aboutsummaryrefslogtreecommitdiff
path: root/delegate/src/test
diff options
context:
space:
mode:
Diffstat (limited to 'delegate/src/test')
-rw-r--r--delegate/src/test/LstmTest.cpp189
-rw-r--r--delegate/src/test/LstmTestHelper.hpp691
2 files changed, 880 insertions, 0 deletions
diff --git a/delegate/src/test/LstmTest.cpp b/delegate/src/test/LstmTest.cpp
new file mode 100644
index 0000000000..1fa9f0c8bf
--- /dev/null
+++ b/delegate/src/test/LstmTest.cpp
@@ -0,0 +1,189 @@
+//
+// Copyright © 2021 Arm Ltd and Contributors. All rights reserved.
+// SPDX-License-Identifier: MIT
+//
+
+#include "LstmTestHelper.hpp"
+
+#include <armnn_delegate.hpp>
+
+#include <flatbuffers/flatbuffers.h>
+#include <tensorflow/lite/schema/schema_generated.h>
+#include <doctest/doctest.h>
+
+namespace armnnDelegate
+{
+
+void LstmTest(std::vector<armnn::BackendId>& backends)
+{
+ int32_t batchSize = 2;
+ int32_t inputSize = 2;
+ int32_t outputSize = 4;
+ // cellSize and outputSize have the same size when there is no projection.
+ int32_t numUnits = outputSize;
+
+ std::vector<int32_t> inputShape {batchSize , inputSize};
+ std::vector<int32_t> cellStateInTensorInfo {batchSize , numUnits};
+ std::vector<int32_t> outputStateInTensorInfo {batchSize , outputSize};
+
+ std::vector<int32_t> scratchBufferTensorInfo {batchSize, numUnits * 4};
+ std::vector<int32_t> cellStateOutTensorInfo {batchSize, numUnits};
+ std::vector<int32_t> outputStateOutTensorInfo {batchSize, outputSize};
+ std::vector<int32_t> outputTensorInfo {batchSize, outputSize};
+
+ std::vector<int32_t> tensorInfo4 {numUnits};
+ std::vector<int32_t> tensorInfo8 {numUnits, 2};
+ std::vector<int32_t> tensorInfo16 {numUnits, 4};
+
+ //tensorInfo8,
+ bool hasInputToInputWeights = true;
+ std::vector<float> inputToInputWeights {-0.45018822f, -0.02338299f, -0.0870589f,
+ -0.34550029f, 0.04266912f, -0.15680569f,
+ -0.34856534f, 0.43890524f};
+
+ std::vector<float> inputToForgetWeights {0.09701663f, 0.20334584f, -0.50592935f,
+ -0.31343272f, -0.40032279f, 0.44781327f,
+ 0.01387155f, -0.35593212f};
+
+ std::vector<float> inputToCellWeights {-0.50013041f, 0.1370284f, 0.11810488f, 0.2013163f,
+ -0.20583314f, 0.44344562f, 0.22077113f,
+ -0.29909778f};
+
+ std::vector<float> inputToOutputWeights {-0.25065863f, -0.28290087f, 0.04613829f,
+ 0.40525138f, 0.44272184f, 0.03897077f,
+ -0.1556896f, 0.19487578f};
+
+ //tensorInfo16,
+ bool hasRecurrentToInputWeights = true;
+ std::vector<float> recurrentToInputWeights {-0.0063535f, -0.2042388f, 0.31454784f,
+ -0.35746509f, 0.28902304f, 0.08183324f,
+ -0.16555229f, 0.02286911f, -0.13566875f,
+ 0.03034258f, 0.48091322f, -0.12528998f,
+ 0.24077177f, -0.51332325f, -0.33502164f,
+ 0.10629296f};
+
+ std::vector<float> recurrentToForgetWeights {-0.48684245f, -0.06655136f, 0.42224967f,
+ 0.2112639f, 0.27654213f, 0.20864892f,
+ -0.07646349f, 0.45877004f, 0.00141793f,
+ -0.14609534f, 0.36447752f, 0.09196436f,
+ 0.28053468f, 0.01560611f, -0.20127171f,
+ -0.01140004f};
+
+ std::vector<float> recurrentToCellWeights {-0.3407414f, 0.24443203f, -0.2078532f,
+ 0.26320225f, 0.05695659f, -0.00123841f,
+ -0.4744786f, -0.35869038f, -0.06418842f,
+ -0.13502428f, -0.501764f, 0.22830659f,
+ -0.46367589f, 0.26016325f, -0.03894562f,
+ -0.16368064f};
+
+ std::vector<float> recurrentToOutputWeights {0.43385774f, -0.17194885f, 0.2718237f,
+ 0.09215671f, 0.24107647f, -0.39835793f,
+ 0.18212086f, 0.01301402f, 0.48572797f,
+ -0.50656658f, 0.20047462f, -0.20607421f,
+ -0.51818722f, -0.15390486f, 0.0468148f,
+ 0.39922136f};
+ // tensorInfo4
+ bool hasCellToInputWeights = false;
+ std::vector<float> cellToInputWeights {};
+ bool hasCellToForgetWeights = false;
+ std::vector<float> cellToForgetWeights {};
+ bool hasCellToOutputWeights = false;
+ std::vector<float> cellToOutputWeights {};
+
+ bool hasInputGateBias = true;
+ std::vector<float> inputGateBias {0., 0., 0., 0.};
+ std::vector<float> forgetGateBias {1., 1., 1., 1.};
+ std::vector<float> cellBias {0., 0., 0., 0.};
+ std::vector<float> outputGateBias {0., 0., 0., 0.};
+
+ bool hasProjectionWeights = false;
+ std::vector<float> projectionWeights;
+ bool hasProjectionBias = false;
+ std::vector<float> projectionBias;
+
+ bool hasInputLayerNormWeights = false;
+ std::vector<float> inputLayerNormWeights;
+ bool hasForgetLayerNormWeights = false;
+ std::vector<float> forgetLayerNormWeights;
+ bool hasCellLayerNormWeights = false;
+ std::vector<float> cellLayerNormWeights;
+ bool hasOutputLayerNormWeights = false;
+ std::vector<float> outputLayerNormWeights;
+
+ std::vector<float> inputValues {2., 3., 3., 4.};
+ std::vector<float> expectedOutputValues {-0.02973187f, 0.1229473f, 0.20885126f, -0.15358765f,
+ -0.0185422f, 0.11281417f, 0.24466537f, -0.1826292f};
+
+ tflite::ActivationFunctionType activationFunction = tflite::ActivationFunctionType_TANH;
+ float clippingThresCell = 0.f;
+ float clippingThresProj = 0.f;
+
+ LstmTestImpl<float>(backends,
+ ::tflite::TensorType_FLOAT32,
+ batchSize,
+ inputSize,
+ outputSize,
+ numUnits,
+ hasInputToInputWeights,
+ inputToInputWeights,
+ inputToForgetWeights,
+ inputToCellWeights,
+ inputToOutputWeights,
+ hasRecurrentToInputWeights,
+ recurrentToInputWeights,
+ recurrentToForgetWeights,
+ recurrentToCellWeights,
+ recurrentToOutputWeights,
+ hasCellToInputWeights,
+ cellToInputWeights,
+ hasCellToForgetWeights,
+ cellToForgetWeights,
+ hasCellToOutputWeights,
+ cellToOutputWeights,
+ hasInputGateBias,
+ inputGateBias,
+ forgetGateBias,
+ cellBias,
+ outputGateBias,
+ hasProjectionWeights,
+ projectionWeights,
+ hasProjectionBias,
+ projectionBias,
+ hasInputLayerNormWeights,
+ inputLayerNormWeights,
+ hasForgetLayerNormWeights,
+ forgetLayerNormWeights,
+ hasCellLayerNormWeights,
+ cellLayerNormWeights,
+ hasOutputLayerNormWeights,
+ outputLayerNormWeights,
+ inputValues,
+ expectedOutputValues,
+ activationFunction,
+ clippingThresCell,
+ clippingThresProj);
+}
+
+TEST_SUITE("LstmTest_CpuRefTests")
+{
+
+TEST_CASE ("LstmTest_CpuRef_Test")
+{
+ std::vector <armnn::BackendId> backends = {armnn::Compute::CpuRef};
+ LstmTest(backends);
+}
+
+} //End of TEST_SUITE("Convolution2dTest_CpuRef")
+
+TEST_SUITE("LstmTest_CpuAccTests")
+{
+
+TEST_CASE ("LstmTest_CpuAcc_Test")
+{
+ std::vector <armnn::BackendId> backends = {armnn::Compute::CpuAcc};
+ LstmTest(backends);
+}
+
+} //End of TEST_SUITE("Convolution2dTest_CpuAcc")
+
+} // namespace armnnDelegate \ No newline at end of file
diff --git a/delegate/src/test/LstmTestHelper.hpp b/delegate/src/test/LstmTestHelper.hpp
new file mode 100644
index 0000000000..36a606119a
--- /dev/null
+++ b/delegate/src/test/LstmTestHelper.hpp
@@ -0,0 +1,691 @@
+//
+// Copyright © 2021 Arm Ltd and Contributors. All rights reserved.
+// SPDX-License-Identifier: MIT
+//
+
+#pragma once
+
+#include "TestUtils.hpp"
+
+#include <armnn_delegate.hpp>
+
+#include <flatbuffers/flatbuffers.h>
+#include <tensorflow/lite/interpreter.h>
+#include <tensorflow/lite/kernels/register.h>
+#include <tensorflow/lite/model.h>
+#include <tensorflow/lite/schema/schema_generated.h>
+#include <tensorflow/lite/version.h>
+#include <tensorflow/lite/c/common.h>
+
+#include <doctest/doctest.h>
+
+namespace
+{
+
+template <typename T>
+std::vector<char> CreateLstmTfLiteModel(tflite::TensorType tensorType,
+ int32_t batchSize,
+ int32_t inputSize,
+ int32_t outputSize,
+ int32_t numUnits,
+ bool hasInputToInputWeights,
+ const std::vector<T>& inputToInputWeights,
+ const std::vector<T>& inputToForgetWeights,
+ const std::vector<T>& inputToCellWeights,
+ const std::vector<T>& inputToOutputWeights,
+ bool hasRecurrentToInputWeights,
+ const std::vector<T>& recurrentToInputWeights,
+ const std::vector<T>& recurrentToForgetWeights,
+ const std::vector<T>& recurrentToCellWeights,
+ const std::vector<T>& recurrentToOutputWeights,
+ bool hasCellToInputWeights,
+ const std::vector<T>& cellToInputWeights,
+ bool hasCellToForgetWeights,
+ const std::vector<T>& cellToForgetWeights,
+ bool hasCellToOutputWeights,
+ const std::vector<T>& cellToOutputWeights,
+ bool hasInputGateBias,
+ const std::vector<T>& inputGateBias,
+ const std::vector<T>& forgetGateBias,
+ const std::vector<T>& cellBias,
+ const std::vector<T>& outputGateBias,
+ bool hasProjectionWeights,
+ const std::vector<T>& projectionWeights,
+ bool hasProjectionBias,
+ const std::vector<T>& projectionBias,
+ bool hasInputLayerNormWeights,
+ const std::vector<T>& inputLayerNormWeights,
+ bool hasForgetLayerNormWeights,
+ const std::vector<T>& forgetLayerNormWeights,
+ bool hasCellLayerNormWeights,
+ const std::vector<T>& cellLayerNormWeights,
+ bool hasOutputLayerNormWeights,
+ const std::vector<T>& outputLayerNormWeights,
+ tflite::ActivationFunctionType activationFunction,
+ float clippingThresCell,
+ float clippingThresProj,
+ float quantScale = 1.0f,
+ int quantOffset = 0,
+ float outputQuantScale = 2.0f,
+ int outputQuantOffset = 0)
+{
+
+ std::vector <int32_t> tensorInfo0 {};
+ std::vector <int32_t> tensorInfo4 {numUnits};
+ std::vector <int32_t> tensorInfo8 {numUnits, static_cast<int32_t>(2)};
+ std::vector <int32_t> tensorInfo16 {numUnits, static_cast<int32_t>(4)};
+
+ std::vector<int32_t> inputShape {batchSize , inputSize};
+ std::vector<int32_t> outputShape {batchSize , outputSize};
+
+ std::vector<int32_t> outputStateInDimensions{batchSize, outputSize};
+ std::vector<int32_t> cellStateInDimensions{batchSize, numUnits};
+
+ std::vector<int> operatorInputs;
+ using namespace tflite;
+ flatbuffers::FlatBufferBuilder flatBufferBuilder;
+ std::vector<flatbuffers::Offset<tflite::Buffer>> buffers;
+ std::vector<flatbuffers::Offset<Tensor>> tensors;
+
+ auto quantizationParameters =
+ CreateQuantizationParameters(flatBufferBuilder,
+ 0,
+ 0,
+ flatBufferBuilder.CreateVector<float>({ quantScale }),
+ flatBufferBuilder.CreateVector<int64_t>({ quantOffset }));
+
+ auto outputQuantizationParameters =
+ CreateQuantizationParameters(flatBufferBuilder,
+ 0,
+ 0,
+ flatBufferBuilder.CreateVector<float>({ outputQuantScale }),
+ flatBufferBuilder.CreateVector<int64_t>({ outputQuantOffset }));
+
+ buffers.push_back(CreateBuffer(flatBufferBuilder, flatBufferBuilder.CreateVector({})));
+ tensors.push_back(CreateTensor(flatBufferBuilder,
+ flatBufferBuilder.CreateVector<int32_t>(inputShape.data(),
+ inputShape.size()),
+ tensorType,
+ buffers.size() - 1,
+ flatBufferBuilder.CreateString("input_0"),
+ quantizationParameters));
+ operatorInputs.push_back(buffers.size() - 1);
+
+ if (hasInputToInputWeights)
+ {
+ buffers.push_back(
+ CreateBuffer(flatBufferBuilder,
+ flatBufferBuilder.CreateVector(reinterpret_cast<const uint8_t *>(inputToInputWeights.data()),
+ sizeof(T) * inputToInputWeights.size())));
+ tensors.push_back(CreateTensor(flatBufferBuilder,
+ flatBufferBuilder.CreateVector<int32_t>(tensorInfo8.data(),
+ tensorInfo8.size()),
+ tensorType,
+ buffers.size() - 1,
+ flatBufferBuilder.CreateString("inputToInputWeights"),
+ outputQuantizationParameters));
+ operatorInputs.push_back(buffers.size() - 1);
+ }
+ else
+ {
+ operatorInputs.push_back(kTfLiteOptionalTensor);
+ }
+
+ buffers.push_back(
+ CreateBuffer(flatBufferBuilder,
+ flatBufferBuilder.CreateVector(reinterpret_cast<const uint8_t *>(inputToForgetWeights.data()),
+ sizeof(T) * inputToForgetWeights.size())));
+ tensors.push_back(CreateTensor(flatBufferBuilder,
+ flatBufferBuilder.CreateVector<int32_t>(tensorInfo8.data(),
+ tensorInfo8.size()),
+ tensorType,
+ buffers.size() - 1,
+ flatBufferBuilder.CreateString("inputToForgetWeights"),
+ outputQuantizationParameters));
+ operatorInputs.push_back(buffers.size() - 1);
+
+ buffers.push_back(
+ CreateBuffer(flatBufferBuilder,
+ flatBufferBuilder.CreateVector(reinterpret_cast<const uint8_t *>(inputToCellWeights.data()),
+ sizeof(T) * inputToCellWeights.size())));
+ tensors.push_back(CreateTensor(flatBufferBuilder,
+ flatBufferBuilder.CreateVector<int32_t>(tensorInfo8.data(),
+ tensorInfo8.size()),
+ tensorType,
+ buffers.size() - 1,
+ flatBufferBuilder.CreateString("inputToCellWeights"),
+ outputQuantizationParameters));
+ operatorInputs.push_back(buffers.size() - 1);
+
+ buffers.push_back(
+ CreateBuffer(flatBufferBuilder,
+ flatBufferBuilder.CreateVector(reinterpret_cast<const uint8_t *>(inputToOutputWeights.data()),
+ sizeof(T) * inputToOutputWeights.size())));
+ tensors.push_back(CreateTensor(flatBufferBuilder,
+ flatBufferBuilder.CreateVector<int32_t>(tensorInfo8.data(),
+ tensorInfo8.size()),
+ tensorType,
+ buffers.size() - 1,
+ flatBufferBuilder.CreateString("inputToOutputWeights"),
+ outputQuantizationParameters));
+ operatorInputs.push_back(buffers.size() - 1);
+
+ if (hasRecurrentToInputWeights)
+ {
+ buffers.push_back(CreateBuffer(
+ flatBufferBuilder,
+ flatBufferBuilder.CreateVector(reinterpret_cast<const uint8_t*>(recurrentToInputWeights.data()),
+ sizeof(T) * recurrentToInputWeights.size())));
+ tensors.push_back(CreateTensor(flatBufferBuilder,
+ flatBufferBuilder.CreateVector<int32_t>(tensorInfo16.data(),
+ tensorInfo16.size()),
+ tensorType,
+ buffers.size() - 1,
+ flatBufferBuilder.CreateString("recurrentToInputWeights"),
+ outputQuantizationParameters));
+ operatorInputs.push_back(buffers.size() - 1);
+ }
+ else
+ {
+ operatorInputs.push_back(kTfLiteOptionalTensor);
+ }
+
+ buffers.push_back(
+ CreateBuffer(flatBufferBuilder,
+ flatBufferBuilder.CreateVector(reinterpret_cast<const uint8_t *>(recurrentToForgetWeights.data()),
+ sizeof(T) * recurrentToForgetWeights.size())));
+ tensors.push_back(CreateTensor(flatBufferBuilder,
+ flatBufferBuilder.CreateVector<int32_t>(tensorInfo16.data(),
+ tensorInfo16.size()),
+ tensorType,
+ buffers.size() - 1,
+ flatBufferBuilder.CreateString("recurrentToForgetWeights"),
+ outputQuantizationParameters));
+ operatorInputs.push_back(buffers.size() - 1);
+
+ buffers.push_back(
+ CreateBuffer(flatBufferBuilder,
+ flatBufferBuilder.CreateVector(reinterpret_cast<const uint8_t *>(recurrentToCellWeights.data()),
+ sizeof(T) * recurrentToCellWeights.size())));
+ tensors.push_back(CreateTensor(flatBufferBuilder,
+ flatBufferBuilder.CreateVector<int32_t>(tensorInfo16.data(),
+ tensorInfo16.size()),
+ tensorType,
+ buffers.size() - 1,
+ flatBufferBuilder.CreateString("recurrentToCellWeights"),
+ outputQuantizationParameters));
+ operatorInputs.push_back(buffers.size() - 1);
+
+ buffers.push_back(
+ CreateBuffer(flatBufferBuilder,
+ flatBufferBuilder.CreateVector(reinterpret_cast<const uint8_t *>(recurrentToOutputWeights.data()),
+ sizeof(T) * recurrentToOutputWeights.size())));
+ tensors.push_back(CreateTensor(flatBufferBuilder,
+ flatBufferBuilder.CreateVector<int32_t>(tensorInfo16.data(),
+ tensorInfo16.size()),
+ tensorType,
+ buffers.size() - 1 ,
+ flatBufferBuilder.CreateString("recurrentToOutputWeights"),
+ outputQuantizationParameters));
+ operatorInputs.push_back(buffers.size() - 1);
+
+ if (hasCellToInputWeights)
+ {
+ buffers.push_back(
+ CreateBuffer(flatBufferBuilder,
+ flatBufferBuilder.CreateVector(reinterpret_cast<const uint8_t*>(cellToInputWeights.data()),
+ sizeof(T) * cellToInputWeights.size())));
+ tensors.push_back(CreateTensor(flatBufferBuilder,
+ flatBufferBuilder.CreateVector<int32_t>(tensorInfo4.data(),
+ tensorInfo4.size()),
+ tensorType,
+ buffers.size() - 1,
+ flatBufferBuilder.CreateString("cellToInputWeights"),
+ outputQuantizationParameters));
+ operatorInputs.push_back(buffers.size() - 1);
+ }
+ else
+ {
+ operatorInputs.push_back(kTfLiteOptionalTensor);
+ }
+
+ if (hasCellToForgetWeights)
+ {
+ buffers.push_back(
+ CreateBuffer(flatBufferBuilder,
+ flatBufferBuilder.CreateVector(reinterpret_cast<const uint8_t*>(cellToForgetWeights.data()),
+ sizeof(T) * cellToForgetWeights.size())));
+ tensors.push_back(CreateTensor(flatBufferBuilder,
+ flatBufferBuilder.CreateVector<int32_t>(tensorInfo4.data(),
+ tensorInfo4.size()),
+ tensorType,
+ buffers.size() - 1,
+ flatBufferBuilder.CreateString("cellToForgetWeights"),
+ outputQuantizationParameters));
+ operatorInputs.push_back(buffers.size() - 1);
+ }
+ else
+ {
+ operatorInputs.push_back(kTfLiteOptionalTensor);
+ }
+
+ if (hasCellToOutputWeights)
+ {
+ buffers.push_back(
+ CreateBuffer(flatBufferBuilder,
+ flatBufferBuilder.CreateVector(reinterpret_cast<const uint8_t*>(cellToOutputWeights.data()),
+ sizeof(T) * cellToOutputWeights.size())));
+ tensors.push_back(CreateTensor(flatBufferBuilder,
+ flatBufferBuilder.CreateVector<int32_t>(tensorInfo4.data(),
+ tensorInfo4.size()),
+ tensorType,
+ buffers.size() - 1,
+ flatBufferBuilder.CreateString("cellToOutputWeights"),
+ outputQuantizationParameters));
+ operatorInputs.push_back(buffers.size() - 1);
+ }
+ else
+ {
+ operatorInputs.push_back(kTfLiteOptionalTensor);
+ }
+
+ if (hasInputGateBias)
+ {
+ buffers.push_back(
+ CreateBuffer(flatBufferBuilder,
+ flatBufferBuilder.CreateVector(reinterpret_cast<const uint8_t*>(inputGateBias.data()),
+ sizeof(T) * inputGateBias.size())));
+ tensors.push_back(CreateTensor(flatBufferBuilder,
+ flatBufferBuilder.CreateVector<int32_t>(tensorInfo4.data(),
+ tensorInfo4.size()),
+ tensorType,
+ buffers.size() - 1,
+ flatBufferBuilder.CreateString("inputGateBias"),
+ outputQuantizationParameters));
+ operatorInputs.push_back(buffers.size() - 1);
+ }
+ else
+ {
+ operatorInputs.push_back(kTfLiteOptionalTensor);
+ }
+
+ buffers.push_back(
+ CreateBuffer(flatBufferBuilder,
+ flatBufferBuilder.CreateVector(reinterpret_cast<const uint8_t *>(forgetGateBias.data()),
+ sizeof(T) * forgetGateBias.size())));
+ tensors.push_back(CreateTensor(flatBufferBuilder,
+ flatBufferBuilder.CreateVector<int32_t>(tensorInfo4.data(),
+ tensorInfo4.size()),
+ tensorType,
+ buffers.size() - 1,
+ flatBufferBuilder.CreateString("forgetGateBias"),
+ outputQuantizationParameters));
+ operatorInputs.push_back(buffers.size() - 1);
+
+ buffers.push_back(
+ CreateBuffer(flatBufferBuilder,
+ flatBufferBuilder.CreateVector(reinterpret_cast<const uint8_t *>(cellBias.data()),
+ sizeof(T) * cellBias.size())));
+ tensors.push_back(CreateTensor(flatBufferBuilder,
+ flatBufferBuilder.CreateVector<int32_t>(tensorInfo4.data(),
+ tensorInfo4.size()),
+ tensorType,
+ buffers.size() - 1,
+ flatBufferBuilder.CreateString("cellBias"),
+ outputQuantizationParameters));
+ operatorInputs.push_back(buffers.size() - 1);
+
+ buffers.push_back(
+ CreateBuffer(flatBufferBuilder,
+ flatBufferBuilder.CreateVector(reinterpret_cast<const uint8_t *>(outputGateBias.data()),
+ sizeof(T) * outputGateBias.size())));
+ tensors.push_back(CreateTensor(flatBufferBuilder,
+ flatBufferBuilder.CreateVector<int32_t>(tensorInfo4.data(),
+ tensorInfo4.size()),
+ tensorType,
+ buffers.size() - 1,
+ flatBufferBuilder.CreateString("outputGateBias"),
+ outputQuantizationParameters));
+ operatorInputs.push_back(buffers.size() - 1);
+
+ if (hasProjectionWeights)
+ {
+ buffers.push_back(
+ CreateBuffer(flatBufferBuilder,
+ flatBufferBuilder.CreateVector(reinterpret_cast<const uint8_t *>(projectionWeights.data()),
+ sizeof(T) * projectionWeights.size())));
+ tensors.push_back(CreateTensor(flatBufferBuilder,
+ flatBufferBuilder.CreateVector<int32_t>(tensorInfo4.data(),
+ tensorInfo4.size()),
+ tensorType,
+ buffers.size() - 1,
+ flatBufferBuilder.CreateString("outputGateBias"),
+ outputQuantizationParameters));
+ operatorInputs.push_back(buffers.size() - 1);
+ }
+ else
+ {
+ operatorInputs.push_back(kTfLiteOptionalTensor);
+ }
+
+ if (hasProjectionBias)
+ {
+ buffers.push_back(
+ CreateBuffer(flatBufferBuilder,
+ flatBufferBuilder.CreateVector(reinterpret_cast<const uint8_t *>(projectionBias.data()),
+ sizeof(T) * projectionBias.size())));
+ tensors.push_back(CreateTensor(flatBufferBuilder,
+ flatBufferBuilder.CreateVector<int32_t>(tensorInfo4.data(),
+ tensorInfo4.size()),
+ tensorType,
+ buffers.size() - 1,
+ flatBufferBuilder.CreateString("projectionBias"),
+ outputQuantizationParameters));
+ operatorInputs.push_back(buffers.size() - 1);
+ }
+ else
+ {
+ operatorInputs.push_back(kTfLiteOptionalTensor);
+ }
+
+ buffers.push_back(CreateBuffer(flatBufferBuilder, flatBufferBuilder.CreateVector({})));
+ tensors.push_back(CreateTensor(flatBufferBuilder,
+ flatBufferBuilder.CreateVector<int32_t>(outputStateInDimensions.data(),
+ outputStateInDimensions.size()),
+ tensorType,
+ buffers.size() - 1,
+ flatBufferBuilder.CreateString("outputStateInInfo"),
+ outputQuantizationParameters,
+ true));
+ operatorInputs.push_back(buffers.size() - 1);
+
+ buffers.push_back(CreateBuffer(flatBufferBuilder, flatBufferBuilder.CreateVector({})));
+ tensors.push_back(CreateTensor(flatBufferBuilder,
+ flatBufferBuilder.CreateVector<int32_t>(cellStateInDimensions.data(),
+ cellStateInDimensions.size()),
+ tensorType,
+ buffers.size() - 1,
+ flatBufferBuilder.CreateString("cellStateInInfo"),
+ outputQuantizationParameters,
+ true));
+ operatorInputs.push_back(buffers.size() - 1);
+
+ if (hasInputLayerNormWeights)
+ {
+ buffers.push_back(
+ CreateBuffer(flatBufferBuilder,
+ flatBufferBuilder.CreateVector(
+ reinterpret_cast<const uint8_t *>(inputLayerNormWeights.data()),
+ sizeof(T) * inputLayerNormWeights.size())));
+ tensors.push_back(CreateTensor(flatBufferBuilder,
+ flatBufferBuilder.CreateVector<int32_t>(tensorInfo4.data(),
+ tensorInfo4.size()),
+ tensorType,
+ buffers.size() - 1,
+ flatBufferBuilder.CreateString("inputLayerNormWeights"),
+ outputQuantizationParameters));
+ operatorInputs.push_back(buffers.size() - 1);
+ }
+ else
+ {
+ operatorInputs.push_back(kTfLiteOptionalTensor);
+ }
+
+ if (hasForgetLayerNormWeights)
+ {
+ buffers.push_back(
+ CreateBuffer(flatBufferBuilder,
+ flatBufferBuilder.CreateVector(
+ reinterpret_cast<const uint8_t *>(forgetLayerNormWeights.data()),
+ sizeof(T) * forgetLayerNormWeights.size())));
+ tensors.push_back(CreateTensor(flatBufferBuilder,
+ flatBufferBuilder.CreateVector<int32_t>(tensorInfo4.data(),
+ tensorInfo4.size()),
+ tensorType,
+ buffers.size() - 1,
+ flatBufferBuilder.CreateString("forgetLayerNormWeights"),
+ outputQuantizationParameters));
+ operatorInputs.push_back(buffers.size() - 1);
+ }
+ else
+ {
+ operatorInputs.push_back(kTfLiteOptionalTensor);
+ }
+
+ if (hasCellLayerNormWeights)
+ {
+ buffers.push_back(
+ CreateBuffer(flatBufferBuilder,
+ flatBufferBuilder.CreateVector(reinterpret_cast<const uint8_t *>(cellLayerNormWeights.data()),
+ sizeof(T) * cellLayerNormWeights.size())));
+ tensors.push_back(CreateTensor(flatBufferBuilder,
+ flatBufferBuilder.CreateVector<int32_t>(tensorInfo4.data(),
+ tensorInfo4.size()),
+ tensorType,
+ buffers.size() - 1,
+ flatBufferBuilder.CreateString("cellLayerNormWeights"),
+ outputQuantizationParameters));
+ operatorInputs.push_back(buffers.size() - 1);
+ }
+ else
+ {
+ operatorInputs.push_back(kTfLiteOptionalTensor);
+ }
+
+ if (hasOutputLayerNormWeights)
+ {
+ buffers.push_back(
+ CreateBuffer(flatBufferBuilder,
+ flatBufferBuilder.CreateVector(
+ reinterpret_cast<const uint8_t *>(outputLayerNormWeights.data()),
+ sizeof(T) * outputLayerNormWeights.size())));
+ tensors.push_back(CreateTensor(flatBufferBuilder,
+ flatBufferBuilder.CreateVector<int32_t>(tensorInfo4.data(),
+ tensorInfo4.size()),
+ tensorType,
+ buffers.size() - 1,
+ flatBufferBuilder.CreateString("outputLayerNormWeights"),
+ outputQuantizationParameters));
+ operatorInputs.push_back(buffers.size() - 1);
+ }
+ else
+ {
+ operatorInputs.push_back(kTfLiteOptionalTensor);
+ }
+ int outputBufferId = buffers.size();
+ buffers.push_back(CreateBuffer(flatBufferBuilder, flatBufferBuilder.CreateVector({})));
+ tensors.push_back(CreateTensor(flatBufferBuilder,
+ flatBufferBuilder.CreateVector<int32_t>(outputShape.data(),
+ outputShape.size()),
+ tensorType,
+ outputBufferId,
+ flatBufferBuilder.CreateString("output"),
+ outputQuantizationParameters));
+ std::vector<int> operatorOutputs;
+ operatorOutputs.push_back(buffers.size() - 1);
+
+ // create operator
+ tflite::BuiltinOptions operatorBuiltinOptionsType = BuiltinOptions_LSTMOptions;
+ flatbuffers::Offset<void> operatorBuiltinOptions =
+ CreateLSTMOptions(flatBufferBuilder,
+ activationFunction,
+ clippingThresCell,
+ clippingThresProj).Union();
+
+ flatbuffers::Offset <Operator> lstmOperator =
+ CreateOperator(flatBufferBuilder,
+ 0,
+ flatBufferBuilder.CreateVector<int32_t>(operatorInputs.data(), operatorInputs.size()),
+ flatBufferBuilder.CreateVector<int32_t>(operatorOutputs.data(), operatorOutputs.size()),
+ operatorBuiltinOptionsType, operatorBuiltinOptions);
+
+ flatbuffers::Offset <SubGraph> subgraph =
+ CreateSubGraph(flatBufferBuilder,
+ flatBufferBuilder.CreateVector(tensors.data(), tensors.size()),
+ flatBufferBuilder.CreateVector<int32_t>(operatorInputs.data(), operatorInputs.size()),
+ flatBufferBuilder.CreateVector<int32_t>(operatorOutputs.data(), operatorOutputs.size()),
+ flatBufferBuilder.CreateVector(&lstmOperator, 1));
+
+ flatbuffers::Offset <flatbuffers::String> modelDescription =
+ flatBufferBuilder.CreateString("ArmnnDelegate: LSTM Operator Model");
+ flatbuffers::Offset <OperatorCode> operatorCode = CreateOperatorCode(flatBufferBuilder,
+ tflite::BuiltinOperator_LSTM);
+
+ flatbuffers::Offset <Model> flatbufferModel =
+ CreateModel(flatBufferBuilder,
+ TFLITE_SCHEMA_VERSION,
+ flatBufferBuilder.CreateVector(&operatorCode, 1),
+ flatBufferBuilder.CreateVector(&subgraph, 1),
+ modelDescription,
+ flatBufferBuilder.CreateVector(buffers.data(), buffers.size()));
+
+ flatBufferBuilder.Finish(flatbufferModel);
+
+ return std::vector<char>(flatBufferBuilder.GetBufferPointer(),
+ flatBufferBuilder.GetBufferPointer() + flatBufferBuilder.GetSize());
+}
+
+template <typename T>
+void LstmTestImpl(std::vector<armnn::BackendId>& backends,
+ tflite::TensorType tensorType,
+ int32_t batchSize,
+ int32_t inputSize,
+ int32_t outputSize,
+ int32_t numUnits,
+ bool hasInputToInputWeights,
+ const std::vector<T>& inputToInputWeights,
+ const std::vector<T>& inputToForgetWeights,
+ const std::vector<T>& inputToCellWeights,
+ const std::vector<T>& inputToOutputWeights,
+ bool hasRecurrentToInputWeights,
+ const std::vector<T>& recurrentToInputWeights,
+ const std::vector<T>& recurrentToForgetWeights,
+ const std::vector<T>& recurrentToCellWeights,
+ const std::vector<T>& recurrentToOutputWeights,
+ bool hasCellToInputWeights,
+ const std::vector<T>& cellToInputWeights,
+ bool hasCellToForgetWeights,
+ const std::vector<T>& cellToForgetWeights,
+ bool hasCellToOutputWeights,
+ const std::vector<T>& cellToOutputWeights,
+ bool hasInputGateBias,
+ const std::vector<T>& inputGateBias,
+ const std::vector<T>& forgetGateBias,
+ const std::vector<T>& cellBias,
+ const std::vector<T>& outputGateBias,
+ bool hasProjectionWeights,
+ const std::vector<T>& projectionWeights,
+ bool hasProjectionBias,
+ const std::vector<T>& projectionBias,
+ bool hasInputLayerNormWeights,
+ const std::vector<T>& inputLayerNormWeights,
+ bool hasForgetLayerNormWeights,
+ const std::vector<T>& forgetLayerNormWeights,
+ bool hasCellLayerNormWeights,
+ const std::vector<T>& cellLayerNormWeights,
+ bool hasOutputLayerNormWeights,
+ const std::vector<T>& outputLayerNormWeights,
+ std::vector<T>& inputValues,
+ std::vector<T>& expectedOutputValues,
+ tflite::ActivationFunctionType activationFunction,
+ float clippingThresCell,
+ float clippingThresProj)
+{
+ using namespace tflite;
+
+ std::vector<char> modelBuffer = CreateLstmTfLiteModel(tensorType,
+ batchSize,
+ inputSize,
+ outputSize,
+ numUnits,
+ hasInputToInputWeights,
+ inputToInputWeights,
+ inputToForgetWeights,
+ inputToCellWeights,
+ inputToOutputWeights,
+ hasRecurrentToInputWeights,
+ recurrentToInputWeights,
+ recurrentToForgetWeights,
+ recurrentToCellWeights,
+ recurrentToOutputWeights,
+ hasCellToInputWeights,
+ cellToInputWeights,
+ hasCellToForgetWeights,
+ cellToForgetWeights,
+ hasCellToOutputWeights,
+ cellToOutputWeights,
+ hasInputGateBias,
+ inputGateBias,
+ forgetGateBias,
+ cellBias,
+ outputGateBias,
+ hasProjectionWeights,
+ projectionWeights,
+ hasProjectionBias,
+ projectionBias,
+ hasInputLayerNormWeights,
+ inputLayerNormWeights,
+ hasForgetLayerNormWeights,
+ forgetLayerNormWeights,
+ hasCellLayerNormWeights,
+ cellLayerNormWeights,
+ hasOutputLayerNormWeights,
+ outputLayerNormWeights,
+ activationFunction,
+ clippingThresCell,
+ clippingThresProj);
+
+ const Model* tfLiteModel = GetModel(modelBuffer.data());
+ // Create TfLite Interpreters
+ std::unique_ptr<Interpreter> armnnDelegateInterpreter;
+ CHECK(InterpreterBuilder(tfLiteModel, ::tflite::ops::builtin::BuiltinOpResolver())
+ (&armnnDelegateInterpreter) == kTfLiteOk);
+ CHECK(armnnDelegateInterpreter != nullptr);
+ CHECK(armnnDelegateInterpreter->AllocateTensors() == kTfLiteOk);
+
+ std::unique_ptr<Interpreter> tfLiteInterpreter;
+ CHECK(InterpreterBuilder(tfLiteModel, ::tflite::ops::builtin::BuiltinOpResolver())
+ (&tfLiteInterpreter) == kTfLiteOk);
+ CHECK(tfLiteInterpreter != nullptr);
+ CHECK(tfLiteInterpreter->AllocateTensors() == kTfLiteOk);
+
+ // Create the ArmNN Delegate
+ armnnDelegate::DelegateOptions delegateOptions(backends);
+ std::unique_ptr<TfLiteDelegate, decltype(&armnnDelegate::TfLiteArmnnDelegateDelete)>
+ theArmnnDelegate(armnnDelegate::TfLiteArmnnDelegateCreate(delegateOptions),
+ armnnDelegate::TfLiteArmnnDelegateDelete);
+ CHECK(theArmnnDelegate != nullptr);
+ // Modify armnnDelegateInterpreter to use armnnDelegate
+ CHECK(armnnDelegateInterpreter->ModifyGraphWithDelegate(theArmnnDelegate.get()) == kTfLiteOk);
+
+ // Set input data
+ auto tfLiteDelegateInputId = tfLiteInterpreter->inputs()[0];
+ auto tfLiteDelageInputData = tfLiteInterpreter->typed_tensor<T>(tfLiteDelegateInputId);
+ for (unsigned int i = 0; i < inputValues.size(); ++i)
+ {
+ tfLiteDelageInputData[i] = inputValues[i];
+ }
+
+ auto armnnDelegateInputId = armnnDelegateInterpreter->inputs()[0];
+ auto armnnDelegateInputData = armnnDelegateInterpreter->typed_tensor<T>(armnnDelegateInputId);
+ for (unsigned int i = 0; i < inputValues.size(); ++i)
+ {
+ armnnDelegateInputData[i] = inputValues[i];
+ }
+
+ // Run EnqueWorkload
+ CHECK(tfLiteInterpreter->Invoke() == kTfLiteOk);
+ CHECK(armnnDelegateInterpreter->Invoke() == kTfLiteOk);
+
+ // Compare output data
+ auto tfLiteDelegateOutputId = tfLiteInterpreter->outputs()[0];
+ auto tfLiteDelagateOutputData = tfLiteInterpreter->typed_tensor<T>(tfLiteDelegateOutputId);
+ auto armnnDelegateOutputId = armnnDelegateInterpreter->outputs()[0];
+ auto armnnDelegateOutputData = armnnDelegateInterpreter->typed_tensor<T>(armnnDelegateOutputId);
+
+ armnnDelegate::CompareData(expectedOutputValues.data(), armnnDelegateOutputData, expectedOutputValues.size());
+ armnnDelegate::CompareData(expectedOutputValues.data(), tfLiteDelagateOutputData, expectedOutputValues.size());
+ armnnDelegate::CompareData(tfLiteDelagateOutputData, armnnDelegateOutputData, expectedOutputValues.size());
+}
+
+} // anonymous namespace \ No newline at end of file
generated by cgit v1.2.1 (git 2.23.0) at 2024-06-21 10:44:09 +0000