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 ;******************************************************************************
22 %include "libavutil/x86/x86util.asm"
26 ; 16777216.0f - used in ff_float_to_fixed24()
27 pf_1_24: times 4 dd 0x4B800000
29 ; used in ff_ac3_compute_mantissa_size()
31 pw_bap_mul1: dw 21846, 21846, 0, 32768, 21846, 21846, 0, 32768
32 pw_bap_mul2: dw 5, 7, 0, 7, 5, 7, 0, 7
34 ; used in ff_ac3_extract_exponents()
36 pd_151: times 4 dd 151
40 ;-----------------------------------------------------------------------------
41 ; void ff_ac3_exponent_min(uint8_t *exp, int num_reuse_blocks, int nb_coefs)
42 ;-----------------------------------------------------------------------------
44 %macro AC3_EXPONENT_MIN 0
45 cglobal ac3_exponent_min, 3, 4, 2, exp, reuse_blks, expn, offset
50 mov offsetq, reuse_blksq
51 mova m0, [expq+offsetq]
55 PMINUB m0, [expq+offsetq], m1
70 %define LOOP_ALIGN ALIGN 16
74 %if HAVE_SSE2_EXTERNAL
80 ;-----------------------------------------------------------------------------
81 ; int ff_ac3_max_msb_abs_int16(const int16_t *src, int len)
83 ; This function uses 2 different methods to calculate a valid result.
84 ; 1) logical 'or' of abs of each element
85 ; This is used for ssse3 because of the pabsw instruction.
86 ; It is also used for mmx because of the lack of min/max instructions.
87 ; 2) calculate min/max for the array, then or(abs(min),abs(max))
88 ; This is used for mmxext and sse2 because they have pminsw/pmaxsw.
89 ;-----------------------------------------------------------------------------
91 ; logical 'or' of 4 or 8 words in an mmx or xmm register into the low word
92 %macro OR_WORDS_HORIZ 2 ; src, tmp
100 %elif cpuflag(mmxext)
115 %macro AC3_MAX_MSB_ABS_INT16 1
116 cglobal ac3_max_msb_abs_int16, 2,2,5, src, len
122 mova m1, [srcq+mmsize]
128 %if notcpuflag(ssse3)
130 mova m1, [srcq+mmsize]
133 ; using memory args is faster for ssse3
135 pabsw m1, [srcq+mmsize]
147 OR_WORDS_HORIZ m2, m0
154 %define ABS2 ABS2_MMX
155 AC3_MAX_MSB_ABS_INT16 or_abs
157 %define ABS2 ABS2_MMXEXT
158 AC3_MAX_MSB_ABS_INT16 min_max
160 AC3_MAX_MSB_ABS_INT16 min_max
162 %define ABS2 ABS2_SSSE3
163 AC3_MAX_MSB_ABS_INT16 or_abs
165 ;-----------------------------------------------------------------------------
166 ; macro used for ff_ac3_lshift_int16() and ff_ac3_rshift_int32()
167 ;-----------------------------------------------------------------------------
169 %macro AC3_SHIFT 3 ; l/r, 16/32, shift instruction, instruction set
170 cglobal ac3_%1shift_int%2, 3, 3, 5, src, len, shift
174 mova m2, [srcq+mmsize ]
175 mova m3, [srcq+mmsize*2]
176 mova m4, [srcq+mmsize*3]
182 mova [srcq+mmsize ], m2
183 mova [srcq+mmsize*2], m3
184 mova [srcq+mmsize*3], m4
186 sub lend, mmsize*32/%2
192 ;-----------------------------------------------------------------------------
193 ; void ff_ac3_lshift_int16(int16_t *src, unsigned int len, unsigned int shift)
194 ;-----------------------------------------------------------------------------
197 AC3_SHIFT l, 16, psllw
199 AC3_SHIFT l, 16, psllw
201 ;-----------------------------------------------------------------------------
202 ; void ff_ac3_rshift_int32(int32_t *src, unsigned int len, unsigned int shift)
203 ;-----------------------------------------------------------------------------
206 AC3_SHIFT r, 32, psrad
208 AC3_SHIFT r, 32, psrad
210 ;-----------------------------------------------------------------------------
211 ; void ff_float_to_fixed24(int32_t *dst, const float *src, unsigned int len)
212 ;-----------------------------------------------------------------------------
214 ; The 3DNow! version is not bit-identical because pf2id uses truncation rather
215 ; than round-to-nearest.
217 cglobal float_to_fixed24, 3, 3, 0, dst, src, len
244 cglobal float_to_fixed24, 3, 3, 3, dst, src, len
269 cglobal float_to_fixed24, 3, 3, 9, dst, src, len
273 movaps m2, [srcq+16 ]
274 movaps m3, [srcq+32 ]
275 movaps m4, [srcq+48 ]
277 movaps m5, [srcq+64 ]
278 movaps m6, [srcq+80 ]
279 movaps m7, [srcq+96 ]
280 movaps m8, [srcq+112]
303 movdqa [dstq+16 ], m2
304 movdqa [dstq+32 ], m3
305 movdqa [dstq+48 ], m4
307 movdqa [dstq+64 ], m5
308 movdqa [dstq+80 ], m6
309 movdqa [dstq+96 ], m7
310 movdqa [dstq+112], m8
322 ;------------------------------------------------------------------------------
323 ; int ff_ac3_compute_mantissa_size(uint16_t mant_cnt[6][16])
324 ;------------------------------------------------------------------------------
326 %macro PHADDD4 2 ; xmm src, xmm tmp
334 cglobal ac3_compute_mantissa_size, 1, 2, 4, mant_cnt, sum
335 movdqa m0, [mant_cntq ]
336 movdqa m1, [mant_cntq+ 1*16]
337 paddw m0, [mant_cntq+ 2*16]
338 paddw m1, [mant_cntq+ 3*16]
339 paddw m0, [mant_cntq+ 4*16]
340 paddw m1, [mant_cntq+ 5*16]
341 paddw m0, [mant_cntq+ 6*16]
342 paddw m1, [mant_cntq+ 7*16]
343 paddw m0, [mant_cntq+ 8*16]
344 paddw m1, [mant_cntq+ 9*16]
345 paddw m0, [mant_cntq+10*16]
346 paddw m1, [mant_cntq+11*16]
347 pmaddwd m0, [ac3_bap_bits ]
348 pmaddwd m1, [ac3_bap_bits+16]
352 movdqa m3, [pw_bap_mul1]
353 movhpd m0, [mant_cntq +2]
354 movlpd m0, [mant_cntq+1*32+2]
355 movhpd m1, [mant_cntq+2*32+2]
356 movlpd m1, [mant_cntq+3*32+2]
357 movhpd m2, [mant_cntq+4*32+2]
358 movlpd m2, [mant_cntq+5*32+2]
364 pmaddwd m0, [pw_bap_mul2]
370 ;------------------------------------------------------------------------------
371 ; void ff_ac3_extract_exponents(uint8_t *exp, int32_t *coef, int nb_coefs)
372 ;------------------------------------------------------------------------------
374 %macro PABSD 1-2 ; src/dst, unused
385 %if HAVE_AMD3DNOW_EXTERNAL
387 cglobal ac3_extract_exponents, 3, 3, 0, exp, coef, len
389 lea coefq, [coefq+4*lenq]
394 movq m0, [coefq+4*lenq ]
395 movq m1, [coefq+4*lenq+8]
421 %macro AC3_EXTRACT_EXPONENTS 0
422 cglobal ac3_extract_exponents, 3, 3, 4, exp, coef, len
424 lea coefq, [coefq+4*lenq]
429 ; move 4 32-bit coefs to xmm0
430 mova m0, [coefq+4*lenq]
433 ; convert to float and extract exponents
440 ; move the lowest byte in each of 4 dwords to the low dword
441 ; NOTE: We cannot just extract the low bytes with pshufb because the dword
442 ; result for 16777215 is -1 due to float inaccuracy. Using packuswb
443 ; clips this to 0, which is the correct exponent.
453 %if HAVE_SSE2_EXTERNAL
455 AC3_EXTRACT_EXPONENTS
457 %if HAVE_SSSE3_EXTERNAL
459 AC3_EXTRACT_EXPONENTS