Chiebot-Mirror
/
FFmpeg
mirror of https://github.com/FFmpeg/FFmpeg.git
8
0
Fork 0
Code Issues Projects Releases Wiki Activity
You can not select more than 25 topics Topics must start with a letter or number, can include dashes ('-') and can be up to 35 characters long.
59690 Commits
37 Branches
400 Tags
465 MiB
Tag: Branch: Tree: 9d098ea32f
Branches Tags
${ item.name }
Create tag ${ searchTerm }
Create branch ${ searchTerm }
from '9d098ea32f'
${ noResults }
FFmpeg/libavcodec/x86/vp9itxfm.asm

1221 lines
45 KiB
Raw Normal View History Unescape

VP9 MC (ssse3) optimizations. Decoding time of ped1080p.webm goes from 20.7sec to 11.3sec.
11 years ago
;******************************************************************************
vp9: split x86 assembly in two files. (And in future, loopfilter or intra pred could be put in their own respective files also.)
11 years ago
;* VP9 IDCT SIMD optimizations
VP9 MC (ssse3) optimizations. Decoding time of ped1080p.webm goes from 20.7sec to 11.3sec.
11 years ago
;*
vp9: split x86 assembly in two files. (And in future, loopfilter or intra pred could be put in their own respective files also.)
11 years ago
;* Copyright (C) 2013 Clément Bœsch <u pkh me>
vp9/x86: idct_add_16x16_ssse3. Currently only dc-only and full 16x16. Other subforms will follow in the near future. Total decoding time of ped1080p.webm goes from 9.7 to 9.3 seconds. DC-only goes from 957 -> 131 cycles, and the full IDCT goes from ~4050 to ~745 cycles.
11 years ago
;* Copyright (C) 2013 Ronald S. Bultje <rsbultje gmail com>
VP9 MC (ssse3) optimizations. Decoding time of ped1080p.webm goes from 20.7sec to 11.3sec.
11 years ago
;*
;* This file is part of FFmpeg.
;*
;* FFmpeg is free software; you can redistribute it and/or
;* modify it under the terms of the GNU Lesser General Public
;* License as published by the Free Software Foundation; either
;* version 2.1 of the License, or (at your option) any later version.
;*
;* FFmpeg is distributed in the hope that it will be useful,
;* but WITHOUT ANY WARRANTY; without even the implied warranty of
;* MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the GNU
;* Lesser General Public License for more details.
;*
;* You should have received a copy of the GNU Lesser General Public
;* License along with FFmpeg; if not, write to the Free Software
;* Foundation, Inc., 51 Franklin Street, Fifth Floor, Boston, MA 02110-1301 USA
;******************************************************************************
%include "libavutil/x86/x86util.asm"
SECTION_RODATA
avcodec/vp9: add ff_vp9_idct_idct_{4x4,8x8}_ssse3(). 1789 decicycles in idct_idct_4x4_add_c, 262136 runs, 8 skips 1839 decicycles in idct_idct_4x4_add_c, 524270 runs, 18 skips 1864 decicycles in idct_idct_4x4_add_c, 1048548 runs, 28 skips 529 decicycles in ff_vp9_idct_idct_4x4_add_ssse3, 262138 runs, 6 skips 516 decicycles in ff_vp9_idct_idct_4x4_add_ssse3, 524282 runs, 6 skips 474 decicycles in ff_vp9_idct_idct_4x4_add_ssse3, 1048565 runs, 11 skips (~3.9x faster) 7726 decicycles in idct_idct_8x8_add_c, 1048433 runs, 143 skips 7732 decicycles in idct_idct_8x8_add_c, 2096882 runs, 270 skips 7731 decicycles in idct_idct_8x8_add_c, 4193772 runs, 532 skips 1145 decicycles in ff_vp9_idct_idct_8x8_add_ssse3, 1048549 runs, 27 skips 1137 decicycles in ff_vp9_idct_idct_8x8_add_ssse3, 2097097 runs, 55 skips 1086 decicycles in ff_vp9_idct_idct_8x8_add_ssse3, 4194188 runs, 116 skips (~7.1x faster) Overall decode time before commit: 16.48s user 0.03s system 99% cpu 16.526 total 16.54s user 0.01s system 99% cpu 16.566 total 16.46s user 0.03s system 99% cpu 16.511 total Overall decode time after commit: 16.34s user 0.02s system 99% cpu 16.378 total 16.28s user 0.02s system 99% cpu 16.315 total 16.32s user 0.03s system 99% cpu 16.366 total Tested on i7 920 with 40s 1080p footage.
11 years ago
pw_11585x2: times 8 dw 23170
vp9/x86: 16x16 sub-IDCT for top-left 8x8 subblock (eob <= 38). Sub8x8 speed (w/o dc-only case) goes from ~750 cycles (inter) or ~735 cycles (intra) to ~415 cycles (inter) or ~430 cycles (intra). Average overall 16x16 idct speed goes from ~635 cycles (inter) or ~720 cycles (intra) to ~415 cycles (inter) or ~545 (intra) - all measurements done using ped1080p.webm.
11 years ago
%macro VP9_IDCT_COEFFS 2
vp9/x86: idct_add_16x16_ssse3. Currently only dc-only and full 16x16. Other subforms will follow in the near future. Total decoding time of ped1080p.webm goes from 9.7 to 9.3 seconds. DC-only goes from 957 -> 131 cycles, and the full IDCT goes from ~4050 to ~745 cycles.
11 years ago
pw_m%1_%2: times 4 dw -%1, %2
pw_%2_%1: times 4 dw %2, %1
pw_m%2_m%1: times 4 dw -%2, -%1
avcodec/vp9: add ff_vp9_idct_idct_{4x4,8x8}_ssse3(). 1789 decicycles in idct_idct_4x4_add_c, 262136 runs, 8 skips 1839 decicycles in idct_idct_4x4_add_c, 524270 runs, 18 skips 1864 decicycles in idct_idct_4x4_add_c, 1048548 runs, 28 skips 529 decicycles in ff_vp9_idct_idct_4x4_add_ssse3, 262138 runs, 6 skips 516 decicycles in ff_vp9_idct_idct_4x4_add_ssse3, 524282 runs, 6 skips 474 decicycles in ff_vp9_idct_idct_4x4_add_ssse3, 1048565 runs, 11 skips (~3.9x faster) 7726 decicycles in idct_idct_8x8_add_c, 1048433 runs, 143 skips 7732 decicycles in idct_idct_8x8_add_c, 2096882 runs, 270 skips 7731 decicycles in idct_idct_8x8_add_c, 4193772 runs, 532 skips 1145 decicycles in ff_vp9_idct_idct_8x8_add_ssse3, 1048549 runs, 27 skips 1137 decicycles in ff_vp9_idct_idct_8x8_add_ssse3, 2097097 runs, 55 skips 1086 decicycles in ff_vp9_idct_idct_8x8_add_ssse3, 4194188 runs, 116 skips (~7.1x faster) Overall decode time before commit: 16.48s user 0.03s system 99% cpu 16.526 total 16.54s user 0.01s system 99% cpu 16.566 total 16.46s user 0.03s system 99% cpu 16.511 total Overall decode time after commit: 16.34s user 0.02s system 99% cpu 16.378 total 16.28s user 0.02s system 99% cpu 16.315 total 16.32s user 0.03s system 99% cpu 16.366 total Tested on i7 920 with 40s 1080p footage.
11 years ago
%endmacro
vp9/x86: 16x16 sub-IDCT for top-left 8x8 subblock (eob <= 38). Sub8x8 speed (w/o dc-only case) goes from ~750 cycles (inter) or ~735 cycles (intra) to ~415 cycles (inter) or ~430 cycles (intra). Average overall 16x16 idct speed goes from ~635 cycles (inter) or ~720 cycles (intra) to ~415 cycles (inter) or ~545 (intra) - all measurements done using ped1080p.webm.
11 years ago
%macro VP9_IDCT_COEFFS_ALL 2
avcodec/vp9: add ff_vp9_idct_idct_{4x4,8x8}_ssse3(). 1789 decicycles in idct_idct_4x4_add_c, 262136 runs, 8 skips 1839 decicycles in idct_idct_4x4_add_c, 524270 runs, 18 skips 1864 decicycles in idct_idct_4x4_add_c, 1048548 runs, 28 skips 529 decicycles in ff_vp9_idct_idct_4x4_add_ssse3, 262138 runs, 6 skips 516 decicycles in ff_vp9_idct_idct_4x4_add_ssse3, 524282 runs, 6 skips 474 decicycles in ff_vp9_idct_idct_4x4_add_ssse3, 1048565 runs, 11 skips (~3.9x faster) 7726 decicycles in idct_idct_8x8_add_c, 1048433 runs, 143 skips 7732 decicycles in idct_idct_8x8_add_c, 2096882 runs, 270 skips 7731 decicycles in idct_idct_8x8_add_c, 4193772 runs, 532 skips 1145 decicycles in ff_vp9_idct_idct_8x8_add_ssse3, 1048549 runs, 27 skips 1137 decicycles in ff_vp9_idct_idct_8x8_add_ssse3, 2097097 runs, 55 skips 1086 decicycles in ff_vp9_idct_idct_8x8_add_ssse3, 4194188 runs, 116 skips (~7.1x faster) Overall decode time before commit: 16.48s user 0.03s system 99% cpu 16.526 total 16.54s user 0.01s system 99% cpu 16.566 total 16.46s user 0.03s system 99% cpu 16.511 total Overall decode time after commit: 16.34s user 0.02s system 99% cpu 16.378 total 16.28s user 0.02s system 99% cpu 16.315 total 16.32s user 0.03s system 99% cpu 16.366 total Tested on i7 920 with 40s 1080p footage.
11 years ago
pw_%1x2: times 8 dw %1*2
vp9/x86: 16x16 sub-IDCT for top-left 8x8 subblock (eob <= 38). Sub8x8 speed (w/o dc-only case) goes from ~750 cycles (inter) or ~735 cycles (intra) to ~415 cycles (inter) or ~430 cycles (intra). Average overall 16x16 idct speed goes from ~635 cycles (inter) or ~720 cycles (intra) to ~415 cycles (inter) or ~545 (intra) - all measurements done using ped1080p.webm.
11 years ago
pw_m%1x2: times 8 dw -%1*2
avcodec/vp9: add ff_vp9_idct_idct_{4x4,8x8}_ssse3(). 1789 decicycles in idct_idct_4x4_add_c, 262136 runs, 8 skips 1839 decicycles in idct_idct_4x4_add_c, 524270 runs, 18 skips 1864 decicycles in idct_idct_4x4_add_c, 1048548 runs, 28 skips 529 decicycles in ff_vp9_idct_idct_4x4_add_ssse3, 262138 runs, 6 skips 516 decicycles in ff_vp9_idct_idct_4x4_add_ssse3, 524282 runs, 6 skips 474 decicycles in ff_vp9_idct_idct_4x4_add_ssse3, 1048565 runs, 11 skips (~3.9x faster) 7726 decicycles in idct_idct_8x8_add_c, 1048433 runs, 143 skips 7732 decicycles in idct_idct_8x8_add_c, 2096882 runs, 270 skips 7731 decicycles in idct_idct_8x8_add_c, 4193772 runs, 532 skips 1145 decicycles in ff_vp9_idct_idct_8x8_add_ssse3, 1048549 runs, 27 skips 1137 decicycles in ff_vp9_idct_idct_8x8_add_ssse3, 2097097 runs, 55 skips 1086 decicycles in ff_vp9_idct_idct_8x8_add_ssse3, 4194188 runs, 116 skips (~7.1x faster) Overall decode time before commit: 16.48s user 0.03s system 99% cpu 16.526 total 16.54s user 0.01s system 99% cpu 16.566 total 16.46s user 0.03s system 99% cpu 16.511 total Overall decode time after commit: 16.34s user 0.02s system 99% cpu 16.378 total 16.28s user 0.02s system 99% cpu 16.315 total 16.32s user 0.03s system 99% cpu 16.366 total Tested on i7 920 with 40s 1080p footage.
11 years ago
pw_%2x2: times 8 dw %2*2
vp9/x86: 16x16 sub-IDCT for top-left 8x8 subblock (eob <= 38). Sub8x8 speed (w/o dc-only case) goes from ~750 cycles (inter) or ~735 cycles (intra) to ~415 cycles (inter) or ~430 cycles (intra). Average overall 16x16 idct speed goes from ~635 cycles (inter) or ~720 cycles (intra) to ~415 cycles (inter) or ~545 (intra) - all measurements done using ped1080p.webm.
11 years ago
pw_m%2x2: times 8 dw -%2*2
VP9_IDCT_COEFFS %1, %2
avcodec/vp9: add ff_vp9_idct_idct_{4x4,8x8}_ssse3(). 1789 decicycles in idct_idct_4x4_add_c, 262136 runs, 8 skips 1839 decicycles in idct_idct_4x4_add_c, 524270 runs, 18 skips 1864 decicycles in idct_idct_4x4_add_c, 1048548 runs, 28 skips 529 decicycles in ff_vp9_idct_idct_4x4_add_ssse3, 262138 runs, 6 skips 516 decicycles in ff_vp9_idct_idct_4x4_add_ssse3, 524282 runs, 6 skips 474 decicycles in ff_vp9_idct_idct_4x4_add_ssse3, 1048565 runs, 11 skips (~3.9x faster) 7726 decicycles in idct_idct_8x8_add_c, 1048433 runs, 143 skips 7732 decicycles in idct_idct_8x8_add_c, 2096882 runs, 270 skips 7731 decicycles in idct_idct_8x8_add_c, 4193772 runs, 532 skips 1145 decicycles in ff_vp9_idct_idct_8x8_add_ssse3, 1048549 runs, 27 skips 1137 decicycles in ff_vp9_idct_idct_8x8_add_ssse3, 2097097 runs, 55 skips 1086 decicycles in ff_vp9_idct_idct_8x8_add_ssse3, 4194188 runs, 116 skips (~7.1x faster) Overall decode time before commit: 16.48s user 0.03s system 99% cpu 16.526 total 16.54s user 0.01s system 99% cpu 16.566 total 16.46s user 0.03s system 99% cpu 16.511 total Overall decode time after commit: 16.34s user 0.02s system 99% cpu 16.378 total 16.28s user 0.02s system 99% cpu 16.315 total 16.32s user 0.03s system 99% cpu 16.366 total Tested on i7 920 with 40s 1080p footage.
11 years ago
%endmacro
vp9/x86: 16x16 sub-IDCT for top-left 8x8 subblock (eob <= 38). Sub8x8 speed (w/o dc-only case) goes from ~750 cycles (inter) or ~735 cycles (intra) to ~415 cycles (inter) or ~430 cycles (intra). Average overall 16x16 idct speed goes from ~635 cycles (inter) or ~720 cycles (intra) to ~415 cycles (inter) or ~545 (intra) - all measurements done using ped1080p.webm.
11 years ago
VP9_IDCT_COEFFS_ALL 15137, 6270
avcodec/vp9: add ff_vp9_idct_idct_{4x4,8x8}_ssse3(). 1789 decicycles in idct_idct_4x4_add_c, 262136 runs, 8 skips 1839 decicycles in idct_idct_4x4_add_c, 524270 runs, 18 skips 1864 decicycles in idct_idct_4x4_add_c, 1048548 runs, 28 skips 529 decicycles in ff_vp9_idct_idct_4x4_add_ssse3, 262138 runs, 6 skips 516 decicycles in ff_vp9_idct_idct_4x4_add_ssse3, 524282 runs, 6 skips 474 decicycles in ff_vp9_idct_idct_4x4_add_ssse3, 1048565 runs, 11 skips (~3.9x faster) 7726 decicycles in idct_idct_8x8_add_c, 1048433 runs, 143 skips 7732 decicycles in idct_idct_8x8_add_c, 2096882 runs, 270 skips 7731 decicycles in idct_idct_8x8_add_c, 4193772 runs, 532 skips 1145 decicycles in ff_vp9_idct_idct_8x8_add_ssse3, 1048549 runs, 27 skips 1137 decicycles in ff_vp9_idct_idct_8x8_add_ssse3, 2097097 runs, 55 skips 1086 decicycles in ff_vp9_idct_idct_8x8_add_ssse3, 4194188 runs, 116 skips (~7.1x faster) Overall decode time before commit: 16.48s user 0.03s system 99% cpu 16.526 total 16.54s user 0.01s system 99% cpu 16.566 total 16.46s user 0.03s system 99% cpu 16.511 total Overall decode time after commit: 16.34s user 0.02s system 99% cpu 16.378 total 16.28s user 0.02s system 99% cpu 16.315 total 16.32s user 0.03s system 99% cpu 16.366 total Tested on i7 920 with 40s 1080p footage.
11 years ago
VP9_IDCT_COEFFS_ALL 16069, 3196
vp9: implement top/left half (4x4) sub-8x8-IDCT. For that specific case (eob>3&&eob<=12), runtime of idct8x8 goes from 668 to 477 cycles. For all idct8x8, runtime goes from 521 to 490 cycles.
11 years ago
VP9_IDCT_COEFFS_ALL 9102, 13623
vp9/x86: idct_add_16x16_ssse3. Currently only dc-only and full 16x16. Other subforms will follow in the near future. Total decoding time of ped1080p.webm goes from 9.7 to 9.3 seconds. DC-only goes from 957 -> 131 cycles, and the full IDCT goes from ~4050 to ~745 cycles.
11 years ago
VP9_IDCT_COEFFS_ALL 16305, 1606
VP9_IDCT_COEFFS_ALL 10394, 12665
VP9_IDCT_COEFFS_ALL 14449, 7723
VP9_IDCT_COEFFS_ALL 4756, 15679
vp9/x86: idct_32x32_add_ssse3. Sub-IDCTs will follow later. ped1080.webm goes from 9.295s to 8.191s (13.5% faster). The IDCT itself goes from 4372 (intra) or 4337 (inter) to 403 (intra) or 329 (inter) cycles for the DC-only form, 23755 (intra) or 23723 (inter) to 3497 (intra) or 3607 (inter) cycles for the no-DC form, which averages from 23393 (intra) or 16612 (inter) to 3449 (intra) or 2392 (inter) for all 32x32s together, i.e. about ~7x faster (all tests done on ped1080p.webm).
11 years ago
VP9_IDCT_COEFFS_ALL 16364, 804
VP9_IDCT_COEFFS_ALL 11003, 12140
VP9_IDCT_COEFFS_ALL 14811, 7005
VP9_IDCT_COEFFS_ALL 5520, 15426
VP9_IDCT_COEFFS_ALL 15893, 3981
VP9_IDCT_COEFFS_ALL 8423, 14053
VP9_IDCT_COEFFS_ALL 13160, 9760
VP9_IDCT_COEFFS_ALL 2404, 16207
avcodec/vp9: add ff_vp9_idct_idct_{4x4,8x8}_ssse3(). 1789 decicycles in idct_idct_4x4_add_c, 262136 runs, 8 skips 1839 decicycles in idct_idct_4x4_add_c, 524270 runs, 18 skips 1864 decicycles in idct_idct_4x4_add_c, 1048548 runs, 28 skips 529 decicycles in ff_vp9_idct_idct_4x4_add_ssse3, 262138 runs, 6 skips 516 decicycles in ff_vp9_idct_idct_4x4_add_ssse3, 524282 runs, 6 skips 474 decicycles in ff_vp9_idct_idct_4x4_add_ssse3, 1048565 runs, 11 skips (~3.9x faster) 7726 decicycles in idct_idct_8x8_add_c, 1048433 runs, 143 skips 7732 decicycles in idct_idct_8x8_add_c, 2096882 runs, 270 skips 7731 decicycles in idct_idct_8x8_add_c, 4193772 runs, 532 skips 1145 decicycles in ff_vp9_idct_idct_8x8_add_ssse3, 1048549 runs, 27 skips 1137 decicycles in ff_vp9_idct_idct_8x8_add_ssse3, 2097097 runs, 55 skips 1086 decicycles in ff_vp9_idct_idct_8x8_add_ssse3, 4194188 runs, 116 skips (~7.1x faster) Overall decode time before commit: 16.48s user 0.03s system 99% cpu 16.526 total 16.54s user 0.01s system 99% cpu 16.566 total 16.46s user 0.03s system 99% cpu 16.511 total Overall decode time after commit: 16.34s user 0.02s system 99% cpu 16.378 total 16.28s user 0.02s system 99% cpu 16.315 total 16.32s user 0.03s system 99% cpu 16.366 total Tested on i7 920 with 40s 1080p footage.
11 years ago
pd_8192: times 4 dd 8192
pw_2048: times 8 dw 2048
pw_1024: times 8 dw 1024
vp9/x86: idct_add_16x16_ssse3. Currently only dc-only and full 16x16. Other subforms will follow in the near future. Total decoding time of ped1080p.webm goes from 9.7 to 9.3 seconds. DC-only goes from 957 -> 131 cycles, and the full IDCT goes from ~4050 to ~745 cycles.
11 years ago
pw_512: times 8 dw 512
avcodec/vp9: add ff_vp9_idct_idct_{4x4,8x8}_ssse3(). 1789 decicycles in idct_idct_4x4_add_c, 262136 runs, 8 skips 1839 decicycles in idct_idct_4x4_add_c, 524270 runs, 18 skips 1864 decicycles in idct_idct_4x4_add_c, 1048548 runs, 28 skips 529 decicycles in ff_vp9_idct_idct_4x4_add_ssse3, 262138 runs, 6 skips 516 decicycles in ff_vp9_idct_idct_4x4_add_ssse3, 524282 runs, 6 skips 474 decicycles in ff_vp9_idct_idct_4x4_add_ssse3, 1048565 runs, 11 skips (~3.9x faster) 7726 decicycles in idct_idct_8x8_add_c, 1048433 runs, 143 skips 7732 decicycles in idct_idct_8x8_add_c, 2096882 runs, 270 skips 7731 decicycles in idct_idct_8x8_add_c, 4193772 runs, 532 skips 1145 decicycles in ff_vp9_idct_idct_8x8_add_ssse3, 1048549 runs, 27 skips 1137 decicycles in ff_vp9_idct_idct_8x8_add_ssse3, 2097097 runs, 55 skips 1086 decicycles in ff_vp9_idct_idct_8x8_add_ssse3, 4194188 runs, 116 skips (~7.1x faster) Overall decode time before commit: 16.48s user 0.03s system 99% cpu 16.526 total 16.54s user 0.01s system 99% cpu 16.566 total 16.46s user 0.03s system 99% cpu 16.511 total Overall decode time after commit: 16.34s user 0.02s system 99% cpu 16.378 total 16.28s user 0.02s system 99% cpu 16.315 total 16.32s user 0.03s system 99% cpu 16.366 total Tested on i7 920 with 40s 1080p footage.
11 years ago
VP9 MC (ssse3) optimizations. Decoding time of ped1080p.webm goes from 20.7sec to 11.3sec.
11 years ago
SECTION .text
avcodec/vp9: add ff_vp9_idct_idct_{4x4,8x8}_ssse3(). 1789 decicycles in idct_idct_4x4_add_c, 262136 runs, 8 skips 1839 decicycles in idct_idct_4x4_add_c, 524270 runs, 18 skips 1864 decicycles in idct_idct_4x4_add_c, 1048548 runs, 28 skips 529 decicycles in ff_vp9_idct_idct_4x4_add_ssse3, 262138 runs, 6 skips 516 decicycles in ff_vp9_idct_idct_4x4_add_ssse3, 524282 runs, 6 skips 474 decicycles in ff_vp9_idct_idct_4x4_add_ssse3, 1048565 runs, 11 skips (~3.9x faster) 7726 decicycles in idct_idct_8x8_add_c, 1048433 runs, 143 skips 7732 decicycles in idct_idct_8x8_add_c, 2096882 runs, 270 skips 7731 decicycles in idct_idct_8x8_add_c, 4193772 runs, 532 skips 1145 decicycles in ff_vp9_idct_idct_8x8_add_ssse3, 1048549 runs, 27 skips 1137 decicycles in ff_vp9_idct_idct_8x8_add_ssse3, 2097097 runs, 55 skips 1086 decicycles in ff_vp9_idct_idct_8x8_add_ssse3, 4194188 runs, 116 skips (~7.1x faster) Overall decode time before commit: 16.48s user 0.03s system 99% cpu 16.526 total 16.54s user 0.01s system 99% cpu 16.566 total 16.46s user 0.03s system 99% cpu 16.511 total Overall decode time after commit: 16.34s user 0.02s system 99% cpu 16.378 total 16.28s user 0.02s system 99% cpu 16.315 total 16.32s user 0.03s system 99% cpu 16.366 total Tested on i7 920 with 40s 1080p footage.
11 years ago
; (a*x + b*y + round) >> shift
%macro VP9_MULSUB_2W_2X 6 ; dst1, dst2, src (unchanged), round, coefs1, coefs2
pmaddwd m%1, m%3, %5
pmaddwd m%2, m%3, %6
vp9: minor refactorings in idct ssse3 assembly. Make register usage in macros explicit; change mulsub_2w_4x to use 2 instead of 3 temp registers.
11 years ago
paddd m%1, %4
paddd m%2, %4
psrad m%1, 14
psrad m%2, 14
avcodec/vp9: add ff_vp9_idct_idct_{4x4,8x8}_ssse3(). 1789 decicycles in idct_idct_4x4_add_c, 262136 runs, 8 skips 1839 decicycles in idct_idct_4x4_add_c, 524270 runs, 18 skips 1864 decicycles in idct_idct_4x4_add_c, 1048548 runs, 28 skips 529 decicycles in ff_vp9_idct_idct_4x4_add_ssse3, 262138 runs, 6 skips 516 decicycles in ff_vp9_idct_idct_4x4_add_ssse3, 524282 runs, 6 skips 474 decicycles in ff_vp9_idct_idct_4x4_add_ssse3, 1048565 runs, 11 skips (~3.9x faster) 7726 decicycles in idct_idct_8x8_add_c, 1048433 runs, 143 skips 7732 decicycles in idct_idct_8x8_add_c, 2096882 runs, 270 skips 7731 decicycles in idct_idct_8x8_add_c, 4193772 runs, 532 skips 1145 decicycles in ff_vp9_idct_idct_8x8_add_ssse3, 1048549 runs, 27 skips 1137 decicycles in ff_vp9_idct_idct_8x8_add_ssse3, 2097097 runs, 55 skips 1086 decicycles in ff_vp9_idct_idct_8x8_add_ssse3, 4194188 runs, 116 skips (~7.1x faster) Overall decode time before commit: 16.48s user 0.03s system 99% cpu 16.526 total 16.54s user 0.01s system 99% cpu 16.566 total 16.46s user 0.03s system 99% cpu 16.511 total Overall decode time after commit: 16.34s user 0.02s system 99% cpu 16.378 total 16.28s user 0.02s system 99% cpu 16.315 total 16.32s user 0.03s system 99% cpu 16.366 total Tested on i7 920 with 40s 1080p footage.
11 years ago
%endmacro
vp9/x86: idct_32x32_add_ssse3 sub-8x8-idct. Runtime of the full 32x32 idct goes from 2446 to 2441 cycles (intra) or from 1425 to 1306 cycles (inter). Overall runtime is not significantly affected.
11 years ago
%macro VP9_UNPACK_MULSUB_2W_4X 7-9 ; dst1, dst2, (src1, src2,) coef1, coef2, rnd, tmp1, tmp2
%if %0 == 7
vp9: minor refactorings in idct ssse3 assembly. Make register usage in macros explicit; change mulsub_2w_4x to use 2 instead of 3 temp registers.
11 years ago
punpckhwd m%6, m%2, m%1
VP9_MULSUB_2W_2X %7, %6, %6, %5, [pw_m%3_%4], [pw_%4_%3]
avcodec/vp9: add ff_vp9_idct_idct_{4x4,8x8}_ssse3(). 1789 decicycles in idct_idct_4x4_add_c, 262136 runs, 8 skips 1839 decicycles in idct_idct_4x4_add_c, 524270 runs, 18 skips 1864 decicycles in idct_idct_4x4_add_c, 1048548 runs, 28 skips 529 decicycles in ff_vp9_idct_idct_4x4_add_ssse3, 262138 runs, 6 skips 516 decicycles in ff_vp9_idct_idct_4x4_add_ssse3, 524282 runs, 6 skips 474 decicycles in ff_vp9_idct_idct_4x4_add_ssse3, 1048565 runs, 11 skips (~3.9x faster) 7726 decicycles in idct_idct_8x8_add_c, 1048433 runs, 143 skips 7732 decicycles in idct_idct_8x8_add_c, 2096882 runs, 270 skips 7731 decicycles in idct_idct_8x8_add_c, 4193772 runs, 532 skips 1145 decicycles in ff_vp9_idct_idct_8x8_add_ssse3, 1048549 runs, 27 skips 1137 decicycles in ff_vp9_idct_idct_8x8_add_ssse3, 2097097 runs, 55 skips 1086 decicycles in ff_vp9_idct_idct_8x8_add_ssse3, 4194188 runs, 116 skips (~7.1x faster) Overall decode time before commit: 16.48s user 0.03s system 99% cpu 16.526 total 16.54s user 0.01s system 99% cpu 16.566 total 16.46s user 0.03s system 99% cpu 16.511 total Overall decode time after commit: 16.34s user 0.02s system 99% cpu 16.378 total 16.28s user 0.02s system 99% cpu 16.315 total 16.32s user 0.03s system 99% cpu 16.366 total Tested on i7 920 with 40s 1080p footage.
11 years ago
punpcklwd m%2, m%1
vp9: minor refactorings in idct ssse3 assembly. Make register usage in macros explicit; change mulsub_2w_4x to use 2 instead of 3 temp registers.
11 years ago
VP9_MULSUB_2W_2X %1, %2, %2, %5, [pw_m%3_%4], [pw_%4_%3]
packssdw m%1, m%7
packssdw m%2, m%6
vp9/x86: idct_32x32_add_ssse3 sub-8x8-idct. Runtime of the full 32x32 idct goes from 2446 to 2441 cycles (intra) or from 1425 to 1306 cycles (inter). Overall runtime is not significantly affected.
11 years ago
%else
punpckhwd m%8, m%4, m%3
VP9_MULSUB_2W_2X %9, %8, %8, %7, [pw_m%5_%6], [pw_%6_%5]
punpcklwd m%2, m%4, m%3
VP9_MULSUB_2W_2X %1, %2, %2, %7, [pw_m%5_%6], [pw_%6_%5]
packssdw m%1, m%9
packssdw m%2, m%8
%endif
avcodec/vp9: add ff_vp9_idct_idct_{4x4,8x8}_ssse3(). 1789 decicycles in idct_idct_4x4_add_c, 262136 runs, 8 skips 1839 decicycles in idct_idct_4x4_add_c, 524270 runs, 18 skips 1864 decicycles in idct_idct_4x4_add_c, 1048548 runs, 28 skips 529 decicycles in ff_vp9_idct_idct_4x4_add_ssse3, 262138 runs, 6 skips 516 decicycles in ff_vp9_idct_idct_4x4_add_ssse3, 524282 runs, 6 skips 474 decicycles in ff_vp9_idct_idct_4x4_add_ssse3, 1048565 runs, 11 skips (~3.9x faster) 7726 decicycles in idct_idct_8x8_add_c, 1048433 runs, 143 skips 7732 decicycles in idct_idct_8x8_add_c, 2096882 runs, 270 skips 7731 decicycles in idct_idct_8x8_add_c, 4193772 runs, 532 skips 1145 decicycles in ff_vp9_idct_idct_8x8_add_ssse3, 1048549 runs, 27 skips 1137 decicycles in ff_vp9_idct_idct_8x8_add_ssse3, 2097097 runs, 55 skips 1086 decicycles in ff_vp9_idct_idct_8x8_add_ssse3, 4194188 runs, 116 skips (~7.1x faster) Overall decode time before commit: 16.48s user 0.03s system 99% cpu 16.526 total 16.54s user 0.01s system 99% cpu 16.566 total 16.46s user 0.03s system 99% cpu 16.511 total Overall decode time after commit: 16.34s user 0.02s system 99% cpu 16.378 total 16.28s user 0.02s system 99% cpu 16.315 total 16.32s user 0.03s system 99% cpu 16.366 total Tested on i7 920 with 40s 1080p footage.
11 years ago
%endmacro
vp9/x86: idct_32x32_add_ssse3. Sub-IDCTs will follow later. ped1080.webm goes from 9.295s to 8.191s (13.5% faster). The IDCT itself goes from 4372 (intra) or 4337 (inter) to 403 (intra) or 329 (inter) cycles for the DC-only form, 23755 (intra) or 23723 (inter) to 3497 (intra) or 3607 (inter) cycles for the no-DC form, which averages from 23393 (intra) or 16612 (inter) to 3449 (intra) or 2392 (inter) for all 32x32s together, i.e. about ~7x faster (all tests done on ped1080p.webm).
11 years ago
%macro VP9_STORE_2X 5-6 dstq ; reg1, reg2, tmp1, tmp2, zero, dst
movh m%3, [%6]
movh m%4, [%6+strideq]
vp9: minor refactorings in idct ssse3 assembly. Make register usage in macros explicit; change mulsub_2w_4x to use 2 instead of 3 temp registers.
11 years ago
punpcklbw m%3, m%5
punpcklbw m%4, m%5
paddw m%3, m%1
paddw m%4, m%2
packuswb m%3, m%5
packuswb m%4, m%5
vp9/x86: idct_32x32_add_ssse3. Sub-IDCTs will follow later. ped1080.webm goes from 9.295s to 8.191s (13.5% faster). The IDCT itself goes from 4372 (intra) or 4337 (inter) to 403 (intra) or 329 (inter) cycles for the DC-only form, 23755 (intra) or 23723 (inter) to 3497 (intra) or 3607 (inter) cycles for the no-DC form, which averages from 23393 (intra) or 16612 (inter) to 3449 (intra) or 2392 (inter) for all 32x32s together, i.e. about ~7x faster (all tests done on ped1080p.webm).
11 years ago
movh [%6], m%3
movh [%6+strideq], m%4
avcodec/vp9: add ff_vp9_idct_idct_{4x4,8x8}_ssse3(). 1789 decicycles in idct_idct_4x4_add_c, 262136 runs, 8 skips 1839 decicycles in idct_idct_4x4_add_c, 524270 runs, 18 skips 1864 decicycles in idct_idct_4x4_add_c, 1048548 runs, 28 skips 529 decicycles in ff_vp9_idct_idct_4x4_add_ssse3, 262138 runs, 6 skips 516 decicycles in ff_vp9_idct_idct_4x4_add_ssse3, 524282 runs, 6 skips 474 decicycles in ff_vp9_idct_idct_4x4_add_ssse3, 1048565 runs, 11 skips (~3.9x faster) 7726 decicycles in idct_idct_8x8_add_c, 1048433 runs, 143 skips 7732 decicycles in idct_idct_8x8_add_c, 2096882 runs, 270 skips 7731 decicycles in idct_idct_8x8_add_c, 4193772 runs, 532 skips 1145 decicycles in ff_vp9_idct_idct_8x8_add_ssse3, 1048549 runs, 27 skips 1137 decicycles in ff_vp9_idct_idct_8x8_add_ssse3, 2097097 runs, 55 skips 1086 decicycles in ff_vp9_idct_idct_8x8_add_ssse3, 4194188 runs, 116 skips (~7.1x faster) Overall decode time before commit: 16.48s user 0.03s system 99% cpu 16.526 total 16.54s user 0.01s system 99% cpu 16.566 total 16.46s user 0.03s system 99% cpu 16.511 total Overall decode time after commit: 16.34s user 0.02s system 99% cpu 16.378 total 16.28s user 0.02s system 99% cpu 16.315 total 16.32s user 0.03s system 99% cpu 16.366 total Tested on i7 920 with 40s 1080p footage.
11 years ago
%endmacro
;-------------------------------------------------------------------------------------------
; void vp9_idct_idct_4x4_add_<opt>(uint8_t *dst, ptrdiff_t stride, int16_t *block, int eob);
;-------------------------------------------------------------------------------------------
%macro VP9_IDCT4_1D_FINALIZE 0
SUMSUB_BA w, 3, 2, 4 ; m3=t3+t0, m2=-t3+t0
SUMSUB_BA w, 1, 0, 4 ; m1=t2+t1, m0=-t2+t1
avcodec/x86/vp9dsp: merge a few SWAP together.
11 years ago
SWAP 0, 3, 2 ; 3102 -> 0123
avcodec/vp9: add ff_vp9_idct_idct_{4x4,8x8}_ssse3(). 1789 decicycles in idct_idct_4x4_add_c, 262136 runs, 8 skips 1839 decicycles in idct_idct_4x4_add_c, 524270 runs, 18 skips 1864 decicycles in idct_idct_4x4_add_c, 1048548 runs, 28 skips 529 decicycles in ff_vp9_idct_idct_4x4_add_ssse3, 262138 runs, 6 skips 516 decicycles in ff_vp9_idct_idct_4x4_add_ssse3, 524282 runs, 6 skips 474 decicycles in ff_vp9_idct_idct_4x4_add_ssse3, 1048565 runs, 11 skips (~3.9x faster) 7726 decicycles in idct_idct_8x8_add_c, 1048433 runs, 143 skips 7732 decicycles in idct_idct_8x8_add_c, 2096882 runs, 270 skips 7731 decicycles in idct_idct_8x8_add_c, 4193772 runs, 532 skips 1145 decicycles in ff_vp9_idct_idct_8x8_add_ssse3, 1048549 runs, 27 skips 1137 decicycles in ff_vp9_idct_idct_8x8_add_ssse3, 2097097 runs, 55 skips 1086 decicycles in ff_vp9_idct_idct_8x8_add_ssse3, 4194188 runs, 116 skips (~7.1x faster) Overall decode time before commit: 16.48s user 0.03s system 99% cpu 16.526 total 16.54s user 0.01s system 99% cpu 16.566 total 16.46s user 0.03s system 99% cpu 16.511 total Overall decode time after commit: 16.34s user 0.02s system 99% cpu 16.378 total 16.28s user 0.02s system 99% cpu 16.315 total 16.32s user 0.03s system 99% cpu 16.366 total Tested on i7 920 with 40s 1080p footage.
11 years ago
%endmacro
%macro VP9_IDCT4_1D 0
SUMSUB_BA w, 2, 0, 4 ; m2=IN(0)+IN(2) m0=IN(0)-IN(2)
vp9: split pre-load of 11585x2 out of 1d idct macro. This allows us to load it only once, instead of twice, in this function.
11 years ago
pmulhrsw m2, m6 ; m2=t0
pmulhrsw m0, m6 ; m0=t1
vp9: minor refactorings in idct ssse3 assembly. Make register usage in macros explicit; change mulsub_2w_4x to use 2 instead of 3 temp registers.
11 years ago
VP9_UNPACK_MULSUB_2W_4X 1, 3, 15137, 6270, m7, 4, 5 ; m1=t2, m3=t3
avcodec/vp9: add ff_vp9_idct_idct_{4x4,8x8}_ssse3(). 1789 decicycles in idct_idct_4x4_add_c, 262136 runs, 8 skips 1839 decicycles in idct_idct_4x4_add_c, 524270 runs, 18 skips 1864 decicycles in idct_idct_4x4_add_c, 1048548 runs, 28 skips 529 decicycles in ff_vp9_idct_idct_4x4_add_ssse3, 262138 runs, 6 skips 516 decicycles in ff_vp9_idct_idct_4x4_add_ssse3, 524282 runs, 6 skips 474 decicycles in ff_vp9_idct_idct_4x4_add_ssse3, 1048565 runs, 11 skips (~3.9x faster) 7726 decicycles in idct_idct_8x8_add_c, 1048433 runs, 143 skips 7732 decicycles in idct_idct_8x8_add_c, 2096882 runs, 270 skips 7731 decicycles in idct_idct_8x8_add_c, 4193772 runs, 532 skips 1145 decicycles in ff_vp9_idct_idct_8x8_add_ssse3, 1048549 runs, 27 skips 1137 decicycles in ff_vp9_idct_idct_8x8_add_ssse3, 2097097 runs, 55 skips 1086 decicycles in ff_vp9_idct_idct_8x8_add_ssse3, 4194188 runs, 116 skips (~7.1x faster) Overall decode time before commit: 16.48s user 0.03s system 99% cpu 16.526 total 16.54s user 0.01s system 99% cpu 16.566 total 16.46s user 0.03s system 99% cpu 16.511 total Overall decode time after commit: 16.34s user 0.02s system 99% cpu 16.378 total 16.28s user 0.02s system 99% cpu 16.315 total 16.32s user 0.03s system 99% cpu 16.366 total Tested on i7 920 with 40s 1080p footage.
11 years ago
VP9_IDCT4_1D_FINALIZE
%endmacro
; 2x2 top left corner
%macro VP9_IDCT4_2x2_1D 0
pmulhrsw m0, m5 ; m0=t1
mova m2, m0 ; m2=t0
mova m3, m1
pmulhrsw m1, m6 ; m1=t2
pmulhrsw m3, m7 ; m3=t3
VP9_IDCT4_1D_FINALIZE
%endmacro
%macro VP9_IDCT4_WRITEOUT 0
mova m5, [pw_2048]
pmulhrsw m0, m5 ; (x*2048 + (1<<14))>>15 <=> (x+8)>>4
pmulhrsw m1, m5
vp9: minor refactorings in idct ssse3 assembly. Make register usage in macros explicit; change mulsub_2w_4x to use 2 instead of 3 temp registers.
11 years ago
VP9_STORE_2X 0, 1, 6, 7, 4
avcodec/vp9: add ff_vp9_idct_idct_{4x4,8x8}_ssse3(). 1789 decicycles in idct_idct_4x4_add_c, 262136 runs, 8 skips 1839 decicycles in idct_idct_4x4_add_c, 524270 runs, 18 skips 1864 decicycles in idct_idct_4x4_add_c, 1048548 runs, 28 skips 529 decicycles in ff_vp9_idct_idct_4x4_add_ssse3, 262138 runs, 6 skips 516 decicycles in ff_vp9_idct_idct_4x4_add_ssse3, 524282 runs, 6 skips 474 decicycles in ff_vp9_idct_idct_4x4_add_ssse3, 1048565 runs, 11 skips (~3.9x faster) 7726 decicycles in idct_idct_8x8_add_c, 1048433 runs, 143 skips 7732 decicycles in idct_idct_8x8_add_c, 2096882 runs, 270 skips 7731 decicycles in idct_idct_8x8_add_c, 4193772 runs, 532 skips 1145 decicycles in ff_vp9_idct_idct_8x8_add_ssse3, 1048549 runs, 27 skips 1137 decicycles in ff_vp9_idct_idct_8x8_add_ssse3, 2097097 runs, 55 skips 1086 decicycles in ff_vp9_idct_idct_8x8_add_ssse3, 4194188 runs, 116 skips (~7.1x faster) Overall decode time before commit: 16.48s user 0.03s system 99% cpu 16.526 total 16.54s user 0.01s system 99% cpu 16.566 total 16.46s user 0.03s system 99% cpu 16.511 total Overall decode time after commit: 16.34s user 0.02s system 99% cpu 16.378 total 16.28s user 0.02s system 99% cpu 16.315 total 16.32s user 0.03s system 99% cpu 16.366 total Tested on i7 920 with 40s 1080p footage.
11 years ago
lea dstq, [dstq+2*strideq]
pmulhrsw m2, m5
pmulhrsw m3, m5
vp9: minor refactorings in idct ssse3 assembly. Make register usage in macros explicit; change mulsub_2w_4x to use 2 instead of 3 temp registers.
11 years ago
VP9_STORE_2X 2, 3, 6, 7, 4
avcodec/vp9: add ff_vp9_idct_idct_{4x4,8x8}_ssse3(). 1789 decicycles in idct_idct_4x4_add_c, 262136 runs, 8 skips 1839 decicycles in idct_idct_4x4_add_c, 524270 runs, 18 skips 1864 decicycles in idct_idct_4x4_add_c, 1048548 runs, 28 skips 529 decicycles in ff_vp9_idct_idct_4x4_add_ssse3, 262138 runs, 6 skips 516 decicycles in ff_vp9_idct_idct_4x4_add_ssse3, 524282 runs, 6 skips 474 decicycles in ff_vp9_idct_idct_4x4_add_ssse3, 1048565 runs, 11 skips (~3.9x faster) 7726 decicycles in idct_idct_8x8_add_c, 1048433 runs, 143 skips 7732 decicycles in idct_idct_8x8_add_c, 2096882 runs, 270 skips 7731 decicycles in idct_idct_8x8_add_c, 4193772 runs, 532 skips 1145 decicycles in ff_vp9_idct_idct_8x8_add_ssse3, 1048549 runs, 27 skips 1137 decicycles in ff_vp9_idct_idct_8x8_add_ssse3, 2097097 runs, 55 skips 1086 decicycles in ff_vp9_idct_idct_8x8_add_ssse3, 4194188 runs, 116 skips (~7.1x faster) Overall decode time before commit: 16.48s user 0.03s system 99% cpu 16.526 total 16.54s user 0.01s system 99% cpu 16.566 total 16.46s user 0.03s system 99% cpu 16.511 total Overall decode time after commit: 16.34s user 0.02s system 99% cpu 16.378 total 16.28s user 0.02s system 99% cpu 16.315 total 16.32s user 0.03s system 99% cpu 16.366 total Tested on i7 920 with 40s 1080p footage.
11 years ago
%endmacro
INIT_MMX ssse3
cglobal vp9_idct_idct_4x4_add, 4,4,0, dst, stride, block, eob
cmp eobd, 4 ; 2x2 or smaller
jg .idctfull
cmp eobd, 1 ; faster path for when only DC is set
jne .idct2x2
movd m0, [blockq]
mova m5, [pw_11585x2]
pmulhrsw m0, m5
pmulhrsw m0, m5
pshufw m0, m0, 0
pxor m4, m4
movh [blockq], m4
vp9: minor refactorings in idct ssse3 assembly. Make register usage in macros explicit; change mulsub_2w_4x to use 2 instead of 3 temp registers.
11 years ago
pmulhrsw m0, [pw_2048] ; (x*2048 + (1<<14))>>15 <=> (x+8)>>4
VP9_STORE_2X 0, 0, 6, 7, 4
avcodec/vp9: add ff_vp9_idct_idct_{4x4,8x8}_ssse3(). 1789 decicycles in idct_idct_4x4_add_c, 262136 runs, 8 skips 1839 decicycles in idct_idct_4x4_add_c, 524270 runs, 18 skips 1864 decicycles in idct_idct_4x4_add_c, 1048548 runs, 28 skips 529 decicycles in ff_vp9_idct_idct_4x4_add_ssse3, 262138 runs, 6 skips 516 decicycles in ff_vp9_idct_idct_4x4_add_ssse3, 524282 runs, 6 skips 474 decicycles in ff_vp9_idct_idct_4x4_add_ssse3, 1048565 runs, 11 skips (~3.9x faster) 7726 decicycles in idct_idct_8x8_add_c, 1048433 runs, 143 skips 7732 decicycles in idct_idct_8x8_add_c, 2096882 runs, 270 skips 7731 decicycles in idct_idct_8x8_add_c, 4193772 runs, 532 skips 1145 decicycles in ff_vp9_idct_idct_8x8_add_ssse3, 1048549 runs, 27 skips 1137 decicycles in ff_vp9_idct_idct_8x8_add_ssse3, 2097097 runs, 55 skips 1086 decicycles in ff_vp9_idct_idct_8x8_add_ssse3, 4194188 runs, 116 skips (~7.1x faster) Overall decode time before commit: 16.48s user 0.03s system 99% cpu 16.526 total 16.54s user 0.01s system 99% cpu 16.566 total 16.46s user 0.03s system 99% cpu 16.511 total Overall decode time after commit: 16.34s user 0.02s system 99% cpu 16.378 total 16.28s user 0.02s system 99% cpu 16.315 total 16.32s user 0.03s system 99% cpu 16.366 total Tested on i7 920 with 40s 1080p footage.
11 years ago
lea dstq, [dstq+2*strideq]
vp9: minor refactorings in idct ssse3 assembly. Make register usage in macros explicit; change mulsub_2w_4x to use 2 instead of 3 temp registers.
11 years ago
VP9_STORE_2X 0, 0, 6, 7, 4
avcodec/vp9: add ff_vp9_idct_idct_{4x4,8x8}_ssse3(). 1789 decicycles in idct_idct_4x4_add_c, 262136 runs, 8 skips 1839 decicycles in idct_idct_4x4_add_c, 524270 runs, 18 skips 1864 decicycles in idct_idct_4x4_add_c, 1048548 runs, 28 skips 529 decicycles in ff_vp9_idct_idct_4x4_add_ssse3, 262138 runs, 6 skips 516 decicycles in ff_vp9_idct_idct_4x4_add_ssse3, 524282 runs, 6 skips 474 decicycles in ff_vp9_idct_idct_4x4_add_ssse3, 1048565 runs, 11 skips (~3.9x faster) 7726 decicycles in idct_idct_8x8_add_c, 1048433 runs, 143 skips 7732 decicycles in idct_idct_8x8_add_c, 2096882 runs, 270 skips 7731 decicycles in idct_idct_8x8_add_c, 4193772 runs, 532 skips 1145 decicycles in ff_vp9_idct_idct_8x8_add_ssse3, 1048549 runs, 27 skips 1137 decicycles in ff_vp9_idct_idct_8x8_add_ssse3, 2097097 runs, 55 skips 1086 decicycles in ff_vp9_idct_idct_8x8_add_ssse3, 4194188 runs, 116 skips (~7.1x faster) Overall decode time before commit: 16.48s user 0.03s system 99% cpu 16.526 total 16.54s user 0.01s system 99% cpu 16.566 total 16.46s user 0.03s system 99% cpu 16.511 total Overall decode time after commit: 16.34s user 0.02s system 99% cpu 16.378 total 16.28s user 0.02s system 99% cpu 16.315 total 16.32s user 0.03s system 99% cpu 16.366 total Tested on i7 920 with 40s 1080p footage.
11 years ago
RET
; faster path for when only top left 2x2 block is set
.idct2x2:
movd m0, [blockq+0]
movd m1, [blockq+8]
mova m5, [pw_11585x2]
mova m6, [pw_6270x2]
mova m7, [pw_15137x2]
VP9_IDCT4_2x2_1D
TRANSPOSE4x4W 0, 1, 2, 3, 4
VP9_IDCT4_2x2_1D
pxor m4, m4 ; used for the block reset, and VP9_STORE_2X
movh [blockq+ 0], m4
movh [blockq+ 8], m4
VP9_IDCT4_WRITEOUT
RET
.idctfull: ; generic full 4x4 idct/idct
mova m0, [blockq+ 0]
mova m1, [blockq+ 8]
mova m2, [blockq+16]
mova m3, [blockq+24]
vp9: split pre-load of 11585x2 out of 1d idct macro. This allows us to load it only once, instead of twice, in this function.
11 years ago
mova m6, [pw_11585x2]
avcodec/vp9: add ff_vp9_idct_idct_{4x4,8x8}_ssse3(). 1789 decicycles in idct_idct_4x4_add_c, 262136 runs, 8 skips 1839 decicycles in idct_idct_4x4_add_c, 524270 runs, 18 skips 1864 decicycles in idct_idct_4x4_add_c, 1048548 runs, 28 skips 529 decicycles in ff_vp9_idct_idct_4x4_add_ssse3, 262138 runs, 6 skips 516 decicycles in ff_vp9_idct_idct_4x4_add_ssse3, 524282 runs, 6 skips 474 decicycles in ff_vp9_idct_idct_4x4_add_ssse3, 1048565 runs, 11 skips (~3.9x faster) 7726 decicycles in idct_idct_8x8_add_c, 1048433 runs, 143 skips 7732 decicycles in idct_idct_8x8_add_c, 2096882 runs, 270 skips 7731 decicycles in idct_idct_8x8_add_c, 4193772 runs, 532 skips 1145 decicycles in ff_vp9_idct_idct_8x8_add_ssse3, 1048549 runs, 27 skips 1137 decicycles in ff_vp9_idct_idct_8x8_add_ssse3, 2097097 runs, 55 skips 1086 decicycles in ff_vp9_idct_idct_8x8_add_ssse3, 4194188 runs, 116 skips (~7.1x faster) Overall decode time before commit: 16.48s user 0.03s system 99% cpu 16.526 total 16.54s user 0.01s system 99% cpu 16.566 total 16.46s user 0.03s system 99% cpu 16.511 total Overall decode time after commit: 16.34s user 0.02s system 99% cpu 16.378 total 16.28s user 0.02s system 99% cpu 16.315 total 16.32s user 0.03s system 99% cpu 16.366 total Tested on i7 920 with 40s 1080p footage.
11 years ago
mova m7, [pd_8192] ; rounding
VP9_IDCT4_1D
TRANSPOSE4x4W 0, 1, 2, 3, 4
VP9_IDCT4_1D
pxor m4, m4 ; used for the block reset, and VP9_STORE_2X
mova [blockq+ 0], m4
mova [blockq+ 8], m4
mova [blockq+16], m4
mova [blockq+24], m4
VP9_IDCT4_WRITEOUT
RET
vp9/x86: idct_add_16x16_ssse3. Currently only dc-only and full 16x16. Other subforms will follow in the near future. Total decoding time of ped1080p.webm goes from 9.7 to 9.3 seconds. DC-only goes from 957 -> 131 cycles, and the full IDCT goes from ~4050 to ~745 cycles.
11 years ago
%if ARCH_X86_64 ; TODO: 32-bit? (32-bit limited to 8 xmm reg, we use more)
avcodec/vp9: add ff_vp9_idct_idct_{4x4,8x8}_ssse3(). 1789 decicycles in idct_idct_4x4_add_c, 262136 runs, 8 skips 1839 decicycles in idct_idct_4x4_add_c, 524270 runs, 18 skips 1864 decicycles in idct_idct_4x4_add_c, 1048548 runs, 28 skips 529 decicycles in ff_vp9_idct_idct_4x4_add_ssse3, 262138 runs, 6 skips 516 decicycles in ff_vp9_idct_idct_4x4_add_ssse3, 524282 runs, 6 skips 474 decicycles in ff_vp9_idct_idct_4x4_add_ssse3, 1048565 runs, 11 skips (~3.9x faster) 7726 decicycles in idct_idct_8x8_add_c, 1048433 runs, 143 skips 7732 decicycles in idct_idct_8x8_add_c, 2096882 runs, 270 skips 7731 decicycles in idct_idct_8x8_add_c, 4193772 runs, 532 skips 1145 decicycles in ff_vp9_idct_idct_8x8_add_ssse3, 1048549 runs, 27 skips 1137 decicycles in ff_vp9_idct_idct_8x8_add_ssse3, 2097097 runs, 55 skips 1086 decicycles in ff_vp9_idct_idct_8x8_add_ssse3, 4194188 runs, 116 skips (~7.1x faster) Overall decode time before commit: 16.48s user 0.03s system 99% cpu 16.526 total 16.54s user 0.01s system 99% cpu 16.566 total 16.46s user 0.03s system 99% cpu 16.511 total Overall decode time after commit: 16.34s user 0.02s system 99% cpu 16.378 total 16.28s user 0.02s system 99% cpu 16.315 total 16.32s user 0.03s system 99% cpu 16.366 total Tested on i7 920 with 40s 1080p footage.
11 years ago
;-------------------------------------------------------------------------------------------
; void vp9_idct_idct_8x8_add_<opt>(uint8_t *dst, ptrdiff_t stride, int16_t *block, int eob);
;-------------------------------------------------------------------------------------------
%macro VP9_IDCT8_1D_FINALIZE 0
SUMSUB_BA w, 3, 10, 4 ; m3=t0+t7, m10=t0-t7
SUMSUB_BA w, 1, 2, 4 ; m1=t1+t6, m2=t1-t6
SUMSUB_BA w, 11, 0, 4 ; m11=t2+t5, m0=t2-t5
SUMSUB_BA w, 9, 8, 4 ; m9=t3+t4, m8=t3-t4
avcodec/x86/vp9dsp: merge a few SWAP together.
11 years ago
SWAP 11, 10, 2
SWAP 3, 9, 0
avcodec/vp9: add ff_vp9_idct_idct_{4x4,8x8}_ssse3(). 1789 decicycles in idct_idct_4x4_add_c, 262136 runs, 8 skips 1839 decicycles in idct_idct_4x4_add_c, 524270 runs, 18 skips 1864 decicycles in idct_idct_4x4_add_c, 1048548 runs, 28 skips 529 decicycles in ff_vp9_idct_idct_4x4_add_ssse3, 262138 runs, 6 skips 516 decicycles in ff_vp9_idct_idct_4x4_add_ssse3, 524282 runs, 6 skips 474 decicycles in ff_vp9_idct_idct_4x4_add_ssse3, 1048565 runs, 11 skips (~3.9x faster) 7726 decicycles in idct_idct_8x8_add_c, 1048433 runs, 143 skips 7732 decicycles in idct_idct_8x8_add_c, 2096882 runs, 270 skips 7731 decicycles in idct_idct_8x8_add_c, 4193772 runs, 532 skips 1145 decicycles in ff_vp9_idct_idct_8x8_add_ssse3, 1048549 runs, 27 skips 1137 decicycles in ff_vp9_idct_idct_8x8_add_ssse3, 2097097 runs, 55 skips 1086 decicycles in ff_vp9_idct_idct_8x8_add_ssse3, 4194188 runs, 116 skips (~7.1x faster) Overall decode time before commit: 16.48s user 0.03s system 99% cpu 16.526 total 16.54s user 0.01s system 99% cpu 16.566 total 16.46s user 0.03s system 99% cpu 16.511 total Overall decode time after commit: 16.34s user 0.02s system 99% cpu 16.378 total 16.28s user 0.02s system 99% cpu 16.315 total 16.32s user 0.03s system 99% cpu 16.366 total Tested on i7 920 with 40s 1080p footage.
11 years ago
%endmacro
%macro VP9_IDCT8_1D 0
SUMSUB_BA w, 8, 0, 4 ; m8=IN(0)+IN(4) m0=IN(0)-IN(4)
pmulhrsw m8, m12 ; m8=t0a
pmulhrsw m0, m12 ; m0=t1a
vp9: minor refactorings in idct ssse3 assembly. Make register usage in macros explicit; change mulsub_2w_4x to use 2 instead of 3 temp registers.
11 years ago
VP9_UNPACK_MULSUB_2W_4X 2, 10, 15137, 6270, m7, 4, 5 ; m2=t2a, m10=t3a
VP9_UNPACK_MULSUB_2W_4X 1, 11, 16069, 3196, m7, 4, 5 ; m1=t4a, m11=t7a
VP9_UNPACK_MULSUB_2W_4X 9, 3, 9102, 13623, m7, 4, 5 ; m9=t5a, m3=t6a
avcodec/vp9: add ff_vp9_idct_idct_{4x4,8x8}_ssse3(). 1789 decicycles in idct_idct_4x4_add_c, 262136 runs, 8 skips 1839 decicycles in idct_idct_4x4_add_c, 524270 runs, 18 skips 1864 decicycles in idct_idct_4x4_add_c, 1048548 runs, 28 skips 529 decicycles in ff_vp9_idct_idct_4x4_add_ssse3, 262138 runs, 6 skips 516 decicycles in ff_vp9_idct_idct_4x4_add_ssse3, 524282 runs, 6 skips 474 decicycles in ff_vp9_idct_idct_4x4_add_ssse3, 1048565 runs, 11 skips (~3.9x faster) 7726 decicycles in idct_idct_8x8_add_c, 1048433 runs, 143 skips 7732 decicycles in idct_idct_8x8_add_c, 2096882 runs, 270 skips 7731 decicycles in idct_idct_8x8_add_c, 4193772 runs, 532 skips 1145 decicycles in ff_vp9_idct_idct_8x8_add_ssse3, 1048549 runs, 27 skips 1137 decicycles in ff_vp9_idct_idct_8x8_add_ssse3, 2097097 runs, 55 skips 1086 decicycles in ff_vp9_idct_idct_8x8_add_ssse3, 4194188 runs, 116 skips (~7.1x faster) Overall decode time before commit: 16.48s user 0.03s system 99% cpu 16.526 total 16.54s user 0.01s system 99% cpu 16.566 total 16.46s user 0.03s system 99% cpu 16.511 total Overall decode time after commit: 16.34s user 0.02s system 99% cpu 16.378 total 16.28s user 0.02s system 99% cpu 16.315 total 16.32s user 0.03s system 99% cpu 16.366 total Tested on i7 920 with 40s 1080p footage.
11 years ago
SUMSUB_BA w, 10, 8, 4 ; m10=t0a+t3a (t0), m8=t0a-t3a (t3)
SUMSUB_BA w, 2, 0, 4 ; m2=t1a+t2a (t1), m0=t1a-t2a (t2)
SUMSUB_BA w, 9, 1, 4 ; m9=t4a+t5a (t4), m1=t4a-t5a (t5a)
SUMSUB_BA w, 3, 11, 4 ; m3=t7a+t6a (t7), m11=t7a-t6a (t6a)
SUMSUB_BA w, 1, 11, 4 ; m1=t6a+t5a (t6), m11=t6a-t5a (t5)
pmulhrsw m1, m12 ; m1=t6
pmulhrsw m11, m12 ; m11=t5
VP9_IDCT8_1D_FINALIZE
%endmacro
vp9: implement top/left half (4x4) sub-8x8-IDCT. For that specific case (eob>3&&eob<=12), runtime of idct8x8 goes from 668 to 477 cycles. For all idct8x8, runtime goes from 521 to 490 cycles.
11 years ago
%macro VP9_IDCT8_4x4_1D 0
pmulhrsw m0, m12 ; m0=t1a/t0a
pmulhrsw m10, m2, [pw_15137x2] ; m10=t3a
pmulhrsw m2, [pw_6270x2] ; m2=t2a
pmulhrsw m11, m1, [pw_16069x2] ; m11=t7a
pmulhrsw m1, [pw_3196x2] ; m1=t4a
pmulhrsw m9, m3, [pw_9102x2] ; m9=-t5a
pmulhrsw m3, [pw_13623x2] ; m3=t6a
psubw m8, m0, m10 ; m8=t0a-t3a (t3)
paddw m10, m0 ; m10=t0a+t3a (t0)
SUMSUB_BA w, 2, 0, 4 ; m2=t1a+t2a (t1), m0=t1a-t2a (t2)
SUMSUB_BA w, 9, 1, 4 ; m1=t4a+t5a (t4), m9=t4a-t5a (t5a)
SWAP 1, 9
SUMSUB_BA w, 3, 11, 4 ; m3=t7a+t6a (t7), m11=t7a-t6a (t6a)
SUMSUB_BA w, 1, 11, 4 ; m1=t6a+t5a (t6), m11=t6a-t5a (t5)
pmulhrsw m1, m12 ; m1=t6
pmulhrsw m11, m12 ; m11=t5
VP9_IDCT8_1D_FINALIZE
%endmacro
avcodec/vp9: add ff_vp9_idct_idct_{4x4,8x8}_ssse3(). 1789 decicycles in idct_idct_4x4_add_c, 262136 runs, 8 skips 1839 decicycles in idct_idct_4x4_add_c, 524270 runs, 18 skips 1864 decicycles in idct_idct_4x4_add_c, 1048548 runs, 28 skips 529 decicycles in ff_vp9_idct_idct_4x4_add_ssse3, 262138 runs, 6 skips 516 decicycles in ff_vp9_idct_idct_4x4_add_ssse3, 524282 runs, 6 skips 474 decicycles in ff_vp9_idct_idct_4x4_add_ssse3, 1048565 runs, 11 skips (~3.9x faster) 7726 decicycles in idct_idct_8x8_add_c, 1048433 runs, 143 skips 7732 decicycles in idct_idct_8x8_add_c, 2096882 runs, 270 skips 7731 decicycles in idct_idct_8x8_add_c, 4193772 runs, 532 skips 1145 decicycles in ff_vp9_idct_idct_8x8_add_ssse3, 1048549 runs, 27 skips 1137 decicycles in ff_vp9_idct_idct_8x8_add_ssse3, 2097097 runs, 55 skips 1086 decicycles in ff_vp9_idct_idct_8x8_add_ssse3, 4194188 runs, 116 skips (~7.1x faster) Overall decode time before commit: 16.48s user 0.03s system 99% cpu 16.526 total 16.54s user 0.01s system 99% cpu 16.566 total 16.46s user 0.03s system 99% cpu 16.511 total Overall decode time after commit: 16.34s user 0.02s system 99% cpu 16.378 total 16.28s user 0.02s system 99% cpu 16.315 total 16.32s user 0.03s system 99% cpu 16.366 total Tested on i7 920 with 40s 1080p footage.
11 years ago
; TODO: a lot of t* copies can probably be removed and merged with
; following SUMSUBs from VP9_IDCT8_1D_FINALIZE with AVX
%macro VP9_IDCT8_2x2_1D 0
pmulhrsw m0, m12 ; m0=t0
mova m3, m1
pmulhrsw m1, m6 ; m1=t4
pmulhrsw m3, m7 ; m3=t7
mova m2, m0 ; m2=t1
mova m10, m0 ; m10=t2
mova m8, m0 ; m8=t3
mova m11, m3 ; t5 = t7a ...
mova m9, m3 ; t6 = t7a ...
psubw m11, m1 ; t5 = t7a - t4a
paddw m9, m1 ; t6 = t7a + t4a
pmulhrsw m11, m12 ; m11=t5
pmulhrsw m9, m12 ; m9=t6
SWAP 0, 10
SWAP 9, 1
VP9_IDCT8_1D_FINALIZE
%endmacro
%macro VP9_IDCT8_WRITEOUT 0
mova m5, [pw_1024]
pmulhrsw m0, m5 ; (x*1024 + (1<<14))>>15 <=> (x+16)>>5
pmulhrsw m1, m5
vp9: minor refactorings in idct ssse3 assembly. Make register usage in macros explicit; change mulsub_2w_4x to use 2 instead of 3 temp registers.
11 years ago
VP9_STORE_2X 0, 1, 6, 7, 4
avcodec/vp9: add ff_vp9_idct_idct_{4x4,8x8}_ssse3(). 1789 decicycles in idct_idct_4x4_add_c, 262136 runs, 8 skips 1839 decicycles in idct_idct_4x4_add_c, 524270 runs, 18 skips 1864 decicycles in idct_idct_4x4_add_c, 1048548 runs, 28 skips 529 decicycles in ff_vp9_idct_idct_4x4_add_ssse3, 262138 runs, 6 skips 516 decicycles in ff_vp9_idct_idct_4x4_add_ssse3, 524282 runs, 6 skips 474 decicycles in ff_vp9_idct_idct_4x4_add_ssse3, 1048565 runs, 11 skips (~3.9x faster) 7726 decicycles in idct_idct_8x8_add_c, 1048433 runs, 143 skips 7732 decicycles in idct_idct_8x8_add_c, 2096882 runs, 270 skips 7731 decicycles in idct_idct_8x8_add_c, 4193772 runs, 532 skips 1145 decicycles in ff_vp9_idct_idct_8x8_add_ssse3, 1048549 runs, 27 skips 1137 decicycles in ff_vp9_idct_idct_8x8_add_ssse3, 2097097 runs, 55 skips 1086 decicycles in ff_vp9_idct_idct_8x8_add_ssse3, 4194188 runs, 116 skips (~7.1x faster) Overall decode time before commit: 16.48s user 0.03s system 99% cpu 16.526 total 16.54s user 0.01s system 99% cpu 16.566 total 16.46s user 0.03s system 99% cpu 16.511 total Overall decode time after commit: 16.34s user 0.02s system 99% cpu 16.378 total 16.28s user 0.02s system 99% cpu 16.315 total 16.32s user 0.03s system 99% cpu 16.366 total Tested on i7 920 with 40s 1080p footage.
11 years ago
lea dstq, [dstq+2*strideq]
pmulhrsw m2, m5
pmulhrsw m3, m5
vp9: minor refactorings in idct ssse3 assembly. Make register usage in macros explicit; change mulsub_2w_4x to use 2 instead of 3 temp registers.
11 years ago
VP9_STORE_2X 2, 3, 6, 7, 4
avcodec/vp9: add ff_vp9_idct_idct_{4x4,8x8}_ssse3(). 1789 decicycles in idct_idct_4x4_add_c, 262136 runs, 8 skips 1839 decicycles in idct_idct_4x4_add_c, 524270 runs, 18 skips 1864 decicycles in idct_idct_4x4_add_c, 1048548 runs, 28 skips 529 decicycles in ff_vp9_idct_idct_4x4_add_ssse3, 262138 runs, 6 skips 516 decicycles in ff_vp9_idct_idct_4x4_add_ssse3, 524282 runs, 6 skips 474 decicycles in ff_vp9_idct_idct_4x4_add_ssse3, 1048565 runs, 11 skips (~3.9x faster) 7726 decicycles in idct_idct_8x8_add_c, 1048433 runs, 143 skips 7732 decicycles in idct_idct_8x8_add_c, 2096882 runs, 270 skips 7731 decicycles in idct_idct_8x8_add_c, 4193772 runs, 532 skips 1145 decicycles in ff_vp9_idct_idct_8x8_add_ssse3, 1048549 runs, 27 skips 1137 decicycles in ff_vp9_idct_idct_8x8_add_ssse3, 2097097 runs, 55 skips 1086 decicycles in ff_vp9_idct_idct_8x8_add_ssse3, 4194188 runs, 116 skips (~7.1x faster) Overall decode time before commit: 16.48s user 0.03s system 99% cpu 16.526 total 16.54s user 0.01s system 99% cpu 16.566 total 16.46s user 0.03s system 99% cpu 16.511 total Overall decode time after commit: 16.34s user 0.02s system 99% cpu 16.378 total 16.28s user 0.02s system 99% cpu 16.315 total 16.32s user 0.03s system 99% cpu 16.366 total Tested on i7 920 with 40s 1080p footage.
11 years ago
lea dstq, [dstq+2*strideq]
pmulhrsw m8, m5
pmulhrsw m9, m5
vp9: minor refactorings in idct ssse3 assembly. Make register usage in macros explicit; change mulsub_2w_4x to use 2 instead of 3 temp registers.
11 years ago
VP9_STORE_2X 8, 9, 6, 7, 4
avcodec/vp9: add ff_vp9_idct_idct_{4x4,8x8}_ssse3(). 1789 decicycles in idct_idct_4x4_add_c, 262136 runs, 8 skips 1839 decicycles in idct_idct_4x4_add_c, 524270 runs, 18 skips 1864 decicycles in idct_idct_4x4_add_c, 1048548 runs, 28 skips 529 decicycles in ff_vp9_idct_idct_4x4_add_ssse3, 262138 runs, 6 skips 516 decicycles in ff_vp9_idct_idct_4x4_add_ssse3, 524282 runs, 6 skips 474 decicycles in ff_vp9_idct_idct_4x4_add_ssse3, 1048565 runs, 11 skips (~3.9x faster) 7726 decicycles in idct_idct_8x8_add_c, 1048433 runs, 143 skips 7732 decicycles in idct_idct_8x8_add_c, 2096882 runs, 270 skips 7731 decicycles in idct_idct_8x8_add_c, 4193772 runs, 532 skips 1145 decicycles in ff_vp9_idct_idct_8x8_add_ssse3, 1048549 runs, 27 skips 1137 decicycles in ff_vp9_idct_idct_8x8_add_ssse3, 2097097 runs, 55 skips 1086 decicycles in ff_vp9_idct_idct_8x8_add_ssse3, 4194188 runs, 116 skips (~7.1x faster) Overall decode time before commit: 16.48s user 0.03s system 99% cpu 16.526 total 16.54s user 0.01s system 99% cpu 16.566 total 16.46s user 0.03s system 99% cpu 16.511 total Overall decode time after commit: 16.34s user 0.02s system 99% cpu 16.378 total 16.28s user 0.02s system 99% cpu 16.315 total 16.32s user 0.03s system 99% cpu 16.366 total Tested on i7 920 with 40s 1080p footage.
11 years ago
lea dstq, [dstq+2*strideq]
pmulhrsw m10, m5
pmulhrsw m11, m5
vp9: minor refactorings in idct ssse3 assembly. Make register usage in macros explicit; change mulsub_2w_4x to use 2 instead of 3 temp registers.
11 years ago
VP9_STORE_2X 10, 11, 6, 7, 4
avcodec/vp9: add ff_vp9_idct_idct_{4x4,8x8}_ssse3(). 1789 decicycles in idct_idct_4x4_add_c, 262136 runs, 8 skips 1839 decicycles in idct_idct_4x4_add_c, 524270 runs, 18 skips 1864 decicycles in idct_idct_4x4_add_c, 1048548 runs, 28 skips 529 decicycles in ff_vp9_idct_idct_4x4_add_ssse3, 262138 runs, 6 skips 516 decicycles in ff_vp9_idct_idct_4x4_add_ssse3, 524282 runs, 6 skips 474 decicycles in ff_vp9_idct_idct_4x4_add_ssse3, 1048565 runs, 11 skips (~3.9x faster) 7726 decicycles in idct_idct_8x8_add_c, 1048433 runs, 143 skips 7732 decicycles in idct_idct_8x8_add_c, 2096882 runs, 270 skips 7731 decicycles in idct_idct_8x8_add_c, 4193772 runs, 532 skips 1145 decicycles in ff_vp9_idct_idct_8x8_add_ssse3, 1048549 runs, 27 skips 1137 decicycles in ff_vp9_idct_idct_8x8_add_ssse3, 2097097 runs, 55 skips 1086 decicycles in ff_vp9_idct_idct_8x8_add_ssse3, 4194188 runs, 116 skips (~7.1x faster) Overall decode time before commit: 16.48s user 0.03s system 99% cpu 16.526 total 16.54s user 0.01s system 99% cpu 16.566 total 16.46s user 0.03s system 99% cpu 16.511 total Overall decode time after commit: 16.34s user 0.02s system 99% cpu 16.378 total 16.28s user 0.02s system 99% cpu 16.315 total 16.32s user 0.03s system 99% cpu 16.366 total Tested on i7 920 with 40s 1080p footage.
11 years ago
%endmacro
INIT_XMM ssse3
cglobal vp9_idct_idct_8x8_add, 4,4,13, dst, stride, block, eob
mova m12, [pw_11585x2] ; often used
vp9: implement top/left half (4x4) sub-8x8-IDCT. For that specific case (eob>3&&eob<=12), runtime of idct8x8 goes from 668 to 477 cycles. For all idct8x8, runtime goes from 521 to 490 cycles.
11 years ago
cmp eobd, 12 ; top left half or less
avcodec/vp9: add ff_vp9_idct_idct_{4x4,8x8}_ssse3(). 1789 decicycles in idct_idct_4x4_add_c, 262136 runs, 8 skips 1839 decicycles in idct_idct_4x4_add_c, 524270 runs, 18 skips 1864 decicycles in idct_idct_4x4_add_c, 1048548 runs, 28 skips 529 decicycles in ff_vp9_idct_idct_4x4_add_ssse3, 262138 runs, 6 skips 516 decicycles in ff_vp9_idct_idct_4x4_add_ssse3, 524282 runs, 6 skips 474 decicycles in ff_vp9_idct_idct_4x4_add_ssse3, 1048565 runs, 11 skips (~3.9x faster) 7726 decicycles in idct_idct_8x8_add_c, 1048433 runs, 143 skips 7732 decicycles in idct_idct_8x8_add_c, 2096882 runs, 270 skips 7731 decicycles in idct_idct_8x8_add_c, 4193772 runs, 532 skips 1145 decicycles in ff_vp9_idct_idct_8x8_add_ssse3, 1048549 runs, 27 skips 1137 decicycles in ff_vp9_idct_idct_8x8_add_ssse3, 2097097 runs, 55 skips 1086 decicycles in ff_vp9_idct_idct_8x8_add_ssse3, 4194188 runs, 116 skips (~7.1x faster) Overall decode time before commit: 16.48s user 0.03s system 99% cpu 16.526 total 16.54s user 0.01s system 99% cpu 16.566 total 16.46s user 0.03s system 99% cpu 16.511 total Overall decode time after commit: 16.34s user 0.02s system 99% cpu 16.378 total 16.28s user 0.02s system 99% cpu 16.315 total 16.32s user 0.03s system 99% cpu 16.366 total Tested on i7 920 with 40s 1080p footage.
11 years ago
jg .idctfull
vp9: implement top/left half (4x4) sub-8x8-IDCT. For that specific case (eob>3&&eob<=12), runtime of idct8x8 goes from 668 to 477 cycles. For all idct8x8, runtime goes from 521 to 490 cycles.
11 years ago
cmp eobd, 3 ; top left corner or less
jg .idcthalf
avcodec/vp9: add ff_vp9_idct_idct_{4x4,8x8}_ssse3(). 1789 decicycles in idct_idct_4x4_add_c, 262136 runs, 8 skips 1839 decicycles in idct_idct_4x4_add_c, 524270 runs, 18 skips 1864 decicycles in idct_idct_4x4_add_c, 1048548 runs, 28 skips 529 decicycles in ff_vp9_idct_idct_4x4_add_ssse3, 262138 runs, 6 skips 516 decicycles in ff_vp9_idct_idct_4x4_add_ssse3, 524282 runs, 6 skips 474 decicycles in ff_vp9_idct_idct_4x4_add_ssse3, 1048565 runs, 11 skips (~3.9x faster) 7726 decicycles in idct_idct_8x8_add_c, 1048433 runs, 143 skips 7732 decicycles in idct_idct_8x8_add_c, 2096882 runs, 270 skips 7731 decicycles in idct_idct_8x8_add_c, 4193772 runs, 532 skips 1145 decicycles in ff_vp9_idct_idct_8x8_add_ssse3, 1048549 runs, 27 skips 1137 decicycles in ff_vp9_idct_idct_8x8_add_ssse3, 2097097 runs, 55 skips 1086 decicycles in ff_vp9_idct_idct_8x8_add_ssse3, 4194188 runs, 116 skips (~7.1x faster) Overall decode time before commit: 16.48s user 0.03s system 99% cpu 16.526 total 16.54s user 0.01s system 99% cpu 16.566 total 16.46s user 0.03s system 99% cpu 16.511 total Overall decode time after commit: 16.34s user 0.02s system 99% cpu 16.378 total 16.28s user 0.02s system 99% cpu 16.315 total 16.32s user 0.03s system 99% cpu 16.366 total Tested on i7 920 with 40s 1080p footage.
11 years ago
cmp eobd, 1 ; faster path for when only DC is set
jne .idcttopleftcorner
movd m0, [blockq]
pmulhrsw m0, m12
pmulhrsw m0, m12
SPLATW m0, m0, 0
pxor m4, m4
movd [blockq], m4
mova m5, [pw_1024]
pmulhrsw m0, m5 ; (x*1024 + (1<<14))>>15 <=> (x+16)>>5
vp9: minor refactorings in idct ssse3 assembly. Make register usage in macros explicit; change mulsub_2w_4x to use 2 instead of 3 temp registers.
11 years ago
VP9_STORE_2X 0, 0, 6, 7, 4
avcodec/vp9: add ff_vp9_idct_idct_{4x4,8x8}_ssse3(). 1789 decicycles in idct_idct_4x4_add_c, 262136 runs, 8 skips 1839 decicycles in idct_idct_4x4_add_c, 524270 runs, 18 skips 1864 decicycles in idct_idct_4x4_add_c, 1048548 runs, 28 skips 529 decicycles in ff_vp9_idct_idct_4x4_add_ssse3, 262138 runs, 6 skips 516 decicycles in ff_vp9_idct_idct_4x4_add_ssse3, 524282 runs, 6 skips 474 decicycles in ff_vp9_idct_idct_4x4_add_ssse3, 1048565 runs, 11 skips (~3.9x faster) 7726 decicycles in idct_idct_8x8_add_c, 1048433 runs, 143 skips 7732 decicycles in idct_idct_8x8_add_c, 2096882 runs, 270 skips 7731 decicycles in idct_idct_8x8_add_c, 4193772 runs, 532 skips 1145 decicycles in ff_vp9_idct_idct_8x8_add_ssse3, 1048549 runs, 27 skips 1137 decicycles in ff_vp9_idct_idct_8x8_add_ssse3, 2097097 runs, 55 skips 1086 decicycles in ff_vp9_idct_idct_8x8_add_ssse3, 4194188 runs, 116 skips (~7.1x faster) Overall decode time before commit: 16.48s user 0.03s system 99% cpu 16.526 total 16.54s user 0.01s system 99% cpu 16.566 total 16.46s user 0.03s system 99% cpu 16.511 total Overall decode time after commit: 16.34s user 0.02s system 99% cpu 16.378 total 16.28s user 0.02s system 99% cpu 16.315 total 16.32s user 0.03s system 99% cpu 16.366 total Tested on i7 920 with 40s 1080p footage.
11 years ago
lea dstq, [dstq+2*strideq]
vp9: minor refactorings in idct ssse3 assembly. Make register usage in macros explicit; change mulsub_2w_4x to use 2 instead of 3 temp registers.
11 years ago
VP9_STORE_2X 0, 0, 6, 7, 4
avcodec/vp9: add ff_vp9_idct_idct_{4x4,8x8}_ssse3(). 1789 decicycles in idct_idct_4x4_add_c, 262136 runs, 8 skips 1839 decicycles in idct_idct_4x4_add_c, 524270 runs, 18 skips 1864 decicycles in idct_idct_4x4_add_c, 1048548 runs, 28 skips 529 decicycles in ff_vp9_idct_idct_4x4_add_ssse3, 262138 runs, 6 skips 516 decicycles in ff_vp9_idct_idct_4x4_add_ssse3, 524282 runs, 6 skips 474 decicycles in ff_vp9_idct_idct_4x4_add_ssse3, 1048565 runs, 11 skips (~3.9x faster) 7726 decicycles in idct_idct_8x8_add_c, 1048433 runs, 143 skips 7732 decicycles in idct_idct_8x8_add_c, 2096882 runs, 270 skips 7731 decicycles in idct_idct_8x8_add_c, 4193772 runs, 532 skips 1145 decicycles in ff_vp9_idct_idct_8x8_add_ssse3, 1048549 runs, 27 skips 1137 decicycles in ff_vp9_idct_idct_8x8_add_ssse3, 2097097 runs, 55 skips 1086 decicycles in ff_vp9_idct_idct_8x8_add_ssse3, 4194188 runs, 116 skips (~7.1x faster) Overall decode time before commit: 16.48s user 0.03s system 99% cpu 16.526 total 16.54s user 0.01s system 99% cpu 16.566 total 16.46s user 0.03s system 99% cpu 16.511 total Overall decode time after commit: 16.34s user 0.02s system 99% cpu 16.378 total 16.28s user 0.02s system 99% cpu 16.315 total 16.32s user 0.03s system 99% cpu 16.366 total Tested on i7 920 with 40s 1080p footage.
11 years ago
lea dstq, [dstq+2*strideq]
vp9: minor refactorings in idct ssse3 assembly. Make register usage in macros explicit; change mulsub_2w_4x to use 2 instead of 3 temp registers.
11 years ago
VP9_STORE_2X 0, 0, 6, 7, 4
avcodec/vp9: add ff_vp9_idct_idct_{4x4,8x8}_ssse3(). 1789 decicycles in idct_idct_4x4_add_c, 262136 runs, 8 skips 1839 decicycles in idct_idct_4x4_add_c, 524270 runs, 18 skips 1864 decicycles in idct_idct_4x4_add_c, 1048548 runs, 28 skips 529 decicycles in ff_vp9_idct_idct_4x4_add_ssse3, 262138 runs, 6 skips 516 decicycles in ff_vp9_idct_idct_4x4_add_ssse3, 524282 runs, 6 skips 474 decicycles in ff_vp9_idct_idct_4x4_add_ssse3, 1048565 runs, 11 skips (~3.9x faster) 7726 decicycles in idct_idct_8x8_add_c, 1048433 runs, 143 skips 7732 decicycles in idct_idct_8x8_add_c, 2096882 runs, 270 skips 7731 decicycles in idct_idct_8x8_add_c, 4193772 runs, 532 skips 1145 decicycles in ff_vp9_idct_idct_8x8_add_ssse3, 1048549 runs, 27 skips 1137 decicycles in ff_vp9_idct_idct_8x8_add_ssse3, 2097097 runs, 55 skips 1086 decicycles in ff_vp9_idct_idct_8x8_add_ssse3, 4194188 runs, 116 skips (~7.1x faster) Overall decode time before commit: 16.48s user 0.03s system 99% cpu 16.526 total 16.54s user 0.01s system 99% cpu 16.566 total 16.46s user 0.03s system 99% cpu 16.511 total Overall decode time after commit: 16.34s user 0.02s system 99% cpu 16.378 total 16.28s user 0.02s system 99% cpu 16.315 total 16.32s user 0.03s system 99% cpu 16.366 total Tested on i7 920 with 40s 1080p footage.
11 years ago
lea dstq, [dstq+2*strideq]
vp9: minor refactorings in idct ssse3 assembly. Make register usage in macros explicit; change mulsub_2w_4x to use 2 instead of 3 temp registers.
11 years ago
VP9_STORE_2X 0, 0, 6, 7, 4
avcodec/vp9: add ff_vp9_idct_idct_{4x4,8x8}_ssse3(). 1789 decicycles in idct_idct_4x4_add_c, 262136 runs, 8 skips 1839 decicycles in idct_idct_4x4_add_c, 524270 runs, 18 skips 1864 decicycles in idct_idct_4x4_add_c, 1048548 runs, 28 skips 529 decicycles in ff_vp9_idct_idct_4x4_add_ssse3, 262138 runs, 6 skips 516 decicycles in ff_vp9_idct_idct_4x4_add_ssse3, 524282 runs, 6 skips 474 decicycles in ff_vp9_idct_idct_4x4_add_ssse3, 1048565 runs, 11 skips (~3.9x faster) 7726 decicycles in idct_idct_8x8_add_c, 1048433 runs, 143 skips 7732 decicycles in idct_idct_8x8_add_c, 2096882 runs, 270 skips 7731 decicycles in idct_idct_8x8_add_c, 4193772 runs, 532 skips 1145 decicycles in ff_vp9_idct_idct_8x8_add_ssse3, 1048549 runs, 27 skips 1137 decicycles in ff_vp9_idct_idct_8x8_add_ssse3, 2097097 runs, 55 skips 1086 decicycles in ff_vp9_idct_idct_8x8_add_ssse3, 4194188 runs, 116 skips (~7.1x faster) Overall decode time before commit: 16.48s user 0.03s system 99% cpu 16.526 total 16.54s user 0.01s system 99% cpu 16.566 total 16.46s user 0.03s system 99% cpu 16.511 total Overall decode time after commit: 16.34s user 0.02s system 99% cpu 16.378 total 16.28s user 0.02s system 99% cpu 16.315 total 16.32s user 0.03s system 99% cpu 16.366 total Tested on i7 920 with 40s 1080p footage.
11 years ago
RET
; faster path for when only left corner is set (3 input: DC, right to DC, below
; to DC). Note: also working with a 2x2 block
.idcttopleftcorner:
movd m0, [blockq+0]
movd m1, [blockq+16]
mova m6, [pw_3196x2]
mova m7, [pw_16069x2]
VP9_IDCT8_2x2_1D
TRANSPOSE8x8W 0, 1, 2, 3, 8, 9, 10, 11, 4
VP9_IDCT8_2x2_1D
pxor m4, m4 ; used for the block reset, and VP9_STORE_2X
movd [blockq+ 0], m4
movd [blockq+16], m4
VP9_IDCT8_WRITEOUT
RET
vp9: implement top/left half (4x4) sub-8x8-IDCT. For that specific case (eob>3&&eob<=12), runtime of idct8x8 goes from 668 to 477 cycles. For all idct8x8, runtime goes from 521 to 490 cycles.
11 years ago
.idcthalf:
movh m0, [blockq + 0]
movh m1, [blockq +16]
movh m2, [blockq +32]
movh m3, [blockq +48]
VP9_IDCT8_4x4_1D
TRANSPOSE8x8W 0, 1, 2, 3, 8, 9, 10, 11, 4
VP9_IDCT8_4x4_1D
pxor m4, m4
movh [blockq+ 0], m4
movh [blockq+16], m4
movh [blockq+32], m4
movh [blockq+48], m4
VP9_IDCT8_WRITEOUT
RET
avcodec/vp9: add ff_vp9_idct_idct_{4x4,8x8}_ssse3(). 1789 decicycles in idct_idct_4x4_add_c, 262136 runs, 8 skips 1839 decicycles in idct_idct_4x4_add_c, 524270 runs, 18 skips 1864 decicycles in idct_idct_4x4_add_c, 1048548 runs, 28 skips 529 decicycles in ff_vp9_idct_idct_4x4_add_ssse3, 262138 runs, 6 skips 516 decicycles in ff_vp9_idct_idct_4x4_add_ssse3, 524282 runs, 6 skips 474 decicycles in ff_vp9_idct_idct_4x4_add_ssse3, 1048565 runs, 11 skips (~3.9x faster) 7726 decicycles in idct_idct_8x8_add_c, 1048433 runs, 143 skips 7732 decicycles in idct_idct_8x8_add_c, 2096882 runs, 270 skips 7731 decicycles in idct_idct_8x8_add_c, 4193772 runs, 532 skips 1145 decicycles in ff_vp9_idct_idct_8x8_add_ssse3, 1048549 runs, 27 skips 1137 decicycles in ff_vp9_idct_idct_8x8_add_ssse3, 2097097 runs, 55 skips 1086 decicycles in ff_vp9_idct_idct_8x8_add_ssse3, 4194188 runs, 116 skips (~7.1x faster) Overall decode time before commit: 16.48s user 0.03s system 99% cpu 16.526 total 16.54s user 0.01s system 99% cpu 16.566 total 16.46s user 0.03s system 99% cpu 16.511 total Overall decode time after commit: 16.34s user 0.02s system 99% cpu 16.378 total 16.28s user 0.02s system 99% cpu 16.315 total 16.32s user 0.03s system 99% cpu 16.366 total Tested on i7 920 with 40s 1080p footage.
11 years ago
.idctfull: ; generic full 8x8 idct/idct
mova m0, [blockq+ 0] ; IN(0)
mova m1, [blockq+ 16] ; IN(1)
mova m2, [blockq+ 32] ; IN(2)
mova m3, [blockq+ 48] ; IN(3)
mova m8, [blockq+ 64] ; IN(4)
mova m9, [blockq+ 80] ; IN(5)
mova m10, [blockq+ 96] ; IN(6)
mova m11, [blockq+112] ; IN(7)
mova m7, [pd_8192] ; rounding
VP9_IDCT8_1D
TRANSPOSE8x8W 0, 1, 2, 3, 8, 9, 10, 11, 4
VP9_IDCT8_1D
pxor m4, m4 ; used for the block reset, and VP9_STORE_2X
mova [blockq+ 0], m4
mova [blockq+ 16], m4
mova [blockq+ 32], m4
mova [blockq+ 48], m4
mova [blockq+ 64], m4
mova [blockq+ 80], m4
mova [blockq+ 96], m4
mova [blockq+112], m4
VP9_IDCT8_WRITEOUT
RET
vp9/x86: idct_add_16x16_ssse3. Currently only dc-only and full 16x16. Other subforms will follow in the near future. Total decoding time of ped1080p.webm goes from 9.7 to 9.3 seconds. DC-only goes from 957 -> 131 cycles, and the full IDCT goes from ~4050 to ~745 cycles.
11 years ago
;---------------------------------------------------------------------------------------------
; void vp9_idct_idct_16x16_add_<opt>(uint8_t *dst, ptrdiff_t stride, int16_t *block, int eob);
;---------------------------------------------------------------------------------------------
vp9/x86: idct_32x32_add_ssse3. Sub-IDCTs will follow later. ped1080.webm goes from 9.295s to 8.191s (13.5% faster). The IDCT itself goes from 4372 (intra) or 4337 (inter) to 403 (intra) or 329 (inter) cycles for the DC-only form, 23755 (intra) or 23723 (inter) to 3497 (intra) or 3607 (inter) cycles for the no-DC form, which averages from 23393 (intra) or 16612 (inter) to 3449 (intra) or 2392 (inter) for all 32x32s together, i.e. about ~7x faster (all tests done on ped1080p.webm).
11 years ago
; at the end of this macro, m7 is stored in stack_scratch
; everything else (t0-6 and t8-15) is stored in m0-6 and m8-15
; the following sumsubs have not been done yet:
; SUMSUB_BA w, 6, 9, 15 ; t6, t9
; SUMSUB_BA w, 7, 8, 15 ; t7, t8
%macro VP9_IDCT16_1D_START 4 ; src, nnzc, stride, stack_scratch
vp9/x86: idct_32x32_add_ssse3 sub-8x8-idct. Runtime of the full 32x32 idct goes from 2446 to 2441 cycles (intra) or from 1425 to 1306 cycles (inter). Overall runtime is not significantly affected.
11 years ago
%if %2 <= 4
mova m3, [%1+ 1*%3] ; IN(1)
mova m12, [%1+ 2*%3] ; IN(2)
mova m0, [%1+ 3*%3] ; IN(3)
pmulhrsw m15, m12, [pw_16069x2] ; t6-7
pmulhrsw m12, [pw_3196x2] ; t4-5
pmulhrsw m4, m3, [pw_16305x2] ; t14-15
pmulhrsw m3, [pw_1606x2] ; t8-9
pmulhrsw m7, m0, [pw_m4756x2] ; t10-11
pmulhrsw m0, [pw_15679x2] ; t12-13
; m8=t0, m9=t1, m10=t2, m11=t3, m12=t4, m14=t5, m13=t6, m15=t7
; m3=t8, m5=t9, m1=t10, m7=t11, m0=t12, m6=t13, m2=t14, m4=t15
paddw m14, m15, m12
psubw m13, m15, m12
pmulhrsw m13, [pw_11585x2] ; t5
pmulhrsw m14, [pw_11585x2] ; t6
VP9_UNPACK_MULSUB_2W_4X 2, 5, 4, 3, 15137, 6270, [pd_8192], 10, 11 ; t9, t14
VP9_UNPACK_MULSUB_2W_4X 6, 1, 0, 7, 6270, m15137, [pd_8192], 10, 11 ; t10, t13
; m15=t0, m14=t1, m13=t2, m12=t3, m11=t4, m10=t5, m9=t6, m8=t7
; m7=t8, m6=t9, m2=t10, m3=t11, m4=t12, m5=t13, m1=t14, m0=t15
%else
vp9/x86: idct_32x32_add_ssse3. Sub-IDCTs will follow later. ped1080.webm goes from 9.295s to 8.191s (13.5% faster). The IDCT itself goes from 4372 (intra) or 4337 (inter) to 403 (intra) or 329 (inter) cycles for the DC-only form, 23755 (intra) or 23723 (inter) to 3497 (intra) or 3607 (inter) cycles for the no-DC form, which averages from 23393 (intra) or 16612 (inter) to 3449 (intra) or 2392 (inter) for all 32x32s together, i.e. about ~7x faster (all tests done on ped1080p.webm).
11 years ago
mova m5, [%1+ 1*%3] ; IN(1)
mova m14, [%1+ 2*%3] ; IN(2)
mova m6, [%1+ 3*%3] ; IN(3)
mova m9, [%1+ 4*%3] ; IN(4)
mova m7, [%1+ 5*%3] ; IN(5)
mova m15, [%1+ 6*%3] ; IN(6)
mova m4, [%1+ 7*%3] ; IN(7)
%if %2 <= 8
vp9/x86: 16x16 sub-IDCT for top-left 8x8 subblock (eob <= 38). Sub8x8 speed (w/o dc-only case) goes from ~750 cycles (inter) or ~735 cycles (intra) to ~415 cycles (inter) or ~430 cycles (intra). Average overall 16x16 idct speed goes from ~635 cycles (inter) or ~720 cycles (intra) to ~415 cycles (inter) or ~545 (intra) - all measurements done using ped1080p.webm.
11 years ago
pmulhrsw m8, m9, [pw_15137x2] ; t3
pmulhrsw m9, [pw_6270x2] ; t2
pmulhrsw m13, m14, [pw_16069x2] ; t7
pmulhrsw m14, [pw_3196x2] ; t4
pmulhrsw m12, m15, [pw_m9102x2] ; t5
pmulhrsw m15, [pw_13623x2] ; t6
pmulhrsw m2, m5, [pw_16305x2] ; t15
pmulhrsw m5, [pw_1606x2] ; t8
pmulhrsw m3, m4, [pw_m10394x2] ; t9
pmulhrsw m4, [pw_12665x2] ; t14
pmulhrsw m0, m7, [pw_14449x2] ; t13
pmulhrsw m7, [pw_7723x2] ; t10
pmulhrsw m1, m6, [pw_m4756x2] ; t11
pmulhrsw m6, [pw_15679x2] ; t12
%else
vp9/x86: idct_32x32_add_ssse3. Sub-IDCTs will follow later. ped1080.webm goes from 9.295s to 8.191s (13.5% faster). The IDCT itself goes from 4372 (intra) or 4337 (inter) to 403 (intra) or 329 (inter) cycles for the DC-only form, 23755 (intra) or 23723 (inter) to 3497 (intra) or 3607 (inter) cycles for the no-DC form, which averages from 23393 (intra) or 16612 (inter) to 3449 (intra) or 2392 (inter) for all 32x32s together, i.e. about ~7x faster (all tests done on ped1080p.webm).
11 years ago
mova m3, [%1+ 9*%3] ; IN(9)
mova m12, [%1+10*%3] ; IN(10)
mova m0, [%1+11*%3] ; IN(11)
mova m8, [%1+12*%3] ; IN(12)
mova m1, [%1+13*%3] ; IN(13)
mova m13, [%1+14*%3] ; IN(14)
mova m2, [%1+15*%3] ; IN(15)
vp9/x86: idct_add_16x16_ssse3. Currently only dc-only and full 16x16. Other subforms will follow in the near future. Total decoding time of ped1080p.webm goes from 9.7 to 9.3 seconds. DC-only goes from 957 -> 131 cycles, and the full IDCT goes from ~4050 to ~745 cycles.
11 years ago
; m10=in0, m5=in1, m14=in2, m6=in3, m9=in4, m7=in5, m15=in6, m4=in7
; m11=in8, m3=in9, m12=in10 m0=in11, m8=in12, m1=in13, m13=in14, m2=in15
VP9_UNPACK_MULSUB_2W_4X 9, 8, 15137, 6270, [pd_8192], 10, 11 ; t2, t3
VP9_UNPACK_MULSUB_2W_4X 14, 13, 16069, 3196, [pd_8192], 10, 11 ; t4, t7
VP9_UNPACK_MULSUB_2W_4X 12, 15, 9102, 13623, [pd_8192], 10, 11 ; t5, t6
VP9_UNPACK_MULSUB_2W_4X 5, 2, 16305, 1606, [pd_8192], 10, 11 ; t8, t15
VP9_UNPACK_MULSUB_2W_4X 3, 4, 10394, 12665, [pd_8192], 10, 11 ; t9, t14
VP9_UNPACK_MULSUB_2W_4X 7, 0, 14449, 7723, [pd_8192], 10, 11 ; t10, t13
VP9_UNPACK_MULSUB_2W_4X 1, 6, 4756, 15679, [pd_8192], 10, 11 ; t11, t12
vp9/x86: 16x16 sub-IDCT for top-left 8x8 subblock (eob <= 38). Sub8x8 speed (w/o dc-only case) goes from ~750 cycles (inter) or ~735 cycles (intra) to ~415 cycles (inter) or ~430 cycles (intra). Average overall 16x16 idct speed goes from ~635 cycles (inter) or ~720 cycles (intra) to ~415 cycles (inter) or ~545 (intra) - all measurements done using ped1080p.webm.
11 years ago
%endif
vp9/x86: idct_add_16x16_ssse3. Currently only dc-only and full 16x16. Other subforms will follow in the near future. Total decoding time of ped1080p.webm goes from 9.7 to 9.3 seconds. DC-only goes from 957 -> 131 cycles, and the full IDCT goes from ~4050 to ~745 cycles.
11 years ago
; m11=t0, m10=t1, m9=t2, m8=t3, m14=t4, m12=t5, m15=t6, m13=t7
; m5=t8, m3=t9, m7=t10, m1=t11, m6=t12, m0=t13, m4=t14, m2=t15
SUMSUB_BA w, 12, 14, 10 ; t4, t5
SUMSUB_BA w, 15, 13, 10 ; t7, t6
SUMSUB_BA w, 3, 5, 10 ; t8, t9
SUMSUB_BA w, 7, 1, 10 ; t11, t10
SUMSUB_BA w, 0, 6, 10 ; t12, t13
SUMSUB_BA w, 4, 2, 10 ; t15, t14
; m8=t0, m9=t1, m10=t2, m11=t3, m12=t4, m14=t5, m13=t6, m15=t7
; m3=t8, m5=t9, m1=t10, m7=t11, m0=t12, m6=t13, m2=t14, m4=t15
SUMSUB_BA w, 14, 13, 10
pmulhrsw m13, [pw_11585x2] ; t5
pmulhrsw m14, [pw_11585x2] ; t6
VP9_UNPACK_MULSUB_2W_4X 2, 5, 15137, 6270, [pd_8192], 10, 11 ; t9, t14
VP9_UNPACK_MULSUB_2W_4X 6, 1, 6270, m15137, [pd_8192], 10, 11 ; t10, t13
vp9/x86: idct_32x32_add_ssse3 sub-8x8-idct. Runtime of the full 32x32 idct goes from 2446 to 2441 cycles (intra) or from 1425 to 1306 cycles (inter). Overall runtime is not significantly affected.
11 years ago
%endif
vp9/x86: idct_add_16x16_ssse3. Currently only dc-only and full 16x16. Other subforms will follow in the near future. Total decoding time of ped1080p.webm goes from 9.7 to 9.3 seconds. DC-only goes from 957 -> 131 cycles, and the full IDCT goes from ~4050 to ~745 cycles.
11 years ago
; m8=t0, m9=t1, m10=t2, m11=t3, m12=t4, m13=t5, m14=t6, m15=t7
; m3=t8, m2=t9, m6=t10, m7=t11, m0=t12, m1=t13, m5=t14, m4=t15
SUMSUB_BA w, 7, 3, 10 ; t8, t11
SUMSUB_BA w, 6, 2, 10 ; t9, t10
SUMSUB_BA w, 0, 4, 10 ; t15, t12
SUMSUB_BA w, 1, 5, 10 ; t14. t13
; m15=t0, m14=t1, m13=t2, m12=t3, m11=t4, m10=t5, m9=t6, m8=t7
; m7=t8, m6=t9, m2=t10, m3=t11, m4=t12, m5=t13, m1=t14, m0=t15
SUMSUB_BA w, 2, 5, 10
SUMSUB_BA w, 3, 4, 10
pmulhrsw m5, [pw_11585x2] ; t10
pmulhrsw m4, [pw_11585x2] ; t11
pmulhrsw m3, [pw_11585x2] ; t12
pmulhrsw m2, [pw_11585x2] ; t13
; backup first register
vp9/x86: idct_32x32_add_ssse3. Sub-IDCTs will follow later. ped1080.webm goes from 9.295s to 8.191s (13.5% faster). The IDCT itself goes from 4372 (intra) or 4337 (inter) to 403 (intra) or 329 (inter) cycles for the DC-only form, 23755 (intra) or 23723 (inter) to 3497 (intra) or 3607 (inter) cycles for the no-DC form, which averages from 23393 (intra) or 16612 (inter) to 3449 (intra) or 2392 (inter) for all 32x32s together, i.e. about ~7x faster (all tests done on ped1080p.webm).
11 years ago
mova [%4], m7
vp9/x86: idct_add_16x16_ssse3. Currently only dc-only and full 16x16. Other subforms will follow in the near future. Total decoding time of ped1080p.webm goes from 9.7 to 9.3 seconds. DC-only goes from 957 -> 131 cycles, and the full IDCT goes from ~4050 to ~745 cycles.
11 years ago
; m15=t0, m14=t1, m13=t2, m12=t3, m11=t4, m10=t5, m9=t6, m8=t7
; m7=t8, m6=t9, m5=t10, m4=t11, m3=t12, m2=t13, m1=t14, m0=t15
; from load/start
vp9/x86: idct_32x32_add_ssse3 sub-8x8-idct. Runtime of the full 32x32 idct goes from 2446 to 2441 cycles (intra) or from 1425 to 1306 cycles (inter). Overall runtime is not significantly affected.
11 years ago
%if %2 <= 4
mova m11, [%1+ 0*%3] ; IN(0)
pmulhrsw m11, [pw_11585x2] ; t0-t3
psubw m8, m11, m15
paddw m15, m11
psubw m9, m11, m14
paddw m14, m11
psubw m10, m11, m13
paddw m13, m11
%else
vp9/x86: idct_32x32_add_ssse3. Sub-IDCTs will follow later. ped1080.webm goes from 9.295s to 8.191s (13.5% faster). The IDCT itself goes from 4372 (intra) or 4337 (inter) to 403 (intra) or 329 (inter) cycles for the DC-only form, 23755 (intra) or 23723 (inter) to 3497 (intra) or 3607 (inter) cycles for the no-DC form, which averages from 23393 (intra) or 16612 (inter) to 3449 (intra) or 2392 (inter) for all 32x32s together, i.e. about ~7x faster (all tests done on ped1080p.webm).
11 years ago
mova m10, [%1+ 0*%3] ; IN(0)
%if %2 <= 8
vp9/x86: 16x16 sub-IDCT for top-left 8x8 subblock (eob <= 38). Sub8x8 speed (w/o dc-only case) goes from ~750 cycles (inter) or ~735 cycles (intra) to ~415 cycles (inter) or ~430 cycles (intra). Average overall 16x16 idct speed goes from ~635 cycles (inter) or ~720 cycles (intra) to ~415 cycles (inter) or ~545 (intra) - all measurements done using ped1080p.webm.
11 years ago
pmulhrsw m10, [pw_11585x2] ; t0 and t1
psubw m11, m10, m8
paddw m8, m10
%else
vp9/x86: idct_32x32_add_ssse3. Sub-IDCTs will follow later. ped1080.webm goes from 9.295s to 8.191s (13.5% faster). The IDCT itself goes from 4372 (intra) or 4337 (inter) to 403 (intra) or 329 (inter) cycles for the DC-only form, 23755 (intra) or 23723 (inter) to 3497 (intra) or 3607 (inter) cycles for the no-DC form, which averages from 23393 (intra) or 16612 (inter) to 3449 (intra) or 2392 (inter) for all 32x32s together, i.e. about ~7x faster (all tests done on ped1080p.webm).
11 years ago
mova m11, [%1+ 8*%3] ; IN(8)
vp9/x86: idct_add_16x16_ssse3. Currently only dc-only and full 16x16. Other subforms will follow in the near future. Total decoding time of ped1080p.webm goes from 9.7 to 9.3 seconds. DC-only goes from 957 -> 131 cycles, and the full IDCT goes from ~4050 to ~745 cycles.
11 years ago
; from 3 stages back
SUMSUB_BA w, 11, 10, 7
pmulhrsw m11, [pw_11585x2] ; t0
pmulhrsw m10, [pw_11585x2] ; t1
; from 2 stages back
SUMSUB_BA w, 8, 11, 7 ; t0, t3
vp9/x86: 16x16 sub-IDCT for top-left 8x8 subblock (eob <= 38). Sub8x8 speed (w/o dc-only case) goes from ~750 cycles (inter) or ~735 cycles (intra) to ~415 cycles (inter) or ~430 cycles (intra). Average overall 16x16 idct speed goes from ~635 cycles (inter) or ~720 cycles (intra) to ~415 cycles (inter) or ~545 (intra) - all measurements done using ped1080p.webm.
11 years ago
%endif
vp9/x86: idct_add_16x16_ssse3. Currently only dc-only and full 16x16. Other subforms will follow in the near future. Total decoding time of ped1080p.webm goes from 9.7 to 9.3 seconds. DC-only goes from 957 -> 131 cycles, and the full IDCT goes from ~4050 to ~745 cycles.
11 years ago
SUMSUB_BA w, 9, 10, 7 ; t1, t2
; from 1 stage back
SUMSUB_BA w, 15, 8, 7 ; t0, t7
SUMSUB_BA w, 14, 9, 7 ; t1, t6
SUMSUB_BA w, 13, 10, 7 ; t2, t5
vp9/x86: idct_32x32_add_ssse3 sub-8x8-idct. Runtime of the full 32x32 idct goes from 2446 to 2441 cycles (intra) or from 1425 to 1306 cycles (inter). Overall runtime is not significantly affected.
11 years ago
%endif
vp9/x86: idct_add_16x16_ssse3. Currently only dc-only and full 16x16. Other subforms will follow in the near future. Total decoding time of ped1080p.webm goes from 9.7 to 9.3 seconds. DC-only goes from 957 -> 131 cycles, and the full IDCT goes from ~4050 to ~745 cycles.
11 years ago
SUMSUB_BA w, 12, 11, 7 ; t3, t4
SUMSUB_BA w, 0, 15, 7 ; t0, t15
SUMSUB_BA w, 1, 14, 7 ; t1, t14
SUMSUB_BA w, 2, 13, 7 ; t2, t13
SUMSUB_BA w, 3, 12, 7 ; t3, t12
SUMSUB_BA w, 4, 11, 7 ; t4, t11
SUMSUB_BA w, 5, 10, 7 ; t5, t10
vp9/x86: idct_32x32_add_ssse3. Sub-IDCTs will follow later. ped1080.webm goes from 9.295s to 8.191s (13.5% faster). The IDCT itself goes from 4372 (intra) or 4337 (inter) to 403 (intra) or 329 (inter) cycles for the DC-only form, 23755 (intra) or 23723 (inter) to 3497 (intra) or 3607 (inter) cycles for the no-DC form, which averages from 23393 (intra) or 16612 (inter) to 3449 (intra) or 2392 (inter) for all 32x32s together, i.e. about ~7x faster (all tests done on ped1080p.webm).
11 years ago
%endmacro
%macro VP9_IDCT16_1D 2-3 16 ; src, pass, nnzc
VP9_IDCT16_1D_START %1, %3, 32, rsp+32
vp9/x86: idct_add_16x16_ssse3. Currently only dc-only and full 16x16. Other subforms will follow in the near future. Total decoding time of ped1080p.webm goes from 9.7 to 9.3 seconds. DC-only goes from 957 -> 131 cycles, and the full IDCT goes from ~4050 to ~745 cycles.
11 years ago
%if %2 == 1
; backup a different register
mova [rsp+16], m15
mova m7, [rsp+32]
SUMSUB_BA w, 6, 9, 15 ; t6, t9
SUMSUB_BA w, 7, 8, 15 ; t7, t8
TRANSPOSE8x8W 0, 1, 2, 3, 4, 5, 6, 7, 15
mova [rsp+ 0], m0
mova [rsp+ 32], m1
mova [rsp+ 64], m2
mova [rsp+ 96], m3
mova [rsp+128], m4
mova [rsp+160], m5
mova [rsp+192], m6
mova [rsp+224], m7
mova m15, [rsp+16]
TRANSPOSE8x8W 8, 9, 10, 11, 12, 13, 14, 15, 0
mova [rsp+ 16], m8
mova [rsp+ 48], m9
mova [rsp+ 80], m10
mova [rsp+112], m11
mova [rsp+144], m12
mova [rsp+176], m13
mova [rsp+208], m14
mova [rsp+240], m15
%else ; %2 == 2
; backup more registers
mova [rsp+64], m8
mova [rsp+96], m9
pxor m7, m7
pmulhrsw m0, [pw_512]
pmulhrsw m1, [pw_512]
VP9_STORE_2X 0, 1, 8, 9, 7
lea dstq, [dstq+strideq*2]
pmulhrsw m2, [pw_512]
pmulhrsw m3, [pw_512]
VP9_STORE_2X 2, 3, 8, 9, 7
lea dstq, [dstq+strideq*2]
pmulhrsw m4, [pw_512]
pmulhrsw m5, [pw_512]
VP9_STORE_2X 4, 5, 8, 9, 7
lea dstq, [dstq+strideq*2]
; restore from cache
SWAP 0, 7 ; move zero from m7 to m0
mova m7, [rsp+32]
mova m8, [rsp+64]
mova m9, [rsp+96]
SUMSUB_BA w, 6, 9, 1 ; t6, t9
SUMSUB_BA w, 7, 8, 1 ; t7, t8
pmulhrsw m6, [pw_512]
pmulhrsw m7, [pw_512]
VP9_STORE_2X 6, 7, 1, 2, 0
lea dstq, [dstq+strideq*2]
pmulhrsw m8, [pw_512]
pmulhrsw m9, [pw_512]
VP9_STORE_2X 8, 9, 1, 2, 0
lea dstq, [dstq+strideq*2]
pmulhrsw m10, [pw_512]
pmulhrsw m11, [pw_512]
VP9_STORE_2X 10, 11, 1, 2, 0
lea dstq, [dstq+strideq*2]
pmulhrsw m12, [pw_512]
pmulhrsw m13, [pw_512]
VP9_STORE_2X 12, 13, 1, 2, 0
lea dstq, [dstq+strideq*2]
pmulhrsw m14, [pw_512]
pmulhrsw m15, [pw_512]
VP9_STORE_2X 14, 15, 1, 2, 0
%endif ; %2 == 1/2
%endmacro
vp9/x86: 16x16 sub-IDCT for top-left 8x8 subblock (eob <= 38). Sub8x8 speed (w/o dc-only case) goes from ~750 cycles (inter) or ~735 cycles (intra) to ~415 cycles (inter) or ~430 cycles (intra). Average overall 16x16 idct speed goes from ~635 cycles (inter) or ~720 cycles (intra) to ~415 cycles (inter) or ~545 (intra) - all measurements done using ped1080p.webm.
11 years ago
%macro ZERO_BLOCK 4 ; mem, stride, nnzcpl, zero_reg
%assign %%y 0
%rep %3
%assign %%x 0
%rep %3*2/mmsize
mova [%1+%%y+%%x], %4
%assign %%x (%%x+mmsize)
%endrep
%assign %%y (%%y+%2)
vp9/x86: idct_add_16x16_ssse3. Currently only dc-only and full 16x16. Other subforms will follow in the near future. Total decoding time of ped1080p.webm goes from 9.7 to 9.3 seconds. DC-only goes from 957 -> 131 cycles, and the full IDCT goes from ~4050 to ~745 cycles.
11 years ago
%endrep
%endmacro
vp9/x86: idct_32x32_add_ssse3. Sub-IDCTs will follow later. ped1080.webm goes from 9.295s to 8.191s (13.5% faster). The IDCT itself goes from 4372 (intra) or 4337 (inter) to 403 (intra) or 329 (inter) cycles for the DC-only form, 23755 (intra) or 23723 (inter) to 3497 (intra) or 3607 (inter) cycles for the no-DC form, which averages from 23393 (intra) or 16612 (inter) to 3449 (intra) or 2392 (inter) for all 32x32s together, i.e. about ~7x faster (all tests done on ped1080p.webm).
11 years ago
%macro VP9_STORE_2XFULL 6-7 strideq; dc, tmp1, tmp2, tmp3, tmp4, zero, stride
vp9/x86: idct_add_16x16_ssse3. Currently only dc-only and full 16x16. Other subforms will follow in the near future. Total decoding time of ped1080p.webm goes from 9.7 to 9.3 seconds. DC-only goes from 957 -> 131 cycles, and the full IDCT goes from ~4050 to ~745 cycles.
11 years ago
mova m%3, [dstq]
vp9/x86: idct_32x32_add_ssse3. Sub-IDCTs will follow later. ped1080.webm goes from 9.295s to 8.191s (13.5% faster). The IDCT itself goes from 4372 (intra) or 4337 (inter) to 403 (intra) or 329 (inter) cycles for the DC-only form, 23755 (intra) or 23723 (inter) to 3497 (intra) or 3607 (inter) cycles for the no-DC form, which averages from 23393 (intra) or 16612 (inter) to 3449 (intra) or 2392 (inter) for all 32x32s together, i.e. about ~7x faster (all tests done on ped1080p.webm).
11 years ago
mova m%5, [dstq+%7]
vp9/x86: idct_add_16x16_ssse3. Currently only dc-only and full 16x16. Other subforms will follow in the near future. Total decoding time of ped1080p.webm goes from 9.7 to 9.3 seconds. DC-only goes from 957 -> 131 cycles, and the full IDCT goes from ~4050 to ~745 cycles.
11 years ago
punpcklbw m%2, m%3, m%6
punpckhbw m%3, m%6
punpcklbw m%4, m%5, m%6
punpckhbw m%5, m%6
paddw m%2, m%1
paddw m%3, m%1
paddw m%4, m%1
paddw m%5, m%1
packuswb m%2, m%3
packuswb m%4, m%5
mova [dstq], m%2
vp9/x86: idct_32x32_add_ssse3. Sub-IDCTs will follow later. ped1080.webm goes from 9.295s to 8.191s (13.5% faster). The IDCT itself goes from 4372 (intra) or 4337 (inter) to 403 (intra) or 329 (inter) cycles for the DC-only form, 23755 (intra) or 23723 (inter) to 3497 (intra) or 3607 (inter) cycles for the no-DC form, which averages from 23393 (intra) or 16612 (inter) to 3449 (intra) or 2392 (inter) for all 32x32s together, i.e. about ~7x faster (all tests done on ped1080p.webm).
11 years ago
mova [dstq+%7], m%4
vp9/x86: idct_add_16x16_ssse3. Currently only dc-only and full 16x16. Other subforms will follow in the near future. Total decoding time of ped1080p.webm goes from 9.7 to 9.3 seconds. DC-only goes from 957 -> 131 cycles, and the full IDCT goes from ~4050 to ~745 cycles.
11 years ago
%endmacro
INIT_XMM ssse3
cglobal vp9_idct_idct_16x16_add, 4, 5, 16, 512, dst, stride, block, eob
vp9/x86: 16x16 sub-IDCT for top-left 8x8 subblock (eob <= 38). Sub8x8 speed (w/o dc-only case) goes from ~750 cycles (inter) or ~735 cycles (intra) to ~415 cycles (inter) or ~430 cycles (intra). Average overall 16x16 idct speed goes from ~635 cycles (inter) or ~720 cycles (intra) to ~415 cycles (inter) or ~545 (intra) - all measurements done using ped1080p.webm.
11 years ago
; 2x2=eob=3, 4x4=eob=10
cmp eobd, 38
jg .idctfull
vp9/x86: idct_add_16x16_ssse3. Currently only dc-only and full 16x16. Other subforms will follow in the near future. Total decoding time of ped1080p.webm goes from 9.7 to 9.3 seconds. DC-only goes from 957 -> 131 cycles, and the full IDCT goes from ~4050 to ~745 cycles.
11 years ago
cmp eobd, 1 ; faster path for when only DC is set
vp9/x86: 16x16 sub-IDCT for top-left 8x8 subblock (eob <= 38). Sub8x8 speed (w/o dc-only case) goes from ~750 cycles (inter) or ~735 cycles (intra) to ~415 cycles (inter) or ~430 cycles (intra). Average overall 16x16 idct speed goes from ~635 cycles (inter) or ~720 cycles (intra) to ~415 cycles (inter) or ~545 (intra) - all measurements done using ped1080p.webm.
11 years ago
jne .idct8x8
vp9/x86: idct_add_16x16_ssse3. Currently only dc-only and full 16x16. Other subforms will follow in the near future. Total decoding time of ped1080p.webm goes from 9.7 to 9.3 seconds. DC-only goes from 957 -> 131 cycles, and the full IDCT goes from ~4050 to ~745 cycles.
11 years ago
; dc-only
movd m0, [blockq]
mova m1, [pw_11585x2]
pmulhrsw m0, m1
pmulhrsw m0, m1
SPLATW m0, m0, q0000
pmulhrsw m0, [pw_512]
pxor m5, m5
movd [blockq], m5
%rep 7
VP9_STORE_2XFULL 0, 1, 2, 3, 4, 5
lea dstq, [dstq+2*strideq]
%endrep
VP9_STORE_2XFULL 0, 1, 2, 3, 4, 5
RET
vp9/x86: 16x16 sub-IDCT for top-left 8x8 subblock (eob <= 38). Sub8x8 speed (w/o dc-only case) goes from ~750 cycles (inter) or ~735 cycles (intra) to ~415 cycles (inter) or ~430 cycles (intra). Average overall 16x16 idct speed goes from ~635 cycles (inter) or ~720 cycles (intra) to ~415 cycles (inter) or ~545 (intra) - all measurements done using ped1080p.webm.
11 years ago
.idct8x8:
DEFINE_ARGS dst, stride, block, cnt, dst_bak
VP9_IDCT16_1D blockq, 1, 8
mov cntd, 2
mov dst_bakq, dstq
.loop2_8x8:
VP9_IDCT16_1D rsp, 2, 8
lea dstq, [dst_bakq+8]
add rsp, 16
dec cntd
jg .loop2_8x8
sub rsp, 32
; at the end of the loop, m0 should still be zero
; use that to zero out block coefficients
ZERO_BLOCK blockq, 32, 8, m0
RET
vp9/x86: idct_add_16x16_ssse3. Currently only dc-only and full 16x16. Other subforms will follow in the near future. Total decoding time of ped1080p.webm goes from 9.7 to 9.3 seconds. DC-only goes from 957 -> 131 cycles, and the full IDCT goes from ~4050 to ~745 cycles.
11 years ago
.idctfull:
DEFINE_ARGS dst, stride, block, cnt, dst_bak
mov cntd, 2
.loop1_full:
VP9_IDCT16_1D blockq, 1
add blockq, 16
add rsp, 256
dec cntd
jg .loop1_full
sub blockq, 32
sub rsp, 512
mov cntd, 2
mov dst_bakq, dstq
.loop2_full:
VP9_IDCT16_1D rsp, 2
lea dstq, [dst_bakq+8]
add rsp, 16
dec cntd
jg .loop2_full
sub rsp, 32
; at the end of the loop, m0 should still be zero
; use that to zero out block coefficients
vp9/x86: 16x16 sub-IDCT for top-left 8x8 subblock (eob <= 38). Sub8x8 speed (w/o dc-only case) goes from ~750 cycles (inter) or ~735 cycles (intra) to ~415 cycles (inter) or ~430 cycles (intra). Average overall 16x16 idct speed goes from ~635 cycles (inter) or ~720 cycles (intra) to ~415 cycles (inter) or ~545 (intra) - all measurements done using ped1080p.webm.
11 years ago
ZERO_BLOCK blockq, 32, 16, m0
vp9/x86: idct_add_16x16_ssse3. Currently only dc-only and full 16x16. Other subforms will follow in the near future. Total decoding time of ped1080p.webm goes from 9.7 to 9.3 seconds. DC-only goes from 957 -> 131 cycles, and the full IDCT goes from ~4050 to ~745 cycles.
11 years ago
RET
vp9/x86: idct_32x32_add_ssse3. Sub-IDCTs will follow later. ped1080.webm goes from 9.295s to 8.191s (13.5% faster). The IDCT itself goes from 4372 (intra) or 4337 (inter) to 403 (intra) or 329 (inter) cycles for the DC-only form, 23755 (intra) or 23723 (inter) to 3497 (intra) or 3607 (inter) cycles for the no-DC form, which averages from 23393 (intra) or 16612 (inter) to 3449 (intra) or 2392 (inter) for all 32x32s together, i.e. about ~7x faster (all tests done on ped1080p.webm).
11 years ago
;---------------------------------------------------------------------------------------------
; void vp9_idct_idct_32x32_add_<opt>(uint8_t *dst, ptrdiff_t stride, int16_t *block, int eob);
;---------------------------------------------------------------------------------------------
%macro VP9_IDCT32_1D 2-3 32 ; src, pass, nnzc
%assign %%str 16*%2*%2
; first do t0-15, this can be done identical to idct16x16
VP9_IDCT16_1D_START %1, %3/2, 64*2, rsp+ 4*%%str
; backup a different register
mova [rsp+30*%%str], m15 ; t15
mova m7, [rsp+ 4*%%str]
SUMSUB_BA w, 6, 9, 15 ; t6, t9
SUMSUB_BA w, 7, 8, 15 ; t7, t8
; store everything on stack to make space available for t16-31
; we store interleaved with the output of the second half (t16-31)
; so we don't need to allocate extra stack space
mova [rsp+ 0*%%str], m0 ; t0
mova [rsp+ 4*%%str], m1 ; t1
mova [rsp+ 8*%%str], m2 ; t2
mova [rsp+12*%%str], m3 ; t3
mova [rsp+16*%%str], m4 ; t4
mova [rsp+20*%%str], m5 ; t5
mova [rsp+24*%%str], m6 ; t6
mova [rsp+28*%%str], m7 ; t7
mova [rsp+ 2*%%str], m8 ; t8
mova [rsp+ 6*%%str], m9 ; t9
mova [rsp+10*%%str], m10 ; t10
mova [rsp+14*%%str], m11 ; t11
mova [rsp+18*%%str], m12 ; t12
mova [rsp+22*%%str], m13 ; t13
mova [rsp+26*%%str], m14 ; t14
; then, secondly, do t16-31
vp9/x86: idct_32x32_add_ssse3 sub-8x8-idct. Runtime of the full 32x32 idct goes from 2446 to 2441 cycles (intra) or from 1425 to 1306 cycles (inter). Overall runtime is not significantly affected.
11 years ago
%if %3 <= 8
mova m4, [%1+ 1*64]
mova m3, [%1+ 3*64]
mova m0, [%1+ 5*64]
mova m7, [%1+ 7*64]
pmulhrsw m11, m4, [pw_16364x2] ;t31
pmulhrsw m4, [pw_804x2] ;t16
pmulhrsw m8, m7, [pw_m5520x2] ;t19
pmulhrsw m7, [pw_15426x2] ;t28
pmulhrsw m15, m0, [pw_15893x2] ;t27
pmulhrsw m0, [pw_3981x2] ;t20
pmulhrsw m12, m3, [pw_m2404x2] ;t23
pmulhrsw m3, [pw_16207x2] ;t24
; m4=t16/17, m8=t18/19, m0=t20/21, m12=t22/23,
; m3=t24/25, m15=t26/27, m7=t28/29, m11=t30/31
VP9_UNPACK_MULSUB_2W_4X 5, 10, 11, 4, 16069, 3196, [pd_8192], 6, 9 ; t17, t30
VP9_UNPACK_MULSUB_2W_4X 9, 6, 7, 8, 3196, m16069, [pd_8192], 1, 14 ; t18, t29
; from 1 stage forward
SUMSUB_BA w, 8, 4, 1
; temporary storage
mova [rsp+17*%%str], m8 ; t16
mova [rsp+21*%%str], m4 ; t19
VP9_UNPACK_MULSUB_2W_4X 1, 14, 15, 0, 9102, 13623, [pd_8192], 4, 8 ; t21, t26
VP9_UNPACK_MULSUB_2W_4X 13, 2, 3, 12, 13623, m9102, [pd_8192], 4, 8 ; t22, t25
; m4=t16, m5=t17, m9=t18, m8=t19, m0=t20, m1=t21, m13=t22, m12=t23,
; m3=t24, m2=t25, m14=t26, m15=t27, m7=t28, m6=t29, m10=t30, m11=t31
%else
vp9/x86: idct_32x32_add_ssse3. Sub-IDCTs will follow later. ped1080.webm goes from 9.295s to 8.191s (13.5% faster). The IDCT itself goes from 4372 (intra) or 4337 (inter) to 403 (intra) or 329 (inter) cycles for the DC-only form, 23755 (intra) or 23723 (inter) to 3497 (intra) or 3607 (inter) cycles for the no-DC form, which averages from 23393 (intra) or 16612 (inter) to 3449 (intra) or 2392 (inter) for all 32x32s together, i.e. about ~7x faster (all tests done on ped1080p.webm).
11 years ago
mova m10, [%1+ 1*64]
mova m13, [%1+ 3*64]
mova m14, [%1+ 5*64]
mova m9, [%1+ 7*64]
mova m8, [%1+ 9*64]
mova m15, [%1+11*64]
mova m12, [%1+13*64]
mova m11, [%1+15*64]
vp9/x86: idct_32x32_add_ssse3 sub-16x16-idct. Runtime of all IDCTs together goes from 3327 to 2473 cycles (intra, i.e. ~35% faster) or from 2312 to 1448 cycles (inter, i.e. ~60% faster). Total decode time of ped1080p.webm goes from 8.086sec to 7.974sec (1.4% faster).
11 years ago
%if %3 <= 16
pmulhrsw m5, m10, [pw_16364x2]
pmulhrsw m10, [pw_804x2]
pmulhrsw m4, m11, [pw_m11003x2]
pmulhrsw m11, [pw_12140x2]
pmulhrsw m7, m8, [pw_14811x2]
pmulhrsw m8, [pw_7005x2]
pmulhrsw m6, m9, [pw_m5520x2]
pmulhrsw m9, [pw_15426x2]
pmulhrsw m1, m14, [pw_15893x2]
pmulhrsw m14, [pw_3981x2]
pmulhrsw m0, m15, [pw_m8423x2]
pmulhrsw m15, [pw_14053x2]
%else
vp9/x86: idct_32x32_add_ssse3. Sub-IDCTs will follow later. ped1080.webm goes from 9.295s to 8.191s (13.5% faster). The IDCT itself goes from 4372 (intra) or 4337 (inter) to 403 (intra) or 329 (inter) cycles for the DC-only form, 23755 (intra) or 23723 (inter) to 3497 (intra) or 3607 (inter) cycles for the no-DC form, which averages from 23393 (intra) or 16612 (inter) to 3449 (intra) or 2392 (inter) for all 32x32s together, i.e. about ~7x faster (all tests done on ped1080p.webm).
11 years ago
mova m4, [%1+17*64]
mova m0, [%1+21*64]
mova m7, [%1+23*64]
mova m6, [%1+25*64]
mova m1, [%1+27*64]
mova m5, [%1+31*64]
; m10=in1, m4=in17, m8=in9, m6=in25, m14=in5, m0=in21, m12=in13, m2=in29,
; m13=in3, m3=in19, m15=in11, m1=in27, m9=in7, m7=in23, m11=in15, m5=in31
VP9_UNPACK_MULSUB_2W_4X 10, 5, 16364, 804, [pd_8192], 2, 3 ; t16, t31
VP9_UNPACK_MULSUB_2W_4X 4, 11, 11003, 12140, [pd_8192], 2, 3 ; t17, t30
VP9_UNPACK_MULSUB_2W_4X 8, 7, 14811, 7005, [pd_8192], 2, 3 ; t18, t29
VP9_UNPACK_MULSUB_2W_4X 6, 9, 5520, 15426, [pd_8192], 2, 3 ; t19, t28
VP9_UNPACK_MULSUB_2W_4X 14, 1, 15893, 3981, [pd_8192], 2, 3 ; t20, t27
VP9_UNPACK_MULSUB_2W_4X 0, 15, 8423, 14053, [pd_8192], 2, 3 ; t21, t26
vp9/x86: idct_32x32_add_ssse3 sub-16x16-idct. Runtime of all IDCTs together goes from 3327 to 2473 cycles (intra, i.e. ~35% faster) or from 2312 to 1448 cycles (inter, i.e. ~60% faster). Total decode time of ped1080p.webm goes from 8.086sec to 7.974sec (1.4% faster).
11 years ago
%endif
vp9/x86: idct_32x32_add_ssse3. Sub-IDCTs will follow later. ped1080.webm goes from 9.295s to 8.191s (13.5% faster). The IDCT itself goes from 4372 (intra) or 4337 (inter) to 403 (intra) or 329 (inter) cycles for the DC-only form, 23755 (intra) or 23723 (inter) to 3497 (intra) or 3607 (inter) cycles for the no-DC form, which averages from 23393 (intra) or 16612 (inter) to 3449 (intra) or 2392 (inter) for all 32x32s together, i.e. about ~7x faster (all tests done on ped1080p.webm).
11 years ago
; from 1 stage forward
SUMSUB_BA w, 4, 10, 2
SUMSUB_BA w, 8, 6, 2
; from 2 stages forward
SUMSUB_BA w, 8, 4, 2
; temporary storage
mova [rsp+17*%%str], m8 ; t16
mova [rsp+21*%%str], m4 ; t19
vp9/x86: idct_32x32_add_ssse3 sub-16x16-idct. Runtime of all IDCTs together goes from 3327 to 2473 cycles (intra, i.e. ~35% faster) or from 2312 to 1448 cycles (inter, i.e. ~60% faster). Total decode time of ped1080p.webm goes from 8.086sec to 7.974sec (1.4% faster).
11 years ago
%if %3 <= 16
pmulhrsw m3, m12, [pw_13160x2]
pmulhrsw m12, [pw_9760x2]
pmulhrsw m2, m13, [pw_m2404x2]
pmulhrsw m13, [pw_16207x2]
%else
vp9/x86: idct_32x32_add_ssse3. Sub-IDCTs will follow later. ped1080.webm goes from 9.295s to 8.191s (13.5% faster). The IDCT itself goes from 4372 (intra) or 4337 (inter) to 403 (intra) or 329 (inter) cycles for the DC-only form, 23755 (intra) or 23723 (inter) to 3497 (intra) or 3607 (inter) cycles for the no-DC form, which averages from 23393 (intra) or 16612 (inter) to 3449 (intra) or 2392 (inter) for all 32x32s together, i.e. about ~7x faster (all tests done on ped1080p.webm).
11 years ago
mova m2, [%1+29*64]
mova m3, [%1+19*64]
VP9_UNPACK_MULSUB_2W_4X 12, 3, 13160, 9760, [pd_8192], 4, 8 ; t22, t25
VP9_UNPACK_MULSUB_2W_4X 2, 13, 2404, 16207, [pd_8192], 4, 8 ; t23, t24
vp9/x86: idct_32x32_add_ssse3 sub-16x16-idct. Runtime of all IDCTs together goes from 3327 to 2473 cycles (intra, i.e. ~35% faster) or from 2312 to 1448 cycles (inter, i.e. ~60% faster). Total decode time of ped1080p.webm goes from 8.086sec to 7.974sec (1.4% faster).
11 years ago
%endif
vp9/x86: idct_32x32_add_ssse3. Sub-IDCTs will follow later. ped1080.webm goes from 9.295s to 8.191s (13.5% faster). The IDCT itself goes from 4372 (intra) or 4337 (inter) to 403 (intra) or 329 (inter) cycles for the DC-only form, 23755 (intra) or 23723 (inter) to 3497 (intra) or 3607 (inter) cycles for the no-DC form, which averages from 23393 (intra) or 16612 (inter) to 3449 (intra) or 2392 (inter) for all 32x32s together, i.e. about ~7x faster (all tests done on ped1080p.webm).
11 years ago
; m10=t16, m4=t17, m8=t18, m6=t19, m14=t20, m0=t21, m12=t22, m2=t23,
; m13=t24, m3=t25, m15=t26, m1=t27, m9=t28, m7=t29, m11=t30, m5=t31
SUMSUB_BA w, 0, 14, 4
SUMSUB_BA w, 12, 2, 4
SUMSUB_BA w, 3, 13, 4
SUMSUB_BA w, 15, 1, 4
SUMSUB_BA w, 7, 9, 4
SUMSUB_BA w, 11, 5, 4
; m4=t16, m10=t17, m6=t18, m8=t19, m0=t20, m14=t21, m2=t22, m12=t23,
; m3=t24, m13=t25, m1=t26, m15=t27, m7=t28, m9=t29, m5=t30, m11=t31
VP9_UNPACK_MULSUB_2W_4X 5, 10, 16069, 3196, [pd_8192], 4, 8 ; t17, t30
VP9_UNPACK_MULSUB_2W_4X 9, 6, 3196, m16069, [pd_8192], 4, 8 ; t18, t29
VP9_UNPACK_MULSUB_2W_4X 1, 14, 9102, 13623, [pd_8192], 4, 8 ; t21, t26
VP9_UNPACK_MULSUB_2W_4X 13, 2, 13623, m9102, [pd_8192], 4, 8 ; t22, t25
vp9/x86: idct_32x32_add_ssse3 sub-8x8-idct. Runtime of the full 32x32 idct goes from 2446 to 2441 cycles (intra) or from 1425 to 1306 cycles (inter). Overall runtime is not significantly affected.
11 years ago
%endif
vp9/x86: idct_32x32_add_ssse3. Sub-IDCTs will follow later. ped1080.webm goes from 9.295s to 8.191s (13.5% faster). The IDCT itself goes from 4372 (intra) or 4337 (inter) to 403 (intra) or 329 (inter) cycles for the DC-only form, 23755 (intra) or 23723 (inter) to 3497 (intra) or 3607 (inter) cycles for the no-DC form, which averages from 23393 (intra) or 16612 (inter) to 3449 (intra) or 2392 (inter) for all 32x32s together, i.e. about ~7x faster (all tests done on ped1080p.webm).
11 years ago
; m4=t16, m5=t17, m9=t18, m8=t19, m0=t20, m1=t21, m13=t22, m12=t23,
; m3=t24, m2=t25, m14=t26, m15=t27, m7=t28, m6=t29, m10=t30, m11=t31
SUMSUB_BA w, 9, 5, 4
SUMSUB_BA w, 1, 13, 4
SUMSUB_BA w, 0, 12, 4
SUMSUB_BA w, 15, 3, 4
SUMSUB_BA w, 14, 2, 4
SUMSUB_BA w, 6, 10, 4
SUMSUB_BA w, 7, 11, 4
; m8[s]=t16, m9=t17, m5=t18, m4[s]=t19, m12=t20, m13=t21, m1=t22, m0=t23,
; m15=t24, m14=t25, m2=t26, m3=t27, m11=t28, m10=t29, m6=t30, m7=t31
mova m8, [rsp+17*%%str] ; t16
; from 2 stages forward
SUMSUB_BA w, 0, 8, 4
SUMSUB_BA w, 15, 7, 4
; from 3 stages forward
SUMSUB_BA w, 8, 7, 4
pmulhrsw m7, [pw_11585x2]
pmulhrsw m8, [pw_11585x2]
; store t16/t23
mova [rsp+ 1*%%str], m0 ; t16
mova [rsp+29*%%str], m7 ; t23
mova m4, [rsp+21*%%str] ; t19
VP9_UNPACK_MULSUB_2W_4X 10, 5, 15137, 6270, [pd_8192], 0, 7 ; t18, t29
VP9_UNPACK_MULSUB_2W_4X 11, 4, 15137, 6270, [pd_8192], 0, 7 ; t19, t28
VP9_UNPACK_MULSUB_2W_4X 3, 12, 6270, m15137, [pd_8192], 0, 7 ; t20, t27
VP9_UNPACK_MULSUB_2W_4X 2, 13, 6270, m15137, [pd_8192], 0, 7 ; t21, t26
; m8=t16, m9=t17, m10=t18, m11=t19, m3=t20, m2=t21, m1=t22, m0=t23,
; m15=t24, m14=t25, m13=t26, m12=t27, m4=t28, m5=t29, m6=t30, m7=t31
SUMSUB_BA w, 1, 9, 0
SUMSUB_BA w, 2, 10, 0
SUMSUB_BA w, 3, 11, 0
SUMSUB_BA w, 12, 4, 0
SUMSUB_BA w, 13, 5, 0
SUMSUB_BA w, 14, 6, 0
; m0=t16, m1=t17, m2=t18, m3=t19, m11=t20, m10=t21, m9=t22, m8=t23,
; m7=t24, m6=t25, m5=t26, m4=t27, m12=t28, m13=t29, m14=t30, m15=t31
SUMSUB_BA w, 9, 6, 0
SUMSUB_BA w, 10, 5, 0
SUMSUB_BA w, 11, 4, 0
pmulhrsw m6, [pw_11585x2]
pmulhrsw m9, [pw_11585x2]
pmulhrsw m5, [pw_11585x2]
pmulhrsw m10, [pw_11585x2]
pmulhrsw m4, [pw_11585x2]
pmulhrsw m11, [pw_11585x2]
; m0=t16, m1=t17, m2=t18, m3=t19, m4=t20, m5=t21, m6=t22, m7=t23,
; m8=t24, m9=t25, m10=t26, m11=t27, m12=t28, m13=t29, m14=t30, m15=t31
; store t17-19 (and t20-22 for pass 1) - keep t24-31 in registers for
; final sumsub in pass 1, or keep t20-22 and t24-31 in registers for
; final sumsub of pass 2
mova [rsp+ 5*%%str], m1 ; t17
mova [rsp+ 9*%%str], m2 ; t18
mova [rsp+13*%%str], m3 ; t19
; then do final pass to sumsub+store the two halves
%if %2 == 1
mova [rsp+17*%%str], m4 ; t20
mova [rsp+21*%%str], m5 ; t21
mova [rsp+25*%%str], m6 ; t22
mova m0, [rsp+ 0*%%str] ; t0
mova m1, [rsp+ 4*%%str] ; t1
mova m2, [rsp+ 8*%%str] ; t2
mova m3, [rsp+12*%%str] ; t3
mova m4, [rsp+16*%%str] ; t4
mova m5, [rsp+20*%%str] ; t5
mova m6, [rsp+24*%%str] ; t6
SUMSUB_BA w, 15, 0, 7
mova [rsp+ 3*%%str], m0 ; t15
mova m7, [rsp+28*%%str] ; t7
SUMSUB_BA w, 14, 1, 0
SUMSUB_BA w, 13, 2, 0
SUMSUB_BA w, 12, 3, 0
SUMSUB_BA w, 11, 4, 0
SUMSUB_BA w, 10, 5, 0
SUMSUB_BA w, 9, 6, 0
SUMSUB_BA w, 8, 7, 0
TRANSPOSE8x8W 15, 14, 13, 12, 11, 10, 9, 8, 0
mova [rsp+ 0*%%str], m15
mova [rsp+ 4*%%str], m14
mova [rsp+ 8*%%str], m13
mova [rsp+12*%%str], m12
mova [rsp+16*%%str], m11
mova [rsp+20*%%str], m10
mova [rsp+24*%%str], m9
mova [rsp+28*%%str], m8
mova m0, [rsp+ 3*%%str] ; t15
TRANSPOSE8x8W 7, 6, 5, 4, 3, 2, 1, 0, 8
mova [rsp+ 3*%%str], m7
mova [rsp+ 7*%%str], m6
mova [rsp+11*%%str], m5
mova [rsp+15*%%str], m4
mova [rsp+19*%%str], m3
mova [rsp+23*%%str], m2
mova [rsp+27*%%str], m1
mova [rsp+31*%%str], m0
mova m15, [rsp+ 2*%%str] ; t8
mova m14, [rsp+ 6*%%str] ; t9
mova m13, [rsp+10*%%str] ; t10
mova m12, [rsp+14*%%str] ; t11
mova m11, [rsp+18*%%str] ; t12
mova m10, [rsp+22*%%str] ; t13
mova m9, [rsp+26*%%str] ; t14
mova m8, [rsp+30*%%str] ; t15
mova m7, [rsp+ 1*%%str] ; t16
mova m6, [rsp+ 5*%%str] ; t17
mova m5, [rsp+ 9*%%str] ; t18
mova m4, [rsp+13*%%str] ; t19
mova m3, [rsp+17*%%str] ; t20
mova m2, [rsp+21*%%str] ; t21
mova m1, [rsp+25*%%str] ; t22
SUMSUB_BA w, 7, 8, 0
mova [rsp+ 2*%%str], m8
mova m0, [rsp+29*%%str] ; t23
SUMSUB_BA w, 6, 9, 8
SUMSUB_BA w, 5, 10, 8
SUMSUB_BA w, 4, 11, 8
SUMSUB_BA w, 3, 12, 8
SUMSUB_BA w, 2, 13, 8
SUMSUB_BA w, 1, 14, 8
SUMSUB_BA w, 0, 15, 8
TRANSPOSE8x8W 0, 1, 2, 3, 4, 5, 6, 7, 8
mova [rsp+ 1*%%str], m0
mova [rsp+ 5*%%str], m1
mova [rsp+ 9*%%str], m2
mova [rsp+13*%%str], m3
mova [rsp+17*%%str], m4
mova [rsp+21*%%str], m5
mova [rsp+25*%%str], m6
mova [rsp+29*%%str], m7
mova m8, [rsp+ 2*%%str]
TRANSPOSE8x8W 8, 9, 10, 11, 12, 13, 14, 15, 0
mova [rsp+ 2*%%str], m8
mova [rsp+ 6*%%str], m9
mova [rsp+10*%%str], m10
mova [rsp+14*%%str], m11
mova [rsp+18*%%str], m12
mova [rsp+22*%%str], m13
mova [rsp+26*%%str], m14
mova [rsp+30*%%str], m15
%else
; t0-7 is in [rsp+{0,4,8,12,16,20,24,28}*%%str]
; t8-15 is in [rsp+{2,6,10,14,18,22,26,30}*%%str]
; t16-19 and t23 is in [rsp+{1,5,9,13,29}*%%str]
; t20-22 is in m4-6
; t24-31 is in m8-15
pxor m7, m7
%macro STORE_2X2 7-8 1 ; src[1-4], tmp[1-2], zero, inc_dst_ptrs
SUMSUB_BA w, %4, %1, %5
SUMSUB_BA w, %3, %2, %5
pmulhrsw m%4, [pw_512]
pmulhrsw m%3, [pw_512]
VP9_STORE_2X %4, %3, %5, %6, %7
%if %8 == 1
add dstq, stride2q
%endif
pmulhrsw m%2, [pw_512]
pmulhrsw m%1, [pw_512]
VP9_STORE_2X %2, %1, %5, %6, %7, dst_endq
%if %8 == 1
sub dst_endq, stride2q
%endif
%endmacro
; store t0-1 and t30-31
mova m0, [rsp+ 0*%%str]
mova m1, [rsp+ 4*%%str]
STORE_2X2 0, 1, 14, 15, 2, 3, 7
; store t2-3 and t28-29
mova m0, [rsp+ 8*%%str]
mova m1, [rsp+12*%%str]
STORE_2X2 0, 1, 12, 13, 2, 3, 7
; store t4-5 and t26-27
mova m0, [rsp+16*%%str]
mova m1, [rsp+20*%%str]
STORE_2X2 0, 1, 10, 11, 2, 3, 7
; store t6-7 and t24-25
mova m0, [rsp+24*%%str]
mova m1, [rsp+28*%%str]
STORE_2X2 0, 1, 8, 9, 2, 3, 7
; store t8-9 and t22-23
mova m0, [rsp+ 2*%%str]
mova m1, [rsp+ 6*%%str]
mova m8, [rsp+29*%%str]
STORE_2X2 0, 1, 6, 8, 2, 3, 7
; store t10-11 and t20-21
mova m0, [rsp+10*%%str]
mova m1, [rsp+14*%%str]
STORE_2X2 0, 1, 4, 5, 2, 3, 7
; store t12-13 and t18-19
mova m0, [rsp+18*%%str]
mova m1, [rsp+22*%%str]
mova m5, [rsp+13*%%str]
mova m4, [rsp+ 9*%%str]
STORE_2X2 0, 1, 4, 5, 2, 3, 7
; store t14-17
mova m0, [rsp+26*%%str]
mova m1, [rsp+30*%%str]
mova m5, [rsp+ 5*%%str]
mova m4, [rsp+ 1*%%str]
STORE_2X2 0, 1, 4, 5, 2, 3, 7, 0
%endif
%endmacro
INIT_XMM ssse3
cglobal vp9_idct_idct_32x32_add, 4, 8, 16, 2048, dst, stride, block, eob
vp9/x86: idct_32x32_add_ssse3 sub-16x16-idct. Runtime of all IDCTs together goes from 3327 to 2473 cycles (intra, i.e. ~35% faster) or from 2312 to 1448 cycles (inter, i.e. ~60% faster). Total decode time of ped1080p.webm goes from 8.086sec to 7.974sec (1.4% faster).
11 years ago
cmp eobd, 135
vp9/x86: idct_32x32_add_ssse3. Sub-IDCTs will follow later. ped1080.webm goes from 9.295s to 8.191s (13.5% faster). The IDCT itself goes from 4372 (intra) or 4337 (inter) to 403 (intra) or 329 (inter) cycles for the DC-only form, 23755 (intra) or 23723 (inter) to 3497 (intra) or 3607 (inter) cycles for the no-DC form, which averages from 23393 (intra) or 16612 (inter) to 3449 (intra) or 2392 (inter) for all 32x32s together, i.e. about ~7x faster (all tests done on ped1080p.webm).
11 years ago
jg .idctfull
vp9/x86: idct_32x32_add_ssse3 sub-8x8-idct. Runtime of the full 32x32 idct goes from 2446 to 2441 cycles (intra) or from 1425 to 1306 cycles (inter). Overall runtime is not significantly affected.
11 years ago
cmp eobd, 34
vp9/x86: idct_32x32_add_ssse3 sub-16x16-idct. Runtime of all IDCTs together goes from 3327 to 2473 cycles (intra, i.e. ~35% faster) or from 2312 to 1448 cycles (inter, i.e. ~60% faster). Total decode time of ped1080p.webm goes from 8.086sec to 7.974sec (1.4% faster).
11 years ago
jg .idct16x16
vp9/x86: idct_32x32_add_ssse3 sub-8x8-idct. Runtime of the full 32x32 idct goes from 2446 to 2441 cycles (intra) or from 1425 to 1306 cycles (inter). Overall runtime is not significantly affected.
11 years ago
cmp eobd, 1
jg .idct8x8
vp9/x86: idct_32x32_add_ssse3. Sub-IDCTs will follow later. ped1080.webm goes from 9.295s to 8.191s (13.5% faster). The IDCT itself goes from 4372 (intra) or 4337 (inter) to 403 (intra) or 329 (inter) cycles for the DC-only form, 23755 (intra) or 23723 (inter) to 3497 (intra) or 3607 (inter) cycles for the no-DC form, which averages from 23393 (intra) or 16612 (inter) to 3449 (intra) or 2392 (inter) for all 32x32s together, i.e. about ~7x faster (all tests done on ped1080p.webm).
11 years ago
; dc-only case
movd m0, [blockq]
mova m1, [pw_11585x2]
pmulhrsw m0, m1
pmulhrsw m0, m1
SPLATW m0, m0, q0000
pmulhrsw m0, [pw_512]
pxor m5, m5
movd [blockq], m5
DEFINE_ARGS dst, stride, block, cnt
%rep 31
VP9_STORE_2XFULL 0, 1, 2, 3, 4, 5, mmsize
add dstq, strideq
%endrep
VP9_STORE_2XFULL 0, 1, 2, 3, 4, 5, mmsize
RET
DEFINE_ARGS dst_bak, stride, block, cnt, dst, stride30, dst_end, stride2
vp9/x86: idct_32x32_add_ssse3 sub-8x8-idct. Runtime of the full 32x32 idct goes from 2446 to 2441 cycles (intra) or from 1425 to 1306 cycles (inter). Overall runtime is not significantly affected.
11 years ago
.idct8x8:
VP9_IDCT32_1D blockq, 1, 8
mov stride30q, strideq ; stride
lea stride2q, [strideq*2] ; stride*2
shl stride30q, 5 ; stride*32
mov cntd, 4
sub stride30q, stride2q ; stride*30
.loop2_8x8:
mov dstq, dst_bakq
lea dst_endq, [dst_bakq+stride30q]
VP9_IDCT32_1D rsp, 2, 8
add dst_bakq, 8
add rsp, 16
dec cntd
jg .loop2_8x8
sub rsp, 64
; at the end of the loop, m7 should still be zero
; use that to zero out block coefficients
ZERO_BLOCK blockq, 64, 8, m7
RET
vp9/x86: idct_32x32_add_ssse3 sub-16x16-idct. Runtime of all IDCTs together goes from 3327 to 2473 cycles (intra, i.e. ~35% faster) or from 2312 to 1448 cycles (inter, i.e. ~60% faster). Total decode time of ped1080p.webm goes from 8.086sec to 7.974sec (1.4% faster).
11 years ago
.idct16x16:
mov cntd, 2
.loop1_16x16:
VP9_IDCT32_1D blockq, 1, 16
add blockq, 16
add rsp, 512
dec cntd
jg .loop1_16x16
sub blockq, 32
sub rsp, 1024
mov stride30q, strideq ; stride
lea stride2q, [strideq*2] ; stride*2
shl stride30q, 5 ; stride*32
mov cntd, 4
sub stride30q, stride2q ; stride*30
.loop2_16x16:
mov dstq, dst_bakq
lea dst_endq, [dst_bakq+stride30q]
VP9_IDCT32_1D rsp, 2, 16
add dst_bakq, 8
add rsp, 16
dec cntd
jg .loop2_16x16
sub rsp, 64
; at the end of the loop, m7 should still be zero
; use that to zero out block coefficients
ZERO_BLOCK blockq, 64, 16, m7
RET
vp9/x86: idct_32x32_add_ssse3. Sub-IDCTs will follow later. ped1080.webm goes from 9.295s to 8.191s (13.5% faster). The IDCT itself goes from 4372 (intra) or 4337 (inter) to 403 (intra) or 329 (inter) cycles for the DC-only form, 23755 (intra) or 23723 (inter) to 3497 (intra) or 3607 (inter) cycles for the no-DC form, which averages from 23393 (intra) or 16612 (inter) to 3449 (intra) or 2392 (inter) for all 32x32s together, i.e. about ~7x faster (all tests done on ped1080p.webm).
11 years ago
.idctfull:
mov cntd, 4
.loop1_full:
VP9_IDCT32_1D blockq, 1
add blockq, 16
add rsp, 512
dec cntd
jg .loop1_full
sub blockq, 64
sub rsp, 2048
mov stride30q, strideq ; stride
lea stride2q, [strideq*2] ; stride*2
shl stride30q, 5 ; stride*32
mov cntd, 4
sub stride30q, stride2q ; stride*30
.loop2_full:
mov dstq, dst_bakq
lea dst_endq, [dst_bakq+stride30q]
VP9_IDCT32_1D rsp, 2
add dst_bakq, 8
add rsp, 16
dec cntd
jg .loop2_full
sub rsp, 64
; at the end of the loop, m7 should still be zero
; use that to zero out block coefficients
ZERO_BLOCK blockq, 64, 32, m7
RET
vp9/x86: idct_add_16x16_ssse3. Currently only dc-only and full 16x16. Other subforms will follow in the near future. Total decoding time of ped1080p.webm goes from 9.7 to 9.3 seconds. DC-only goes from 957 -> 131 cycles, and the full IDCT goes from ~4050 to ~745 cycles.
11 years ago
%endif ; x86-64