avfilter/vf_bilateral: add slice threading support

1 files changed, 295 insertions, 166 deletions

diff --git a/libavfilter/vf_bilateral.c b/libavfilter/vf_bilateral.c
index b236efe6c3..41dc38d6f2 100644
--- a/libavfilter/vf_bilateral.c
+++ b/libavfilter/vf_bilateral.c
@@ -36,6 +36,7 @@ typedef struct BilateralContext {
     float sigmaR;
     int planes;
 
+    int nb_threads;
     int nb_planes;
     int depth;
     int planewidth[4];
@@ -44,14 +45,14 @@ typedef struct BilateralContext {
     float alpha;
     float range_table[65536];
 
-    float *img_out_f;
-    float *img_temp;
-    float *map_factor_a;
-    float *map_factor_b;
-    float *slice_factor_a;
-    float *slice_factor_b;
-    float *line_factor_a;
-    float *line_factor_b;
+    float *img_out_f[4];
+    float *img_temp[4];
+    float *map_factor_a[4];
+    float *map_factor_b[4];
+    float *slice_factor_a[4];
+    float *slice_factor_b[4];
+    float *line_factor_a[4];
+    float *line_factor_b[4];
 } BilateralContext;
 
 #define OFFSET(x) offsetof(BilateralContext, x)
@@ -102,6 +103,10 @@ static int config_params(AVFilterContext *ctx)
     return 0;
 }
 
+typedef struct ThreadData {
+    AVFrame *in, *out;
+} ThreadData;
+
 static int config_input(AVFilterLink *inlink)
 {
     AVFilterContext *ctx = inlink->dst;
@@ -109,7 +114,6 @@ static int config_input(AVFilterLink *inlink)
     const AVPixFmtDescriptor *desc = av_pix_fmt_desc_get(inlink->format);
 
     s->depth = desc->comp[0].depth;
-
     config_params(ctx);
 
     s->planewidth[1] = s->planewidth[2] = AV_CEIL_RSHIFT(inlink->w, desc->log2_chroma_w);
@@ -118,133 +122,169 @@ static int config_input(AVFilterLink *inlink)
     s->planeheight[0] = s->planeheight[3] = inlink->h;
 
     s->nb_planes = av_pix_fmt_count_planes(inlink->format);
-
-    s->img_out_f = av_calloc(inlink->w * inlink->h, sizeof(float));
-    s->img_temp = av_calloc(inlink->w * inlink->h, sizeof(float));
-    s->map_factor_a = av_calloc(inlink->w * inlink->h, sizeof(float));
-    s->map_factor_b = av_calloc(inlink->w * inlink->h, sizeof(float));
-    s->slice_factor_a = av_calloc(inlink->w, sizeof(float));
-    s->slice_factor_b = av_calloc(inlink->w, sizeof(float));
-    s->line_factor_a = av_calloc(inlink->w, sizeof(float));
-    s->line_factor_b = av_calloc(inlink->w, sizeof(float));
-
-    if (!s->img_out_f ||
-        !s->img_temp ||
-        !s->map_factor_a ||
-        !s->map_factor_b ||
-        !s->slice_factor_a ||
-        !s->slice_factor_a ||
-        !s->line_factor_a ||
-        !s->line_factor_a)
-        return AVERROR(ENOMEM);
+    s->nb_threads = ff_filter_get_nb_threads(ctx);
+
+    for (int p = 0; p < s->nb_planes; p++) {
+        const int w = s->planewidth[p];
+        const int h = s->planeheight[p];
+
+        s->img_out_f[p] = av_calloc(w * h, sizeof(float));
+        s->img_temp[p] = av_calloc(w * h, sizeof(float));
+        s->map_factor_a[p] = av_calloc(w * h, sizeof(float));
+        s->map_factor_b[p] = av_calloc(w * h, sizeof(float));
+        s->slice_factor_a[p] = av_calloc(w, sizeof(float));
+        s->slice_factor_b[p] = av_calloc(w, sizeof(float));
+        s->line_factor_a[p] = av_calloc(w, sizeof(float));
+        s->line_factor_b[p] = av_calloc(w, sizeof(float));
+
+        if (!s->img_out_f[p] ||
+            !s->img_temp[p] ||
+            !s->map_factor_a[p] ||
+            !s->map_factor_b[p] ||
+            !s->slice_factor_a[p] ||
+            !s->slice_factor_a[p] ||
+            !s->line_factor_a[p] ||
+            !s->line_factor_a[p])
+            return AVERROR(ENOMEM);
+    }
 
     return 0;
 }
 
-#define BILATERAL(type, name)                                                           \
-static void bilateral_##name(BilateralContext *s, const uint8_t *ssrc, uint8_t *ddst,   \
-                             int width, int height, int src_linesize, int dst_linesize) \
-{                                                                                       \
-    type *dst = (type *)ddst;                                                           \
-    const type *src = (const type *)ssrc;                                               \
-    float *img_out_f = s->img_out_f, *img_temp = s->img_temp;                           \
-    float *map_factor_a = s->map_factor_a, *map_factor_b = s->map_factor_b;             \
-    float *slice_factor_a = s->slice_factor_a, *slice_factor_b = s->slice_factor_b;     \
-    float *line_factor_a = s->line_factor_a, *line_factor_b = s->line_factor_b;         \
-    const float *range_table = s->range_table;                                          \
-    const float alpha = s->alpha;                                                       \
-    float ypr, ycr, *ycy, *ypy, *xcy, fp, fc;                                           \
-    const float inv_alpha_ = 1.f - alpha;                                               \
-    float *ycf, *ypf, *xcf, *in_factor;                                                 \
-    const type *tcy, *tpy;                                                                \
-    int h1;                                                                               \
-                                                                                          \
-    for (int y = 0; y < height; y++) {                                                    \
-        float *temp_factor_x, *temp_x = &img_temp[y * width];                             \
-        const type *in_x = &src[y * src_linesize];                                        \
-        const type *texture_x = &src[y * src_linesize];                                   \
-        type tpr;                                                                         \
-                                                                                          \
-        *temp_x++ = ypr = *in_x++;                                                        \
-        tpr = *texture_x++;                                                               \
-                                                                                          \
-        temp_factor_x = &map_factor_a[y * width];                                         \
-        *temp_factor_x++ = fp = 1;                                                        \
-                                                                                          \
-        for (int x = 1; x < width; x++) {                                                 \
-            float alpha_;                                                                 \
-            int range_dist;                                                               \
-            type tcr = *texture_x++;                                                      \
-            type dr = abs(tcr - tpr);                                                     \
-                                                                                          \
-            range_dist = dr;                                                              \
-            alpha_ = range_table[range_dist];                                             \
-            *temp_x++ = ycr = inv_alpha_*(*in_x++) + alpha_*ypr;                          \
-            tpr = tcr;                                                                    \
-            ypr = ycr;                                                                    \
-            *temp_factor_x++ = fc = inv_alpha_ + alpha_ * fp;                             \
-            fp = fc;                                                                      \
-        }                                                                                 \
-        --temp_x; *temp_x = ((*temp_x) + (*--in_x));                                      \
-        tpr = *--texture_x;                                                               \
-        ypr = *in_x;                                                                      \
-                                                                                          \
-        --temp_factor_x; *temp_factor_x = ((*temp_factor_x) + 1);                         \
-        fp = 1;                                                                           \
-                                                                                          \
-        for (int x = width - 2; x >= 0; x--) {                                            \
-            type tcr = *--texture_x;                                                      \
-            type dr = abs(tcr - tpr);                                                     \
-            int range_dist = dr;                                                          \
-            float alpha_ = range_table[range_dist];                                       \
-                                                                                          \
-            ycr = inv_alpha_ * (*--in_x) + alpha_ * ypr;                                  \
-            --temp_x; *temp_x = ((*temp_x) + ycr);                                        \
-            tpr = tcr;                                                                    \
-            ypr = ycr;                                                                    \
-                                                                                          \
-            fc = inv_alpha_ + alpha_*fp;                                                  \
-            --temp_factor_x;                                                              \
-            *temp_factor_x = ((*temp_factor_x) + fc);                                     \
-            fp = fc;                                                                      \
-        }                                                                                 \
-    }                                                                                     \
-    memcpy(img_out_f, img_temp, sizeof(float) * width);                                   \
-                                                                                          \
-    in_factor = map_factor_a;                                                             \
-    memcpy(map_factor_b, in_factor, sizeof(float) * width);                               \
-    for (int y = 1; y < height; y++) {                                                    \
-        tpy = &src[(y - 1) * src_linesize];                                               \
-        tcy = &src[y * src_linesize];                                                     \
-        xcy = &img_temp[y * width];                                                       \
-        ypy = &img_out_f[(y - 1) * width];                                                \
-        ycy = &img_out_f[y * width];                                                      \
-                                                                                          \
-        xcf = &in_factor[y * width];                                                      \
-        ypf = &map_factor_b[(y - 1) * width];                                             \
-        ycf = &map_factor_b[y * width];                                                   \
-        for (int x = 0; x < width; x++) {                                                 \
-            type dr = abs((*tcy++) - (*tpy++));                                           \
-            int range_dist = dr;                                                          \
-            float alpha_ = range_table[range_dist];                                       \
-                                                                                          \
-            *ycy++ = inv_alpha_*(*xcy++) + alpha_*(*ypy++);                               \
-            *ycf++ = inv_alpha_*(*xcf++) + alpha_*(*ypf++);                               \
-        }                                                                                 \
-    }                                                                                     \
-    h1 = height - 1;                                                                      \
-    ycf = line_factor_a;                                                                  \
-    ypf = line_factor_b;                                                                  \
-    memcpy(ypf, &in_factor[h1 * width], sizeof(float) * width);                           \
-    for (int x = 0; x < width; x++)                                                       \
-        map_factor_b[h1 * width + x] = (map_factor_b[h1 * width + x] + ypf[x]);           \
-                                                                                          \
-    ycy = slice_factor_a;                                                                 \
-    ypy = slice_factor_b;                                                                 \
-    memcpy(ypy, &img_temp[h1 * width], sizeof(float) * width);                            \
-    for (int x = 0, k = 0; x < width; x++) {                                              \
-        int idx = h1 * width + x;                                                         \
-        img_out_f[idx] = (img_out_f[idx] + ypy[k++]) / map_factor_b[h1 * width + x];      \
+#define BILATERAL_H(type, name)                                               \
+static void bilateralh_##name(BilateralContext *s, AVFrame *out, AVFrame *in, \
+                              int jobnr, int nb_jobs, int plane)              \
+{                                                                             \
+    const int width = s->planewidth[plane];                                   \
+    const int height = s->planeheight[plane];                                 \
+    const int slice_start = (height * jobnr) / nb_jobs;                       \
+    const int slice_end = (height * (jobnr+1)) / nb_jobs;                     \
+    const int src_linesize = in->linesize[plane] / sizeof(type);              \
+    const type *src = (const type *)in->data[plane];                          \
+    float *img_temp = s->img_temp[plane];                                     \
+    float *map_factor_a = s->map_factor_a[plane];                             \
+    const float *const range_table = s->range_table;                          \
+    const float alpha = s->alpha;                                             \
+    float ypr, ycr, fp, fc;                                                   \
+    const float inv_alpha_ = 1.f - alpha;                                     \
+                                                                              \
+    for (int y = slice_start; y < slice_end; y++) {                           \
+        float *temp_factor_x, *temp_x = &img_temp[y * width];                 \
+        const type *in_x = &src[y * src_linesize];                            \
+        const type *texture_x = &src[y * src_linesize];                       \
+        type tpr;                                                             \
+                                                                              \
+        *temp_x++ = ypr = *in_x++;                                            \
+        tpr = *texture_x++;                                                   \
+                                                                              \
+        temp_factor_x = &map_factor_a[y * width];                             \
+        *temp_factor_x++ = fp = 1;                                            \
+                                                                              \
+        for (int x = 1; x < width; x++) {                                     \
+            float alpha_;                                                     \
+            int range_dist;                                                   \
+            type tcr = *texture_x++;                                          \
+            type dr = abs(tcr - tpr);                                         \
+                                                                              \
+            range_dist = dr;                                                  \
+            alpha_ = range_table[range_dist];                                 \
+            *temp_x++ = ycr = inv_alpha_*(*in_x++) + alpha_*ypr;              \
+            tpr = tcr;                                                        \
+            ypr = ycr;                                                        \
+            *temp_factor_x++ = fc = inv_alpha_ + alpha_ * fp;                 \
+            fp = fc;                                                          \
+        }                                                                     \
+        --temp_x; *temp_x = ((*temp_x) + (*--in_x));                          \
+        tpr = *--texture_x;                                                   \
+        ypr = *in_x;                                                          \
+                                                                              \
+        --temp_factor_x; *temp_factor_x = ((*temp_factor_x) + 1);             \
+        fp = 1;                                                               \
+                                                                              \
+        for (int x = width - 2; x >= 0; x--) {                                \
+            type tcr = *--texture_x;                                          \
+            type dr = abs(tcr - tpr);                                         \
+            int range_dist = dr;                                              \
+            float alpha_ = range_table[range_dist];                           \
+                                                                              \
+            ycr = inv_alpha_ * (*--in_x) + alpha_ * ypr;                      \
+            --temp_x; *temp_x = ((*temp_x) + ycr);                            \
+            tpr = tcr;                                                        \
+            ypr = ycr;                                                        \
+                                                                              \
+            fc = inv_alpha_ + alpha_*fp;                                      \
+            --temp_factor_x;                                                  \
+            *temp_factor_x = ((*temp_factor_x) + fc);                         \
+            fp = fc;                                                          \
+        }                                                                     \
+    }                                                                         \
+}
+
+BILATERAL_H(uint8_t, byte)
+BILATERAL_H(uint16_t, word)
+
+#define BILATERAL_V(type, name)                                               \
+static void bilateralv_##name(BilateralContext *s, AVFrame *out, AVFrame *in, \
+                              int jobnr, int nb_jobs, int plane)              \
+{                                                                             \
+    const int width = s->planewidth[plane];                                   \
+    const int height = s->planeheight[plane];                                 \
+    const int slice_start = (width * jobnr) / nb_jobs;                        \
+    const int slice_end = (width * (jobnr+1)) / nb_jobs;                      \
+    const int src_linesize = in->linesize[plane] / sizeof(type);              \
+    const type *src = (const type *)in->data[plane] + slice_start;            \
+    float *img_out_f = s->img_out_f[plane] + slice_start;                     \
+    float *img_temp = s->img_temp[plane] + slice_start;                       \
+    float *map_factor_a = s->map_factor_a[plane] + slice_start;               \
+    float *map_factor_b = s->map_factor_b[plane] + slice_start;               \
+    float *slice_factor_a = s->slice_factor_a[plane] + slice_start;           \
+    float *slice_factor_b = s->slice_factor_b[plane] + slice_start;           \
+    float *line_factor_a = s->line_factor_a[plane] + slice_start;             \
+    float *line_factor_b = s->line_factor_b[plane] + slice_start;             \
+    const float *const range_table = s->range_table;                          \
+    const float alpha = s->alpha;                                             \
+    float *ycy, *ypy, *xcy;                                                   \
+    const float inv_alpha_ = 1.f - alpha;                                     \
+    float *ycf, *ypf, *xcf, *in_factor;                                       \
+    const type *tcy, *tpy;                                                    \
+    int h1;                                                                   \
+                                                                              \
+    memcpy(img_out_f, img_temp, sizeof(float) * (slice_end - slice_start));   \
+                                                                              \
+    in_factor = map_factor_a;                                                   \
+    memcpy(map_factor_b, in_factor, sizeof(float) * (slice_end - slice_start)); \
+    for (int y = 1; y < height; y++) {                                          \
+        tpy = &src[(y - 1) * src_linesize];                                   \
+        tcy = &src[y * src_linesize];                                         \
+        xcy = &img_temp[y * width];                                           \
+        ypy = &img_out_f[(y - 1) * width];                                    \
+        ycy = &img_out_f[y * width];                                          \
+                                                                              \
+        xcf = &in_factor[y * width];                                          \
+        ypf = &map_factor_b[(y - 1) * width];                                 \
+        ycf = &map_factor_b[y * width];                                       \
+        for (int x = 0; x < slice_end - slice_start; x++) {                   \
+            type dr = abs((*tcy++) - (*tpy++));                               \
+            int range_dist = dr;                                              \
+            float alpha_ = range_table[range_dist];                           \
+                                                                              \
+            *ycy++ = inv_alpha_*(*xcy++) + alpha_*(*ypy++);                   \
+            *ycf++ = inv_alpha_*(*xcf++) + alpha_*(*ypf++);                   \
+        }                                                                     \
+    }                                                                         \
+    h1 = height - 1;                                                          \
+    ycf = line_factor_a;                                                      \
+    ypf = line_factor_b;                                                            \
+    memcpy(ypf, &in_factor[h1 * width], sizeof(float) * (slice_end - slice_start)); \
+    for (int x = 0, k = 0; x < slice_end - slice_start; x++)                        \
+        map_factor_b[h1 * width + x] = (map_factor_b[h1 * width + x] + ypf[k++]); \
+                                                                                  \
+    ycy = slice_factor_a;                                                         \
+    ypy = slice_factor_b;                                                         \
+    memcpy(ypy, &img_temp[h1 * width], sizeof(float) * (slice_end - slice_start)); \
+    for (int x = 0, k = 0; x < slice_end - slice_start; x++) {                \
+        int idx = h1 * width + x;                                             \
+        img_out_f[idx] = (img_out_f[idx] + ypy[k++]) / map_factor_b[h1 * width + x]; \
     }                                                                                     \
                                                                                           \
     for (int y = h1 - 1; y >= 0; y--) {                                                   \
@@ -262,7 +302,7 @@ static void bilateral_##name(BilateralContext *s, const uint8_t *ssrc, uint8_t *
         ycf_ = ycf;                                                                       \
         ypf_ = ypf;                                                                       \
         factor_ = &map_factor_b[y * width];                                               \
-        for (int x = 0; x < width; x++) {                                                 \
+        for (int x = 0; x < slice_end - slice_start; x++) {                               \
             type dr = abs((*tcy++) - (*tpy++));                                           \
             int range_dist = dr;                                                          \
             float alpha_ = range_table[range_dist];                                       \
@@ -281,20 +321,119 @@ static void bilateral_##name(BilateralContext *s, const uint8_t *ssrc, uint8_t *
         ypy = ycy;                                                                        \
         ypf = ycf;                                                                        \
     }                                                                                     \
-                                                                                          \
-    for (int i = 0; i < height; i++)                                                      \
-        for (int j = 0; j < width; j++)                                                   \
-            dst[j + i * dst_linesize] = lrintf(img_out_f[i * width + j]);                 \
 }
 
-BILATERAL(uint8_t, byte)
-BILATERAL(uint16_t, word)
+BILATERAL_V(uint8_t, byte)
+BILATERAL_V(uint16_t, word)
+
+#define BILATERAL_O(type, name)                                               \
+static void bilateralo_##name(BilateralContext *s, AVFrame *out, AVFrame *in, \
+                              int jobnr, int nb_jobs, int plane)              \
+{                                                                             \
+    const int width = s->planewidth[plane];                                   \
+    const int height = s->planeheight[plane];                                 \
+    const int slice_start = (height * jobnr) / nb_jobs;                       \
+    const int slice_end = (height * (jobnr+1)) / nb_jobs;                     \
+    const int dst_linesize = out->linesize[plane] / sizeof(type);             \
+                                                                              \
+    for (int i = slice_start; i < slice_end; i++) {                           \
+        type *dst = (type *)out->data[plane] + i * dst_linesize;              \
+        const float *const img_out_f = s->img_out_f[plane] + i * width;       \
+        for (int j = 0; j < width; j++)                                       \
+            dst[j] = lrintf(img_out_f[j]);                                    \
+    }                                                                         \
+}
+
+BILATERAL_O(uint8_t, byte)
+BILATERAL_O(uint16_t, word)
+
+static int bilateralh_planes(AVFilterContext *ctx, void *arg,
+                             int jobnr, int nb_jobs)
+{
+    BilateralContext *s = ctx->priv;
+    ThreadData *td = arg;
+    AVFrame *out = td->out;
+    AVFrame *in = td->in;
+
+    for (int plane = 0; plane < s->nb_planes; plane++) {
+        if (!(s->planes & (1 << plane)))
+            continue;
+
+        if (s->depth <= 8)
+           bilateralh_byte(s, out, in, jobnr, nb_jobs, plane);
+        else
+           bilateralh_word(s, out, in, jobnr, nb_jobs, plane);
+    }
+
+    return 0;
+}
+
+static int bilateralv_planes(AVFilterContext *ctx, void *arg,
+                             int jobnr, int nb_jobs)
+{
+    BilateralContext *s = ctx->priv;
+    ThreadData *td = arg;
+    AVFrame *out = td->out;
+    AVFrame *in = td->in;
+
+    for (int plane = 0; plane < s->nb_planes; plane++) {
+        if (!(s->planes & (1 << plane)))
+            continue;
+
+        if (s->depth <= 8)
+           bilateralv_byte(s, out, in, jobnr, nb_jobs, plane);
+        else
+           bilateralv_word(s, out, in, jobnr, nb_jobs, plane);
+    }
+
+    return 0;
+}
+
+static int bilateralo_planes(AVFilterContext *ctx, void *arg,
+                            int jobnr, int nb_jobs)
+{
+    BilateralContext *s = ctx->priv;
+    ThreadData *td = arg;
+    AVFrame *out = td->out;
+    AVFrame *in = td->in;
+
+    for (int plane = 0; plane < s->nb_planes; plane++) {
+        if (!(s->planes & (1 << plane))) {
+            if (out != in) {
+                const int height = s->planeheight[plane];
+                const int slice_start = (height * jobnr) / nb_jobs;
+                const int slice_end = (height * (jobnr+1)) / nb_jobs;
+                const int width = s->planewidth[plane];
+                const int linesize = in->linesize[plane];
+                const int dst_linesize = out->linesize[plane];
+                const uint8_t *src = in->data[plane];
+                uint8_t *dst = out->data[plane];
+
+                av_image_copy_plane(dst + slice_start * dst_linesize,
+                                    dst_linesize,
+                                    src + slice_start * linesize,
+                                    linesize,
+                                    width * ((s->depth + 7) / 8),
+                                    slice_end - slice_start);
+            }
+            continue;
+        }
+
+        if (s->depth <= 8)
+           bilateralo_byte(s, out, in, jobnr, nb_jobs, plane);
+        else
+           bilateralo_word(s, out, in, jobnr, nb_jobs, plane);
+    }
+
+    return 0;
+}
 
 static int filter_frame(AVFilterLink *inlink, AVFrame *in)
 {
     AVFilterContext *ctx = inlink->dst;
     BilateralContext *s = ctx->priv;
     AVFilterLink *outlink = ctx->outputs[0];
+    ThreadData td;
     AVFrame *out;
 
     if (av_frame_is_writable(in)) {
@@ -308,24 +447,11 @@ static int filter_frame(AVFilterLink *inlink, AVFrame *in)
         av_frame_copy_props(out, in);
     }
 
-    for (int plane = 0; plane < s->nb_planes; plane++) {
-        if (!(s->planes & (1 << plane))) {
-            if (out != in)
-                av_image_copy_plane(out->data[plane], out->linesize[plane],
-                                    in->data[plane], in->linesize[plane],
-                                    s->planewidth[plane] * ((s->depth + 7) / 8), s->planeheight[plane]);
-            continue;
-        }
-
-        if (s->depth <= 8)
-           bilateral_byte(s, in->data[plane], out->data[plane],
-                      s->planewidth[plane], s->planeheight[plane],
-                      in->linesize[plane], out->linesize[plane]);
-        else
-           bilateral_word(s, in->data[plane], out->data[plane],
-                      s->planewidth[plane], s->planeheight[plane],
-                      in->linesize[plane] / 2, out->linesize[plane] / 2);
-    }
+    td.in = in;
+    td.out = out;
+    ff_filter_execute(ctx, bilateralh_planes, &td, NULL, s->nb_threads);
+    ff_filter_execute(ctx, bilateralv_planes, &td, NULL, s->nb_threads);
+    ff_filter_execute(ctx, bilateralo_planes, &td, NULL, s->nb_threads);
 
     if (out != in)
         av_frame_free(&in);
@@ -336,14 +462,16 @@ static av_cold void uninit(AVFilterContext *ctx)
 {
     BilateralContext *s = ctx->priv;
 
-    av_freep(&s->img_out_f);
-    av_freep(&s->img_temp);
-    av_freep(&s->map_factor_a);
-    av_freep(&s->map_factor_b);
-    av_freep(&s->slice_factor_a);
-    av_freep(&s->slice_factor_b);
-    av_freep(&s->line_factor_a);
-    av_freep(&s->line_factor_b);
+    for (int p = 0; p < s->nb_planes; p++) {
+        av_freep(&s->img_out_f[p]);
+        av_freep(&s->img_temp[p]);
+        av_freep(&s->map_factor_a[p]);
+        av_freep(&s->map_factor_b[p]);
+        av_freep(&s->slice_factor_a[p]);
+        av_freep(&s->slice_factor_b[p]);
+        av_freep(&s->line_factor_a[p]);
+        av_freep(&s->line_factor_b[p]);
+    }
 }
 
 static int process_command(AVFilterContext *ctx,
@@ -386,6 +514,7 @@ const AVFilter ff_vf_bilateral = {
     FILTER_INPUTS(bilateral_inputs),
     FILTER_OUTPUTS(bilateral_outputs),
     FILTER_PIXFMTS_ARRAY(pix_fmts),
-    .flags         = AVFILTER_FLAG_SUPPORT_TIMELINE_GENERIC,
+    .flags         = AVFILTER_FLAG_SUPPORT_TIMELINE_GENERIC |
+                     AVFILTER_FLAG_SLICE_THREADS,
     .process_command = process_command,
 };