From 002e6530f6218b00a28aef9be8b21efb08cf3602 Mon Sep 17 00:00:00 2001 From: Ramy Elgammal Date: Wed, 11 Jan 2023 18:48:04 +0000 Subject: Implement dynamic fusion softmax operator - Return aux tensorInfo by get_aux_tensors() at runtime to init the aux tensor with the right size. - Keep softmax unfusable for this commit - Hence, added Tensor3D to template writer arguments declaration, for sake of keeping dynamic fusion softmax componenets' kernels matching their cl counterparts. Resolves: COMPMID-5523 Change-Id: I667f39545db925f667036ef448302c79a0330373 Signed-off-by: Ramy Elgammal Reviewed-on: https://eu-gerrit-1.euhpc.arm.com/c/VisualCompute/ComputeLibrary/+/483924 Tested-by: bsgcomp Reviewed-by: Gunes Bayir Comments-Addressed: bsgcomp Reviewed-on: https://review.mlplatform.org/c/ml/ComputeLibrary/+/8986 Comments-Addressed: Arm Jenkins Reviewed-by: Jakub Sujak Tested-by: Arm Jenkins Benchmark: Arm Jenkins --- .../cl/ClComponentLogits1DMaxShiftExpSum.cpp | 16 +-- .../cl/ClComponentLogits1DMaxShiftExpSum.h | 8 +- .../gpu/components/cl/ClComponentLogits1DNorm.cpp | 95 +++++++++++++++ .../gpu/components/cl/ClComponentLogits1DNorm.h | 128 +++++++++++++++++++++ 4 files changed, 233 insertions(+), 14 deletions(-) create mode 100644 src/dynamic_fusion/sketch/gpu/components/cl/ClComponentLogits1DNorm.cpp create mode 100644 src/dynamic_fusion/sketch/gpu/components/cl/ClComponentLogits1DNorm.h (limited to 'src/dynamic_fusion/sketch/gpu/components') diff --git a/src/dynamic_fusion/sketch/gpu/components/cl/ClComponentLogits1DMaxShiftExpSum.cpp b/src/dynamic_fusion/sketch/gpu/components/cl/ClComponentLogits1DMaxShiftExpSum.cpp index 8ab1853e84..279c77e227 100644 --- a/src/dynamic_fusion/sketch/gpu/components/cl/ClComponentLogits1DMaxShiftExpSum.cpp +++ b/src/dynamic_fusion/sketch/gpu/components/cl/ClComponentLogits1DMaxShiftExpSum.cpp @@ -1,5 +1,5 @@ /* - * Copyright (c) 2022 Arm Limited. + * Copyright (c) 2022-2023 Arm Limited. * * SPDX-License-Identifier: MIT * @@ -27,9 +27,8 @@ #include "arm_compute/core/CL/CLHelpers.h" #include "arm_compute/core/Validate.h" #include "arm_compute/core/utils/misc/ShapeCalculator.h" -#include "arm_compute/dynamic_fusion/sketch/attributes/DepthwiseConv2dAttributes.h" +#include "arm_compute/dynamic_fusion/sketch/attributes/SoftmaxAttributes.h" #include "src/core/CL/CLValidate.h" -#include "src/dynamic_fusion/sketch/gpu/template_writer/cl/ClTemplateDepthwiseConv2d.h" #include "src/dynamic_fusion/sketch/gpu/template_writer/cl/ClTemplateLogits1DMaxShiftExpSum.h" namespace arm_compute @@ -49,7 +48,9 @@ Status ClComponentLogits1DMaxShiftExpSum::validate( const ITensorInfo *sum = tensors.get_const_tensor(TensorType::ACL_DST_0); const ITensorInfo *dst = tensors.get_const_tensor(TensorType::ACL_DST_1); - ARM_COMPUTE_RETURN_ERROR_ON_NULLPTR(src, sum, dst); + ARM_COMPUTE_RETURN_ERROR_ON_NULLPTR(src); + ARM_COMPUTE_RETURN_ERROR_ON_NULLPTR(sum); + ARM_COMPUTE_RETURN_ERROR_ON_NULLPTR(dst); // 1. Check validity // All tensor infos are initialized @@ -61,11 +62,6 @@ Status ClComponentLogits1DMaxShiftExpSum::validate( ARM_COMPUTE_RETURN_ERROR_ON_MISMATCHING_DATA_TYPES(src, dst, sum); ARM_COMPUTE_RETURN_ERROR_ON_MISMATCHING_SHAPES(src, dst); - const uint32_t src_dimensions = src->tensor_shape().num_dimensions(); - const uint32_t sum_dimensions = sum->tensor_shape().num_dimensions(); - - ARM_COMPUTE_RETURN_ERROR_ON(src_dimensions != sum_dimensions + 1); - // Device requirements are met ARM_COMPUTE_RETURN_ERROR_ON_F16_UNSUPPORTED(src); @@ -94,4 +90,4 @@ const IGpuTemplateComponentWriter *ClComponentLogits1DMaxShiftExpSum::template_w } } // namespace dynamic_fusion } // namespace experimental -} // namespace arm_compute \ No newline at end of file +} // namespace arm_compute diff --git a/src/dynamic_fusion/sketch/gpu/components/cl/ClComponentLogits1DMaxShiftExpSum.h b/src/dynamic_fusion/sketch/gpu/components/cl/ClComponentLogits1DMaxShiftExpSum.h index af47609a2c..b5db458248 100644 --- a/src/dynamic_fusion/sketch/gpu/components/cl/ClComponentLogits1DMaxShiftExpSum.h +++ b/src/dynamic_fusion/sketch/gpu/components/cl/ClComponentLogits1DMaxShiftExpSum.h @@ -1,5 +1,5 @@ /* - * Copyright (c) 2022 Arm Limited. + * Copyright (c) 2022-2023 Arm Limited. * * SPDX-License-Identifier: MIT * @@ -64,9 +64,9 @@ public: /** Validate the component * - * @param[in] properties Component properties @ref Properties - * @param[in,out] tensors Tensor arguments to the component - * @param[in] attributes Component attributes @ref Attributes + * @param[in] properties Component properties @ref Properties + * @param[in] tensors Tensor arguments to the component + * @param[in] attributes Component attributes @ref Attributes * * @return Status Validation results * diff --git a/src/dynamic_fusion/sketch/gpu/components/cl/ClComponentLogits1DNorm.cpp b/src/dynamic_fusion/sketch/gpu/components/cl/ClComponentLogits1DNorm.cpp new file mode 100644 index 0000000000..7864d56d29 --- /dev/null +++ b/src/dynamic_fusion/sketch/gpu/components/cl/ClComponentLogits1DNorm.cpp @@ -0,0 +1,95 @@ +/* + * Copyright (c) 2023 Arm Limited. + * + * SPDX-License-Identifier: MIT + * + * Permission is hereby granted, free of charge, to any person obtaining a copy + * of this software and associated documentation files (the "Software"), to + * deal in the Software without restriction, including without limitation the + * rights to use, copy, modify, merge, publish, distribute, sublicense, and/or + * sell copies of the Software, and to permit persons to whom the Software is + * furnished to do so, subject to the following conditions: + * + * The above copyright notice and this permission notice shall be included in all + * copies or substantial portions of the Software. + * + * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR + * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY, + * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE + * AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER + * LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM, + * OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE + * SOFTWARE. + */ + +#include "src/dynamic_fusion/sketch/gpu/components/cl/ClComponentLogits1DNorm.h" + +#include "arm_compute/core/CL/CLHelpers.h" +#include "arm_compute/core/Validate.h" +#include "arm_compute/core/utils/misc/ShapeCalculator.h" +#include "arm_compute/dynamic_fusion/sketch/attributes/SoftmaxAttributes.h" +#include "src/core/CL/CLValidate.h" +#include "src/dynamic_fusion/sketch/gpu/template_writer/cl/ClTemplateLogits1DNorm.h" + +namespace arm_compute +{ +namespace experimental +{ +namespace dynamic_fusion +{ +Status ClComponentLogits1DNorm::validate( + const Properties &properties, + const ArgumentPack &tensors, + const Attributes &attributes) +{ + ARM_COMPUTE_UNUSED(properties, attributes); + + const ITensorInfo *src = tensors.get_const_tensor(TensorType::ACL_SRC_0); + const ITensorInfo *sum = tensors.get_const_tensor(TensorType::ACL_SRC_1); + const ITensorInfo *dst = tensors.get_const_tensor(TensorType::ACL_DST_0); + + ARM_COMPUTE_RETURN_ERROR_ON_NULLPTR(src); + ARM_COMPUTE_RETURN_ERROR_ON_NULLPTR(sum); + ARM_COMPUTE_RETURN_ERROR_ON_NULLPTR(dst); + + // 1. Check validity + // All tensor infos are initialized + ARM_COMPUTE_RETURN_ERROR_ON(src->tensor_shape().total_size() == 0); + ARM_COMPUTE_RETURN_ERROR_ON(sum->tensor_shape().total_size() == 0); + ARM_COMPUTE_RETURN_ERROR_ON(dst->tensor_shape().total_size() == 0); + + // Check for mismatches in shapes and data types + ARM_COMPUTE_RETURN_ERROR_ON_MISMATCHING_DATA_TYPES(src, dst, sum); + ARM_COMPUTE_RETURN_ERROR_ON_MISMATCHING_SHAPES(src, dst); + + ARM_COMPUTE_RETURN_ERROR_ON(attributes.is_log_softmax() && !is_data_type_float(src->data_type())); + + // Device requirements are met + ARM_COMPUTE_RETURN_ERROR_ON_F16_UNSUPPORTED(src); + + // 2. Check support level + ARM_COMPUTE_RETURN_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(src, 1, DataType::F16, DataType::F32); + + return Status{}; +} + +ClComponentLogits1DNorm::ClComponentLogits1DNorm(ComponentId id, + const Properties &properties, + const ArgumentPack &tensors, + const Attributes &attributes) + : IGpuKernelComponent{ id, properties, tensors }, + _component_writer{ std::make_unique(id, tensors, attributes) } +{ +} + +ClComponentLogits1DNorm::~ClComponentLogits1DNorm() +{ +} + +const IGpuTemplateComponentWriter *ClComponentLogits1DNorm::template_writer() const +{ + return _component_writer.get(); +} +} // namespace dynamic_fusion +} // namespace experimental +} // namespace arm_compute diff --git a/src/dynamic_fusion/sketch/gpu/components/cl/ClComponentLogits1DNorm.h b/src/dynamic_fusion/sketch/gpu/components/cl/ClComponentLogits1DNorm.h new file mode 100644 index 0000000000..5bd350b9bd --- /dev/null +++ b/src/dynamic_fusion/sketch/gpu/components/cl/ClComponentLogits1DNorm.h @@ -0,0 +1,128 @@ +/* + * Copyright (c) 2023 Arm Limited. + * + * SPDX-License-Identifier: MIT + * + * Permission is hereby granted, free of charge, to any person obtaining a copy + * of this software and associated documentation files (the "Software"), to + * deal in the Software without restriction, including without limitation the + * rights to use, copy, modify, merge, publish, distribute, sublicense, and/or + * sell copies of the Software, and to permit persons to whom the Software is + * furnished to do so, subject to the following conditions: + * + * The above copyright notice and this permission notice shall be included in all + * copies or substantial portions of the Software. + * + * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR + * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY, + * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE + * AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER + * LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM, + * OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE + * SOFTWARE. + */ +#ifndef SRC_DYNAMIC_FUSION_SKETCH_GPU_COMPONENTS_CL_CLCOMPONENTLOGITS1DNORM +#define SRC_DYNAMIC_FUSION_SKETCH_GPU_COMPONENTS_CL_CLCOMPONENTLOGITS1DNORM + +#include "arm_compute/dynamic_fusion/sketch/attributes/SoftmaxAttributes.h" +#include "src/dynamic_fusion/sketch/gpu/components/IGpuKernelComponent.h" + +namespace arm_compute +{ +/** Forward declaration */ +class ITensorInfo; +namespace experimental +{ +namespace dynamic_fusion +{ +/** Forward declaration */ +template +class ArgumentPack; + +/** Forward declaration */ +class ClTemplateLogits1DNorm; + +/** Component to calculate the final step of the Softmax Layer + * where each logit value is multiplied by the inverse of the sum of the logits. + * + * 1D example: + * + * (input) src: [x1 x2 ... xn], shape: (1 x d) + * (input) sum: [x1 + x2 + ... + xn], shape: (1 x 1) + * (output) dst: [x1/sum x2/sum ... xn/sum], shape: (1 x d) + * + * This component is used by the softmax operator to get the final result. +*/ +class ClComponentLogits1DNorm final : public IGpuKernelComponent +{ +public: + /** Attributes are a set of backend-agnostic parameters that define what a component does */ + using Attributes = SoftmaxAttributes; + + /** Validate the component + * + * @param[in] properties Component properties @ref Properties + * @param[in] tensors Tensor arguments to the component + * @param[in] attributes Component attributes @ref Attributes + * + * @return Status Validation results + * + * Tensor argument names: + * - ACL_SRC_0: Input + * - ACL_SRC_1: Input + * - ACL_DST_0: Output + * + * Tensor argument constness: + * - ACL_SRC_0: Const + * - ACL_SRC_1: Const + * - ACL_DST_0: Const + * + * Valid data layouts: + * - All + * + ** Valid data type configurations: + * |ACL_SRC_0 |ACL_SRC_1 |ACL_DST_0 | + * |:----------|:----------|:----------| + * |F16 | F16 | F16 | + * |F32 | F32 | F32 | + */ + static Status validate( + const Properties &properties, + const ArgumentPack &tensors, + const Attributes &attributes); + + /** Constructor + * + * Similar to @ref ClComponentLogits1DNorm::validate() + */ + ClComponentLogits1DNorm(ComponentId id, + const Properties &properties, + const ArgumentPack &tensors, + const Attributes &attributes); + + /** Destructor */ + ~ClComponentLogits1DNorm() override; + /** Prevent instances of this class from being copy constructed */ + ClComponentLogits1DNorm(const ClComponentLogits1DNorm &component) = delete; + /** Prevent instances of this class from being copied */ + ClComponentLogits1DNorm &operator=(const ClComponentLogits1DNorm &component) = delete; + /** Allow instances of this class to be move constructed */ + ClComponentLogits1DNorm(ClComponentLogits1DNorm &&component) = default; + /** Allow instances of this class to be moved */ + ClComponentLogits1DNorm &operator=(ClComponentLogits1DNorm &&component) = default; + /** Get template writer for the component */ + const IGpuTemplateComponentWriter *template_writer() const override; + /** Get component type */ + GpuComponentType type() const override + { + return GpuComponentType::Unfusable; + } + +private: + std::unique_ptr _component_writer; +}; +} // namespace dynamic_fusion +} // namespace experimental +} // namespace arm_compute + +#endif /* SRC_DYNAMIC_FUSION_SKETCH_GPU_COMPONENTS_CL_CLCOMPONENTLOGITS1DNORM */ -- cgit v1.2.1