Unified · Split
| @@ -37,8 +37,8 @@ static const uint8_t NNEDI_XDIM[] = { 8, 16, 32, 48, 8, 16, 32 }; · 36 unmodified lines | ||
| 37 | 37 | static const uint8_t NNEDI_YDIM[] = { 6, 6, 6, 6, 4, 4, 4 }; |
| 38 | 38 | static const uint16_t NNEDI_NNS[] = { 16, 32, 64, 128, 256 }; |
| 39 | 39 | |
| 40 | −typedef struct PrescreenerOldCoefficients { | |
| 41 | − DECLARE_ALIGNED(32, float, kernel_l0)[4][14 * 4]; | |
| 40 | +typedef struct PrescreenerCoefficients { | |
| 41 | + DECLARE_ALIGNED(32, float, kernel_l0)[4][16 * 4]; | |
| 42 | 42 | DECLARE_ALIGNED(32, float, bias_l0)[4]; |
| 43 | 43 | |
| 44 | 44 | DECLARE_ALIGNED(32, float, kernel_l1)[4][4]; |
| @@ -46,15 +46,7 @@ typedef struct PrescreenerOldCoefficients { · 1 unmodified line | ||
| 46 | 46 | |
| 47 | 47 | DECLARE_ALIGNED(32, float, kernel_l2)[4][8]; |
| 48 | 48 | DECLARE_ALIGNED(32, float, bias_l2)[4]; |
| 49 | −} PrescreenerOldCoefficients; | |
| 50 | − | |
| 51 | −typedef struct PrescreenerNewCoefficients { | |
| 52 | − DECLARE_ALIGNED(32, float, kernel_l0)[4][16 * 4]; | |
| 53 | − DECLARE_ALIGNED(32, float, bias_l0)[4]; | |
| 54 | − | |
| 55 | − DECLARE_ALIGNED(32, float, kernel_l1)[4][4]; | |
| 56 | − DECLARE_ALIGNED(32, float, bias_l1)[4]; | |
| 57 | −} PrescreenerNewCoefficients; | |
| 49 | +} PrescreenerCoefficients; | |
| 58 | 50 | |
| 59 | 51 | typedef struct PredictorCoefficients { |
| 60 | 52 | int xdim, ydim, nns, nsize; |
| @@ -89,8 +81,7 @@ typedef struct NNEDIContext { · 28 unmodified lines | ||
| 89 | 81 | int planeheight[4]; |
| 90 | 82 | int field_n; |
| 91 | 83 | |
| 92 | − PrescreenerOldCoefficients prescreener_old; | |
| 93 | − PrescreenerNewCoefficients prescreener_new[3]; | |
| 84 | + PrescreenerCoefficients prescreener[4]; | |
| 94 | 85 | PredictorCoefficients coeffs[2][5][7]; |
| 95 | 86 | |
| 96 | 87 | float half; |
| @@ -108,9 +99,9 @@ typedef struct NNEDIContext { · 11 unmodified lines | ||
| 108 | 99 | int pscrn; |
| 109 | 100 | |
| 110 | 101 | int input_size; |
| 111 | − uint8_t *prescreen_buf; | |
| 112 | − float *input_buf; | |
| 113 | − float *output_buf; | |
| 102 | + uint8_t **prescreen_buf; | |
| 103 | + float **input_buf; | |
| 104 | + float **output_buf; | |
| 114 | 105 | |
| 115 | 106 | void (*read)(const uint8_t *src, float *dst, |
| 116 | 107 | int src_stride, int dst_stride, |
| @@ -120,7 +111,8 @@ typedef struct NNEDIContext { · 3 unmodified lines | ||
| 120 | 111 | int width, int height, int depth, float scale); |
| 121 | 112 | void (*prescreen[2])(AVFilterContext *ctx, |
| 122 | 113 | const void *src, ptrdiff_t src_stride, |
| 123 | − uint8_t *prescreen, int N, void *data); | |
| 114 | + uint8_t *prescreen, int N, | |
| 115 | + const PrescreenerCoefficients *const coeffs); | |
| 124 | 116 | } NNEDIContext; |
| 125 | 117 | |
| 126 | 118 | #define OFFSET(x) offsetof(NNEDIContext, x) |
| @@ -222,7 +214,7 @@ static int query_formats(AVFilterContext *ctx) · 95 unmodified lines | ||
| 222 | 214 | return ff_set_common_formats(ctx, fmts_list); |
| 223 | 215 | } |
| 224 | 216 | |
| 225 | −static float dot_dsp(NNEDIContext *s, const float *kernel, const float *input, | |
| 217 | +static float dot_dsp(const NNEDIContext *const s, const float *kernel, const float *input, | |
| 226 | 218 | int n, float scale, float bias) |
| 227 | 219 | { |
| 228 | 220 | float sum; |
| @@ -246,10 +238,9 @@ static void transform_elliott(float *input, int size) · 17 unmodified lines | ||
| 246 | 238 | static void process_old(AVFilterContext *ctx, |
| 247 | 239 | const void *src, ptrdiff_t src_stride, |
| 248 | 240 | uint8_t *prescreen, int N, |
| 249 | − void *data) | |
| 241 | + const PrescreenerCoefficients *const m_data) | |
| 250 | 242 | { |
| 251 | 243 | NNEDIContext *s = ctx->priv; |
| 252 | − const PrescreenerOldCoefficients *const m_data = data; | |
| 253 | 244 | const float *src_p = src; |
| 254 | 245 | |
| 255 | 246 | // Adjust source pointer to point to top-left of filter window. |
| @@ -283,10 +274,9 @@ static void process_old(AVFilterContext *ctx, · 27 unmodified lines | ||
| 283 | 274 | static void process_new(AVFilterContext *ctx, |
| 284 | 275 | const void *src, ptrdiff_t src_stride, |
| 285 | 276 | uint8_t *prescreen, int N, |
| 286 | − void *data) | |
| 277 | + const PrescreenerCoefficients *const m_data) | |
| 287 | 278 | { |
| 288 | 279 | NNEDIContext *s = ctx->priv; |
| 289 | − const PrescreenerNewCoefficients *const m_data = data; | |
| 290 | 280 | const float *src_p = src; |
| 291 | 281 | |
| 292 | 282 | // Adjust source pointer to point to top-left of filter window. |
| @@ -344,8 +334,8 @@ static void gather_input(const float *src, ptrdiff_t src_stride, · 51 unmodified lines | ||
| 344 | 334 | float *buf, float mstd[4], |
| 345 | 335 | const PredictorCoefficients *const model) |
| 346 | 336 | { |
| 347 | − float sum = 0; | |
| 348 | − float sum_sq = 0; | |
| 337 | + float sum = 0.f; | |
| 338 | + float sum_sq = 0.f; | |
| 349 | 339 | float tmp; |
| 350 | 340 | |
| 351 | 341 | for (int i = 0; i < model->ydim; i++) { |
| @@ -405,17 +395,16 @@ static void wae5(const float *softmax, const float *el, · 53 unmodified lines | ||
| 405 | 395 | static void predictor(AVFilterContext *ctx, |
| 406 | 396 | const void *src, ptrdiff_t src_stride, void *dst, |
| 407 | 397 | const uint8_t *prescreen, int N, |
| 408 | − void *data, int use_q2) | |
| 398 | + const PredictorCoefficients *const model, int use_q2) | |
| 409 | 399 | { |
| 410 | − NNEDIContext *s = ctx->priv; | |
| 411 | − const PredictorCoefficients *const model = data; | |
| 400 | + const NNEDIContext *const s = ctx->priv; | |
| 412 | 401 | const float *src_p = src; |
| 413 | 402 | float *dst_p = dst; |
| 414 | 403 | |
| 415 | 404 | // Adjust source pointer to point to top-left of filter window. |
| 416 | 405 | 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; | |
| 406 | + const int filter_size = model->nsize; | |
| 407 | + const int nns = model->nns; | |
| 419 | 408 | |
| 420 | 409 | for (int i = 0; i < N; i++) { |
| 421 | 410 | LOCAL_ALIGNED_32(float, input, [48 * 6]); |
| @@ -433,7 +422,7 @@ static void predictor(AVFilterContext *ctx, · 11 unmodified lines | ||
| 433 | 422 | activation[nn] = dot_dsp(s, softmax_q1_filter(nn, model), input, filter_size, scale, model->softmax_bias_q1[nn]); |
| 434 | 423 | |
| 435 | 424 | 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]); | |
| 425 | + activation[nns + nn] = dot_dsp(s, elliott_q1_filter(nn, model), input, filter_size, scale, model->elliott_bias_q1[nn]); | |
| 437 | 426 | |
| 438 | 427 | transform_softmax_exp(activation, nns); |
| 439 | 428 | wae5(activation, activation + nns, nns, mstd); |
| @@ -449,7 +438,7 @@ static void predictor(AVFilterContext *ctx, · 9 unmodified lines | ||
| 449 | 438 | wae5(activation, activation + nns, nns, mstd); |
| 450 | 439 | } |
| 451 | 440 | |
| 452 | − dst_p[i] = mstd[3] / (use_q2 ? 2 : 1); | |
| 441 | + dst_p[i] = mstd[3] * (use_q2 ? 0.5f : 1.f); | |
| 453 | 442 | } |
| 454 | 443 | } |
| 455 | 444 | |
| @@ -551,7 +540,7 @@ static void interpolation(const void *src, ptrdiff_t src_stride, · 95 unmodified lines | ||
| 551 | 540 | |
| 552 | 541 | static int filter_slice(AVFilterContext *ctx, void *arg, int jobnr, int nb_jobs) |
| 553 | 542 | { |
| 554 | − NNEDIContext *s = ctx->priv; | |
| 543 | + const NNEDIContext *const s = ctx->priv; | |
| 555 | 544 | AVFrame *out = s->dst; |
| 556 | 545 | AVFrame *in = s->src; |
| 557 | 546 | const float in_scale = s->in_scale; |
| @@ -572,10 +561,10 @@ static int filter_slice(AVFilterContext *ctx, void *arg, int jobnr, int nb_jobs) · 14 unmodified lines | ||
| 572 | 561 | uint8_t *dst = out->data[p] + slice_start * out->linesize[p]; |
| 573 | 562 | const int src_linesize = in->linesize[p]; |
| 574 | 563 | 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; | |
| 564 | + uint8_t *prescreen_buf = s->prescreen_buf[jobnr]; | |
| 565 | + float *srcbuf = s->input_buf[jobnr]; | |
| 577 | 566 | const int srcbuf_stride = width + 64; |
| 578 | − float *dstbuf = s->output_buf + s->input_size * jobnr; | |
| 567 | + float *dstbuf = s->output_buf[jobnr]; | |
| 579 | 568 | const int dstbuf_stride = width; |
| 580 | 569 | const int slice_height = (slice_end - slice_start) / 2; |
| 581 | 570 | const int last_slice = slice_end == height; |
| @@ -646,15 +635,10 @@ static int filter_slice(AVFilterContext *ctx, void *arg, int jobnr, int nb_jobs) · 64 unmodified lines | ||
| 646 | 635 | width, 1, in_scale); |
| 647 | 636 | |
| 648 | 637 | for (int y = 0; y < slice_end - slice_start; y += 2) { |
| 649 | − if (s->pscrn > 1) { | |
| 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); | |
| 657 | − } | |
| 638 | + if (s->prescreen > 0) | |
| 639 | + s->prescreen[s->pscrn > 1](ctx, srcbuf + (y / 2) * srcbuf_stride + 32, | |
| 640 | + srcbuf_stride, prescreen_buf, width, | |
| 641 | + &s->prescreener[s->pscrn - 1]); | |
| 658 | 642 | |
| 659 | 643 | predictor(ctx, |
| 660 | 644 | srcbuf + (y / 2) * srcbuf_stride + 32, |
| @@ -831,7 +815,7 @@ static int allocate_model(PredictorCoefficients *coeffs, int xdim, int ydim, int · 170 unmodified lines | ||
| 831 | 815 | int bias_size = nns; |
| 832 | 816 | float *data; |
| 833 | 817 | |
| 834 | − data = av_malloc_array(filter_size + bias_size, 4 * sizeof(float)); | |
| 818 | + data = av_calloc(filter_size + bias_size, 4 * sizeof(float)); | |
| 835 | 819 | if (!data) |
| 836 | 820 | return AVERROR(ENOMEM); |
| 837 | 821 | |
| @@ -859,17 +843,17 @@ static int read_weights(AVFilterContext *ctx, const float *bdata) · 21 unmodified lines | ||
| 859 | 843 | NNEDIContext *s = ctx->priv; |
| 860 | 844 | int ret; |
| 861 | 845 | |
| 862 | − copy_weights(&s->prescreener_old.kernel_l0[0][0], 4 * 48, &bdata); | |
| 863 | − copy_weights(s->prescreener_old.bias_l0, 4, &bdata); | |
| 846 | + copy_weights(&s->prescreener[0].kernel_l0[0][0], 4 * 48, &bdata); | |
| 847 | + copy_weights(s->prescreener[0].bias_l0, 4, &bdata); | |
| 864 | 848 | |
| 865 | − copy_weights(&s->prescreener_old.kernel_l1[0][0], 4 * 4, &bdata); | |
| 866 | − copy_weights(s->prescreener_old.bias_l1, 4, &bdata); | |
| 849 | + copy_weights(&s->prescreener[0].kernel_l1[0][0], 4 * 4, &bdata); | |
| 850 | + copy_weights(s->prescreener[0].bias_l1, 4, &bdata); | |
| 867 | 851 | |
| 868 | − copy_weights(&s->prescreener_old.kernel_l2[0][0], 4 * 8, &bdata); | |
| 869 | − copy_weights(s->prescreener_old.bias_l2, 4, &bdata); | |
| 852 | + copy_weights(&s->prescreener[0].kernel_l2[0][0], 4 * 8, &bdata); | |
| 853 | + copy_weights(s->prescreener[0].bias_l2, 4, &bdata); | |
| 870 | 854 | |
| 871 | 855 | for (int i = 0; i < 3; i++) { |
| 872 | − PrescreenerNewCoefficients *data = &s->prescreener_new[i]; | |
| 856 | + PrescreenerCoefficients *data = &s->prescreener[i + 1]; | |
| 873 | 857 | float kernel_l0_shuffled[4 * 64]; |
| 874 | 858 | float kernel_l1_shuffled[4 * 4]; |
| 875 | 859 | |
| @@ -890,14 +874,14 @@ static int read_weights(AVFilterContext *ctx, const float *bdata) · 14 unmodified lines | ||
| 890 | 874 | for (int m = 0; m < 2; m++) { |
| 891 | 875 | // Grouping by neuron count. |
| 892 | 876 | for (int i = 0; i < 5; i++) { |
| 893 | − int nns = NNEDI_NNS[i]; | |
| 877 | + const int nns = NNEDI_NNS[i]; | |
| 894 | 878 | |
| 895 | 879 | // Grouping by window size. |
| 896 | 880 | for (int j = 0; j < 7; j++) { |
| 897 | 881 | 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; | |
| 882 | + const int xdim = NNEDI_XDIM[j]; | |
| 883 | + const int ydim = NNEDI_YDIM[j]; | |
| 884 | + const int filter_size = xdim * ydim; | |
| 901 | 885 | |
| 902 | 886 | ret = allocate_model(model, xdim, ydim, nns); |
| 903 | 887 | if (ret < 0) |
| @@ -927,7 +911,7 @@ static int read_weights(AVFilterContext *ctx, const float *bdata) · 23 unmodified lines | ||
| 927 | 911 | |
| 928 | 912 | static float mean(const float *input, int size) |
| 929 | 913 | { |
| 930 | − float sum = 0.; | |
| 914 | + float sum = 0.f; | |
| 931 | 915 | |
| 932 | 916 | for (int i = 0; i < size; i++) |
| 933 | 917 | sum += input[i]; |
| @@ -941,7 +925,7 @@ static void transform(float *input, int size, float mean, float half) · 7 unmodified lines | ||
| 941 | 925 | input[i] = (input[i] - mean) / half; |
| 942 | 926 | } |
| 943 | 927 | |
| 944 | −static void subtract_mean_old(PrescreenerOldCoefficients *coeffs, float half) | |
| 928 | +static void subtract_mean_old(PrescreenerCoefficients *coeffs, float half) | |
| 945 | 929 | { |
| 946 | 930 | for (int n = 0; n < 4; n++) { |
| 947 | 931 | float m = mean(coeffs->kernel_l0[n], 48); |
| @@ -950,7 +934,7 @@ static void subtract_mean_old(PrescreenerOldCoefficients *coeffs, float half) · 2 unmodified lines | ||
| 950 | 934 | } |
| 951 | 935 | } |
| 952 | 936 | |
| 953 | −static void subtract_mean_new(PrescreenerNewCoefficients *coeffs, float half) | |
| 937 | +static void subtract_mean_new(PrescreenerCoefficients *coeffs, float half) | |
| 954 | 938 | { |
| 955 | 939 | for (int n = 0; n < 4; n++) { |
| 956 | 940 | float m = mean(coeffs->kernel_l0[n], 64); |
| @@ -1121,10 +1105,10 @@ static int config_input(AVFilterLink *inlink) · 164 unmodified lines | ||
| 1121 | 1105 | break; |
| 1122 | 1106 | } |
| 1123 | 1107 | |
| 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); | |
| 1108 | + subtract_mean_old(&s->prescreener[0], s->half); | |
| 1109 | + subtract_mean_new(&s->prescreener[1], s->half); | |
| 1110 | + subtract_mean_new(&s->prescreener[2], s->half); | |
| 1111 | + subtract_mean_new(&s->prescreener[3], s->half); | |
| 1128 | 1112 | |
| 1129 | 1113 | s->prescreen[0] = process_old; |
| 1130 | 1114 | s->prescreen[1] = process_new; |
| @@ -1136,19 +1120,37 @@ static int config_input(AVFilterLink *inlink) · 5 unmodified lines | ||
| 1136 | 1120 | } |
| 1137 | 1121 | } |
| 1138 | 1122 | |
| 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); | |
| 1142 | − | |
| 1143 | 1123 | 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)); | |
| 1124 | + s->input_buf = av_calloc(s->nb_threads, sizeof(*s->input_buf)); | |
| 1145 | 1125 | if (!s->input_buf) |
| 1146 | 1126 | return AVERROR(ENOMEM); |
| 1147 | 1127 | |
| 1148 | − s->output_buf = av_calloc(s->nb_threads * s->input_size, sizeof(*s->output_buf)); | |
| 1128 | + for (int i = 0; i < s->nb_threads; i++) { | |
| 1129 | + s->input_buf[i] = av_calloc(s->input_size, sizeof(**s->input_buf)); | |
| 1130 | + if (!s->input_buf[i]) | |
| 1131 | + return AVERROR(ENOMEM); | |
| 1132 | + } | |
| 1133 | + | |
| 1134 | + s->output_buf = av_calloc(s->nb_threads, sizeof(*s->output_buf)); | |
| 1149 | 1135 | if (!s->output_buf) |
| 1150 | 1136 | return AVERROR(ENOMEM); |
| 1151 | 1137 | |
| 1138 | + for (int i = 0; i < s->nb_threads; i++) { | |
| 1139 | + s->output_buf[i] = av_calloc(s->input_size, sizeof(**s->output_buf)); | |
| 1140 | + if (!s->output_buf[i]) | |
| 1141 | + return AVERROR(ENOMEM); | |
| 1142 | + } | |
| 1143 | + | |
| 1144 | + s->prescreen_buf = av_calloc(s->nb_threads, sizeof(*s->prescreen_buf)); | |
| 1145 | + if (!s->prescreen_buf) | |
| 1146 | + return AVERROR(ENOMEM); | |
| 1147 | + | |
| 1148 | + for (int i = 0; i < s->nb_threads; i++) { | |
| 1149 | + s->prescreen_buf[i] = av_calloc(s->planewidth[0], sizeof(**s->prescreen_buf)); | |
| 1150 | + if (!s->prescreen_buf[i]) | |
| 1151 | + return AVERROR(ENOMEM); | |
| 1152 | + } | |
| 1153 | + | |
| 1152 | 1154 | return 0; |
| 1153 | 1155 | } |
| 1154 | 1156 | |
| @@ -1156,8 +1158,19 @@ static av_cold void uninit(AVFilterContext *ctx) · 1 unmodified line | ||
| 1156 | 1158 | { |
| 1157 | 1159 | NNEDIContext *s = ctx->priv; |
| 1158 | 1160 | |
| 1161 | + for (int i = 0; i < s->nb_threads && s->prescreen_buf; i++) | |
| 1162 | + av_freep(&s->prescreen_buf[i]); | |
| 1163 | + | |
| 1159 | 1164 | av_freep(&s->prescreen_buf); |
| 1165 | + | |
| 1166 | + for (int i = 0; i < s->nb_threads && s->input_buf; i++) | |
| 1167 | + av_freep(&s->input_buf[i]); | |
| 1168 | + | |
| 1160 | 1169 | av_freep(&s->input_buf); |
| 1170 | + | |
| 1171 | + for (int i = 0; i < s->nb_threads && s->output_buf; i++) | |
| 1172 | + av_freep(&s->output_buf[i]); | |
| 1173 | + | |
| 1161 | 1174 | av_freep(&s->output_buf); |
| 1162 | 1175 | av_freep(&s->fdsp); |
| 1163 | 1176 | |