IVGCVSW-5741 Update FullyConnected in TfLiteParser to support NonConstWeights

!armnn:5180

 * Remove unnecessary memcopy for non permeuted const tensors

Signed-off-by: Finn Williams <Finn.Williams@arm.com>
Change-Id: Idc3ce2ac001e7d6be61819279de486f093730383

1 files changed, 171 insertions, 6 deletions

diff --git a/src/armnnTfLiteParser/test/FullyConnected.cpp b/src/armnnTfLiteParser/test/FullyConnected.cpp
index e7aa9082e2..333e17fafd 100644
--- a/src/armnnTfLiteParser/test/FullyConnected.cpp
+++ b/src/armnnTfLiteParser/test/FullyConnected.cpp
@@ -8,18 +8,17 @@
 #include "../TfLiteParser.hpp"
 
 #include <string>
-#include <iostream>
 
 BOOST_AUTO_TEST_SUITE(TensorflowLiteParser)
 
 struct FullyConnectedFixture : public ParserFlatbuffersFixture
 {
     explicit FullyConnectedFixture(const std::string& inputShape,
-                                           const std::string& outputShape,
-                                           const std::string& filterShape,
-                                           const std::string& filterData,
-                                           const std::string biasShape = "",
-                                           const std::string biasData = "")
+                                   const std::string& outputShape,
+                                   const std::string& filterShape,
+                                   const std::string& filterData,
+                                   const std::string biasShape = "",
+                                   const std::string biasData = "")
     {
         std::string inputTensors = "[ 0, 2 ]";
         std::string biasTensor = "";
@@ -195,4 +194,170 @@ BOOST_FIXTURE_TEST_CASE(
                                       true);
 }
 
+
+struct FullyConnectedNonConstWeightsFixture : public ParserFlatbuffersFixture
+{
+    explicit FullyConnectedNonConstWeightsFixture(const std::string& inputShape,
+                                                  const std::string& outputShape,
+                                                  const std::string& filterShape,
+                                                  const std::string biasShape = "")
+    {
+        std::string inputTensors = "[ 0, 1 ]";
+        std::string biasTensor = "";
+        std::string biasBuffer = "";
+        std::string outputs = "2";
+        if (biasShape.size() > 0)
+        {
+            inputTensors = "[ 0, 1, 2 ]";
+            biasTensor = R"(
+                       {
+                      "shape": )" + biasShape + R"(,
+                      "type": "INT32",
+                      "buffer": 2,
+                      "name": "bias",
+                      "quantization": {
+                        "scale": [ 1.0 ],
+                        "zero_point": [ 0 ],
+                        "details_type": 0,
+                        "quantized_dimension": 0
+                      },
+                      "is_variable": true
+                    }, )";
+
+            biasBuffer = R"(,{ "data": [ 10, 0, 0, 0 ] } )";
+            outputs = "3";
+        }
+        m_JsonString = R"(
+            {
+              "version": 3,
+              "operator_codes": [
+                {
+                  "builtin_code": "FULLY_CONNECTED",
+                  "version": 1
+                }
+              ],
+              "subgraphs": [
+                {
+                  "tensors": [
+                    {
+                      "shape": )" + inputShape + R"(,
+                      "type": "INT8",
+                      "buffer": 0,
+                      "name": "input_0",
+                      "quantization": {
+                        "scale": [ 1.0 ],
+                        "zero_point": [ 0 ],
+                        "details_type": 0,
+                        "quantized_dimension": 0
+                      },
+                      "is_variable": false
+                    },
+                    {
+                      "shape": )" + filterShape + R"(,
+                      "type": "INT8",
+                      "buffer": 1,
+                      "name": "weights",
+                      "quantization": {
+                        "scale": [ 1.0 ],
+                        "zero_point": [ 0 ],
+                        "details_type": 0,
+                        "quantized_dimension": 0
+                      },
+                      "is_variable": true
+                    },
+                    )" + biasTensor + R"(
+                    {
+                      "shape": )" + outputShape + R"(,
+                      "type": "INT8",
+                      "buffer": 0,
+                      "name": "output",
+                      "quantization": {
+                        "scale": [
+                          2.0
+                        ],
+                        "zero_point": [
+                          0
+                        ],
+                        "details_type": 0,
+                        "quantized_dimension": 0
+                      },
+                      "is_variable": false
+                    }
+                  ],
+                  "inputs": )" + inputTensors + R"(,
+                  "outputs": [ )" + outputs + R"( ],
+                  "operators": [
+                    {
+                      "opcode_index": 0,
+                      "inputs": )" + inputTensors + R"(,
+                      "outputs": [ )" + outputs + R"( ],
+                      "builtin_options_type": "FullyConnectedOptions",
+                      "builtin_options": {
+                        "fused_activation_function": "NONE",
+                        "weights_format": "DEFAULT",
+                        "keep_num_dims": false,
+                        "asymmetric_quantize_inputs": false
+                      },
+                      "custom_options_format": "FLEXBUFFERS"
+                    }
+                  ]
+                }
+              ],
+              "description": "ArmnnDelegate: FullyConnected Operator Model",
+              "buffers": [
+                {
+                  "data": []
+                },
+                {
+                  "data": [ 2, 3, 4, 5 ]
+                }
+                )" + biasBuffer + R"(
+              ]
+            }
+            )";
+        Setup();
+    }
+};
+
+struct FullyConnectedNonConstWeights : FullyConnectedNonConstWeightsFixture
+{
+    FullyConnectedNonConstWeights()
+            : FullyConnectedNonConstWeightsFixture("[ 1, 4, 1, 1 ]",     // inputShape
+                                                   "[ 1, 1 ]",           // outputShape
+                                                   "[ 1, 4 ]",           // filterShape
+                                                   "[ 1 ]" )             // biasShape
+
+    {}
+};
+
+BOOST_FIXTURE_TEST_CASE(ParseFullyConnectedNonConstWeights, FullyConnectedNonConstWeights)
+{
+    RunTest<2, armnn::DataType::QAsymmS8,
+            armnn::DataType::Signed32,
+            armnn::DataType::QAsymmS8>(
+            0,
+            {{{"input_0", { 1, 2, 3, 4 }},{"weights", { 2, 3, 4, 5 }}}},
+            {{"bias", { 10 }}},
+            {{"output", { 25 }}});
+}
+
+struct FullyConnectedNonConstWeightsNoBias : FullyConnectedNonConstWeightsFixture
+{
+    FullyConnectedNonConstWeightsNoBias()
+            : FullyConnectedNonConstWeightsFixture("[ 1, 4, 1, 1 ]",     // inputShape
+                                                   "[ 1, 1 ]",           // outputShape
+                                                   "[ 1, 4 ]")           // filterShape
+
+    {}
+};
+
+BOOST_FIXTURE_TEST_CASE(ParseFullyConnectedNonConstWeightsNoBias, FullyConnectedNonConstWeightsNoBias)
+{
+    RunTest<2, armnn::DataType::QAsymmS8,
+            armnn::DataType::QAsymmS8>(
+            0,
+            {{{"input_0", { 1, 2, 3, 4 }},{"weights", { 2, 3, 4, 5 }}}},
+            {{"output", { 20 }}});
+}
+
 BOOST_AUTO_TEST_SUITE_END()