aboutsummaryrefslogtreecommitdiff
path: root/src/armnn/test/NetworkTests.cpp
diff options
context:
space:
mode:
Diffstat (limited to 'src/armnn/test/NetworkTests.cpp')
-rw-r--r--src/armnn/test/NetworkTests.cpp640
1 files changed, 4 insertions, 636 deletions
diff --git a/src/armnn/test/NetworkTests.cpp b/src/armnn/test/NetworkTests.cpp
index 4f8dd7ea7b..91ff7fa983 100644
--- a/src/armnn/test/NetworkTests.cpp
+++ b/src/armnn/test/NetworkTests.cpp
@@ -2,16 +2,13 @@
// Copyright © 2017 Arm Ltd. All rights reserved.
// SPDX-License-Identifier: MIT
//
-#include <boost/test/unit_test.hpp>
+
+#include "GraphUtils.hpp"
#include <armnn/ArmNN.hpp>
-#include <Network.hpp>
-#include <Graph.hpp>
-#include <backends/reference/RefWorkloadFactory.hpp>
-#include <backends/neon/NeonWorkloadFactory.hpp>
-#include <backends/cl/ClWorkloadFactory.hpp>
+#include <armnn/Network.hpp>
-#include "GraphUtils.hpp"
+#include <boost/test/unit_test.hpp>
namespace
{
@@ -43,54 +40,6 @@ BOOST_AUTO_TEST_CASE(LayerGuids)
BOOST_TEST(inputId != outputId);
}
-BOOST_AUTO_TEST_CASE(SerializeToDot)
-{
- armnn::Network net;
-
- //Defines layers.
- auto input = net.AddInputLayer(0);
- auto add = net.AddAdditionLayer();
- auto output = net.AddOutputLayer(0);
-
- // Connects layers.
- input->GetOutputSlot(0).Connect(add->GetInputSlot(0));
- input->GetOutputSlot(0).Connect(add->GetInputSlot(1));
- add->GetOutputSlot(0).Connect(output->GetInputSlot(0));
-
- armnn::TensorShape shape({4});
- armnn::TensorInfo info(shape, armnn::DataType::Float32);
- input->GetOutputSlot(0).SetTensorInfo(info);
- add->GetOutputSlot(0).SetTensorInfo(info);
-
- armnn::IRuntime::CreationOptions options;
- armnn::IRuntimePtr runtime(armnn::IRuntime::Create(options));
-
- std::vector<armnn::BackendId> backends = {armnn::Compute::CpuRef};
- armnn::IOptimizedNetworkPtr optimizedNet = armnn::Optimize(net, backends, runtime->GetDeviceSpec());
-
- std::ostringstream ss;
- optimizedNet->SerializeToDot(ss);
-
- auto inputId = input->GetGuid();
- auto addId = add->GetGuid();
- auto outputId = output->GetGuid();
-
- std::stringstream expected;
- expected <<
- "digraph Optimized {\n"
- " node [shape=\"record\"];\n"
- " edge [fontsize=8 fontcolor=\"blue\" fontname=\"arial-bold\"];\n"
- " " << inputId << " [label=\"{Input}\"];\n"
- " " << addId << " [label=\"{Addition}\"];\n"
- " " << outputId << " [label=\"{Output}\"];\n"
- " " << inputId << " -> " << addId << " [label=< [4] >];\n"
- " " << inputId << " -> " << addId << " [label=< [4] >];\n"
- " " << addId << " -> " << outputId << " [label=< [4] >];\n"
- "}\n";
-
- BOOST_TEST(ss.str() == expected.str());
-}
-
BOOST_AUTO_TEST_CASE(NetworkBasic)
{
armnn::Network net;
@@ -417,585 +366,4 @@ BOOST_AUTO_TEST_CASE(NetworkModification_SplitterMultiplication)
prevLayer->GetOutputSlot(0).Connect(layer->GetInputSlot(0));
}
-BOOST_AUTO_TEST_CASE(OptimizeValidateCpuRefWorkloads)
-{
- const armnn::TensorInfo desc({3, 5}, armnn::DataType::Float32);
-
- armnn::Network net;
-
- armnn::NormalizationDescriptor nmDesc;
- armnn::ActivationDescriptor acDesc;
-
- // in
- // |
- // nm
- // / |
- // ac |
- // \ |
- // ml
- // |
- // sm
- // |
- // ot
- armnn::IConnectableLayer* layer = net.AddInputLayer(0, "in");
- layer->GetOutputSlot(0).SetTensorInfo(desc);
-
- armnn::IConnectableLayer* const normLayer = net.AddNormalizationLayer(nmDesc, "nm");
-
- layer->GetOutputSlot(0).Connect(normLayer->GetInputSlot(0));
- normLayer->GetOutputSlot(0).SetTensorInfo(desc);
-
- layer = net.AddActivationLayer(acDesc, "ac");
-
- normLayer->GetOutputSlot(0).Connect(layer->GetInputSlot(0));
- layer->GetOutputSlot(0).SetTensorInfo(desc);
-
- armnn::IConnectableLayer* prevLayer = layer;
- layer = net.AddMultiplicationLayer("ml");
-
- prevLayer->GetOutputSlot(0).Connect(layer->GetInputSlot(0));
- normLayer->GetOutputSlot(0).Connect(layer->GetInputSlot(1));
- layer->GetOutputSlot(0).SetTensorInfo(desc);
-
- prevLayer = layer;
- armnn::SoftmaxDescriptor softmaxDescriptor;
- layer = net.AddSoftmaxLayer(softmaxDescriptor, "sm");
-
- prevLayer->GetOutputSlot(0).Connect(layer->GetInputSlot(0));
- layer->GetOutputSlot(0).SetTensorInfo(desc);
-
- prevLayer = layer;
- layer = net.AddOutputLayer(0, "ot");
-
- prevLayer->GetOutputSlot(0).Connect(layer->GetInputSlot(0));
-
- armnn::IRuntime::CreationOptions options;
- armnn::IRuntimePtr runtime(armnn::IRuntime::Create(options));
-
- std::vector<armnn::BackendId> backends = { armnn::Compute::CpuRef };
- armnn::IOptimizedNetworkPtr optNet = armnn::Optimize(net, backends, runtime->GetDeviceSpec());
- static_cast<armnn::OptimizedNetwork*>(optNet.get())->GetGraph().AllocateDynamicBuffers();
- BOOST_CHECK(optNet);
-
- // Validates workloads.
- armnn::RefWorkloadFactory fact;
- for (auto&& layer : static_cast<armnn::OptimizedNetwork*>(optNet.get())->GetGraph())
- {
- BOOST_CHECK_NO_THROW(
- layer->CreateWorkload(static_cast<armnn::OptimizedNetwork*>(optNet.get())->GetGraph(), fact));
- }
-}
-
-#if ARMCOMPUTENEON_ENABLED
-BOOST_AUTO_TEST_CASE(OptimizeValidateCpuAccDeviceSupportLayerNoFallback)
-{
- // build up the structure of the network
- armnn::INetworkPtr net(armnn::INetwork::Create());
-
- armnn::IConnectableLayer* input = net->AddInputLayer(0);
-
- armnn::IConnectableLayer* output = net->AddOutputLayer(0);
-
- input->GetOutputSlot(0).Connect(output->GetInputSlot(0));
-
- input->GetOutputSlot(0).SetTensorInfo(armnn::TensorInfo({ 1, 1, 4, 4 }, armnn::DataType::Float32));
-
- armnn::IRuntime::CreationOptions options;
- armnn::IRuntimePtr runtime(armnn::IRuntime::Create(options));
-
- std::vector<armnn::BackendId> backends = { armnn::Compute::CpuAcc };
- armnn::IOptimizedNetworkPtr optNet = armnn::Optimize(*net, backends, runtime->GetDeviceSpec());
- BOOST_CHECK(optNet);
- // validate workloads
- armnn::NeonWorkloadFactory fact;
- for (auto&& layer : static_cast<armnn::OptimizedNetwork*>(optNet.get())->GetGraph())
- {
- BOOST_CHECK(layer->GetBackendId() == armnn::Compute::CpuAcc);
- BOOST_CHECK_NO_THROW(
- layer->CreateWorkload(static_cast<armnn::OptimizedNetwork*>(optNet.get())->GetGraph(), fact));
- }
-}
-#endif // ARMCOMPUTENEON_ENABLED
-
-#if ARMCOMPUTECL_ENABLED
-BOOST_AUTO_TEST_CASE(OptimizeValidateGpuDeviceSupportLayerNoFallback)
-{
- // build up the structure of the network
- armnn::INetworkPtr net(armnn::INetwork::Create());
-
- armnn::IConnectableLayer* input = net->AddInputLayer(0);
-
- armnn::IConnectableLayer* output = net->AddOutputLayer(0);
-
- input->GetOutputSlot(0).Connect(output->GetInputSlot(0));
-
- input->GetOutputSlot(0).SetTensorInfo(armnn::TensorInfo({ 1, 1, 4, 4 }, armnn::DataType::Float32));
-
- armnn::IRuntime::CreationOptions options;
- armnn::IRuntimePtr runtime(armnn::IRuntime::Create(options));
-
- std::vector<armnn::BackendId> backends = { armnn::Compute::GpuAcc };
- armnn::IOptimizedNetworkPtr optNet = armnn::Optimize(*net, backends, runtime->GetDeviceSpec());
- BOOST_CHECK(optNet);
- // validate workloads
- armnn::ClWorkloadFactory fact;
- for (auto&& layer : static_cast<armnn::OptimizedNetwork*>(optNet.get())->GetGraph())
- {
- BOOST_CHECK(layer->GetBackendId() == armnn::Compute::GpuAcc);
- BOOST_CHECK_NO_THROW(
- layer->CreateWorkload(static_cast<armnn::OptimizedNetwork*>(optNet.get())->GetGraph(), fact));
- }
-}
-#endif // ARMCOMPUTECL_ENABLED
-
-BOOST_AUTO_TEST_CASE(OptimizeValidateDeviceNonSupportLayerNoFallback)
-{
- // build up the structure of the network
- armnn::INetworkPtr net(armnn::INetwork::Create());
-
- armnn::IConnectableLayer* input = net->AddInputLayer(0);
-
- // This layer configuration isn't supported by CpuAcc and isn't allowed to fall back, so Optimize will return null.
- armnn::NormalizationDescriptor descriptor;
- armnn::IConnectableLayer* normalize = net->AddNormalizationLayer(descriptor);
-
- armnn::IConnectableLayer* output = net->AddOutputLayer(0);
-
- input->GetOutputSlot(0).Connect(normalize->GetInputSlot(0));
- normalize->GetOutputSlot(0).Connect(output->GetInputSlot(0));
-
- input->GetOutputSlot(0).SetTensorInfo(armnn::TensorInfo({ 1, 1, 4, 4 }, armnn::DataType::Float32));
- normalize->GetOutputSlot(0).SetTensorInfo(armnn::TensorInfo({ 1, 1, 4, 4 }, armnn::DataType::Float32));
-
- armnn::IRuntime::CreationOptions options;
- armnn::IRuntimePtr runtime(armnn::IRuntime::Create(options));
-
- std::vector<armnn::BackendId> backends = { armnn::Compute::CpuAcc };
- armnn::IOptimizedNetworkPtr optNet = armnn::Optimize(*net, backends, runtime->GetDeviceSpec());
- BOOST_CHECK(!optNet);
-}
-
-BOOST_AUTO_TEST_CASE(OptimizeValidateDeviceNonSupportLayerWithFallback)
-{
- // build up the structure of the network
- armnn::INetworkPtr net(armnn::INetwork::Create());
-
- armnn::IConnectableLayer* input = net->AddInputLayer(0);
-
- // This layer configuration isn't supported by CpuAcc but it allows to fallback to CpuRef.
- armnn::NormalizationDescriptor descriptor;
- armnn::IConnectableLayer* normalize = net->AddNormalizationLayer(descriptor);
-
- armnn::IConnectableLayer* output = net->AddOutputLayer(0);
-
- input->GetOutputSlot(0).Connect(normalize->GetInputSlot(0));
- normalize->GetOutputSlot(0).Connect(output->GetInputSlot(0));
-
- input->GetOutputSlot(0).SetTensorInfo(armnn::TensorInfo({ 1, 1, 4, 4 }, armnn::DataType::Float32));
- normalize->GetOutputSlot(0).SetTensorInfo(armnn::TensorInfo({ 1, 1, 4, 4 }, armnn::DataType::Float32));
-
- armnn::IRuntime::CreationOptions options;
- armnn::IRuntimePtr runtime(armnn::IRuntime::Create(options));
-
- std::vector<armnn::BackendId> backends = { armnn::Compute::CpuAcc, armnn::Compute::CpuRef };
- armnn::IOptimizedNetworkPtr optNet = armnn::Optimize(*net, backends, runtime->GetDeviceSpec());
- BOOST_REQUIRE(optNet);
-
- for (auto&& layer : static_cast<armnn::OptimizedNetwork*>(optNet.get())->GetGraph())
- {
- // If NEON is enabled, Input and Output layers are supported by CpuAcc,
- // the other layers are supported by CpuRef.
- // If NEON is not enabled, all layers are supported by CpuRef.
-#if ARMCOMPUTENEON_ENABLED
- if (layer->GetType() == armnn::LayerType::Input || layer->GetType() == armnn::LayerType::Output)
- {
- BOOST_CHECK(layer->GetBackendId() == armnn::Compute::CpuAcc);
- }
- else if (layer->GetType() == armnn::LayerType::Normalization)
- {
- BOOST_CHECK(layer->GetBackendId() == armnn::Compute::CpuRef);
- }
-#else
- BOOST_CHECK(layer->GetBackendId() == armnn::Compute::CpuRef);
-#endif
- }
-}
-
-BOOST_AUTO_TEST_CASE(OptimizeValidateWorkloadsUndefinedComputeDevice)
-{
- const armnn::TensorInfo desc({3, 5}, armnn::DataType::Float32);
-
- armnn::Network net;
-
- armnn::NormalizationDescriptor nmDesc;
- armnn::ActivationDescriptor acDesc;
-
- // in
- // |
- // nm
- // / |
- // ac |
- // \ |
- // ml
- // |
- // sm
- // |
- // ot
- armnn::IConnectableLayer* layer = net.AddInputLayer(0, "in");
- layer->GetOutputSlot(0).SetTensorInfo(desc);
-
- armnn::IConnectableLayer* const normLayer = net.AddNormalizationLayer(nmDesc, "nm");
-
- layer->GetOutputSlot(0).Connect(normLayer->GetInputSlot(0));
- normLayer->GetOutputSlot(0).SetTensorInfo(desc);
-
- layer = net.AddActivationLayer(acDesc, "ac");
-
- normLayer->GetOutputSlot(0).Connect(layer->GetInputSlot(0));
- layer->GetOutputSlot(0).SetTensorInfo(desc);
-
- armnn::IConnectableLayer* prevLayer = layer;
- layer = net.AddMultiplicationLayer("ml");
-
- prevLayer->GetOutputSlot(0).Connect(layer->GetInputSlot(0));
- normLayer->GetOutputSlot(0).Connect(layer->GetInputSlot(1));
- layer->GetOutputSlot(0).SetTensorInfo(desc);
-
- prevLayer = layer;
- armnn::SoftmaxDescriptor softmaxDescriptor;
- layer = net.AddSoftmaxLayer(softmaxDescriptor, "sm");
-
- prevLayer->GetOutputSlot(0).Connect(layer->GetInputSlot(0));
- layer->GetOutputSlot(0).SetTensorInfo(desc);
-
- prevLayer = layer;
- layer = net.AddOutputLayer(0, "ot");
-
- prevLayer->GetOutputSlot(0).Connect(layer->GetInputSlot(0));
-
- armnn::IRuntime::CreationOptions options;
- armnn::IRuntimePtr runtime(armnn::IRuntime::Create(options));
-
- std::vector<armnn::BackendId> backends = { armnn::Compute::Undefined };
-
- armnn::IOptimizedNetworkPtr optNet = armnn::Optimize(net, backends, runtime->GetDeviceSpec());
- BOOST_CHECK(!optNet);
-
-}
-
-BOOST_AUTO_TEST_CASE(OptimizeValidateWorkloadsUndefinedComputeDeviceWithFallback)
-{
- const armnn::TensorInfo desc({3, 5}, armnn::DataType::Float32);
-
- armnn::Network net;
-
- armnn::NormalizationDescriptor nmDesc;
- armnn::ActivationDescriptor acDesc;
-
- // in
- // |
- // nm
- // / |
- // ac |
- // \ |
- // ml
- // |
- // sm
- // |
- // ot
- armnn::IConnectableLayer* layer = net.AddInputLayer(0, "in");
- layer->GetOutputSlot(0).SetTensorInfo(desc);
-
- armnn::IConnectableLayer* const normLayer = net.AddNormalizationLayer(nmDesc, "nm");
-
- layer->GetOutputSlot(0).Connect(normLayer->GetInputSlot(0));
- normLayer->GetOutputSlot(0).SetTensorInfo(desc);
-
- layer = net.AddActivationLayer(acDesc, "ac");
-
- normLayer->GetOutputSlot(0).Connect(layer->GetInputSlot(0));
- layer->GetOutputSlot(0).SetTensorInfo(desc);
-
- armnn::IConnectableLayer* prevLayer = layer;
- layer = net.AddMultiplicationLayer("ml");
-
- prevLayer->GetOutputSlot(0).Connect(layer->GetInputSlot(0));
- normLayer->GetOutputSlot(0).Connect(layer->GetInputSlot(1));
- layer->GetOutputSlot(0).SetTensorInfo(desc);
-
- prevLayer = layer;
- armnn::SoftmaxDescriptor softmaxDescriptor;
- layer = net.AddSoftmaxLayer(softmaxDescriptor, "sm");
-
- prevLayer->GetOutputSlot(0).Connect(layer->GetInputSlot(0));
- layer->GetOutputSlot(0).SetTensorInfo(desc);
-
- prevLayer = layer;
- layer = net.AddOutputLayer(0, "ot");
-
- prevLayer->GetOutputSlot(0).Connect(layer->GetInputSlot(0));
-
- armnn::IRuntime::CreationOptions options;
- armnn::IRuntimePtr runtime(armnn::IRuntime::Create(options));
-
- std::vector<armnn::BackendId> backends = { armnn::Compute::Undefined, armnn::Compute::CpuRef };
-
- armnn::IOptimizedNetworkPtr optNet = armnn::Optimize(net, backends, runtime->GetDeviceSpec());
- BOOST_CHECK(optNet);
-
- // validate workloads
- armnn::RefWorkloadFactory fact;
- for (auto&& layer : static_cast<armnn::OptimizedNetwork*>(optNet.get())->GetGraph())
- {
- BOOST_CHECK(layer->GetBackendId() == armnn::Compute::CpuRef);
- BOOST_CHECK_NO_THROW(
- layer->CreateWorkload(static_cast<armnn::OptimizedNetwork*>(optNet.get())->GetGraph(), fact));
- }
-}
-BOOST_AUTO_TEST_CASE(OptimizeValidateWorkloadsDuplicateComputeDeviceWithFallback)
-{
- // build up the structure of the network
- armnn::INetworkPtr net(armnn::INetwork::Create());
-
- armnn::IConnectableLayer* input = net->AddInputLayer(0);
-
- // This layer configuration isn't supported by CpuAcc but it allows to fallback to CpuRef.
- armnn::NormalizationDescriptor descriptor;
- armnn::IConnectableLayer* normalize = net->AddNormalizationLayer(descriptor);
-
- armnn::IConnectableLayer* output = net->AddOutputLayer(0);
-
- input->GetOutputSlot(0).Connect(normalize->GetInputSlot(0));
- normalize->GetOutputSlot(0).Connect(output->GetInputSlot(0));
-
- input->GetOutputSlot(0).SetTensorInfo(armnn::TensorInfo({ 1, 1, 4, 4 }, armnn::DataType::Float32));
- normalize->GetOutputSlot(0).SetTensorInfo(armnn::TensorInfo({ 1, 1, 4, 4 }, armnn::DataType::Float32));
-
- armnn::IRuntime::CreationOptions options;
- armnn::IRuntimePtr runtime(armnn::IRuntime::Create(options));
-
- std::vector<armnn::BackendId> backends = { armnn::Compute::CpuAcc,
- armnn::Compute::GpuAcc,
- armnn::Compute::CpuRef };
-
- armnn::IOptimizedNetworkPtr optNet = armnn::Optimize(*net, backends, runtime->GetDeviceSpec());
- BOOST_REQUIRE(optNet);
-
- for (auto&& layer : static_cast<armnn::OptimizedNetwork*>(optNet.get())->GetGraph())
- {
- // If NEON is enabled, Input and Output layers are supported by CpuAcc,
- // the other layers are supported by CpuRef.
- // If only CL is enabled, Input and Output layers are supported by GpuAcc,
- // the other layers are supported by CpuRef.
- // If neither NEON, nor CL is enabled, all layers are supported by CpuRef.
-#if ARMCOMPUTENEON_ENABLED
- if (layer->GetType() == armnn::LayerType::Input || layer->GetType() == armnn::LayerType::Output)
- {
- BOOST_CHECK(layer->GetBackendId() == armnn::Compute::CpuAcc);
- }
- else if (layer->GetType() == armnn::LayerType::Normalization)
- {
- BOOST_CHECK(layer->GetBackendId() == armnn::Compute::CpuRef);
- }
-#elif ARMCOMPUTECL_ENABLED
- if (layer->GetType() == armnn::LayerType::Input || layer->GetType() == armnn::LayerType::Output)
- {
- BOOST_CHECK(layer->GetBackendId() == armnn::Compute::GpuAcc);
- }
- else if (layer->GetType() == armnn::LayerType::Normalization)
- {
- BOOST_CHECK(layer->GetBackendId() == armnn::Compute::CpuRef);
- }
-#else
- BOOST_CHECK(layer->GetBackendId() == armnn::Compute::CpuRef);
-#endif
- }
-}
-
-BOOST_AUTO_TEST_CASE(OptimizeValidateWorkloadsCpuRefPermuteLayer)
-{
- // Create runtime in which test will run
- armnn::IRuntime::CreationOptions options;
- armnn::IRuntimePtr runtime(armnn::IRuntime::Create(options));
-
- std::vector<armnn::BackendId> backends = {armnn::Compute::CpuRef};
-
- // build up the structure of the network
- armnn::INetworkPtr net(armnn::INetwork::Create());
-
- armnn::IConnectableLayer* input = net->AddInputLayer(0);
-
- armnn::PermuteDescriptor descriptor({0, 2, 3, 1});
- armnn::IConnectableLayer* permute = net->AddPermuteLayer(descriptor);
-
- armnn::IConnectableLayer* output = net->AddOutputLayer(0);
-
- input->GetOutputSlot(0).Connect(permute->GetInputSlot(0));
- permute->GetOutputSlot(0).Connect(output->GetInputSlot(0));
-
- input->GetOutputSlot(0).SetTensorInfo(armnn::TensorInfo({ 1, 1, 4, 4 }, armnn::DataType::Float32));
- permute->GetOutputSlot(0).SetTensorInfo(armnn::TensorInfo({ 1, 4, 1, 4 }, armnn::DataType::Float32));
-
- // optimize the network
- armnn::IOptimizedNetworkPtr optNet = armnn::Optimize(*net, backends, runtime->GetDeviceSpec());
-
- for (auto&& layer : static_cast<armnn::OptimizedNetwork*>(optNet.get())->GetGraph())
- {
- BOOST_CHECK(layer->GetBackendId() == armnn::Compute::CpuRef);
- }
-}
-
-BOOST_AUTO_TEST_CASE(OptimizeValidateWorkloadsCpuRefMeanLayer)
-{
- // Create runtime in which test will run
- armnn::IRuntime::CreationOptions options;
- armnn::IRuntimePtr runtime(armnn::IRuntime::Create(options));
-
- std::vector<armnn::BackendId> backends = {armnn::Compute::CpuRef};
-
- // build up the structure of the network
- armnn::INetworkPtr net(armnn::INetwork::Create());
-
- armnn::IConnectableLayer* input = net->AddInputLayer(0);
-
- armnn::MeanDescriptor descriptor({ 0, 1 }, false);
- armnn::IConnectableLayer* meanLayer = net->AddMeanLayer(descriptor);
-
- armnn::IConnectableLayer* output = net->AddOutputLayer(0);
-
- input->GetOutputSlot(0).Connect(meanLayer->GetInputSlot(0));
- meanLayer->GetOutputSlot(0).Connect(output->GetInputSlot(0));
-
- input->GetOutputSlot(0).SetTensorInfo(armnn::TensorInfo({ 4, 3, 2 }, armnn::DataType::Float32));
- meanLayer->GetOutputSlot(0).SetTensorInfo(armnn::TensorInfo({ 2 }, armnn::DataType::Float32));
-
- // optimize the network
- armnn::IOptimizedNetworkPtr optNet = armnn::Optimize(*net, backends, runtime->GetDeviceSpec());
-
- for (auto&& layer : static_cast<armnn::OptimizedNetwork*>(optNet.get())->GetGraph())
- {
- BOOST_CHECK(layer->GetBackendId() == armnn::Compute::CpuRef);
- }
-}
-
-BOOST_AUTO_TEST_CASE(FP16TurboModeTestOnCpuRef)
-{
- // Test to check when FP16 Turbo mode set
- // it converts the FP32 network to FP16 Network
- // add FP32ToFP16 conversion layer after the InputLayer
- // add FP16ToFP32 conversion layer after the OutputLayer
- // checks the other layers if they are supported in FP16
- // if they are not put the conversion layers before and after
- // if they are not supported in FP16 use FP32 instead
- // if there are inverse conversion layers remove them with optimization
- // at the moment FloorLayer is not supported in FP16 so it rolls back to FP32
- // and inverse conversion layers are removed by the optimizer
- armnn::Network net;
-
- // Defines layers.
- auto input = net.AddInputLayer(0);
- auto floor = net.AddFloorLayer();
- auto output = net.AddOutputLayer(0);
-
- // Connects layers.
- input->GetOutputSlot(0).Connect(floor->GetInputSlot(0));
- floor->GetOutputSlot(0).Connect(output->GetInputSlot(0));
-
- armnn::TensorShape shape({4});
- armnn::TensorInfo info(shape, armnn::DataType::Float32);
- input->GetOutputSlot(0).SetTensorInfo(info);
- floor->GetOutputSlot(0).SetTensorInfo(info);
-
- armnn::IRuntime::CreationOptions options;
- armnn::IRuntimePtr runtime(armnn::IRuntime::Create(options));
-
- std::vector<armnn::BackendId> backends = {armnn::Compute::CpuRef};
-
- armnn::OptimizerOptions optimizerOptions;
- optimizerOptions.m_ReduceFp32ToFp16 = true;
-
- armnn::IOptimizedNetworkPtr optimizedNet = armnn::Optimize(net, backends, runtime->GetDeviceSpec(),
- optimizerOptions);
-
- std::ostringstream ss;
- optimizedNet->SerializeToDot(ss);
-
- auto inputId = input->GetGuid();
- auto floorId = floor->GetGuid();
- auto outputId = output->GetGuid();
-
- std::stringstream expected;
- expected <<
- "digraph Optimized {\n"
- " node [shape=\"record\"];\n"
- " edge [fontsize=8 fontcolor=\"blue\" fontname=\"arial-bold\"];\n"
- " " << inputId << " [label=\"{Input}\"];\n"
- " " << floorId << " [label=\"{Floor}\"];\n"
- " " << outputId << " [label=\"{Output}\"];\n"
- " " << inputId << " -> " << floorId << " [label=< [4] >];\n"
- " " << floorId << " -> " << outputId << " [label=< [4] >];\n"
- "}\n";
-
- BOOST_TEST(ss.str() == expected.str());
-}
-
-#if ARMCOMPUTECL_ENABLED
-BOOST_AUTO_TEST_CASE(FP16TurboModeTestOnGpuAcc)
-{
- // Test to check when Fp16 Turbo mode set
- // it converts the Fp32 network to Fp16 Network
- // add Fp32ToFp16 conversion layer after the InputLayer
- // add Fp16ToFp32 conversion layer after the OutputLayer
- // checks the other layers if they are supported in Fp16
- // if they are not put the conversion layers before and after
- // if they are not supported in Fp16 use Fp32 instead
- // if there are inverse conversion layers remove them with optimization
- // at the moment FloorLayer is not supported in Fp16 so it rolls back to Fp32
- // and inverse conversion layers are removed by the optimizer
- armnn::Network net;
-
- // Defines layers.
- auto input = net.AddInputLayer(0, "input layer");
- // ReLu1
- armnn::ActivationDescriptor activation1Descriptor;
- activation1Descriptor.m_Function = armnn::ActivationFunction::BoundedReLu;
- activation1Descriptor.m_A = 1.f;
- activation1Descriptor.m_B = -1.f;
- auto activation = net.AddActivationLayer(activation1Descriptor, "activation layer");
- auto output = net.AddOutputLayer(0, "output layer");
-
- // Connects layers.
- input->GetOutputSlot(0).Connect(activation->GetInputSlot(0));
- activation->GetOutputSlot(0).Connect(output->GetInputSlot(0));
-
- armnn::TensorShape shape({4});
- armnn::TensorInfo info(shape, armnn::DataType::Float32);
- input->GetOutputSlot(0).SetTensorInfo(info);
- activation->GetOutputSlot(0).SetTensorInfo(info);
-
- armnn::IRuntime::CreationOptions options;
- armnn::IRuntimePtr runtime(armnn::IRuntime::Create(options));
-
- std::vector<armnn::BackendId> backends = {armnn::Compute::GpuAcc};
-
- armnn::OptimizerOptions optimizerOptions;
- optimizerOptions.m_ReduceFp32ToFp16 = true;
-
- armnn::IOptimizedNetworkPtr optimizedNet = armnn::Optimize(
- net, backends, runtime->GetDeviceSpec(), optimizerOptions);
-
- const armnn::Graph& graph = static_cast<armnn::OptimizedNetwork*>(optimizedNet.get())->GetGraph();
-
- // Tests that all layers are present in the graph.
- BOOST_TEST(graph.GetNumLayers() == 5);
-
- // Tests that the vertices exist and have correct names.
- BOOST_TEST(GraphHasNamedLayer(graph, "input layer"));
- BOOST_TEST(GraphHasNamedLayer(graph, "convert_fp32_to_fp16-0-input layer"));
- BOOST_TEST(GraphHasNamedLayer(graph, "activation layer"));
- BOOST_TEST(GraphHasNamedLayer(graph, "convert_fp16_to_fp32-0-output layer"));
- BOOST_TEST(GraphHasNamedLayer(graph, "output layer"));
-}
-#endif
-
BOOST_AUTO_TEST_SUITE_END()
generated by cgit v1.2.1 (git 2.23.0) at 2024-05-15 02:58:10 +0000