[x265] [PATCH 129 of 307] x86: AVX512 interp_8tap_horiz_ps_16xN
mythreyi at multicorewareinc.com
mythreyi at multicorewareinc.com
Sat Apr 7 04:32:07 CEST 2018
# HG changeset patch
# User Jayashri Murugan <jayashri at multicorewareinc.com>
# Date 1504095976 -19800
# Wed Aug 30 17:56:16 2017 +0530
# Node ID be9e1235a7e3ddc6d4de0b38d0253edf1726fa6d
# Parent 578a08347d3e92db9300f5e28baacd72a71a6423
x86: AVX512 interp_8tap_horiz_ps_16xN
Luma
Size | AVX2 performance | AVX512 performance
----------------------------------------------
16x4 | 12.28x | 20.48x
16x8 | 12.07x | 21.72x
16x12 | 11.20x | 21.73x
16x16 | 11.89x | 22.39x
16x32 | 11.99x | 21.35x
16x64 | 12.35x | 21.93x
diff -r 578a08347d3e -r be9e1235a7e3 source/common/x86/asm-primitives.cpp
--- a/source/common/x86/asm-primitives.cpp Wed Aug 30 11:15:08 2017 +0530
+++ b/source/common/x86/asm-primitives.cpp Wed Aug 30 17:56:16 2017 +0530
@@ -4433,6 +4433,13 @@
p.chroma[X265_CSP_I444].pu[LUMA_16x12].filter_hps = PFX(interp_4tap_horiz_ps_16x12_avx512);
p.chroma[X265_CSP_I444].pu[LUMA_16x4].filter_hps = PFX(interp_4tap_horiz_ps_16x4_avx512);
p.chroma[X265_CSP_I444].pu[LUMA_16x64].filter_hps = PFX(interp_4tap_horiz_ps_16x64_avx512);
+
+ p.pu[LUMA_16x8].luma_hps = PFX(interp_8tap_horiz_ps_16x8_avx512);
+ p.pu[LUMA_16x16].luma_hps = PFX(interp_8tap_horiz_ps_16x16_avx512);
+ p.pu[LUMA_16x12].luma_hps = PFX(interp_8tap_horiz_ps_16x12_avx512);
+ p.pu[LUMA_16x4].luma_hps = PFX(interp_8tap_horiz_ps_16x4_avx512);
+ p.pu[LUMA_16x32].luma_hps = PFX(interp_8tap_horiz_ps_16x32_avx512);
+ p.pu[LUMA_16x64].luma_hps = PFX(interp_8tap_horiz_ps_16x64_avx512);
}
#endif
}
diff -r 578a08347d3e -r be9e1235a7e3 source/common/x86/ipfilter8.asm
--- a/source/common/x86/ipfilter8.asm Wed Aug 30 11:15:08 2017 +0530
+++ b/source/common/x86/ipfilter8.asm Wed Aug 30 17:56:16 2017 +0530
@@ -11285,6 +11285,115 @@
IPFILTER_LUMA_PS_32xN_AVX512 24
IPFILTER_LUMA_PS_32xN_AVX512 32
IPFILTER_LUMA_PS_32xN_AVX512 64
+
+%macro PROCESS_IPFILTER_LUMA_PS_8TAP_16x2_AVX512 0
+ movu xm7, [r0]
+ vinserti32x4 m7, [r0 + 8], 1
+ vinserti32x4 m7, [r0 + r1], 2
+ vinserti32x4 m7, [r0 + r1 + 8], 3
+ pshufb m8, m7, m3
+ pshufb m9, m7, m4
+ pshufb m7, m2
+
+ pmaddubsw m7, m0
+ pmaddubsw m10, m8, m1
+ pmaddwd m7, m5
+ pmaddwd m10, m5
+ paddd m7, m10
+
+ pmaddubsw m8, m0
+ pmaddubsw m10, m9, m1
+ pmaddwd m8, m5
+ pmaddwd m10, m5
+ paddd m8, m10
+
+ packssdw m7, m8
+ psubw m7, m6
+ movu [r2], ym7
+ vextracti32x8 [r2 + r3], m7, 1
+%endmacro
+
+%macro PROCESS_IPFILTER_LUMA_PS_8TAP_16x1_AVX512 0
+ movu xm7, [r0]
+ vinserti32x4 m7, [r0 + 8], 1
+ pshufb ym8, ym7, ym3
+ pshufb ym9, ym7, ym4
+ pshufb ym7, ym2
+
+ pmaddubsw ym7, ym0
+ pmaddubsw ym10, ym8, ym1
+ pmaddwd ym7, ym5
+ pmaddwd ym10, ym5
+ paddd ym7, ym10
+
+ pmaddubsw ym8, ym0
+ pmaddubsw ym10, ym9, ym1
+ pmaddwd ym8, ym5
+ pmaddwd ym10, ym5
+ paddd ym8, ym10
+
+ packssdw ym7, ym8
+ psubw ym7, ym6
+ movu [r2], ym7
+%endmacro
+
+;-------------------------------------------------------------------------------------------------------------
+; void interp_horiz_ps_16xN(const pixel* src, intptr_t srcStride, int16_t* dst, intptr_t dstStride, int coeffIdx, int isRowExt)
+;-------------------------------------------------------------------------------------------------------------
+%macro IPFILTER_LUMA_PS_8TAP_16xN_AVX512 1
+INIT_ZMM avx512
+cglobal interp_8tap_horiz_ps_16x%1, 4,7,11
+ mov r4d, r4m
+ mov r5d, r5m
+ add r3, r3
+
+%ifdef PIC
+ lea r6, [tab_LumaCoeff]
+ vpbroadcastd m0, [r6 + r4 * 8]
+ vpbroadcastd m1, [r6 + 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_2000]
+
+ ; register map
+ ; m0 , m1 - interpolate coeff
+ ; m2 , m3, m4 - load shuffle order table
+ ; m5 - pw_1
+ ; m6 - pw_2000
+
+ mov r4d, %1
+ sub r0, 3
+ test r5d, r5d
+ jz .loop
+ lea r6, [r1 * 3]
+ sub r0, r6 ; r0(src)-r6
+ add r4d, 7 ; blkheight += N - 1
+ PROCESS_IPFILTER_LUMA_PS_8TAP_16x1_AVX512
+ lea r0, [r0 + r1]
+ lea r2, [r2 + r3]
+ dec r4d
+
+.loop:
+ PROCESS_IPFILTER_LUMA_PS_8TAP_16x2_AVX512
+ lea r0, [r0 + 2 * r1]
+ lea r2, [r2 + 2 * r3]
+ sub r4d, 2
+ jnz .loop
+ RET
+%endmacro
+
+IPFILTER_LUMA_PS_8TAP_16xN_AVX512 4
+IPFILTER_LUMA_PS_8TAP_16xN_AVX512 8
+IPFILTER_LUMA_PS_8TAP_16xN_AVX512 12
+IPFILTER_LUMA_PS_8TAP_16xN_AVX512 16
+IPFILTER_LUMA_PS_8TAP_16xN_AVX512 32
+IPFILTER_LUMA_PS_8TAP_16xN_AVX512 64
;-------------------------------------------------------------------------------------------------------------
;ipfilter_luma_avx512 code end
;-------------------------------------------------------------------------------------------------------------
\ No newline at end of file
More information about the x265-devel
mailing list