diff options
Diffstat (limited to 'examples/gemm_tuner/cl_gemm_reshaped_rhs_only.cpp')
| -rw-r--r-- | examples/gemm_tuner/cl_gemm_reshaped_rhs_only.cpp | 66 |
1 files changed, 36 insertions, 30 deletions
diff --git a/examples/gemm_tuner/cl_gemm_reshaped_rhs_only.cpp b/examples/gemm_tuner/cl_gemm_reshaped_rhs_only.cpp index dbaaca6048..cfea2c9bac 100644 --- a/examples/gemm_tuner/cl_gemm_reshaped_rhs_only.cpp +++ b/examples/gemm_tuner/cl_gemm_reshaped_rhs_only.cpp @@ -1,5 +1,5 @@ /* - * Copyright (c) 2019-2021 Arm Limited. + * Copyright (c) 2019-2021, 2023 Arm Limited. * * SPDX-License-Identifier: MIT * @@ -25,20 +25,21 @@ #error "This example needs to be built with -DARM_COMPUTE_CL" #endif /* ARM_COMPUTE_CL */ -#include "CommonGemmExampleOptions.h" -#include "GemmTunerHelpers.h" #include "arm_compute/core/Helpers.h" #include "arm_compute/core/KernelDescriptors.h" #include "arm_compute/core/Types.h" #include "arm_compute/core/utils/misc/ShapeCalculator.h" #include "arm_compute/runtime/CL/CLScheduler.h" #include "arm_compute/runtime/CL/CLTuner.h" -#include "src/core/gpu/cl/kernels/ClGemmMatrixMultiplyReshapedOnlyRhsKernel.h" + +#include "src/gpu/cl/kernels/ClGemmMatrixMultiplyReshapedOnlyRhsKernel.h" #include "tests/CL/Helper.h" -#include "utils/Utils.h" #include "utils/command_line/CommandLineOptions.h" #include "utils/command_line/CommandLineParser.h" +#include "utils/Utils.h" +#include "CommonGemmExampleOptions.h" +#include "GemmTunerHelpers.h" #include <cstdlib> using namespace arm_compute; @@ -52,13 +53,13 @@ namespace /** Structure holding all tunable gemm configs specific to this example/strategy */ struct GemmConfigs { - size_t m0{ 4 }; /**< Number of rows processed by the matrix multiplication */ - size_t n0{ 4 }; /**< Number of columns processed by the matrix multiplication */ - size_t k0{ 4 }; /**< Number of partial accumulations performed by the matrix multiplication */ - size_t h0{ 1 }; /**< Number of horizontal blocks of size (k0xn0) stored on the same output row */ - bool interleave_rhs{ true }; /**< Interleave rhs matrix */ - bool transpose_rhs{ true }; /**< Transpose rhs matrix */ - bool export_to_cl_image_rhs{ true }; /**< Export rhs matrix to cl_image.*/ + size_t m0{4}; /**< Number of rows processed by the matrix multiplication */ + size_t n0{4}; /**< Number of columns processed by the matrix multiplication */ + size_t k0{4}; /**< Number of partial accumulations performed by the matrix multiplication */ + size_t h0{1}; /**< Number of horizontal blocks of size (k0xn0) stored on the same output row */ + bool interleave_rhs{true}; /**< Interleave rhs matrix */ + bool transpose_rhs{true}; /**< Transpose rhs matrix */ + bool export_to_cl_image_rhs{true}; /**< Export rhs matrix to cl_image.*/ }; /** Formatted output of the GemmConfigs type @@ -106,7 +107,8 @@ public: h0->set_help("Number of horizontal blocks of size (k0xn0) stored on the same output row"); interleave_rhs->set_help("Interleave rhs matrix (1) / Do not interleave rhs matrix (0)"); transpose_rhs->set_help("Transpose rhs matrix (1) / Do not transpose rhs matrix (0)"); - export_to_cl_image_rhs->set_help("Export rhs matrix to cl_image (1) / Do not export rhs matrix to cl_image (0)"); + export_to_cl_image_rhs->set_help( + "Export rhs matrix to cl_image (1) / Do not export rhs matrix to cl_image (0)"); } /** Prevent instances of this class from being copied (As this class contains pointers) */ GemmConfigOptions(const GemmConfigOptions &) = delete; @@ -119,10 +121,10 @@ public: /** Default destructor */ ~GemmConfigOptions() = default; - SimpleOption<size_t> *m0; /**< Number of rows processed by the matrix multiplication option */ - SimpleOption<size_t> *n0; /**< Number of columns processed by the matrix multiplication option */ - SimpleOption<size_t> *k0; /**< Number of partial accumulations performed by the matrix multiplication option */ - SimpleOption<size_t> *h0; /**< Number of horizontal blocks of size (k0xn0) stored on the same output row option */ + SimpleOption<size_t> *m0; /**< Number of rows processed by the matrix multiplication option */ + SimpleOption<size_t> *n0; /**< Number of columns processed by the matrix multiplication option */ + SimpleOption<size_t> *k0; /**< Number of partial accumulations performed by the matrix multiplication option */ + SimpleOption<size_t> *h0; /**< Number of horizontal blocks of size (k0xn0) stored on the same output row option */ SimpleOption<size_t> *interleave_rhs; /**< Interleave rhs matrix option (1 enable; 0 disable) */ SimpleOption<size_t> *transpose_rhs; /**< Transpose rhs matrix option (1 enable; 0 disable) */ SimpleOption<size_t> *export_to_cl_image_rhs; /**< Export rhs matrix to cl_image.*/ @@ -170,13 +172,13 @@ public: // Parse command line options parser.parse(argc, argv); - if(param_options.help->is_set() && param_options.help->value()) + if (param_options.help->is_set() && param_options.help->value()) { // Print help message parser.print_help(argv[0]); return false; } - if(!parser.validate()) + if (!parser.validate()) { // Invalid arguments. Use default parameters and configs std::cerr << "Invalid arguments." << std::endl; @@ -225,12 +227,18 @@ public: kernel_info.broadcast_bias = true; kernel_info.activation_info = act_info; + if (rhs_info.h0 == 0) + { + rhs_info.h0 = std::max(kernel_info.n / rhs_info.n0, 1U); + } + // Initialise rhs_reshaped tensor info - rhs_reshaped.allocator()->init(TensorInfo(compute_rhs_reshaped_shape(*rhs.info(), rhs_info), 1, params.data_type)); + rhs_reshaped.allocator()->init( + TensorInfo(compute_rhs_reshaped_shape(*rhs.info(), rhs_info), 1, params.data_type)); - if(rhs_info.export_to_cl_image) + if (rhs_info.export_to_cl_image) { - if(!examples::gemm_tuner_helpers::update_padding_for_cl_image(rhs_reshaped.info())) + if (!examples::gemm_tuner_helpers::update_padding_for_cl_image(rhs_reshaped.info())) { std::cerr << "cl_image is not supported on the device, disable export_to_cl_image" << std::endl; return false; @@ -239,8 +247,9 @@ public: // Validate argments Status status{}; - status = gemm.validate(lhs.info(), rhs_reshaped.info(), bias.info(), dst.info(), alpha, beta, lhs_info, rhs_info, kernel_info); - if(!status) + status = gemm.validate(lhs.info(), rhs_reshaped.info(), bias.info(), dst.info(), alpha, beta, lhs_info, + rhs_info, kernel_info); + if (!status) { // Unsupported arguments std::cerr << "Unsupported arguments." << std::endl; @@ -249,7 +258,8 @@ public: } // Configure function - gemm.configure(lhs.info(), rhs_reshaped.info(), bias.info(), dst.info(), alpha, beta, lhs_info, rhs_info, kernel_info); + gemm.configure(lhs.info(), rhs_reshaped.info(), bias.info(), dst.info(), alpha, beta, lhs_info, rhs_info, + kernel_info); // Allocate tensors lhs.allocator()->allocate(); @@ -263,11 +273,7 @@ public: void do_run() override { // Execute the function - ITensorPack gemm_pack({ { ACL_SRC_0, &lhs }, - { ACL_SRC_1, &rhs_reshaped }, - { ACL_SRC_2, &bias }, - { ACL_DST, &dst } - }); + ITensorPack gemm_pack({{ACL_SRC_0, &lhs}, {ACL_SRC_1, &rhs_reshaped}, {ACL_SRC_2, &bias}, {ACL_DST, &dst}}); gemm.run(gemm_pack); // Make sure all the OpenCL jobs are done executing: |