IVGCVSW-8276 GpuFsa Op: Add MatMul
Signed-off-by: Teresa Charlin <teresa.charlinreyes@arm.com>
Change-Id: Ib95eb0fd71106e684cb7652917b8de9f0ac73f9c
diff --git a/src/backends/gpuFsa/test/GpuFsaEndToEndTests.cpp b/src/backends/gpuFsa/test/GpuFsaEndToEndTests.cpp
index a2708c0..7503c46 100644
--- a/src/backends/gpuFsa/test/GpuFsaEndToEndTests.cpp
+++ b/src/backends/gpuFsa/test/GpuFsaEndToEndTests.cpp
@@ -5,9 +5,9 @@
#include "backendsCommon/test/EndToEndTestImpl.hpp"
+#include "backendsCommon/test/BatchMatMulEndToEndTestImpl.hpp"
#include "backendsCommon/test/Convolution2dEndToEndTestImpl.hpp"
#include "backendsCommon/test/layerTests/CastTestImpl.hpp"
-
#include "backendsCommon/test/DepthwiseConvolution2dEndToEndTests.hpp"
#include "backendsCommon/test/ElementwiseBinaryEndToEndTestImpl.hpp"
#include "backendsCommon/test/Pooling2dEndToEndTestImpl.hpp"
@@ -20,6 +20,13 @@
std::vector<BackendId> gpuFsaDefaultBackends = {"GpuFsa"};
+// BatchMatMul
+TEST_CASE("RefBatchMatMulEndToEndFloat32Test")
+{
+ BatchMatMulEndToEnd<armnn::DataType::Float32>(gpuFsaDefaultBackends);
+}
+
+// Cast
TEST_CASE("GpuFsaCastEndtoEndTestFloat32ToFloat16")
{
using namespace half_float::literal;
diff --git a/src/backends/gpuFsa/test/GpuFsaLayerSupportTests.cpp b/src/backends/gpuFsa/test/GpuFsaLayerSupportTests.cpp
index 34af190..b6f7f32 100644
--- a/src/backends/gpuFsa/test/GpuFsaLayerSupportTests.cpp
+++ b/src/backends/gpuFsa/test/GpuFsaLayerSupportTests.cpp
@@ -17,6 +17,25 @@
TEST_SUITE("GpuFsaLayerSupport")
{
+TEST_CASE("IsLayerSupportedGpuFsaBatchMatMul")
+{
+ TensorInfo input0Info({ 2, 2 }, DataType::Float32);
+ TensorInfo input1Info({ 2, 2 }, DataType::Float32);
+ TensorInfo outputInfo({ 2, 2 }, DataType::Float32);
+
+ BatchMatMulDescriptor desc{};
+
+ GpuFsaLayerSupport supportChecker;
+ std::string reasonIfNotSupported;
+ auto supported = supportChecker.IsLayerSupported(LayerType::BatchMatMul,
+ {input0Info, input1Info, outputInfo},
+ desc,
+ EmptyOptional(),
+ EmptyOptional(),
+ reasonIfNotSupported);
+ CHECK(supported);
+}
+
TEST_CASE("IsLayerSupportedCast")
{
armnn::TensorInfo inputTensorInfo({1, 3, 2, 3}, armnn::DataType::Float32);
diff --git a/src/backends/gpuFsa/test/GpuFsaOptimizedNetworkTests.cpp b/src/backends/gpuFsa/test/GpuFsaOptimizedNetworkTests.cpp
index 6ddb942..1e5c976 100644
--- a/src/backends/gpuFsa/test/GpuFsaOptimizedNetworkTests.cpp
+++ b/src/backends/gpuFsa/test/GpuFsaOptimizedNetworkTests.cpp
@@ -15,6 +15,52 @@
TEST_SUITE("GpuFsaOptimizedNetwork")
{
+TEST_CASE("BatchMatMulSupportedOptimizedNetwork")
+{
+ using namespace armnn;
+
+ const float qScale = 1.0f;
+ const int32_t qOffset = 0;
+
+ const TensorShape& input1Shape = { 2, 2 };
+ const TensorShape& input2Shape = { 2, 2 };
+ const TensorShape& outputShape = { 2, 2 };
+
+ TensorInfo input1TensorInfo(input1Shape, DataType::Float32, qScale, qOffset, true);
+ TensorInfo input2TensorInfo(input2Shape, DataType::Float32, qScale, qOffset, true);
+ TensorInfo outputTensorInfo(outputShape, DataType::Float32, qScale, qOffset);
+
+ IRuntime::CreationOptions options;
+ IRuntimePtr runtime(IRuntime::Create(options));
+ INetworkPtr network(INetwork::Create());
+
+ BatchMatMulDescriptor desc{};
+
+ IConnectableLayer* input1 = network->AddInputLayer(0, "input0");
+ IConnectableLayer* input2 = network->AddInputLayer(1, "input1");
+ IConnectableLayer* batchMatMulLayer = network->AddBatchMatMulLayer(desc, "batchMatMul");
+ IConnectableLayer* output = network->AddOutputLayer(2, "output");
+
+ Connect(input1, batchMatMulLayer, input1TensorInfo, 0, 0);
+ Connect(input2, batchMatMulLayer, input2TensorInfo, 0, 1);
+ Connect(batchMatMulLayer, output, outputTensorInfo, 0, 0);
+
+ std::vector<BackendId> backends = { "GpuFsa" };
+
+ OptimizerOptionsOpaque optimizedOptions;
+ IOptimizedNetworkPtr optNet = Optimize(*network, backends, runtime->GetDeviceSpec(), optimizedOptions);
+ CHECK(optNet);
+
+ Graph& graph = GetGraphForTesting(optNet.get());
+
+ // Check graph layer sequence to ensure that the network has been replaced with a PreCompiledLayer
+ CHECK(CheckSequence(graph.cbegin(), graph.cend(),
+ &IsLayerOfType<InputLayer>,
+ &IsLayerOfType<InputLayer>,
+ &IsLayerOfType<PreCompiledLayer>,
+ &IsLayerOfType<OutputLayer>));
+}
+
TEST_CASE("CastSupportedOptimizedNetwork")
{
using namespace armnn;