[x265] [PATCH 093 of 307] x86: AVX512 addAvg_32xN
mythreyi at multicorewareinc.com
mythreyi at multicorewareinc.com
Sat Apr 7 04:31:31 CEST 2018
# HG changeset patch
# User Vignesh Vijayakumar
# Date 1503385834 -19800
# Tue Aug 22 12:40:34 2017 +0530
# Node ID 738f07186eb1d4bca84e9acdf70921ee9e2fee92
# Parent ed1932a414bf5962bbeccfd5c9e208b7db90f77f
x86: AVX512 addAvg_32xN
Size | AVX2 performance | AVX512 performance
----------------------------------------------
32x8 | 15.31x | 19.98x
32x16 | 15.14x | 23.25x
32x24 | 14.65x | 23.95x
32x32 | 15.41x | 24.76x
32x64 | 14.56x | 24.53x
diff -r ed1932a414bf -r 738f07186eb1 source/common/x86/asm-primitives.cpp
--- a/source/common/x86/asm-primitives.cpp Sun Aug 13 18:18:28 2017 +0530
+++ b/source/common/x86/asm-primitives.cpp Tue Aug 22 12:40:34 2017 +0530
@@ -3964,6 +3964,19 @@
p.pu[LUMA_64x32].addAvg = PFX(addAvg_64x32_avx512);
p.pu[LUMA_64x48].addAvg = PFX(addAvg_64x48_avx512);
p.pu[LUMA_64x64].addAvg = PFX(addAvg_64x64_avx512);
+ p.pu[LUMA_32x8].addAvg = PFX(addAvg_32x8_avx512);
+ p.pu[LUMA_32x16].addAvg = PFX(addAvg_32x16_avx512);
+ p.pu[LUMA_32x24].addAvg = PFX(addAvg_32x24_avx512);
+ p.pu[LUMA_32x32].addAvg = PFX(addAvg_32x32_avx512);
+ p.pu[LUMA_32x64].addAvg = PFX(addAvg_32x64_avx512);
+ p.chroma[X265_CSP_I420].pu[CHROMA_420_32x8].addAvg = PFX(addAvg_32x8_avx512);
+ p.chroma[X265_CSP_I420].pu[CHROMA_420_32x16].addAvg = PFX(addAvg_32x16_avx512);
+ p.chroma[X265_CSP_I420].pu[CHROMA_420_32x24].addAvg = PFX(addAvg_32x24_avx512);
+ p.chroma[X265_CSP_I420].pu[CHROMA_420_32x32].addAvg = PFX(addAvg_32x32_avx512);
+ p.chroma[X265_CSP_I422].pu[CHROMA_422_32x16].addAvg = PFX(addAvg_32x16_avx512);
+ p.chroma[X265_CSP_I422].pu[CHROMA_422_32x48].addAvg = PFX(addAvg_32x48_avx512);
+ p.chroma[X265_CSP_I422].pu[CHROMA_422_32x64].addAvg = PFX(addAvg_32x64_avx512);
+ p.chroma[X265_CSP_I422].pu[CHROMA_422_32x32].addAvg = PFX(addAvg_32x32_avx512);
p.cu[BLOCK_32x32].blockfill_s = PFX(blockfill_s_32x32_avx512);
diff -r ed1932a414bf -r 738f07186eb1 source/common/x86/mc-a.asm
--- a/source/common/x86/mc-a.asm Sun Aug 13 18:18:28 2017 +0530
+++ b/source/common/x86/mc-a.asm Tue Aug 22 12:40:34 2017 +0530
@@ -3317,6 +3317,24 @@
movu [r2 + r5], m0
%endmacro
+%macro PROCESS_ADDAVG_32x2_AVX512 0
+ movu m0, [r0]
+ movu m1, [r1]
+ movu m2, [r0 + r3]
+ movu m3, [r1 + r4]
+
+ paddw m0, m1
+ pmulhrsw m0, m4
+ paddw m0, m5
+ paddw m2, m3
+ pmulhrsw m2, m4
+ paddw m2, m5
+
+ packuswb m0, m2
+ vpermq m0, m6, m0
+ movu [r2], ym0
+ vextracti32x8 [r2 + r5], m0, 1
+%endmacro
;--------------------------------------------------------------------------------------------------------------------
;void addAvg (int16_t* src0, int16_t* src1, pixel* dst, intptr_t src0Stride, intptr_t src1Stride, intptr_t dstStride)
;--------------------------------------------------------------------------------------------------------------------
@@ -3344,6 +3362,32 @@
ADDAVG_W64_AVX512 32
ADDAVG_W64_AVX512 48
ADDAVG_W64_AVX512 64
+
+%macro ADDAVG_W32_AVX512 1
+INIT_ZMM avx512
+cglobal addAvg_32x%1, 6,6,7
+ vbroadcasti32x8 m4, [pw_256]
+ vbroadcasti32x8 m5, [pw_128]
+ mova m6, [shuf_avx512]
+ add r3, r3
+ add r4, r4
+
+%rep %1/2 - 1
+ PROCESS_ADDAVG_32x2_AVX512
+ lea r2, [r2 + 2 * r5]
+ lea r0, [r0 + 2 * r3]
+ lea r1, [r1 + 2 * r4]
+%endrep
+ PROCESS_ADDAVG_32x2_AVX512
+ RET
+%endmacro
+
+ADDAVG_W32_AVX512 8
+ADDAVG_W32_AVX512 16
+ADDAVG_W32_AVX512 24
+ADDAVG_W32_AVX512 32
+ADDAVG_W32_AVX512 48
+ADDAVG_W32_AVX512 64
;-----------------------------------------------------------------------------
; addAvg avx512 code end
;-----------------------------------------------------------------------------
More information about the x265-devel
mailing list