[x265] [PATCH] asm: chroma_vpp[32x32] for colorspace i420 in avx2: improve 3881c->3648c
Divya Manivannan
divya at multicorewareinc.com
Tue Dec 9 10:11:09 CET 2014
# HG changeset patch
# User Divya Manivannan <divya at multicorewareinc.com>
# Date 1418116177 -19800
# Tue Dec 09 14:39:37 2014 +0530
# Node ID 2e6c4518f7083d79202a28f739650278e5c0d88d
# Parent 88498ec9b10ba25a01c983a3f67c17bf470349fa
asm: chroma_vpp[32x32] for colorspace i420 in avx2: improve 3881c->3648c
diff -r 88498ec9b10b -r 2e6c4518f708 source/common/x86/asm-primitives.cpp
--- a/source/common/x86/asm-primitives.cpp Tue Dec 09 10:03:35 2014 +0530
+++ b/source/common/x86/asm-primitives.cpp Tue Dec 09 14:39:37 2014 +0530
@@ -1886,6 +1886,7 @@
p.chroma[X265_CSP_I420].filter_vpp[CHROMA_8x8] = x265_interp_4tap_vert_pp_8x8_avx2;
#if X86_64
p.chroma[X265_CSP_I420].filter_vpp[CHROMA_16x16] = x265_interp_4tap_vert_pp_16x16_avx2;
+ p.chroma[X265_CSP_I420].filter_vpp[CHROMA_32x32] = x265_interp_4tap_vert_pp_32x32_avx2;
#endif
}
#endif // if HIGH_BIT_DEPTH
diff -r 88498ec9b10b -r 2e6c4518f708 source/common/x86/ipfilter8.asm
--- a/source/common/x86/ipfilter8.asm Tue Dec 09 10:03:35 2014 +0530
+++ b/source/common/x86/ipfilter8.asm Tue Dec 09 14:39:37 2014 +0530
@@ -4043,7 +4043,406 @@
FILTER_V4_W32 32, 48
FILTER_V4_W32 32, 64
-
+%macro PROCESS_CHROMA_AVX2_W16_32R 0
+ movu xm0, [r0] ; m0 = row 0
+ movu xm1, [r0 + r1] ; m1 = row 1
+ punpckhbw xm2, xm0, xm1
+ punpcklbw xm0, xm1
+ vinserti128 m0, m0, xm2, 1
+ pmaddubsw m0, m12
+ movu xm2, [r0 + r1 * 2] ; m2 = row 2
+ punpckhbw xm3, xm1, xm2
+ punpcklbw xm1, xm2
+ vinserti128 m1, m1, xm3, 1
+ pmaddubsw m1, m12
+ movu xm3, [r0 + r4] ; m3 = row 3
+ punpckhbw xm4, xm2, xm3
+ punpcklbw xm2, xm3
+ vinserti128 m2, m2, xm4, 1
+ pmaddubsw m4, m2, m13
+ paddw m0, m4
+ pmaddubsw m2, m12
+ lea r7, [r0 + r1 * 4]
+ movu xm4, [r7] ; m4 = row 4
+ punpckhbw xm5, xm3, xm4
+ punpcklbw xm3, xm4
+ vinserti128 m3, m3, xm5, 1
+ pmaddubsw m5, m3, m13
+ paddw m1, m5
+ pmaddubsw m3, m12
+ movu xm5, [r7 + r1] ; m5 = row 5
+ punpckhbw xm6, xm4, xm5
+ punpcklbw xm4, xm5
+ vinserti128 m4, m4, xm6, 1
+ pmaddubsw m6, m4, m13
+ paddw m2, m6
+ pmaddubsw m4, m12
+ movu xm6, [r7 + r1 * 2] ; m6 = row 6
+ punpckhbw xm7, xm5, xm6
+ punpcklbw xm5, xm6
+ vinserti128 m5, m5, xm7, 1
+ pmaddubsw m7, m5, m13
+ paddw m3, m7
+ pmaddubsw m5, m12
+ movu xm7, [r7 + r4] ; m7 = row 7
+ punpckhbw xm8, xm6, xm7
+ punpcklbw xm6, xm7
+ vinserti128 m6, m6, xm8, 1
+ pmaddubsw m8, m6, m13
+ paddw m4, m8
+ pmaddubsw m6, m12
+ lea r7, [r7 + r1 * 4]
+ movu xm8, [r7] ; m8 = row 8
+ punpckhbw xm9, xm7, xm8
+ punpcklbw xm7, xm8
+ vinserti128 m7, m7, xm9, 1
+ pmaddubsw m9, m7, m13
+ paddw m5, m9
+ pmaddubsw m7, m12
+ movu xm9, [r7 + r1] ; m9 = row 9
+ punpckhbw xm10, xm8, xm9
+ punpcklbw xm8, xm9
+ vinserti128 m8, m8, xm10, 1
+ pmaddubsw m10, m8, m13
+ paddw m6, m10
+ pmaddubsw m8, m12
+ movu xm10, [r7 + r1 * 2] ; m10 = row 10
+ punpckhbw xm11, xm9, xm10
+ punpcklbw xm9, xm10
+ vinserti128 m9, m9, xm11, 1
+ pmaddubsw m11, m9, m13
+ paddw m7, m11
+ pmaddubsw m9, m12
+
+ pmulhrsw m0, m14 ; m0 = word: row 0
+ pmulhrsw m1, m14 ; m1 = word: row 1
+ pmulhrsw m2, m14 ; m2 = word: row 2
+ pmulhrsw m3, m14 ; m3 = word: row 3
+ pmulhrsw m4, m14 ; m4 = word: row 4
+ pmulhrsw m5, m14 ; m5 = word: row 5
+ pmulhrsw m6, m14 ; m6 = word: row 6
+ pmulhrsw m7, m14 ; m7 = word: row 7
+ packuswb m0, m1
+ packuswb m2, m3
+ packuswb m4, m5
+ packuswb m6, m7
+ vpermq m0, m0, 11011000b
+ vpermq m2, m2, 11011000b
+ vpermq m4, m4, 11011000b
+ vpermq m6, m6, 11011000b
+ vextracti128 xm1, m0, 1
+ vextracti128 xm3, m2, 1
+ vextracti128 xm5, m4, 1
+ vextracti128 xm7, m6, 1
+ movu [r2], xm0
+ movu [r2 + r3], xm1
+ movu [r2 + r3 * 2], xm2
+ movu [r2 + r5], xm3
+ lea r8, [r2 + r3 * 4]
+ movu [r8], xm4
+ movu [r8 + r3], xm5
+ movu [r8 + r3 * 2], xm6
+ movu [r8 + r5], xm7
+ lea r8, [r8 + r3 * 4]
+
+ movu xm11, [r7 + r4] ; m11 = row 11
+ punpckhbw xm6, xm10, xm11
+ punpcklbw xm10, xm11
+ vinserti128 m10, m10, xm6, 1
+ pmaddubsw m6, m10, m13
+ paddw m8, m6
+ pmaddubsw m10, m12
+ lea r7, [r7 + r1 * 4]
+ movu xm6, [r7] ; m6 = row 12
+ punpckhbw xm7, xm11, xm6
+ punpcklbw xm11, xm6
+ vinserti128 m11, m11, xm7, 1
+ pmaddubsw m7, m11, m13
+ paddw m9, m7
+ pmaddubsw m11, m12
+
+ movu xm7, [r7 + r1] ; m7 = row 13
+ punpckhbw xm0, xm6, xm7
+ punpcklbw xm6, xm7
+ vinserti128 m6, m6, xm0, 1
+ pmaddubsw m0, m6, m13
+ paddw m10, m0
+ pmaddubsw m6, m12
+ movu xm0, [r7 + r1 * 2] ; m0 = row 14
+ punpckhbw xm1, xm7, xm0
+ punpcklbw xm7, xm0
+ vinserti128 m7, m7, xm1, 1
+ pmaddubsw m1, m7, m13
+ paddw m11, m1
+ pmaddubsw m7, m12
+ movu xm1, [r7 + r4] ; m1 = row 15
+ punpckhbw xm2, xm0, xm1
+ punpcklbw xm0, xm1
+ vinserti128 m0, m0, xm2, 1
+ pmaddubsw m2, m0, m13
+ paddw m6, m2
+ pmaddubsw m0, m12
+ lea r7, [r7 + r1 * 4]
+ movu xm2, [r7] ; m2 = row 16
+ punpckhbw xm3, xm1, xm2
+ punpcklbw xm1, xm2
+ vinserti128 m1, m1, xm3, 1
+ pmaddubsw m3, m1, m13
+ paddw m7, m3
+ pmaddubsw m1, m12
+ movu xm3, [r7 + r1] ; m3 = row 17
+ punpckhbw xm4, xm2, xm3
+ punpcklbw xm2, xm3
+ vinserti128 m2, m2, xm4, 1
+ pmaddubsw m5, m2, m13
+ paddw m0, m5
+ pmaddubsw m2, m12
+ movu xm4, [r7 + r1 * 2] ; m4 = row 18
+ punpckhbw xm5, xm3, xm4
+ punpcklbw xm3, xm4
+ vinserti128 m3, m3, xm5, 1
+ pmaddubsw m5, m3, m13
+ paddw m1, m5
+ pmaddubsw m3, m12
+
+ pmulhrsw m8, m14 ; m8 = word: row 8
+ pmulhrsw m9, m14 ; m9 = word: row 9
+ pmulhrsw m10, m14 ; m10 = word: row 10
+ pmulhrsw m11, m14 ; m11 = word: row 11
+ pmulhrsw m6, m14 ; m6 = word: row 12
+ pmulhrsw m7, m14 ; m7 = word: row 13
+ pmulhrsw m0, m14 ; m0 = word: row 14
+ pmulhrsw m1, m14 ; m1 = word: row 15
+ packuswb m8, m9
+ packuswb m10, m11
+ packuswb m6, m7
+ packuswb m0, m1
+ vpermq m8, m8, 11011000b
+ vpermq m10, m10, 11011000b
+ vpermq m6, m6, 11011000b
+ vpermq m0, m0, 11011000b
+ vextracti128 xm9, m8, 1
+ vextracti128 xm11, m10, 1
+ vextracti128 xm7, m6, 1
+ vextracti128 xm1, m0, 1
+ movu [r8], xm8
+ movu [r8 + r3], xm9
+ movu [r8 + r3 * 2], xm10
+ movu [r8 + r5], xm11
+ lea r8, [r8 + r3 * 4]
+ movu [r8], xm6
+ movu [r8 + r3], xm7
+ movu [r8 + r3 * 2], xm0
+ movu [r8 + r5], xm1
+ lea r8, [r8 + r3 * 4]
+
+ movu xm5, [r7 + r4] ; m5 = row 19
+ punpckhbw xm6, xm4, xm5
+ punpcklbw xm4, xm5
+ vinserti128 m4, m4, xm6, 1
+ pmaddubsw m6, m4, m13
+ paddw m2, m6
+ pmaddubsw m4, m12
+ lea r7, [r7 + r1 * 4]
+ movu xm6, [r7] ; m6 = row 20
+ punpckhbw xm7, xm5, xm6
+ punpcklbw xm5, xm6
+ vinserti128 m5, m5, xm7, 1
+ pmaddubsw m7, m5, m13
+ paddw m3, m7
+ pmaddubsw m5, m12
+ movu xm7, [r7 + r1] ; m7 = row 21
+ punpckhbw xm8, xm6, xm7
+ punpcklbw xm6, xm7
+ vinserti128 m6, m6, xm8, 1
+ pmaddubsw m8, m6, m13
+ paddw m4, m8
+ pmaddubsw m6, m12
+ movu xm8, [r7 + r1 * 2] ; m8 = row 22
+ punpckhbw xm9, xm7, xm8
+ punpcklbw xm7, xm8
+ vinserti128 m7, m7, xm9, 1
+ pmaddubsw m9, m7, m13
+ paddw m5, m9
+ pmaddubsw m7, m12
+ movu xm9, [r7 + r4] ; m9 = row 23
+ punpckhbw xm10, xm8, xm9
+ punpcklbw xm8, xm9
+ vinserti128 m8, m8, xm10, 1
+ pmaddubsw m10, m8, m13
+ paddw m6, m10
+ pmaddubsw m8, m12
+ lea r7, [r7 + r1 * 4]
+ movu xm10, [r7] ; m10 = row 24
+ punpckhbw xm11, xm9, xm10
+ punpcklbw xm9, xm10
+ vinserti128 m9, m9, xm11, 1
+ pmaddubsw m11, m9, m13
+ paddw m7, m11
+ pmaddubsw m9, m12
+ movu xm11, [r7 + r1] ; m11 = row 25
+ punpckhbw xm0, xm10, xm11
+ punpcklbw xm10, xm11
+ vinserti128 m10, m10, xm0, 1
+ pmaddubsw m0, m10, m13
+ paddw m8, m0
+ pmaddubsw m10, m12
+ movu xm0, [r7 + r1 * 2] ; m0 = row 26
+ punpckhbw xm1, xm11, xm0
+ punpcklbw xm11, xm0
+ vinserti128 m11, m11, xm1, 1
+ pmaddubsw m1, m11, m13
+ paddw m9, m1
+ pmaddubsw m11, m12
+
+ pmulhrsw m2, m14 ; m2 = word: row 16
+ pmulhrsw m3, m14 ; m3 = word: row 17
+ pmulhrsw m4, m14 ; m4 = word: row 18
+ pmulhrsw m5, m14 ; m5 = word: row 19
+ pmulhrsw m6, m14 ; m6 = word: row 20
+ pmulhrsw m7, m14 ; m7 = word: row 21
+ pmulhrsw m8, m14 ; m8 = word: row 22
+ pmulhrsw m9, m14 ; m9 = word: row 23
+ packuswb m2, m3
+ packuswb m4, m5
+ packuswb m6, m7
+ packuswb m8, m9
+ vpermq m2, m2, 11011000b
+ vpermq m4, m4, 11011000b
+ vpermq m6, m6, 11011000b
+ vpermq m8, m8, 11011000b
+ vextracti128 xm3, m2, 1
+ vextracti128 xm5, m4, 1
+ vextracti128 xm7, m6, 1
+ vextracti128 xm9, m8, 1
+ movu [r8], xm2
+ movu [r8 + r3], xm3
+ movu [r8 + r3 * 2], xm4
+ movu [r8 + r5], xm5
+ lea r8, [r8 + r3 * 4]
+ movu [r8], xm6
+ movu [r8 + r3], xm7
+ movu [r8 + r3 * 2], xm8
+ movu [r8 + r5], xm9
+ lea r8, [r8 + r3 * 4]
+
+ movu xm1, [r7 + r4] ; m1 = row 27
+ punpckhbw xm2, xm0, xm1
+ punpcklbw xm0, xm1
+ vinserti128 m0, m0, xm2, 1
+ pmaddubsw m2, m0, m13
+ paddw m10, m2
+ pmaddubsw m0, m12
+ lea r7, [r7 + r1 * 4]
+ movu xm2, [r7] ; m2 = row 28
+ punpckhbw xm3, xm1, xm2
+ punpcklbw xm1, xm2
+ vinserti128 m1, m1, xm3, 1
+ pmaddubsw m3, m1, m13
+ paddw m11, m3
+ pmaddubsw m1, m12
+ movu xm3, [r7 + r1] ; m3 = row 29
+ punpckhbw xm4, xm2, xm3
+ punpcklbw xm2, xm3
+ vinserti128 m2, m2, xm4, 1
+ pmaddubsw m4, m2, m13
+ paddw m0, m4
+ pmaddubsw m2, m12
+ movu xm4, [r7 + r1 * 2] ; m4 = row 30
+ punpckhbw xm5, xm3, xm4
+ punpcklbw xm3, xm4
+ vinserti128 m3, m3, xm5, 1
+ pmaddubsw m5, m3, m13
+ paddw m1, m5
+ pmaddubsw m3, m12
+ movu xm5, [r7 + r4] ; m5 = row 31
+ punpckhbw xm6, xm4, xm5
+ punpcklbw xm4, xm5
+ vinserti128 m4, m4, xm6, 1
+ pmaddubsw m6, m4, m13
+ paddw m2, m6
+ pmaddubsw m4, m12
+ lea r7, [r7 + r1 * 4]
+ movu xm6, [r7] ; m6 = row 32
+ punpckhbw xm7, xm5, xm6
+ punpcklbw xm5, xm6
+ vinserti128 m5, m5, xm7, 1
+ pmaddubsw m7, m5, m13
+ paddw m3, m7
+ pmaddubsw m5, m12
+ movu xm7, [r7 + r1] ; m7 = row 33
+ punpckhbw xm8, xm6, xm7
+ punpcklbw xm6, xm7
+ vinserti128 m6, m6, xm8, 1
+ pmaddubsw m8, m6, m13
+ paddw m4, m8
+ movu xm8, [r7 + r1 * 2] ; m8 = row 34
+ punpckhbw xm6, xm7, xm8
+ punpcklbw xm7, xm8
+ vinserti128 m7, m7, xm6, 1
+ pmaddubsw m6, m7, m13
+ paddw m5, m6
+
+ pmulhrsw m10, m14 ; m10 = word: row 24
+ pmulhrsw m11, m14 ; m11 = word: row 25
+ pmulhrsw m0, m14 ; m0 = word: row 26
+ pmulhrsw m1, m14 ; m1 = word: row 27
+ pmulhrsw m2, m14 ; m2 = word: row 28
+ pmulhrsw m3, m14 ; m3 = word: row 29
+ pmulhrsw m4, m14 ; m4 = word: row 30
+ pmulhrsw m5, m14 ; m5 = word: row 31
+ packuswb m10, m11
+ packuswb m0, m1
+ packuswb m2, m3
+ packuswb m4, m5
+ vpermq m10, m10, 11011000b
+ vpermq m0, m0, 11011000b
+ vpermq m2, m2, 11011000b
+ vpermq m4, m4, 11011000b
+ vextracti128 xm11, m10, 1
+ vextracti128 xm1, m0, 1
+ vextracti128 xm3, m2, 1
+ vextracti128 xm5, m4, 1
+ movu [r8], xm10
+ movu [r8 + r3], xm11
+ movu [r8 + r3 * 2], xm0
+ movu [r8 + r5], xm1
+ lea r8, [r8 + r3 * 4]
+ movu [r8], xm2
+ movu [r8 + r3], xm3
+ movu [r8 + r3 * 2], xm4
+ movu [r8 + r5], xm5
+%endmacro
+
+INIT_YMM avx2
+%if ARCH_X86_64 == 1
+cglobal interp_4tap_vert_pp_32x32, 4, 9, 15
+ mov r4d, r4m
+ shl r4d, 6
+
+%ifdef PIC
+ lea r5, [tab_ChromaCoeffVer_32]
+ add r5, r4
+%else
+ lea r5, [tab_ChromaCoeffVer_32 + r4]
+%endif
+
+ mova m12, [r5]
+ mova m13, [r5 + mmsize]
+ lea r4, [r1 * 3]
+ sub r0, r1
+ lea r5, [r3 * 3]
+ mova m14, [pw_512]
+ mov r6d, 2
+.loopW:
+ PROCESS_CHROMA_AVX2_W16_32R
+ add r2, 16
+ add r0, 16
+ dec r6d
+ jnz .loopW
+ RET
+%endif
+
;-----------------------------------------------------------------------------
; void interp_4tap_vert_pp_%1x%2(pixel *src, intptr_t srcStride, pixel *dst, intptr_t dstStride, int coeffIdx)
;-----------------------------------------------------------------------------
More information about the x265-devel
mailing list