[x265] [PATCH] asm: intrapred dc8 sse2 high bit

[Steve Borho]

On 02/26, dave wrote:
> On 02/25/2015 05:59 AM, chen wrote:
> >At 2015-02-25 08:16:44,dtyx265 at gmail.com wrote:
> >># HG changeset patch
> >># User David T Yuen <dtyx265 at gmail.com>
> >># Date 1424823199 28800
> >># Node ID bd1d713da87bb1e3022c80f462398bd78a95ce48
> >># Parent  644d27ca0b197455393171ba705b1190f3d9b420
> >>asm: intrapred dc8 sse2 high bit
> >>
> >>This replaces c code for systems using ssse3 to sse2 processors
> >>The code is backported from intrapred dc8 sse4 high bit
> >>
> >>./test/TestBench --testbench intrapred | grep 8x8
> >>intra_dc_8x8[f=0]	1.62x 	 443.94   	 717.32
> >>intra_dc_8x8[f=1]	1.59x 	 729.98   	 1157.66
> >>
> >>diff -r 644d27ca0b19 -r bd1d713da87b source/common/x86/asm-primitives.cpp
> >>--- a/source/common/x86/asm-primitives.cpp	Tue Feb 24 14:47:01 2015 -0800
> >>+++ b/source/common/x86/asm-primitives.cpp	Tue Feb 24 16:13:19 2015 -0800
> >>@@ -864,6 +864,7 @@
> >>         ALL_LUMA_TU_S(transpose, transpose, sse2);
> >>
> >>         p.cu[BLOCK_4x4].intra_pred[DC_IDX] = x265_intra_pred_dc4_sse2;
> >>+        p.cu[BLOCK_8x8].intra_pred[DC_IDX] = x265_intra_pred_dc8_sse2;
> >>
> >>         p.cu[BLOCK_4x4].sse_ss = x265_pixel_ssd_ss_4x4_mmx2;
> >>         ALL_LUMA_CU(sse_ss, pixel_ssd_ss, sse2);
> >>diff -r 644d27ca0b19 -r bd1d713da87b source/common/x86/intrapred16.asm
> >>--- a/source/common/x86/intrapred16.asm	Tue Feb 24 14:47:01 2015 -0800
> >>+++ b/source/common/x86/intrapred16.asm	Tue Feb 24 16:13:19 2015 -0800
> >>@@ -86,6 +86,7 @@
> >> cextern pw_8
> >> cextern pw_16
> >> cextern pw_1023
> >>+cextern pd_2
> >> cextern pd_16
> >> cextern pd_32
> >> cextern pw_4096
> >>@@ -166,6 +167,89 @@
> >> ;-----------------------------------------------------------------------------------
> >> ; void intra_pred_dc(pixel* dst, intptr_t dstStride, pixel* above, int, int filter)
> >> ;-----------------------------------------------------------------------------------
> >>+INIT_XMM sse2
> >>+cglobal intra_pred_dc8, 5, 8, 2
> >>+    add             r2,            2
> >>+    add             r1,            r1
> >>+    movu            m0,            [r2 + 32]
> >>+    movu            m1,            [r2]
> >>+
> >>+    paddw           m0,            m1
> >>+    movhlps         m1,            m0
> >>+    paddw           m0,            m1
> >>+    pshufd          m1,            m0, 1
> >>+    paddw           m0,            m1
> >>+    pmaddwd         m0,            [pw_1]
> >>+
> >>+    paddw           m0,            [pw_8]
> >>+    psraw           m0,            4              ; sum = sum / 16
> >>+    pshuflw         m0,            m0, 0
> >>+    pshufd          m0,            m0, 0          ; m0 = word [dc_val ...]
> >>+
> >>+    test            r4d,           r4d
> >>+
> >>+    ; store DC 8x8
> >>+    lea             r6,            [r1 + r1 * 2]
> >>+    lea             r5,            [r6 + r1 * 2]
> >>+    lea             r7,            [r6 + r1 * 4]
> >>+    movu            [r0],          m0
> >>+    movu            [r0 + r1],     m0
> >>+    movu            [r0 + r1 * 2], m0
> >>+    movu            [r0 + r6],     m0
> >>+    movu            [r0 + r1 * 4], m0
> >>+    movu            [r0 + r5],     m0
> >>+    movu            [r0 + r6 * 2], m0
> >>+    movu            [r0 + r7],     m0
> >>+
> >>+    ; Do DC Filter
> >>+    jz              .end
> >>+    mova            m1,            [pd_2]
> >>+    pmullw          m1,            m0
> >[x x x x] * [0 2 0 2] = [0 dc*2 0 dc*2], am I right?
> >Why not PSLL?
> For some reason using psllw in place of pmullw hurts performance of
> unfiltered in 8 and 10 bit far more than you would expect for such a
> small change.  8 bit speed is cut by more than half
> 
> ./test/TestBench --testbench intrapred | grep 8x8
> 
> intra_dc_8x8[f=0]    1.84x      492.58        906.63
> intra_dc_8x8[f=1]    2.38x      537.68        1280.02
> 
>  and 10 bit is barely faster than c code.
> 
> ./test/TestBench --testbench intrapred | grep 8x8
> intra_dc_8x8[f=0]    1.06x      677.54        719.80
> intra_dc_8x8[f=1]    1.60x      725.05        1157.54
> 
> I will resubmit with other adjustments

You may find this useful, if you have not seen it before:

https://software.intel.com/en-us/articles/intel-architecture-code-analyzer

-- 
Steve Borho