1 ;*****************************************************************************
2 ;* x86-optimized AC-3 DSP utils
3 ;* Copyright (c) 2011 Justin Ruggles
5 ;* This file is part of Libav.
7 ;* Libav is free software; you can redistribute it and/or
8 ;* modify it under the terms of the GNU Lesser General Public
9 ;* License as published by the Free Software Foundation; either
10 ;* version 2.1 of the License, or (at your option) any later version.
12 ;* Libav is distributed in the hope that it will be useful,
13 ;* but WITHOUT ANY WARRANTY; without even the implied warranty of
14 ;* MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the GNU
15 ;* Lesser General Public License for more details.
17 ;* You should have received a copy of the GNU Lesser General Public
18 ;* License along with Libav; if not, write to the Free Software
19 ;* Foundation, Inc., 51 Franklin Street, Fifth Floor, Boston, MA 02110-1301 USA
20 ;******************************************************************************
23 %include "x86util.asm"
27 ; 16777216.0f - used in ff_float_to_fixed24()
28 pf_1_24: times 4 dd 0x4B800000
30 ; used in ff_ac3_compute_mantissa_size()
32 pw_bap_mul1: dw 21846, 21846, 0, 32768, 21846, 21846, 0, 32768
33 pw_bap_mul2: dw 5, 7, 0, 7, 5, 7, 0, 7
35 ; used in ff_ac3_extract_exponents()
37 pd_151: times 4 dd 151
41 ;-----------------------------------------------------------------------------
42 ; void ff_ac3_exponent_min(uint8_t *exp, int num_reuse_blocks, int nb_coefs)
43 ;-----------------------------------------------------------------------------
45 %macro AC3_EXPONENT_MIN 1
46 cglobal ac3_exponent_min_%1, 3,4,2, exp, reuse_blks, expn, offset
51 mov offsetq, reuse_blksq
52 mova m0, [expq+offsetq]
56 PMINUB m0, [expq+offsetq], m1
67 %define PMINUB PMINUB_MMX
72 %define PMINUB PMINUB_MMXEXT
73 %define LOOP_ALIGN ALIGN 16
74 AC3_EXPONENT_MIN mmxext
83 ;-----------------------------------------------------------------------------
84 ; int ff_ac3_max_msb_abs_int16(const int16_t *src, int len)
86 ; This function uses 2 different methods to calculate a valid result.
87 ; 1) logical 'or' of abs of each element
88 ; This is used for ssse3 because of the pabsw instruction.
89 ; It is also used for mmx because of the lack of min/max instructions.
90 ; 2) calculate min/max for the array, then or(abs(min),abs(max))
91 ; This is used for mmxext and sse2 because they have pminsw/pmaxsw.
92 ;-----------------------------------------------------------------------------
94 ; logical 'or' of 4 or 8 words in an mmx or xmm register into the low word
95 %macro OR_WORDS_HORIZ 2 ; src, tmp
101 pshuflw %2, %1, q0001
118 %macro AC3_MAX_MSB_ABS_INT16 1
119 cglobal ac3_max_msb_abs_int16, 2,2,5, src, len
125 mova m1, [srcq+mmsize]
131 %if notcpuflag(ssse3)
133 mova m1, [srcq+mmsize]
136 ; using memory args is faster for ssse3
138 pabsw m1, [srcq+mmsize]
150 OR_WORDS_HORIZ m2, m0
157 %define ABS2 ABS2_MMX
158 AC3_MAX_MSB_ABS_INT16 or_abs
160 %define ABS2 ABS2_MMX2
161 AC3_MAX_MSB_ABS_INT16 min_max
163 AC3_MAX_MSB_ABS_INT16 min_max
165 %define ABS2 ABS2_SSSE3
166 AC3_MAX_MSB_ABS_INT16 or_abs
168 ;-----------------------------------------------------------------------------
169 ; macro used for ff_ac3_lshift_int16() and ff_ac3_rshift_int32()
170 ;-----------------------------------------------------------------------------
172 %macro AC3_SHIFT 4 ; l/r, 16/32, shift instruction, instruction set
173 cglobal ac3_%1shift_int%2_%4, 3,3,5, src, len, shift
177 mova m2, [srcq+mmsize ]
178 mova m3, [srcq+mmsize*2]
179 mova m4, [srcq+mmsize*3]
185 mova [srcq+mmsize ], m2
186 mova [srcq+mmsize*2], m3
187 mova [srcq+mmsize*3], m4
189 sub lend, mmsize*32/%2
195 ;-----------------------------------------------------------------------------
196 ; void ff_ac3_lshift_int16(int16_t *src, unsigned int len, unsigned int shift)
197 ;-----------------------------------------------------------------------------
200 AC3_SHIFT l, 16, psllw, mmx
202 AC3_SHIFT l, 16, psllw, sse2
204 ;-----------------------------------------------------------------------------
205 ; void ff_ac3_rshift_int32(int32_t *src, unsigned int len, unsigned int shift)
206 ;-----------------------------------------------------------------------------
209 AC3_SHIFT r, 32, psrad, mmx
211 AC3_SHIFT r, 32, psrad, sse2
213 ;-----------------------------------------------------------------------------
214 ; void ff_float_to_fixed24(int32_t *dst, const float *src, unsigned int len)
215 ;-----------------------------------------------------------------------------
217 ; The 3DNow! version is not bit-identical because pf2id uses truncation rather
218 ; than round-to-nearest.
220 cglobal float_to_fixed24_3dnow, 3,3,0, dst, src, len
247 cglobal float_to_fixed24_sse, 3,3,3, dst, src, len
272 cglobal float_to_fixed24_sse2, 3,3,9, dst, src, len
276 movaps m2, [srcq+16 ]
277 movaps m3, [srcq+32 ]
278 movaps m4, [srcq+48 ]
280 movaps m5, [srcq+64 ]
281 movaps m6, [srcq+80 ]
282 movaps m7, [srcq+96 ]
283 movaps m8, [srcq+112]
306 movdqa [dstq+16 ], m2
307 movdqa [dstq+32 ], m3
308 movdqa [dstq+48 ], m4
310 movdqa [dstq+64 ], m5
311 movdqa [dstq+80 ], m6
312 movdqa [dstq+96 ], m7
313 movdqa [dstq+112], m8
325 ;------------------------------------------------------------------------------
326 ; int ff_ac3_compute_mantissa_size(uint16_t mant_cnt[6][16])
327 ;------------------------------------------------------------------------------
329 %macro PHADDD4 2 ; xmm src, xmm tmp
337 cglobal ac3_compute_mantissa_size_sse2, 1,2,4, mant_cnt, sum
338 movdqa m0, [mant_cntq ]
339 movdqa m1, [mant_cntq+ 1*16]
340 paddw m0, [mant_cntq+ 2*16]
341 paddw m1, [mant_cntq+ 3*16]
342 paddw m0, [mant_cntq+ 4*16]
343 paddw m1, [mant_cntq+ 5*16]
344 paddw m0, [mant_cntq+ 6*16]
345 paddw m1, [mant_cntq+ 7*16]
346 paddw m0, [mant_cntq+ 8*16]
347 paddw m1, [mant_cntq+ 9*16]
348 paddw m0, [mant_cntq+10*16]
349 paddw m1, [mant_cntq+11*16]
350 pmaddwd m0, [ac3_bap_bits ]
351 pmaddwd m1, [ac3_bap_bits+16]
355 movdqa m3, [pw_bap_mul1]
356 movhpd m0, [mant_cntq +2]
357 movlpd m0, [mant_cntq+1*32+2]
358 movhpd m1, [mant_cntq+2*32+2]
359 movlpd m1, [mant_cntq+3*32+2]
360 movhpd m2, [mant_cntq+4*32+2]
361 movlpd m2, [mant_cntq+5*32+2]
367 pmaddwd m0, [pw_bap_mul2]
373 ;------------------------------------------------------------------------------
374 ; void ff_ac3_extract_exponents(uint8_t *exp, int32_t *coef, int nb_coefs)
375 ;------------------------------------------------------------------------------
377 %macro PABSD_MMX 2 ; src/dst, tmp
384 %macro PABSD_SSSE3 1-2 ; src/dst, unused
390 cglobal ac3_extract_exponents_3dnow, 3,3,0, exp, coef, len
392 lea coefq, [coefq+4*lenq]
397 movq m0, [coefq+4*lenq ]
398 movq m1, [coefq+4*lenq+8]
424 %macro AC3_EXTRACT_EXPONENTS 1
425 cglobal ac3_extract_exponents_%1, 3,3,4, exp, coef, len
427 lea coefq, [coefq+4*lenq]
432 ; move 4 32-bit coefs to xmm0
433 mova m0, [coefq+4*lenq]
436 ; convert to float and extract exponents
443 ; move the lowest byte in each of 4 dwords to the low dword
444 ; NOTE: We cannot just extract the low bytes with pshufb because the dword
445 ; result for 16777215 is -1 due to float inaccuracy. Using packuswb
446 ; clips this to 0, which is the correct exponent.
458 %define PABSD PABSD_MMX
459 AC3_EXTRACT_EXPONENTS sse2
461 %define PABSD PABSD_SSSE3
462 AC3_EXTRACT_EXPONENTS ssse3