diff options
| author | Teresa Charlin <teresa.charlinreyes@arm.com> | 2020-09-29 08:38:47 +0100 |
|---|---|---|
| committer | Teresa Charlin <teresa.charlinreyes@arm.com> | 2020-10-01 11:41:33 +0100 |
| commit | a2a512c60e852a82b9d1e85bb7b8874723dda161 (patch) | |
| tree | ed441ed52311f4be67b146f280a3b6f1d8aa2d80 /src/backends/cl/workloads/ClMultiplicationWorkload.cpp | |
| parent | ea8ce7040476da46e145705b0b08e9449144a3b2 (diff) | |
| download | armnn-a2a512c60e852a82b9d1e85bb7b8874723dda161.tar.gz | |
COMPMID-3784 Fix 1 CTS MUL INT32 failure due to using SATURATE
* LargeGraph_TENSOR_INT32_Rank4/26
Signed-off-by: Teresa Charlin <teresa.charlinreyes@arm.com>
Change-Id: I9d07444db56e26c13a77bf022938644ed7953d6b
Diffstat (limited to 'src/backends/cl/workloads/ClMultiplicationWorkload.cpp')
| -rw-r--r-- | src/backends/cl/workloads/ClMultiplicationWorkload.cpp | 14 |
1 files changed, 12 insertions, 2 deletions
diff --git a/src/backends/cl/workloads/ClMultiplicationWorkload.cpp b/src/backends/cl/workloads/ClMultiplicationWorkload.cpp index f4e7b836ed..e9b75c3f10 100644 --- a/src/backends/cl/workloads/ClMultiplicationWorkload.cpp +++ b/src/backends/cl/workloads/ClMultiplicationWorkload.cpp @@ -19,6 +19,10 @@ arm_compute::Status ClMultiplicationWorkloadValidate(const TensorInfo& input0, const arm_compute::TensorInfo aclInput2 = armcomputetensorutils::BuildArmComputeTensorInfo(input1); const arm_compute::TensorInfo aclOutput = armcomputetensorutils::BuildArmComputeTensorInfo(output); + auto convertPolicy = (IsQuantizedType(input0.GetDataType()) || IsQuantizedType(input1.GetDataType())) ? + arm_compute::ConvertPolicy::SATURATE : + arm_compute::ConvertPolicy::WRAP; + // At the time of writing, configure() will fail if a rounding policy other than TO_ZERO is supplied to it, // when providing a scale of 1.0 for F32 tensors, even though the provided rounding policy appears to be // ignored for F32 tensors. @@ -26,7 +30,7 @@ arm_compute::Status ClMultiplicationWorkloadValidate(const TensorInfo& input0, &aclInput2, &aclOutput, 1.0f, - arm_compute::ConvertPolicy::SATURATE, + convertPolicy, arm_compute::RoundingPolicy::TO_ZERO); } @@ -40,12 +44,18 @@ ClMultiplicationWorkload::ClMultiplicationWorkload(const MultiplicationQueueDesc arm_compute::ICLTensor& input0 = static_cast<IClTensorHandle*>(m_Data.m_Inputs[0])->GetTensor(); arm_compute::ICLTensor& input1 = static_cast<IClTensorHandle*>(m_Data.m_Inputs[1])->GetTensor(); arm_compute::ICLTensor& output = static_cast<IClTensorHandle*>(m_Data.m_Outputs[0])->GetTensor(); + + auto convertPolicy = (IsQuantizedType(info.m_InputTensorInfos[0].GetDataType()) || + IsQuantizedType(info.m_InputTensorInfos[1].GetDataType())) ? + arm_compute::ConvertPolicy::SATURATE : + arm_compute::ConvertPolicy::WRAP; + // Construct m_PixelWiseMultiplication.configure(&input0, &input1, &output, 1.0f, - arm_compute::ConvertPolicy::SATURATE, + convertPolicy, arm_compute::RoundingPolicy::TO_NEAREST_EVEN); } |