2 * Copyright (C) 2010-2011 Kevin Stone
3 * Copyright (C) 2016 Paul B Mahol
5 * This file is part of FFmpeg.
7 * FFmpeg is free software; you can redistribute it and/or modify
8 * it under the terms of the GNU General Public License as published by
9 * the Free Software Foundation; either version 2 of the License, or
10 * (at your option) any later version.
12 * FFmpeg is distributed in the hope that it will be useful,
13 * but WITHOUT ANY WARRANTY; without even the implied warranty of
14 * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the
15 * GNU General Public License for more details.
17 * You should have received a copy of the GNU General Public License along
18 * with FFmpeg; if not, write to the Free Software Foundation, Inc.,
19 * 51 Franklin Street, Fifth Floor, Boston, MA 02110-1301 USA.
24 #include "libavutil/avassert.h"
25 #include "libavutil/common.h"
26 #include "libavutil/float_dsp.h"
27 #include "libavutil/imgutils.h"
28 #include "libavutil/mem_internal.h"
29 #include "libavutil/opt.h"
30 #include "libavutil/pixdesc.h"
36 static const size_t NNEDI_WEIGHTS_SIZE = 13574928;
37 static const uint8_t NNEDI_XDIM[] = { 8, 16, 32, 48, 8, 16, 32 };
38 static const uint8_t NNEDI_YDIM[] = { 6, 6, 6, 6, 4, 4, 4 };
39 static const uint16_t NNEDI_NNS[] = { 16, 32, 64, 128, 256 };
41 typedef struct PrescreenerCoefficients {
42 DECLARE_ALIGNED(32, float, kernel_l0)[4][16 * 4];
43 DECLARE_ALIGNED(32, float, bias_l0)[4];
45 DECLARE_ALIGNED(32, float, kernel_l1)[4][4];
46 DECLARE_ALIGNED(32, float, bias_l1)[4];
48 DECLARE_ALIGNED(32, float, kernel_l2)[4][8];
49 DECLARE_ALIGNED(32, float, bias_l2)[4];
50 } PrescreenerCoefficients;
52 typedef struct PredictorCoefficients {
53 int xdim, ydim, nns, nsize;
57 float *softmax_bias_q1;
58 float *elliott_bias_q1;
61 float *softmax_bias_q2;
62 float *elliott_bias_q2;
63 } PredictorCoefficients;
65 typedef struct NNEDIContext {
74 AVFloatDSPContext *fdsp;
83 PrescreenerCoefficients prescreener[4];
84 PredictorCoefficients coeffs[2][5][7];
101 uint8_t **prescreen_buf;
105 void (*read)(const uint8_t *src, float *dst,
106 int src_stride, int dst_stride,
107 int width, int height, float scale);
108 void (*write)(const float *src, uint8_t *dst,
109 int src_stride, int dst_stride,
110 int width, int height, int depth, float scale);
111 void (*prescreen[2])(AVFilterContext *ctx,
112 const void *src, ptrdiff_t src_stride,
113 uint8_t *prescreen, int N,
114 const PrescreenerCoefficients *const coeffs);
117 #define OFFSET(x) offsetof(NNEDIContext, x)
118 #define RFLAGS AV_OPT_FLAG_VIDEO_PARAM|AV_OPT_FLAG_FILTERING_PARAM|AV_OPT_FLAG_RUNTIME_PARAM
119 #define FLAGS AV_OPT_FLAG_VIDEO_PARAM|AV_OPT_FLAG_FILTERING_PARAM
121 static const AVOption nnedi_options[] = {
122 {"weights", "set weights file", OFFSET(weights_file), AV_OPT_TYPE_STRING, {.str="nnedi3_weights.bin"}, 0, 0, FLAGS },
123 {"deint", "set which frames to deinterlace", OFFSET(deint), AV_OPT_TYPE_INT, {.i64=0}, 0, 1, RFLAGS, "deint" },
124 {"all", "deinterlace all frames", 0, AV_OPT_TYPE_CONST, {.i64=0}, 0, 0, RFLAGS, "deint" },
125 {"interlaced", "only deinterlace frames marked as interlaced", 0, AV_OPT_TYPE_CONST, {.i64=1}, 0, 0, RFLAGS, "deint" },
126 {"field", "set mode of operation", OFFSET(field), AV_OPT_TYPE_INT, {.i64=-1}, -2, 3, RFLAGS, "field" },
127 {"af", "use frame flags, both fields", 0, AV_OPT_TYPE_CONST, {.i64=-2}, 0, 0, RFLAGS, "field" },
128 {"a", "use frame flags, single field", 0, AV_OPT_TYPE_CONST, {.i64=-1}, 0, 0, RFLAGS, "field" },
129 {"t", "use top field only", 0, AV_OPT_TYPE_CONST, {.i64=0}, 0, 0, RFLAGS, "field" },
130 {"b", "use bottom field only", 0, AV_OPT_TYPE_CONST, {.i64=1}, 0, 0, RFLAGS, "field" },
131 {"tf", "use both fields, top first", 0, AV_OPT_TYPE_CONST, {.i64=2}, 0, 0, RFLAGS, "field" },
132 {"bf", "use both fields, bottom first", 0, AV_OPT_TYPE_CONST, {.i64=3}, 0, 0, RFLAGS, "field" },
133 {"planes", "set which planes to process", OFFSET(process_plane), AV_OPT_TYPE_INT, {.i64=7}, 0, 15, RFLAGS },
134 {"nsize", "set size of local neighborhood around each pixel, used by the predictor neural network", OFFSET(nsize), AV_OPT_TYPE_INT, {.i64=6}, 0, 6, RFLAGS, "nsize" },
135 {"s8x6", NULL, 0, AV_OPT_TYPE_CONST, {.i64=0}, 0, 0, RFLAGS, "nsize" },
136 {"s16x6", NULL, 0, AV_OPT_TYPE_CONST, {.i64=1}, 0, 0, RFLAGS, "nsize" },
137 {"s32x6", NULL, 0, AV_OPT_TYPE_CONST, {.i64=2}, 0, 0, RFLAGS, "nsize" },
138 {"s48x6", NULL, 0, AV_OPT_TYPE_CONST, {.i64=3}, 0, 0, RFLAGS, "nsize" },
139 {"s8x4", NULL, 0, AV_OPT_TYPE_CONST, {.i64=4}, 0, 0, RFLAGS, "nsize" },
140 {"s16x4", NULL, 0, AV_OPT_TYPE_CONST, {.i64=5}, 0, 0, RFLAGS, "nsize" },
141 {"s32x4", NULL, 0, AV_OPT_TYPE_CONST, {.i64=6}, 0, 0, RFLAGS, "nsize" },
142 {"nns", "set number of neurons in predictor neural network", OFFSET(nnsparam), AV_OPT_TYPE_INT, {.i64=1}, 0, 4, RFLAGS, "nns" },
143 {"n16", NULL, 0, AV_OPT_TYPE_CONST, {.i64=0}, 0, 0, RFLAGS, "nns" },
144 {"n32", NULL, 0, AV_OPT_TYPE_CONST, {.i64=1}, 0, 0, RFLAGS, "nns" },
145 {"n64", NULL, 0, AV_OPT_TYPE_CONST, {.i64=2}, 0, 0, RFLAGS, "nns" },
146 {"n128", NULL, 0, AV_OPT_TYPE_CONST, {.i64=3}, 0, 0, RFLAGS, "nns" },
147 {"n256", NULL, 0, AV_OPT_TYPE_CONST, {.i64=4}, 0, 0, RFLAGS, "nns" },
148 {"qual", "set quality", OFFSET(qual), AV_OPT_TYPE_INT, {.i64=1}, 1, 2, RFLAGS, "qual" },
149 {"fast", NULL, 0, AV_OPT_TYPE_CONST, {.i64=1}, 0, 0, RFLAGS, "qual" },
150 {"slow", NULL, 0, AV_OPT_TYPE_CONST, {.i64=2}, 0, 0, RFLAGS, "qual" },
151 {"etype", "set which set of weights to use in the predictor", OFFSET(etype), AV_OPT_TYPE_INT, {.i64=0}, 0, 1, RFLAGS, "etype" },
152 {"a", "weights trained to minimize absolute error", 0, AV_OPT_TYPE_CONST, {.i64=0}, 0, 0, RFLAGS, "etype" },
153 {"abs","weights trained to minimize absolute error", 0, AV_OPT_TYPE_CONST, {.i64=0}, 0, 0, RFLAGS, "etype" },
154 {"s", "weights trained to minimize squared error", 0, AV_OPT_TYPE_CONST, {.i64=1}, 0, 0, RFLAGS, "etype" },
155 {"mse","weights trained to minimize squared error", 0, AV_OPT_TYPE_CONST, {.i64=1}, 0, 0, RFLAGS, "etype" },
156 {"pscrn", "set prescreening", OFFSET(pscrn), AV_OPT_TYPE_INT, {.i64=2}, 0, 4, RFLAGS, "pscrn" },
157 {"none", NULL, 0, AV_OPT_TYPE_CONST, {.i64=0}, 0, 0, RFLAGS, "pscrn" },
158 {"original", NULL, 0, AV_OPT_TYPE_CONST, {.i64=1}, 0, 0, RFLAGS, "pscrn" },
159 {"new", NULL, 0, AV_OPT_TYPE_CONST, {.i64=2}, 0, 0, RFLAGS, "pscrn" },
160 {"new2", NULL, 0, AV_OPT_TYPE_CONST, {.i64=3}, 0, 0, RFLAGS, "pscrn" },
161 {"new3", NULL, 0, AV_OPT_TYPE_CONST, {.i64=4}, 0, 0, RFLAGS, "pscrn" },
165 AVFILTER_DEFINE_CLASS(nnedi);
167 static int config_output(AVFilterLink *outlink)
169 AVFilterContext *ctx = outlink->src;
171 outlink->time_base.num = ctx->inputs[0]->time_base.num;
172 outlink->time_base.den = ctx->inputs[0]->time_base.den * 2;
173 outlink->w = ctx->inputs[0]->w;
174 outlink->h = ctx->inputs[0]->h;
176 outlink->frame_rate = av_mul_q(ctx->inputs[0]->frame_rate,
182 static int query_formats(AVFilterContext *ctx)
184 static const enum AVPixelFormat pix_fmts[] = {
186 AV_PIX_FMT_GRAY9, AV_PIX_FMT_GRAY10, AV_PIX_FMT_GRAY12, AV_PIX_FMT_GRAY14, AV_PIX_FMT_GRAY16,
187 AV_PIX_FMT_YUV410P, AV_PIX_FMT_YUV411P,
188 AV_PIX_FMT_YUV420P, AV_PIX_FMT_YUV422P,
189 AV_PIX_FMT_YUV440P, AV_PIX_FMT_YUV444P,
190 AV_PIX_FMT_YUVJ444P, AV_PIX_FMT_YUVJ440P,
191 AV_PIX_FMT_YUVJ422P, AV_PIX_FMT_YUVJ420P,
193 AV_PIX_FMT_YUVA420P, AV_PIX_FMT_YUVA422P, AV_PIX_FMT_YUVA444P,
194 AV_PIX_FMT_GBRP, AV_PIX_FMT_GBRAP,
195 AV_PIX_FMT_YUV420P9, AV_PIX_FMT_YUV422P9, AV_PIX_FMT_YUV444P9,
196 AV_PIX_FMT_YUV420P10, AV_PIX_FMT_YUV422P10, AV_PIX_FMT_YUV444P10,
197 AV_PIX_FMT_YUV440P10,
198 AV_PIX_FMT_YUV420P12, AV_PIX_FMT_YUV422P12, AV_PIX_FMT_YUV444P12,
199 AV_PIX_FMT_YUV440P12,
200 AV_PIX_FMT_YUV420P14, AV_PIX_FMT_YUV422P14, AV_PIX_FMT_YUV444P14,
201 AV_PIX_FMT_YUV420P16, AV_PIX_FMT_YUV422P16, AV_PIX_FMT_YUV444P16,
202 AV_PIX_FMT_GBRP9, AV_PIX_FMT_GBRP10, AV_PIX_FMT_GBRP12, AV_PIX_FMT_GBRP14, AV_PIX_FMT_GBRP16,
203 AV_PIX_FMT_YUVA444P9, AV_PIX_FMT_YUVA444P10, AV_PIX_FMT_YUVA444P12, AV_PIX_FMT_YUVA444P16,
204 AV_PIX_FMT_YUVA422P9, AV_PIX_FMT_YUVA422P10, AV_PIX_FMT_YUVA422P12, AV_PIX_FMT_YUVA422P16,
205 AV_PIX_FMT_YUVA420P9, AV_PIX_FMT_YUVA420P10, AV_PIX_FMT_YUVA420P16,
206 AV_PIX_FMT_GBRAP10, AV_PIX_FMT_GBRAP12, AV_PIX_FMT_GBRAP16,
210 AVFilterFormats *fmts_list = ff_make_format_list(pix_fmts);
212 return AVERROR(ENOMEM);
213 return ff_set_common_formats(ctx, fmts_list);
216 static float dot_dsp(const NNEDIContext *const s, const float *kernel, const float *input,
217 int n, float scale, float bias)
221 sum = s->fdsp->scalarproduct_float(kernel, input, n);
223 y = sum * scale + bias + 1e-20f;
228 static float elliott(float x)
230 return x / (1.0f + fabsf(x));
233 static void transform_elliott(float *input, int size)
235 for (int i = 0; i < size; i++)
236 input[i] = elliott(input[i]);
239 static void process_old(AVFilterContext *ctx,
240 const void *src, ptrdiff_t src_stride,
241 uint8_t *prescreen, int N,
242 const PrescreenerCoefficients *const m_data)
244 NNEDIContext *s = ctx->priv;
245 const float *src_p = src;
247 // Adjust source pointer to point to top-left of filter window.
248 const float *window = src_p - 2 * src_stride - 5;
250 for (int j = 0; j < N; j++) {
251 LOCAL_ALIGNED_32(float, input, [48]);
254 for (int i = 0; i < 4; i++)
255 memcpy(input + i * 12, window + i * src_stride + j, 12 * sizeof(float));
258 for (int n = 0; n < 4; n++)
259 state[n] = dot_dsp(s, m_data->kernel_l0[n], input, 48, 1.0f, m_data->bias_l0[n]);
260 transform_elliott(state + 1, 3);
263 for (int n = 0; n < 4; n++)
264 state[n + 4] = dot_dsp(s, m_data->kernel_l1[n], state, 4, 1.0f, m_data->bias_l1[n]);
265 transform_elliott(state + 4, 3);
268 for (int n = 0; n < 4; n++)
269 state[n + 8] = dot_dsp(s, m_data->kernel_l2[n], state, 8, 1.0f, m_data->bias_l2[n]);
271 prescreen[j] = FFMAX(state[10], state[11]) <= FFMAX(state[8], state[9]) ? 255 : 0;
275 static void process_new(AVFilterContext *ctx,
276 const void *src, ptrdiff_t src_stride,
277 uint8_t *prescreen, int N,
278 const PrescreenerCoefficients *const m_data)
280 NNEDIContext *s = ctx->priv;
281 const float *src_p = src;
283 // Adjust source pointer to point to top-left of filter window.
284 const float *window = src_p - 2 * src_stride - 6;
286 for (int j = 0; j < N; j += 4) {
287 LOCAL_ALIGNED_32(float, input, [64]);
290 for (int i = 0; i < 4; i++)
291 memcpy(input + i * 16, window + i * src_stride + j, 16 * sizeof(float));
293 for (int n = 0; n < 4; n++)
294 state[n] = dot_dsp(s, m_data->kernel_l0[n], input, 64, 1.0f, m_data->bias_l0[n]);
295 transform_elliott(state, 4);
297 for (int n = 0; n < 4; n++)
298 state[n + 4] = dot_dsp(s, m_data->kernel_l1[n], state, 4, 1.0f, m_data->bias_l1[n]);
300 for (int n = 0; n < 4; n++)
301 prescreen[j + n] = state[n + 4] > 0.f;
305 static int filter_offset(int nn, const PredictorCoefficients *const model)
307 return nn * model->nsize;
310 static const float *softmax_q1_filter(int nn,
311 const PredictorCoefficients *const model)
313 return model->softmax_q1 + filter_offset(nn, model);
316 static const float *elliott_q1_filter(int nn,
317 const PredictorCoefficients *const model)
319 return model->elliott_q1 + filter_offset(nn, model);
322 static const float *softmax_q2_filter(int nn,
323 const PredictorCoefficients *const model)
325 return model->softmax_q2 + filter_offset(nn, model);
328 static const float *elliott_q2_filter(int nn,
329 const PredictorCoefficients *const model)
331 return model->elliott_q2 + filter_offset(nn, model);
334 static void gather_input(const float *src, ptrdiff_t src_stride,
335 float *buf, float mstd[4],
336 const PredictorCoefficients *const model)
338 const float scale = 1.f / model->nsize;
343 for (int i = 0; i < model->ydim; i++) {
344 memcpy(buf, src, model->xdim * sizeof(float));
346 for (int j = 0; j < model->xdim; j++) {
347 const float val = src[j];
357 mstd[0] = sum * scale;
360 tmp = sum_sq * scale - mstd[0] * mstd[0];
361 if (tmp < FLT_EPSILON) {
365 mstd[1] = sqrtf(tmp);
366 mstd[2] = 1.0f / mstd[1];
370 static float softmax_exp(float x)
372 return expf(av_clipf(x, -80.f, 80.f));
375 static void transform_softmax_exp(float *input, int size)
377 for (int i = 0; i < size; i++)
378 input[i] = softmax_exp(input[i]);
381 static void wae5(const float *softmax, const float *el,
382 int n, float mstd[4])
384 float vsum = 0.0f, wsum = 0.0f;
386 for (int i = 0; i < n; i++) {
387 vsum += softmax[i] * elliott(el[i]);
392 mstd[3] += (5.0f * vsum) / wsum * mstd[1] + mstd[0];
397 static void predictor(AVFilterContext *ctx,
398 const void *src, ptrdiff_t src_stride, void *dst,
399 const uint8_t *prescreen, int N,
400 const PredictorCoefficients *const model, int use_q2)
402 const NNEDIContext *const s = ctx->priv;
403 const float *src_p = src;
406 // Adjust source pointer to point to top-left of filter window.
407 const float *window = src_p - (model->ydim / 2) * src_stride - (model->xdim / 2 - 1);
408 const int filter_size = model->nsize;
409 const int nns = model->nns;
411 for (int i = 0; i < N; i++) {
412 LOCAL_ALIGNED_32(float, input, [48 * 6]);
413 float activation[256 * 2];
420 gather_input(window + i, src_stride, input, mstd, model);
423 for (int nn = 0; nn < nns; nn++)
424 activation[nn] = dot_dsp(s, softmax_q1_filter(nn, model), input, filter_size, scale, model->softmax_bias_q1[nn]);
426 for (int nn = 0; nn < nns; nn++)
427 activation[nns + nn] = dot_dsp(s, elliott_q1_filter(nn, model), input, filter_size, scale, model->elliott_bias_q1[nn]);
429 transform_softmax_exp(activation, nns);
430 wae5(activation, activation + nns, nns, mstd);
433 for (int nn = 0; nn < nns; nn++)
434 activation[nn] = dot_dsp(s, softmax_q2_filter(nn, model), input, filter_size, scale, model->softmax_bias_q2[nn]);
436 for (int nn = 0; nn < nns; nn++)
437 activation[nns + nn] = dot_dsp(s, elliott_q2_filter(nn, model), input, filter_size, scale, model->elliott_bias_q2[nn]);
439 transform_softmax_exp(activation, nns);
440 wae5(activation, activation + nns, nns, mstd);
443 dst_p[i] = mstd[3] * (use_q2 ? 0.5f : 1.f);
447 static void read_bytes(const uint8_t *src, float *dst,
448 int src_stride, int dst_stride,
449 int width, int height, float scale)
451 for (int y = 0; y < height; y++) {
452 for (int x = 0; x < 32; x++)
453 dst[-x - 1] = src[x];
455 for (int x = 0; x < width; x++)
458 for (int x = 0; x < 32; x++)
459 dst[width + x] = src[width - x - 1];
466 static void read_words(const uint8_t *srcp, float *dst,
467 int src_stride, int dst_stride,
468 int width, int height, float scale)
470 const uint16_t *src = (const uint16_t *)srcp;
474 for (int y = 0; y < height; y++) {
475 for (int x = 0; x < 32; x++)
476 dst[-x - 1] = src[x] * scale;
478 for (int x = 0; x < width; x++)
479 dst[x] = src[x] * scale;
481 for (int x = 0; x < 32; x++)
482 dst[width + x] = src[width - x - 1] * scale;
489 static void write_bytes(const float *src, uint8_t *dst,
490 int src_stride, int dst_stride,
491 int width, int height, int depth,
494 for (int y = 0; y < height; y++) {
495 for (int x = 0; x < width; x++)
496 dst[x] = av_clip_uint8(src[x]);
503 static void write_words(const float *src, uint8_t *dstp,
504 int src_stride, int dst_stride,
505 int width, int height, int depth,
508 uint16_t *dst = (uint16_t *)dstp;
512 for (int y = 0; y < height; y++) {
513 for (int x = 0; x < width; x++)
514 dst[x] = av_clip_uintp2_c(src[x] * scale, depth);
521 static void interpolation(const void *src, ptrdiff_t src_stride,
522 void *dst, const uint8_t *prescreen, int n)
524 const float *src_p = src;
526 const float *window = src_p - 2 * src_stride;
528 for (int i = 0; i < n; i++) {
534 accum += (-3.0f / 32.0f) * window[0 * src_stride + i];
535 accum += (19.0f / 32.0f) * window[1 * src_stride + i];
536 accum += (19.0f / 32.0f) * window[2 * src_stride + i];
537 accum += (-3.0f / 32.0f) * window[3 * src_stride + i];
543 static int filter_slice(AVFilterContext *ctx, void *arg, int jobnr, int nb_jobs)
545 const NNEDIContext *const s = ctx->priv;
547 AVFrame *in = s->prev;
548 const float in_scale = s->in_scale;
549 const float out_scale = s->out_scale;
550 const int depth = s->depth;
551 const int interlaced = in->interlaced_frame;
552 const int tff = s->field_n == (s->field < 0 ? interlaced ? in->top_field_first : 1 :
556 for (int p = 0; p < s->nb_planes; p++) {
557 const int height = s->planeheight[p];
558 const int width = s->planewidth[p];
559 const int slice_start = 2 * ((height / 2 * jobnr) / nb_jobs);
560 const int slice_end = 2 * ((height / 2 * (jobnr+1)) / nb_jobs);
561 const uint8_t *src_data = in->data[p];
562 uint8_t *dst_data = out->data[p];
563 uint8_t *dst = out->data[p] + slice_start * out->linesize[p];
564 const int src_linesize = in->linesize[p];
565 const int dst_linesize = out->linesize[p];
566 uint8_t *prescreen_buf = s->prescreen_buf[jobnr];
567 float *srcbuf = s->input_buf[jobnr];
568 const int srcbuf_stride = width + 64;
569 float *dstbuf = s->output_buf[jobnr];
570 const int dstbuf_stride = width;
571 const int slice_height = (slice_end - slice_start) / 2;
572 const int last_slice = slice_end == height;
573 const uint8_t *in_line;
577 if (!(s->process_plane & (1 << p))) {
578 av_image_copy_plane(dst, out->linesize[p],
579 in->data[p] + slice_start * in->linesize[p],
581 s->linesize[p], slice_end - slice_start);
585 y_out = slice_start + (tff ^ (slice_start & 1));
586 in_line = src_data + (y_out * src_linesize);
587 out_line = dst_data + (y_out * dst_linesize);
589 while (y_out < slice_end) {
590 memcpy(out_line, in_line, s->linesize[p]);
592 in_line += src_linesize * 2;
593 out_line += dst_linesize * 2;
596 y_out = slice_start + ((!tff) ^ (slice_start & 1));
598 s->read(src_data + FFMAX(y_out - 5, tff) * src_linesize,
600 src_linesize * 2, srcbuf_stride,
602 srcbuf += srcbuf_stride;
604 s->read(src_data + FFMAX(y_out - 3, tff) * src_linesize,
606 src_linesize * 2, srcbuf_stride,
608 srcbuf += srcbuf_stride;
610 s->read(src_data + FFMAX(y_out - 1, tff) * src_linesize,
612 src_linesize * 2, srcbuf_stride,
614 srcbuf += srcbuf_stride;
616 in_line = src_data + FFMIN(y_out + 1, height - 1 - !tff) * src_linesize;
617 out_line = dst_data + (y_out * dst_linesize);
619 s->read(in_line, srcbuf + 32, src_linesize * 2, srcbuf_stride,
620 width, slice_height - last_slice, in_scale);
622 y_out += (slice_height - last_slice) * 2;
624 s->read(src_data + FFMIN(y_out + 1, height - 1 - !tff) * src_linesize,
625 srcbuf + 32 + srcbuf_stride * (slice_height - last_slice),
626 src_linesize * 2, srcbuf_stride,
629 s->read(src_data + FFMIN(y_out + 3, height - 1 - !tff) * src_linesize,
630 srcbuf + 32 + srcbuf_stride * (slice_height + 1 - last_slice),
631 src_linesize * 2, srcbuf_stride,
634 s->read(src_data + FFMIN(y_out + 5, height - 1 - !tff) * src_linesize,
635 srcbuf + 32 + srcbuf_stride * (slice_height + 2 - last_slice),
636 src_linesize * 2, srcbuf_stride,
639 for (int y = 0; y < slice_end - slice_start; y += 2) {
641 s->prescreen[s->pscrn > 1](ctx, srcbuf + (y / 2) * srcbuf_stride + 32,
642 srcbuf_stride, prescreen_buf, width,
643 &s->prescreener[s->pscrn - 1]);
646 srcbuf + (y / 2) * srcbuf_stride + 32,
648 dstbuf + (y / 2) * dstbuf_stride,
649 prescreen_buf, width,
650 &s->coeffs[s->etype][s->nnsparam][s->nsize], s->qual == 2);
653 interpolation(srcbuf + (y / 2) * srcbuf_stride + 32,
655 dstbuf + (y / 2) * dstbuf_stride,
656 prescreen_buf, width);
659 s->write(dstbuf, out_line, dstbuf_stride, dst_linesize * 2,
660 width, slice_height, depth, out_scale);
666 static int get_frame(AVFilterContext *ctx, int is_second)
668 NNEDIContext *s = ctx->priv;
669 AVFilterLink *outlink = ctx->outputs[0];
672 dst = ff_get_video_buffer(outlink, outlink->w, outlink->h);
674 return AVERROR(ENOMEM);
675 av_frame_copy_props(dst, s->prev);
676 dst->interlaced_frame = 0;
679 ctx->internal->execute(ctx, filter_slice, dst, NULL, FFMIN(s->planeheight[1] / 2, s->nb_threads));
681 if (s->field == -2 || s->field > 1)
682 s->field_n = !s->field_n;
684 return ff_filter_frame(outlink, dst);
687 static int filter_frame(AVFilterLink *inlink, AVFrame *in)
689 AVFilterContext *ctx = inlink->dst;
690 NNEDIContext *s = ctx->priv;
698 if ((s->deint && !in->interlaced_frame) || ctx->is_disabled) {
700 ret = ff_filter_frame(ctx->outputs[0], s->prev);
705 s->pts = s->prev->pts * 2;
706 ret = get_frame(ctx, 0);
707 if (ret < 0 || (s->field > -2 && s->field < 2)) {
708 av_frame_free(&s->prev);
713 s->pts = s->prev->pts + in->pts;
714 ret = get_frame(ctx, 1);
715 av_frame_free(&s->prev);
720 static int request_frame(AVFilterLink *link)
722 AVFilterContext *ctx = link->src;
723 NNEDIContext *s = ctx->priv;
729 ret = ff_request_frame(ctx->inputs[0]);
731 if (ret == AVERROR_EOF && s->prev) {
732 AVFrame *next = av_frame_clone(s->prev);
735 return AVERROR(ENOMEM);
737 next->pts = s->prev->pts + av_rescale_q(1, av_inv_q(ctx->outputs[0]->frame_rate),
738 ctx->outputs[0]->time_base);
741 ret = filter_frame(ctx->inputs[0], next);
742 } else if (ret < 0) {
749 static void copy_weights(float *dst, int n, const float **data)
751 memcpy(dst, *data, n * sizeof(float));
755 static float *allocate(float **ptr, int size)
764 static int allocate_model(PredictorCoefficients *coeffs, int xdim, int ydim, int nns)
766 int filter_size = nns * xdim * ydim;
770 data = av_calloc(filter_size + bias_size, 4 * sizeof(float));
772 return AVERROR(ENOMEM);
777 coeffs->nsize = xdim * ydim;
780 coeffs->softmax_q1 = allocate(&data, filter_size);
781 coeffs->elliott_q1 = allocate(&data, filter_size);
782 coeffs->softmax_bias_q1 = allocate(&data, bias_size);
783 coeffs->elliott_bias_q1 = allocate(&data, bias_size);
785 coeffs->softmax_q2 = allocate(&data, filter_size);
786 coeffs->elliott_q2 = allocate(&data, filter_size);
787 coeffs->softmax_bias_q2 = allocate(&data, bias_size);
788 coeffs->elliott_bias_q2 = allocate(&data, bias_size);
793 static int read_weights(AVFilterContext *ctx, const float *bdata)
795 NNEDIContext *s = ctx->priv;
798 copy_weights(&s->prescreener[0].kernel_l0[0][0], 4 * 48, &bdata);
799 copy_weights(s->prescreener[0].bias_l0, 4, &bdata);
801 copy_weights(&s->prescreener[0].kernel_l1[0][0], 4 * 4, &bdata);
802 copy_weights(s->prescreener[0].bias_l1, 4, &bdata);
804 copy_weights(&s->prescreener[0].kernel_l2[0][0], 4 * 8, &bdata);
805 copy_weights(s->prescreener[0].bias_l2, 4, &bdata);
807 for (int i = 0; i < 3; i++) {
808 PrescreenerCoefficients *data = &s->prescreener[i + 1];
809 float kernel_l0_shuffled[4 * 64];
810 float kernel_l1_shuffled[4 * 4];
812 copy_weights(kernel_l0_shuffled, 4 * 64, &bdata);
813 copy_weights(data->bias_l0, 4, &bdata);
815 copy_weights(kernel_l1_shuffled, 4 * 4, &bdata);
816 copy_weights(data->bias_l1, 4, &bdata);
818 for (int n = 0; n < 4; n++) {
819 for (int k = 0; k < 64; k++)
820 data->kernel_l0[n][k] = kernel_l0_shuffled[(k / 8) * 32 + n * 8 + k % 8];
821 for (int k = 0; k < 4; k++)
822 data->kernel_l1[n][k] = kernel_l1_shuffled[k * 4 + n];
826 for (int m = 0; m < 2; m++) {
827 // Grouping by neuron count.
828 for (int i = 0; i < 5; i++) {
829 const int nns = NNEDI_NNS[i];
831 // Grouping by window size.
832 for (int j = 0; j < 7; j++) {
833 PredictorCoefficients *model = &s->coeffs[m][i][j];
834 const int xdim = NNEDI_XDIM[j];
835 const int ydim = NNEDI_YDIM[j];
836 const int filter_size = xdim * ydim;
838 ret = allocate_model(model, xdim, ydim, nns);
842 // Quality 1 model. NNS[i] * (XDIM[j] * YDIM[j]) * 2 coefficients.
843 copy_weights(model->softmax_q1, nns * filter_size, &bdata);
844 copy_weights(model->elliott_q1, nns * filter_size, &bdata);
846 // Quality 1 model bias. NNS[i] * 2 coefficients.
847 copy_weights(model->softmax_bias_q1, nns, &bdata);
848 copy_weights(model->elliott_bias_q1, nns, &bdata);
850 // Quality 2 model. NNS[i] * (XDIM[j] * YDIM[j]) * 2 coefficients.
851 copy_weights(model->softmax_q2, nns * filter_size, &bdata);
852 copy_weights(model->elliott_q2, nns * filter_size, &bdata);
854 // Quality 2 model bias. NNS[i] * 2 coefficients.
855 copy_weights(model->softmax_bias_q2, nns, &bdata);
856 copy_weights(model->elliott_bias_q2, nns, &bdata);
864 static float mean(const float *input, int size)
868 for (int i = 0; i < size; i++)
874 static void transform(float *input, int size, float mean, float half)
876 for (int i = 0; i < size; i++)
877 input[i] = (input[i] - mean) / half;
880 static void subtract_mean_old(PrescreenerCoefficients *coeffs, float half)
882 for (int n = 0; n < 4; n++) {
883 float m = mean(coeffs->kernel_l0[n], 48);
885 transform(coeffs->kernel_l0[n], 48, m, half);
889 static void subtract_mean_new(PrescreenerCoefficients *coeffs, float half)
891 for (int n = 0; n < 4; n++) {
892 float m = mean(coeffs->kernel_l0[n], 64);
894 transform(coeffs->kernel_l0[n], 64, m, half);
898 static void subtract_mean_predictor(PredictorCoefficients *model)
900 const int filter_size = model->nsize;
901 const int nns = model->nns;
902 const float scale = 1.f / nns;
904 double softmax_means[256]; // Average of individual softmax filters.
905 double elliott_means[256]; // Average of individual elliott filters.
906 double mean_filter[48 * 6] = { 0 }; // Pointwise average of all softmax filters.
910 for (int nn = 0; nn < nns; nn++) {
911 softmax_means[nn] = mean(model->softmax_q1 + nn * filter_size, filter_size);
912 elliott_means[nn] = mean(model->elliott_q1 + nn * filter_size, filter_size);
914 for (int k = 0; k < filter_size; k++)
915 mean_filter[k] += model->softmax_q1[nn * filter_size + k] - softmax_means[nn];
918 for (int k = 0; k < filter_size; k++)
919 mean_filter[k] *= scale;
921 mean_bias = mean(model->softmax_bias_q1, nns);
923 for (int nn = 0; nn < nns; nn++) {
924 for (int k = 0; k < filter_size; k++) {
925 model->softmax_q1[nn * filter_size + k] -= softmax_means[nn] + mean_filter[k];
926 model->elliott_q1[nn * filter_size + k] -= elliott_means[nn];
928 model->softmax_bias_q1[nn] -= mean_bias;
932 memset(mean_filter, 0, sizeof(mean_filter));
934 for (int nn = 0; nn < nns; nn++) {
935 softmax_means[nn] = mean(model->softmax_q2 + nn * filter_size, filter_size);
936 elliott_means[nn] = mean(model->elliott_q2 + nn * filter_size, filter_size);
938 for (int k = 0; k < filter_size; k++) {
939 mean_filter[k] += model->softmax_q2[nn * filter_size + k] - softmax_means[nn];
943 for (int k = 0; k < filter_size; k++)
944 mean_filter[k] *= scale;
946 mean_bias = mean(model->softmax_bias_q2, nns);
948 for (int nn = 0; nn < nns; nn++) {
949 for (int k = 0; k < filter_size; k++) {
950 model->softmax_q2[nn * filter_size + k] -= softmax_means[nn] + mean_filter[k];
951 model->elliott_q2[nn * filter_size + k] -= elliott_means[nn];
954 model->softmax_bias_q2[nn] -= mean_bias;
958 static av_cold int init(AVFilterContext *ctx)
960 NNEDIContext *s = ctx->priv;
961 FILE *weights_file = NULL;
962 int64_t weights_size;
967 weights_file = av_fopen_utf8(s->weights_file, "rb");
969 av_log(ctx, AV_LOG_ERROR, "No weights file provided, aborting!\n");
970 return AVERROR(EINVAL);
973 if (fseek(weights_file, 0, SEEK_END)) {
974 av_log(ctx, AV_LOG_ERROR, "Couldn't seek to the end of weights file.\n");
975 fclose(weights_file);
976 return AVERROR(EINVAL);
979 weights_size = ftell(weights_file);
981 if (weights_size == -1) {
982 fclose(weights_file);
983 av_log(ctx, AV_LOG_ERROR, "Couldn't get size of weights file.\n");
984 return AVERROR(EINVAL);
985 } else if (weights_size != NNEDI_WEIGHTS_SIZE) {
986 fclose(weights_file);
987 av_log(ctx, AV_LOG_ERROR, "Unexpected weights file size.\n");
988 return AVERROR(EINVAL);
991 if (fseek(weights_file, 0, SEEK_SET)) {
992 fclose(weights_file);
993 av_log(ctx, AV_LOG_ERROR, "Couldn't seek to the start of weights file.\n");
994 return AVERROR(EINVAL);
997 bdata = av_malloc(NNEDI_WEIGHTS_SIZE);
999 fclose(weights_file);
1000 return AVERROR(ENOMEM);
1003 bytes_read = fread(bdata, 1, NNEDI_WEIGHTS_SIZE, weights_file);
1004 if (bytes_read != NNEDI_WEIGHTS_SIZE) {
1005 fclose(weights_file);
1006 ret = AVERROR_INVALIDDATA;
1007 av_log(ctx, AV_LOG_ERROR, "Couldn't read weights file.\n");
1011 fclose(weights_file);
1013 s->fdsp = avpriv_float_dsp_alloc(0);
1015 ret = AVERROR(ENOMEM);
1019 ret = read_weights(ctx, bdata);
1028 static int config_input(AVFilterLink *inlink)
1030 AVFilterContext *ctx = inlink->dst;
1031 NNEDIContext *s = ctx->priv;
1032 const AVPixFmtDescriptor *desc = av_pix_fmt_desc_get(inlink->format);
1035 s->depth = desc->comp[0].depth;
1036 s->nb_threads = ff_filter_get_nb_threads(ctx);
1037 s->nb_planes = av_pix_fmt_count_planes(inlink->format);
1038 if ((ret = av_image_fill_linesizes(s->linesize, inlink->format, inlink->w)) < 0)
1041 s->planewidth[1] = s->planewidth[2] = AV_CEIL_RSHIFT(inlink->w, desc->log2_chroma_w);
1042 s->planewidth[0] = s->planewidth[3] = inlink->w;
1043 s->planeheight[1] = s->planeheight[2] = AV_CEIL_RSHIFT(inlink->h, desc->log2_chroma_h);
1044 s->planeheight[0] = s->planeheight[3] = inlink->h;
1046 s->half = ((1 << 8) - 1) / 2.f;
1047 s->out_scale = 1 << (s->depth - 8);
1048 s->in_scale = 1.f / s->out_scale;
1052 s->read = read_bytes;
1053 s->write = write_bytes;
1056 s->read = read_words;
1057 s->write = write_words;
1061 subtract_mean_old(&s->prescreener[0], s->half);
1062 subtract_mean_new(&s->prescreener[1], s->half);
1063 subtract_mean_new(&s->prescreener[2], s->half);
1064 subtract_mean_new(&s->prescreener[3], s->half);
1066 s->prescreen[0] = process_old;
1067 s->prescreen[1] = process_new;
1069 for (int i = 0; i < 2; i++) {
1070 for (int j = 0; j < 5; j++) {
1071 for (int k = 0; k < 7; k++)
1072 subtract_mean_predictor(&s->coeffs[i][j][k]);
1076 s->input_size = (s->planewidth[0] + 64) * (s->planeheight[0] + 6);
1077 s->input_buf = av_calloc(s->nb_threads, sizeof(*s->input_buf));
1079 return AVERROR(ENOMEM);
1081 for (int i = 0; i < s->nb_threads; i++) {
1082 s->input_buf[i] = av_calloc(s->input_size, sizeof(**s->input_buf));
1083 if (!s->input_buf[i])
1084 return AVERROR(ENOMEM);
1087 s->output_buf = av_calloc(s->nb_threads, sizeof(*s->output_buf));
1089 return AVERROR(ENOMEM);
1091 for (int i = 0; i < s->nb_threads; i++) {
1092 s->output_buf[i] = av_calloc(s->input_size, sizeof(**s->output_buf));
1093 if (!s->output_buf[i])
1094 return AVERROR(ENOMEM);
1097 s->prescreen_buf = av_calloc(s->nb_threads, sizeof(*s->prescreen_buf));
1098 if (!s->prescreen_buf)
1099 return AVERROR(ENOMEM);
1101 for (int i = 0; i < s->nb_threads; i++) {
1102 s->prescreen_buf[i] = av_calloc(s->planewidth[0], sizeof(**s->prescreen_buf));
1103 if (!s->prescreen_buf[i])
1104 return AVERROR(ENOMEM);
1110 static av_cold void uninit(AVFilterContext *ctx)
1112 NNEDIContext *s = ctx->priv;
1114 for (int i = 0; i < s->nb_threads && s->prescreen_buf; i++)
1115 av_freep(&s->prescreen_buf[i]);
1117 av_freep(&s->prescreen_buf);
1119 for (int i = 0; i < s->nb_threads && s->input_buf; i++)
1120 av_freep(&s->input_buf[i]);
1122 av_freep(&s->input_buf);
1124 for (int i = 0; i < s->nb_threads && s->output_buf; i++)
1125 av_freep(&s->output_buf[i]);
1127 av_freep(&s->output_buf);
1130 for (int i = 0; i < 2; i++) {
1131 for (int j = 0; j < 5; j++) {
1132 for (int k = 0; k < 7; k++) {
1133 av_freep(&s->coeffs[i][j][k].data);
1138 av_frame_free(&s->prev);
1141 static const AVFilterPad inputs[] = {
1144 .type = AVMEDIA_TYPE_VIDEO,
1145 .filter_frame = filter_frame,
1146 .config_props = config_input,
1151 static const AVFilterPad outputs[] = {
1154 .type = AVMEDIA_TYPE_VIDEO,
1155 .config_props = config_output,
1156 .request_frame = request_frame,
1161 AVFilter ff_vf_nnedi = {
1163 .description = NULL_IF_CONFIG_SMALL("Apply neural network edge directed interpolation intra-only deinterlacer."),
1164 .priv_size = sizeof(NNEDIContext),
1165 .priv_class = &nnedi_class,
1168 .query_formats = query_formats,
1171 .flags = AVFILTER_FLAG_SUPPORT_TIMELINE_INTERNAL | AVFILTER_FLAG_SLICE_THREADS,
1172 .process_command = ff_filter_process_command,