[FFmpeg-devel] [PATCH] avcodec/vp9: add 64-bit ipred_dr_32x32_16 avx2 implementation
Ilia Valiakhmetov
zakne0ne at gmail.com
Sun Jun 25 17:42:08 EEST 2017
vp9_diag_downright_32x32_12bpp_c: 429.7
vp9_diag_downright_32x32_12bpp_sse2: 158.9
vp9_diag_downright_32x32_12bpp_ssse3: 144.6
vp9_diag_downright_32x32_12bpp_avx: 141.0
vp9_diag_downright_32x32_12bpp_avx2: 73.8
Almost 50% faster than avx implementation
---
libavcodec/x86/vp9dsp_init_16bpp.c | 6 +-
libavcodec/x86/vp9intrapred_16bpp.asm | 103 +++++++++++++++++++++++++++++++++-
2 files changed, 106 insertions(+), 3 deletions(-)
diff --git a/libavcodec/x86/vp9dsp_init_16bpp.c b/libavcodec/x86/vp9dsp_init_16bpp.c
index 8d1aa13..54216f0 100644
--- a/libavcodec/x86/vp9dsp_init_16bpp.c
+++ b/libavcodec/x86/vp9dsp_init_16bpp.c
@@ -52,8 +52,9 @@ decl_ipred_fns(dc, 16, mmxext, sse2);
decl_ipred_fns(dc_top, 16, mmxext, sse2);
decl_ipred_fns(dc_left, 16, mmxext, sse2);
decl_ipred_fn(dl, 16, 16, avx2);
-decl_ipred_fn(dr, 16, 16, avx2);
decl_ipred_fn(dl, 32, 16, avx2);
+decl_ipred_fn(dr, 16, 16, avx2);
+decl_ipred_fn(dr, 32, 16, avx2);
#define decl_ipred_dir_funcs(type) \
decl_ipred_fns(type, 16, sse2, sse2); \
@@ -137,8 +138,9 @@ av_cold void ff_vp9dsp_init_16bpp_x86(VP9DSPContext *dsp)
init_fpel_func(1, 1, 64, avg, _16, avx2);
init_fpel_func(0, 1, 128, avg, _16, avx2);
init_ipred_func(dl, DIAG_DOWN_LEFT, 16, 16, avx2);
- init_ipred_func(dr, DIAG_DOWN_RIGHT, 16, 16, avx2);
init_ipred_func(dl, DIAG_DOWN_LEFT, 32, 16, avx2);
+ init_ipred_func(dr, DIAG_DOWN_RIGHT, 16, 16, avx2);
+ init_ipred_func(dr, DIAG_DOWN_RIGHT, 32, 16, avx2);
}
#endif /* HAVE_YASM */
diff --git a/libavcodec/x86/vp9intrapred_16bpp.asm b/libavcodec/x86/vp9intrapred_16bpp.asm
index 6d4400b..32b6982 100644
--- a/libavcodec/x86/vp9intrapred_16bpp.asm
+++ b/libavcodec/x86/vp9intrapred_16bpp.asm
@@ -1221,8 +1221,109 @@ cglobal vp9_ipred_dr_16x16_16, 4, 5, 6, dst, stride, l, a
mova [dstq+strideq*0], m4 ; 0
mova [dst3q+strideq*4], m5 ; 7
RET
-%endif
+%if ARCH_X86_64
+cglobal vp9_ipred_dr_32x32_16, 4, 7, 10, dst, stride, l, a
+ mova m0, [lq+mmsize*0+0] ; l[0-15]
+ mova m1, [lq+mmsize*1+0] ; l[16-31]
+ movu m2, [aq+mmsize*0-2] ; *abcdefghijklmno
+ mova m3, [aq+mmsize*0+0] ; abcdefghijklmnop
+ mova m4, [aq+mmsize*1+0] ; qrstuvwxyz012345
+ vperm2i128 m5, m0, m1, q0201 ; lmnopqrstuvwxyz0
+ vpalignr m6, m5, m0, 2 ; mnopqrstuvwxyz01
+ vpalignr m7, m5, m0, 4 ; nopqrstuvwxyz012
+ LOWPASS 0, 6, 7 ; L[0-15]
+ vperm2i128 m7, m1, m2, q0201 ; stuvwxyz*abcdefg
+ vpalignr m5, m7, m1, 2 ; lmnopqrstuvwxyz*
+ vpalignr m6, m7, m1, 4 ; mnopqrstuvwxyz*a
+ LOWPASS 1, 5, 6 ; L[16-31]#
+ vperm2i128 m5, m3, m4, q0201 ; ijklmnopqrstuvwx
+ vpalignr m6, m5, m3, 2 ; bcdefghijklmnopq
+ LOWPASS 2, 3, 6 ; A[0-15]
+ movu m3, [aq+mmsize*1-2] ; pqrstuvwxyz01234
+ vperm2i128 m6, m4, m4, q2001 ; yz012345........
+ vpalignr m7, m6, m4, 2 ; rstuvwxyz012345.
+ LOWPASS 3, 4, 7 ; A[16-31].
+ vperm2i128 m4, m1, m2, q0201 ; TUVWXYZ#ABCDEFGH
+ vperm2i128 m5, m0, m1, q0201 ; L[7-15]L[16-23]
+ vperm2i128 m8, m2, m3, q0201 ; IJKLMNOPQRSTUVWX
+ DEFINE_ARGS dst8, stride, stride3, stride7, stride5, dst24, cnt
+ lea stride3q, [strideq*3]
+ lea stride5q, [stride3q+strideq*2]
+ lea stride7q, [strideq*4+stride3q]
+ lea dst24q, [dst8q+stride3q*8]
+ lea dst8q, [dst8q+strideq*8]
+ mov cntd, 2
+
+.loop:
+ mova [dst24q+stride7q+0 ], m0 ; 31 23 15 7
+ mova [dst24q+stride7q+32], m1
+ mova [dst8q+stride7q+0], m1
+ mova [dst8q+stride7q+32], m2
+ vpalignr m6, m4, m1, 2
+ vpalignr m7, m5, m0, 2
+ vpalignr m9, m8, m2, 2
+ mova [dst24q+stride3q*2+0], m7 ; 30 22 14 6
+ mova [dst24q+stride3q*2+32], m6
+ mova [dst8q+stride3q*2+0], m6
+ mova [dst8q+stride3q*2+32], m9
+ vpalignr m6, m4, m1, 4
+ vpalignr m7, m5, m0, 4
+ vpalignr m9, m8, m2, 4
+ mova [dst24q+stride5q+0], m7 ; 29 21 13 5
+ mova [dst24q+stride5q+32], m6
+ mova [dst8q+stride5q+0], m6
+ mova [dst8q+stride5q+32], m9
+ vpalignr m6, m4, m1, 6
+ vpalignr m7, m5, m0, 6
+ vpalignr m9, m8, m2, 6
+ mova [dst24q+strideq*4+0 ], m7 ; 28 20 12 4
+ mova [dst24q+strideq*4+32], m6
+ mova [dst8q+strideq*4+0], m6
+ mova [dst8q+strideq*4+32], m9
+ vpalignr m6, m4, m1, 8
+ vpalignr m7, m5, m0, 8
+ vpalignr m9, m8, m2, 8
+ mova [dst24q+stride3q+0 ], m7 ; 27 19 11 3
+ mova [dst24q+stride3q+32], m6
+ mova [dst8q+stride3q+0], m6
+ mova [dst8q+stride3q+32], m9
+ vpalignr m6, m4, m1, 10
+ vpalignr m7, m5, m0, 10
+ vpalignr m9, m8, m2, 10
+ mova [dst24q+strideq*2+0 ], m7 ; 26 18 10 2
+ mova [dst24q+strideq*2+32], m6
+ mova [dst8q+strideq*2+0], m6
+ mova [dst8q+strideq*2+32], m9
+ vpalignr m6, m4, m1, 12
+ vpalignr m7, m5, m0, 12
+ vpalignr m9, m8, m2, 12
+ mova [dst24q+strideq+0 ], m7 ; 25 17 9 1
+ mova [dst24q+strideq+32], m6
+ mova [dst8q+strideq+0], m6
+ mova [dst8q+strideq+32], m9
+ vpalignr m6, m4, m1, 14
+ vpalignr m7, m5, m0, 14
+ vpalignr m9, m8, m2, 14
+ mova [dst24q+strideq*0+0 ], m7 ; 24 16 8 0
+ mova [dst24q+strideq*0+32], m6
+ mova [dst8q+strideq*0+0], m6
+ mova [dst8q+strideq*0+32], m9
+ mova m0, m5
+ mova m5, m1
+ mova m1, m4
+ mova m4, m2
+ mova m2, m8
+ mova m8, m3
+ sub dst24q, stride7q
+ sub dst24q, strideq
+ sub dst8q, stride7q
+ sub dst8q, strideq
+ dec cntd
+ jg .loop
+ RET
+%endif
+%endif
%macro VL_FUNCS 1 ; stack_mem_for_32x32_32bit_function
cglobal vp9_ipred_vl_4x4_16, 2, 4, 3, dst, stride, l, a
--
2.8.3
More information about the ffmpeg-devel
mailing list