libavfilter/afir_template.c

   1 /*
   2  * Copyright (c) 2017 Paul B Mahol
   3  *
   4  * This file is part of FFmpeg.
   5  *
   6  * FFmpeg is free software; you can redistribute it and/or
   7  * modify it under the terms of the GNU Lesser General Public
   8  * License as published by the Free Software Foundation; either
   9  * version 2.1 of the License, or (at your option) any later version.
  10  *
  11  * FFmpeg is distributed in the hope that it will be useful,
  12  * but WITHOUT ANY WARRANTY; without even the implied warranty of
  13  * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the GNU
  14  * Lesser General Public License for more details.
  15  *
  16  * You should have received a copy of the GNU Lesser General Public
  17  * License along with FFmpeg; if not, write to the Free Software
  18  * Foundation, Inc., 51 Franklin Street, Fifth Floor, Boston, MA 02110-1301 USA
  19  */
  20
  21 #include "libavutil/tx.h"
  22 #include "avfilter.h"
  23 #include "audio.h"
  24
  25 #undef ctype
  26 #undef ftype
  27 #undef SQRT
  28 #undef HYPOT
  29 #undef SAMPLE_FORMAT
  30 #undef TX_TYPE
  31 #undef FABS
  32 #undef POW
  33 #if DEPTH == 32
  34 #define SAMPLE_FORMAT float
  35 #define SQRT sqrtf
  36 #define HYPOT hypotf
  37 #define ctype AVComplexFloat
  38 #define ftype float
  39 #define TX_TYPE AV_TX_FLOAT_RDFT
  40 #define FABS fabsf
  41 #define POW powf
  42 #else
  43 #define SAMPLE_FORMAT double
  44 #define SQRT sqrt
  45 #define HYPOT hypot
  46 #define ctype AVComplexDouble
  47 #define ftype double
  48 #define TX_TYPE AV_TX_DOUBLE_RDFT
  49 #define FABS fabs
  50 #define POW pow
  51 #endif
  52
  53 #define fn3(a,b)   a##_##b
  54 #define fn2(a,b)   fn3(a,b)
  55 #define fn(a)      fn2(a, SAMPLE_FORMAT)
  56
  57 static ftype fn(ir_gain)(AVFilterContext *ctx, AudioFIRContext *s,
  58                          int cur_nb_taps, const ftype *time)
  59 {
  60     ftype ch_gain, sum = 0;
  61
  62     if (s->ir_norm < 0.f) {
  63         ch_gain = 1;
  64     } else if (s->ir_norm == 0.f) {
  65         for (int i = 0; i < cur_nb_taps; i++)
  66             sum += time[i];
  67         ch_gain = 1. / sum;
  68     } else {
  69         ftype ir_norm = s->ir_norm;
  70
  71         for (int i = 0; i < cur_nb_taps; i++)
  72             sum += POW(FABS(time[i]), ir_norm);
  73         ch_gain = 1. / POW(sum, 1. / ir_norm);
  74     }
  75
  76     return ch_gain;
  77 }
  78
  79 static void fn(ir_scale)(AVFilterContext *ctx, AudioFIRContext *s,
  80                          int cur_nb_taps, int ch,
  81                          ftype *time, ftype ch_gain)
  82 {
  83     if (ch_gain != 1. || s->ir_gain != 1.) {
  84         ftype gain = ch_gain * s->ir_gain;
  85
  86         av_log(ctx, AV_LOG_DEBUG, "ch%d gain %f\n", ch, gain);
  87 #if DEPTH == 32
  88         s->fdsp->vector_fmul_scalar(time, time, gain, FFALIGN(cur_nb_taps, 4));
  89 #else
  90         s->fdsp->vector_dmul_scalar(time, time, gain, FFALIGN(cur_nb_taps, 8));
  91 #endif
  92     }
  93 }
  94
  95 static void fn(convert_channel)(AVFilterContext *ctx, AudioFIRContext *s, int ch,
  96                                 AudioFIRSegment *seg, int coeff_partition, int selir)
  97 {
  98     const int coffset = coeff_partition * seg->coeff_size;
  99     const int nb_taps = s->nb_taps[selir];
 100     ftype *time = (ftype *)s->norm_ir[selir]->extended_data[ch];
 101     ftype *tempin = (ftype *)seg->tempin->extended_data[ch];
 102     ftype *tempout = (ftype *)seg->tempout->extended_data[ch];
 103     ctype *coeff = (ctype *)seg->coeff->extended_data[ch];
 104     const int remaining = nb_taps - (seg->input_offset + coeff_partition * seg->part_size);
 105     const int size = remaining >= seg->part_size ? seg->part_size : remaining;
 106
 107     memset(tempin + size, 0, sizeof(*tempin) * (seg->block_size - size));
 108     memcpy(tempin, time + seg->input_offset + coeff_partition * seg->part_size,
 109            size * sizeof(*tempin));
 110     seg->ctx_fn(seg->ctx[ch], tempout, tempin, sizeof(*tempin));
 111     memcpy(coeff + coffset, tempout, seg->coeff_size * sizeof(*coeff));
 112
 113     av_log(ctx, AV_LOG_DEBUG, "channel: %d\n", ch);
 114     av_log(ctx, AV_LOG_DEBUG, "nb_partitions: %d\n", seg->nb_partitions);
 115     av_log(ctx, AV_LOG_DEBUG, "partition size: %d\n", seg->part_size);
 116     av_log(ctx, AV_LOG_DEBUG, "block size: %d\n", seg->block_size);
 117     av_log(ctx, AV_LOG_DEBUG, "fft_length: %d\n", seg->fft_length);
 118     av_log(ctx, AV_LOG_DEBUG, "coeff_size: %d\n", seg->coeff_size);
 119     av_log(ctx, AV_LOG_DEBUG, "input_size: %d\n", seg->input_size);
 120     av_log(ctx, AV_LOG_DEBUG, "input_offset: %d\n", seg->input_offset);
 121 }
 122
 123 static void fn(fir_fadd)(AudioFIRContext *s, ftype *dst, const ftype *src, int nb_samples)
 124 {
 125     if ((nb_samples & 15) == 0 && nb_samples >= 8) {
 126 #if DEPTH == 32
 127         s->fdsp->vector_fmac_scalar(dst, src, 1.f, nb_samples);
 128 #else
 129         s->fdsp->vector_dmac_scalar(dst, src, 1.0, nb_samples);
 130 #endif
 131     } else {
 132         for (int n = 0; n < nb_samples; n++)
 133             dst[n] += src[n];
 134     }
 135 }
 136
 137 static int fn(fir_quantum)(AVFilterContext *ctx, AVFrame *out, int ch, int ioffset, int offset, int selir)
 138 {
 139     AudioFIRContext *s = ctx->priv;
 140     const ftype *in = (const ftype *)s->in->extended_data[ch] + ioffset;
 141     ftype *blockout, *ptr = (ftype *)out->extended_data[ch] + offset;
 142     const int min_part_size = s->min_part_size;
 143     const int nb_samples = FFMIN(min_part_size, out->nb_samples - offset);
 144     const int nb_segments = s->nb_segments[selir];
 145     const float dry_gain = s->dry_gain;
 146     const float wet_gain = s->wet_gain;
 147
 148     for (int segment = 0; segment < nb_segments; segment++) {
 149         AudioFIRSegment *seg = &s->seg[selir][segment];
 150         ftype *src = (ftype *)seg->input->extended_data[ch];
 151         ftype *dst = (ftype *)seg->output->extended_data[ch];
 152         ftype *sumin = (ftype *)seg->sumin->extended_data[ch];
 153         ftype *sumout = (ftype *)seg->sumout->extended_data[ch];
 154         ftype *tempin = (ftype *)seg->tempin->extended_data[ch];
 155         ftype *buf = (ftype *)seg->buffer->extended_data[ch];
 156         int *output_offset = &seg->output_offset[ch];
 157         const int nb_partitions = seg->nb_partitions;
 158         const int input_offset = seg->input_offset;
 159         const int part_size = seg->part_size;
 160         int j;
 161
 162         seg->part_index[ch] = seg->part_index[ch] % nb_partitions;
 163         if (dry_gain == 1.f) {
 164             memcpy(src + input_offset, in, nb_samples * sizeof(*src));
 165         } else if (min_part_size >= 8) {
 166 #if DEPTH == 32
 167             s->fdsp->vector_fmul_scalar(src + input_offset, in, dry_gain, FFALIGN(nb_samples, 4));
 168 #else
 169             s->fdsp->vector_dmul_scalar(src + input_offset, in, dry_gain, FFALIGN(nb_samples, 8));
 170 #endif
 171         } else {
 172             ftype *src2 = src + input_offset;
 173             for (int n = 0; n < nb_samples; n++)
 174                 src2[n] = in[n] * dry_gain;
 175         }
 176
 177         output_offset[0] += min_part_size;
 178         if (output_offset[0] >= part_size) {
 179             output_offset[0] = 0;
 180         } else {
 181             memmove(src, src + min_part_size, (seg->input_size - min_part_size) * sizeof(*src));
 182
 183             dst += output_offset[0];
 184             fn(fir_fadd)(s, ptr, dst, nb_samples);
 185             continue;
 186         }
 187
 188         memset(sumin, 0, sizeof(*sumin) * seg->fft_length);
 189
 190         blockout = (ftype *)seg->blockout->extended_data[ch] + seg->part_index[ch] * seg->block_size;
 191         memset(tempin + part_size, 0, sizeof(*tempin) * (seg->block_size - part_size));
 192         memcpy(tempin, src, sizeof(*src) * part_size);
 193         seg->tx_fn(seg->tx[ch], blockout, tempin, sizeof(ftype));
 194
 195         j = seg->part_index[ch];
 196         for (int i = 0; i < nb_partitions; i++) {
 197             const int input_partition = j;
 198             const int coeff_partition = i;
 199             const int coffset = coeff_partition * seg->coeff_size;
 200             const ftype *blockout = (const ftype *)seg->blockout->extended_data[ch] + input_partition * seg->block_size;
 201             const ctype *coeff = ((const ctype *)seg->coeff->extended_data[ch]) + coffset;
 202
 203             if (j == 0)
 204                 j = nb_partitions;
 205             j--;
 206
 207 #if DEPTH == 32
 208             s->afirdsp.fcmul_add(sumin, blockout, (const ftype *)coeff, part_size);
 209 #else
 210             s->afirdsp.dcmul_add(sumin, blockout, (const ftype *)coeff, part_size);
 211 #endif
 212         }
 213
 214         seg->itx_fn(seg->itx[ch], sumout, sumin, sizeof(ctype));
 215
 216         fn(fir_fadd)(s, buf, sumout, part_size);
 217         memcpy(dst, buf, part_size * sizeof(*dst));
 218         memcpy(buf, sumout + part_size, part_size * sizeof(*buf));
 219
 220         fn(fir_fadd)(s, ptr, dst, nb_samples);
 221
 222         if (part_size != min_part_size)
 223             memmove(src, src + min_part_size, (seg->input_size - min_part_size) * sizeof(*src));
 224
 225         seg->part_index[ch] = (seg->part_index[ch] + 1) % nb_partitions;
 226     }
 227
 228     if (wet_gain == 1.f)
 229         return 0;
 230
 231     if (min_part_size >= 8) {
 232 #if DEPTH == 32
 233         s->fdsp->vector_fmul_scalar(ptr, ptr, wet_gain, FFALIGN(nb_samples, 4));
 234 #else
 235         s->fdsp->vector_dmul_scalar(ptr, ptr, wet_gain, FFALIGN(nb_samples, 8));
 236 #endif
 237     } else {
 238         for (int n = 0; n < nb_samples; n++)
 239             ptr[n] *= wet_gain;
 240     }
 241
 242     return 0;
 243 }
 244
 245 static void fn(fir_quantums)(AVFilterContext *ctx, AudioFIRContext *s, AVFrame *out,
 246                              int min_part_size, int ch, int offset,
 247                              int prev_selir, int selir)
 248 {
 249     if (ctx->is_disabled || s->prev_is_disabled) {
 250         const ftype *in = (const ftype *)s->in->extended_data[ch] + offset;
 251         const ftype *xfade0 = (const ftype *)s->xfade[0]->extended_data[ch];
 252         const ftype *xfade1 = (const ftype *)s->xfade[1]->extended_data[ch];
 253         ftype *src0 = (ftype *)s->fadein[0]->extended_data[ch];
 254         ftype *src1 = (ftype *)s->fadein[1]->extended_data[ch];
 255         ftype *dst = ((ftype *)out->extended_data[ch]) + offset;
 256
 257         if (ctx->is_disabled && !s->prev_is_disabled) {
 258             memset(src0, 0, min_part_size * sizeof(ftype));
 259             fn(fir_quantum)(ctx, s->fadein[0], ch, offset, 0, selir);
 260             for (int n = 0; n < min_part_size; n++)
 261                 dst[n] = xfade1[n] * src0[n] + xfade0[n] * in[n];
 262         } else if (!ctx->is_disabled && s->prev_is_disabled) {
 263             memset(src1, 0, min_part_size * sizeof(ftype));
 264             fn(fir_quantum)(ctx, s->fadein[1], ch, offset, 0, selir);
 265             for (int n = 0; n < min_part_size; n++)
 266                 dst[n] = xfade1[n] * in[n] + xfade0[n] * src1[n];
 267         } else {
 268             memcpy(dst, in, sizeof(ftype) * min_part_size);
 269         }
 270     } else if (prev_selir != selir && s->loading[ch] != 0) {
 271         const ftype *xfade0 = (const ftype *)s->xfade[0]->extended_data[ch];
 272         const ftype *xfade1 = (const ftype *)s->xfade[1]->extended_data[ch];
 273         ftype *src0 = (ftype *)s->fadein[0]->extended_data[ch];
 274         ftype *src1 = (ftype *)s->fadein[1]->extended_data[ch];
 275         ftype *dst = ((ftype *)out->extended_data[ch]) + offset;
 276
 277         memset(src0, 0, min_part_size * sizeof(ftype));
 278         memset(src1, 0, min_part_size * sizeof(ftype));
 279
 280         fn(fir_quantum)(ctx, s->fadein[0], ch, offset, 0, prev_selir);
 281         fn(fir_quantum)(ctx, s->fadein[1], ch, offset, 0, selir);
 282
 283         if (s->loading[ch] > s->max_offset[selir]) {
 284             for (int n = 0; n < min_part_size; n++)
 285                 dst[n] = xfade1[n] * src0[n] + xfade0[n] * src1[n];
 286             s->loading[ch] = 0;
 287         } else {
 288             memcpy(dst, src0, min_part_size * sizeof(ftype));
 289         }
 290     } else {
 291         fn(fir_quantum)(ctx, out, ch, offset, offset, selir);
 292     }
 293 }