From 34702479adcf3559e8ebe27915179af509b070d2 Mon Sep 17 00:00:00 2001 From: steniu01 Date: Tue, 11 Jul 2017 09:22:58 +0100 Subject: COMPMID-417 Checking CL non uniform support at runtime. What have been done in the ticket are: 1. Add support to check whether cl-non-unform-workgroup is supported at runtime 2. Add helper function to check the CL version at runtime 3. Add boolen to check whether CLSecheduler's init has been called. Change-Id: I6e6df8eb5cebfac7229aa406242bb183477fd191 Reviewed-on: http://mpd-gerrit.cambridge.arm.com/80265 Tested-by: Kaizen Reviewed-by: Anthony Barbier --- src/core/CL/CLHelpers.cpp | 53 +++++++++++++++++++++++++++++++++++++++++ src/core/CL/CLKernelLibrary.cpp | 23 +++++++++++++++--- 2 files changed, 73 insertions(+), 3 deletions(-) (limited to 'src/core') diff --git a/src/core/CL/CLHelpers.cpp b/src/core/CL/CLHelpers.cpp index 835260d35a..dd87e778d7 100644 --- a/src/core/CL/CLHelpers.cpp +++ b/src/core/CL/CLHelpers.cpp @@ -166,4 +166,57 @@ GPUTarget get_arch_from_target(GPUTarget target) { return (target & GPUTarget::GPU_ARCH_MASK); } + +bool non_uniform_workgroup_support(const cl::Device &device) +{ + std::vector extension; + size_t extension_size = 0; + cl_int err = clGetDeviceInfo(device.get(), CL_DEVICE_EXTENSIONS, 0, nullptr, &extension_size); + ARM_COMPUTE_ERROR_ON_MSG((err != 0) || (extension_size == 0), "clGetDeviceInfo failed to return valid information"); + // Resize vector + extension.resize(extension_size); + // Query extension + err = clGetDeviceInfo(device.get(), CL_DEVICE_EXTENSIONS, extension_size, extension.data(), nullptr); + ARM_COMPUTE_ERROR_ON_MSG(err != 0, "clGetDeviceInfo failed to return valid information"); + ARM_COMPUTE_UNUSED(err); + + std::string extension_str(extension.begin(), extension.end()); + auto pos = extension_str.find("cl_arm_non_uniform_work_group_size"); + return (pos != std::string::npos); +} + +CLVersion get_cl_version(const cl::Device &device) +{ + std::vector version; + size_t version_size = 0; + cl_int err = clGetDeviceInfo(device.get(), CL_DEVICE_VERSION, 0, nullptr, &version_size); + ARM_COMPUTE_ERROR_ON_MSG((err != 0) || (version_size == 0), "clGetDeviceInfo failed to return valid information"); + // Resize vector + version.resize(version_size); + // Query version + err = clGetDeviceInfo(device.get(), CL_DEVICE_VERSION, version_size, version.data(), nullptr); + ARM_COMPUTE_ERROR_ON_MSG(err != 0, "clGetDeviceInfo failed to return valid information"); + ARM_COMPUTE_UNUSED(err); + + std::string version_str(version.begin(), version.end()); + if(version_str.find("OpenCL 2") != std::string::npos) + { + return CLVersion::CL20; + } + else if(version_str.find("OpenCL 1.2") != std::string::npos) + { + return CLVersion::CL12; + } + else if(version_str.find("OpenCL 1.1") != std::string::npos) + { + return CLVersion::CL11; + } + else if(version_str.find("OpenCL 1.0") != std::string::npos) + { + return CLVersion::CL10; + } + + return CLVersion::UNKNOWN; +} + } // namespace arm_compute diff --git a/src/core/CL/CLKernelLibrary.cpp b/src/core/CL/CLKernelLibrary.cpp index 72230435d8..769d50992d 100644 --- a/src/core/CL/CLKernelLibrary.cpp +++ b/src/core/CL/CLKernelLibrary.cpp @@ -23,6 +23,7 @@ */ #include "arm_compute/core/CL/CLKernelLibrary.h" +#include "arm_compute/core/CL/CLHelpers.h" #include "arm_compute/core/Error.h" #include "arm_compute/core/Utils.h" @@ -514,9 +515,25 @@ Kernel CLKernelLibrary::create_kernel(const std::string &kernel_name, const Stri ARM_COMPUTE_ERROR("Kernel %s not found in the CLKernelLibrary", kernel_name.c_str()); } + std::string concat_str; + + if(non_uniform_workgroup_support(_device)) + { + concat_str += " -cl-arm-non-uniform-work-group-size "; + } + else if(get_cl_version(_device) == CLVersion::CL20) + { + concat_str += " -cl-std=CL2.0 "; + } + else + { + ARM_COMPUTE_ERROR("Non uniform workgroup size is not supported!!"); + } + // Check if the program has been built before with same build options. - const std::string program_name = kernel_program_it->second; - const std::string build_options = stringify_set(build_options_set); + const std::string program_name = kernel_program_it->second; + const std::string build_options = stringify_set(build_options_set) + concat_str; + const std::string built_program_name = program_name + "_" + build_options; auto built_program_it = _built_programs_map.find(built_program_name); @@ -591,7 +608,7 @@ const Program &CLKernelLibrary::load_program(const std::string &program_name) co std::string CLKernelLibrary::stringify_set(const StringSet &s) const { - std::string concat_set = "-cl-arm-non-uniform-work-group-size "; + std::string concat_set; #ifndef EMBEDDED_KERNELS concat_set += "-I" + _kernel_path + " "; -- cgit v1.2.1