[x265] [PATCH] asm: chroma_hps[6x8, 6x16] for high bit depth

[aasaipriya at multicorewareinc.com]

# HG changeset patch
# User Aasaipriya Chandran <aasaipriya at multicorewareinc.com>
# Date 1433428704 -19800
#      Thu Jun 04 20:08:24 2015 +0530
# Node ID 394b62c338ede516aaab252d5e207b4ed20efbb3
# Parent  093618ce0b26ea4703b5928f618d2895cf6daf32
asm: chroma_hps[6x8, 6x16] for high bit depth

chroma_hps[ 6x16] - 3.57x   1825c->1716c
chroma_hps[  6x8] - 3.46x   1104c->1005c

diff -r 093618ce0b26 -r 394b62c338ed source/common/x86/asm-primitives.cpp
--- a/source/common/x86/asm-primitives.cpp	Tue Jun 02 17:21:24 2015 +0800
+++ b/source/common/x86/asm-primitives.cpp	Thu Jun 04 20:08:24 2015 +0530
@@ -1579,6 +1579,7 @@
         p.chroma[X265_CSP_I420].pu[CHROMA_420_32x8].filter_hps = x265_interp_4tap_horiz_ps_32x8_avx2;
         p.chroma[X265_CSP_I420].pu[CHROMA_420_24x32].filter_hps = x265_interp_4tap_horiz_ps_24x32_avx2;
         p.chroma[X265_CSP_I420].pu[CHROMA_420_12x16].filter_hps = x265_interp_4tap_horiz_ps_12x16_avx2;
+        p.chroma[X265_CSP_I420].pu[CHROMA_420_6x8].filter_hps = x265_interp_4tap_horiz_ps_6x8_avx2;
         p.chroma[X265_CSP_I422].pu[CHROMA_422_8x8].filter_hps = x265_interp_4tap_horiz_ps_8x8_avx2;
         p.chroma[X265_CSP_I422].pu[CHROMA_422_8x16].filter_hps = x265_interp_4tap_horiz_ps_8x16_avx2;
         p.chroma[X265_CSP_I422].pu[CHROMA_422_8x32].filter_hps = x265_interp_4tap_horiz_ps_8x32_avx2;
@@ -1596,6 +1597,7 @@
         p.chroma[X265_CSP_I422].pu[CHROMA_422_32x16].filter_hps = x265_interp_4tap_horiz_ps_32x16_avx2;
         p.chroma[X265_CSP_I422].pu[CHROMA_422_12x32].filter_hps = x265_interp_4tap_horiz_ps_12x32_avx2;
         p.chroma[X265_CSP_I422].pu[CHROMA_422_24x64].filter_hps = x265_interp_4tap_horiz_ps_24x64_avx2;
+        p.chroma[X265_CSP_I422].pu[CHROMA_422_6x16].filter_hps = x265_interp_4tap_horiz_ps_6x16_avx2;
         p.chroma[X265_CSP_I444].pu[LUMA_8x8].filter_hps = x265_interp_4tap_horiz_ps_8x8_avx2;
         p.chroma[X265_CSP_I444].pu[LUMA_8x4].filter_hps = x265_interp_4tap_horiz_ps_8x4_avx2;
         p.chroma[X265_CSP_I444].pu[LUMA_8x16].filter_hps = x265_interp_4tap_horiz_ps_8x16_avx2;
@@ -1678,6 +1680,7 @@
         p.chroma[X265_CSP_I444].pu[LUMA_64x64].filter_hpp = x265_interp_4tap_horiz_pp_64x64_avx2;
         p.chroma[X265_CSP_I444].pu[LUMA_48x64].filter_hpp = x265_interp_4tap_horiz_pp_48x64_avx2;
 
+
         if (cpuMask & X265_CPU_BMI2)
             p.scanPosLast = x265_scanPosLast_avx2_bmi2;
     }
diff -r 093618ce0b26 -r 394b62c338ed source/common/x86/ipfilter16.asm
--- a/source/common/x86/ipfilter16.asm	Tue Jun 02 17:21:24 2015 +0800
+++ b/source/common/x86/ipfilter16.asm	Thu Jun 04 20:08:24 2015 +0530
@@ -9847,3 +9847,57 @@
     jnz                 .loop0
     RET
 %endif
+
+%macro IPFILTER_CHROMA_PS_6xN_AVX2 1
+INIT_YMM avx2
+%if ARCH_X86_64 == 1
+cglobal interp_4tap_horiz_ps_6x%1, 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, %1
+    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
+    movq                [r2], xm4
+    pextrd              [r2 + 8], xm4, 2
+    add                 r2, r3
+    add                 r0, r1
+    dec                 r4d
+    jnz                 .loop0
+    RET
+%endif
+%endmacro
+
+    IPFILTER_CHROMA_PS_6xN_AVX2 8
+    IPFILTER_CHROMA_PS_6xN_AVX2 16