[x265] [PATCH] Luma_hpp[64x64 , 64x48 , 64x32 , 64x16] avx2 asm code : improved 33137c->22606c , 24826c->17202c , 16726c->11560c , 7830c->5534c

aasaipriya at multicorewareinc.com aasaipriya at multicorewareinc.com
Tue Nov 25 07:54:06 CET 2014


# HG changeset patch
# User Aasaipriya Chandran <aasaipriya at multicorewareinc.com>
# Date 1416898324 -19800
#      Tue Nov 25 12:22:04 2014 +0530
# Node ID 507aedba064f7a1f2ee27d945ee25f17b2365548
# Parent  98fb658f3229ab10e808204c265a12e18d71638e
Luma_hpp[64x64 , 64x48 , 64x32 , 64x16] avx2 asm code : improved 33137c->22606c , 24826c->17202c , 16726c->11560c , 7830c->5534c

diff -r 98fb658f3229 -r 507aedba064f source/common/x86/asm-primitives.cpp
--- a/source/common/x86/asm-primitives.cpp	Tue Nov 11 19:30:19 2014 +0900
+++ b/source/common/x86/asm-primitives.cpp	Tue Nov 25 12:22:04 2014 +0530
@@ -1799,6 +1799,12 @@
         p.transpose[BLOCK_64x64] = x265_transpose64_avx2;
 #endif
         p.luma_hpp[LUMA_4x4] = x265_interp_8tap_horiz_pp_4x4_avx2;
+
+        p.luma_hpp[LUMA_64x64] = x265_interp_8tap_horiz_pp_64x64_avx2;
+        p.luma_hpp[LUMA_64x48] = x265_interp_8tap_horiz_pp_64x48_avx2;
+        p.luma_hpp[LUMA_64x32] = x265_interp_8tap_horiz_pp_64x32_avx2;
+        p.luma_hpp[LUMA_64x16] = x265_interp_8tap_horiz_pp_64x16_avx2;
+
         p.luma_vpp[LUMA_4x4] = x265_interp_8tap_vert_pp_4x4_avx2;
     }
 #endif // if HIGH_BIT_DEPTH
diff -r 98fb658f3229 -r 507aedba064f source/common/x86/ipfilter8.asm
--- a/source/common/x86/ipfilter8.asm	Tue Nov 11 19:30:19 2014 +0900
+++ b/source/common/x86/ipfilter8.asm	Tue Nov 25 12:22:04 2014 +0530
@@ -854,6 +854,117 @@
     RET
 
 
+%macro IPFILTER_LUMA_64x_avx2 2
+INIT_YMM avx2
+cglobal interp_8tap_horiz_pp_%1x%2, 4,6,8
+    sub               r0,    3
+    mov               r4d,   r4m
+%ifdef PIC
+    lea               r5,        [tab_LumaCoeff]
+    vpbroadcastd      m0,        [r5 + r4 * 8]
+    vpbroadcastd      m1,        [r5 + r4 * 8 + 4]
+%else
+    vpbroadcastd      m0,        [tab_LumaCoeff + r4 * 8]
+    vpbroadcastd      m1,        [tab_LumaCoeff + r4 * 8 + 4]
+%endif
+    movu              m3,        [tab_Tm + 16]
+    vpbroadcastd      m7,        [pw_1]
+
+    ; register map
+    ; m0 , m1 interpolate coeff
+    ; m2 , m2  shuffle order table
+    ; m7 - pw_1
+
+    mov               r4d,   %2
+.loop:
+    ; Row 0
+    vbroadcasti128    m4,        [r0]                        ; [x E D C B A 9 8 7 6 5 4 3 2 1 0]
+    pshufb            m5,        m4,    m3
+    pshufb            m4,        [tab_Tm]
+    pmaddubsw         m4,        m0
+    pmaddubsw         m5,        m1
+    paddw             m4,        m5
+    pmaddwd           m4,        m7
+    vbroadcasti128    m5,        [r0 + 8]
+    pshufb            m6,        m5,    m3
+    pshufb            m5,        [tab_Tm]
+    pmaddubsw         m5,        m0
+    pmaddubsw         m6,        m1
+    paddw             m5,        m6
+    pmaddwd           m5,        m7
+    packssdw          m4,        m5                          ; [17 16 15 14 07 06 05 04 13 12 11 10 03 02 01 00]
+    pmulhrsw          m4,        [pw_512]
+    vbroadcasti128    m2,        [r0 + 16]
+    pshufb            m5,        m2,    m3
+    pshufb            m2,        [tab_Tm]
+    pmaddubsw         m2,        m0
+    pmaddubsw         m5,        m1
+    paddw             m2,        m5
+    pmaddwd           m2,        m7
+    vbroadcasti128    m5,        [r0 + 24]
+    pshufb            m6,        m5,    m3
+    pshufb            m5,        [tab_Tm]
+    pmaddubsw         m5,        m0
+    pmaddubsw         m6,        m1
+    paddw             m5,        m6
+    pmaddwd           m5,        m7
+    packssdw          m2,        m5
+    pmulhrsw          m2,        [pw_512]
+    packuswb          m4,        m2
+    vpermq            m4,        m4,    11011000b
+    vextracti128      xm5,       m4,    1
+    pshufd            xm4,       xm4,   11011000b
+    pshufd            xm5,       xm5,   11011000b
+    movu              [r2],      xm4
+    movu              [r2 + 16], xm5
+
+    vbroadcasti128    m4,        [r0 + 32]
+    pshufb            m5,        m4,    m3
+    pshufb            m4,        [tab_Tm]
+    pmaddubsw         m4,        m0
+    pmaddubsw         m5,        m1
+    paddw             m4,        m5
+    pmaddwd           m4,        m7
+    vbroadcasti128    m5,        [r0 + 40]
+    pshufb            m6,        m5,    m3
+    pshufb            m5,        [tab_Tm]
+    pmaddubsw         m5,        m0
+    pmaddubsw         m6,        m1
+    paddw             m5,        m6
+    pmaddwd           m5,        m7
+    packssdw          m4,        m5
+    pmulhrsw          m4,        [pw_512]
+    vbroadcasti128    m2,        [r0 + 48]
+    pshufb            m5,        m2,    m3
+    pshufb            m2,        [tab_Tm]
+    pmaddubsw         m2,        m0
+    pmaddubsw         m5,        m1
+    paddw             m2,        m5
+    pmaddwd           m2,        m7
+    vbroadcasti128    m5,        [r0 + 56]
+    pshufb            m6,        m5,    m3
+    pshufb            m5,        [tab_Tm]
+    pmaddubsw         m5,        m0
+    pmaddubsw         m6,        m1
+    paddw             m5,        m6
+    pmaddwd           m5,        m7
+    packssdw          m2,        m5
+    pmulhrsw          m2,        [pw_512]
+    packuswb          m4,        m2
+    vpermq            m4,        m4,    11011000b
+    vextracti128      xm5,       m4,    1
+    pshufd            xm4,       xm4,   11011000b
+    pshufd            xm5,       xm5,   11011000b
+    movu              [r2 +32],  xm4
+    movu              [r2 + 48], xm5
+
+    lea               r0,        [r0 + r1]
+    lea               r2,        [r2 + r3]
+    dec               r4d
+    jnz               .loop
+    RET
+%endmacro
+
 ;--------------------------------------------------------------------------------------------------------------
 ; void interp_8tap_horiz_pp_%1x%2(pixel *src, intptr_t srcStride, pixel *dst, intptr_t dstStride, int coeffIdx)
 ;--------------------------------------------------------------------------------------------------------------
@@ -862,6 +973,11 @@
     IPFILTER_LUMA 12, 16, pp
     IPFILTER_LUMA 4, 16, pp
 
+    IPFILTER_LUMA_64x_avx2 64 , 64
+    IPFILTER_LUMA_64x_avx2 64 , 48
+    IPFILTER_LUMA_64x_avx2 64 , 32
+    IPFILTER_LUMA_64x_avx2 64 , 16
+
 ;--------------------------------------------------------------------------------------------------------------
 ; void interp_8tap_horiz_pp_%1x%2(pixel *src, intptr_t srcStride, pixel *dst, intptr_t dstStride, int coeffIdx)
 ;--------------------------------------------------------------------------------------------------------------


More information about the x265-devel mailing list