[x265] [PATCH] asm: avx2 10bit code for convert_p2s[16xN]

[rajesh at multicorewareinc.com]

# HG changeset patch
# User Rajesh Paulraj<rajesh at multicorewareinc.com>
# Date 1429188456 -19800
#      Thu Apr 16 18:17:36 2015 +0530
# Node ID 8d14e23666ec5fb90a85d90032f52d3fb3e6fe4a
# Parent  7be1172ec816298c32f588908e1b6f0fa214d349
asm: avx2 10bit code for convert_p2s[16xN]

     convert_p2s[16x4](10.44x), convert_p2s[16x8](15.30x),
     convert_p2s[16x12](16.55x), convert_p2s[16x16](17.48x),
     convert_p2s[16x32](17.57x), convert_p2s[16x64](20.21x)

diff -r 7be1172ec816 -r 8d14e23666ec source/common/x86/asm-primitives.cpp
--- a/source/common/x86/asm-primitives.cpp	Thu Apr 16 11:38:32 2015 +0530
+++ b/source/common/x86/asm-primitives.cpp	Thu Apr 16 18:17:36 2015 +0530
@@ -1223,6 +1223,13 @@
         ALL_LUMA_PU(luma_vps, interp_8tap_vert_ps, avx2);
         ALL_LUMA_PU(luma_vsp, interp_8tap_vert_sp, avx2);
         ALL_LUMA_PU(luma_vss, interp_8tap_vert_ss, avx2);
+
+        p.pu[LUMA_16x4].convert_p2s = x265_filterPixelToShort_16x4_avx2;
+        p.pu[LUMA_16x8].convert_p2s = x265_filterPixelToShort_16x8_avx2;
+        p.pu[LUMA_16x12].convert_p2s = x265_filterPixelToShort_16x12_avx2;
+        p.pu[LUMA_16x16].convert_p2s = x265_filterPixelToShort_16x16_avx2;
+        p.pu[LUMA_16x32].convert_p2s = x265_filterPixelToShort_16x32_avx2;
+        p.pu[LUMA_16x64].convert_p2s = x265_filterPixelToShort_16x64_avx2;
     }
 }
 #else // if HIGH_BIT_DEPTH
diff -r 7be1172ec816 -r 8d14e23666ec source/common/x86/ipfilter16.asm
--- a/source/common/x86/ipfilter16.asm	Thu Apr 16 11:38:32 2015 +0530
+++ b/source/common/x86/ipfilter16.asm	Thu Apr 16 18:17:36 2015 +0530
@@ -5878,6 +5878,60 @@
 ;-----------------------------------------------------------------------------
 ; void filterPixelToShort(pixel *src, intptr_t srcStride, int16_t *dst, intptr_t dstStride)
 ;-----------------------------------------------------------------------------
+%macro P2S_H_16xN_avx2 1
+INIT_YMM avx2
+cglobal filterPixelToShort_16x%1, 3, 7, 3
+    add        r1d, r1d
+    mov        r3d, r3m
+    add        r3d, r3d
+    lea        r4, [r3 * 3]
+    lea        r5, [r1 * 3]
+
+    ; load height
+    mov        r6d, %1/4
+
+    ; load constant
+    mova       m2, [pw_2000]
+
+.loop
+    movu       m0, [r0]
+    movu       m1, [r0 + r1]
+    psllw      m0, 4
+    psubw      m0, m2
+    psllw      m1, 4
+    psubw      m1, m2
+
+    movu       [r2 + r3 * 0], m0
+    movu       [r2 + r3 * 1], m1
+
+    movu       m0, [r0 + r1 * 2]
+    movu       m1, [r0 + r5]
+    psllw      m0, 4
+    psubw      m0, m2
+    psllw      m1, 4
+    psubw      m1, m2
+
+    movu       [r2 + r3 * 2], m0
+    movu       [r2 + r4], m1
+
+    lea        r0, [r0 + r1 * 4]
+    lea        r2, [r2 + r3 * 4]
+
+    dec        r6d
+    jnz        .loop
+    RET
+%endmacro
+P2S_H_16xN_avx2 16
+P2S_H_16xN_avx2 4
+P2S_H_16xN_avx2 8
+P2S_H_16xN_avx2 12
+P2S_H_16xN_avx2 32
+P2S_H_16xN_avx2 64
+P2S_H_16xN_avx2 24
+
+;-----------------------------------------------------------------------------
+; void filterPixelToShort(pixel *src, intptr_t srcStride, int16_t *dst, intptr_t dstStride)
+;-----------------------------------------------------------------------------
 %macro P2S_H_32xN 1
 INIT_XMM ssse3
 cglobal filterPixelToShort_32x%1, 3, 7, 5
diff -r 7be1172ec816 -r 8d14e23666ec source/common/x86/ipfilter8.h
--- a/source/common/x86/ipfilter8.h	Thu Apr 16 11:38:32 2015 +0530
+++ b/source/common/x86/ipfilter8.h	Thu Apr 16 18:17:36 2015 +0530
@@ -314,6 +314,12 @@
 void x265_filterPixelToShort_24x32_ssse3(const pixel* src, intptr_t srcStride, int16_t* dst, intptr_t dstStride);
 void x265_filterPixelToShort_12x16_ssse3(const pixel* src, intptr_t srcStride, int16_t* dst, intptr_t dstStride);
 void x265_filterPixelToShort_48x64_ssse3(const pixel* src, intptr_t srcStride, int16_t* dst, intptr_t dstStride);
+void x265_filterPixelToShort_16x4_avx2(const pixel* src, intptr_t srcStride, int16_t* dst, intptr_t dstStride);
+void x265_filterPixelToShort_16x8_avx2(const pixel* src, intptr_t srcStride, int16_t* dst, intptr_t dstStride);
+void x265_filterPixelToShort_16x12_avx2(const pixel* src, intptr_t srcStride, int16_t* dst, intptr_t dstStride);
+void x265_filterPixelToShort_16x16_avx2(const pixel* src, intptr_t srcStride, int16_t* dst, intptr_t dstStride);
+void x265_filterPixelToShort_16x32_avx2(const pixel* src, intptr_t srcStride, int16_t* dst, intptr_t dstStride);
+void x265_filterPixelToShort_16x64_avx2(const pixel* src, intptr_t srcStride, int16_t* dst, intptr_t dstStride);
 
 #define SETUP_CHROMA_P2S_FUNC_DEF(W, H, cpu) \
     void x265_filterPixelToShort_ ## W ## x ## H ## cpu(const pixel* src, intptr_t srcStride, int16_t* dst, intptr_t dstStride);