[x265] [PATCH 161 of 307] x86: AVX512 interp_4tap_horiz_ps_8xN for high bit depth
mythreyi at multicorewareinc.com
mythreyi at multicorewareinc.com
Sat Apr 7 04:32:39 CEST 2018
# HG changeset patch
# User Jayashri Murugan <jayashri at multicorewareinc.com>
# Date 1509605321 -19800
# Thu Nov 02 12:18:41 2017 +0530
# Node ID f4cd489d06cfbbf66c6f0f7dc684606c80615c5e
# Parent b45e6581fdf689f4b17e4e0dae733a7df795de07
x86: AVX512 interp_4tap_horiz_ps_8xN for high bit depth
Color Space i420
Size | AVX2 performance | AVX512 performance
----------------------------------------------
8x4 | 21.70x | 35.85x
8x8 | 22.91x | 33.85x
8x16 | 20.03x | 39.94x
8x32 | 22.08x | 47.20x
Color Space i422
Size | AVX2 performance | AVX512 performance
----------------------------------------------
8x4 | 21.37x | 35.57x
8x8 | 21.69x | 33.72x
8x12 | 19.50x | 37.37x
8x16 | 21.05x | 39.71x
8x32 | 22.01x | 47.10x
8x64 | 23.50x | 47.07x
Color Space i444
Size | AVX2 performance | AVX512 performance
----------------------------------------------
8x4 | 21.94x | 35.60x
8x8 | 21.62x | 35.04x
8x16 | 20.17x | 39.90x
8x32 | 22.03x | 46.81x
diff -r b45e6581fdf6 -r f4cd489d06cf source/common/x86/asm-primitives.cpp
--- a/source/common/x86/asm-primitives.cpp Mon Nov 13 22:43:09 2017 -0800
+++ b/source/common/x86/asm-primitives.cpp Thu Nov 02 12:18:41 2017 +0530
@@ -2778,6 +2778,23 @@
p.chroma[X265_CSP_I444].pu[LUMA_48x64].filter_hps = PFX(interp_4tap_horiz_ps_48x64_avx512);
+ p.chroma[X265_CSP_I420].pu[CHROMA_420_8x8].filter_hps = PFX(interp_4tap_horiz_ps_8x8_avx512);
+ p.chroma[X265_CSP_I420].pu[CHROMA_420_8x4].filter_hps = PFX(interp_4tap_horiz_ps_8x4_avx512);
+ p.chroma[X265_CSP_I420].pu[CHROMA_420_8x16].filter_hps = PFX(interp_4tap_horiz_ps_8x16_avx512);
+ p.chroma[X265_CSP_I420].pu[CHROMA_420_8x32].filter_hps = PFX(interp_4tap_horiz_ps_8x32_avx512);
+
+ p.chroma[X265_CSP_I422].pu[CHROMA_422_8x8].filter_hps = PFX(interp_4tap_horiz_ps_8x8_avx512);
+ p.chroma[X265_CSP_I422].pu[CHROMA_422_8x16].filter_hps = PFX(interp_4tap_horiz_ps_8x16_avx512);
+ p.chroma[X265_CSP_I422].pu[CHROMA_422_8x32].filter_hps = PFX(interp_4tap_horiz_ps_8x32_avx512);
+ p.chroma[X265_CSP_I422].pu[CHROMA_422_8x12].filter_hps = PFX(interp_4tap_horiz_ps_8x12_avx512);
+ p.chroma[X265_CSP_I422].pu[CHROMA_422_8x64].filter_hps = PFX(interp_4tap_horiz_ps_8x64_avx512);
+ p.chroma[X265_CSP_I422].pu[CHROMA_422_8x4].filter_hps = PFX(interp_4tap_horiz_ps_8x4_avx512);
+
+ p.chroma[X265_CSP_I444].pu[LUMA_8x8].filter_hps = PFX(interp_4tap_horiz_ps_8x8_avx512);
+ p.chroma[X265_CSP_I444].pu[LUMA_8x4].filter_hps = PFX(interp_4tap_horiz_ps_8x4_avx512);
+ p.chroma[X265_CSP_I444].pu[LUMA_8x16].filter_hps = PFX(interp_4tap_horiz_ps_8x16_avx512);
+ p.chroma[X265_CSP_I444].pu[LUMA_8x32].filter_hps = PFX(interp_4tap_horiz_ps_8x32_avx512);
+
}
#endif
}
diff -r b45e6581fdf6 -r f4cd489d06cf source/common/x86/ipfilter16.asm
--- a/source/common/x86/ipfilter16.asm Mon Nov 13 22:43:09 2017 -0800
+++ b/source/common/x86/ipfilter16.asm Thu Nov 02 12:18:41 2017 +0530
@@ -8146,6 +8146,132 @@
FILTER_VER_PS_CHROMA_64xN_AVX512 48
FILTER_VER_PS_CHROMA_64xN_AVX512 64
%endif
+
+%macro PROCESS_IPFILTER_CHROMA_PS_8x4_AVX512 0
+ ; register map
+ ; m0 , m1 - interpolate coeff
+ ; m2 , m3 - shuffle load order table
+ ; m4 - INTERP_OFFSET_PS
+ ; m5 - shuffle store order table
+
+ movu xm6, [r0]
+ vinserti32x4 m6, [r0 + r1], 1
+ vinserti32x4 m6, [r0 + 2 * r1], 2
+ vinserti32x4 m6, [r0 + r6], 3
+
+ pshufb m8, m6, m3
+ pshufb m6, m2
+ pmaddwd m6, m0
+ pmaddwd m8, m1
+ paddd m6, m8
+ paddd m6, m4
+ psrad m6, INTERP_SHIFT_PS
+
+ movu xm7, [r0 + 8]
+ vinserti32x4 m7, [r0 + r1 + 8], 1
+ vinserti32x4 m7, [r0 + 2 * r1 + 8], 2
+ vinserti32x4 m7, [r0 + r6 + 8], 3
+
+ pshufb m8, m7, m3
+ pshufb m7, m2
+ pmaddwd m7, m0
+ pmaddwd m8, m1
+ paddd m7, m8
+ paddd m7, m4
+ psrad m7, INTERP_SHIFT_PS
+
+ packssdw m6, m7
+ pshufb m6, m5
+ movu [r2], xm6
+ vextracti32x4 [r2 + r3], m6, 1
+ vextracti32x4 [r2 + 2 * r3], m6, 2
+ vextracti32x4 [r2 + r7], m6, 3
+%endmacro
+
+%macro PROCESS_IPFILTER_CHROMA_PS_8x3_AVX512 0
+ movu xm6, [r0]
+ vinserti32x4 m6, [r0 + r1], 1
+ vinserti32x4 m6, [r0 + 2 * r1], 2
+
+ pshufb m8, m6, m3
+ pshufb m6, m2
+ pmaddwd m6, m0
+ pmaddwd m8, m1
+ paddd m6, m8
+ paddd m6, m4
+ psrad m6, INTERP_SHIFT_PS
+
+ movu xm7, [r0 + 8]
+ vinserti32x4 m7, [r0 + r1 + 8], 1
+ vinserti32x4 m7, [r0 + 2 * r1 + 8], 2
+
+ pshufb m8, m7, m3
+ pshufb m7, m2
+ pmaddwd m7, m0
+ pmaddwd m8, m1
+ paddd m7, m8
+ paddd m7, m4
+ psrad m7, INTERP_SHIFT_PS
+
+ packssdw m6, m7
+ pshufb m6, m5
+ movu [r2], xm6
+ vextracti32x4 [r2 + r3], m6, 1
+ vextracti32x4 [r2 + 2 * r3], m6, 2
+%endmacro
+
+%macro IPFILTER_CHROMA_PS_AVX512_8xN 1
+INIT_ZMM avx512
+cglobal interp_4tap_horiz_ps_8x%1, 4,9,9
+ add r1d, r1d
+ add r3d, r3d
+ mov r4d, r4m
+ mov r5d, r5m
+
+ lea r6, [3 * r1]
+ lea r7, [3 * r3]
+%ifdef PIC
+ lea r8, [tab_ChromaCoeff]
+ vpbroadcastd m0, [r8 + r4 * 8]
+ vpbroadcastd m1, [r8 + r4 * 8 + 4]
+%else
+ vpbroadcastd m0, [tab_ChromaCoeff + r4 * 8]
+ vpbroadcastd m1, [tab_ChromaCoeff + r4 * 8 + 4]
+%endif
+ vbroadcasti32x8 m2, [interp8_hpp_shuf1_load_avx512]
+ vbroadcasti32x8 m3, [interp8_hpp_shuf2_load_avx512]
+ vbroadcasti32x4 m4, [INTERP_OFFSET_PS]
+ vbroadcasti32x8 m5, [interp8_hpp_shuf1_store_avx512]
+
+ mov r8d, %1
+ sub r0, 2
+ test r5d, r5d
+ jz .loop
+ sub r0, r1
+ add r8d, 3
+ PROCESS_IPFILTER_CHROMA_PS_8x3_AVX512
+ lea r0, [r0 + r6]
+ lea r2, [r2 + r7]
+ sub r8d, 3
+
+.loop:
+ PROCESS_IPFILTER_CHROMA_PS_8x4_AVX512
+ lea r0, [r0 + 4 * r1]
+ lea r2, [r2 + 4 * r3]
+ sub r8d, 4
+ jnz .loop
+ RET
+%endmacro
+
+%if ARCH_X86_64
+IPFILTER_CHROMA_PS_AVX512_8xN 4
+IPFILTER_CHROMA_PS_AVX512_8xN 8
+IPFILTER_CHROMA_PS_AVX512_8xN 12
+IPFILTER_CHROMA_PS_AVX512_8xN 16
+IPFILTER_CHROMA_PS_AVX512_8xN 32
+IPFILTER_CHROMA_PS_AVX512_8xN 64
+%endif
+
;-------------------------------------------------------------------------------------------------------------
;ipfilter_chroma_avx512 code end
;-------------------------------------------------------------------------------------------------------------
More information about the x265-devel
mailing list