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/common.h"
25 #include "libavutil/float_dsp.h"
26 #include "libavutil/imgutils.h"
27 #include "libavutil/mem_internal.h"
28 #include "libavutil/opt.h"
29 #include "libavutil/pixdesc.h"
35 static const size_t NNEDI_WEIGHTS_SIZE = 13574928;
36 static const uint8_t NNEDI_XDIM[] = { 8, 16, 32, 48, 8, 16, 32 };
37 static const uint8_t NNEDI_YDIM[] = { 6, 6, 6, 6, 4, 4, 4 };
38 static const uint16_t NNEDI_NNS[] = { 16, 32, 64, 128, 256 };
40 typedef struct PrescreenerOldCoefficients {
41 DECLARE_ALIGNED(32, float, kernel_l0)[4][14 * 4];
42 DECLARE_ALIGNED(32, float, bias_l0)[4];
44 DECLARE_ALIGNED(32, float, kernel_l1)[4][4];
45 DECLARE_ALIGNED(32, float, bias_l1)[4];
47 DECLARE_ALIGNED(32, float, kernel_l2)[4][8];
48 DECLARE_ALIGNED(32, float, bias_l2)[4];
49 } PrescreenerOldCoefficients;
51 typedef struct PrescreenerNewCoefficients {
52 DECLARE_ALIGNED(32, float, kernel_l0)[4][16 * 4];
53 DECLARE_ALIGNED(32, float, bias_l0)[4];
55 DECLARE_ALIGNED(32, float, kernel_l1)[4][4];
56 DECLARE_ALIGNED(32, float, bias_l1)[4];
57 } PrescreenerNewCoefficients;
59 typedef struct PredictorCoefficients {
60 int xdim, ydim, nns, nsize;
64 float *softmax_bias_q1;
65 float *elliott_bias_q1;
68 float *softmax_bias_q2;
69 float *elliott_bias_q2;
70 } PredictorCoefficients;
72 typedef struct NNEDIContext {
83 AVFloatDSPContext *fdsp;
92 PrescreenerOldCoefficients prescreener_old;
93 PrescreenerNewCoefficients prescreener_new[3];
94 PredictorCoefficients coeffs[2][5][7];
111 uint8_t *prescreen_buf;
115 void (*read)(const uint8_t *src, float *dst,
116 int src_stride, int dst_stride,
117 int width, int height, float scale);
118 void (*write)(const float *src, uint8_t *dst,
119 int src_stride, int dst_stride,
120 int width, int height, int depth, float scale);
121 void (*prescreen[2])(AVFilterContext *ctx,
122 const void *src, ptrdiff_t src_stride,
123 uint8_t *prescreen, int N, void *data);
126 #define OFFSET(x) offsetof(NNEDIContext, x)
127 #define RFLAGS AV_OPT_FLAG_VIDEO_PARAM|AV_OPT_FLAG_FILTERING_PARAM|AV_OPT_FLAG_RUNTIME_PARAM
128 #define FLAGS AV_OPT_FLAG_VIDEO_PARAM|AV_OPT_FLAG_FILTERING_PARAM
130 static const AVOption nnedi_options[] = {
131 {"weights", "set weights file", OFFSET(weights_file), AV_OPT_TYPE_STRING, {.str="nnedi3_weights.bin"}, 0, 0, FLAGS },
132 {"deint", "set which frames to deinterlace", OFFSET(deint), AV_OPT_TYPE_INT, {.i64=0}, 0, 1, RFLAGS, "deint" },
133 {"all", "deinterlace all frames", 0, AV_OPT_TYPE_CONST, {.i64=0}, 0, 0, RFLAGS, "deint" },
134 {"interlaced", "only deinterlace frames marked as interlaced", 0, AV_OPT_TYPE_CONST, {.i64=1}, 0, 0, RFLAGS, "deint" },
135 {"field", "set mode of operation", OFFSET(field), AV_OPT_TYPE_INT, {.i64=-1}, -2, 3, RFLAGS, "field" },
136 {"af", "use frame flags, both fields", 0, AV_OPT_TYPE_CONST, {.i64=-2}, 0, 0, RFLAGS, "field" },
137 {"a", "use frame flags, single field", 0, AV_OPT_TYPE_CONST, {.i64=-1}, 0, 0, RFLAGS, "field" },
138 {"t", "use top field only", 0, AV_OPT_TYPE_CONST, {.i64=0}, 0, 0, RFLAGS, "field" },
139 {"b", "use bottom field only", 0, AV_OPT_TYPE_CONST, {.i64=1}, 0, 0, RFLAGS, "field" },
140 {"tf", "use both fields, top first", 0, AV_OPT_TYPE_CONST, {.i64=2}, 0, 0, RFLAGS, "field" },
141 {"bf", "use both fields, bottom first", 0, AV_OPT_TYPE_CONST, {.i64=3}, 0, 0, RFLAGS, "field" },
142 {"planes", "set which planes to process", OFFSET(process_plane), AV_OPT_TYPE_INT, {.i64=7}, 0, 15, RFLAGS },
143 {"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" },
144 {"s8x6", NULL, 0, AV_OPT_TYPE_CONST, {.i64=0}, 0, 0, RFLAGS, "nsize" },
145 {"s16x6", NULL, 0, AV_OPT_TYPE_CONST, {.i64=1}, 0, 0, RFLAGS, "nsize" },
146 {"s32x6", NULL, 0, AV_OPT_TYPE_CONST, {.i64=2}, 0, 0, RFLAGS, "nsize" },
147 {"s48x6", NULL, 0, AV_OPT_TYPE_CONST, {.i64=3}, 0, 0, RFLAGS, "nsize" },
148 {"s8x4", NULL, 0, AV_OPT_TYPE_CONST, {.i64=4}, 0, 0, RFLAGS, "nsize" },
149 {"s16x4", NULL, 0, AV_OPT_TYPE_CONST, {.i64=5}, 0, 0, RFLAGS, "nsize" },
150 {"s32x4", NULL, 0, AV_OPT_TYPE_CONST, {.i64=6}, 0, 0, RFLAGS, "nsize" },
151 {"nns", "set number of neurons in predictor neural network", OFFSET(nnsparam), AV_OPT_TYPE_INT, {.i64=1}, 0, 4, RFLAGS, "nns" },
152 {"n16", NULL, 0, AV_OPT_TYPE_CONST, {.i64=0}, 0, 0, RFLAGS, "nns" },
153 {"n32", NULL, 0, AV_OPT_TYPE_CONST, {.i64=1}, 0, 0, RFLAGS, "nns" },
154 {"n64", NULL, 0, AV_OPT_TYPE_CONST, {.i64=2}, 0, 0, RFLAGS, "nns" },
155 {"n128", NULL, 0, AV_OPT_TYPE_CONST, {.i64=3}, 0, 0, RFLAGS, "nns" },
156 {"n256", NULL, 0, AV_OPT_TYPE_CONST, {.i64=4}, 0, 0, RFLAGS, "nns" },
157 {"qual", "set quality", OFFSET(qual), AV_OPT_TYPE_INT, {.i64=1}, 1, 2, RFLAGS, "qual" },
158 {"fast", NULL, 0, AV_OPT_TYPE_CONST, {.i64=1}, 0, 0, RFLAGS, "qual" },
159 {"slow", NULL, 0, AV_OPT_TYPE_CONST, {.i64=2}, 0, 0, RFLAGS, "qual" },
160 {"etype", "set which set of weights to use in the predictor", OFFSET(etype), AV_OPT_TYPE_INT, {.i64=0}, 0, 1, RFLAGS, "etype" },
161 {"a", "weights trained to minimize absolute error", 0, AV_OPT_TYPE_CONST, {.i64=0}, 0, 0, RFLAGS, "etype" },
162 {"abs","weights trained to minimize absolute error", 0, AV_OPT_TYPE_CONST, {.i64=0}, 0, 0, RFLAGS, "etype" },
163 {"s", "weights trained to minimize squared error", 0, AV_OPT_TYPE_CONST, {.i64=1}, 0, 0, RFLAGS, "etype" },
164 {"mse","weights trained to minimize squared error", 0, AV_OPT_TYPE_CONST, {.i64=1}, 0, 0, RFLAGS, "etype" },
165 {"pscrn", "set prescreening", OFFSET(pscrn), AV_OPT_TYPE_INT, {.i64=2}, 0, 4, RFLAGS, "pscrn" },
166 {"none", NULL, 0, AV_OPT_TYPE_CONST, {.i64=0}, 0, 0, RFLAGS, "pscrn" },
167 {"original", NULL, 0, AV_OPT_TYPE_CONST, {.i64=1}, 0, 0, RFLAGS, "pscrn" },
168 {"new", NULL, 0, AV_OPT_TYPE_CONST, {.i64=2}, 0, 0, RFLAGS, "pscrn" },
169 {"new2", NULL, 0, AV_OPT_TYPE_CONST, {.i64=3}, 0, 0, RFLAGS, "pscrn" },
170 {"new3", NULL, 0, AV_OPT_TYPE_CONST, {.i64=4}, 0, 0, RFLAGS, "pscrn" },
174 AVFILTER_DEFINE_CLASS(nnedi);
176 static int config_output(AVFilterLink *outlink)
178 AVFilterContext *ctx = outlink->src;
180 outlink->time_base.num = ctx->inputs[0]->time_base.num;
181 outlink->time_base.den = ctx->inputs[0]->time_base.den * 2;
182 outlink->w = ctx->inputs[0]->w;
183 outlink->h = ctx->inputs[0]->h;
185 outlink->frame_rate = av_mul_q(ctx->inputs[0]->frame_rate,
191 static int query_formats(AVFilterContext *ctx)
193 static const enum AVPixelFormat pix_fmts[] = {
195 AV_PIX_FMT_GRAY9, AV_PIX_FMT_GRAY10, AV_PIX_FMT_GRAY12, AV_PIX_FMT_GRAY14, AV_PIX_FMT_GRAY16,
196 AV_PIX_FMT_YUV410P, AV_PIX_FMT_YUV411P,
197 AV_PIX_FMT_YUV420P, AV_PIX_FMT_YUV422P,
198 AV_PIX_FMT_YUV440P, AV_PIX_FMT_YUV444P,
199 AV_PIX_FMT_YUVJ444P, AV_PIX_FMT_YUVJ440P,
200 AV_PIX_FMT_YUVJ422P, AV_PIX_FMT_YUVJ420P,
202 AV_PIX_FMT_YUVA420P, AV_PIX_FMT_YUVA422P, AV_PIX_FMT_YUVA444P,
203 AV_PIX_FMT_GBRP, AV_PIX_FMT_GBRAP,
204 AV_PIX_FMT_YUV420P9, AV_PIX_FMT_YUV422P9, AV_PIX_FMT_YUV444P9,
205 AV_PIX_FMT_YUV420P10, AV_PIX_FMT_YUV422P10, AV_PIX_FMT_YUV444P10,
206 AV_PIX_FMT_YUV440P10,
207 AV_PIX_FMT_YUV420P12, AV_PIX_FMT_YUV422P12, AV_PIX_FMT_YUV444P12,
208 AV_PIX_FMT_YUV440P12,
209 AV_PIX_FMT_YUV420P14, AV_PIX_FMT_YUV422P14, AV_PIX_FMT_YUV444P14,
210 AV_PIX_FMT_YUV420P16, AV_PIX_FMT_YUV422P16, AV_PIX_FMT_YUV444P16,
211 AV_PIX_FMT_GBRP9, AV_PIX_FMT_GBRP10, AV_PIX_FMT_GBRP12, AV_PIX_FMT_GBRP14, AV_PIX_FMT_GBRP16,
212 AV_PIX_FMT_YUVA444P9, AV_PIX_FMT_YUVA444P10, AV_PIX_FMT_YUVA444P12, AV_PIX_FMT_YUVA444P16,
213 AV_PIX_FMT_YUVA422P9, AV_PIX_FMT_YUVA422P10, AV_PIX_FMT_YUVA422P12, AV_PIX_FMT_YUVA422P16,
214 AV_PIX_FMT_YUVA420P9, AV_PIX_FMT_YUVA420P10, AV_PIX_FMT_YUVA420P16,
215 AV_PIX_FMT_GBRAP10, AV_PIX_FMT_GBRAP12, AV_PIX_FMT_GBRAP16,
219 AVFilterFormats *fmts_list = ff_make_format_list(pix_fmts);
221 return AVERROR(ENOMEM);
222 return ff_set_common_formats(ctx, fmts_list);
225 static float dot_dsp(NNEDIContext *s, const float *kernel, const float *input,
226 int n, float scale, float bias)
230 sum = s->fdsp->scalarproduct_float(kernel, input, n);
232 return sum * scale + bias;
235 static float elliott(float x)
237 return x / (1.0f + fabsf(x));
240 static void transform_elliott(float *input, int size)
242 for (int i = 0; i < size; i++)
243 input[i] = elliott(input[i]);
246 static void process_old(AVFilterContext *ctx,
247 const void *src, ptrdiff_t src_stride,
248 uint8_t *prescreen, int N,
251 NNEDIContext *s = ctx->priv;
252 const PrescreenerOldCoefficients *const m_data = data;
253 const float *src_p = src;
255 // Adjust source pointer to point to top-left of filter window.
256 const float *window = src_p - 2 * src_stride - 5;
258 for (int j = 0; j < N; j++) {
259 LOCAL_ALIGNED_32(float, input, [48]);
262 for (int i = 0; i < 4; i++)
263 memcpy(input + i * 12, window + i * src_stride + j, 12 * sizeof(float));
266 for (int n = 0; n < 4; n++)
267 state[n] = dot_dsp(s, m_data->kernel_l0[n], input, 48, 1.0f, m_data->bias_l0[n]);
268 transform_elliott(state + 1, 3);
271 for (int n = 0; n < 4; n++)
272 state[n + 4] = dot_dsp(s, m_data->kernel_l1[n], state, 4, 1.0f, m_data->bias_l1[n]);
273 transform_elliott(state + 4, 3);
276 for (int n = 0; n < 4; n++)
277 state[n + 8] = dot_dsp(s, m_data->kernel_l2[n], state, 8, 1.0f, m_data->bias_l2[n]);
279 prescreen[j] = FFMAX(state[10], state[11]) <= FFMAX(state[8], state[9]) ? 255 : 0;
283 static void process_new(AVFilterContext *ctx,
284 const void *src, ptrdiff_t src_stride,
285 uint8_t *prescreen, int N,
288 NNEDIContext *s = ctx->priv;
289 const PrescreenerNewCoefficients *const m_data = data;
290 const float *src_p = src;
292 // Adjust source pointer to point to top-left of filter window.
293 const float *window = src_p - 2 * src_stride - 6;
295 for (int j = 0; j < N; j += 4) {
296 LOCAL_ALIGNED_32(float, input, [64]);
299 for (int i = 0; i < 4; i++)
300 memcpy(input + i * 16, window + i * src_stride + j, 16 * sizeof(float));
302 for (int n = 0; n < 4; n++)
303 state[n] = dot_dsp(s, m_data->kernel_l0[n], input, 64, 1.0f, m_data->bias_l0[n]);
304 transform_elliott(state, 4);
306 for (int n = 0; n < 4; n++)
307 state[n + 4] = dot_dsp(s, m_data->kernel_l1[n], state, 4, 1.0f, m_data->bias_l1[n]);
309 for (int n = 0; n < 4; n++)
310 prescreen[j + n] = state[n + 4] > 0.f;
314 static int filter_offset(int nn, const PredictorCoefficients *const model)
316 return nn * model->nsize;
319 static const float *softmax_q1_filter(int nn,
320 const PredictorCoefficients *const model)
322 return model->softmax_q1 + filter_offset(nn, model);
325 static const float *elliott_q1_filter(int nn,
326 const PredictorCoefficients *const model)
328 return model->elliott_q1 + filter_offset(nn, model);
331 static const float *softmax_q2_filter(int nn,
332 const PredictorCoefficients *const model)
334 return model->softmax_q2 + filter_offset(nn, model);
337 static const float *elliott_q2_filter(int nn,
338 const PredictorCoefficients *const model)
340 return model->elliott_q2 + filter_offset(nn, model);
343 static void gather_input(const float *src, ptrdiff_t src_stride,
344 float *buf, float mstd[4],
345 const PredictorCoefficients *const model)
351 for (int i = 0; i < model->ydim; i++) {
352 memcpy(buf, src, model->xdim * sizeof(float));
354 for (int j = 0; j < model->xdim; j++) {
355 const float val = src[j];
365 mstd[0] = sum / model->nsize;
368 tmp = sum_sq / model->nsize - mstd[0] * mstd[0];
369 if (tmp < FLT_EPSILON) {
373 mstd[1] = sqrtf(tmp);
374 mstd[2] = 1.0f / mstd[1];
378 static float softmax_exp(float x)
380 return expf(av_clipf(x, -80.f, 80.f));
383 static void transform_softmax_exp(float *input, int size)
385 for (int i = 0; i < size; i++)
386 input[i] = softmax_exp(input[i]);
389 static void wae5(const float *softmax, const float *el,
390 int n, float mstd[4])
392 float vsum = 0.0f, wsum = 0.0f;
394 for (int i = 0; i < n; i++) {
395 vsum += softmax[i] * elliott(el[i]);
400 mstd[3] += (5.0f * vsum) / wsum * mstd[1] + mstd[0];
405 static void predictor(AVFilterContext *ctx,
406 const void *src, ptrdiff_t src_stride, void *dst,
407 const uint8_t *prescreen, int N,
408 void *data, int use_q2)
410 NNEDIContext *s = ctx->priv;
411 const PredictorCoefficients *const model = data;
412 const float *src_p = src;
415 // Adjust source pointer to point to top-left of filter window.
416 const float *window = src_p - (model->ydim / 2) * src_stride - (model->xdim / 2 - 1);
417 int filter_size = model->nsize;
418 int nns = model->nns;
420 for (int i = 0; i < N; i++) {
421 LOCAL_ALIGNED_32(float, input, [48 * 6]);
422 float activation[256 * 2];
429 gather_input(window + i, src_stride, input, mstd, model);
432 for (int nn = 0; nn < nns; nn++)
433 activation[nn] = dot_dsp(s, softmax_q1_filter(nn, model), input, filter_size, scale, model->softmax_bias_q1[nn]);
435 for (int nn = 0; nn < nns; nn++)
436 activation[model->nns + nn] = dot_dsp(s, elliott_q1_filter(nn, model), input, filter_size, scale, model->elliott_bias_q1[nn]);
438 transform_softmax_exp(activation, nns);
439 wae5(activation, activation + nns, nns, mstd);
442 for (int nn = 0; nn < nns; nn++)
443 activation[nn] = dot_dsp(s, softmax_q2_filter(nn, model), input, filter_size, scale, model->softmax_bias_q2[nn]);
445 for (int nn = 0; nn < nns; nn++)
446 activation[nns + nn] = dot_dsp(s, elliott_q2_filter(nn, model), input, filter_size, scale, model->elliott_bias_q2[nn]);
448 transform_softmax_exp(activation, nns);
449 wae5(activation, activation + nns, nns, mstd);
452 dst_p[i] = mstd[3] / (use_q2 ? 2 : 1);
456 static void read_bytes(const uint8_t *src, float *dst,
457 int src_stride, int dst_stride,
458 int width, int height, float scale)
460 for (int y = 0; y < height; y++) {
461 for (int x = 0; x < 32; x++)
462 dst[-x - 1] = src[x];
464 for (int x = 0; x < width; x++)
467 for (int x = 0; x < 32; x++)
468 dst[width + x] = src[width - x - 1];
475 static void read_words(const uint8_t *srcp, float *dst,
476 int src_stride, int dst_stride,
477 int width, int height, float scale)
479 const uint16_t *src = (const uint16_t *)srcp;
483 for (int y = 0; y < height; y++) {
484 for (int x = 0; x < 32; x++)
485 dst[-x - 1] = src[x] * scale;
487 for (int x = 0; x < width; x++)
488 dst[x] = src[x] * scale;
490 for (int x = 0; x < 32; x++)
491 dst[width + x] = src[width - x - 1] * scale;
498 static void write_bytes(const float *src, uint8_t *dst,
499 int src_stride, int dst_stride,
500 int width, int height, int depth,
503 for (int y = 0; y < height; y++) {
504 for (int x = 0; x < width; x++)
505 dst[x] = av_clip_uint8(src[x]);
512 static void write_words(const float *src, uint8_t *dstp,
513 int src_stride, int dst_stride,
514 int width, int height, int depth,
517 uint16_t *dst = (uint16_t *)dstp;
521 for (int y = 0; y < height; y++) {
522 for (int x = 0; x < width; x++)
523 dst[x] = av_clip_uintp2_c(src[x] * scale, depth);
530 static void interpolation(const void *src, ptrdiff_t src_stride,
531 void *dst, const uint8_t *prescreen, int n)
533 const float *src_p = src;
535 const float *window = src_p - 2 * src_stride;
537 for (int i = 0; i < n; i++) {
543 accum += (-3.0f / 32.0f) * window[0 * src_stride + i];
544 accum += (19.0f / 32.0f) * window[1 * src_stride + i];
545 accum += (19.0f / 32.0f) * window[2 * src_stride + i];
546 accum += (-3.0f / 32.0f) * window[3 * src_stride + i];
552 static int filter_slice(AVFilterContext *ctx, void *arg, int jobnr, int nb_jobs)
554 NNEDIContext *s = ctx->priv;
555 AVFrame *out = s->dst;
556 AVFrame *in = s->src;
557 const float in_scale = s->in_scale;
558 const float out_scale = s->out_scale;
559 const int depth = s->depth;
560 const int interlaced = in->interlaced_frame;
561 const int tff = s->field_n == (s->field < 0 ? interlaced ? in->top_field_first : 1 :
565 for (int p = 0; p < s->nb_planes; p++) {
566 const int height = s->planeheight[p];
567 const int width = s->planewidth[p];
568 const int slice_start = 2 * ((height / 2 * jobnr) / nb_jobs);
569 const int slice_end = 2 * ((height / 2 * (jobnr+1)) / nb_jobs);
570 const uint8_t *src_data = in->data[p];
571 uint8_t *dst_data = out->data[p];
572 uint8_t *dst = out->data[p] + slice_start * out->linesize[p];
573 const int src_linesize = in->linesize[p];
574 const int dst_linesize = out->linesize[p];
575 uint8_t *prescreen_buf = s->prescreen_buf + s->planewidth[0] * jobnr;
576 float *srcbuf = s->input_buf + s->input_size * jobnr;
577 const int srcbuf_stride = width + 64;
578 float *dstbuf = s->output_buf + s->input_size * jobnr;
579 const int dstbuf_stride = width;
580 const int slice_height = (slice_end - slice_start) / 2;
581 const int last_slice = slice_end == height;
582 const uint8_t *in_line;
586 if (!(s->process_plane & (1 << p))) {
587 av_image_copy_plane(dst, out->linesize[p],
588 in->data[p] + slice_start * in->linesize[p],
590 s->linesize[p], slice_end - slice_start);
594 y_out = slice_start + (tff ^ (slice_start & 1));
595 in_line = src_data + (y_out * src_linesize);
596 out_line = dst_data + (y_out * dst_linesize);
598 while (y_out < slice_end) {
599 memcpy(out_line, in_line, s->linesize[p]);
601 in_line += src_linesize * 2;
602 out_line += dst_linesize * 2;
605 y_out = slice_start + ((!tff) ^ (slice_start & 1));
607 s->read(src_data + FFMAX(y_out - 5, tff) * src_linesize,
609 src_linesize * 2, srcbuf_stride,
611 srcbuf += srcbuf_stride;
613 s->read(src_data + FFMAX(y_out - 3, tff) * src_linesize,
615 src_linesize * 2, srcbuf_stride,
617 srcbuf += srcbuf_stride;
619 s->read(src_data + FFMAX(y_out - 1, tff) * src_linesize,
621 src_linesize * 2, srcbuf_stride,
623 srcbuf += srcbuf_stride;
625 in_line = src_data + FFMIN(y_out + 1, height - 1 - !tff) * src_linesize;
626 out_line = dst_data + (y_out * dst_linesize);
628 s->read(in_line, srcbuf + 32, src_linesize * 2, srcbuf_stride,
629 width, slice_height - last_slice, in_scale);
631 y_out += (slice_height - last_slice) * 2;
633 s->read(src_data + FFMIN(y_out + 1, height - 1 - !tff) * src_linesize,
634 srcbuf + 32 + srcbuf_stride * (slice_height - last_slice),
635 src_linesize * 2, srcbuf_stride,
638 s->read(src_data + FFMIN(y_out + 3, height - 1 - !tff) * src_linesize,
639 srcbuf + 32 + srcbuf_stride * (slice_height + 1 - last_slice),
640 src_linesize * 2, srcbuf_stride,
643 s->read(src_data + FFMIN(y_out + 5, height - 1 - !tff) * src_linesize,
644 srcbuf + 32 + srcbuf_stride * (slice_height + 2 - last_slice),
645 src_linesize * 2, srcbuf_stride,
648 for (int y = 0; y < slice_end - slice_start; y += 2) {
650 s->prescreen[1](ctx, srcbuf + (y / 2) * srcbuf_stride + 32,
651 srcbuf_stride, prescreen_buf, width,
652 &s->prescreener_new[s->pscrn - 2]);
653 } else if (s->pscrn == 1) {
654 s->prescreen[0](ctx, srcbuf + (y / 2) * srcbuf_stride + 32,
655 srcbuf_stride, prescreen_buf, width,
656 &s->prescreener_old);
660 srcbuf + (y / 2) * srcbuf_stride + 32,
662 dstbuf + (y / 2) * dstbuf_stride,
663 prescreen_buf, width,
664 &s->coeffs[s->etype][s->nnsparam][s->nsize], s->qual == 2);
666 if (s->prescreen > 0)
667 interpolation(srcbuf + (y / 2) * srcbuf_stride + 32,
669 dstbuf + (y / 2) * dstbuf_stride,
670 prescreen_buf, width);
673 s->write(dstbuf, out_line, dstbuf_stride, dst_linesize * 2,
674 width, slice_height, depth, out_scale);
680 static int get_frame(AVFilterContext *ctx, int is_second)
682 NNEDIContext *s = ctx->priv;
683 AVFilterLink *outlink = ctx->outputs[0];
684 AVFrame *src = s->src;
686 s->dst = ff_get_video_buffer(outlink, outlink->w, outlink->h);
688 return AVERROR(ENOMEM);
689 av_frame_copy_props(s->dst, src);
690 s->dst->interlaced_frame = 0;
692 ctx->internal->execute(ctx, filter_slice, NULL, NULL, FFMIN(s->planeheight[1] / 2, s->nb_threads));
694 if (s->field == -2 || s->field > 1)
695 s->field_n = !s->field_n;
700 static int filter_frame(AVFilterLink *inlink, AVFrame *src)
702 AVFilterContext *ctx = inlink->dst;
703 AVFilterLink *outlink = ctx->outputs[0];
704 NNEDIContext *s = ctx->priv;
708 s->field == -2) && !s->second) {
710 } else if (s->field > 1 ||
715 ret = get_frame(ctx, 1);
717 av_frame_free(&s->dst);
718 av_frame_free(&s->second);
724 if (src->pts != AV_NOPTS_VALUE &&
725 dst->pts != AV_NOPTS_VALUE)
726 dst->pts += src->pts;
728 dst->pts = AV_NOPTS_VALUE;
730 ret = ff_filter_frame(outlink, dst);
735 s->cur_pts = s->second->pts;
736 av_frame_free(&s->second);
738 if ((s->deint && src->interlaced_frame &&
739 !ctx->is_disabled) ||
740 (!s->deint && !ctx->is_disabled)) {
745 if ((s->deint && !src->interlaced_frame) || ctx->is_disabled) {
746 AVFrame *dst = av_frame_clone(src);
749 av_frame_free(&s->second);
750 return AVERROR(ENOMEM);
753 if (s->field > 1 || s->field == -2) {
754 av_frame_free(&s->second);
755 if ((s->deint && src->interlaced_frame) ||
761 if (dst->pts != AV_NOPTS_VALUE)
763 return ff_filter_frame(outlink, dst);
767 ret = get_frame(ctx, 0);
769 av_frame_free(&s->dst);
770 av_frame_free(&s->src);
771 av_frame_free(&s->second);
775 if (src->pts != AV_NOPTS_VALUE)
776 s->dst->pts = src->pts * 2;
777 if (s->field <= 1 && s->field > -2) {
782 return ff_filter_frame(outlink, s->dst);
785 static int request_frame(AVFilterLink *link)
787 AVFilterContext *ctx = link->src;
788 NNEDIContext *s = ctx->priv;
794 ret = ff_request_frame(ctx->inputs[0]);
796 if (ret == AVERROR_EOF && s->second) {
797 AVFrame *next = av_frame_clone(s->second);
800 return AVERROR(ENOMEM);
802 next->pts = s->second->pts * 2 - s->cur_pts;
805 filter_frame(ctx->inputs[0], next);
806 } else if (ret < 0) {
813 static void copy_weights(float *dst, int n, const float **data)
815 memcpy(dst, *data, n * sizeof(float));
819 static float *allocate(float **ptr, int size)
828 static int allocate_model(PredictorCoefficients *coeffs, int xdim, int ydim, int nns)
830 int filter_size = nns * xdim * ydim;
834 data = av_malloc_array(filter_size + bias_size, 4 * sizeof(float));
836 return AVERROR(ENOMEM);
841 coeffs->nsize = xdim * ydim;
844 coeffs->softmax_q1 = allocate(&data, filter_size);
845 coeffs->elliott_q1 = allocate(&data, filter_size);
846 coeffs->softmax_bias_q1 = allocate(&data, bias_size);
847 coeffs->elliott_bias_q1 = allocate(&data, bias_size);
849 coeffs->softmax_q2 = allocate(&data, filter_size);
850 coeffs->elliott_q2 = allocate(&data, filter_size);
851 coeffs->softmax_bias_q2 = allocate(&data, bias_size);
852 coeffs->elliott_bias_q2 = allocate(&data, bias_size);
857 static int read_weights(AVFilterContext *ctx, const float *bdata)
859 NNEDIContext *s = ctx->priv;
862 copy_weights(&s->prescreener_old.kernel_l0[0][0], 4 * 48, &bdata);
863 copy_weights(s->prescreener_old.bias_l0, 4, &bdata);
865 copy_weights(&s->prescreener_old.kernel_l1[0][0], 4 * 4, &bdata);
866 copy_weights(s->prescreener_old.bias_l1, 4, &bdata);
868 copy_weights(&s->prescreener_old.kernel_l2[0][0], 4 * 8, &bdata);
869 copy_weights(s->prescreener_old.bias_l2, 4, &bdata);
871 for (int i = 0; i < 3; i++) {
872 PrescreenerNewCoefficients *data = &s->prescreener_new[i];
873 float kernel_l0_shuffled[4 * 64];
874 float kernel_l1_shuffled[4 * 4];
876 copy_weights(kernel_l0_shuffled, 4 * 64, &bdata);
877 copy_weights(data->bias_l0, 4, &bdata);
879 copy_weights(kernel_l1_shuffled, 4 * 4, &bdata);
880 copy_weights(data->bias_l1, 4, &bdata);
882 for (int n = 0; n < 4; n++) {
883 for (int k = 0; k < 64; k++)
884 data->kernel_l0[n][k] = kernel_l0_shuffled[(k / 8) * 32 + n * 8 + k % 8];
885 for (int k = 0; k < 4; k++)
886 data->kernel_l1[n][k] = kernel_l1_shuffled[k * 4 + n];
890 for (int m = 0; m < 2; m++) {
891 // Grouping by neuron count.
892 for (int i = 0; i < 5; i++) {
893 int nns = NNEDI_NNS[i];
895 // Grouping by window size.
896 for (int j = 0; j < 7; j++) {
897 PredictorCoefficients *model = &s->coeffs[m][i][j];
898 int xdim = NNEDI_XDIM[j];
899 int ydim = NNEDI_YDIM[j];
900 int filter_size = xdim * ydim;
902 ret = allocate_model(model, xdim, ydim, nns);
906 // Quality 1 model. NNS[i] * (XDIM[j] * YDIM[j]) * 2 coefficients.
907 copy_weights(model->softmax_q1, nns * filter_size, &bdata);
908 copy_weights(model->elliott_q1, nns * filter_size, &bdata);
910 // Quality 1 model bias. NNS[i] * 2 coefficients.
911 copy_weights(model->softmax_bias_q1, nns, &bdata);
912 copy_weights(model->elliott_bias_q1, nns, &bdata);
914 // Quality 2 model. NNS[i] * (XDIM[j] * YDIM[j]) * 2 coefficients.
915 copy_weights(model->softmax_q2, nns * filter_size, &bdata);
916 copy_weights(model->elliott_q2, nns * filter_size, &bdata);
918 // Quality 2 model bias. NNS[i] * 2 coefficients.
919 copy_weights(model->softmax_bias_q2, nns, &bdata);
920 copy_weights(model->elliott_bias_q2, nns, &bdata);
928 static float mean(const float *input, int size)
932 for (int i = 0; i < size; i++)
938 static void transform(float *input, int size, float mean, float half)
940 for (int i = 0; i < size; i++)
941 input[i] = (input[i] - mean) / half;
944 static void subtract_mean_old(PrescreenerOldCoefficients *coeffs, float half)
946 for (int n = 0; n < 4; n++) {
947 float m = mean(coeffs->kernel_l0[n], 48);
949 transform(coeffs->kernel_l0[n], 48, m, half);
953 static void subtract_mean_new(PrescreenerNewCoefficients *coeffs, float half)
955 for (int n = 0; n < 4; n++) {
956 float m = mean(coeffs->kernel_l0[n], 64);
958 transform(coeffs->kernel_l0[n], 64, m, half);
962 static void subtract_mean_predictor(PredictorCoefficients *model)
964 int filter_size = model->nsize;
965 int nns = model->nns;
967 float softmax_means[256]; // Average of individual softmax filters.
968 float elliott_means[256]; // Average of individual elliott filters.
969 float mean_filter[48 * 6]; // Pointwise average of all softmax filters.
973 for (int nn = 0; nn < nns; nn++) {
974 softmax_means[nn] = mean(model->softmax_q1 + nn * filter_size, filter_size);
975 elliott_means[nn] = mean(model->elliott_q1 + nn * filter_size, filter_size);
977 for (int k = 0; k < filter_size; k++)
978 mean_filter[k] += model->softmax_q1[nn * filter_size + k] - softmax_means[nn];
981 for (int k = 0; k < filter_size; k++)
982 mean_filter[k] /= nns;
984 mean_bias = mean(model->softmax_bias_q1, nns);
986 for (int nn = 0; nn < nns; nn++) {
987 for (int k = 0; k < filter_size; k++) {
988 model->softmax_q1[nn * filter_size + k] -= softmax_means[nn] + mean_filter[k];
989 model->elliott_q1[nn * filter_size + k] -= elliott_means[nn];
991 model->softmax_bias_q1[nn] -= mean_bias;
995 memset(mean_filter, 0, 48 * 6 * sizeof(float));
997 for (int nn = 0; nn < nns; nn++) {
998 softmax_means[nn] = mean(model->softmax_q2 + nn * filter_size, filter_size);
999 elliott_means[nn] = mean(model->elliott_q2 + nn * filter_size, filter_size);
1001 for (int k = 0; k < filter_size; k++) {
1002 mean_filter[k] += model->softmax_q2[nn * filter_size + k] - softmax_means[nn];
1006 for (int k = 0; k < filter_size; k++)
1007 mean_filter[k] /= nns;
1009 mean_bias = mean(model->softmax_bias_q2, nns);
1011 for (int nn = 0; nn < nns; nn++) {
1012 for (int k = 0; k < filter_size; k++) {
1013 model->softmax_q2[nn * filter_size + k] -= softmax_means[nn] + mean_filter[k];
1014 model->elliott_q2[nn * filter_size + k] -= elliott_means[nn];
1017 model->softmax_bias_q2[nn] -= mean_bias;
1021 static av_cold int init(AVFilterContext *ctx)
1023 NNEDIContext *s = ctx->priv;
1024 FILE *weights_file = NULL;
1025 int64_t weights_size;
1030 weights_file = av_fopen_utf8(s->weights_file, "rb");
1031 if (!weights_file) {
1032 av_log(ctx, AV_LOG_ERROR, "No weights file provided, aborting!\n");
1033 return AVERROR(EINVAL);
1036 if (fseek(weights_file, 0, SEEK_END)) {
1037 av_log(ctx, AV_LOG_ERROR, "Couldn't seek to the end of weights file.\n");
1038 fclose(weights_file);
1039 return AVERROR(EINVAL);
1042 weights_size = ftell(weights_file);
1044 if (weights_size == -1) {
1045 fclose(weights_file);
1046 av_log(ctx, AV_LOG_ERROR, "Couldn't get size of weights file.\n");
1047 return AVERROR(EINVAL);
1048 } else if (weights_size != NNEDI_WEIGHTS_SIZE) {
1049 fclose(weights_file);
1050 av_log(ctx, AV_LOG_ERROR, "Unexpected weights file size.\n");
1051 return AVERROR(EINVAL);
1054 if (fseek(weights_file, 0, SEEK_SET)) {
1055 fclose(weights_file);
1056 av_log(ctx, AV_LOG_ERROR, "Couldn't seek to the start of weights file.\n");
1057 return AVERROR(EINVAL);
1060 bdata = av_malloc(NNEDI_WEIGHTS_SIZE);
1062 fclose(weights_file);
1063 return AVERROR(ENOMEM);
1066 bytes_read = fread(bdata, 1, NNEDI_WEIGHTS_SIZE, weights_file);
1067 if (bytes_read != NNEDI_WEIGHTS_SIZE) {
1068 fclose(weights_file);
1069 ret = AVERROR_INVALIDDATA;
1070 av_log(ctx, AV_LOG_ERROR, "Couldn't read weights file.\n");
1074 fclose(weights_file);
1076 s->fdsp = avpriv_float_dsp_alloc(0);
1078 ret = AVERROR(ENOMEM);
1082 ret = read_weights(ctx, bdata);
1091 static int config_input(AVFilterLink *inlink)
1093 AVFilterContext *ctx = inlink->dst;
1094 NNEDIContext *s = ctx->priv;
1095 const AVPixFmtDescriptor *desc = av_pix_fmt_desc_get(inlink->format);
1098 s->depth = desc->comp[0].depth;
1099 s->nb_threads = ff_filter_get_nb_threads(ctx);
1100 s->nb_planes = av_pix_fmt_count_planes(inlink->format);
1101 if ((ret = av_image_fill_linesizes(s->linesize, inlink->format, inlink->w)) < 0)
1104 s->planewidth[1] = s->planewidth[2] = AV_CEIL_RSHIFT(inlink->w, desc->log2_chroma_w);
1105 s->planewidth[0] = s->planewidth[3] = inlink->w;
1106 s->planeheight[1] = s->planeheight[2] = AV_CEIL_RSHIFT(inlink->h, desc->log2_chroma_h);
1107 s->planeheight[0] = s->planeheight[3] = inlink->h;
1109 s->half = ((1 << 8) - 1) / 2.f;
1110 s->out_scale = 1 << (s->depth - 8);
1111 s->in_scale = 1.f / s->out_scale;
1115 s->read = read_bytes;
1116 s->write = write_bytes;
1119 s->read = read_words;
1120 s->write = write_words;
1124 subtract_mean_old(&s->prescreener_old, s->half);
1125 subtract_mean_new(&s->prescreener_new[0], s->half);
1126 subtract_mean_new(&s->prescreener_new[1], s->half);
1127 subtract_mean_new(&s->prescreener_new[2], s->half);
1129 s->prescreen[0] = process_old;
1130 s->prescreen[1] = process_new;
1132 for (int i = 0; i < 2; i++) {
1133 for (int j = 0; j < 5; j++) {
1134 for (int k = 0; k < 7; k++)
1135 subtract_mean_predictor(&s->coeffs[i][j][k]);
1139 s->prescreen_buf = av_calloc(s->nb_threads * s->planewidth[0], sizeof(*s->prescreen_buf));
1140 if (!s->prescreen_buf)
1141 return AVERROR(ENOMEM);
1143 s->input_size = (s->planewidth[0] + 64) * (s->planeheight[0] + 6);
1144 s->input_buf = av_calloc(s->nb_threads * s->input_size, sizeof(*s->input_buf));
1146 return AVERROR(ENOMEM);
1148 s->output_buf = av_calloc(s->nb_threads * s->input_size, sizeof(*s->output_buf));
1150 return AVERROR(ENOMEM);
1155 static av_cold void uninit(AVFilterContext *ctx)
1157 NNEDIContext *s = ctx->priv;
1159 av_freep(&s->prescreen_buf);
1160 av_freep(&s->input_buf);
1161 av_freep(&s->output_buf);
1164 for (int i = 0; i < 2; i++) {
1165 for (int j = 0; j < 5; j++) {
1166 for (int k = 0; k < 7; k++) {
1167 av_freep(&s->coeffs[i][j][k].data);
1172 av_frame_free(&s->second);
1175 static const AVFilterPad inputs[] = {
1178 .type = AVMEDIA_TYPE_VIDEO,
1179 .filter_frame = filter_frame,
1180 .config_props = config_input,
1185 static const AVFilterPad outputs[] = {
1188 .type = AVMEDIA_TYPE_VIDEO,
1189 .config_props = config_output,
1190 .request_frame = request_frame,
1195 AVFilter ff_vf_nnedi = {
1197 .description = NULL_IF_CONFIG_SMALL("Apply neural network edge directed interpolation intra-only deinterlacer."),
1198 .priv_size = sizeof(NNEDIContext),
1199 .priv_class = &nnedi_class,
1202 .query_formats = query_formats,
1205 .flags = AVFILTER_FLAG_SUPPORT_TIMELINE_INTERNAL | AVFILTER_FLAG_SLICE_THREADS,
1206 .process_command = ff_filter_process_command,