[x265] [PATCH] asm: chroma_hps[48x64 for i444] high bit depth - 4.88x 38394c->32817c

[aasaipriya at multicorewareinc.com]

# HG changeset patch
# User Aasaipriya Chandran <aasaipriya at multicorewareinc.com>
# Date 1432875116 -19800
#      Fri May 29 10:21:56 2015 +0530
# Node ID 7fccb7d165b1c11af9e3d329fad2d818175fc8c4
# Parent  d1672c92a283beb80c2fbea91a3cd098f26a7263
asm: chroma_hps[48x64 for i444] high bit depth - 4.88x    38394c->32817c

diff -r d1672c92a283 -r 7fccb7d165b1 source/common/x86/asm-primitives.cpp
--- a/source/common/x86/asm-primitives.cpp	Fri May 29 10:11:44 2015 +0530
+++ b/source/common/x86/asm-primitives.cpp	Fri May 29 10:21:56 2015 +0530
@@ -1589,6 +1589,7 @@
         p.chroma[X265_CSP_I444].pu[LUMA_64x48].filter_hps = x265_interp_4tap_horiz_ps_64x48_avx2;
         p.chroma[X265_CSP_I444].pu[LUMA_64x32].filter_hps = x265_interp_4tap_horiz_ps_64x32_avx2;
         p.chroma[X265_CSP_I444].pu[LUMA_64x16].filter_hps = x265_interp_4tap_horiz_ps_64x16_avx2;
+        p.chroma[X265_CSP_I444].pu[LUMA_48x64].filter_hps = x265_interp_4tap_horiz_ps_48x64_avx2;
 
         p.chroma[X265_CSP_I420].pu[CHROMA_420_6x8].filter_hpp = x265_interp_4tap_horiz_pp_6x8_avx2;
         p.chroma[X265_CSP_I420].pu[CHROMA_420_8x2].filter_hpp = x265_interp_4tap_horiz_pp_8x2_avx2;
diff -r d1672c92a283 -r 7fccb7d165b1 source/common/x86/ipfilter16.asm
--- a/source/common/x86/ipfilter16.asm	Fri May 29 10:11:44 2015 +0530
+++ b/source/common/x86/ipfilter16.asm	Fri May 29 10:21:56 2015 +0530
@@ -9522,3 +9522,122 @@
 IPFILTER_CHROMA_PS_64xN_AVX2 48
 IPFILTER_CHROMA_PS_64xN_AVX2 32
 IPFILTER_CHROMA_PS_64xN_AVX2 16
+
+INIT_YMM avx2
+%if ARCH_X86_64 == 1
+cglobal interp_4tap_horiz_ps_48x64, 4, 7, 6
+    add                 r1d, r1d
+    add                 r3d, r3d
+    mov                 r4d, r4m
+    mov                 r5d, r5m
+
+%ifdef PIC
+    lea                 r6, [tab_ChromaCoeff]
+    vpbroadcastq        m0, [r6 + r4 * 8]
+%else
+    vpbroadcastq        m0, [tab_ChromaCoeff + r4 * 8]
+%endif
+    mova                m3, [pb_shuf]
+    vbroadcasti128      m2, [pd_n32768]
+
+    ; register map
+    ; m0 , m1 interpolate coeff
+
+    sub                 r0, 2
+    test                r5d, r5d
+    mov                 r4d, 64
+    jz                  .loop0
+    sub                 r0, r1
+    add                 r4d, 3
+
+.loop0:
+    vbroadcasti128      m4, [r0]
+    vbroadcasti128      m5, [r0 + 8]
+    pshufb              m4, m3
+    pshufb              m5, m3
+    pmaddwd             m4, m0
+    pmaddwd             m5, m0
+    phaddd              m4, m5
+    paddd               m4, m2
+    vpermq              m4, m4, q3120
+    psrad               m4, 2
+    vextracti128        xm5, m4, 1
+    packssdw            xm4, xm5
+    movu                [r2], xm4
+
+    vbroadcasti128      m4, [r0 + 16]
+    vbroadcasti128      m5, [r0 + 24]
+    pshufb              m4, m3
+    pshufb              m5, m3
+    pmaddwd             m4, m0
+    pmaddwd             m5, m0
+    phaddd              m4, m5
+    paddd               m4, m2
+    vpermq              m4, m4, q3120
+    psrad               m4, 2
+    vextracti128        xm5, m4, 1
+    packssdw            xm4, xm5
+    movu                [r2 + 16], xm4
+
+    vbroadcasti128      m4, [r0 + 32]
+    vbroadcasti128      m5, [r0 + 40]
+    pshufb              m4, m3
+    pshufb              m5, m3
+    pmaddwd             m4, m0
+    pmaddwd             m5, m0
+    phaddd              m4, m5
+    paddd               m4, m2
+    vpermq              m4, m4, q3120
+    psrad               m4, 2
+    vextracti128        xm5, m4, 1
+    packssdw            xm4, xm5
+    movu                [r2 + 32], xm4
+
+    vbroadcasti128      m4, [r0 + 48]
+    vbroadcasti128      m5, [r0 + 56]
+    pshufb              m4, m3
+    pshufb              m5, m3
+    pmaddwd             m4, m0
+    pmaddwd             m5, m0
+    phaddd              m4, m5
+    paddd               m4, m2
+    vpermq              m4, m4, q3120
+    psrad               m4, 2
+    vextracti128        xm5, m4, 1
+    packssdw            xm4, xm5
+    movu                [r2 + 48], xm4
+
+    vbroadcasti128      m4, [r0 + 64]
+    vbroadcasti128      m5, [r0 + 72]
+    pshufb              m4, m3
+    pshufb              m5, m3
+    pmaddwd             m4, m0
+    pmaddwd             m5, m0
+    phaddd              m4, m5
+    paddd               m4, m2
+    vpermq              m4, m4, q3120
+    psrad               m4, 2
+    vextracti128        xm5, m4, 1
+    packssdw            xm4, xm5
+    movu                [r2 + 64], xm4
+
+    vbroadcasti128      m4, [r0 + 80]
+    vbroadcasti128      m5, [r0 + 88]
+    pshufb              m4, m3
+    pshufb              m5, m3
+    pmaddwd             m4, m0
+    pmaddwd             m5, m0
+    phaddd              m4, m5
+    paddd               m4, m2
+    vpermq              m4, m4, q3120
+    psrad               m4, 2
+    vextracti128        xm5, m4, 1
+    packssdw            xm4, xm5
+    movu                [r2 + 80], xm4
+
+    add                 r2, r3
+    add                 r0, r1
+    dec                 r4d
+    jnz                 .loop0
+    RET
+%endif