上传完整项目文件

This commit is contained in:
xhgit committed 2026-06-21 14:26:15 +08:00
1 parent e63d58df8d
commit aa0dd45839
8663 files changed
+2393125

No files matched your search

+11
View File
@@ -0,0 +1,11 @@
OBJS-$(CONFIG_DNN) += dnn/dnn_interface.o
OBJS-$(CONFIG_DNN) += dnn/dnn_io_proc.o
OBJS-$(CONFIG_DNN) += dnn/queue.o
OBJS-$(CONFIG_DNN) += dnn/safe_queue.o
OBJS-$(CONFIG_DNN) += dnn/dnn_backend_common.o
DNN-OBJS-$(CONFIG_LIBTENSORFLOW) += dnn/dnn_backend_tf.o
DNN-OBJS-$(CONFIG_LIBOPENVINO) += dnn/dnn_backend_openvino.o
DNN-OBJS-$(CONFIG_LIBTORCH) += dnn/dnn_backend_torch.o
OBJS-$(CONFIG_DNN) += $(DNN-OBJS-yes)
+180
View File
@@ -0,0 +1,180 @@
/*
* This file is part of FFmpeg.
*
* FFmpeg is free software; you can redistribute it and/or
* modify it under the terms of the GNU Lesser General Public
* License as published by the Free Software Foundation; either
* version 2.1 of the License, or (at your option) any later version.
*
* FFmpeg is distributed in the hope that it will be useful,
* but WITHOUT ANY WARRANTY; without even the implied warranty of
* MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the GNU
* Lesser General Public License for more details.
*
* You should have received a copy of the GNU Lesser General Public
* License along with FFmpeg; if not, write to the Free Software
* Foundation, Inc., 51 Franklin Street, Fifth Floor, Boston, MA 02110-1301 USA
*/
/**
* @file
* DNN common functions different backends.
*/
#include "libavutil/mem.h"
#include "dnn_backend_common.h"
#define DNN_ASYNC_SUCCESS (void *)0
#define DNN_ASYNC_FAIL (void *)-1
int ff_check_exec_params(void *ctx, DNNBackendType backend, DNNFunctionType func_type, DNNExecBaseParams *exec_params)
{
if (!exec_params) {
av_log(ctx, AV_LOG_ERROR, "exec_params is null when execute model.\n");
return AVERROR(EINVAL);
}
if (!exec_params->in_frame) {
av_log(ctx, AV_LOG_ERROR, "in frame is NULL when execute model.\n");
return AVERROR(EINVAL);
}
if (!exec_params->out_frame && func_type == DFT_PROCESS_FRAME) {
av_log(ctx, AV_LOG_ERROR, "out frame is NULL when execute model.\n");
return AVERROR(EINVAL);
}
return 0;
}
int ff_dnn_fill_task(TaskItem *task, DNNExecBaseParams *exec_params, void *backend_model, int async, int do_ioproc) {
if (task == NULL || exec_params == NULL || backend_model == NULL)
return AVERROR(EINVAL);
if (do_ioproc != 0 && do_ioproc != 1)
return AVERROR(EINVAL);
if (async != 0 && async != 1)
return AVERROR(EINVAL);
task->do_ioproc = do_ioproc;
task->async = async;
task->input_name = exec_params->input_name;
task->in_frame = exec_params->in_frame;
task->out_frame = exec_params->out_frame;
task->model = backend_model;
task->nb_output = exec_params->nb_output;
task->output_names = exec_params->output_names;
return 0;
}
/**
* Thread routine for async execution.
* @param args pointer to DNNAsyncExecModule module
*/
static void *async_thread_routine(void *args)
{
DNNAsyncExecModule *async_module = args;
void *request = async_module->args;
if (async_module->start_inference(request) != 0) {
return DNN_ASYNC_FAIL;
}
async_module->callback(request);
return DNN_ASYNC_SUCCESS;
}
int ff_dnn_async_module_cleanup(DNNAsyncExecModule *async_module)
{
void *status = 0;
if (!async_module) {
return AVERROR(EINVAL);
}
#if HAVE_PTHREAD_CANCEL
pthread_join(async_module->thread_id, &status);
if (status == DNN_ASYNC_FAIL) {
av_log(NULL, AV_LOG_ERROR, "Last Inference Failed.\n");
return DNN_GENERIC_ERROR;
}
#endif
async_module->start_inference = NULL;
async_module->callback = NULL;
async_module->args = NULL;
return 0;
}
int ff_dnn_start_inference_async(void *ctx, DNNAsyncExecModule *async_module)
{
int ret;
void *status = 0;
if (!async_module) {
av_log(ctx, AV_LOG_ERROR, "async_module is null when starting async inference.\n");
return AVERROR(EINVAL);
}
#if HAVE_PTHREAD_CANCEL
pthread_join(async_module->thread_id, &status);
if (status == DNN_ASYNC_FAIL) {
av_log(ctx, AV_LOG_ERROR, "Unable to start inference as previous inference failed.\n");
return DNN_GENERIC_ERROR;
}
ret = pthread_create(&async_module->thread_id, NULL, async_thread_routine, async_module);
if (ret != 0) {
av_log(ctx, AV_LOG_ERROR, "Unable to start async inference.\n");
return ret;
}
#else
ret = async_module->start_inference(async_module->args);
if (ret != 0) {
return ret;
}
async_module->callback(async_module->args);
#endif
return 0;
}
DNNAsyncStatusType ff_dnn_get_result_common(Queue *task_queue, AVFrame **in, AVFrame **out)
{
TaskItem *task = ff_queue_peek_front(task_queue);
if (!task) {
return DAST_EMPTY_QUEUE;
}
if (task->inference_done != task->inference_todo) {
return DAST_NOT_READY;
}
*in = task->in_frame;
*out = task->out_frame;
ff_queue_pop_front(task_queue);
av_freep(&task);
return DAST_SUCCESS;
}
int ff_dnn_fill_gettingoutput_task(TaskItem *task, DNNExecBaseParams *exec_params, void *backend_model, int input_height, int input_width, void *ctx)
{
AVFrame *in_frame = NULL;
AVFrame *out_frame = NULL;
in_frame = av_frame_alloc();
if (!in_frame) {
av_log(ctx, AV_LOG_ERROR, "Failed to allocate memory for input frame\n");
return AVERROR(ENOMEM);
}
out_frame = av_frame_alloc();
if (!out_frame) {
av_frame_free(&in_frame);
av_log(ctx, AV_LOG_ERROR, "Failed to allocate memory for output frame\n");
return AVERROR(ENOMEM);
}
in_frame->width = input_width;
in_frame->height = input_height;
exec_params->in_frame = in_frame;
exec_params->out_frame = out_frame;
return ff_dnn_fill_task(task, exec_params, backend_model, 0, 0);
}
+160
View File
@@ -0,0 +1,160 @@
/*
* This file is part of FFmpeg.
*
* FFmpeg is free software; you can redistribute it and/or
* modify it under the terms of the GNU Lesser General Public
* License as published by the Free Software Foundation; either
* version 2.1 of the License, or (at your option) any later version.
*
* FFmpeg is distributed in the hope that it will be useful,
* but WITHOUT ANY WARRANTY; without even the implied warranty of
* MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the GNU
* Lesser General Public License for more details.
*
* You should have received a copy of the GNU Lesser General Public
* License along with FFmpeg; if not, write to the Free Software
* Foundation, Inc., 51 Franklin Street, Fifth Floor, Boston, MA 02110-1301 USA
*/
/**
* @file
* DNN common functions different backends.
*/
#ifndef AVFILTER_DNN_DNN_BACKEND_COMMON_H
#define AVFILTER_DNN_DNN_BACKEND_COMMON_H
#include "queue.h"
#include "../dnn_interface.h"
#include "libavutil/thread.h"
#define DNN_DEFINE_CLASS_EXT(name, desc, options) \
{ \
.class_name = desc, \
.item_name = av_default_item_name, \
.option = options, \
.version = LIBAVUTIL_VERSION_INT, \
.category = AV_CLASS_CATEGORY_FILTER, \
}
#define DNN_DEFINE_CLASS(fname) \
DNN_DEFINE_CLASS_EXT(fname, #fname, fname##_options)
// one task for one function call from dnn interface
typedef struct TaskItem {
void *model; // model for the backend
AVFrame *in_frame;
AVFrame *out_frame;
const char *input_name;
const char **output_names;
uint8_t async;
uint8_t do_ioproc;
uint32_t nb_output;
uint32_t inference_todo;
uint32_t inference_done;
} TaskItem;
// one task might have multiple inferences
typedef struct LastLevelTaskItem {
TaskItem *task;
uint32_t bbox_index;
} LastLevelTaskItem;
/**
* Common Async Execution Mechanism for the DNN Backends.
*/
typedef struct DNNAsyncExecModule {
/**
* Synchronous inference function for the backend
* with corresponding request item as the argument.
*/
int (*start_inference)(void *request);
/**
* Completion Callback for the backend.
* Expected argument type of callback must match that
* of the inference function.
*/
void (*callback)(void *args);
/**
* Argument for the execution functions.
* i.e. Request item for the backend.
*/
void *args;
#if HAVE_PTHREAD_CANCEL
pthread_t thread_id;
pthread_attr_t thread_attr;
#endif
} DNNAsyncExecModule;
int ff_check_exec_params(void *ctx, DNNBackendType backend, DNNFunctionType func_type, DNNExecBaseParams *exec_params);
/**
* Fill the Task for Backend Execution. It should be called after
* checking execution parameters using ff_check_exec_params.
*
* @param task pointer to the allocated task
* @param exec_param pointer to execution parameters
* @param backend_model void pointer to the backend model
* @param async flag for async execution. Must be 0 or 1
* @param do_ioproc flag for IO processing. Must be 0 or 1
*
* @returns 0 if successful or error code otherwise.
*/
int ff_dnn_fill_task(TaskItem *task, DNNExecBaseParams *exec_params, void *backend_model, int async, int do_ioproc);
/**
* Join the Async Execution thread and set module pointers to NULL.
*
* @param async_module pointer to DNNAsyncExecModule module
*
* @returns 0 if successful or error code otherwise.
*/
int ff_dnn_async_module_cleanup(DNNAsyncExecModule *async_module);
/**
* Start asynchronous inference routine for the TensorFlow
* model on a detached thread. It calls the completion callback
* after the inference completes. Completion callback and inference
* function must be set before calling this function.
*
* If POSIX threads aren't supported, the execution rolls back
* to synchronous mode, calling completion callback after inference.
*
* @param ctx pointer to the backend context
* @param async_module pointer to DNNAsyncExecModule module
*
* @returns 0 on the start of async inference or error code otherwise.
*/
int ff_dnn_start_inference_async(void *ctx, DNNAsyncExecModule *async_module);
/**
* Extract input and output frame from the Task Queue after
* asynchronous inference.
*
* @param task_queue pointer to the task queue of the backend
* @param in double pointer to the input frame
* @param out double pointer to the output frame
*
* @retval DAST_EMPTY_QUEUE if task queue is empty
* @retval DAST_NOT_READY if inference not completed yet.
* @retval DAST_SUCCESS if result successfully extracted
*/
DNNAsyncStatusType ff_dnn_get_result_common(Queue *task_queue, AVFrame **in, AVFrame **out);
/**
* Allocate input and output frames and fill the Task
* with execution parameters.
*
* @param task pointer to the allocated task
* @param exec_params pointer to execution parameters
* @param backend_model void pointer to the backend model
* @param input_height height of input frame
* @param input_width width of input frame
* @param ctx pointer to the backend context
*
* @returns 0 if successful or error code otherwise.
*/
int ff_dnn_fill_gettingoutput_task(TaskItem *task, DNNExecBaseParams *exec_params, void *backend_model, int input_height, int input_width, void *ctx);
#endif
File diff suppressed because it is too large. Load diff
+892
View File
@@ -0,0 +1,892 @@
/*
* Copyright (c) 2018 Sergey Lavrushkin
*
* This file is part of FFmpeg.
*
* FFmpeg is free software; you can redistribute it and/or
* modify it under the terms of the GNU Lesser General Public
* License as published by the Free Software Foundation; either
* version 2.1 of the License, or (at your option) any later version.
*
* FFmpeg is distributed in the hope that it will be useful,
* but WITHOUT ANY WARRANTY; without even the implied warranty of
* MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the GNU
* Lesser General Public License for more details.
*
* You should have received a copy of the GNU Lesser General Public
* License along with FFmpeg; if not, write to the Free Software
* Foundation, Inc., 51 Franklin Street, Fifth Floor, Boston, MA 02110-1301 USA
*/
/**
* @file
* DNN tensorflow backend implementation.
*/
#include "libavformat/avio.h"
#include "libavutil/avassert.h"
#include "libavutil/avstring.h"
#include "libavutil/cpu.h"
#include "libavutil/mem.h"
#include "libavutil/opt.h"
#include "libavcodec/defs.h"
#include "dnn_io_proc.h"
#include "dnn_backend_common.h"
#include "safe_queue.h"
#include <tensorflow/c/c_api.h>
typedef struct TFModel {
DNNModel model;
DnnContext *ctx;
TF_Graph *graph;
TF_Session *session;
TF_Status *status;
SafeQueue *request_queue;
Queue *lltask_queue;
Queue *task_queue;
} TFModel;
/**
* Stores execution parameters for single
* call to the TensorFlow C API
*/
typedef struct TFInferRequest {
TF_Output *tf_outputs;
TF_Tensor **output_tensors;
TF_Output *tf_input;
TF_Tensor *input_tensor;
} TFInferRequest;
typedef struct TFRequestItem {
TFInferRequest *infer_request;
LastLevelTaskItem *lltask;
TF_Status *status;
DNNAsyncExecModule exec_module;
} TFRequestItem;
#define OFFSET(x) offsetof(TFOptions, x)
#define FLAGS AV_OPT_FLAG_FILTERING_PARAM
static const AVOption dnn_tensorflow_options[] = {
{ "sess_config", "config for SessionOptions", OFFSET(sess_config), AV_OPT_TYPE_STRING, { .str = NULL }, 0, 0, FLAGS },
{ NULL }
};
static int execute_model_tf(TFRequestItem *request, Queue *lltask_queue);
static void infer_completion_callback(void *args);
static inline void destroy_request_item(TFRequestItem **arg);
static void free_buffer(void *data, size_t length)
{
av_freep(&data);
}
/**
* Free the contents of TensorFlow inference request.
* It does not free the TFInferRequest instance.
*
* @param request pointer to TFInferRequest instance.
* NULL pointer is allowed.
*/
static void tf_free_request(TFInferRequest *request)
{
if (!request)
return;
if (request->input_tensor) {
TF_DeleteTensor(request->input_tensor);
request->input_tensor = NULL;
}
av_freep(&request->tf_input);
av_freep(&request->tf_outputs);
if (request->output_tensors) {
int nb_output = sizeof(*request->output_tensors)/sizeof(request->output_tensors[0]);
for (uint32_t i = 0; i < nb_output; ++i) {
if (request->output_tensors[i]) {
TF_DeleteTensor(request->output_tensors[i]);
request->output_tensors[i] = NULL;
}
}
av_freep(&request->output_tensors);
}
}
/**
* Create a TensorFlow inference request. All properties
* are initially unallocated and set as NULL.
*
* @return pointer to the allocated TFInferRequest instance.
*/
static TFInferRequest *tf_create_inference_request(void)
{
TFInferRequest *infer_request = av_malloc(sizeof(TFInferRequest));
if (!infer_request) {
return NULL;
}
infer_request->tf_outputs = NULL;
infer_request->tf_input = NULL;
infer_request->input_tensor = NULL;
infer_request->output_tensors = NULL;
return infer_request;
}
/**
* Start synchronous inference for the TensorFlow model.
*
* @param request pointer to the TFRequestItem for inference
* @retval 0 if execution is successful
* @retval AVERROR(EINVAL) if request is NULL
* @retval DNN_GENERIC_ERROR if execution fails
*/
static int tf_start_inference(void *args)
{
TFRequestItem *request = args;
TFInferRequest *infer_request = request->infer_request;
LastLevelTaskItem *lltask = request->lltask;
TaskItem *task = lltask->task;
TFModel *tf_model = task->model;
if (!request) {
av_log(tf_model->ctx, AV_LOG_ERROR, "TFRequestItem is NULL\n");
return AVERROR(EINVAL);
}
TF_SessionRun(tf_model->session, NULL,
infer_request->tf_input, &infer_request->input_tensor, 1,
infer_request->tf_outputs, infer_request->output_tensors,
task->nb_output, NULL, 0, NULL,
request->status);
if (TF_GetCode(request->status) != TF_OK) {
av_log(tf_model->ctx, AV_LOG_ERROR, "%s", TF_Message(request->status));
return DNN_GENERIC_ERROR;
}
return 0;
}
/**
* Free the TFRequestItem completely.
*
* @param arg Address of the TFInferRequest instance.
*/
static inline void destroy_request_item(TFRequestItem **arg) {
TFRequestItem *request;
if (!arg) {
return;
}
request = *arg;
tf_free_request(request->infer_request);
av_freep(&request->infer_request);
av_freep(&request->lltask);
TF_DeleteStatus(request->status);
ff_dnn_async_module_cleanup(&request->exec_module);
av_freep(arg);
}
static int extract_lltask_from_task(TaskItem *task, Queue *lltask_queue)
{
TFModel *tf_model = task->model;
DnnContext *ctx = tf_model->ctx;
LastLevelTaskItem *lltask = av_malloc(sizeof(*lltask));
if (!lltask) {
av_log(ctx, AV_LOG_ERROR, "Unable to allocate space for LastLevelTaskItem\n");
return AVERROR(ENOMEM);
}
task->inference_todo = 1;
task->inference_done = 0;
lltask->task = task;
if (ff_queue_push_back(lltask_queue, lltask) < 0) {
av_log(ctx, AV_LOG_ERROR, "Failed to push back lltask_queue.\n");
av_freep(&lltask);
return AVERROR(ENOMEM);
}
return 0;
}
static TF_Buffer *read_graph(const char *model_filename)
{
TF_Buffer *graph_buf;
unsigned char *graph_data = NULL;
AVIOContext *model_file_context;
long size, bytes_read;
if (avio_open(&model_file_context, model_filename, AVIO_FLAG_READ) < 0){
return NULL;
}
size = avio_size(model_file_context);
graph_data = av_malloc(size);
if (!graph_data){
avio_closep(&model_file_context);
return NULL;
}
bytes_read = avio_read(model_file_context, graph_data, size);
avio_closep(&model_file_context);
if (bytes_read != size){
av_freep(&graph_data);
return NULL;
}
graph_buf = TF_NewBuffer();
graph_buf->data = graph_data;
graph_buf->length = size;
graph_buf->data_deallocator = free_buffer;
return graph_buf;
}
static TF_Tensor *allocate_input_tensor(const DNNData *input)
{
TF_DataType dt;
size_t size;
int64_t input_dims[4] = { 0 };
input_dims[0] = 1;
input_dims[1] = input->dims[dnn_get_height_idx_by_layout(input->layout)];
input_dims[2] = input->dims[dnn_get_width_idx_by_layout(input->layout)];
input_dims[3] = input->dims[dnn_get_channel_idx_by_layout(input->layout)];
switch (input->dt) {
case DNN_FLOAT:
dt = TF_FLOAT;
size = sizeof(float);
break;
case DNN_UINT8:
dt = TF_UINT8;
size = 1;
break;
default:
av_assert0(!"should not reach here");
}
return TF_AllocateTensor(dt, input_dims, 4,
input_dims[1] * input_dims[2] * input_dims[3] * size);
}
static int get_input_tf(DNNModel *model, DNNData *input, const char *input_name)
{
TFModel *tf_model = (TFModel *)model;
DnnContext *ctx = tf_model->ctx;
TF_Status *status;
TF_DataType dt;
int64_t dims[4];
TF_Output tf_output;
tf_output.oper = TF_GraphOperationByName(tf_model->graph, input_name);
if (!tf_output.oper) {
av_log(ctx, AV_LOG_ERROR, "Could not find \"%s\" in model\n", input_name);
return AVERROR(EINVAL);
}
tf_output.index = 0;
dt = TF_OperationOutputType(tf_output);
switch (dt) {
case TF_FLOAT:
input->dt = DNN_FLOAT;
break;
case TF_UINT8:
input->dt = DNN_UINT8;
break;
default:
av_log(ctx, AV_LOG_ERROR, "Unsupported output type %d in model\n", dt);
return AVERROR(EINVAL);
}
input->order = DCO_RGB;
status = TF_NewStatus();
TF_GraphGetTensorShape(tf_model->graph, tf_output, dims, 4, status);
if (TF_GetCode(status) != TF_OK){
TF_DeleteStatus(status);
av_log(ctx, AV_LOG_ERROR, "Failed to get input tensor shape: number of dimension incorrect\n");
return DNN_GENERIC_ERROR;
}
TF_DeleteStatus(status);
// currently only NHWC is supported
av_assert0(dims[0] == 1 || dims[0] == -1);
for (int i = 0; i < 4; i++)
input->dims[i] = dims[i];
input->layout = DL_NHWC;
return 0;
}
static int get_output_tf(DNNModel *model, const char *input_name, int input_width, int input_height,
const char *output_name, int *output_width, int *output_height)
{
int ret;
TFModel *tf_model = (TFModel *)model;
DnnContext *ctx = tf_model->ctx;
TaskItem task;
TFRequestItem *request;
DNNExecBaseParams exec_params = {
.input_name = input_name,
.output_names = &output_name,
.nb_output = 1,
.in_frame = NULL,
.out_frame = NULL,
};
ret = ff_dnn_fill_gettingoutput_task(&task, &exec_params, tf_model, input_height, input_width, ctx);
if (ret != 0) {
goto err;
}
ret = extract_lltask_from_task(&task, tf_model->lltask_queue);
if (ret != 0) {
av_log(ctx, AV_LOG_ERROR, "unable to extract inference from task.\n");
goto err;
}
request = ff_safe_queue_pop_front(tf_model->request_queue);
if (!request) {
av_log(ctx, AV_LOG_ERROR, "unable to get infer request.\n");
ret = AVERROR(EINVAL);
goto err;
}
ret = execute_model_tf(request, tf_model->lltask_queue);
*output_width = task.out_frame->width;
*output_height = task.out_frame->height;
err:
av_frame_free(&task.out_frame);
av_frame_free(&task.in_frame);
return ret;
}
#define SPACE_CHARS " \t\r\n"
static int hex_to_data(uint8_t *data, const char *p)
{
int c, len, v;
len = 0;
v = 1;
for (;;) {
p += strspn(p, SPACE_CHARS);
if (*p == '\0')
break;
c = av_toupper((unsigned char) *p++);
if (c >= '0' && c <= '9')
c = c - '0';
else if (c >= 'A' && c <= 'F')
c = c - 'A' + 10;
else
break;
v = (v << 4) | c;
if (v & 0x100) {
if (data) {
data[len] = v;
}
len++;
v = 1;
}
}
return len;
}
static int load_tf_model(TFModel *tf_model, const char *model_filename)
{
DnnContext *ctx = tf_model->ctx;
TF_Buffer *graph_def;
TF_ImportGraphDefOptions *graph_opts;
TF_SessionOptions *sess_opts;
const TF_Operation *init_op;
uint8_t *sess_config = NULL;
int sess_config_length = 0;
// prepare the sess config data
if (ctx->tf_option.sess_config != NULL) {
const char *config;
/*
tf_model->ctx.options.sess_config is hex to present the serialized proto
required by TF_SetConfig below, so we need to first generate the serialized
proto in a python script, tools/python/tf_sess_config.py is a script example
to generate the configs of sess_config.
*/
if (strncmp(ctx->tf_option.sess_config, "0x", 2) != 0) {
av_log(ctx, AV_LOG_ERROR, "sess_config should start with '0x'\n");
return AVERROR(EINVAL);
}
config = ctx->tf_option.sess_config + 2;
sess_config_length = hex_to_data(NULL, config);
sess_config = av_mallocz(sess_config_length + AV_INPUT_BUFFER_PADDING_SIZE);
if (!sess_config) {
av_log(ctx, AV_LOG_ERROR, "failed to allocate memory\n");
return AVERROR(ENOMEM);
}
if (hex_to_data(sess_config, config) < 0) {
av_log(ctx, AV_LOG_ERROR, "failed to convert hex to data\n");
return AVERROR(EINVAL);
}
}
graph_def = read_graph(model_filename);
if (!graph_def){
av_log(ctx, AV_LOG_ERROR, "Failed to read model \"%s\" graph\n", model_filename);
av_freep(&sess_config);
return AVERROR(EINVAL);
}
tf_model->graph = TF_NewGraph();
tf_model->status = TF_NewStatus();
graph_opts = TF_NewImportGraphDefOptions();
TF_GraphImportGraphDef(tf_model->graph, graph_def, graph_opts, tf_model->status);
TF_DeleteImportGraphDefOptions(graph_opts);
TF_DeleteBuffer(graph_def);
if (TF_GetCode(tf_model->status) != TF_OK){
av_log(ctx, AV_LOG_ERROR, "Failed to import serialized graph to model graph\n");
av_freep(&sess_config);
return DNN_GENERIC_ERROR;
}
init_op = TF_GraphOperationByName(tf_model->graph, "init");
sess_opts = TF_NewSessionOptions();
if (sess_config) {
TF_SetConfig(sess_opts, sess_config, sess_config_length,tf_model->status);
av_freep(&sess_config);
if (TF_GetCode(tf_model->status) != TF_OK) {
TF_DeleteSessionOptions(sess_opts);
av_log(ctx, AV_LOG_ERROR, "Failed to set config for sess options with %s\n",
ctx->tf_option.sess_config);
return DNN_GENERIC_ERROR;
}
}
tf_model->session = TF_NewSession(tf_model->graph, sess_opts, tf_model->status);
TF_DeleteSessionOptions(sess_opts);
if (TF_GetCode(tf_model->status) != TF_OK)
{
av_freep(&sess_config);
av_log(ctx, AV_LOG_ERROR, "Failed to create new session with model graph\n");
return DNN_GENERIC_ERROR;
}
// Run initialization operation with name "init" if it is present in graph
if (init_op){
TF_SessionRun(tf_model->session, NULL,
NULL, NULL, 0,
NULL, NULL, 0,
&init_op, 1, NULL, tf_model->status);
if (TF_GetCode(tf_model->status) != TF_OK)
{
av_freep(&sess_config);
av_log(ctx, AV_LOG_ERROR, "Failed to run session when initializing\n");
return DNN_GENERIC_ERROR;
}
}
return 0;
}
static void dnn_free_model_tf(DNNModel **model)
{
TFModel *tf_model;
if (!model || !*model)
return;
tf_model = (TFModel *)(*model);
while (ff_safe_queue_size(tf_model->request_queue) != 0) {
TFRequestItem *item = ff_safe_queue_pop_front(tf_model->request_queue);
destroy_request_item(&item);
}
ff_safe_queue_destroy(tf_model->request_queue);
while (ff_queue_size(tf_model->lltask_queue) != 0) {
LastLevelTaskItem *item = ff_queue_pop_front(tf_model->lltask_queue);
av_freep(&item);
}
ff_queue_destroy(tf_model->lltask_queue);
while (ff_queue_size(tf_model->task_queue) != 0) {
TaskItem *item = ff_queue_pop_front(tf_model->task_queue);
av_frame_free(&item->in_frame);
av_frame_free(&item->out_frame);
av_freep(&item);
}
ff_queue_destroy(tf_model->task_queue);
if (tf_model->graph){
TF_DeleteGraph(tf_model->graph);
}
if (tf_model->session){
TF_CloseSession(tf_model->session, tf_model->status);
TF_DeleteSession(tf_model->session, tf_model->status);
}
if (tf_model->status){
TF_DeleteStatus(tf_model->status);
}
av_freep(&tf_model);
*model = NULL;
}
static DNNModel *dnn_load_model_tf(DnnContext *ctx, DNNFunctionType func_type, AVFilterContext *filter_ctx)
{
DNNModel *model = NULL;
TFModel *tf_model = NULL;
tf_model = av_mallocz(sizeof(TFModel));
if (!tf_model)
return NULL;
model = &tf_model->model;
tf_model->ctx = ctx;
if (load_tf_model(tf_model, ctx->model_filename) != 0){
av_log(ctx, AV_LOG_ERROR, "Failed to load TensorFlow model: \"%s\"\n", ctx->model_filename);
goto err;
}
if (ctx->nireq <= 0) {
ctx->nireq = av_cpu_count() / 2 + 1;
}
#if !HAVE_PTHREAD_CANCEL
if (ctx->options.async) {
ctx->options.async = 0;
av_log(filter_ctx, AV_LOG_WARNING, "pthread is not supported, roll back to sync.\n");
}
#endif
tf_model->request_queue = ff_safe_queue_create();
if (!tf_model->request_queue) {
goto err;
}
for (int i = 0; i < ctx->nireq; i++) {
TFRequestItem *item = av_mallocz(sizeof(*item));
if (!item) {
goto err;
}
item->lltask = NULL;
item->infer_request = tf_create_inference_request();
if (!item->infer_request) {
av_log(ctx, AV_LOG_ERROR, "Failed to allocate memory for TensorFlow inference request\n");
av_freep(&item);
goto err;
}
item->status = TF_NewStatus();
item->exec_module.start_inference = &tf_start_inference;
item->exec_module.callback = &infer_completion_callback;
item->exec_module.args = item;
if (ff_safe_queue_push_back(tf_model->request_queue, item) < 0) {
destroy_request_item(&item);
goto err;
}
}
tf_model->lltask_queue = ff_queue_create();
if (!tf_model->lltask_queue) {
goto err;
}
tf_model->task_queue = ff_queue_create();
if (!tf_model->task_queue) {
goto err;
}
model->get_input = &get_input_tf;
model->get_output = &get_output_tf;
model->filter_ctx = filter_ctx;
model->func_type = func_type;
return model;
err:
dnn_free_model_tf(&model);
return NULL;
}
static int fill_model_input_tf(TFModel *tf_model, TFRequestItem *request) {
DNNData input = { 0 };
LastLevelTaskItem *lltask;
TaskItem *task;
TFInferRequest *infer_request = NULL;
DnnContext *ctx = tf_model->ctx;
int ret = 0;
lltask = ff_queue_pop_front(tf_model->lltask_queue);
av_assert0(lltask);
task = lltask->task;
request->lltask = lltask;
ret = get_input_tf(&tf_model->model, &input, task->input_name);
if (ret != 0) {
goto err;
}
infer_request = request->infer_request;
input.dims[1] = task->in_frame->height;
input.dims[2] = task->in_frame->width;
infer_request->tf_input = av_malloc(sizeof(TF_Output));
if (!infer_request->tf_input) {
av_log(ctx, AV_LOG_ERROR, "Failed to allocate memory for input tensor\n");
ret = AVERROR(ENOMEM);
goto err;
}
infer_request->tf_input->oper = TF_GraphOperationByName(tf_model->graph, task->input_name);
if (!infer_request->tf_input->oper){
av_log(ctx, AV_LOG_ERROR, "Could not find \"%s\" in model\n", task->input_name);
ret = DNN_GENERIC_ERROR;
goto err;
}
infer_request->tf_input->index = 0;
infer_request->input_tensor = allocate_input_tensor(&input);
if (!infer_request->input_tensor){
av_log(ctx, AV_LOG_ERROR, "Failed to allocate memory for input tensor\n");
ret = AVERROR(ENOMEM);
goto err;
}
input.data = (float *)TF_TensorData(infer_request->input_tensor);
switch (tf_model->model.func_type) {
case DFT_PROCESS_FRAME:
if (task->do_ioproc) {
if (tf_model->model.frame_pre_proc != NULL) {
tf_model->model.frame_pre_proc(task->in_frame, &input, tf_model->model.filter_ctx);
} else {
ff_proc_from_frame_to_dnn(task->in_frame, &input, ctx);
}
}
break;
case DFT_ANALYTICS_DETECT:
ff_frame_to_dnn_detect(task->in_frame, &input, ctx);
break;
default:
avpriv_report_missing_feature(ctx, "model function type %d", tf_model->model.func_type);
break;
}
infer_request->tf_outputs = av_malloc_array(task->nb_output, sizeof(TF_Output));
if (infer_request->tf_outputs == NULL) {
av_log(ctx, AV_LOG_ERROR, "Failed to allocate memory for *tf_outputs\n");
ret = AVERROR(ENOMEM);
goto err;
}
infer_request->output_tensors = av_calloc(task->nb_output, sizeof(*infer_request->output_tensors));
if (!infer_request->output_tensors) {
av_log(ctx, AV_LOG_ERROR, "Failed to allocate memory for output tensor\n");
ret = AVERROR(ENOMEM);
goto err;
}
for (int i = 0; i < task->nb_output; ++i) {
infer_request->output_tensors[i] = NULL;
infer_request->tf_outputs[i].oper = TF_GraphOperationByName(tf_model->graph, task->output_names[i]);
if (!infer_request->tf_outputs[i].oper) {
av_log(ctx, AV_LOG_ERROR, "Could not find output \"%s\" in model\n", task->output_names[i]);
ret = DNN_GENERIC_ERROR;
goto err;
}
infer_request->tf_outputs[i].index = 0;
}
return 0;
err:
tf_free_request(infer_request);
return ret;
}
static void infer_completion_callback(void *args) {
TFRequestItem *request = args;
LastLevelTaskItem *lltask = request->lltask;
TaskItem *task = lltask->task;
DNNData *outputs;
TFInferRequest *infer_request = request->infer_request;
TFModel *tf_model = task->model;
DnnContext *ctx = tf_model->ctx;
outputs = av_calloc(task->nb_output, sizeof(*outputs));
if (!outputs) {
av_log(ctx, AV_LOG_ERROR, "Failed to allocate memory for *outputs\n");
goto err;
}
for (uint32_t i = 0; i < task->nb_output; ++i) {
outputs[i].dims[dnn_get_height_idx_by_layout(outputs[i].layout)] =
TF_Dim(infer_request->output_tensors[i], 1);
outputs[i].dims[dnn_get_width_idx_by_layout(outputs[i].layout)] =
TF_Dim(infer_request->output_tensors[i], 2);
outputs[i].dims[dnn_get_channel_idx_by_layout(outputs[i].layout)] =
TF_Dim(infer_request->output_tensors[i], 3);
outputs[i].data = TF_TensorData(infer_request->output_tensors[i]);
outputs[i].dt = (DNNDataType)TF_TensorType(infer_request->output_tensors[i]);
}
switch (tf_model->model.func_type) {
case DFT_PROCESS_FRAME:
//it only support 1 output if it's frame in & frame out
if (task->do_ioproc) {
if (tf_model->model.frame_post_proc != NULL) {
tf_model->model.frame_post_proc(task->out_frame, outputs, tf_model->model.filter_ctx);
} else {
ff_proc_from_dnn_to_frame(task->out_frame, outputs, ctx);
}
} else {
task->out_frame->width =
outputs[0].dims[dnn_get_width_idx_by_layout(outputs[0].layout)];
task->out_frame->height =
outputs[0].dims[dnn_get_height_idx_by_layout(outputs[0].layout)];
}
break;
case DFT_ANALYTICS_DETECT:
if (!tf_model->model.detect_post_proc) {
av_log(ctx, AV_LOG_ERROR, "Detect filter needs provide post proc\n");
return;
}
tf_model->model.detect_post_proc(task->in_frame, outputs, task->nb_output, tf_model->model.filter_ctx);
break;
default:
av_log(ctx, AV_LOG_ERROR, "Tensorflow backend does not support this kind of dnn filter now\n");
goto err;
}
task->inference_done++;
err:
tf_free_request(infer_request);
av_freep(&outputs);
if (ff_safe_queue_push_back(tf_model->request_queue, request) < 0) {
destroy_request_item(&request);
av_log(ctx, AV_LOG_ERROR, "Failed to push back request_queue.\n");
}
}
static int execute_model_tf(TFRequestItem *request, Queue *lltask_queue)
{
TFModel *tf_model;
DnnContext *ctx;
LastLevelTaskItem *lltask;
TaskItem *task;
int ret = 0;
if (ff_queue_size(lltask_queue) == 0) {
destroy_request_item(&request);
return 0;
}
lltask = ff_queue_peek_front(lltask_queue);
task = lltask->task;
tf_model = task->model;
ctx = tf_model->ctx;
ret = fill_model_input_tf(tf_model, request);
if (ret != 0) {
goto err;
}
if (task->async) {
if (ff_dnn_start_inference_async(ctx, &request->exec_module) != 0) {
goto err;
}
return 0;
}
else {
ret = tf_start_inference(request);
if (ret != 0) {
goto err;
}
infer_completion_callback(request);
return (task->inference_done == task->inference_todo) ? 0 : DNN_GENERIC_ERROR;
}
err:
tf_free_request(request->infer_request);
if (ff_safe_queue_push_back(tf_model->request_queue, request) < 0) {
destroy_request_item(&request);
}
return ret;
}
static int dnn_execute_model_tf(const DNNModel *model, DNNExecBaseParams *exec_params)
{
TFModel *tf_model = (TFModel *)model;
DnnContext *ctx = tf_model->ctx;
TaskItem *task;
TFRequestItem *request;
int ret = 0;
ret = ff_check_exec_params(ctx, DNN_TF, model->func_type, exec_params);
if (ret != 0) {
return ret;
}
task = av_malloc(sizeof(*task));
if (!task) {
av_log(ctx, AV_LOG_ERROR, "unable to alloc memory for task item.\n");
return AVERROR(ENOMEM);
}
ret = ff_dnn_fill_task(task, exec_params, tf_model, ctx->async, 1);
if (ret != 0) {
av_log(ctx, AV_LOG_ERROR, "Fill task with invalid parameter(s).\n");
av_freep(&task);
return ret;
}
if (ff_queue_push_back(tf_model->task_queue, task) < 0) {
av_freep(&task);
av_log(ctx, AV_LOG_ERROR, "unable to push back task_queue.\n");
return AVERROR(ENOMEM);
}
ret = extract_lltask_from_task(task, tf_model->lltask_queue);
if (ret != 0) {
av_log(ctx, AV_LOG_ERROR, "unable to extract last level task from task.\n");
return ret;
}
request = ff_safe_queue_pop_front(tf_model->request_queue);
if (!request) {
av_log(ctx, AV_LOG_ERROR, "unable to get infer request.\n");
return AVERROR(EINVAL);
}
return execute_model_tf(request, tf_model->lltask_queue);
}
static DNNAsyncStatusType dnn_get_result_tf(const DNNModel *model, AVFrame **in, AVFrame **out)
{
TFModel *tf_model = (TFModel *)model;
return ff_dnn_get_result_common(tf_model->task_queue, in, out);
}
static int dnn_flush_tf(const DNNModel *model)
{
TFModel *tf_model = (TFModel *)model;
DnnContext *ctx = tf_model->ctx;
TFRequestItem *request;
int ret;
if (ff_queue_size(tf_model->lltask_queue) == 0) {
// no pending task need to flush
return 0;
}
request = ff_safe_queue_pop_front(tf_model->request_queue);
if (!request) {
av_log(ctx, AV_LOG_ERROR, "unable to get infer request.\n");
return AVERROR(EINVAL);
}
ret = fill_model_input_tf(tf_model, request);
if (ret != 0) {
av_log(ctx, AV_LOG_ERROR, "Failed to fill model input.\n");
if (ff_safe_queue_push_back(tf_model->request_queue, request) < 0) {
destroy_request_item(&request);
}
return ret;
}
return ff_dnn_start_inference_async(ctx, &request->exec_module);
}
const DNNModule ff_dnn_backend_tf = {
.clazz = DNN_DEFINE_CLASS(dnn_tensorflow),
.type = DNN_TF,
.load_model = dnn_load_model_tf,
.execute_model = dnn_execute_model_tf,
.get_result = dnn_get_result_tf,
.flush = dnn_flush_tf,
.free_model = dnn_free_model_tf,
};
@@ -0,0 +1,583 @@
/*
* Copyright (c) 2024
*
* This file is part of FFmpeg.
*
* FFmpeg is free software; you can redistribute it and/or
* modify it under the terms of the GNU Lesser General Public
* License as published by the Free Software Foundation; either
* version 2.1 of the License, or (at your option) any later version.
*
* FFmpeg is distributed in the hope that it will be useful,
* but WITHOUT ANY WARRANTY; without even the implied warranty of
* MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the GNU
* Lesser General Public License for more details.
*
* You should have received a copy of the GNU Lesser General Public
* License along with FFmpeg; if not, write to the Free Software
* Foundation, Inc., 51 Franklin Street, Fifth Floor, Boston, MA 02110-1301 USA
*/
/**
* @file
* DNN Torch backend implementation.
*/
#include <torch/torch.h>
#include <torch/script.h>
extern "C" {
#include "dnn_io_proc.h"
#include "dnn_backend_common.h"
#include "libavutil/opt.h"
#include "libavutil/mem.h"
#include "queue.h"
#include "safe_queue.h"
}
typedef struct THModel {
DNNModel model;
DnnContext *ctx;
torch::jit::Module *jit_model;
SafeQueue *request_queue;
Queue *task_queue;
Queue *lltask_queue;
} THModel;
typedef struct THInferRequest {
torch::Tensor *output;
torch::Tensor *input_tensor;
} THInferRequest;
typedef struct THRequestItem {
THInferRequest *infer_request;
LastLevelTaskItem *lltask;
DNNAsyncExecModule exec_module;
} THRequestItem;
#define OFFSET(x) offsetof(THOptions, x)
#define FLAGS AV_OPT_FLAG_FILTERING_PARAM
static const AVOption dnn_th_options[] = {
{ "optimize", "turn on graph executor optimization", OFFSET(optimize), AV_OPT_TYPE_INT, { .i64 = 0 }, 0, 1, FLAGS},
{ NULL }
};
static int extract_lltask_from_task(TaskItem *task, Queue *lltask_queue)
{
THModel *th_model = (THModel *)task->model;
DnnContext *ctx = th_model->ctx;
LastLevelTaskItem *lltask = (LastLevelTaskItem *)av_malloc(sizeof(*lltask));
if (!lltask) {
av_log(ctx, AV_LOG_ERROR, "Failed to allocate memory for LastLevelTaskItem\n");
return AVERROR(ENOMEM);
}
task->inference_todo = 1;
task->inference_done = 0;
lltask->task = task;
if (ff_queue_push_back(lltask_queue, lltask) < 0) {
av_log(ctx, AV_LOG_ERROR, "Failed to push back lltask_queue.\n");
av_freep(&lltask);
return AVERROR(ENOMEM);
}
return 0;
}
static void th_free_request(THInferRequest *request)
{
if (!request)
return;
if (request->output) {
delete(request->output);
request->output = NULL;
}
if (request->input_tensor) {
delete(request->input_tensor);
request->input_tensor = NULL;
}
return;
}
static inline void destroy_request_item(THRequestItem **arg)
{
THRequestItem *item;
if (!arg || !*arg) {
return;
}
item = *arg;
th_free_request(item->infer_request);
av_freep(&item->infer_request);
av_freep(&item->lltask);
ff_dnn_async_module_cleanup(&item->exec_module);
av_freep(arg);
}
static void dnn_free_model_th(DNNModel **model)
{
THModel *th_model;
if (!model || !*model)
return;
th_model = (THModel *) (*model);
while (ff_safe_queue_size(th_model->request_queue) != 0) {
THRequestItem *item = (THRequestItem *)ff_safe_queue_pop_front(th_model->request_queue);
destroy_request_item(&item);
}
ff_safe_queue_destroy(th_model->request_queue);
while (ff_queue_size(th_model->lltask_queue) != 0) {
LastLevelTaskItem *item = (LastLevelTaskItem *)ff_queue_pop_front(th_model->lltask_queue);
av_freep(&item);
}
ff_queue_destroy(th_model->lltask_queue);
while (ff_queue_size(th_model->task_queue) != 0) {
TaskItem *item = (TaskItem *)ff_queue_pop_front(th_model->task_queue);
av_frame_free(&item->in_frame);
av_frame_free(&item->out_frame);
av_freep(&item);
}
ff_queue_destroy(th_model->task_queue);
delete th_model->jit_model;
av_freep(&th_model);
*model = NULL;
}
static int get_input_th(DNNModel *model, DNNData *input, const char *input_name)
{
input->dt = DNN_FLOAT;
input->order = DCO_RGB;
input->layout = DL_NCHW;
input->dims[0] = 1;
input->dims[1] = 3;
input->dims[2] = -1;
input->dims[3] = -1;
return 0;
}
static void deleter(void *arg)
{
av_freep(&arg);
}
static int fill_model_input_th(THModel *th_model, THRequestItem *request)
{
LastLevelTaskItem *lltask = NULL;
TaskItem *task = NULL;
THInferRequest *infer_request = NULL;
DNNData input = { 0 };
DnnContext *ctx = th_model->ctx;
int ret, width_idx, height_idx, channel_idx;
lltask = (LastLevelTaskItem *)ff_queue_pop_front(th_model->lltask_queue);
if (!lltask) {
ret = AVERROR(EINVAL);
goto err;
}
request->lltask = lltask;
task = lltask->task;
infer_request = request->infer_request;
ret = get_input_th(&th_model->model, &input, NULL);
if ( ret != 0) {
goto err;
}
width_idx = dnn_get_width_idx_by_layout(input.layout);
height_idx = dnn_get_height_idx_by_layout(input.layout);
channel_idx = dnn_get_channel_idx_by_layout(input.layout);
input.dims[height_idx] = task->in_frame->height;
input.dims[width_idx] = task->in_frame->width;
input.data = av_malloc(input.dims[height_idx] * input.dims[width_idx] *
input.dims[channel_idx] * sizeof(float));
if (!input.data)
return AVERROR(ENOMEM);
infer_request->input_tensor = new torch::Tensor();
infer_request->output = new torch::Tensor();
switch (th_model->model.func_type) {
case DFT_PROCESS_FRAME:
input.scale = 255;
if (task->do_ioproc) {
if (th_model->model.frame_pre_proc != NULL) {
th_model->model.frame_pre_proc(task->in_frame, &input, th_model->model.filter_ctx);
} else {
ff_proc_from_frame_to_dnn(task->in_frame, &input, ctx);
}
}
break;
default:
avpriv_report_missing_feature(NULL, "model function type %d", th_model->model.func_type);
break;
}
*infer_request->input_tensor = torch::from_blob(input.data,
{1, input.dims[channel_idx], input.dims[height_idx], input.dims[width_idx]},
deleter, torch::kFloat32);
return 0;
err:
th_free_request(infer_request);
return ret;
}
static int th_start_inference(void *args)
{
THRequestItem *request = (THRequestItem *)args;
THInferRequest *infer_request = NULL;
LastLevelTaskItem *lltask = NULL;
TaskItem *task = NULL;
THModel *th_model = NULL;
DnnContext *ctx = NULL;
std::vector<torch::jit::IValue> inputs;
torch::NoGradGuard no_grad;
if (!request) {
av_log(NULL, AV_LOG_ERROR, "THRequestItem is NULL\n");
return AVERROR(EINVAL);
}
infer_request = request->infer_request;
lltask = request->lltask;
task = lltask->task;
th_model = (THModel *)task->model;
ctx = th_model->ctx;
if (ctx->torch_option.optimize)
torch::jit::setGraphExecutorOptimize(true);
else
torch::jit::setGraphExecutorOptimize(false);
if (!infer_request->input_tensor || !infer_request->output) {
av_log(ctx, AV_LOG_ERROR, "input or output tensor is NULL\n");
return DNN_GENERIC_ERROR;
}
// Transfer tensor to the same device as model
c10::Device device = (*th_model->jit_model->parameters().begin()).device();
if (infer_request->input_tensor->device() != device)
*infer_request->input_tensor = infer_request->input_tensor->to(device);
inputs.push_back(*infer_request->input_tensor);
*infer_request->output = th_model->jit_model->forward(inputs).toTensor();
return 0;
}
static void infer_completion_callback(void *args) {
THRequestItem *request = (THRequestItem*)args;
LastLevelTaskItem *lltask = request->lltask;
TaskItem *task = lltask->task;
DNNData outputs = { 0 };
THInferRequest *infer_request = request->infer_request;
THModel *th_model = (THModel *)task->model;
torch::Tensor *output = infer_request->output;
c10::IntArrayRef sizes = output->sizes();
outputs.order = DCO_RGB;
outputs.layout = DL_NCHW;
outputs.dt = DNN_FLOAT;
if (sizes.size() == 4) {
// 4 dimensions: [batch_size, channel, height, width]
// this format of data is normally used for video frame SR
outputs.dims[0] = sizes.at(0); // N
outputs.dims[1] = sizes.at(1); // C
outputs.dims[2] = sizes.at(2); // H
outputs.dims[3] = sizes.at(3); // W
} else {
avpriv_report_missing_feature(th_model->ctx, "Support of this kind of model");
goto err;
}
switch (th_model->model.func_type) {
case DFT_PROCESS_FRAME:
if (task->do_ioproc) {
// Post process can only deal with CPU memory.
if (output->device() != torch::kCPU)
*output = output->to(torch::kCPU);
outputs.scale = 255;
outputs.data = output->data_ptr();
if (th_model->model.frame_post_proc != NULL) {
th_model->model.frame_post_proc(task->out_frame, &outputs, th_model->model.filter_ctx);
} else {
ff_proc_from_dnn_to_frame(task->out_frame, &outputs, th_model->ctx);
}
} else {
task->out_frame->width = outputs.dims[dnn_get_width_idx_by_layout(outputs.layout)];
task->out_frame->height = outputs.dims[dnn_get_height_idx_by_layout(outputs.layout)];
}
break;
default:
avpriv_report_missing_feature(th_model->ctx, "model function type %d", th_model->model.func_type);
goto err;
}
task->inference_done++;
av_freep(&request->lltask);
err:
th_free_request(infer_request);
if (ff_safe_queue_push_back(th_model->request_queue, request) < 0) {
destroy_request_item(&request);
av_log(th_model->ctx, AV_LOG_ERROR, "Unable to push back request_queue when failed to start inference.\n");
}
}
static int execute_model_th(THRequestItem *request, Queue *lltask_queue)
{
THModel *th_model = NULL;
LastLevelTaskItem *lltask;
TaskItem *task = NULL;
int ret = 0;
if (ff_queue_size(lltask_queue) == 0) {
destroy_request_item(&request);
return 0;
}
lltask = (LastLevelTaskItem *)ff_queue_peek_front(lltask_queue);
if (lltask == NULL) {
av_log(NULL, AV_LOG_ERROR, "Failed to get LastLevelTaskItem\n");
ret = AVERROR(EINVAL);
goto err;
}
task = lltask->task;
th_model = (THModel *)task->model;
ret = fill_model_input_th(th_model, request);
if ( ret != 0) {
goto err;
}
if (task->async) {
avpriv_report_missing_feature(th_model->ctx, "LibTorch async");
} else {
ret = th_start_inference((void *)(request));
if (ret != 0) {
goto err;
}
infer_completion_callback(request);
return (task->inference_done == task->inference_todo) ? 0 : DNN_GENERIC_ERROR;
}
err:
th_free_request(request->infer_request);
if (ff_safe_queue_push_back(th_model->request_queue, request) < 0) {
destroy_request_item(&request);
}
return ret;
}
static int get_output_th(DNNModel *model, const char *input_name, int input_width, int input_height,
const char *output_name, int *output_width, int *output_height)
{
int ret = 0;
THModel *th_model = (THModel*) model;
DnnContext *ctx = th_model->ctx;
TaskItem task = { 0 };
THRequestItem *request = NULL;
DNNExecBaseParams exec_params = {
.input_name = input_name,
.output_names = &output_name,
.nb_output = 1,
.in_frame = NULL,
.out_frame = NULL,
};
ret = ff_dnn_fill_gettingoutput_task(&task, &exec_params, th_model, input_height, input_width, ctx);
if ( ret != 0) {
goto err;
}
ret = extract_lltask_from_task(&task, th_model->lltask_queue);
if ( ret != 0) {
av_log(ctx, AV_LOG_ERROR, "unable to extract last level task from task.\n");
goto err;
}
request = (THRequestItem*) ff_safe_queue_pop_front(th_model->request_queue);
if (!request) {
av_log(ctx, AV_LOG_ERROR, "unable to get infer request.\n");
ret = AVERROR(EINVAL);
goto err;
}
ret = execute_model_th(request, th_model->lltask_queue);
*output_width = task.out_frame->width;
*output_height = task.out_frame->height;
err:
av_frame_free(&task.out_frame);
av_frame_free(&task.in_frame);
return ret;
}
static THInferRequest *th_create_inference_request(void)
{
THInferRequest *request = (THInferRequest *)av_malloc(sizeof(THInferRequest));
if (!request) {
return NULL;
}
request->input_tensor = NULL;
request->output = NULL;
return request;
}
static DNNModel *dnn_load_model_th(DnnContext *ctx, DNNFunctionType func_type, AVFilterContext *filter_ctx)
{
DNNModel *model = NULL;
THModel *th_model = NULL;
THRequestItem *item = NULL;
const char *device_name = ctx->device ? ctx->device : "cpu";
th_model = (THModel *)av_mallocz(sizeof(THModel));
if (!th_model)
return NULL;
model = &th_model->model;
th_model->ctx = ctx;
c10::Device device = c10::Device(device_name);
if (device.is_xpu()) {
if (!at::hasXPU()) {
av_log(ctx, AV_LOG_ERROR, "No XPU device found\n");
goto fail;
}
at::detail::getXPUHooks().initXPU();
} else if (!device.is_cpu()) {
av_log(ctx, AV_LOG_ERROR, "Not supported device:\"%s\"\n", device_name);
goto fail;
}
try {
th_model->jit_model = new torch::jit::Module;
(*th_model->jit_model) = torch::jit::load(ctx->model_filename);
th_model->jit_model->to(device);
} catch (const c10::Error& e) {
av_log(ctx, AV_LOG_ERROR, "Failed to load torch model\n");
goto fail;
}
th_model->request_queue = ff_safe_queue_create();
if (!th_model->request_queue) {
goto fail;
}
item = (THRequestItem *)av_mallocz(sizeof(THRequestItem));
if (!item) {
goto fail;
}
item->lltask = NULL;
item->infer_request = th_create_inference_request();
if (!item->infer_request) {
av_log(NULL, AV_LOG_ERROR, "Failed to allocate memory for Torch inference request\n");
goto fail;
}
item->exec_module.start_inference = &th_start_inference;
item->exec_module.callback = &infer_completion_callback;
item->exec_module.args = item;
if (ff_safe_queue_push_back(th_model->request_queue, item) < 0) {
goto fail;
}
item = NULL;
th_model->task_queue = ff_queue_create();
if (!th_model->task_queue) {
goto fail;
}
th_model->lltask_queue = ff_queue_create();
if (!th_model->lltask_queue) {
goto fail;
}
model->get_input = &get_input_th;
model->get_output = &get_output_th;
model->filter_ctx = filter_ctx;
model->func_type = func_type;
return model;
fail:
if (item) {
destroy_request_item(&item);
av_freep(&item);
}
dnn_free_model_th(&model);
return NULL;
}
static int dnn_execute_model_th(const DNNModel *model, DNNExecBaseParams *exec_params)
{
THModel *th_model = (THModel *)model;
DnnContext *ctx = th_model->ctx;
TaskItem *task;
THRequestItem *request;
int ret = 0;
ret = ff_check_exec_params(ctx, DNN_TH, model->func_type, exec_params);
if (ret != 0) {
av_log(ctx, AV_LOG_ERROR, "exec parameter checking fail.\n");
return ret;
}
task = (TaskItem *)av_malloc(sizeof(TaskItem));
if (!task) {
av_log(ctx, AV_LOG_ERROR, "unable to alloc memory for task item.\n");
return AVERROR(ENOMEM);
}
ret = ff_dnn_fill_task(task, exec_params, th_model, 0, 1);
if (ret != 0) {
av_freep(&task);
av_log(ctx, AV_LOG_ERROR, "unable to fill task.\n");
return ret;
}
ret = ff_queue_push_back(th_model->task_queue, task);
if (ret < 0) {
av_freep(&task);
av_log(ctx, AV_LOG_ERROR, "unable to push back task_queue.\n");
return ret;
}
ret = extract_lltask_from_task(task, th_model->lltask_queue);
if (ret != 0) {
av_log(ctx, AV_LOG_ERROR, "unable to extract last level task from task.\n");
return ret;
}
request = (THRequestItem *)ff_safe_queue_pop_front(th_model->request_queue);
if (!request) {
av_log(ctx, AV_LOG_ERROR, "unable to get infer request.\n");
return AVERROR(EINVAL);
}
return execute_model_th(request, th_model->lltask_queue);
}
static DNNAsyncStatusType dnn_get_result_th(const DNNModel *model, AVFrame **in, AVFrame **out)
{
THModel *th_model = (THModel *)model;
return ff_dnn_get_result_common(th_model->task_queue, in, out);
}
static int dnn_flush_th(const DNNModel *model)
{
THModel *th_model = (THModel *)model;
THRequestItem *request;
if (ff_queue_size(th_model->lltask_queue) == 0)
// no pending task need to flush
return 0;
request = (THRequestItem *)ff_safe_queue_pop_front(th_model->request_queue);
if (!request) {
av_log(th_model->ctx, AV_LOG_ERROR, "unable to get infer request.\n");
return AVERROR(EINVAL);
}
return execute_model_th(request, th_model->lltask_queue);
}
extern const DNNModule ff_dnn_backend_torch = {
.clazz = DNN_DEFINE_CLASS(dnn_th),
.type = DNN_TH,
.load_model = dnn_load_model_th,
.execute_model = dnn_execute_model_th,
.get_result = dnn_get_result_th,
.flush = dnn_flush_th,
.free_model = dnn_free_model_th,
};
+139
View File
@@ -0,0 +1,139 @@
/*
* Copyright (c) 2018 Sergey Lavrushkin
*
* This file is part of FFmpeg.
*
* FFmpeg is free software; you can redistribute it and/or
* modify it under the terms of the GNU Lesser General Public
* License as published by the Free Software Foundation; either
* version 2.1 of the License, or (at your option) any later version.
*
* FFmpeg is distributed in the hope that it will be useful,
* but WITHOUT ANY WARRANTY; without even the implied warranty of
* MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the GNU
* Lesser General Public License for more details.
*
* You should have received a copy of the GNU Lesser General Public
* License along with FFmpeg; if not, write to the Free Software
* Foundation, Inc., 51 Franklin Street, Fifth Floor, Boston, MA 02110-1301 USA
*/
/**
* @file
* Implements DNN module initialization with specified backend.
*/
#include "../dnn_interface.h"
#include "libavutil/avassert.h"
#include "libavutil/mem.h"
#include "libavutil/opt.h"
#include "libavfilter/filters.h"
extern const DNNModule ff_dnn_backend_openvino;
extern const DNNModule ff_dnn_backend_tf;
extern const DNNModule ff_dnn_backend_torch;
#define OFFSET(x) offsetof(DnnContext, x)
#define FLAGS AV_OPT_FLAG_FILTERING_PARAM
static const AVOption dnn_base_options[] = {
{"model", "path to model file",
OFFSET(model_filename), AV_OPT_TYPE_STRING, {.str = NULL}, 0, 0, FLAGS},
{"input", "input name of the model",
OFFSET(model_inputname), AV_OPT_TYPE_STRING, {.str = NULL}, 0, 0, FLAGS},
{"output", "output name of the model",
OFFSET(model_outputnames_string), AV_OPT_TYPE_STRING, {.str = NULL}, 0, 0, FLAGS},
{"backend_configs", "backend configs (deprecated)",
OFFSET(backend_options), AV_OPT_TYPE_STRING, {.str = NULL}, 0, 0, FLAGS | AV_OPT_FLAG_DEPRECATED},
{"options", "backend configs (deprecated)",
OFFSET(backend_options), AV_OPT_TYPE_STRING, {.str = NULL}, 0, 0, FLAGS | AV_OPT_FLAG_DEPRECATED},
{"nireq", "number of request",
OFFSET(nireq), AV_OPT_TYPE_INT, {.i64 = 0}, 0, INT_MAX, FLAGS},
{"async", "use DNN async inference",
OFFSET(async), AV_OPT_TYPE_BOOL, {.i64 = 1}, 0, 1, FLAGS},
{"device", "device to run model",
OFFSET(device), AV_OPT_TYPE_STRING, {.str = NULL}, 0, 0, FLAGS},
{NULL}
};
AVFILTER_DEFINE_CLASS(dnn_base);
typedef struct DnnBackendInfo {
const size_t offset;
union {
const AVClass *class;
const DNNModule *module;
};
} DnnBackendInfo;
static const DnnBackendInfo dnn_backend_info_list[] = {
{0, .class = &dnn_base_class},
// Must keep the same order as in DNNOptions, so offset value in incremental order
#if CONFIG_LIBTENSORFLOW
{offsetof(DnnContext, tf_option), .module = &ff_dnn_backend_tf},
#endif
#if CONFIG_LIBOPENVINO
{offsetof(DnnContext, ov_option), .module = &ff_dnn_backend_openvino},
#endif
#if CONFIG_LIBTORCH
{offsetof(DnnContext, torch_option), .module = &ff_dnn_backend_torch},
#endif
};
const DNNModule *ff_get_dnn_module(DNNBackendType backend_type, void *log_ctx)
{
for (int i = 1; i < FF_ARRAY_ELEMS(dnn_backend_info_list); i++) {
if (dnn_backend_info_list[i].module->type == backend_type)
return dnn_backend_info_list[i].module;
}
av_log(log_ctx, AV_LOG_ERROR,
"Module backend_type %d is not supported or enabled.\n",
backend_type);
return NULL;
}
void ff_dnn_init_child_class(DnnContext *ctx)
{
for (int i = 0; i < FF_ARRAY_ELEMS(dnn_backend_info_list); i++) {
const AVClass **ptr = (const AVClass **) ((char *) ctx + dnn_backend_info_list[i].offset);
*ptr = dnn_backend_info_list[i].class;
}
}
void *ff_dnn_child_next(DnnContext *obj, void *prev) {
size_t pre_offset;
if (!prev) {
av_assert0(obj->clazz);
return obj;
}
pre_offset = (char *)prev - (char *)obj;
for (int i = 0; i < FF_ARRAY_ELEMS(dnn_backend_info_list) - 1; i++) {
if (dnn_backend_info_list[i].offset == pre_offset) {
const AVClass **ptr = (const AVClass **) ((char *) obj + dnn_backend_info_list[i + 1].offset);
av_assert0(*ptr);
return ptr;
}
}
return NULL;
}
const AVClass *ff_dnn_child_class_iterate_with_mask(void **iter, uint32_t backend_mask)
{
for (uintptr_t i = (uintptr_t)*iter; i < FF_ARRAY_ELEMS(dnn_backend_info_list); i++) {
if (i > 0) {
const DNNModule *module = dnn_backend_info_list[i].module;
if (!(module->type & backend_mask))
continue;
}
*iter = (void *)(i + 1);
return dnn_backend_info_list[i].class;
}
return NULL;
}
+471
View File
@@ -0,0 +1,471 @@
/*
* Copyright (c) 2020
*
* This file is part of FFmpeg.
*
* FFmpeg is free software; you can redistribute it and/or
* modify it under the terms of the GNU Lesser General Public
* License as published by the Free Software Foundation; either
* version 2.1 of the License, or (at your option) any later version.
*
* FFmpeg is distributed in the hope that it will be useful,
* but WITHOUT ANY WARRANTY; without even the implied warranty of
* MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the GNU
* Lesser General Public License for more details.
*
* You should have received a copy of the GNU Lesser General Public
* License along with FFmpeg; if not, write to the Free Software
* Foundation, Inc., 51 Franklin Street, Fifth Floor, Boston, MA 02110-1301 USA
*/
#include "dnn_io_proc.h"
#include "libavutil/imgutils.h"
#include "libavutil/mem.h"
#include "libswscale/swscale.h"
#include "libavutil/avassert.h"
#include "libavutil/detection_bbox.h"
static int get_datatype_size(DNNDataType dt)
{
switch (dt)
{
case DNN_FLOAT:
return sizeof(float);
case DNN_UINT8:
return sizeof(uint8_t);
default:
av_assert0(!"not supported yet.");
return 1;
}
}
int ff_proc_from_dnn_to_frame(AVFrame *frame, DNNData *output, void *log_ctx)
{
struct SwsContext *sws_ctx;
int ret = 0;
int linesize[4] = { 0 };
void **dst_data = NULL;
void *middle_data = NULL;
uint8_t *planar_data[4] = { 0 };
int plane_size = frame->width * frame->height * sizeof(uint8_t);
enum AVPixelFormat src_fmt = AV_PIX_FMT_NONE;
int src_datatype_size = get_datatype_size(output->dt);
int bytewidth = av_image_get_linesize(frame->format, frame->width, 0);
if (bytewidth < 0) {
return AVERROR(EINVAL);
}
/* scale == 1 and mean == 0 and dt == UINT8: passthrough */
if (fabsf(output->scale - 1) < 1e-6f && fabsf(output->mean) < 1e-6 && output->dt == DNN_UINT8)
src_fmt = AV_PIX_FMT_GRAY8;
/* (scale == 255 or scale == 0) and mean == 0 and dt == FLOAT: normalization */
else if ((fabsf(output->scale - 255) < 1e-6f || fabsf(output->scale) < 1e-6f) &&
fabsf(output->mean) < 1e-6 && output->dt == DNN_FLOAT)
src_fmt = AV_PIX_FMT_GRAYF32;
else {
av_log(log_ctx, AV_LOG_ERROR, "dnn_process output data doesn't type: UINT8 "
"scale: %f, mean: %f\n", output->scale, output->mean);
return AVERROR(ENOSYS);
}
dst_data = (void **)frame->data;
linesize[0] = frame->linesize[0];
if (output->layout == DL_NCHW) {
middle_data = av_malloc(plane_size * output->dims[1]);
if (!middle_data) {
ret = AVERROR(ENOMEM);
goto err;
}
dst_data = &middle_data;
linesize[0] = frame->width * 3;
}
switch (frame->format) {
case AV_PIX_FMT_RGB24:
case AV_PIX_FMT_BGR24:
sws_ctx = sws_getContext(frame->width * 3,
frame->height,
src_fmt,
frame->width * 3,
frame->height,
AV_PIX_FMT_GRAY8,
0, NULL, NULL, NULL);
if (!sws_ctx) {
av_log(log_ctx, AV_LOG_ERROR, "Impossible to create scale context for the conversion "
"fmt:%s s:%dx%d -> fmt:%s s:%dx%d\n",
av_get_pix_fmt_name(src_fmt), frame->width * 3, frame->height,
av_get_pix_fmt_name(AV_PIX_FMT_GRAY8), frame->width * 3, frame->height);
ret = AVERROR(EINVAL);
goto err;
}
sws_scale(sws_ctx, (const uint8_t *[4]){(const uint8_t *)output->data, 0, 0, 0},
(const int[4]){frame->width * 3 * src_datatype_size, 0, 0, 0}, 0, frame->height,
(uint8_t * const*)dst_data, linesize);
sws_freeContext(sws_ctx);
// convert data from planar to packed
if (output->layout == DL_NCHW) {
sws_ctx = sws_getContext(frame->width,
frame->height,
AV_PIX_FMT_GBRP,
frame->width,
frame->height,
frame->format,
0, NULL, NULL, NULL);
if (!sws_ctx) {
av_log(log_ctx, AV_LOG_ERROR, "Impossible to create scale context for the conversion "
"fmt:%s s:%dx%d -> fmt:%s s:%dx%d\n",
av_get_pix_fmt_name(AV_PIX_FMT_GBRP), frame->width, frame->height,
av_get_pix_fmt_name(frame->format),frame->width, frame->height);
ret = AVERROR(EINVAL);
goto err;
}
if (frame->format == AV_PIX_FMT_RGB24) {
planar_data[0] = (uint8_t *)middle_data + plane_size;
planar_data[1] = (uint8_t *)middle_data + plane_size * 2;
planar_data[2] = (uint8_t *)middle_data;
} else if (frame->format == AV_PIX_FMT_BGR24) {
planar_data[0] = (uint8_t *)middle_data + plane_size;
planar_data[1] = (uint8_t *)middle_data;
planar_data[2] = (uint8_t *)middle_data + plane_size * 2;
}
sws_scale(sws_ctx, (const uint8_t * const *)planar_data,
(const int [4]){frame->width * sizeof(uint8_t),
frame->width * sizeof(uint8_t),
frame->width * sizeof(uint8_t), 0},
0, frame->height, frame->data, frame->linesize);
sws_freeContext(sws_ctx);
}
break;
case AV_PIX_FMT_GRAYF32:
av_image_copy_plane(frame->data[0], frame->linesize[0],
output->data, bytewidth,
bytewidth, frame->height);
break;
case AV_PIX_FMT_YUV420P:
case AV_PIX_FMT_YUV422P:
case AV_PIX_FMT_YUV444P:
case AV_PIX_FMT_YUV410P:
case AV_PIX_FMT_YUV411P:
case AV_PIX_FMT_GRAY8:
case AV_PIX_FMT_NV12:
sws_ctx = sws_getContext(frame->width,
frame->height,
AV_PIX_FMT_GRAYF32,
frame->width,
frame->height,
AV_PIX_FMT_GRAY8,
0, NULL, NULL, NULL);
if (!sws_ctx) {
av_log(log_ctx, AV_LOG_ERROR, "Impossible to create scale context for the conversion "
"fmt:%s s:%dx%d -> fmt:%s s:%dx%d\n",
av_get_pix_fmt_name(src_fmt), frame->width, frame->height,
av_get_pix_fmt_name(AV_PIX_FMT_GRAY8), frame->width, frame->height);
ret = AVERROR(EINVAL);
goto err;
}
sws_scale(sws_ctx, (const uint8_t *[4]){(const uint8_t *)output->data, 0, 0, 0},
(const int[4]){frame->width * src_datatype_size, 0, 0, 0}, 0, frame->height,
(uint8_t * const*)frame->data, frame->linesize);
sws_freeContext(sws_ctx);
break;
default:
avpriv_report_missing_feature(log_ctx, "%s", av_get_pix_fmt_name(frame->format));
ret = AVERROR(ENOSYS);
goto err;
}
err:
av_free(middle_data);
return ret;
}
int ff_proc_from_frame_to_dnn(AVFrame *frame, DNNData *input, void *log_ctx)
{
struct SwsContext *sws_ctx;
int ret = 0;
int linesize[4] = { 0 };
void **src_data = NULL;
void *middle_data = NULL;
uint8_t *planar_data[4] = { 0 };
int plane_size = frame->width * frame->height * sizeof(uint8_t);
enum AVPixelFormat dst_fmt = AV_PIX_FMT_NONE;
int dst_datatype_size = get_datatype_size(input->dt);
int bytewidth = av_image_get_linesize(frame->format, frame->width, 0);
if (bytewidth < 0) {
return AVERROR(EINVAL);
}
/* scale == 1 and mean == 0 and dt == UINT8: passthrough */
if (fabsf(input->scale - 1) < 1e-6f && fabsf(input->mean) < 1e-6 && input->dt == DNN_UINT8)
dst_fmt = AV_PIX_FMT_GRAY8;
/* (scale == 255 or scale == 0) and mean == 0 and dt == FLOAT: normalization */
else if ((fabsf(input->scale - 255) < 1e-6f || fabsf(input->scale) < 1e-6f) &&
fabsf(input->mean) < 1e-6 && input->dt == DNN_FLOAT)
dst_fmt = AV_PIX_FMT_GRAYF32;
else {
av_log(log_ctx, AV_LOG_ERROR, "dnn_process input data doesn't support type: UINT8 "
"scale: %f, mean: %f\n", input->scale, input->mean);
return AVERROR(ENOSYS);
}
src_data = (void **)frame->data;
linesize[0] = frame->linesize[0];
if (input->layout == DL_NCHW) {
middle_data = av_malloc(plane_size * input->dims[1]);
if (!middle_data) {
ret = AVERROR(ENOMEM);
goto err;
}
src_data = &middle_data;
linesize[0] = frame->width * 3;
}
switch (frame->format) {
case AV_PIX_FMT_RGB24:
case AV_PIX_FMT_BGR24:
// convert data from planar to packed
if (input->layout == DL_NCHW) {
sws_ctx = sws_getContext(frame->width,
frame->height,
frame->format,
frame->width,
frame->height,
AV_PIX_FMT_GBRP,
0, NULL, NULL, NULL);
if (!sws_ctx) {
av_log(log_ctx, AV_LOG_ERROR, "Impossible to create scale context for the conversion "
"fmt:%s s:%dx%d -> fmt:%s s:%dx%d\n",
av_get_pix_fmt_name(frame->format), frame->width, frame->height,
av_get_pix_fmt_name(AV_PIX_FMT_GBRP),frame->width, frame->height);
ret = AVERROR(EINVAL);
goto err;
}
if (frame->format == AV_PIX_FMT_RGB24) {
planar_data[0] = (uint8_t *)middle_data + plane_size;
planar_data[1] = (uint8_t *)middle_data + plane_size * 2;
planar_data[2] = (uint8_t *)middle_data;
} else if (frame->format == AV_PIX_FMT_BGR24) {
planar_data[0] = (uint8_t *)middle_data + plane_size;
planar_data[1] = (uint8_t *)middle_data;
planar_data[2] = (uint8_t *)middle_data + plane_size * 2;
}
sws_scale(sws_ctx, (const uint8_t * const *)frame->data,
frame->linesize, 0, frame->height, planar_data,
(const int [4]){frame->width * sizeof(uint8_t),
frame->width * sizeof(uint8_t),
frame->width * sizeof(uint8_t), 0});
sws_freeContext(sws_ctx);
}
sws_ctx = sws_getContext(frame->width * 3,
frame->height,
AV_PIX_FMT_GRAY8,
frame->width * 3,
frame->height,
dst_fmt,
0, NULL, NULL, NULL);
if (!sws_ctx) {
av_log(log_ctx, AV_LOG_ERROR, "Impossible to create scale context for the conversion "
"fmt:%s s:%dx%d -> fmt:%s s:%dx%d\n",
av_get_pix_fmt_name(AV_PIX_FMT_GRAY8), frame->width * 3, frame->height,
av_get_pix_fmt_name(dst_fmt),frame->width * 3, frame->height);
ret = AVERROR(EINVAL);
goto err;
}
sws_scale(sws_ctx, (const uint8_t **)src_data,
linesize, 0, frame->height,
(uint8_t * const [4]){input->data, 0, 0, 0},
(const int [4]){frame->width * 3 * dst_datatype_size, 0, 0, 0});
sws_freeContext(sws_ctx);
break;
case AV_PIX_FMT_GRAYF32:
av_image_copy_plane(input->data, bytewidth,
frame->data[0], frame->linesize[0],
bytewidth, frame->height);
break;
case AV_PIX_FMT_YUV420P:
case AV_PIX_FMT_YUV422P:
case AV_PIX_FMT_YUV444P:
case AV_PIX_FMT_YUV410P:
case AV_PIX_FMT_YUV411P:
case AV_PIX_FMT_GRAY8:
case AV_PIX_FMT_NV12:
sws_ctx = sws_getContext(frame->width,
frame->height,
AV_PIX_FMT_GRAY8,
frame->width,
frame->height,
dst_fmt,
0, NULL, NULL, NULL);
if (!sws_ctx) {
av_log(log_ctx, AV_LOG_ERROR, "Impossible to create scale context for the conversion "
"fmt:%s s:%dx%d -> fmt:%s s:%dx%d\n",
av_get_pix_fmt_name(AV_PIX_FMT_GRAY8), frame->width, frame->height,
av_get_pix_fmt_name(dst_fmt),frame->width, frame->height);
ret = AVERROR(EINVAL);
goto err;
}
sws_scale(sws_ctx, (const uint8_t **)frame->data,
frame->linesize, 0, frame->height,
(uint8_t * const [4]){input->data, 0, 0, 0},
(const int [4]){frame->width * dst_datatype_size, 0, 0, 0});
sws_freeContext(sws_ctx);
break;
default:
avpriv_report_missing_feature(log_ctx, "%s", av_get_pix_fmt_name(frame->format));
ret = AVERROR(ENOSYS);
goto err;
}
err:
av_free(middle_data);
return ret;
}
static enum AVPixelFormat get_pixel_format(DNNData *data)
{
if (data->dt == DNN_UINT8) {
switch (data->order) {
case DCO_BGR:
return AV_PIX_FMT_BGR24;
case DCO_RGB:
return AV_PIX_FMT_RGB24;
default:
av_assert0(!"unsupported data pixel format.\n");
return AV_PIX_FMT_BGR24;
}
}
av_assert0(!"unsupported data type.\n");
return AV_PIX_FMT_BGR24;
}
int ff_frame_to_dnn_classify(AVFrame *frame, DNNData *input, uint32_t bbox_index, void *log_ctx)
{
const AVPixFmtDescriptor *desc;
int offsetx[4], offsety[4];
uint8_t *bbox_data[4];
struct SwsContext *sws_ctx;
int linesizes[4];
int ret = 0;
enum AVPixelFormat fmt;
int left, top, width, height;
int width_idx, height_idx;
const AVDetectionBBoxHeader *header;
const AVDetectionBBox *bbox;
AVFrameSideData *sd = av_frame_get_side_data(frame, AV_FRAME_DATA_DETECTION_BBOXES);
int max_step[4] = { 0 };
av_assert0(sd);
/* (scale != 1 and scale != 0) or mean != 0 */
if ((fabsf(input->scale - 1) > 1e-6f && fabsf(input->scale) > 1e-6f) ||
fabsf(input->mean) > 1e-6f) {
av_log(log_ctx, AV_LOG_ERROR, "dnn_classify input data doesn't support "
"scale: %f, mean: %f\n", input->scale, input->mean);
return AVERROR(ENOSYS);
}
if (input->layout == DL_NCHW) {
av_log(log_ctx, AV_LOG_ERROR, "dnn_classify input data doesn't support layout: NCHW\n");
return AVERROR(ENOSYS);
}
width_idx = dnn_get_width_idx_by_layout(input->layout);
height_idx = dnn_get_height_idx_by_layout(input->layout);
header = (const AVDetectionBBoxHeader *)sd->data;
bbox = av_get_detection_bbox(header, bbox_index);
left = bbox->x;
width = bbox->w;
top = bbox->y;
height = bbox->h;
fmt = get_pixel_format(input);
sws_ctx = sws_getContext(width, height, frame->format,
input->dims[width_idx],
input->dims[height_idx], fmt,
SWS_FAST_BILINEAR, NULL, NULL, NULL);
if (!sws_ctx) {
av_log(log_ctx, AV_LOG_ERROR, "Failed to create scale context for the conversion "
"fmt:%s s:%dx%d -> fmt:%s s:%dx%d\n",
av_get_pix_fmt_name(frame->format), width, height,
av_get_pix_fmt_name(fmt),
input->dims[width_idx],
input->dims[height_idx]);
return AVERROR(EINVAL);
}
ret = av_image_fill_linesizes(linesizes, fmt, input->dims[width_idx]);
if (ret < 0) {
av_log(log_ctx, AV_LOG_ERROR, "unable to get linesizes with av_image_fill_linesizes");
sws_freeContext(sws_ctx);
return ret;
}
desc = av_pix_fmt_desc_get(frame->format);
offsetx[1] = offsetx[2] = AV_CEIL_RSHIFT(left, desc->log2_chroma_w);
offsetx[0] = offsetx[3] = left;
offsety[1] = offsety[2] = AV_CEIL_RSHIFT(top, desc->log2_chroma_h);
offsety[0] = offsety[3] = top;
av_image_fill_max_pixsteps(max_step, NULL, desc);
for (int k = 0; frame->data[k]; k++)
bbox_data[k] = frame->data[k] + offsety[k] * frame->linesize[k] + offsetx[k] * max_step[k];
sws_scale(sws_ctx, (const uint8_t *const *)&bbox_data, frame->linesize,
0, height,
(uint8_t *const [4]){input->data, 0, 0, 0}, linesizes);
sws_freeContext(sws_ctx);
return ret;
}
int ff_frame_to_dnn_detect(AVFrame *frame, DNNData *input, void *log_ctx)
{
struct SwsContext *sws_ctx;
int linesizes[4];
int ret = 0, width_idx, height_idx;
enum AVPixelFormat fmt = get_pixel_format(input);
/* (scale != 1 and scale != 0) or mean != 0 */
if ((fabsf(input->scale - 1) > 1e-6f && fabsf(input->scale) > 1e-6f) ||
fabsf(input->mean) > 1e-6f) {
av_log(log_ctx, AV_LOG_ERROR, "dnn_detect input data doesn't support "
"scale: %f, mean: %f\n", input->scale, input->mean);
return AVERROR(ENOSYS);
}
if (input->layout == DL_NCHW) {
av_log(log_ctx, AV_LOG_ERROR, "dnn_detect input data doesn't support layout: NCHW\n");
return AVERROR(ENOSYS);
}
width_idx = dnn_get_width_idx_by_layout(input->layout);
height_idx = dnn_get_height_idx_by_layout(input->layout);
sws_ctx = sws_getContext(frame->width, frame->height, frame->format,
input->dims[width_idx],
input->dims[height_idx], fmt,
SWS_FAST_BILINEAR, NULL, NULL, NULL);
if (!sws_ctx) {
av_log(log_ctx, AV_LOG_ERROR, "Impossible to create scale context for the conversion "
"fmt:%s s:%dx%d -> fmt:%s s:%dx%d\n",
av_get_pix_fmt_name(frame->format), frame->width, frame->height,
av_get_pix_fmt_name(fmt), input->dims[width_idx],
input->dims[height_idx]);
return AVERROR(EINVAL);
}
ret = av_image_fill_linesizes(linesizes, fmt, input->dims[width_idx]);
if (ret < 0) {
av_log(log_ctx, AV_LOG_ERROR, "unable to get linesizes with av_image_fill_linesizes");
sws_freeContext(sws_ctx);
return ret;
}
sws_scale(sws_ctx, (const uint8_t *const *)frame->data, frame->linesize, 0, frame->height,
(uint8_t *const [4]){input->data, 0, 0, 0}, linesizes);
sws_freeContext(sws_ctx);
return ret;
}
+38
View File
@@ -0,0 +1,38 @@
/*
* Copyright (c) 2020
*
* This file is part of FFmpeg.
*
* FFmpeg is free software; you can redistribute it and/or
* modify it under the terms of the GNU Lesser General Public
* License as published by the Free Software Foundation; either
* version 2.1 of the License, or (at your option) any later version.
*
* FFmpeg is distributed in the hope that it will be useful,
* but WITHOUT ANY WARRANTY; without even the implied warranty of
* MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the GNU
* Lesser General Public License for more details.
*
* You should have received a copy of the GNU Lesser General Public
* License along with FFmpeg; if not, write to the Free Software
* Foundation, Inc., 51 Franklin Street, Fifth Floor, Boston, MA 02110-1301 USA
*/
/**
* @file
* DNN input&output process between AVFrame and DNNData.
*/
#ifndef AVFILTER_DNN_DNN_IO_PROC_H
#define AVFILTER_DNN_DNN_IO_PROC_H
#include "../dnn_interface.h"
#include "libavutil/frame.h"
int ff_proc_from_frame_to_dnn(AVFrame *frame, DNNData *input, void *log_ctx);
int ff_proc_from_dnn_to_frame(AVFrame *frame, DNNData *output, void *log_ctx);
int ff_frame_to_dnn_detect(AVFrame *frame, DNNData *input, void *log_ctx);
int ff_frame_to_dnn_classify(AVFrame *frame, DNNData *input, uint32_t bbox_index, void *log_ctx);
#endif
+191
View File
@@ -0,0 +1,191 @@
/*
* Copyright (c) 2020
*
* This file is part of FFmpeg.
*
* FFmpeg is free software; you can redistribute it and/or
* modify it under the terms of the GNU Lesser General Public
* License as published by the Free Software Foundation; either
* version 2.1 of the License, or (at your option) any later version.
*
* FFmpeg is distributed in the hope that it will be useful,
* but WITHOUT ANY WARRANTY; without even the implied warranty of
* MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the GNU
* Lesser General Public License for more details.
*
* You should have received a copy of the GNU Lesser General Public
* License along with FFmpeg; if not, write to the Free Software
* Foundation, Inc., 51 Franklin Street, Fifth Floor, Boston, MA 02110-1301 USA
*/
#include <stdio.h>
#include "queue.h"
#include "libavutil/mem.h"
typedef struct QueueEntry QueueEntry;
struct QueueEntry {
void *value;
QueueEntry *prev;
QueueEntry *next;
};
struct Queue {
QueueEntry *head;
QueueEntry *tail;
size_t length;
};
static inline QueueEntry *create_entry(void *val)
{
QueueEntry *entry = av_malloc(sizeof(*entry));
if (entry)
entry->value = val;
return entry;
}
Queue* ff_queue_create(void)
{
Queue *q = av_malloc(sizeof(*q));
if (!q)
return NULL;
q->head = create_entry(q);
q->tail = create_entry(q);
if (!q->head || !q->tail) {
av_freep(&q->head);
av_freep(&q->tail);
av_freep(&q);
return NULL;
}
q->head->next = q->tail;
q->tail->prev = q->head;
q->head->prev = NULL;
q->tail->next = NULL;
q->length = 0;
return q;
}
void ff_queue_destroy(Queue *q)
{
QueueEntry *entry;
if (!q)
return;
entry = q->head;
while (entry != NULL) {
QueueEntry *temp = entry;
entry = entry->next;
av_freep(&temp);
}
av_freep(&q);
}
size_t ff_queue_size(Queue *q)
{
return q ? q->length : 0;
}
void *ff_queue_peek_front(Queue *q)
{
if (!q || q->length == 0)
return NULL;
return q->head->next->value;
}
void *ff_queue_peek_back(Queue *q)
{
if (!q || q->length == 0)
return NULL;
return q->tail->prev->value;
}
int ff_queue_push_front(Queue *q, void *v)
{
QueueEntry *new_entry;
QueueEntry *original_next;
if (!q)
return 0;
new_entry = create_entry(v);
if (!new_entry)
return -1;
original_next = q->head->next;
q->head->next = new_entry;
original_next->prev = new_entry;
new_entry->prev = q->head;
new_entry->next = original_next;
q->length++;
return q->length;
}
int ff_queue_push_back(Queue *q, void *v)
{
QueueEntry *new_entry;
QueueEntry *original_prev;
if (!q)
return 0;
new_entry = create_entry(v);
if (!new_entry)
return -1;
original_prev = q->tail->prev;
q->tail->prev = new_entry;
original_prev->next = new_entry;
new_entry->next = q->tail;
new_entry->prev = original_prev;
q->length++;
return q->length;
}
void *ff_queue_pop_front(Queue *q)
{
QueueEntry *front;
QueueEntry *new_head_next;
void *ret;
if (!q || q->length == 0)
return NULL;
front = q->head->next;
new_head_next = front->next;
ret = front->value;
q->head->next = new_head_next;
new_head_next->prev = q->head;
av_freep(&front);
q->length--;
return ret;
}
void *ff_queue_pop_back(Queue *q)
{
QueueEntry *back;
QueueEntry *new_tail_prev;
void *ret;
if (!q || q->length == 0)
return NULL;
back = q->tail->prev;
new_tail_prev = back->prev;
ret = back->value;
q->tail->prev = new_tail_prev;
new_tail_prev->next = q->tail;
av_freep(&back);
q->length--;
return ret;
}
+117
View File
@@ -0,0 +1,117 @@
/*
* Copyright (c) 2020
*
* This file is part of FFmpeg.
*
* FFmpeg is free software; you can redistribute it and/or
* modify it under the terms of the GNU Lesser General Public
* License as published by the Free Software Foundation; either
* version 2.1 of the License, or (at your option) any later version.
*
* FFmpeg is distributed in the hope that it will be useful,
* but WITHOUT ANY WARRANTY; without even the implied warranty of
* MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the GNU
* Lesser General Public License for more details.
*
* You should have received a copy of the GNU Lesser General Public
* License along with FFmpeg; if not, write to the Free Software
* Foundation, Inc., 51 Franklin Street, Fifth Floor, Boston, MA 02110-1301 USA
*/
#include <stddef.h>
#ifndef AVFILTER_DNN_QUEUE_H
#define AVFILTER_DNN_QUEUE_H
/**
* Linear double-ended data structure
*/
typedef struct Queue Queue;
/**
* @brief Create a Queue instance.
* It initializes the length of the Queue as 0.
*
* @return Pointer to the Queue
* @retval NULL if allocation fails
*/
Queue *ff_queue_create(void);
/**
* @brief Destroy the Queue instance.
* It also frees all elements of the Queue.
*/
void ff_queue_destroy(Queue *q);
/**
* @brief Return the length of the Queue
*/
size_t ff_queue_size(Queue *q);
/**
* @brief Return a pointer to the data
* at the head of the queue.
*
* @retval NULL if null pointer was passed
* or queue is empty
*/
void *ff_queue_peek_front(Queue *q);
/**
* @brief Return a pointer to the data at
* the tail of the queue.
*
* @retval NULL if null pointer was passed
* or queue is empty
*/
void *ff_queue_peek_back(Queue *q);
/**
* @brief Add data to the head of the queue.
* It increases the length of Queue by one.
*
* @param q pointer to the Queue.
* @param v data to be added
* @return The length of the Queue
* @retval 0 if the pointer to queue is NULL
* @retval -1 if new entry cannot be created
*/
int ff_queue_push_front(Queue *q, void *v);
/**
* @brief Add data to the tail of the queue.
* It increases the length of Queue by one.
*
* @param q pointer to the Queue
* @param v data to be added
* @return The length of the Queue
* @retval 0 if the pointer to queue is NULL
* @retval -1 if new entry cannot be created
*/
int ff_queue_push_back(Queue *q, void *v);
/**
* @brief Remove and free first element from
* the Queue. It shrinks the length of Queue
* by one.
*
* @param q pointer to the Queue.
* @return The value of first element as void.
* If a null pointer or empty queue is passed,
* it returns NULL
*/
void *ff_queue_pop_front(Queue *q);
/**
* @brief Remove and free last element from
* the Queue. It shrinks the length of Queue
* by one.
*
* @param q pointer to the Queue.
* @return The value of last element as void.
* If a null pointer or empty queue is passed,
* it returns NULL
*/
void *ff_queue_pop_back(Queue *q);
#endif
+116
View File
@@ -0,0 +1,116 @@
/*
* Copyright (c) 2020
*
* This file is part of FFmpeg.
*
* FFmpeg is free software; you can redistribute it and/or
* modify it under the terms of the GNU Lesser General Public
* License as published by the Free Software Foundation; either
* version 2.1 of the License, or (at your option) any later version.
*
* FFmpeg is distributed in the hope that it will be useful,
* but WITHOUT ANY WARRANTY; without even the implied warranty of
* MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the GNU
* Lesser General Public License for more details.
*
* You should have received a copy of the GNU Lesser General Public
* License along with FFmpeg; if not, write to the Free Software
* Foundation, Inc., 51 Franklin Street, Fifth Floor, Boston, MA 02110-1301 USA
*/
#include <stdio.h>
#include "queue.h"
#include "safe_queue.h"
#include "libavutil/mem.h"
#include "libavutil/avassert.h"
#include "libavutil/thread.h"
#if HAVE_PTHREAD_CANCEL
#define DNNCond pthread_cond_t
#define dnn_cond_init pthread_cond_init
#define dnn_cond_destroy pthread_cond_destroy
#define dnn_cond_signal pthread_cond_signal
#define dnn_cond_wait pthread_cond_wait
#else
#define DNNCond char
static inline int dnn_cond_init(DNNCond *cond, const void *attr) { return 0; }
static inline int dnn_cond_destroy(DNNCond *cond) { return 0; }
static inline int dnn_cond_signal(DNNCond *cond) { return 0; }
static inline int dnn_cond_wait(DNNCond *cond, AVMutex *mutex)
{
av_assert0(!"should not reach here");
return 0;
}
#endif
struct SafeQueue {
Queue *q;
AVMutex mutex;
DNNCond cond;
};
SafeQueue *ff_safe_queue_create(void)
{
SafeQueue *sq = av_malloc(sizeof(*sq));
if (!sq)
return NULL;
sq->q = ff_queue_create();
if (!sq->q) {
av_freep(&sq);
return NULL;
}
ff_mutex_init(&sq->mutex, NULL);
dnn_cond_init(&sq->cond, NULL);
return sq;
}
void ff_safe_queue_destroy(SafeQueue *sq)
{
if (!sq)
return;
ff_queue_destroy(sq->q);
ff_mutex_destroy(&sq->mutex);
dnn_cond_destroy(&sq->cond);
av_freep(&sq);
}
size_t ff_safe_queue_size(SafeQueue *sq)
{
return sq ? ff_queue_size(sq->q) : 0;
}
int ff_safe_queue_push_front(SafeQueue *sq, void *v)
{
int ret;
ff_mutex_lock(&sq->mutex);
ret = ff_queue_push_front(sq->q, v);
dnn_cond_signal(&sq->cond);
ff_mutex_unlock(&sq->mutex);
return ret;
}
int ff_safe_queue_push_back(SafeQueue *sq, void *v)
{
int ret;
ff_mutex_lock(&sq->mutex);
ret = ff_queue_push_back(sq->q, v);
dnn_cond_signal(&sq->cond);
ff_mutex_unlock(&sq->mutex);
return ret;
}
void *ff_safe_queue_pop_front(SafeQueue *sq)
{
void *value;
ff_mutex_lock(&sq->mutex);
while (ff_queue_size(sq->q) == 0) {
dnn_cond_wait(&sq->cond, &sq->mutex);
}
value = ff_queue_pop_front(sq->q);
dnn_cond_signal(&sq->cond);
ff_mutex_unlock(&sq->mutex);
return value;
}
+96
View File
@@ -0,0 +1,96 @@
/*
* Copyright (c) 2020
*
* This file is part of FFmpeg.
*
* FFmpeg is free software; you can redistribute it and/or
* modify it under the terms of the GNU Lesser General Public
* License as published by the Free Software Foundation; either
* version 2.1 of the License, or (at your option) any later version.
*
* FFmpeg is distributed in the hope that it will be useful,
* but WITHOUT ANY WARRANTY; without even the implied warranty of
* MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the GNU
* Lesser General Public License for more details.
*
* You should have received a copy of the GNU Lesser General Public
* License along with FFmpeg; if not, write to the Free Software
* Foundation, Inc., 51 Franklin Street, Fifth Floor, Boston, MA 02110-1301 USA
*/
#ifndef AVFILTER_DNN_SAFE_QUEUE_H
#define AVFILTER_DNN_SAFE_QUEUE_H
/**
* Double-ended queue with mutex locks ensuring
* data consistency while multithreading.
*/
typedef struct SafeQueue SafeQueue;
/**
* @brief Create and initialize a SafeQueue instance.
*
* @return Pointer to the SafeQueue
* @retval NULL if initialization fails
*/
SafeQueue *ff_safe_queue_create(void);
/**
* @brief Destroy the SafeQueue instance.
* It also frees all elements of the queue,
* destroys the mutex and condition variable.
*/
void ff_safe_queue_destroy(SafeQueue *sq);
/**
* @brief Return the length of the SafeQueue
*/
size_t ff_safe_queue_size(SafeQueue *sq);
/**
* @brief Add data to the head of queue in the
* SafeQueue after locking mutex. After adding
* the data, it signals the condition variable
* and unlocks the mutex. It increases the length
* of queue in the SafeQueue by one.
*
* @param sq pointer to the SafeQueue
* @param v data to be added
* @return The length of the queue
* @retval 0 if the queue is not initialized
* @retval -1 if new entry cannot be created
*/
int ff_safe_queue_push_front(SafeQueue *sq, void *v);
/**
* @brief Add data to the tail of queue in the
* SafeQueue after locking mutex. After adding
* the data, it signals the condition variable
* and unlocks the mutex. It increases the length
* of queue in the SafeQueue by one.
*
* @param sq pointer to the SafeQueue
* @param v data to be added
* @return The length of the queue
* @retval 0 if the queue is not initialized
* @retval -1 if new entry cannot be created
*/
int ff_safe_queue_push_back(SafeQueue *sq, void *v);
/**
* @brief Remove and free first element from
* the queue in SafeQueue. Before removing, it
* waits for the condition variable to signal and
* acquires the mutex. Finally, it signals the
* condition and unlocks the mutex.
* It shrinks the length of queue in the SafeQueue
* by one.
*
* @param sq pointer to the SafeQueue.
* @return The value of first element as void.
* If a null pointer or empty queue is passed,
* it returns NULL
*/
void *ff_safe_queue_pop_front(SafeQueue *sq);
#endif