[x265] [PATCH] asm: chroma_hps[32xN for i420] high bit depth

[aasaipriya at multicorewareinc.com]

# HG changeset patch
# User Aasaipriya Chandran <aasaipriya at multicorewareinc.com>
# Date 1432620759 -19800
#      Tue May 26 11:42:39 2015 +0530
# Node ID f67f224a17c47b7c72c024f94a4a2e71cc6b7e3b
# Parent  6ec17b5bec01a4f135d7af7c9aaeac5dd98c6376
asm: chroma_hps[32xN for i420] high bit depth

chroma_hps[32x32] - 7.40x 21402c->15441c
chroma_hps[32x8]  - 6.34c 8647c->5632c
chroma_hps[32x16] - 6.71c 13101c->8885c
chroma_hps[32x24] - 6.93c 18954c->12368c

diff -r 6ec17b5bec01 -r f67f224a17c4 source/common/x86/asm-primitives.cpp
--- a/source/common/x86/asm-primitives.cpp	Tue May 26 11:24:42 2015 +0530
+++ b/source/common/x86/asm-primitives.cpp	Tue May 26 11:42:39 2015 +0530
@@ -1518,6 +1518,10 @@
         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;
+        p.chroma[X265_CSP_I420].pu[CHROMA_420_32x32].filter_hps = x265_interp_4tap_horiz_ps_32x32_avx2;
+        p.chroma[X265_CSP_I420].pu[CHROMA_420_32x16].filter_hps = x265_interp_4tap_horiz_ps_32x16_avx2;
+        p.chroma[X265_CSP_I420].pu[CHROMA_420_32x24].filter_hps = x265_interp_4tap_horiz_ps_32x24_avx2;
+        p.chroma[X265_CSP_I420].pu[CHROMA_420_32x8].filter_hps = x265_interp_4tap_horiz_ps_32x8_avx2;
 
         if (cpuMask & X265_CPU_BMI2)
             p.scanPosLast = x265_scanPosLast_avx2_bmi2;
diff -r 6ec17b5bec01 -r f67f224a17c4 source/common/x86/ipfilter16.asm
--- a/source/common/x86/ipfilter16.asm	Tue May 26 11:24:42 2015 +0530
+++ b/source/common/x86/ipfilter16.asm	Tue May 26 11:42:39 2015 +0530
@@ -8196,3 +8196,101 @@
 IPFILTER_CHROMA_PS_16xN_AVX2 32
 IPFILTER_CHROMA_PS_16xN_AVX2 12
 IPFILTER_CHROMA_PS_16xN_AVX2 4
+
+%macro IPFILTER_CHROMA_PS_32xN_AVX2 1
+INIT_YMM avx2
+%if ARCH_X86_64 == 1
+cglobal interp_4tap_horiz_ps_32x%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
+
+    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
+
+    add                 r2, r3
+    add                 r0, r1
+    dec                 r4d
+    jnz                 .loop0
+    RET
+%endif
+%endmacro
+
+IPFILTER_CHROMA_PS_32xN_AVX2 32
+IPFILTER_CHROMA_PS_32xN_AVX2 16
+IPFILTER_CHROMA_PS_32xN_AVX2 24
+IPFILTER_CHROMA_PS_32xN_AVX2 8