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

1 files changed, 0 insertions, 229 deletions

diff --git a/src/armnn/test/optimizations/Fp32NetworkToBf16ConverterTests.cpp b/src/armnn/test/optimizations/Fp32NetworkToBf16ConverterTests.cpp
deleted file mode 100644
index 66893ce1f5..0000000000
--- a/src/armnn/test/optimizations/Fp32NetworkToBf16ConverterTests.cpp
+++ /dev/null
@@ -1,229 +0,0 @@
-//
-// Copyright © 2020 Arm Ltd. All rights reserved.
-// SPDX-License-Identifier: MIT
-//
-
-#include <TestUtils.hpp>
-
-#include <Optimizer.hpp>
-
-#include <doctest/doctest.h>
-
-TEST_SUITE("Optimizer")
-{
-using namespace armnn::optimizations;
-
-TEST_CASE("Fp32NetworkToBf16OptimizationNoConversionTest")
-{
-    armnn::Graph graph;
-
-    const armnn::TensorInfo infoFP32({ 2, 2, 1, 3 }, armnn::DataType::Float32);
-
-    // Create the simple test network without Conv2D/FullyConnected.
-    auto input = graph.AddLayer<armnn::InputLayer>(0, "input");
-    input->GetOutputSlot().SetTensorInfo(infoFP32);
-
-    auto floor = graph.AddLayer<armnn::FloorLayer>("floor");
-    floor->GetOutputSlot().SetTensorInfo(infoFP32);
-
-    auto output = graph.AddLayer<armnn::OutputLayer>(1, "output");
-
-    // Connect up the layers
-    input->GetOutputSlot().Connect(floor->GetInputSlot(0));
-    floor->GetOutputSlot().Connect(output->GetInputSlot(0));
-
-    CHECK(CheckSequence(graph.cbegin(), graph.cend(), &IsLayerOfType<armnn::InputLayer>,
-                             &IsLayerOfType<armnn::FloorLayer>, &IsLayerOfType<armnn::OutputLayer>));
-
-    // Run the optimizer
-    armnn::Optimizer::Pass(graph, armnn::MakeOptimizations(Fp32NetworkToBf16Converter()));
-
-    CHECK(CheckSequence(graph.cbegin(), graph.cend(), &IsLayerOfType<armnn::InputLayer>,
-                             &IsLayerOfType<armnn::FloorLayer>,
-                             &IsLayerOfType<armnn::OutputLayer>));
-}
-
-TEST_CASE("Fp32NetworkToBf16OptimizationConv2DTest")
-{
-    armnn::Graph graph;
-
-    const armnn::TensorInfo infoFP32({ 2, 3, 8, 1 }, armnn::DataType::Float32);
-
-    // Create const tensor fp32 data
-    unsigned int dims[] = { 4, 2, 1, 1 };
-    std::vector<float> floatWeights{ 0.0f, -1.0f,
-                                     3.8f, // 0x40733333 Round down
-                                     3.1055E+29f, // 0x707ADC3C Round up
-                                     9.149516E-10f, // 0x307B7FFF Round down
-                                    -3.8f, // 0xC0733333 Round down
-                                    -3.1055E+29f, // 0xF07ADC3C Round up
-                                    -9.149516E-10f // 0xB07B7FFF Round down
-                                   };
-    armnn::ConstTensor weights(armnn::TensorInfo(4, dims, armnn::DataType::Float32, 0.0f, 0, true), floatWeights);
-
-    // Create const bias fp32 data
-    unsigned int biasDims[] {4};
-    std::vector<float> floatBias{ 1.0f, 2.0f, 3.0f, 4.0f };
-    armnn::ConstTensor bias(armnn::TensorInfo(1, biasDims, armnn::DataType::Float32, 0.0f, 0, true), floatBias);
-
-    // A network with Convolution2d layer
-    auto input = graph.AddLayer<armnn::InputLayer>(0, "input");
-    input->GetOutputSlot().SetTensorInfo(infoFP32);
-
-    armnn::Convolution2dDescriptor descriptor;
-    descriptor.m_BiasEnabled = true;
-    auto conv = graph.AddLayer<armnn::Convolution2dLayer>(descriptor, "conv2d");
-    conv->GetOutputSlot().SetTensorInfo(infoFP32);
-
-    auto weightsLayer = graph.AddLayer<armnn::ConstantLayer>("Weights");
-    weightsLayer->m_LayerOutput = std::make_shared<armnn::ScopedTensorHandle>(weights);
-    weightsLayer->GetOutputSlot(0).SetTensorInfo(weightsLayer->m_LayerOutput->GetTensorInfo());
-
-    auto biasLayer = graph.AddLayer<armnn::ConstantLayer>("Bias");
-    biasLayer->m_LayerOutput = std::make_shared<armnn::ScopedTensorHandle>(bias);
-    biasLayer->GetOutputSlot(0).SetTensorInfo(biasLayer->m_LayerOutput->GetTensorInfo());
-
-    auto output = graph.AddLayer<armnn::OutputLayer>(1, "output");
-
-    // Connect up the layers
-    input->GetOutputSlot().Connect(conv->GetInputSlot(0));
-    weightsLayer->GetOutputSlot(0).Connect(conv->GetInputSlot(1));
-    biasLayer->GetOutputSlot(0).Connect(conv->GetInputSlot(2));
-    conv->GetOutputSlot().Connect(output->GetInputSlot(0));
-
-    CHECK(CheckSequence(graph.cbegin(), graph.cend(), &IsLayerOfType<armnn::InputLayer>,
-                                                      &IsLayerOfType<armnn::ConstantLayer>,
-                                                      &IsLayerOfType<armnn::ConstantLayer>,
-                                                      &IsLayerOfType<armnn::Convolution2dLayer>,
-                                                      &IsLayerOfType<armnn::OutputLayer>));
-
-    // Run the optimizer
-    armnn::Optimizer::Pass(graph, armnn::MakeOptimizations(RedirectMembersToConstantInputs(),
-                                                           Fp32NetworkToBf16Converter()));
-
-    CHECK(7 == graph.GetNumLayers());
-    CHECK(CheckSequence(graph.cbegin(), graph.cend(), &IsLayerOfType<armnn::InputLayer>,
-                                                      &IsLayerOfType<armnn::ConstantLayer>,
-                                                      &IsLayerOfType<armnn::ConstantLayer>,
-                                                      &IsLayerOfType<armnn::ConvertFp32ToBf16Layer>,
-                                                      &IsLayerOfType<armnn::ConvertFp32ToBf16Layer>,
-                                                      &IsLayerOfType<armnn::Convolution2dLayer>,
-                                                      &IsLayerOfType<armnn::OutputLayer>));
-
-    armnn::TensorInfo inputTensor = conv->GetInputSlot(0).GetConnectedOutputSlot()->GetTensorInfo();
-    armnn::TensorInfo weightTensor = conv->GetInputSlot(1).GetConnectedOutputSlot()->GetTensorInfo();
-    armnn::TensorInfo biasTensor = conv->GetInputSlot(2).GetConnectedOutputSlot()->GetTensorInfo();
-    armnn::TensorInfo outputTensor = conv->GetOutputSlot(0).GetTensorInfo();
-    CHECK((conv->GetDataType() == armnn::DataType::BFloat16));
-    CHECK((conv->m_Weight->GetTensorInfo().GetDataType() == armnn::DataType::BFloat16));
-    CHECK((conv->m_Bias->GetTensorInfo().GetDataType() == armnn::DataType::Float32));
-    CHECK((inputTensor.GetDataType() == armnn::DataType::BFloat16));
-    CHECK((weightTensor.GetDataType() == armnn::DataType::BFloat16));
-    CHECK((biasTensor.GetDataType() == armnn::DataType::Float32));
-    CHECK((outputTensor.GetDataType() == armnn::DataType::Float32));
-
-    // Check whether data matches expected Bf16 data
-    const armnn::BFloat16* data = conv->m_Weight->GetConstTensor<armnn::BFloat16>();
-    CHECK(data[0] == armnn::BFloat16(0.0f));
-    CHECK(data[1] == armnn::BFloat16(-1.0f));
-    CHECK(data[2] == armnn::BFloat16(3.796875f)); // 0x4073
-    CHECK(data[3] == armnn::BFloat16(3.1072295E29f)); // 0x707B
-    CHECK(data[4] == armnn::BFloat16(9.131327E-10f)); // 0x307B
-    CHECK(data[5] == armnn::BFloat16(-3.796875f)); // 0xC073
-    CHECK(data[6] == armnn::BFloat16(-3.1072295E29f)); // 0xF07B
-    CHECK(data[7] == armnn::BFloat16(-9.131327E-10f)); // 0xB07B
-}
-
-TEST_CASE("Fp32NetworkToBf16OptimizationFullyConnectedTest")
-{
-    armnn::Graph graph;
-
-    const armnn::TensorInfo infoFP32({ 2, 3, 8, 1 }, armnn::DataType::Float32);
-
-    // Create const tensor fp32 data
-    unsigned int dims[] = { 4, 2, 1, 1 };
-    std::vector<float> floatWeights{ 0.0f, -1.0f,
-                                     3.8f, // 0x40733333 Round down
-                                     3.1055E+29f, // 0x707ADC3C Round up
-                                     9.149516E-10f, // 0x307B7FFF Round down
-                                    -3.8f, // 0xC0733333 Round down
-                                    -3.1055E+29f, // 0xF07ADC3C Round up
-                                    -9.149516E-10f // 0xB07B7FFF Round down
-                                   };
-    armnn::ConstTensor weights(armnn::TensorInfo(4, dims, armnn::DataType::Float32, 0.0f, 0, true), floatWeights);
-
-    // Create const bias fp32 data
-    unsigned int biasDims[] {4};
-    std::vector<float> floatBias{ 1.0f, 2.0f, 3.0f, 4.0f };
-    armnn::ConstTensor bias(armnn::TensorInfo(1, biasDims, armnn::DataType::Float32, 0.0f, 0, true), floatBias);
-
-    // A network with FullyConnected layer
-    auto input = graph.AddLayer<armnn::InputLayer>(0, "input");
-    input->GetOutputSlot().SetTensorInfo(infoFP32);
-
-    armnn::FullyConnectedDescriptor descriptor;
-    descriptor.m_BiasEnabled = true;
-
-    auto fc = graph.AddLayer<armnn::FullyConnectedLayer>(descriptor, "fully");
-    fc->GetOutputSlot().SetTensorInfo(infoFP32);
-
-    auto weightsLayer = graph.AddLayer<armnn::ConstantLayer>("Weights");
-    weightsLayer->m_LayerOutput = std::make_shared<armnn::ScopedTensorHandle>(weights);
-    weightsLayer->GetOutputSlot(0).SetTensorInfo(weightsLayer->m_LayerOutput->GetTensorInfo());
-
-    auto biasLayer = graph.AddLayer<armnn::ConstantLayer>("Bias");
-    biasLayer->m_LayerOutput = std::make_shared<armnn::ScopedTensorHandle>(bias);
-    biasLayer->GetOutputSlot(0).SetTensorInfo(biasLayer->m_LayerOutput->GetTensorInfo());
-
-    auto output = graph.AddLayer<armnn::OutputLayer>(1, "output");
-
-    // Connect up the layers
-    input->GetOutputSlot().Connect(fc->GetInputSlot(0));
-    weightsLayer->GetOutputSlot(0).Connect(fc->GetInputSlot(1));
-    biasLayer->GetOutputSlot(0).Connect(fc->GetInputSlot(2));
-    fc->GetOutputSlot().Connect(output->GetInputSlot(0));
-
-    CHECK(CheckSequence(graph.cbegin(), graph.cend(), &IsLayerOfType<armnn::InputLayer>,
-                                                      &IsLayerOfType<armnn::ConstantLayer>,
-                                                      &IsLayerOfType<armnn::ConstantLayer>,
-                                                      &IsLayerOfType<armnn::FullyConnectedLayer>,
-                                                      &IsLayerOfType<armnn::OutputLayer>));
-
-    // Run the optimizer
-    armnn::Optimizer::Pass(graph, armnn::MakeOptimizations(RedirectMembersToConstantInputs(),
-                                                           Fp32NetworkToBf16Converter()));
-
-    CHECK(7 == graph.GetNumLayers());
-    CHECK(CheckSequence(graph.cbegin(), graph.cend(), &IsLayerOfType<armnn::InputLayer>,
-                                                      &IsLayerOfType<armnn::ConstantLayer>,
-                                                      &IsLayerOfType<armnn::ConstantLayer>,
-                                                      &IsLayerOfType<armnn::ConvertFp32ToBf16Layer>,
-                                                      &IsLayerOfType<armnn::ConvertFp32ToBf16Layer>,
-                                                      &IsLayerOfType<armnn::FullyConnectedLayer>,
-                                                      &IsLayerOfType<armnn::OutputLayer>));
-
-    armnn::TensorInfo inputTensor = fc->GetInputSlot(0).GetConnectedOutputSlot()->GetTensorInfo();
-    armnn::TensorInfo weightTensor = fc->GetInputSlot(1).GetConnectedOutputSlot()->GetTensorInfo();
-    armnn::TensorInfo biasTensor = fc->GetInputSlot(2).GetConnectedOutputSlot()->GetTensorInfo();
-    armnn::TensorInfo outputTensor = fc->GetOutputSlot(0).GetTensorInfo();
-    CHECK((fc->GetDataType() == armnn::DataType::BFloat16));
-    CHECK((fc->m_Weight->GetTensorInfo().GetDataType() == armnn::DataType::BFloat16));
-    CHECK((fc->m_Bias->GetTensorInfo().GetDataType() == armnn::DataType::Float32));
-    CHECK((inputTensor.GetDataType() == armnn::DataType::BFloat16));
-    CHECK((weightTensor.GetDataType() == armnn::DataType::BFloat16));
-    CHECK((biasTensor.GetDataType() == armnn::DataType::Float32));
-    CHECK((outputTensor.GetDataType() == armnn::DataType::Float32));
-
-    // Check whether data matches expected Bf16 data
-    const armnn::BFloat16* data = fc->m_Weight->GetConstTensor<armnn::BFloat16>();
-    CHECK(data[0] == armnn::BFloat16(0.0f));
-    CHECK(data[1] == armnn::BFloat16(-1.0f));
-    CHECK(data[2] == armnn::BFloat16(3.796875f)); // 0x4073
-    CHECK(data[3] == armnn::BFloat16(3.1072295E29f)); // 0x707B
-    CHECK(data[4] == armnn::BFloat16(9.131327E-10f)); // 0x307B
-    CHECK(data[5] == armnn::BFloat16(-3.796875f)); // 0xC073
-    CHECK(data[6] == armnn::BFloat16(-3.1072295E29f)); // 0xF07B
-    CHECK(data[7] == armnn::BFloat16(-9.131327E-10f)); // 0xB07B
-}
-
-}
\ No newline at end of file