From 31441595009182c985dacbedc70c41ee6664d070 Mon Sep 17 00:00:00 2001 From: Ryan OShea Date: Mon, 7 Nov 2022 16:20:48 +0000 Subject: 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 Change-Id: Ibaa6da9d29c96a1ce32ff5196b0847fde9f04a1c --- delegate/src/DelegateOptions.cpp | 5 -- delegate/src/test/DelegateOptionsTest.cpp | 13 ---- delegate/src/test/DelegateOptionsTestHelper.hpp | 89 ------------------------- 3 files changed, 107 deletions(-) (limited to 'delegate/src') 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 @@ -145,11 +145,6 @@ DelegateOptions::DelegateOptions(char const* const* options_keys, else if (std::string(options_keys[i]) == std::string("reduce-fp32-to-fp16")) { 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 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 CreateCeilTfLiteModel(tflite::TensorType tensorType, flatBufferBuilder.GetBufferPointer() + flatBufferBuilder.GetSize()); } -void ReduceFp32ToBf16TestImpl() -{ - using namespace tflite; - // Set input data - std::vector inputShape{ 1, 5, 5, 1 }; - std::vector filterShape{ 1, 3, 3, 1 }; - std::vector biasShape{ 1 }; - std::vector outputShape{ 1, 3, 3, 1 }; - - std::vector 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 filterValues = - { - 4, 5, 6, - 0, 0, 0, - 3, 2, 1 - }; - - std::vector biasValues = { 5 }; - - std::vector expectedResult = - { - 28, 38, 29, - 96, 104, 53, - 31, 55, 24 - }; - - tflite::Padding padding = Padding_SAME; - - std::vector modelBuffer; - modelBuffer = CreateConv2dTfLiteModel(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 armnnDelegateInterpreter; - CHECK(InterpreterBuilder(tfLiteModel, ::tflite::ops::builtin::BuiltinOpResolver()) - (&armnnDelegateInterpreter) == kTfLiteOk); - CHECK(armnnDelegateInterpreter != nullptr); - CHECK(armnnDelegateInterpreter->AllocateTensors() == kTfLiteOk); - - // Create the Armnn Delegate - std::vector backends = {armnn::Compute::CpuRef}; - std::vector backendOptions; - - // Enable debug with BF16 enabled - armnn::OptimizerOptions optimizerOptions(false, true, true, false); - - armnnDelegate::DelegateOptions delegateOptions(backends, optimizerOptions); - std::unique_ptr - 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(armnnDelegateOutputId); - armnnDelegate::CompareData(expectedResult.data(), armnnDelegateOutputData, expectedResult.size()); - armnnDelegateInterpreter.reset(nullptr); -} - template void DelegateOptionTest(tflite::TensorType tensorType, const std::vector& backends, -- cgit v1.2.1