yes. It looks like we need to optimize sse2 code. I will work on this. On Fri, Jun 26, 2015 at 5:31 PM, Praveen Tiwari < prav...@multicorewareinc.com> wrote:
> > > > ---------- Forwarded message ---------- > From: <raj...@multicorewareinc.com> > Date: Fri, Jun 26, 2015 at 3:14 PM > Subject: [x265] [PATCH] asm: pixelavg_pp[8xN] avx2 code for 10bpp > To: x265-devel@videolan.org > > > # HG changeset patch > # User Rajesh Paulraj<raj...@multicorewareinc.com> > # Date 1435311076 -19800 > # Fri Jun 26 15:01:16 2015 +0530 > # Node ID 956401f1a679f1e71181b704d64e4acdb6f1a93f > # Parent d64227e54233d1646c55bcb4b0b831e5340009ed > asm: pixelavg_pp[8xN] avx2 code for 10bpp > > avx2: > avg_pp[ 8x4] 4.39x 145.09 636.75 > avg_pp[ 8x8] 5.33x 215.27 1146.55 > avg_pp[ 8x16] 6.50x 336.88 2190.68 > avg_pp[ 8x32] 7.71x 579.86 4470.84 > > sse2: > avg_pp[ 8x4] 2.31x 287.63 663.94 > avg_pp[ 8x8] 3.26x 370.21 1205.26 > avg_pp[ 8x16] 3.99x 581.63 2323.25 > avg_pp[ 8x32] 4.78x 995.79 4755.58 > > > Basically, our macro "pixel_avg_8xN" just SSE (just simple syntax > conversion for avx2, not using 256 bit capability) so, fundamentally there > should be no major improvement in speed. But improvements 287.63c > -> 145.09c, 370.21c -> 215.27 etc are quite good. Does it means SSE2 codes > are not optimize well ? Can you revisit SSE code using this algorithm? > > > > diff -r d64227e54233 -r 956401f1a679 source/common/x86/asm-primitives.cpp > --- a/source/common/x86/asm-primitives.cpp Thu Jun 25 16:25:51 2015 > +0530 > +++ b/source/common/x86/asm-primitives.cpp Fri Jun 26 15:01:16 2015 > +0530 > @@ -1362,6 +1362,10 @@ > p.cu[BLOCK_32x32].intra_pred[33] = > PFX(intra_pred_ang32_33_avx2); > p.cu[BLOCK_32x32].intra_pred[34] = > PFX(intra_pred_ang32_2_avx2); > > + p.pu[LUMA_8x4].pixelavg_pp = PFX(pixel_avg_8x4_avx2); > + p.pu[LUMA_8x8].pixelavg_pp = PFX(pixel_avg_8x8_avx2); > + p.pu[LUMA_8x16].pixelavg_pp = PFX(pixel_avg_8x16_avx2); > + p.pu[LUMA_8x32].pixelavg_pp = PFX(pixel_avg_8x32_avx2); > p.pu[LUMA_12x16].pixelavg_pp = PFX(pixel_avg_12x16_avx2); > p.pu[LUMA_16x4].pixelavg_pp = PFX(pixel_avg_16x4_avx2); > p.pu[LUMA_16x8].pixelavg_pp = PFX(pixel_avg_16x8_avx2); > diff -r d64227e54233 -r 956401f1a679 source/common/x86/mc-a.asm > --- a/source/common/x86/mc-a.asm Thu Jun 25 16:25:51 2015 +0530 > +++ b/source/common/x86/mc-a.asm Fri Jun 26 15:01:16 2015 +0530 > @@ -4490,6 +4490,88 @@ > RET > %endif > > +%macro pixel_avg_W8 0 > + movu xm0, [r2] > + movu xm1, [r4] > + pavgw xm0, xm1 > + movu [r0], xm0 > + movu xm2, [r2 + r3] > + movu xm3, [r4 + r5] > + pavgw xm2, xm3 > + movu [r0 + r1], xm2 > + > + movu xm0, [r2 + r3 * 2] > + movu xm1, [r4 + r5 * 2] > + pavgw xm0, xm1 > + movu [r0 + r1 * 2], xm0 > + movu xm2, [r2 + r6] > + movu xm3, [r4 + r7] > + pavgw xm2, xm3 > + movu [r0 + r8], xm2 > + > + lea r0, [r0 + 4 * r1] > + lea r2, [r2 + 4 * r3] > + lea r4, [r4 + 4 * r5] > +%endmacro > + > > +;------------------------------------------------------------------------------------------------------------------------------- > +;void pixelavg_pp(pixel dst, intptr_t dstride, const pixel src0, intptr_t > sstride0, const pixel* src1, intptr_t sstride1, int) > > +;------------------------------------------------------------------------------------------------------------------------------- > +%if ARCH_X86_64 > +INIT_YMM avx2 > +cglobal pixel_avg_8x4, 6,10,4 > + add r1d, r1d > + add r3d, r3d > + add r5d, r5d > + lea r6, [r3 * 3] > + lea r7, [r5 * 3] > + lea r8, [r1 * 3] > + pixel_avg_W8 > + RET > + > +cglobal pixel_avg_8x8, 6,10,4 > + add r1d, r1d > + add r3d, r3d > + add r5d, r5d > + lea r6, [r3 * 3] > + lea r7, [r5 * 3] > + lea r8, [r1 * 3] > + mov r9d, 2 > +.loop > + pixel_avg_W8 > + dec r9d > + jnz .loop > + RET > + > +cglobal pixel_avg_8x16, 6,10,4 > + add r1d, r1d > + add r3d, r3d > + add r5d, r5d > + lea r6, [r3 * 3] > + lea r7, [r5 * 3] > + lea r8, [r1 * 3] > + mov r9d, 4 > +.loop > + pixel_avg_W8 > + dec r9d > + jnz .loop > + RET > + > +cglobal pixel_avg_8x32, 6,10,4 > + add r1d, r1d > + add r3d, r3d > + add r5d, r5d > + lea r6, [r3 * 3] > + lea r7, [r5 * 3] > + lea r8, [r1 * 3] > + mov r9d, 8 > +.loop > + pixel_avg_W8 > + dec r9d > + jnz .loop > + RET > +%endif > + > %macro pixel_avg_H4 0 > movu m0, [r2] > movu m1, [r4] > _______________________________________________ > x265-devel mailing list > x265-devel@videolan.org > https://mailman.videolan.org/listinfo/x265-devel > > > _______________________________________________ > x265-devel mailing list > x265-devel@videolan.org > https://mailman.videolan.org/listinfo/x265-devel > >
_______________________________________________ x265-devel mailing list x265-devel@videolan.org https://mailman.videolan.org/listinfo/x265-devel
