<div style="line-height:1.7;color:#000000;font-size:14px;font-family:arial"><div><br> </div>At 2015-03-11 17:11:19,"Praveen Tiwari" <praveen@multicorewareinc.com> wrote:<br> <blockquote id="isReplyContent" style="margin: 0px 0px 0px 0.8ex; padding-left: 1ex; border-left-color: rgb(204, 204, 204); border-left-width: 1px; border-left-style: solid;"><div dir="ltr"><br><div class="gmail_quote">---------- Forwarded message ----------<br>From: <b class="gmail_sendername"></b> <span dir="ltr"><<a href="mailto:sumalatha@multicorewareinc.com">sumalatha@multicorewareinc.com</a>></span><br>Date: Wed, Mar 11, 2015 at 2:24 PM<br>Subject: [x265] [PATCH] asm: avx2 code for sad[32x32] for 8bpp<br>To: <a href="mailto:x265-devel@videolan.org">x265-devel@videolan.org</a><br><br><br># HG changeset patch<br>
# User Sumalatha Polureddy<<a href="mailto:sumalatha@multicorewareinc.com">sumalatha@multicorewareinc.com</a>><br>
# Date 1426064050 -19800<br>
# Node ID 01bfd365bf5f5317874b5c0315736ca76176f3df<br>
# Parent  800f8ecd1e7393756f4bb58e536497162dc32150<br>
asm: avx2 code for sad[32x32] for 8bpp<br>
<br>>>SSE3<br>>>sad[32x32]  230.81x          745.76          172131.92<br>
<br>>>AVX2<br>>>sad[32x32]  330.38x          496.68          164091.02</div><div class="gmail_quote"><br></div><div class="gmail_quote"><br></div><div class="gmail_quote">Are you comparing the debug mode performance numbers? 230.81x  ???</div><div class="gmail_quote"> <br></div><div class="gmail_quote">SSE3<br></div><div class="gmail_quote">sad[32x32]  31.96x   770.39          24623.33<br></div><div class="gmail_quote"><br>
on i7-4770k CPU. Please check the issue.<br><br><br>
diff -r 800f8ecd1e73 -r 01bfd365bf5f source/common/x86/asm-primitives.cpp<br>
--- a/source/common/x86/asm-primitives.cpp      Tue Mar 10 10:49:11 2015 +0530<br>
+++ b/source/common/x86/asm-primitives.cpp      Wed Mar 11 14:24:10 2015 +0530<br>
@@ -1442,6 +1442,8 @@<br>
         p.pu[LUMA_8x16].satd  = x265_pixel_satd_8x16_avx2;<br>
         p.pu[LUMA_8x8].satd   = x265_pixel_satd_8x8_avx2;<br>
<br>
+        p.pu[LUMA_32x32].sad = x265_pixel_sad_32x32_avx2;<br>
+<br>
         p.pu[LUMA_8x4].sad_x3 = x265_pixel_sad_x3_8x4_avx2;<br>
         p.pu[LUMA_8x8].sad_x3 = x265_pixel_sad_x3_8x8_avx2;<br>
         p.pu[LUMA_8x16].sad_x3 = x265_pixel_sad_x3_8x16_avx2;<br>
diff -r 800f8ecd1e73 -r 01bfd365bf5f source/common/x86/sad-a.asm<br>
--- a/source/common/x86/sad-a.asm       Tue Mar 10 10:49:11 2015 +0530<br>
+++ b/source/common/x86/sad-a.asm       Wed Mar 11 14:24:10 2015 +0530<br>
@@ -3897,5 +3897,31 @@<br>
     movq            [r6 + 8], xm1<br>
     RET<br>
<br>
+INIT_YMM avx2<br>
+cglobal pixel_sad_32x32, 4,4,5<br>
+    xorps           m0, m0<br>
+%assign x 0<br>
+%rep 16<br>
+    movu           m1, [r0]               ; row 0 of pix0<br>
+    movu           m2, [r2]               ; row 0 of pix1<br>
+    movu           m3, [r0 + r1]          ; row 1 of pix0<br>
+    movu           m4, [r2 + r3]          ; row 1 of pix1<br>
+<br>
+    psadbw         m1, m2<br>
+    psadbw         m3, m4<br>
+    paddd          m0, m1<br>
+    paddd          m0, m3<br>
+%assign x x+1<br>
+  %if x < 16<br>
+    lea     r2,     [r2 + 2 * r3]<br>
+    lea     r0,     [r0 + 2 * r1]<br>
+  %endif<br>
+%endrep<br>
+    vextracti128   xm1, m0, 1<br>
+    paddd          xm0, xm1<br>
+    pshufd         xm1, xm0, 2<br>
+    paddd          xm0,xm1<br>
+    movd           eax, xm0<br>
+    RET<br>
<br>
 %endif<br>
</div><div class="gmail_quote">unroll 16 times?</div><div class="gmail_quote">for this large function, could you try the loop and compare performance to choice best factor of unroll</div></div>
</blockquote></div>