# HG changeset patch
# User Vignesh Vijayakumar
# Date 1504171458 -19800
# Thu Aug 31 14:54:18 2017 +0530
# Node ID 1fb1948309a0a9218a07e060300b9d5a7ff58321
# Parent 9928b3e5b4d4235bea9ffb22434446e68c3aacdb
x86: AVX512 interp_4tap_horiz_pp_8xN for high bit depth
Color Space i444
Size | AVX2 performance | AVX512 performance
----------------------------------------------
8x4 | 5.14x | 9.51x
8x8 | 6.20x | 12.75x
8x16 | 6.32x | 12.44x
8x32 | 6.01x | 13.68x
diff -r 9928b3e5b4d4 -r 1fb1948309a0 source/common/x86/asm-primitives.cpp
--- a/source/common/x86/asm-primitives.cpp Thu Aug 31 14:24:24 2017 +0530
+++ b/source/common/x86/asm-primitives.cpp Thu Aug 31 14:54:18 2017 +0530
@@ -2354,6 +2354,10 @@
p.cu[BLOCK_32x32].copy_cnt = PFX(copy_cnt_32_avx512);
p.cu[BLOCK_16x16].copy_cnt = PFX(copy_cnt_16_avx512);
+ p.chroma[X265_CSP_I420].pu[CHROMA_420_8x4].filter_hpp =
PFX(interp_4tap_horiz_pp_8x4_avx512);
+ p.chroma[X265_CSP_I420].pu[CHROMA_420_8x8].filter_hpp =
PFX(interp_4tap_horiz_pp_8x8_avx512);
+ p.chroma[X265_CSP_I420].pu[CHROMA_420_8x16].filter_hpp =
PFX(interp_4tap_horiz_pp_8x16_avx512);
+ p.chroma[X265_CSP_I420].pu[CHROMA_420_8x32].filter_hpp =
PFX(interp_4tap_horiz_pp_8x32_avx512);
p.chroma[X265_CSP_I420].pu[CHROMA_420_16x4].filter_hpp =
PFX(interp_4tap_horiz_pp_16x4_avx512);
p.chroma[X265_CSP_I420].pu[CHROMA_420_16x8].filter_hpp =
PFX(interp_4tap_horiz_pp_16x8_avx512);
p.chroma[X265_CSP_I420].pu[CHROMA_420_16x12].filter_hpp =
PFX(interp_4tap_horiz_pp_16x12_avx512);
@@ -2364,6 +2368,12 @@
p.chroma[X265_CSP_I420].pu[CHROMA_420_32x24].filter_hpp =
PFX(interp_4tap_horiz_pp_32x24_avx512);
p.chroma[X265_CSP_I420].pu[CHROMA_420_32x32].filter_hpp =
PFX(interp_4tap_horiz_pp_32x32_avx512);
+ p.chroma[X265_CSP_I422].pu[CHROMA_422_8x4].filter_hpp =
PFX(interp_4tap_horiz_pp_8x4_avx512);
+ p.chroma[X265_CSP_I422].pu[CHROMA_422_8x8].filter_hpp =
PFX(interp_4tap_horiz_pp_8x8_avx512);
+ p.chroma[X265_CSP_I422].pu[CHROMA_422_8x12].filter_hpp =
PFX(interp_4tap_horiz_pp_8x12_avx512);
+ p.chroma[X265_CSP_I422].pu[CHROMA_422_8x16].filter_hpp =
PFX(interp_4tap_horiz_pp_8x16_avx512);
+ p.chroma[X265_CSP_I422].pu[CHROMA_422_8x32].filter_hpp =
PFX(interp_4tap_horiz_pp_8x32_avx512);
+ p.chroma[X265_CSP_I422].pu[CHROMA_422_8x64].filter_hpp =
PFX(interp_4tap_horiz_pp_8x64_avx512);
p.chroma[X265_CSP_I422].pu[CHROMA_422_16x8].filter_hpp =
PFX(interp_4tap_horiz_pp_16x8_avx512);
p.chroma[X265_CSP_I422].pu[CHROMA_422_16x16].filter_hpp =
PFX(interp_4tap_horiz_pp_16x16_avx512);
p.chroma[X265_CSP_I422].pu[CHROMA_422_16x24].filter_hpp =
PFX(interp_4tap_horiz_pp_16x24_avx512);
@@ -2374,6 +2384,10 @@
p.chroma[X265_CSP_I422].pu[CHROMA_422_32x48].filter_hpp =
PFX(interp_4tap_horiz_pp_32x48_avx512);
p.chroma[X265_CSP_I422].pu[CHROMA_422_32x64].filter_hpp =
PFX(interp_4tap_horiz_pp_32x64_avx512);
+ p.chroma[X265_CSP_I444].pu[LUMA_8x4].filter_hpp =
PFX(interp_4tap_horiz_pp_8x4_avx512);
+ p.chroma[X265_CSP_I444].pu[LUMA_8x8].filter_hpp =
PFX(interp_4tap_horiz_pp_8x8_avx512);
+ p.chroma[X265_CSP_I444].pu[LUMA_8x16].filter_hpp =
PFX(interp_4tap_horiz_pp_8x16_avx512);
+ p.chroma[X265_CSP_I444].pu[LUMA_8x32].filter_hpp =
PFX(interp_4tap_horiz_pp_8x32_avx512);
p.chroma[X265_CSP_I444].pu[LUMA_16x4].filter_hpp =
PFX(interp_4tap_horiz_pp_16x4_avx512);
p.chroma[X265_CSP_I444].pu[LUMA_16x8].filter_hpp =
PFX(interp_4tap_horiz_pp_16x8_avx512);
p.chroma[X265_CSP_I444].pu[LUMA_16x12].filter_hpp =
PFX(interp_4tap_horiz_pp_16x12_avx512);
diff -r 9928b3e5b4d4 -r 1fb1948309a0 source/common/x86/ipfilter16.asm
--- a/source/common/x86/ipfilter16.asm Thu Aug 31 14:24:24 2017 +0530
+++ b/source/common/x86/ipfilter16.asm Thu Aug 31 14:54:18 2017 +0530
@@ -5082,6 +5082,49 @@
;-------------------------------------------------------------------------------------------------------------
;ipfilter_chroma_avx512 code start
;-------------------------------------------------------------------------------------------------------------
+%macro PROCESS_IPFILTER_CHROMA_PP_8x4_AVX512 0
+ ; register map
+ ; m0 , m1 interpolate coeff
+ ; m2 , m3 shuffle order table
+ ; m4 - pd_32
+ ; m5 - zero
+ ; m6 - pw_pixel_max
+
+ movu xm7, [r0]
+ vinserti32x4 m7, [r0 + r1], 1
+ vinserti32x4 m7, [r0 + 2 * r1], 2
+ vinserti32x4 m7, [r0 + r6], 3
+
+ pshufb m9, m7, m3
+ pshufb m7, m2
+ pmaddwd m7, m0
+ pmaddwd m9, m1
+ paddd m7, m9
+ paddd m7, m4
+ psrad m7, 6
+
+ movu xm8, [r0 + 8]
+ vinserti32x4 m8, [r0 + r1 + 8], 1
+ vinserti32x4 m8, [r0 + 2 * r1 + 8], 2
+ vinserti32x4 m8, [r0 + r6 + 8], 3
+
+ pshufb m9, m8, m3
+ pshufb m8, m2
+ pmaddwd m8, m0
+ pmaddwd m9, m1
+ paddd m8, m9
+ paddd m8, m4
+ psrad m8, 6
+
+ packusdw m7, m8
+ CLIPW m7, m5, m6
+ pshufb m7, m10
+ movu [r2], xm7
+ vextracti32x4 [r2 + r3], m7, 1
+ vextracti32x4 [r2 + 2 * r3], m7, 2
+ vextracti32x4 [r2 + r7], m7, 3
+%endmacro
+
%macro PROCESS_IPFILTER_CHROMA_PP_16x2_AVX512 0
; register map
; m0 , m1 interpolate coeff
@@ -5367,6 +5410,71 @@
; void interp_4tap_horiz_pp(pixel *src, intptr_t srcStride, int16_t *dst,
intptr_t dstStride, int coeffIdx
;-------------------------------------------------------------------------------------------------------------
INIT_ZMM avx512
+cglobal interp_4tap_horiz_pp_8x4, 5,8,11
+ add r1d, r1d
+ add r3d, r3d
+ sub r0, 2
+ mov r4d, r4m
+ lea r6, [3 * r1]
+ lea r7, [3 * r3]
+%ifdef PIC
+ lea r5, [tab_ChromaCoeff]
+ vpbroadcastd m0, [r5 + r4 * 8]
+ vpbroadcastd m1, [r5 + 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]
+ vbroadcasti32x8 m4, [pd_32]
+ pxor m5, m5
+ vbroadcasti32x8 m6, [pw_pixel_max]
+ vbroadcasti32x8 m10, [interp8_hpp_shuf1_store_avx512]
+
+ PROCESS_IPFILTER_CHROMA_PP_8x4_AVX512
+ RET
+
+INIT_ZMM avx512
+%macro IPFILTER_CHROMA_AVX512_8xN 1
+cglobal interp_4tap_horiz_pp_8x%1, 5,8,11
+ add r1d, r1d
+ add r3d, r3d
+ sub r0, 2
+ mov r4d, r4m
+ lea r6, [3 * r1]
+ lea r7, [3 * r3]
+%ifdef PIC
+ lea r5, [tab_ChromaCoeff]
+ vpbroadcastd m0, [r5 + r4 * 8]
+ vpbroadcastd m1, [r5 + 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]
+ vbroadcasti32x8 m4, [pd_32]
+ pxor m5, m5
+ vbroadcasti32x8 m6, [pw_pixel_max]
+ vbroadcasti32x8 m10, [interp8_hpp_shuf1_store_avx512]
+
+%rep %1/4 - 1
+ PROCESS_IPFILTER_CHROMA_PP_8x4_AVX512
+ lea r0, [r0 + 4 * r1]
+ lea r2, [r2 + 4 * r3]
+%endrep
+ PROCESS_IPFILTER_CHROMA_PP_8x4_AVX512
+ RET
+%endmacro
+
+IPFILTER_CHROMA_AVX512_8xN 8
+IPFILTER_CHROMA_AVX512_8xN 12
+IPFILTER_CHROMA_AVX512_8xN 16
+IPFILTER_CHROMA_AVX512_8xN 32
+IPFILTER_CHROMA_AVX512_8xN 64
+
+INIT_ZMM avx512
%macro IPFILTER_CHROMA_AVX512_16xN 1
cglobal interp_4tap_horiz_pp_16x%1, 5,6,11
add r1d, r1d
_______________________________________________
x265-devel mailing list
x265-devel@videolan.org
https://mailman.videolan.org/listinfo/x265-devel
