avfilter/vf_unsharp: enable slice threading

benchmarking with a simple command:
ffmpeg -i 1080p.mp4 -vf unsharp=la=3:ca=3 -an -f null /dev/null
with the patch, the fps increase from 50 to 120 on my local machine (i7-6770HQ).

Signed-off-by: Ruiling Song <ruiling.song@intel.com>

2 files changed, 81 insertions, 25 deletions

diff --git a/libavfilter/unsharp.h b/libavfilter/unsharp.h
index caff986fc1..a60b30f31a 100644
--- a/libavfilter/unsharp.h
+++ b/libavfilter/unsharp.h
@@ -37,7 +37,8 @@ typedef struct UnsharpFilterParam {
     int steps_y;                             ///< vertical step count
     int scalebits;                           ///< bits to shift pixel
     int32_t halfscale;                       ///< amount to add to pixel
-    uint32_t *sc[MAX_MATRIX_SIZE - 1];       ///< finite state machine storage
+    uint32_t *sr;        ///< finite state machine storage within a row
+    uint32_t **sc;       ///< finite state machine storage across rows
 } UnsharpFilterParam;
 
 typedef struct UnsharpContext {
@@ -47,6 +48,7 @@ typedef struct UnsharpContext {
     UnsharpFilterParam luma;   ///< luma parameters (width, height, amount)
     UnsharpFilterParam chroma; ///< chroma parameters (width, height, amount)
     int hsub, vsub;
+    int nb_threads;
     int opencl;
     int (* apply_unsharp)(AVFilterContext *ctx, AVFrame *in, AVFrame *out);
 } UnsharpContext;
diff --git a/libavfilter/vf_unsharp.c b/libavfilter/vf_unsharp.c
index 41ccc56942..af05833a5d 100644
--- a/libavfilter/vf_unsharp.c
+++ b/libavfilter/vf_unsharp.c
@@ -47,15 +47,22 @@
 #include "libavutil/pixdesc.h"
 #include "unsharp.h"
 
-static void apply_unsharp(      uint8_t *dst, int dst_stride,
-                          const uint8_t *src, int src_stride,
-                          int width, int height, UnsharpFilterParam *fp)
+typedef struct TheadData {
+    UnsharpFilterParam *fp;
+    uint8_t       *dst;
+    const uint8_t *src;
+    int dst_stride;
+    int src_stride;
+    int width;
+    int height;
+} ThreadData;
+
+static int unsharp_slice(AVFilterContext *ctx, void *arg, int jobnr, int nb_jobs)
 {
+    ThreadData *td = arg;
+    UnsharpFilterParam *fp = td->fp;
     uint32_t **sc = fp->sc;
-    uint32_t sr[MAX_MATRIX_SIZE - 1], tmp1, tmp2;
-
-    int32_t res;
-    int x, y, z;
+    uint32_t *sr = fp->sr;
     const uint8_t *src2 = NULL;  //silence a warning
     const int amount = fp->amount;
     const int steps_x = fp->steps_x;
@@ -63,30 +70,54 @@ static void apply_unsharp(      uint8_t *dst, int dst_stride,
     const int scalebits = fp->scalebits;
     const int32_t halfscale = fp->halfscale;
 
+    uint8_t *dst = td->dst;
+    const uint8_t *src = td->src;
+    const int dst_stride = td->dst_stride;
+    const int src_stride = td->src_stride;
+    const int width = td->width;
+    const int height = td->height;
+    const int sc_offset = jobnr * 2 * steps_y;
+    const int sr_offset = jobnr * (MAX_MATRIX_SIZE - 1);
+    const int slice_start = (height * jobnr) / nb_jobs;
+    const int slice_end = (height * (jobnr+1)) / nb_jobs;
+
+    int32_t res;
+    int x, y, z;
+    uint32_t tmp1, tmp2;
+
     if (!amount) {
-        av_image_copy_plane(dst, dst_stride, src, src_stride, width, height);
-        return;
+        av_image_copy_plane(dst + slice_start * dst_stride, dst_stride,
+                            src + slice_start * src_stride, src_stride,
+                            width, slice_end - slice_start);
+        return 0;
     }
 
     for (y = 0; y < 2 * steps_y; y++)
-        memset(sc[y], 0, sizeof(sc[y][0]) * (width + 2 * steps_x));
+        memset(sc[sc_offset + y], 0, sizeof(sc[y][0]) * (width + 2 * steps_x));
 
-    for (y = -steps_y; y < height + steps_y; y++) {
+    // if this is not the first tile, we start from (slice_start - steps_y),
+    // so we can get smooth result at slice boundary
+    if (slice_start > steps_y) {
+        src += (slice_start - steps_y) * src_stride;
+        dst += (slice_start - steps_y) * dst_stride;
+    }
+
+    for (y = -steps_y + slice_start; y < steps_y + slice_end; y++) {
         if (y < height)
             src2 = src;
 
-        memset(sr, 0, sizeof(sr[0]) * (2 * steps_x - 1));
+        memset(sr + sr_offset, 0, sizeof(sr[0]) * (2 * steps_x - 1));
         for (x = -steps_x; x < width + steps_x; x++) {
             tmp1 = x <= 0 ? src2[0] : x >= width ? src2[width-1] : src2[x];
             for (z = 0; z < steps_x * 2; z += 2) {
-                tmp2 = sr[z + 0] + tmp1; sr[z + 0] = tmp1;
-                tmp1 = sr[z + 1] + tmp2; sr[z + 1] = tmp2;
+                tmp2 = sr[sr_offset + z + 0] + tmp1; sr[sr_offset + z + 0] = tmp1;
+                tmp1 = sr[sr_offset + z + 1] + tmp2; sr[sr_offset + z + 1] = tmp2;
             }
             for (z = 0; z < steps_y * 2; z += 2) {
-                tmp2 = sc[z + 0][x + steps_x] + tmp1; sc[z + 0][x + steps_x] = tmp1;
-                tmp1 = sc[z + 1][x + steps_x] + tmp2; sc[z + 1][x + steps_x] = tmp2;
+                tmp2 = sc[sc_offset + z + 0][x + steps_x] + tmp1; sc[sc_offset + z + 0][x + steps_x] = tmp1;
+                tmp1 = sc[sc_offset + z + 1][x + steps_x] + tmp2; sc[sc_offset + z + 1][x + steps_x] = tmp2;
             }
-            if (x >= steps_x && y >= steps_y) {
+            if (x >= steps_x && y >= (steps_y + slice_start)) {
                 const uint8_t *srx = src - steps_y * src_stride + x - steps_x;
                 uint8_t *dsx       = dst - steps_y * dst_stride + x - steps_x;
 
@@ -99,6 +130,7 @@ static void apply_unsharp(      uint8_t *dst, int dst_stride,
             src += src_stride;
         }
     }
+    return 0;
 }
 
 static int apply_unsharp_c(AVFilterContext *ctx, AVFrame *in, AVFrame *out)
@@ -107,6 +139,8 @@ static int apply_unsharp_c(AVFilterContext *ctx, AVFrame *in, AVFrame *out)
     UnsharpContext *s = ctx->priv;
     int i, plane_w[3], plane_h[3];
     UnsharpFilterParam *fp[3];
+    ThreadData td;
+
     plane_w[0] = inlink->w;
     plane_w[1] = plane_w[2] = AV_CEIL_RSHIFT(inlink->w, s->hsub);
     plane_h[0] = inlink->h;
@@ -114,7 +148,14 @@ static int apply_unsharp_c(AVFilterContext *ctx, AVFrame *in, AVFrame *out)
     fp[0] = &s->luma;
     fp[1] = fp[2] = &s->chroma;
     for (i = 0; i < 3; i++) {
-        apply_unsharp(out->data[i], out->linesize[i], in->data[i], in->linesize[i], plane_w[i], plane_h[i], fp[i]);
+        td.fp = fp[i];
+        td.dst = out->data[i];
+        td.src = in->data[i];
+        td.width = plane_w[i];
+        td.height = plane_h[i];
+        td.dst_stride = out->linesize[i];
+        td.src_stride = in->linesize[i];
+        ctx->internal->execute(ctx, unsharp_slice, &td, NULL, FFMIN(plane_h[i], s->nb_threads));
     }
     return 0;
 }
@@ -163,6 +204,7 @@ static int query_formats(AVFilterContext *ctx)
 static int init_filter_param(AVFilterContext *ctx, UnsharpFilterParam *fp, const char *effect_type, int width)
 {
     int z;
+    UnsharpContext *s = ctx->priv;
     const char *effect = fp->amount == 0 ? "none" : fp->amount < 0 ? "blur" : "sharpen";
 
     if  (!(fp->msize_x & fp->msize_y & 1)) {
@@ -175,7 +217,12 @@ static int init_filter_param(AVFilterContext *ctx, UnsharpFilterParam *fp, const
     av_log(ctx, AV_LOG_VERBOSE, "effect:%s type:%s msize_x:%d msize_y:%d amount:%0.2f\n",
            effect, effect_type, fp->msize_x, fp->msize_y, fp->amount / 65535.0);
 
-    for (z = 0; z < 2 * fp->steps_y; z++)
+    fp->sr = av_malloc_array((MAX_MATRIX_SIZE - 1) * s->nb_threads, sizeof(uint32_t));
+    fp->sc = av_malloc_array(2 * fp->steps_y * s->nb_threads, sizeof(uint32_t **));
+    if (!fp->sr || !fp->sc)
+        return AVERROR(ENOMEM);
+
+    for (z = 0; z < 2 * fp->steps_y * s->nb_threads; z++)
         if (!(fp->sc[z] = av_malloc_array(width + 2 * fp->steps_x,
                                           sizeof(*(fp->sc[z])))))
             return AVERROR(ENOMEM);
@@ -192,6 +239,11 @@ static int config_props(AVFilterLink *link)
     s->hsub = desc->log2_chroma_w;
     s->vsub = desc->log2_chroma_h;
 
+    // ensure (height / nb_threads) > 4 * steps_y,
+    // so that we don't have too much overlap between two threads
+    s->nb_threads = FFMIN(ff_filter_get_nb_threads(link->dst),
+                          link->h / (4 * s->luma.steps_y));
+
     ret = init_filter_param(link->dst, &s->luma,   "luma",   link->w);
     if (ret < 0)
         return ret;
@@ -202,20 +254,22 @@ static int config_props(AVFilterLink *link)
     return 0;
 }
 
-static void free_filter_param(UnsharpFilterParam *fp)
+static void free_filter_param(UnsharpFilterParam *fp, int nb_threads)
 {
     int z;
 
-    for (z = 0; z < 2 * fp->steps_y; z++)
+    for (z = 0; z < 2 * fp->steps_y * nb_threads; z++)
         av_freep(&fp->sc[z]);
+    av_freep(&fp->sc);
+    av_freep(&fp->sr);
 }
 
 static av_cold void uninit(AVFilterContext *ctx)
 {
     UnsharpContext *s = ctx->priv;
 
-    free_filter_param(&s->luma);
-    free_filter_param(&s->chroma);
+    free_filter_param(&s->luma, s->nb_threads);
+    free_filter_param(&s->chroma, s->nb_threads);
 }
 
 static int filter_frame(AVFilterLink *link, AVFrame *in)
@@ -294,5 +348,5 @@ AVFilter ff_vf_unsharp = {
     .query_formats = query_formats,
     .inputs        = avfilter_vf_unsharp_inputs,
     .outputs       = avfilter_vf_unsharp_outputs,
-    .flags         = AVFILTER_FLAG_SUPPORT_TIMELINE_GENERIC,
+    .flags         = AVFILTER_FLAG_SUPPORT_TIMELINE_GENERIC | AVFILTER_FLAG_SLICE_THREADS,
 };