Unified · Split
| @@ -325,7 +325,7 @@ static void process_new(AVFilterContext *ctx, · 324 unmodified lines | |||
| 325 | } | 325 | } |
| 326 | } | 326 | } |
| 327 | 327 | ||
| 328 | static size_t filter_offset(unsigned nn, PredictorCoefficients *model) | 328 | static int filter_offset(unsigned nn, PredictorCoefficients *model) |
| 329 | { | 329 | { |
| 330 | return nn * model->xdim * model->ydim; | 330 | return nn * model->xdim * model->ydim; |
| 331 | } | 331 | } |
| @@ -420,8 +420,8 @@ static void predictor(AVFilterContext *ctx, · 88 unmodified lines | |||
| 420 | 420 | ||
| 421 | // Adjust source pointer to point to top-left of filter window. | 421 | // Adjust source pointer to point to top-left of filter window. |
| 422 | const float *window = src_p - (model->ydim / 2) * src_stride - (model->xdim / 2 - 1); | 422 | const float *window = src_p - (model->ydim / 2) * src_stride - (model->xdim / 2 - 1); |
| 423 | unsigned filter_size = model->xdim * model->ydim; | 423 | int filter_size = model->xdim * model->ydim; |
| 424 | unsigned nns = model->nns; | 424 | int nns = model->nns; |
| 425 | 425 | ||
| 426 | for (int i = 0; i < N; i++) { | 426 | for (int i = 0; i < N; i++) { |
| 427 | LOCAL_ALIGNED_32(float, input, [48 * 6]); | 427 | LOCAL_ALIGNED_32(float, input, [48 * 6]); |
| @@ -816,13 +816,13 @@ static int request_frame(AVFilterLink *link) · 388 unmodified lines | |||
| 816 | return 0; | 816 | return 0; |
| 817 | } | 817 | } |
| 818 | 818 | ||
| 819 | static void read(float *dst, size_t n, const float **data) | 819 | static void copy_weights(float *dst, int n, const float **data) |
| 820 | { | 820 | { |
| 821 | memcpy(dst, *data, n * sizeof(float)); | 821 | memcpy(dst, *data, n * sizeof(float)); |
| 822 | *data += n; | 822 | *data += n; |
| 823 | } | 823 | } |
| 824 | 824 | ||
| 825 | static float *allocate(float **ptr, size_t size) | 825 | static float *allocate(float **ptr, int size) |
| 826 | { | 826 | { |
| 827 | float *ret = *ptr; | 827 | float *ret = *ptr; |
| 828 | 828 | ||
| @@ -833,8 +833,8 @@ static float *allocate(float **ptr, size_t size) · 4 unmodified lines | |||
| 833 | 833 | ||
| 834 | static int allocate_model(PredictorCoefficients *coeffs, int xdim, int ydim, int nns) | 834 | static int allocate_model(PredictorCoefficients *coeffs, int xdim, int ydim, int nns) |
| 835 | { | 835 | { |
| 836 | size_t filter_size = nns * xdim * ydim; | 836 | int filter_size = nns * xdim * ydim; |
| 837 | size_t bias_size = nns; | 837 | int bias_size = nns; |
| 838 | float *data; | 838 | float *data; |
| 839 | 839 | ||
| 840 | data = av_malloc_array(filter_size + bias_size, 4 * sizeof(float)); | 840 | data = av_malloc_array(filter_size + bias_size, 4 * sizeof(float)); |
| @@ -864,25 +864,25 @@ static int read_weights(AVFilterContext *ctx, const float *bdata) · 23 unmodified lines | |||
| 864 | NNEDIContext *s = ctx->priv; | 864 | NNEDIContext *s = ctx->priv; |
| 865 | int ret; | 865 | int ret; |
| 866 | 866 | ||
| 867 | read(&s->prescreener_old.kernel_l0[0][0], 4 * 48, &bdata); | 867 | copy_weights(&s->prescreener_old.kernel_l0[0][0], 4 * 48, &bdata); |
| 868 | read(s->prescreener_old.bias_l0, 4, &bdata); | 868 | copy_weights(s->prescreener_old.bias_l0, 4, &bdata); |
| 869 | 869 | ||
| 870 | read(&s->prescreener_old.kernel_l1[0][0], 4 * 4, &bdata); | 870 | copy_weights(&s->prescreener_old.kernel_l1[0][0], 4 * 4, &bdata); |
| 871 | read(s->prescreener_old.bias_l1, 4, &bdata); | 871 | copy_weights(s->prescreener_old.bias_l1, 4, &bdata); |
| 872 | 872 | ||
| 873 | read(&s->prescreener_old.kernel_l2[0][0], 4 * 8, &bdata); | 873 | copy_weights(&s->prescreener_old.kernel_l2[0][0], 4 * 8, &bdata); |
| 874 | read(s->prescreener_old.bias_l2, 4, &bdata); | 874 | copy_weights(s->prescreener_old.bias_l2, 4, &bdata); |
| 875 | 875 | ||
| 876 | for (int i = 0; i < 3; i++) { | 876 | for (int i = 0; i < 3; i++) { |
| 877 | PrescreenerNewCoefficients *data = &s->prescreener_new[i]; | 877 | PrescreenerNewCoefficients *data = &s->prescreener_new[i]; |
| 878 | float kernel_l0_shuffled[4 * 64]; | 878 | float kernel_l0_shuffled[4 * 64]; |
| 879 | float kernel_l1_shuffled[4 * 4]; | 879 | float kernel_l1_shuffled[4 * 4]; |
| 880 | 880 | ||
| 881 | read(kernel_l0_shuffled, 4 * 64, &bdata); | 881 | copy_weights(kernel_l0_shuffled, 4 * 64, &bdata); |
| 882 | read(data->bias_l0, 4, &bdata); | 882 | copy_weights(data->bias_l0, 4, &bdata); |
| 883 | 883 | ||
| 884 | read(kernel_l1_shuffled, 4 * 4, &bdata); | 884 | copy_weights(kernel_l1_shuffled, 4 * 4, &bdata); |
| 885 | read(data->bias_l1, 4, &bdata); | 885 | copy_weights(data->bias_l1, 4, &bdata); |
| 886 | 886 | ||
| 887 | for (int n = 0; n < 4; n++) { | 887 | for (int n = 0; n < 4; n++) { |
| 888 | for (int k = 0; k < 64; k++) | 888 | for (int k = 0; k < 64; k++) |
| @@ -902,27 +902,27 @@ static int read_weights(AVFilterContext *ctx, const float *bdata) · 13 unmodified lines | |||
| 902 | PredictorCoefficients *model = &s->coeffs[m][i][j]; | 902 | PredictorCoefficients *model = &s->coeffs[m][i][j]; |
| 903 | int xdim = NNEDI_XDIM[j]; | 903 | int xdim = NNEDI_XDIM[j]; |
| 904 | int ydim = NNEDI_YDIM[j]; | 904 | int ydim = NNEDI_YDIM[j]; |
| 905 | size_t filter_size = xdim * ydim; | 905 | int filter_size = xdim * ydim; |
| 906 | 906 | ||
| 907 | ret = allocate_model(model, xdim, ydim, nns); | 907 | ret = allocate_model(model, xdim, ydim, nns); |
| 908 | if (ret < 0) | 908 | if (ret < 0) |
| 909 | return ret; | 909 | return ret; |
| 910 | 910 | ||
| 911 | // Quality 1 model. NNS[i] * (XDIM[j] * YDIM[j]) * 2 coefficients. | 911 | // Quality 1 model. NNS[i] * (XDIM[j] * YDIM[j]) * 2 coefficients. |
| 912 | read(model->softmax_q1, nns * filter_size, &bdata); | 912 | copy_weights(model->softmax_q1, nns * filter_size, &bdata); |
| 913 | read(model->elliott_q1, nns * filter_size, &bdata); | 913 | copy_weights(model->elliott_q1, nns * filter_size, &bdata); |
| 914 | 914 | ||
| 915 | // Quality 1 model bias. NNS[i] * 2 coefficients. | 915 | // Quality 1 model bias. NNS[i] * 2 coefficients. |
| 916 | read(model->softmax_bias_q1, nns, &bdata); | 916 | copy_weights(model->softmax_bias_q1, nns, &bdata); |
| 917 | read(model->elliott_bias_q1, nns, &bdata); | 917 | copy_weights(model->elliott_bias_q1, nns, &bdata); |
| 918 | 918 | ||
| 919 | // Quality 2 model. NNS[i] * (XDIM[j] * YDIM[j]) * 2 coefficients. | 919 | // Quality 2 model. NNS[i] * (XDIM[j] * YDIM[j]) * 2 coefficients. |
| 920 | read(model->softmax_q2, nns * filter_size, &bdata); | 920 | copy_weights(model->softmax_q2, nns * filter_size, &bdata); |
| 921 | read(model->elliott_q2, nns * filter_size, &bdata); | 921 | copy_weights(model->elliott_q2, nns * filter_size, &bdata); |
| 922 | 922 | ||
| 923 | // Quality 2 model bias. NNS[i] * 2 coefficients. | 923 | // Quality 2 model bias. NNS[i] * 2 coefficients. |
| 924 | read(model->softmax_bias_q2, nns, &bdata); | 924 | copy_weights(model->softmax_bias_q2, nns, &bdata); |
| 925 | read(model->elliott_bias_q2, nns, &bdata); | 925 | copy_weights(model->elliott_bias_q2, nns, &bdata); |
| 926 | } | 926 | } |
| 927 | } | 927 | } |
| 928 | } | 928 | } |
| @@ -966,7 +966,7 @@ static void subtract_mean_new(PrescreenerNewCoefficients *coeffs, float half) · 37 unmodified lines | |||
| 966 | 966 | ||
| 967 | static void subtract_mean_predictor(PredictorCoefficients *model) | 967 | static void subtract_mean_predictor(PredictorCoefficients *model) |
| 968 | { | 968 | { |
| 969 | size_t filter_size = model->xdim * model->ydim; | 969 | int filter_size = model->xdim * model->ydim; |
| 970 | int nns = model->nns; | 970 | int nns = model->nns; |
| 971 | 971 | ||
| 972 | float softmax_means[256]; // Average of individual softmax filters. | 972 | float softmax_means[256]; // Average of individual softmax filters. |