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 FFmpeg.
10 ;* FFmpeg 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 ;* FFmpeg 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 FFmpeg; 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/x86util.asm"
29 pw_pixel_max: times 8 dw ((1 << 10)-1)
34 ;-----------------------------------------------------------------------------
35 ; void h264_idct_add(pixel *dst, dctcoef *block, int stride)
36 ;-----------------------------------------------------------------------------
44 CLIPW %1, %4, [pw_pixel_max]
64 IDCT4_1D d,0,1,2,3,4,5
65 TRANSPOSE4x4D 0,1,2,3,4
67 IDCT4_1D d,0,1,2,3,4,5
73 STORE_DIFFx2 m0, m1, m4, m5, %1, %3
75 STORE_DIFFx2 m2, m3, m4, m5, %1, %3
79 cglobal h264_idct_add_10, 3,3
80 IDCT4_ADD_10 r0, r1, r2
91 ;-----------------------------------------------------------------------------
92 ; h264_idct_add16(pixel *dst, const int *block_offset, dctcoef *block, int stride, const uint8_t nnzc[6*8])
93 ;-----------------------------------------------------------------------------
94 ;;;;;;; NO FATE SAMPLES TRIGGER THIS
96 add4x4_idct %+ SUFFIX:
102 IDCT4_1D d,0,1,2,3,4,5
103 TRANSPOSE4x4D 0,1,2,3,4
105 IDCT4_1D d,0,1,2,3,4,5
111 STORE_DIFFx2 m0, m1, m4, m5, r5, r3
113 STORE_DIFFx2 m2, m3, m4, m5, r5, r3
120 %if HAVE_AVX_EXTERNAL
130 call add4x4_idct %+ SUFFIX
137 %macro IDCT_ADD16_10 0
138 cglobal h264_idct_add16_10, 5,6
160 %if HAVE_AVX_EXTERNAL
165 ;-----------------------------------------------------------------------------
166 ; void h264_idct_dc_add(pixel *dst, dctcoef *block, int stride)
167 ;-----------------------------------------------------------------------------
168 %macro IDCT_DC_ADD_OP_10 3
171 paddw m1, m0, [%1+0 ]
172 paddw m2, m0, [%1+%2 ]
173 paddw m3, m0, [%1+%2*2]
174 paddw m4, m0, [%1+%3 ]
196 cglobal h264_idct_dc_add_10,3,3
203 mova m6, [pw_pixel_max]
204 IDCT_DC_ADD_OP_10 r0, r2, r1
207 ;-----------------------------------------------------------------------------
208 ; void h264_idct8_dc_add(pixel *dst, dctcoef *block, int stride)
209 ;-----------------------------------------------------------------------------
210 %macro IDCT8_DC_ADD 0
211 cglobal h264_idct8_dc_add_10,3,4,7
218 mova m6, [pw_pixel_max]
219 IDCT_DC_ADD_OP_10 r0, r2, r1
221 IDCT_DC_ADD_OP_10 r0, r2, r1
227 %if HAVE_AVX_EXTERNAL
232 ;-----------------------------------------------------------------------------
233 ; h264_idct_add16intra(pixel *dst, const int *block_offset, dctcoef *block, int stride, const uint8_t nnzc[6*8])
234 ;-----------------------------------------------------------------------------
237 mov r5d, [r1+(%1+0)*4]
238 call add4x4_idct %+ SUFFIX
239 mov r5d, [r1+(%1+1)*4]
241 call add4x4_idct %+ SUFFIX
246 %assign last_block 16
247 %macro ADD16_OP_INTRA 2
253 mov r5d, [r1+(%1+0)*4]
254 call idct_dc_add %+ SUFFIX
262 %macro IDCT_ADD16INTRA_10 0
263 idct_dc_add %+ SUFFIX:
274 mova m6, [pw_pixel_max]
275 IDCT_DC_ADD_OP_10 r5, r3, r6
278 cglobal h264_idct_add16intra_10,5,7,8
279 ADD16_OP_INTRA 0, 4+1*8
280 ADD16_OP_INTRA 2, 4+2*8
281 ADD16_OP_INTRA 4, 6+1*8
282 ADD16_OP_INTRA 6, 6+2*8
283 ADD16_OP_INTRA 8, 4+3*8
284 ADD16_OP_INTRA 10, 4+4*8
285 ADD16_OP_INTRA 12, 6+3*8
286 ADD16_OP_INTRA 14, 6+4*8
300 %if HAVE_AVX_EXTERNAL
305 %assign last_block 36
306 ;-----------------------------------------------------------------------------
307 ; h264_idct_add8(pixel **dst, const int *block_offset, dctcoef *block, int stride, const uint8_t nnzc[6*8])
308 ;-----------------------------------------------------------------------------
310 cglobal h264_idct_add8_10,5,8,7
316 ADD16_OP_INTRA 16, 4+ 6*8
317 ADD16_OP_INTRA 18, 4+ 7*8
325 ADD16_OP_INTRA 32, 4+11*8
326 ADD16_OP_INTRA 34, 4+12*8
333 %endmacro ; IDCT_ADD8
337 %if HAVE_AVX_EXTERNAL
342 ;-----------------------------------------------------------------------------
343 ; void h264_idct8_add(pixel *dst, dctcoef *block, int stride)
344 ;-----------------------------------------------------------------------------
390 SWAP 7, 6, 4, 5, 2, 3, 1, 0 ; 70315246 -> 01234567
393 %macro IDCT8_1D_FULL 1
400 IDCT8_1D [%1], [%1+ 64*2]
403 ; %1=int16_t *block, %2=int16_t *dstblock
404 %macro IDCT8_ADD_SSE_START 2
407 TRANSPOSE4x4D 0,1,2,3,8
409 TRANSPOSE4x4D 4,5,6,7,8
413 TRANSPOSE4x4D 0,1,2,3,7
419 TRANSPOSE4x4D 4,5,6,7,3
427 ; %1=uint8_t *dst, %2=int16_t *block, %3=int stride
428 %macro IDCT8_ADD_SSE_END 3
434 STORE_DIFFx2 m0, m1, m6, m7, %1, %3
436 STORE_DIFFx2 m2, m3, m6, m7, %1, %3
440 STORE_DIFFx2 m4, m5, m6, m7, %1, %3
442 STORE_DIFFx2 m0, m1, m6, m7, %1, %3
446 cglobal h264_idct8_add_10, 3,4,16
448 %assign pad 16-gprsize-(stack_offset&15)
450 call h264_idct8_add1_10 %+ SUFFIX
456 ; TODO: does not need to use stack
457 h264_idct8_add1_10 %+ SUFFIX:
458 %assign pad 256+16-gprsize
463 IDCT8_ADD_SSE_START r1, rsp
470 IDCT8_ADD_SSE_START r1+16, rsp+128
471 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
472 IDCT8_1D [rsp], [rsp+128]
481 IDCT8_1D [rsp+16], [rsp+144]
503 CLIPW m8, m0, [pw_pixel_max]
505 mova m8, [pw_pixel_max]
506 STORE_DIFF16 m9, m1, m0, m8, r0+r2
508 STORE_DIFF16 m10, m2, m0, m8, r0
509 STORE_DIFF16 m11, m3, m0, m8, r0+r2
511 STORE_DIFF16 m12, m4, m0, m8, r0
512 STORE_DIFF16 m13, m5, m0, m8, r0+r2
514 STORE_DIFF16 m14, m6, m0, m8, r0
515 STORE_DIFF16 m15, m7, m0, m8, r0+r2
517 IDCT8_ADD_SSE_START r1, rsp
518 IDCT8_ADD_SSE_START r1+16, rsp+128
520 IDCT8_ADD_SSE_END r0, rsp, r2
521 IDCT8_ADD_SSE_END r3, rsp+16, r2
546 %if HAVE_AVX_EXTERNAL
551 ;-----------------------------------------------------------------------------
552 ; h264_idct8_add4(pixel **dst, const int *block_offset, dctcoef *block, int stride, const uint8_t nnzc[6*8])
553 ;-----------------------------------------------------------------------------
554 ;;;;;;; NO FATE SAMPLES TRIGGER THIS
555 %macro IDCT8_ADD4_OP 2
560 call h264_idct8_add1_10 %+ SUFFIX
568 cglobal h264_idct8_add4_10, 0,7,16
569 %assign pad 16-gprsize-(stack_offset&15)
576 IDCT8_ADD4_OP 0, 4+1*8
577 IDCT8_ADD4_OP 4, 6+1*8
578 IDCT8_ADD4_OP 8, 4+3*8
579 IDCT8_ADD4_OP 12, 6+3*8
582 %endmacro ; IDCT8_ADD4
586 %if HAVE_AVX_EXTERNAL