[x265] [PATCH 233 of 307] x86: AVX512 optimise interp_4tap_vert_ss_8xN

mythreyi at multicorewareinc.com mythreyi at multicorewareinc.com
Sat Apr 7 04:33:51 CEST 2018


# HG changeset patch
# User Vignesh Vijayakumar<vignesh at multicorewareinc.com>
# Date 1512380104 -19800
#      Mon Dec 04 15:05:04 2017 +0530
# Node ID ae75b2d09d10f28391d573507c13512360593386
# Parent  3e8615bc86537e07754a1c023ade702a837042a8
x86: AVX512 optimise interp_4tap_vert_ss_8xN

i444 8x4
AVX2 performance   : 10.61x
AVX512 performance : 18.93x

diff -r 3e8615bc8653 -r ae75b2d09d10 source/common/x86/asm-primitives.cpp
--- a/source/common/x86/asm-primitives.cpp	Mon Dec 04 14:23:30 2017 +0530
+++ b/source/common/x86/asm-primitives.cpp	Mon Dec 04 15:05:04 2017 +0530
@@ -4903,6 +4903,7 @@
         p.chroma[X265_CSP_I420].pu[CHROMA_420_32x24].filter_vpp = PFX(interp_4tap_vert_pp_32x24_avx512);
         p.chroma[X265_CSP_I420].pu[CHROMA_420_32x32].filter_vpp = PFX(interp_4tap_vert_pp_32x32_avx512);
 
+        p.chroma[X265_CSP_I420].pu[CHROMA_420_8x4].filter_vss = PFX(interp_4tap_vert_ss_8x4_avx512);
         p.chroma[X265_CSP_I420].pu[CHROMA_420_8x8].filter_vss = PFX(interp_4tap_vert_ss_8x8_avx512);
         p.chroma[X265_CSP_I420].pu[CHROMA_420_8x16].filter_vss = PFX(interp_4tap_vert_ss_8x16_avx512);
         p.chroma[X265_CSP_I420].pu[CHROMA_420_8x32].filter_vss = PFX(interp_4tap_vert_ss_8x32_avx512);
@@ -4938,7 +4939,9 @@
         p.chroma[X265_CSP_I422].pu[CHROMA_422_32x48].filter_vpp = PFX(interp_4tap_vert_pp_32x48_avx512);
         p.chroma[X265_CSP_I422].pu[CHROMA_422_32x64].filter_vpp = PFX(interp_4tap_vert_pp_32x64_avx512);
 
+        p.chroma[X265_CSP_I422].pu[CHROMA_422_8x4].filter_vss = PFX(interp_4tap_vert_ss_8x4_avx512);
         p.chroma[X265_CSP_I422].pu[CHROMA_422_8x8].filter_vss = PFX(interp_4tap_vert_ss_8x8_avx512);
+        p.chroma[X265_CSP_I422].pu[CHROMA_422_8x12].filter_vss = PFX(interp_4tap_vert_ss_8x12_avx512);
         p.chroma[X265_CSP_I422].pu[CHROMA_422_8x16].filter_vss = PFX(interp_4tap_vert_ss_8x16_avx512);
         p.chroma[X265_CSP_I422].pu[CHROMA_422_8x32].filter_vss = PFX(interp_4tap_vert_ss_8x32_avx512);
         p.chroma[X265_CSP_I422].pu[CHROMA_422_8x64].filter_vss = PFX(interp_4tap_vert_ss_8x64_avx512);
@@ -4979,6 +4982,7 @@
         p.chroma[X265_CSP_I444].pu[LUMA_64x32].filter_vpp = PFX(interp_4tap_vert_pp_64x32_avx512);
         p.chroma[X265_CSP_I444].pu[LUMA_64x16].filter_vpp = PFX(interp_4tap_vert_pp_64x16_avx512);
 
+        p.chroma[X265_CSP_I444].pu[LUMA_8x4].filter_vss = PFX(interp_4tap_vert_ss_8x4_avx512);
         p.chroma[X265_CSP_I444].pu[LUMA_8x8].filter_vss = PFX(interp_4tap_vert_ss_8x8_avx512);
         p.chroma[X265_CSP_I444].pu[LUMA_8x16].filter_vss = PFX(interp_4tap_vert_ss_8x16_avx512);
         p.chroma[X265_CSP_I444].pu[LUMA_8x32].filter_vss = PFX(interp_4tap_vert_ss_8x32_avx512);
diff -r 3e8615bc8653 -r ae75b2d09d10 source/common/x86/ipfilter8.asm
--- a/source/common/x86/ipfilter8.asm	Mon Dec 04 14:23:30 2017 +0530
+++ b/source/common/x86/ipfilter8.asm	Mon Dec 04 15:05:04 2017 +0530
@@ -11146,78 +11146,54 @@
 ;-------------------------------------------------------------------------------------------------------------
 ;avx512 chroma_vss code start
 ;-------------------------------------------------------------------------------------------------------------
-%macro PROCESS_CHROMA_VERT_SS_8x8_AVX512 0
+%macro PROCESS_CHROMA_VERT_SS_8x4_AVX512 0
+    lea                   r5,                 [r0 + 4 * r1]
     movu                  xm1,                [r0]
-    lea                   r6,                 [r0 + 2 * r1]
-    lea                   r8,                 [r0 + 4 * r1]
-    lea                   r9,                 [r8 + 2 * r1]
-    vinserti32x4          m1,                 [r6],                1
-    vinserti32x4          m1,                 [r8],                2
-    vinserti32x4          m1,                 [r9],                3
     movu                  xm3,                [r0 + r1]
-    vinserti32x4          m3,                 [r6 + r1],           1
-    vinserti32x4          m3,                 [r8 + r1],           2
-    vinserti32x4          m3,                 [r9 + r1],           3
+    vinserti32x4          m1,                 [r0 + r1],           1
+    vinserti32x4          m3,                 [r0 + 2 * r1],       1
+    vinserti32x4          m1,                 [r0 + 2 * r1],       2
+    vinserti32x4          m3,                 [r0 + r6],           2
+    vinserti32x4          m1,                 [r0 + r6],           3
+    vinserti32x4          m3,                 [r0 + 4 * r1],       3
+
     punpcklwd             m0,                 m1,                  m3
     pmaddwd               m0,                 m8
     punpckhwd             m1,                 m3
     pmaddwd               m1,                 m8
 
     movu                  xm4,                [r0 + 2 * r1]
-    vinserti32x4          m4,                 [r6 + 2 * r1],       1
-    vinserti32x4          m4,                 [r8 + 2 * r1],       2
-    vinserti32x4          m4,                 [r9 + 2 * r1],       3
-    punpcklwd             m2,                 m3,                  m4
-    pmaddwd               m2,                 m8
-    punpckhwd             m3,                 m4
-    pmaddwd               m3,                 m8
-
-    movu                  xm5,                [r0 + r10]
-    vinserti32x4          m5,                 [r6 + r10],          1
-    vinserti32x4          m5,                 [r8 + r10],          2
-    vinserti32x4          m5,                 [r9 + r10],          3
-    punpcklwd             m6,                 m4,                  m5
-    pmaddwd               m6,                 m9
-    paddd                 m0,                 m6
+    movu                  xm5,                [r0 + r6]
+    vinserti32x4          m4,                 [r0 + r6],           1
+    vinserti32x4          m5,                 [r5],                1
+    vinserti32x4          m4,                 [r5],                2
+    vinserti32x4          m5,                 [r5 + r1],           2
+    vinserti32x4          m4,                 [r5 + r1],           3
+    vinserti32x4          m5,                 [r5 + 2 * r1],       3
+
+    punpcklwd             m3,                 m4,                  m5
+    pmaddwd               m3,                 m9
     punpckhwd             m4,                 m5
     pmaddwd               m4,                 m9
+
+    paddd                 m0,                 m3
     paddd                 m1,                 m4
 
-    movu                  xm4,                [r0 + 4 * r1]
-    vinserti32x4          m4,                 [r6 + 4 * r1],       1
-    vinserti32x4          m4,                 [r8 + 4 * r1],       2
-    vinserti32x4          m4,                 [r9 + 4 * r1],       3
-    punpcklwd             m6,                 m5,                  m4
-    pmaddwd               m6,                 m9
-    paddd                 m2,                 m6
-    punpckhwd             m5,                 m4
-    pmaddwd               m5,                 m9
-    paddd                 m3,                 m5
-
     psrad                 m0,                 6
     psrad                 m1,                 6
-    psrad                 m2,                 6
-    psrad                 m3,                 6
     packssdw              m0,                 m1
-    packssdw              m2,                 m3
-
     movu                  [r2],               xm0
-    movu                  [r2 + r3],          xm2
-    vextracti32x4         [r2 + 2 * r3],      m0,                  1
-    vextracti32x4         [r2 + r7],          m2,                  1
-    lea                   r2,                 [r2 + 4 * r3]
-    vextracti32x4         [r2],               m0,                  2
-    vextracti32x4         [r2 + r3],          m2,                  2
-    vextracti32x4         [r2 + 2 * r3],      m0,                  3
-    vextracti32x4         [r2 + r7],          m2,                  3
+    vextracti32x4         [r2 + r3],          m0,                  1
+    vextracti32x4         [r2 + 2 * r3],      m0,                  2
+    vextracti32x4         [r2 + r7],          m0,                  3
 %endmacro
 
 ;-----------------------------------------------------------------------------------------------------------------
 ; void interp_4tap_vert(int16_t *src, intptr_t srcStride, int16_t *dst, intptr_t dstStride, int coeffIdx)
 ;-----------------------------------------------------------------------------------------------------------------
-%if ARCH_X86_64
+%macro FILTER_VER_SS_CHROMA_8xN_AVX512 1
 INIT_ZMM avx512
-cglobal interp_4tap_vert_ss_8x8, 5, 11, 10
+cglobal interp_4tap_vert_ss_8x%1, 5, 8, 10
     add                   r1d,                r1d
     add                   r3d,                r3d
     sub                   r0,                 r1
@@ -11231,41 +11207,22 @@
     mova                  m8,                 [r5]
     mova                  m9,                 [r5 + mmsize]
 %endif
-    lea                   r10,                [3 * r1]
+    lea                   r6,                 [3 * r1]
     lea                   r7,                 [3 * r3]
-    PROCESS_CHROMA_VERT_SS_8x8_AVX512
-    RET
-%endif
-
-%macro FILTER_VER_SS_CHROMA_8xN_AVX512 1
-INIT_ZMM avx512
-cglobal interp_4tap_vert_ss_8x%1, 5, 11, 10
-    add                   r1d,                r1d
-    add                   r3d,                r3d
-    sub                   r0,                 r1
-    shl                   r4d,                7
-%ifdef PIC
-    lea                   r5,                 [pw_ChromaCoeffVer_32_avx512]
-    mova                  m8,                 [r5 + r4]
-    mova                  m9,                 [r5 + r4 + mmsize]
-%else
-    lea                   r5,                 [pw_ChromaCoeffVer_32_avx512 + r4]
-    mova                  m8,                 [r5]
-    mova                  m9,                 [r5 + mmsize]
-%endif
-    lea                   r10,                [3 * r1]
-    lea                   r7,                 [3 * r3]
-
-%rep %1/8 - 1
-    PROCESS_CHROMA_VERT_SS_8x8_AVX512
-    lea                   r0,                 [r8 + 4 * r1]
+
+%rep %1/4 - 1
+    PROCESS_CHROMA_VERT_SS_8x4_AVX512
+    lea                   r0,                 [r0 + 4 * r1]
     lea                   r2,                 [r2 + 4 * r3]
 %endrep
-    PROCESS_CHROMA_VERT_SS_8x8_AVX512
+    PROCESS_CHROMA_VERT_SS_8x4_AVX512
     RET
 %endmacro
 
 %if ARCH_X86_64
+    FILTER_VER_SS_CHROMA_8xN_AVX512 4
+    FILTER_VER_SS_CHROMA_8xN_AVX512 8
+    FILTER_VER_SS_CHROMA_8xN_AVX512 12
     FILTER_VER_SS_CHROMA_8xN_AVX512 16
     FILTER_VER_SS_CHROMA_8xN_AVX512 32
     FILTER_VER_SS_CHROMA_8xN_AVX512 64


More information about the x265-devel mailing list