[x265] [PATCH 171 of 307] x86: AVX512 interp_4tap_vert_sp_16xN and interp_4tap_vert_ss_16xN for high bit depth

mythreyi at multicorewareinc.com mythreyi at multicorewareinc.com
Sat Apr 7 04:32:49 CEST 2018


# HG changeset patch
# User Vignesh Vijayakumar<vignesh at multicorewareinc.com>
# Date 1510735444 -19800
#      Wed Nov 15 14:14:04 2017 +0530
# Node ID 06da2926583d766da027e141ed6c9cf2f77a208d
# Parent  487307659c367f26096d0da0c81a89ca89b2ffbe
x86: AVX512 interp_4tap_vert_sp_16xN and interp_4tap_vert_ss_16xN for high bit depth

i444 chroma_vsp
Size  |  AVX2 performance | AVX512 performance
----------------------------------------------
16x4  |      21.82x       |      33.15x
16x8  |      22.65x       |      32.41x
16x12 |      23.56x       |      32.00x
16x16 |      24.08x       |      34.31x
16x32 |      23.28x       |      35.39x
16x64 |      24.43x       |      36.54x

i444 chroma_vss
Size  |  AVX2 performance | AVX512 performance
----------------------------------------------
16x4  |      19.90x       |      30.10x
16x8  |      22.65x       |      30.27x
16x12 |      24.85x       |      30.14x
16x16 |      25.33x       |      31.05x
16x32 |      25.08x       |      35.79x
16x64 |      26.20x       |      36.01x

diff -r 487307659c36 -r 06da2926583d source/common/x86/asm-primitives.cpp
--- a/source/common/x86/asm-primitives.cpp	Tue Nov 14 19:16:49 2017 +0530
+++ b/source/common/x86/asm-primitives.cpp	Wed Nov 15 14:14:04 2017 +0530
@@ -2677,6 +2677,7 @@
         p.chroma[X265_CSP_I444].pu[LUMA_32x24].filter_vsp = PFX(interp_4tap_vert_sp_32x24_avx512);
         p.chroma[X265_CSP_I444].pu[LUMA_32x32].filter_vsp = PFX(interp_4tap_vert_sp_32x32_avx512);
         p.chroma[X265_CSP_I444].pu[LUMA_32x64].filter_vsp = PFX(interp_4tap_vert_sp_32x64_avx512);
+
         p.chroma[X265_CSP_I444].pu[LUMA_16x4].filter_vpp = PFX(interp_4tap_vert_pp_16x4_avx512);
         p.chroma[X265_CSP_I444].pu[LUMA_16x8].filter_vpp = PFX(interp_4tap_vert_pp_16x8_avx512);
         p.chroma[X265_CSP_I444].pu[LUMA_16x12].filter_vpp = PFX(interp_4tap_vert_pp_16x12_avx512);
@@ -2689,6 +2690,19 @@
         p.chroma[X265_CSP_I444].pu[LUMA_16x16].filter_vps = PFX(interp_4tap_vert_ps_16x16_avx512);
         p.chroma[X265_CSP_I444].pu[LUMA_16x32].filter_vps = PFX(interp_4tap_vert_ps_16x32_avx512);
         p.chroma[X265_CSP_I444].pu[LUMA_16x64].filter_vps = PFX(interp_4tap_vert_ps_16x64_avx512);
+        p.chroma[X265_CSP_I444].pu[LUMA_16x4].filter_vss = PFX(interp_4tap_vert_ss_16x4_avx512);
+        p.chroma[X265_CSP_I444].pu[LUMA_16x8].filter_vss = PFX(interp_4tap_vert_ss_16x8_avx512);
+        p.chroma[X265_CSP_I444].pu[LUMA_16x12].filter_vss = PFX(interp_4tap_vert_ss_16x12_avx512);
+        p.chroma[X265_CSP_I444].pu[LUMA_16x16].filter_vss = PFX(interp_4tap_vert_ss_16x16_avx512);
+        p.chroma[X265_CSP_I444].pu[LUMA_16x32].filter_vss = PFX(interp_4tap_vert_ss_16x32_avx512);
+        p.chroma[X265_CSP_I444].pu[LUMA_16x64].filter_vss = PFX(interp_4tap_vert_ss_16x64_avx512);
+        p.chroma[X265_CSP_I444].pu[LUMA_16x4].filter_vsp = PFX(interp_4tap_vert_sp_16x4_avx512);
+        p.chroma[X265_CSP_I444].pu[LUMA_16x8].filter_vsp = PFX(interp_4tap_vert_sp_16x8_avx512);
+        p.chroma[X265_CSP_I444].pu[LUMA_16x12].filter_vsp = PFX(interp_4tap_vert_sp_16x12_avx512);
+        p.chroma[X265_CSP_I444].pu[LUMA_16x16].filter_vsp = PFX(interp_4tap_vert_sp_16x16_avx512);
+        p.chroma[X265_CSP_I444].pu[LUMA_16x32].filter_vsp = PFX(interp_4tap_vert_sp_16x32_avx512);
+        p.chroma[X265_CSP_I444].pu[LUMA_16x64].filter_vsp = PFX(interp_4tap_vert_sp_16x64_avx512);
+
         p.chroma[X265_CSP_I444].pu[LUMA_8x8].filter_vpp = PFX(interp_4tap_vert_pp_8x8_avx512);
         p.chroma[X265_CSP_I444].pu[LUMA_8x16].filter_vpp = PFX(interp_4tap_vert_pp_8x16_avx512);
         p.chroma[X265_CSP_I444].pu[LUMA_8x32].filter_vpp = PFX(interp_4tap_vert_pp_8x32_avx512);
@@ -2712,6 +2726,7 @@
         p.chroma[X265_CSP_I422].pu[CHROMA_422_32x32].filter_vsp = PFX(interp_4tap_vert_sp_32x32_avx512);
         p.chroma[X265_CSP_I422].pu[CHROMA_422_32x48].filter_vsp = PFX(interp_4tap_vert_sp_32x48_avx512);
         p.chroma[X265_CSP_I422].pu[CHROMA_422_32x64].filter_vsp = PFX(interp_4tap_vert_sp_32x64_avx512);
+
         p.chroma[X265_CSP_I422].pu[CHROMA_422_16x8].filter_vpp = PFX(interp_4tap_vert_pp_16x8_avx512);
         p.chroma[X265_CSP_I422].pu[CHROMA_422_16x16].filter_vpp = PFX(interp_4tap_vert_pp_16x16_avx512);
         p.chroma[X265_CSP_I422].pu[CHROMA_422_16x24].filter_vpp = PFX(interp_4tap_vert_pp_16x24_avx512);
@@ -2722,6 +2737,17 @@
         p.chroma[X265_CSP_I422].pu[CHROMA_422_16x24].filter_vps = PFX(interp_4tap_vert_ps_16x24_avx512);
         p.chroma[X265_CSP_I422].pu[CHROMA_422_16x32].filter_vps = PFX(interp_4tap_vert_ps_16x32_avx512);
         p.chroma[X265_CSP_I422].pu[CHROMA_422_16x64].filter_vps = PFX(interp_4tap_vert_ps_16x64_avx512);
+        p.chroma[X265_CSP_I422].pu[CHROMA_422_16x8].filter_vss = PFX(interp_4tap_vert_ss_16x8_avx512);
+        p.chroma[X265_CSP_I422].pu[CHROMA_422_16x16].filter_vss = PFX(interp_4tap_vert_ss_16x16_avx512);
+        p.chroma[X265_CSP_I422].pu[CHROMA_422_16x24].filter_vss = PFX(interp_4tap_vert_ss_16x24_avx512);
+        p.chroma[X265_CSP_I422].pu[CHROMA_422_16x32].filter_vss = PFX(interp_4tap_vert_ss_16x32_avx512);
+        p.chroma[X265_CSP_I422].pu[CHROMA_422_16x64].filter_vss = PFX(interp_4tap_vert_ss_16x64_avx512);
+        p.chroma[X265_CSP_I422].pu[CHROMA_422_16x8].filter_vsp = PFX(interp_4tap_vert_sp_16x8_avx512);
+        p.chroma[X265_CSP_I422].pu[CHROMA_422_16x16].filter_vsp = PFX(interp_4tap_vert_sp_16x16_avx512);
+        p.chroma[X265_CSP_I422].pu[CHROMA_422_16x24].filter_vsp = PFX(interp_4tap_vert_sp_16x24_avx512);
+        p.chroma[X265_CSP_I422].pu[CHROMA_422_16x32].filter_vsp = PFX(interp_4tap_vert_sp_16x32_avx512);
+        p.chroma[X265_CSP_I422].pu[CHROMA_422_16x64].filter_vsp = PFX(interp_4tap_vert_sp_16x64_avx512);
+
         p.chroma[X265_CSP_I422].pu[CHROMA_422_8x8].filter_vpp = PFX(interp_4tap_vert_pp_8x8_avx512);
         p.chroma[X265_CSP_I422].pu[CHROMA_422_8x16].filter_vpp = PFX(interp_4tap_vert_pp_8x16_avx512);
         p.chroma[X265_CSP_I422].pu[CHROMA_422_8x32].filter_vpp = PFX(interp_4tap_vert_pp_8x32_avx512);
@@ -2747,6 +2773,7 @@
         p.chroma[X265_CSP_I420].pu[CHROMA_420_32x16].filter_vsp = PFX(interp_4tap_vert_sp_32x16_avx512);
         p.chroma[X265_CSP_I420].pu[CHROMA_420_32x24].filter_vsp = PFX(interp_4tap_vert_sp_32x24_avx512);
         p.chroma[X265_CSP_I420].pu[CHROMA_420_32x32].filter_vsp = PFX(interp_4tap_vert_sp_32x32_avx512);
+
         p.chroma[X265_CSP_I420].pu[CHROMA_420_16x4].filter_vpp = PFX(interp_4tap_vert_pp_16x4_avx512);
         p.chroma[X265_CSP_I420].pu[CHROMA_420_16x8].filter_vpp = PFX(interp_4tap_vert_pp_16x8_avx512);
         p.chroma[X265_CSP_I420].pu[CHROMA_420_16x12].filter_vpp = PFX(interp_4tap_vert_pp_16x12_avx512);
@@ -2757,6 +2784,17 @@
         p.chroma[X265_CSP_I420].pu[CHROMA_420_16x12].filter_vps = PFX(interp_4tap_vert_ps_16x12_avx512);
         p.chroma[X265_CSP_I420].pu[CHROMA_420_16x16].filter_vps = PFX(interp_4tap_vert_ps_16x16_avx512);
         p.chroma[X265_CSP_I420].pu[CHROMA_420_16x32].filter_vps = PFX(interp_4tap_vert_ps_16x32_avx512);
+        p.chroma[X265_CSP_I420].pu[CHROMA_420_16x4].filter_vss = PFX(interp_4tap_vert_ss_16x4_avx512);
+        p.chroma[X265_CSP_I420].pu[CHROMA_420_16x8].filter_vss = PFX(interp_4tap_vert_ss_16x8_avx512);
+        p.chroma[X265_CSP_I420].pu[CHROMA_420_16x12].filter_vss = PFX(interp_4tap_vert_ss_16x12_avx512);
+        p.chroma[X265_CSP_I420].pu[CHROMA_420_16x16].filter_vss = PFX(interp_4tap_vert_ss_16x16_avx512);
+        p.chroma[X265_CSP_I420].pu[CHROMA_420_16x32].filter_vss = PFX(interp_4tap_vert_ss_16x32_avx512);
+        p.chroma[X265_CSP_I420].pu[CHROMA_420_16x4].filter_vsp = PFX(interp_4tap_vert_sp_16x4_avx512);
+        p.chroma[X265_CSP_I420].pu[CHROMA_420_16x8].filter_vsp = PFX(interp_4tap_vert_sp_16x8_avx512);
+        p.chroma[X265_CSP_I420].pu[CHROMA_420_16x12].filter_vsp = PFX(interp_4tap_vert_sp_16x12_avx512);
+        p.chroma[X265_CSP_I420].pu[CHROMA_420_16x16].filter_vsp = PFX(interp_4tap_vert_sp_16x16_avx512);
+        p.chroma[X265_CSP_I420].pu[CHROMA_420_16x32].filter_vsp = PFX(interp_4tap_vert_sp_16x32_avx512);
+
         p.chroma[X265_CSP_I420].pu[CHROMA_420_8x8].filter_vpp = PFX(interp_4tap_vert_pp_8x8_avx512);
         p.chroma[X265_CSP_I420].pu[CHROMA_420_8x16].filter_vpp = PFX(interp_4tap_vert_pp_8x16_avx512);
         p.chroma[X265_CSP_I420].pu[CHROMA_420_8x32].filter_vpp = PFX(interp_4tap_vert_pp_8x32_avx512);
diff -r 487307659c36 -r 06da2926583d source/common/x86/ipfilter16.asm
--- a/source/common/x86/ipfilter16.asm	Tue Nov 14 19:16:49 2017 +0530
+++ b/source/common/x86/ipfilter16.asm	Wed Nov 15 14:14:04 2017 +0530
@@ -8272,6 +8272,147 @@
 FILTER_VER_PS_CHROMA_64xN_AVX512 64
 %endif
 
+%macro PROCESS_CHROMA_VERT_S_16x4_AVX512 1
+    movu                  ym1,                [r0]
+    lea                   r6,                 [r0 + 2 * r1]
+    vinserti32x8          m1,                 [r6],                1
+    movu                  ym3,                [r0 + r1]
+    vinserti32x8          m3,                 [r6 + r1],           1
+    punpcklwd             m0,                 m1,                  m3
+    pmaddwd               m0,                 m8
+    punpckhwd             m1,                 m3
+    pmaddwd               m1,                 m8
+
+    movu                  ym4,                [r0 + 2 * r1]
+    vinserti32x8          m4,                 [r6 + 2 * r1],       1
+    punpcklwd             m2,                 m3,                  m4
+    pmaddwd               m2,                 m8
+    punpckhwd             m3,                 m4
+    pmaddwd               m3,                 m8
+
+    movu                  ym5,                [r0 + r8]
+    vinserti32x8          m5,                 [r6 + r8],           1
+    punpcklwd             m6,                 m4,                  m5
+    pmaddwd               m6,                 m9
+    paddd                 m0,                 m6
+    punpckhwd             m4,                 m5
+    pmaddwd               m4,                 m9
+    paddd                 m1,                 m4
+
+    movu                  ym4,                [r0 + 4 * r1]
+    vinserti32x8          m4,                 [r6 + 4 * r1],       1
+    punpcklwd             m6,                 m5,                  m4
+    pmaddwd               m6,                 m9
+    paddd                 m2,                 m6
+    punpckhwd             m5,                 m4
+    pmaddwd               m5,                 m9
+    paddd                 m3,                 m5
+
+%ifidn %1,sp
+    paddd                 m0,                 m7
+    paddd                 m1,                 m7
+    paddd                 m2,                 m7
+    paddd                 m3,                 m7
+
+    psrad                 m0,                 INTERP_SHIFT_SP
+    psrad                 m1,                 INTERP_SHIFT_SP
+    psrad                 m2,                 INTERP_SHIFT_SP
+    psrad                 m3,                 INTERP_SHIFT_SP
+%else
+    psrad                 m0,                 6
+    psrad                 m1,                 6
+    psrad                 m2,                 6
+    psrad                 m3,                 6
+%endif
+
+    packssdw              m0,                 m1
+    packssdw              m2,                 m3
+    movu                  [r2],               ym0
+    movu                  [r2 + r3],          ym2
+    vextracti32x8         [r2 + 2 * r3],      m0,                  1
+    vextracti32x8         [r2 + r7],          m2,                  1
+%endmacro
+
+;-----------------------------------------------------------------------------------------------------------------
+; void interp_4tap_vert(int16_t *src, intptr_t srcStride, int16_t *dst, intptr_t dstStride, int coeffIdx)
+;-----------------------------------------------------------------------------------------------------------------
+%macro CHROMA_VERT_S_16x4_AVX512 1
+%if ARCH_X86_64
+INIT_ZMM avx512
+cglobal interp_4tap_vert_%1_16x4, 5, 9, 10
+    add                   r1d,                r1d
+    add                   r3d,                r3d
+    sub                   r0,                 r1
+    shl                   r4d,                7
+
+%ifdef PIC
+    lea                   r5,                 [tab_ChromaCoeffV_avx512]
+    lea                   r5,                 [r5 + r4]
+%else
+    lea                   r5,                 [tab_ChromaCoeffV_avx512 + r4]
+%endif
+
+    mova                  m8,                 [r5]
+    mova                  m9,                 [r5 + mmsize]
+%ifidn %1, sp
+    vbroadcasti32x4       m7,                 [INTERP_OFFSET_SP]
+%endif
+    lea                   r7,                 [3 * r3]
+    lea                   r8,                 [3 * r1]
+    PROCESS_CHROMA_VERT_S_16x4_AVX512 %1
+    RET
+%endif
+%endmacro
+
+CHROMA_VERT_S_16x4_AVX512 ss
+CHROMA_VERT_S_16x4_AVX512 sp
+
+%macro FILTER_VER_S_CHROMA_16xN_AVX512 2
+INIT_ZMM avx512
+cglobal interp_4tap_vert_%1_16x%2, 5, 9, 10
+    add                   r1d,                r1d
+    add                   r3d,                r3d
+    sub                   r0,                 r1
+    shl                   r4d,                7
+
+%ifdef PIC
+    lea                   r5,                 [tab_ChromaCoeffV_avx512]
+    lea                   r5,                 [r5 + r4]
+%else
+    lea                   r5,                 [tab_ChromaCoeffV_avx512 + r4]
+%endif
+
+    mova                  m8,                 [r5]
+    mova                  m9,                 [r5 + mmsize]
+%ifidn %1, sp
+    vbroadcasti32x4       m7,                 [INTERP_OFFSET_SP]
+%endif
+    lea                   r7,                 [3 * r3]
+    lea                   r8,                 [3 * r1]
+%rep %2/4 - 1
+    PROCESS_CHROMA_VERT_S_16x4_AVX512 %1
+    lea                   r0,                 [r0 + 4 * r1]
+    lea                   r2,                 [r2 + 4 * r3]
+%endrep
+    PROCESS_CHROMA_VERT_S_16x4_AVX512 %1
+    RET
+%endmacro
+
+%if ARCH_X86_64
+    FILTER_VER_S_CHROMA_16xN_AVX512 ss, 8
+    FILTER_VER_S_CHROMA_16xN_AVX512 ss, 12
+    FILTER_VER_S_CHROMA_16xN_AVX512 ss, 16
+    FILTER_VER_S_CHROMA_16xN_AVX512 ss, 24
+    FILTER_VER_S_CHROMA_16xN_AVX512 ss, 32
+    FILTER_VER_S_CHROMA_16xN_AVX512 ss, 64
+    FILTER_VER_S_CHROMA_16xN_AVX512 sp, 8
+    FILTER_VER_S_CHROMA_16xN_AVX512 sp, 12
+    FILTER_VER_S_CHROMA_16xN_AVX512 sp, 16
+    FILTER_VER_S_CHROMA_16xN_AVX512 sp, 24
+    FILTER_VER_S_CHROMA_16xN_AVX512 sp, 32
+    FILTER_VER_S_CHROMA_16xN_AVX512 sp, 64
+%endif
+
 %macro PROCESS_CHROMA_VERT_S_24x8_AVX512 1
     movu                  ym1,                [r0]
     lea                   r6,                 [r0 + 2 * r1]


More information about the x265-devel mailing list