[x265] Fwd: [PATCH] asm: pixelavg_pp[8xN] avx2 code for 10bpp

Praveen Tiwari praveen at multicorewareinc.com
Fri Jun 26 12:19:13 CEST 2015


ahh, width is just 8*16 = 128, two rows at a time will need vextracti128 as
well while storing, which goes to port5, a bottleneck port. pavgw is much
cheaper than it. You may try to combine 16XN sizes.

Regards,
Praveen

On Fri, Jun 26, 2015 at 3:40 PM, Rajesh Paulraj <rajesh at multicorewareinc.com
> wrote:

> I tried using vinserti128. But that reduces the performance than this one.
> So i kept this version.
>
> On Fri, Jun 26, 2015 at 3:37 PM, Praveen Tiwari <
> praveen at multicorewareinc.com> wrote:
>
>>
>>
>>
>> ---------- Forwarded message ----------
>> From: <rajesh at 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 at videolan.org
>>
>>
>> # HG changeset patch
>> # User Rajesh Paulraj<rajesh at 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
>>
>> 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
>> +
>> >> Your macro is not using avx2 capabilities, did you check the
>> performance of two rows combined ? It will reduce your  pavgw and movu
>> instruction by half. You can use vinserti128 to combine two rows at a
>> time.
>>
>> +    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 at videolan.org
>> https://mailman.videolan.org/listinfo/x265-devel
>>
>>
>> _______________________________________________
>> x265-devel mailing list
>> x265-devel at videolan.org
>> https://mailman.videolan.org/listinfo/x265-devel
>>
>>
>
> _______________________________________________
> x265-devel mailing list
> x265-devel at videolan.org
> https://mailman.videolan.org/listinfo/x265-devel
>
>
-------------- next part --------------
An HTML attachment was scrubbed...
URL: <http://mailman.videolan.org/pipermail/x265-devel/attachments/20150626/1dc3fe25/attachment.html>


More information about the x265-devel mailing list