[x265] [PATCH 107 of 307] x86: AVX512 interp_4tap_horiz_pp_24xN for high bit depth

mythreyi at multicorewareinc.com mythreyi at multicorewareinc.com
Sat Apr 7 04:31:45 CEST 2018


# HG changeset patch
# User Vignesh Vijayakumar
# Date 1504173085 -19800
#      Thu Aug 31 15:21:25 2017 +0530
# Node ID c726239a07580fd13c4177f0206d615ee02c5975
# Parent  1fb1948309a0a9218a07e060300b9d5a7ff58321
x86: AVX512 interp_4tap_horiz_pp_24xN for high bit depth

i444 24x32
AVX2 performance     : 8.85x
AVX512 performance   : 19.37x

diff -r 1fb1948309a0 -r c726239a0758 source/common/x86/asm-primitives.cpp
--- a/source/common/x86/asm-primitives.cpp	Thu Aug 31 14:54:18 2017 +0530
+++ b/source/common/x86/asm-primitives.cpp	Thu Aug 31 15:21:25 2017 +0530
@@ -2367,6 +2367,7 @@
         p.chroma[X265_CSP_I420].pu[CHROMA_420_32x16].filter_hpp = PFX(interp_4tap_horiz_pp_32x16_avx512);
         p.chroma[X265_CSP_I420].pu[CHROMA_420_32x24].filter_hpp = PFX(interp_4tap_horiz_pp_32x24_avx512);
         p.chroma[X265_CSP_I420].pu[CHROMA_420_32x32].filter_hpp = PFX(interp_4tap_horiz_pp_32x32_avx512);
+        p.chroma[X265_CSP_I420].pu[CHROMA_420_24x32].filter_hpp = PFX(interp_4tap_horiz_pp_24x32_avx512);
 
         p.chroma[X265_CSP_I422].pu[CHROMA_422_8x4].filter_hpp = PFX(interp_4tap_horiz_pp_8x4_avx512);
         p.chroma[X265_CSP_I422].pu[CHROMA_422_8x8].filter_hpp = PFX(interp_4tap_horiz_pp_8x8_avx512);
@@ -2383,6 +2384,7 @@
         p.chroma[X265_CSP_I422].pu[CHROMA_422_32x32].filter_hpp = PFX(interp_4tap_horiz_pp_32x32_avx512);
         p.chroma[X265_CSP_I422].pu[CHROMA_422_32x48].filter_hpp = PFX(interp_4tap_horiz_pp_32x48_avx512);
         p.chroma[X265_CSP_I422].pu[CHROMA_422_32x64].filter_hpp = PFX(interp_4tap_horiz_pp_32x64_avx512);
+        p.chroma[X265_CSP_I422].pu[CHROMA_422_24x64].filter_hpp = PFX(interp_4tap_horiz_pp_24x64_avx512);
 
         p.chroma[X265_CSP_I444].pu[LUMA_8x4].filter_hpp = PFX(interp_4tap_horiz_pp_8x4_avx512);
         p.chroma[X265_CSP_I444].pu[LUMA_8x8].filter_hpp = PFX(interp_4tap_horiz_pp_8x8_avx512);
@@ -2404,6 +2406,7 @@
         p.chroma[X265_CSP_I444].pu[LUMA_64x48].filter_hpp = PFX(interp_4tap_horiz_pp_64x48_avx512);
         p.chroma[X265_CSP_I444].pu[LUMA_64x64].filter_hpp = PFX(interp_4tap_horiz_pp_64x64_avx512);
         p.chroma[X265_CSP_I444].pu[LUMA_48x64].filter_hpp = PFX(interp_4tap_horiz_pp_48x64_avx512);
+        p.chroma[X265_CSP_I444].pu[LUMA_24x32].filter_hpp = PFX(interp_4tap_horiz_pp_24x32_avx512);
 
     }
 }
diff -r 1fb1948309a0 -r c726239a0758 source/common/x86/ipfilter16.asm
--- a/source/common/x86/ipfilter16.asm	Thu Aug 31 14:54:18 2017 +0530
+++ b/source/common/x86/ipfilter16.asm	Thu Aug 31 15:21:25 2017 +0530
@@ -5161,6 +5161,103 @@
     vextracti32x8   [r2 + r3], m7,        1
 %endmacro
 
+%macro PROCESS_IPFILTER_CHROMA_PP_24x4_AVX512 0
+    ; register map
+    ; m0 , m1 interpolate coeff
+    ; m2 , m3  shuffle order table
+    ; m4 - pd_32
+    ; m5 - zero
+    ; m6 - pw_pixel_max
+
+    movu            ym7,       [r0]
+    vinserti32x8    m7,        [r0 + r1],      1
+    movu            ym8,       [r0 + 8]
+    vinserti32x8    m8,        [r0 + r1 + 8],  1
+
+    pshufb          m9,        m7,        m3
+    pshufb          m7,        m2
+    pmaddwd         m7,        m0
+    pmaddwd         m9,        m1
+    paddd           m7,        m9
+    paddd           m7,        m4
+    psrad           m7,        6
+
+    pshufb          m9,        m8,        m3
+    pshufb          m8,        m2
+    pmaddwd         m8,        m0
+    pmaddwd         m9,        m1
+    paddd           m8,        m9
+    paddd           m8,        m4
+    psrad           m8,        6
+
+    packusdw        m7,        m8
+    CLIPW           m7,        m5,        m6
+    pshufb          m7,        m10
+    movu            [r2],      ym7
+    vextracti32x8   [r2 + r3], m7,        1
+
+    movu            ym7,       [r0 + 2 * r1]
+    vinserti32x8    m7,        [r0 + r6],      1
+    movu            ym8,       [r0 + 2 * r1 + 8]
+    vinserti32x8    m8,        [r0 + r6 + 8],  1
+
+    pshufb          m9,        m7,        m3
+    pshufb          m7,        m2
+    pmaddwd         m7,        m0
+    pmaddwd         m9,        m1
+    paddd           m7,        m9
+    paddd           m7,        m4
+    psrad           m7,        6
+
+    pshufb          m9,        m8,        m3
+    pshufb          m8,        m2
+    pmaddwd         m8,        m0
+    pmaddwd         m9,        m1
+    paddd           m8,        m9
+    paddd           m8,        m4
+    psrad           m8,        6
+
+    packusdw        m7,        m8
+    CLIPW           m7,        m5,        m6
+    pshufb          m7,        m10
+    movu            [r2 + 2 * r3],        ym7
+    vextracti32x8   [r2 + r7], m7,        1
+
+    movu            xm7,       [r0 + mmsize/2]
+    vinserti32x4    m7,        [r0 + r1 + mmsize/2],      1
+    vinserti32x4    m7,        [r0 + 2 * r1 + mmsize/2],  2
+    vinserti32x4    m7,        [r0 + r6 + mmsize/2],      3
+
+    pshufb          m9,        m7,        m3
+    pshufb          m7,        m2
+    pmaddwd         m7,        m0
+    pmaddwd         m9,        m1
+    paddd           m7,        m9
+    paddd           m7,        m4
+    psrad           m7,        6
+
+    movu            xm8,       [r0 + mmsize/2 + 8]
+    vinserti32x4    m8,        [r0 + r1 + mmsize/2 + 8],      1
+    vinserti32x4    m8,        [r0 + 2 * r1 + mmsize/2 + 8],  2
+    vinserti32x4    m8,        [r0 + r6 + mmsize/2 + 8],      3
+
+    pshufb          m9,        m8,        m3
+    pshufb          m8,        m2
+    pmaddwd         m8,        m0
+    pmaddwd         m9,        m1
+    paddd           m8,        m9
+    paddd           m8,        m4
+    psrad           m8,        6
+
+    packusdw        m7,        m8
+    CLIPW           m7,        m5,        m6
+    pshufb          m7,        m10
+    movu            [r2 + mmsize/2],      xm7
+    vextracti32x4   [r2 + r3 + mmsize/2],     m7,        1
+    vextracti32x4   [r2 + 2 * r3 + mmsize/2], m7,        2
+    vextracti32x4   [r2 + r7 + mmsize/2],     m7,        3
+%endmacro
+
 %macro PROCESS_IPFILTER_CHROMA_PP_32x2_AVX512 0
     ; register map
     ; m0 , m1 interpolate coeff
@@ -5514,6 +5611,42 @@
 IPFILTER_CHROMA_AVX512_16xN 64
 
 INIT_ZMM avx512
+%macro IPFILTER_CHROMA_AVX512_24xN 1
+cglobal interp_4tap_horiz_pp_24x%1, 5,8,11
+    add             r1d, r1d
+    add             r3d, r3d
+    sub             r0, 2
+    mov             r4d, r4m
+    lea             r6, [3 * r1]
+    lea             r7, [3 * r3]
+%ifdef PIC
+    lea             r5, [tab_ChromaCoeff]
+    vpbroadcastd    m0, [r5 + r4 * 8]
+    vpbroadcastd    m1, [r5 + r4 * 8 + 4]
+%else
+    vpbroadcastd    m0, [tab_ChromaCoeff + r4 * 8]
+    vpbroadcastd    m1, [tab_ChromaCoeff + r4 * 8 + 4]
+%endif
+    vbroadcasti32x8 m2, [interp8_hpp_shuf1_load_avx512]
+    vbroadcasti32x8 m3, [interp8_hpp_shuf2_load_avx512]
+    vbroadcasti32x8 m4, [pd_32]
+    pxor            m5, m5
+    vbroadcasti32x8 m6, [pw_pixel_max]
+    vbroadcasti32x8 m10, [interp8_hpp_shuf1_store_avx512]
+
+%rep %1/4 - 1
+    PROCESS_IPFILTER_CHROMA_PP_24x4_AVX512
+    lea             r0, [r0 + 4 * r1]
+    lea             r2, [r2 + 4 * r3]
+%endrep
+    PROCESS_IPFILTER_CHROMA_PP_24x4_AVX512
+    RET
+%endmacro
+
+IPFILTER_CHROMA_AVX512_24xN 32
+IPFILTER_CHROMA_AVX512_24xN 64
+
+INIT_ZMM avx512
 %macro IPFILTER_CHROMA_AVX512_32xN 1
 cglobal interp_4tap_horiz_pp_32x%1, 5,6,11
     add             r1d, r1d


More information about the x265-devel mailing list