From 2e4d889fb036d1c0a34503400a3f45cfc6f9f3e1 Mon Sep 17 00:00:00 2001 From: Matthew Sloyan Date: Tue, 18 Oct 2022 18:02:48 +0100 Subject: Add FP16 support to IModelRunner * Added specific FP16 readfromVector and writeToVector methods. * Added FP16 support to float readfromVector and writeToVector methods. * Added missing reference to IModelRunner::setInput. Signed-off-by: Matthew Sloyan Change-Id: I6b66468737e672afc925ccad4fb710fbb9427c14 --- reference_model/include/model_runner.h | 2 +- reference_model/src/model_runner.cc | 12 +++--- reference_model/src/model_runner_impl.cc | 14 ++++--- reference_model/src/model_runner_impl.h | 2 +- reference_model/src/tensor.cc | 65 ++++++++++++++++++++++++++++++++ reference_model/src/tensor.h | 2 + 6 files changed, 84 insertions(+), 13 deletions(-) diff --git a/reference_model/include/model_runner.h b/reference_model/include/model_runner.h index 4629467..4335794 100644 --- a/reference_model/include/model_runner.h +++ b/reference_model/include/model_runner.h @@ -68,7 +68,7 @@ public: * NOTE: setInput() must be called for each input tensor before run() is called. */ template - int setInput(std::string input_name, std::vector vals); + int setInput(std::string input_name, std::vector& vals); /* * Retrieve the output tensors from the graph after running. diff --git a/reference_model/src/model_runner.cc b/reference_model/src/model_runner.cc index 2395a85..5c086e6 100644 --- a/reference_model/src/model_runner.cc +++ b/reference_model/src/model_runner.cc @@ -53,7 +53,7 @@ GraphStatus IModelRunner::run() } template -int IModelRunner::setInput(std::string input_name, std::vector vals) +int IModelRunner::setInput(std::string input_name, std::vector& vals) { return model_runner_impl->setInput(input_name, vals); } @@ -65,12 +65,14 @@ std::vector IModelRunner::getOutput(std::string output_name) } // Template explicit specialization -template int IModelRunner::setInput(std::string input_name, std::vector vals); -template int IModelRunner::setInput(std::string input_name, std::vector vals); -template int IModelRunner::setInput(std::string input_name, std::vector vals); -template int IModelRunner::setInput(std::string input_name, std::vector vals); +template int IModelRunner::setInput(std::string input_name, std::vector& vals); +template int IModelRunner::setInput(std::string input_name, std::vector& vals); +template int IModelRunner::setInput(std::string input_name, std::vector& vals); +template int IModelRunner::setInput(std::string input_name, std::vector& vals); +template int IModelRunner::setInput(std::string input_name, std::vector& vals); template std::vector IModelRunner::getOutput(std::string output_name); +template std::vector IModelRunner::getOutput(std::string output_name); template std::vector IModelRunner::getOutput(std::string output_name); template std::vector IModelRunner::getOutput(std::string output_name); template std::vector IModelRunner::getOutput(std::string output_name); \ No newline at end of file diff --git a/reference_model/src/model_runner_impl.cc b/reference_model/src/model_runner_impl.cc index e0fdc49..8427150 100644 --- a/reference_model/src/model_runner_impl.cc +++ b/reference_model/src/model_runner_impl.cc @@ -156,7 +156,7 @@ done: } template -int ModelRunnerImpl::setInput(std::string input_name, std::vector vals) +int ModelRunnerImpl::setInput(std::string input_name, std::vector& vals) { if (_main_gt == nullptr) { @@ -214,7 +214,7 @@ std::vector ModelRunnerImpl::getOutput(std::string output_name) return std::vector(); } - std::vector outputs(tensor->getElementCount(), 0); + std::vector outputs(tensor->getElementCount()); if (tensor->writeToVector(outputs)) { @@ -266,12 +266,14 @@ void ModelRunnerImpl::checkGraphStatus(SubgraphTraverser& main_gt) } // Template explicit specialization -template int ModelRunnerImpl::setInput(std::string input_name, std::vector vals); -template int ModelRunnerImpl::setInput(std::string input_name, std::vector vals); -template int ModelRunnerImpl::setInput(std::string input_name, std::vector vals); -template int ModelRunnerImpl::setInput(std::string input_name, std::vector vals); +template int ModelRunnerImpl::setInput(std::string input_name, std::vector& vals); +template int ModelRunnerImpl::setInput(std::string input_name, std::vector& vals); +template int ModelRunnerImpl::setInput(std::string input_name, std::vector& vals); +template int ModelRunnerImpl::setInput(std::string input_name, std::vector& vals); +template int ModelRunnerImpl::setInput(std::string input_name, std::vector& vals); template std::vector ModelRunnerImpl::getOutput(std::string output_name); +template std::vector ModelRunnerImpl::getOutput(std::string output_name); template std::vector ModelRunnerImpl::getOutput(std::string output_name); template std::vector ModelRunnerImpl::getOutput(std::string output_name); template std::vector ModelRunnerImpl::getOutput(std::string output_name); \ No newline at end of file diff --git a/reference_model/src/model_runner_impl.h b/reference_model/src/model_runner_impl.h index 7a91bfe..f26c484 100644 --- a/reference_model/src/model_runner_impl.h +++ b/reference_model/src/model_runner_impl.h @@ -46,7 +46,7 @@ public: GraphStatus run(); template - int setInput(std::string input_name, std::vector vals); + int setInput(std::string input_name, std::vector& vals); template std::vector getOutput(std::string output_name); diff --git a/reference_model/src/tensor.cc b/reference_model/src/tensor.cc index 8e65a27..8d192ca 100644 --- a/reference_model/src/tensor.cc +++ b/reference_model/src/tensor.cc @@ -429,6 +429,7 @@ int TosaReference::Tensor::readfromVector(const std::vector& vals) uint32_t elements = getElementCount(); switch (getDtype()) { + case DType_FP16: case DType_FP32: if (vals.size() != elements) { @@ -448,6 +449,38 @@ int TosaReference::Tensor::readfromVector(const std::vector& vals) return 0; } +int TosaReference::Tensor::readfromVector(const std::vector& vals) +{ + uint32_t elements = getElementCount(); + std::vector tensor(elements); + + switch (getDtype()) + { + case DType_FP16: + if (vals.size() != elements) + { + WARNING("The input size (%ld) doesn't match the number of elements (%d) assigned to the tensor.", + vals.size(), elements); + return -1; + } + + // Convert from fp16 to fp32 + for (uint32_t i=0; i < elements; i++) + { + tensor[i] = half_float::half_cast(vals[i]); + } + + setTensorValueFloat(elements, tensor.data()); + break; + default: + WARNING("The input type doesn't match the data type assigned to the tensor (%s).", + EnumNameDType(getDtype())); + return -2; + } + setIsValid(); + return 0; +} + int TosaReference::Tensor::readfromVector(const std::vector& vals) { uint32_t elements = getElementCount(); @@ -532,6 +565,7 @@ int TosaReference::Tensor::writeToVector(std::vector& vals) switch (getDtype()) { + case DType_FP16: case DType_FP32: if (vals.size() != elements) { @@ -550,6 +584,37 @@ int TosaReference::Tensor::writeToVector(std::vector& vals) return 0; } +int TosaReference::Tensor::writeToVector(std::vector& vals) +{ + uint32_t elements = getElementCount(); + std::vector tensor(elements); + + switch (getDtype()) + { + case DType_FP16: + if (vals.size() != elements) + { + WARNING("The output size (%ld) doesn't match the number of elements (%d) assigned to the tensor.", + vals.size(), elements); + return -1; + } + + getTensorValueFloat(elements, tensor.data()); + + // Convert fp32 to fp16 + for (uint32_t i=0; i < elements; i++) + { + vals[i] = half_float::half_cast(tensor[i]); + } + break; + default: + WARNING("The output type doesn't match the data type assigned to the tensor (%s).", + EnumNameDType(getDtype())); + return -2; + } + return 0; +} + int TosaReference::Tensor::writeToVector(std::vector& vals) { uint32_t elements = getElementCount(); diff --git a/reference_model/src/tensor.h b/reference_model/src/tensor.h index efd7e62..4efbf84 100644 --- a/reference_model/src/tensor.h +++ b/reference_model/src/tensor.h @@ -229,11 +229,13 @@ public: virtual int copyValueFrom(Tensor* tensor) = 0; virtual int readfromVector(const std::vector& vals); + virtual int readfromVector(const std::vector& vals); virtual int readfromVector(const std::vector& vals); virtual int readfromVector(const std::vector& vals); virtual int readfromVector(const std::vector& vals); virtual int writeToVector(std::vector& vals); + virtual int writeToVector(std::vector& vals); virtual int writeToVector(std::vector& vals); virtual int writeToVector(std::vector& vals); virtual int writeToVector(std::vector& vals); -- cgit v1.2.1