OpenVINO is a Deep Learning Deployment Toolkit at https://github.com/openvinotoolkit/openvino, it supports CPU, GPU and heterogeneous plugins to accelerate deep learning inferencing. Please refer to https://github.com/openvinotoolkit/openvino/blob/master/build-instruction.md to build openvino (c library is built at the same time). Please add option -DENABLE_MKL_DNN=ON for cmake to enable CPU path. The header files and libraries are installed to /usr/local/deployment_tools/inference_engine/ with default options on my system. To build FFmpeg with openvion, take my system as an example, run with: $ export LD_LIBRARY_PATH=$LD_LIBRARY_PATH:/usr/local/deployment_tools/inference_engine/lib/intel64/:/usr/local/deployment_tools/inference_engine/external/tbb/lib/ $ ../ffmpeg/configure --enable-libopenvino --extra-cflags=-I/usr/local/deployment_tools/inference_engine/include/ --extra-ldflags=-L/usr/local/deployment_tools/inference_engine/lib/intel64 $ make Here are the features provided by OpenVINO inference engine: - support more DNN model formats It supports TensorFlow, Caffe, ONNX, MXNet and Kaldi by converting them into OpenVINO format with a python script. And torth model can be first converted into ONNX and then to OpenVINO format. see the script at https://github.com/openvinotoolkit/openvino/tree/master/model-optimizer/mo.py which also does some optimization at model level. - optimize at inference stage It optimizes for X86 CPUs with SSE, AVX etc. It also optimizes based on OpenCL for Intel GPUs. (only Intel GPU supported becuase Intel OpenCL extension is used for optimization)
Unified · Split
| @@ -253,6 +253,8 @@ External library support: · 252 unmodified lines | |||
| 253 | --enable-libopenh264 enable H.264 encoding via OpenH264 [no] | 253 | --enable-libopenh264 enable H.264 encoding via OpenH264 [no] |
| 254 | --enable-libopenjpeg enable JPEG 2000 de/encoding via OpenJPEG [no] | 254 | --enable-libopenjpeg enable JPEG 2000 de/encoding via OpenJPEG [no] |
| 255 | --enable-libopenmpt enable decoding tracked files via libopenmpt [no] | 255 | --enable-libopenmpt enable decoding tracked files via libopenmpt [no] |
| 256 | --enable-libopenvino enable OpenVINO as a DNN module backend | ||
| 257 | for DNN based filters like dnn_processing [no] | ||
| 256 | --enable-libopus enable Opus de/encoding via libopus [no] | 258 | --enable-libopus enable Opus de/encoding via libopus [no] |
| 257 | --enable-libpulse enable Pulseaudio input via libpulse [no] | 259 | --enable-libpulse enable Pulseaudio input via libpulse [no] |
| 258 | --enable-librabbitmq enable RabbitMQ library [no] | 260 | --enable-librabbitmq enable RabbitMQ library [no] |
| @@ -1790,6 +1792,7 @@ EXTERNAL_LIBRARY_LIST=" · 1531 unmodified lines | |||
| 1790 | libopenh264 | 1792 | libopenh264 |
| 1791 | libopenjpeg | 1793 | libopenjpeg |
| 1792 | libopenmpt | 1794 | libopenmpt |
| 1795 | libopenvino | ||
| 1793 | libopus | 1796 | libopus |
| 1794 | libpulse | 1797 | libpulse |
| 1795 | librabbitmq | 1798 | librabbitmq |
| @@ -2620,7 +2623,7 @@ cbs_mpeg2_select="cbs" · 824 unmodified lines | |||
| 2620 | cbs_vp9_select="cbs" | 2623 | cbs_vp9_select="cbs" |
| 2621 | dct_select="rdft" | 2624 | dct_select="rdft" |
| 2622 | dirac_parse_select="golomb" | 2625 | dirac_parse_select="golomb" |
| 2623 | dnn_suggest="libtensorflow" | 2626 | dnn_suggest="libtensorflow libopenvino" |
| 2624 | error_resilience_select="me_cmp" | 2627 | error_resilience_select="me_cmp" |
| 2625 | faandct_deps="faan" | 2628 | faandct_deps="faan" |
| 2626 | faandct_select="fdctdsp" | 2629 | faandct_select="fdctdsp" |
| @@ -6350,6 +6353,7 @@ enabled libopenh264 && require_pkg_config libopenh264 openh264 wels/codec_ · 3723 unmodified lines | |||
| 6350 | enabled libopenjpeg && { check_pkg_config libopenjpeg "libopenjp2 >= 2.1.0" openjpeg.h opj_version || | 6353 | enabled libopenjpeg && { check_pkg_config libopenjpeg "libopenjp2 >= 2.1.0" openjpeg.h opj_version || |
| 6351 | { require_pkg_config libopenjpeg "libopenjp2 >= 2.1.0" openjpeg.h opj_version -DOPJ_STATIC && add_cppflags -DOPJ_STATIC; } } | 6354 | { require_pkg_config libopenjpeg "libopenjp2 >= 2.1.0" openjpeg.h opj_version -DOPJ_STATIC && add_cppflags -DOPJ_STATIC; } } |
| 6352 | enabled libopenmpt && require_pkg_config libopenmpt "libopenmpt >= 0.2.6557" libopenmpt/libopenmpt.h openmpt_module_create -lstdc++ && append libopenmpt_extralibs "-lstdc++" | 6355 | enabled libopenmpt && require_pkg_config libopenmpt "libopenmpt >= 0.2.6557" libopenmpt/libopenmpt.h openmpt_module_create -lstdc++ && append libopenmpt_extralibs "-lstdc++" |
| 6356 | enabled libopenvino && require libopenvino c_api/ie_c_api.h ie_c_api_version -linference_engine_c_api | ||
| 6353 | enabled libopus && { | 6357 | enabled libopus && { |
| 6354 | enabled libopus_decoder && { | 6358 | enabled libopus_decoder && { |
| 6355 | require_pkg_config libopus opus opus_multistream.h opus_multistream_decoder_create | 6359 | require_pkg_config libopus opus opus_multistream.h opus_multistream_decoder_create |
| @@ -9,5 +9,6 @@ OBJS-$(CONFIG_DNN) += dnn/dnn_backend_native_layer_mat · 8 unmodified lines | |||
| 9 | OBJS-$(CONFIG_DNN) += dnn/dnn_backend_native_layer_mathunary.o | 9 | OBJS-$(CONFIG_DNN) += dnn/dnn_backend_native_layer_mathunary.o |
| 10 | 10 | ||
| 11 | DNN-OBJS-$(CONFIG_LIBTENSORFLOW) += dnn/dnn_backend_tf.o | 11 | DNN-OBJS-$(CONFIG_LIBTENSORFLOW) += dnn/dnn_backend_tf.o |
| 12 | DNN-OBJS-$(CONFIG_LIBOPENVINO) += dnn/dnn_backend_openvino.o | ||
| 12 | 13 | ||
| 13 | OBJS-$(CONFIG_DNN) += $(DNN-OBJS-yes) | 14 | OBJS-$(CONFIG_DNN) += $(DNN-OBJS-yes) |
| @@ -0,0 +1,261 @@ | |||
| 1 | /* | ||
| 2 | * Copyright (c) 2020 | ||
| 3 | * | ||
| 4 | * This file is part of FFmpeg. | ||
| 5 | * | ||
| 6 | * FFmpeg is free software; you can redistribute it and/or | ||
| 7 | * modify it under the terms of the GNU Lesser General Public | ||
| 8 | * License as published by the Free Software Foundation; either | ||
| 9 | * version 2.1 of the License, or (at your option) any later version. | ||
| 10 | * | ||
| 11 | * FFmpeg is distributed in the hope that it will be useful, | ||
| 12 | * but WITHOUT ANY WARRANTY; without even the implied warranty of | ||
| 13 | * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the GNU | ||
| 14 | * Lesser General Public License for more details. | ||
| 15 | * | ||
| 16 | * You should have received a copy of the GNU Lesser General Public | ||
| 17 | * License along with FFmpeg; if not, write to the Free Software | ||
| 18 | * Foundation, Inc., 51 Franklin Street, Fifth Floor, Boston, MA 02110-1301 USA | ||
| 19 | */ | ||
| 20 | |||
| 21 | /** | ||
| 22 | * @file | ||
| 23 | * DNN OpenVINO backend implementation. | ||
| 24 | */ | ||
| 25 | |||
| 26 | #include "dnn_backend_openvino.h" | ||
| 27 | #include "libavformat/avio.h" | ||
| 28 | #include "libavutil/avassert.h" | ||
| 29 | #include <c_api/ie_c_api.h> | ||
| 30 | |||
| 31 | typedef struct OVModel{ | ||
| 32 | ie_core_t *core; | ||
| 33 | ie_network_t *network; | ||
| 34 | ie_executable_network_t *exe_network; | ||
| 35 | ie_infer_request_t *infer_request; | ||
| 36 | ie_blob_t *input_blob; | ||
| 37 | ie_blob_t **output_blobs; | ||
| 38 | uint32_t nb_output; | ||
| 39 | } OVModel; | ||
| 40 | |||
| 41 | static DNNDataType precision_to_datatype(precision_e precision) | ||
| 42 | { | ||
| 43 | switch (precision) | ||
| 44 | { | ||
| 45 | case FP32: | ||
| 46 | return DNN_FLOAT; | ||
| 47 | default: | ||
| 48 | av_assert0(!"not supported yet."); | ||
| 49 | return DNN_FLOAT; | ||
| 50 | } | ||
| 51 | } | ||
| 52 | |||
| 53 | static DNNReturnType get_input_ov(void *model, DNNData *input, const char *input_name) | ||
| 54 | { | ||
| 55 | OVModel *ov_model = (OVModel *)model; | ||
| 56 | char *model_input_name = NULL; | ||
| 57 | IEStatusCode status; | ||
| 58 | size_t model_input_count = 0; | ||
| 59 | dimensions_t dims; | ||
| 60 | precision_e precision; | ||
| 61 | |||
| 62 | status = ie_network_get_inputs_number(ov_model->network, &model_input_count); | ||
| 63 | if (status != OK) | ||
| 64 | return DNN_ERROR; | ||
| 65 | |||
| 66 | for (size_t i = 0; i < model_input_count; i++) { | ||
| 67 | status = ie_network_get_input_name(ov_model->network, i, &model_input_name); | ||
| 68 | if (status != OK) | ||
| 69 | return DNN_ERROR; | ||
| 70 | if (strcmp(model_input_name, input_name) == 0) { | ||
| 71 | ie_network_name_free(&model_input_name); | ||
| 72 | status |= ie_network_get_input_dims(ov_model->network, input_name, &dims); | ||
| 73 | status |= ie_network_get_input_precision(ov_model->network, input_name, &precision); | ||
| 74 | if (status != OK) | ||
| 75 | return DNN_ERROR; | ||
| 76 | |||
| 77 | // The order of dims in the openvino is fixed and it is always NCHW for 4-D data. | ||
| 78 | // while we pass NHWC data from FFmpeg to openvino | ||
| 79 | status = ie_network_set_input_layout(ov_model->network, input_name, NHWC); | ||
| 80 | if (status != OK) | ||
| 81 | return DNN_ERROR; | ||
| 82 | |||
| 83 | input->channels = dims.dims[1]; | ||
| 84 | input->height = dims.dims[2]; | ||
| 85 | input->width = dims.dims[3]; | ||
| 86 | input->dt = precision_to_datatype(precision); | ||
| 87 | return DNN_SUCCESS; | ||
| 88 | } | ||
| 89 | |||
| 90 | ie_network_name_free(&model_input_name); | ||
| 91 | } | ||
| 92 | |||
| 93 | return DNN_ERROR; | ||
| 94 | } | ||
| 95 | |||
| 96 | static DNNReturnType set_input_output_ov(void *model, DNNData *input, const char *input_name, const char **output_names, uint32_t nb_output) | ||
| 97 | { | ||
| 98 | OVModel *ov_model = (OVModel *)model; | ||
| 99 | IEStatusCode status; | ||
| 100 | dimensions_t dims; | ||
| 101 | precision_e precision; | ||
| 102 | ie_blob_buffer_t blob_buffer; | ||
| 103 | |||
| 104 | status = ie_exec_network_create_infer_request(ov_model->exe_network, &ov_model->infer_request); | ||
| 105 | if (status != OK) | ||
| 106 | goto err; | ||
| 107 | |||
| 108 | status = ie_infer_request_get_blob(ov_model->infer_request, input_name, &ov_model->input_blob); | ||
| 109 | if (status != OK) | ||
| 110 | goto err; | ||
| 111 | |||
| 112 | status |= ie_blob_get_dims(ov_model->input_blob, &dims); | ||
| 113 | status |= ie_blob_get_precision(ov_model->input_blob, &precision); | ||
| 114 | if (status != OK) | ||
| 115 | goto err; | ||
| 116 | |||
| 117 | av_assert0(input->channels == dims.dims[1]); | ||
| 118 | av_assert0(input->height == dims.dims[2]); | ||
| 119 | av_assert0(input->width == dims.dims[3]); | ||
| 120 | av_assert0(input->dt == precision_to_datatype(precision)); | ||
| 121 | |||
| 122 | status = ie_blob_get_buffer(ov_model->input_blob, &blob_buffer); | ||
| 123 | if (status != OK) | ||
| 124 | goto err; | ||
| 125 | input->data = blob_buffer.buffer; | ||
| 126 | |||
| 127 | // outputs | ||
| 128 | ov_model->nb_output = 0; | ||
| 129 | av_freep(&ov_model->output_blobs); | ||
| 130 | ov_model->output_blobs = av_mallocz_array(nb_output, sizeof(*ov_model->output_blobs)); | ||
| 131 | if (!ov_model->output_blobs) | ||
| 132 | goto err; | ||
| 133 | |||
| 134 | for (int i = 0; i < nb_output; i++) { | ||
| 135 | const char *output_name = output_names[i]; | ||
| 136 | status = ie_infer_request_get_blob(ov_model->infer_request, output_name, &(ov_model->output_blobs[i])); | ||
| 137 | if (status != OK) | ||
| 138 | goto err; | ||
| 139 | ov_model->nb_output++; | ||
| 140 | } | ||
| 141 | |||
| 142 | return DNN_SUCCESS; | ||
| 143 | |||
| 144 | err: | ||
| 145 | if (ov_model->output_blobs) { | ||
| 146 | for (uint32_t i = 0; i < ov_model->nb_output; i++) { | ||
| 147 | ie_blob_free(&(ov_model->output_blobs[i])); | ||
| 148 | } | ||
| 149 | av_freep(&ov_model->output_blobs); | ||
| 150 | } | ||
| 151 | if (ov_model->input_blob) | ||
| 152 | ie_blob_free(&ov_model->input_blob); | ||
| 153 | if (ov_model->infer_request) | ||
| 154 | ie_infer_request_free(&ov_model->infer_request); | ||
| 155 | return DNN_ERROR; | ||
| 156 | } | ||
| 157 | |||
| 158 | DNNModel *ff_dnn_load_model_ov(const char *model_filename) | ||
| 159 | { | ||
| 160 | DNNModel *model = NULL; | ||
| 161 | OVModel *ov_model = NULL; | ||
| 162 | IEStatusCode status; | ||
| 163 | ie_config_t config = {NULL, NULL, NULL}; | ||
| 164 | |||
| 165 | model = av_malloc(sizeof(DNNModel)); | ||
| 166 | if (!model){ | ||
| 167 | return NULL; | ||
| 168 | } | ||
| 169 | |||
| 170 | ov_model = av_mallocz(sizeof(OVModel)); | ||
| 171 | if (!ov_model) | ||
| 172 | goto err; | ||
| 173 | |||
| 174 | status = ie_core_create("", &ov_model->core); | ||
| 175 | if (status != OK) | ||
| 176 | goto err; | ||
| 177 | |||
| 178 | status = ie_core_read_network(ov_model->core, model_filename, NULL, &ov_model->network); | ||
| 179 | if (status != OK) | ||
| 180 | goto err; | ||
| 181 | |||
| 182 | status = ie_core_load_network(ov_model->core, ov_model->network, "CPU", &config, &ov_model->exe_network); | ||
| 183 | if (status != OK) | ||
| 184 | goto err; | ||
| 185 | |||
| 186 | model->model = (void *)ov_model; | ||
| 187 | model->set_input_output = &set_input_output_ov; | ||
| 188 | model->get_input = &get_input_ov; | ||
| 189 | |||
| 190 | return model; | ||
| 191 | |||
| 192 | err: | ||
| 193 | if (model) | ||
| 194 | av_freep(&model); | ||
| 195 | if (ov_model) { | ||
| 196 | if (ov_model->exe_network) | ||
| 197 | ie_exec_network_free(&ov_model->exe_network); | ||
| 198 | if (ov_model->network) | ||
| 199 | ie_network_free(&ov_model->network); | ||
| 200 | if (ov_model->core) | ||
| 201 | ie_core_free(&ov_model->core); | ||
| 202 | av_freep(&ov_model); | ||
| 203 | } | ||
| 204 | return NULL; | ||
| 205 | } | ||
| 206 | |||
| 207 | DNNReturnType ff_dnn_execute_model_ov(const DNNModel *model, DNNData *outputs, uint32_t nb_output) | ||
| 208 | { | ||
| 209 | dimensions_t dims; | ||
| 210 | precision_e precision; | ||
| 211 | ie_blob_buffer_t blob_buffer; | ||
| 212 | OVModel *ov_model = (OVModel *)model->model; | ||
| 213 | uint32_t nb = FFMIN(nb_output, ov_model->nb_output); | ||
| 214 | IEStatusCode status = ie_infer_request_infer(ov_model->infer_request); | ||
| 215 | if (status != OK) | ||
| 216 | return DNN_ERROR; | ||
| 217 | |||
| 218 | for (uint32_t i = 0; i < nb; ++i) { | ||
| 219 | status = ie_blob_get_buffer(ov_model->output_blobs[i], &blob_buffer); | ||
| 220 | if (status != OK) | ||
| 221 | return DNN_ERROR; | ||
| 222 | |||
| 223 | status |= ie_blob_get_dims(ov_model->output_blobs[i], &dims); | ||
| 224 | status |= ie_blob_get_precision(ov_model->output_blobs[i], &precision); | ||
| 225 | if (status != OK) | ||
| 226 | return DNN_ERROR; | ||
| 227 | |||
| 228 | outputs[i].channels = dims.dims[1]; | ||
| 229 | outputs[i].height = dims.dims[2]; | ||
| 230 | outputs[i].width = dims.dims[3]; | ||
| 231 | outputs[i].dt = precision_to_datatype(precision); | ||
| 232 | outputs[i].data = blob_buffer.buffer; | ||
| 233 | } | ||
| 234 | |||
| 235 | return DNN_SUCCESS; | ||
| 236 | } | ||
| 237 | |||
| 238 | void ff_dnn_free_model_ov(DNNModel **model) | ||
| 239 | { | ||
| 240 | if (*model){ | ||
| 241 | OVModel *ov_model = (OVModel *)(*model)->model; | ||
| 242 | if (ov_model->output_blobs) { | ||
| 243 | for (uint32_t i = 0; i < ov_model->nb_output; i++) { | ||
| 244 | ie_blob_free(&(ov_model->output_blobs[i])); | ||
| 245 | } | ||
| 246 | av_freep(&ov_model->output_blobs); | ||
| 247 | } | ||
| 248 | if (ov_model->input_blob) | ||
| 249 | ie_blob_free(&ov_model->input_blob); | ||
| 250 | if (ov_model->infer_request) | ||
| 251 | ie_infer_request_free(&ov_model->infer_request); | ||
| 252 | if (ov_model->exe_network) | ||
| 253 | ie_exec_network_free(&ov_model->exe_network); | ||
| 254 | if (ov_model->network) | ||
| 255 | ie_network_free(&ov_model->network); | ||
| 256 | if (ov_model->core) | ||
| 257 | ie_core_free(&ov_model->core); | ||
| 258 | av_freep(&ov_model); | ||
| 259 | av_freep(model); | ||
| 260 | } | ||
| 261 | } | ||
| @@ -0,0 +1,38 @@ | |||
| 1 | /* | ||
| 2 | * Copyright (c) 2020 | ||
| 3 | * | ||
| 4 | * This file is part of FFmpeg. | ||
| 5 | * | ||
| 6 | * FFmpeg is free software; you can redistribute it and/or | ||
| 7 | * modify it under the terms of the GNU Lesser General Public | ||
| 8 | * License as published by the Free Software Foundation; either | ||
| 9 | * version 2.1 of the License, or (at your option) any later version. | ||
| 10 | * | ||
| 11 | * FFmpeg is distributed in the hope that it will be useful, | ||
| 12 | * but WITHOUT ANY WARRANTY; without even the implied warranty of | ||
| 13 | * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the GNU | ||
| 14 | * Lesser General Public License for more details. | ||
| 15 | * | ||
| 16 | * You should have received a copy of the GNU Lesser General Public | ||
| 17 | * License along with FFmpeg; if not, write to the Free Software | ||
| 18 | * Foundation, Inc., 51 Franklin Street, Fifth Floor, Boston, MA 02110-1301 USA | ||
| 19 | */ | ||
| 20 | |||
| 21 | /** | ||
| 22 | * @file | ||
| 23 | * DNN inference functions interface for OpenVINO backend. | ||
| 24 | */ | ||
| 25 | |||
| 26 | |||
| 27 | #ifndef AVFILTER_DNN_DNN_BACKEND_OPENVINO_H | ||
| 28 | #define AVFILTER_DNN_DNN_BACKEND_OPENVINO_H | ||
| 29 | |||
| 30 | #include "../dnn_interface.h" | ||
| 31 | |||
| 32 | DNNModel *ff_dnn_load_model_ov(const char *model_filename); | ||
| 33 | |||
| 34 | DNNReturnType ff_dnn_execute_model_ov(const DNNModel *model, DNNData *outputs, uint32_t nb_output); | ||
| 35 | |||
| 36 | void ff_dnn_free_model_ov(DNNModel **model); | ||
| 37 | |||
| 38 | #endif | ||
| @@ -26,6 +26,7 @@ · 25 unmodified lines | |||
| 26 | #include "../dnn_interface.h" | 26 | #include "../dnn_interface.h" |
| 27 | #include "dnn_backend_native.h" | 27 | #include "dnn_backend_native.h" |
| 28 | #include "dnn_backend_tf.h" | 28 | #include "dnn_backend_tf.h" |
| 29 | #include "dnn_backend_openvino.h" | ||
| 29 | #include "libavutil/mem.h" | 30 | #include "libavutil/mem.h" |
| 30 | 31 | ||
| 31 | DNNModule *ff_get_dnn_module(DNNBackendType backend_type) | 32 | DNNModule *ff_get_dnn_module(DNNBackendType backend_type) |
| @@ -53,6 +54,16 @@ DNNModule *ff_get_dnn_module(DNNBackendType backend_type) · 21 unmodified lines | |||
| 53 | return NULL; | 54 | return NULL; |
| 54 | #endif | 55 | #endif |
| 55 | break; | 56 | break; |
| 57 | case DNN_OV: | ||
| 58 | #if (CONFIG_LIBOPENVINO == 1) | ||
| 59 | dnn_module->load_model = &ff_dnn_load_model_ov; | ||
| 60 | dnn_module->execute_model = &ff_dnn_execute_model_ov; | ||
| 61 | dnn_module->free_model = &ff_dnn_free_model_ov; | ||
| 62 | #else | ||
| 63 | av_freep(&dnn_module); | ||
| 64 | return NULL; | ||
| 65 | #endif | ||
| 66 | break; | ||
| 56 | default: | 67 | default: |
| 57 | av_log(NULL, AV_LOG_ERROR, "Module backend_type is not native or tensorflow\n"); | 68 | av_log(NULL, AV_LOG_ERROR, "Module backend_type is not native or tensorflow\n"); |
| 58 | av_freep(&dnn_module); | 69 | av_freep(&dnn_module); |
| @@ -30,7 +30,7 @@ · 29 unmodified lines | |||
| 30 | 30 | ||
| 31 | typedef enum {DNN_SUCCESS, DNN_ERROR} DNNReturnType; | 31 | typedef enum {DNN_SUCCESS, DNN_ERROR} DNNReturnType; |
| 32 | 32 | ||
| 33 | typedef enum {DNN_NATIVE, DNN_TF} DNNBackendType; | 33 | typedef enum {DNN_NATIVE, DNN_TF, DNN_OV} DNNBackendType; |
| 34 | 34 | ||
| 35 | typedef enum {DNN_FLOAT = 1, DNN_UINT8 = 4} DNNDataType; | 35 | typedef enum {DNN_FLOAT = 1, DNN_UINT8 = 4} DNNDataType; |
| 36 | 36 | ||