[x265] [PATCH 072 of 307] x86: AVX512 cpy2Dto1D_shl_32 and cpy2Dto1D_shl_16

[mythreyi at multicorewareinc.com]

# HG changeset patch
# User Vignesh Vijayakumar
# Date 1501663291 -19800
#      Wed Aug 02 14:11:31 2017 +0530
# Node ID ce93c1b1894ae7d789e451f65479f018ba90ec76
# Parent  aac415b7223acced7fc844c4a07225704b811df0
x86: AVX512 cpy2Dto1D_shl_32 and cpy2Dto1D_shl_16

Size | BitDepth | AVX2 performance | AVX512 performance
-------------------------------------------------------
16x16|    8     |     15.09x       |       21.16
16x16|    10    |     16.05x       |       17.86
32x32|    8     |     13.90x       |       25.62
32x32|    10    |     11.69x       |       23.24

diff -r aac415b7223a -r ce93c1b1894a source/common/x86/asm-primitives.cpp
--- a/source/common/x86/asm-primitives.cpp	Tue Aug 01 17:37:05 2017 +0530
+++ b/source/common/x86/asm-primitives.cpp	Wed Aug 02 14:11:31 2017 +0530
@@ -2309,6 +2309,8 @@
         p.pu[LUMA_32x32].sad_x4 = PFX(pixel_sad_x4_32x32_avx512);
         p.pu[LUMA_32x64].sad_x4 = PFX(pixel_sad_x4_32x64_avx512);
 
+        p.cu[BLOCK_16x16].cpy2Dto1D_shl = PFX(cpy2Dto1D_shl_16_avx512);
+        p.cu[BLOCK_32x32].cpy2Dto1D_shl = PFX(cpy2Dto1D_shl_32_avx512);
     }
 }
 #else // if HIGH_BIT_DEPTH
@@ -3988,6 +3990,8 @@
         p.cu[BLOCK_64x64].copy_ss = PFX(blockcopy_ss_64x64_avx512);
 
         p.cu[BLOCK_32x32].calcresidual = PFX(getResidual32_avx512);
+        p.cu[BLOCK_16x16].cpy2Dto1D_shl = PFX(cpy2Dto1D_shl_16_avx512);
+        p.cu[BLOCK_32x32].cpy2Dto1D_shl = PFX(cpy2Dto1D_shl_32_avx512);
 
     }
 #endif
diff -r aac415b7223a -r ce93c1b1894a source/common/x86/blockcopy8.asm
--- a/source/common/x86/blockcopy8.asm	Tue Aug 01 17:37:05 2017 +0530
+++ b/source/common/x86/blockcopy8.asm	Wed Aug 02 14:11:31 2017 +0530
@@ -6140,6 +6140,102 @@
     RET
 
 ;--------------------------------------------------------------------------------------
+; cpy2Dto1D_shl avx512 code start
+;--------------------------------------------------------------------------------------
+%macro PROCESS_CPY2Dto1D_SHL_16x8_AVX512 0
+    movu             m1,    [r1]
+    vinserti32x8     m1,    [r1 + r2], 1
+    movu             m2,    [r1 + 2 * r2]
+    vinserti32x8     m2,    [r1 + r3], 1
+
+    psllw    m1, xm0
+    psllw    m2, xm0
+    movu     [r0], m1
+    movu     [r0 + mmsize], m2
+
+    add      r0, 2 * mmsize
+    lea      r1, [r1 + r2 * 4]
+
+    movu             m1,    [r1]
+    vinserti32x8     m1,    [r1 + r2], 1
+    movu             m2,    [r1 + 2 * r2]
+    vinserti32x8     m2,    [r1 + r3], 1
+
+    psllw    m1, xm0
+    psllw    m2, xm0
+    movu     [r0], m1
+    movu     [r0 + mmsize], m2
+%endmacro
+
+%macro PROCESS_CPY2Dto1D_SHL_32x8_AVX512 0
+    movu     m1, [r1]
+    movu     m2, [r1 + r2]
+    movu     m3, [r1 + 2 * r2]
+    movu     m4, [r1 + r3]
+
+    psllw    m1, xm0
+    psllw    m2, xm0
+    psllw    m3, xm0
+    psllw    m4, xm0
+    movu     [r0], m1
+    movu     [r0 + mmsize], m2
+    movu     [r0 + 2 * mmsize], m3
+    movu     [r0 + 3 * mmsize], m4
+
+    add      r0, 4 * mmsize
+    lea      r1, [r1 + r2 * 4]
+
+    movu     m1, [r1]
+    movu     m2, [r1 + r2]
+    movu     m3, [r1 + 2 * r2]
+    movu     m4, [r1 + r3]
+
+    psllw    m1, xm0
+    psllw    m2, xm0
+    psllw    m3, xm0
+    psllw    m4, xm0
+    movu     [r0], m1
+    movu     [r0 + mmsize], m2
+    movu     [r0 + 2 * mmsize], m3
+    movu     [r0 + 3 * mmsize], m4
+%endmacro
+
+;--------------------------------------------------------------------------------------
+; void cpy2Dto1D_shl(int16_t* dst, const int16_t* src, intptr_t srcStride, int shift);
+;--------------------------------------------------------------------------------------
+INIT_ZMM avx512
+cglobal cpy2Dto1D_shl_32, 4, 4, 5
+    add     r2d, r2d
+    movd    xm0, r3d
+    lea     r3, [3 * r2]
+
+    PROCESS_CPY2Dto1D_SHL_32x8_AVX512
+    add      r0, 4 * mmsize
+    lea      r1, [r1 + r2 * 4]
+    PROCESS_CPY2Dto1D_SHL_32x8_AVX512
+    add      r0, 4 * mmsize
+    lea      r1, [r1 + r2 * 4]
+    PROCESS_CPY2Dto1D_SHL_32x8_AVX512
+    add      r0, 4 * mmsize
+    lea      r1, [r1 + r2 * 4]
+    PROCESS_CPY2Dto1D_SHL_32x8_AVX512
+    RET
+
+INIT_ZMM avx512
+cglobal cpy2Dto1D_shl_16, 4, 4, 3
+    add     r2d, r2d
+    movd    xm0, r3d
+    lea     r3, [3 * r2]
+
+    PROCESS_CPY2Dto1D_SHL_16x8_AVX512
+    add      r0, 2 * mmsize
+    lea      r1, [r1 + r2 * 4]
+    PROCESS_CPY2Dto1D_SHL_16x8_AVX512
+    RET
+;--------------------------------------------------------------------------------------
+; cpy2Dto1D_shl avx512 code end
+;--------------------------------------------------------------------------------------
+;--------------------------------------------------------------------------------------
 ; void cpy1Dto2D_shr(int16_t* dst, const int16_t* src, intptr_t dstStride, int shift)
 ;--------------------------------------------------------------------------------------
 INIT_XMM sse2
diff -r aac415b7223a -r ce93c1b1894a source/common/x86/blockcopy8.h
--- a/source/common/x86/blockcopy8.h	Tue Aug 01 17:37:05 2017 +0530
+++ b/source/common/x86/blockcopy8.h	Wed Aug 02 14:11:31 2017 +0530
@@ -28,6 +28,7 @@
 FUNCDEF_TU_S(void, cpy2Dto1D_shl, sse2, int16_t* dst, const int16_t* src, intptr_t srcStride, int shift);
 FUNCDEF_TU_S(void, cpy2Dto1D_shl, sse4, int16_t* dst, const int16_t* src, intptr_t srcStride, int shift);
 FUNCDEF_TU_S(void, cpy2Dto1D_shl, avx2, int16_t* dst, const int16_t* src, intptr_t srcStride, int shift);
+FUNCDEF_TU_S(void, cpy2Dto1D_shl, avx512, int16_t* dst, const int16_t* src, intptr_t srcStride, int shift);
 
 FUNCDEF_TU_S(void, cpy2Dto1D_shr, sse2, int16_t* dst, const int16_t* src, intptr_t srcStride, int shift);
 FUNCDEF_TU_S(void, cpy2Dto1D_shr, sse4, int16_t* dst, const int16_t* src, intptr_t srcStride, int shift);