[x265] [PATCH 100 of 307] x86: AVX512 interp_8tap_horiz_pp_48x64
mythreyi at multicorewareinc.com
mythreyi at multicorewareinc.com
Sat Apr 7 04:31:38 CEST 2018
# HG changeset patch
# User Vignesh Vijayakumar
# Date 1503912578 -19800
# Mon Aug 28 14:59:38 2017 +0530
# Node ID 562c00d2153193eec85ab907b60eeb5aca7cc609
# Parent a7bf0a24cfc8eb8edc95d340b240b91d03dac5bd
x86: AVX512 interp_8tap_horiz_pp_48x64
AVX2 performance : 19.57x
AVX512 perfornamce : 35.25x
diff -r a7bf0a24cfc8 -r 562c00d21531 source/common/x86/asm-primitives.cpp
--- a/source/common/x86/asm-primitives.cpp Mon Aug 28 14:46:28 2017 +0530
+++ b/source/common/x86/asm-primitives.cpp Mon Aug 28 14:59:38 2017 +0530
@@ -4159,6 +4159,7 @@
p.pu[LUMA_64x32].luma_hpp = PFX(interp_8tap_horiz_pp_64x32_avx512);
p.pu[LUMA_64x48].luma_hpp = PFX(interp_8tap_horiz_pp_64x48_avx512);
p.pu[LUMA_64x64].luma_hpp = PFX(interp_8tap_horiz_pp_64x64_avx512);
+ p.pu[LUMA_48x64].luma_hpp = PFX(interp_8tap_horiz_pp_48x64_avx512);
p.pu[LUMA_64x16].pixelavg_pp = PFX(pixel_avg_64x16_avx512);
p.pu[LUMA_64x32].pixelavg_pp = PFX(pixel_avg_64x32_avx512);
diff -r a7bf0a24cfc8 -r 562c00d21531 source/common/x86/ipfilter8.asm
--- a/source/common/x86/ipfilter8.asm Mon Aug 28 14:46:28 2017 +0530
+++ b/source/common/x86/ipfilter8.asm Mon Aug 28 14:59:38 2017 +0530
@@ -10489,6 +10489,151 @@
vextracti32x4 [r2 + r7], m7, 3
%endmacro
+%macro PROCESS_IPFILTER_LUMA_PP_48x4_AVX512 0
+ ; register map
+ ; m0 , m1 interpolate coeff
+ ; m2 , m3, m4 shuffle order table
+ ; m5 - pw_1
+ ; m6 - pw_512
+
+ movu ym7, [r0]
+ vinserti32x8 m7, [r0 + r1], 1
+ movu ym9, [r0 + 8]
+ vinserti32x8 m9, [r0 + r1 + 8], 1
+
+ pshufb m8, m7, m3
+ pshufb m7, m2
+ pshufb m10, m9, m3
+ pshufb m11, m9, m4
+ pshufb m9, m2
+
+ pmaddubsw m7, m0
+ pmaddubsw m12, m8, m1
+ pmaddwd m7, m5
+ pmaddwd m12, m5
+ paddd m7, m12
+
+ pmaddubsw m8, m0
+ pmaddubsw m12, m9, m1
+ pmaddwd m8, m5
+ pmaddwd m12, m5
+ paddd m8, m12
+
+ pmaddubsw m9, m0
+ pmaddubsw m12, m10, m1
+ pmaddwd m9, m5
+ pmaddwd m12, m5
+ paddd m9, m12
+
+ pmaddubsw m10, m0
+ pmaddubsw m12, m11, m1
+ pmaddwd m10, m5
+ pmaddwd m12, m5
+ paddd m10, m12
+
+ packssdw m7, m8
+ packssdw m9, m10
+ pmulhrsw m7, m6
+ pmulhrsw m9, m6
+ packuswb m7, m9
+ movu [r2], ym7
+ vextracti32x8 [r2 + r3], m7, 1
+
+ movu ym7, [r0 + 2 * r1]
+ vinserti32x8 m7, [r0 + r6], 1
+ movu ym9, [r0 + 2 * r1 + 8]
+ vinserti32x8 m9, [r0 + r6 + 8], 1
+
+ pshufb m8, m7, m3
+ pshufb m7, m2
+ pshufb m10, m9, m3
+ pshufb m11, m9, m4
+ pshufb m9, m2
+
+ pmaddubsw m7, m0
+ pmaddubsw m12, m8, m1
+ pmaddwd m7, m5
+ pmaddwd m12, m5
+ paddd m7, m12
+
+ pmaddubsw m8, m0
+ pmaddubsw m12, m9, m1
+ pmaddwd m8, m5
+ pmaddwd m12, m5
+ paddd m8, m12
+
+ pmaddubsw m9, m0
+ pmaddubsw m12, m10, m1
+ pmaddwd m9, m5
+ pmaddwd m12, m5
+ paddd m9, m12
+
+ pmaddubsw m10, m0
+ pmaddubsw m12, m11, m1
+ pmaddwd m10, m5
+ pmaddwd m12, m5
+ paddd m10, m12
+
+ packssdw m7, m8
+ packssdw m9, m10
+ pmulhrsw m7, m6
+ pmulhrsw m9, m6
+ packuswb m7, m9
+ movu [r2 + 2 * r3], ym7
+ vextracti32x8 [r2 + r7], m7, 1
+
+ movu xm7, [r0 + mmsize/2]
+ vinserti32x4 m7, [r0 + r1 + mmsize/2], 1
+ vinserti32x4 m7, [r0 + 2 * r1 + mmsize/2], 2
+ vinserti32x4 m7, [r0 + r6 + mmsize/2], 3
+
+ pshufb m8, m7, m3
+ pshufb m7, m2
+
+ movu xm9, [r0 + 40]
+ vinserti32x4 m9, [r0 + r1 + 40], 1
+ vinserti32x4 m9, [r0 + 2 * r1 + 40], 2
+ vinserti32x4 m9, [r0 + r6 + 40], 3
+
+ pshufb m10, m9, m3
+ pshufb m11, m9, m4
+ pshufb m9, m2
+
+ pmaddubsw m7, m0
+ pmaddubsw m12, m8, m1
+ pmaddwd m7, m5
+ pmaddwd m12, m5
+ paddd m7, m12
+
+ pmaddubsw m8, m0
+ pmaddubsw m12, m9, m1
+ pmaddwd m8, m5
+ pmaddwd m12, m5
+ paddd m8, m12
+
+ pmaddubsw m9, m0
+ pmaddubsw m12, m10, m1
+ pmaddwd m9, m5
+ pmaddwd m12, m5
+ paddd m9, m12
+
+ pmaddubsw m10, m0
+ pmaddubsw m12, m11, m1
+ pmaddwd m10, m5
+ pmaddwd m12, m5
+ paddd m10, m12
+
+ packssdw m7, m8
+ packssdw m9, m10
+ pmulhrsw m7, m6
+ pmulhrsw m9, m6
+ packuswb m7, m9
+ movu [r2 + mmsize/2], xm7
+ vextracti32x4 [r2 + r3 + mmsize/2], m7, 1
+ vextracti32x4 [r2 + 2 * r3 + mmsize/2], m7, 2
+ vextracti32x4 [r2 + r7 + mmsize/2], m7, 3
+%endmacro
+
%macro IPFILTER_LUMA_64xN_AVX512 1
INIT_ZMM avx512
cglobal interp_8tap_horiz_pp_64x%1, 4,6,13
@@ -10592,6 +10737,34 @@
IPFILTER_LUMA_16xN_AVX512 16
IPFILTER_LUMA_16xN_AVX512 32
IPFILTER_LUMA_16xN_AVX512 64
+
+INIT_ZMM avx512
+cglobal interp_8tap_horiz_pp_48x64, 4,8,14
+ sub r0, 3
+ mov r4d, r4m
+ lea r6, [3 * r1]
+ lea r7, [3 * r3]
+%ifdef PIC
+ lea r5, [tab_LumaCoeff]
+ vpbroadcastd m0, [r5 + r4 * 8]
+ vpbroadcastd m1, [r5 + r4 * 8 + 4]
+%else
+ vpbroadcastd m0, [tab_LumaCoeff + r4 * 8]
+ vpbroadcastd m1, [tab_LumaCoeff + r4 * 8 + 4]
+%endif
+ vbroadcasti32x8 m2, [interp4_horiz_shuf_load1_avx512]
+ vbroadcasti32x8 m3, [interp4_horiz_shuf_load3_avx512]
+ vbroadcasti32x8 m4, [interp4_horiz_shuf_load2_avx512]
+ vpbroadcastd m5, [pw_1]
+ vbroadcasti32x8 m6, [pw_512]
+
+%rep 15
+ PROCESS_IPFILTER_LUMA_PP_48x4_AVX512
+ lea r0, [r0 + 4 * r1]
+ lea r2, [r2 + 4 * r3]
+%endrep
+ PROCESS_IPFILTER_LUMA_PP_48x4_AVX512
+ RET
;-------------------------------------------------------------------------------------------------------------
;ipfilter_luma_avx512 code end
;-------------------------------------------------------------------------------------------------------------
\ No newline at end of file
More information about the x265-devel
mailing list