[x265] [PATCH 2 of 2] asm:intra pred planar32 sse2 high bit

dave dtyx265 at gmail.com
Tue Mar 10 02:24:02 CET 2015


On 03/09/2015 04:34 PM, chen wrote:
>
> At 2015-03-10 07:31:13,dave <dtyx265 at gmail.com> wrote:
>
>     On 03/09/2015 04:09 PM, chen wrote:
>>
>>     At 2015-03-10 07:06:26,dave <dtyx265 at gmail.com> wrote:
>>
>>         On 03/09/2015 03:23 PM, chen wrote:
>>>
>>>         At 2015-03-10 05:24:09,dtyx265 at gmail.com  wrote:
>>>         ># HG changeset patch
>>>         ># User David T Yuen<dtyx265 at gmail.com>
>>>         ># Date 1425935926 25200
>>>         ># Node ID 4d1d54d28cb1635448ceec764c793d1b37cef7a4
>>>         ># Parent  ef383507c21ca704b58e759e440f4ae2e177c499
>>>         >asm:intra pred planar32 sse2 high bit
>>>         >
>>>         >This replaces c code for systems using ssse3 to sse2 processors
>>>         >The code is backported from intrapred planar32 sse4
>>>         >
>>>         >Unlike the sse4 high bit version which operates with 32-bit values
>>>         >this version uses 16-bits to calculate the prediction.  16-bits
>>>         >is just enough for 10 bit luma depth, or at least it's enough for
>>>         >the testbench.  Anything more and this primitive could overflow.
>>>         >
>>>         >./test/TestBench --testbench intrapred | grep intra_planar_32x32
>>>         >intra_planar_32x32	13.80x 	 8289.69  	 114379.01
>>>         >
>>>         >diff -r ef383507c21c -r 4d1d54d28cb1 source/common/x86/asm-primitives.cpp
>>>         >--- a/source/common/x86/asm-primitives.cpp	Mon Mar 09 13:50:57 2015 -0700
>>>         >+++ b/source/common/x86/asm-primitives.cpp	Mon Mar 09 14:18:46 2015 -0700
>>>         >@@ -876,6 +876,7 @@
>>>         >         p.cu[BLOCK_4x4].intra_pred[PLANAR_IDX] = x265_intra_pred_planar4_sse2;
>>>         >         p.cu[BLOCK_8x8].intra_pred[PLANAR_IDX] = x265_intra_pred_planar8_sse2;
>>>         >         p.cu[BLOCK_16x16].intra_pred[PLANAR_IDX] = x265_intra_pred_planar16_sse2;
>>>         >+        p.cu[BLOCK_32x32].intra_pred[PLANAR_IDX] = x265_intra_pred_planar32_sse2;
>>>         >
>>>         >         p.cu[BLOCK_4x4].sse_ss = x265_pixel_ssd_ss_4x4_mmx2;
>>>         >         ALL_LUMA_CU(sse_ss, pixel_ssd_ss, sse2);
>>>         >diff -r ef383507c21c -r 4d1d54d28cb1 source/common/x86/intrapred16.asm
>>>         >--- a/source/common/x86/intrapred16.asm	Mon Mar 09 13:50:57 2015 -0700
>>>         >+++ b/source/common/x86/intrapred16.asm	Mon Mar 09 14:18:46 2015 -0700
>>>         >@@ -65,6 +65,10 @@
>>>         > pw_planar16_1:        dw 15, 15, 15, 15, 15, 15, 15, 15
>>>         > pd_planar32_1:        dd 31, 31, 31, 31
>>>         >
>>>         >+pw_planar32_1:        dw 31, 31, 31, 31, 31, 31, 31, 31
>>>         >+pw_planar32_L:        dw 31, 30, 29, 28, 27, 26, 25, 24
>>>         >+pw_planar32_H:        dw 23, 22, 21, 20, 19, 18, 17, 16
>>>         >+
>>>         > const planar32_table
>>>         > %assign x 31
>>>         > %rep 8
>>>         >@@ -86,12 +90,15 @@
>>>         > cextern pw_4
>>>         > cextern pw_8
>>>         > cextern pw_16
>>>         >+cextern pw_32
>>>         > cextern pw_1023
>>>         > cextern pd_16
>>>         > cextern pd_32
>>>         > cextern pw_4096
>>>         > cextern multiL
>>>         > cextern multiH
>>>         >+cextern multiH2
>>>         >+cextern multiH3
>>>         > cextern multi_2Row
>>>         > cextern pw_swap
>>>         > cextern pb_unpackwq1
>>>         >@@ -575,6 +582,133 @@
>>>         >     INTRA_PRED_PLANAR_16 15
>>>         >     RET
>>>         >
>>>         >+;---------------------------------------------------------------------------------------
>>>         >+; void intra_pred_planar(pixel* dst, intptr_t dstStride, pixel*srcPix, int, int filter)
>>>         >+;---------------------------------------------------------------------------------------
>>>         >+INIT_XMM sse2
>>>         >+cglobal intra_pred_planar32, 3,3,13
>>>         >+    movd            m3, [r2 + 66]               ; topRight   = above[32]
>>>         >+
>>>         >+    pshuflw         m3, m3, 0x00
>>>         >+    pshufd          m3, m3, 0x44
>>>         >+
>>>         >+    pmullw          m0, m3, [multiL]            ; (x + 1) * topRight
>>>         >+    pmullw          m1, m3, [multiH]            ; (x + 1) * topRight
>>>         >+    pmullw          m2, m3, [multiH2]           ; (x + 1) * topRight
>>>         >+    pmullw          m3, [multiH3]               ; (x + 1) * topRight
>>>         >+
>>>         >+    movd            m6, [r2 + 194]               ; bottomLeft = left[32]
>>>         >+    pshuflw         m6, m6, 0x00
>>>         >+    pshufd          m6, m6, 0x44
>>>         >+    mova            m5, m6
>>>         >+    paddw           m5, [pw_32]
>>>         >+
>>>         >+    paddw           m0, m5
>>>         >+    paddw           m1, m5
>>>         >+    paddw           m2, m5
>>>         >+    paddw           m3, m5
>>>         >+    mova            m8, m6
>>>         >+    mova            m9, m6
>>>         >+    mova            m10, m6
>>>         >+
>>>         >+    movu            m4, [r2 + 2]
>>>         >+    psubw           m8, m4
>>>         >+    pmullw          m4, [pw_planar32_1]
>>>         >+    paddw           m0, m4
>>>         >+
>>>         >+    movu            m5, [r2 + 18]
>>>         >+    psubw           m9, m5
>>>         >+    pmullw          m5, [pw_planar32_1]
>>>         >+    paddw           m1, m5
>>>         >+
>>>         >+    movu            m4, [r2 + 34]
>>>         >+    psubw           m10, m4
>>>         >+    pmullw          m4, [pw_planar32_1]
>>>         >+    paddw           m2, m4
>>>         >+
>>>         >+    movu            m5, [r2 + 50]
>>>         >+    psubw           m6, m5
>>>         >+    pmullw          m5, [pw_planar32_1]
>>>         the constant use more than 2 times, need buffer into register
>>>         >+    paddw           m3, m5
>>>         >+
>>>         >+%macro PROCESS 1
>>>         >+    pmullw          m5, %1, [pw_planar32_L]
>>>         >+    pmullw          m11, %1, [pw_planar32_H]
>>>         buffer into register
>>
>>         Buffering these constants in registers slows performance on
>>         my machine.
>>
>>         ./test/TestBench --testbench intrapred | grep intra_planar_32x32
>>         intra_planar_32x32    12.01x      9521.82  114371.95
>>
>>         And no, I didn't set the registers inside the PROCESS macro...
>>
>>         [MC] your code work on x64 platform, you have at least two
>>         free register to buffer these constant, it reduce code size
>>         and improve performance
>>
>         mova            m13, [pw_planar32_L]
>         mova            m14, [pw_planar32_H]
>         mova            m15, [pw_planar16_0]
>     %macro PROCESS 1
>         pmullw          m5, %1, m13 ;[pw_planar32_L]
>         pmullw          m11, %1, m14 ;[pw_planar32_H]
>         paddw           m5, m0
>         paddw           m11, m1
>         psrlw           m5, 6
>         psrlw           m11, 6
>         movu            [r0], m5
>         movu            [r0 + 16], m11
>
>         pmullw          m5, %1, m15 ;[pw_planar16_0]
>
>     This runs slower
>
>     [MC] just 4 cycles? could you test the really encoder speed? the
>     really have lower cache hit rate.
>
Interesting.  Performance is almost identical

original code

/x265 -I 1 --input ~/Videos/bridge-close-cif/bridge-close.y4m -o 
bridge-close.y4m
y4m  [info]: 352x288 fps 30/1 i420p8 frames 0 - 1999 of 2000
x265 [info]: HEVC encoder version 1.5+162-4d1d54d28cb1
x265 [info]: build info [Linux][GCC 4.7.2][64 bit] 16bpp
x265 [info]: using cpu capabilities: MMX2 SSE2Slow SlowCTZ
x265 [info]: Main 10 profile, Level-2 (Main tier)
x265 [info]: Thread pool created using 2 threads
x265 [info]: frame threads / pool features       : 1 / wpp(5 rows)
x265 [info]: Internal bit depth                  : 10
x265 [info]: Coding QT: max CU size, min CU size : 64 / 8
x265 [info]: Residual QT: max TU size, max depth : 32 / 1 inter / 1 intra
x265 [info]: ME / range / subpel / merge         : hex / 57 / 2 / 2
x265 [info]: Keyframe min / max / scenecut       : 1 / 1 / 40
x265 [info]: Lookahead / bframes / badapt        : 20 / 0 / 0
x265 [info]: b-pyramid / weightp / weightb / refs: 0 / 1 / 0 / 3
x265 [info]: Rate Control / AQ-Strength / CUTree : CRF-28.0 / 1.0 / 1
x265 [info]: tools: rd=3 psy-rd=0.30 deblock sao signhide tmvp
x265 [info]: frame I:   2000, Avg QP:33.08  kb/s: 973.45
x265 [info]: global :   2000, Avg QP:33.08  kb/s: 973.45
x265 [info]: consecutive B-frames: 100.0%

encoded 2000 frames in 414.26s (4.83 fps), 973.45 kb/s

and using registers to hold constants

./x265 -I 1 --input ~/Videos/bridge-close-cif/bridge-close.y4m -o 
bridge-close.y4m
y4m  [info]: 352x288 fps 30/1 i420p8 frames 0 - 1999 of 2000
x265 [info]: HEVC encoder version 1.5+162-4d1d54d28cb1
x265 [info]: build info [Linux][GCC 4.7.2][64 bit] 16bpp
x265 [info]: using cpu capabilities: MMX2 SSE2Slow SlowCTZ
x265 [info]: Main 10 profile, Level-2 (Main tier)
x265 [info]: Thread pool created using 2 threads
x265 [info]: frame threads / pool features       : 1 / wpp(5 rows)
x265 [info]: Internal bit depth                  : 10
x265 [info]: Coding QT: max CU size, min CU size : 64 / 8
x265 [info]: Residual QT: max TU size, max depth : 32 / 1 inter / 1 intra
x265 [info]: ME / range / subpel / merge         : hex / 57 / 2 / 2
x265 [info]: Keyframe min / max / scenecut       : 1 / 1 / 40
x265 [info]: Lookahead / bframes / badapt        : 20 / 0 / 0
x265 [info]: b-pyramid / weightp / weightb / refs: 0 / 1 / 0 / 3
x265 [info]: Rate Control / AQ-Strength / CUTree : CRF-28.0 / 1.0 / 1
x265 [info]: tools: rd=3 psy-rd=0.30 deblock sao signhide tmvp
x265 [info]: frame I:   2000, Avg QP:33.08  kb/s: 973.45
x265 [info]: global :   2000, Avg QP:33.08  kb/s: 973.45
x265 [info]: consecutive B-frames: 100.0%

encoded 2000 frames in 414.28s (4.83 fps), 973.45 kb/s

The closest I could find to forcing intra planar32 to be used is -I 1
>
>
>     ./test/TestBench --testbench intrapred | grep intra_planar_32x32
>     intra_planar_32x32    12.04x      9517.47       114580.26
>
>     over 1200 worse
>
>>
>>>         >+    paddw           m5, m0
>>>         >+    paddw           m11, m1
>>>         >+    psrlw           m5, 6
>>>         >+    psrlw           m11, 6
>>>         >+    movu            [r0], m5
>>>         >+    movu            [r0 + 16], m11
>>>         >+
>>>         >+    pmullw          m5, %1, [pw_planar16_0]
>>>         >+    pmullw          %1, [pw_planar8_0]
>>>         >+    paddw           m5, m2
>>>         >+    paddw           %1, m3
>>>         >+    psrlw           m5, 6
>>>         >+    psrlw           %1, 6
>>>         >+    movu            [r0 + 32], m5
>>>         >+    movu            [r0 + 48], %1
>>>         >+%endmacro
>>>         >+
>>>         >+%macro  INCREMENT 0
>>>         >+    paddw           m2, m10
>>>         >+    paddw           m3, m6
>>>         >+    paddw           m0, m8
>>>         >+    paddw           m1, m9
>>>         >+    lea             r0, [r0 + r1 * 2]
>>>         we have free register, so may buffer r1*2 and replace by ADD to avoid execute port bottleneck
>>         This provides a small performance improvement, I will submit
>>         a new patch.
>>>         >+%endmacro
>>>         >+
>>>         >+%assign x 0
>>>         >+%rep 4
>>>         >+    movu            m4, [r2 + 130 + x * 16]     ;130 = 32*sizeof(pixel)*2 + 1*sizeof(pixel)
>>>         >+
>>>         >+    pshuflw         m7, m4, 0x00
>>>         >+    pshufd          m7, m7, 0x44
>>>         >+    pshuflw         m12, m4, 0x55
>>>         >+    pshufd          m12, m12, 0x44
>>>         >+
>>>         >+    PROCESS m7
>>>         >+    INCREMENT
>>>         >+    PROCESS m12
>>>         >+    INCREMENT
>>>         >+
>>>         >+    pshuflw         m7, m4, 0xAA
>>>         >+    pshufd          m7, m7, 0x44
>>>         >+    pshuflw         m12, m4, 0xFF
>>>         >+    pshufd          m12, m12, 0x44
>>>         >+
>>>         >+    PROCESS m7
>>>         >+    INCREMENT
>>>         >+    PROCESS m12
>>>         >+    INCREMENT
>>>         >+
>>>         >+    pshufhw         m7, m4, 0x00
>>>         >+    pshufd          m7, m7, 0xEE
>>>         >+    pshufhw         m12, m4, 0x55
>>>         >+    pshufd          m12, m12, 0xEE
>>>         >+
>>>         >+    PROCESS m7
>>>         >+    INCREMENT
>>>         >+    PROCESS m12
>>>         >+    INCREMENT
>>>         >+
>>>         >+    pshufhw         m7, m4, 0xAA
>>>         >+    pshufd          m7, m7, 0xEE
>>>         >+    pshufhw         m12, m4, 0xFF
>>>         >+    pshufd          m12, m12, 0xEE
>>>         >+
>>>         >+    PROCESS m7
>>>         >+    INCREMENT
>>>         >+    PROCESS m12
>>>         >+
>>>         >+    %if x < 3
>>>         >+    INCREMENT
>>>         >+    %endif
>>>         >+%assign x x+1
>>>         >+%endrep
>>>         >+    RET
>>>         >+
>>>         > ;-----------------------------------------------------------------------------------
>>>         > ; void intra_pred_dc(pixel* dst, intptr_t dstStride, pixel* above, int, int filter)
>>>         > ;-----------------------------------------------------------------------------------
>>
>
>
> _______________________________________________
> 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/20150309/b5fff37e/attachment-0001.html>


More information about the x265-devel mailing list