[x265] [PATCH 284 of 307] x86: AVX512 cpy2Dto1D_shr_16

[mythreyi at multicorewareinc.com]

# HG changeset patch
# User Vignesh Vijayakumar<vignesh at multicorewareinc.com>
# Date 1515409019 -19800
#      Mon Jan 08 16:26:59 2018 +0530
# Node ID a5d29083237f28a944143862f980960c3f2b15ff
# Parent  6d1c8baabf78dba44a500f2039eace3e31abf69c
x86: AVX512 cpy2Dto1D_shr_16

AVX2 performance   : 21.14x
AVX512 performance : 28.50x

diff -r 6d1c8baabf78 -r a5d29083237f source/common/x86/asm-primitives.cpp
--- a/source/common/x86/asm-primitives.cpp	Mon Jan 08 16:14:29 2018 +0530
+++ b/source/common/x86/asm-primitives.cpp	Mon Jan 08 16:26:59 2018 +0530
@@ -2559,6 +2559,7 @@
         p.cu[BLOCK_16x16].cpy1Dto2D_shr = PFX(cpy1Dto2D_shr_16_avx512);
         p.cu[BLOCK_32x32].cpy1Dto2D_shr = PFX(cpy1Dto2D_shr_32_avx512);
 
+        p.cu[BLOCK_16x16].cpy2Dto1D_shr = PFX(cpy2Dto1D_shr_16_avx512);
         p.cu[BLOCK_32x32].cpy2Dto1D_shr = PFX(cpy2Dto1D_shr_32_avx512);
 
         p.weight_pp = PFX(weight_pp_avx512);
@@ -4914,6 +4915,7 @@
         p.cu[BLOCK_16x16].cpy1Dto2D_shr = PFX(cpy1Dto2D_shr_16_avx512);
         p.cu[BLOCK_32x32].cpy1Dto2D_shr = PFX(cpy1Dto2D_shr_32_avx512);
 
+        p.cu[BLOCK_16x16].cpy2Dto1D_shr = PFX(cpy2Dto1D_shr_16_avx512);
         p.cu[BLOCK_32x32].cpy2Dto1D_shr = PFX(cpy2Dto1D_shr_32_avx512);
 
         p.cu[BLOCK_32x32].copy_cnt = PFX(copy_cnt_32_avx512);
diff -r 6d1c8baabf78 -r a5d29083237f source/common/x86/blockcopy8.asm
--- a/source/common/x86/blockcopy8.asm	Mon Jan 08 16:14:29 2018 +0530
+++ b/source/common/x86/blockcopy8.asm	Mon Jan 08 16:26:59 2018 +0530
@@ -5199,6 +5199,53 @@
     jnz        .loop
     RET
 
+INIT_ZMM avx512
+cglobal cpy2Dto1D_shr_16, 4, 5, 4
+    shl                 r2d,             1
+    movd                xm0,             r3d
+    pcmpeqw             ymm1,            ymm1
+    psllw               ym1,             ymm1,        xm0
+    psraw               ym1,             1
+    vinserti32x8        m1,              ym1,         1
+    lea                 r3,              [r2 * 3]
+    mov                 r4d,             2
+
+.loop:
+    ; Row 0-1
+    movu                ym2,             [r1]
+    vinserti32x8        m2,              [r1 + r2],   1
+    psubw               m2,              m1
+    psraw               m2,              xm0
+    movu                [r0],            m2
+
+    ; Row 2-3
+    movu                ym2,             [r1 + 2 * r2]
+    vinserti32x8        m2,              [r1 + r3],   1
+    psubw               m2,              m1
+    psraw               m2,              xm0
+    movu                [r0 + mmsize],   m2
+
+    lea        r1, [r1 + 4 * r2]
+    ; Row 4-5
+
+    movu                ym2,             [r1]
+    vinserti32x8        m2,              [r1 + r2],   1
+    psubw               m2,              m1
+    psraw               m2,              xm0
+    movu                [r0 + 2 * mmsize], m2
+
+    ; Row 6-7
+    movu                ym2,             [r1 + 2 * r2]
+    vinserti32x8        m2,              [r1 + r3],   1
+    psubw               m2,              m1
+    psraw               m2,              xm0
+    movu                [r0 + 3 * mmsize], m2
+
+    add                 r0,              4 * mmsize
+    lea                 r1,              [r1 + 4 * r2]
+    dec                 r4d
+    jnz                 .loop
+    RET
 
 ;--------------------------------------------------------------------------------------
 ; void cpy2Dto1D_shr(int16_t* dst, const int16_t* src, intptr_t srcStride, int shift);