[x265] [PATCH 061 of 307] x86: AVX512 addAvg_W64 for high bit depth

[mythreyi at multicorewareinc.com]

# HG changeset patch
# User Vignesh Vijayakumar
# Date 1501588310 -19800
#      Tue Aug 01 17:21:50 2017 +0530
# Node ID 465b4925d622ba66e2536c9f79eaaffcdd26d5fc
# Parent  73ee464e136910a95d7b3070a1c736dedeaa6278
x86: AVX512 addAvg_W64 for high bit depth

Size  |  AVX2 performance | AVX512 performance
----------------------------------------------
64x16 |      11.13x       |      18.48x
64x32 |      11.04x       |      17.75x
64x48 |      10.97x       |      17.85x
64x64 |      10.93x       |      17.37x

diff -r 73ee464e1369 -r 465b4925d622 source/common/x86/asm-primitives.cpp
--- a/source/common/x86/asm-primitives.cpp	Tue Aug 01 16:45:51 2017 +0530
+++ b/source/common/x86/asm-primitives.cpp	Tue Aug 01 17:21:50 2017 +0530
@@ -2278,6 +2278,10 @@
         p.pu[LUMA_64x48].sad = PFX(pixel_sad_64x48_avx512);
         p.pu[LUMA_64x64].sad = PFX(pixel_sad_64x64_avx512);
 
+        p.pu[LUMA_64x16].addAvg = PFX(addAvg_64x16_avx512);
+        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);
diff -r 73ee464e1369 -r 465b4925d622 source/common/x86/mc-a.asm
--- a/source/common/x86/mc-a.asm	Tue Aug 01 16:45:51 2017 +0530
+++ b/source/common/x86/mc-a.asm	Tue Aug 01 17:21:50 2017 +0530
@@ -1738,6 +1738,80 @@
     movu        [r2 + r8],       m0
 %endmacro
 
+%macro PROCESS_ADDAVG_64x4_HBD_AVX512 0
+    movu        m0,              [r0]
+    movu        m1,              [r1]
+    paddw       m0,              m1
+    pmulhrsw    m0,              m3
+    paddw       m0,              m4
+    pmaxsw      m0,              m2
+    pminsw      m0,              m5
+    movu        [r2],            m0
+
+    movu        m0,              [r0 + mmsize]
+    movu        m1,              [r1 + mmsize]
+    paddw       m0,              m1
+    pmulhrsw    m0,              m3
+    paddw       m0,              m4
+    pmaxsw      m0,              m2
+    pminsw      m0,              m5
+    movu        [r2 + mmsize],   m0
+
+    movu        m0,              [r0 + r3]
+    movu        m1,              [r1 + r4]
+    paddw       m0,              m1
+    pmulhrsw    m0,              m3
+    paddw       m0,              m4
+    pmaxsw      m0,              m2
+    pminsw      m0,              m5
+    movu        [r2 + r5],       m0
+
+    movu        m0,              [r0 + r3 + mmsize]
+    movu        m1,              [r1 + r4 + mmsize]
+    paddw       m0,              m1
+    pmulhrsw    m0,              m3
+    paddw       m0,              m4
+    pmaxsw      m0,              m2
+    pminsw      m0,              m5
+    movu        [r2 + r5 + mmsize],       m0
+
+    movu        m0,              [r0 + 2 * r3]
+    movu        m1,              [r1 + 2 * r4]
+    paddw       m0,              m1
+    pmulhrsw    m0,              m3
+    paddw       m0,              m4
+    pmaxsw      m0,              m2
+    pminsw      m0,              m5
+    movu        [r2 + 2 * r5],   m0
+
+    movu        m0,              [r0 + 2 * r3 + mmsize]
+    movu        m1,              [r1 + 2 * r4 + mmsize]
+    paddw       m0,              m1
+    pmulhrsw    m0,              m3
+    paddw       m0,              m4
+    pmaxsw      m0,              m2
+    pminsw      m0,              m5
+    movu        [r2 + 2 * r5 + mmsize],   m0
+
+    movu        m0,              [r0 + r6]
+    movu        m1,              [r1 + r7]
+    paddw       m0,              m1
+    pmulhrsw    m0,              m3
+    paddw       m0,              m4
+    pmaxsw      m0,              m2
+    pminsw      m0,              m5
+    movu        [r2 + r8],       m0
+
+    movu        m0,              [r0 + r6 + mmsize]
+    movu        m1,              [r1 + r7 + mmsize]
+    paddw       m0,              m1
+    pmulhrsw    m0,              m3
+    paddw       m0,              m4
+    pmaxsw      m0,              m2
+    pminsw      m0,              m5
+    movu        [r2 + r8 + mmsize],       m0
+%endmacro
+
 ;-----------------------------------------------------------------------------
 ;void addAvg (int16_t* src0, int16_t* src1, pixel* dst, intptr_t src0Stride, intptr_t src1Stride, intptr_t dstStride)
 ;-----------------------------------------------------------------------------
@@ -1771,6 +1845,35 @@
 ADDAVG_W32_HBD_AVX512 32
 ADDAVG_W32_HBD_AVX512 48
 ADDAVG_W32_HBD_AVX512 64
+
+%macro ADDAVG_W64_HBD_AVX512 1
+INIT_ZMM avx512
+cglobal addAvg_64x%1, 6,9,6
+    vbroadcasti32x8        m4,              [pw_ %+ ADDAVG_ROUND]
+    vbroadcasti32x8        m5,              [pw_pixel_max]
+    vbroadcasti32x8        m3,              [pw_ %+ ADDAVG_FACTOR]
+    pxor        m2,              m2
+    add         r3,              r3
+    add         r4,              r4
+    add         r5,              r5
+    lea         r6,              [3 * r3]
+    lea         r7,              [3 * r4]
+    lea         r8,              [3 * r5]
+
+%rep %1/4 - 1
+    PROCESS_ADDAVG_64x4_HBD_AVX512
+    lea         r2,              [r2 + 4 * r5]
+    lea         r0,              [r0 + 4 * r3]
+    lea         r1,              [r1 + 4 * r4]
+%endrep
+    PROCESS_ADDAVG_64x4_HBD_AVX512
+    RET
+%endmacro
+
+ADDAVG_W64_HBD_AVX512 16
+ADDAVG_W64_HBD_AVX512 32
+ADDAVG_W64_HBD_AVX512 48
+ADDAVG_W64_HBD_AVX512 64
 ;-----------------------------------------------------------------------------
 ;addAvg avx512 high bit depth code end
 ;-----------------------------------------------------------------------------