[x265] [PATCH 103 of 307] x86: AVX512 interp_4tap_horiz_pp_64xN for high bit depth

mythreyi at multicorewareinc.com mythreyi at multicorewareinc.com
Sat Apr 7 04:31:41 CEST 2018


# HG changeset patch
# User Vignesh Vijayakumar
# Date 1504160983 -19800
#      Thu Aug 31 11:59:43 2017 +0530
# Node ID e5efe0bdbe0feef4588e29c81bee16651351c50f
# Parent  a98d83b53b2e32b4060b05342ee041c44eff2045
x86: AVX512 interp_4tap_horiz_pp_64xN for high bit depth

i444
Size  |  AVX2 performance | AVX512 performance
----------------------------------------------
64x16 |       9.07x       |      20.51x
64x32 |       9.51x       |      19.58x
64x48 |       9.13x       |      20.72x
64x64 |       9.12x       |      20.52x

diff -r a98d83b53b2e -r e5efe0bdbe0f source/common/x86/asm-primitives.cpp
--- a/source/common/x86/asm-primitives.cpp	Thu Aug 31 11:51:39 2017 +0530
+++ b/source/common/x86/asm-primitives.cpp	Thu Aug 31 11:59:43 2017 +0530
@@ -2369,6 +2369,10 @@
         p.chroma[X265_CSP_I444].pu[LUMA_32x24].filter_hpp = PFX(interp_4tap_horiz_pp_32x24_avx512);
         p.chroma[X265_CSP_I444].pu[LUMA_32x32].filter_hpp = PFX(interp_4tap_horiz_pp_32x32_avx512);
         p.chroma[X265_CSP_I444].pu[LUMA_32x64].filter_hpp = PFX(interp_4tap_horiz_pp_32x64_avx512);
+        p.chroma[X265_CSP_I444].pu[LUMA_64x16].filter_hpp = PFX(interp_4tap_horiz_pp_64x16_avx512);
+        p.chroma[X265_CSP_I444].pu[LUMA_64x32].filter_hpp = PFX(interp_4tap_horiz_pp_64x32_avx512);
+        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);
 
     }
 }
diff -r a98d83b53b2e -r e5efe0bdbe0f source/common/x86/ipfilter16.asm
--- a/source/common/x86/ipfilter16.asm	Thu Aug 31 11:51:39 2017 +0530
+++ b/source/common/x86/ipfilter16.asm	Thu Aug 31 11:59:43 2017 +0530
@@ -5139,6 +5139,110 @@
     movu            [r2 + r3], m7
 %endmacro
 
+%macro PROCESS_IPFILTER_CHROMA_PP_64x2_AVX512 0
+    ; register map
+    ; m0 , m1 interpolate coeff
+    ; m2 , m3  shuffle order table
+    ; m4 - pd_32
+    ; m5 - zero
+    ; m6 - pw_pixel_max
+
+    movu            m7,        [r0]
+    movu            m8,        [r0 + 8]
+
+    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],      m7
+
+    movu            m7,        [r0 + mmsize]
+    movu            m8,        [r0 + mmsize + 8]
+
+    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 + mmsize],        m7
+
+    movu            m7,        [r0 + r1]
+    movu            m8,        [r0 + r1 + 8]
+
+    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 + r3], m7
+
+    movu            m7,        [r0 + r1 + mmsize]
+    movu            m8,        [r0 + r1 + mmsize + 8]
+
+    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 + r3 + mmsize],   m7
+%endmacro
 ;-------------------------------------------------------------------------------------------------------------
 ; void interp_4tap_horiz_pp(pixel *src, intptr_t srcStride, int16_t *dst, intptr_t dstStride, int coeffIdx
 ;-------------------------------------------------------------------------------------------------------------
@@ -5179,6 +5283,42 @@
 IPFILTER_CHROMA_AVX512_32xN 32
 IPFILTER_CHROMA_AVX512_32xN 48
 IPFILTER_CHROMA_AVX512_32xN 64
+
+INIT_ZMM avx512
+%macro IPFILTER_CHROMA_AVX512_64xN 1
+cglobal interp_4tap_horiz_pp_64x%1, 5,6,11
+    add             r1d, r1d
+    add             r3d, r3d
+    sub             r0, 2
+    mov             r4d, r4m
+%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 %1/2 - 1
+    PROCESS_IPFILTER_CHROMA_PP_64x2_AVX512
+    lea             r0, [r0 + 2 * r1]
+    lea             r2, [r2 + 2 * r3]
+%endrep
+    PROCESS_IPFILTER_CHROMA_PP_64x2_AVX512
+    RET
+%endmacro
+
+IPFILTER_CHROMA_AVX512_64xN 16
+IPFILTER_CHROMA_AVX512_64xN 32
+IPFILTER_CHROMA_AVX512_64xN 48
+IPFILTER_CHROMA_AVX512_64xN 64
 ;-------------------------------------------------------------------------------------------------------------
 ;ipfilter_chroma_avx512 code end
 ;-------------------------------------------------------------------------------------------------------------
\ No newline at end of file


More information about the x265-devel mailing list