diff options
| author | Ryan OShea <ryan.oshea3@arm.com> | 2022-11-07 16:20:48 +0000 |
|---|---|---|
| committer | ryan.oshea3 <ryan.oshea3@arm.com> | 2022-11-16 15:22:50 +0000 |
| commit | 31441595009182c985dacbedc70c41ee6664d070 (patch) | |
| tree | 248a85295aeff4022c9b395fc97748b0a0aa6b35 /src/backends/reference/workloads/RefConvertBf16ToFp32Workload.cpp | |
| parent | bd18eab07a8f30492de1e462b1815189014cb8d5 (diff) | |
| download | armnn-31441595009182c985dacbedc70c41ee6664d070.tar.gz | |
IVGCVSW-7214 Disable BF16-Turbo-Mode and remove conversion layers
- Remove Bf16ToFp32 Conversion Layer
- Remove Fp32ToBf16 Conversion Layer
- Remove B16 Conversion tests
* Throw exception if m_ReduceFp32ToBf16 optimzer option is set to true
* Provide comments to enable fast math in order to use bf16
* Update docs to inform users to enable fast math for bf16
Execute Network Changes
* Require bf16_turbo_mode to also have fast_math_enabled set to true
- Remove setting m_ReduceFp32ToBf16 optimizer option
Signed-off-by: Ryan OShea <ryan.oshea3@arm.com>
Change-Id: Ibaa6da9d29c96a1ce32ff5196b0847fde9f04a1c
Diffstat (limited to 'src/backends/reference/workloads/RefConvertBf16ToFp32Workload.cpp')
| -rw-r--r-- | src/backends/reference/workloads/RefConvertBf16ToFp32Workload.cpp | 39 |
1 files changed, 0 insertions, 39 deletions
diff --git a/src/backends/reference/workloads/RefConvertBf16ToFp32Workload.cpp b/src/backends/reference/workloads/RefConvertBf16ToFp32Workload.cpp deleted file mode 100644 index 2fe2eafb9b..0000000000 --- a/src/backends/reference/workloads/RefConvertBf16ToFp32Workload.cpp +++ /dev/null @@ -1,39 +0,0 @@ -// -// Copyright © 2022 Arm Ltd and Contributors. All rights reserved. -// SPDX-License-Identifier: MIT -// - -#include "RefConvertBf16ToFp32Workload.hpp" -#include "RefWorkloadUtils.hpp" - -#include <armnnUtils/FloatingPointConverter.hpp> - -#include <BFloat16.hpp> - -namespace armnn -{ - -void RefConvertBf16ToFp32Workload::Execute() const -{ - Execute(m_Data.m_Inputs, m_Data.m_Outputs); -} - -void RefConvertBf16ToFp32Workload::ExecuteAsync(ExecutionData& executionData) -{ - WorkingMemDescriptor* workingMemDescriptor = static_cast<WorkingMemDescriptor*>(executionData.m_Data); - Execute(workingMemDescriptor->m_Inputs, workingMemDescriptor->m_Outputs); -} - -void RefConvertBf16ToFp32Workload::Execute(std::vector<ITensorHandle*> inputs, - std::vector<ITensorHandle*> outputs) const -{ - ARMNN_SCOPED_PROFILING_EVENT(Compute::CpuRef, "RefConvertBf16ToFp32Workload_Execute"); - - const BFloat16* const input = reinterpret_cast<const BFloat16*>(inputs[0]->Map()); - float* const output = reinterpret_cast<float*>(outputs[0]->Map()); - - unsigned int numElements = GetTensorInfo(inputs[0]).GetNumElements(); - armnnUtils::FloatingPointConverter::ConvertBFloat16ToFloat32(input, numElements, output); -} - -} //namespace armnn |