Narumol Prangnawarat | 9ef3614 | 2022-01-25 15:15:34 +0000 | [diff] [blame] | 1 | // |
| 2 | // Copyright © 2022 Arm Ltd and Contributors. All rights reserved. |
| 3 | // SPDX-License-Identifier: MIT |
| 4 | // |
| 5 | |
| 6 | #include <arm_compute/runtime/CL/functions/CLActivationLayer.h> |
| 7 | |
| 8 | #include <armnnTestUtils/TensorCopyUtils.hpp> |
| 9 | |
| 10 | #include <cl/ClImportTensorHandle.hpp> |
| 11 | #include <cl/ClImportTensorHandleFactory.hpp> |
| 12 | #include <cl/ClTensorHandle.hpp> |
| 13 | #include <cl/ClTensorHandleFactory.hpp> |
| 14 | #include <cl/ICLTensorProxy.hpp> |
| 15 | #include <cl/test/ClContextControlFixture.hpp> |
| 16 | #include <cl/test/ClWorkloadFactoryHelper.hpp> |
| 17 | |
| 18 | #include <doctest/doctest.h> |
| 19 | |
| 20 | using namespace armnn; |
| 21 | |
| 22 | TEST_SUITE("ICLTensorProxyTests") |
| 23 | { |
| 24 | |
| 25 | TEST_CASE_FIXTURE(ClContextControlFixture, "ICLTensorProxyTest") |
| 26 | { |
| 27 | ClTensorHandleFactory handleFactory = |
| 28 | ClWorkloadFactoryHelper::GetTensorHandleFactory(ClWorkloadFactoryHelper::GetMemoryManager()); |
| 29 | |
| 30 | TensorInfo info({ 1, 3, 4, 1 }, DataType::Float32); |
| 31 | |
| 32 | // create TensorHandle for memory import |
| 33 | auto handle = handleFactory.CreateTensorHandle(info, true); |
| 34 | |
| 35 | std::vector<float> inputData |
| 36 | { |
| 37 | -5, -2, 1, 2, |
| 38 | 3, 10, -20, 8, |
| 39 | 0, -12, 7, -9 |
| 40 | }; |
| 41 | |
| 42 | handle->Allocate(); |
| 43 | |
| 44 | CopyDataToITensorHandle(handle.get(), inputData.data()); |
| 45 | |
| 46 | // Get CLtensor |
| 47 | arm_compute::CLTensor& tensor = PolymorphicDowncast<ClTensorHandle*>(handle.get())->GetTensor(); |
| 48 | ICLTensorProxy iclTensorProxy(&tensor); |
| 49 | |
| 50 | // Check that the ICLTensorProxy get correct information from the delegate tensor |
| 51 | CHECK((iclTensorProxy.info() == tensor.info())); |
| 52 | CHECK((iclTensorProxy.buffer() == tensor.buffer())); |
| 53 | CHECK((iclTensorProxy.cl_buffer() == tensor.cl_buffer())); |
| 54 | CHECK((iclTensorProxy.quantization().scale == tensor.quantization().scale)); |
| 55 | CHECK((iclTensorProxy.quantization().offset == tensor.quantization().offset)); |
| 56 | } |
| 57 | |
| 58 | TEST_CASE_FIXTURE(ClContextControlFixture, "ChangeICLTensorProxyExecutionTest") |
| 59 | { |
| 60 | // Start execution with with copied tensor |
| 61 | ClTensorHandleFactory handleFactory = |
| 62 | ClWorkloadFactoryHelper::GetTensorHandleFactory(ClWorkloadFactoryHelper::GetMemoryManager()); |
| 63 | |
| 64 | TensorInfo info({ 1, 3, 4, 1 }, DataType::Float32); |
| 65 | unsigned int numElements = info.GetNumElements(); |
| 66 | |
| 67 | // create TensorHandle for memory import |
| 68 | auto handle = handleFactory.CreateTensorHandle(info, true); |
| 69 | |
| 70 | std::vector<float> inputData |
| 71 | { |
| 72 | -5, -2, 1, 2, |
| 73 | 3, 10, -20, 8, |
| 74 | 0, -12, 7, -9 |
| 75 | }; |
| 76 | |
| 77 | std::vector<float> ExpectedOutput |
| 78 | { |
| 79 | 0, 0, 1, 2, |
| 80 | 3, 10, 0, 8, |
| 81 | 0, 0, 7, 0 |
| 82 | }; |
| 83 | |
| 84 | handle->Allocate(); |
| 85 | |
| 86 | CopyDataToITensorHandle(handle.get(), inputData.data()); |
| 87 | |
| 88 | // Get CLtensor |
| 89 | arm_compute::CLTensor& tensor = PolymorphicDowncast<ClTensorHandle*>(handle.get())->GetTensor(); |
| 90 | |
| 91 | // Set a proxy tensor to allocated tensor |
| 92 | std::unique_ptr<ICLTensorProxy> iclTensorProxy; |
| 93 | iclTensorProxy = std::make_unique<ICLTensorProxy>(&tensor); |
| 94 | |
| 95 | // Create and configure activation function |
| 96 | const arm_compute::ActivationLayerInfo act_info(arm_compute::ActivationLayerInfo::ActivationFunction::RELU); |
| 97 | arm_compute::CLActivationLayer act_func; |
| 98 | act_func.configure(iclTensorProxy.get(), nullptr, act_info); |
| 99 | |
| 100 | act_func.run(); |
| 101 | arm_compute::CLScheduler::get().sync(); |
| 102 | |
| 103 | std::vector<float> actualOutput(info.GetNumElements()); |
| 104 | |
| 105 | CopyDataFromITensorHandle(actualOutput.data(), handle.get()); |
| 106 | |
| 107 | // Validate result as expected output |
| 108 | for(unsigned int i = 0; i < numElements; ++i) |
| 109 | { |
| 110 | CHECK((actualOutput[i] == ExpectedOutput[i])); |
| 111 | } |
| 112 | |
| 113 | // Change to execute with imported tensor |
| 114 | ClImportTensorHandleFactory importHandleFactory(static_cast<MemorySourceFlags>(MemorySource::Malloc), |
| 115 | static_cast<MemorySourceFlags>(MemorySource::Malloc)); |
| 116 | // create TensorHandle for memory import |
| 117 | auto importHandle = importHandleFactory.CreateTensorHandle(info); |
| 118 | |
| 119 | // Get CLtensor |
| 120 | arm_compute::CLTensor& importTensor = PolymorphicDowncast<ClImportTensorHandle*>(importHandle.get())->GetTensor(); |
| 121 | |
| 122 | // Allocate user memory |
| 123 | const size_t totalBytes = importTensor.info()->total_size(); |
| 124 | const size_t alignment = |
| 125 | arm_compute::CLKernelLibrary::get().get_device().getInfo<CL_DEVICE_GLOBAL_MEM_CACHELINE_SIZE>(); |
| 126 | size_t space = totalBytes + alignment + alignment; |
| 127 | auto testData = std::make_unique<uint8_t[]>(space); |
| 128 | void* alignedPtr = testData.get(); |
| 129 | CHECK(std::align(alignment, totalBytes, alignedPtr, space)); |
| 130 | |
| 131 | // Import memory |
| 132 | CHECK(importHandle->Import(alignedPtr, armnn::MemorySource::Malloc)); |
| 133 | |
| 134 | // Input with negative values |
| 135 | auto* typedPtr = reinterpret_cast<float*>(alignedPtr); |
| 136 | std::fill_n(typedPtr, numElements, -5.0f); |
| 137 | |
| 138 | // Set the import Tensor to TensorProxy to change Tensor in the CLActivationLayer without calling configure function |
| 139 | iclTensorProxy->set(&importTensor); |
| 140 | |
| 141 | // Execute function and sync |
| 142 | act_func.run(); |
| 143 | arm_compute::CLScheduler::get().sync(); |
| 144 | |
| 145 | // Validate result by checking that the output has no negative values |
| 146 | for(unsigned int i = 0; i < numElements; ++i) |
| 147 | { |
| 148 | CHECK(typedPtr[i] == 0); |
| 149 | } |
| 150 | } |
| 151 | } |