From c85e0ca68a3d6d32fd64353da9293712d008fa87 Mon Sep 17 00:00:00 2001 From: =?UTF-8?q?Cl=C3=A9ment=20B=C5=93sch?= Date: Sun, 10 Aug 2014 21:49:55 +0200 Subject: [PATCH] avfilter/dctdnoiz: add slice threading --- libavfilter/vf_dctdnoiz.c | 114 ++++++++++++++++++++++++++++---------- 1 file changed, 84 insertions(+), 30 deletions(-) diff --git a/libavfilter/vf_dctdnoiz.c b/libavfilter/vf_dctdnoiz.c index fe1c5d78bf..a9017b1f1c 100644 --- a/libavfilter/vf_dctdnoiz.c +++ b/libavfilter/vf_dctdnoiz.c @@ -36,18 +36,22 @@ static const char *const var_names[] = { "c", NULL }; enum { VAR_C, VAR_VARS_NB }; +#define MAX_THREADS 8 + typedef struct DCTdnoizContext { const AVClass *class; /* coefficient factor expression */ char *expr_str; - AVExpr *expr; - double var_values[VAR_VARS_NB]; + AVExpr *expr[MAX_THREADS]; + double var_values[MAX_THREADS][VAR_VARS_NB]; + int nb_threads; int pr_width, pr_height; // width and height to process float sigma; // used when no expression are st float th; // threshold (3*sigma) float *cbuf[2][3]; // two planar rgb color buffers + float *slices[MAX_THREADS]; // slices buffers (1 slice buffer per thread) float *weights; // dct coeff are cumulated with overlapping; these values are used for averaging int p_linesize; // line sizes for color and weights int overlap; // number of block overlapping pixels @@ -56,7 +60,8 @@ typedef struct DCTdnoizContext { int bsize; // block size, 1<th); \ } \ \ static void filter_freq_expr_##bsize(DCTdnoizContext *s, \ const float *src, int src_linesize, \ - float *dst, int dst_linesize) \ + float *dst, int dst_linesize, int thread_id) \ { \ - filter_freq_##bsize(src, src_linesize, dst, dst_linesize, s->expr, s->var_values, 0); \ + filter_freq_##bsize(src, src_linesize, dst, dst_linesize, \ + s->expr[thread_id], s->var_values[thread_id], 0); \ } DEF_FILTER_FREQ_FUNCS(8) @@ -475,7 +481,7 @@ static int config_input(AVFilterLink *inlink) { AVFilterContext *ctx = inlink->dst; DCTdnoizContext *s = ctx->priv; - int i, x, y, bx, by, linesize, *iweights; + int i, x, y, bx, by, linesize, *iweights, max_slice_h, slice_h; const int bsize = 1 << s->n; switch (inlink->format) { @@ -500,6 +506,11 @@ static int config_input(AVFilterLink *inlink) av_log(ctx, AV_LOG_WARNING, "The last %d vertical pixels won't be denoised\n", inlink->h - s->pr_height); + max_slice_h = s->pr_height / ((s->bsize - 1) * 2); + s->nb_threads = FFMIN3(MAX_THREADS, ctx->graph->nb_threads, max_slice_h); + av_log(ctx, AV_LOG_DEBUG, "threads: [max=%d hmax=%d user=%d] => %d\n", + MAX_THREADS, max_slice_h, ctx->graph->nb_threads, s->nb_threads); + s->p_linesize = linesize = FFALIGN(s->pr_width, 32); for (i = 0; i < 2; i++) { s->cbuf[i][0] = av_malloc(linesize * s->pr_height * sizeof(*s->cbuf[i][0])); @@ -509,6 +520,27 @@ static int config_input(AVFilterLink *inlink) return AVERROR(ENOMEM); } + /* eval expressions are probably not thread safe when the eval internal + * state can be changed (typically through load & store operations) */ + if (s->expr_str) { + for (i = 0; i < s->nb_threads; i++) { + int ret = av_expr_parse(&s->expr[i], s->expr_str, var_names, + NULL, NULL, NULL, NULL, 0, ctx); + if (ret < 0) + return ret; + } + } + + /* each slice will need to (pre & re)process the top and bottom block of + * the previous one in in addition to its processing area. This is because + * each pixel is averaged by all the surrounding blocks */ + slice_h = (int)ceilf(s->pr_height / s->nb_threads) + (s->bsize - 1) * 2; + for (i = 0; i < s->nb_threads; i++) { + s->slices[i] = av_malloc_array(linesize, slice_h * sizeof(*s->slices[i])); + if (!s->slices[i]) + return AVERROR(ENOMEM); + } + s->weights = av_malloc(s->pr_height * linesize * sizeof(*s->weights)); if (!s->weights) return AVERROR(ENOMEM); @@ -544,10 +576,6 @@ static av_cold int init(AVFilterContext *ctx) } if (s->expr_str) { - int ret = av_expr_parse(&s->expr, s->expr_str, var_names, - NULL, NULL, NULL, NULL, 0, ctx); - if (ret < 0) - return ret; switch (s->n) { case 3: s->filter_freq_func = filter_freq_expr_8; break; case 4: s->filter_freq_func = filter_freq_expr_16; break; @@ -576,36 +604,56 @@ static int query_formats(AVFilterContext *ctx) return 0; } -static void filter_plane(AVFilterContext *ctx, - float *dst, int dst_linesize, - const float *src, int src_linesize, - int w, int h) +typedef struct ThreadData { + float *src, *dst; +} ThreadData; + +static int filter_slice(AVFilterContext *ctx, + void *arg, int jobnr, int nb_jobs) { int x, y; DCTdnoizContext *s = ctx->priv; - float *dst0 = dst; - const float *weights = s->weights; + const ThreadData *td = arg; + const int w = s->pr_width; + const int h = s->pr_height; + const int slice_start = (h * jobnr ) / nb_jobs; + const int slice_end = (h * (jobnr+1)) / nb_jobs; + const int slice_start_ctx = FFMAX(slice_start - s->bsize + 1, 0); + const int slice_end_ctx = FFMIN(slice_end, h - s->bsize + 1); + const int slice_h = slice_end_ctx - slice_start_ctx; + const int src_linesize = s->p_linesize; + const int dst_linesize = s->p_linesize; + const int slice_linesize = s->p_linesize; + float *dst; + const float *src = td->src + slice_start_ctx * src_linesize; + const float *weights = s->weights + slice_start * dst_linesize; + float *slice = s->slices[jobnr]; // reset block sums - memset(dst, 0, h * dst_linesize * sizeof(*dst)); + memset(slice, 0, (slice_h + s->bsize - 1) * dst_linesize * sizeof(*slice)); // block dct sums - for (y = 0; y < h - s->bsize + 1; y += s->step) { + for (y = 0; y < slice_h; y += s->step) { for (x = 0; x < w - s->bsize + 1; x += s->step) s->filter_freq_func(s, src + x, src_linesize, - dst + x, dst_linesize); + slice + x, slice_linesize, + jobnr); src += s->step * src_linesize; - dst += s->step * dst_linesize; + slice += s->step * slice_linesize; } // average blocks - dst = dst0; - for (y = 0; y < h; y++) { + slice = s->slices[jobnr] + (slice_start - slice_start_ctx) * slice_linesize; + dst = td->dst + slice_start * dst_linesize; + for (y = slice_start; y < slice_end; y++) { for (x = 0; x < w; x++) - dst[x] *= weights[x]; + dst[x] = slice[x] * weights[x]; + slice += slice_linesize; dst += dst_linesize; weights += dst_linesize; } + + return 0; } static int filter_frame(AVFilterLink *inlink, AVFrame *in) @@ -632,10 +680,13 @@ static int filter_frame(AVFilterLink *inlink, AVFrame *in) s->color_decorrelation(s->cbuf[0], s->p_linesize, in->data[0], in->linesize[0], s->pr_width, s->pr_height); - for (plane = 0; plane < 3; plane++) - filter_plane(ctx, s->cbuf[1][plane], s->p_linesize, - s->cbuf[0][plane], s->p_linesize, - s->pr_width, s->pr_height); + for (plane = 0; plane < 3; plane++) { + ThreadData td = { + .src = s->cbuf[0][plane], + .dst = s->cbuf[1][plane], + }; + ctx->internal->execute(ctx, filter_slice, &td, NULL, s->nb_threads); + } s->color_correlation(out->data[0], out->linesize[0], s->cbuf[1], s->p_linesize, s->pr_width, s->pr_height); @@ -687,7 +738,10 @@ static av_cold void uninit(AVFilterContext *ctx) av_free(s->cbuf[i][1]); av_free(s->cbuf[i][2]); } - av_expr_free(s->expr); + for (i = 0; i < s->nb_threads; i++) { + av_free(s->slices[i]); + av_expr_free(s->expr[i]); + } } static const AVFilterPad dctdnoiz_inputs[] = { @@ -718,5 +772,5 @@ AVFilter ff_vf_dctdnoiz = { .inputs = dctdnoiz_inputs, .outputs = dctdnoiz_outputs, .priv_class = &dctdnoiz_class, - .flags = AVFILTER_FLAG_SUPPORT_TIMELINE_GENERIC, + .flags = AVFILTER_FLAG_SUPPORT_TIMELINE_GENERIC | AVFILTER_FLAG_SLICE_THREADS, };