diff options
| -rw-r--r-- | SConscript | 1 | ||||
| -rw-r--r-- | arm_compute/core/CL/CLTypes.h | 13 | ||||
| -rw-r--r-- | arm_compute/core/CL/ICLKernel.h | 9 | ||||
| -rw-r--r-- | arm_compute/core/CL/kernels/CLDirectConvolutionLayerKernel.h | 4 | ||||
| -rw-r--r-- | arm_compute/core/utils/misc/Signal.h | 99 | ||||
| -rw-r--r-- | arm_compute/runtime/CL/CLScheduler.h | 18 | ||||
| -rw-r--r-- | arm_compute/runtime/CL/CLTuner.h | 3 | ||||
| -rw-r--r-- | arm_compute/runtime/CL/ICLTuner.h | 16 | ||||
| -rw-r--r-- | arm_compute/runtime/CL/tuners/BifrostTuner.h | 43 | ||||
| -rw-r--r-- | src/core/CL/kernels/CLDirectConvolutionLayerKernel.cpp | 73 | ||||
| -rw-r--r-- | src/graph/Graph.cpp | 1 | ||||
| -rw-r--r-- | src/runtime/CL/CLScheduler.cpp | 2 | ||||
| -rw-r--r-- | src/runtime/CL/CLTuner.cpp | 7 | ||||
| -rw-r--r-- | src/runtime/CL/functions/CLDirectConvolutionLayer.cpp | 5 | ||||
| -rw-r--r-- | src/runtime/CL/tuners/BifrostTuner.cpp | 143 | ||||
| -rw-r--r-- | tests/validation/CL/UNIT/Tuner.cpp | 80 |
16 files changed, 426 insertions, 91 deletions
diff --git a/SConscript b/SConscript index f90ee3e536..5a4b6ce5ed 100644 --- a/SConscript +++ b/SConscript @@ -186,6 +186,7 @@ if env['opencl']: runtime_files += Glob('src/runtime/CL/*.cpp') runtime_files += Glob('src/runtime/CL/functions/*.cpp') + runtime_files += Glob('src/runtime/CL/tuners/*.cpp') graph2_files += Glob('src/graph2/backends/CL/*.cpp') diff --git a/arm_compute/core/CL/CLTypes.h b/arm_compute/core/CL/CLTypes.h index a9d5fdd063..14c31fa396 100644 --- a/arm_compute/core/CL/CLTypes.h +++ b/arm_compute/core/CL/CLTypes.h @@ -60,5 +60,16 @@ enum class CLVersion CL20, /* the OpenCL 2.0 and above */ UNKNOWN /* unkown version */ }; -} + +/** OpenCL device options */ +struct CLDeviceOptions +{ + std::string name; /**< Device name */ + std::string extensions; /**< List of supported extensions */ + std::string ddk_version; /**< DDK version */ + GPUTarget gpu_target; /**< GPU target architecture/instance */ + size_t num_cores; /**< Number of cores */ + size_t cache_size; /**< Cache size */ +}; +} // namespace arm_compute #endif /* __ARM_COMPUTE_CL_TYPES_H__ */ diff --git a/arm_compute/core/CL/ICLKernel.h b/arm_compute/core/CL/ICLKernel.h index c7d0c2156b..e6700cd61e 100644 --- a/arm_compute/core/CL/ICLKernel.h +++ b/arm_compute/core/CL/ICLKernel.h @@ -198,6 +198,15 @@ public: _lws_hint = lws_hint; } + /** Return the Local-Workgroup-Size hint + * + * @return Current lws hint + */ + cl::NDRange lws_hint() const + { + return _lws_hint; + } + /** Get the configuration ID * * @note The configuration ID can be used by the caller to distinguish different calls of the same OpenCL kernel diff --git a/arm_compute/core/CL/kernels/CLDirectConvolutionLayerKernel.h b/arm_compute/core/CL/kernels/CLDirectConvolutionLayerKernel.h index d47b7da213..eb1bf58b1b 100644 --- a/arm_compute/core/CL/kernels/CLDirectConvolutionLayerKernel.h +++ b/arm_compute/core/CL/kernels/CLDirectConvolutionLayerKernel.h @@ -1,5 +1,5 @@ /* - * Copyright (c) 2017 ARM Limited. + * Copyright (c) 2017-2018 ARM Limited. * * SPDX-License-Identifier: MIT * @@ -88,7 +88,7 @@ public: void run(const Window &window, cl::CommandQueue &queue) override; BorderSize border_size() const override; -private: +public: const ICLTensor *_input; const ICLTensor *_biases; const ICLTensor *_weights; diff --git a/arm_compute/core/utils/misc/Signal.h b/arm_compute/core/utils/misc/Signal.h new file mode 100644 index 0000000000..71f13951b9 --- /dev/null +++ b/arm_compute/core/utils/misc/Signal.h @@ -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. + */ +#ifndef __ARM_COMPUTE_MISC_SIGNAL_H__ +#define __ARM_COMPUTE_MISC_SIGNAL_H__ + +#include <functional> + +namespace arm_compute +{ +namespace utils +{ +namespace signal +{ +namespace detail +{ +/** Base signal class */ +template <typename SignalType> +class SignalImpl; + +/** Signal class function specialization */ +template <typename ReturnType, typename... Args> +class SignalImpl<ReturnType(Args...)> +{ +public: + using Callback = std::function<ReturnType(Args...)>; + +public: + /** Default Constructor */ + SignalImpl() = default; + + /** Connects signal + * + * @param[in] cb Callback to connect the signal with + */ + void connect(const Callback &cb) + { + _cb = cb; + } + + /** Disconnects the signal */ + void disconnect() + { + _cb = nullptr; + } + + /** Checks if the signal is connected + * + * @return True if there is a connection else false + */ + bool connected() const + { + return (_cb != nullptr); + } + + /** Calls the connected callback + * + * @param[in] args Callback arguments + */ + void operator()(Args &&... args) + { + if(_cb) + { + _cb(std::forward<Args>(args)...); + } + } + +private: + Callback _cb{}; /**< Signal callback */ +}; +} // namespace detail + +/** Signal alias */ +template <class T> +using Signal = detail::SignalImpl<T>; +} // namespace signal +} // namespace utils +} // namespace arm_compute +#endif /* __ARM_COMPUTE_MISC_SIGNAL_H__ */ diff --git a/arm_compute/runtime/CL/CLScheduler.h b/arm_compute/runtime/CL/CLScheduler.h index 89c3bc1553..e1e7ff637f 100644 --- a/arm_compute/runtime/CL/CLScheduler.h +++ b/arm_compute/runtime/CL/CLScheduler.h @@ -30,7 +30,7 @@ #include "arm_compute/core/CL/OpenCL.h" #include "arm_compute/core/Error.h" #include "arm_compute/core/Types.h" -#include "arm_compute/runtime/CL/CLTuner.h" +#include "arm_compute/runtime/CL/ICLTuner.h" #if defined(ARM_COMPUTE_DEBUG_ENABLED) namespace @@ -194,17 +194,19 @@ public: return event; } -private: - /** Tune OpenCL kernel - * - * @note This method uses a brute force approach to find the optimal LWS + /** Tunes OpenCL kernel * * @param[in] kernel Kernel to tune - * - * @return The optimal LWS for the specified kernel */ - cl::NDRange tune_kernel(ICLKernel &kernel); + void tune_kernel_static(ICLKernel &kernel) + { + if(_cl_tuner != nullptr) + { + _cl_tuner->tune_kernel_static(kernel); + } + } +private: /** Flag to ensure symbols initialisation is happening before Scheduler creation */ static std::once_flag _initialize_symbols; diff --git a/arm_compute/runtime/CL/CLTuner.h b/arm_compute/runtime/CL/CLTuner.h index c1fbfd249c..f789500de3 100644 --- a/arm_compute/runtime/CL/CLTuner.h +++ b/arm_compute/runtime/CL/CLTuner.h @@ -99,7 +99,8 @@ public: void save_to_file(const std::string &filename) const; // Inherited methods overridden: - void tune_kernel(ICLKernel &kernel) override; + void tune_kernel_static(ICLKernel &kernel) override; + void tune_kernel_dynamic(ICLKernel &kernel) override; /** Is the kernel_event set ? * diff --git a/arm_compute/runtime/CL/ICLTuner.h b/arm_compute/runtime/CL/ICLTuner.h index c71835c0aa..95b22b5b7e 100644 --- a/arm_compute/runtime/CL/ICLTuner.h +++ b/arm_compute/runtime/CL/ICLTuner.h @@ -1,5 +1,5 @@ /* - * Copyright (c) 2017 ARM Limited. + * Copyright (c) 2017-2018 ARM Limited. * * SPDX-License-Identifier: MIT * @@ -34,11 +34,21 @@ class ICLTuner public: /** Virtual destructor */ virtual ~ICLTuner() = default; - /** Tune OpenCL kernel + /** Tune OpenCL kernel statically + * + * @note Tuning is performed using only kernel and tensor metadata, + * thus can be performed when memory is not available + * + * @param[in] kernel Kernel to tune + */ + virtual void tune_kernel_static(ICLKernel &kernel) = 0; + /** Tune OpenCL kernel dynamically + * + * @note Tuning requires memory to be available on all kernel tensors and objects in order to be performed * * @param[in] kernel Kernel to tune */ - virtual void tune_kernel(ICLKernel &kernel) = 0; + virtual void tune_kernel_dynamic(ICLKernel &kernel) = 0; }; } #endif /*__ARM_COMPUTE_ICLTUNER_H__ */ diff --git a/arm_compute/runtime/CL/tuners/BifrostTuner.h b/arm_compute/runtime/CL/tuners/BifrostTuner.h new file mode 100644 index 0000000000..080a37c152 --- /dev/null +++ b/arm_compute/runtime/CL/tuners/BifrostTuner.h @@ -0,0 +1,43 @@ +/* + * 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_TUNERS_BIFROST_TUNER_H__ +#define __ARM_COMPUTE_TUNERS_BIFROST_TUNER_H__ + +#include "arm_compute/runtime/CL/ICLTuner.h" + +namespace arm_compute +{ +namespace tuners +{ +/** Bifrost based OpenCL tuner implementation */ +class BifrostTuner final : public ICLTuner +{ +public: + // Inherited overriden methods + void tune_kernel_static(ICLKernel &kernel) override; + void tune_kernel_dynamic(ICLKernel &kernel) override; +}; +} // namespace tuners +} // namespace arm_compute +#endif /*__ARM_COMPUTE_TUNERS_BIFROST_TUNER_H__ */ diff --git a/src/core/CL/kernels/CLDirectConvolutionLayerKernel.cpp b/src/core/CL/kernels/CLDirectConvolutionLayerKernel.cpp index 56ac0c7250..b5526c4fca 100644 --- a/src/core/CL/kernels/CLDirectConvolutionLayerKernel.cpp +++ b/src/core/CL/kernels/CLDirectConvolutionLayerKernel.cpp @@ -315,79 +315,6 @@ void CLDirectConvolutionLayerKernel::configure(const ICLTensor *input, const ICL kernel_name << "_f32_bifrost"; _kernel = static_cast<cl::Kernel>(CLKernelLibrary::get().create_kernel(kernel_name.str(), build_options.options())); - - // Through extensive experimentation with over 30 representative tensor - // shapes, we found a small number of local work size configurations - // that result in nearly optimal execution times. Selecting the right - // lws for a given shape, however, required a complex decision tree, - // until we constructed a simple feature as described below. - // - // We started from the number of multiply-accumulate operations for a - // convolution layer, which is equal to the product of the input - // dimensions 0..2 and the weights dimensions 0..2. Unfortunately, - // this resulted in ties between distinct shapes that required distinct - // lws configurations. Replacing the width of the input with the kernel - // size, however, resulted in nearly optimal predictions. We use underscores - // in variable names to indicate when they are intentionally misleading. - const size_t product_of_weights_dimensions = weights->info()->dimension(0) * weights->info()->dimension(1) * weights->info()->dimension(2); - const size_t product_of_input_dimensions_ = input->info()->dimension(0) * weights->info()->dimension(1) * input->info()->dimension(2); - const float mega_ops_ = 1e-6 * product_of_weights_dimensions * product_of_input_dimensions_; - - switch(kernel_size) - { - case 1: - { - if(mega_ops_ < 1.f) - { - _lws_hint = cl::NDRange(1, 1, 8); - } - else if(mega_ops_ < 7.f) - { - _lws_hint = cl::NDRange(1, 1, 4); - } - else - { - _lws_hint = cl::NDRange(1, 1, 2); - } - break; - } - case 3: - { - if(mega_ops_ < 1.f) - { - _lws_hint = cl::NDRange(1, 1, 8); - } - else if(mega_ops_ < 13.f) - { - _lws_hint = cl::NDRange(2, 1, 4); - } - else if(mega_ops_ < 50.f) - { - _lws_hint = cl::NDRange(3, 1, 4); - } - else - { - _lws_hint = cl::NDRange(2, 1, 6); - } - break; - } - case 5: - { - if(mega_ops_ < 2.f || mega_ops_ > 80.f) - { - _lws_hint = cl::NDRange(2, 1, 4); - } - else - { - _lws_hint = cl::NDRange(2, 1, 8); - } - break; - } - default: - { - ARM_COMPUTE_ERROR("Kernel size not optimized for Bifrost"); - } - } } else { diff --git a/src/graph/Graph.cpp b/src/graph/Graph.cpp index 2fe3a90aef..47bd672114 100644 --- a/src/graph/Graph.cpp +++ b/src/graph/Graph.cpp @@ -30,6 +30,7 @@ #include "arm_compute/graph/Tensor.h" #include "arm_compute/runtime/CL/CLScheduler.h" #include "arm_compute/runtime/CL/CLTensor.h" +#include "arm_compute/runtime/CL/CLTuner.h" #include "arm_compute/runtime/Tensor.h" #include "support/ToolchainSupport.h" diff --git a/src/runtime/CL/CLScheduler.cpp b/src/runtime/CL/CLScheduler.cpp index 65292fe837..2a5d836c33 100644 --- a/src/runtime/CL/CLScheduler.cpp +++ b/src/runtime/CL/CLScheduler.cpp @@ -52,7 +52,7 @@ void CLScheduler::enqueue(ICLKernel &kernel, bool flush) if(_cl_tuner != nullptr) { // Tune the OpenCL kernel - _cl_tuner->tune_kernel(kernel); + _cl_tuner->tune_kernel_dynamic(kernel); } // Run kernel diff --git a/src/runtime/CL/CLTuner.cpp b/src/runtime/CL/CLTuner.cpp index df8e255356..17a62ab46e 100644 --- a/src/runtime/CL/CLTuner.cpp +++ b/src/runtime/CL/CLTuner.cpp @@ -113,7 +113,12 @@ bool CLTuner::tune_new_kernels() const return _tune_new_kernels; } -void CLTuner::tune_kernel(ICLKernel &kernel) +void CLTuner::tune_kernel_static(ICLKernel &kernel) +{ + ARM_COMPUTE_UNUSED(kernel); +} + +void CLTuner::tune_kernel_dynamic(ICLKernel &kernel) { // Get the configuration ID from the kernel const std::string &config_id = kernel.config_id(); diff --git a/src/runtime/CL/functions/CLDirectConvolutionLayer.cpp b/src/runtime/CL/functions/CLDirectConvolutionLayer.cpp index d6a335c1ec..c48865a0cc 100644 --- a/src/runtime/CL/functions/CLDirectConvolutionLayer.cpp +++ b/src/runtime/CL/functions/CLDirectConvolutionLayer.cpp @@ -1,5 +1,5 @@ /* - * Copyright (c) 2017 ARM Limited. + * Copyright (c) 2017-2018 ARM Limited. * * SPDX-License-Identifier: MIT * @@ -52,6 +52,9 @@ void CLDirectConvolutionLayer::configure(ICLTensor *input, const ICLTensor *weig zero_value = PixelValue(static_cast<uint8_t>(input->info()->quantization_info().offset)); } _input_border_handler.configure(input, _direct_conv_kernel.border_size(), BorderMode::CONSTANT, zero_value); + + // Tune kernels + CLScheduler::get().tune_kernel_static(_direct_conv_kernel); } Status CLDirectConvolutionLayer::validate(const ITensorInfo *input, const ITensorInfo *weights, const ITensorInfo *biases, const ITensorInfo *output, const PadStrideInfo &conv_info) diff --git a/src/runtime/CL/tuners/BifrostTuner.cpp b/src/runtime/CL/tuners/BifrostTuner.cpp new file mode 100644 index 0000000000..c0ebd24afe --- /dev/null +++ b/src/runtime/CL/tuners/BifrostTuner.cpp @@ -0,0 +1,143 @@ +/* + * 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/tuners/BifrostTuner.h" + +#include "arm_compute/core/CL/CLHelpers.h" +#include "arm_compute/core/CL/CLKernels.h" +#include "arm_compute/core/utils/misc/Cast.h" + +namespace arm_compute +{ +namespace tuners +{ +namespace +{ +/** Tunes a @ref CLDirectConvolutionLayerKernel for a bifrost target + * + * @param[in] k Kernels to tune + */ +void tune_direct_convolution_kernel(CLDirectConvolutionLayerKernel &k) +{ + cl::NDRange lws_hint = k.lws_hint(); + + const GPUTarget gpu_target = k.get_target(); + const DataType dt = k._input->info()->data_type(); + const TensorShape weights_shape = k._weights->info()->tensor_shape(); + const TensorShape inputs_shape = k._input->info()->tensor_shape(); + const size_t kernel_size = weights_shape.x(); + const unsigned int stride_x = k._conv_stride_x; + const unsigned int stride_y = k._conv_stride_y; + + if(gpu_target_is_in(gpu_target, GPUTarget::G71, GPUTarget::G72) && (kernel_size <= 5) && (stride_x == 1) && (stride_y == 1) && (dt == DataType::F32)) + { + // Through extensive experimentation with over 30 representative tensor + // shapes, we found a small number of local work size configurations + // that result in nearly optimal execution times. Selecting the right + // lws for a given shape, however, required a complex decision tree, + // until we constructed a simple feature as described below. + // + // We started from the number of multiply-accumulate operations for a + // convolution layer, which is equal to the product of the input + // dimensions 0..2 and the weights dimensions 0..2. Unfortunately, + // this resulted in ties between distinct shapes that required distinct + // lws configurations. Replacing the width of the input with the kernel + // size, however, resulted in nearly optimal predictions. We use underscores + // in variable names to indicate when they are intentionally misleading. + const size_t product_of_weights_dimensions = weights_shape[0] * weights_shape[1] * weights_shape[2]; + const size_t product_of_input_dimensions_ = inputs_shape[0] * inputs_shape[1] * inputs_shape[2]; + const float mega_ops_ = 1e-6 * product_of_weights_dimensions * product_of_input_dimensions_; + + switch(kernel_size) + { + case 1: + { + if(mega_ops_ < 1.f) + { + lws_hint = cl::NDRange(1, 1, 8); + } + else if(mega_ops_ < 7.f) + { + lws_hint = cl::NDRange(1, 1, 4); + } + else + { + lws_hint = cl::NDRange(1, 1, 2); + } + break; + } + case 3: + { + if(mega_ops_ < 1.f) + { + lws_hint = cl::NDRange(1, 1, 8); + } + else if(mega_ops_ < 13.f) + { + lws_hint = cl::NDRange(2, 1, 4); + } + else if(mega_ops_ < 50.f) + { + lws_hint = cl::NDRange(3, 1, 4); + } + else + { + lws_hint = cl::NDRange(2, 1, 6); + } + break; + } + case 5: + { + if(mega_ops_ < 2.f || mega_ops_ > 80.f) + { + lws_hint = cl::NDRange(2, 1, 4); + } + else + { + lws_hint = cl::NDRange(2, 1, 8); + } + break; + } + default: + break; + } + k.set_lws_hint(lws_hint); + } +} +} // namespace + +void BifrostTuner::tune_kernel_static(ICLKernel &kernel) +{ + // Continue on tuning if dynamic tuning + if(dynamic_cast<CLDirectConvolutionLayerKernel *>(&kernel) != nullptr) + { + tune_direct_convolution_kernel(*utils::cast::polymorphic_downcast<CLDirectConvolutionLayerKernel *>(&kernel)); + } +} + +void BifrostTuner::tune_kernel_dynamic(ICLKernel &kernel) +{ + ARM_COMPUTE_UNUSED(kernel); +} +} // namespace tuners +} // namespace arm_compute
\ No newline at end of file diff --git a/tests/validation/CL/UNIT/Tuner.cpp b/tests/validation/CL/UNIT/Tuner.cpp new file mode 100644 index 0000000000..26d21b54f2 --- /dev/null +++ b/tests/validation/CL/UNIT/Tuner.cpp @@ -0,0 +1,80 @@ +/* + * 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/core/CL/kernels/CLDirectConvolutionLayerKernel.h" +#include "arm_compute/runtime/CL/CLScheduler.h" +#include "arm_compute/runtime/CL/CLTensor.h" +#include "arm_compute/runtime/CL/tuners/BifrostTuner.h" +#include "support/ToolchainSupport.h" +#include "tests/Utils.h" +#include "tests/framework/Asserts.h" +#include "tests/framework/Macros.h" + +namespace arm_compute +{ +namespace test +{ +namespace validation +{ +TEST_SUITE(CL) +TEST_SUITE(UNIT) +TEST_SUITE(Tuner) + +/** Validates static tuning of Bifrost tuner */ +TEST_CASE(BifrostTunerSimple, framework::DatasetMode::ALL) +{ + // Create tuner + tuners::BifrostTuner tuner; + + // Create tensors + auto src = create_tensor<CLTensor>(TensorShape(13U, 13U, 16U), DataType::F32); + auto weights = create_tensor<CLTensor>(TensorShape(3U, 3U, 16U, 3U), DataType::F32); + auto bias = create_tensor<CLTensor>(TensorShape(3U), DataType::F32); + auto dst = create_tensor<CLTensor>(TensorShape(13U, 13U, 3U), DataType::F32); + + // Create kernel + cl::NDRange fake_lws(2000); + CLDirectConvolutionLayerKernel conv; + conv.set_target(GPUTarget::G72); + + // Hard-wire lws to kernel and validate lws + conv.set_lws_hint(fake_lws); + ARM_COMPUTE_EXPECT(conv.lws_hint()[0] == 2000, framework::LogLevel::ERRORS); + + // Configure + conv.configure(&src, &weights, &bias, &dst, PadStrideInfo(1, 1, 1, 1)); + ARM_COMPUTE_EXPECT(conv.lws_hint()[0] == 2000, framework::LogLevel::ERRORS); + + // Tune kernel and validate + tuner.tune_kernel_static(conv); + ARM_COMPUTE_EXPECT(conv.lws_hint()[0] != 2000, framework::LogLevel::ERRORS); + + // Clear tuner + CLScheduler::get().default_init(); +} +TEST_SUITE_END() +TEST_SUITE_END() +TEST_SUITE_END() +} // namespace validation +} // namespace test +} // namespace arm_compute |