x86: AVX2 high bit-depth denoise_dct

28->15 cycles

Also reorder instructions to use fewer registers, 3 cycles faster on Ivy Bridge with 64-bit Windows.

diff --git a/common/quant.c b/common/quant.c
index 24da5f999cd7b76ef027e0db62d04b038bf66398..d1f7b2fd54ce15060b3a311f84eb5f652fd09ec2 100644 (file)
--- a/common/quant.c
+++ b/common/quant.c
@@ -543,6 +543,7 @@ void x264_quant_init( x264_t *h, int cpu, x264_quant_function_t *pf )
         pf->quant_4x4_dc = x264_quant_4x4_dc_avx2;
         pf->quant_8x8 = x264_quant_8x8_avx2;
         pf->quant_4x4x4 = x264_quant_4x4x4_avx2;
+        pf->denoise_dct = x264_denoise_dct_avx2;
     }
 #endif // HAVE_MMX
 #else // !HIGH_BIT_DEPTH

diff --git a/common/x86/quant-a.asm b/common/x86/quant-a.asm
index b693181598d24280055f9fc7d7595f48883954a8..7c2557b190d12fc76206eacbd43d1b1de90ac834 100644 (file)
--- a/common/x86/quant-a.asm
+++ b/common/x86/quant-a.asm
@@ -946,31 +946,29 @@ OPTIMIZE_CHROMA_2x2_DC
 ; void denoise_dct( int32_t *dct, uint32_t *sum, uint32_t *offset, int size )
 ;-----------------------------------------------------------------------------
 %macro DENOISE_DCT 0
-cglobal denoise_dct, 4,4,8
-    pxor      m6, m6
+cglobal denoise_dct, 4,4,6
+    pxor      m5, m5
     movsxdifnidn r3, r3d
 .loop:
     mova      m2, [r0+r3*4-2*mmsize]
     mova      m3, [r0+r3*4-1*mmsize]
     ABSD      m0, m2
     ABSD      m1, m3
-    mova      m4, m0
-    mova      m5, m1
+    paddd     m4, m0, [r1+r3*4-2*mmsize]
     psubd     m0, [r2+r3*4-2*mmsize]
+    mova      [r1+r3*4-2*mmsize], m4
+    paddd     m4, m1, [r1+r3*4-1*mmsize]
     psubd     m1, [r2+r3*4-1*mmsize]
-    pcmpgtd   m7, m0, m6
-    pand      m0, m7
-    pcmpgtd   m7, m1, m6
-    pand      m1, m7
+    mova      [r1+r3*4-1*mmsize], m4
+    pcmpgtd   m4, m0, m5
+    pand      m0, m4
+    pcmpgtd   m4, m1, m5
+    pand      m1, m4
     PSIGND    m0, m2
     PSIGND    m1, m3
     mova      [r0+r3*4-2*mmsize], m0
     mova      [r0+r3*4-1*mmsize], m1
-    paddd     m4, [r1+r3*4-2*mmsize]
-    paddd     m5, [r1+r3*4-1*mmsize]
-    mova      [r1+r3*4-2*mmsize], m4
-    mova      [r1+r3*4-1*mmsize], m5
-    sub       r3, mmsize/2
+    sub      r3d, mmsize/2
     jg .loop
     RET
 %endmacro
@@ -985,6 +983,8 @@ INIT_XMM ssse3
 DENOISE_DCT
 INIT_XMM avx
 DENOISE_DCT
+INIT_YMM avx2
+DENOISE_DCT
 
 %else ; !HIGH_BIT_DEPTH