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 ff_h264_idct_add_10(pixel *dst, int16_t *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 ; void ff_h264_idct_add16_10(pixel *dst, const int *block_offset,
91 ; int16_t *block, int stride,
92 ; 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
128 call add4x4_idct %+ SUFFIX
135 %macro IDCT_ADD16_10 0
136 cglobal h264_idct_add16_10, 5,6
161 ;-----------------------------------------------------------------------------
162 ; void ff_h264_idct_dc_add_10(pixel *dst, int16_t *block, int stride)
163 ;-----------------------------------------------------------------------------
164 %macro IDCT_DC_ADD_OP_10 3
167 paddw m1, m0, [%1+0 ]
168 paddw m2, m0, [%1+%2 ]
169 paddw m3, m0, [%1+%2*2]
170 paddw m4, m0, [%1+%3 ]
192 cglobal h264_idct_dc_add_10,3,3
199 mova m6, [pw_pixel_max]
200 IDCT_DC_ADD_OP_10 r0, r2, r1
203 ;-----------------------------------------------------------------------------
204 ; void ff_h264_idct8_dc_add_10(pixel *dst, int16_t *block, int stride)
205 ;-----------------------------------------------------------------------------
206 %macro IDCT8_DC_ADD 0
207 cglobal h264_idct8_dc_add_10,3,4,7
214 mova m6, [pw_pixel_max]
215 IDCT_DC_ADD_OP_10 r0, r2, r1
217 IDCT_DC_ADD_OP_10 r0, r2, r1
226 ;-----------------------------------------------------------------------------
227 ; void ff_h264_idct_add16intra_10(pixel *dst, const int *block_offset,
228 ; int16_t *block, int stride,
229 ; const uint8_t nnzc[6*8])
230 ;-----------------------------------------------------------------------------
233 mov r5d, [r1+(%1+0)*4]
234 call add4x4_idct %+ SUFFIX
235 mov r5d, [r1+(%1+1)*4]
237 call add4x4_idct %+ SUFFIX
242 %assign last_block 16
243 %macro ADD16_OP_INTRA 2
249 mov r5d, [r1+(%1+0)*4]
250 call idct_dc_add %+ SUFFIX
258 %macro IDCT_ADD16INTRA_10 0
259 idct_dc_add %+ SUFFIX:
270 mova m6, [pw_pixel_max]
271 IDCT_DC_ADD_OP_10 r5, r3, r6
274 cglobal h264_idct_add16intra_10,5,7,8
275 ADD16_OP_INTRA 0, 4+1*8
276 ADD16_OP_INTRA 2, 4+2*8
277 ADD16_OP_INTRA 4, 6+1*8
278 ADD16_OP_INTRA 6, 6+2*8
279 ADD16_OP_INTRA 8, 4+3*8
280 ADD16_OP_INTRA 10, 4+4*8
281 ADD16_OP_INTRA 12, 6+3*8
282 ADD16_OP_INTRA 14, 6+4*8
299 %assign last_block 36
300 ;-----------------------------------------------------------------------------
301 ; void ff_h264_idct_add8_10(pixel **dst, const int *block_offset,
302 ; int16_t *block, int stride,
303 ; const uint8_t nnzc[6*8])
304 ;-----------------------------------------------------------------------------
306 cglobal h264_idct_add8_10,5,8,7
312 ADD16_OP_INTRA 16, 4+ 6*8
313 ADD16_OP_INTRA 18, 4+ 7*8
321 ADD16_OP_INTRA 32, 4+11*8
322 ADD16_OP_INTRA 34, 4+12*8
329 %endmacro ; IDCT_ADD8
336 ;-----------------------------------------------------------------------------
337 ; void ff_h264_idct8_add_10(pixel *dst, int16_t *block, int stride)
338 ;-----------------------------------------------------------------------------
384 SWAP 7, 6, 4, 5, 2, 3, 1, 0 ; 70315246 -> 01234567
387 %macro IDCT8_1D_FULL 1
394 IDCT8_1D [%1], [%1+ 64*2]
397 ; %1=int16_t *block, %2=int16_t *dstblock
398 %macro IDCT8_ADD_SSE_START 2
401 TRANSPOSE4x4D 0,1,2,3,8
403 TRANSPOSE4x4D 4,5,6,7,8
407 TRANSPOSE4x4D 0,1,2,3,7
413 TRANSPOSE4x4D 4,5,6,7,3
421 ; %1=uint8_t *dst, %2=int16_t *block, %3=int stride
422 %macro IDCT8_ADD_SSE_END 3
428 STORE_DIFFx2 m0, m1, m6, m7, %1, %3
430 STORE_DIFFx2 m2, m3, m6, m7, %1, %3
434 STORE_DIFFx2 m4, m5, m6, m7, %1, %3
436 STORE_DIFFx2 m0, m1, m6, m7, %1, %3
440 cglobal h264_idct8_add_10, 3,4,16
442 %assign pad 16-gprsize-(stack_offset&15)
444 call h264_idct8_add1_10 %+ SUFFIX
450 ; TODO: does not need to use stack
451 h264_idct8_add1_10 %+ SUFFIX:
452 %assign pad 256+16-gprsize
457 IDCT8_ADD_SSE_START r1, rsp
464 IDCT8_ADD_SSE_START r1+16, rsp+128
465 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
466 IDCT8_1D [rsp], [rsp+128]
475 IDCT8_1D [rsp+16], [rsp+144]
497 CLIPW m8, m0, [pw_pixel_max]
499 mova m8, [pw_pixel_max]
500 STORE_DIFF16 m9, m1, m0, m8, r0+r2
502 STORE_DIFF16 m10, m2, m0, m8, r0
503 STORE_DIFF16 m11, m3, m0, m8, r0+r2
505 STORE_DIFF16 m12, m4, m0, m8, r0
506 STORE_DIFF16 m13, m5, m0, m8, r0+r2
508 STORE_DIFF16 m14, m6, m0, m8, r0
509 STORE_DIFF16 m15, m7, m0, m8, r0+r2
511 IDCT8_ADD_SSE_START r1, rsp
512 IDCT8_ADD_SSE_START r1+16, rsp+128
514 IDCT8_ADD_SSE_END r0, rsp, r2
515 IDCT8_ADD_SSE_END r3, rsp+16, r2
543 ;-----------------------------------------------------------------------------
544 ; void ff_h264_idct8_add4_10(pixel **dst, const int *block_offset,
545 ; int16_t *block, int stride,
546 ; const uint8_t nnzc[6*8])
547 ;-----------------------------------------------------------------------------
548 ;;;;;;; NO FATE SAMPLES TRIGGER THIS
549 %macro IDCT8_ADD4_OP 2
554 call h264_idct8_add1_10 %+ SUFFIX
562 cglobal h264_idct8_add4_10, 0,7,16
563 %assign pad 16-gprsize-(stack_offset&15)
570 IDCT8_ADD4_OP 0, 4+1*8
571 IDCT8_ADD4_OP 4, 6+1*8
572 IDCT8_ADD4_OP 8, 4+3*8
573 IDCT8_ADD4_OP 12, 6+3*8
576 %endmacro ; IDCT8_ADD4