Numpy reader changes

rocAL/include/api/rocal_api_data_loaders.h

@@ -576,6 +576,46 @@ extern "C" RocalTensor ROCAL_API_CALL rocalRawTFRecordSourceSingleShard(RocalCon
                                                                         unsigned out_width = 0, unsigned out_height = 0,
                                                                         const char* record_name_prefix = "");
 
+/*! \brief Creates Numpy raw data reader and loader. It allocates the resources and objects required to read raw data stored on the numpy arrays.
+ * \ingroup group_rocal_data_loaders
+ * \param [in] context Rocal context
+ * \param [in] source_path A NULL terminated char string pointing to the location on the disk
+ * \param [in] internal_shard_count Defines the parallelism level by internally sharding the input dataset and load/decode using multiple decoder/loader instances. Using shard counts bigger than 1 improves the load/decode performance if compute resources (CPU cores) are available.
+ * \param [in] tensor_layout The tensor layout for the numpy loader tensor specified by the user.
+ * \param [in] is_output Determines if the user wants the loaded numpy arrays to be part of the output or not.
+ * \param [in] shuffle Determines if the user wants to shuffle the dataset or not.
+ * \param [in] loop Determines if the user wants to indefinitely loops through images or not.
+ * \return Reference to the output tensor
+ */
+extern "C"  RocalTensor  ROCAL_API_CALL rocalNumpyFileSource(RocalContext p_context,
+                                                             const char* source_path,
+                                                             unsigned internal_shard_count,
+                                                             RocalTensorLayout tensor_layout = ROCAL_NONE,
+                                                             bool is_output = false,
+                                                             bool shuffle = false,
+                                                             bool loop = false);
+
+/*! \brief Creates Numpy raw data reader and loader. It allocates the resources and objects required to read raw data stored on the numpy arrays.
+ * \ingroup group_rocal_data_loaders
+ * \param [in] context Rocal context
+ * \param [in] source_path A NULL terminated char string pointing to the location on the disk
+ * \param [in] tensor_layout The tensor layout for the numpy loader tensor specified by the user.
+ * \param [in] is_output Determines if the user wants the loaded numpy arrays to be part of the output or not.
+ * \param [in] shuffle Determines if the user wants to shuffle the dataset or not.
+ * \param [in] loop Determines if the user wants to indefinitely loops through images or not.
+ * \param [in] shard_id Shard id for this loader
+ * \param [in] shard_count Total shard count
+ * \return Reference to the output tensor
+ */
+extern "C"  RocalTensor  rocalNumpyFileSourceSingleShard(RocalContext p_context,
+                                                         const char* source_path,
+                                                         RocalTensorLayout tensor_layout = ROCAL_NONE,
+                                                         bool is_output = false,
+                                                         bool shuffle = false,
+                                                         bool loop = false,
+                                                         unsigned shard_id = 0,
+                                                         unsigned shard_count = 1);
+
 /*!
  * \brief Creates a video reader and decoder as a source. It allocates the resources and objects required to read and decode mp4 videos stored on the file systems.
  * \ingroup group_rocal_data_loaders

rocAL/include/api/rocal_api_types.h

@@ -218,9 +218,15 @@ enum RocalTensorLayout {
     /*! \brief AMD ROCAL_NFCHW
      */
     ROCAL_NFCHW = 3,
+    /*! \brief AMD ROCAL_NDHWC
+     */
+    ROCAL_NDHWC = 4,
+    /*! \brief AMD ROCAL_NCDHW
+     */
+    ROCAL_NCDHW = 5,
     /*! \brief AMD ROCAL_NONE
      */
-    ROCAL_NONE = 4  // Layout for generic tensors (Non-Image or Non-Video)
+    ROCAL_NONE = 6  // Layout for generic tensors (Non-Image or Non-Video)
 };
 
 /*! \brief rocAL Tensor Output Type enum

rocAL/include/loaders/image/node_numpy_loader.h

@@ -0,0 +1,49 @@
+/*
+Copyright (c) 2019 - 2024 Advanced Micro Devices, Inc. All rights reserved.
+
+Permission is hereby granted, free of charge, to any person obtaining a copy
+of this software and associated documentation files (the "Software"), to deal
+in the Software without restriction, including without limitation the rights
+to use, copy, modify, merge, publish, distribute, sublicense, and/or sell
+copies of the Software, and to permit persons to whom the Software is
+furnished to do so, subject to the following conditions:
+
+The above copyright notice and this permission notice shall be included in
+all copies or substantial portions of the Software.
+
+THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
+IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
+FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT.  IN NO EVENT SHALL THE
+AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
+LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM,
+OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN
+THE SOFTWARE.
+*/
+
+#pragma once
+#include "node.h"
+#include "numpy_loader_sharded.h"
+
+class NumpyLoaderNode : public Node {
+   public:
+    NumpyLoaderNode(Tensor *output, void *device_resources);
+    ~NumpyLoaderNode() override;
+    NumpyLoaderNode() = delete;
+
+    /// \param internal_shard_count Defines the amount of parallelism user wants for the load process to be handled internally.
+    /// \param source_path Defines the path that includes the numpy array dataset
+    /// \param load_batch_count Defines the quantum count of the numpy arrays to be loaded. It's usually equal to the user's batch size.
+    /// The loader will repeat numpy arrays if necessary to be able to have numpy arrays in multiples of the load_batch_count,
+    /// for example if there are 10 numpy arrays in the dataset and load_batch_count is 3, the loader repeats 2 numpy arrays as if there are 12 numpy arrays available.
+    void init(unsigned internal_shard_count, const std::string &source_path, StorageType storage_type, DecoderType decoder_type, bool shuffle, bool loop,
+              size_t load_batch_count, RocalMemType mem_type);
+
+    std::shared_ptr<LoaderModule> get_loader_module();
+
+   protected:
+    void create_node() override{};
+    void update_node() override{};
+
+   private:
+    std::shared_ptr<NumpyLoaderSharded> _loader_module = nullptr;
+};

rocAL/include/loaders/image/node_numpy_loader_single_shard.h

@@ -0,0 +1,50 @@
+/*
+Copyright (c) 2019 - 2024 Advanced Micro Devices, Inc. All rights reserved.
+
+Permission is hereby granted, free of charge, to any person obtaining a copy
+of this software and associated documentation files (the "Software"), to deal
+in the Software without restriction, including without limitation the rights
+to use, copy, modify, merge, publish, distribute, sublicense, and/or sell
+copies of the Software, and to permit persons to whom the Software is
+furnished to do so, subject to the following conditions:
+
+The above copyright notice and this permission notice shall be included in
+all copies or substantial portions of the Software.
+
+THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
+IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
+FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT.  IN NO EVENT SHALL THE
+AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
+LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM,
+OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN
+THE SOFTWARE.
+*/
+
+#pragma once
+#include "node.h"
+#include "numpy_loader_sharded.h"
+
+class NumpyLoaderSingleShardNode : public Node {
+   public:
+    NumpyLoaderSingleShardNode(Tensor *output, void *device_resources);
+    ~NumpyLoaderSingleShardNode() override;
+
+    /// \param user_shard_count shard count from user
+    /// \param  user_shard_id shard id from user
+    /// \param source_path Defines the path that includes the numpy array dataset
+    /// \param load_batch_count Defines the quantum count of the numpy arrays to be loaded. It's usually equal to the user's batch size.
+    /// The loader will repeat numpy arrays if necessary to be able to have numpy arrays in multiples of the load_batch_count,
+    /// for example if there are 10 numpy arrays in the dataset and load_batch_count is 3, the loader repeats 2 numpy arrays as if there are 12 numpy arrays available.
+    void init(unsigned shard_id, unsigned shard_count, const std::string &source_path,
+              StorageType storage_type, DecoderType decoder_type, bool shuffle, bool loop,
+              size_t load_batch_count, RocalMemType mem_type);
+
+    std::shared_ptr<LoaderModule> get_loader_module();
+
+   protected:
+    void create_node() override{};
+    void update_node() override{};
+
+   private:
+    std::shared_ptr<NumpyLoader> _loader_module = nullptr;
+};

rocAL/include/loaders/image/numpy_loader.h

@@ -0,0 +1,90 @@
+/*
+Copyright (c) 2019 - 2024 Advanced Micro Devices, Inc. All rights reserved.
+
+Permission is hereby granted, free of charge, to any person obtaining a copy
+of this software and associated documentation files (the "Software"), to deal
+in the Software without restriction, including without limitation the rights
+to use, copy, modify, merge, publish, distribute, sublicense, and/or sell
+copies of the Software, and to permit persons to whom the Software is
+furnished to do so, subject to the following conditions:
+
+The above copyright notice and this permission notice shall be included in
+all copies or substantial portions of the Software.
+
+THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
+IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
+FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT.  IN NO EVENT SHALL THE
+AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
+LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM,
+OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN
+THE SOFTWARE.
+*/
+
+#pragma once
+
+#include <string>
+#include <thread>
+#include <vector>
+
+#include "circular_buffer.h"
+#include "commons.h"
+#include "image_read_and_decode.h"
+
+// NumpyLoader runs an internal thread for loading of numpy arrays asynchronously
+// it uses a circular buffer to store loaded numpy arrays for the user
+class NumpyLoader : public LoaderModule {
+   public:
+    explicit NumpyLoader(void* dev_resources);
+    ~NumpyLoader() override;
+    LoaderModuleStatus load_next() override;
+    void initialize(ReaderConfig reader_cfg, DecoderConfig decoder_cfg, RocalMemType mem_type, unsigned batch_size, bool keep_orig_size = false) override;
+    void set_output(Tensor* output_image) override;
+    void set_random_bbox_data_reader(std::shared_ptr<RandomBBoxCrop_MetaDataReader> randombboxcrop_meta_data_reader) override;
+    size_t remaining_count() override;  // returns number of remaining items to be loaded
+    void reset() override;              // Resets the loader to load from the beginning of the media
+    Timing timing() override;
+    void start_loading() override;
+    LoaderModuleStatus set_cpu_affinity(cpu_set_t cpu_mask);
+    LoaderModuleStatus set_cpu_sched_policy(struct sched_param sched_policy);
+    void set_gpu_device_id(int device_id);
+    std::vector<std::string> get_id() override;
+    decoded_image_info get_decode_image_info() override;
+    crop_image_info get_crop_image_info() override;
+    void set_prefetch_queue_depth(size_t prefetch_queue_depth) override;
+    void shut_down() override;
+    void feed_external_input(const std::vector<std::string>& input_images_names, const std::vector<unsigned char*>& input_buffer,
+                             const std::vector<ROIxywh>& roi_xywh, unsigned int max_width, unsigned int max_height, unsigned int channels, ExternalSourceFileMode mode, bool eos) override {}
+
+   private:
+    bool is_out_of_data();
+    void de_init();
+    void stop_internal_thread();
+    LoaderModuleStatus update_output_tensor();
+    LoaderModuleStatus load_routine();
+    std::shared_ptr<Reader> _reader;
+    std::shared_ptr<RandomBBoxCrop_MetaDataReader> _randombboxcrop_meta_data_reader = nullptr;
+    Tensor* _output_tensor;
+    std::vector<std::string> _output_names;  //!< image name/ids that are stores in the _output_image
+    size_t _output_mem_size;
+    MetaDataBatch* _meta_data = nullptr;  //!< The output of the meta_data_graph,
+    bool _internal_thread_running;
+    size_t _batch_size;
+    size_t _image_size;
+    std::thread _load_thread;
+    RocalMemType _mem_type;
+    decoded_image_info _decoded_img_info;
+    crop_image_info _crop_image_info;
+    decoded_image_info _output_decoded_img_info;
+    crop_image_info _output_cropped_img_info;
+    CircularBuffer _circ_buff;
+    TimingDBG _file_load_time, _swap_handle_time;
+    bool _is_initialized;
+    bool _stopped = false;
+    bool _loop;                     //<! If true the reader will wrap around at the end of the media (files/samples/...) and wouldn't stop
+    size_t _prefetch_queue_depth;   // Used for circular buffer's internal buffer
+    size_t _image_counter = 0;      //!< How many samples have been loaded already
+    size_t _remaining_count;  //!< How many samples are there yet to be loaded
+    bool _decoder_keep_original = false;
+    int _device_id;
+    std::vector<std::vector<unsigned>> _tensor_roi;
+};

rocAL/include/loaders/image/numpy_loader_sharded.h

@@ -0,0 +1,61 @@
+/*
+Copyright (c) 2019 - 2024 Advanced Micro Devices, Inc. All rights reserved.
+
+Permission is hereby granted, free of charge, to any person obtaining a copy
+of this software and associated documentation files (the "Software"), to deal
+in the Software without restriction, including without limitation the rights
+to use, copy, modify, merge, publish, distribute, sublicense, and/or sell
+copies of the Software, and to permit persons to whom the Software is
+furnished to do so, subject to the following conditions:
+
+The above copyright notice and this permission notice shall be included in
+all copies or substantial portions of the Software.
+
+THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
+IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
+FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT.  IN NO EVENT SHALL THE
+AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
+LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM,
+OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN
+THE SOFTWARE.
+*/
+
+#pragma once
+
+#include "numpy_loader.h"
+//
+// NumpyLoaderSharded Can be used to run load and decode in multiple shards, each shard by a single loader instance,
+// It improves load and decode performance since each loader loads the numpy arrays in parallel using an internal thread
+//
+class NumpyLoaderSharded : public LoaderModule {
+   public:
+    explicit NumpyLoaderSharded(void *dev_resources);
+    ~NumpyLoaderSharded() override;
+    LoaderModuleStatus load_next() override;
+    void initialize(ReaderConfig reader_cfg, DecoderConfig decoder_cfg, RocalMemType mem_type, unsigned batch_size, bool keep_orig_size = false) override;
+    void set_output(Tensor *output_image) override;
+    void set_random_bbox_data_reader(std::shared_ptr<RandomBBoxCrop_MetaDataReader> randombboxcrop_meta_data_reader) override;
+    size_t remaining_count() override;
+    void reset() override;
+    void start_loading() override;
+    std::vector<std::string> get_id() override;
+    decoded_image_info get_decode_image_info() override;
+    crop_image_info get_crop_image_info() override;
+    Timing timing() override;
+    void set_prefetch_queue_depth(size_t prefetch_queue_depth) override;
+    void shut_down() override;
+    void feed_external_input(const std::vector<std::string>& input_images_names, const std::vector<unsigned char*>& input_buffer,
+                             const std::vector<ROIxywh>& roi_xywh, unsigned int max_width, unsigned int max_height, unsigned int channels, ExternalSourceFileMode mode, bool eos) override {}
+
+   private:
+    void increment_loader_idx();
+    void *_dev_resources;
+    bool _initialized = false;
+    std::vector<std::shared_ptr<NumpyLoader>> _loaders;
+    size_t _loader_idx;
+    size_t _shard_count = 1;
+    void fast_forward_through_empty_loaders();
+    size_t _prefetch_queue_depth;
+    Tensor *_output_tensor;
+    std::shared_ptr<RandomBBoxCrop_MetaDataReader> _randombboxcrop_meta_data_reader = nullptr;
+};

rocAL/include/loaders/numpy_source_evaluator.h

@@ -0,0 +1,46 @@
+/*
+Copyright (c) 2019 - 2024 Advanced Micro Devices, Inc. All rights reserved.
+
+Permission is hereby granted, free of charge, to any person obtaining a copy
+of this software and associated documentation files (the "Software"), to deal
+in the Software without restriction, including without limitation the rights
+to use, copy, modify, merge, publish, distribute, sublicense, and/or sell
+copies of the Software, and to permit persons to whom the Software is
+furnished to do so, subject to the following conditions:
+
+The above copyright notice and this permission notice shall be included in
+all copies or substantial portions of the Software.
+
+THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
+IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
+FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT.  IN NO EVENT SHALL THE
+AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
+LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM,
+OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN
+THE SOFTWARE.
+*/
+
+#pragma once
+#include <map>
+#include <memory>
+
+#include "loader_module.h"
+#include "reader_factory.h"
+#include "timing_debug.h"
+
+enum class NumpySourceEvaluatorStatus {
+    OK = 0,
+    UNSUPPORTED_STORAGE_TYPE,
+};
+
+class NumpySourceEvaluator {
+   public:
+    NumpySourceEvaluatorStatus create(ReaderConfig reader_cfg);
+    void find_max_numpy_dimensions();
+    std::vector<size_t> max_numpy_dims() { return _max_numpy_dims; }
+    RocalTensorDataType get_numpy_dtype() { return _numpy_dtype; }
+
+    std::vector<size_t> _max_numpy_dims;
+    RocalTensorDataType _numpy_dtype;
+    std::shared_ptr<Reader> _reader;
+};

rocAL/include/pipeline/commons.h

@@ -46,6 +46,8 @@ enum class RocalTensorlayout {
     NCHW,
     NFHWC,
     NFCHW,
+    NDHWC,
+    NCDHW,
     NONE
 };
 
@@ -105,6 +107,17 @@ enum class RocalROIType {
     XYWH
 };
 
+/*! \brief Tensor 3D ROI type
+ *
+ * currently supports following formats
+ * LTFRBB - Left-Top-Front-Right-Bottom-Back
+ * XYZWHD - X-Y-Z-Width-Height-Depth
+ */
+enum class Rocal3DROIType {
+    LTFRBB = 0,
+    XYZWHD
+};
+
 /*! \brief Tensor ROI in LTRB format
  *
  */