[x265] [PATCH] asm: avx2 10bit code for luma_hpp[4xN]
rajesh at multicorewareinc.com
rajesh at multicorewareinc.com
Wed May 20 10:54:56 CEST 2015
# HG changeset patch
# User Rajesh Paulraj<rajesh at multicorewareinc.com>
# Date 1432111013 -19800
# Wed May 20 14:06:53 2015 +0530
# Node ID 453218d87a4ee8667b013b1f36465533df06330b
# Parent 9b31a8a7bd57efededcc3884ee7777c09f649394
asm: avx2 10bit code for luma_hpp[4xN]
avx2:
luma_hpp[ 4x4] 4.59x 423.90 1944.66
luma_hpp[ 4x8] 4.74x 803.53 3806.63
luma_hpp[ 4x16] 4.73x 1574.01 7442.57
sse4:
luma_hpp[ 4x4] 3.69x 527.97 1946.47
luma_hpp[ 4x8] 3.93x 961.48 3780.20
luma_hpp[ 4x16] 4.06x 1833.63 7445.62
diff -r 9b31a8a7bd57 -r 453218d87a4e source/common/x86/asm-primitives.cpp
--- a/source/common/x86/asm-primitives.cpp Tue May 19 19:51:56 2015 -0500
+++ b/source/common/x86/asm-primitives.cpp Wed May 20 14:06:53 2015 +0530
@@ -1407,6 +1407,9 @@
p.pu[LUMA_4x8].luma_hps = x265_interp_8tap_horiz_ps_4x8_avx2;
p.pu[LUMA_4x16].luma_hps = x265_interp_8tap_horiz_ps_4x16_avx2;
+ p.pu[LUMA_4x4].luma_hpp = x265_interp_8tap_horiz_pp_4x4_avx2;
+ p.pu[LUMA_4x8].luma_hpp = x265_interp_8tap_horiz_pp_4x8_avx2;
+ p.pu[LUMA_4x16].luma_hpp = x265_interp_8tap_horiz_pp_4x16_avx2;
p.pu[LUMA_8x4].luma_hpp = x265_interp_8tap_horiz_pp_8x4_avx2;
p.pu[LUMA_8x8].luma_hpp = x265_interp_8tap_horiz_pp_8x8_avx2;
p.pu[LUMA_8x16].luma_hpp = x265_interp_8tap_horiz_pp_8x16_avx2;
diff -r 9b31a8a7bd57 -r 453218d87a4e source/common/x86/ipfilter16.asm
--- a/source/common/x86/ipfilter16.asm Tue May 19 19:51:56 2015 -0500
+++ b/source/common/x86/ipfilter16.asm Wed May 20 14:06:53 2015 +0530
@@ -862,6 +862,84 @@
movhps [r2 + r3], m3
%endmacro
+;-------------------------------------------------------------------------------------------------------------
+; void interp_8tap_horiz_pp(pixel *src, intptr_t srcStride, pixel *dst, intptr_t dstStride, int coeffIdx
+;-------------------------------------------------------------------------------------------------------------
+%macro FILTER_HOR_LUMA_W4_avx2 1
+INIT_YMM avx2
+cglobal interp_8tap_horiz_pp_4x%1, 4,7,7
+ add r1d, r1d
+ add r3d, r3d
+ sub r0, 6
+ mov r4d, r4m
+ shl r4d, 4
+%ifdef PIC
+ lea r5, [tab_LumaCoeff]
+ vpbroadcastq m0, [r5 + r4]
+ vpbroadcastq m1, [r5 + r4 + 8]
+%else
+ vpbroadcastq m0, [tab_LumaCoeff + r4]
+ vpbroadcastq m1, [tab_LumaCoeff + r4 + 8]
+%endif
+ lea r6, [pw_pixel_max]
+ mova m3, [interp8_hpp_shuf]
+ mova m6, [pd_32]
+ pxor m2, m2
+
+ ; register map
+ ; m0 , m1 interpolate coeff
+
+ mov r4d, %1/2
+
+.loop:
+ vbroadcasti128 m4, [r0]
+ vbroadcasti128 m5, [r0 + 8]
+ pshufb m4, m3
+ pshufb m5, m3
+
+ pmaddwd m4, m0
+ pmaddwd m5, m1
+ paddd m4, m5
+
+ phaddd m4, m4
+ vpermq m4, m4, q3120
+ paddd m4, m6
+ psrad m4, 6
+
+ packusdw m4, m4
+ vpermq m4, m4, q2020
+ CLIPW m4, m2, [r6]
+ movq [r2], xm4
+
+ vbroadcasti128 m4, [r0 + r1]
+ vbroadcasti128 m5, [r0 + r1 + 8]
+ pshufb m4, m3
+ pshufb m5, m3
+
+ pmaddwd m4, m0
+ pmaddwd m5, m1
+ paddd m4, m5
+
+ phaddd m4, m4
+ vpermq m4, m4, q3120
+ paddd m4, m6
+ psrad m4, 6
+
+ packusdw m4, m4
+ vpermq m4, m4, q2020
+ CLIPW m4, m2, [r6]
+ movq [r2 + r3], xm4
+
+ lea r2, [r2 + 2 * r3]
+ lea r0, [r0 + 2 * r1]
+ dec r4d
+ jnz .loop
+ RET
+%endmacro
+FILTER_HOR_LUMA_W4_avx2 4
+FILTER_HOR_LUMA_W4_avx2 8
+FILTER_HOR_LUMA_W4_avx2 16
+
;-------------------------------------------------------------------------------------------------------------
; void interp_8tap_horiz_pp(pixel *src, intptr_t srcStride, pixel *dst, intptr_t dstStride, int coeffIdx
;-------------------------------------------------------------------------------------------------------------
More information about the x265-devel
mailing list