[x265] [PATCH 224 of 307] x86: AVX512 interp_8tap_vert_sp_16xN
mythreyi at multicorewareinc.com
mythreyi at multicorewareinc.com
Sat Apr 7 04:33:42 CEST 2018
# HG changeset patch
# User Vignesh Vijayakumar<vignesh at multicorewareinc.com>
# Date 1512018320 -19800
# Thu Nov 30 10:35:20 2017 +0530
# Node ID a78e09e144582bd52c52d3475aa1922fc2ae8893
# Parent 3e14c3f607d0f9ec6dd3735d21fc2e698217fe71
x86: AVX512 interp_8tap_vert_sp_16xN
Size | AVX2 performance | AVX512 performance
----------------------------------------------
16x4 | 9.68x | 16.45x
16x8 | 11.69x | 16.93x
16x16 | 13.26x | 18.58x
16x32 | 12.96x | 19.23x
16x64 | 13.12x | 16.84x
diff -r 3e14c3f607d0 -r a78e09e14458 source/common/x86/asm-primitives.cpp
--- a/source/common/x86/asm-primitives.cpp Thu Nov 30 16:00:14 2017 +0530
+++ b/source/common/x86/asm-primitives.cpp Thu Nov 30 10:35:20 2017 +0530
@@ -5002,7 +5002,12 @@
p.pu[LUMA_64x48].luma_vpp = PFX(interp_8tap_vert_pp_64x48_avx512);
p.pu[LUMA_64x32].luma_vpp = PFX(interp_8tap_vert_pp_64x32_avx512);
p.pu[LUMA_64x16].luma_vpp = PFX(interp_8tap_vert_pp_64x16_avx512);
-
+ p.pu[LUMA_16x4].luma_vsp = PFX(interp_8tap_vert_sp_16x4_avx512);
+ p.pu[LUMA_16x8].luma_vsp = PFX(interp_8tap_vert_sp_16x8_avx512);
+ p.pu[LUMA_16x12].luma_vsp = PFX(interp_8tap_vert_sp_16x12_avx512);
+ p.pu[LUMA_16x16].luma_vsp = PFX(interp_8tap_vert_sp_16x16_avx512);
+ p.pu[LUMA_16x32].luma_vsp = PFX(interp_8tap_vert_sp_16x32_avx512);
+ p.pu[LUMA_16x64].luma_vsp = PFX(interp_8tap_vert_sp_16x64_avx512);
p.pu[LUMA_32x64].luma_vsp = PFX(interp_8tap_vert_sp_32x64_avx512);
p.pu[LUMA_32x32].luma_vsp = PFX(interp_8tap_vert_sp_32x32_avx512);
p.pu[LUMA_32x24].luma_vsp = PFX(interp_8tap_vert_sp_32x24_avx512);
diff -r 3e14c3f607d0 -r a78e09e14458 source/common/x86/ipfilter8.asm
--- a/source/common/x86/ipfilter8.asm Thu Nov 30 16:00:14 2017 +0530
+++ b/source/common/x86/ipfilter8.asm Thu Nov 30 10:35:20 2017 +0530
@@ -12985,8 +12985,7 @@
FILTER_VER_SS_LUMA_8xN_AVX512 16
FILTER_VER_SS_LUMA_8xN_AVX512 32
%endif
-
-%macro PROCESS_LUMA_VERT_SS_16x4_AVX512 0
+%macro PROCESS_LUMA_VERT_S_16x4_AVX512 1
movu ym1, [r0]
movu ym3, [r0 + r1]
vinserti32x8 m1, [r0 + 2 * r1], 1
@@ -13062,7 +13061,26 @@
paddd m11, m13
paddd m2, m10
paddd m3, m11
-
+%ifidn %1, sp
+ paddd m0, m19
+ paddd m1, m19
+ paddd m2, m19
+ paddd m3, m19
+
+ psrad m0, 12
+ psrad m1, 12
+ psrad m2, 12
+ psrad m3, 12
+
+ packssdw m0, m1
+ packssdw m2, m3
+ packuswb m0, m2
+ vpermq m0, m20, m0
+ movu [r2], xm0
+ vextracti32x4 [r2 + r3], m0, 2
+ vextracti32x4 [r2 + 2 * r3], m0, 1
+ vextracti32x4 [r2 + r5], m0, 3
+%else
psrad m0, 6
psrad m1, 6
psrad m2, 6
@@ -13075,15 +13093,15 @@
movu [r2 + r3], ym2
vextracti32x8 [r2 + 2 * r3], m0, 1
vextracti32x8 [r2 + r5], m2, 1
+%endif
%endmacro
;-----------------------------------------------------------------------------------------------------------------
; void interp_8tap_vert(int16_t *src, intptr_t srcStride, int16_t *dst, intptr_t dstStride, int coeffIdx)
;-----------------------------------------------------------------------------------------------------------------
-%macro FILTER_VER_SS_LUMA_16xN_AVX512 1
+%macro FILTER_VER_S_LUMA_16xN_AVX512 2
INIT_ZMM avx512
-cglobal interp_8tap_vert_ss_16x%1, 5, 8, 19
+cglobal interp_8tap_vert_%1_16x%2, 5, 8, 21
add r1d, r1d
- add r3d, r3d
lea r7, [3 * r1]
sub r0, r7
shl r4d, 8
@@ -13100,28 +13118,39 @@
mova m17, [r5 + 2 * mmsize]
mova m18, [r5 + 3 * mmsize]
%endif
+%ifidn %1, sp
+ vbroadcasti32x4 m19, [pd_526336]
+ mova m20, [interp8_vsp_store_avx512]
+%else
+ add r3d, r3d
+%endif
lea r5, [3 * r3]
-%rep %1/4 - 1
- PROCESS_LUMA_VERT_SS_16x4_AVX512
+%rep %2/4 - 1
+ PROCESS_LUMA_VERT_S_16x4_AVX512 %1
lea r0, [r0 + 4 * r1]
lea r2, [r2 + 4 * r3]
%endrep
- PROCESS_LUMA_VERT_SS_16x4_AVX512
+ PROCESS_LUMA_VERT_S_16x4_AVX512 %1
RET
%endmacro
%if ARCH_X86_64
- FILTER_VER_SS_LUMA_16xN_AVX512 4
- FILTER_VER_SS_LUMA_16xN_AVX512 8
- FILTER_VER_SS_LUMA_16xN_AVX512 12
- FILTER_VER_SS_LUMA_16xN_AVX512 16
- FILTER_VER_SS_LUMA_16xN_AVX512 32
- FILTER_VER_SS_LUMA_16xN_AVX512 64
-%endif
-
+ FILTER_VER_S_LUMA_16xN_AVX512 ss, 4
+ FILTER_VER_S_LUMA_16xN_AVX512 ss, 8
+ FILTER_VER_S_LUMA_16xN_AVX512 ss, 12
+ FILTER_VER_S_LUMA_16xN_AVX512 ss, 16
+ FILTER_VER_S_LUMA_16xN_AVX512 ss, 32
+ FILTER_VER_S_LUMA_16xN_AVX512 ss, 64
+ FILTER_VER_S_LUMA_16xN_AVX512 sp, 4
+ FILTER_VER_S_LUMA_16xN_AVX512 sp, 8
+ FILTER_VER_S_LUMA_16xN_AVX512 sp, 12
+ FILTER_VER_S_LUMA_16xN_AVX512 sp, 16
+ FILTER_VER_S_LUMA_16xN_AVX512 sp, 32
+ FILTER_VER_S_LUMA_16xN_AVX512 sp, 64
+%endif
%macro PROCESS_LUMA_VERT_SS_24x8_AVX512 0
- PROCESS_LUMA_VERT_SS_16x4_AVX512
+ PROCESS_LUMA_VERT_S_16x4_AVX512 ss
lea r4, [r6 + 4 * r1]
lea r8, [r4 + 4 * r1]
movu ym1, [r6]
More information about the x265-devel
mailing list