[x265] [PATCH Review Only] luma_hps[32x8, 32x16, 32x24, 32x32, 32x64] avx2 asm : improved 6058c->3579c, 9208c->5449c, 12504c->7155c, 15974c->9537c, 28644c->16476c

chen chenm003 at 163.com
Tue Mar 3 16:20:35 CET 2015


updated 

At 2015-03-03 13:24:46,aasaipriya at multicorewareinc.com wrote:
># HG changeset patch
># User Aasaipriya Chandran <aasaipriya at multicorewareinc.com>
># Date 1425360254 -19800
>#      Tue Mar 03 10:54:14 2015 +0530
># Node ID 88824fdb34aefbe9fec3c1da8d30375e00fe7d14
># Parent  04861917b7b343216ac1b9a9f70e38396af53275
>luma_hps[32x8, 32x16, 32x24, 32x32, 32x64] avx2 asm : improved 6058c->3579c, 9208c->5449c, 12504c->7155c, 15974c->9537c, 28644c->16476c
>
>diff -r 04861917b7b3 -r 88824fdb34ae source/common/x86/asm-primitives.cpp
>--- a/source/common/x86/asm-primitives.cpp	Thu Feb 26 19:11:18 2015 +0530
>+++ b/source/common/x86/asm-primitives.cpp	Tue Mar 03 10:54:14 2015 +0530
>@@ -1786,6 +1786,12 @@
> 
>         p.pu[LUMA_48x64].luma_hpp = x265_interp_8tap_horiz_pp_48x64_avx2;
> 
>+        p.pu[LUMA_32x32].luma_hps = x265_interp_8tap_horiz_ps_32x32_avx2;
>+        p.pu[LUMA_32x16].luma_hps = x265_interp_8tap_horiz_ps_32x16_avx2;
>+        p.pu[LUMA_32x24].luma_hps = x265_interp_8tap_horiz_ps_32x24_avx2;
>+        p.pu[LUMA_32x8].luma_hps = x265_interp_8tap_horiz_ps_32x8_avx2;
>+        p.pu[LUMA_32x64].luma_hps = x265_interp_8tap_horiz_ps_32x64_avx2;
>+
>         p.chroma[X265_CSP_I420].pu[CHROMA_420_8x8].filter_hpp = x265_interp_4tap_horiz_pp_8x8_avx2;
>         p.chroma[X265_CSP_I420].pu[CHROMA_420_4x4].filter_hpp = x265_interp_4tap_horiz_pp_4x4_avx2;
>         p.chroma[X265_CSP_I420].pu[CHROMA_420_32x32].filter_hpp = x265_interp_4tap_horiz_pp_32x32_avx2;
>diff -r 04861917b7b3 -r 88824fdb34ae source/common/x86/ipfilter8.asm
>--- a/source/common/x86/ipfilter8.asm	Thu Feb 26 19:11:18 2015 +0530
>+++ b/source/common/x86/ipfilter8.asm	Tue Mar 03 10:54:14 2015 +0530
>@@ -1994,6 +1994,92 @@
> IPFILTER_LUMA_PS_16x_AVX2 16 , 32
> IPFILTER_LUMA_PS_16x_AVX2 16 , 64
> 
>+%macro IPFILTER_LUMA_PS_32x_AVX2 2
>+INIT_YMM avx2
>+cglobal interp_8tap_horiz_ps_%1x%2, 6, 7, 8
>+%ifdef PIC
>+    lea                         r6,                [tab_LumaCoeff]
>+    vpbroadcastq                m0,                [r6 + r4 * 8]
>+%else
>+    vpbroadcastq                m0,                [tab_LumaCoeff + r4 * 8]
>+%endif
>+    mova                        m6,                [tab_Lm + 32]
>+    mova                        m1,                [tab_Lm]
>+    mov                         r4,                %2                           ;height
>+    add                         r3d,               r3d
>+    vbroadcasti128              m2,                [pw_2000]
>+    mova                        m7,                [interp8_hps_shuf]
>+    ; register map
>+    ; m0      - interpolate coeff
>+    ; m1 , m6 - shuffle order table
>+    ; m2      - pw_2000
>+
>+    sub                         r0,                3
>+    test                        r5d,               r5d
>+    jz                          .label
>+    lea                         r6,                [r1 * 3]                     ; r8 = (N / 2 - 1) * srcStride
>+    sub                         r0,                r6
>+    ;sub                         r0,                [r1 * 3]                           ; r0(src)-[r1*3]
>+    add                         r4,                7                            ; blkheight += N - 1  (7 - 1 = 6 ; since the last one row not in loop)
>+
>+.label
use .loop is better

>+    ; Row 0
>+    vbroadcasti128              m3,                [r0]                         ; [x x x x x A 9 8 7 6 5 4 3 2 1 0]
>+    pshufb                      m4,                m3,             m6           ; row 0 (col 4 to 7)
>+    pshufb                      m3,                m1                           ; shuffled based on the col order tab_Lm row 0 (col 0 to 3)
>+    pmaddubsw                   m3,                m0
>+    pmaddubsw                   m4,                m0
>+    phaddw                      m3,                m4                           ; DWORD [R1D R1C R0D R0C R1B R1A R0B R0A]
PMADDWD is faster here

>+
>+    vbroadcasti128              m4,                [r0 + 8]
>+    pshufb                      m5,                m4,            m6            ;row 0 (col 12 to 15)
>+    pshufb                      m4,                m1                           ;row 0 (col 8 to 11)
>+    pmaddubsw                   m4,                m0
>+    pmaddubsw                   m5,                m0
>+    phaddw                      m4,                m5                           ; DWORD [R3D R3C R2D R2C R3B R3A R2B R2A]
>+
>+    phaddw                      m3,                m4                           ; col 0 - 16
>+    mova                        m5,                m7
>+    vpermd                      m3,                m5,               m3
>+    psubw                       m3,                m2
>+
>+    movu                        [r2],              m3                          ;row 0
>+
>+    vbroadcasti128              m3,                [r0 + 16]
>+    pshufb                      m4,                m3,             m6           ; row 0 (col 20 to 23)
>+    pshufb                      m3,                m1                           ; row 0 (col 16 to 19)
>+    pmaddubsw                   m3,                m0
>+    pmaddubsw                   m4,                m0
>+    phaddw                      m3,                m4                           ; DWORD [R1D R1C R0D R0C R1B R1A R0
>+
>+    vbroadcasti128              m4,                [r0 + 24]
>+    pshufb                      m5,                m4,            m6            ;row 0 (col 28 to 31)
>+    pshufb                      m4,                m1                           ;row 0 (col 24 to 27)
>+    pmaddubsw                   m4,                m0
>+    pmaddubsw                   m5,                m0
>+    phaddw                      m4,                m5                           ; DWORD [R3D R3C R2D R2C R3B R3A R2B R2
>+
>+    phaddw                      m3,                m4                           ; all rows and col completed.
>+    mova                        m5,                m7
>+    vpermd                      m3,                m5,               m3
>+    psubw                       m3,                m2
>+
>+    movu                        [r2 + 32],         m3                          ;row 0
>+
>+    lea                         r0,                [r0 + r1]                ; first loop src ->5th row(i.e 4)
>+    lea                         r2,                [r2 + r3]                ; first loop dst ->5th row(i.e 4)
ADD is execute in port6, it is faster

>+    dec                         r4d
>+    jnz                         .label
>+
>+RET
>+%endmacro
>+
>+
>+IPFILTER_LUMA_PS_32x_AVX2 32 , 32
>+IPFILTER_LUMA_PS_32x_AVX2 32 , 16
>+IPFILTER_LUMA_PS_32x_AVX2 32 , 24
>+IPFILTER_LUMA_PS_32x_AVX2 32 , 8
>+IPFILTER_LUMA_PS_32x_AVX2 32 , 64
> 
> ;--------------------------------------------------------------------------------------------------------------
> ; void interp_8tap_horiz_pp_%1x%2(pixel *src, intptr_t srcStride, pixel *dst, intptr_t dstStride, int coeffIdx)
>_______________________________________________
>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/20150303/172b5787/attachment-0001.html>


More information about the x265-devel mailing list