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"
34 %define pw_pixel_max pw_1023
36 ;-----------------------------------------------------------------------------
37 ; void ff_h264_idct_add_10(pixel *dst, int16_t *block, int stride)
38 ;-----------------------------------------------------------------------------
46 CLIPW %1, %4, [pw_pixel_max]
66 IDCT4_1D d,0,1,2,3,4,5
67 TRANSPOSE4x4D 0,1,2,3,4
69 IDCT4_1D d,0,1,2,3,4,5
75 STORE_DIFFx2 m0, m1, m4, m5, %1, %3
77 STORE_DIFFx2 m2, m3, m4, m5, %1, %3
81 cglobal h264_idct_add_10, 3,3
82 IDCT4_ADD_10 r0, r1, r2
93 ;-----------------------------------------------------------------------------
94 ; void ff_h264_idct_add16_10(pixel *dst, const int *block_offset,
95 ; int16_t *block, int stride,
96 ; const uint8_t nnzc[6*8])
97 ;-----------------------------------------------------------------------------
98 ;;;;;;; NO FATE SAMPLES TRIGGER THIS
100 add4x4_idct %+ SUFFIX:
106 IDCT4_1D d,0,1,2,3,4,5
107 TRANSPOSE4x4D 0,1,2,3,4
109 IDCT4_1D d,0,1,2,3,4,5
115 STORE_DIFFx2 m0, m1, m4, m5, r5, r3
117 STORE_DIFFx2 m2, m3, m4, m5, r5, r3
124 %if HAVE_AVX_EXTERNAL
134 call add4x4_idct %+ SUFFIX
141 %macro IDCT_ADD16_10 0
142 cglobal h264_idct_add16_10, 5,6
164 %if HAVE_AVX_EXTERNAL
169 ;-----------------------------------------------------------------------------
170 ; void ff_h264_idct_dc_add_10(pixel *dst, int16_t *block, int stride)
171 ;-----------------------------------------------------------------------------
172 %macro IDCT_DC_ADD_OP_10 3
175 paddw m1, m0, [%1+0 ]
176 paddw m2, m0, [%1+%2 ]
177 paddw m3, m0, [%1+%2*2]
178 paddw m4, m0, [%1+%3 ]
200 cglobal h264_idct_dc_add_10,3,3
207 mova m6, [pw_pixel_max]
208 IDCT_DC_ADD_OP_10 r0, r2, r1
211 ;-----------------------------------------------------------------------------
212 ; void ff_h264_idct8_dc_add_10(pixel *dst, int16_t *block, int stride)
213 ;-----------------------------------------------------------------------------
214 %macro IDCT8_DC_ADD 0
215 cglobal h264_idct8_dc_add_10,3,4,7
222 mova m6, [pw_pixel_max]
223 IDCT_DC_ADD_OP_10 r0, r2, r1
225 IDCT_DC_ADD_OP_10 r0, r2, r1
231 %if HAVE_AVX_EXTERNAL
236 ;-----------------------------------------------------------------------------
237 ; void ff_h264_idct_add16intra_10(pixel *dst, const int *block_offset,
238 ; int16_t *block, int stride,
239 ; const uint8_t nnzc[6*8])
240 ;-----------------------------------------------------------------------------
243 mov r5d, [r1+(%1+0)*4]
244 call add4x4_idct %+ SUFFIX
245 mov r5d, [r1+(%1+1)*4]
247 call add4x4_idct %+ SUFFIX
252 %assign last_block 16
253 %macro ADD16_OP_INTRA 2
259 mov r5d, [r1+(%1+0)*4]
260 call idct_dc_add %+ SUFFIX
268 %macro IDCT_ADD16INTRA_10 0
269 idct_dc_add %+ SUFFIX:
280 mova m6, [pw_pixel_max]
281 IDCT_DC_ADD_OP_10 r5, r3, r6
284 cglobal h264_idct_add16intra_10,5,7,8
285 ADD16_OP_INTRA 0, 4+1*8
286 ADD16_OP_INTRA 2, 4+2*8
287 ADD16_OP_INTRA 4, 6+1*8
288 ADD16_OP_INTRA 6, 6+2*8
289 ADD16_OP_INTRA 8, 4+3*8
290 ADD16_OP_INTRA 10, 4+4*8
291 ADD16_OP_INTRA 12, 6+3*8
292 ADD16_OP_INTRA 14, 6+4*8
306 %if HAVE_AVX_EXTERNAL
311 %assign last_block 36
312 ;-----------------------------------------------------------------------------
313 ; void ff_h264_idct_add8_10(pixel **dst, const int *block_offset,
314 ; int16_t *block, int stride,
315 ; const uint8_t nnzc[6*8])
316 ;-----------------------------------------------------------------------------
318 cglobal h264_idct_add8_10,5,8,7
324 ADD16_OP_INTRA 16, 4+ 6*8
325 ADD16_OP_INTRA 18, 4+ 7*8
333 ADD16_OP_INTRA 32, 4+11*8
334 ADD16_OP_INTRA 34, 4+12*8
341 %endmacro ; IDCT_ADD8
345 %if HAVE_AVX_EXTERNAL
350 ;-----------------------------------------------------------------------------
351 ; void ff_h264_idct8_add_10(pixel *dst, int16_t *block, int stride)
352 ;-----------------------------------------------------------------------------
398 SWAP 7, 6, 4, 5, 2, 3, 1, 0 ; 70315246 -> 01234567
401 %macro IDCT8_1D_FULL 1
408 IDCT8_1D [%1], [%1+ 64*2]
411 ; %1=int16_t *block, %2=int16_t *dstblock
412 %macro IDCT8_ADD_SSE_START 2
415 TRANSPOSE4x4D 0,1,2,3,8
417 TRANSPOSE4x4D 4,5,6,7,8
421 TRANSPOSE4x4D 0,1,2,3,7
427 TRANSPOSE4x4D 4,5,6,7,3
435 ; %1=uint8_t *dst, %2=int16_t *block, %3=int stride
436 %macro IDCT8_ADD_SSE_END 3
442 STORE_DIFFx2 m0, m1, m6, m7, %1, %3
444 STORE_DIFFx2 m2, m3, m6, m7, %1, %3
448 STORE_DIFFx2 m4, m5, m6, m7, %1, %3
450 STORE_DIFFx2 m0, m1, m6, m7, %1, %3
454 cglobal h264_idct8_add_10, 3,4,16
456 %assign pad 16-gprsize-(stack_offset&15)
458 call h264_idct8_add1_10 %+ SUFFIX
464 ; TODO: does not need to use stack
465 h264_idct8_add1_10 %+ SUFFIX:
466 %assign pad 256+16-gprsize
471 IDCT8_ADD_SSE_START r1, rsp
478 IDCT8_ADD_SSE_START r1+16, rsp+128
479 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
480 IDCT8_1D [rsp], [rsp+128]
489 IDCT8_1D [rsp+16], [rsp+144]
511 CLIPW m8, m0, [pw_pixel_max]
513 mova m8, [pw_pixel_max]
514 STORE_DIFF16 m9, m1, m0, m8, r0+r2
516 STORE_DIFF16 m10, m2, m0, m8, r0
517 STORE_DIFF16 m11, m3, m0, m8, r0+r2
519 STORE_DIFF16 m12, m4, m0, m8, r0
520 STORE_DIFF16 m13, m5, m0, m8, r0+r2
522 STORE_DIFF16 m14, m6, m0, m8, r0
523 STORE_DIFF16 m15, m7, m0, m8, r0+r2
525 IDCT8_ADD_SSE_START r1, rsp
526 IDCT8_ADD_SSE_START r1+16, rsp+128
528 IDCT8_ADD_SSE_END r0, rsp, r2
529 IDCT8_ADD_SSE_END r3, rsp+16, r2
554 %if HAVE_AVX_EXTERNAL
559 ;-----------------------------------------------------------------------------
560 ; void ff_h264_idct8_add4_10(pixel **dst, const int *block_offset,
561 ; int16_t *block, int stride,
562 ; const uint8_t nnzc[6*8])
563 ;-----------------------------------------------------------------------------
564 ;;;;;;; NO FATE SAMPLES TRIGGER THIS
565 %macro IDCT8_ADD4_OP 2
570 call h264_idct8_add1_10 %+ SUFFIX
578 cglobal h264_idct8_add4_10, 0,7,16
579 %assign pad 16-gprsize-(stack_offset&15)
586 IDCT8_ADD4_OP 0, 4+1*8
587 IDCT8_ADD4_OP 4, 6+1*8
588 IDCT8_ADD4_OP 8, 4+3*8
589 IDCT8_ADD4_OP 12, 6+3*8
592 %endmacro ; IDCT8_ADD4
596 %if HAVE_AVX_EXTERNAL