From 2b5f0f2574551f59970bb9d710bafad2bc4bbd4a Mon Sep 17 00:00:00 2001 From: Michalis Spyrou Date: Wed, 10 Jan 2018 14:08:50 +0000 Subject: COMPMID-782 Port examples to the new format Change-Id: Ib178a97c080ff650094d02ee49e2a0aa22376dd0 Reviewed-on: https://eu-gerrit-1.euhpc.arm.com/115717 Reviewed-by: Anthony Barbier Tested-by: Jenkins --- examples/cl_sgemm.cpp | 272 ++++++++++++++++++++++++++------------------------ 1 file changed, 141 insertions(+), 131 deletions(-) (limited to 'examples/cl_sgemm.cpp') diff --git a/examples/cl_sgemm.cpp b/examples/cl_sgemm.cpp index 939870fabf..f2c63985f6 100644 --- a/examples/cl_sgemm.cpp +++ b/examples/cl_sgemm.cpp @@ -36,174 +36,184 @@ using namespace arm_compute; using namespace utils; -void main_cl_sgemm(int argc, char **argv) +class CLSGEMMExample : public Example { - NPYLoader npy0, npy1, npy2; - CLTensor src0, src1, src2, dst; - float alpha = 1.0f, beta = 0.0f; - - CLTuner tuner; - CLScheduler::get().default_init(&tuner); - - std::ifstream stream; - if(argc > 1) +public: + void do_setup(int argc, char **argv) override { - stream.open(argv[1], std::fstream::in); - } + NPYLoader npy0, npy1, npy2; + alpha = 1.0f; + beta = 0.0f; - if(argc < 3 || (argc < 4 && stream.bad())) - { - // Print help - std::cout << "Usage: 1) ./build/cl_sgemm input_matrix_1.npy input_matrix_2.npy [input_matrix_3.npy] [alpha = 1] [beta = 0]\n"; - std::cout << " 2) ./build/cl_sgemm M N K [alpha = 1.0f] [beta = 0.0f]\n\n"; - 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"; - - src0.allocator()->init(TensorInfo(TensorShape(5U, 7U), 1, DataType::F32)); - src1.allocator()->init(TensorInfo(TensorShape(3U, 5U), 1, DataType::F32)); - src2.allocator()->init(TensorInfo(TensorShape(3U, 7U), 1, DataType::F32)); - } - else - { - if(stream.good()) /* case file1.npy file2.npy [file3.npy] [alpha = 1.0f] [beta = 0.0f] */ + CLScheduler::get().default_init(&tuner); + + std::ifstream stream; + if(argc > 1) { - npy0.open(argv[1]); - npy0.init_tensor(src0, DataType::F32); - npy1.open(argv[2]); - npy1.init_tensor(src1, DataType::F32); + stream.open(argv[1], std::fstream::in); + } - if(argc > 3) + if(argc < 3 || (argc < 4 && stream.bad())) + { + // Print help + std::cout << "Usage: 1) ./build/cl_sgemm input_matrix_1.npy input_matrix_2.npy [input_matrix_3.npy] [alpha = 1] [beta = 0]\n"; + std::cout << " 2) ./build/cl_sgemm M N K [alpha = 1.0f] [beta = 0.0f]\n\n"; + 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"; + + src0.allocator()->init(TensorInfo(TensorShape(5U, 7U), 1, DataType::F32)); + src1.allocator()->init(TensorInfo(TensorShape(3U, 5U), 1, DataType::F32)); + src2.allocator()->init(TensorInfo(TensorShape(3U, 7U), 1, DataType::F32)); + } + else + { + if(stream.good()) /* case file1.npy file2.npy [file3.npy] [alpha = 1.0f] [beta = 0.0f] */ { - stream.close(); - stream.clear(); - stream.open(argv[3], std::fstream::in); - if(stream.good()) /* case with third file */ - { - npy2.open(argv[3]); - npy2.init_tensor(src2, DataType::F32); + npy0.open(argv[1]); + npy0.init_tensor(src0, DataType::F32); + npy1.open(argv[2]); + npy1.init_tensor(src1, DataType::F32); - if(argc > 4) + if(argc > 3) + { + stream.close(); + stream.clear(); + stream.open(argv[3], std::fstream::in); + if(stream.good()) /* case with third file */ { - // Convert string to float - alpha = strtof(argv[4], nullptr); + npy2.open(argv[3]); + npy2.init_tensor(src2, DataType::F32); - if(argc > 5) + if(argc > 4) { // Convert string to float - beta = strtof(argv[5], nullptr); + alpha = strtof(argv[4], nullptr); + + if(argc > 5) + { + // Convert string to float + beta = strtof(argv[5], nullptr); + } } } - } - else /* case without third file */ - { - alpha = strtof(argv[3], nullptr); - - if(argc > 4) + else /* case without third file */ { - beta = strtof(argv[4], nullptr); + alpha = strtof(argv[3], nullptr); + + if(argc > 4) + { + beta = strtof(argv[4], nullptr); + } } } } - } - else /* case M N K [alpha = 1.0f] [beta = 0.0f] */ - { - size_t M = strtol(argv[1], nullptr, 10); - size_t N = strtol(argv[2], nullptr, 10); - size_t K = strtol(argv[3], nullptr, 10); - - src0.allocator()->init(TensorInfo(TensorShape(K, M), 1, DataType::F32)); - src1.allocator()->init(TensorInfo(TensorShape(N, K), 1, DataType::F32)); - src2.allocator()->init(TensorInfo(TensorShape(N, M), 1, DataType::F32)); - - if(argc > 4) + else /* case M N K [alpha = 1.0f] [beta = 0.0f] */ { - alpha = strtof(argv[4], nullptr); + size_t M = strtol(argv[1], nullptr, 10); + size_t N = strtol(argv[2], nullptr, 10); + size_t K = strtol(argv[3], nullptr, 10); + + src0.allocator()->init(TensorInfo(TensorShape(K, M), 1, DataType::F32)); + src1.allocator()->init(TensorInfo(TensorShape(N, K), 1, DataType::F32)); + src2.allocator()->init(TensorInfo(TensorShape(N, M), 1, DataType::F32)); - if(argc > 5) + if(argc > 4) { - beta = strtof(argv[5], nullptr); + alpha = strtof(argv[4], nullptr); + + if(argc > 5) + { + beta = strtof(argv[5], nullptr); + } } } } - } - init_sgemm_output(dst, src0, src1, DataType::F32); + init_sgemm_output(dst, src0, src1, DataType::F32); - // Configure function - CLGEMM sgemm; - sgemm.configure(&src0, &src1, (src2.info()->total_size() > 0) ? &src2 : nullptr, &dst, alpha, beta); + // Configure function + sgemm.configure(&src0, &src1, (src2.info()->total_size() > 0) ? &src2 : nullptr, &dst, alpha, beta); - // Allocate all the images - src0.allocator()->allocate(); - src1.allocator()->allocate(); - dst.allocator()->allocate(); + // Allocate all the images + src0.allocator()->allocate(); + src1.allocator()->allocate(); + dst.allocator()->allocate(); - // Fill the input images with either the data provided or random data - if(npy0.is_open()) - { - npy0.fill_tensor(src0); - npy1.fill_tensor(src1); + // Fill the input images with either the data provided or random data + if(npy0.is_open()) + { + npy0.fill_tensor(src0); + npy1.fill_tensor(src1); + + output_filename = "sgemm_out.npy"; + is_fortran = npy0.is_fortran(); - if(npy2.is_open()) + if(npy2.is_open()) + { + src2.allocator()->allocate(); + npy2.fill_tensor(src2); + } + } + else { src2.allocator()->allocate(); - npy2.fill_tensor(src2); + + fill_random_tensor(src0, -1.f, 1.f); + fill_random_tensor(src1, -1.f, 1.f); + fill_random_tensor(src2, -1.f, 1.f); } - } - else - { - src2.allocator()->allocate(); - fill_random_tensor(src0, -1.f, 1.f); - fill_random_tensor(src1, -1.f, 1.f); - fill_random_tensor(src2, -1.f, 1.f); + // Dummy run for CLTuner + sgemm.run(); } - - // Dummy run for CLTuner - sgemm.run(); - - auto start = std::chrono::high_resolution_clock::now(); - - // Execute the function - sgemm.run(); - - // Make sure all the OpenCL jobs are done executing: - CLScheduler::get().sync(); - - auto stop = std::chrono::high_resolution_clock::now(); - - if(!npy0.is_open()) /* If the inputs were not files, print the results */ + void do_run() override { - std::cout << "\nMatrix 1:" << std::endl; - src0.map(true); - src0.print(std::cout, IOFormatInfo()); - src0.unmap(); - - std::cout << "Matrix 2:" << std::endl; - src1.map(true); - src1.print(std::cout, IOFormatInfo()); - src1.unmap(); - - std::cout << "Matrix 3:" << std::endl; - src2.map(true); - src2.print(std::cout, IOFormatInfo()); - src2.unmap(); - - std::cout << "Alpha:" << alpha << "\n\n"; - std::cout << "Beta:" << beta << "\n\n"; - - std::cout << "Output Matrix:" << std::endl; - dst.map(true); - dst.print(std::cout, IOFormatInfo()); - dst.unmap(); + // Execute the function + sgemm.run(); + + // Make sure all the OpenCL jobs are done executing: + CLScheduler::get().sync(); } - else /* Save to .npy file */ + void do_teardown() override { - save_to_npy(dst, "sgemm_out.npy", npy0.is_fortran()); + if(output_filename.empty()) /* If the inputs were not files, print the results */ + { + std::cout << "\nMatrix 1:" << std::endl; + src0.map(true); + src0.print(std::cout, IOFormatInfo()); + src0.unmap(); + + std::cout << "Matrix 2:" << std::endl; + src1.map(true); + src1.print(std::cout, IOFormatInfo()); + src1.unmap(); + + std::cout << "Matrix 3:" << std::endl; + src2.map(true); + src2.print(std::cout, IOFormatInfo()); + src2.unmap(); + + std::cout << "Alpha:" << alpha << "\n\n"; + std::cout << "Beta:" << beta << "\n\n"; + + std::cout << "Output Matrix:" << std::endl; + dst.map(true); + dst.print(std::cout, IOFormatInfo()); + dst.unmap(); + } + else /* Save to .npy file */ + { + save_to_npy(dst, output_filename, is_fortran); + } } - auto delta = std::chrono::duration_cast(stop - start); - std::cout << "Time elapsed: " << delta.count() << "us." << std::endl; -} +private: + CLTensor src0{}, src1{}, src2{}, dst{}; + CLGEMM sgemm{}; + CLTuner tuner{}; + float alpha{}, beta{}; + bool is_fortran{}; + std::string output_filename{}; +}; /** Main program for sgemm test * @@ -212,5 +222,5 @@ void main_cl_sgemm(int argc, char **argv) */ int main(int argc, char **argv) { - return utils::run_example(argc, argv, main_cl_sgemm); + return utils::run_example(argc, argv); } -- cgit v1.2.1