[x265] [PATCH] asm: chroma_hps[16xN for i420] high bit depth

[aasaipriya at multicorewareinc.com]

# HG changeset patch
# User Aasaipriya Chandran <aasaipriya at multicorewareinc.com>
# Date 1432619682 -19800
#      Tue May 26 11:24:42 2015 +0530
# Node ID 6ec17b5bec01a4f135d7af7c9aaeac5dd98c6376
# Parent  413177b405e03e27030328846441578995bd1f85
asm: chroma_hps[16xN for i420] high bit depth

chroma_hps[16x16] - 6.84x 7044c->4379c
chroma_hps[16x4]  - 7.50x 2377c->1508c
chroma_hps[16x8]  - 6.00x 4644c->3935c
chroma_hps[16x12] - 6.57c 5818c->3615c
chroma_hps[16x32] - 7.37c 12290c->7536c

diff -r 413177b405e0 -r 6ec17b5bec01 source/common/x86/asm-primitives.cpp
--- a/source/common/x86/asm-primitives.cpp	Tue May 26 11:01:01 2015 +0530
+++ b/source/common/x86/asm-primitives.cpp	Tue May 26 11:24:42 2015 +0530
@@ -1513,6 +1513,11 @@
         p.chroma[X265_CSP_I420].pu[CHROMA_420_8x6].filter_hps = x265_interp_4tap_horiz_ps_8x6_avx2;
         p.chroma[X265_CSP_I420].pu[CHROMA_420_8x2].filter_hps = x265_interp_4tap_horiz_ps_8x2_avx2;
         p.chroma[X265_CSP_I420].pu[CHROMA_420_8x32].filter_hps = x265_interp_4tap_horiz_ps_8x32_avx2;
+        p.chroma[X265_CSP_I420].pu[CHROMA_420_16x16].filter_hps = x265_interp_4tap_horiz_ps_16x16_avx2;
+        p.chroma[X265_CSP_I420].pu[CHROMA_420_16x8].filter_hps = x265_interp_4tap_horiz_ps_16x8_avx2;
+        p.chroma[X265_CSP_I420].pu[CHROMA_420_16x32].filter_hps = x265_interp_4tap_horiz_ps_16x32_avx2;
+        p.chroma[X265_CSP_I420].pu[CHROMA_420_16x12].filter_hps = x265_interp_4tap_horiz_ps_16x12_avx2;
+        p.chroma[X265_CSP_I420].pu[CHROMA_420_16x4].filter_hps = x265_interp_4tap_horiz_ps_16x4_avx2;
 
         if (cpuMask & X265_CPU_BMI2)
             p.scanPosLast = x265_scanPosLast_avx2_bmi2;
diff -r 413177b405e0 -r 6ec17b5bec01 source/common/x86/ipfilter16.asm
--- a/source/common/x86/ipfilter16.asm	Tue May 26 11:01:01 2015 +0530
+++ b/source/common/x86/ipfilter16.asm	Tue May 26 11:24:42 2015 +0530
@@ -8125,3 +8125,74 @@
     IPFILTER_CHROMA_PS_8xN_AVX2 32
     IPFILTER_CHROMA_PS_8xN_AVX2 6
     IPFILTER_CHROMA_PS_8xN_AVX2 2
+
+%macro IPFILTER_CHROMA_PS_16xN_AVX2 1
+INIT_YMM avx2
+%if ARCH_X86_64 == 1
+cglobal interp_4tap_horiz_ps_16x%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
+    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
+
+    add                 r2, r3
+    add                 r0, r1
+    dec                 r4d
+    jnz                 .loop0
+    RET
+%endif
+%endmacro
+
+IPFILTER_CHROMA_PS_16xN_AVX2 16
+IPFILTER_CHROMA_PS_16xN_AVX2 8
+IPFILTER_CHROMA_PS_16xN_AVX2 32
+IPFILTER_CHROMA_PS_16xN_AVX2 12
+IPFILTER_CHROMA_PS_16xN_AVX2 4