[x265] [PATCH 108 of 307] [x265-avx512]x86: AVX512 dequant_normal

[mythreyi at multicorewareinc.com]

# HG changeset patch
# User Gopi Satykrishna Akisetty <gopi.satykrishna at multicorewareinc.com>
# Date 1507109312 -19800
#      Wed Oct 04 14:58:32 2017 +0530
# Node ID 2221c70ef3b9b416b0ad491cd2325ccb595df8bb
# Parent  c726239a07580fd13c4177f0206d615ee02c5975
[x265-avx512]x86: AVX512 dequant_normal

AVX2 Performance    :    9.81x
AVX512 Performance  :   15.37x

diff -r c726239a0758 -r 2221c70ef3b9 source/common/x86/asm-primitives.cpp
--- a/source/common/x86/asm-primitives.cpp	Thu Aug 31 15:21:25 2017 +0530
+++ b/source/common/x86/asm-primitives.cpp	Wed Oct 04 14:58:32 2017 +0530
@@ -2350,6 +2350,7 @@
         p.cu[BLOCK_32x32].cpy2Dto1D_shl = PFX(cpy2Dto1D_shl_32_avx512);
         p.cu[BLOCK_32x32].cpy1Dto2D_shl = PFX(cpy1Dto2D_shl_32_avx512);
         p.weight_pp = PFX(weight_pp_avx512);
+        p.dequant_normal = PFX(dequant_normal_avx512);
 
         p.cu[BLOCK_32x32].copy_cnt = PFX(copy_cnt_32_avx512);
         p.cu[BLOCK_16x16].copy_cnt = PFX(copy_cnt_16_avx512);
@@ -4132,6 +4133,8 @@
         p.cu[BLOCK_32x32].copy_cnt = PFX(copy_cnt_32_avx512);
         p.cu[BLOCK_16x16].copy_cnt = PFX(copy_cnt_16_avx512);
 
+        p.dequant_normal = PFX(dequant_normal_avx512);
+
         //i444 chroma_hpp
         p.chroma[X265_CSP_I444].pu[LUMA_64x64].filter_hpp = PFX(interp_4tap_horiz_pp_64x64_avx512);
         p.chroma[X265_CSP_I444].pu[LUMA_64x32].filter_hpp = PFX(interp_4tap_horiz_pp_64x32_avx512);
diff -r c726239a0758 -r 2221c70ef3b9 source/common/x86/pixel-util8.asm
--- a/source/common/x86/pixel-util8.asm	Thu Aug 31 15:21:25 2017 +0530
+++ b/source/common/x86/pixel-util8.asm	Wed Oct 04 14:58:32 2017 +0530
@@ -1237,6 +1237,58 @@
     jnz            .loop
     RET
 
+INIT_ZMM avx512
+cglobal dequant_normal, 5,5,7
+    vpbroadcastd    m2, [pw_1]          ; m2 = word [1]
+    vpbroadcastd    m5, [pd_32767]      ; m5 = dword [32767]
+    vpbroadcastd    m6, [pd_n32768]     ; m6 = dword [-32768]
+%if HIGH_BIT_DEPTH
+    cmp             r3d, 32767
+    jle            .skip
+    shr             r3d, (BIT_DEPTH - 8)
+    sub             r4d, (BIT_DEPTH - 8)
+.skip:
+%endif
+    movd            xm0, r4d            ; m0 = shift
+    add             r4d, -1+16
+    bts             r3d, r4d
+
+    movd            xm1, r3d
+    vpbroadcastd    m1, xm1             ; m1 = dword [add scale]
+
+    ; m0 = shift
+    ; m1 = scale
+    ; m2 = word [1]
+    mov             r3d, r2d
+    shr             r2d, 5
+.loop:
+    movu            m3, [r0]
+    punpckhwd       m4, m3, m2
+    punpcklwd       m3, m2
+    pmaddwd         m3, m1              ; m3 = dword (clipQCoef * scale + add)
+    pmaddwd         m4, m1
+    psrad           m3, xm0
+    psrad           m4, xm0
+    pminsd          m3, m5
+    pmaxsd          m3, m6
+    pminsd          m4, m5
+    pmaxsd          m4, m6
+    packssdw        m3, m4
+
+    mova             [r1 + 0 * mmsize/2], ym3
+    cmp              r3d, 16
+    je               .num16
+    vextracti32x8    [r1 + 1 * mmsize/2], m3, 1
+
+    add             r0, mmsize
+    add             r1, mmsize
+
+    dec             r2d
+    jnz            .loop
+    RET
+.num16:
+    RET
+
 
 ;-----------------------------------------------------------------------------
 ; int x265_count_nonzero_4x4_sse2(const int16_t *quantCoeff);