ce188d4d |
1 | /* libFLAC - Free Lossless Audio Codec library |
2 | * Copyright (C) 2000-2009 Josh Coalson |
3 | * Copyright (C) 2011-2016 Xiph.Org Foundation |
4 | * |
5 | * Redistribution and use in source and binary forms, with or without |
6 | * modification, are permitted provided that the following conditions |
7 | * are met: |
8 | * |
9 | * - Redistributions of source code must retain the above copyright |
10 | * notice, this list of conditions and the following disclaimer. |
11 | * |
12 | * - Redistributions in binary form must reproduce the above copyright |
13 | * notice, this list of conditions and the following disclaimer in the |
14 | * documentation and/or other materials provided with the distribution. |
15 | * |
16 | * - Neither the name of the Xiph.org Foundation nor the names of its |
17 | * contributors may be used to endorse or promote products derived from |
18 | * this software without specific prior written permission. |
19 | * |
20 | * THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS |
21 | * ``AS IS'' AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT |
22 | * LIMITED TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR |
23 | * A PARTICULAR PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL THE FOUNDATION OR |
24 | * CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, |
25 | * EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT LIMITED TO, |
26 | * PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE, DATA, OR |
27 | * PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY THEORY OF |
28 | * LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT (INCLUDING |
29 | * NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE OF THIS |
30 | * SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE. |
31 | */ |
32 | |
33 | #ifdef HAVE_CONFIG_H |
34 | # include <config.h> |
35 | #endif |
36 | |
37 | #include "private/cpu.h" |
38 | |
39 | #ifndef FLAC__INTEGER_ONLY_LIBRARY |
40 | #ifndef FLAC__NO_ASM |
41 | #if (defined FLAC__CPU_IA32 || defined FLAC__CPU_X86_64) && defined FLAC__HAS_X86INTRIN |
42 | #include "private/fixed.h" |
43 | #ifdef FLAC__SSE2_SUPPORTED |
44 | |
45 | #include <emmintrin.h> /* SSE2 */ |
46 | #include <math.h> |
47 | #include "private/macros.h" |
48 | #include "share/compat.h" |
49 | #include "FLAC/assert.h" |
50 | |
51 | #ifdef FLAC__CPU_IA32 |
52 | #define m128i_to_i64(dest, src) _mm_storel_epi64((__m128i*)&dest, src) |
53 | #else |
54 | #define m128i_to_i64(dest, src) dest = _mm_cvtsi128_si64(src) |
55 | #endif |
56 | |
57 | FLAC__SSE_TARGET("sse2") |
58 | unsigned FLAC__fixed_compute_best_predictor_intrin_sse2(const FLAC__int32 data[], unsigned data_len, float residual_bits_per_sample[FLAC__MAX_FIXED_ORDER + 1]) |
59 | { |
60 | FLAC__uint32 total_error_0, total_error_1, total_error_2, total_error_3, total_error_4; |
61 | unsigned i, order; |
62 | |
63 | __m128i total_err0, total_err1, total_err2; |
64 | |
65 | { |
66 | FLAC__int32 itmp; |
67 | __m128i last_error; |
68 | |
69 | last_error = _mm_cvtsi32_si128(data[-1]); // 0 0 0 le0 |
70 | itmp = data[-2]; |
71 | last_error = _mm_shuffle_epi32(last_error, _MM_SHUFFLE(2,1,0,0)); |
72 | last_error = _mm_sub_epi32(last_error, _mm_cvtsi32_si128(itmp)); // 0 0 le0 le1 |
73 | itmp -= data[-3]; |
74 | last_error = _mm_shuffle_epi32(last_error, _MM_SHUFFLE(2,1,0,0)); |
75 | last_error = _mm_sub_epi32(last_error, _mm_cvtsi32_si128(itmp)); // 0 le0 le1 le2 |
76 | itmp -= data[-3] - data[-4]; |
77 | last_error = _mm_shuffle_epi32(last_error, _MM_SHUFFLE(2,1,0,0)); |
78 | last_error = _mm_sub_epi32(last_error, _mm_cvtsi32_si128(itmp)); // le0 le1 le2 le3 |
79 | |
80 | total_err0 = total_err1 = _mm_setzero_si128(); |
81 | for(i = 0; i < data_len; i++) { |
82 | __m128i err0, err1, tmp; |
83 | err0 = _mm_cvtsi32_si128(data[i]); // 0 0 0 e0 |
84 | err1 = _mm_shuffle_epi32(err0, _MM_SHUFFLE(0,0,0,0)); // e0 e0 e0 e0 |
85 | #if 1 /* OPT_SSE */ |
86 | err1 = _mm_sub_epi32(err1, last_error); |
87 | last_error = _mm_srli_si128(last_error, 4); // 0 le0 le1 le2 |
88 | err1 = _mm_sub_epi32(err1, last_error); |
89 | last_error = _mm_srli_si128(last_error, 4); // 0 0 le0 le1 |
90 | err1 = _mm_sub_epi32(err1, last_error); |
91 | last_error = _mm_srli_si128(last_error, 4); // 0 0 0 le0 |
92 | err1 = _mm_sub_epi32(err1, last_error); // e1 e2 e3 e4 |
93 | #else |
94 | last_error = _mm_add_epi32(last_error, _mm_srli_si128(last_error, 8)); // le0 le1 le2+le0 le3+le1 |
95 | last_error = _mm_add_epi32(last_error, _mm_srli_si128(last_error, 4)); // le0 le1+le0 le2+le0+le1 le3+le1+le2+le0 |
96 | err1 = _mm_sub_epi32(err1, last_error); // e1 e2 e3 e4 |
97 | #endif |
98 | tmp = _mm_slli_si128(err0, 12); // e0 0 0 0 |
99 | last_error = _mm_srli_si128(err1, 4); // 0 e1 e2 e3 |
100 | last_error = _mm_or_si128(last_error, tmp); // e0 e1 e2 e3 |
101 | |
102 | tmp = _mm_srai_epi32(err0, 31); |
103 | err0 = _mm_xor_si128(err0, tmp); |
104 | err0 = _mm_sub_epi32(err0, tmp); |
105 | tmp = _mm_srai_epi32(err1, 31); |
106 | err1 = _mm_xor_si128(err1, tmp); |
107 | err1 = _mm_sub_epi32(err1, tmp); |
108 | |
109 | total_err0 = _mm_add_epi32(total_err0, err0); // 0 0 0 te0 |
110 | total_err1 = _mm_add_epi32(total_err1, err1); // te1 te2 te3 te4 |
111 | } |
112 | } |
113 | |
114 | total_error_0 = _mm_cvtsi128_si32(total_err0); |
115 | total_err2 = total_err1; // te1 te2 te3 te4 |
116 | total_err1 = _mm_srli_si128(total_err1, 8); // 0 0 te1 te2 |
117 | total_error_4 = _mm_cvtsi128_si32(total_err2); |
118 | total_error_2 = _mm_cvtsi128_si32(total_err1); |
119 | total_err2 = _mm_srli_si128(total_err2, 4); // 0 te1 te2 te3 |
120 | total_err1 = _mm_srli_si128(total_err1, 4); // 0 0 0 te1 |
121 | total_error_3 = _mm_cvtsi128_si32(total_err2); |
122 | total_error_1 = _mm_cvtsi128_si32(total_err1); |
123 | |
124 | /* prefer higher order */ |
125 | if(total_error_0 < flac_min(flac_min(flac_min(total_error_1, total_error_2), total_error_3), total_error_4)) |
126 | order = 0; |
127 | else if(total_error_1 < flac_min(flac_min(total_error_2, total_error_3), total_error_4)) |
128 | order = 1; |
129 | else if(total_error_2 < flac_min(total_error_3, total_error_4)) |
130 | order = 2; |
131 | else if(total_error_3 < total_error_4) |
132 | order = 3; |
133 | else |
134 | order = 4; |
135 | |
136 | /* Estimate the expected number of bits per residual signal sample. */ |
137 | /* 'total_error*' is linearly related to the variance of the residual */ |
138 | /* signal, so we use it directly to compute E(|x|) */ |
139 | FLAC__ASSERT(data_len > 0 || total_error_0 == 0); |
140 | FLAC__ASSERT(data_len > 0 || total_error_1 == 0); |
141 | FLAC__ASSERT(data_len > 0 || total_error_2 == 0); |
142 | FLAC__ASSERT(data_len > 0 || total_error_3 == 0); |
143 | FLAC__ASSERT(data_len > 0 || total_error_4 == 0); |
144 | |
145 | residual_bits_per_sample[0] = (float)((total_error_0 > 0) ? log(M_LN2 * (double)total_error_0 / (double)data_len) / M_LN2 : 0.0); |
146 | residual_bits_per_sample[1] = (float)((total_error_1 > 0) ? log(M_LN2 * (double)total_error_1 / (double)data_len) / M_LN2 : 0.0); |
147 | residual_bits_per_sample[2] = (float)((total_error_2 > 0) ? log(M_LN2 * (double)total_error_2 / (double)data_len) / M_LN2 : 0.0); |
148 | residual_bits_per_sample[3] = (float)((total_error_3 > 0) ? log(M_LN2 * (double)total_error_3 / (double)data_len) / M_LN2 : 0.0); |
149 | residual_bits_per_sample[4] = (float)((total_error_4 > 0) ? log(M_LN2 * (double)total_error_4 / (double)data_len) / M_LN2 : 0.0); |
150 | |
151 | return order; |
152 | } |
153 | |
154 | FLAC__SSE_TARGET("sse2") |
155 | unsigned FLAC__fixed_compute_best_predictor_wide_intrin_sse2(const FLAC__int32 data[], unsigned data_len, float residual_bits_per_sample[FLAC__MAX_FIXED_ORDER + 1]) |
156 | { |
157 | FLAC__uint64 total_error_0, total_error_1, total_error_2, total_error_3, total_error_4; |
158 | unsigned i, order; |
159 | |
160 | __m128i total_err0, total_err1, total_err3; |
161 | |
162 | { |
163 | FLAC__int32 itmp; |
164 | __m128i last_error, zero = _mm_setzero_si128(); |
165 | |
166 | last_error = _mm_cvtsi32_si128(data[-1]); // 0 0 0 le0 |
167 | itmp = data[-2]; |
168 | last_error = _mm_shuffle_epi32(last_error, _MM_SHUFFLE(2,1,0,0)); |
169 | last_error = _mm_sub_epi32(last_error, _mm_cvtsi32_si128(itmp)); // 0 0 le0 le1 |
170 | itmp -= data[-3]; |
171 | last_error = _mm_shuffle_epi32(last_error, _MM_SHUFFLE(2,1,0,0)); |
172 | last_error = _mm_sub_epi32(last_error, _mm_cvtsi32_si128(itmp)); // 0 le0 le1 le2 |
173 | itmp -= data[-3] - data[-4]; |
174 | last_error = _mm_shuffle_epi32(last_error, _MM_SHUFFLE(2,1,0,0)); |
175 | last_error = _mm_sub_epi32(last_error, _mm_cvtsi32_si128(itmp)); // le0 le1 le2 le3 |
176 | |
177 | total_err0 = total_err1 = total_err3 = _mm_setzero_si128(); |
178 | for(i = 0; i < data_len; i++) { |
179 | __m128i err0, err1, tmp; |
180 | err0 = _mm_cvtsi32_si128(data[i]); // 0 0 0 e0 |
181 | err1 = _mm_shuffle_epi32(err0, _MM_SHUFFLE(0,0,0,0)); // e0 e0 e0 e0 |
182 | #if 1 /* OPT_SSE */ |
183 | err1 = _mm_sub_epi32(err1, last_error); |
184 | last_error = _mm_srli_si128(last_error, 4); // 0 le0 le1 le2 |
185 | err1 = _mm_sub_epi32(err1, last_error); |
186 | last_error = _mm_srli_si128(last_error, 4); // 0 0 le0 le1 |
187 | err1 = _mm_sub_epi32(err1, last_error); |
188 | last_error = _mm_srli_si128(last_error, 4); // 0 0 0 le0 |
189 | err1 = _mm_sub_epi32(err1, last_error); // e1 e2 e3 e4 |
190 | #else |
191 | last_error = _mm_add_epi32(last_error, _mm_srli_si128(last_error, 8)); // le0 le1 le2+le0 le3+le1 |
192 | last_error = _mm_add_epi32(last_error, _mm_srli_si128(last_error, 4)); // le0 le1+le0 le2+le0+le1 le3+le1+le2+le0 |
193 | err1 = _mm_sub_epi32(err1, last_error); // e1 e2 e3 e4 |
194 | #endif |
195 | tmp = _mm_slli_si128(err0, 12); // e0 0 0 0 |
196 | last_error = _mm_srli_si128(err1, 4); // 0 e1 e2 e3 |
197 | last_error = _mm_or_si128(last_error, tmp); // e0 e1 e2 e3 |
198 | |
199 | tmp = _mm_srai_epi32(err0, 31); |
200 | err0 = _mm_xor_si128(err0, tmp); |
201 | err0 = _mm_sub_epi32(err0, tmp); |
202 | tmp = _mm_srai_epi32(err1, 31); |
203 | err1 = _mm_xor_si128(err1, tmp); |
204 | err1 = _mm_sub_epi32(err1, tmp); |
205 | |
206 | total_err0 = _mm_add_epi64(total_err0, err0); // 0 te0 |
207 | err0 = _mm_unpacklo_epi32(err1, zero); // 0 |e3| 0 |e4| |
208 | err1 = _mm_unpackhi_epi32(err1, zero); // 0 |e1| 0 |e2| |
209 | total_err3 = _mm_add_epi64(total_err3, err0); // te3 te4 |
210 | total_err1 = _mm_add_epi64(total_err1, err1); // te1 te2 |
211 | } |
212 | } |
213 | |
214 | m128i_to_i64(total_error_0, total_err0); |
215 | m128i_to_i64(total_error_4, total_err3); |
216 | m128i_to_i64(total_error_2, total_err1); |
217 | total_err3 = _mm_srli_si128(total_err3, 8); // 0 te3 |
218 | total_err1 = _mm_srli_si128(total_err1, 8); // 0 te1 |
219 | m128i_to_i64(total_error_3, total_err3); |
220 | m128i_to_i64(total_error_1, total_err1); |
221 | |
222 | /* prefer higher order */ |
223 | if(total_error_0 < flac_min(flac_min(flac_min(total_error_1, total_error_2), total_error_3), total_error_4)) |
224 | order = 0; |
225 | else if(total_error_1 < flac_min(flac_min(total_error_2, total_error_3), total_error_4)) |
226 | order = 1; |
227 | else if(total_error_2 < flac_min(total_error_3, total_error_4)) |
228 | order = 2; |
229 | else if(total_error_3 < total_error_4) |
230 | order = 3; |
231 | else |
232 | order = 4; |
233 | |
234 | /* Estimate the expected number of bits per residual signal sample. */ |
235 | /* 'total_error*' is linearly related to the variance of the residual */ |
236 | /* signal, so we use it directly to compute E(|x|) */ |
237 | FLAC__ASSERT(data_len > 0 || total_error_0 == 0); |
238 | FLAC__ASSERT(data_len > 0 || total_error_1 == 0); |
239 | FLAC__ASSERT(data_len > 0 || total_error_2 == 0); |
240 | FLAC__ASSERT(data_len > 0 || total_error_3 == 0); |
241 | FLAC__ASSERT(data_len > 0 || total_error_4 == 0); |
242 | |
243 | residual_bits_per_sample[0] = (float)((total_error_0 > 0) ? log(M_LN2 * (double)total_error_0 / (double)data_len) / M_LN2 : 0.0); |
244 | residual_bits_per_sample[1] = (float)((total_error_1 > 0) ? log(M_LN2 * (double)total_error_1 / (double)data_len) / M_LN2 : 0.0); |
245 | residual_bits_per_sample[2] = (float)((total_error_2 > 0) ? log(M_LN2 * (double)total_error_2 / (double)data_len) / M_LN2 : 0.0); |
246 | residual_bits_per_sample[3] = (float)((total_error_3 > 0) ? log(M_LN2 * (double)total_error_3 / (double)data_len) / M_LN2 : 0.0); |
247 | residual_bits_per_sample[4] = (float)((total_error_4 > 0) ? log(M_LN2 * (double)total_error_4 / (double)data_len) / M_LN2 : 0.0); |
248 | |
249 | return order; |
250 | } |
251 | |
252 | #endif /* FLAC__SSE2_SUPPORTED */ |
253 | #endif /* (FLAC__CPU_IA32 || FLAC__CPU_X86_64) && FLAC__HAS_X86INTRIN */ |
254 | #endif /* FLAC__NO_ASM */ |
255 | #endif /* FLAC__INTEGER_ONLY_LIBRARY */ |