[x265] [PATCH] copy_cnt_32, AVX2 asm code as per new interface, performance improved from 16.81x to 32.16x on HASWELL-I5

[praveen at multicorewareinc.com]

# HG changeset patch
# User Praveen Tiwari
# Date 1410340533 -19800
# Node ID 7bc4db02ccc728f6e2ddedd036c96e3d37b90f22
# Parent  c5b3e04e4eba2fcc4298c225d11ab25e0da82558
copy_cnt_32, AVX2 asm code as per new interface, performance improved from 16.81x to 32.16x on HASWELL-I5

diff -r c5b3e04e4eba -r 7bc4db02ccc7 source/common/x86/blockcopy8.asm
--- a/source/common/x86/blockcopy8.asm	Wed Sep 10 13:35:30 2014 +0530
+++ b/source/common/x86/blockcopy8.asm	Wed Sep 10 14:45:33 2014 +0530
@@ -4317,42 +4317,68 @@
 
 
 INIT_YMM avx2
-cglobal copy_cnt_32, 3,4,5
+cglobal copy_cnt_32, 3,5,7
     add         r2d, r2d
-    mov         r3d, 32/1
-    xorpd       m3, m3
-    xorpd       m4, m4
-
-.loop
+    lea         r3,  [r2 * 3]
+    mov         r4d, 32/4
+
+    xorpd       m5, m5
+    xorpd       m6, m6
+
+.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
-    dec         r3d
-    jnz        .loop
+    movu        m0, [r1]
+    movu        [r0], m0
+    movu        m1, [r1 + 32]
+    movu        [r0 + 32], m1
+
+    vpacksswb   m0, m1
+    pminub      m0, [pb_1]
+
+    ; row 1
+    movu        m1, [r1 + r2]
+    movu        [r0 + 64], m1
+    movu        m2, [r1 + r2 + 32]
+    movu        [r0 + 96], m2
+
+    vpacksswb   m1, m2
+    pminub      m1, [pb_1]
+    paddb       m0, m1
+
+    ; row 2
+    movu        m2, [r1 + r2 * 2]
+    movu        [r0 + 128], m2
+    movu        m3, [r1 + r2 * 2 + 32]
+    movu        [r0 + 160], m3
+
+    vpacksswb   m2, m3
+    pminub      m2, [pb_1]
+
+    ; row 3
+    movu        m3, [r1 + r3]
+    movu        [r0 + 192], m3
+    movu        m4, [r1 + r3 + 32]
+    movu        [r0 + 224], m4
+
+    vpacksswb   m3, m4
+    pminub      m3, [pb_1]
+    paddb       m2, m3
+
+    paddb       m0, m2
+    paddb       m6, m0
+
+    add         r0, 256
+    lea         r1, [r1 + 4 * r2]
+    dec         r4d
+    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
+    vextracti128 xm1, m6, 1
+    paddb        xm6, xm1
+    psadbw       xm6, xm5
+    movhlps      xm1, xm6
+    paddd        xm6, xm1
+    movd         eax, xm6
     RET
 
 ;-----------------------------------------------------------------------------