#include <float.h>
+#include "libavutil/avassert.h"
#include "libavutil/common.h"
#include "libavutil/float_dsp.h"
#include "libavutil/imgutils.h"
static const uint8_t NNEDI_YDIM[] = { 6, 6, 6, 6, 4, 4, 4 };
static const uint16_t NNEDI_NNS[] = { 16, 32, 64, 128, 256 };
-typedef struct PrescreenerOldCoefficients {
- DECLARE_ALIGNED(32, float, kernel_l0)[4][14 * 4];
+typedef struct PrescreenerCoefficients {
+ DECLARE_ALIGNED(32, float, kernel_l0)[4][16 * 4];
DECLARE_ALIGNED(32, float, bias_l0)[4];
DECLARE_ALIGNED(32, float, kernel_l1)[4][4];
DECLARE_ALIGNED(32, float, kernel_l2)[4][8];
DECLARE_ALIGNED(32, float, bias_l2)[4];
-} PrescreenerOldCoefficients;
-
-typedef struct PrescreenerNewCoefficients {
- DECLARE_ALIGNED(32, float, kernel_l0)[4][16 * 4];
- DECLARE_ALIGNED(32, float, bias_l0)[4];
-
- DECLARE_ALIGNED(32, float, kernel_l1)[4][4];
- DECLARE_ALIGNED(32, float, bias_l1)[4];
-} PrescreenerNewCoefficients;
+} PrescreenerCoefficients;
typedef struct PredictorCoefficients {
int xdim, ydim, nns, nsize;
char *weights_file;
- AVFrame *src;
- AVFrame *second;
- AVFrame *dst;
+ AVFrame *prev;
int eof;
- int64_t cur_pts;
+ int64_t pts;
AVFloatDSPContext *fdsp;
int depth;
int planeheight[4];
int field_n;
- PrescreenerOldCoefficients prescreener_old;
- PrescreenerNewCoefficients prescreener_new[3];
+ PrescreenerCoefficients prescreener[4];
PredictorCoefficients coeffs[2][5][7];
float half;
int pscrn;
int input_size;
- uint8_t *prescreen_buf;
- float *input_buf;
- float *output_buf;
+ uint8_t **prescreen_buf;
+ float **input_buf;
+ float **output_buf;
void (*read)(const uint8_t *src, float *dst,
int src_stride, int dst_stride,
int width, int height, int depth, float scale);
void (*prescreen[2])(AVFilterContext *ctx,
const void *src, ptrdiff_t src_stride,
- uint8_t *prescreen, int N, void *data);
+ uint8_t *prescreen, int N,
+ const PrescreenerCoefficients *const coeffs);
} NNEDIContext;
#define OFFSET(x) offsetof(NNEDIContext, x)
return ff_set_common_formats(ctx, fmts_list);
}
-static float dot_dsp(NNEDIContext *s, const float *kernel, const float *input,
+static float dot_dsp(const NNEDIContext *const s, const float *kernel, const float *input,
int n, float scale, float bias)
{
- float sum;
+ float sum, y;
sum = s->fdsp->scalarproduct_float(kernel, input, n);
- return sum * scale + bias;
+ y = sum * scale + bias + 1e-20f;
+
+ return y;
}
static float elliott(float x)
static void process_old(AVFilterContext *ctx,
const void *src, ptrdiff_t src_stride,
uint8_t *prescreen, int N,
- void *data)
+ const PrescreenerCoefficients *const m_data)
{
NNEDIContext *s = ctx->priv;
- const PrescreenerOldCoefficients *const m_data = data;
const float *src_p = src;
// Adjust source pointer to point to top-left of filter window.
static void process_new(AVFilterContext *ctx,
const void *src, ptrdiff_t src_stride,
uint8_t *prescreen, int N,
- void *data)
+ const PrescreenerCoefficients *const m_data)
{
NNEDIContext *s = ctx->priv;
- const PrescreenerNewCoefficients *const m_data = data;
const float *src_p = src;
// Adjust source pointer to point to top-left of filter window.
float *buf, float mstd[4],
const PredictorCoefficients *const model)
{
- float sum = 0;
- float sum_sq = 0;
+ const float scale = 1.f / model->nsize;
+ float sum = 0.f;
+ float sum_sq = 0.f;
float tmp;
for (int i = 0; i < model->ydim; i++) {
buf += model->xdim;
}
- mstd[0] = sum / model->nsize;
+ mstd[0] = sum * scale;
mstd[3] = 0.f;
- tmp = sum_sq / model->nsize - mstd[0] * mstd[0];
+ tmp = sum_sq * scale - mstd[0] * mstd[0];
if (tmp < FLT_EPSILON) {
mstd[1] = 0.0f;
mstd[2] = 0.0f;
static void predictor(AVFilterContext *ctx,
const void *src, ptrdiff_t src_stride, void *dst,
const uint8_t *prescreen, int N,
- void *data, int use_q2)
+ const PredictorCoefficients *const model, int use_q2)
{
- NNEDIContext *s = ctx->priv;
- const PredictorCoefficients *const model = data;
+ const NNEDIContext *const s = ctx->priv;
const float *src_p = src;
float *dst_p = dst;
// Adjust source pointer to point to top-left of filter window.
const float *window = src_p - (model->ydim / 2) * src_stride - (model->xdim / 2 - 1);
- int filter_size = model->nsize;
- int nns = model->nns;
+ const int filter_size = model->nsize;
+ const int nns = model->nns;
for (int i = 0; i < N; i++) {
LOCAL_ALIGNED_32(float, input, [48 * 6]);
activation[nn] = dot_dsp(s, softmax_q1_filter(nn, model), input, filter_size, scale, model->softmax_bias_q1[nn]);
for (int nn = 0; nn < nns; nn++)
- activation[model->nns + nn] = dot_dsp(s, elliott_q1_filter(nn, model), input, filter_size, scale, model->elliott_bias_q1[nn]);
+ activation[nns + nn] = dot_dsp(s, elliott_q1_filter(nn, model), input, filter_size, scale, model->elliott_bias_q1[nn]);
transform_softmax_exp(activation, nns);
wae5(activation, activation + nns, nns, mstd);
wae5(activation, activation + nns, nns, mstd);
}
- dst_p[i] = mstd[3] / (use_q2 ? 2 : 1);
+ dst_p[i] = mstd[3] * (use_q2 ? 0.5f : 1.f);
}
}
static int filter_slice(AVFilterContext *ctx, void *arg, int jobnr, int nb_jobs)
{
- NNEDIContext *s = ctx->priv;
- AVFrame *out = s->dst;
- AVFrame *in = s->src;
+ const NNEDIContext *const s = ctx->priv;
+ AVFrame *out = arg;
+ AVFrame *in = s->prev;
const float in_scale = s->in_scale;
const float out_scale = s->out_scale;
const int depth = s->depth;
uint8_t *dst = out->data[p] + slice_start * out->linesize[p];
const int src_linesize = in->linesize[p];
const int dst_linesize = out->linesize[p];
- uint8_t *prescreen_buf = s->prescreen_buf + s->planewidth[0] * jobnr;
- float *srcbuf = s->input_buf + s->input_size * jobnr;
+ uint8_t *prescreen_buf = s->prescreen_buf[jobnr];
+ float *srcbuf = s->input_buf[jobnr];
const int srcbuf_stride = width + 64;
- float *dstbuf = s->output_buf + s->input_size * jobnr;
+ float *dstbuf = s->output_buf[jobnr];
const int dstbuf_stride = width;
const int slice_height = (slice_end - slice_start) / 2;
const int last_slice = slice_end == height;
width, 1, in_scale);
for (int y = 0; y < slice_end - slice_start; y += 2) {
- if (s->pscrn > 1) {
- s->prescreen[1](ctx, srcbuf + (y / 2) * srcbuf_stride + 32,
- srcbuf_stride, prescreen_buf, width,
- &s->prescreener_new[s->pscrn - 2]);
- } else if (s->pscrn == 1) {
- s->prescreen[0](ctx, srcbuf + (y / 2) * srcbuf_stride + 32,
- srcbuf_stride, prescreen_buf, width,
- &s->prescreener_old);
- }
+ if (s->pscrn > 0)
+ s->prescreen[s->pscrn > 1](ctx, srcbuf + (y / 2) * srcbuf_stride + 32,
+ srcbuf_stride, prescreen_buf, width,
+ &s->prescreener[s->pscrn - 1]);
predictor(ctx,
srcbuf + (y / 2) * srcbuf_stride + 32,
prescreen_buf, width,
&s->coeffs[s->etype][s->nnsparam][s->nsize], s->qual == 2);
- if (s->prescreen > 0)
+ if (s->pscrn > 0)
interpolation(srcbuf + (y / 2) * srcbuf_stride + 32,
srcbuf_stride,
dstbuf + (y / 2) * dstbuf_stride,
{
NNEDIContext *s = ctx->priv;
AVFilterLink *outlink = ctx->outputs[0];
- AVFrame *src = s->src;
+ AVFrame *dst;
- s->dst = ff_get_video_buffer(outlink, outlink->w, outlink->h);
- if (!s->dst)
+ dst = ff_get_video_buffer(outlink, outlink->w, outlink->h);
+ if (!dst)
return AVERROR(ENOMEM);
- av_frame_copy_props(s->dst, src);
- s->dst->interlaced_frame = 0;
+ av_frame_copy_props(dst, s->prev);
+ dst->interlaced_frame = 0;
+ dst->pts = s->pts;
- ctx->internal->execute(ctx, filter_slice, NULL, NULL, FFMIN(s->planeheight[1] / 2, s->nb_threads));
+ ctx->internal->execute(ctx, filter_slice, dst, NULL, FFMIN(s->planeheight[1] / 2, s->nb_threads));
if (s->field == -2 || s->field > 1)
s->field_n = !s->field_n;
- return 0;
+ return ff_filter_frame(outlink, dst);
}
-static int filter_frame(AVFilterLink *inlink, AVFrame *src)
+static int filter_frame(AVFilterLink *inlink, AVFrame *in)
{
AVFilterContext *ctx = inlink->dst;
- AVFilterLink *outlink = ctx->outputs[0];
NNEDIContext *s = ctx->priv;
int ret;
- if ((s->field > 1 ||
- s->field == -2) && !s->second) {
- goto second;
- } else if (s->field > 1 ||
- s->field == -2) {
- AVFrame *dst;
-
- s->src = s->second;
- ret = get_frame(ctx, 1);
- if (ret < 0) {
- av_frame_free(&s->dst);
- av_frame_free(&s->second);
- s->src = NULL;
- return ret;
- }
- dst = s->dst;
-
- if (src->pts != AV_NOPTS_VALUE &&
- dst->pts != AV_NOPTS_VALUE)
- dst->pts += src->pts;
- else
- dst->pts = AV_NOPTS_VALUE;
-
- ret = ff_filter_frame(outlink, dst);
- if (ret < 0)
- return ret;
- if (s->eof)
- return 0;
- s->cur_pts = s->second->pts;
- av_frame_free(&s->second);
-second:
- if ((s->deint && src->interlaced_frame &&
- !ctx->is_disabled) ||
- (!s->deint && !ctx->is_disabled)) {
- s->second = src;
- }
+ if (!s->prev) {
+ s->prev = in;
+ return 0;
}
- if ((s->deint && !src->interlaced_frame) || ctx->is_disabled) {
- AVFrame *dst = av_frame_clone(src);
- if (!dst) {
- av_frame_free(&src);
- av_frame_free(&s->second);
- return AVERROR(ENOMEM);
- }
-
- if (s->field > 1 || s->field == -2) {
- av_frame_free(&s->second);
- if ((s->deint && src->interlaced_frame) ||
- (!s->deint))
- s->second = src;
- } else {
- av_frame_free(&src);
- }
- if (dst->pts != AV_NOPTS_VALUE)
- dst->pts *= 2;
- return ff_filter_frame(outlink, dst);
+ if ((s->deint && !in->interlaced_frame) || ctx->is_disabled) {
+ s->prev->pts *= 2;
+ ret = ff_filter_frame(ctx->outputs[0], s->prev);
+ s->prev = in;
+ return ret;
}
- s->src = src;
+ s->pts = s->prev->pts * 2;
ret = get_frame(ctx, 0);
- if (ret < 0) {
- av_frame_free(&s->dst);
- av_frame_free(&s->src);
- av_frame_free(&s->second);
+ if (ret < 0 || (s->field > -2 && s->field < 2)) {
+ av_frame_free(&s->prev);
+ s->prev = in;
return ret;
}
- if (src->pts != AV_NOPTS_VALUE)
- s->dst->pts = src->pts * 2;
- if (s->field <= 1 && s->field > -2) {
- av_frame_free(&src);
- s->src = NULL;
- }
-
- return ff_filter_frame(outlink, s->dst);
+ s->pts = s->prev->pts + in->pts;
+ ret = get_frame(ctx, 1);
+ av_frame_free(&s->prev);
+ s->prev = in;
+ return ret;
}
static int request_frame(AVFilterLink *link)
ret = ff_request_frame(ctx->inputs[0]);
- if (ret == AVERROR_EOF && s->second) {
- AVFrame *next = av_frame_clone(s->second);
+ if (ret == AVERROR_EOF && s->prev) {
+ AVFrame *next = av_frame_clone(s->prev);
if (!next)
return AVERROR(ENOMEM);
- next->pts = s->second->pts * 2 - s->cur_pts;
+ next->pts = s->prev->pts + av_rescale_q(1, av_inv_q(ctx->outputs[0]->frame_rate),
+ ctx->outputs[0]->time_base);
s->eof = 1;
- filter_frame(ctx->inputs[0], next);
+ ret = filter_frame(ctx->inputs[0], next);
} else if (ret < 0) {
return ret;
}
- return 0;
+ return ret;
}
static void copy_weights(float *dst, int n, const float **data)
int bias_size = nns;
float *data;
- data = av_malloc_array(filter_size + bias_size, 4 * sizeof(float));
+ data = av_calloc(filter_size + bias_size, 4 * sizeof(float));
if (!data)
return AVERROR(ENOMEM);
NNEDIContext *s = ctx->priv;
int ret;
- copy_weights(&s->prescreener_old.kernel_l0[0][0], 4 * 48, &bdata);
- copy_weights(s->prescreener_old.bias_l0, 4, &bdata);
+ copy_weights(&s->prescreener[0].kernel_l0[0][0], 4 * 48, &bdata);
+ copy_weights(s->prescreener[0].bias_l0, 4, &bdata);
- copy_weights(&s->prescreener_old.kernel_l1[0][0], 4 * 4, &bdata);
- copy_weights(s->prescreener_old.bias_l1, 4, &bdata);
+ copy_weights(&s->prescreener[0].kernel_l1[0][0], 4 * 4, &bdata);
+ copy_weights(s->prescreener[0].bias_l1, 4, &bdata);
- copy_weights(&s->prescreener_old.kernel_l2[0][0], 4 * 8, &bdata);
- copy_weights(s->prescreener_old.bias_l2, 4, &bdata);
+ copy_weights(&s->prescreener[0].kernel_l2[0][0], 4 * 8, &bdata);
+ copy_weights(s->prescreener[0].bias_l2, 4, &bdata);
for (int i = 0; i < 3; i++) {
- PrescreenerNewCoefficients *data = &s->prescreener_new[i];
+ PrescreenerCoefficients *data = &s->prescreener[i + 1];
float kernel_l0_shuffled[4 * 64];
float kernel_l1_shuffled[4 * 4];
for (int m = 0; m < 2; m++) {
// Grouping by neuron count.
for (int i = 0; i < 5; i++) {
- int nns = NNEDI_NNS[i];
+ const int nns = NNEDI_NNS[i];
// Grouping by window size.
for (int j = 0; j < 7; j++) {
PredictorCoefficients *model = &s->coeffs[m][i][j];
- int xdim = NNEDI_XDIM[j];
- int ydim = NNEDI_YDIM[j];
- int filter_size = xdim * ydim;
+ const int xdim = NNEDI_XDIM[j];
+ const int ydim = NNEDI_YDIM[j];
+ const int filter_size = xdim * ydim;
ret = allocate_model(model, xdim, ydim, nns);
if (ret < 0)
static float mean(const float *input, int size)
{
- float sum = 0.;
+ float sum = 0.f;
for (int i = 0; i < size; i++)
sum += input[i];
input[i] = (input[i] - mean) / half;
}
-static void subtract_mean_old(PrescreenerOldCoefficients *coeffs, float half)
+static void subtract_mean_old(PrescreenerCoefficients *coeffs, float half)
{
for (int n = 0; n < 4; n++) {
float m = mean(coeffs->kernel_l0[n], 48);
}
}
-static void subtract_mean_new(PrescreenerNewCoefficients *coeffs, float half)
+static void subtract_mean_new(PrescreenerCoefficients *coeffs, float half)
{
for (int n = 0; n < 4; n++) {
float m = mean(coeffs->kernel_l0[n], 64);
static void subtract_mean_predictor(PredictorCoefficients *model)
{
- int filter_size = model->nsize;
- int nns = model->nns;
+ const int filter_size = model->nsize;
+ const int nns = model->nns;
+ const float scale = 1.f / nns;
- float softmax_means[256]; // Average of individual softmax filters.
- float elliott_means[256]; // Average of individual elliott filters.
- float mean_filter[48 * 6]; // Pointwise average of all softmax filters.
- float mean_bias;
+ double softmax_means[256]; // Average of individual softmax filters.
+ double elliott_means[256]; // Average of individual elliott filters.
+ double mean_filter[48 * 6] = { 0 }; // Pointwise average of all softmax filters.
+ double mean_bias;
// Quality 1.
for (int nn = 0; nn < nns; nn++) {
}
for (int k = 0; k < filter_size; k++)
- mean_filter[k] /= nns;
+ mean_filter[k] *= scale;
mean_bias = mean(model->softmax_bias_q1, nns);
}
// Quality 2.
- memset(mean_filter, 0, 48 * 6 * sizeof(float));
+ memset(mean_filter, 0, sizeof(mean_filter));
for (int nn = 0; nn < nns; nn++) {
softmax_means[nn] = mean(model->softmax_q2 + nn * filter_size, filter_size);
}
for (int k = 0; k < filter_size; k++)
- mean_filter[k] /= nns;
+ mean_filter[k] *= scale;
mean_bias = mean(model->softmax_bias_q2, nns);
break;
}
- subtract_mean_old(&s->prescreener_old, s->half);
- subtract_mean_new(&s->prescreener_new[0], s->half);
- subtract_mean_new(&s->prescreener_new[1], s->half);
- subtract_mean_new(&s->prescreener_new[2], s->half);
+ subtract_mean_old(&s->prescreener[0], s->half);
+ subtract_mean_new(&s->prescreener[1], s->half);
+ subtract_mean_new(&s->prescreener[2], s->half);
+ subtract_mean_new(&s->prescreener[3], s->half);
s->prescreen[0] = process_old;
s->prescreen[1] = process_new;
}
}
- s->prescreen_buf = av_calloc(s->nb_threads * s->planewidth[0], sizeof(*s->prescreen_buf));
- if (!s->prescreen_buf)
- return AVERROR(ENOMEM);
-
s->input_size = (s->planewidth[0] + 64) * (s->planeheight[0] + 6);
- s->input_buf = av_calloc(s->nb_threads * s->input_size, sizeof(*s->input_buf));
+ s->input_buf = av_calloc(s->nb_threads, sizeof(*s->input_buf));
if (!s->input_buf)
return AVERROR(ENOMEM);
- s->output_buf = av_calloc(s->nb_threads * s->input_size, sizeof(*s->output_buf));
+ for (int i = 0; i < s->nb_threads; i++) {
+ s->input_buf[i] = av_calloc(s->input_size, sizeof(**s->input_buf));
+ if (!s->input_buf[i])
+ return AVERROR(ENOMEM);
+ }
+
+ s->output_buf = av_calloc(s->nb_threads, sizeof(*s->output_buf));
if (!s->output_buf)
return AVERROR(ENOMEM);
+ for (int i = 0; i < s->nb_threads; i++) {
+ s->output_buf[i] = av_calloc(s->input_size, sizeof(**s->output_buf));
+ if (!s->output_buf[i])
+ return AVERROR(ENOMEM);
+ }
+
+ s->prescreen_buf = av_calloc(s->nb_threads, sizeof(*s->prescreen_buf));
+ if (!s->prescreen_buf)
+ return AVERROR(ENOMEM);
+
+ for (int i = 0; i < s->nb_threads; i++) {
+ s->prescreen_buf[i] = av_calloc(s->planewidth[0], sizeof(**s->prescreen_buf));
+ if (!s->prescreen_buf[i])
+ return AVERROR(ENOMEM);
+ }
+
return 0;
}
{
NNEDIContext *s = ctx->priv;
+ for (int i = 0; i < s->nb_threads && s->prescreen_buf; i++)
+ av_freep(&s->prescreen_buf[i]);
+
av_freep(&s->prescreen_buf);
+
+ for (int i = 0; i < s->nb_threads && s->input_buf; i++)
+ av_freep(&s->input_buf[i]);
+
av_freep(&s->input_buf);
+
+ for (int i = 0; i < s->nb_threads && s->output_buf; i++)
+ av_freep(&s->output_buf[i]);
+
av_freep(&s->output_buf);
av_freep(&s->fdsp);
}
}
- av_frame_free(&s->second);
+ av_frame_free(&s->prev);
}
static const AVFilterPad inputs[] = {
{ NULL }
};
-AVFilter ff_vf_nnedi = {
+const AVFilter ff_vf_nnedi = {
.name = "nnedi",
.description = NULL_IF_CONFIG_SMALL("Apply neural network edge directed interpolation intra-only deinterlacer."),
.priv_size = sizeof(NNEDIContext),