Sheri Zhang | 1efed92 | 2021-03-10 22:43:38 +0000 | [diff] [blame] | 1 | /* |
| 2 | * Copyright (c) 2021 Arm Limited. |
| 3 | * |
| 4 | * SPDX-License-Identifier: MIT |
| 5 | * |
| 6 | * Permission is hereby granted, free of charge, to any person obtaining a copy |
| 7 | * of this software and associated documentation files (the "Software"), to |
| 8 | * deal in the Software without restriction, including without limitation the |
| 9 | * rights to use, copy, modify, merge, publish, distribute, sublicense, and/or |
| 10 | * sell copies of the Software, and to permit persons to whom the Software is |
| 11 | * furnished to do so, subject to the following conditions: |
| 12 | * |
| 13 | * The above copyright notice and this permission notice shall be included in all |
| 14 | * copies or substantial portions of the Software. |
| 15 | * |
| 16 | * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR |
| 17 | * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY, |
| 18 | * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE |
| 19 | * AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER |
| 20 | * LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM, |
| 21 | * OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE |
| 22 | * SOFTWARE. |
| 23 | */ |
Georgios Pinitas | 7891a73 | 2021-08-20 21:39:25 +0100 | [diff] [blame] | 24 | #include "src/gpu/cl/operators/ClDirectConv2d.h" |
Sheri Zhang | 1efed92 | 2021-03-10 22:43:38 +0000 | [diff] [blame] | 25 | |
| 26 | #include "arm_compute/runtime/CL/CLScheduler.h" |
| 27 | #include "src/core/CL/kernels/CLFillBorderKernel.h" |
Georgios Pinitas | 7891a73 | 2021-08-20 21:39:25 +0100 | [diff] [blame] | 28 | #include "src/gpu/cl/ClCompileContext.h" |
| 29 | #include "src/gpu/cl/kernels/ClActivationKernel.h" |
| 30 | #include "src/gpu/cl/kernels/ClDirectConv2dKernel.h" |
Sheri Zhang | 1efed92 | 2021-03-10 22:43:38 +0000 | [diff] [blame] | 31 | |
| 32 | namespace arm_compute |
| 33 | { |
| 34 | namespace opencl |
| 35 | { |
| 36 | namespace |
| 37 | { |
| 38 | ITensorPack select_activation_src_dst(ITensorPack &tensors) |
| 39 | { |
| 40 | ITensorPack pack; |
| 41 | pack.add_tensor(TensorType::ACL_SRC, tensors.get_tensor(TensorType::ACL_DST)); |
| 42 | pack.add_tensor(TensorType::ACL_DST, tensors.get_tensor(TensorType::ACL_DST)); |
| 43 | return pack; |
| 44 | } |
| 45 | } // namespace |
| 46 | |
Manuel Bottini | b4bb6a0 | 2021-05-24 16:01:32 +0100 | [diff] [blame] | 47 | void ClDirectConv2d::configure(const CLCompileContext &compile_context, ITensorInfo *src, ITensorInfo *weights, ITensorInfo *biases, ITensorInfo *dst, |
| 48 | const PadStrideInfo &conv_info, const ActivationLayerInfo &act_info) |
Sheri Zhang | 1efed92 | 2021-03-10 22:43:38 +0000 | [diff] [blame] | 49 | { |
Georgios Pinitas | 9fc3be6 | 2021-05-29 04:01:51 +0100 | [diff] [blame] | 50 | ARM_COMPUTE_ERROR_ON_NULLPTR(src); |
| 51 | |
Sheri Zhang | 1efed92 | 2021-03-10 22:43:38 +0000 | [diff] [blame] | 52 | // Configure direct convolution kernel |
Georgios Pinitas | 9fc3be6 | 2021-05-29 04:01:51 +0100 | [diff] [blame] | 53 | const ActivationLayerInfo conv2d_act_info = (src->data_layout() == DataLayout::NHWC && is_data_type_float(src->data_type())) ? act_info : ActivationLayerInfo(); |
| 54 | auto k = std::make_unique<kernels::ClDirectConv2dKernel>(); |
Sheri Zhang | 1efed92 | 2021-03-10 22:43:38 +0000 | [diff] [blame] | 55 | k->set_target(CLScheduler::get().target()); |
Georgios Pinitas | 9fc3be6 | 2021-05-29 04:01:51 +0100 | [diff] [blame] | 56 | k->configure(compile_context, src, weights, biases, dst, conv_info, conv2d_act_info); |
Sheri Zhang | 1efed92 | 2021-03-10 22:43:38 +0000 | [diff] [blame] | 57 | _direct_conv_kernel = std::move(k); |
| 58 | |
| 59 | // Configure border handler |
| 60 | PixelValue zero_value(0.f); |
| 61 | if(is_data_type_quantized_asymmetric(src->data_type())) |
| 62 | { |
| 63 | zero_value = PixelValue(0, src->data_type(), src->quantization_info()); |
| 64 | } |
| 65 | auto b = std::make_unique<CLFillBorderKernel>(); |
| 66 | b->configure(compile_context, src, _direct_conv_kernel->border_size(), BorderMode::CONSTANT, zero_value); |
| 67 | _src_border_handler = std::move(b); |
| 68 | |
Georgios Pinitas | 9fc3be6 | 2021-05-29 04:01:51 +0100 | [diff] [blame] | 69 | // Fused activation is currently supported for NHWC and floating point types |
| 70 | if(act_info.enabled() && !conv2d_act_info.enabled()) |
Sheri Zhang | 1efed92 | 2021-03-10 22:43:38 +0000 | [diff] [blame] | 71 | { |
| 72 | auto a = std::make_unique<kernels::ClActivationKernel>(); |
| 73 | a->configure(compile_context, dst, dst, act_info); |
| 74 | _activation_kernel = std::move(a); |
| 75 | } |
| 76 | |
| 77 | // Tune kernels |
| 78 | CLScheduler::get().tune_kernel_static(*_direct_conv_kernel); |
| 79 | } |
| 80 | |
Manuel Bottini | b4bb6a0 | 2021-05-24 16:01:32 +0100 | [diff] [blame] | 81 | Status ClDirectConv2d::validate(const ITensorInfo *src, const ITensorInfo *weights, const ITensorInfo *biases, const ITensorInfo *dst, |
| 82 | const PadStrideInfo &conv_info, const ActivationLayerInfo &act_info) |
Sheri Zhang | 1efed92 | 2021-03-10 22:43:38 +0000 | [diff] [blame] | 83 | { |
Georgios Pinitas | 9fc3be6 | 2021-05-29 04:01:51 +0100 | [diff] [blame] | 84 | ARM_COMPUTE_RETURN_ON_ERROR(kernels::ClDirectConv2dKernel::validate(src, weights, biases, dst, conv_info, ActivationLayerInfo(), CLScheduler::get().target())); |
Sheri Zhang | 1efed92 | 2021-03-10 22:43:38 +0000 | [diff] [blame] | 85 | if(act_info.enabled()) |
| 86 | { |
| 87 | ARM_COMPUTE_RETURN_ON_ERROR(kernels::ClActivationKernel::validate(dst, dst, act_info)); |
| 88 | } |
| 89 | return Status{}; |
| 90 | } |
| 91 | |
Manuel Bottini | b4bb6a0 | 2021-05-24 16:01:32 +0100 | [diff] [blame] | 92 | void ClDirectConv2d::run(ITensorPack &tensors) |
Sheri Zhang | 1efed92 | 2021-03-10 22:43:38 +0000 | [diff] [blame] | 93 | { |
| 94 | // Run border handler |
| 95 | CLScheduler::get().enqueue_op(*_src_border_handler.get(), tensors, false); |
| 96 | // Run direct convolution |
| 97 | CLScheduler::get().enqueue_op(*_direct_conv_kernel.get(), tensors, false); |
| 98 | // Run activation kernel |
| 99 | if(_activation_kernel) |
| 100 | { |
| 101 | auto act_pack = select_activation_src_dst(tensors); |
| 102 | CLScheduler::get().enqueue_op(*_activation_kernel.get(), act_pack, false); |
| 103 | } |
| 104 | } |
| 105 | } // namespace opencl |
| 106 | } // namespace arm_compute |