[x265] [PATCH 01 of 12] asm: interp_4tap_vert_ps_2xN sse2

dtyx265 at gmail.com dtyx265 at gmail.com
Mon May 18 04:48:52 CEST 2015


# HG changeset patch
# User David T Yuen <dtyx265 at gmail.com>
# Date 1431911009 25200
# Node ID 465fb4340a241e501b53a6241f5ae81c29ba073a
# Parent  8592bf81d0848279fa79cd1487406cb516dffe99
asm: interp_4tap_vert_ps_2xN sse2

Updated vert_pp_2xN macro to also create ps.  This replaces c code for ps with minimal impact on pp.

64-bit

/test/TestBench --testbench interp | grep vp | grep " 2x"
chroma_vpp[  2x4]	1.80x 	 644.93   	 1159.97
chroma_vps[  2x4]	1.42x 	 630.00   	 894.95
chroma_vpp[  2x8]	1.72x 	 1204.99  	 2067.47
chroma_vps[  2x8]	1.49x 	 1152.50  	 1712.50
chroma_vpp[ 2x16]	1.94x 	 2314.96  	 4482.00
chroma_vps[ 2x16]	1.91x 	 2222.45  	 4252.86

32-bit

./test/TestBench --testbench interp | grep vp | grep " 2x"
chroma_vpp[  2x4]	1.74x 	 849.94   	 1479.88
chroma_vps[  2x4]	1.64x 	 762.49   	 1247.46
chroma_vpp[  2x8]	1.89x 	 1482.47  	 2807.46
chroma_vps[  2x8]	1.93x 	 1392.49  	 2682.46
chroma_vpp[ 2x16]	2.26x 	 2769.98  	 6249.80
chroma_vps[ 2x16]	1.91x 	 2632.49  	 5028.81

diff -r 8592bf81d084 -r 465fb4340a24 source/common/x86/asm-primitives.cpp
--- a/source/common/x86/asm-primitives.cpp	Thu May 14 17:12:14 2015 +0530
+++ b/source/common/x86/asm-primitives.cpp	Sun May 17 18:03:29 2015 -0700
@@ -1448,6 +1448,9 @@
         p.chroma[X265_CSP_I444].pu[LUMA_4x4].filter_vpp = x265_interp_4tap_vert_pp_4x4_sse2;
         p.chroma[X265_CSP_I444].pu[LUMA_4x8].filter_vpp = x265_interp_4tap_vert_pp_4x8_sse2;
         p.chroma[X265_CSP_I444].pu[LUMA_4x16].filter_vpp = x265_interp_4tap_vert_pp_4x16_sse2;
+        p.chroma[X265_CSP_I420].pu[CHROMA_420_2x4].filter_vps = x265_interp_4tap_vert_ps_2x4_sse2;
+        p.chroma[X265_CSP_I420].pu[CHROMA_420_2x8].filter_vps = x265_interp_4tap_vert_ps_2x8_sse2;
+        p.chroma[X265_CSP_I422].pu[CHROMA_422_2x16].filter_vps = x265_interp_4tap_vert_ps_2x16_sse2;
 #if X86_64
         p.chroma[X265_CSP_I420].pu[CHROMA_420_6x8].filter_vpp = x265_interp_4tap_vert_pp_6x8_sse2;
         p.chroma[X265_CSP_I420].pu[CHROMA_420_8x2].filter_vpp = x265_interp_4tap_vert_pp_8x2_sse2;
diff -r 8592bf81d084 -r 465fb4340a24 source/common/x86/ipfilter8.asm
--- a/source/common/x86/ipfilter8.asm	Thu May 14 17:12:14 2015 +0530
+++ b/source/common/x86/ipfilter8.asm	Sun May 17 18:03:29 2015 -0700
@@ -878,19 +878,26 @@
 %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
+; void interp_4tap_vert_%1_2x%2(pixel *src, intptr_t srcStride, pixel *dst, intptr_t dstStride, int coeffIdx)
+;-----------------------------------------------------------------------------
+%macro FILTER_V4_W2_H4_sse2 2
 INIT_XMM sse2
 %if ARCH_X86_64
-cglobal interp_4tap_vert_pp_2x%1, 4, 6, 9
+cglobal interp_4tap_vert_%1_2x%2, 4, 6, 9
     pxor        m8,        m8
 %else
-cglobal interp_4tap_vert_pp_2x%1, 4, 6, 8
+cglobal interp_4tap_vert_%1_2x%2, 4, 6, 8
 %endif
     mov         r4d,       r4m
     sub         r0,        r1
 
+%ifidn %1,pp
+    mova        m1,        [pw_32]
+%elifidn %1,ps
+    mova        m1,        [pw_2000]
+    add         r3d,       r3d
+%endif
+
 %ifdef PIC
     lea         r5,        [tabw_ChromaCoeff]
     movh        m0,        [r5 + r4 * 8]
@@ -899,11 +906,10 @@
 %endif
 
     punpcklqdq  m0,        m0
-    mova        m1,        [pw_32]
     lea         r5,        [3 * r1]
 
 %assign x 1
-%rep %1/4
+%rep %2/4
     movd        m2,        [r0]
     movd        m3,        [r0 + r1]
     movd        m4,        [r0 + 2 * r1]
@@ -930,7 +936,6 @@
     pshuflw     m3,        m2,          q2301
     pshufhw     m3,        m3,          q2301
     paddw       m2,        m3
-    psrld       m2,        16
 
     movd        m7,        [r0 + r1]
 
@@ -954,8 +959,10 @@
     pshuflw     m5,        m4,          q2301
     pshufhw     m5,        m5,          q2301
     paddw       m4,        m5
+
+%ifidn %1,pp
+    psrld       m2,        16
     psrld       m4,        16
-
     packssdw    m2,        m4
     paddw       m2,        m1
     psraw       m2,        6
@@ -983,8 +990,24 @@
     shr         r4,        16
     mov         [r2 + r3], r4w
 %endif
-
-%if x < %1/4
+%elifidn %1,ps
+    psrldq      m2,        2
+    psrldq      m4,        2
+    pshufd      m2,        m2, q3120
+    pshufd      m4,        m4, q3120
+    psubw       m4,        m1
+    psubw       m2,        m1
+
+    movd        [r2],      m2
+    psrldq      m2,        4
+    movd        [r2 + r3], m2
+    lea         r2,        [r2 + 2 * r3]
+    movd        [r2],      m4
+    psrldq      m4,        4
+    movd        [r2 + r3], m4
+%endif
+
+%if x < %2/4
     lea         r2,        [r2 + 2 * r3]
 %endif
 %assign x x+1
@@ -993,9 +1016,13 @@
 
 %endmacro
 
-    FILTER_V4_W2_H4_sse2 4
-    FILTER_V4_W2_H4_sse2 8
-    FILTER_V4_W2_H4_sse2 16
+    FILTER_V4_W2_H4_sse2 pp, 4
+    FILTER_V4_W2_H4_sse2 pp, 8
+    FILTER_V4_W2_H4_sse2 pp, 16
+
+    FILTER_V4_W2_H4_sse2 ps, 4
+    FILTER_V4_W2_H4_sse2 ps, 8
+    FILTER_V4_W2_H4_sse2 ps, 16
 
 ;-----------------------------------------------------------------------------
 ; void interp_4tap_vert_pp_4x2(pixel *src, intptr_t srcStride, pixel *dst, intptr_t dstStride, int coeffIdx)
diff -r 8592bf81d084 -r 465fb4340a24 source/common/x86/ipfilter8.h
--- a/source/common/x86/ipfilter8.h	Thu May 14 17:12:14 2015 +0530
+++ b/source/common/x86/ipfilter8.h	Sun May 17 18:03:29 2015 -0700
@@ -913,6 +913,9 @@
 void x265_interp_4tap_vert_pp_4x8_sse2(const pixel *src, intptr_t srcStride, pixel *dst, intptr_t dstStride, int coeffIdx);
 void x265_interp_4tap_vert_pp_4x16_sse2(const pixel *src, intptr_t srcStride, pixel *dst, intptr_t dstStride, int coeffIdx);
 void x265_interp_4tap_vert_pp_4x32_sse2(const pixel *src, intptr_t srcStride, pixel *dst, intptr_t dstStride, int coeffIdx);
+void x265_interp_4tap_vert_ps_2x4_sse2(const pixel* src, intptr_t srcStride, int16_t* dst, intptr_t dstStride, int coeffIdx);
+void x265_interp_4tap_vert_ps_2x8_sse2(const pixel* src, intptr_t srcStride, int16_t* dst, intptr_t dstStride, int coeffIdx);
+void x265_interp_4tap_vert_ps_2x16_sse2(const pixel* src, intptr_t srcStride, int16_t* dst, intptr_t dstStride, int coeffIdx);
 #ifdef X86_64
 void x265_interp_4tap_vert_pp_6x8_sse2(const pixel *src, intptr_t srcStride, pixel *dst, intptr_t dstStride, int coeffIdx);
 void x265_interp_4tap_vert_pp_6x16_sse2(const pixel *src, intptr_t srcStride, pixel *dst, intptr_t dstStride, int coeffIdx);


More information about the x265-devel mailing list