IVGCVSW-4193 Add QUANTIZE to the TfLite parser

Signed-off-by: Sadik Armagan <sadik.armagan@arm.com>
Change-Id: I53431c70edb64c9283ae556407cc07e54469d8cc

3 files changed, 155 insertions, 8 deletions

diff --git a/src/armnnTfLiteParser/TfLiteParser.cpp b/src/armnnTfLiteParser/TfLiteParser.cpp
index 22d65645a3..69bd59a21f 100644
--- a/src/armnnTfLiteParser/TfLiteParser.cpp
+++ b/src/armnnTfLiteParser/TfLiteParser.cpp
@@ -442,36 +442,37 @@ TfLiteParser::TfLiteParser(const Optional<ITfLiteParser::TfLiteParserOptions>& o
 , m_ParserFunctions(tflite::BuiltinOperator_MAX+1, &TfLiteParser::ParseUnsupportedOperator)
 {
     // register supported operators
+    m_ParserFunctions[tflite::BuiltinOperator_ADD]                     = &TfLiteParser::ParseAdd;
     m_ParserFunctions[tflite::BuiltinOperator_AVERAGE_POOL_2D]         = &TfLiteParser::ParseAveragePool2D;
     m_ParserFunctions[tflite::BuiltinOperator_BATCH_TO_SPACE_ND]       = &TfLiteParser::ParseBatchToSpaceND;
     m_ParserFunctions[tflite::BuiltinOperator_CONCATENATION]           = &TfLiteParser::ParseConcatenation;
     m_ParserFunctions[tflite::BuiltinOperator_CONV_2D]                 = &TfLiteParser::ParseConv2D;
+    m_ParserFunctions[tflite::BuiltinOperator_CUSTOM]                  = &TfLiteParser::ParseCustomOperator;
     m_ParserFunctions[tflite::BuiltinOperator_DEPTHWISE_CONV_2D]       = &TfLiteParser::ParseDepthwiseConv2D;
     m_ParserFunctions[tflite::BuiltinOperator_DEQUANTIZE]              = &TfLiteParser::ParseDequantize;
-    m_ParserFunctions[tflite::BuiltinOperator_CUSTOM]                  = &TfLiteParser::ParseCustomOperator;
     m_ParserFunctions[tflite::BuiltinOperator_FULLY_CONNECTED]         = &TfLiteParser::ParseFullyConnected;
     m_ParserFunctions[tflite::BuiltinOperator_LOGISTIC]                = &TfLiteParser::ParseLogistic;
     m_ParserFunctions[tflite::BuiltinOperator_L2_NORMALIZATION]        = &TfLiteParser::ParseL2Normalization;
     m_ParserFunctions[tflite::BuiltinOperator_MAX_POOL_2D]             = &TfLiteParser::ParseMaxPool2D;
     m_ParserFunctions[tflite::BuiltinOperator_MAXIMUM]                 = &TfLiteParser::ParseMaximum;
+    m_ParserFunctions[tflite::BuiltinOperator_MEAN]                    = &TfLiteParser::ParseMean;
     m_ParserFunctions[tflite::BuiltinOperator_MINIMUM]                 = &TfLiteParser::ParseMinimum;
+    m_ParserFunctions[tflite::BuiltinOperator_MUL]                     = &TfLiteParser::ParseMul;
+    m_ParserFunctions[tflite::BuiltinOperator_PACK]                    = &TfLiteParser::ParsePack;
+    m_ParserFunctions[tflite::BuiltinOperator_PAD]                     = &TfLiteParser::ParsePad;
+    m_ParserFunctions[tflite::BuiltinOperator_QUANTIZE]                = &TfLiteParser::ParseQuantize;
     m_ParserFunctions[tflite::BuiltinOperator_RELU]                    = &TfLiteParser::ParseRelu;
     m_ParserFunctions[tflite::BuiltinOperator_RELU6]                   = &TfLiteParser::ParseRelu6;
     m_ParserFunctions[tflite::BuiltinOperator_RESHAPE]                 = &TfLiteParser::ParseReshape;
     m_ParserFunctions[tflite::BuiltinOperator_RESIZE_BILINEAR]         = &TfLiteParser::ParseResizeBilinear;
     m_ParserFunctions[tflite::BuiltinOperator_RESIZE_NEAREST_NEIGHBOR] = &TfLiteParser::ParseResizeNearestNeighbor;
+    m_ParserFunctions[tflite::BuiltinOperator_SLICE]                   = &TfLiteParser::ParseSlice;
     m_ParserFunctions[tflite::BuiltinOperator_SOFTMAX]                 = &TfLiteParser::ParseSoftmax;
     m_ParserFunctions[tflite::BuiltinOperator_SPACE_TO_BATCH_ND]       = &TfLiteParser::ParseSpaceToBatchND;
+    m_ParserFunctions[tflite::BuiltinOperator_SPLIT]                   = &TfLiteParser::ParseSplit;
     m_ParserFunctions[tflite::BuiltinOperator_SQUEEZE]                 = &TfLiteParser::ParseSqueeze;
     m_ParserFunctions[tflite::BuiltinOperator_STRIDED_SLICE]           = &TfLiteParser::ParseStridedSlice;
     m_ParserFunctions[tflite::BuiltinOperator_SUB]                     = &TfLiteParser::ParseSub;
-    m_ParserFunctions[tflite::BuiltinOperator_ADD]                     = &TfLiteParser::ParseAdd;
-    m_ParserFunctions[tflite::BuiltinOperator_MUL]                     = &TfLiteParser::ParseMul;
-    m_ParserFunctions[tflite::BuiltinOperator_MEAN]                    = &TfLiteParser::ParseMean;
-    m_ParserFunctions[tflite::BuiltinOperator_PACK]                    = &TfLiteParser::ParsePack;
-    m_ParserFunctions[tflite::BuiltinOperator_PAD]                     = &TfLiteParser::ParsePad;
-    m_ParserFunctions[tflite::BuiltinOperator_SLICE]                   = &TfLiteParser::ParseSlice;
-    m_ParserFunctions[tflite::BuiltinOperator_SPLIT]                   = &TfLiteParser::ParseSplit;
     m_ParserFunctions[tflite::BuiltinOperator_TANH]                    = &TfLiteParser::ParseTanH;
     m_ParserFunctions[tflite::BuiltinOperator_TRANSPOSE]               = &TfLiteParser::ParseTranspose;
     m_ParserFunctions[tflite::BuiltinOperator_TRANSPOSE_CONV]          = &TfLiteParser::ParseTransposeConv;
@@ -1746,6 +1747,30 @@ void TfLiteParser::ParsePad(size_t subgraphIndex, size_t operatorIndex)
     RegisterOutputSlots(subgraphIndex, operatorIndex, layer, {outputTensorIndexes[0]});
 }
 
+void TfLiteParser::ParseQuantize(size_t subgraphIndex, size_t operatorIndex)
+{
+    CHECK_MODEL(m_Model, subgraphIndex, operatorIndex);
+
+    auto inputs = GetInputs(m_Model, subgraphIndex, operatorIndex);
+    CHECK_VALID_SIZE(inputs.size(), 1);
+
+    auto outputs = GetOutputs(m_Model, subgraphIndex, operatorIndex);
+    CHECK_VALID_SIZE(outputs.size(), 1);
+
+    auto layerName = boost::str(boost::format("Quantize:%1%:%2%") % subgraphIndex % operatorIndex);
+
+    IConnectableLayer* layer = m_Network->AddQuantizeLayer(layerName.c_str());
+    BOOST_ASSERT(layer != nullptr);
+
+    TensorInfo outputTensorInfo = ToTensorInfo(outputs[0]);
+    layer->GetOutputSlot(0).SetTensorInfo(outputTensorInfo);
+
+    auto inputTensorIndexes = AsUnsignedVector(GetInputTensorIds(m_Model, subgraphIndex, operatorIndex));
+    RegisterInputSlots(subgraphIndex, operatorIndex, layer, {inputTensorIndexes[0]});
+
+    auto outputTensorIndexes = AsUnsignedVector(GetOutputTensorIds(m_Model, subgraphIndex, operatorIndex));
+    RegisterOutputSlots(subgraphIndex, operatorIndex, layer, outputTensorIndexes);
+}
 
 void TfLiteParser::ParseRelu(size_t subgraphIndex, size_t operatorIndex)
 {
diff --git a/src/armnnTfLiteParser/TfLiteParser.hpp b/src/armnnTfLiteParser/TfLiteParser.hpp
index a8241f6650..42ea1a0372 100644
--- a/src/armnnTfLiteParser/TfLiteParser.hpp
+++ b/src/armnnTfLiteParser/TfLiteParser.hpp
@@ -113,6 +113,7 @@ private:
     void ParsePack(size_t subgraphIndex, size_t operatorIndex);
     void ParsePad(size_t subgraphIndex, size_t operatorIndex);
     void ParsePool(size_t subgraphIndex, size_t operatorIndex, armnn::PoolingAlgorithm algorithm);
+    void ParseQuantize(size_t subgraphIndex, size_t operatorIndex);
     void ParseRelu(size_t subgraphIndex, size_t operatorIndex);
     void ParseRelu6(size_t subgraphIndex, size_t operatorIndex);
     void ParseReshape(size_t subgraphIndex, size_t operatorIndex);
diff --git a/src/armnnTfLiteParser/test/Quantize.cpp b/src/armnnTfLiteParser/test/Quantize.cpp
new file mode 100644
index 0000000000..ca5e6d5091
--- /dev/null
+++ b/src/armnnTfLiteParser/test/Quantize.cpp
@@ -0,0 +1,121 @@
+//
+// Copyright © 2019 Arm Ltd. All rights reserved.
+// SPDX-License-Identifier: MIT
+//
+
+#include <boost/test/unit_test.hpp>
+#include "ParserFlatbuffersFixture.hpp"
+#include "../TfLiteParser.hpp"
+
+#include <string>
+#include <iostream>
+
+BOOST_AUTO_TEST_SUITE(TensorflowLiteParser)
+
+    struct QuantizeFixture : public ParserFlatbuffersFixture
+    {
+        explicit QuantizeFixture(const std::string & inputShape,
+                                 const std::string & outputShape,
+                                 const std::string & dataType)
+        {
+            m_JsonString = R"(
+            {
+                "version": 3,
+                "operator_codes": [ { "builtin_code": "QUANTIZE" } ],
+                "subgraphs": [ {
+                    "tensors": [
+                        {
+                            "shape": )" + inputShape + R"(,
+                            "type": "FLOAT32",
+                            "buffer": 0,
+                            "name": "inputTensor",
+                            "quantization": {
+                                "min": [ 0.0 ],
+                                "max": [ 255.0 ],
+                                "scale": [ 1.0 ],
+                                "zero_point": [ 0 ],
+                            }
+                        },
+                        {
+                            "shape": )" + outputShape + R"( ,
+                            "type": )" + dataType + R"(,
+                            "buffer": 1,
+                            "name": "outputTensor",
+                            "quantization": {
+                                "min": [ 0.0 ],
+                                "max": [ 255.0 ],
+                                "scale": [ 1.5 ],
+                                "zero_point": [ 0 ],
+                            }
+                        }
+                    ],
+                    "inputs": [ 0 ],
+                    "outputs": [ 1 ],
+                    "operators": [
+                        {
+                            "opcode_index": 0,
+                            "inputs": [ 0 ],
+                            "outputs": [ 1 ],
+                            "builtin_options_type": "QuantizeOptions",
+                            "builtin_options": {
+                            },
+                            "custom_options_format": "FLEXBUFFERS"
+                        }
+                ],
+                } ],
+                "buffers" : [
+                    { },
+                    { },
+                ]
+            }
+        )";
+            SetupSingleInputSingleOutput("inputTensor", "outputTensor");
+        }
+    };
+
+    struct SimpleQuantizeFixtureQAsymm8 : QuantizeFixture
+    {
+        SimpleQuantizeFixtureQAsymm8() : QuantizeFixture("[ 1, 6 ]",
+                                                         "[ 1, 6 ]",
+                                                         "UINT8") {}
+    };
+
+    BOOST_FIXTURE_TEST_CASE(SimpleQuantizeQAsymm8, SimpleQuantizeFixtureQAsymm8)
+    {
+        RunTest<2, armnn::DataType::Float32, armnn::DataType::QuantisedAsymm8>(
+                0,
+                {{"inputTensor",  { 0.0f, 1.5f, 7.5f, 150.0f, 300.0f, 382.5f }}},
+                {{"outputTensor", { 0u,   1u,   5u,   100u,   200u,   255u }}});
+    }
+
+    struct SimpleQuantizeFixtureQSymm16 : QuantizeFixture
+    {
+        SimpleQuantizeFixtureQSymm16() : QuantizeFixture("[ 1, 6 ]",
+                                                         "[ 1, 6 ]",
+                                                         "INT16") {}
+    };
+
+    BOOST_FIXTURE_TEST_CASE(SimpleQuantizeQsymm16, SimpleQuantizeFixtureQSymm16)
+    {
+        RunTest<2, armnn::DataType::Float32, armnn::DataType::QuantisedSymm16>(
+                0,
+                {{"inputTensor",  { 0.0f, 1.5f, 7.5f, 49150.5f, -1.5f,-49152.0f }}},
+                {{"outputTensor", { 0,    1,    5,    32767,    -1,   -32768 }}});
+    }
+
+    struct SimpleQuantizeFixtureQSymmS8 : QuantizeFixture
+    {
+        SimpleQuantizeFixtureQSymmS8() : QuantizeFixture("[ 1, 6 ]",
+                                                         "[ 1, 6 ]",
+                                                         "INT8") {}
+    };
+
+    BOOST_FIXTURE_TEST_CASE(SimpleQuantizeQSymmS8, SimpleQuantizeFixtureQSymmS8)
+    {
+        RunTest<2, armnn::DataType::Float32, armnn::DataType::QSymmS8>(
+                0,
+                {{"inputTensor",  { 0.0f, 1.5f, 7.5f, 190.5f, -192.0f, -1.5f }}},
+                {{"outputTensor", { 0,    1,    5,    127,    -128,   -1 }}});
+    }
+
+BOOST_AUTO_TEST_SUITE_END()