diff options
| author | Ronald S. Bultje <rsbultje@gmail.com> | 2015-09-16 16:07:46 -0400 |
|---|---|---|
| committer | Ronald S. Bultje <rsbultje@gmail.com> | 2015-09-16 21:11:34 -0400 |
| commit | 344d519040f03515bb915bf418f3ab9e38ada33b (patch) | |
| tree | 5ec8ab2d2d64ff4cd16c1f55cade61ab9f0be885 /libavcodec/x86/vp9mc.asm | |
| parent | 77f359670fdbfcd7ac2a464b3bc000b1ceb54763 (diff) | |
| download | ffmpeg-344d519040f03515bb915bf418f3ab9e38ada33b.tar.gz | |
vp9: add subpel MC SIMD for 10/12bpp.
Diffstat (limited to 'libavcodec/x86/vp9mc.asm')
| -rw-r--r-- | libavcodec/x86/vp9mc.asm | 26 |
1 files changed, 18 insertions, 8 deletions
diff --git a/libavcodec/x86/vp9mc.asm b/libavcodec/x86/vp9mc.asm index bc61c12841..d8283126eb 100644 --- a/libavcodec/x86/vp9mc.asm +++ b/libavcodec/x86/vp9mc.asm @@ -45,6 +45,13 @@ times 8 dw %7 times 8 dw %8 %endmacro +%macro F8_16BPP_TAPS 8 +times 8 dw %1, %2 +times 8 dw %3, %4 +times 8 dw %5, %6 +times 8 dw %7, %8 +%endmacro + %macro FILTER 1 const filters_%1 ; smooth F8_TAPS -3, -1, 32, 64, 38, 1, -3, 0 @@ -102,12 +109,15 @@ FILTER ssse3 %define F8_TAPS F8_SSE2_TAPS ; int16_t ff_filters_sse2[3][15][8][8] FILTER sse2 +%define F8_TAPS F8_16BPP_TAPS +; int16_t ff_filters_16bpp[3][15][4][16] +FILTER 16bpp SECTION .text %macro filter_sse2_h_fn 1 %assign %%px mmsize/2 -cglobal vp9_%1_8tap_1d_h_ %+ %%px, 6, 6, 15, dst, dstride, src, sstride, h, filtery +cglobal vp9_%1_8tap_1d_h_ %+ %%px %+ _8, 6, 6, 15, dst, dstride, src, sstride, h, filtery pxor m5, m5 mova m6, [pw_64] mova m7, [filteryq+ 0] @@ -192,7 +202,7 @@ filter_sse2_h_fn avg %macro filter_h_fn 1 %assign %%px mmsize/2 -cglobal vp9_%1_8tap_1d_h_ %+ %%px, 6, 6, 11, dst, dstride, src, sstride, h, filtery +cglobal vp9_%1_8tap_1d_h_ %+ %%px %+ _8, 6, 6, 11, dst, dstride, src, sstride, h, filtery mova m6, [pw_256] mova m7, [filteryq+ 0] %if ARCH_X86_64 && mmsize > 8 @@ -253,7 +263,7 @@ filter_h_fn avg %if ARCH_X86_64 %macro filter_hx2_fn 1 %assign %%px mmsize -cglobal vp9_%1_8tap_1d_h_ %+ %%px, 6, 6, 14, dst, dstride, src, sstride, h, filtery +cglobal vp9_%1_8tap_1d_h_ %+ %%px %+ _8, 6, 6, 14, dst, dstride, src, sstride, h, filtery mova m13, [pw_256] mova m8, [filteryq+ 0] mova m9, [filteryq+32] @@ -315,9 +325,9 @@ filter_hx2_fn avg %macro filter_sse2_v_fn 1 %assign %%px mmsize/2 %if ARCH_X86_64 -cglobal vp9_%1_8tap_1d_v_ %+ %%px, 6, 8, 15, dst, dstride, src, sstride, h, filtery, src4, sstride3 +cglobal vp9_%1_8tap_1d_v_ %+ %%px %+ _8, 6, 8, 15, dst, dstride, src, sstride, h, filtery, src4, sstride3 %else -cglobal vp9_%1_8tap_1d_v_ %+ %%px, 4, 7, 15, dst, dstride, src, sstride, filtery, src4, sstride3 +cglobal vp9_%1_8tap_1d_v_ %+ %%px %+ _8, 4, 7, 15, dst, dstride, src, sstride, filtery, src4, sstride3 mov filteryq, r5mp %define hd r4mp %endif @@ -413,9 +423,9 @@ filter_sse2_v_fn avg %macro filter_v_fn 1 %assign %%px mmsize/2 %if ARCH_X86_64 -cglobal vp9_%1_8tap_1d_v_ %+ %%px, 6, 8, 11, dst, dstride, src, sstride, h, filtery, src4, sstride3 +cglobal vp9_%1_8tap_1d_v_ %+ %%px %+ _8, 6, 8, 11, dst, dstride, src, sstride, h, filtery, src4, sstride3 %else -cglobal vp9_%1_8tap_1d_v_ %+ %%px, 4, 7, 11, dst, dstride, src, sstride, filtery, src4, sstride3 +cglobal vp9_%1_8tap_1d_v_ %+ %%px %+ _8, 4, 7, 11, dst, dstride, src, sstride, filtery, src4, sstride3 mov filteryq, r5mp %define hd r4mp %endif @@ -487,7 +497,7 @@ filter_v_fn avg %macro filter_vx2_fn 1 %assign %%px mmsize -cglobal vp9_%1_8tap_1d_v_ %+ %%px, 6, 8, 14, dst, dstride, src, sstride, h, filtery, src4, sstride3 +cglobal vp9_%1_8tap_1d_v_ %+ %%px %+ _8, 6, 8, 14, dst, dstride, src, sstride, h, filtery, src4, sstride3 mova m13, [pw_256] lea sstride3q, [sstrideq*3] lea src4q, [srcq+sstrideq] |