[x265] [PATCH 296 of 307] x86 : AVX512 intra_pred_ang16 mode 4 and 32 high bit depth

mythreyi at multicorewareinc.com mythreyi at multicorewareinc.com
Sat Apr 7 04:34:54 CEST 2018


# HG changeset patch
# User Jayashree
# Date 1516014532 -19800
#      Mon Jan 15 16:38:52 2018 +0530
# Node ID ab1c3fd4fc64e0239e4e632f03cd616e857718ed
# Parent  a2b347ed81f90ac82f59d891deba7fa876df7f62
x86 : AVX512 intra_pred_ang16 mode 4 and 32 high bit depth
Mode | AVX2 performance | AVX512 performance
---------------------------------------------------
 4   |    9.5x         |       14.2x
 32  |    12.x         |       19.7x

diff -r a2b347ed81f9 -r ab1c3fd4fc64 source/common/x86/asm-primitives.cpp
--- a/source/common/x86/asm-primitives.cpp	Mon Jan 15 17:52:26 2018 +0530
+++ b/source/common/x86/asm-primitives.cpp	Mon Jan 15 16:38:52 2018 +0530
@@ -3116,6 +3116,9 @@
         p.cu[BLOCK_16x16].intra_pred[27]    = PFX(intra_pred_ang16_27_avx512);
         p.cu[BLOCK_16x16].intra_pred[5] = PFX(intra_pred_ang16_5_avx512);
         p.cu[BLOCK_16x16].intra_pred[31] = PFX(intra_pred_ang16_31_avx512);
+        p.cu[BLOCK_16x16].intra_pred[4] = PFX(intra_pred_ang16_4_avx512);
+        p.cu[BLOCK_16x16].intra_pred[32] = PFX(intra_pred_ang16_32_avx512);
+
         p.pu[LUMA_64x64].luma_hvpp = interp_8tap_hv_pp_cpu<LUMA_64x64>;
         p.pu[LUMA_64x48].luma_hvpp = interp_8tap_hv_pp_cpu<LUMA_64x48>;
         p.pu[LUMA_64x32].luma_hvpp = interp_8tap_hv_pp_cpu<LUMA_64x32>;
diff -r a2b347ed81f9 -r ab1c3fd4fc64 source/common/x86/intrapred16.asm
--- a/source/common/x86/intrapred16.asm	Mon Jan 15 17:52:26 2018 +0530
+++ b/source/common/x86/intrapred16.asm	Mon Jan 15 16:38:52 2018 +0530
@@ -19659,6 +19659,27 @@
 
     call        ang32_mode_4_32
     RET
+cglobal intra_pred_ang16_4, 3,7,13
+    add         r2,        64
+    xor         r6d,       r6d
+    vbroadcasti32x8  m15,  [pd_16]
+    lea         r3,        [ang_table_avx2 + 18 * 32]
+    add         r1d,       r1d
+    lea         r4,        [r1 * 3]
+
+    call        ang16_mode_4_32
+    RET
+
+cglobal intra_pred_ang16_32, 3,7,13
+    xor         r6d,       r6d
+    inc         r6d
+    vbroadcasti32x8  m15,  [pd_16]
+    lea         r3,        [ang_table_avx2 + 18 * 32]
+    shl         r1d,       1
+    lea         r4,        [r1 * 3]
+
+    call        ang16_mode_4_32
+    RET
 
 ;-------------------------------------------------------------------------------------------------------
 ; avx512 code for intra_pred_ang32 mode 2 to 34 end


More information about the x265-devel mailing list