[x265] [PATCH 1 of 5] asm: interp_4tap_vert_pp sse2

dtyx265 at gmail.com dtyx265 at gmail.com
Fri May 8 22:57:12 CEST 2015


# HG changeset patch
# User David T Yuen <dtyx265 at gmail.com>
# Date 1431115282 25200
# Node ID 59e2e1b446e14adc8a7dae184bae9b241cc7a2f1
# Parent  57fce553135247caaafd415458d457d7f8696cd0
asm: interp_4tap_vert_pp sse2

This replaces c code for 2x4, 2x8 and 2x16

64-bit

./test/TestBench --testbench interp | grep vpp
chroma_vpp[  2x4]	1.80x 	 644.93   	 1159.95
chroma_vpp[  2x8]	1.72x 	 1204.84  	 2067.43
chroma_vpp[ 2x16]	1.99x 	 2252.45  	 4480.33

32-bit

./test/TestBench --testbench interp | grep vpp
chroma_vpp[  2x4]	1.80x 	 822.45   	 1479.92
chroma_vpp[  2x8]	1.90x 	 1477.46  	 2807.42
chroma_vpp[ 2x16]	2.26x 	 2770.04  	 6247.41

diff -r 57fce5531352 -r 59e2e1b446e1 source/common/x86/asm-primitives.cpp
--- a/source/common/x86/asm-primitives.cpp	Thu May 07 14:58:39 2015 +0530
+++ b/source/common/x86/asm-primitives.cpp	Fri May 08 13:01:22 2015 -0700
@@ -1370,6 +1370,9 @@
         CHROMA_420_VSP_FILTERS(_sse2);
         CHROMA_422_VSP_FILTERS(_sse2);
         CHROMA_444_VSP_FILTERS(_sse2);
+        p.chroma[X265_CSP_I420].pu[CHROMA_420_2x4].filter_vpp = x265_interp_4tap_vert_pp_2x4_sse2;
+        p.chroma[X265_CSP_I420].pu[CHROMA_420_2x8].filter_vpp = x265_interp_4tap_vert_pp_2x8_sse2;
+        p.chroma[X265_CSP_I422].pu[CHROMA_422_2x16].filter_vpp = x265_interp_4tap_vert_pp_2x16_sse2;
 
         ALL_LUMA_PU(luma_hpp, interp_8tap_horiz_pp, sse2);
         p.pu[LUMA_4x4].luma_hpp = x265_interp_8tap_horiz_pp_4x4_sse2;
diff -r 57fce5531352 -r 59e2e1b446e1 source/common/x86/ipfilter8.asm
--- a/source/common/x86/ipfilter8.asm	Thu May 07 14:58:39 2015 +0530
+++ b/source/common/x86/ipfilter8.asm	Fri May 08 13:01:22 2015 -0700
@@ -1042,6 +1042,135 @@
     IPFILTER_LUMA_sse2 64, 16, ps
     IPFILTER_LUMA_sse2 16, 64, ps
 
+%macro  WORD_TO_DOUBLE 1
+%if ARCH_X86_64
+    punpcklbw   %1,     m8
+%else
+    punpcklbw   %1,     %1
+    psrlw       %1,     8
+%endif
+%endmacro
+
+;-----------------------------------------------------------------------------
+; void interp_4tap_vert_pp_2xn(pixel *src, intptr_t srcStride, pixel *dst, intptr_t dstStride, int coeffIdx)
+;-----------------------------------------------------------------------------
+%macro FILTER_V4_W2_H4_sse2 1
+INIT_XMM sse2
+%if ARCH_X86_64
+cglobal interp_4tap_vert_pp_2x%1, 4, 6, 9
+    pxor        m8,        m8
+%else
+cglobal interp_4tap_vert_pp_2x%1, 4, 6, 8
+%endif
+    mov         r4d,       r4m
+    sub         r0,        r1
+
+%ifdef PIC
+    lea         r5,        [tabw_ChromaCoeff]
+    movh        m0,        [r5 + r4 * 8]
+%else
+    movh        m0,        [tabw_ChromaCoeff + r4 * 8]
+%endif
+
+    punpcklqdq  m0,        m0
+    mova        m1,        [pw_32]
+    lea         r5,        [3 * r1]
+
+%assign x 1
+%rep %1/4
+    movd        m2,        [r0]
+    movd        m3,        [r0 + r1]
+    movd        m4,        [r0 + 2 * r1]
+    movd        m5,        [r0 + r5]
+
+    punpcklbw   m2,        m3
+    punpcklbw   m6,        m4,        m5
+    punpcklwd   m2,        m6
+
+    WORD_TO_DOUBLE         m2
+    pmaddwd     m2,        m0
+
+    lea         r0,        [r0 + 4 * r1]
+    movd        m6,        [r0]
+
+    punpcklbw   m3,        m4
+    punpcklbw   m7,        m5,        m6
+    punpcklwd   m3,        m7
+
+    WORD_TO_DOUBLE         m3
+    pmaddwd     m3,        m0
+
+    packssdw    m2,        m3
+    pshuflw     m3,        m2,          q2301
+    pshufhw     m3,        m3,          q2301
+    paddw       m2,        m3
+    psrld       m2,        16
+
+    movd        m7,        [r0 + r1]
+
+    punpcklbw   m4,        m5
+    punpcklbw   m3,        m6,        m7
+    punpcklwd   m4,        m3
+
+    WORD_TO_DOUBLE         m4
+    pmaddwd     m4,        m0
+
+    movd        m3,        [r0 + 2 * r1]
+
+    punpcklbw   m5,        m6
+    punpcklbw   m7,        m3
+    punpcklwd   m5,        m7
+
+    WORD_TO_DOUBLE         m5
+    pmaddwd     m5,        m0
+
+    packssdw    m4,        m5
+    pshuflw     m5,        m4,          q2301
+    pshufhw     m5,        m5,          q2301
+    paddw       m4,        m5
+    psrld       m4,        16
+
+    packssdw    m2,        m4
+    paddw       m2,        m1
+    psraw       m2,        6
+    packuswb    m2,        m2
+
+%if ARCH_X86_64
+    movq        r4,        m2
+    mov         [r2],      r4w
+    shr         r4,        16
+    mov         [r2 + r3], r4w
+    lea         r2,        [r2 + 2 * r3]
+    shr         r4,        16
+    mov         [r2],      r4w
+    shr         r4,        16
+    mov         [r2 + r3], r4w
+%else
+    movd        r4,        m2
+    mov         [r2],      r4w
+    shr         r4,        16
+    mov         [r2 + r3], r4w
+    lea         r2,        [r2 + 2 * r3]
+    psrldq      m2,        4
+    movd        r4,        m2
+    mov         [r2],      r4w
+    shr         r4,        16
+    mov         [r2 + r3], r4w
+%endif
+
+%if x < %1/4
+    lea         r2,        [r2 + 2 * r3]
+%endif
+%assign x x+1
+%endrep
+    RET
+
+%endmacro
+
+    FILTER_V4_W2_H4_sse2 4
+    FILTER_V4_W2_H4_sse2 8
+    FILTER_V4_W2_H4_sse2 16
+
 ;-----------------------------------------------------------------------------
 ; void interp_4tap_horiz_pp_2x4(pixel *src, intptr_t srcStride, pixel *dst, intptr_t dstStride, int coeffIdx)
 ;-----------------------------------------------------------------------------
diff -r 57fce5531352 -r 59e2e1b446e1 source/common/x86/ipfilter8.h
--- a/source/common/x86/ipfilter8.h	Thu May 07 14:58:39 2015 +0530
+++ b/source/common/x86/ipfilter8.h	Fri May 08 13:01:22 2015 -0700
@@ -905,6 +905,9 @@
 void x265_interp_8tap_horiz_ps_64x48_sse2(const pixel* src, intptr_t srcStride, int16_t* dst, intptr_t dstStride, int coeffIdx, int isRowExt);
 void x265_interp_8tap_horiz_ps_64x64_sse2(const pixel* src, intptr_t srcStride, int16_t* dst, intptr_t dstStride, int coeffIdx, int isRowExt);
 void x265_interp_8tap_hv_pp_8x8_sse3(const pixel* src, intptr_t srcStride, pixel* dst, intptr_t dstStride, int idxX, int idxY);
+void x265_interp_4tap_vert_pp_2x4_sse2(const pixel *src, intptr_t srcStride, pixel *dst, intptr_t dstStride, int coeffIdx);
+void x265_interp_4tap_vert_pp_2x8_sse2(const pixel *src, intptr_t srcStride, pixel *dst, intptr_t dstStride, int coeffIdx);
+void x265_interp_4tap_vert_pp_2x16_sse2(const pixel *src, intptr_t srcStride, pixel *dst, intptr_t dstStride, int coeffIdx);
 #undef LUMA_FILTERS
 #undef LUMA_SP_FILTERS
 #undef LUMA_SS_FILTERS


More information about the x265-devel mailing list