From a35980546c00ae1647ce033b061530607a5ad1e4 Mon Sep 17 00:00:00 2001 From: morgolock Date: Tue, 31 Dec 2019 12:20:47 +0000 Subject: COMPMID-2994: Add support QASYMM8_SIGNED in NEElementwiseMax Change-Id: I8261558384ae028f7f016dfd6715de140f0b6445 Signed-off-by: morgolock Reviewed-on: https://review.mlplatform.org/c/2528 Comments-Addressed: Arm Jenkins Tested-by: Arm Jenkins Reviewed-by: Michele Di Giorgio --- .../NEON/functions/NEElementwiseOperations.h | 6 +-- .../NEON/kernels/NEElementwiseOperationKernel.cpp | 6 +-- .../NEON/functions/NEElementwiseOperators.cpp | 6 +++ tests/validation/NEON/ElementwiseMax.cpp | 44 +++++++++++++++++++--- 4 files changed, 50 insertions(+), 12 deletions(-) diff --git a/arm_compute/runtime/NEON/functions/NEElementwiseOperations.h b/arm_compute/runtime/NEON/functions/NEElementwiseOperations.h index 983a9a5d4e..d2281f8bea 100644 --- a/arm_compute/runtime/NEON/functions/NEElementwiseOperations.h +++ b/arm_compute/runtime/NEON/functions/NEElementwiseOperations.h @@ -1,5 +1,5 @@ /* - * Copyright (c) 2018-2019 ARM Limited. + * Copyright (c) 2018-2020 ARM Limited. * * SPDX-License-Identifier: MIT * @@ -41,14 +41,14 @@ class NEElementwiseMax : public INESimpleFunction public: /** Initialise the kernel's inputs, output and conversion policy. * - * @param[in, out] input1 First tensor input. Data types supported: QASYMM8/S16/F16/S32/F32. + * @param[in, out] input1 First tensor input. Data types supported: QASYMM8/QASYMM8_SIGNED/S16/F16/S32/F32. * @param[in, out] input2 Second tensor input. Data types supported: Same as @p input1. * @param[out] output Output tensor. Data types supported: Same as @p input1. */ void configure(ITensor *input1, ITensor *input2, ITensor *output); /** Static function to check if given info will lead to a valid configuration of @ref NEArithmeticOperationKernel for max * - * @param[in] input1 First tensor input info. Data types supported: QASYMM8/S16/F16/S32/F32. + * @param[in] input1 First tensor input info. Data types supported: QASYMM8/QASYMM8_SIGNED/S16/F16/S32/F32. * @param[in] input2 Second tensor input info. Data types supported: Same as @p input1. * @param[in] output Output tensor info. Data types supported: Same as @p input1. * diff --git a/src/core/NEON/kernels/NEElementwiseOperationKernel.cpp b/src/core/NEON/kernels/NEElementwiseOperationKernel.cpp index 4928ae9bdd..b8e6f0cc69 100644 --- a/src/core/NEON/kernels/NEElementwiseOperationKernel.cpp +++ b/src/core/NEON/kernels/NEElementwiseOperationKernel.cpp @@ -1,5 +1,5 @@ /* - * Copyright (c) 2018-2019 ARM Limited. + * Copyright (c) 2018-2020 ARM Limited. * * SPDX-License-Identifier: MIT * @@ -795,8 +795,7 @@ void elementwise_op_quantized_signed(const ITensor *in1, const ITensor *in2, ITe const UniformQuantizationInfo output_qinfo = out->info()->quantization_info().uniform(); - // Output quantization info (add 0.5 to round toward the nearest integer - 0.5 rounds away from zero) - const float32x4_t voffseto = vdupq_n_f32(output_qinfo.offset + 0.5f); + const float32x4_t voffseto = vdupq_n_f32(output_qinfo.offset); const float32x4_t invvscaleo = vdupq_n_f32(1.f / output_qinfo.scale); if(is_broadcast_across_x) @@ -1000,7 +999,6 @@ NEElementwiseOperationKernel::NEElementwiseOperationKernel() Status NEElementwiseOperationKernel::validate_arguments_common(const ITensorInfo &input1, const ITensorInfo &input2, const ITensorInfo &output) { ARM_COMPUTE_RETURN_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(&input1, 1, DataType::QASYMM8, DataType::QASYMM8_SIGNED, DataType::S16, DataType::F16, DataType::S32, DataType::F32); - ARM_COMPUTE_RETURN_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(&input2, 1, DataType::QASYMM8, DataType::QASYMM8_SIGNED, DataType::S16, DataType::F16, DataType::S32, DataType::F32); ARM_COMPUTE_RETURN_ERROR_ON_CPU_F16_UNSUPPORTED(&input1); ARM_COMPUTE_RETURN_ERROR_ON_MISMATCHING_DATA_TYPES(&input1, &input2); diff --git a/src/runtime/NEON/functions/NEElementwiseOperators.cpp b/src/runtime/NEON/functions/NEElementwiseOperators.cpp index 699363111d..ede8c20428 100644 --- a/src/runtime/NEON/functions/NEElementwiseOperators.cpp +++ b/src/runtime/NEON/functions/NEElementwiseOperators.cpp @@ -23,6 +23,7 @@ */ #include "arm_compute/runtime/NEON/functions/NEElementwiseOperations.h" #include +#include "arm_compute/core/Validate.h" #include "arm_compute/core/ITensor.h" #include "support/ToolchainSupport.h" @@ -40,6 +41,11 @@ void NEElementwiseMax::configure(ITensor *input1, ITensor *input2, ITensor *outp Status NEElementwiseMax::validate(const ITensorInfo *input1, const ITensorInfo *input2, const ITensorInfo *output) { + if( input1->data_type() == DataType::QASYMM8_SIGNED) + { + ARM_COMPUTE_RETURN_ERROR_ON_MISMATCHING_DATA_TYPES(input1, output); + ARM_COMPUTE_RETURN_ERROR_ON_MISMATCHING_QUANTIZATION_INFO(input1, input2, output); + } return NEArithmeticOperationKernel::validate(ArithmeticOperation::MAX, input1, input2, output); } diff --git a/tests/validation/NEON/ElementwiseMax.cpp b/tests/validation/NEON/ElementwiseMax.cpp index 5c8a12ea96..fc6bb9454d 100644 --- a/tests/validation/NEON/ElementwiseMax.cpp +++ b/tests/validation/NEON/ElementwiseMax.cpp @@ -1,5 +1,5 @@ /* - * Copyright (c) 2018-2019 ARM Limited. + * Copyright (c) 2018-2020 ARM Limited. * * SPDX-License-Identifier: MIT * @@ -46,6 +46,10 @@ RelativeTolerance tolerance_fp32(0.000001f); const auto ElementwiseMaxQASYMM8Dataset = combine(combine(framework::dataset::make("DataType", DataType::QASYMM8), framework::dataset::make("DataType", DataType::QASYMM8)), framework::dataset::make("DataType", DataType::QASYMM8)); +const auto ElementwiseMaxQASYMM8SignedDataset = combine(combine(framework::dataset::make("DataType", DataType::QASYMM8_SIGNED), framework::dataset::make("DataType", DataType::QASYMM8_SIGNED)), + framework::dataset::make("DataType", + DataType::QASYMM8_SIGNED)); + /** Input data sets **/ const auto ElementwiseMaxS32Dataset = combine(combine(framework::dataset::make("DataType", DataType::S32), framework::dataset::make("DataType", DataType::S32)), framework::dataset::make("DataType", DataType::S32)); @@ -71,25 +75,41 @@ DATA_TEST_CASE(Validate, framework::DatasetMode::ALL, zip(zip(zip( framework::dataset::make("Input1Info", { TensorInfo(TensorShape(32U, 13U, 2U), 1, DataType::F32), TensorInfo(TensorShape(32U, 13U, 2U), 1, DataType::S32), TensorInfo(TensorShape(27U, 13U, 2U), 1, DataType::S32), - TensorInfo(TensorShape(32U, 13U, 2U), 1, DataType::S32), // Invalid data type combination - TensorInfo(TensorShape(32U, 13U, 2U), 1, DataType::F32), // Mismatching shapes + TensorInfo(TensorShape(32U, 13U, 2U), 1, DataType::S32), // Invalid data type combination + TensorInfo(TensorShape(32U, 13U, 2U), 1, DataType::F32), // Mismatching shapes + TensorInfo(TensorShape(8U, 8U, 3U), 1, DataType::QASYMM8_SIGNED), // OK + TensorInfo(TensorShape(8U, 8U, 3U), 1, DataType::QASYMM8_SIGNED), // Mismatching data types + TensorInfo(TensorShape(8U, 8U, 3U), 1, DataType::QASYMM8_SIGNED,QuantizationInfo(0.1f, 2) ), // Mismatching qinfo + }), framework::dataset::make("Input2Info",{ TensorInfo(TensorShape(32U, 13U, 2U), 1, DataType::F32), TensorInfo(TensorShape(32U, 13U, 2U), 1, DataType::S32), TensorInfo(TensorShape(27U, 13U, 2U), 1, DataType::S32), TensorInfo(TensorShape(32U, 13U, 2U), 1, DataType::S16), TensorInfo(TensorShape(48U, 11U, 2U), 1, DataType::F32), + TensorInfo(TensorShape(8U, 8U, 3U), 1, DataType::QASYMM8_SIGNED), + TensorInfo(TensorShape(8U, 8U, 3U), 1, DataType::QASYMM8), + TensorInfo(TensorShape(8U, 8U, 3U), 1, DataType::QASYMM8_SIGNED,QuantizationInfo(0.4f, 5) ), // Mismatching qinfo + })), framework::dataset::make("OutputInfo",{ TensorInfo(TensorShape(32U, 13U, 2U), 1, DataType::F32), TensorInfo(TensorShape(32U, 13U, 2U), 1, DataType::S32), TensorInfo(TensorShape(27U, 13U, 2U), 1, DataType::S32), TensorInfo(TensorShape(32U, 13U, 2U), 1, DataType::S32), TensorInfo(TensorShape(48U, 11U, 2U), 1, DataType::F32), + TensorInfo(TensorShape(8U, 8U, 3U), 1, DataType::QASYMM8_SIGNED), + TensorInfo(TensorShape(8U, 8U, 3U), 1, DataType::QASYMM8_SIGNED), + TensorInfo(TensorShape(8U, 8U, 3U), 1, DataType::QASYMM8_SIGNED,QuantizationInfo(0.3f,4 ) ), // Mismatching qinfo })), - framework::dataset::make("Expected", { true, true, true, false, false})), + + framework::dataset::make("Expected", { true, true, true, false, false, true, false, false,false})), input1_info, input2_info, output_info, expected) { - ARM_COMPUTE_EXPECT(bool(NEElementwiseMax::validate(&input1_info.clone()->set_is_resizable(false), &input2_info.clone()->set_is_resizable(false), &output_info.clone()->set_is_resizable(false))) == expected, framework::LogLevel::ERRORS); + ARM_COMPUTE_EXPECT(bool(NEElementwiseMax::validate( + &input1_info.clone()->set_is_resizable(false), + &input2_info.clone()->set_is_resizable(false), + &output_info.clone()->set_is_resizable(false)) + ) == expected, framework::LogLevel::ERRORS); } // clang-format on // *INDENT-ON* @@ -138,6 +158,20 @@ FIXTURE_DATA_TEST_CASE(RunSmallBroadcast, NEElementwiseMaxQuantizedBroadcastFixt validate(Accessor(_target), _reference); } TEST_SUITE_END() + +TEST_SUITE(QASYMM8_SIGNED) +FIXTURE_DATA_TEST_CASE(RunSmall, NEElementwiseMaxQuantizedFixture, framework::DatasetMode::PRECOMMIT, combine(combine(combine(combine(datasets::SmallShapes(), + ElementwiseMaxQASYMM8SignedDataset), + framework::dataset::make("QuantizationInfo", { QuantizationInfo(10.f, 2) })), + framework::dataset::make("QuantizationInfo", { QuantizationInfo(10.f, 2) })), + framework::dataset::make("QuantizationInfo", { QuantizationInfo(10.f, 2) }))) +{ + // Validate output + validate(Accessor(_target), _reference); +} + +TEST_SUITE_END() + TEST_SUITE_END() TEST_SUITE(Float) -- cgit v1.2.1