IVGCVSW-7211 Fix float16 operators being wrongly unsupported with android-nn-driver.
!armnn:8862
* Added float16 mean test cases.
* Float16 CTS/CTS pass on CpuAcc.
Signed-off-by: Cathal Corbett <cathal.corbett@arm.com>
Change-Id: Ibd9021d0ae4a205cc2c91555f3ae00c6dba84609
diff --git a/test/1.2/Mean.cpp b/test/1.2/Mean.cpp
new file mode 100644
index 0000000..a2a8b7a
--- /dev/null
+++ b/test/1.2/Mean.cpp
@@ -0,0 +1,204 @@
+//
+// Copyright © 2022 Arm Ltd and Contributors. All rights reserved.
+// SPDX-License-Identifier: MIT
+//
+
+#include "../DriverTestHelpers.hpp"
+#include "../TestHalfTensor.hpp"
+
+#include <1.2/HalPolicy.hpp>
+
+#include <array>
+
+using Half = half_float::half;
+
+using namespace android::hardware;
+using namespace driverTestHelpers;
+using namespace armnn_driver;
+
+using HalPolicy = hal_1_2::HalPolicy;
+using RequestArgument = V1_0::RequestArgument;
+
+namespace
+{
+
+void MeanTestImpl(const TestHalfTensor& input,
+ const hidl_vec<uint32_t>& axisDimensions,
+ const int32_t* axisValues,
+ int32_t keepDims,
+ const TestHalfTensor& expectedOutput,
+ bool fp16Enabled,
+ armnn::Compute computeDevice)
+{
+ auto driver = std::make_unique<ArmnnDriver>(DriverOptions(computeDevice, fp16Enabled));
+
+ HalPolicy::Model model = {};
+
+ AddInputOperand<HalPolicy>(model, input.GetDimensions(), V1_2::OperandType::TENSOR_FLOAT16);
+
+ AddTensorOperand<HalPolicy>(model,
+ axisDimensions,
+ const_cast<int32_t*>(axisValues),
+ HalPolicy::OperandType::TENSOR_INT32);
+
+ AddIntOperand<HalPolicy>(model, keepDims);
+
+ AddOutputOperand<HalPolicy>(model, expectedOutput.GetDimensions(), V1_2::OperandType::TENSOR_FLOAT16);
+
+ model.operations.resize(1);
+ model.operations[0].type = HalPolicy::OperationType::MEAN;
+ model.operations[0].inputs = hidl_vec<uint32_t>{ 0, 1, 2 };
+ model.operations[0].outputs = hidl_vec<uint32_t>{ 3 };
+ model.relaxComputationFloat32toFloat16 = fp16Enabled;
+
+ //android::sp<V1_0::IPreparedModel> preparedModel = PrepareModel(model, *driver);
+ android::sp<V1_2::IPreparedModel> preparedModel = PrepareModel_1_2(model, *driver);
+
+ // The request's memory pools will follow the same order as the inputs
+ V1_0::DataLocation inLoc = {};
+ inLoc.poolIndex = 0;
+ inLoc.offset = 0;
+ inLoc.length = input.GetNumElements() * sizeof(Half);
+ RequestArgument inArg = {};
+ inArg.location = inLoc;
+ inArg.dimensions = input.GetDimensions();
+
+ // An additional memory pool is needed for the output
+ V1_0::DataLocation outLoc = {};
+ outLoc.poolIndex = 1;
+ outLoc.offset = 0;
+ outLoc.length = expectedOutput.GetNumElements() * sizeof(Half);
+ RequestArgument outArg = {};
+ outArg.location = outLoc;
+ outArg.dimensions = expectedOutput.GetDimensions();
+
+ // Make the request based on the arguments
+ V1_0::Request request = {};
+ request.inputs = hidl_vec<RequestArgument>{ inArg };
+ request.outputs = hidl_vec<RequestArgument>{ outArg };
+
+ // Set the input data
+ AddPoolAndSetData(input.GetNumElements(), request, input.GetData());
+
+ // Add memory for the output
+ android::sp<IMemory> outMemory = AddPoolAndGetData<Half>(expectedOutput.GetNumElements(), request);
+ const Half* outputData = static_cast<const Half*>(static_cast<void*>(outMemory->getPointer()));
+
+ if (preparedModel.get() != nullptr)
+ {
+ V1_0::ErrorStatus execStatus = Execute(preparedModel, request);
+ DOCTEST_CHECK((int)execStatus == (int)V1_0::ErrorStatus::NONE);
+ }
+
+ const Half* expectedOutputData = expectedOutput.GetData();
+ for (unsigned int i = 0; i < expectedOutput.GetNumElements(); i++)
+ {
+ DOCTEST_CHECK(outputData[i] == expectedOutputData[i]);
+ }
+}
+
+} // anonymous namespace
+
+DOCTEST_TEST_SUITE("MeanTests_1.2_CpuRef")
+{
+
+DOCTEST_TEST_CASE("MeanFp16NoKeepDimsTest_CpuRef")
+{
+ using namespace half_float::literal;
+
+ TestHalfTensor input{ armnn::TensorShape{ 4, 3, 2 },
+ { 1.0_h, 2.0_h, 3.0_h, 4.0_h, 5.0_h, 6.0_h, 7.0_h, 8.0_h, 9.0_h, 10.0_h,
+ 11.0_h, 12.0_h, 13.0_h, 14.0_h, 15.0_h, 16.0_h, 17.0_h, 18.0_h, 19.0_h,
+ 20.0_h, 21.0_h, 22.0_h, 23.0_h, 24.0_h } };
+ hidl_vec<uint32_t> axisDimensions = { 2 };
+ int32_t axisValues[] = { 0, 1 };
+ int32_t keepDims = 0;
+ TestHalfTensor expectedOutput{ armnn::TensorShape{ 2 }, { 12.0_h, 13.0_h } };
+
+ MeanTestImpl(input, axisDimensions, axisValues, keepDims, expectedOutput, true, armnn::Compute::CpuRef);
+}
+
+DOCTEST_TEST_CASE("MeanFp16KeepDimsTest_CpuRef")
+{
+ using namespace half_float::literal;
+
+ TestHalfTensor input{ armnn::TensorShape{ 1, 1, 3, 2 }, { 1.0_h, 1.0_h, 2.0_h, 2.0_h, 3.0_h, 3.0_h } };
+ hidl_vec<uint32_t> axisDimensions = { 1 };
+ int32_t axisValues[] = { 2 };
+ int32_t keepDims = 1;
+ TestHalfTensor expectedOutput{ armnn::TensorShape{ 1, 1, 1, 2 }, { 2.0_h, 2.0_h } };
+
+ MeanTestImpl(input, axisDimensions, axisValues, keepDims, expectedOutput, true, armnn::Compute::CpuRef);
+}
+
+}
+
+#ifdef ARMCOMPUTECL_ENABLED
+DOCTEST_TEST_SUITE("MeanTests_1.2_CpuAcc")
+{
+ DOCTEST_TEST_CASE("MeanFp16NoKeepDimsTest_CpuAcc")
+ {
+ using namespace half_float::literal;
+
+ std::vector<Half> in = { 1.0_h, 2.0_h, 3.0_h, 4.0_h, 5.0_h, 6.0_h, 7.0_h, 8.0_h, 9.0_h, 10.0_h,
+ 11.0_h, 12.0_h, 13.0_h, 14.0_h, 15.0_h, 16.0_h, 17.0_h, 18.0_h, 19.0_h,
+ 20.0_h, 21.0_h, 22.0_h, 23.0_h, 24.0_h };
+ TestHalfTensor input{ armnn::TensorShape{ 4, 3, 2 },
+ in};
+ hidl_vec<uint32_t> axisDimensions = { 2 };
+ int32_t axisValues[] = { 0, 1 };
+ int32_t keepDims = 0;
+ std::vector<Half> out = { 12.0_h, 13.0_h };
+ TestHalfTensor expectedOutput{ armnn::TensorShape{ 2 }, out };
+
+ MeanTestImpl(input, axisDimensions, axisValues, keepDims, expectedOutput, true, armnn::Compute::CpuAcc);
+ }
+
+ DOCTEST_TEST_CASE("MeanFp16KeepDimsTest_CpuAcc")
+ {
+ using namespace half_float::literal;
+
+ std::vector<Half> in = { 1.0_h, 1.0_h, 2.0_h, 2.0_h, 3.0_h, 3.0_h };
+ TestHalfTensor input{ armnn::TensorShape{ 1, 1, 3, 2 }, in };
+ hidl_vec<uint32_t> axisDimensions = { 1 };
+ int32_t axisValues[] = { 2 };
+ int32_t keepDims = 1;
+ std::vector<Half> out = { 2.0_h, 2.0_h };
+ TestHalfTensor expectedOutput{ armnn::TensorShape{ 1, 1, 1, 2 }, out };
+
+ MeanTestImpl(input, axisDimensions, axisValues, keepDims, expectedOutput, true, armnn::Compute::CpuAcc);
+ }
+}
+
+DOCTEST_TEST_SUITE("MeanTests_1.2_GpuAcc")
+{
+ DOCTEST_TEST_CASE("MeanFp16NoKeepDimsTest_GpuAcc")
+ {
+ using namespace half_float::literal;
+
+ TestHalfTensor input{ armnn::TensorShape{ 4, 3, 2 },
+ { 1.0_h, 2.0_h, 3.0_h, 4.0_h, 5.0_h, 6.0_h, 7.0_h, 8.0_h, 9.0_h, 10.0_h,
+ 11.0_h, 12.0_h, 13.0_h, 14.0_h, 15.0_h, 16.0_h, 17.0_h, 18.0_h, 19.0_h,
+ 20.0_h, 21.0_h, 22.0_h, 23.0_h, 24.0_h } };
+ hidl_vec<uint32_t> axisDimensions = { 2 };
+ int32_t axisValues[] = { 0, 1 };
+ int32_t keepDims = 0;
+ TestHalfTensor expectedOutput{ armnn::TensorShape{ 2 }, { 12.0_h, 13.0_h } };
+
+ MeanTestImpl(input, axisDimensions, axisValues, keepDims, expectedOutput, true, armnn::Compute::GpuAcc);
+ }
+
+ DOCTEST_TEST_CASE("MeanFp16KeepDimsTest_GpuAcc")
+ {
+ using namespace half_float::literal;
+
+ TestHalfTensor input{ armnn::TensorShape{ 1, 1, 3, 2 }, { 1.0_h, 1.0_h, 2.0_h, 2.0_h, 3.0_h, 3.0_h } };
+ hidl_vec<uint32_t> axisDimensions = { 1 };
+ int32_t axisValues[] = { 2 };
+ int32_t keepDims = 1;
+ TestHalfTensor expectedOutput{ armnn::TensorShape{ 1, 1, 1, 2 }, { 2.0_h, 2.0_h } };
+
+ MeanTestImpl(input, axisDimensions, axisValues, keepDims, expectedOutput, true, armnn::Compute::GpuAcc);
+ }
+}
+#endif