#include "dnn_backend_native_layer_conv2d.h"
#include "dnn_backend_native_layers.h"
-static DNNReturnType set_input_output_native(void *model, DNNInputData *input, const char *input_name, const char **output_names, uint32_t nb_output)
+#define OFFSET(x) offsetof(NativeContext, x)
+#define FLAGS AV_OPT_FLAG_FILTERING_PARAM
+static const AVOption dnn_native_options[] = {
+ { "conv2d_threads", "threads num for conv2d layer", OFFSET(options.conv2d_threads), AV_OPT_TYPE_INT, { .i64 = 0 }, INT_MIN, INT_MAX, FLAGS },
+ { NULL },
+};
+
+const AVClass dnn_native_class = {
+ .class_name = "dnn_native",
+ .item_name = av_default_item_name,
+ .option = dnn_native_options,
+ .version = LIBAVUTIL_VERSION_INT,
+ .category = AV_CLASS_CATEGORY_FILTER,
+};
+
+static DNNReturnType get_input_native(void *model, DNNData *input, const char *input_name)
{
- ConvolutionalNetwork *network = (ConvolutionalNetwork *)model;
+ NativeModel *native_model = (NativeModel *)model;
+ NativeContext *ctx = &native_model->ctx;
+
+ for (int i = 0; i < native_model->operands_num; ++i) {
+ DnnOperand *oprd = &native_model->operands[i];
+ if (strcmp(oprd->name, input_name) == 0) {
+ if (oprd->type != DOT_INPUT) {
+ av_log(ctx, AV_LOG_ERROR, "Found \"%s\" in model, but it is not input node\n", input_name);
+ return DNN_ERROR;
+ }
+ input->dt = oprd->data_type;
+ av_assert0(oprd->dims[0] == 1);
+ input->height = oprd->dims[1];
+ input->width = oprd->dims[2];
+ input->channels = oprd->dims[3];
+ return DNN_SUCCESS;
+ }
+ }
+
+ // do not find the input operand
+ av_log(ctx, AV_LOG_ERROR, "Could not find \"%s\" in model\n", input_name);
+ return DNN_ERROR;
+}
+
+static DNNReturnType set_input_native(void *model, DNNData *input, const char *input_name)
+{
+ NativeModel *native_model = (NativeModel *)model;
+ NativeContext *ctx = &native_model->ctx;
DnnOperand *oprd = NULL;
- if (network->layers_num <= 0 || network->operands_num <= 0)
+ if (native_model->layers_num <= 0 || native_model->operands_num <= 0) {
+ av_log(ctx, AV_LOG_ERROR, "No operands or layers in model\n");
return DNN_ERROR;
+ }
/* inputs */
- av_assert0(input->dt == DNN_FLOAT);
- for (int i = 0; i < network->operands_num; ++i) {
- oprd = &network->operands[i];
+ for (int i = 0; i < native_model->operands_num; ++i) {
+ oprd = &native_model->operands[i];
if (strcmp(oprd->name, input_name) == 0) {
- if (oprd->type != DOT_INPUT)
+ if (oprd->type != DOT_INPUT) {
+ av_log(ctx, AV_LOG_ERROR, "Found \"%s\" in model, but it is not input node\n", input_name);
return DNN_ERROR;
+ }
break;
}
oprd = NULL;
}
-
- if (!oprd)
+ if (!oprd) {
+ av_log(ctx, AV_LOG_ERROR, "Could not find \"%s\" in model\n", input_name);
return DNN_ERROR;
+ }
oprd->dims[0] = 1;
oprd->dims[1] = input->height;
av_freep(&oprd->data);
oprd->length = calculate_operand_data_length(oprd);
- oprd->data = av_malloc(oprd->length);
- if (!oprd->data)
+ if (oprd->length <= 0) {
+ av_log(ctx, AV_LOG_ERROR, "The input data length overflow\n");
return DNN_ERROR;
-
- input->data = oprd->data;
-
- /* outputs */
- network->nb_output = 0;
- av_freep(&network->output_indexes);
- network->output_indexes = av_mallocz_array(nb_output, sizeof(*network->output_indexes));
- if (!network->output_indexes)
+ }
+ oprd->data = av_malloc(oprd->length);
+ if (!oprd->data) {
+ av_log(ctx, AV_LOG_ERROR, "Failed to malloc memory for input data\n");
return DNN_ERROR;
-
- for (uint32_t i = 0; i < nb_output; ++i) {
- const char *output_name = output_names[i];
- for (int j = 0; j < network->operands_num; ++j) {
- oprd = &network->operands[j];
- if (strcmp(oprd->name, output_name) == 0) {
- network->output_indexes[network->nb_output++] = j;
- break;
- }
- }
}
- if (network->nb_output != nb_output)
- return DNN_ERROR;
+ input->data = oprd->data;
return DNN_SUCCESS;
}
// layers_num,layer_type,layer_parameterss,layer_type,layer_parameters...
// For CONV layer: activation_function, input_num, output_num, kernel_size, kernel, biases
// For DEPTH_TO_SPACE layer: block_size
-DNNModel *ff_dnn_load_model_native(const char *model_filename)
+DNNModel *ff_dnn_load_model_native(const char *model_filename, const char *options, void *userdata)
{
DNNModel *model = NULL;
char header_expected[] = "FFMPEGDNNNATIVE";
char *buf;
size_t size;
int version, header_size, major_version_expected = 1;
- ConvolutionalNetwork *network = NULL;
+ NativeModel *native_model = NULL;
AVIOContext *model_file_context;
int file_size, dnn_size, parsed_size;
int32_t layer;
DNNLayerType layer_type;
- model = av_malloc(sizeof(DNNModel));
- if (!model){
- return NULL;
- }
-
if (avio_open(&model_file_context, model_filename, AVIO_FLAG_READ) < 0){
- av_freep(&model);
return NULL;
}
file_size = avio_size(model_file_context);
+ model = av_mallocz(sizeof(DNNModel));
+ if (!model){
+ goto fail;
+ }
+
/**
* check file header with string and version
*/
size = sizeof(header_expected);
buf = av_malloc(size);
if (!buf) {
- avio_closep(&model_file_context);
- av_freep(&model);
- return NULL;
+ goto fail;
}
// size - 1 to skip the ending '\0' which is not saved in file
dnn_size = size - 1;
if (strncmp(buf, header_expected, size) != 0) {
av_freep(&buf);
- avio_closep(&model_file_context);
- av_freep(&model);
- return NULL;
+ goto fail;
}
av_freep(&buf);
version = (int32_t)avio_rl32(model_file_context);
dnn_size += 4;
if (version != major_version_expected) {
- avio_closep(&model_file_context);
- av_freep(&model);
- return NULL;
+ goto fail;
}
// currently no need to check minor version
dnn_size += 4;
header_size = dnn_size;
- network = av_mallocz(sizeof(ConvolutionalNetwork));
- if (!network){
- avio_closep(&model_file_context);
- av_freep(&model);
- return NULL;
+ native_model = av_mallocz(sizeof(NativeModel));
+ if (!native_model){
+ goto fail;
+ }
+
+ native_model->ctx.class = &dnn_native_class;
+ model->options = options;
+ if (av_opt_set_from_string(&native_model->ctx, model->options, NULL, "=", "&") < 0)
+ goto fail;
+ model->model = (void *)native_model;
+
+#if !HAVE_PTHREAD_CANCEL
+ if (native_model->ctx.options.conv2d_threads > 1){
+ av_log(&native_model->ctx, AV_LOG_WARNING, "'conv2d_threads' option was set but it is not supported "
+ "on this build (pthread support is required)\n");
}
- model->model = (void *)network;
+#endif
avio_seek(model_file_context, file_size - 8, SEEK_SET);
- network->layers_num = (int32_t)avio_rl32(model_file_context);
- network->operands_num = (int32_t)avio_rl32(model_file_context);
+ native_model->layers_num = (int32_t)avio_rl32(model_file_context);
+ native_model->operands_num = (int32_t)avio_rl32(model_file_context);
dnn_size += 8;
avio_seek(model_file_context, header_size, SEEK_SET);
- network->layers = av_mallocz(network->layers_num * sizeof(Layer));
- if (!network->layers){
- avio_closep(&model_file_context);
- ff_dnn_free_model_native(&model);
- return NULL;
+ native_model->layers = av_mallocz(native_model->layers_num * sizeof(Layer));
+ if (!native_model->layers){
+ goto fail;
}
- network->operands = av_mallocz(network->operands_num * sizeof(DnnOperand));
- if (!network->operands){
- avio_closep(&model_file_context);
- ff_dnn_free_model_native(&model);
- return NULL;
+ native_model->operands = av_mallocz(native_model->operands_num * sizeof(DnnOperand));
+ if (!native_model->operands){
+ goto fail;
}
- for (layer = 0; layer < network->layers_num; ++layer){
+ for (layer = 0; layer < native_model->layers_num; ++layer){
layer_type = (int32_t)avio_rl32(model_file_context);
dnn_size += 4;
if (layer_type >= DLT_COUNT) {
- avio_closep(&model_file_context);
- ff_dnn_free_model_native(&model);
- return NULL;
+ goto fail;
}
- network->layers[layer].type = layer_type;
- parsed_size = layer_funcs[layer_type].pf_load(&network->layers[layer], model_file_context, file_size);
+ native_model->layers[layer].type = layer_type;
+ parsed_size = layer_funcs[layer_type].pf_load(&native_model->layers[layer], model_file_context, file_size, native_model->operands_num);
if (!parsed_size) {
- avio_closep(&model_file_context);
- ff_dnn_free_model_native(&model);
- return NULL;
+ goto fail;
}
dnn_size += parsed_size;
}
- for (int32_t i = 0; i < network->operands_num; ++i){
+ for (int32_t i = 0; i < native_model->operands_num; ++i){
DnnOperand *oprd;
int32_t name_len;
int32_t operand_index = (int32_t)avio_rl32(model_file_context);
dnn_size += 4;
- oprd = &network->operands[operand_index];
+ if (operand_index >= native_model->operands_num) {
+ goto fail;
+ }
+
+ oprd = &native_model->operands[operand_index];
name_len = (int32_t)avio_rl32(model_file_context);
dnn_size += 4;
return NULL;
}
- model->set_input_output = &set_input_output_native;
+ model->set_input = &set_input_native;
+ model->get_input = &get_input_native;
+ model->userdata = userdata;
return model;
+
+fail:
+ ff_dnn_free_model_native(&model);
+ avio_closep(&model_file_context);
+ return NULL;
}
-DNNReturnType ff_dnn_execute_model_native(const DNNModel *model, DNNData *outputs, uint32_t nb_output)
+DNNReturnType ff_dnn_execute_model_native(const DNNModel *model, DNNData *outputs, const char **output_names, uint32_t nb_output)
{
- ConvolutionalNetwork *network = (ConvolutionalNetwork *)model->model;
+ NativeModel *native_model = (NativeModel *)model->model;
+ NativeContext *ctx = &native_model->ctx;
int32_t layer;
- uint32_t nb = FFMIN(nb_output, network->nb_output);
- if (network->layers_num <= 0 || network->operands_num <= 0)
+ if (native_model->layers_num <= 0 || native_model->operands_num <= 0) {
+ av_log(ctx, AV_LOG_ERROR, "No operands or layers in model\n");
return DNN_ERROR;
- if (!network->operands[0].data)
+ }
+ if (!native_model->operands[0].data) {
+ av_log(ctx, AV_LOG_ERROR, "Empty model input data\n");
return DNN_ERROR;
+ }
- for (layer = 0; layer < network->layers_num; ++layer){
- DNNLayerType layer_type = network->layers[layer].type;
- layer_funcs[layer_type].pf_exec(network->operands,
- network->layers[layer].input_operand_indexes,
- network->layers[layer].output_operand_index,
- network->layers[layer].params);
+ for (layer = 0; layer < native_model->layers_num; ++layer){
+ DNNLayerType layer_type = native_model->layers[layer].type;
+ if (layer_funcs[layer_type].pf_exec(native_model->operands,
+ native_model->layers[layer].input_operand_indexes,
+ native_model->layers[layer].output_operand_index,
+ native_model->layers[layer].params,
+ &native_model->ctx) == DNN_ERROR) {
+ av_log(ctx, AV_LOG_ERROR, "Failed to execuet model\n");
+ return DNN_ERROR;
+ }
}
- for (uint32_t i = 0; i < nb; ++i) {
- DnnOperand *oprd = &network->operands[network->output_indexes[i]];
+ for (uint32_t i = 0; i < nb_output; ++i) {
+ DnnOperand *oprd = NULL;
+ const char *output_name = output_names[i];
+ for (int j = 0; j < native_model->operands_num; ++j) {
+ if (strcmp(native_model->operands[j].name, output_name) == 0) {
+ oprd = &native_model->operands[j];
+ break;
+ }
+ }
+
+ if (oprd == NULL) {
+ av_log(ctx, AV_LOG_ERROR, "Could not find output in model\n");
+ return DNN_ERROR;
+ }
+
outputs[i].data = oprd->data;
outputs[i].height = oprd->dims[1];
outputs[i].width = oprd->dims[2];
outputs[i].channels = oprd->dims[3];
+ outputs[i].dt = oprd->data_type;
}
return DNN_SUCCESS;
int32_t calculate_operand_data_length(const DnnOperand* oprd)
{
// currently, we just support DNN_FLOAT
- return oprd->dims[0] * oprd->dims[1] * oprd->dims[2] * oprd->dims[3] * sizeof(float);
+ uint64_t len = sizeof(float);
+ for (int i = 0; i < 4; i++) {
+ len *= oprd->dims[i];
+ if (len > INT32_MAX)
+ return 0;
+ }
+ return len;
}
void ff_dnn_free_model_native(DNNModel **model)
{
- ConvolutionalNetwork *network;
+ NativeModel *native_model;
ConvolutionalParams *conv_params;
int32_t layer;
if (*model)
{
- network = (ConvolutionalNetwork *)(*model)->model;
- for (layer = 0; layer < network->layers_num; ++layer){
- if (network->layers[layer].type == DLT_CONV2D){
- conv_params = (ConvolutionalParams *)network->layers[layer].params;
- av_freep(&conv_params->kernel);
- av_freep(&conv_params->biases);
+ if ((*model)->model) {
+ native_model = (NativeModel *)(*model)->model;
+ if (native_model->layers) {
+ for (layer = 0; layer < native_model->layers_num; ++layer){
+ if (native_model->layers[layer].type == DLT_CONV2D){
+ conv_params = (ConvolutionalParams *)native_model->layers[layer].params;
+ av_freep(&conv_params->kernel);
+ av_freep(&conv_params->biases);
+ }
+ av_freep(&native_model->layers[layer].params);
+ }
+ av_freep(&native_model->layers);
}
- av_freep(&network->layers[layer].params);
- }
- av_freep(&network->layers);
- for (uint32_t operand = 0; operand < network->operands_num; ++operand)
- av_freep(&network->operands[operand].data);
- av_freep(&network->operands);
+ if (native_model->operands) {
+ for (uint32_t operand = 0; operand < native_model->operands_num; ++operand)
+ av_freep(&native_model->operands[operand].data);
+ av_freep(&native_model->operands);
+ }
- av_freep(&network->output_indexes);
- av_freep(&network);
+ av_freep(&native_model);
+ }
av_freep(model);
}
}