[x265] [PATCH] asm: avx2 10bit code for luma_hpp[12x16] (5154.47 -> 3632.88)
rajesh at multicorewareinc.com
rajesh at multicorewareinc.com
Tue May 19 10:58:41 CEST 2015
# HG changeset patch
# User Rajesh Paulraj<rajesh at multicorewareinc.com>
# Date 1432021220 -19800
# Tue May 19 13:10:20 2015 +0530
# Node ID b7f9e65a33ade32c1f14b04d69cce50cecde8ab5
# Parent 569f678f36d731690115b27ed244970f3bc822a8
asm: avx2 10bit code for luma_hpp[12x16] (5154.47 -> 3632.88)
diff -r 569f678f36d7 -r b7f9e65a33ad source/common/x86/asm-primitives.cpp
--- a/source/common/x86/asm-primitives.cpp Tue May 19 12:55:49 2015 +0530
+++ b/source/common/x86/asm-primitives.cpp Tue May 19 13:10:20 2015 +0530
@@ -1407,6 +1407,7 @@
p.pu[LUMA_64x32].luma_hpp = x265_interp_8tap_horiz_pp_64x32_avx2;
p.pu[LUMA_64x48].luma_hpp = x265_interp_8tap_horiz_pp_64x48_avx2;
p.pu[LUMA_64x64].luma_hpp = x265_interp_8tap_horiz_pp_64x64_avx2;
+ p.pu[LUMA_12x16].luma_hpp = x265_interp_8tap_horiz_pp_12x16_avx2;
if (cpuMask & X265_CPU_BMI2)
p.scanPosLast = x265_scanPosLast_avx2_bmi2;
diff -r 569f678f36d7 -r b7f9e65a33ad source/common/x86/ipfilter16.asm
--- a/source/common/x86/ipfilter16.asm Tue May 19 12:55:49 2015 +0530
+++ b/source/common/x86/ipfilter16.asm Tue May 19 13:10:20 2015 +0530
@@ -1162,6 +1162,96 @@
FILTER_HOR_LUMA_W32 64, 48
FILTER_HOR_LUMA_W32 64, 64
+;-------------------------------------------------------------------------------------------------------------
+; void interp_8tap_horiz_pp(pixel *src, intptr_t srcStride, pixel *dst, intptr_t dstStride, int coeffIdx
+;-------------------------------------------------------------------------------------------------------------
+INIT_YMM avx2
+cglobal interp_8tap_horiz_pp_12x16, 4,6,8
+ 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
+ mova m3, [interp8_hpp_shuf]
+ mova m7, [pd_32]
+ pxor m2, m2
+
+ ; register map
+ ; m0 , m1 interpolate coeff
+
+ mov r4d, 16
+
+.loop:
+ vbroadcasti128 m4, [r0]
+ vbroadcasti128 m5, [r0 + 8]
+ pshufb m4, m3
+ pshufb m5, m3
+
+ pmaddwd m4, m0
+ pmaddwd m5, m1
+ paddd m4, m5
+
+ vbroadcasti128 m5, [r0 + 8]
+ vbroadcasti128 m6, [r0 + 16]
+ pshufb m5, m3
+ pshufb m6, m3
+
+ pmaddwd m5, m0
+ pmaddwd m6, m1
+ paddd m5, m6
+
+ phaddd m4, m5
+ vpermq m4, m4, q3120
+ paddd m4, m7
+ psrad m4, 6
+
+ packusdw m4, m4
+ vpermq m4, m4, q2020
+ CLIPW m4, m2, [pw_pixel_max]
+ movu [r2], xm4
+
+ vbroadcasti128 m4, [r0 + 16]
+ vbroadcasti128 m5, [r0 + 24]
+ pshufb m4, m3
+ pshufb m5, m3
+
+ pmaddwd m4, m0
+ pmaddwd m5, m1
+ paddd m4, m5
+
+ vbroadcasti128 m5, [r0 + 24]
+ vbroadcasti128 m6, [r0 + 32]
+ pshufb m5, m3
+ pshufb m6, m3
+
+ pmaddwd m5, m0
+ pmaddwd m6, m1
+ paddd m5, m6
+
+ phaddd m4, m5
+ vpermq m4, m4, q3120
+ paddd m4, m7
+ psrad m4, 6
+
+ packusdw m4, m4
+ vpermq m4, m4, q2020
+ CLIPW m4, m2, [pw_pixel_max]
+ movq [r2 + 16], xm4
+
+ add r2, r3
+ add r0, r1
+ dec r4d
+ jnz .loop
+ RET
+
;-----------------------------------------------------------------------------
; void interp_4tap_horiz_%3_%1x%2(pixel *src, intptr_t srcStride, pixel *dst, intptr_t dstStride, int coeffIdx)
;-----------------------------------------------------------------------------
More information about the x265-devel
mailing list