From ceff0f9a991b693f568c25b1e0933582301082e7 Mon Sep 17 00:00:00 2001 From: Georgios Pinitas Date: Mon, 19 Mar 2018 19:57:01 +0000 Subject: COMPMID-1016: Optimize kernel reconfiguration Optimizes kernel reconfiguration when memory manager is used. Note that this works only if every sub-sequent reconfigurations leads to sizes less than the first one. Change-Id: I08898e99929c3756147a02979b726c2380b6e11d Reviewed-on: https://eu-gerrit-1.euhpc.arm.com/125114 Reviewed-by: Anthony Barbier Tested-by: Jenkins --- tests/validation/CL/UNIT/MemoryManager.cpp | 70 ++++ .../validation/GLES_COMPUTE/UNIT/MemoryManager.cpp | 99 +++++ .../fixtures/UNIT/MemoryManagerFixture.h | 411 +++++++++++++++++++++ 3 files changed, 580 insertions(+) create mode 100644 tests/validation/CL/UNIT/MemoryManager.cpp create mode 100644 tests/validation/GLES_COMPUTE/UNIT/MemoryManager.cpp create mode 100644 tests/validation/fixtures/UNIT/MemoryManagerFixture.h (limited to 'tests/validation') diff --git a/tests/validation/CL/UNIT/MemoryManager.cpp b/tests/validation/CL/UNIT/MemoryManager.cpp new file mode 100644 index 0000000000..2129c03243 --- /dev/null +++ b/tests/validation/CL/UNIT/MemoryManager.cpp @@ -0,0 +1,70 @@ +/* + * 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. + */ +#include "arm_compute/runtime/CL/CLBufferAllocator.h" +#include "arm_compute/runtime/CL/CLMemoryGroup.h" +#include "arm_compute/runtime/CL/CLTensorAllocator.h" +#include "arm_compute/runtime/CL/functions/CLFullyConnectedLayer.h" +#include "support/ToolchainSupport.h" +#include "tests/AssetsLibrary.h" +#include "tests/CL/CLAccessor.h" +#include "tests/Globals.h" +#include "tests/Utils.h" +#include "tests/framework/Asserts.h" +#include "tests/framework/Macros.h" +#include "tests/validation/Validation.h" +#include "tests/validation/fixtures/UNIT/MemoryManagerFixture.h" + +namespace arm_compute +{ +namespace test +{ +namespace validation +{ +namespace +{ +RelativeTolerance tolerance_f32(0.05f); +} // namespace + +TEST_SUITE(CL) +TEST_SUITE(UNIT) +TEST_SUITE(MemoryManager) + +using CLBlobMemoryManagerSimpleWithinFunctionLevelFixture = BlobMemoryManagerSimpleTestCaseFixture; +FIXTURE_TEST_CASE(BlobMemoryManagerSimpleWithinFunctionLevel, + CLBlobMemoryManagerSimpleWithinFunctionLevelFixture, + framework::DatasetMode::ALL) +{ + // Validate output + validate(CLAccessor(_target), _reference, tolerance_f32); +} + +TEST_SUITE_END() +TEST_SUITE_END() +TEST_SUITE_END() +} // namespace validation +} // namespace test +} // namespace arm_compute diff --git a/tests/validation/GLES_COMPUTE/UNIT/MemoryManager.cpp b/tests/validation/GLES_COMPUTE/UNIT/MemoryManager.cpp new file mode 100644 index 0000000000..8f59a05b87 --- /dev/null +++ b/tests/validation/GLES_COMPUTE/UNIT/MemoryManager.cpp @@ -0,0 +1,99 @@ +/* + * 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. + */ +#include "arm_compute/runtime/GLES_COMPUTE/GCBufferAllocator.h" +#include "arm_compute/runtime/GLES_COMPUTE/GCMemoryGroup.h" +#include "arm_compute/runtime/GLES_COMPUTE/GCTensorAllocator.h" +#include "arm_compute/runtime/GLES_COMPUTE/functions/GCFullyConnectedLayer.h" +#include "arm_compute/runtime/GLES_COMPUTE/functions/GCSoftmaxLayer.h" +#include "support/ToolchainSupport.h" +#include "tests/AssetsLibrary.h" +#include "tests/GLES_COMPUTE/GCAccessor.h" +#include "tests/Globals.h" +#include "tests/Utils.h" +#include "tests/framework/Asserts.h" +#include "tests/framework/Macros.h" +#include "tests/validation/Validation.h" +#include "tests/validation/fixtures/UNIT/MemoryManagerFixture.h" + +namespace arm_compute +{ +namespace test +{ +namespace validation +{ +namespace +{ +RelativeTolerance tolerance_f32(0.05f); +} // namespace + +TEST_SUITE(GC) +TEST_SUITE(UNIT) +TEST_SUITE(MemoryManager) + +// Setting BlobMemoryManagerSimpleWithinFunctionLevel test +using GCBlobMemoryManagerSimpleWithinFunctionLevelFixture = BlobMemoryManagerSimpleTestCaseFixture; +FIXTURE_TEST_CASE(BlobMemoryManagerSimpleWithinFunctionLevel, + GCBlobMemoryManagerSimpleWithinFunctionLevelFixture, + framework::DatasetMode::ALL) +{ + // Validate output + validate(GCAccessor(_target), _reference, tolerance_f32); +} + +// Setting BlobMemoryManagerReconfigure test +using GCBlobMemoryManagerReconfigureFixture = BlobMemoryManagerReconfigureTestCaseFixture; +FIXTURE_TEST_CASE(BlobMemoryManagerReconfigure, + GCBlobMemoryManagerReconfigureFixture, + framework::DatasetMode::ALL) +{ + // Validate output + validate(GCAccessor(_target), _reference, tolerance_f32); +} + +// Setting BlobMemoryManagerReconfigure2 test +using GCBlobMemoryManagerReconfigure2Fixture = BlobMemoryManagerReconfigure2TestCaseFixture; +FIXTURE_TEST_CASE(BlobMemoryManagerReconfigure2, + GCBlobMemoryManagerReconfigure2Fixture, + framework::DatasetMode::ALL) +{ + // Validate output + validate(GCAccessor(_target), _reference, tolerance_f32); +} + +TEST_SUITE_END() +TEST_SUITE_END() +TEST_SUITE_END() +} // namespace validation +} // namespace test +} // namespace arm_compute diff --git a/tests/validation/fixtures/UNIT/MemoryManagerFixture.h b/tests/validation/fixtures/UNIT/MemoryManagerFixture.h new file mode 100644 index 0000000000..21ad42bf77 --- /dev/null +++ b/tests/validation/fixtures/UNIT/MemoryManagerFixture.h @@ -0,0 +1,411 @@ +/* + * Copyright (c) 2017-2018 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_TEST_UNIT_MEMORY_MANAGER +#define ARM_COMPUTE_TEST_UNIT_MEMORY_MANAGER + +#include "arm_compute/core/TensorShape.h" +#include "arm_compute/core/Types.h" +#include "arm_compute/runtime/BlobLifetimeManager.h" +#include "arm_compute/runtime/MemoryManagerOnDemand.h" +#include "arm_compute/runtime/PoolManager.h" +#include "tests/AssetsLibrary.h" +#include "tests/Globals.h" +#include "tests/IAccessor.h" +#include "tests/framework/Asserts.h" +#include "tests/framework/Fixture.h" +#include "tests/validation/Helpers.h" +#include "tests/validation/reference/FullyConnectedLayer.h" +#include "tests/validation/reference/SoftmaxLayer.h" + +namespace arm_compute +{ +namespace test +{ +namespace validation +{ +/** Simple test case to run two fully connected layers using a blob affinity memory manager + * + * Runs two fully connected layers back to back + */ +template +class BlobMemoryManagerSimpleTestCaseFixture : public framework::Fixture +{ + using T = float; + +public: + void setup() + { + _target = compute_target(); + _reference = compute_reference(); + }; + +protected: + template + void fill(U &&tensor, int i) + { + std::uniform_real_distribution<> distribution(0.5f, 1.f); + library->fill(tensor, distribution, i); + } + + TensorType compute_target() + { + auto lifetime_mgr = std::make_shared(); + auto pool_mgr = std::make_shared(); + auto mm = std::make_shared(lifetime_mgr, pool_mgr); + + // Create tensors + TensorType w1 = create_tensor(TensorShape(128U, 128U), DataType::F32, 1); + TensorType b1 = create_tensor(TensorShape(128U), DataType::F32, 1); + TensorType w2 = create_tensor(TensorShape(128U, 24U), DataType::F32, 1); + TensorType b2 = create_tensor(TensorShape(24U), DataType::F32, 1); + TensorType src = create_tensor(TensorShape(128U), DataType::F32, 1); + TensorType fc1 = create_tensor(TensorShape(128U), DataType::F32, 1); + TensorType dst = create_tensor(TensorShape(24U), DataType::F32, 1); + + // Create and configure function + FullyConnectedFunction fc_layer_1(mm); + FullyConnectedFunction fc_layer_2(mm); + fc_layer_1.configure(&src, &w1, &b1, &fc1); + fc_layer_2.configure(&fc1, &w2, &b2, &dst); + + // Allocate tensors + w1.allocator()->allocate(); + b1.allocator()->allocate(); + w2.allocator()->allocate(); + b2.allocator()->allocate(); + src.allocator()->allocate(); + fc1.allocator()->allocate(); + dst.allocator()->allocate(); + + // Finalize memory manager + mm->set_allocator(&_allocator); + mm->set_num_pools(1); + mm->finalize(); + ARM_COMPUTE_EXPECT(mm->is_finalized(), framework::LogLevel::ERRORS); + ARM_COMPUTE_EXPECT(mm->lifetime_manager()->are_all_finalized(), framework::LogLevel::ERRORS); + + // Fill tensors + fill(AccessorType(src), 0); + fill(AccessorType(w1), 1); + fill(AccessorType(b1), 2); + fill(AccessorType(w2), 3); + fill(AccessorType(b2), 4); + + // Compute functions + fc_layer_1.run(); + fc_layer_2.run(); + + return dst; + } + + SimpleTensor compute_reference() + { + // Create reference + SimpleTensor w1{ TensorShape(128U, 128U), DataType::F32 }; + SimpleTensor b1{ TensorShape(128U), DataType::F32 }; + SimpleTensor w2{ TensorShape(128U, 24U), DataType::F32 }; + SimpleTensor b2{ TensorShape(24U), DataType::F32 }; + SimpleTensor src{ TensorShape(128U), DataType::F32 }; + + // Fill reference + fill(src, 0); + fill(w1, 1); + fill(b1, 2); + fill(w2, 3); + fill(b2, 4); + + auto fc1 = reference::fully_connected_layer(src, w1, b1, TensorShape(128U)); + return reference::fully_connected_layer(fc1, w2, b2, TensorShape(24U)); + } + +protected: + TensorType _target{}; + SimpleTensor _reference{}; + AllocatorType _allocator{}; +}; + +/** Test case to run two fully connected layers using a blob affinity memory manager, + * reconfigure with different shapes and rerun + * + * Runs two fully connected layers back to back then reconfigures with different batch size and reruns + * Shapes of the reconfigure step are smaller that the initial configured step + */ +template +class BlobMemoryManagerReconfigureTestCaseFixture : public framework::Fixture +{ + using T = float; + +public: + void setup() + { + _max_batches = 8; + _cur_batches = 6; + _target = compute_target(); + _reference = compute_reference(); + }; + +protected: + template + void fill(U &&tensor, int i) + { + std::uniform_real_distribution<> distribution(0.5f, 1.f); + library->fill(tensor, distribution, i); + } + + TensorType compute_target() + { + AllocatorType allocator{}; + auto lifetime_mgr = std::make_shared(); + auto pool_mgr = std::make_shared(); + auto mm = std::make_shared(lifetime_mgr, pool_mgr); + + // Create tensors + TensorType w1 = create_tensor(TensorShape(128U, 128U), DataType::F32, 1); + TensorType b1 = create_tensor(TensorShape(128U), DataType::F32, 1); + TensorType w2 = create_tensor(TensorShape(128U, 24U), DataType::F32, 1); + TensorType b2 = create_tensor(TensorShape(24U), DataType::F32, 1); + TensorType src = create_tensor(TensorShape(128U, _max_batches), DataType::F32, 1); + TensorType fc1 = create_tensor(TensorShape(128U, _max_batches), DataType::F32, 1); + TensorType dst = create_tensor(TensorShape(24U, _max_batches), DataType::F32, 1); + + // Create and configure function + FullyConnectedFunction fc_layer_1(mm); + FullyConnectedFunction fc_layer_2(mm); + fc_layer_1.configure(&src, &w1, &b1, &fc1); + fc_layer_2.configure(&fc1, &w2, &b2, &dst); + + // Allocate persistent tensors + w1.allocator()->allocate(); + b1.allocator()->allocate(); + w2.allocator()->allocate(); + b2.allocator()->allocate(); + + // Allocate tensors (1st iteration) + src.allocator()->allocate(); + fc1.allocator()->allocate(); + dst.allocator()->allocate(); + + // Finalize memory manager + mm->set_allocator(&allocator); + mm->set_num_pools(1); + mm->finalize(); + ARM_COMPUTE_EXPECT(mm->is_finalized(), framework::LogLevel::ERRORS); + ARM_COMPUTE_EXPECT(mm->lifetime_manager()->are_all_finalized(), framework::LogLevel::ERRORS); + + // Fill tensors (1st iteration) + fill(AccessorType(src), 0); + fill(AccessorType(w1), 1); + fill(AccessorType(b1), 2); + fill(AccessorType(w2), 3); + fill(AccessorType(b2), 4); + + // Compute functions (1st iteration) + fc_layer_1.run(); + fc_layer_2.run(); + + // Update tensor shapes (2nd iteration) + auto src_padding = src.allocator()->info().padding(); + auto fc1_padding = fc1.allocator()->info().padding(); + auto dst_padding = dst.allocator()->info().padding(); + int diff = _max_batches - _cur_batches; + auto new_src_padding = PaddingSize(src_padding.top, src_padding.right, src_padding.bottom + diff, src_padding.left); + auto new_fc1_padding = PaddingSize(fc1_padding.top, fc1_padding.right, fc1_padding.bottom + diff, fc1_padding.left); + auto new_dst_padding = PaddingSize(dst_padding.top, dst_padding.right, dst_padding.bottom + diff, dst_padding.left); + src.allocator()->info().set_tensor_shape(TensorShape(128U, _cur_batches)).set_is_resizable(true).extend_padding(new_src_padding); + src.allocator()->info().set_is_resizable(false); + fc1.allocator()->info().set_tensor_shape(TensorShape(128U, _cur_batches)).set_is_resizable(true).extend_padding(new_fc1_padding); + fc1.allocator()->info().set_is_resizable(false); + dst.allocator()->info().set_tensor_shape(TensorShape(24U, _cur_batches)).set_is_resizable(true).extend_padding(new_dst_padding); + dst.allocator()->info().set_is_resizable(false); + + // Configure functions (2nd iteration) + fc_layer_1.configure(&src, &w1, &b1, &fc1, true, false, true); + fc_layer_2.configure(&fc1, &w2, &b2, &dst, true, false, true); + + // Fill tensors (2nd iteration) + fill(AccessorType(src), 5); + + // Compute functions (2nd iteration) + fc_layer_1.run(); + fc_layer_2.run(); + + return dst; + } + + SimpleTensor compute_reference() + { + // Create reference + SimpleTensor w1{ TensorShape(128U, 128U), DataType::F32 }; + SimpleTensor b1{ TensorShape(128U), DataType::F32 }; + SimpleTensor w2{ TensorShape(128U, 24U), DataType::F32 }; + SimpleTensor b2{ TensorShape(24U), DataType::F32 }; + SimpleTensor src{ TensorShape(128U, _cur_batches), DataType::F32 }; + + // Fill reference + fill(src, 5); + fill(w1, 1); + fill(b1, 2); + fill(w2, 3); + fill(b2, 4); + + auto fc1 = reference::fully_connected_layer(src, w1, b1, TensorShape(128U, _cur_batches)); + return reference::fully_connected_layer(fc1, w2, b2, TensorShape(24U, _cur_batches)); + } + +protected: + TensorType _target{}; + SimpleTensor _reference{}; + AllocatorType _allocator{}; + unsigned int _max_batches{}; + unsigned int _cur_batches{}; +}; + +/** Test case to run a fully connected layer followed by a softmax layer using a blob affinity memory manager, + * reconfigure with different shapes and rerun + * + * Runs a fully connected convolution layer followed by a softmax layer then reconfigures with different batch size and reruns + * Shapes of the reconfigure step are smaller that the initial configured step + */ +template +class BlobMemoryManagerReconfigure2TestCaseFixture : public framework::Fixture +{ + using T = float; + +public: + void setup() + { + _max_batches = 30; + _cur_batches = 3; + _target = compute_target(); + _reference = compute_reference(); + }; + +protected: + template + void fill(U &&tensor, int i) + { + std::uniform_real_distribution<> distribution(0.5f, 1.f); + library->fill(tensor, distribution, i); + } + + TensorType compute_target() + { + AllocatorType allocator{}; + auto lifetime_mgr = std::make_shared(); + auto pool_mgr = std::make_shared(); + auto mm = std::make_shared(lifetime_mgr, pool_mgr); + + // Create tensors + TensorType w = create_tensor(TensorShape(112U, 8U), DataType::F32, 1); + TensorType b = create_tensor(TensorShape(8U), DataType::F32, 1); + TensorType src = create_tensor(TensorShape(1U, 1U, 112U, _max_batches), DataType::F32, 1); + TensorType fc = create_tensor(TensorShape(8U, _max_batches), DataType::F32, 1); + TensorType dst = create_tensor(TensorShape(8U, _max_batches), DataType::F32, 1); + + // Create and configure function + FullyConnectedFunction fc_layer(mm); + SoftmaxFunction smx_layer(mm); + fc_layer.configure(&src, &w, &b, &fc); + smx_layer.configure(&fc, &dst); + + // Allocate persistent tensors + w.allocator()->allocate(); + b.allocator()->allocate(); + + // Allocate tensors (1st iteration) + src.allocator()->allocate(); + fc.allocator()->allocate(); + dst.allocator()->allocate(); + + // Finalize memory manager + mm->set_allocator(&allocator); + mm->set_num_pools(1); + mm->finalize(); + ARM_COMPUTE_EXPECT(mm->is_finalized(), framework::LogLevel::ERRORS); + ARM_COMPUTE_EXPECT(mm->lifetime_manager()->are_all_finalized(), framework::LogLevel::ERRORS); + + // Fill tensors (1st iteration) + fill(AccessorType(src), 0); + fill(AccessorType(w), 1); + fill(AccessorType(b), 2); + + // Compute functions (1st iteration) + fc_layer.run(); + smx_layer.run(); + + // Get padding requirements + auto fc_padding = fc.allocator()->info().padding(); + + // Run rest iterations + for(int i = _max_batches; i >= static_cast(_cur_batches); --i) + { + int diff = _max_batches - i; + auto new_fc_padding = PaddingSize(fc_padding.top, fc_padding.right, fc_padding.bottom + diff, fc_padding.left); + src.allocator()->info().set_tensor_shape(TensorShape(1U, 1U, 112U, i)); + fc.allocator()->info().set_tensor_shape(TensorShape(8U, i)).set_is_resizable(true).extend_padding(new_fc_padding); + fc.allocator()->info().set_is_resizable(false); + dst.allocator()->info().set_tensor_shape(TensorShape(8U, i)); + + // Configure functions + fc_layer.configure(&src, &w, &b, &fc, true, false, true); + smx_layer.configure(&fc, &dst); + + // Fill tensors + fill(AccessorType(src), 3); + + // Compute functions + fc_layer.run(); + smx_layer.run(); + } + + return dst; + } + + SimpleTensor compute_reference() + { + // Create reference + SimpleTensor w{ TensorShape(112U, 8U), DataType::F32 }; + SimpleTensor b{ TensorShape(8U), DataType::F32 }; + SimpleTensor src{ TensorShape(1U, 1U, 112U, _cur_batches), DataType::F32 }; + + // Fill reference + fill(src, 3); + fill(w, 1); + fill(b, 2); + + auto fc = reference::fully_connected_layer(src, w, b, TensorShape(8U, _cur_batches)); + return reference::softmax_layer(fc, 1.f); + } + +protected: + TensorType _target{}; + SimpleTensor _reference{}; + AllocatorType _allocator{}; + unsigned int _max_batches{}; + unsigned int _cur_batches{}; +}; +} // namespace validation +} // namespace test +} // namespace arm_compute +#endif /* ARM_COMPUTE_TEST_UNIT_MEMORY_MANAGER */ -- cgit v1.2.1