FLAC/src/stream_encoder_intrin_ssse3.c

   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__NO_ASM
  40 #if (defined FLAC__CPU_IA32 || defined FLAC__CPU_X86_64) && FLAC__HAS_X86INTRIN
  41 #include "private/stream_encoder.h"
  42 #include "private/bitmath.h"
  43 #ifdef FLAC__SSSE3_SUPPORTED
  44
  45 #include <stdlib.h>    /* for abs() */
  46 #include <tmmintrin.h> /* SSSE3 */
  47 #include "FLAC/assert.h"
  48
  49 FLAC__SSE_TARGET("ssse3")
  50 void FLAC__precompute_partition_info_sums_intrin_ssse3(const FLAC__int32 residual[], FLAC__uint64 abs_residual_partition_sums[],
  51                 uint32_t residual_samples, uint32_t predictor_order, uint32_t min_partition_order, uint32_t max_partition_order, uint32_t bps)
  52 {
  53         const uint32_t default_partition_samples = (residual_samples + predictor_order) >> max_partition_order;
  54         uint32_t partitions = 1u << max_partition_order;
  55
  56         FLAC__ASSERT(default_partition_samples > predictor_order);
  57
  58         /* first do max_partition_order */
  59         {
  60                 const uint32_t threshold = 32 - FLAC__bitmath_ilog2(default_partition_samples);
  61                 uint32_t partition, residual_sample, end = (uint32_t)(-(int32_t)predictor_order);
  62
  63                 if(bps + FLAC__MAX_EXTRA_RESIDUAL_BPS < threshold) {
  64                         for(partition = residual_sample = 0; partition < partitions; partition++) {
  65                                 __m128i mm_sum = _mm_setzero_si128();
  66                                 uint32_t e1, e3;
  67                                 end += default_partition_samples;
  68
  69                                 e1 = (residual_sample + 3) & ~3; e3 = end & ~3;
  70                                 if(e1 > end)
  71                                         e1 = end; /* try flac -l 1 -b 16 and you'll be here */
  72
  73                                 /* assumption: residual[] is properly aligned so (residual + e1) is properly aligned too and _mm_loadu_si128() is fast */
  74                                 for( ; residual_sample < e1; residual_sample++) {
  75                                         __m128i mm_res = _mm_abs_epi32(_mm_cvtsi32_si128(residual[residual_sample]));
  76                                         mm_sum = _mm_add_epi32(mm_sum, mm_res);
  77                                 }
  78
  79                                 for( ; residual_sample < e3; residual_sample+=4) {
  80                                         __m128i mm_res = _mm_abs_epi32(_mm_loadu_si128((const __m128i*)(residual+residual_sample)));
  81                                         mm_sum = _mm_add_epi32(mm_sum, mm_res);
  82                                 }
  83
  84                                 for( ; residual_sample < end; residual_sample++) {
  85                                         __m128i mm_res = _mm_abs_epi32(_mm_cvtsi32_si128(residual[residual_sample]));
  86                                         mm_sum = _mm_add_epi32(mm_sum, mm_res);
  87                                 }
  88
  89                                 mm_sum = _mm_add_epi32(mm_sum, _mm_shuffle_epi32(mm_sum, _MM_SHUFFLE(1,0,3,2)));
  90                                 mm_sum = _mm_add_epi32(mm_sum, _mm_shufflelo_epi16(mm_sum, _MM_SHUFFLE(1,0,3,2)));
  91                                 abs_residual_partition_sums[partition] = (FLAC__uint32)_mm_cvtsi128_si32(mm_sum);
  92 /* workaround for MSVC bugs (at least versions 2015 and 2017 are affected) */
  93 #if (defined _MSC_VER) && (defined FLAC__CPU_X86_64)
  94                                 abs_residual_partition_sums[partition] &= 0xFFFFFFFF;
  95 #endif
  96                         }
  97                 }
  98                 else { /* have to pessimistically use 64 bits for accumulator */
  99                         for(partition = residual_sample = 0; partition < partitions; partition++) {
 100                                 __m128i mm_sum = _mm_setzero_si128();
 101                                 uint32_t e1, e3;
 102                                 end += default_partition_samples;
 103
 104                                 e1 = (residual_sample + 1) & ~1; e3 = end & ~1;
 105                                 FLAC__ASSERT(e1 <= end);
 106
 107                                 for( ; residual_sample < e1; residual_sample++) {
 108                                         __m128i mm_res = _mm_abs_epi32(_mm_cvtsi32_si128(residual[residual_sample])); /*  0   0   0  |r0|  ==   00   |r0_64| */
 109                                         mm_sum = _mm_add_epi64(mm_sum, mm_res);
 110                                 }
 111
 112                                 for( ; residual_sample < e3; residual_sample+=2) {
 113                                         __m128i mm_res = _mm_abs_epi32(_mm_loadl_epi64((const __m128i*)(residual+residual_sample))); /*  0   0  |r1|   |r0| */
 114                                         mm_res = _mm_shuffle_epi32(mm_res, _MM_SHUFFLE(3,1,2,0)); /* 0  |r1|  0  |r0|  ==  |r1_64|  |r0_64|  */
 115                                         mm_sum = _mm_add_epi64(mm_sum, mm_res);
 116                                 }
 117
 118                                 for( ; residual_sample < end; residual_sample++) {
 119                                         __m128i mm_res = _mm_abs_epi32(_mm_cvtsi32_si128(residual[residual_sample]));
 120                                         mm_sum = _mm_add_epi64(mm_sum, mm_res);
 121                                 }
 122
 123                                 mm_sum = _mm_add_epi64(mm_sum, _mm_srli_si128(mm_sum, 8));
 124                                 _mm_storel_epi64((__m128i*)(abs_residual_partition_sums+partition), mm_sum);
 125                         }
 126                 }
 127         }
 128
 129         /* now merge partitions for lower orders */
 130         {
 131                 uint32_t from_partition = 0, to_partition = partitions;
 132                 int partition_order;
 133                 for(partition_order = (int)max_partition_order - 1; partition_order >= (int)min_partition_order; partition_order--) {
 134                         uint32_t i;
 135                         partitions >>= 1;
 136                         for(i = 0; i < partitions; i++) {
 137                                 abs_residual_partition_sums[to_partition++] =
 138                                         abs_residual_partition_sums[from_partition  ] +
 139                                         abs_residual_partition_sums[from_partition+1];
 140                                 from_partition += 2;
 141                         }
 142                 }
 143         }
 144 }
 145
 146 #endif /* FLAC__SSSE3_SUPPORTED */
 147 #endif /* (FLAC__CPU_IA32 || FLAC__CPU_X86_64) && FLAC__HAS_X86INTRIN */
 148 #endif /* FLAC__NO_ASM */