mbaret commented on a change in pull request #6222: URL: https://github.com/apache/incubator-tvm/pull/6222#discussion_r467867668
########## File path: src/runtime/contrib/ethosn/ethosn_device.cc ########## @@ -0,0 +1,222 @@ +/* + * Licensed to the Apache Software Foundation (ASF) under one + * or more contributor license agreements. See the NOTICE file + * distributed with this work for additional information + * regarding copyright ownership. The ASF licenses this file + * to you 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. + */ + +/*! + * \file ethosn_device.cc + * \brief Ethos-N NPU device integration. + */ + +#include <dlpack/dlpack.h> +#include <poll.h> +#include <tvm/tir/expr.h> +#include <unistd.h> + +#include <algorithm> +#include <memory> + +#include "ethosn_driver_library/Buffer.hpp" +#include "ethosn_support_library/Support.hpp" + +#if defined ETHOSN_HW + +#include "ethosn_driver_library/Inference.hpp" +#include "ethosn_driver_library/Network.hpp" + +namespace tvm { +namespace runtime { +namespace ethosn { + +namespace sl = ::ethosn::support_library; +namespace dl = ::ethosn::driver_library; + +int64_t GetTensorSize(const DLTensor& tensor) { + int64_t size = 1; + for (int i = 0; i < tensor.ndim; i++) { + size *= tensor.shape[i]; + } + return size; +} + +bool WaitForInference(dl::Inference* inference, int timeout) { + // Wait for inference to complete + int fd = inference->GetFileDescriptor(); + struct pollfd fds; + memset(&fds, 0, sizeof(fds)); + fds.fd = fd; + fds.events = POLLIN; // Wait for any available input. + + const int ms_per_seconds = 1000; + int poll_result = poll(&fds, 1, timeout * ms_per_seconds); + if (poll_result > 0) { + dl::InferenceResult result; + if (read(fd, &result, sizeof(result)) != sizeof(result)) { + return false; + } + if (result != dl::InferenceResult::Completed) { + return false; + } + } else if (poll_result == 0) { + return false; + } else { + return false; + } + return true; +} + +template <typename T> +void CopyOutput(dl::Buffer* source_buffers[], std::vector<DLTensor*>* outputs) { + for (DLTensor* tensor : *outputs) { + dl::Buffer* source_buffer = source_buffers[0]; + uint8_t* source_buffer_data = source_buffer->GetMappedBuffer(); + size_t size = source_buffer->GetSize(); + T* dest_pointer = static_cast<T*>(tensor->data); + std::copy_backward(source_buffer_data, source_buffer_data + size, dest_pointer + size); + source_buffers++; + } +} + +void CreateBuffers(std::vector<std::shared_ptr<dl::Buffer> >* fm, + const std::vector<DLTensor*>& tensors) { + int index = 0; + for (auto buffer : tensors) { + auto* data = static_cast<uint8_t*>(buffer->data); + // The NPU only needs the size of the tensor * uint8_t. + auto data_size = static_cast<uint32_t>(GetTensorSize(*buffer)); + (*fm)[index++] = std::make_shared<dl::Buffer>(data, data_size, dl::DataFormat::NHWC); + } +} + +bool Inference(tvm::runtime::TVMArgs args, sl::CompiledNetwork* network, + std::vector<uint32_t> input_order, std::vector<uint32_t> output_order) { + // Unpack parameters + uint8_t argc = 0; + std::vector<DLTensor*> inputs(input_order.size()); + for (uint8_t i = 0; i < network->GetInputBufferInfos().size(); i++) { + inputs[input_order[i]] = args[argc++]; + } + auto out_infos = network->GetOutputBufferInfos(); + std::vector<DLTensor*> outputs(output_order.size()); + for (uint8_t i = 0; i < network->GetOutputBufferInfos().size(); i++) { + outputs[output_order[i]] = args[argc++]; + } + + // Set up input buffers + std::vector<std::shared_ptr<dl::Buffer> > ifm(inputs.size()); + CreateBuffers(&ifm, inputs); + + // Set up output buffers + std::vector<std::shared_ptr<dl::Buffer> > ofm(outputs.size()); + CreateBuffers(&ofm, outputs); + + // Raw pointers for the inference + dl::Buffer* ifm_raw[inputs.size()]; + for (size_t i = 0; i < inputs.size(); i++) { + ifm_raw[i] = ifm[i].get(); + } + dl::Buffer* ofm_raw[outputs.size()]; + for (size_t i = 0; i < outputs.size(); i++) { + ofm_raw[i] = ofm[i].get(); + } + + auto npu = std::make_unique<dl::Network>(*network); + + // Execute the inference. + std::unique_ptr<dl::Inference> result( + npu->ScheduleInference(ifm_raw, sizeof(ifm_raw) / sizeof(ifm_raw[0]), ofm_raw, + sizeof(ofm_raw) / sizeof(ofm_raw[0]))); + bool inferenceCompleted = WaitForInference(result.get(), 60); + if (inferenceCompleted) { + switch ((outputs)[0]->dtype.bits) { + case 8: { + dl::Buffer** ofms = &ofm_raw[0]; + for (DLTensor* tensor : outputs) { + uint8_t* source_buffer_data = (*ofms++)->GetMappedBuffer(); + uint8_t* dest_pointer = static_cast<uint8_t*>(tensor->data); + if (source_buffer_data != dest_pointer) { + CopyOutput<uint8_t>(ofm_raw, &outputs); + break; + } + } + break; + } + case 16: + CopyOutput<uint16_t>(ofm_raw, &outputs); + break; + case 32: + CopyOutput<uint32_t>(ofm_raw, &outputs); + break; + default: + break; + } + } + + return inferenceCompleted; +} + +} // namespace ethosn +} // namespace runtime +} // namespace tvm + +#else Review comment: I've added a comment here. Let me know if it's useful. ---------------------------------------------------------------- This is an automated message from the Apache Git Service. To respond to the message, please log on to GitHub and use the URL above to go to the specific comment. For queries about this service, please contact Infrastructure at: [email protected]
