecc1e5c99c451cbc8034514b9dc590a2d4ef
# Parent  c4b689f6050231e99b9663b7504cd7fff90bdafb
denoise_dct: avx2 asm code

diff -r c4b689f60502 -r e723ecc1e5c9 source/common/x86/asm-primitives.cpp
--- a/source/common/x86/asm-primitives.cpp	Thu Sep 18 15:11:26 2014 +0530
+++ b/source/common/x86/asm-primitives.cpp	Thu Sep 18 15:30:18 2014 +0530
@@ -1738,7 +1738,7 @@
         p.cvt32to16_shl[BLOCK_8x8] = x265_cvt32to16_shl_8_avx2;
         p.cvt32to16_shl[BLOCK_16x16] = x265_cvt32to16_shl_16_avx2;
         p.cvt32to16_shl[BLOCK_32x32] = x265_cvt32to16_shl_32_avx2;
-        //p.denoiseDct = x265_denoise_dct_avx2;
+        p.denoiseDct = x265_denoise_dct_avx2;
         p.dct[DCT_4x4] = x265_dct4_avx2;
         p.quant = x265_quant_avx2;
         p.nquant = x265_nquant_avx2;
diff -r c4b689f60502 -r e723ecc1e5c9 source/common/x86/dct8.asm
--- a/source/common/x86/dct8.asm	Thu Sep 18 15:11:26 2014 +0530
+++ b/source/common/x86/dct8.asm	Thu Sep 18 15:30:18 2014 +0530
@@ -1082,24 +1082,26 @@
     RET
 
 INIT_YMM avx2
-cglobal denoise_dct, 4,4,4
-    pxor      m3, m3
-    movsxdifnidn r3, r3d
+cglobal denoise_dct, 4, 4, 6
+    pxor     m5,  m5
+    shr      r3d, 3
 .loop:
-    mova      m1, [r0+r3*2-mmsize]
-    pabsw     m0, m1
-    psubusw   m2, m0, [r2+r3*2-mmsize]
-    vpermq    m0, m0, q3120
-    psignw    m2, m1
-    mova [r0+r3*2-mmsize], m2
-    punpcklwd m1, m0, m3
-    punpckhwd m0, m3
-    paddd     m1, [r1+r3*4-2*mmsize]
-    paddd     m0, [r1+r3*4-1*mmsize]
-    mova      [r1+r3*4-2*mmsize], m1
-    mova      [r1+r3*4-1*mmsize], m0
-    sub       r3, mmsize/2
-    jg .loop
+    mova     m0, [r0]
+    pabsd    m1, m0
+    mova     m2, [r1]
+    paddd    m2, m1
+    mova     [r1], m2
+    pmovzxwd m3, [r2]
+    psubd    m1, m3
+    pcmpgtd  m4, m1, m5
+    pand     m1, m4
+    psignd   m1, m0
+    mova     [r0], m1
+    add      r0, 32
+    add      r1, 32
+    add      r2, 16
+    dec      r3d
+    jnz .loop
     RET
 
 


More information about the x265-devel mailing list