diff options
Diffstat (limited to 'runtimes/neurun/src/backend/acl_cl/StageGenerator.cc')
| -rw-r--r-- | runtimes/neurun/src/backend/acl_cl/StageGenerator.cc | 357 |
1 files changed, 206 insertions, 151 deletions
diff --git a/runtimes/neurun/src/backend/acl_cl/StageGenerator.cc b/runtimes/neurun/src/backend/acl_cl/StageGenerator.cc index c63698fd8..89bbd7bd2 100644 --- a/runtimes/neurun/src/backend/acl_cl/StageGenerator.cc +++ b/runtimes/neurun/src/backend/acl_cl/StageGenerator.cc @@ -16,6 +16,8 @@ #include "backend/acl_cl/StageGenerator.h" +#include "kernel/acl_cl/CLFunction.h" + #include <arm_compute/runtime/CL/functions/CLConvolutionLayer.h> #include <arm_compute/runtime/CL/functions/CLPoolingLayer.h> #include <arm_compute/runtime/CL/functions/CLActivationLayer.h> @@ -25,20 +27,27 @@ #include "kernel/acl_cl/ConcatLayer.h" -#include "internal/Padding.h" +#include "util/Padding.h" -#include "graph/operand/Index.h" +#include "model/operand/Index.h" -#include "logging.h" +#include "util/logging.h" #include "NeuralNetworks.h" -#include "support/nnapi/Utils.h" +#include "util/Utils.h" template <typename T> std::unique_ptr<T> make_layer(void) { return std::unique_ptr<T>{new T}; } -::arm_compute::PadStrideInfo asPadStringInfo(const ::internal::Padding &padding, - const ::internal::Stride &stride) +std::unique_ptr<::neurun::kernel::acl_cl::CLFunction> +make_cl_function(std::unique_ptr<::arm_compute::IFunction> &&layer) +{ + return std::unique_ptr<::neurun::kernel::acl_cl::CLFunction>( + new ::neurun::kernel::acl_cl::CLFunction(std::move(layer))); +} + +::arm_compute::PadStrideInfo asPadStringInfo(const neurun::util::Padding &padding, + const neurun::util::Stride &stride) { return ::arm_compute::PadStrideInfo{stride.horizontal, stride.vertical, @@ -86,7 +95,9 @@ void ActivationBuilder::appendReLU(::arm_compute::ICLTensor *ifm_alloc) fn->configure(ifm_alloc, nullptr, act_info); - _builder.append(std::move(fn)); + auto acl_fn = make_cl_function(std::move(fn)); + + _builder.append(std::move(acl_fn)); } void ActivationBuilder::append(FuseCode code, ::arm_compute::ICLTensor *ifm_alloc) @@ -113,25 +124,27 @@ void ActivationBuilder::append(FuseCode code, ::arm_compute::ICLTensor *ifm_allo // // StageGenerator // -StageGenerator::StageGenerator(const neurun::graph::operand::Set &ctx, +StageGenerator::StageGenerator(const neurun::model::operand::Set &ctx, const std::shared_ptr<TensorBuilder> &tensor_builder) : _ctx(ctx), _tensor_builder(tensor_builder) { // DO NOTHING } -Stage StageGenerator::generate(const graph::operation::Conv2D::Implicit::Node &node) +void StageGenerator::visit(const model::operation::Conv2DNode &node) { - const ::neurun::graph::operand::Index ofm_index{node.getOutputs().at(0)}; - const ::neurun::graph::operand::Index ifm_index{node.getInputs().at(0)}; - const ::neurun::graph::operand::Index ker_index{node.getInputs().at(1)}; - const ::neurun::graph::operand::Index bias_index{node.getInputs().at(2)}; + using model::operation::Conv2DNode; + + const auto ofm_index{node.getOutputs().at(0)}; + const auto ifm_index{node.getInputs().at(Conv2DNode::Input::INPUT)}; + const auto ker_index{node.getInputs().at(Conv2DNode::Input::KERNEL)}; + const auto bias_index{node.getInputs().at(Conv2DNode::Input::BIAS)}; - const ::neurun::graph::operand::Index vstride_index{node.param().vstride_index}; - const ::neurun::graph::operand::Index hstride_index{node.param().hstride_index}; + const auto vstride_index{node.param().vstride_index}; + const auto hstride_index{node.param().hstride_index}; - const ::neurun::graph::operand::Index padding_index{node.param().padding_index}; - const ::neurun::graph::operand::Index activation_index{node.param().activation_index}; + const auto padding_index{node.param().padding_index}; + const auto activation_index{node.param().activation_index}; const auto ofm_shape = _ctx.at(ofm_index).shape().asFeature(); const auto ifm_shape = _ctx.at(ifm_index).shape().asFeature(); @@ -143,7 +156,7 @@ Stage StageGenerator::generate(const graph::operation::Conv2D::Implicit::Node &n assert((ANEURALNETWORKS_PADDING_SAME == padding_type) || (ANEURALNETWORKS_PADDING_VALID == padding_type)); - ::internal::Stride stride; + neurun::util::Stride stride; stride.vertical = _ctx.at(vstride_index).asScalar<int32_t>(); stride.horizontal = _ctx.at(hstride_index).asScalar<int32_t>(); @@ -151,64 +164,67 @@ Stage StageGenerator::generate(const graph::operation::Conv2D::Implicit::Node &n // Construct operation parameters struct Param { - int ofm_index; - int ifm_index; - int ker_index; - int bias_index; + model::operand::Index ofm_index; + model::operand::Index ifm_index; + model::operand::Index ker_index; + model::operand::Index bias_index; - ::internal::Padding padding; - ::internal::Stride stride; + neurun::util::Padding padding; + neurun::util::Stride stride; FuseCode activation; }; Param param; - param.ofm_index = ofm_index.asInt(); - param.ifm_index = ifm_index.asInt(); - param.ker_index = ker_index.asInt(); - param.bias_index = bias_index.asInt(); + param.ofm_index = ofm_index; + param.ifm_index = ifm_index; + param.ker_index = ker_index; + param.bias_index = bias_index; param.stride = stride; param.padding = (padding_type == ANEURALNETWORKS_PADDING_SAME) - ? ::internal::same_padding(ifm_shape, ofm_shape, stride, ker_shape.W, ker_shape.H) - : ::internal::valid_padding(); + ? neurun::util::same_padding(ifm_shape, ofm_shape, stride, ker_shape.W, ker_shape.H) + : neurun::util::valid_padding(); param.activation = static_cast<FuseCode>(_ctx.at(activation_index).asScalar<int32_t>()); auto tensors = _tensor_builder; - return [tensors, param](IExecutionBuilder &builder) { - auto ofm_alloc = tensors->at(::neurun::graph::operand::Index{param.ofm_index}).get(); - auto ifm_alloc = tensors->at(::neurun::graph::operand::Index{param.ifm_index}).get(); - auto ker_alloc = tensors->at(::neurun::graph::operand::Index{param.ker_index}).get(); - auto bias_alloc = tensors->at(::neurun::graph::operand::Index{param.bias_index}).get(); + returnStage([tensors, param](IExecutionBuilder &builder) { + auto ofm_alloc = tensors->at(param.ofm_index).get(); + auto ifm_alloc = tensors->at(param.ifm_index).get(); + auto ker_alloc = tensors->at(param.ker_index).get(); + auto bias_alloc = tensors->at(param.bias_index).get(); const auto conv_info = asPadStringInfo(param.padding, param.stride); std::unique_ptr<::arm_compute::CLConvolutionLayer> fn{new ::arm_compute::CLConvolutionLayer}; - fn->configure(ifm_alloc, ker_alloc, bias_alloc, ofm_alloc, conv_info); + fn->configure(ifm_alloc->handle(), ker_alloc->handle(), bias_alloc->handle(), + ofm_alloc->handle(), conv_info); - builder.append(std::move(fn)); + auto acl_fn = make_cl_function(std::move(fn)); - ActivationBuilder{builder}.append(param.activation, ofm_alloc); - }; + builder.append(std::move(acl_fn)); + + ActivationBuilder{builder}.append(param.activation, ofm_alloc->handle()); + }); } -Stage StageGenerator::generate(const graph::operation::MaxPool2D::Implicit::Node &node) +void StageGenerator::visit(const model::operation::MaxPool2DNode &node) { - const ::neurun::graph::operand::Index ofm_index{node.getOutputs().at(0)}; - const ::neurun::graph::operand::Index ifm_index{node.getInputs().at(0)}; + const auto ofm_index{node.getOutputs().at(0)}; + const auto ifm_index{node.getInputs().at(model::operation::MaxPool2DNode::Input::INPUT)}; - const ::neurun::graph::operand::Index kh_index{node.param().kh_index}; - const ::neurun::graph::operand::Index kw_index{node.param().kw_index}; + const auto kh_index{node.param().kh_index}; + const auto kw_index{node.param().kw_index}; - const ::neurun::graph::operand::Index vstride_index{node.param().vstride_index}; - const ::neurun::graph::operand::Index hstride_index{node.param().hstride_index}; + const auto vstride_index{node.param().vstride_index}; + const auto hstride_index{node.param().hstride_index}; - const ::neurun::graph::operand::Index padding_index{node.param().padding_index}; + const auto padding_index{node.param().padding_index}; const auto ofm_shape = _ctx.at(ofm_index).shape().asFeature(); const auto ifm_shape = _ctx.at(ifm_index).shape().asFeature(); @@ -225,22 +241,22 @@ Stage StageGenerator::generate(const graph::operation::MaxPool2D::Implicit::Node // Construct operation parameters struct Param { - int ofm_index; - int ifm_index; + model::operand::Index ofm_index; + model::operand::Index ifm_index; uint32_t kw; uint32_t kh; - ::internal::Padding padding; - ::internal::Stride stride; + neurun::util::Padding padding; + neurun::util::Stride stride; // TODO Add 'activation' field }; Param param; - param.ofm_index = ofm_index.asInt(); - param.ifm_index = ifm_index.asInt(); + param.ofm_index = ofm_index; + param.ifm_index = ifm_index; param.kh = kh; param.kw = kw; @@ -249,8 +265,8 @@ Stage StageGenerator::generate(const graph::operation::MaxPool2D::Implicit::Node param.stride.horizontal = hstride; param.padding = (padding_type == ANEURALNETWORKS_PADDING_SAME) - ? ::internal::same_padding(ifm_shape, ofm_shape, param.stride, kw, kh) - : ::internal::valid_padding(); + ? neurun::util::same_padding(ifm_shape, ofm_shape, param.stride, kw, kh) + : neurun::util::valid_padding(); VERBOSE(MaxPool2D) << "IFM_H: " << ifm_shape.H << std::endl; VERBOSE(MaxPool2D) << "IFM_W: " << ifm_shape.W << std::endl; @@ -267,9 +283,9 @@ Stage StageGenerator::generate(const graph::operation::MaxPool2D::Implicit::Node auto tensors = _tensor_builder; - return [tensors, param](IExecutionBuilder &builder) { - auto ofm_alloc = tensors->at(::neurun::graph::operand::Index{param.ofm_index}).get(); - auto ifm_alloc = tensors->at(::neurun::graph::operand::Index{param.ifm_index}).get(); + returnStage([tensors, param](IExecutionBuilder &builder) { + auto ofm_alloc = tensors->at(param.ofm_index).get(); + auto ifm_alloc = tensors->at(param.ifm_index).get(); ::arm_compute::PoolingLayerInfo info{::arm_compute::PoolingType::MAX, ::arm_compute::Size2D{param.kw, param.kh}, @@ -277,24 +293,26 @@ Stage StageGenerator::generate(const graph::operation::MaxPool2D::Implicit::Node std::unique_ptr<::arm_compute::CLPoolingLayer> fn{new ::arm_compute::CLPoolingLayer}; - fn->configure(ifm_alloc, ofm_alloc, info); + fn->configure(ifm_alloc->handle(), ofm_alloc->handle(), info); - builder.append(std::move(fn)); - }; + auto acl_fn = make_cl_function(std::move(fn)); + + builder.append((std::move(acl_fn))); + }); } -Stage StageGenerator::generate(const graph::operation::AvgPool2D::Implicit::Node &node) +void StageGenerator::visit(const model::operation::AvgPool2DNode &node) { - const ::neurun::graph::operand::Index ofm_index{node.getOutputs().at(0)}; - const ::neurun::graph::operand::Index ifm_index{node.getInputs().at(0)}; + const auto ofm_index{node.getOutputs().at(0)}; + const auto ifm_index{node.getInputs().at(model::operation::AvgPool2DNode::Input::INPUT)}; - const ::neurun::graph::operand::Index kh_index{node.param().kh_index}; - const ::neurun::graph::operand::Index kw_index{node.param().kw_index}; + const auto kh_index{node.param().kh_index}; + const auto kw_index{node.param().kw_index}; - const ::neurun::graph::operand::Index vstride_index{node.param().vstride_index}; - const ::neurun::graph::operand::Index hstride_index{node.param().hstride_index}; + const auto vstride_index{node.param().vstride_index}; + const auto hstride_index{node.param().hstride_index}; - const ::neurun::graph::operand::Index padding_index{node.param().padding_index}; + const auto padding_index{node.param().padding_index}; const auto ofm_shape = _ctx.at(ofm_index).shape().asFeature(); const auto ifm_shape = _ctx.at(ifm_index).shape().asFeature(); @@ -314,22 +332,22 @@ Stage StageGenerator::generate(const graph::operation::AvgPool2D::Implicit::Node // Construct operation parameters struct Param { - int ofm_index; - int ifm_index; + model::operand::Index ofm_index; + model::operand::Index ifm_index; uint32_t kw; uint32_t kh; - ::internal::Padding padding; - ::internal::Stride stride; + neurun::util::Padding padding; + neurun::util::Stride stride; // TODO Add 'activation' field }; Param param; - param.ofm_index = ofm_index.asInt(); - param.ifm_index = ifm_index.asInt(); + param.ofm_index = ofm_index; + param.ifm_index = ifm_index; param.kh = kh; param.kw = kw; @@ -338,8 +356,8 @@ Stage StageGenerator::generate(const graph::operation::AvgPool2D::Implicit::Node param.stride.horizontal = hstride; param.padding = (padding_type == ANEURALNETWORKS_PADDING_SAME) - ? ::internal::same_padding(ifm_shape, ofm_shape, param.stride, kw, kh) - : ::internal::valid_padding(); + ? neurun::util::same_padding(ifm_shape, ofm_shape, param.stride, kw, kh) + : neurun::util::valid_padding(); VERBOSE(AvgPool2D) << "IFM_H: " << ifm_shape.H << std::endl; VERBOSE(AvgPool2D) << "IFM_W: " << ifm_shape.W << std::endl; @@ -349,7 +367,7 @@ Stage StageGenerator::generate(const graph::operation::AvgPool2D::Implicit::Node VERBOSE(AvgPool2D) << "KER_W: " << kw << std::endl; VERBOSE(AvgPool2D) << "STRIDE_H: " << vstride << std::endl; VERBOSE(AvgPool2D) << "STRIDE_W: " << hstride << std::endl; - VERBOSE(AvgPool2D) << "PAD: " << ::nnfw::support::nnapi::to_string(padding_type) << std::endl; + VERBOSE(AvgPool2D) << "PAD: " << neurun::util::to_string(padding_type) << std::endl; VERBOSE(AvgPool2D) << "PAD(T): " << param.padding.top << std::endl; VERBOSE(AvgPool2D) << "PAD(B): " << param.padding.bottom << std::endl; VERBOSE(AvgPool2D) << "PAD(L): " << param.padding.left << std::endl; @@ -357,9 +375,9 @@ Stage StageGenerator::generate(const graph::operation::AvgPool2D::Implicit::Node auto tensors = _tensor_builder; - return [tensors, param](IExecutionBuilder &builder) { - auto ofm_alloc = tensors->at(::neurun::graph::operand::Index{param.ofm_index}).get(); - auto ifm_alloc = tensors->at(::neurun::graph::operand::Index{param.ifm_index}).get(); + returnStage([tensors, param](IExecutionBuilder &builder) { + auto ofm_alloc = tensors->at(param.ofm_index).get(); + auto ifm_alloc = tensors->at(param.ifm_index).get(); ::arm_compute::PoolingLayerInfo info{ ::arm_compute::PoolingType::AVG, ::arm_compute::Size2D{param.kw, param.kh}, @@ -367,170 +385,207 @@ Stage StageGenerator::generate(const graph::operation::AvgPool2D::Implicit::Node std::unique_ptr<::arm_compute::CLPoolingLayer> fn{new ::arm_compute::CLPoolingLayer}; - fn->configure(ifm_alloc, ofm_alloc, info); + fn->configure(ifm_alloc->handle(), ofm_alloc->handle(), info); - builder.append(std::move(fn)); - }; + auto acl_fn = make_cl_function(std::move(fn)); + + builder.append((std::move(acl_fn))); + }); } -Stage StageGenerator::generate(const graph::operation::Concat::Node &node) +void StageGenerator::visit(const model::operation::ConcatNode &node) { - const ::neurun::graph::operand::Index ofm_index{node.getOutputs().at(0)}; - const ::neurun::graph::operand::Index axis_index{node.param().axis_index}; + const auto ofm_index{node.getOutputs().at(0)}; + const auto axis_index{node.param().axis_index}; struct Param { - int32_t output_index; - std::vector<int32_t> input_indexes; + model::operand::Index output_index; + std::vector<model::operand::Index> input_indexes; int32_t axis; }; Param param; - param.output_index = ofm_index.asInt(); + param.output_index = ofm_index; for (const auto &e : node.getInputs()) { - param.input_indexes.emplace_back(e.asInt()); + param.input_indexes.emplace_back(e); } param.axis = _ctx.at(axis_index).asScalar<int32_t>(); auto tensors = _tensor_builder; - return [tensors, param](IExecutionBuilder &builder) { - auto output_alloc = tensors->at(::neurun::graph::operand::Index{param.output_index}).get(); + returnStage([tensors, param](IExecutionBuilder &builder) { + // If tensor allocator allocate as subtensor + bool canEliminate = true; + for (auto ifm_ind : param.input_indexes) + { + if (!tensors->isSubTensorOf(param.output_index, ifm_ind)) + { + canEliminate = false; + break; + } + } + if (canEliminate) + { + // If concat eliminated, return with nothing to do + return; + } + + auto output_alloc = tensors->at(param.output_index).get(); - std::vector<::arm_compute::ICLTensor *> input_allocs; + std::vector<::neurun::backend::acl_cl::operand::ICLTensor *> input_allocs; for (auto ifm_ind : param.input_indexes) { - input_allocs.emplace_back(tensors->at(::neurun::graph::operand::Index{ifm_ind}).get()); + input_allocs.emplace_back( + dynamic_cast<::neurun::backend::acl_cl::operand::CLTensor *>(tensors->at(ifm_ind).get())); } std::unique_ptr<::neurun::kernel::acl_cl::ConcatLayer> fn{ new ::neurun::kernel::acl_cl::ConcatLayer}; - fn->configure(input_allocs, param.axis, output_alloc); + fn->configure(input_allocs, param.axis, + dynamic_cast<::neurun::backend::acl_cl::operand::CLTensor *>(output_alloc)); - builder.append(std::move(fn)); - }; + auto acl_fn = make_cl_function(std::move(fn)); + + builder.append(std::move(acl_fn)); + }); } -Stage StageGenerator::generate(const graph::operation::FullyConnected::Node &node) +void StageGenerator::visit(const model::operation::FullyConnectedNode &node) { - const ::neurun::graph::operand::Index output_index{node.getOutputs().at(0)}; - const ::neurun::graph::operand::Index input_index{node.getInputs().at(0)}; - const ::neurun::graph::operand::Index weight_index{node.getInputs().at(1)}; - const ::neurun::graph::operand::Index bias_index{node.getInputs().at(2)}; - const ::neurun::graph::operand::Index activation_index{node.param().activation_index}; + using model::operation::FullyConnectedNode; + + const auto output_index{node.getOutputs().at(0)}; + const auto input_index{node.getInputs().at(FullyConnectedNode::Input::INPUT)}; + const auto weight_index{node.getInputs().at(FullyConnectedNode::Input::WEIGHT)}; + const auto bias_index{node.getInputs().at(FullyConnectedNode::Input::BIAS)}; + const auto activation_index{node.param().activation_index}; // Construct operation parameters struct Param { - int output_index; + model::operand::Index output_index; - int input_index; - int weight_index; - int bias_index; + model::operand::Index input_index; + model::operand::Index weight_index; + model::operand::Index bias_index; FuseCode activation; }; Param param; - param.output_index = output_index.asInt(); - param.input_index = input_index.asInt(); - param.weight_index = weight_index.asInt(); - param.bias_index = bias_index.asInt(); + param.output_index = output_index; + param.input_index = input_index; + param.weight_index = weight_index; + param.bias_index = bias_index; param.activation = static_cast<FuseCode>(_ctx.at(activation_index).asScalar<int32_t>()); auto tensors = _tensor_builder; - return [tensors, param](IExecutionBuilder &builder) { - auto output_alloc = tensors->at(::neurun::graph::operand::Index{param.output_index}).get(); - auto input_alloc = tensors->at(::neurun::graph::operand::Index{param.input_index}).get(); - auto weight_alloc = tensors->at(::neurun::graph::operand::Index{param.weight_index}).get(); - auto bias_alloc = tensors->at(::neurun::graph::operand::Index{param.bias_index}).get(); + returnStage([tensors, param](IExecutionBuilder &builder) { + auto output_alloc = tensors->at(param.output_index).get(); + auto input_alloc = tensors->at(param.input_index).get(); + auto weight_alloc = tensors->at(param.weight_index).get(); + auto bias_alloc = tensors->at(param.bias_index).get(); auto fn = make_layer<::arm_compute::CLFullyConnectedLayer>(); - fn->configure(input_alloc, weight_alloc, bias_alloc, output_alloc); + fn->configure(input_alloc->handle(), weight_alloc->handle(), bias_alloc->handle(), + output_alloc->handle()); - builder.append(std::move(fn)); + auto acl_fn = make_cl_function(std::move(fn)); - ActivationBuilder{builder}.append(param.activation, output_alloc); - }; + builder.append(std::move(acl_fn)); + + ActivationBuilder{builder}.append(param.activation, output_alloc->handle()); + }); } -Stage StageGenerator::generate(const graph::operation::Reshape::Node &node) +void StageGenerator::visit(const model::operation::ReshapeNode &node) { - const ::neurun::graph::operand::Index output_index{node.getOutputs().at(0)}; - const ::neurun::graph::operand::Index input_index{node.getInputs().at(0)}; + const auto output_index{node.getOutputs().at(0)}; + const auto input_index{node.getInputs().at(model::operation::ReshapeNode::Input::INPUT)}; struct Param { - int output_index; - int input_index; + model::operand::Index output_index; + model::operand::Index input_index; }; Param param; - param.output_index = output_index.asInt(); - param.input_index = input_index.asInt(); + param.output_index = output_index; + param.input_index = input_index; auto tensors = _tensor_builder; - return [tensors, param](IExecutionBuilder &builder) { - auto output_alloc = tensors->at(::neurun::graph::operand::Index{param.output_index}).get(); - auto input_alloc = tensors->at(::neurun::graph::operand::Index{param.input_index}).get(); + returnStage([tensors, param](IExecutionBuilder &builder) { + auto output_alloc = tensors->at(param.output_index).get(); + auto input_alloc = tensors->at(param.input_index).get(); auto fn = make_layer<::arm_compute::CLReshapeLayer>(); - fn->configure(input_alloc, output_alloc); + fn->configure(input_alloc->handle(), output_alloc->handle()); - builder.append(std::move(fn)); - }; + auto acl_fn = make_cl_function(std::move(fn)); + + builder.append(std::move(acl_fn)); + }); } -Stage StageGenerator::generate(const graph::operation::Softmax::Node &node) +void StageGenerator::visit(const model::operation::SoftmaxNode &node) { - const ::neurun::graph::operand::Index output_index{node.getOutputs().at(0)}; - const ::neurun::graph::operand::Index input_index{node.getInputs().at(0)}; - const ::neurun::graph::operand::Index scale_index{node.param().scale_index}; + const auto output_index{node.getOutputs().at(0)}; + const auto input_index{node.getInputs().at(model::operation::SoftmaxNode::Input::INPUT)}; + const auto scale_index{node.param().scale_index}; assert(_ctx.at(scale_index).shape().rank() == 0); struct Param { - int output_index; - int input_index; + model::operand::Index output_index; + model::operand::Index input_index; float scale; }; Param param; - param.output_index = output_index.asInt(); - param.input_index = input_index.asInt(); + param.output_index = output_index; + param.input_index = input_index; param.scale = _ctx.at(scale_index).asScalar<float>(); auto tensors = _tensor_builder; - return [tensors, param](IExecutionBuilder &builder) { - auto output_alloc = tensors->at(::neurun::graph::operand::Index{param.output_index}).get(); - auto input_alloc = tensors->at(::neurun::graph::operand::Index{param.input_index}).get(); + returnStage([tensors, param](IExecutionBuilder &builder) { + auto output_alloc = tensors->at(param.output_index).get(); + auto input_alloc = tensors->at(param.input_index).get(); auto fn = make_layer<::arm_compute::CLSoftmaxLayer>(); - fn->configure(input_alloc, output_alloc, param.scale); + fn->configure(input_alloc->handle(), output_alloc->handle(), param.scale); - builder.append(std::move(fn)); - }; + auto acl_fn = make_cl_function(std::move(fn)); + + builder.append(std::move(acl_fn)); + }); } -Stage StageGenerator::generate(const graph::operation::NOP::Node & /* node */) +void StageGenerator::visit(const model::operation::PermuteNode & /* node */) { - // DO NOTHING - return nullptr; + throw "Unsupported"; +} + +void StageGenerator::visit(const model::operation::AddNode &) +{ + VERBOSE(Add) << "generate CPU Add" << std::endl; + + throw std::runtime_error("NYI"); } } // namespace acl_cl |