From 9e9cbafa9e6cc6b543c89a96d52fc9c5fde04ceb Mon Sep 17 00:00:00 2001 From: Michalis Spyrou Date: Thu, 15 Mar 2018 14:41:34 +0000 Subject: COMPMID-1004 GLES: Add memory manager to GLES functions Change-Id: I80fc9c0dd02afd79b501abde751036f9599b7bf2 Reviewed-on: https://eu-gerrit-1.euhpc.arm.com/125103 Tested-by: Jenkins Reviewed-by: Georgios Pinitas --- .../runtime/GLES_COMPUTE/GCBufferAllocator.h | 48 ++++++++++++++++++++ arm_compute/runtime/GLES_COMPUTE/GCMemoryGroup.h | 48 ++++++++++++++++++++ arm_compute/runtime/GLES_COMPUTE/GCTensor.h | 4 +- .../runtime/GLES_COMPUTE/GCTensorAllocator.h | 52 ++++++++++++++-------- .../GLES_COMPUTE/functions/GCConvolutionLayer.h | 4 +- .../GLES_COMPUTE/functions/GCFullyConnectedLayer.h | 4 +- .../runtime/GLES_COMPUTE/functions/GCGEMM.h | 4 +- .../GLES_COMPUTE/functions/GCNormalizationLayer.h | 6 ++- .../GLES_COMPUTE/functions/GCSoftmaxLayer.h | 6 ++- 9 files changed, 148 insertions(+), 28 deletions(-) create mode 100644 arm_compute/runtime/GLES_COMPUTE/GCBufferAllocator.h create mode 100644 arm_compute/runtime/GLES_COMPUTE/GCMemoryGroup.h (limited to 'arm_compute/runtime') diff --git a/arm_compute/runtime/GLES_COMPUTE/GCBufferAllocator.h b/arm_compute/runtime/GLES_COMPUTE/GCBufferAllocator.h new file mode 100644 index 0000000000..103ac37bdd --- /dev/null +++ b/arm_compute/runtime/GLES_COMPUTE/GCBufferAllocator.h @@ -0,0 +1,48 @@ +/* + * Copyright (c) 2018 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. + *gc + * 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 __ARM_COMPUTE_GCBUFFERALLOCATOR_H__ +#define __ARM_COMPUTE_GCBUFFERALLOCATOR_H__ + +#include "arm_compute/runtime/IAllocator.h" + +#include "arm_compute/core/GLES_COMPUTE/OpenGLES.h" +#include "arm_compute/runtime/GLES_COMPUTE/GCScheduler.h" + +#include + +namespace arm_compute +{ +/** Default gles buffer allocator implementation */ +class GCBufferAllocator : public IAllocator +{ +public: + /** Default constructor */ + GCBufferAllocator() = default; + + // Inherited methods overridden: + void *allocate(size_t size, size_t alignment) override; + void free(void *ptr) override; +}; +} // arm_compute +#endif /*__ARM_COMPUTE_GCBUFFERALLOCATOR_H__ */ diff --git a/arm_compute/runtime/GLES_COMPUTE/GCMemoryGroup.h b/arm_compute/runtime/GLES_COMPUTE/GCMemoryGroup.h new file mode 100644 index 0000000000..485aa0e024 --- /dev/null +++ b/arm_compute/runtime/GLES_COMPUTE/GCMemoryGroup.h @@ -0,0 +1,48 @@ +/* + * Copyright (c) 2018 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 __ARM_COMPUTE_GCMEMORYGROUP_H__ +#define __ARM_COMPUTE_GCMEMORYGROUP_H__ + +#include "arm_compute/runtime/MemoryGroupBase.h" + +#include "arm_compute/core/GLES_COMPUTE/OpenGLES.h" +#include "arm_compute/core/utils/misc/Cast.h" +#include "arm_compute/runtime/GLES_COMPUTE/GCTensor.h" + +namespace arm_compute +{ +using GCMemoryGroup = MemoryGroupBase; + +template <> +inline void MemoryGroupBase::associate_memory_group(GCTensor *obj) +{ + ARM_COMPUTE_ERROR_ON(obj == nullptr); + ARM_COMPUTE_ERROR_ON(dynamic_cast(obj->allocator()) == nullptr); + + auto allocator = arm_compute::utils::cast::polymorphic_downcast(obj->allocator()); + ARM_COMPUTE_ERROR_ON(allocator == nullptr); + allocator->set_associated_memory_group(this); +} +} // arm_compute +#endif /*__ARM_COMPUTE_GCMEMORYGROUP_H__ */ diff --git a/arm_compute/runtime/GLES_COMPUTE/GCTensor.h b/arm_compute/runtime/GLES_COMPUTE/GCTensor.h index 3e51f9908f..0f5f194266 100644 --- a/arm_compute/runtime/GLES_COMPUTE/GCTensor.h +++ b/arm_compute/runtime/GLES_COMPUTE/GCTensor.h @@ -1,5 +1,5 @@ /* - * Copyright (c) 2017 ARM Limited. + * Copyright (c) 2017-2018 ARM Limited. * * SPDX-License-Identifier: MIT * @@ -91,7 +91,7 @@ protected: void do_unmap() override; private: - mutable GCTensorAllocator _allocator; + mutable GCTensorAllocator _allocator; /**< Instance of the OpenGL ES tensor allocator */ }; using GCImage = GCTensor; diff --git a/arm_compute/runtime/GLES_COMPUTE/GCTensorAllocator.h b/arm_compute/runtime/GLES_COMPUTE/GCTensorAllocator.h index ce52cbbbdc..fc14f04ac2 100644 --- a/arm_compute/runtime/GLES_COMPUTE/GCTensorAllocator.h +++ b/arm_compute/runtime/GLES_COMPUTE/GCTensorAllocator.h @@ -1,5 +1,5 @@ /* - * Copyright (c) 2017 ARM Limited. + * Copyright (c) 2017-2018 ARM Limited. * * SPDX-License-Identifier: MIT * @@ -27,17 +27,37 @@ #include "arm_compute/core/GLES_COMPUTE/OpenGLES.h" #include "arm_compute/runtime/ITensorAllocator.h" +#include "arm_compute/runtime/MemoryGroupBase.h" #include namespace arm_compute { +class GCTensor; +template +class MemoryGroupBase; +using GCMemoryGroup = MemoryGroupBase; + +class GLBufferWrapper +{ +public: + GLBufferWrapper() + : _ssbo_name(0) + { + ARM_COMPUTE_GL_CHECK(glGenBuffers(1, &_ssbo_name)); + } + ~GLBufferWrapper() + { + ARM_COMPUTE_GL_CHECK(glDeleteBuffers(1, &_ssbo_name)); + } + GLuint _ssbo_name; +}; /** Basic implementation of a GLES memory tensor allocator. */ class GCTensorAllocator : public ITensorAllocator { public: /** Default constructor. */ - GCTensorAllocator(); + GCTensorAllocator(GCTensor *owner = nullptr); /** Prevent instances of this class from being copied (As this class contains pointers). */ GCTensorAllocator(const GCTensorAllocator &) = delete; @@ -52,7 +72,7 @@ public: GCTensorAllocator &operator=(GCTensorAllocator &&) = default; /** Default destructor */ - ~GCTensorAllocator() = default; + ~GCTensorAllocator(); /** Interface to be implemented by the child class to return the pointer to the mapped data. */ uint8_t *data(); @@ -95,6 +115,12 @@ public: */ void free() override; + /** Associates the tensor with a memory group + * + * @param[in] associated_memory_group Memory group to associate the tensor with + */ + void set_associated_memory_group(GCMemoryGroup *associated_memory_group); + protected: /** Call map() on the SSBO. * @@ -106,22 +132,10 @@ protected: void unlock() override; private: - class GLBufferWrapper - { - public: - GLBufferWrapper() - : _ssbo_name(0) - { - ARM_COMPUTE_GL_CHECK(glGenBuffers(1, &_ssbo_name)); - } - ~GLBufferWrapper() - { - ARM_COMPUTE_GL_CHECK(glDeleteBuffers(1, &_ssbo_name)); - } - GLuint _ssbo_name; - }; - std::unique_ptr _gl_buffer; - uint8_t *_mapping; + GCMemoryGroup *_associated_memory_group; /**< Registered memory group */ + std::unique_ptr _gl_buffer; /**< OpenGL ES object containing the tensor data. */ + uint8_t *_mapping; /**< Pointer to the CPU mapping of the OpenGL ES buffer. */ + GCTensor *_owner; /**< Owner of the allocator */ }; } diff --git a/arm_compute/runtime/GLES_COMPUTE/functions/GCConvolutionLayer.h b/arm_compute/runtime/GLES_COMPUTE/functions/GCConvolutionLayer.h index e3fa98e6e7..2bac982d0c 100644 --- a/arm_compute/runtime/GLES_COMPUTE/functions/GCConvolutionLayer.h +++ b/arm_compute/runtime/GLES_COMPUTE/functions/GCConvolutionLayer.h @@ -33,6 +33,7 @@ #include "arm_compute/core/GLES_COMPUTE/kernels/GCIm2ColKernel.h" #include "arm_compute/core/GLES_COMPUTE/kernels/GCWeightsReshapeKernel.h" #include "arm_compute/core/Types.h" +#include "arm_compute/runtime/GLES_COMPUTE/GCMemoryGroup.h" #include "arm_compute/runtime/GLES_COMPUTE/GCTensor.h" #include "arm_compute/runtime/IFunction.h" @@ -83,7 +84,7 @@ class GCConvolutionLayer : public IFunction { public: /** Default constructor */ - GCConvolutionLayer(); + GCConvolutionLayer(std::shared_ptr memory_manager = nullptr); /** Set the input and output tensors. * @@ -115,6 +116,7 @@ private: void configure_mm(const IGCTensor *input, const IGCTensor *weights, IGCTensor *output, bool is_interleaved_transposed = true); private: + GCMemoryGroup _memory_group; GCConvolutionLayerReshapeWeights _reshape_weights; GCIm2ColKernel _input_im2col_kernel; GCGEMMInterleave4x4Kernel _input_interleave_kernel; diff --git a/arm_compute/runtime/GLES_COMPUTE/functions/GCFullyConnectedLayer.h b/arm_compute/runtime/GLES_COMPUTE/functions/GCFullyConnectedLayer.h index 3ba44f59cb..81be1de21a 100644 --- a/arm_compute/runtime/GLES_COMPUTE/functions/GCFullyConnectedLayer.h +++ b/arm_compute/runtime/GLES_COMPUTE/functions/GCFullyConnectedLayer.h @@ -28,6 +28,7 @@ #include "arm_compute/core/GLES_COMPUTE/kernels/GCGEMMMatrixMultiplyKernel.h" #include "arm_compute/core/GLES_COMPUTE/kernels/GCIm2ColKernel.h" #include "arm_compute/core/GLES_COMPUTE/kernels/GCTransposeKernel.h" +#include "arm_compute/runtime/GLES_COMPUTE/GCMemoryGroup.h" #include "arm_compute/runtime/GLES_COMPUTE/GCTensor.h" #include "arm_compute/runtime/GLES_COMPUTE/IGCSimpleFunction.h" @@ -63,7 +64,7 @@ class GCFullyConnectedLayer : public IFunction { public: /** Constructor */ - GCFullyConnectedLayer(); + GCFullyConnectedLayer(std::shared_ptr memory_manager = nullptr); /** Set the input and output tensors. * * @param[in] input Source tensor. Data type supported: F16/F32. @@ -82,6 +83,7 @@ private: void configure_fc_fc(const IGCTensor *input, const IGCTensor *weights, IGCTensor *output); void configure_conv_fc(const IGCTensor *input, const IGCTensor *weights, IGCTensor *output); + GCMemoryGroup _memory_group; GCIm2ColKernel _im2col_kernel; GCFullyConnectedLayerReshapeWeights _reshape_weights_kernel; GCGEMMMatrixMultiplyKernel _mm_kernel; diff --git a/arm_compute/runtime/GLES_COMPUTE/functions/GCGEMM.h b/arm_compute/runtime/GLES_COMPUTE/functions/GCGEMM.h index 8ddfae1169..31ad0abaa0 100644 --- a/arm_compute/runtime/GLES_COMPUTE/functions/GCGEMM.h +++ b/arm_compute/runtime/GLES_COMPUTE/functions/GCGEMM.h @@ -29,6 +29,7 @@ #include "arm_compute/core/GLES_COMPUTE/kernels/GCGEMMMatrixAdditionKernel.h" #include "arm_compute/core/GLES_COMPUTE/kernels/GCGEMMMatrixMultiplyKernel.h" #include "arm_compute/core/GLES_COMPUTE/kernels/GCGEMMTranspose1xWKernel.h" +#include "arm_compute/runtime/GLES_COMPUTE/GCMemoryGroup.h" #include "arm_compute/runtime/GLES_COMPUTE/GCTensor.h" #include "arm_compute/runtime/IFunction.h" @@ -48,7 +49,7 @@ class GCGEMM : public IFunction { public: /** Default constructor. */ - GCGEMM(); + GCGEMM(std::shared_ptr memory_manager = nullptr); /** Initialise the kernel's inputs and output * @@ -73,6 +74,7 @@ public: void run() override; private: + GCMemoryGroup _memory_group; GCGEMMInterleave4x4Kernel _interleave_kernel; GCGEMMTranspose1xWKernel _transpose_kernel; GCGEMMMatrixMultiplyKernel _mm_kernel; diff --git a/arm_compute/runtime/GLES_COMPUTE/functions/GCNormalizationLayer.h b/arm_compute/runtime/GLES_COMPUTE/functions/GCNormalizationLayer.h index d080a2f7b9..adc8157f0e 100644 --- a/arm_compute/runtime/GLES_COMPUTE/functions/GCNormalizationLayer.h +++ b/arm_compute/runtime/GLES_COMPUTE/functions/GCNormalizationLayer.h @@ -1,5 +1,5 @@ /* - * Copyright (c) 2017 ARM Limited. + * Copyright (c) 2017-2018 ARM Limited. * * SPDX-License-Identifier: MIT * @@ -29,6 +29,7 @@ #include "arm_compute/core/GLES_COMPUTE/kernels/GCFillBorderKernel.h" #include "arm_compute/core/GLES_COMPUTE/kernels/GCNormalizationLayerKernel.h" #include "arm_compute/core/GLES_COMPUTE/kernels/GCPixelWiseMultiplicationKernel.h" +#include "arm_compute/runtime/GLES_COMPUTE/GCMemoryGroup.h" #include "arm_compute/runtime/GLES_COMPUTE/GCTensor.h" #include "arm_compute/core/Types.h" @@ -48,7 +49,7 @@ class GCNormalizationLayer : public IFunction { public: /** Default constructor */ - GCNormalizationLayer(); + GCNormalizationLayer(std::shared_ptr memory_manager = nullptr); /** Set the input and output tensors. * * @param[in] input Source tensor. 3 lower dims represent a single input with dimensions [width, height, IFM], @@ -62,6 +63,7 @@ public: void run() override; private: + GCMemoryGroup _memory_group; GCTensor _squared_input; /**< The intermediate buffer which stores results of squaring input*/ GCNormalizationLayerKernel _norm_kernel; /**< Normalization layer kernel to run */ GCPixelWiseMultiplicationKernel _multiply_kernel; /**< Pixel multiplication kernel to run */ diff --git a/arm_compute/runtime/GLES_COMPUTE/functions/GCSoftmaxLayer.h b/arm_compute/runtime/GLES_COMPUTE/functions/GCSoftmaxLayer.h index e7f8d5053a..1011c9a2ef 100644 --- a/arm_compute/runtime/GLES_COMPUTE/functions/GCSoftmaxLayer.h +++ b/arm_compute/runtime/GLES_COMPUTE/functions/GCSoftmaxLayer.h @@ -1,5 +1,5 @@ /* - * Copyright (c) 2017 ARM Limited. + * Copyright (c) 2017-2018 ARM Limited. * * SPDX-License-Identifier: MIT * @@ -25,6 +25,7 @@ #define __ARM_COMPUTE_GCSOFTMAXLAYER_H__ #include "arm_compute/core/GLES_COMPUTE/kernels/GCSoftmaxLayerKernel.h" +#include "arm_compute/runtime/GLES_COMPUTE/GCMemoryGroup.h" #include "arm_compute/runtime/GLES_COMPUTE/GCTensor.h" #include "arm_compute/runtime/IFunction.h" @@ -46,7 +47,7 @@ class GCSoftmaxLayer : public IFunction { public: /** Constructor */ - GCSoftmaxLayer(); + GCSoftmaxLayer(std::shared_ptr memory_manager = nullptr); /** Set the input and output tensors. * * @param[in] input Source tensor. Data types supported: F16/F32 @@ -59,6 +60,7 @@ public: void run() override; private: + GCMemoryGroup _memory_group; GCLogits1DMaxKernel _max_kernel; GCLogits1DShiftExpSumKernel _shift_exp_sum_kernel; GCLogits1DNormKernel _norm_kernel; -- cgit v1.2.1