2 * Copyright (c) 2017-2018 ARM Limited.
4 * SPDX-License-Identifier: MIT
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:
13 * The above copyright notice and this permission notice shall be included in all
14 * copies or substantial portions of the Software.
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
24 #ifndef ARM_COMPUTE_CL /* Needed by Utils.cpp to handle OpenCL exceptions properly */
25 #error "This example needs to be built with -DARM_COMPUTE_CL"
26 #endif /* ARM_COMPUTE_CL */
28 #include "arm_compute/core/Types.h"
29 #include "arm_compute/runtime/CL/CLFunctions.h"
30 #include "arm_compute/runtime/CL/CLScheduler.h"
31 #include "arm_compute/runtime/CL/CLTuner.h"
32 #include "utils/Utils.h"
36 using namespace arm_compute;
37 using namespace utils;
39 class CLSGEMMExample : public Example
42 void do_setup(int argc, char **argv) override
44 NPYLoader npy0, npy1, npy2;
48 CLScheduler::get().default_init(&tuner);
53 stream.open(argv[1], std::fstream::in);
56 if(argc < 3 || (argc < 4 && stream.bad()))
59 std::cout << "Usage: 1) ./build/cl_sgemm input_matrix_1.npy input_matrix_2.npy [input_matrix_3.npy] [alpha = 1] [beta = 0]\n";
60 std::cout << " 2) ./build/cl_sgemm M N K [alpha = 1.0f] [beta = 0.0f]\n\n";
61 std::cout << "Too few or no input_matrices provided. Using M=7, N=3, K=5, alpha=1.0f and beta=0.0f\n\n";
63 src0.allocator()->init(TensorInfo(TensorShape(5U, 7U), 1, DataType::F32));
64 src1.allocator()->init(TensorInfo(TensorShape(3U, 5U), 1, DataType::F32));
65 src2.allocator()->init(TensorInfo(TensorShape(3U, 7U), 1, DataType::F32));
69 if(stream.good()) /* case file1.npy file2.npy [file3.npy] [alpha = 1.0f] [beta = 0.0f] */
72 npy0.init_tensor(src0, DataType::F32);
74 npy1.init_tensor(src1, DataType::F32);
80 stream.open(argv[3], std::fstream::in);
81 if(stream.good()) /* case with third file */
84 npy2.init_tensor(src2, DataType::F32);
88 // Convert string to float
89 alpha = strtof(argv[4], nullptr);
93 // Convert string to float
94 beta = strtof(argv[5], nullptr);
98 else /* case without third file */
100 alpha = strtof(argv[3], nullptr);
104 beta = strtof(argv[4], nullptr);
109 else /* case M N K [alpha = 1.0f] [beta = 0.0f] */
111 size_t M = strtol(argv[1], nullptr, 10);
112 size_t N = strtol(argv[2], nullptr, 10);
113 size_t K = strtol(argv[3], nullptr, 10);
115 src0.allocator()->init(TensorInfo(TensorShape(K, M), 1, DataType::F32));
116 src1.allocator()->init(TensorInfo(TensorShape(N, K), 1, DataType::F32));
117 src2.allocator()->init(TensorInfo(TensorShape(N, M), 1, DataType::F32));
121 alpha = strtof(argv[4], nullptr);
125 beta = strtof(argv[5], nullptr);
131 init_sgemm_output(dst, src0, src1, DataType::F32);
133 // Configure function
134 sgemm.configure(&src0, &src1, (src2.info()->total_size() > 0) ? &src2 : nullptr, &dst, alpha, beta);
136 // Allocate all the images
137 src0.allocator()->allocate();
138 src1.allocator()->allocate();
139 dst.allocator()->allocate();
141 // Fill the input images with either the data provided or random data
144 npy0.fill_tensor(src0);
145 npy1.fill_tensor(src1);
147 output_filename = "sgemm_out.npy";
148 is_fortran = npy0.is_fortran();
152 src2.allocator()->allocate();
153 npy2.fill_tensor(src2);
158 src2.allocator()->allocate();
160 fill_random_tensor(src0, -1.f, 1.f);
161 fill_random_tensor(src1, -1.f, 1.f);
162 fill_random_tensor(src2, -1.f, 1.f);
165 // Dummy run for CLTuner
168 void do_run() override
170 // Execute the function
173 // Make sure all the OpenCL jobs are done executing:
174 CLScheduler::get().sync();
176 void do_teardown() override
178 if(!output_filename.empty()) /* Save to .npy file */
180 save_to_npy(dst, output_filename, is_fortran);
185 CLTensor src0{}, src1{}, src2{}, dst{};
188 float alpha{}, beta{};
190 std::string output_filename{};
193 /** Main program for sgemm test
195 * @param[in] argc Number of arguments
196 * @param[in] argv Arguments ( [optional] Matrix A, [optional] Matrix B, [optional] Matrix C, [optional] alpha, [optional] beta )
198 int main(int argc, char **argv)
200 return utils::run_example<CLSGEMMExample>(argc, argv);