/* * Advanced Simulation Library * * Copyright 2015 Avtech Scientific * * * This file is part of Advanced Simulation Library (ASL). * * ASL is free software: you can redistribute it and/or modify it * under the terms of the GNU Affero General Public License as * published by the Free Software Foundation, version 3 of the License. * * ASL is distributed in the hope that it will be useful, * but WITHOUT ANY WARRANTY; without even the implied warranty of * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the * GNU Affero General Public License for more details. * * You should have received a copy of the GNU Affero General Public License * along with ASL. If not, see . * */ #include "aclKernel.h" #include "../aclHardware.h" #include #include #include "../../aslUtilities.h" #include using namespace std; using namespace asl; namespace acl { unsigned int Kernel::kernelNum = 0; Kernel::Kernel(const KernelConfiguration kernelConfig_): groupsNumber(0), kernelConfig(kernelConfig_), kernelSource("") { id = kernelNum; ++kernelNum; } cl_uint Kernel::detectVectorWidth() { vector widths; // get all possible vector widths for this queue for (unsigned int i = 0; i < TYPE.size(); ++i) widths.push_back(getVectorWidth(queue, (TypeID)i)); cl_uint absoluteMin = *min_element(widths.begin(), widths.end()); cl_uint min = *max_element(widths.begin(), widths.end()); unsigned int i = 0; while ((min > absoluteMin) && (i < arguments.size())) { if (min > widths[arguments[i]->getTypeID()]) min = widths[arguments[i]->getTypeID()]; ++i; } i = 0; while ((min > absoluteMin) && (i < localDeclarations.size())) { if (min > widths[localDeclarations[i]->getTypeID()]) min = widths[localDeclarations[i]->getTypeID()]; ++i; } // Workaround for the case when double vector width is 0. // The code still works, however with acceleration // only for types other than double. min = min == 0 ? 1 : min; return min; } // Enables double extension only if there are doubles in the source, // enables right extension (workaround for the incomplete AMD cl_amd_fp64) void enableDoubleExtension(string & kernelSource, const CommandQueue & queue) { if (kernelSource.find(TYPE[TYPE_DOUBLE]) != string::npos) { string availableExtensions = getDevice(queue).getInfo(); // workaround for the incomplete AMD cl_amd_fp64. // once AMD implements this extension completely // only the else-block should stay if (availableExtensions.find("cl_amd_fp64") != string::npos) { kernelSource = "#pragma OPENCL EXTENSION cl_amd_fp64 : enable\n" + kernelSource; } else { kernelSource = "#pragma OPENCL EXTENSION cl_khr_fp64 : enable\n" + kernelSource; } } else { string availableExtensions = getDevice(queue).getInfo(); // workaround for the incomplete AMD cl_amd_fp64. // once AMD implements this extension completely // only the else-block should stay if (availableExtensions.find("cl_amd_fp64") != string::npos) { kernelSource = "#pragma OPENCL EXTENSION cl_amd_fp64 : disable\n" + kernelSource; } else { kernelSource = "#pragma OPENCL EXTENSION cl_khr_fp64 : disable\n" + kernelSource; } } } void Kernel::generateExtensions() { kernelSource = "\n" + kernelSource; for (unsigned int i = 0; i < kernelConfig.extensions.size(); i++) { kernelSource = "#pragma OPENCL EXTENSION " + kernelConfig.extensions[i] + " : enable\n" + kernelSource; } enableDoubleExtension(kernelSource, queue); } void Kernel::generateArguments() { for (unsigned int i = 0; i < arguments.size(); i++) { kernelSource += arguments[i]->getTypeSignature(kernelConfig) + ",\n "; for (unsigned int j = 0; j < numToStr(id).size(); ++j) { kernelSource += " "; } } // drop all after the ',' of last parameter kernelSource.erase(kernelSource.size() - 25 - numToStr(id).size()); kernelSource += ")\n{\n\t"; } void Kernel::generateIndex() { if (kernelConfig.local) { kernelSource += "uint " + INDEX + " = get_local_id(0);\n\t"; kernelSource += "uint groupID = get_group_id(0);\n"; } else { if ((kernelConfig.vectorWidth > 1) && (kernelConfig.unaligned)) { kernelSource += "uint " + INDEX + " = " + numToStr(kernelConfig.vectorWidth) + " * get_global_id(0);\n"; } else { kernelSource += "uint " + INDEX + " = get_global_id(0);\n"; } } } void Kernel::generateLocalDeclarations() { for (unsigned int i = 0; i < localDeclarations.size(); i++) { kernelSource += "\t" + localDeclarations[i]->getLocalDeclaration(kernelConfig) + ";\n"; } } void Kernel::generateExpressions() { for (unsigned int i = 0; i < expression.size(); i++) { kernelSource += "\t" + expression[i]->str(kernelConfig) + ";\n"; } kernelSource += "}"; } void Kernel::generateKernelSource() { kernelSource = "__kernel void compute_" + numToStr(id) + "("; filterDeclarations(); generateArguments(); generateIndex(); generateLocalDeclarations(); generateExpressions(); // enable extensions in the last step after // the kernel source is available for analysis generateExtensions(); regenerateKernelSource = false; } void Kernel::buildKernel() { cl_int status = 0; cl::Program::Sources source(1, make_pair(kernelSource.data(), kernelSource.size())); cl::Program program = cl::Program(getContext(queue), source, &status); errorMessage(status, "Program::Program()"); status = program.build(std::vector(1, getDevice(queue))); if (status == CL_BUILD_PROGRAM_FAILURE) { string str = program.getBuildInfo(getDevice(queue)); cout << " \n\t\t\tBUILD LOG\n"; cout << " ************************************************\n"; cout << str << endl; cout << " ************************************************\n"; cout << " \n\t\t\tKERNEL SOURCE CODE\n"; cout << " ------------------------------------------------\n"; cout << kernelSource << endl; cout << " ------------------------------------------------\n"; } errorMessage(status, "Program::build()"); string kernelName; kernelName = "compute_" + numToStr(id); kernel = cl::Kernel(program, kernelName.c_str(), &status); errorMessage(status, "Kernel() - kernelBase_" + numToStr(id)); // cout<<" "< getMaxItemSize(queue))) errorMessage("Kernel::setup() - requested group size is larger than supported by device"); if (kernelConfig.local && (groupsNumber == 0)) errorMessage("Kernel::setup() - groups number was not set"); updateKernelConfiguration(); generateKernelSource(); buildKernel(); } inline unsigned int paddingElementsSIMD(unsigned int size, unsigned int vectorWidth); void Kernel::compute() { if (regenerateKernelSource) setup(); cl_int status = 0; cl::Event event; setKernelArguments(); size_t paddedSize = (size + paddingElementsSIMD(size, kernelConfig.vectorWidth)) / kernelConfig.vectorWidth; // if 4th argument is cl::NullRange - OpenCL implementation // will determine how to break the global work-items into // appropriate work-group instances status = queue->enqueueNDRangeKernel(kernel, cl::NullRange, kernelConfig.local ? cl::NDRange(groupsNumber * paddedSize) : cl::NDRange(paddedSize), kernelConfig.local ? cl::NDRange(paddedSize) : cl::NullRange, NULL, &event); errorMessage(status, "CommandQueue::enqueueNDRangeKernel() - kernel"); status = event.wait(); errorMessage(status, "Event::wait() - event"); } void Kernel::setGroupsNumber(unsigned int n) { groupsNumber = n; } unsigned int Kernel::getGroupsNumber() { return groupsNumber; } string Kernel::getKernelSource() { return kernelSource; } unsigned int Kernel::getKernelID() { return id; } const cl::Kernel & Kernel::getKernel() const { return kernel; } void Kernel::clear() { kernelSource = "__kernel void compute_" + numToStr(id) + "("; regenerateKernelSource = true; expression.clear(); localDeclarations.clear(); arguments.clear(); size = 0; } inline unsigned int paddingElementsSIMD(unsigned int size, unsigned int vectorWidth) { // second modulo is added in order to make padding = 0 in the case // that size is divisible by vectorWidth return (vectorWidth - (size % vectorWidth)) % vectorWidth; } } // namespace acl