[x265] [PATCH 298 of 307] X86:AVX512 intra_pred_ang16 mode 6 and 30 high bit depth

mythreyi at multicorewareinc.com mythreyi at multicorewareinc.com
Sat Apr 7 04:34:56 CEST 2018


# HG changeset patch
# User Jayashree
# Date 1516097338 -19800
#      Tue Jan 16 15:38:58 2018 +0530
# Node ID b0d00ca83af0cb2053d6eda82b6d4081236a0f5f
# Parent  ce088a0f4c0ede8ff2e5f0ed6faa005c280acd84
X86:AVX512 intra_pred_ang16 mode 6 and 30 high bit depth
Mode | AVX2 performance | AVX512 performance
---------------------------------------------------
 6   |    10.6x         |    17.5x
 30  |    12.62x        |    21.0x

diff -r ce088a0f4c0e -r b0d00ca83af0 source/common/x86/asm-primitives.cpp
--- a/source/common/x86/asm-primitives.cpp	Tue Jan 16 14:55:31 2018 +0530
+++ b/source/common/x86/asm-primitives.cpp	Tue Jan 16 15:38:58 2018 +0530
@@ -3119,7 +3119,8 @@
         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.cu[BLOCK_16x16].intra_pred[6] = PFX(intra_pred_ang16_6_avx512);
+        p.cu[BLOCK_16x16].intra_pred[30] = PFX(intra_pred_ang16_30_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 ce088a0f4c0e -r b0d00ca83af0 source/common/x86/intrapred16.asm
--- a/source/common/x86/intrapred16.asm	Tue Jan 16 14:55:31 2018 +0530
+++ b/source/common/x86/intrapred16.asm	Tue Jan 16 15:38:58 2018 +0530
@@ -19996,9 +19996,29 @@
     call        ang16_mode_6_30
 
     add         r2,        12
-
     call        ang32_mode_6_30
     RET
+cglobal intra_pred_ang16_6, 3,7,14
+    add         r2,        64
+    xor         r6d,       r6d
+    vbroadcasti32x8  m15,  [pd_16]
+    lea         r3,        [ang_table_avx2 + 15 * 32]
+    shl         r1d,       1
+    lea         r4,        [r1 * 3]
+
+    call        ang16_mode_6_30
+    RET
+
+cglobal intra_pred_ang16_30, 3,7,14
+    xor         r6d,       r6d
+    inc         r6d
+    vbroadcasti32x8  m15,  [pd_16]
+    lea         r3,        [ang_table_avx2 + 15 * 32]
+     shl         r1d,       1
+    lea         r4,        [r1 * 3]
+
+    call        ang16_mode_6_30
+    RET
 ;-------------------------------------------------------------------------------------------------------
 ; avx512 code for intra_pred_ang32 mode 2 to 34 end
 ;-------------------------------------------------------------------------------------------------------


More information about the x265-devel mailing list