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 #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 */
27
28 #include "arm_compute/core/Helpers.h"
29 #include "arm_compute/core/KernelDescriptors.h"
30 #include "arm_compute/core/Types.h"
31 #include "arm_compute/core/utils/misc/ShapeCalculator.h"
32 #include "arm_compute/runtime/CL/CLScheduler.h"
33 #include "arm_compute/runtime/CL/CLTuner.h"
34 #include "examples/gemm_tuner/CommonGemmExampleOptions.h"
35 #include "examples/gemm_tuner/GemmTunerHelpers.h"
36 #include "src/core/CL/kernels/CLGEMMMatrixMultiplyReshapedKernel.h"
37 #include "src/core/CL/kernels/CLGEMMReshapeLHSMatrixKernel.h"
38 #include "tests/CL/Helper.h"
39 #include "utils/Utils.h"
40 #include "utils/command_line/CommandLineOptions.h"
41 #include "utils/command_line/CommandLineParser.h"
42
43 #include <cstdlib>
44
45 using namespace arm_compute;
46 using namespace utils;
47 using namespace arm_compute::misc::shape_calculator;
48 using namespace gemm_tuner;
49
50 namespace
51 {
52 /** Structure holding all tunable gemm configs specific to this example/strategy */
53 struct GemmConfigs
54 {
55 size_t m0{ 4 }; /**< Number of rows processed by the matrix multiplication */
56 size_t n0{ 4 }; /**< Number of columns processed by the matrix multiplication */
57 size_t k0{ 4 }; /**< Number of partial accumulations performed by the matrix multiplication */
58 size_t v0{ 1 }; /**< Number of vertical blocks of size (m0xk0) stored on the same output row */
59 size_t h0{ 1 }; /**< Number of horizontal blocks of size (k0xn0) stored on the same output row */
60 bool interleave_lhs{ true }; /**< Interleave lhs matrix */
61 bool transpose_lhs{ true }; /**< Transpose lhs matrix. */
62 bool interleave_rhs{ true }; /**< Interleave rhs matrix */
63 bool transpose_rhs{ true }; /**< Transpose rhs matrix. */
64 bool export_to_cl_image_rhs{ true }; /**< Export rhs matrix to cl_image. */
65 };
66
67 /** Formatted output of the GemmConfigs type
68 *
69 * @param[out] os Output stream.
70 * @param[in] configs Tunable configurations to output
71 *
72 * @return Modified output stream.
73 */
operator <<(::std::ostream & os,const GemmConfigs & configs)74 ::std::ostream &operator<<(::std::ostream &os, const GemmConfigs &configs)
75 {
76 std::string false_str = std::string("false");
77 std::string true_str = std::string("true");
78
79 os << "m0 : " << configs.m0 << std::endl;
80 os << "n0 : " << configs.n0 << std::endl;
81 os << "k0 : " << configs.k0 << std::endl;
82 os << "v0 : " << configs.v0 << std::endl;
83 os << "h0 : " << configs.h0 << std::endl;
84 os << "interleave_lhs : " << (configs.interleave_lhs ? true_str : false_str) << std::endl;
85 os << "transpose_lhs : " << (configs.transpose_lhs ? true_str : false_str) << std::endl;
86 os << "interleave_rhs : " << (configs.interleave_rhs ? true_str : false_str) << std::endl;
87 os << "transpose_rhs : " << (configs.transpose_rhs ? true_str : false_str) << std::endl;
88 os << "export_to_cl_image_rhs : " << (configs.export_to_cl_image_rhs ? true_str : false_str) << std::endl;
89 return os;
90 }
91
92 /** Command line options for gemm configs */
93 class GemmConfigOptions
94 {
95 public:
96 /** Constructor
97 *
98 * @param[in,out] parser A parser on which "parse()" hasn't been called yet.
99 */
GemmConfigOptions(CommandLineParser & parser)100 GemmConfigOptions(CommandLineParser &parser)
101 : m0(parser.add_positional_option<SimpleOption<size_t>>("m0", 4)),
102 n0(parser.add_positional_option<SimpleOption<size_t>>("n0", 4)),
103 k0(parser.add_positional_option<SimpleOption<size_t>>("k0", 4)),
104 v0(parser.add_positional_option<SimpleOption<size_t>>("v0", 1)),
105 h0(parser.add_positional_option<SimpleOption<size_t>>("h0", 1)),
106 interleave_lhs(parser.add_positional_option<SimpleOption<size_t>>("interleave_lhs", 1)),
107 interleave_rhs(parser.add_positional_option<SimpleOption<size_t>>("interleave_rhs", 1)),
108 transpose_rhs(parser.add_positional_option<SimpleOption<size_t>>("transpose_rhs", 1)),
109 export_to_cl_image_rhs(parser.add_positional_option<SimpleOption<size_t>>("export_to_cl_image_rhs", 1))
110 {
111 m0->set_help("Number of rows processed by the matrix multiplication");
112 n0->set_help("Number of columns processed by the matrix multiplication");
113 k0->set_help("Number of partial accumulations performed by the matrix multiplication");
114 v0->set_help("Number of vertical blocks of size (m0xk0) stored on the same output row");
115 h0->set_help("Number of horizontal blocks of size (k0xn0) stored on the same output row");
116 interleave_lhs->set_help("Interleave lhs matrix (1) / Do not interleave lhs matrix (0)");
117 interleave_rhs->set_help("Interleave rhs matrix (1) / Do not interleave rhs matrix (0)");
118 // FIXME: Currently we only support 2 variants of the gemm reshaped kernels in which transpose_lhs and
119 // transpose_rhs are the opposites of each other. In the future we may extend the kernels to include the other
120 // 2 variants (both transposed and none transposed)
121 transpose_rhs->set_help("Transpose rhs matrix but not lhs matrix (1) / Do not transpose rhs matrix but do transpose lhs matrix (0)");
122 export_to_cl_image_rhs->set_help("Export rhs matrix to cl_image (1) / Do not export rhs matrix to cl_image (0)");
123 }
124 /** Prevent instances of this class from being copied (As this class contains pointers) */
125 GemmConfigOptions(const GemmConfigOptions &) = delete;
126 /** Prevent instances of this class from being copied (As this class contains pointers) */
127 GemmConfigOptions &operator=(const GemmConfigOptions &) = delete;
128 /** Allow instances of this class to be moved */
129 GemmConfigOptions(GemmConfigOptions &&) = default;
130 /** Allow instances of this class to be moved */
131 GemmConfigOptions &operator=(GemmConfigOptions &&) = default;
132 /** Default destructor */
133 ~GemmConfigOptions() = default;
134
135 SimpleOption<size_t> *m0; /**< Number of rows processed by the matrix multiplication option */
136 SimpleOption<size_t> *n0; /**< Number of columns processed by the matrix multiplication option */
137 SimpleOption<size_t> *k0; /**< Number of partial accumulations performed by the matrix multiplication option */
138 SimpleOption<size_t> *v0; /**< Number of vertical blocks of size (m0xk0) stored on the same output row option */
139 SimpleOption<size_t> *h0; /**< Number of horizontal blocks of size (k0xn0) stored on the same output row option */
140 SimpleOption<size_t> *interleave_lhs; /**< Interleave lhs matrix option (1 enable; 0 disable) */
141 SimpleOption<size_t> *interleave_rhs; /**< Interleave rhs matrix option (1 enable; 0 disable) */
142 // FIXME: Currently we only support 2 variants of the gemm reshaped kernels in which transpose_lhs and
143 // transpose_rhs are the opposites of each other. In the future we may extend the kernels to include the other
144 // 2 variants (both transposed and none transposed)
145 SimpleOption<size_t> *transpose_rhs; /**< Transpose rhs matrix option (1 enable; 0 disable). Also set the lhs matrix transpose option to the opposite. */
146 SimpleOption<size_t> *export_to_cl_image_rhs; /**< Export rhs matrix to cl_image.*/
147 };
148
149 /** Consumes the gemm configuration options and creates a structure containing all information
150 *
151 * @param[in] options Options to consume
152 *
153 * @return Structure containing the gemm configurations
154 */
consume_gemm_configs(const GemmConfigOptions & options)155 GemmConfigs consume_gemm_configs(const GemmConfigOptions &options)
156 {
157 GemmConfigs configs;
158 configs.m0 = options.m0->value();
159 configs.n0 = options.n0->value();
160 configs.k0 = options.k0->value();
161 configs.v0 = options.v0->value();
162 configs.h0 = options.h0->value();
163 configs.interleave_lhs = options.interleave_lhs->value() != 0;
164 // FIXME: Currently we only support 2 variants of the gemm reshaped kernels in which transpose_lhs and
165 // transpose_rhs are the opposites of each other. In the future we may extend the kernels to include the other
166 // 2 variants (both transposed and none transposed)
167 configs.transpose_lhs = options.transpose_rhs->value() == 0;
168 configs.interleave_rhs = options.interleave_rhs->value() != 0;
169 configs.transpose_rhs = options.transpose_rhs->value() != 0;
170 configs.export_to_cl_image_rhs = options.export_to_cl_image_rhs->value() != 0;
171 return configs;
172 }
173
174 } // namespace
175 // Create function for CLGEMMReshapeLHSMatrixKernel
176 using CLGEMMReshapeLHSMatrix = test::CLSynthetizeFunction<CLGEMMReshapeLHSMatrixKernel>;
177 // Create function for CLGEMMMatrixMultiplyReshapedKernel
178 using CLGEMMMatrixMultiplyReshaped = test::CLSynthetizeFunction<CLGEMMMatrixMultiplyReshapedKernel>;
179
180 class CLGEMMMatrixMultiplyReshapedExample : public Example
181 {
182 public:
do_setup(int argc,char ** argv)183 bool do_setup(int argc, char **argv) override
184 {
185 // Default parameters
186 const float alpha = 1.0f;
187 const float beta = 0.0f;
188 const ActivationLayerInfo act_info = ActivationLayerInfo();
189 CommonGemmExampleParams params;
190 GemmConfigs configs;
191
192 // Set up command line parser and options
193 CommandLineParser parser;
194 CommonGemmExampleOptions param_options(parser);
195 GemmConfigOptions config_options(parser);
196
197 // Parse command line options
198 parser.parse(argc, argv);
199 if(param_options.help->is_set() && param_options.help->value())
200 {
201 // Print help message
202 parser.print_help(argv[0]);
203 return false;
204 }
205 if(!parser.validate())
206 {
207 // Invalid arguments. Use default parameters and configs
208 std::cerr << "Invalid arguments." << std::endl;
209 parser.print_help(argv[0]);
210 std::cerr << "Falling back to default parameters and configs" << std::endl;
211 }
212 else
213 {
214 // Get parameters and configs from command-line options
215 params = consume_common_gemm_example_parameters(param_options);
216 configs = consume_gemm_configs(config_options);
217 }
218
219 // Print gemm parameters and configurations
220 std::cout << "Gemm parameters:" << std::endl;
221 std::cout << params << std::endl;
222 std::cout << "Gemm configurations:" << std::endl;
223 std::cout << configs << std::endl;
224
225 CLScheduler::get().default_init(&tuner);
226
227 lhs.allocator()->init(TensorInfo(TensorShape(params.K, params.M, params.B), 1, params.data_type));
228 rhs.allocator()->init(TensorInfo(TensorShape(params.N, params.K, params.B), 1, params.data_type));
229 bias.allocator()->init(TensorInfo(TensorShape(params.N, 1, params.B), 1, params.data_type));
230
231 GEMMLHSMatrixInfo lhs_info;
232 lhs_info.m0 = configs.m0;
233 lhs_info.k0 = configs.k0;
234 lhs_info.v0 = configs.v0;
235 lhs_info.interleave = configs.interleave_lhs;
236 lhs_info.transpose = configs.transpose_lhs;
237
238 GEMMRHSMatrixInfo rhs_info;
239 rhs_info.n0 = configs.n0;
240 rhs_info.k0 = configs.k0;
241 rhs_info.h0 = configs.h0;
242 rhs_info.interleave = configs.interleave_rhs;
243 rhs_info.transpose = configs.transpose_rhs;
244 rhs_info.export_to_cl_image = configs.export_to_cl_image_rhs;
245
246 GEMMKernelInfo kernel_info;
247 kernel_info.m = params.M;
248 kernel_info.n = params.N;
249 kernel_info.k = params.K;
250 kernel_info.depth_output_gemm3d = 0;
251 kernel_info.reinterpret_input_as_3d = false;
252 kernel_info.broadcast_bias = true;
253 kernel_info.activation_info = act_info;
254
255 // Initialise lhs_reshaped tensor info
256 lhs_reshaped.allocator()->init(TensorInfo(compute_lhs_reshaped_shape(*lhs.info(), lhs_info), 1, params.data_type));
257
258 // Initialise rhs_reshaped tensor info
259 rhs_reshaped.allocator()->init(TensorInfo(compute_rhs_reshaped_shape(*rhs.info(), rhs_info), 1, params.data_type));
260
261 if(rhs_info.export_to_cl_image)
262 {
263 examples::gemm_tuner_helpers::update_padding_for_cl_image(rhs_reshaped.info());
264 }
265
266 // Validate argments
267 Status status{};
268 status = reshape_lhs.validate((&lhs)->info(), (&lhs_reshaped)->info(), lhs_info, kernel_info.reinterpret_input_as_3d);
269 if(!status)
270 {
271 // Unsupported arguments
272 std::cerr << "Unsupported arguments." << std::endl;
273 std::cerr << "Check documentation for supported/unsupported combinations" << std::endl;
274 return false;
275 }
276
277 status = gemm.validate((&lhs_reshaped)->info(), (&rhs_reshaped)->info(), (&bias)->info(), (&dst)->info(), alpha, beta, lhs_info, rhs_info, kernel_info);
278 if(!status)
279 {
280 // Unsupported arguments
281 std::cerr << "Unsupported arguments." << std::endl;
282 std::cerr << "Check documentation for supported/unsupported combinations" << std::endl;
283 return false;
284 }
285
286 // Configure reshape lhs function
287 reshape_lhs.configure(&lhs, &lhs_reshaped, lhs_info);
288
289 // Configure function
290 gemm.configure(&lhs_reshaped, &rhs_reshaped, &bias, &dst, alpha, beta, lhs_info, rhs_info, kernel_info);
291
292 // Allocate tensors
293 lhs.allocator()->allocate();
294 rhs.allocator()->allocate();
295 lhs_reshaped.allocator()->allocate();
296 rhs_reshaped.allocator()->allocate();
297 bias.allocator()->allocate();
298 dst.allocator()->allocate();
299
300 return true;
301 }
do_run()302 void do_run() override
303 {
304 // Execute the function
305 reshape_lhs.run();
306 gemm.run();
307
308 // Make sure all the OpenCL jobs are done executing:
309 CLScheduler::get().sync();
310 }
311
do_teardown()312 void do_teardown() override
313 {
314 }
315
316 private:
317 CLTensor lhs{};
318 CLTensor rhs{};
319 CLTensor lhs_reshaped{};
320 CLTensor rhs_reshaped{};
321 CLTensor bias{};
322 CLTensor dst{};
323 CLTuner tuner{};
324 CLGEMMReshapeLHSMatrix reshape_lhs{};
325 CLGEMMMatrixMultiplyReshaped gemm{};
326 };
327
328 /** Main program for gemm reshaped test
329 *
330 * @param[in] argc Number of arguments
331 * @param[in] argv Arguments ( [optional] M, [optional] N, [optional] K, [optional] B, [optional] m0, [optional] n0, [optional] k0, [optional] v0, [optional] h0, [optional] interleave_lhs, [optional] interleave_rhs, [optional] transpose_rhs )
332 */
main(int argc,char ** argv)333 int main(int argc, char **argv)
334 {
335 return run_example<CLGEMMMatrixMultiplyReshapedExample>(argc, argv);
336 }
337