[x265] [PATCH 202 of 307] x86: AVX512 interp_8tap_vert_ss_24x32
mythreyi at multicorewareinc.com
mythreyi at multicorewareinc.com
Sat Apr 7 04:33:20 CEST 2018
# HG changeset patch
# User Vignesh Vijayakumar<vignesh at multicorewareinc.com>
# Date 1511776850 -19800
# Mon Nov 27 15:30:50 2017 +0530
# Node ID 06cda1cbd637555e9c243623a2ccb7f0158554bc
# Parent 10ab896094d6e96495e76dfd24c46886cc49c3f4
x86: AVX512 interp_8tap_vert_ss_24x32
AVX2 performance : 10.98x
AVX512 performance : 16.36x
diff -r 10ab896094d6 -r 06cda1cbd637 source/common/x86/asm-primitives.cpp
--- a/source/common/x86/asm-primitives.cpp Mon Nov 27 15:24:06 2017 +0530
+++ b/source/common/x86/asm-primitives.cpp Mon Nov 27 15:30:50 2017 +0530
@@ -4968,6 +4968,7 @@
p.pu[LUMA_16x16].luma_vss = PFX(interp_8tap_vert_ss_16x16_avx512);
p.pu[LUMA_16x32].luma_vss = PFX(interp_8tap_vert_ss_16x32_avx512);
p.pu[LUMA_16x64].luma_vss = PFX(interp_8tap_vert_ss_16x64_avx512);
+ p.pu[LUMA_24x32].luma_vss = PFX(interp_8tap_vert_ss_24x32_avx512);
p.pu[LUMA_32x64].luma_vss = PFX(interp_8tap_vert_ss_32x64_avx512);
p.pu[LUMA_32x32].luma_vss = PFX(interp_8tap_vert_ss_32x32_avx512);
p.pu[LUMA_32x24].luma_vss = PFX(interp_8tap_vert_ss_32x24_avx512);
diff -r 10ab896094d6 -r 06cda1cbd637 source/common/x86/ipfilter8.asm
--- a/source/common/x86/ipfilter8.asm Mon Nov 27 15:24:06 2017 +0530
+++ b/source/common/x86/ipfilter8.asm Mon Nov 27 15:30:50 2017 +0530
@@ -12945,6 +12945,243 @@
FILTER_VER_SS_LUMA_16xN_AVX512 64
%endif
+%macro PROCESS_LUMA_VERT_SS_24x8_AVX512 0
+ PROCESS_LUMA_VERT_SS_16x4_AVX512
+ lea r4, [r6 + 4 * r1]
+ lea r8, [r4 + 4 * r1]
+ movu ym1, [r6]
+ movu ym3, [r6 + r1]
+ vinserti32x8 m1, [r6 + 2 * r1], 1
+ vinserti32x8 m3, [r6 + r7], 1
+ punpcklwd m0, m1, m3
+ pmaddwd m0, m15
+ punpckhwd m1, m3
+ pmaddwd m1, m15
+
+ movu ym4, [r6 + 2 * r1]
+ vinserti32x8 m4, [r4], 1
+ punpcklwd m2, m3, m4
+ pmaddwd m2, m15
+ punpckhwd m3, m4
+ pmaddwd m3, m15
+
+ movu ym5, [r6 + r7]
+ vinserti32x8 m5, [r4 + r1], 1
+ punpcklwd m6, m4, m5
+ pmaddwd m6, m16
+ punpckhwd m4, m5
+ pmaddwd m4, m16
+
+ paddd m0, m6
+ paddd m1, m4
+
+ movu ym4, [r4]
+ vinserti32x8 m4, [r4 + 2 * r1], 1
+ punpcklwd m6, m5, m4
+ pmaddwd m6, m16
+ punpckhwd m5, m4
+ pmaddwd m5, m16
+
+ paddd m2, m6
+ paddd m3, m5
+
+ movu ym11, [r4 + r1]
+ vinserti32x8 m11, [r4 + r7], 1
+ punpcklwd m8, m4, m11
+ pmaddwd m8, m17
+ punpckhwd m4, m11
+ pmaddwd m4, m17
+
+ movu ym12, [r4 + 2 * r1]
+ vinserti32x8 m12, [r4 + 4 * r1], 1
+ punpcklwd m10, m11, m12
+ pmaddwd m10, m17
+ punpckhwd m11, m12
+ pmaddwd m11, m17
+
+ movu ym13, [r4 + r7]
+ vinserti32x8 m13, [r8 + r1], 1
+ punpcklwd m14, m12, m13
+ pmaddwd m14, m18
+ punpckhwd m12, m13
+ pmaddwd m12, m18
+
+ paddd m8, m14
+ paddd m4, m12
+ paddd m0, m8
+ paddd m1, m4
+
+ movu ym12, [r4 + 4 * r1]
+ vinserti32x8 m12, [r8 + 2 * r1], 1
+ punpcklwd m14, m13, m12
+ pmaddwd m14, m18
+ punpckhwd m13, m12
+ pmaddwd m13, m18
+
+ paddd m10, m14
+ paddd m11, m13
+ paddd m2, m10
+ paddd m3, m11
+
+ psrad m0, 6
+ psrad m1, 6
+ psrad m2, 6
+ psrad m3, 6
+
+ packssdw m0, m1
+ packssdw m2, m3
+
+ lea r9, [r2 + 4 * r3]
+ movu [r9], ym0
+ movu [r9 + r3], ym2
+ vextracti32x8 [r9 + 2 * r3], m0, 1
+ vextracti32x8 [r9 + r5], m2, 1
+
+ movu xm1, [r0 + mmsize/2]
+ vinserti32x4 m1, [r0 + 2 * r1 + mmsize/2], 1
+ vinserti32x4 m1, [r0 + 4 * r1 + mmsize/2], 2
+ vinserti32x4 m1, [r6 + 2 * r1 + mmsize/2], 3
+ movu xm3, [r0 + r1 + mmsize/2]
+ vinserti32x4 m3, [r0 + r7 + mmsize/2], 1
+ vinserti32x4 m3, [r6 + r1 + mmsize/2], 2
+ vinserti32x4 m3, [r6 + r7 + mmsize/2], 3
+ punpcklwd m0, m1, m3
+ pmaddwd m0, m15
+ punpckhwd m1, m3
+ pmaddwd m1, m15
+
+ movu xm4, [r0 + 2 * r1 + mmsize/2]
+ vinserti32x4 m4, [r0 + 4 * r1 + mmsize/2], 1
+ vinserti32x4 m4, [r6 + 2 * r1 + mmsize/2], 2
+ vinserti32x4 m4, [r6 + 4 * r1 + mmsize/2], 3
+ punpcklwd m2, m3, m4
+ pmaddwd m2, m15
+ punpckhwd m3, m4
+ pmaddwd m3, m15
+
+ movu xm5, [r0 + r7 + mmsize/2]
+ vinserti32x4 m5, [r6 + r1 + mmsize/2], 1
+ vinserti32x4 m5, [r6 + r7 + mmsize/2], 2
+ vinserti32x4 m5, [r4 + r1 + mmsize/2], 3
+ punpcklwd m6, m4, m5
+ pmaddwd m6, m16
+ punpckhwd m4, m5
+ pmaddwd m4, m16
+
+ paddd m0, m6
+ paddd m1, m4
+
+ movu xm4, [r0 + 4 * r1 + mmsize/2]
+ vinserti32x4 m4, [r6 + 2 * r1 + mmsize/2], 1
+ vinserti32x4 m4, [r6 + 4 * r1 + mmsize/2], 2
+ vinserti32x4 m4, [r4 + 2 * r1 + mmsize/2], 3
+ punpcklwd m6, m5, m4
+ pmaddwd m6, m16
+ punpckhwd m5, m4
+ pmaddwd m5, m16
+
+ paddd m2, m6
+ paddd m3, m5
+
+ movu xm11, [r6 + r1 + mmsize/2]
+ vinserti32x4 m11, [r6 + r7 + mmsize/2], 1
+ vinserti32x4 m11, [r4 + r1 + mmsize/2], 2
+ vinserti32x4 m11, [r4 + r7 + mmsize/2], 3
+ punpcklwd m8, m4, m11
+ pmaddwd m8, m17
+ punpckhwd m4, m11
+ pmaddwd m4, m17
+
+ movu xm12, [r6 + 2 * r1 + mmsize/2]
+ vinserti32x4 m12, [r6 + 4 * r1 + mmsize/2], 1
+ vinserti32x4 m12, [r4 + 2 * r1 + mmsize/2], 2
+ vinserti32x4 m12, [r4 + 4 * r1 + mmsize/2], 3
+ punpcklwd m10, m11, m12
+ pmaddwd m10, m17
+ punpckhwd m11, m12
+ pmaddwd m11, m17
+
+ movu xm13, [r6 + r7 + mmsize/2]
+ vinserti32x4 m13, [r4 + r1 + mmsize/2], 1
+ vinserti32x4 m13, [r4 + r7 + mmsize/2], 2
+ vinserti32x4 m13, [r8 + r1 + mmsize/2], 3
+ punpcklwd m14, m12, m13
+ pmaddwd m14, m18
+ punpckhwd m12, m13
+ pmaddwd m12, m18
+
+ paddd m8, m14
+ paddd m4, m12
+ paddd m0, m8
+ paddd m1, m4
+
+ movu xm12, [r6 + 4 * r1 + mmsize/2]
+ vinserti32x4 m12, [r4 + 2 * r1 + mmsize/2], 1
+ vinserti32x4 m12, [r4 + 4 * r1 + mmsize/2], 2
+ vinserti32x4 m12, [r8 + 2 * r1 + mmsize/2], 3
+ punpcklwd m14, m13, m12
+ pmaddwd m14, m18
+ punpckhwd m13, m12
+ pmaddwd m13, m18
+
+ paddd m10, m14
+ paddd m11, m13
+ paddd m2, m10
+ paddd m3, m11
+
+ psrad m0, 6
+ psrad m1, 6
+ psrad m2, 6
+ psrad m3, 6
+
+ packssdw m0, m1
+ packssdw m2, m3
+
+ movu [r2 + mmsize/2], xm0
+ movu [r2 + r3 + mmsize/2], xm2
+ vextracti32x4 [r2 + 2 * r3 + mmsize/2], m0, 1
+ vextracti32x4 [r2 + r5 + mmsize/2], m2, 1
+ lea r2, [r2 + 4 * r3]
+ vextracti32x4 [r2 + mmsize/2], m0, 2
+ vextracti32x4 [r2 + r3 + mmsize/2], m2, 2
+ vextracti32x4 [r2 + 2 * r3 + mmsize/2], m0, 3
+ vextracti32x4 [r2 + r5 + mmsize/2], m2, 3
+%endmacro
+;-----------------------------------------------------------------------------------------------------------------
+; void interp_8tap_vert(int16_t *src, intptr_t srcStride, int16_t *dst, intptr_t dstStride, int coeffIdx)
+;-----------------------------------------------------------------------------------------------------------------
+%if ARCH_X86_64
+INIT_ZMM avx512
+cglobal interp_8tap_vert_ss_24x32, 5, 10, 19
+ add r1d, r1d
+ add r3d, r3d
+ lea r7, [3 * r1]
+ sub r0, r7
+ shl r4d, 8
+%ifdef PIC
+ lea r5, [pw_LumaCoeffVer_avx512]
+ mova m15, [r5 + r4]
+ mova m16, [r5 + r4 + 1 * mmsize]
+ mova m17, [r5 + r4 + 2 * mmsize]
+ mova m18, [r5 + r4 + 3 * mmsize]
+%else
+ lea r5, [pw_LumaCoeffVer_avx512 + r4]
+ mova m15, [r5]
+ mova m16, [r5 + 1 * mmsize]
+ mova m17, [r5 + 2 * mmsize]
+ mova m18, [r5 + 3 * mmsize]
+%endif
+
+ lea r5, [3 * r3]
+%rep 3
+ PROCESS_LUMA_VERT_SS_24x8_AVX512
+ lea r0, [r4]
+ lea r2, [r2 + 4 * r3]
+%endrep
+ PROCESS_LUMA_VERT_SS_24x8_AVX512
+ RET
+%endif
+
%macro PROCESS_LUMA_VERT_SS_32x2_AVX512 0
movu m1, [r0] ;0 row
movu m3, [r0 + r1] ;1 row
More information about the x265-devel
mailing list