| @@ -38,11 +38,6 @@ · 37 unmodified lines |
| 38 | 38 | #include "vf_nlmeans.h" |
| 39 | 39 | #include "video.h" |
| 40 | 40 | |
| 41 | | −struct weighted_avg { |
| 42 | | − float total_weight; |
| 43 | | − float sum; |
| 44 | | −}; |
| 45 | | − |
| 46 | 41 | typedef struct NLMeansContext { |
| 47 | 42 | const AVClass *class; |
| 48 | 43 | int nb_planes; |
| @@ -329,6 +324,58 @@ struct thread_data { · 280 unmodified lines |
| 329 | 324 | int p; |
| 330 | 325 | }; |
| 331 | 326 | |
| 327 | +static void compute_weights_line_c(const uint32_t *const iia, |
| 328 | + const uint32_t *const iib, |
| 329 | + const uint32_t *const iid, |
| 330 | + const uint32_t *const iie, |
| 331 | + const uint8_t *const src, |
| 332 | + struct weighted_avg *wa, |
| 333 | + const float *const weight_lut, |
| 334 | + int max_meaningful_diff, |
| 335 | + int startx, int endx) |
| 336 | +{ |
| 337 | + for (int x = startx; x < endx; x++) { |
| 338 | + /* |
| 339 | + * M is a discrete map where every entry contains the sum of all the entries |
| 340 | + * in the rectangle from the top-left origin of M to its coordinate. In the |
| 341 | + * following schema, "i" contains the sum of the whole map: |
| 342 | + * |
| 343 | + * M = +----------+-----------------+----+ |
| 344 | + * | | | | |
| 345 | + * | | | | |
| 346 | + * | a| b| c| |
| 347 | + * +----------+-----------------+----+ |
| 348 | + * | | | | |
| 349 | + * | | | | |
| 350 | + * | | X | | |
| 351 | + * | | | | |
| 352 | + * | d| e| f| |
| 353 | + * +----------+-----------------+----+ |
| 354 | + * | | | | |
| 355 | + * | g| h| i| |
| 356 | + * +----------+-----------------+----+ |
| 357 | + * |
| 358 | + * The sum of the X box can be calculated with: |
| 359 | + * X = e-d-b+a |
| 360 | + * |
| 361 | + * See https://en.wikipedia.org/wiki/Summed_area_table |
| 362 | + * |
| 363 | + * The compute*_ssd functions compute the integral image M where every entry |
| 364 | + * contains the sum of the squared difference of every corresponding pixels of |
| 365 | + * two input planes of the same size as M. |
| 366 | + */ |
| 367 | + const uint32_t a = iia[x]; |
| 368 | + const uint32_t b = iib[x]; |
| 369 | + const uint32_t d = iid[x]; |
| 370 | + const uint32_t e = iie[x]; |
| 371 | + const uint32_t patch_diff_sq = FFMIN(e - d - b + a, max_meaningful_diff); |
| 372 | + const float weight = weight_lut[patch_diff_sq]; // exp(-patch_diff_sq * s->pdiff_scale) |
| 373 | + |
| 374 | + wa[x].total_weight += weight; |
| 375 | + wa[x].sum += weight * src[x]; |
| 376 | + } |
| 377 | +} |
| 378 | + |
| 332 | 379 | static int nlmeans_slice(AVFilterContext *ctx, void *arg, int jobnr, int nb_jobs) |
| 333 | 380 | { |
| 334 | 381 | NLMeansContext *s = ctx->priv; |
| @@ -346,50 +393,19 @@ static int nlmeans_slice(AVFilterContext *ctx, void *arg, int jobnr, int nb_jobs · 11 unmodified lines |
| 346 | 393 | const int dist_d = dist_b * s->ii_lz_32; |
| 347 | 394 | const int dist_e = dist_d + dist_b; |
| 348 | 395 | const float *const weight_lut = s->weight_lut; |
| 396 | + NLMeansDSPContext *dsp = &s->dsp; |
| 349 | 397 | |
| 350 | 398 | for (int y = starty; y < endy; y++) { |
| 351 | | − const uint8_t *src = td->src + y*src_linesize; |
| 399 | + const uint8_t *const src = td->src + y*src_linesize; |
| 352 | 400 | struct weighted_avg *wa = s->wa + y*s->wa_linesize; |
| 353 | | − for (int x = td->startx; x < td->endx; x++) { |
| 354 | | − /* |
| 355 | | − * M is a discrete map where every entry contains the sum of all the entries |
| 356 | | − * in the rectangle from the top-left origin of M to its coordinate. In the |
| 357 | | − * following schema, "i" contains the sum of the whole map: |
| 358 | | − * |
| 359 | | − * M = +----------+-----------------+----+ |
| 360 | | − * | | | | |
| 361 | | − * | | | | |
| 362 | | − * | a| b| c| |
| 363 | | − * +----------+-----------------+----+ |
| 364 | | − * | | | | |
| 365 | | − * | | | | |
| 366 | | − * | | X | | |
| 367 | | − * | | | | |
| 368 | | − * | d| e| f| |
| 369 | | − * +----------+-----------------+----+ |
| 370 | | − * | | | | |
| 371 | | − * | g| h| i| |
| 372 | | − * +----------+-----------------+----+ |
| 373 | | − * |
| 374 | | − * The sum of the X box can be calculated with: |
| 375 | | − * X = e-d-b+a |
| 376 | | − * |
| 377 | | − * See https://en.wikipedia.org/wiki/Summed_area_table |
| 378 | | − * |
| 379 | | − * The compute*_ssd functions compute the integral image M where every entry |
| 380 | | − * contains the sum of the squared difference of every corresponding pixels of |
| 381 | | − * two input planes of the same size as M. |
| 382 | | − */ |
| 383 | | − const uint32_t a = ii[x]; |
| 384 | | − const uint32_t b = ii[x + dist_b]; |
| 385 | | − const uint32_t d = ii[x + dist_d]; |
| 386 | | − const uint32_t e = ii[x + dist_e]; |
| 387 | | − const uint32_t patch_diff_sq = FFMIN(e - d - b + a, max_meaningful_diff); |
| 388 | | − const float weight = weight_lut[patch_diff_sq]; // exp(-patch_diff_sq * s->pdiff_scale) |
| 389 | | − |
| 390 | | − wa[x].total_weight += weight; |
| 391 | | − wa[x].sum += weight * src[x]; |
| 392 | | − } |
| 401 | + const uint32_t *const iia = ii; |
| 402 | + const uint32_t *const iib = ii + dist_b; |
| 403 | + const uint32_t *const iid = ii + dist_d; |
| 404 | + const uint32_t *const iie = ii + dist_e; |
| 405 | + |
| 406 | + dsp->compute_weights_line(iia, iib, iid, iie, src, wa, |
| 407 | + weight_lut, max_meaningful_diff, |
| 408 | + td->startx, td->endx); |
| 393 | 409 | ii += s->ii_lz_32; |
| 394 | 410 | } |
| 395 | 411 | return 0; |
| @@ -493,6 +509,7 @@ static int filter_frame(AVFilterLink *inlink, AVFrame *in) · 97 unmodified lines |
| 493 | 509 | void ff_nlmeans_init(NLMeansDSPContext *dsp) |
| 494 | 510 | { |
| 495 | 511 | dsp->compute_safe_ssd_integral_image = compute_safe_ssd_integral_image_c; |
| 512 | + dsp->compute_weights_line = compute_weights_line_c; |
| 496 | 513 | |
| 497 | 514 | if (ARCH_AARCH64) |
| 498 | 515 | ff_nlmeans_init_aarch64(dsp); |