1 /*
2 * Copyright (c) 2019-2020 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/runtime/CL/functions/CLFFT1D.h"
25
26 #include "arm_compute/core/CL/ICLTensor.h"
27 #include "arm_compute/core/Validate.h"
28 #include "arm_compute/runtime/CL/CLScheduler.h"
29 #include "src/core/CL/kernels/CLFFTDigitReverseKernel.h"
30 #include "src/core/CL/kernels/CLFFTRadixStageKernel.h"
31 #include "src/core/CL/kernels/CLFFTScaleKernel.h"
32 #include "src/core/utils/helpers/fft.h"
33 #include "support/MemorySupport.h"
34
35 namespace arm_compute
36 {
CLFFT1D(std::shared_ptr<IMemoryManager> memory_manager)37 CLFFT1D::CLFFT1D(std::shared_ptr<IMemoryManager> memory_manager)
38 : _memory_group(std::move(memory_manager)),
39 _digit_reverse_kernel(support::cpp14::make_unique<CLFFTDigitReverseKernel>()),
40 _fft_kernels(),
41 _scale_kernel(support::cpp14::make_unique<CLFFTScaleKernel>()),
42 _digit_reversed_input(),
43 _digit_reverse_indices(),
44 _num_ffts(0),
45 _run_scale(false)
46 {
47 }
48
49 CLFFT1D::~CLFFT1D() = default;
50
configure(const ICLTensor * input,ICLTensor * output,const FFT1DInfo & config)51 void CLFFT1D::configure(const ICLTensor *input, ICLTensor *output, const FFT1DInfo &config)
52 {
53 configure(CLKernelLibrary::get().get_compile_context(), input, output, config);
54 }
55
configure(const CLCompileContext & compile_context,const ICLTensor * input,ICLTensor * output,const FFT1DInfo & config)56 void CLFFT1D::configure(const CLCompileContext &compile_context, const ICLTensor *input, ICLTensor *output, const FFT1DInfo &config)
57 {
58 ARM_COMPUTE_ERROR_ON_NULLPTR(input, output);
59 ARM_COMPUTE_ERROR_THROW_ON(CLFFT1D::validate(input->info(), output->info(), config));
60
61 // Decompose size to radix factors
62 const auto supported_radix = CLFFTRadixStageKernel::supported_radix();
63 const unsigned int N = input->info()->tensor_shape()[config.axis];
64 const auto decomposed_vector = arm_compute::helpers::fft::decompose_stages(N, supported_radix);
65 ARM_COMPUTE_ERROR_ON(decomposed_vector.empty());
66
67 // Flags
68 _run_scale = config.direction == FFTDirection::Inverse;
69 const bool is_c2r = input->info()->num_channels() == 2 && output->info()->num_channels() == 1;
70
71 // Configure digit reverse
72 FFTDigitReverseKernelInfo digit_reverse_config;
73 digit_reverse_config.axis = config.axis;
74 digit_reverse_config.conjugate = config.direction == FFTDirection::Inverse;
75 TensorInfo digit_reverse_indices_info(TensorShape(input->info()->tensor_shape()[config.axis]), 1, DataType::U32);
76 _digit_reverse_indices.allocator()->init(digit_reverse_indices_info);
77 _memory_group.manage(&_digit_reversed_input);
78 _digit_reverse_kernel->configure(compile_context, input, &_digit_reversed_input, &_digit_reverse_indices, digit_reverse_config);
79
80 // Create and configure FFT kernels
81 unsigned int Nx = 1;
82 _num_ffts = decomposed_vector.size();
83 _fft_kernels.reserve(_num_ffts);
84 for(unsigned int i = 0; i < _num_ffts; ++i)
85 {
86 const unsigned int radix_for_stage = decomposed_vector.at(i);
87
88 FFTRadixStageKernelInfo fft_kernel_info;
89 fft_kernel_info.axis = config.axis;
90 fft_kernel_info.radix = radix_for_stage;
91 fft_kernel_info.Nx = Nx;
92 fft_kernel_info.is_first_stage = (i == 0);
93 _fft_kernels.emplace_back(support::cpp14::make_unique<CLFFTRadixStageKernel>());
94 _fft_kernels.back()->configure(compile_context, &_digit_reversed_input, ((i == (_num_ffts - 1)) && !is_c2r) ? output : nullptr, fft_kernel_info);
95
96 Nx *= radix_for_stage;
97 }
98
99 // Configure scale kernel
100 if(_run_scale)
101 {
102 FFTScaleKernelInfo scale_config;
103 scale_config.scale = static_cast<float>(N);
104 scale_config.conjugate = config.direction == FFTDirection::Inverse;
105 is_c2r ? _scale_kernel->configure(compile_context, &_digit_reversed_input, output, scale_config) : _scale_kernel->configure(output, nullptr, scale_config);
106 }
107
108 // Allocate tensors
109 _digit_reversed_input.allocator()->allocate();
110 _digit_reverse_indices.allocator()->allocate();
111
112 // Init digit reverse indices
113 const auto digit_reverse_cpu = arm_compute::helpers::fft::digit_reverse_indices(N, decomposed_vector);
114 _digit_reverse_indices.map(CLScheduler::get().queue(), true);
115 std::copy_n(digit_reverse_cpu.data(), N, reinterpret_cast<unsigned int *>(_digit_reverse_indices.buffer()));
116 _digit_reverse_indices.unmap(CLScheduler::get().queue());
117 }
118
validate(const ITensorInfo * input,const ITensorInfo * output,const FFT1DInfo & config)119 Status CLFFT1D::validate(const ITensorInfo *input, const ITensorInfo *output, const FFT1DInfo &config)
120 {
121 ARM_COMPUTE_RETURN_ERROR_ON_NULLPTR(input, output);
122 ARM_COMPUTE_RETURN_ERROR_ON(input->data_type() != DataType::F32);
123 ARM_COMPUTE_RETURN_ERROR_ON(input->num_channels() != 1 && input->num_channels() != 2);
124 ARM_COMPUTE_RETURN_ERROR_ON(std::set<unsigned int>({ 0, 1 }).count(config.axis) == 0);
125
126 // Check if FFT is decomposable
127 const auto supported_radix = CLFFTRadixStageKernel::supported_radix();
128 const unsigned int N = input->tensor_shape()[config.axis];
129 const auto decomposed_vector = arm_compute::helpers::fft::decompose_stages(N, supported_radix);
130 ARM_COMPUTE_RETURN_ERROR_ON(decomposed_vector.empty());
131
132 // Checks performed when output is configured
133 if((output != nullptr) && (output->total_size() != 0))
134 {
135 ARM_COMPUTE_RETURN_ERROR_ON(output->num_channels() == 1 && input->num_channels() == 1);
136 ARM_COMPUTE_RETURN_ERROR_ON(output->num_channels() != 1 && output->num_channels() != 2);
137 ARM_COMPUTE_RETURN_ERROR_ON_MISMATCHING_SHAPES(input, output);
138 ARM_COMPUTE_RETURN_ERROR_ON_MISMATCHING_DATA_TYPES(input, output);
139 }
140
141 return Status{};
142 }
143
run()144 void CLFFT1D::run()
145 {
146 MemoryGroupResourceScope scope_mg(_memory_group);
147
148 // Run digit reverse
149 CLScheduler::get().enqueue(*_digit_reverse_kernel, false);
150
151 // Run radix kernels
152 for(unsigned int i = 0; i < _num_ffts; ++i)
153 {
154 CLScheduler::get().enqueue(*_fft_kernels[i], i == (_num_ffts - 1) && !_run_scale);
155 }
156
157 // Run output scaling
158 if(_run_scale)
159 {
160 CLScheduler::get().enqueue(*_scale_kernel, true);
161 }
162 }
163 } // namespace arm_compute
164