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 ;******************************************************************************
26 %include "x86util.asm"
30 pw_pixel_max: times 8 dw ((1 << 10)-1)
32 scan8_mem: db 4+1*8, 5+1*8, 4+2*8, 5+2*8
33 db 6+1*8, 7+1*8, 6+2*8, 7+2*8
34 db 4+3*8, 5+3*8, 4+4*8, 5+4*8
35 db 6+3*8, 7+3*8, 6+4*8, 7+4*8
44 %define scan8 scan8_mem
49 ;-----------------------------------------------------------------------------
50 ; void h264_idct_add(pixel *dst, dctcoef *block, int stride)
51 ;-----------------------------------------------------------------------------
59 CLIPW %1, %4, [pw_pixel_max]
79 IDCT4_1D d,0,1,2,3,4,5
80 TRANSPOSE4x4D 0,1,2,3,4
82 IDCT4_1D d,0,1,2,3,4,5
84 STORE_DIFFx2 m0, m1, m4, m5, %1, %3
86 STORE_DIFFx2 m2, m3, m4, m5, %1, %3
90 cglobal h264_idct_add_10_%1, 3,3
91 IDCT4_ADD_10 r0, r1, r2
102 ;-----------------------------------------------------------------------------
103 ; h264_idct_add16(pixel *dst, const int *block_offset, dctcoef *block, int stride, const uint8_t nnzc[6*8])
104 ;-----------------------------------------------------------------------------
105 ;;;;;;; NO FATE SAMPLES TRIGGER THIS
113 IDCT4_1D d,0,1,2,3,4,5
114 TRANSPOSE4x4D 0,1,2,3,4
116 IDCT4_1D d,0,1,2,3,4,5
118 STORE_DIFFx2 m0, m1, m4, m5, r5, r3
120 STORE_DIFFx2 m2, m3, m4, m5, r5, r3
136 mov r5d, dword [r1+%2*4]
144 %macro IDCT_ADD16_10 1
145 cglobal h264_idct_add16_10_%1, 5,6
146 ADD16_OP %1, 0, 4+1*8
147 ADD16_OP %1, 1, 5+1*8
148 ADD16_OP %1, 2, 4+2*8
149 ADD16_OP %1, 3, 5+2*8
150 ADD16_OP %1, 4, 6+1*8
151 ADD16_OP %1, 5, 7+1*8
152 ADD16_OP %1, 6, 6+2*8
153 ADD16_OP %1, 7, 7+2*8
154 ADD16_OP %1, 8, 4+3*8
155 ADD16_OP %1, 9, 5+3*8
156 ADD16_OP %1, 10, 4+4*8
157 ADD16_OP %1, 11, 5+4*8
158 ADD16_OP %1, 12, 6+3*8
159 ADD16_OP %1, 13, 7+3*8
160 ADD16_OP %1, 14, 6+4*8
161 ADD16_OP %1, 15, 7+4*8
172 ;-----------------------------------------------------------------------------
173 ; void h264_idct_dc_add(pixel *dst, dctcoef *block, int stride)
174 ;-----------------------------------------------------------------------------
175 %macro IDCT_DC_ADD_OP_10 3
178 paddw m1, m0, [%1+0 ]
179 paddw m2, m0, [%1+%2 ]
180 paddw m3, m0, [%1+%2*2]
181 paddw m4, m0, [%1+%3 ]
203 cglobal h264_idct_dc_add_10_mmx2,3,3
209 mova m6, [pw_pixel_max]
210 IDCT_DC_ADD_OP_10 r0, r2, r1
213 ;-----------------------------------------------------------------------------
214 ; void h264_idct8_dc_add(pixel *dst, dctcoef *block, int stride)
215 ;-----------------------------------------------------------------------------
216 %macro IDCT8_DC_ADD 1
217 cglobal h264_idct8_dc_add_10_%1,3,3,7
224 mova m6, [pw_pixel_max]
225 IDCT_DC_ADD_OP_10 r0, r2, r1
227 IDCT_DC_ADD_OP_10 r0, r2, r1
238 ;-----------------------------------------------------------------------------
239 ; h264_idct_add16intra(pixel *dst, const int *block_offset, dctcoef *block, int stride, const uint8_t nnzc[6*8])
240 ;-----------------------------------------------------------------------------
243 mov r5d, dword [r1+(%2+0)*4]
245 mov r5d, dword [r1+(%2+1)*4]
252 %assign last_block 16
253 %macro ADD16_OP_INTRA 3
256 mov r5d, dword [r2+ 0]
257 or r5d, dword [r2+64]
259 mov r5d, dword [r1+(%2+0)*4]
268 %macro IDCT_ADD16INTRA_10 1
278 mova m6, [pw_pixel_max]
279 IDCT_DC_ADD_OP_10 r5, r3, r6
282 cglobal h264_idct_add16intra_10_%1,5,7,8
283 ADD16_OP_INTRA %1, 0, 4+1*8
284 ADD16_OP_INTRA %1, 2, 4+2*8
285 ADD16_OP_INTRA %1, 4, 6+1*8
286 ADD16_OP_INTRA %1, 6, 6+2*8
287 ADD16_OP_INTRA %1, 8, 4+3*8
288 ADD16_OP_INTRA %1, 10, 4+4*8
289 ADD16_OP_INTRA %1, 12, 6+3*8
290 ADD16_OP_INTRA %1, 14, 6+4*8
300 IDCT_ADD16INTRA_10 sse2
303 IDCT_ADD16INTRA_10 avx
306 %assign last_block 24
307 ;-----------------------------------------------------------------------------
308 ; h264_idct_add8(pixel **dst, const int *block_offset, dctcoef *block, int stride, const uint8_t nnzc[6*8])
309 ;-----------------------------------------------------------------------------
311 cglobal h264_idct_add8_10_%1,5,7
317 ADD16_OP_INTRA %1, 16, 1+1*8
318 ADD16_OP_INTRA %1, 18, 1+2*8
320 mov r0, [r10+gprsize]
325 ADD16_OP_INTRA %1, 20, 1+4*8
326 ADD16_OP_INTRA %1, 22, 1+5*8
333 %endmacro ; IDCT_ADD8
342 ;-----------------------------------------------------------------------------
343 ; void h264_idct8_add(pixel *dst, dctcoef *block, int stride)
344 ;-----------------------------------------------------------------------------
390 SWAP 7, 6, 4, 5, 2, 3, 1, 0 ; 70315246 -> 01234567
393 %macro IDCT8_1D_FULL 1
400 IDCT8_1D [%1], [%1+ 64*2]
403 ; %1=int16_t *block, %2=int16_t *dstblock
404 %macro IDCT8_ADD_SSE_START 2
407 TRANSPOSE4x4D 0,1,2,3,8
409 TRANSPOSE4x4D 4,5,6,7,8
413 TRANSPOSE4x4D 0,1,2,3,7
419 TRANSPOSE4x4D 4,5,6,7,3
427 ; %1=uint8_t *dst, %2=int16_t *block, %3=int stride
428 %macro IDCT8_ADD_SSE_END 3
434 STORE_DIFFx2 m0, m1, m6, m7, %1, %3
436 STORE_DIFFx2 m2, m3, m6, m7, %1, %3
440 STORE_DIFFx2 m4, m5, m6, m7, %1, %3
442 STORE_DIFFx2 m0, m1, m6, m7, %1, %3
446 cglobal h264_idct8_add_10_%1, 3,4,16
448 %assign pad 16-gprsize-(stack_offset&15)
450 call h264_idct8_add1_10_%1
456 ; TODO: does not need to use stack
457 h264_idct8_add1_10_%1:
458 %assign pad 256+16-gprsize
463 IDCT8_ADD_SSE_START r1, rsp
470 IDCT8_ADD_SSE_START r1+16, rsp+128
471 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
472 IDCT8_1D [rsp], [rsp+128]
481 IDCT8_1D [rsp+16], [rsp+144]
487 CLIPW m8, m0, [pw_pixel_max]
489 mova m8, [pw_pixel_max]
490 STORE_DIFF16 m9, m1, m0, m8, r0+r2
492 STORE_DIFF16 m10, m2, m0, m8, r0
493 STORE_DIFF16 m11, m3, m0, m8, r0+r2
495 STORE_DIFF16 m12, m4, m0, m8, r0
496 STORE_DIFF16 m13, m5, m0, m8, r0+r2
498 STORE_DIFF16 m14, m6, m0, m8, r0
499 STORE_DIFF16 m15, m7, m0, m8, r0+r2
501 IDCT8_ADD_SSE_START r1, rsp
502 IDCT8_ADD_SSE_START r1+16, rsp+128
504 IDCT8_ADD_SSE_END r0, rsp, r2
505 IDCT8_ADD_SSE_END r3, rsp+16, r2
519 ;-----------------------------------------------------------------------------
520 ; h264_idct8_add4(pixel **dst, const int *block_offset, dctcoef *block, int stride, const uint8_t nnzc[6*8])
521 ;-----------------------------------------------------------------------------
522 ;;;;;;; NO FATE SAMPLES TRIGGER THIS
523 %macro IDCT8_ADD4_OP 3
526 mov r0d, dword [r6+%2*4]
528 call h264_idct8_add1_10_%1
536 cglobal h264_idct8_add4_10_%1, 0,7,16
537 %assign pad 16-gprsize-(stack_offset&15)
544 IDCT8_ADD4_OP %1, 0, 4+1*8
545 IDCT8_ADD4_OP %1, 4, 6+1*8
546 IDCT8_ADD4_OP %1, 8, 4+3*8
547 IDCT8_ADD4_OP %1, 12, 6+3*8
550 %endmacro ; IDCT8_ADD4