From ea54a01f6bd30f013cbe88ae1751985bc86b6af5 Mon Sep 17 00:00:00 2001 From: Narumol Prangnawarat Date: Mon, 16 Mar 2020 16:36:10 +0000 Subject: IVGCVSW-4516 Add ConvertFp32ToBf16Layer and Ref workload support Signed-off-by: Narumol Prangnawarat Change-Id: I9099a4f840fb747336f77d20a0868b64e801a310 --- src/backends/reference/RefLayerSupport.cpp | 15 ++++++++++ src/backends/reference/RefLayerSupport.hpp | 4 +++ src/backends/reference/RefWorkloadFactory.cpp | 7 +++++ src/backends/reference/RefWorkloadFactory.hpp | 3 ++ src/backends/reference/backend.mk | 1 + .../reference/test/RefLayerSupportTests.cpp | 32 ++++++++++++++++++++++ src/backends/reference/test/RefLayerTests.cpp | 3 ++ src/backends/reference/workloads/CMakeLists.txt | 2 ++ .../workloads/RefConvertFp32ToBf16Workload.cpp | 27 ++++++++++++++++++ .../workloads/RefConvertFp32ToBf16Workload.hpp | 21 ++++++++++++++ .../reference/workloads/RefWorkloadUtils.hpp | 6 ++++ src/backends/reference/workloads/RefWorkloads.hpp | 1 + 12 files changed, 122 insertions(+) create mode 100644 src/backends/reference/workloads/RefConvertFp32ToBf16Workload.cpp create mode 100644 src/backends/reference/workloads/RefConvertFp32ToBf16Workload.hpp (limited to 'src/backends/reference') diff --git a/src/backends/reference/RefLayerSupport.cpp b/src/backends/reference/RefLayerSupport.cpp index a4f4efd92a..9f22b9ef0e 100644 --- a/src/backends/reference/RefLayerSupport.cpp +++ b/src/backends/reference/RefLayerSupport.cpp @@ -411,6 +411,21 @@ bool RefLayerSupport::IsConvertFp16ToFp32Supported(const TensorInfo& input, &FalseFuncU8<>)); } +bool RefLayerSupport::IsConvertFp32ToBf16Supported(const TensorInfo& input, + const TensorInfo& output, + Optional reasonIfUnsupported) const +{ + bool supported = true; + + supported &= CheckSupportRule(TypeIs(input, DataType::Float32), reasonIfUnsupported, + "Reference for ConvertFp32ToBf16 layer: input type not supported"); + + supported &= CheckSupportRule(TypeIs(output, DataType::BFloat16), reasonIfUnsupported, + "Reference for ConvertFp32ToBf16 layer: output type not supported"); + + return supported; +} + bool RefLayerSupport::IsConvertFp32ToFp16Supported(const TensorInfo& input, const TensorInfo& output, Optional reasonIfUnsupported) const diff --git a/src/backends/reference/RefLayerSupport.hpp b/src/backends/reference/RefLayerSupport.hpp index ff3478142f..30f45c37f2 100644 --- a/src/backends/reference/RefLayerSupport.hpp +++ b/src/backends/reference/RefLayerSupport.hpp @@ -68,6 +68,10 @@ public: const TensorInfo& output, Optional reasonIfUnsupported = EmptyOptional()) const override; + bool IsConvertFp32ToBf16Supported(const TensorInfo& input, + const TensorInfo& output, + Optional reasonIfUnsupported = EmptyOptional()) const override; + bool IsConvertFp32ToFp16Supported(const TensorInfo& input, const TensorInfo& output, Optional reasonIfUnsupported = EmptyOptional()) const override; diff --git a/src/backends/reference/RefWorkloadFactory.cpp b/src/backends/reference/RefWorkloadFactory.cpp index c9fc626bc1..5d3775a59d 100644 --- a/src/backends/reference/RefWorkloadFactory.cpp +++ b/src/backends/reference/RefWorkloadFactory.cpp @@ -194,6 +194,13 @@ std::unique_ptr RefWorkloadFactory::CreateConvertFp16ToFp32( return std::make_unique(descriptor, info); } +std::unique_ptr RefWorkloadFactory::CreateConvertFp32ToBf16( + const ConvertFp32ToBf16QueueDescriptor& descriptor, + const WorkloadInfo& info) const +{ + return std::make_unique(descriptor, info); +} + std::unique_ptr RefWorkloadFactory::CreateConvertFp32ToFp16( const ConvertFp32ToFp16QueueDescriptor& descriptor, const WorkloadInfo& info) const diff --git a/src/backends/reference/RefWorkloadFactory.hpp b/src/backends/reference/RefWorkloadFactory.hpp index 119605b998..9a53ae2e5a 100644 --- a/src/backends/reference/RefWorkloadFactory.hpp +++ b/src/backends/reference/RefWorkloadFactory.hpp @@ -92,6 +92,9 @@ public: std::unique_ptr CreateConvertFp16ToFp32(const ConvertFp16ToFp32QueueDescriptor& descriptor, const WorkloadInfo& info) const override; + std::unique_ptr CreateConvertFp32ToBf16(const ConvertFp32ToBf16QueueDescriptor& descriptor, + const WorkloadInfo& info) const override; + std::unique_ptr CreateConvertFp32ToFp16(const ConvertFp32ToFp16QueueDescriptor& descriptor, const WorkloadInfo& info) const override; diff --git a/src/backends/reference/backend.mk b/src/backends/reference/backend.mk index aeac2e28b9..239863f2c7 100644 --- a/src/backends/reference/backend.mk +++ b/src/backends/reference/backend.mk @@ -51,6 +51,7 @@ BACKEND_SOURCES := \ workloads/RefConstantWorkload.cpp \ workloads/RefConvertBf16ToFp32Workload.cpp \ workloads/RefConvertFp16ToFp32Workload.cpp \ + workloads/RefConvertFp32ToBf16Workload.cpp \ workloads/RefConvertFp32ToFp16Workload.cpp \ workloads/RefConvolution2dWorkload.cpp \ workloads/RefDebugWorkload.cpp \ diff --git a/src/backends/reference/test/RefLayerSupportTests.cpp b/src/backends/reference/test/RefLayerSupportTests.cpp index 0b549dbb77..f0c69f92cc 100644 --- a/src/backends/reference/test/RefLayerSupportTests.cpp +++ b/src/backends/reference/test/RefLayerSupportTests.cpp @@ -148,6 +148,38 @@ BOOST_AUTO_TEST_CASE(IsConvertBf16ToFp32SupportedBf16OutputReference) BOOST_CHECK_EQUAL(reasonIfUnsupported, "Reference for ConvertBf16ToFp32 layer: output type not supported\n"); } +BOOST_AUTO_TEST_CASE(IsConvertFp32ToBf16SupportedReference) +{ + std::string reasonIfUnsupported; + + bool result = IsConvertLayerSupportedTests(reasonIfUnsupported); + + BOOST_CHECK(result); +} + +BOOST_AUTO_TEST_CASE(IsConvertFp32ToBf16SupportedBf16InputReference) +{ + std::string reasonIfUnsupported; + + bool result = IsConvertLayerSupportedTests(reasonIfUnsupported); + + BOOST_CHECK(!result); + BOOST_CHECK_EQUAL(reasonIfUnsupported, "Reference for ConvertFp32ToBf16 layer: input type not supported\n"); +} + +BOOST_AUTO_TEST_CASE(IsConvertFp32ToBf16SupportedFp32OutputReference) +{ + std::string reasonIfUnsupported; + + bool result = IsConvertLayerSupportedTests(reasonIfUnsupported); + + BOOST_CHECK(!result); + BOOST_CHECK_EQUAL(reasonIfUnsupported, "Reference for ConvertFp32ToBf16 layer: output type not supported\n"); +} + BOOST_AUTO_TEST_CASE(IsConvertFp32ToFp16SupportedReference) { std::string reasonIfUnsupported; diff --git a/src/backends/reference/test/RefLayerTests.cpp b/src/backends/reference/test/RefLayerTests.cpp index 08cfa2843f..f5c1c65c88 100644 --- a/src/backends/reference/test/RefLayerTests.cpp +++ b/src/backends/reference/test/RefLayerTests.cpp @@ -1145,6 +1145,9 @@ ARMNN_AUTO_TEST_CASE(LstmLayerInt16NoCifgNoPeepholeNoProjectionInt16Constant, // Convert from BFloat16 to Float32 ARMNN_AUTO_TEST_CASE(ConvertBf16ToFp32, ConvertBf16ToFp32Test) +// Convert from Float32 to BFloat16 +ARMNN_AUTO_TEST_CASE(ConvertFp32ToBf16, ConvertFp32ToBf16Test) + // Convert from Float16 to Float32 ARMNN_AUTO_TEST_CASE(SimpleConvertFp16ToFp32, SimpleConvertFp16ToFp32Test) // Convert from Float32 to Float16 diff --git a/src/backends/reference/workloads/CMakeLists.txt b/src/backends/reference/workloads/CMakeLists.txt index 86764d8a01..9f3880e077 100644 --- a/src/backends/reference/workloads/CMakeLists.txt +++ b/src/backends/reference/workloads/CMakeLists.txt @@ -73,6 +73,8 @@ list(APPEND armnnRefBackendWorkloads_sources RefConvertBf16ToFp32Workload.hpp RefConvertFp16ToFp32Workload.cpp RefConvertFp16ToFp32Workload.hpp + RefConvertFp32ToBf16Workload.cpp + RefConvertFp32ToBf16Workload.hpp RefConvertFp32ToFp16Workload.cpp RefConvertFp32ToFp16Workload.hpp RefConvolution2dWorkload.cpp diff --git a/src/backends/reference/workloads/RefConvertFp32ToBf16Workload.cpp b/src/backends/reference/workloads/RefConvertFp32ToBf16Workload.cpp new file mode 100644 index 0000000000..181b236e83 --- /dev/null +++ b/src/backends/reference/workloads/RefConvertFp32ToBf16Workload.cpp @@ -0,0 +1,27 @@ +// +// Copyright © 2020 Arm Ltd. All rights reserved. +// SPDX-License-Identifier: MIT +// + +#include "RefConvertFp32ToBf16Workload.hpp" +#include "RefWorkloadUtils.hpp" + +#include + +#include + +namespace armnn +{ + +void RefConvertFp32ToBf16Workload::Execute() const +{ + ARMNN_SCOPED_PROFILING_EVENT(Compute::CpuRef, "RefConvertFp32ToBf16Workload_Execute"); + + const float* const input = GetInputTensorDataFloat(0, m_Data); + BFloat16* const output = GetOutputTensorDataBFloat16(0, m_Data); + + unsigned int numElements = GetTensorInfo(m_Data.m_Inputs[0]).GetNumElements(); + armnnUtils::FloatingPointConverter::ConvertFloat32ToBFloat16(input, numElements, output); +} + +} //namespace armnn diff --git a/src/backends/reference/workloads/RefConvertFp32ToBf16Workload.hpp b/src/backends/reference/workloads/RefConvertFp32ToBf16Workload.hpp new file mode 100644 index 0000000000..409603bb6c --- /dev/null +++ b/src/backends/reference/workloads/RefConvertFp32ToBf16Workload.hpp @@ -0,0 +1,21 @@ +// +// Copyright © 2020 Arm Ltd. All rights reserved. +// SPDX-License-Identifier: MIT +// + +#pragma once + +#include +#include + +namespace armnn +{ + +class RefConvertFp32ToBf16Workload : public Float32ToBFloat16Workload +{ +public: + using Float32ToBFloat16Workload::Float32ToBFloat16Workload; + virtual void Execute() const override; +}; + +} //namespace armnn diff --git a/src/backends/reference/workloads/RefWorkloadUtils.hpp b/src/backends/reference/workloads/RefWorkloadUtils.hpp index 69713145c0..f1b31571db 100644 --- a/src/backends/reference/workloads/RefWorkloadUtils.hpp +++ b/src/backends/reference/workloads/RefWorkloadUtils.hpp @@ -75,6 +75,12 @@ const BFloat16* GetInputTensorDataBFloat16(unsigned int idx, const PayloadType& return GetInputTensorData(idx, data); } +template +BFloat16* GetOutputTensorDataBFloat16(unsigned int idx, const PayloadType& data) +{ + return GetOutputTensorData(idx, data); +} + //////////////////////////////////////////// /// u8 helpers //////////////////////////////////////////// diff --git a/src/backends/reference/workloads/RefWorkloads.hpp b/src/backends/reference/workloads/RefWorkloads.hpp index 37d79f0fc0..cbfade3c02 100644 --- a/src/backends/reference/workloads/RefWorkloads.hpp +++ b/src/backends/reference/workloads/RefWorkloads.hpp @@ -24,6 +24,7 @@ #include "RefConcatWorkload.hpp" #include "RefConvertBf16ToFp32Workload.hpp" #include "RefConvertFp16ToFp32Workload.hpp" +#include "RefConvertFp32ToBf16Workload.hpp" #include "RefConvertFp32ToFp16Workload.hpp" #include "RefDebugWorkload.hpp" #include "RefDepthToSpaceWorkload.hpp" -- cgit v1.2.1