[x265] [PATCH] copy_cnt_32: avx2 asm code, improved 1521.17 cycles -> 934.46 cycles

[praveen at multicorewareinc.com]

# HG changeset patch
# User Praveen Tiwari
# Date 1411039927 -19800
# Node ID 2fae54f872d000b308c2122295bc8df7320ef135
# Parent  532f798f98d7c7f5c493a819046a45e29b2da16a
copy_cnt_32: avx2 asm code, improved 1521.17 cycles -> 934.46 cycles

diff -r 532f798f98d7 -r 2fae54f872d0 source/common/x86/asm-primitives.cpp
--- a/source/common/x86/asm-primitives.cpp	Thu Sep 18 16:37:55 2014 +0530
+++ b/source/common/x86/asm-primitives.cpp	Thu Sep 18 17:02:07 2014 +0530
@@ -1731,7 +1731,7 @@
 
         p.copy_cnt[BLOCK_8x8] = x265_copy_cnt_8_avx2;
         p.copy_cnt[BLOCK_16x16] = x265_copy_cnt_16_avx2;
-        // p.copy_cnt[BLOCK_32x32] = x265_copy_cnt_32_avx2;
+        p.copy_cnt[BLOCK_32x32] = x265_copy_cnt_32_avx2;
 
 
         p.cvt32to16_shl[BLOCK_4x4] = x265_cvt32to16_shl_4_avx2;
diff -r 532f798f98d7 -r 2fae54f872d0 source/common/x86/blockcopy8.asm
--- a/source/common/x86/blockcopy8.asm	Thu Sep 18 16:37:55 2014 +0530
+++ b/source/common/x86/blockcopy8.asm	Thu Sep 18 17:02:07 2014 +0530
@@ -4268,42 +4268,47 @@
 
 
 INIT_YMM avx2
-cglobal copy_cnt_32, 3,4,5
+cglobal copy_cnt_32, 3, 5, 5
     add         r2d, r2d
-    mov         r3d, 32/1
-    xorpd       m3, m3
+    mov         r3d, 32/2
+
+    mova        m3, [pb_1]
     xorpd       m4, m4
 
-.loop
+.loop:
     ; row 0
-    movu        m0, [r1 + 0 * mmsize]
-    movu        m1, [r1 + 1 * mmsize]
-    packsswb    m2, m0, m1
-    pcmpeqb     m2, m4
-    paddb       m3, m2
-
-    pmovsxwd    m2, xm0
-    pmovsxwd    m0, [r1 + 0 * mmsize + mmsize/2]
-    movu        [r0 + 0 * mmsize], m2
-    movu        [r0 + 1 * mmsize], m0
-    pmovsxwd    m0, xm1
-    pmovsxwd    m1, [r1 + 1 * mmsize + mmsize/2]
-    movu        [r0 + 2 * mmsize], m0
-    movu        [r0 + 3 * mmsize], m1
-
-    add         r0, 4 * mmsize
-    add         r1, r2
+    movu        m0, [r1]
+    movu        [r0], m0
+    movu        m1, [r1 + 32]
+    movu        [r0 + 32], m1
+
+    packsswb    m0, m1
+    pminub      m0, m3
+
+    ; row 1
+    movu        m1, [r1 + r2]
+    movu        [r0 + 64], m1
+    movu        m2, [r1 + r2 + 32]
+    movu        [r0 + 96], m2
+
+    packsswb    m1, m2
+    pminub      m1, m3
+    paddb       m0, m1
+    paddb       m4, m0
+
+    add         r0, 128
+    lea         r1, [r1 + 2 * r2]
     dec         r3d
-    jnz        .loop
+    jnz         .loop
 
     ; get count
-    vextracti128 xm0, m3, 1
-    paddb       xm0, xm3
-    movhlps     xm3, xm0
-    paddb       xm0, xm3
-    paddb       xm0, [pb_128]
-    psadbw      xm0, xm4
-    movd        eax, xm0
+    xorpd        m0,  m0
+    vextracti128 xm1, m4, 1
+    paddb        xm4, xm1
+    psadbw       xm4, xm0
+    movhlps      xm1, xm4
+    paddd        xm4, xm1
+    movd         eax, xm4
     RET
 
 ;-----------------------------------------------------------------------------