Gunes Bayir | ae72a46 | 2023-01-29 13:24:24 +0000 | [diff] [blame] | 1 | /* |
| 2 | * Copyright (c) 2023 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 | */ |
| 24 | #include "arm_compute/core/experimental/Types.h" |
| 25 | #include "arm_compute/runtime/NEON/NEScheduler.h" |
| 26 | |
| 27 | #include "src/common/utils/Log.h" |
| 28 | #include "src/core/helpers/MemoryHelpers.h" |
| 29 | #include "src/cpu/kernels/CpuAddMulAddKernel.h" |
| 30 | #include "src/cpu/operators/CpuAddMulAdd.h" |
| 31 | #include "src/cpu/utils/CpuAuxTensorHandler.h" |
| 32 | |
| 33 | namespace arm_compute |
| 34 | { |
| 35 | namespace cpu |
| 36 | { |
| 37 | void CpuAddMulAdd::configure(const ITensorInfo *input1, const ITensorInfo *input2, |
| 38 | const ITensorInfo *bn_mul, const ITensorInfo *bn_add, |
| 39 | ITensorInfo *add_output, ITensorInfo *final_output, |
| 40 | ConvertPolicy policy, const ActivationLayerInfo &act_info) |
| 41 | { |
| 42 | ARM_COMPUTE_LOG_PARAMS(input1, input2, bn_mul, bn_add, add_output, final_output, policy, act_info); |
| 43 | |
| 44 | auto k = std::make_unique<kernels::CpuAddMulAddKernel>(); |
| 45 | |
| 46 | const DataType data_type = input1->data_type(); |
| 47 | if(is_data_type_quantized(data_type)) |
| 48 | { |
| 49 | _dequantize_bn_mul.configure(bn_mul, &_dequantized_bn_mul); |
| 50 | _dequantize_bn_add.configure(bn_add, &_dequantized_bn_add); |
| 51 | |
| 52 | k->configure(input1, input2, &_dequantized_bn_mul, &_dequantized_bn_add, add_output, final_output, policy, act_info); |
| 53 | |
| 54 | // Save auxilary memory requirements after configuration |
| 55 | _aux_mem[DequantizedBnMul] = experimental::MemoryInfo(offset_int_vec(DequantizedBnMul), experimental::MemoryLifetime::Temporary, _dequantized_bn_mul.total_size()); |
| 56 | _aux_mem[DequantizedBnAdd] = experimental::MemoryInfo(offset_int_vec(DequantizedBnAdd), experimental::MemoryLifetime::Temporary, _dequantized_bn_add.total_size()); |
| 57 | } |
| 58 | else |
| 59 | { |
| 60 | k->configure(input1, input2, bn_mul, bn_add, add_output, final_output, policy, act_info); |
| 61 | } |
| 62 | |
| 63 | _kernel = std::move(k); |
| 64 | } |
| 65 | |
| 66 | Status CpuAddMulAdd::validate(const ITensorInfo *input1, const ITensorInfo *input2, |
| 67 | const ITensorInfo *bn_mul, const ITensorInfo *bn_add, |
| 68 | const ITensorInfo *add_output, const ITensorInfo *final_output, |
| 69 | ConvertPolicy policy, const ActivationLayerInfo &act_info) |
| 70 | { |
| 71 | const DataType data_type = input1->data_type(); |
| 72 | if(is_data_type_quantized(data_type)) |
| 73 | { |
| 74 | TensorInfo dequantized_bn_mul; |
| 75 | TensorInfo dequantized_bn_add; |
| 76 | |
| 77 | ARM_COMPUTE_RETURN_ON_ERROR(CpuDequantize::validate(bn_mul, &dequantized_bn_mul)); |
| 78 | ARM_COMPUTE_RETURN_ON_ERROR(CpuDequantize::validate(bn_add, &dequantized_bn_add)); |
| 79 | |
| 80 | return kernels::CpuAddMulAddKernel::validate(input1, input2, &dequantized_bn_mul, &dequantized_bn_add, add_output, final_output, policy, act_info); |
| 81 | } |
| 82 | else |
| 83 | { |
| 84 | return kernels::CpuAddMulAddKernel::validate(input1, input2, bn_mul, bn_add, add_output, final_output, policy, act_info); |
| 85 | } |
| 86 | } |
| 87 | |
| 88 | void CpuAddMulAdd::run(ITensorPack &tensors) |
| 89 | { |
| 90 | const DataType data_type = tensors.get_const_tensor(TensorType::ACL_SRC_0)->info()->data_type(); |
| 91 | |
| 92 | if(is_data_type_quantized(data_type)) |
| 93 | { |
| 94 | const ITensor *bn_mul = tensors.get_const_tensor(TensorType::ACL_SRC_2); |
| 95 | const ITensor *bn_add = tensors.get_const_tensor(TensorType::ACL_SRC_3); |
| 96 | |
| 97 | CpuAuxTensorHandler dequantized_bn_mul_handler(offset_int_vec(DequantizedBnMul), _dequantized_bn_mul, tensors, true); |
| 98 | CpuAuxTensorHandler dequantized_bn_add_handler(offset_int_vec(DequantizedBnAdd), _dequantized_bn_add, tensors, true); |
| 99 | |
| 100 | ITensorPack dequantize_mul_pack = |
| 101 | { |
| 102 | { TensorType::ACL_SRC_0, bn_mul }, |
| 103 | { TensorType::ACL_DST_0, dequantized_bn_mul_handler.get() } |
| 104 | }; |
| 105 | |
| 106 | ITensorPack dequantize_add_pack = |
| 107 | { |
| 108 | { TensorType::ACL_SRC_0, bn_add }, |
| 109 | { TensorType::ACL_DST_0, dequantized_bn_add_handler.get() } |
| 110 | }; |
| 111 | |
| 112 | _dequantize_bn_mul.run(dequantize_mul_pack); |
| 113 | _dequantize_bn_add.run(dequantize_add_pack); |
| 114 | |
| 115 | ITensorPack add_mul_add_pack = |
| 116 | { |
| 117 | { TensorType::ACL_SRC_0, tensors.get_const_tensor(TensorType::ACL_SRC_0) }, |
| 118 | { TensorType::ACL_SRC_1, tensors.get_const_tensor(TensorType::ACL_SRC_1) }, |
| 119 | { TensorType::ACL_SRC_2, dequantized_bn_mul_handler.get() }, |
| 120 | { TensorType::ACL_SRC_3, dequantized_bn_add_handler.get() }, |
| 121 | { TensorType::ACL_DST_0, tensors.get_tensor(TensorType::ACL_DST_0) }, |
| 122 | { TensorType::ACL_DST_1, tensors.get_tensor(TensorType::ACL_DST_1) }, |
| 123 | }; |
| 124 | |
| 125 | NEScheduler::get().schedule_op(_kernel.get(), Window::DimY, _kernel->window(), add_mul_add_pack); |
| 126 | } |
| 127 | else |
| 128 | { |
| 129 | NEScheduler::get().schedule_op(_kernel.get(), Window::DimY, _kernel->window(), tensors); |
| 130 | } |
| 131 | } |
| 132 | |
| 133 | experimental::MemoryRequirements CpuAddMulAdd::workspace() const |
| 134 | { |
| 135 | return _aux_mem; |
| 136 | } |
| 137 | |
| 138 | } // namespace cpu |
| 139 | } // namespace arm_compute |