[x265] [PATCH 145 of 307] x86: AVX512 interp_4tap_vert_pp_8xN for high bit depth

mythreyi at multicorewareinc.com mythreyi at multicorewareinc.com
Sat Apr 7 04:32:23 CEST 2018


# HG changeset patch
# User Vignesh Vijayakumar<vignesh at multicorewareinc.com>
# Date 1509612018 -19800
#      Thu Nov 02 14:10:18 2017 +0530
# Node ID d3a1db4790b662306a3f1222cde66c006e10f604
# Parent  ba385c1b80c649750ce66dda40a6364d50230b34
x86: AVX512 interp_4tap_vert_pp_8xN for high bit depth

i444
Size  |  AVX2 performance | AVX512 performance
----------------------------------------------
8x8   |      17.75x        |      28.01x
8x16  |      21.54x        |      25.32x
8x32  |      22.78x        |      25.94x

diff -r ba385c1b80c6 -r d3a1db4790b6 source/common/x86/asm-primitives.cpp
--- a/source/common/x86/asm-primitives.cpp	Thu Nov 02 14:07:05 2017 +0530
+++ b/source/common/x86/asm-primitives.cpp	Thu Nov 02 14:10:18 2017 +0530
@@ -2650,6 +2650,10 @@
         p.chroma[X265_CSP_I444].pu[LUMA_16x16].filter_vpp = PFX(interp_4tap_vert_pp_16x16_avx512);
         p.chroma[X265_CSP_I444].pu[LUMA_16x32].filter_vpp = PFX(interp_4tap_vert_pp_16x32_avx512);
         p.chroma[X265_CSP_I444].pu[LUMA_16x64].filter_vpp = PFX(interp_4tap_vert_pp_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);
+
         p.chroma[X265_CSP_I422].pu[CHROMA_422_32x16].filter_vpp = PFX(interp_4tap_vert_pp_32x16_avx512);
         p.chroma[X265_CSP_I422].pu[CHROMA_422_32x32].filter_vpp = PFX(interp_4tap_vert_pp_32x32_avx512);
         p.chroma[X265_CSP_I422].pu[CHROMA_422_32x48].filter_vpp = PFX(interp_4tap_vert_pp_32x48_avx512);
@@ -2659,6 +2663,11 @@
         p.chroma[X265_CSP_I422].pu[CHROMA_422_16x24].filter_vpp = PFX(interp_4tap_vert_pp_16x24_avx512);
         p.chroma[X265_CSP_I422].pu[CHROMA_422_16x32].filter_vpp = PFX(interp_4tap_vert_pp_16x32_avx512);
         p.chroma[X265_CSP_I422].pu[CHROMA_422_16x64].filter_vpp = PFX(interp_4tap_vert_pp_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);
+        p.chroma[X265_CSP_I422].pu[CHROMA_422_8x64].filter_vpp = PFX(interp_4tap_vert_pp_8x64_avx512);
+
         p.chroma[X265_CSP_I420].pu[CHROMA_420_32x8].filter_vpp = PFX(interp_4tap_vert_pp_32x8_avx512);
         p.chroma[X265_CSP_I420].pu[CHROMA_420_32x16].filter_vpp = PFX(interp_4tap_vert_pp_32x16_avx512);
         p.chroma[X265_CSP_I420].pu[CHROMA_420_32x24].filter_vpp = PFX(interp_4tap_vert_pp_32x24_avx512);
@@ -2668,6 +2677,9 @@
         p.chroma[X265_CSP_I420].pu[CHROMA_420_16x12].filter_vpp = PFX(interp_4tap_vert_pp_16x12_avx512);
         p.chroma[X265_CSP_I420].pu[CHROMA_420_16x16].filter_vpp = PFX(interp_4tap_vert_pp_16x16_avx512);
         p.chroma[X265_CSP_I420].pu[CHROMA_420_16x32].filter_vpp = PFX(interp_4tap_vert_pp_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);
 
         p.cu[BLOCK_8x8].dct = PFX(dct8_avx512);
 
diff -r ba385c1b80c6 -r d3a1db4790b6 source/common/x86/ipfilter16.asm
--- a/source/common/x86/ipfilter16.asm	Thu Nov 02 14:07:05 2017 +0530
+++ b/source/common/x86/ipfilter16.asm	Thu Nov 02 14:10:18 2017 +0530
@@ -5904,6 +5904,140 @@
     RET
 %endif
 
+%macro PROCESS_CHROMA_VERT_PP_8x8_AVX512 0
+    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
+    punpcklwd             m0,                 m1,                  m3
+    pmaddwd               m0,                 [r5]
+    punpckhwd             m1,                 m3
+    pmaddwd               m1,                 [r5]
+
+    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,                 [r5]
+    punpckhwd             m3,                 m4
+    pmaddwd               m3,                 [r5]
+
+    lea                   r0,                 [r0 + 2 * r1]
+    lea                   r6,                 [r6 + 2 * r1]
+    lea                   r8,                 [r8 + 2 * r1]
+    lea                   r9,                 [r9 + 2 * r1]
+
+    movu                  xm5,                [r0 + r1]
+    vinserti32x4          m5,                 [r6 + r1],           1
+    vinserti32x4          m5,                 [r8 + r1],           2
+    vinserti32x4          m5,                 [r9 + r1],           3
+    punpcklwd             m6,                 m4,                  m5
+    pmaddwd               m6,                 [r5 + mmsize]
+    paddd                 m0,                 m6
+    punpckhwd             m4,                 m5
+    pmaddwd               m4,                 [r5 + mmsize]
+    paddd                 m1,                 m4
+
+    movu                  xm4,                [r0 + 2 * r1]
+    vinserti32x4          m4,                 [r6 + 2 * r1],       1
+    vinserti32x4          m4,                 [r8 + 2 * r1],       2
+    vinserti32x4          m4,                 [r9 + 2 * r1],       3
+    punpcklwd             m6,                 m5,                  m4
+    pmaddwd               m6,                 [r5 + mmsize]
+    paddd                 m2,                 m6
+    punpckhwd             m5,                 m4
+    pmaddwd               m5,                 [r5 + mmsize]
+    paddd                 m3,                 m5
+
+    paddd                 m0,                 m7
+    paddd                 m1,                 m7
+    paddd                 m2,                 m7
+    paddd                 m3,                 m7
+
+    psrad                 m0,                 INTERP_SHIFT_PP
+    psrad                 m1,                 INTERP_SHIFT_PP
+    psrad                 m2,                 INTERP_SHIFT_PP
+    psrad                 m3,                 INTERP_SHIFT_PP
+
+    packssdw              m0,                 m1
+    packssdw              m2,                 m3
+    pxor                  m5,                 m5
+    CLIPW2                m0,                 m2,                  m5,                 m8
+    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
+%endmacro
+
+;-----------------------------------------------------------------------------------------------------------------
+; void interp_4tap_vert(int16_t *src, intptr_t srcStride, int16_t *dst, intptr_t dstStride, int coeffIdx)
+;-----------------------------------------------------------------------------------------------------------------
+%if ARCH_X86_64
+INIT_ZMM avx512
+cglobal interp_4tap_vert_pp_8x8, 5, 10, 9
+    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
+    vbroadcasti32x8       m7,                 [INTERP_OFFSET_PP]
+    vbroadcasti32x8       m8,                 [pw_pixel_max]
+    lea                   r7,                 [3 * r3]
+    PROCESS_CHROMA_VERT_PP_8x8_AVX512
+    RET
+%endif
+
+%macro FILTER_VER_PP_CHROMA_8xN_AVX512 1
+INIT_ZMM avx512
+cglobal interp_4tap_vert_pp_8x%1, 5, 10, 9
+    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
+    vbroadcasti32x8       m7,                 [INTERP_OFFSET_PP]
+    vbroadcasti32x8       m8,                 [pw_pixel_max]
+    lea                   r7,                 [3 * r3]
+%rep %1/8 - 1
+    PROCESS_CHROMA_VERT_PP_8x8_AVX512
+    lea                   r0,                 [r9]
+    lea                   r2,                 [r2 + 4 * r3]
+%endrep
+    PROCESS_CHROMA_VERT_PP_8x8_AVX512
+    RET
+%endmacro
+
+%if ARCH_X86_64
+FILTER_VER_PP_CHROMA_8xN_AVX512 16
+FILTER_VER_PP_CHROMA_8xN_AVX512 32
+FILTER_VER_PP_CHROMA_8xN_AVX512 64
+%endif
+
 %macro PROCESS_CHROMA_VERT_PP_16x4_AVX512 0
     movu                  ym1,                [r0]
     lea                   r6,                 [r0 + 2 * r1]


More information about the x265-devel mailing list