IVGCVSW-7214 Disable BF16-Turbo-Mode and remove conversion layers

 - Remove Bf16ToFp32 Conversion Layer
 - Remove Fp32ToBf16 Conversion Layer
 - Remove B16 Conversion tests
 * Throw exception if m_ReduceFp32ToBf16 optimzer option is set to true
 * Provide comments to enable fast math in order to use bf16
 * Update docs to inform users to enable fast math for bf16

 Execute Network Changes
 * Require bf16_turbo_mode to also have fast_math_enabled set to true
 - Remove setting m_ReduceFp32ToBf16 optimizer option

Signed-off-by: Ryan OShea <ryan.oshea3@arm.com>
Change-Id: Ibaa6da9d29c96a1ce32ff5196b0847fde9f04a1c

5 files changed, 1 insertions, 141 deletions

diff --git a/delegate/include/DelegateOptions.hpp b/delegate/include/DelegateOptions.hpp
index 2b0107e834..4f157db73b 100644
--- a/delegate/include/DelegateOptions.hpp
+++ b/delegate/include/DelegateOptions.hpp
@@ -113,7 +113,7 @@ public:
      *
      *    Option key: "reduce-fp32-to-bf16" \n
      *    Possible values: ["true"/"false"] \n
-     *    Description: Reduce Fp32 data to Bf16 for faster processing
+     *    Description: This option is currently ignored. Please enable Fast Math in the CpuAcc or GpuAcc backends.
      *
      *    Option key: "debug-data" \n
      *    Possible values: ["true"/"false"] \n
diff --git a/delegate/python/test/test_external_delegate.py b/delegate/python/test/test_external_delegate.py
index a8dd8e6d3e..fe58d57e90 100644
--- a/delegate/python/test/test_external_delegate.py
+++ b/delegate/python/test/test_external_delegate.py
@@ -218,39 +218,6 @@ def test_external_delegate_options_fp32_to_fp16(capfd, delegate_dir, test_data_f
     assert "convert_fp32_to_fp16" in captured.out
     assert "convert_fp16_to_fp32" in captured.out
 
-def test_external_delegate_options_fp32_to_bf16(capfd, delegate_dir, test_data_folder):
-    # create armnn delegate with reduce-fp32-to-bf16 option
-    armnn_delegate = tflite.load_delegate(delegate_dir, options = {"backends": "CpuRef",
-                                                                   "debug-data": "1",
-                                                                   "reduce-fp32-to-bf16": "1"})
-
-    model_file_name = "conv2d.tflite"
-
-    inputShape = [ 1, 5, 5, 1 ]
-    outputShape = [ 1, 3, 3, 1 ]
-
-    inputValues = [ 1, 5, 2, 3, 5,
-                    8, 7, 3, 6, 3,
-                    3, 3, 9, 1, 9,
-                    4, 1, 8, 1, 3,
-                    6, 8, 1, 9, 2 ]
-
-    expectedResult = [ 28, 38, 29,
-                       96, 104, 53,
-                       31, 55, 24 ]
-
-    input = np.array(inputValues, dtype=np.float32).reshape(inputShape)
-    expected_output = np.array(expectedResult, dtype=np.float32).reshape(outputShape)
-
-    # run the inference
-    armnn_outputs = run_inference(test_data_folder, model_file_name, [input], [armnn_delegate])
-
-    # check results
-    compare_outputs(armnn_outputs, [expected_output])
-
-    captured = capfd.readouterr()
-    assert "convert_fp32_to_bf16" in captured.out
-
 def test_external_delegate_options_memory_import(delegate_dir, test_data_folder):
     # create armnn delegate with memory-import option
     armnn_delegate = tflite.load_delegate(delegate_dir, options = {"backends": "CpuAcc,CpuRef",
diff --git a/delegate/src/DelegateOptions.cpp b/delegate/src/DelegateOptions.cpp
index a55a579333..bb1edab009 100644
--- a/delegate/src/DelegateOptions.cpp
+++ b/delegate/src/DelegateOptions.cpp
@@ -146,11 +146,6 @@ DelegateOptions::DelegateOptions(char const* const* options_keys,
         {
             optimizerOptions.m_ReduceFp32ToFp16 = armnn::stringUtils::StringToBool(options_values[i]);
         }
-            // Process reduce-fp32-to-bf16 option
-        else if (std::string(options_keys[i]) == std::string("reduce-fp32-to-bf16"))
-        {
-            optimizerOptions.m_ReduceFp32ToBf16 = armnn::stringUtils::StringToBool(options_values[i]);
-        }
             // Process debug-data
         else if (std::string(options_keys[i]) == std::string("debug-data"))
         {
diff --git a/delegate/src/test/DelegateOptionsTest.cpp b/delegate/src/test/DelegateOptionsTest.cpp
index 50d3f78563..98323131f9 100644
--- a/delegate/src/test/DelegateOptionsTest.cpp
+++ b/delegate/src/test/DelegateOptionsTest.cpp
@@ -116,19 +116,6 @@ TEST_CASE ("ArmnnDelegateOptimizerOptionsDebugFunction")
     CHECK(callback);
 }
 
-TEST_CASE ("ArmnnDelegateOptimizerOptionsReduceFp32ToBf16")
-{
-    std::stringstream ss;
-    {
-        StreamRedirector redirect(std::cout, ss.rdbuf());
-
-        ReduceFp32ToBf16TestImpl();
-    }
-
-    // ReduceFp32ToBf16 option is enabled
-    CHECK(ss.str().find("convert_fp32_to_bf16") != std::string::npos);
-}
-
 TEST_CASE ("ArmnnDelegateOptimizerOptionsImport")
 {
     std::vector<armnn::BackendId> backends = {  armnn::Compute::CpuAcc, armnn::Compute::CpuRef };
diff --git a/delegate/src/test/DelegateOptionsTestHelper.hpp b/delegate/src/test/DelegateOptionsTestHelper.hpp
index 87bf0d6c3d..7e147de31f 100644
--- a/delegate/src/test/DelegateOptionsTestHelper.hpp
+++ b/delegate/src/test/DelegateOptionsTestHelper.hpp
@@ -219,95 +219,6 @@ std::vector<char> CreateCeilTfLiteModel(tflite::TensorType tensorType,
                              flatBufferBuilder.GetBufferPointer() + flatBufferBuilder.GetSize());
 }
 
-void ReduceFp32ToBf16TestImpl()
-{
-    using namespace tflite;
-    // Set input data
-    std::vector<int32_t> inputShape{ 1, 5, 5, 1 };
-    std::vector<int32_t> filterShape{ 1, 3, 3, 1 };
-    std::vector<int32_t> biasShape{ 1 };
-    std::vector<int32_t> outputShape{ 1, 3, 3, 1 };
-
-    std::vector<float> inputValues =
-        {
-            1, 5, 2, 3, 5,
-            8, 7, 3, 6, 3,
-            3, 3, 9, 1, 9,
-            4, 1, 8, 1, 3,
-            6, 8, 1, 9, 2
-        };
-
-    std::vector<float> filterValues =
-        {
-            4, 5, 6,
-            0, 0, 0,
-            3, 2, 1
-        };
-
-    std::vector<float> biasValues = { 5 };
-
-    std::vector<float> expectedResult =
-        {
-            28, 38, 29,
-            96, 104, 53,
-            31, 55, 24
-        };
-
-    tflite::Padding padding = Padding_SAME;
-
-    std::vector<char> modelBuffer;
-    modelBuffer = CreateConv2dTfLiteModel<float>(BuiltinOperator_CONV_2D,
-                                                 ::tflite::TensorType_FLOAT32,
-                                                 2,
-                                                 2,
-                                                 1,
-                                                 1,
-                                                 padding,
-                                                 ActivationFunctionType_NONE,
-                                                 inputShape,
-                                                 filterShape,
-                                                 biasShape,
-                                                 outputShape,
-                                                 filterValues,
-                                                 biasValues);
-
-
-    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);
-
-    // Create the Armnn Delegate
-    std::vector<armnn::BackendId> backends = {armnn::Compute::CpuRef};
-    std::vector<armnn::BackendOptions> backendOptions;
-
-    // Enable debug with BF16 enabled
-    armnn::OptimizerOptions optimizerOptions(false, true, true, false);
-
-    armnnDelegate::DelegateOptions delegateOptions(backends, optimizerOptions);
-    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
-    armnnDelegate::FillInput(armnnDelegateInterpreter, 0, inputValues);
-
-    // Run EnqueueWorkload
-    CHECK(armnnDelegateInterpreter->Invoke() == kTfLiteOk);
-
-    // Compare output data
-    auto armnnDelegateOutputId = armnnDelegateInterpreter->outputs()[0];
-    auto armnnDelegateOutputData = armnnDelegateInterpreter->typed_tensor<float>(armnnDelegateOutputId);
-    armnnDelegate::CompareData(expectedResult.data(), armnnDelegateOutputData, expectedResult.size());
-    armnnDelegateInterpreter.reset(nullptr);
-}
-
 template <typename T>
 void DelegateOptionTest(tflite::TensorType tensorType,
                         const std::vector<armnn::BackendId>& backends,