Merge commit 'c0bd865ad60da31282c5d8e1000c98366249c31e'

[ffmpeg] / libavfilter / af_afir.c

diff --git a/libavfilter/af_afir.c b/libavfilter/af_afir.c
index 3e13edf96ad475d242af7f8a5ab3567a1b1ec604..31919f62e9c93d684dbf6b78140d6c71bd672e3e 100644 (file)
--- a/libavfilter/af_afir.c
+++ b/libavfilter/af_afir.c
@@ -56,12 +56,12 @@ static void fcmul_add_c(float *sum, const float *t, const float *c, ptrdiff_t le
     sum[2 * n] += t[2 * n] * c[2 * n];
 }
 
-static int fir_channel(AVFilterContext *ctx, void *arg, int ch, int nb_jobs)
+static int fir_quantum(AVFilterContext *ctx, AVFrame *out, int ch, int offset)
 {
     AudioFIRContext *s = ctx->priv;
-    const float *in = (const float *)s->in[0]->extended_data[ch];
-    AVFrame *out = arg;
-    float *block, *buf, *ptr = (float *)out->extended_data[ch];
+    const float *in = (const float *)s->in[0]->extended_data[ch] + offset;
+    float *block, *buf, *ptr = (float *)out->extended_data[ch] + offset;
+    const int nb_samples = FFMIN(s->min_part_size, out->nb_samples - offset);
     int n, i, j;
 
     for (int segment = 0; segment < s->nb_segments; segment++) {
@@ -70,7 +70,7 @@ static int fir_channel(AVFilterContext *ctx, void *arg, int ch, int nb_jobs)
         float *dst = (float *)seg->output->extended_data[ch];
         float *sum = (float *)seg->sum->extended_data[ch];
 
-        s->fdsp->vector_fmul_scalar(src + seg->input_offset, in, s->dry_gain, FFALIGN(out->nb_samples, 4));
+        s->fdsp->vector_fmul_scalar(src + seg->input_offset, in, s->dry_gain, FFALIGN(nb_samples, 4));
         emms_c();
 
         seg->output_offset[ch] += s->min_part_size;
@@ -80,7 +80,7 @@ static int fir_channel(AVFilterContext *ctx, void *arg, int ch, int nb_jobs)
             memmove(src, src + s->min_part_size, (seg->input_size - s->min_part_size) * sizeof(*src));
 
             dst += seg->output_offset[ch];
-            for (n = 0; n < out->nb_samples; n++) {
+            for (n = 0; n < nb_samples; n++) {
                 ptr[n] += dst[n];
             }
             continue;
@@ -103,7 +103,7 @@ static int fir_channel(AVFilterContext *ctx, void *arg, int ch, int nb_jobs)
             const float *block = (const float *)seg->block->extended_data[ch] + i * seg->block_size;
             const FFTComplex *coeff = (const FFTComplex *)seg->coeff->extended_data[ch * !s->one2many] + coffset;
 
-            s->fcmul_add(sum, block, (const float *)coeff, seg->part_size);
+            s->afirdsp.fcmul_add(sum, block, (const float *)coeff, seg->part_size);
 
             if (j == 0)
                 j = seg->nb_partitions;
@@ -127,17 +127,41 @@ static int fir_channel(AVFilterContext *ctx, void *arg, int ch, int nb_jobs)
 
         memmove(src, src + s->min_part_size, (seg->input_size - s->min_part_size) * sizeof(*src));
 
-        for (n = 0; n < out->nb_samples; n++) {
+        for (n = 0; n < nb_samples; n++) {
             ptr[n] += dst[n];
         }
     }
 
-    s->fdsp->vector_fmul_scalar(ptr, ptr, s->wet_gain, FFALIGN(out->nb_samples, 4));
+    s->fdsp->vector_fmul_scalar(ptr, ptr, s->wet_gain, FFALIGN(nb_samples, 4));
     emms_c();
 
     return 0;
 }
 
+static int fir_channel(AVFilterContext *ctx, AVFrame *out, int ch)
+{
+    AudioFIRContext *s = ctx->priv;
+
+    for (int offset = 0; offset < out->nb_samples; offset += s->min_part_size) {
+        fir_quantum(ctx, out, ch, offset);
+    }
+
+    return 0;
+}
+
+static int fir_channels(AVFilterContext *ctx, void *arg, int jobnr, int nb_jobs)
+{
+    AVFrame *out = arg;
+    const int start = (out->channels * jobnr) / nb_jobs;
+    const int end = (out->channels * (jobnr+1)) / nb_jobs;
+
+    for (int ch = start; ch < end; ch++) {
+        fir_channel(ctx, out, ch);
+    }
+
+    return 0;
+}
+
 static int fir_frame(AudioFIRContext *s, AVFrame *in, AVFilterLink *outlink)
 {
     AVFilterContext *ctx = outlink->src;
@@ -152,7 +176,8 @@ static int fir_frame(AudioFIRContext *s, AVFrame *in, AVFilterLink *outlink)
     if (s->pts == AV_NOPTS_VALUE)
         s->pts = in->pts;
     s->in[0] = in;
-    ctx->internal->execute(ctx, fir_channel, out, NULL, outlink->channels);
+    ctx->internal->execute(ctx, fir_channels, out, NULL, FFMIN(outlink->channels,
+                                                               ff_filter_get_nb_threads(ctx)));
 
     out->pts = s->pts;
     if (s->pts != AV_NOPTS_VALUE)
@@ -312,7 +337,7 @@ static int init_segment(AVFilterContext *ctx, AudioFIRSegment *seg,
     if (!seg->rdft || !seg->irdft)
         return AVERROR(ENOMEM);
 
-    seg->fft_length    = part_size * 4 + 1;
+    seg->fft_length    = part_size * 2 + 1;
     seg->part_size     = part_size;
     seg->block_size    = FFALIGN(seg->fft_length, 32);
     seg->coeff_size    = FFALIGN(seg->part_size + 1, 32);
@@ -511,8 +536,8 @@ static int activate(AVFilterContext *ctx)
 {
     AudioFIRContext *s = ctx->priv;
     AVFilterLink *outlink = ctx->outputs[0];
+    int ret, status, available, wanted;
     AVFrame *in = NULL;
-    int ret, status;
     int64_t pts;
 
     FF_FILTER_FORWARD_STATUS_BACK_ALL(ctx->outputs[0], ctx);
@@ -543,7 +568,9 @@ static int activate(AVFilterContext *ctx)
             return ret;
     }
 
-    ret = ff_inlink_consume_samples(ctx->inputs[0], s->min_part_size, s->min_part_size, &in);
+    available = ff_inlink_queued_samples(ctx->inputs[0]);
+    wanted = FFMAX(s->min_part_size, (available / s->min_part_size) * s->min_part_size);
+    ret = ff_inlink_consume_samples(ctx->inputs[0], wanted, wanted, &in);
     if (ret > 0)
         ret = fir_frame(s, in, outlink);
 
@@ -726,6 +753,14 @@ static int config_video(AVFilterLink *outlink)
     return 0;
 }
 
+void ff_afir_init(AudioFIRDSPContext *dsp)
+{
+    dsp->fcmul_add = fcmul_add_c;
+
+    if (ARCH_X86)
+        ff_afir_init_x86(dsp);
+}
+
 static av_cold int init(AVFilterContext *ctx)
 {
     AudioFIRContext *s = ctx->priv;
@@ -765,14 +800,11 @@ static av_cold int init(AVFilterContext *ctx)
         }
     }
 
-    s->fcmul_add = fcmul_add_c;
-
     s->fdsp = avpriv_float_dsp_alloc(0);
     if (!s->fdsp)
         return AVERROR(ENOMEM);
 
-    if (ARCH_X86)
-        ff_afir_init_x86(s);
+    ff_afir_init(&s->afirdsp);
 
     return 0;
 }
@@ -810,8 +842,8 @@ static const AVOption afir_options[] = {
     { "channel", "set IR channel to display frequency response", OFFSET(ir_channel), AV_OPT_TYPE_INT, {.i64=0}, 0, 1024, VF },
     { "size",   "set video size",    OFFSET(w),          AV_OPT_TYPE_IMAGE_SIZE, {.str = "hd720"}, 0, 0, VF },
     { "rate",   "set video rate",    OFFSET(frame_rate), AV_OPT_TYPE_VIDEO_RATE, {.str = "25"}, 0, INT32_MAX, VF },
-    { "minp",   "set min partition size", OFFSET(minp),  AV_OPT_TYPE_INT,   {.i64=8192},  16, 32768, AF },
-    { "maxp",   "set max partition size", OFFSET(maxp),  AV_OPT_TYPE_INT,   {.i64=8192},  16, 32768, AF },
+    { "minp",   "set min partition size", OFFSET(minp),  AV_OPT_TYPE_INT,   {.i64=8192}, 8, 32768, AF },
+    { "maxp",   "set max partition size", OFFSET(maxp),  AV_OPT_TYPE_INT,   {.i64=8192}, 8, 32768, AF },
     { NULL }
 };