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
69 STORE_DIFFx2 m0, m1, m4, m5, %1, %3
71 STORE_DIFFx2 m2, m3, m4, m5, %1, %3
75 cglobal h264_idct_add_10, 3,3
76 IDCT4_ADD_10 r0, r1, r2
85 ;-----------------------------------------------------------------------------
86 ; h264_idct_add16(pixel *dst, const int *block_offset, dctcoef *block, int stride, const uint8_t nnzc[6*8])
87 ;-----------------------------------------------------------------------------
88 ;;;;;;; NO FATE SAMPLES TRIGGER THIS
90 add4x4_idct %+ SUFFIX:
96 IDCT4_1D d,0,1,2,3,4,5
97 TRANSPOSE4x4D 0,1,2,3,4
99 IDCT4_1D d,0,1,2,3,4,5
101 STORE_DIFFx2 m0, m1, m4, m5, r5, r3
103 STORE_DIFFx2 m2, m3, m4, m5, r5, r3
118 call add4x4_idct %+ SUFFIX
125 %macro IDCT_ADD16_10 0
126 cglobal h264_idct_add16_10, 5,6
151 ;-----------------------------------------------------------------------------
152 ; void h264_idct_dc_add(pixel *dst, dctcoef *block, int stride)
153 ;-----------------------------------------------------------------------------
154 %macro IDCT_DC_ADD_OP_10 3
157 paddw m1, m0, [%1+0 ]
158 paddw m2, m0, [%1+%2 ]
159 paddw m3, m0, [%1+%2*2]
160 paddw m4, m0, [%1+%3 ]
182 cglobal h264_idct_dc_add_10,3,3
188 mova m6, [pw_pixel_max]
189 IDCT_DC_ADD_OP_10 r0, r2, r1
192 ;-----------------------------------------------------------------------------
193 ; void h264_idct8_dc_add(pixel *dst, dctcoef *block, int stride)
194 ;-----------------------------------------------------------------------------
195 %macro IDCT8_DC_ADD 0
196 cglobal h264_idct8_dc_add_10,3,3,7
203 mova m6, [pw_pixel_max]
204 IDCT_DC_ADD_OP_10 r0, r2, r1
206 IDCT_DC_ADD_OP_10 r0, r2, r1
215 ;-----------------------------------------------------------------------------
216 ; h264_idct_add16intra(pixel *dst, const int *block_offset, dctcoef *block, int stride, const uint8_t nnzc[6*8])
217 ;-----------------------------------------------------------------------------
220 mov r5d, [r1+(%1+0)*4]
221 call add4x4_idct %+ SUFFIX
222 mov r5d, [r1+(%1+1)*4]
224 call add4x4_idct %+ SUFFIX
229 %assign last_block 16
230 %macro ADD16_OP_INTRA 2
236 mov r5d, [r1+(%1+0)*4]
237 call idct_dc_add %+ SUFFIX
245 %macro IDCT_ADD16INTRA_10 0
246 idct_dc_add %+ SUFFIX:
255 mova m6, [pw_pixel_max]
256 IDCT_DC_ADD_OP_10 r5, r3, r6
259 cglobal h264_idct_add16intra_10,5,7,8
260 ADD16_OP_INTRA 0, 4+1*8
261 ADD16_OP_INTRA 2, 4+2*8
262 ADD16_OP_INTRA 4, 6+1*8
263 ADD16_OP_INTRA 6, 6+2*8
264 ADD16_OP_INTRA 8, 4+3*8
265 ADD16_OP_INTRA 10, 4+4*8
266 ADD16_OP_INTRA 12, 6+3*8
267 ADD16_OP_INTRA 14, 6+4*8
284 %assign last_block 36
285 ;-----------------------------------------------------------------------------
286 ; h264_idct_add8(pixel **dst, const int *block_offset, dctcoef *block, int stride, const uint8_t nnzc[6*8])
287 ;-----------------------------------------------------------------------------
289 cglobal h264_idct_add8_10,5,8,7
295 ADD16_OP_INTRA 16, 4+ 6*8
296 ADD16_OP_INTRA 18, 4+ 7*8
304 ADD16_OP_INTRA 32, 4+11*8
305 ADD16_OP_INTRA 34, 4+12*8
312 %endmacro ; IDCT_ADD8
319 ;-----------------------------------------------------------------------------
320 ; void h264_idct8_add(pixel *dst, dctcoef *block, int stride)
321 ;-----------------------------------------------------------------------------
367 SWAP 7, 6, 4, 5, 2, 3, 1, 0 ; 70315246 -> 01234567
370 %macro IDCT8_1D_FULL 1
377 IDCT8_1D [%1], [%1+ 64*2]
380 ; %1=int16_t *block, %2=int16_t *dstblock
381 %macro IDCT8_ADD_SSE_START 2
384 TRANSPOSE4x4D 0,1,2,3,8
386 TRANSPOSE4x4D 4,5,6,7,8
390 TRANSPOSE4x4D 0,1,2,3,7
396 TRANSPOSE4x4D 4,5,6,7,3
404 ; %1=uint8_t *dst, %2=int16_t *block, %3=int stride
405 %macro IDCT8_ADD_SSE_END 3
411 STORE_DIFFx2 m0, m1, m6, m7, %1, %3
413 STORE_DIFFx2 m2, m3, m6, m7, %1, %3
417 STORE_DIFFx2 m4, m5, m6, m7, %1, %3
419 STORE_DIFFx2 m0, m1, m6, m7, %1, %3
423 cglobal h264_idct8_add_10, 3,4,16
425 %assign pad 16-gprsize-(stack_offset&15)
427 call h264_idct8_add1_10 %+ SUFFIX
433 ; TODO: does not need to use stack
434 h264_idct8_add1_10 %+ SUFFIX:
435 %assign pad 256+16-gprsize
440 IDCT8_ADD_SSE_START r1, rsp
447 IDCT8_ADD_SSE_START r1+16, rsp+128
448 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
449 IDCT8_1D [rsp], [rsp+128]
458 IDCT8_1D [rsp+16], [rsp+144]
464 CLIPW m8, m0, [pw_pixel_max]
466 mova m8, [pw_pixel_max]
467 STORE_DIFF16 m9, m1, m0, m8, r0+r2
469 STORE_DIFF16 m10, m2, m0, m8, r0
470 STORE_DIFF16 m11, m3, m0, m8, r0+r2
472 STORE_DIFF16 m12, m4, m0, m8, r0
473 STORE_DIFF16 m13, m5, m0, m8, r0+r2
475 STORE_DIFF16 m14, m6, m0, m8, r0
476 STORE_DIFF16 m15, m7, m0, m8, r0+r2
478 IDCT8_ADD_SSE_START r1, rsp
479 IDCT8_ADD_SSE_START r1+16, rsp+128
481 IDCT8_ADD_SSE_END r0, rsp, r2
482 IDCT8_ADD_SSE_END r3, rsp+16, r2
494 ;-----------------------------------------------------------------------------
495 ; h264_idct8_add4(pixel **dst, const int *block_offset, dctcoef *block, int stride, const uint8_t nnzc[6*8])
496 ;-----------------------------------------------------------------------------
497 ;;;;;;; NO FATE SAMPLES TRIGGER THIS
498 %macro IDCT8_ADD4_OP 2
503 call h264_idct8_add1_10 %+ SUFFIX
511 cglobal h264_idct8_add4_10, 0,7,16
512 %assign pad 16-gprsize-(stack_offset&15)
519 IDCT8_ADD4_OP 0, 4+1*8
520 IDCT8_ADD4_OP 4, 6+1*8
521 IDCT8_ADD4_OP 8, 4+3*8
522 IDCT8_ADD4_OP 12, 6+3*8
525 %endmacro ; IDCT8_ADD4