src/libFLAC/stream_decoder.c : Fix NULL de-reference.
[flac.git] / src / libFLAC / stream_encoder_intrin_sse2.c
1 /* libFLAC - Free Lossless Audio Codec library
2  * Copyright (C) 2000-2009  Josh Coalson
3  * Copyright (C) 2011-2014  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 #ifndef FLAC__NO_ASM
38 #if (defined FLAC__CPU_IA32 || defined FLAC__CPU_X86_64) && defined FLAC__HAS_X86INTRIN
39 #include "private/stream_encoder.h"
40 #include "private/bitmath.h"
41 #ifdef FLAC__SSE2_SUPPORTED
42
43 #include <stdlib.h>    /* for abs() */
44 #include <emmintrin.h> /* SSE2 */
45 #include "FLAC/assert.h"
46
47 FLAC__SSE_TARGET("sse2")
48 void FLAC__precompute_partition_info_sums_intrin_sse2(const FLAC__int32 residual[], FLAC__uint64 abs_residual_partition_sums[],
49                 unsigned residual_samples, unsigned predictor_order, unsigned min_partition_order, unsigned max_partition_order, unsigned bps)
50 {
51         const unsigned default_partition_samples = (residual_samples + predictor_order) >> max_partition_order;
52         unsigned partitions = 1u << max_partition_order;
53
54         FLAC__ASSERT(default_partition_samples > predictor_order);
55
56         /* first do max_partition_order */
57         {
58                 unsigned partition, residual_sample, end = (unsigned)(-(int)predictor_order);
59                 unsigned e1, e3;
60                 __m128i mm_res, mm_sum, mm_mask;
61
62                 if(FLAC__bitmath_ilog2(default_partition_samples) + bps + FLAC__MAX_EXTRA_RESIDUAL_BPS < 32) {
63                         for(partition = residual_sample = 0; partition < partitions; partition++) {
64                                 end += default_partition_samples;
65                                 mm_sum = _mm_setzero_si128();
66
67                                 e1 = (residual_sample + 3) & ~3; e3 = end & ~3;
68                                 if(e1 > end)
69                                         e1 = end; /* try flac -l 1 -b 16 and you'll be here */
70
71                                 /* assumption: residual[] is properly aligned so (residual + e1) is properly aligned too and _mm_loadu_si128() is fast */
72                                 for( ; residual_sample < e1; residual_sample++) {
73                                         mm_res = _mm_cvtsi32_si128(residual[residual_sample]);
74                                         mm_mask = _mm_srai_epi32(mm_res, 31);
75                                         mm_res = _mm_xor_si128(mm_res, mm_mask);
76                                         mm_res = _mm_sub_epi32(mm_res, mm_mask); /* abs(INT_MIN) is undefined, but if the residual is INT_MIN we have bigger problems */
77                                         mm_sum = _mm_add_epi32(mm_sum, mm_res);
78                                 }
79
80                                 for( ; residual_sample < e3; residual_sample+=4) {
81                                         mm_res = _mm_loadu_si128((const __m128i*)(residual+residual_sample));
82                                         mm_mask = _mm_srai_epi32(mm_res, 31);
83                                         mm_res = _mm_xor_si128(mm_res, mm_mask);
84                                         mm_res = _mm_sub_epi32(mm_res, mm_mask);
85                                         mm_sum = _mm_add_epi32(mm_sum, mm_res);
86                                 }
87
88                                 for( ; residual_sample < end; residual_sample++) {
89                                         mm_res = _mm_cvtsi32_si128(residual[residual_sample]);
90                                         mm_mask = _mm_srai_epi32(mm_res, 31);
91                                         mm_res = _mm_xor_si128(mm_res, mm_mask);
92                                         mm_res = _mm_sub_epi32(mm_res, mm_mask);
93                                         mm_sum = _mm_add_epi32(mm_sum, mm_res);
94                                 }
95
96                                 mm_sum = _mm_add_epi32(mm_sum, _mm_srli_si128(mm_sum, 8));
97                                 mm_sum = _mm_add_epi32(mm_sum, _mm_srli_si128(mm_sum, 4));
98                                 abs_residual_partition_sums[partition] = (FLAC__uint32)_mm_cvtsi128_si32(mm_sum);
99                         }
100                 }
101                 else { /* have to pessimistically use 64 bits for accumulator */
102                         for(partition = residual_sample = 0; partition < partitions; partition++) {
103                                 end += default_partition_samples;
104                                 mm_sum = _mm_setzero_si128();
105
106                                 e1 = (residual_sample + 1) & ~1; e3 = end & ~1;
107                                 FLAC__ASSERT(e1 <= end);
108
109                                 for( ; residual_sample < e1; residual_sample++) {
110                                         mm_res = _mm_cvtsi32_si128(residual[residual_sample]); /*  0   0   0   r0 */
111                                         mm_mask = _mm_srai_epi32(mm_res, 31);
112                                         mm_res = _mm_xor_si128(mm_res, mm_mask);
113                                         mm_res = _mm_sub_epi32(mm_res, mm_mask); /*  0   0   0  |r0|  ==   00   |r0_64| */
114                                         mm_sum = _mm_add_epi64(mm_sum, mm_res);
115                                 }
116
117                                 for( ; residual_sample < e3; residual_sample+=2) {
118                                         mm_res = _mm_loadl_epi64((const __m128i*)(residual+residual_sample)); /*  0   0   r1  r0 */
119                                         mm_mask = _mm_srai_epi32(mm_res, 31);
120                                         mm_res = _mm_xor_si128(mm_res, mm_mask);
121                                         mm_res = _mm_sub_epi32(mm_res, mm_mask); /*  0   0  |r1|   |r0| */
122                                         mm_res = _mm_shuffle_epi32(mm_res, _MM_SHUFFLE(3,1,2,0)); /* 0  |r1|  0  |r0|  ==  |r1_64|  |r0_64|  */
123                                         mm_sum = _mm_add_epi64(mm_sum, mm_res);
124                                 }
125
126                                 for( ; residual_sample < end; residual_sample++) {
127                                         mm_res = _mm_cvtsi32_si128(residual[residual_sample]);
128                                         mm_mask = _mm_srai_epi32(mm_res, 31);
129                                         mm_res = _mm_xor_si128(mm_res, mm_mask);
130                                         mm_res = _mm_sub_epi32(mm_res, mm_mask);
131                                         mm_sum = _mm_add_epi64(mm_sum, mm_res);
132                                 }
133
134                                 mm_sum = _mm_add_epi64(mm_sum, _mm_srli_si128(mm_sum, 8));
135                                 _mm_storel_epi64((__m128i*)(abs_residual_partition_sums+partition), mm_sum);
136                         }
137                 }
138         }
139
140         /* now merge partitions for lower orders */
141         {
142                 unsigned from_partition = 0, to_partition = partitions;
143                 int partition_order;
144                 for(partition_order = (int)max_partition_order - 1; partition_order >= (int)min_partition_order; partition_order--) {
145                         unsigned i;
146                         partitions >>= 1;
147                         for(i = 0; i < partitions; i++) {
148                                 abs_residual_partition_sums[to_partition++] =
149                                         abs_residual_partition_sums[from_partition  ] +
150                                         abs_residual_partition_sums[from_partition+1];
151                                 from_partition += 2;
152                         }
153                 }
154         }
155 }
156
157 #endif /* FLAC__SSE2_SUPPORTED */
158 #endif /* (FLAC__CPU_IA32 || FLAC__CPU_X86_64) && FLAC__HAS_X86INTRIN */
159 #endif /* FLAC__NO_ASM */