Laurent Carlier | 749294b | 2020-06-01 09:03:17 +0100 | [diff] [blame] | 1 | // |
Mike Kelly | 7cbe781 | 2023-07-25 17:37:33 +0100 | [diff] [blame] | 2 | // Copyright © 2017-2023 Arm Ltd and Contributors. All rights reserved. |
David Beck | ecb56cd | 2018-09-05 12:52:57 +0100 | [diff] [blame] | 3 | // SPDX-License-Identifier: MIT |
telsoa01 | 4fcda01 | 2018-03-09 14:13:49 +0000 | [diff] [blame] | 4 | // |
| 5 | #pragma once |
| 6 | |
Francis Murtagh | d301c0a | 2022-08-02 19:42:29 +0100 | [diff] [blame] | 7 | #include <BFloat16.hpp> |
Aron Virginas-Tar | c9cc804 | 2018-11-01 16:15:57 +0000 | [diff] [blame] | 8 | #include <Half.hpp> |
Matthew Bentham | 14e4669 | 2018-09-20 15:35:30 +0100 | [diff] [blame] | 9 | |
Aron Virginas-Tar | c9cc804 | 2018-11-01 16:15:57 +0000 | [diff] [blame] | 10 | #include <aclCommon/ArmComputeTensorUtils.hpp> |
| 11 | #include <cl/OpenClTimer.hpp> |
Colm Donelan | 0c47974 | 2021-12-10 12:43:54 +0000 | [diff] [blame] | 12 | #include <armnn/backends/TensorHandle.hpp> |
telsoa01 | c577f2c | 2018-08-31 09:22:23 +0100 | [diff] [blame] | 13 | |
Derek Lamberti | d466a54 | 2020-01-22 15:37:29 +0000 | [diff] [blame] | 14 | #include <armnn/Utils.hpp> |
| 15 | |
Matthew Bentham | 9b3e738 | 2020-02-05 21:39:55 +0000 | [diff] [blame] | 16 | #include <arm_compute/runtime/CL/CLTensor.h> |
| 17 | #include <arm_compute/runtime/IFunction.h> |
Aron Virginas-Tar | a8e06ed | 2018-10-19 16:46:15 +0100 | [diff] [blame] | 18 | |
| 19 | #include <sstream> |
| 20 | |
telsoa01 | c577f2c | 2018-08-31 09:22:23 +0100 | [diff] [blame] | 21 | #define ARMNN_SCOPED_PROFILING_EVENT_CL(name) \ |
| 22 | ARMNN_SCOPED_PROFILING_EVENT_WITH_INSTRUMENTS(armnn::Compute::GpuAcc, \ |
Keith Davis | 5a64f22 | 2021-08-04 10:35:20 +0100 | [diff] [blame] | 23 | armnn::EmptyOptional(), \ |
| 24 | name, \ |
| 25 | armnn::OpenClTimer(), \ |
| 26 | armnn::WallClockTimer()) |
| 27 | |
| 28 | #define ARMNN_SCOPED_PROFILING_EVENT_CL_GUID(name, guid) \ |
| 29 | ARMNN_SCOPED_PROFILING_EVENT_WITH_INSTRUMENTS(armnn::Compute::GpuAcc, \ |
| 30 | guid, \ |
Mike Kelly | 7cbe781 | 2023-07-25 17:37:33 +0100 | [diff] [blame] | 31 | GetName() + "_" + name, \ |
| 32 | armnn::OpenClTimer(), \ |
| 33 | armnn::WallClockTimer()) |
| 34 | |
| 35 | /// Creates a profiling event that uses GetGuid() and GetName() from the calling class |
| 36 | #define ARMNN_SCOPED_PROFILING_EVENT_CL_NAME_GUID(label) \ |
| 37 | ARMNN_SCOPED_PROFILING_EVENT_WITH_INSTRUMENTS(armnn::Compute::GpuAcc, \ |
| 38 | this->GetGuid(), \ |
| 39 | this->GetName() + "_" + label, \ |
telsoa01 | c577f2c | 2018-08-31 09:22:23 +0100 | [diff] [blame] | 40 | armnn::OpenClTimer(), \ |
| 41 | armnn::WallClockTimer()) |
telsoa01 | 4fcda01 | 2018-03-09 14:13:49 +0000 | [diff] [blame] | 42 | |
| 43 | namespace armnn |
| 44 | { |
| 45 | |
Keith Davis | 5a64f22 | 2021-08-04 10:35:20 +0100 | [diff] [blame] | 46 | inline std::string GetConvolutionMethodString(arm_compute::ConvolutionMethod& convolutionMethod) |
| 47 | { |
| 48 | switch (convolutionMethod) |
| 49 | { |
| 50 | case arm_compute::ConvolutionMethod::FFT: |
| 51 | return "FFT"; |
| 52 | case arm_compute::ConvolutionMethod::DIRECT: |
| 53 | return "Direct"; |
| 54 | case arm_compute::ConvolutionMethod::GEMM: |
| 55 | return "GEMM"; |
| 56 | case arm_compute::ConvolutionMethod::WINOGRAD: |
| 57 | return "Winograd"; |
| 58 | default: |
| 59 | return "Unknown"; |
| 60 | } |
| 61 | } |
| 62 | |
telsoa01 | 4fcda01 | 2018-03-09 14:13:49 +0000 | [diff] [blame] | 63 | template <typename T> |
Matthew Bentham | ca6616c | 2018-09-21 15:16:53 +0100 | [diff] [blame] | 64 | void CopyArmComputeClTensorData(arm_compute::CLTensor& dstTensor, const T* srcData) |
telsoa01 | 4fcda01 | 2018-03-09 14:13:49 +0000 | [diff] [blame] | 65 | { |
| 66 | { |
telsoa01 | c577f2c | 2018-08-31 09:22:23 +0100 | [diff] [blame] | 67 | ARMNN_SCOPED_PROFILING_EVENT_CL("MapClTensorForWriting"); |
telsoa01 | 4fcda01 | 2018-03-09 14:13:49 +0000 | [diff] [blame] | 68 | dstTensor.map(true); |
| 69 | } |
| 70 | |
| 71 | { |
telsoa01 | c577f2c | 2018-08-31 09:22:23 +0100 | [diff] [blame] | 72 | ARMNN_SCOPED_PROFILING_EVENT_CL("CopyToClTensor"); |
telsoa01 | 4fcda01 | 2018-03-09 14:13:49 +0000 | [diff] [blame] | 73 | armcomputetensorutils::CopyArmComputeITensorData<T>(srcData, dstTensor); |
| 74 | } |
| 75 | |
| 76 | dstTensor.unmap(); |
| 77 | } |
| 78 | |
keidav01 | d74dc91 | 2018-12-10 18:16:07 +0000 | [diff] [blame] | 79 | inline auto SetClStridedSliceData(const std::vector<int>& m_begin, |
| 80 | const std::vector<int>& m_end, |
| 81 | const std::vector<int>& m_stride) |
| 82 | { |
| 83 | arm_compute::Coordinates starts; |
| 84 | arm_compute::Coordinates ends; |
| 85 | arm_compute::Coordinates strides; |
| 86 | |
| 87 | unsigned int num_dims = static_cast<unsigned int>(m_begin.size()); |
| 88 | |
| 89 | for (unsigned int i = 0; i < num_dims; i++) { |
| 90 | unsigned int revertedIndex = num_dims - i - 1; |
| 91 | |
| 92 | starts.set(i, static_cast<int>(m_begin[revertedIndex])); |
| 93 | ends.set(i, static_cast<int>(m_end[revertedIndex])); |
| 94 | strides.set(i, static_cast<int>(m_stride[revertedIndex])); |
| 95 | } |
| 96 | |
| 97 | return std::make_tuple(starts, ends, strides); |
| 98 | } |
| 99 | |
Aron Virginas-Tar | 94c4fef | 2019-11-25 15:37:08 +0000 | [diff] [blame] | 100 | inline auto SetClSliceData(const std::vector<unsigned int>& m_begin, |
| 101 | const std::vector<unsigned int>& m_size) |
| 102 | { |
| 103 | // This function must translate the size vector given to an end vector |
| 104 | // expected by the ACL NESlice workload |
| 105 | arm_compute::Coordinates starts; |
| 106 | arm_compute::Coordinates ends; |
| 107 | |
| 108 | unsigned int num_dims = static_cast<unsigned int>(m_begin.size()); |
| 109 | |
| 110 | // For strided slices, we have the relationship size = (end - begin) / stride |
| 111 | // For slice, we assume stride to be a vector of all ones, yielding the formula |
| 112 | // size = (end - begin) therefore we know end = size + begin |
| 113 | for (unsigned int i = 0; i < num_dims; i++) |
| 114 | { |
| 115 | unsigned int revertedIndex = num_dims - i - 1; |
| 116 | |
| 117 | starts.set(i, static_cast<int>(m_begin[revertedIndex])); |
| 118 | ends.set(i, static_cast<int>(m_begin[revertedIndex] + m_size[revertedIndex])); |
| 119 | } |
| 120 | |
| 121 | return std::make_tuple(starts, ends); |
| 122 | } |
| 123 | |
Matthew Bentham | 785df50 | 2018-09-21 10:29:58 +0100 | [diff] [blame] | 124 | inline void InitializeArmComputeClTensorData(arm_compute::CLTensor& clTensor, |
James Conroy | 1f58f03 | 2021-04-27 17:13:27 +0100 | [diff] [blame] | 125 | const ConstTensorHandle* handle) |
telsoa01 | c577f2c | 2018-08-31 09:22:23 +0100 | [diff] [blame] | 126 | { |
Narumol Prangnawarat | ac2770a | 2020-04-01 16:51:23 +0100 | [diff] [blame] | 127 | ARMNN_ASSERT(handle); |
Matthew Bentham | ca6616c | 2018-09-21 15:16:53 +0100 | [diff] [blame] | 128 | |
| 129 | armcomputetensorutils::InitialiseArmComputeTensorEmpty(clTensor); |
telsoa01 | c577f2c | 2018-08-31 09:22:23 +0100 | [diff] [blame] | 130 | switch(handle->GetTensorInfo().GetDataType()) |
| 131 | { |
| 132 | case DataType::Float16: |
Matthew Bentham | ca6616c | 2018-09-21 15:16:53 +0100 | [diff] [blame] | 133 | CopyArmComputeClTensorData(clTensor, handle->GetConstTensor<armnn::Half>()); |
telsoa01 | c577f2c | 2018-08-31 09:22:23 +0100 | [diff] [blame] | 134 | break; |
| 135 | case DataType::Float32: |
Matthew Bentham | ca6616c | 2018-09-21 15:16:53 +0100 | [diff] [blame] | 136 | CopyArmComputeClTensorData(clTensor, handle->GetConstTensor<float>()); |
telsoa01 | c577f2c | 2018-08-31 09:22:23 +0100 | [diff] [blame] | 137 | break; |
Derek Lamberti | f90c56d | 2020-01-10 17:14:08 +0000 | [diff] [blame] | 138 | case DataType::QAsymmU8: |
Matthew Bentham | ca6616c | 2018-09-21 15:16:53 +0100 | [diff] [blame] | 139 | CopyArmComputeClTensorData(clTensor, handle->GetConstTensor<uint8_t>()); |
Matthew Bentham | 785df50 | 2018-09-21 10:29:58 +0100 | [diff] [blame] | 140 | break; |
Narumol Prangnawarat | 33d2c78 | 2020-11-13 18:00:23 +0000 | [diff] [blame] | 141 | case DataType::QAsymmS8: |
Derek Lamberti | d466a54 | 2020-01-22 15:37:29 +0000 | [diff] [blame] | 142 | case DataType::QSymmS8: |
Keith Davis | 899f64f | 2019-11-26 16:01:18 +0000 | [diff] [blame] | 143 | CopyArmComputeClTensorData(clTensor, handle->GetConstTensor<int8_t>()); |
| 144 | break; |
Ryan OShea | 2323af4 | 2020-05-13 16:36:19 +0100 | [diff] [blame] | 145 | case DataType::QSymmS16: |
| 146 | CopyArmComputeClTensorData(clTensor, handle->GetConstTensor<int16_t>()); |
| 147 | break; |
Matthew Bentham | 785df50 | 2018-09-21 10:29:58 +0100 | [diff] [blame] | 148 | case DataType::Signed32: |
Matthew Bentham | ca6616c | 2018-09-21 15:16:53 +0100 | [diff] [blame] | 149 | CopyArmComputeClTensorData(clTensor, handle->GetConstTensor<int32_t>()); |
Matthew Bentham | 785df50 | 2018-09-21 10:29:58 +0100 | [diff] [blame] | 150 | break; |
Francis Murtagh | d301c0a | 2022-08-02 19:42:29 +0100 | [diff] [blame] | 151 | case DataType::BFloat16: |
| 152 | CopyArmComputeClTensorData(clTensor, handle->GetConstTensor<armnn::BFloat16>()); |
| 153 | break; |
telsoa01 | c577f2c | 2018-08-31 09:22:23 +0100 | [diff] [blame] | 154 | default: |
Francis Murtagh | d301c0a | 2022-08-02 19:42:29 +0100 | [diff] [blame] | 155 | // Throw exception; assertion not called in release build. |
| 156 | throw Exception("Unexpected tensor type during InitializeArmComputeClTensorData()."); |
telsoa01 | c577f2c | 2018-08-31 09:22:23 +0100 | [diff] [blame] | 157 | } |
| 158 | }; |
| 159 | |
Aron Virginas-Tar | a8e06ed | 2018-10-19 16:46:15 +0100 | [diff] [blame] | 160 | inline RuntimeException WrapClError(const cl::Error& clError, const CheckLocation& location) |
| 161 | { |
| 162 | std::stringstream message; |
| 163 | message << "CL error: " << clError.what() << ". Error code: " << clError.err(); |
| 164 | |
| 165 | return RuntimeException(message.str(), location); |
| 166 | } |
| 167 | |
| 168 | inline void RunClFunction(arm_compute::IFunction& function, const CheckLocation& location) |
| 169 | { |
| 170 | try |
| 171 | { |
| 172 | function.run(); |
| 173 | } |
| 174 | catch (cl::Error& error) |
| 175 | { |
| 176 | throw WrapClError(error, location); |
| 177 | } |
| 178 | } |
| 179 | |
David Monahan | c11ba46 | 2020-12-03 11:09:46 +0000 | [diff] [blame] | 180 | template <typename DataType, typename PayloadType> |
| 181 | DataType* GetOutputTensorData(unsigned int idx, const PayloadType& data) |
| 182 | { |
| 183 | ITensorHandle* tensorHandle = data.m_Outputs[idx]; |
| 184 | return reinterpret_cast<DataType*>(tensorHandle->Map()); |
| 185 | } |
| 186 | |
telsoa01 | 4fcda01 | 2018-03-09 14:13:49 +0000 | [diff] [blame] | 187 | } //namespace armnn |