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/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
89 ;-----------------------------------------------------------------------------
90 ; h264_idct_add16(pixel *dst, const int *block_offset, dctcoef *block, int stride, const uint8_t nnzc[6*8])
91 ;-----------------------------------------------------------------------------
92 ;;;;;;; NO FATE SAMPLES TRIGGER THIS
94 add4x4_idct %+ SUFFIX:
100 IDCT4_1D d,0,1,2,3,4,5
101 TRANSPOSE4x4D 0,1,2,3,4
103 IDCT4_1D d,0,1,2,3,4,5
109 STORE_DIFFx2 m0, m1, m4, m5, r5, r3
111 STORE_DIFFx2 m2, m3, m4, m5, r5, r3
126 call add4x4_idct %+ SUFFIX
133 %macro IDCT_ADD16_10 0
134 cglobal h264_idct_add16_10, 5,6
159 ;-----------------------------------------------------------------------------
160 ; void h264_idct_dc_add(pixel *dst, dctcoef *block, int stride)
161 ;-----------------------------------------------------------------------------
162 %macro IDCT_DC_ADD_OP_10 3
165 paddw m1, m0, [%1+0 ]
166 paddw m2, m0, [%1+%2 ]
167 paddw m3, m0, [%1+%2*2]
168 paddw m4, m0, [%1+%3 ]
190 cglobal h264_idct_dc_add_10,3,3
197 mova m6, [pw_pixel_max]
198 IDCT_DC_ADD_OP_10 r0, r2, r1
201 ;-----------------------------------------------------------------------------
202 ; void h264_idct8_dc_add(pixel *dst, dctcoef *block, int stride)
203 ;-----------------------------------------------------------------------------
204 %macro IDCT8_DC_ADD 0
205 cglobal h264_idct8_dc_add_10,3,4,7
212 mova m6, [pw_pixel_max]
213 IDCT_DC_ADD_OP_10 r0, r2, r1
215 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+(%1+0)*4]
230 call add4x4_idct %+ SUFFIX
231 mov r5d, [r1+(%1+1)*4]
233 call add4x4_idct %+ SUFFIX
238 %assign last_block 16
239 %macro ADD16_OP_INTRA 2
245 mov r5d, [r1+(%1+0)*4]
246 call idct_dc_add %+ SUFFIX
254 %macro IDCT_ADD16INTRA_10 0
255 idct_dc_add %+ SUFFIX:
266 mova m6, [pw_pixel_max]
267 IDCT_DC_ADD_OP_10 r5, r3, r6
270 cglobal h264_idct_add16intra_10,5,7,8
271 ADD16_OP_INTRA 0, 4+1*8
272 ADD16_OP_INTRA 2, 4+2*8
273 ADD16_OP_INTRA 4, 6+1*8
274 ADD16_OP_INTRA 6, 6+2*8
275 ADD16_OP_INTRA 8, 4+3*8
276 ADD16_OP_INTRA 10, 4+4*8
277 ADD16_OP_INTRA 12, 6+3*8
278 ADD16_OP_INTRA 14, 6+4*8
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_10,5,8,7
306 ADD16_OP_INTRA 16, 4+ 6*8
307 ADD16_OP_INTRA 18, 4+ 7*8
315 ADD16_OP_INTRA 32, 4+11*8
316 ADD16_OP_INTRA 34, 4+12*8
323 %endmacro ; IDCT_ADD8
330 ;-----------------------------------------------------------------------------
331 ; void h264_idct8_add(pixel *dst, dctcoef *block, int stride)
332 ;-----------------------------------------------------------------------------
378 SWAP 7, 6, 4, 5, 2, 3, 1, 0 ; 70315246 -> 01234567
381 %macro IDCT8_1D_FULL 1
388 IDCT8_1D [%1], [%1+ 64*2]
391 ; %1=int16_t *block, %2=int16_t *dstblock
392 %macro IDCT8_ADD_SSE_START 2
395 TRANSPOSE4x4D 0,1,2,3,8
397 TRANSPOSE4x4D 4,5,6,7,8
401 TRANSPOSE4x4D 0,1,2,3,7
407 TRANSPOSE4x4D 4,5,6,7,3
415 ; %1=uint8_t *dst, %2=int16_t *block, %3=int stride
416 %macro IDCT8_ADD_SSE_END 3
422 STORE_DIFFx2 m0, m1, m6, m7, %1, %3
424 STORE_DIFFx2 m2, m3, m6, m7, %1, %3
428 STORE_DIFFx2 m4, m5, m6, m7, %1, %3
430 STORE_DIFFx2 m0, m1, m6, m7, %1, %3
434 cglobal h264_idct8_add_10, 3,4,16
436 %assign pad 16-gprsize-(stack_offset&15)
438 call h264_idct8_add1_10 %+ SUFFIX
444 ; TODO: does not need to use stack
445 h264_idct8_add1_10 %+ SUFFIX:
446 %assign pad 256+16-gprsize
451 IDCT8_ADD_SSE_START r1, rsp
458 IDCT8_ADD_SSE_START r1+16, rsp+128
459 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
460 IDCT8_1D [rsp], [rsp+128]
469 IDCT8_1D [rsp+16], [rsp+144]
491 CLIPW m8, m0, [pw_pixel_max]
493 mova m8, [pw_pixel_max]
494 STORE_DIFF16 m9, m1, m0, m8, r0+r2
496 STORE_DIFF16 m10, m2, m0, m8, r0
497 STORE_DIFF16 m11, m3, m0, m8, r0+r2
499 STORE_DIFF16 m12, m4, m0, m8, r0
500 STORE_DIFF16 m13, m5, m0, m8, r0+r2
502 STORE_DIFF16 m14, m6, m0, m8, r0
503 STORE_DIFF16 m15, m7, m0, m8, r0+r2
505 IDCT8_ADD_SSE_START r1, rsp
506 IDCT8_ADD_SSE_START r1+16, rsp+128
508 IDCT8_ADD_SSE_END r0, rsp, r2
509 IDCT8_ADD_SSE_END r3, rsp+16, r2
537 ;-----------------------------------------------------------------------------
538 ; h264_idct8_add4(pixel **dst, const int *block_offset, dctcoef *block, int stride, const uint8_t nnzc[6*8])
539 ;-----------------------------------------------------------------------------
540 ;;;;;;; NO FATE SAMPLES TRIGGER THIS
541 %macro IDCT8_ADD4_OP 2
546 call h264_idct8_add1_10 %+ SUFFIX
554 cglobal h264_idct8_add4_10, 0,7,16
555 %assign pad 16-gprsize-(stack_offset&15)
562 IDCT8_ADD4_OP 0, 4+1*8
563 IDCT8_ADD4_OP 4, 6+1*8
564 IDCT8_ADD4_OP 8, 4+3*8
565 IDCT8_ADD4_OP 12, 6+3*8
568 %endmacro ; IDCT8_ADD4