[x265] [PATCH 097 of 307] x86: AVX512 convert_p2s link 32xN and 64xN chroma_444 primitives

[mythreyi at multicorewareinc.com]

# HG changeset patch
# User Vignesh Vijayakumar
# Date 1503901717 -19800
#      Mon Aug 28 11:58:37 2017 +0530
# Node ID bf199a5eca5be148be8a0c91cd9f2e8e0e908059
# Parent  0355f0128b7d713c4a21c91d3cc5bed1e8b43c47
x86: AVX512 convert_p2s link 32xN and 64xN chroma_444 primitives

diff -r 0355f0128b7d -r bf199a5eca5b source/common/x86/asm-primitives.cpp
--- a/source/common/x86/asm-primitives.cpp	Thu Aug 24 12:20:07 2017 +0530
+++ b/source/common/x86/asm-primitives.cpp	Mon Aug 28 11:58:37 2017 +0530
@@ -2253,6 +2253,15 @@
         p.chroma[X265_CSP_I422].pu[CHROMA_422_32x32].p2s = PFX(filterPixelToShort_32x32_avx512);
         p.chroma[X265_CSP_I422].pu[CHROMA_422_32x48].p2s = PFX(filterPixelToShort_32x48_avx512);
         p.chroma[X265_CSP_I422].pu[CHROMA_422_32x64].p2s = PFX(filterPixelToShort_32x64_avx512);
+        p.chroma[X265_CSP_I444].pu[LUMA_32x8].p2s = PFX(filterPixelToShort_32x8_avx2);
+        p.chroma[X265_CSP_I444].pu[LUMA_32x16].p2s = PFX(filterPixelToShort_32x16_avx512);
+        p.chroma[X265_CSP_I444].pu[LUMA_32x24].p2s = PFX(filterPixelToShort_32x24_avx512);
+        p.chroma[X265_CSP_I444].pu[LUMA_32x32].p2s = PFX(filterPixelToShort_32x32_avx512);
+        p.chroma[X265_CSP_I444].pu[LUMA_32x64].p2s = PFX(filterPixelToShort_32x64_avx512);
+        p.chroma[X265_CSP_I444].pu[LUMA_64x16].p2s = PFX(filterPixelToShort_64x16_avx512);
+        p.chroma[X265_CSP_I444].pu[LUMA_64x32].p2s = PFX(filterPixelToShort_64x32_avx512);
+        p.chroma[X265_CSP_I444].pu[LUMA_64x48].p2s = PFX(filterPixelToShort_64x48_avx512);
+        p.chroma[X265_CSP_I444].pu[LUMA_64x64].p2s = PFX(filterPixelToShort_64x64_avx512);
 
         p.cu[BLOCK_32x32].ssd_s = PFX(pixel_ssd_s_32_avx512);
 
@@ -4041,6 +4050,15 @@
         p.chroma[X265_CSP_I422].pu[CHROMA_422_32x32].p2s = PFX(filterPixelToShort_32x32_avx512);
         p.chroma[X265_CSP_I422].pu[CHROMA_422_32x48].p2s = PFX(filterPixelToShort_32x48_avx512);
         p.chroma[X265_CSP_I422].pu[CHROMA_422_32x64].p2s = PFX(filterPixelToShort_32x64_avx512);
+        p.chroma[X265_CSP_I444].pu[LUMA_32x8].p2s = PFX(filterPixelToShort_32x8_avx2);
+        p.chroma[X265_CSP_I444].pu[LUMA_32x16].p2s = PFX(filterPixelToShort_32x16_avx512);
+        p.chroma[X265_CSP_I444].pu[LUMA_32x24].p2s = PFX(filterPixelToShort_32x24_avx512);
+        p.chroma[X265_CSP_I444].pu[LUMA_32x32].p2s = PFX(filterPixelToShort_32x32_avx512);
+        p.chroma[X265_CSP_I444].pu[LUMA_32x64].p2s = PFX(filterPixelToShort_32x64_avx512);
+        p.chroma[X265_CSP_I444].pu[LUMA_64x16].p2s = PFX(filterPixelToShort_64x16_avx512);
+        p.chroma[X265_CSP_I444].pu[LUMA_64x32].p2s = PFX(filterPixelToShort_64x32_avx512);
+        p.chroma[X265_CSP_I444].pu[LUMA_64x48].p2s = PFX(filterPixelToShort_64x48_avx512);
+        p.chroma[X265_CSP_I444].pu[LUMA_64x64].p2s = PFX(filterPixelToShort_64x64_avx512);
 
         p.cu[BLOCK_64x64].sse_ss = (pixel_sse_ss_t)PFX(pixel_ssd_ss_64x64_avx512);
         p.cu[BLOCK_32x32].sse_ss = (pixel_sse_ss_t)PFX(pixel_ssd_ss_32x32_avx512);