1 ;*****************************************************************************
2 ;* MMX/SSE2/AVX-optimized 10-bit H.264 iDCT code
3 ;*****************************************************************************
4 ;* Copyright (C) 2005-2011 x264 project
6 ;* Authors: Daniel Kang <daniel.d.kang@gmail.com>
8 ;* This file is part of Libav.
10 ;* Libav is free software; you can redistribute it and/or
11 ;* modify it under the terms of the GNU Lesser General Public
12 ;* License as published by the Free Software Foundation; either
13 ;* version 2.1 of the License, or (at your option) any later version.
15 ;* Libav is distributed in the hope that it will be useful,
16 ;* but WITHOUT ANY WARRANTY; without even the implied warranty of
17 ;* MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the GNU
18 ;* Lesser General Public License for more details.
20 ;* You should have received a copy of the GNU Lesser General Public
21 ;* License along with Libav; if not, write to the Free Software
22 ;* Foundation, Inc., 51 Franklin Street, Fifth Floor, Boston, MA 02110-1301 USA
23 ;******************************************************************************
25 %include "libavutil/x86/x86inc.asm"
26 %include "libavutil/x86/x86util.asm"
30 pw_pixel_max: times 8 dw ((1 << 10)-1)
35 ;-----------------------------------------------------------------------------
36 ; void h264_idct_add(pixel *dst, dctcoef *block, int stride)
37 ;-----------------------------------------------------------------------------
45 CLIPW %1, %4, [pw_pixel_max]
65 IDCT4_1D d,0,1,2,3,4,5
66 TRANSPOSE4x4D 0,1,2,3,4
68 IDCT4_1D d,0,1,2,3,4,5
70 STORE_DIFFx2 m0, m1, m4, m5, %1, %3
72 STORE_DIFFx2 m2, m3, m4, m5, %1, %3
76 cglobal h264_idct_add_10_%1, 3,3
77 IDCT4_ADD_10 r0, r1, r2
88 ;-----------------------------------------------------------------------------
89 ; h264_idct_add16(pixel *dst, const int *block_offset, dctcoef *block, int stride, const uint8_t nnzc[6*8])
90 ;-----------------------------------------------------------------------------
91 ;;;;;;; NO FATE SAMPLES TRIGGER THIS
99 IDCT4_1D d,0,1,2,3,4,5
100 TRANSPOSE4x4D 0,1,2,3,4
102 IDCT4_1D d,0,1,2,3,4,5
104 STORE_DIFFx2 m0, m1, m4, m5, r5, r3
106 STORE_DIFFx2 m2, m3, m4, m5, r5, r3
130 %macro IDCT_ADD16_10 1
131 cglobal h264_idct_add16_10_%1, 5,6
132 ADD16_OP %1, 0, 4+1*8
133 ADD16_OP %1, 1, 5+1*8
134 ADD16_OP %1, 2, 4+2*8
135 ADD16_OP %1, 3, 5+2*8
136 ADD16_OP %1, 4, 6+1*8
137 ADD16_OP %1, 5, 7+1*8
138 ADD16_OP %1, 6, 6+2*8
139 ADD16_OP %1, 7, 7+2*8
140 ADD16_OP %1, 8, 4+3*8
141 ADD16_OP %1, 9, 5+3*8
142 ADD16_OP %1, 10, 4+4*8
143 ADD16_OP %1, 11, 5+4*8
144 ADD16_OP %1, 12, 6+3*8
145 ADD16_OP %1, 13, 7+3*8
146 ADD16_OP %1, 14, 6+4*8
147 ADD16_OP %1, 15, 7+4*8
158 ;-----------------------------------------------------------------------------
159 ; void h264_idct_dc_add(pixel *dst, dctcoef *block, int stride)
160 ;-----------------------------------------------------------------------------
161 %macro IDCT_DC_ADD_OP_10 3
164 paddw m1, m0, [%1+0 ]
165 paddw m2, m0, [%1+%2 ]
166 paddw m3, m0, [%1+%2*2]
167 paddw m4, m0, [%1+%3 ]
189 cglobal h264_idct_dc_add_10_mmx2,3,3
195 mova m6, [pw_pixel_max]
196 IDCT_DC_ADD_OP_10 r0, r2, r1
199 ;-----------------------------------------------------------------------------
200 ; void h264_idct8_dc_add(pixel *dst, dctcoef *block, int stride)
201 ;-----------------------------------------------------------------------------
202 %macro IDCT8_DC_ADD 1
203 cglobal h264_idct8_dc_add_10_%1,3,3,7
210 mova m6, [pw_pixel_max]
211 IDCT_DC_ADD_OP_10 r0, r2, r1
213 IDCT_DC_ADD_OP_10 r0, r2, r1
224 ;-----------------------------------------------------------------------------
225 ; h264_idct_add16intra(pixel *dst, const int *block_offset, dctcoef *block, int stride, const uint8_t nnzc[6*8])
226 ;-----------------------------------------------------------------------------
229 mov r5d, [r1+(%2+0)*4]
231 mov r5d, [r1+(%2+1)*4]
238 %assign last_block 16
239 %macro ADD16_OP_INTRA 3
245 mov r5d, [r1+(%2+0)*4]
254 %macro IDCT_ADD16INTRA_10 1
264 mova m6, [pw_pixel_max]
265 IDCT_DC_ADD_OP_10 r5, r3, r6
268 cglobal h264_idct_add16intra_10_%1,5,7,8
269 ADD16_OP_INTRA %1, 0, 4+1*8
270 ADD16_OP_INTRA %1, 2, 4+2*8
271 ADD16_OP_INTRA %1, 4, 6+1*8
272 ADD16_OP_INTRA %1, 6, 6+2*8
273 ADD16_OP_INTRA %1, 8, 4+3*8
274 ADD16_OP_INTRA %1, 10, 4+4*8
275 ADD16_OP_INTRA %1, 12, 6+3*8
276 ADD16_OP_INTRA %1, 14, 6+4*8
289 IDCT_ADD16INTRA_10 sse2
292 IDCT_ADD16INTRA_10 avx
295 %assign last_block 36
296 ;-----------------------------------------------------------------------------
297 ; h264_idct_add8(pixel **dst, const int *block_offset, dctcoef *block, int stride, const uint8_t nnzc[6*8])
298 ;-----------------------------------------------------------------------------
300 cglobal h264_idct_add8_10_%1,5,8,7
306 ADD16_OP_INTRA %1, 16, 4+ 6*8
307 ADD16_OP_INTRA %1, 18, 4+ 7*8
315 ADD16_OP_INTRA %1, 32, 4+11*8
316 ADD16_OP_INTRA %1, 34, 4+12*8
323 %endmacro ; IDCT_ADD8
332 ;-----------------------------------------------------------------------------
333 ; void h264_idct8_add(pixel *dst, dctcoef *block, int stride)
334 ;-----------------------------------------------------------------------------
380 SWAP 7, 6, 4, 5, 2, 3, 1, 0 ; 70315246 -> 01234567
383 %macro IDCT8_1D_FULL 1
390 IDCT8_1D [%1], [%1+ 64*2]
393 ; %1=int16_t *block, %2=int16_t *dstblock
394 %macro IDCT8_ADD_SSE_START 2
397 TRANSPOSE4x4D 0,1,2,3,8
399 TRANSPOSE4x4D 4,5,6,7,8
403 TRANSPOSE4x4D 0,1,2,3,7
409 TRANSPOSE4x4D 4,5,6,7,3
417 ; %1=uint8_t *dst, %2=int16_t *block, %3=int stride
418 %macro IDCT8_ADD_SSE_END 3
424 STORE_DIFFx2 m0, m1, m6, m7, %1, %3
426 STORE_DIFFx2 m2, m3, m6, m7, %1, %3
430 STORE_DIFFx2 m4, m5, m6, m7, %1, %3
432 STORE_DIFFx2 m0, m1, m6, m7, %1, %3
436 cglobal h264_idct8_add_10_%1, 3,4,16
438 %assign pad 16-gprsize-(stack_offset&15)
440 call h264_idct8_add1_10_%1
446 ; TODO: does not need to use stack
447 h264_idct8_add1_10_%1:
448 %assign pad 256+16-gprsize
453 IDCT8_ADD_SSE_START r1, rsp
460 IDCT8_ADD_SSE_START r1+16, rsp+128
461 PERMUTE 1,9, 2,10, 3,11, 5,1, 6,2, 7,3, 9,13, 10,14, 11,15, 13,5, 14,6, 15,7
462 IDCT8_1D [rsp], [rsp+128]
471 IDCT8_1D [rsp+16], [rsp+144]
477 CLIPW m8, m0, [pw_pixel_max]
479 mova m8, [pw_pixel_max]
480 STORE_DIFF16 m9, m1, m0, m8, r0+r2
482 STORE_DIFF16 m10, m2, m0, m8, r0
483 STORE_DIFF16 m11, m3, m0, m8, r0+r2
485 STORE_DIFF16 m12, m4, m0, m8, r0
486 STORE_DIFF16 m13, m5, m0, m8, r0+r2
488 STORE_DIFF16 m14, m6, m0, m8, r0
489 STORE_DIFF16 m15, m7, m0, m8, r0+r2
491 IDCT8_ADD_SSE_START r1, rsp
492 IDCT8_ADD_SSE_START r1+16, rsp+128
494 IDCT8_ADD_SSE_END r0, rsp, r2
495 IDCT8_ADD_SSE_END r3, rsp+16, r2
509 ;-----------------------------------------------------------------------------
510 ; h264_idct8_add4(pixel **dst, const int *block_offset, dctcoef *block, int stride, const uint8_t nnzc[6*8])
511 ;-----------------------------------------------------------------------------
512 ;;;;;;; NO FATE SAMPLES TRIGGER THIS
513 %macro IDCT8_ADD4_OP 3
518 call h264_idct8_add1_10_%1
526 cglobal h264_idct8_add4_10_%1, 0,7,16
527 %assign pad 16-gprsize-(stack_offset&15)
534 IDCT8_ADD4_OP %1, 0, 4+1*8
535 IDCT8_ADD4_OP %1, 4, 6+1*8
536 IDCT8_ADD4_OP %1, 8, 4+3*8
537 IDCT8_ADD4_OP %1, 12, 6+3*8
540 %endmacro ; IDCT8_ADD4