Remove direct convolution implementation as its give worse results. Simplifies code too.
Unified · Split
| @@ -25,6 +25,7 @@ · 24 unmodified lines | ||
| 25 | 25 | |
| 26 | 26 | #include <float.h> |
| 27 | 27 | |
| 28 | +#include "libavutil/cpu.h" | |
| 28 | 29 | #include "libavutil/tx.h" |
| 29 | 30 | #include "libavutil/avstring.h" |
| 30 | 31 | #include "libavutil/channel_layout.h" |
| @@ -158,6 +159,7 @@ static int init_segment(AVFilterContext *ctx, AudioFIRSegment *seg, · 127 unmodified lines | ||
| 158 | 159 | int offset, int nb_partitions, int part_size) |
| 159 | 160 | { |
| 160 | 161 | AudioFIRContext *s = ctx->priv; |
| 162 | + const size_t cpu_align = av_cpu_max_align(); | |
| 161 | 163 | int ret; |
| 162 | 164 | |
| 163 | 165 | seg->tx = av_calloc(ctx->inputs[0]->ch_layout.nb_channels, sizeof(*seg->tx)); |
| @@ -167,8 +169,8 @@ static int init_segment(AVFilterContext *ctx, AudioFIRSegment *seg, · 3 unmodified lines | ||
| 167 | 169 | |
| 168 | 170 | seg->fft_length = part_size * 2 + 1; |
| 169 | 171 | seg->part_size = part_size; |
| 170 | − seg->block_size = FFALIGN(seg->fft_length, 32); | |
| 171 | − seg->coeff_size = FFALIGN(seg->part_size + 1, 32); | |
| 172 | + seg->block_size = FFALIGN(seg->fft_length, cpu_align); | |
| 173 | + seg->coeff_size = FFALIGN(seg->part_size + 1, cpu_align); | |
| 172 | 174 | seg->nb_partitions = nb_partitions; |
| 173 | 175 | seg->input_size = offset + s->min_part_size; |
| 174 | 176 | seg->input_offset = offset; |
| @@ -178,7 +180,7 @@ static int init_segment(AVFilterContext *ctx, AudioFIRSegment *seg, · 3 unmodified lines | ||
| 178 | 180 | if (!seg->part_index || !seg->output_offset) |
| 179 | 181 | return AVERROR(ENOMEM); |
| 180 | 182 | |
| 181 | − for (int ch = 0; ch < ctx->inputs[0]->ch_layout.nb_channels && part_size >= 8; ch++) { | |
| 183 | + for (int ch = 0; ch < ctx->inputs[0]->ch_layout.nb_channels && part_size >= 1; ch++) { | |
| 182 | 184 | union { double d; float f; } scale, iscale; |
| 183 | 185 | enum AVTXType tx_type; |
| 184 | 186 | |
| @@ -156,14 +156,6 @@ static void fn(convert_channels)(AVFilterContext *ctx, AudioFIRContext *s) · 155 unmodified lines | ||
| 156 | 156 | const int remaining = s->nb_taps - toffset; |
| 157 | 157 | const int size = remaining >= seg->part_size ? seg->part_size : remaining; |
| 158 | 158 | |
| 159 | − if (size < 8) { | |
| 160 | − for (int n = 0; n < size; n++) | |
| 161 | − coeff[coffset + n].re = time[toffset + n]; | |
| 162 | − | |
| 163 | − toffset += size; | |
| 164 | − continue; | |
| 165 | − } | |
| 166 | − | |
| 167 | 159 | memset(blockin, 0, sizeof(*blockin) * seg->fft_length); |
| 168 | 160 | memcpy(blockin, time + toffset, size * sizeof(*blockin)); |
| 169 | 161 | |
| @@ -245,16 +237,9 @@ static int fn(get_power)(AVFilterContext *ctx, AudioFIRContext *s, int cur_nb_ta · 75 unmodified lines | ||
| 245 | 237 | return 0; |
| 246 | 238 | } |
| 247 | 239 | |
| 248 | −static void fn(direct)(const ftype *in, const ctype *ir, int len, ftype *out) | |
| 249 | −{ | |
| 250 | − for (int n = 0; n < len; n++) | |
| 251 | − for (int m = 0; m <= n; m++) | |
| 252 | − out[n] += ir[m].re * in[n - m]; | |
| 253 | −} | |
| 254 | − | |
| 255 | 240 | static void fn(fir_fadd)(AudioFIRContext *s, ftype *dst, const ftype *src, int nb_samples) |
| 256 | 241 | { |
| 257 | − if ((nb_samples & 15) == 0 && nb_samples >= 16) { | |
| 242 | + if ((nb_samples & 15) == 0 && nb_samples >= 8) { | |
| 258 | 243 | #if DEPTH == 32 |
| 259 | 244 | s->fdsp->vector_fmac_scalar(dst, src, 1.f, nb_samples); |
| 260 | 245 | #else |
| @@ -272,7 +257,8 @@ static int fn(fir_quantum)(AVFilterContext *ctx, AVFrame *out, int ch, int offse · 11 unmodified lines | ||
| 272 | 257 | const ftype *in = (const ftype *)s->in->extended_data[ch] + offset; |
| 273 | 258 | ftype *blockin, *blockout, *buf, *ptr = (ftype *)out->extended_data[ch] + offset; |
| 274 | 259 | const int nb_samples = FFMIN(s->min_part_size, out->nb_samples - offset); |
| 275 | − int n, i, j; | |
| 260 | + const int min_part_size = s->min_part_size; | |
| 261 | + const float dry_gain = s->dry_gain; | |
| 276 | 262 | |
| 277 | 263 | for (int segment = 0; segment < s->nb_segments; segment++) { |
| 278 | 264 | AudioFIRSegment *seg = &s->seg[segment]; |
| @@ -280,76 +266,55 @@ static int fn(fir_quantum)(AVFilterContext *ctx, AVFrame *out, int ch, int offse · 1 unmodified line | ||
| 280 | 266 | ftype *dst = (ftype *)seg->output->extended_data[ch]; |
| 281 | 267 | ftype *sumin = (ftype *)seg->sumin->extended_data[ch]; |
| 282 | 268 | ftype *sumout = (ftype *)seg->sumout->extended_data[ch]; |
| 269 | + int *output_offset = &seg->output_offset[ch]; | |
| 270 | + const int input_offset = seg->input_offset; | |
| 271 | + const int part_size = seg->part_size; | |
| 272 | + int j; | |
| 283 | 273 | |
| 284 | − if (s->min_part_size >= 8) { | |
| 274 | + if (min_part_size >= 8) { | |
| 285 | 275 | #if DEPTH == 32 |
| 286 | − s->fdsp->vector_fmul_scalar(src + seg->input_offset, in, s->dry_gain, FFALIGN(nb_samples, 4)); | |
| 276 | + s->fdsp->vector_fmul_scalar(src + input_offset, in, dry_gain, FFALIGN(nb_samples, 4)); | |
| 287 | 277 | #else |
| 288 | − s->fdsp->vector_dmul_scalar(src + seg->input_offset, in, s->dry_gain, FFALIGN(nb_samples, 8)); | |
| 278 | + s->fdsp->vector_dmul_scalar(src + input_offset, in, dry_gain, FFALIGN(nb_samples, 8)); | |
| 289 | 279 | #endif |
| 290 | 280 | emms_c(); |
| 291 | 281 | } else { |
| 292 | − for (n = 0; n < nb_samples; n++) | |
| 293 | − src[seg->input_offset + n] = in[n] * s->dry_gain; | |
| 282 | + ftype *src2 = src + input_offset; | |
| 283 | + for (int n = 0; n < nb_samples; n++) | |
| 284 | + src2[n] = in[n] * dry_gain; | |
| 294 | 285 | } |
| 295 | 286 | |
| 296 | − seg->output_offset[ch] += s->min_part_size; | |
| 297 | − if (seg->output_offset[ch] == seg->part_size) { | |
| 298 | − seg->output_offset[ch] = 0; | |
| 287 | + output_offset[0] += min_part_size; | |
| 288 | + if (output_offset[0] == part_size) { | |
| 289 | + output_offset[0] = 0; | |
| 299 | 290 | } else { |
| 300 | − memmove(src, src + s->min_part_size, (seg->input_size - s->min_part_size) * sizeof(*src)); | |
| 291 | + memmove(src, src + min_part_size, (seg->input_size - min_part_size) * sizeof(*src)); | |
| 301 | 292 | |
| 302 | − dst += seg->output_offset[ch]; | |
| 293 | + dst += output_offset[0]; | |
| 303 | 294 | fn(fir_fadd)(s, ptr, dst, nb_samples); |
| 304 | 295 | continue; |
| 305 | 296 | } |
| 306 | 297 | |
| 307 | − if (seg->part_size < 8) { | |
| 308 | − memset(dst, 0, sizeof(*dst) * seg->part_size * seg->nb_partitions); | |
| 309 | − | |
| 310 | − j = seg->part_index[ch]; | |
| 311 | − | |
| 312 | − for (i = 0; i < seg->nb_partitions; i++) { | |
| 313 | − const int coffset = j * seg->coeff_size; | |
| 314 | − const ctype *coeff = (const ctype *)seg->coeff->extended_data[ch * !s->one2many] + coffset; | |
| 315 | − | |
| 316 | − fn(direct)(src, coeff, nb_samples, dst); | |
| 317 | − | |
| 318 | − if (j == 0) | |
| 319 | − j = seg->nb_partitions; | |
| 320 | − j--; | |
| 321 | − } | |
| 322 | − | |
| 323 | − seg->part_index[ch] = (seg->part_index[ch] + 1) % seg->nb_partitions; | |
| 324 | − | |
| 325 | − memmove(src, src + s->min_part_size, (seg->input_size - s->min_part_size) * sizeof(*src)); | |
| 326 | − | |
| 327 | − for (n = 0; n < nb_samples; n++) { | |
| 328 | − ptr[n] += dst[n]; | |
| 329 | − } | |
| 330 | − continue; | |
| 331 | − } | |
| 332 | − | |
| 333 | 298 | memset(sumin, 0, sizeof(*sumin) * seg->fft_length); |
| 334 | 299 | blockin = (ftype *)seg->blockin->extended_data[ch] + seg->part_index[ch] * seg->block_size; |
| 335 | 300 | blockout = (ftype *)seg->blockout->extended_data[ch] + seg->part_index[ch] * seg->block_size; |
| 336 | − memset(blockin + seg->part_size, 0, sizeof(*blockin) * (seg->fft_length - seg->part_size)); | |
| 301 | + memset(blockin + part_size, 0, sizeof(*blockin) * (seg->fft_length - part_size)); | |
| 337 | 302 | |
| 338 | − memcpy(blockin, src, sizeof(*src) * seg->part_size); | |
| 303 | + memcpy(blockin, src, sizeof(*src) * part_size); | |
| 339 | 304 | |
| 340 | 305 | seg->tx_fn(seg->tx[ch], blockout, blockin, sizeof(ftype)); |
| 341 | 306 | |
| 342 | 307 | j = seg->part_index[ch]; |
| 343 | 308 | |
| 344 | − for (i = 0; i < seg->nb_partitions; i++) { | |
| 309 | + for (int i = 0; i < seg->nb_partitions; i++) { | |
| 345 | 310 | const int coffset = j * seg->coeff_size; |
| 346 | 311 | const ftype *blockout = (const ftype *)seg->blockout->extended_data[ch] + i * seg->block_size; |
| 347 | 312 | const ctype *coeff = (const ctype *)seg->coeff->extended_data[ch * !s->one2many] + coffset; |
| 348 | 313 | |
| 349 | 314 | #if DEPTH == 32 |
| 350 | − s->afirdsp.fcmul_add(sumin, blockout, (const ftype *)coeff, seg->part_size); | |
| 315 | + s->afirdsp.fcmul_add(sumin, blockout, (const ftype *)coeff, part_size); | |
| 351 | 316 | #else |
| 352 | − s->afirdsp.dcmul_add(sumin, blockout, (const ftype *)coeff, seg->part_size); | |
| 317 | + s->afirdsp.dcmul_add(sumin, blockout, (const ftype *)coeff, part_size); | |
| 353 | 318 | #endif |
| 354 | 319 | |
| 355 | 320 | if (j == 0) |
| @@ -360,21 +325,21 @@ static int fn(fir_quantum)(AVFilterContext *ctx, AVFrame *out, int ch, int offse · 4 unmodified lines | ||
| 360 | 325 | seg->itx_fn(seg->itx[ch], sumout, sumin, sizeof(ctype)); |
| 361 | 326 | |
| 362 | 327 | buf = (ftype *)seg->buffer->extended_data[ch]; |
| 363 | − fn(fir_fadd)(s, buf, sumout, seg->part_size); | |
| 328 | + fn(fir_fadd)(s, buf, sumout, part_size); | |
| 364 | 329 | |
| 365 | − memcpy(dst, buf, seg->part_size * sizeof(*dst)); | |
| 330 | + memcpy(dst, buf, part_size * sizeof(*dst)); | |
| 366 | 331 | |
| 367 | 332 | buf = (ftype *)seg->buffer->extended_data[ch]; |
| 368 | − memcpy(buf, sumout + seg->part_size, seg->part_size * sizeof(*buf)); | |
| 333 | + memcpy(buf, sumout + part_size, part_size * sizeof(*buf)); | |
| 369 | 334 | |
| 370 | 335 | seg->part_index[ch] = (seg->part_index[ch] + 1) % seg->nb_partitions; |
| 371 | 336 | |
| 372 | − memmove(src, src + s->min_part_size, (seg->input_size - s->min_part_size) * sizeof(*src)); | |
| 337 | + memmove(src, src + min_part_size, (seg->input_size - min_part_size) * sizeof(*src)); | |
| 373 | 338 | |
| 374 | 339 | fn(fir_fadd)(s, ptr, dst, nb_samples); |
| 375 | 340 | } |
| 376 | 341 | |
| 377 | − if (s->min_part_size >= 8) { | |
| 342 | + if (min_part_size >= 8) { | |
| 378 | 343 | #if DEPTH == 32 |
| 379 | 344 | s->fdsp->vector_fmul_scalar(ptr, ptr, s->wet_gain, FFALIGN(nb_samples, 4)); |
| 380 | 345 | #else |
| @@ -382,11 +347,9 @@ static int fn(fir_quantum)(AVFilterContext *ctx, AVFrame *out, int ch, int offse · 1 unmodified line | ||
| 382 | 347 | #endif |
| 383 | 348 | emms_c(); |
| 384 | 349 | } else { |
| 385 | − for (n = 0; n < nb_samples; n++) | |
| 350 | + for (int n = 0; n < nb_samples; n++) | |
| 386 | 351 | ptr[n] *= s->wet_gain; |
| 387 | 352 | } |
| 388 | 353 | |
| 389 | 354 | return 0; |
| 390 | 355 | } |
| 391 | − | |
| 392 | − | |