[x265] [PATCH 192 of 307] x86: AVX512 interp_8tap_vert_sp_16xN and interp_8tap_vert_ss_16xN

mythreyi at multicorewareinc.com mythreyi at multicorewareinc.com
Sat Apr 7 04:33:10 CEST 2018


# HG changeset patch
# User Vignesh Vijayakumar<vignesh at multicorewareinc.com>
# Date 1511433052 -19800
#      Thu Nov 23 16:00:52 2017 +0530
# Node ID 66679e6701cf73584ec0db224927ee9eb24893a7
# Parent  b7ebc01ecbfe2510fbe6a2ce305bdb479267decf
x86: AVX512 interp_8tap_vert_sp_16xN and interp_8tap_vert_ss_16xN

luma_vsp
Size  |  AVX2 performance | AVX512 performance
----------------------------------------------
16x4  |       9.68x       |      16.85x
16x8  |      11.69x       |      18.47x
16x12 |      13.43x       |      18.34x
16x16 |      13.26x       |      18.96x
16x32 |      12.96x       |      19.66x
16x64 |      13.12x       |      19.72x

luma_vss
Size  |  AVX2 performance | AVX512 performance
----------------------------------------------
16x4  |       7.90x       |      15.87x
16x8  |       8.58x       |      15.10x
16x12 |      10.78x       |      15.38x
16x16 |      11.51x       |      16.27x
16x32 |      11.92x       |      16.93x
16x64 |      10.17x       |      17.24x

diff -r b7ebc01ecbfe -r 66679e6701cf source/common/x86/asm-primitives.cpp
--- a/source/common/x86/asm-primitives.cpp	Thu Nov 23 13:52:24 2017 +0530
+++ b/source/common/x86/asm-primitives.cpp	Thu Nov 23 16:00:52 2017 +0530
@@ -2835,6 +2835,12 @@
         p.chroma[X265_CSP_I422].pu[CHROMA_422_24x64].filter_vsp = PFX(interp_4tap_vert_sp_24x64_avx512);
         p.chroma[X265_CSP_I444].pu[LUMA_24x32].filter_vsp = PFX(interp_4tap_vert_sp_24x32_avx512);
 
+        p.pu[LUMA_16x4].luma_vss = PFX(interp_8tap_vert_ss_16x4_avx512);
+        p.pu[LUMA_16x8].luma_vss = PFX(interp_8tap_vert_ss_16x8_avx512);
+        p.pu[LUMA_16x12].luma_vss = PFX(interp_8tap_vert_ss_16x12_avx512);
+        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_32x8].luma_vss = PFX(interp_8tap_vert_ss_32x8_avx512);
         p.pu[LUMA_32x16].luma_vss = PFX(interp_8tap_vert_ss_32x16_avx512);
         p.pu[LUMA_32x32].luma_vss = PFX(interp_8tap_vert_ss_32x32_avx512);
@@ -2845,6 +2851,12 @@
         p.pu[LUMA_64x48].luma_vss = PFX(interp_8tap_vert_ss_64x48_avx512);
         p.pu[LUMA_64x64].luma_vss = PFX(interp_8tap_vert_ss_64x64_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_32x8].luma_vsp = PFX(interp_8tap_vert_sp_32x8_avx512);
         p.pu[LUMA_32x16].luma_vsp = PFX(interp_8tap_vert_sp_32x16_avx512);
         p.pu[LUMA_32x32].luma_vsp = PFX(interp_8tap_vert_sp_32x32_avx512);
diff -r b7ebc01ecbfe -r 66679e6701cf source/common/x86/ipfilter16.asm
--- a/source/common/x86/ipfilter16.asm	Thu Nov 23 13:52:24 2017 +0530
+++ b/source/common/x86/ipfilter16.asm	Thu Nov 23 16:00:52 2017 +0530
@@ -10795,6 +10795,166 @@
 ;-------------------------------------------------------------------------------------------------------------
 ;avx512 luma_vss and luma_vsp code start
 ;-------------------------------------------------------------------------------------------------------------
+%macro PROCESS_LUMA_VERT_S_16x4_AVX512 1
+    movu                 ym1,                 [r0]
+    movu                 ym3,                 [r0 + r1]
+    vinserti32x8         m1,                  [r0 + 2 * r1],          1
+    vinserti32x8         m3,                  [r0 + r7],              1
+    punpcklwd            m0,                  m1,                     m3
+    pmaddwd              m0,                  m15
+    punpckhwd            m1,                  m3
+    pmaddwd              m1,                  m15
+
+    lea                  r6,                  [r0 + 4 * r1]
+    movu                 ym4,                 [r0 + 2 * r1]
+    vinserti32x8         m4,                  [r6],                   1
+    punpcklwd            m2,                  m3,                     m4
+    pmaddwd              m2,                  m15
+    punpckhwd            m3,                  m4
+    pmaddwd              m3,                  m15
+
+    movu                 ym5,                 [r0 + r7]
+    vinserti32x8         m5,                  [r6 + r1],              1
+    punpcklwd            m6,                  m4,                     m5
+    pmaddwd              m6,                  m16
+    punpckhwd            m4,                  m5
+    pmaddwd              m4,                  m16
+
+    paddd                m0,                  m6
+    paddd                m1,                  m4
+
+    movu                 ym4,                 [r6]
+    vinserti32x8         m4,                  [r6 + 2 * r1],          1
+    punpcklwd            m6,                  m5,                     m4
+    pmaddwd              m6,                  m16
+    punpckhwd            m5,                  m4
+    pmaddwd              m5,                  m16
+
+    paddd                m2,                  m6
+    paddd                m3,                  m5
+
+    movu                 ym11,                [r6 + r1]
+    vinserti32x8         m11,                 [r6 + r7],              1
+    punpcklwd            m8,                  m4,                     m11
+    pmaddwd              m8,                  m17
+    punpckhwd            m4,                  m11
+    pmaddwd              m4,                  m17
+
+    movu                 ym12,                [r6 + 2 * r1]
+    vinserti32x8         m12,                 [r6 + 4 * r1],          1
+    punpcklwd            m10,                 m11,                    m12
+    pmaddwd              m10,                 m17
+    punpckhwd            m11,                 m12
+    pmaddwd              m11,                 m17
+
+    lea                  r4,                  [r6 + 4 * r1]
+    movu                 ym13,                [r6 + r7]
+    vinserti32x8         m13,                 [r4 + 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,                [r6 + 4 * r1]
+    vinserti32x8         m12,                 [r4 + 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
+
+%ifidn %1, sp
+    paddd                m0,                  m19
+    paddd                m1,                  m19
+    paddd                m2,                  m19
+    paddd                m3,                  m19
+
+    psrad                m0,                  INTERP_SHIFT_SP
+    psrad                m1,                  INTERP_SHIFT_SP
+    psrad                m2,                  INTERP_SHIFT_SP
+    psrad                m3,                  INTERP_SHIFT_SP
+
+    packssdw             m0,                  m1
+    packssdw             m2,                  m3
+    CLIPW2               m0,                  m2,                   m20,                 m21
+%else
+    psrad                m0,                  6
+    psrad                m1,                  6
+    psrad                m2,                  6
+    psrad                m3,                  6
+
+    packssdw             m0,                  m1
+    packssdw             m2,                  m3
+%endif
+
+    movu                 [r2],                ym0
+    movu                 [r2 + r3],           ym2
+    vextracti32x8        [r2 + 2 * r3],       m0,                1
+    vextracti32x8        [r2 + r5],           m2,                1
+%endmacro
+;-----------------------------------------------------------------------------------------------------------------
+; void interp_8tap_vert(int16_t *src, intptr_t srcStride, int16_t *dst, intptr_t dstStride, int coeffIdx)
+;-----------------------------------------------------------------------------------------------------------------
+%macro FILTER_VER_S_LUMA_16xN_AVX512 2
+INIT_ZMM avx512
+cglobal interp_8tap_vert_%1_16x%2, 5, 8, 22
+    add                   r1d,                r1d
+    add                   r3d,                r3d
+    lea                   r7,                 [3 * r1]
+    sub                   r0,                 r7
+    shl                   r4d,                8
+%ifdef PIC
+    lea                   r5,                 [tab_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,                 [tab_LumaCoeffVer_avx512 + r4]
+    mova                  m15,                [r5]
+    mova                  m16,                [r5 + 1 * mmsize]
+    mova                  m17,                [r5 + 2 * mmsize]
+    mova                  m18,                [r5 + 3 * mmsize]
+%endif
+%ifidn %1, sp
+    vbroadcasti32x4       m19,                [INTERP_OFFSET_SP]
+    pxor                  m20,                m20
+    vbroadcasti32x8       m21,                [pw_pixel_max]
+%endif
+    lea                   r5,                 [3 * r3]
+%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_S_16x4_AVX512 %1
+    RET
+%endmacro
+
+%if ARCH_X86_64
+    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_S_32x2_AVX512 1
     movu                 m1,                  [r0]                           ;0 row
     movu                 m3,                  [r0 + r1]                      ;1 row


More information about the x265-devel mailing list