diff options
Diffstat (limited to 'runtimes/contrib/labs')
| -rw-r--r-- | runtimes/contrib/labs/CMakeLists.txt | 5 | ||||
| -rw-r--r-- | runtimes/contrib/labs/jniacl/CMakeLists.txt | 18 | ||||
| -rw-r--r-- | runtimes/contrib/labs/jniacl/src/io_accessor.cc | 96 | ||||
| -rw-r--r-- | runtimes/contrib/labs/jniacl/src/io_accessor.h | 93 | ||||
| -rw-r--r-- | runtimes/contrib/labs/jniacl/src/jniacl_main.cc | 37 | ||||
| -rw-r--r-- | runtimes/contrib/labs/opencl_test/CMakeLists.txt | 11 | ||||
| -rw-r--r-- | runtimes/contrib/labs/opencl_test/README.md | 8 | ||||
| -rw-r--r-- | runtimes/contrib/labs/opencl_test/src/opencl_test.cc | 386 | ||||
| -rw-r--r-- | runtimes/contrib/labs/tflite_examples/CMakeLists.txt | 2 | ||||
| -rw-r--r-- | runtimes/contrib/labs/tflite_examples/src/conv.cpp | 330 |
10 files changed, 986 insertions, 0 deletions
diff --git a/runtimes/contrib/labs/CMakeLists.txt b/runtimes/contrib/labs/CMakeLists.txt new file mode 100644 index 000000000..57e28c11a --- /dev/null +++ b/runtimes/contrib/labs/CMakeLists.txt @@ -0,0 +1,5 @@ +if(NOT BUILD_LABS) + return() +endif(NOT BUILD_LABS) + +add_subdirectories() diff --git a/runtimes/contrib/labs/jniacl/CMakeLists.txt b/runtimes/contrib/labs/jniacl/CMakeLists.txt new file mode 100644 index 000000000..f66127b84 --- /dev/null +++ b/runtimes/contrib/labs/jniacl/CMakeLists.txt @@ -0,0 +1,18 @@ +# +# Simple Android JNI execution test of ACL +# + +if(NOT "${TARGET_OS}" STREQUAL "android") + return() +endif(NOT "${TARGET_OS}" STREQUAL "android") + +nnfw_find_package(ARMCompute REQUIRED) + +link_directories(${CMAKE_INSTALL_PREFIX}/lib) + +set(JNIACL_SRCS src/jniacl_main.cc + src/io_accessor.cc) + +add_library(jniacl_jni SHARED ${JNIACL_SRCS}) +target_include_directories(jniacl_jni PUBLIC ${TFLITE_JNI_INCLUDES} src) +target_link_libraries(jniacl_jni arm_compute_graph log) diff --git a/runtimes/contrib/labs/jniacl/src/io_accessor.cc b/runtimes/contrib/labs/jniacl/src/io_accessor.cc new file mode 100644 index 000000000..076c93f3d --- /dev/null +++ b/runtimes/contrib/labs/jniacl/src/io_accessor.cc @@ -0,0 +1,96 @@ +/* + * Copyright (c) 2018 Samsung Electronics Co., Ltd. All Rights Reserved + * + * Licensed under the Apache License, Version 2.0 (the "License"); + * you may not use this file except in compliance with the License. + * You may obtain a copy of the License at + * + * http://www.apache.org/licenses/LICENSE-2.0 + * + * Unless required by applicable law or agreed to in writing, software + * distributed under the License is distributed on an "AS IS" BASIS, + * WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. + * See the License for the specific language governing permissions and + * limitations under the License. + */ + +/* + * Copyright (c) 2018 ARM Limited. + * + * SPDX-License-Identifier: MIT + * + * 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. + */ +#include "io_accessor.h" +#include <ostream> +#include <android/log.h> + +bool InputAccessor::access_tensor(arm_compute::ITensor &tensor) +{ + // Subtract the mean value from each channel + arm_compute::Window window; + window.use_tensor_dimensions(tensor.info()->tensor_shape()); + + execute_window_loop(window, [&](const arm_compute::Coordinates &id) { + *reinterpret_cast<float *>(tensor.ptr_to_element(id)) = _test_input; + _test_input += _inc ? 1.0 : 0.0; + + __android_log_print(ANDROID_LOG_DEBUG, "LOG_TAG", "Input %d, %d = %lf\r\n", id.y(), id.x(), + *reinterpret_cast<float *>(tensor.ptr_to_element(id))); + }); + return true; +} + +bool OutputAccessor::access_tensor(arm_compute::ITensor &tensor) +{ + // Subtract the mean value from each channel + arm_compute::Window window; + window.use_tensor_dimensions(tensor.info()->tensor_shape()); + + execute_window_loop(window, [&](const arm_compute::Coordinates &id) { + __android_log_print(ANDROID_LOG_DEBUG, "Output", "Input %d, %d = %lf\r\n", id.y(), id.x(), + *reinterpret_cast<float *>(tensor.ptr_to_element(id))); + }); + return false; // end the network +} + +bool WeightAccessor::access_tensor(arm_compute::ITensor &tensor) +{ + // Subtract the mean value from each channel + arm_compute::Window window; + window.use_tensor_dimensions(tensor.info()->tensor_shape()); + + execute_window_loop(window, [&](const arm_compute::Coordinates &id) { + *reinterpret_cast<float *>(tensor.ptr_to_element(id)) = _test_weight; + _test_weight += _inc ? 1.0 : 0.0; + }); + return true; +} + +bool BiasAccessor::access_tensor(arm_compute::ITensor &tensor) +{ + // Subtract the mean value from each channel + arm_compute::Window window; + window.use_tensor_dimensions(tensor.info()->tensor_shape()); + + execute_window_loop(window, [&](const arm_compute::Coordinates &id) { + *reinterpret_cast<float *>(tensor.ptr_to_element(id)) = 0.0; + }); + return true; +} diff --git a/runtimes/contrib/labs/jniacl/src/io_accessor.h b/runtimes/contrib/labs/jniacl/src/io_accessor.h new file mode 100644 index 000000000..bc4376644 --- /dev/null +++ b/runtimes/contrib/labs/jniacl/src/io_accessor.h @@ -0,0 +1,93 @@ +/* + * Copyright (c) 2018 Samsung Electronics Co., Ltd. All Rights Reserved + * + * Licensed under the Apache License, Version 2.0 (the "License"); + * you may not use this file except in compliance with the License. + * You may obtain a copy of the License at + * + * http://www.apache.org/licenses/LICENSE-2.0 + * + * Unless required by applicable law or agreed to in writing, software + * distributed under the License is distributed on an "AS IS" BASIS, + * WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. + * See the License for the specific language governing permissions and + * limitations under the License. + */ + +/* + * Copyright (c) 2018 ARM Limited. + * + * SPDX-License-Identifier: MIT + * + * 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. + */ +#ifndef __IO_ACCESSOR_H__ +#define __IO_ACCESSOR_H__ + +#include <arm_compute/graph/ITensorAccessor.h> + +class InputAccessor : public arm_compute::graph::ITensorAccessor +{ +public: + InputAccessor(bool inc) : _inc(inc) { _test_input = 1.0; } + InputAccessor(InputAccessor &&) = default; + + // Inherited methods overriden: + bool access_tensor(arm_compute::ITensor &tensor) override; + +private: + bool _inc; + float _test_input; +}; + +class OutputAccessor : public arm_compute::graph::ITensorAccessor +{ +public: + OutputAccessor() = default; + OutputAccessor(OutputAccessor &&) = default; + + // Inherited methods overriden: + bool access_tensor(arm_compute::ITensor &tensor) override; +}; + +class WeightAccessor : public arm_compute::graph::ITensorAccessor +{ +public: + WeightAccessor(bool inc) : _inc(inc) { _test_weight = 1.0; } + WeightAccessor(WeightAccessor &&) = default; + + // Inherited methods overriden: + bool access_tensor(arm_compute::ITensor &tensor) override; + +private: + bool _inc; + float _test_weight; +}; + +class BiasAccessor : public arm_compute::graph::ITensorAccessor +{ +public: + BiasAccessor() = default; + BiasAccessor(BiasAccessor &&) = default; + + // Inherited methods overriden: + bool access_tensor(arm_compute::ITensor &tensor) override; +}; + +#endif // __IO_ACCESSOR_H__ diff --git a/runtimes/contrib/labs/jniacl/src/jniacl_main.cc b/runtimes/contrib/labs/jniacl/src/jniacl_main.cc new file mode 100644 index 000000000..4e5f10d1f --- /dev/null +++ b/runtimes/contrib/labs/jniacl/src/jniacl_main.cc @@ -0,0 +1,37 @@ +#include <jni.h> +#include <string> + +#include <arm_compute/graph/Graph.h> +#include <arm_compute/graph/Nodes.h> + +#include "io_accessor.h" + +extern "C" JNIEXPORT jstring JNICALL +Java_com_samsung_testaclexec_ActivityMain_RunACLJNI(JNIEnv *env, jobject) +{ + using arm_compute::DataType; + using arm_compute::graph::Tensor; + using arm_compute::graph::TargetHint; + using arm_compute::graph::Graph; + using arm_compute::TensorInfo; + using arm_compute::TensorShape; + + arm_compute::graph::Graph graph; + TargetHint target_hint = TargetHint::OPENCL; + bool autoinc = true; + + graph << target_hint << Tensor(TensorInfo(TensorShape(3U, 3U, 1U, 1U), 1, DataType::F32), + std::unique_ptr<InputAccessor>(new InputAccessor(autoinc))) + << arm_compute::graph::ConvolutionLayer( + 3U, 3U, 1U, std::unique_ptr<WeightAccessor>(new WeightAccessor(autoinc)), + std::unique_ptr<BiasAccessor>(new BiasAccessor()), + arm_compute::PadStrideInfo(1, 1, 0, 0)) + << Tensor(std::unique_ptr<OutputAccessor>(new OutputAccessor())); + ; + + graph.run(); + + std::string hello = "SoftMax Run OK"; + + return env->NewStringUTF(hello.c_str()); +} diff --git a/runtimes/contrib/labs/opencl_test/CMakeLists.txt b/runtimes/contrib/labs/opencl_test/CMakeLists.txt new file mode 100644 index 000000000..dc8f5f661 --- /dev/null +++ b/runtimes/contrib/labs/opencl_test/CMakeLists.txt @@ -0,0 +1,11 @@ +if(NOT ${TARGET_ARCH_BASE} STREQUAL "arm") + return() +endif(NOT ${TARGET_ARCH_BASE} STREQUAL "arm") + +list(APPEND OPENCL_INFO_SOURCE "src/opencl_test.cc") + +nnfw_find_package(ARMCompute REQUIRED) + +add_executable(opencl_test ${OPENCL_INFO_SOURCE}) +target_link_libraries(opencl_test arm_compute) +target_link_libraries(opencl_test arm_compute_ex) diff --git a/runtimes/contrib/labs/opencl_test/README.md b/runtimes/contrib/labs/opencl_test/README.md new file mode 100644 index 000000000..950528f81 --- /dev/null +++ b/runtimes/contrib/labs/opencl_test/README.md @@ -0,0 +1,8 @@ +This directory contains experients of OpenCL code. + +How to run: +``` +LD_LIBRARY_PATH=Product/out/lib Product/obj/contrib/opencl_test/opencl_test [option] +``` + - `[option]` + - `-g`: prints devices inside GPU and check if they use same memory address diff --git a/runtimes/contrib/labs/opencl_test/src/opencl_test.cc b/runtimes/contrib/labs/opencl_test/src/opencl_test.cc new file mode 100644 index 000000000..1faa91478 --- /dev/null +++ b/runtimes/contrib/labs/opencl_test/src/opencl_test.cc @@ -0,0 +1,386 @@ +/* + * Copyright (c) 2018 Samsung Electronics Co., Ltd. All Rights Reserved + * + * Licensed under the Apache License, Version 2.0 (the "License"); + * you may not use this file except in compliance with the License. + * You may obtain a copy of the License at + * + * http://www.apache.org/licenses/LICENSE-2.0 + * + * Unless required by applicable law or agreed to in writing, software + * distributed under the License is distributed on an "AS IS" BASIS, + * WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. + * See the License for the specific language governing permissions and + * limitations under the License. + */ + +/******************************************************************************* + * Copyright (c) 2008-2015 The Khronos Group Inc. + * + * Permission is hereby granted, free of charge, to any person obtaining a + * copy of this software and/or associated documentation files (the + * "Materials"), to deal in the Materials without restriction, including + * without limitation the rights to use, copy, modify, merge, publish, + * distribute, sublicense, and/or sell copies of the Materials, and to + * permit persons to whom the Materials are 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 Materials. + * + * THE MATERIALS ARE 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 + * MATERIALS OR THE USE OR OTHER DEALINGS IN THE MATERIALS. + ******************************************************************************/ + +#include "arm_compute/core/CL/OpenCL.h" + +#include <iostream> +#include <vector> + +void printDeviceInfo(int n, cl::Device &device, cl::Device &default_device) +{ + bool is_default = (device() == default_device()); + std::cout << "\t\t\t#" << n << " Device: (id: " << device() << ") " + << (is_default ? " -> default" : "") << "\n"; + + const auto name = device.getInfo<CL_DEVICE_NAME>(); + std::cout << "\t\t\t\tName: " << name << "\n"; + + const auto compute_unit = device.getInfo<CL_DEVICE_MAX_COMPUTE_UNITS>(); + std::cout << "\t\t\t\tMax Compute Unit: " << compute_unit << "\n"; + + const auto max_work_item_size = device.getInfo<CL_DEVICE_MAX_WORK_ITEM_SIZES>(); + std::cout << "\t\t\t\tMax Work Item Size: ["; + for (auto size : max_work_item_size) + std::cout << size << ","; + std::cout << "]\n"; + + const auto max_work_group_size = device.getInfo<CL_DEVICE_MAX_WORK_GROUP_SIZE>(); + std::cout << "\t\t\t\tMax Work Grpup Size: " << max_work_group_size << "\n"; + + const auto max_clock_frequency = device.getInfo<CL_DEVICE_MAX_CLOCK_FREQUENCY>(); + std::cout << "\t\t\t\tMax Clock Frequency: " << max_clock_frequency << "\n"; + + std::cout << "\n"; +} + +class OpenCLGpu +{ +public: + cl::Platform platform_; + cl::Context context_; + cl::vector<cl::Device> devices_; + std::vector<cl::CommandQueue *> q_; + cl::Program program_; + + OpenCLGpu() + { + cl_int cl_error; + + platform_ = cl::Platform::getDefault(); + + try + { + cl_context_properties properties[3] = {CL_CONTEXT_PLATFORM, + (cl_context_properties)platform_(), 0}; + + context_ = cl::Context(CL_DEVICE_TYPE_GPU, properties, NULL, NULL, &cl_error); + } + catch (cl::Error &err) // thrown when there is no Context for this platform + { + std::cout << "\t\t No Context Found\n"; + return; + } + + devices_ = context_.getInfo<CL_CONTEXT_DEVICES>(); + + for (int dev_id = 0; dev_id < devices_.size(); dev_id++) + { + cl::CommandQueue *que = new cl::CommandQueue(context_, devices_[dev_id]); + q_.emplace_back(que); + } + } + + ~OpenCLGpu() + { + for (auto each_q : q_) + delete each_q; + } + + void buildProgram(std::string &kernel_source_code) + { + std::vector<std::string> programStrings{kernel_source_code}; + + program_ = cl::Program(context_, programStrings); + + try + { + program_.build("-cl-std=CL1.2"); + } + catch (cl::Error &err) + { + cl_int buildErr = CL_SUCCESS; + auto buildInfo = program_.getBuildInfo<CL_PROGRAM_BUILD_LOG>(&buildErr); + for (auto &pair : buildInfo) + { + std::cerr << pair.second << std::endl << std::endl; + } + } + } +}; + +void checkContextMem() +{ + cl_int cl_error; + + // get context, devices + // + std::cout << "\nChecking if devices in GPU shares the same memory address:\n\n"; + + OpenCLGpu gpu; + + std::cout << "\nDevices in GPU:\n\n"; + + auto &devices = gpu.devices_; + auto default_device = cl::Device::getDefault(); + + int d = 0; + for (auto device : devices) + printDeviceInfo(++d, device, default_device); + + if (d < 2) + { + std::cout << "\t\t This options works when there are n (>= 2) devices.\n"; + return; + } + + // allocate and map memory + + typedef cl_int T; + const int items_per_device = 128; + const int length = items_per_device * devices.size(); + + std::vector<T> input(length); + std::vector<T> output(length, 0); + + for (int i = 0; i < length; i++) + input[i] = i; + + cl::Buffer input_buf(gpu.context_, (cl_mem_flags)CL_MEM_USE_HOST_PTR, length * sizeof(T), + input.data(), &cl_error); + cl::Buffer output_buf(gpu.context_, (cl_mem_flags)CL_MEM_USE_HOST_PTR, length * sizeof(T), + output.data(), &cl_error); + + // compile test cl code + + std::string kernel_source{"typedef int T; \n" + "kernel void memory_test( \n" + " const int dev_id, \n" + " global T* input, \n" + " global T* output, \n" + " const int start_idx, \n" + " const int count) \n" + "{ \n" + " int input_idx = get_global_id(0); \n" + " if(input_idx < count) \n" + " { \n" + " int output_idx = start_idx + input_idx; \n" + " output[output_idx] = input[input_idx] + dev_id; \n" + " } \n" + "} \n"}; + + gpu.buildProgram(kernel_source); + + try + { + auto kernel_functor = cl::KernelFunctor<cl_int, cl::Buffer, cl::Buffer, cl_int, cl_int>( + gpu.program_, "memory_test"); // name should be same as cl function name + + // create a queue per device and queue a kernel job + + for (int dev_id = 0; dev_id < devices.size(); dev_id++) + { + kernel_functor(cl::EnqueueArgs(*(gpu.q_[dev_id]), cl::NDRange(items_per_device)), + (cl_int)dev_id, // dev id + input_buf, output_buf, + (cl_int)(items_per_device * dev_id), // start index + (cl_int)(items_per_device), // count + cl_error); + } + + // sync + + for (d = 0; d < devices.size(); d++) + (gpu.q_[d])->finish(); + + // check if memory state changed by all devices + + cl::copy(*(gpu.q_[0]), output_buf, begin(output), end(output)); + + bool use_same_memory = true; + + for (int dev_id = 0; dev_id < devices.size(); dev_id++) + { + for (int i = 0; i < items_per_device; ++i) + { + int output_idx = items_per_device * dev_id + i; + if (output[output_idx] != input[i] + dev_id) + { + std::cout << "Output[" << output_idx << "] : " + << "expected = " << input[i] + dev_id << "; actual = " << output[output_idx] + << "\n"; + use_same_memory = false; + break; + } + } + } + + if (use_same_memory) + std::cout << "\n=> Mapped memory addresses used by devices in GPU are same.\n\n"; + else + std::cout << "\n=> Mapped memory addresses used by devices in GPU are different.\n\n"; + } + catch (cl::Error &err) + { + std::cerr << "error: code: " << err.err() << ", what: " << err.what() << std::endl; + } +} + +void printHelp() +{ + std::cout << "opencl information: \n\n"; + std::cout << "\t -h : help\n"; + std::cout + << "\t -g : print if memory map is shared among devices in GPU (in default platform)\n\n"; + std::cout << "\t -s : test for synchronized work by two devices in a GPU\n\n"; +} + +#include <mutex> +#include <chrono> +#include <thread> +#include <condition_variable> + +#define MAX_DEVICE_NUM 8 // just for testing + +int kernel_idx[MAX_DEVICE_NUM]; +unsigned char kernel_completed = 0x00; // bit 0 = 1 means kernel by device[0] was completed. +unsigned char + kernel_completed_flag; // if comparing kernel_completed with this var, all kernels are completed +int device_num; +std::mutex kernel_complete_handler_mutex; + +std::condition_variable wakeup_main; +std::mutex wakeup_main_mutex; + +void notifyKernelFinished(cl_event ev, cl_int ev_info, void *device_idx) +{ + std::cout << "callback from device[" << *((int *)device_idx) << "] : ==> completed.\n"; + + std::unique_lock<std::mutex> lock(kernel_complete_handler_mutex); + + kernel_completed |= 0x01 << *((int *)device_idx); + if (kernel_completed == kernel_completed_flag) + wakeup_main.notify_one(); +} + +void testSync() +{ + OpenCLGpu gpu; + + cl_int cl_error; + typedef cl_int T; + const int items_per_device = 1024 * 768; + const int length = items_per_device * gpu.devices_.size(); + + std::vector<T> output(length, 0); + + cl::Buffer output_buf(gpu.context_, (cl_mem_flags)CL_MEM_USE_HOST_PTR, length * sizeof(T), + output.data(), &cl_error); + + std::string kernel_source{"kernel void test(global float* output, const int count) \n" + "{ \n" + " int idx = get_global_id(0); \n" + " if(idx < count) \n" + " { \n" + " float x = hypot(idx/1.111, idx*1.111); \n" + " for (int y = 0; y < 200; y++) \n" + " x = rootn(log(pown(rootn(log(pown(x, 20)), 5), 20)), 5); \n" + " output[idx] = x; \n" + " } \n" + "} \n"}; + + gpu.buildProgram(kernel_source); + + try + { + auto kernel_functor = cl::KernelFunctor<cl::Buffer, cl_int>( + gpu.program_, "test"); // name should be same as cl function name + + // variable init + cl::Event ev[MAX_DEVICE_NUM]; + + device_num = gpu.devices_.size(); + + kernel_completed = 0; + kernel_completed_flag = 0; + for (int i = 0; i < device_num; i++) + { + kernel_idx[i] = i; + kernel_completed_flag |= 0x01 << i; + } + + // create a queue per device and queue a kernel job + // queueing with callback function + for (int dev_id = 0; dev_id < gpu.devices_.size(); dev_id++) + { + ev[dev_id] = kernel_functor(cl::EnqueueArgs(*(gpu.q_[dev_id]), cl::NDRange(items_per_device)), + output_buf, + (cl_int)(items_per_device), // count + cl_error); + ev[dev_id].setCallback(CL_COMPLETE, notifyKernelFinished, (void *)(kernel_idx + dev_id)); + + // how to check kernel execution status + // + // auto status = ev[dev_id].getInfo<CL_EVENT_COMMAND_EXECUTION_STATUS>(); + // std::cout << "Event status = " << (status == CL_QUEUED ? "CL_QUEUED" : status == + // CL_SUBMITTED ? "CL_SUBMITTED" : status == CL_COMPLETE ? "CL_COMPLETE" : "unknown") + // << std::endl; + // std::cout << "Event status code = " << status << std::endl; + } + + // long wait until kernels are over + { + std::unique_lock<std::mutex> lk(wakeup_main_mutex); + wakeup_main.wait(lk, [] { return (kernel_completed == kernel_completed_flag); }); + + std::cout << "all devices were completed.\n"; + } + } + catch (cl::Error &err) + { + std::cerr << "error: code: " << err.err() << ", what: " << err.what() << std::endl; + } +} + +int main(const int argc, char **argv) +{ + if (argc < 2) + printHelp(); + else + { + std::string option = argv[1]; + + if (option == "-h") // help + printHelp(); + else if (option == "-g") // check if devices in GPU uses same memory address + checkContextMem(); + else if (option == "-s") // check synchronization between devices in GPU + testSync(); + } + return 0; +} diff --git a/runtimes/contrib/labs/tflite_examples/CMakeLists.txt b/runtimes/contrib/labs/tflite_examples/CMakeLists.txt new file mode 100644 index 000000000..463bc5531 --- /dev/null +++ b/runtimes/contrib/labs/tflite_examples/CMakeLists.txt @@ -0,0 +1,2 @@ +add_executable(tflite_conv_example "src/conv.cpp") +target_link_libraries(tflite_conv_example tensorflow-lite ${LIB_PTHREAD} dl nnfw_lib_tflite) diff --git a/runtimes/contrib/labs/tflite_examples/src/conv.cpp b/runtimes/contrib/labs/tflite_examples/src/conv.cpp new file mode 100644 index 000000000..3117c316c --- /dev/null +++ b/runtimes/contrib/labs/tflite_examples/src/conv.cpp @@ -0,0 +1,330 @@ +/* + * Copyright (c) 2018 Samsung Electronics Co., Ltd. All Rights Reserved + * + * Licensed under the Apache License, Version 2.0 (the "License"); + * you may not use this file except in compliance with the License. + * You may obtain a copy of the License at + * + * http://www.apache.org/licenses/LICENSE-2.0 + * + * Unless required by applicable law or agreed to in writing, software + * distributed under the License is distributed on an "AS IS" BASIS, + * WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. + * See the License for the specific language governing permissions and + * limitations under the License. + */ + +#include "tflite/ext/kernels/register.h" +#include "tensorflow/lite/model.h" +#include "tensorflow/lite/builtin_op_data.h" + +#include <iostream> + +using namespace tflite; +using namespace nnfw::tflite; + +namespace vector +{ + +template <typename T> struct View +{ + virtual ~View() = default; + + virtual int32_t size(void) const = 0; + virtual T at(uint32_t off) const = 0; +}; +} + +namespace feature +{ + +struct Shape +{ + int32_t C; + int32_t H; + int32_t W; +}; + +template <typename T> struct View +{ + virtual ~View() = default; + + virtual const Shape &shape(void) const = 0; + virtual T at(uint32_t ch, uint32_t row, uint32_t col) const = 0; +}; +} + +namespace kernel +{ + +struct Shape +{ + int32_t N; + int32_t C; + int32_t H; + int32_t W; +}; + +template <typename T> struct View +{ + virtual ~View() = default; + + virtual const Shape &shape(void) const = 0; + virtual T at(uint32_t nth, uint32_t ch, uint32_t row, uint32_t col) const = 0; +}; +} + +const int32_t N = 1; +const int32_t C = 2; + +class SampleBiasObject final : public vector::View<float> +{ +public: + SampleBiasObject() : _size(N) + { + // DO NOTHING + } + +public: + int32_t size(void) const override { return _size; } + + float at(uint32_t off) const override { return 0.0f; } + +private: + int32_t _size; +}; + +class SampleFeatureObject final : public feature::View<float> +{ +public: + SampleFeatureObject() + { + _shape.C = C; + _shape.H = 3; + _shape.W = 4; + + const uint32_t size = _shape.C * _shape.H * _shape.W; + + for (uint32_t off = 0; off < size; ++off) + { + _value.emplace_back(off); + } + + assert(_value.size() == size); + } + +public: + const feature::Shape &shape(void) const override { return _shape; }; + + float at(uint32_t ch, uint32_t row, uint32_t col) const override + { + return _value.at(ch * _shape.H * _shape.W + row * _shape.W + col); + } + +public: + float &at(uint32_t ch, uint32_t row, uint32_t col) + { + return _value.at(ch * _shape.H * _shape.W + row * _shape.W + col); + } + +private: + feature::Shape _shape; + std::vector<float> _value; +}; + +class SampleKernelObject final : public kernel::View<float> +{ +public: + SampleKernelObject() + { + _shape.N = N; + _shape.C = C; + _shape.H = 3; + _shape.W = 4; + + const uint32_t size = _shape.N * _shape.C * _shape.H * _shape.W; + + for (uint32_t off = 0; off < size; ++off) + { + _value.emplace_back(off); + } + + assert(_value.size() == size); + } + +public: + const kernel::Shape &shape(void) const override { return _shape; }; + + float at(uint32_t nth, uint32_t ch, uint32_t row, uint32_t col) const override + { + return _value.at(nth * _shape.C * _shape.H * _shape.W + ch * _shape.H * _shape.W + + row * _shape.W + col); + } + +private: + kernel::Shape _shape; + std::vector<float> _value; +}; + +int main(int argc, char **argv) +{ + const SampleFeatureObject ifm; + const SampleKernelObject kernel; + const SampleBiasObject bias; + + const int32_t IFM_C = ifm.shape().C; + const int32_t IFM_H = ifm.shape().H; + const int32_t IFM_W = ifm.shape().W; + + const int32_t KER_N = kernel.shape().N; + const int32_t KER_C = kernel.shape().C; + const int32_t KER_H = kernel.shape().H; + const int32_t KER_W = kernel.shape().W; + + const int32_t OFM_C = kernel.shape().N; + const int32_t OFM_H = (IFM_H - KER_H) + 1; + const int32_t OFM_W = (IFM_W - KER_W) + 1; + + // Assumption on this example + assert(IFM_C == KER_C); + assert(KER_N == bias.size()); + + // Comment from 'context.h' + // + // Parameters for asymmetric quantization. Quantized values can be converted + // back to float using: + // real_value = scale * (quantized_value - zero_point); + // + // Q: Is this necessary? + TfLiteQuantizationParams quantization; + + quantization.scale = 1; + quantization.zero_point = 0; + + Interpreter interp; + + // On AddTensors(N) call, T/F Lite interpreter creates N tensors whose index is [0 ~ N) + interp.AddTensors(5); + + // Configure OFM + interp.SetTensorParametersReadWrite(0, kTfLiteFloat32 /* type */, "output" /* name */, + {1 /*N*/, OFM_H, OFM_W, OFM_C} /* dims */, quantization); + + // Configure IFM + interp.SetTensorParametersReadWrite(1, kTfLiteFloat32 /* type */, "input" /* name */, + {1 /*N*/, IFM_H, IFM_W, IFM_C} /* dims */, quantization); + + // Configure Filter + const uint32_t kernel_size = KER_N * KER_C * KER_H * KER_W; + float kernel_data[kernel_size] = { + 0.0f, + }; + + // Fill kernel data in NHWC order + { + uint32_t off = 0; + + for (uint32_t nth = 0; nth < KER_N; ++nth) + { + for (uint32_t row = 0; row < KER_H; ++row) + { + for (uint32_t col = 0; col < KER_W; ++col) + { + for (uint32_t ch = 0; ch < KER_C; ++ch) + { + const auto value = kernel.at(nth, ch, row, col); + kernel_data[off++] = value; + } + } + } + } + + assert(kernel_size == off); + } + + interp.SetTensorParametersReadOnly( + 2, kTfLiteFloat32 /* type */, "filter" /* name */, {KER_N, KER_H, KER_W, KER_C} /* dims */, + quantization, reinterpret_cast<const char *>(kernel_data), sizeof(kernel_data)); + + // Configure Bias + const uint32_t bias_size = bias.size(); + float bias_data[bias_size] = { + 0.0f, + }; + + // Fill bias data + for (uint32_t off = 0; off < bias.size(); ++off) + { + bias_data[off] = bias.at(off); + } + + interp.SetTensorParametersReadOnly(3, kTfLiteFloat32 /* type */, "bias" /* name */, + {bias.size()} /* dims */, quantization, + reinterpret_cast<const char *>(bias_data), sizeof(bias_data)); + + // Add Convolution Node + // + // NOTE AddNodeWithParameters take the ownership of param, and deallocate it with free + // So, param should be allocated with malloc + TfLiteConvParams *param = reinterpret_cast<TfLiteConvParams *>(malloc(sizeof(TfLiteConvParams))); + + param->padding = kTfLitePaddingValid; + param->stride_width = 1; + param->stride_height = 1; + param->activation = kTfLiteActRelu; + + // Run Convolution and store its result into Tensor #0 + // - Read IFM from Tensor #1 + // - Read Filter from Tensor #2, + // - Read Bias from Tensor #3 + interp.AddNodeWithParameters({1, 2, 3}, {0}, nullptr, 0, reinterpret_cast<void *>(param), + BuiltinOpResolver().FindOp(BuiltinOperator_CONV_2D, 1)); + + // Set Tensor #1 as Input #0, and Tensor #0 as Output #0 + interp.SetInputs({1}); + interp.SetOutputs({0}); + + // Let's use NNAPI (if possible) + interp.UseNNAPI(true); + + // Allocate Tensor + interp.AllocateTensors(); + + // Fill IFM data in HWC order + { + uint32_t off = 0; + + for (uint32_t row = 0; row < ifm.shape().H; ++row) + { + for (uint32_t col = 0; col < ifm.shape().W; ++col) + { + for (uint32_t ch = 0; ch < ifm.shape().C; ++ch) + { + const auto value = ifm.at(ch, row, col); + interp.typed_input_tensor<float>(0)[off++] = value; + } + } + } + } + + // Let's Rock-n-Roll! + interp.Invoke(); + + // Print OFM + { + uint32_t off = 0; + + for (uint32_t row = 0; row < OFM_H; ++row) + { + for (uint32_t col = 0; col < OFM_W; ++col) + { + for (uint32_t ch = 0; ch < kernel.shape().N; ++ch) + { + std::cout << interp.typed_output_tensor<float>(0)[off++] << std::endl; + } + } + } + } + + return 0; +} |