<div dir="ltr"><br><div class="gmail_extra"><br><br><div class="gmail_quote">On Tue, Oct 8, 2013 at 2:33 AM, <span dir="ltr"><<a href="mailto:praveen@multicorewareinc.com" target="_blank">praveen@multicorewareinc.com</a>></span> wrote:<br>
<blockquote class="gmail_quote" style="margin:0 0 0 .8ex;border-left:1px #ccc solid;padding-left:1ex"># HG changeset patch<br>
# User Praveen Tiwari<br>
# Date 1381217602 -19800<br>
# Node ID 4f728eeab74a089c86068663baf522c40a136981<br>
# Parent 2b2fc4a46c7dcf8720b1b9872c0f3b86c048ffcd<br>
filterHorizontal_p_p_4, 48x48 asm code<br></blockquote><div><br></div><div>For luma, the only width-48 block used in the encoder is 48x64.</div><div><br></div><div>And at width 64 there is only 64x16, 64x32, 64x48, 64x64 (1/4, 1/2, 3/4, 4/4).</div>
<div><br></div><div>The same applies to width 32 (8, 16, 24, 32) and 16 (4, 8, 12, 16). (width 24 only has height 32, width 12 only has height 16)</div><div><br></div><div>width 8 only has 8x4 and 8x8</div><div><br></div>
<div>So to minimize your work effort you should be writing 8-tap luma macros that interpolate:</div><div><br></div><div>* 64x16</div><div>* 32x8<br></div><div>* 16x4<br></div><div>* 8x4<br></div><div><br></div><div>The 48x64, 24x32, and 12x16 blocks are rarely used (AMP) and could be built from 16x4 or 4x4.</div>
<div><br></div><div>These 4-tap filters are only used for 4:2:0 chroma and they will have different block-size requirements, but you need to figure out exactly which chroma blocks are needed before writing 4-tap block intrinsics.</div>
<div><br></div><blockquote class="gmail_quote" style="margin:0 0 0 .8ex;border-left:1px #ccc solid;padding-left:1ex">
diff -r 2b2fc4a46c7d -r 4f728eeab74a source/common/x86/ipfilter8.asm<br>
--- a/source/common/x86/ipfilter8.asm Tue Oct 08 12:53:44 2013 +0530<br>
+++ b/source/common/x86/ipfilter8.asm Tue Oct 08 13:03:22 2013 +0530<br>
@@ -530,3 +530,101 @@<br>
FILTER_H4_w32 x0, x1, x2, x3<br>
movu [dstq + 16], x1<br>
RET<br>
+<br>
+ SECTION_RODATA 32<br>
+tab_Tm: db 0, 1, 2, 3, 1, 2, 3, 4, 2, 3, 4, 5, 3, 4, 5, 6<br>
+ db 4, 5, 6, 7, 5, 6, 7, 8, 6, 7, 8, 9, 7, 8, 9, 10<br>
+<br>
+tab_c_512: times 8 dw 512<br>
+<br>
+SECTION .text<br>
+<br>
+%macro FILTER_H4_w48 4<br>
+ movu %1, [srcq - 1]<br>
+ pshufb %2, %1, Tm0<br>
+ pmaddubsw %2, coef2<br>
+ pshufb %1, %1, Tm1<br>
+ pmaddubsw %1, coef2<br>
+ phaddw %2, %1<br>
+ movu %1, [srcq - 1 + 8]<br>
+ pshufb %4, %1, Tm0<br>
+ pmaddubsw %4, coef2<br>
+ pshufb %1, %1, Tm1<br>
+ pmaddubsw %1, coef2<br>
+ phaddw %4, %1<br>
+ pmulhrsw %2, %3<br>
+ pmulhrsw %4, %3<br>
+ packuswb %2, %4<br>
+ movu [dstq], %2<br>
+ movu %1, [srcq - 1 + 16]<br>
+ pshufb %2, %1, Tm0<br>
+ pmaddubsw %2, coef2<br>
+ pshufb %1, %1, Tm1<br>
+ pmaddubsw %1, coef2<br>
+ phaddw %2, %1<br>
+ movu %1, [srcq - 1 + 24]<br>
+ pshufb %4, %1, Tm0<br>
+ pmaddubsw %4, coef2<br>
+ pshufb %1, %1, Tm1<br>
+ pmaddubsw %1, coef2<br>
+ phaddw %4, %1<br>
+ pmulhrsw %2, %3<br>
+ pmulhrsw %4, %3<br>
+ packuswb %2, %4<br>
+ movu [dstq + 16], x1<br>
+ movu %1, [srcq - 1 + 32]<br>
+ pshufb %2, %1, Tm0<br>
+ pmaddubsw %2, coef2<br>
+ pshufb %1, %1, Tm1<br>
+ pmaddubsw %1, coef2<br>
+ phaddw %2, %1<br>
+ movu %1, [srcq - 1 + 40]<br>
+ pshufb %4, %1, Tm0<br>
+ pmaddubsw %4, coef2<br>
+ pshufb %1, %1, Tm1<br>
+ pmaddubsw %1, coef2<br>
+ phaddw %4, %1<br>
+ pmulhrsw %2, %3<br>
+ pmulhrsw %4, %3<br>
+ packuswb %2, %4<br>
+%endmacro<br>
+<br>
+%macro FILTER_H4_w48_CALL 0<br>
+ FILTER_H4_w48 x0, x1, x2, x3<br>
+<br>
+ movu [dstq + 32], x1<br>
+<br>
+ add srcq, srcstrideq<br>
+ add dstq, dststrideq<br>
+%endmacro<br>
+<br>
+;-----------------------------------------------------------------------------<br>
+; void filterHorizontal_p_p_4(pixel *src, intptr_t srcStride, pixel *dst, intptr_t dstStride, int width, int height, short const *coeff)<br>
+;-----------------------------------------------------------------------------<br>
+INIT_XMM sse4<br>
+cglobal filterHorizontal_p_p_4, 4, 5, 6, src, srcstride, dst, dststride<br>
+%define coef2 m6<br>
+%define Tm0 m5<br>
+%define Tm1 m4<br>
+%define x3 m3<br>
+%define x2 m2<br>
+%define x1 m1<br>
+%define x0 m0<br>
+<br>
+ mov r4, r6m<br>
+ movu coef2, [r4]<br>
+ packsswb coef2, coef2<br>
+ pshufd coef2, coef2, 0<br>
+<br>
+ mova x2, [tab_c_512]<br>
+<br>
+ mova Tm0, [tab_Tm]<br>
+ mova Tm1, [tab_Tm + 16]<br>
+<br>
+ %rep 47<br>
+ FILTER_H4_w48_CALL<br>
+ %endrep<br>
+<br>
+ FILTER_H4_w48 x0, x1, x2, x3<br>
+ movu [dstq + 32], x1<br>
+ RET<br>
_______________________________________________<br>
x265-devel mailing list<br>
<a href="mailto:x265-devel@videolan.org">x265-devel@videolan.org</a><br>
<a href="https://mailman.videolan.org/listinfo/x265-devel" target="_blank">https://mailman.videolan.org/listinfo/x265-devel</a><br>
</blockquote></div><br><br clear="all"><div><br></div>-- <br>Steve Borho
</div></div>