diff options
Diffstat (limited to 'src/backends/gpuFsa/workloads/GpuFsaPreCompiledWorkload.cpp')
| -rw-r--r-- | src/backends/gpuFsa/workloads/GpuFsaPreCompiledWorkload.cpp | 40 |
1 files changed, 26 insertions, 14 deletions
diff --git a/src/backends/gpuFsa/workloads/GpuFsaPreCompiledWorkload.cpp b/src/backends/gpuFsa/workloads/GpuFsaPreCompiledWorkload.cpp index 20386b5d86..1663044a17 100644 --- a/src/backends/gpuFsa/workloads/GpuFsaPreCompiledWorkload.cpp +++ b/src/backends/gpuFsa/workloads/GpuFsaPreCompiledWorkload.cpp @@ -10,6 +10,7 @@ #include <gpuFsa/GpuFsaTensorHandle.hpp> #include <gpuFsa/GpuFsaBackend.hpp> #include <aclCommon/ArmComputeTensorUtils.hpp> +#include <fmt/format.h> #include <aclCommon/ArmComputeTensorUtils.hpp> #include <arm_compute/runtime/CL/CLTensor.h> @@ -21,7 +22,6 @@ #include <arm_compute/dynamic_fusion/runtime/gpu/cl/ClWorkloadRuntime.h> #include <arm_compute/dynamic_fusion/sketch/gpu/GpuWorkloadContext.h> -#include <src/dynamic_fusion/sketch/gpu/GpuWorkloadContextImpl.h> #include <arm_compute/dynamic_fusion/sketch/gpu/operators/GpuConv2d.h> #include <arm_compute/dynamic_fusion/sketch/gpu/operators/GpuOutput.h> #include <arm_compute/dynamic_fusion/sketch/gpu/GpuWorkloadSketch.h> @@ -48,26 +48,38 @@ void GpuFsaPreCompiledWorkload::Execute() const * First we need all of the data stored within the PreCompiled blob which was used to setup the workload, namely: * The GpuWorkloadContext, this is a context which contains the TensorInfos and is unique to the graph being run * The Sketch, this can contain one or many ops and acts as a subgraph within the context - * The TensorInfoIds, These are the ids of the TensorInfos used during the creation of the Sketch and stored within - * the context. + * The inputTensorInfos / outputTensorInfos, These are vectors containing the TensorInfos used when creating the sketch + * * It is very important that the Tensors passed into the Runtime being used to execute this sketch are created with * the same TensorInfos as used when creating the sketch. We do this by creating new tensors, getting the original - * TensorInfos from the GpuWorkloadContext via their ids, and then importing the buffers from our own TensorHandles - * directly into these newly created Tensors. This allows us to link the externally visible Tensors from ArmNN to the - * Tensors which are needed to execute with the Sketch. + * TensorInfos from the vectors of tensorInfos stored in the blob, and then importing the buffers from our own + * TensorHandles directly into these newly created Tensors. This allows us to link the externally visible Tensors + * from ArmNN to the Tensors which are needed to execute with the Sketch. * */ using namespace arm_compute::experimental::dynamic_fusion; // Get the runtime and configure it with the precompiled sketch ClWorkloadRuntime runtime; GpuFsaPreCompiledBlob *preCompiledBlob = static_cast<GpuFsaPreCompiledBlob*>(m_Data.m_PreCompiledObject); - auto workloadContext = - &(preCompiledBlob->workloadContext->implementation()); auto sketch = preCompiledBlob->sketch.release(); - std::vector<int32_t> inputIds = *(preCompiledBlob->inputIds.get()); - std::vector<int32_t> outputIds = *(preCompiledBlob->outputIds.get()); auto status = runtime.configure(*sketch); + // Get the TensorInfos stored within the PreCompiledBlob and check they're the right size + auto inputTensorInfos = preCompiledBlob->inputTensorInfos.get(); + auto outputTensorInfos = preCompiledBlob->outputTensorInfos.get(); + if (inputTensorInfos->size() != m_Data.m_Inputs.size()) + { + throw InvalidArgumentException(fmt::format("GpuFsaPreCompiledWorkload::Execute: The number of inputTensorInfos" + " {} does not match the number of inputs {}.", + inputTensorInfos->size(), m_Data.m_Inputs.size())); + } + if (outputTensorInfos->size() != m_Data.m_Outputs.size()) + { + throw InvalidArgumentException(fmt::format("GpuFsaPreCompiledWorkload::Execute: The number of outputTensorInfos" + " {} does not match the number of outputs {}.", + outputTensorInfos->size(), m_Data.m_Outputs.size())); + } + // (Important) Allocate auxiliary tensor memory if there are any for(auto &data : runtime.get_auxiliary_tensors()) { @@ -85,8 +97,8 @@ void GpuFsaPreCompiledWorkload::Execute() const for (uint32_t inputSlotIdx = 0; inputSlotIdx < m_Data.m_Inputs.size(); ++inputSlotIdx) { arm_compute::CLTensor* input = new arm_compute::CLTensor{}; - input->allocator()->init(*(dynamic_cast<arm_compute::TensorInfo*>( - workloadContext->get_tensor_info(inputIds[inputSlotIdx])))); + // inputTensorInfos is a ptr to a vector of ptrs, so we need to do a double dereference + input->allocator()->init(*((*inputTensorInfos)[inputSlotIdx])); auto* inputHandle = PolymorphicDowncast<GpuFsaTensorHandle*>(m_Data.m_Inputs[inputSlotIdx]); input->allocator()->import_memory(inputHandle->GetTensor().cl_buffer()); inputsWeightsOutputs.emplace_back(std::move(input)); @@ -95,8 +107,8 @@ void GpuFsaPreCompiledWorkload::Execute() const for (uint32_t outputSlotIdx = 0; outputSlotIdx < m_Data.m_Outputs.size(); ++outputSlotIdx) { arm_compute::CLTensor* output = new arm_compute::CLTensor{}; - output->allocator()->init(*(dynamic_cast<arm_compute::TensorInfo*>( - workloadContext->get_tensor_info(outputIds[outputSlotIdx])))); + // outputTensorInfos is a ptr to a vector of ptrs, so we need to do a double dereference + output->allocator()->init(*((*outputTensorInfos)[outputSlotIdx])); auto* outputHandle = PolymorphicDowncast<GpuFsaTensorHandle*>(m_Data.m_Outputs[outputSlotIdx]); output->allocator()->import_memory(outputHandle->GetTensor().cl_buffer()); inputsWeightsOutputs.emplace_back(std::move(output)); |