[x265] [PATCH] asm: filter_vsp, filter_vss for 4x32 in avx2

[Divya Manivannan]

# HG changeset patch
# User Divya Manivannan <divya at multicorewareinc.com>
# Date 1430217656 -19800
#      Tue Apr 28 16:10:56 2015 +0530
# Node ID 0b564e4a433980c0d029e90d118ad3f7c45afa9c
# Parent  8d34f596917997d680252647380bba1aada97d37
asm: filter_vsp, filter_vss for 4x32 in avx2

filter_vsp[4x32]: 1750c->1159c
filter_vss[4x32]: 1409c->969c

diff -r 8d34f5969179 -r 0b564e4a4339 source/common/x86/asm-primitives.cpp
--- a/source/common/x86/asm-primitives.cpp	Tue Apr 28 15:47:21 2015 +0530
+++ b/source/common/x86/asm-primitives.cpp	Tue Apr 28 16:10:56 2015 +0530
@@ -2288,6 +2288,7 @@
         p.chroma[X265_CSP_I422].pu[CHROMA_422_2x16].filter_vss = x265_interp_4tap_vert_ss_2x16_avx2;
         p.chroma[X265_CSP_I422].pu[CHROMA_422_16x24].filter_vss = x265_interp_4tap_vert_ss_16x24_avx2;
         p.chroma[X265_CSP_I422].pu[CHROMA_422_12x32].filter_vss = x265_interp_4tap_vert_ss_12x32_avx2;
+        p.chroma[X265_CSP_I422].pu[CHROMA_422_4x32].filter_vss = x265_interp_4tap_vert_ss_4x32_avx2;
 
         //i444 for chroma_vss
         p.chroma[X265_CSP_I444].pu[LUMA_4x4].filter_vss = x265_interp_4tap_vert_ss_4x4_avx2;
@@ -2477,6 +2478,7 @@
         p.chroma[X265_CSP_I422].pu[CHROMA_422_2x16].filter_vsp = x265_interp_4tap_vert_sp_2x16_avx2;
         p.chroma[X265_CSP_I422].pu[CHROMA_422_16x24].filter_vsp = x265_interp_4tap_vert_sp_16x24_avx2;
         p.chroma[X265_CSP_I422].pu[CHROMA_422_12x32].filter_vsp = x265_interp_4tap_vert_sp_12x32_avx2;
+        p.chroma[X265_CSP_I422].pu[CHROMA_422_4x32].filter_vsp = x265_interp_4tap_vert_sp_4x32_avx2;
 
         //i444 for chroma_vsp
         p.chroma[X265_CSP_I444].pu[LUMA_4x4].filter_vsp = x265_interp_4tap_vert_sp_4x4_avx2;
diff -r 8d34f5969179 -r 0b564e4a4339 source/common/x86/ipfilter8.asm
--- a/source/common/x86/ipfilter8.asm	Tue Apr 28 15:47:21 2015 +0530
+++ b/source/common/x86/ipfilter8.asm	Tue Apr 28 16:10:56 2015 +0530
@@ -15303,6 +15303,38 @@
     FILTER_VER_CHROMA_S_AVX2_4x16 sp
     FILTER_VER_CHROMA_S_AVX2_4x16 ss
 
+%macro FILTER_VER_CHROMA_S_AVX2_4x32 1
+INIT_YMM avx2
+cglobal interp_4tap_vert_%1_4x32, 4, 7, 8
+    mov             r4d, r4m
+    shl             r4d, 6
+    add             r1d, r1d
+    sub             r0, r1
+
+%ifdef PIC
+    lea             r5, [pw_ChromaCoeffV]
+    add             r5, r4
+%else
+    lea             r5, [pw_ChromaCoeffV + r4]
+%endif
+
+    lea             r4, [r1 * 3]
+%ifidn %1,sp
+    mova            m7, [pd_526336]
+%else
+    add             r3d, r3d
+%endif
+    lea             r6, [r3 * 3]
+%rep 2
+    PROCESS_CHROMA_AVX2_W4_16R %1
+    lea             r2, [r2 + r3 * 4]
+%endrep
+    RET
+%endmacro
+
+    FILTER_VER_CHROMA_S_AVX2_4x32 sp
+    FILTER_VER_CHROMA_S_AVX2_4x32 ss
+
 %macro FILTER_VER_CHROMA_S_AVX2_4x2 1
 INIT_YMM avx2
 cglobal interp_4tap_vert_%1_4x2, 4, 6, 6