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 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
91 ;-----------------------------------------------------------------------------
92 ; void ff_h264_idct_add16_10(pixel *dst, const int *block_offset,
93 ; int16_t *block, int stride,
94 ; const uint8_t nnzc[6*8])
95 ;-----------------------------------------------------------------------------
96 ;;;;;;; NO FATE SAMPLES TRIGGER THIS
98 add4x4_idct %+ SUFFIX:
104 IDCT4_1D d,0,1,2,3,4,5
105 TRANSPOSE4x4D 0,1,2,3,4
107 IDCT4_1D d,0,1,2,3,4,5
113 STORE_DIFFx2 m0, m1, m4, m5, r5, r3
115 STORE_DIFFx2 m2, m3, m4, m5, r5, r3
122 %if HAVE_AVX_EXTERNAL
132 call add4x4_idct %+ SUFFIX
139 %macro IDCT_ADD16_10 0
140 cglobal h264_idct_add16_10, 5,6
162 %if HAVE_AVX_EXTERNAL
167 ;-----------------------------------------------------------------------------
168 ; void ff_h264_idct_dc_add_10(pixel *dst, int16_t *block, int stride)
169 ;-----------------------------------------------------------------------------
170 %macro IDCT_DC_ADD_OP_10 3
173 paddw m1, m0, [%1+0 ]
174 paddw m2, m0, [%1+%2 ]
175 paddw m3, m0, [%1+%2*2]
176 paddw m4, m0, [%1+%3 ]
198 cglobal h264_idct_dc_add_10,3,3
205 mova m6, [pw_pixel_max]
206 IDCT_DC_ADD_OP_10 r0, r2, r1
209 ;-----------------------------------------------------------------------------
210 ; void ff_h264_idct8_dc_add_10(pixel *dst, int16_t *block, int stride)
211 ;-----------------------------------------------------------------------------
212 %macro IDCT8_DC_ADD 0
213 cglobal h264_idct8_dc_add_10,3,4,7
220 mova m6, [pw_pixel_max]
221 IDCT_DC_ADD_OP_10 r0, r2, r1
223 IDCT_DC_ADD_OP_10 r0, r2, r1
229 %if HAVE_AVX_EXTERNAL
234 ;-----------------------------------------------------------------------------
235 ; void ff_h264_idct_add16intra_10(pixel *dst, const int *block_offset,
236 ; int16_t *block, int stride,
237 ; const uint8_t nnzc[6*8])
238 ;-----------------------------------------------------------------------------
241 mov r5d, [r1+(%1+0)*4]
242 call add4x4_idct %+ SUFFIX
243 mov r5d, [r1+(%1+1)*4]
245 call add4x4_idct %+ SUFFIX
250 %assign last_block 16
251 %macro ADD16_OP_INTRA 2
257 mov r5d, [r1+(%1+0)*4]
258 call idct_dc_add %+ SUFFIX
266 %macro IDCT_ADD16INTRA_10 0
267 idct_dc_add %+ SUFFIX:
278 mova m6, [pw_pixel_max]
279 IDCT_DC_ADD_OP_10 r5, r3, r6
282 cglobal h264_idct_add16intra_10,5,7,8
283 ADD16_OP_INTRA 0, 4+1*8
284 ADD16_OP_INTRA 2, 4+2*8
285 ADD16_OP_INTRA 4, 6+1*8
286 ADD16_OP_INTRA 6, 6+2*8
287 ADD16_OP_INTRA 8, 4+3*8
288 ADD16_OP_INTRA 10, 4+4*8
289 ADD16_OP_INTRA 12, 6+3*8
290 ADD16_OP_INTRA 14, 6+4*8
304 %if HAVE_AVX_EXTERNAL
309 %assign last_block 36
310 ;-----------------------------------------------------------------------------
311 ; void ff_h264_idct_add8_10(pixel **dst, const int *block_offset,
312 ; int16_t *block, int stride,
313 ; const uint8_t nnzc[6*8])
314 ;-----------------------------------------------------------------------------
316 cglobal h264_idct_add8_10,5,8,7
322 ADD16_OP_INTRA 16, 4+ 6*8
323 ADD16_OP_INTRA 18, 4+ 7*8
331 ADD16_OP_INTRA 32, 4+11*8
332 ADD16_OP_INTRA 34, 4+12*8
339 %endmacro ; IDCT_ADD8
343 %if HAVE_AVX_EXTERNAL
348 ;-----------------------------------------------------------------------------
349 ; void ff_h264_idct8_add_10(pixel *dst, int16_t *block, int stride)
350 ;-----------------------------------------------------------------------------
396 SWAP 7, 6, 4, 5, 2, 3, 1, 0 ; 70315246 -> 01234567
399 %macro IDCT8_1D_FULL 1
406 IDCT8_1D [%1], [%1+ 64*2]
409 ; %1=int16_t *block, %2=int16_t *dstblock
410 %macro IDCT8_ADD_SSE_START 2
413 TRANSPOSE4x4D 0,1,2,3,8
415 TRANSPOSE4x4D 4,5,6,7,8
419 TRANSPOSE4x4D 0,1,2,3,7
425 TRANSPOSE4x4D 4,5,6,7,3
433 ; %1=uint8_t *dst, %2=int16_t *block, %3=int stride
434 %macro IDCT8_ADD_SSE_END 3
440 STORE_DIFFx2 m0, m1, m6, m7, %1, %3
442 STORE_DIFFx2 m2, m3, m6, m7, %1, %3
446 STORE_DIFFx2 m4, m5, m6, m7, %1, %3
448 STORE_DIFFx2 m0, m1, m6, m7, %1, %3
452 cglobal h264_idct8_add_10, 3,4,16
454 %assign pad 16-gprsize-(stack_offset&15)
456 call h264_idct8_add1_10 %+ SUFFIX
462 ; TODO: does not need to use stack
463 h264_idct8_add1_10 %+ SUFFIX:
464 %assign pad 256+16-gprsize
469 IDCT8_ADD_SSE_START r1, rsp
476 IDCT8_ADD_SSE_START r1+16, rsp+128
477 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
478 IDCT8_1D [rsp], [rsp+128]
487 IDCT8_1D [rsp+16], [rsp+144]
509 CLIPW m8, m0, [pw_pixel_max]
511 mova m8, [pw_pixel_max]
512 STORE_DIFF16 m9, m1, m0, m8, r0+r2
514 STORE_DIFF16 m10, m2, m0, m8, r0
515 STORE_DIFF16 m11, m3, m0, m8, r0+r2
517 STORE_DIFF16 m12, m4, m0, m8, r0
518 STORE_DIFF16 m13, m5, m0, m8, r0+r2
520 STORE_DIFF16 m14, m6, m0, m8, r0
521 STORE_DIFF16 m15, m7, m0, m8, r0+r2
523 IDCT8_ADD_SSE_START r1, rsp
524 IDCT8_ADD_SSE_START r1+16, rsp+128
526 IDCT8_ADD_SSE_END r0, rsp, r2
527 IDCT8_ADD_SSE_END r3, rsp+16, r2
552 %if HAVE_AVX_EXTERNAL
557 ;-----------------------------------------------------------------------------
558 ; void ff_h264_idct8_add4_10(pixel **dst, const int *block_offset,
559 ; int16_t *block, int stride,
560 ; const uint8_t nnzc[6*8])
561 ;-----------------------------------------------------------------------------
562 ;;;;;;; NO FATE SAMPLES TRIGGER THIS
563 %macro IDCT8_ADD4_OP 2
568 call h264_idct8_add1_10 %+ SUFFIX
576 cglobal h264_idct8_add4_10, 0,7,16
577 %assign pad 16-gprsize-(stack_offset&15)
584 IDCT8_ADD4_OP 0, 4+1*8
585 IDCT8_ADD4_OP 4, 6+1*8
586 IDCT8_ADD4_OP 8, 4+3*8
587 IDCT8_ADD4_OP 12, 6+3*8
590 %endmacro ; IDCT8_ADD4
594 %if HAVE_AVX_EXTERNAL