1 /*
2 * Copyright (c) 2019-2021 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 */
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
34 #include "src/common/utils/Log.h"
35
36 namespace arm_compute
37 {
CLFFT1D(std::shared_ptr<IMemoryManager> memory_manager)38 CLFFT1D::CLFFT1D(std::shared_ptr<IMemoryManager> memory_manager)
39 : _memory_group(std::move(memory_manager)),
40 _digit_reverse_kernel(std::make_unique<CLFFTDigitReverseKernel>()),
41 _fft_kernels(),
42 _scale_kernel(std::make_unique<CLFFTScaleKernel>()),
43 _digit_reversed_input(),
44 _digit_reverse_indices(),
45 _num_ffts(0),
46 _run_scale(false)
47 {
48 }
49
50 CLFFT1D::~CLFFT1D() = default;
51
configure(const ICLTensor * input,ICLTensor * output,const FFT1DInfo & config)52 void CLFFT1D::configure(const ICLTensor *input, ICLTensor *output, const FFT1DInfo &config)
53 {
54 configure(CLKernelLibrary::get().get_compile_context(), input, output, config);
55 }
56
configure(const CLCompileContext & compile_context,const ICLTensor * input,ICLTensor * output,const FFT1DInfo & config)57 void CLFFT1D::configure(const CLCompileContext &compile_context, const ICLTensor *input, ICLTensor *output, const FFT1DInfo &config)
58 {
59 ARM_COMPUTE_ERROR_ON_NULLPTR(input, output);
60 ARM_COMPUTE_ERROR_THROW_ON(CLFFT1D::validate(input->info(), output->info(), config));
61 ARM_COMPUTE_LOG_PARAMS(input, output, config);
62
63 // Decompose size to radix factors
64 const auto supported_radix = CLFFTRadixStageKernel::supported_radix();
65 const unsigned int N = input->info()->tensor_shape()[config.axis];
66 const auto decomposed_vector = arm_compute::helpers::fft::decompose_stages(N, supported_radix);
67 ARM_COMPUTE_ERROR_ON(decomposed_vector.empty());
68
69 // Flags
70 _run_scale = config.direction == FFTDirection::Inverse;
71 const bool is_c2r = input->info()->num_channels() == 2 && output->info()->num_channels() == 1;
72
73 // Configure digit reverse
74 FFTDigitReverseKernelInfo digit_reverse_config;
75 digit_reverse_config.axis = config.axis;
76 digit_reverse_config.conjugate = config.direction == FFTDirection::Inverse;
77 TensorInfo digit_reverse_indices_info(TensorShape(input->info()->tensor_shape()[config.axis]), 1, DataType::U32);
78 _digit_reverse_indices.allocator()->init(digit_reverse_indices_info);
79 _memory_group.manage(&_digit_reversed_input);
80 _digit_reverse_kernel->configure(compile_context, input, &_digit_reversed_input, &_digit_reverse_indices, digit_reverse_config);
81
82 // Create and configure FFT kernels
83 unsigned int Nx = 1;
84 _num_ffts = decomposed_vector.size();
85 _fft_kernels.reserve(_num_ffts);
86 for(unsigned int i = 0; i < _num_ffts; ++i)
87 {
88 const unsigned int radix_for_stage = decomposed_vector.at(i);
89
90 FFTRadixStageKernelInfo fft_kernel_info;
91 fft_kernel_info.axis = config.axis;
92 fft_kernel_info.radix = radix_for_stage;
93 fft_kernel_info.Nx = Nx;
94 fft_kernel_info.is_first_stage = (i == 0);
95 _fft_kernels.emplace_back(std::make_unique<CLFFTRadixStageKernel>());
96 _fft_kernels.back()->configure(compile_context, &_digit_reversed_input, ((i == (_num_ffts - 1)) && !is_c2r) ? output : nullptr, fft_kernel_info);
97
98 Nx *= radix_for_stage;
99 }
100
101 // Configure scale kernel
102 if(_run_scale)
103 {
104 FFTScaleKernelInfo scale_config;
105 scale_config.scale = static_cast<float>(N);
106 scale_config.conjugate = config.direction == FFTDirection::Inverse;
107 is_c2r ? _scale_kernel->configure(compile_context, &_digit_reversed_input, output, scale_config) : _scale_kernel->configure(output, nullptr, scale_config);
108 }
109
110 // Allocate tensors
111 _digit_reversed_input.allocator()->allocate();
112 _digit_reverse_indices.allocator()->allocate();
113
114 // Init digit reverse indices
115 const auto digit_reverse_cpu = arm_compute::helpers::fft::digit_reverse_indices(N, decomposed_vector);
116 _digit_reverse_indices.map(CLScheduler::get().queue(), true);
117 std::copy_n(digit_reverse_cpu.data(), N, reinterpret_cast<unsigned int *>(_digit_reverse_indices.buffer()));
118 _digit_reverse_indices.unmap(CLScheduler::get().queue());
119 }
120
validate(const ITensorInfo * input,const ITensorInfo * output,const FFT1DInfo & config)121 Status CLFFT1D::validate(const ITensorInfo *input, const ITensorInfo *output, const FFT1DInfo &config)
122 {
123 ARM_COMPUTE_RETURN_ERROR_ON_NULLPTR(input, output);
124 ARM_COMPUTE_RETURN_ERROR_ON_DATA_TYPE_NOT_IN(input, DataType::F16, DataType::F32);
125 ARM_COMPUTE_RETURN_ERROR_ON(input->num_channels() != 1 && input->num_channels() != 2);
126 ARM_COMPUTE_RETURN_ERROR_ON(std::set<unsigned int>({ 0, 1 }).count(config.axis) == 0);
127
128 // Check if FFT is decomposable
129 const auto supported_radix = CLFFTRadixStageKernel::supported_radix();
130 const unsigned int N = input->tensor_shape()[config.axis];
131 const auto decomposed_vector = arm_compute::helpers::fft::decompose_stages(N, supported_radix);
132 ARM_COMPUTE_RETURN_ERROR_ON(decomposed_vector.empty());
133
134 // Checks performed when output is configured
135 if((output != nullptr) && (output->total_size() != 0))
136 {
137 ARM_COMPUTE_RETURN_ERROR_ON(output->num_channels() == 1 && input->num_channels() == 1);
138 ARM_COMPUTE_RETURN_ERROR_ON(output->num_channels() != 1 && output->num_channels() != 2);
139 ARM_COMPUTE_RETURN_ERROR_ON_MISMATCHING_SHAPES(input, output);
140 ARM_COMPUTE_RETURN_ERROR_ON_MISMATCHING_DATA_TYPES(input, output);
141 }
142
143 return Status{};
144 }
145
run()146 void CLFFT1D::run()
147 {
148 MemoryGroupResourceScope scope_mg(_memory_group);
149
150 // Run digit reverse
151 CLScheduler::get().enqueue(*_digit_reverse_kernel, false);
152
153 // Run radix kernels
154 for(unsigned int i = 0; i < _num_ffts; ++i)
155 {
156 CLScheduler::get().enqueue(*_fft_kernels[i], i == (_num_ffts - 1) && !_run_scale);
157 }
158
159 // Run output scaling
160 if(_run_scale)
161 {
162 CLScheduler::get().enqueue(*_scale_kernel, true);
163 }
164 }
165 } // namespace arm_compute
166