[x265] [PATCH 282 of 307] x86: AVX512 cpy1Dto2D_shr_32

[mythreyi at multicorewareinc.com]

# HG changeset patch
# User Vignesh Vijayakumar<vignesh at multicorewareinc.com>
# Date 1515407502 -19800
#      Mon Jan 08 16:01:42 2018 +0530
# Node ID 48917be3e409f917468ff2f73302b62afef492fb
# Parent  c9f8c315a900c488e41bb39955a1c893e35a66d4
x86: AVX512 cpy1Dto2D_shr_32

AVX2 performance   : 21.03x
AVX512 performance : 34.55x

diff -r c9f8c315a900 -r 48917be3e409 source/common/x86/asm-primitives.cpp
--- a/source/common/x86/asm-primitives.cpp	Mon Jan 08 15:55:34 2018 +0530
+++ b/source/common/x86/asm-primitives.cpp	Mon Jan 08 16:01:42 2018 +0530
@@ -2557,6 +2557,8 @@
         p.cu[BLOCK_32x32].cpy1Dto2D_shl[NONALIGNED] = PFX(cpy1Dto2D_shl_32_avx512);
         p.cu[BLOCK_32x32].cpy1Dto2D_shl[ALIGNED] = PFX(cpy1Dto2D_shl_aligned_32_avx512);
         p.cu[BLOCK_16x16].cpy1Dto2D_shr = PFX(cpy1Dto2D_shr_16_avx512);
+        p.cu[BLOCK_32x32].cpy1Dto2D_shr = PFX(cpy1Dto2D_shr_32_avx512);
+
         p.weight_pp = PFX(weight_pp_avx512);
         p.weight_sp = PFX(weight_sp_avx512);
         p.dequant_normal = PFX(dequant_normal_avx512);
@@ -4908,6 +4910,7 @@
         p.cu[BLOCK_32x32].cpy1Dto2D_shl[NONALIGNED] = PFX(cpy1Dto2D_shl_32_avx512);
         p.cu[BLOCK_32x32].cpy1Dto2D_shl[ALIGNED] = PFX(cpy1Dto2D_shl_aligned_32_avx512);
         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_32x32].copy_cnt = PFX(copy_cnt_32_avx512);
         p.cu[BLOCK_16x16].copy_cnt = PFX(copy_cnt_16_avx512);
diff -r c9f8c315a900 -r 48917be3e409 source/common/x86/blockcopy8.asm
--- a/source/common/x86/blockcopy8.asm	Mon Jan 08 15:55:34 2018 +0530
+++ b/source/common/x86/blockcopy8.asm	Mon Jan 08 16:01:42 2018 +0530
@@ -6809,3 +6809,30 @@
     dec         r3d
     jnz        .loop
     RET
+
+INIT_ZMM avx512
+cglobal cpy1Dto2D_shr_32, 3, 4, 6
+    shl                 r2d,             1
+    movd                xm0,             r3m
+    pcmpeqw             ymm1,            ymm1
+    psllw               ym1,             ymm1,       xm0
+    psraw               ym1,             1
+    vinserti32x8        m1,              ym1,        1
+    mov                 r3d,             16
+
+.loop:
+    ; Row 0-1
+    movu                m2,              [r1]
+    movu                m3,              [r1 + mmsize]
+    psubw               m2,              m1
+    psubw               m3,              m1
+    psraw               m2,              xm0
+    psraw               m3,              xm0
+    movu                [r0],            m2
+    movu                [r0 + r2],       m3
+
+    add                 r1,              2 * mmsize
+    lea                 r0,              [r0 + r2 * 2]
+    dec                 r3d
+    jnz                 .loop
+    RET