2 * This file is part of the GROMACS molecular simulation package.
4 * Copyright (c) 2018, by the GROMACS development team, led by
5 * Mark Abraham, David van der Spoel, Berk Hess, and Erik Lindahl,
6 * and including many others, as listed in the AUTHORS file in the
7 * top-level source directory and at http://www.gromacs.org.
9 * GROMACS is free software; you can redistribute it and/or
10 * modify it under the terms of the GNU Lesser General Public License
11 * as published by the Free Software Foundation; either version 2.1
12 * of the License, or (at your option) any later version.
14 * GROMACS is distributed in the hope that it will be useful,
15 * but WITHOUT ANY WARRANTY; without even the implied warranty of
16 * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the GNU
17 * Lesser General Public License for more details.
19 * You should have received a copy of the GNU Lesser General Public
20 * License along with GROMACS; if not, see
21 * http://www.gnu.org/licenses, or write to the Free Software Foundation,
22 * Inc., 51 Franklin Street, Fifth Floor, Boston, MA 02110-1301 USA.
24 * If you want to redistribute modifications to GROMACS, please
25 * consider that scientific software is very special. Version
26 * control is crucial - bugs must be traceable. We will be happy to
27 * consider code for inclusion in the official distribution, but
28 * derived work must not be called official GROMACS. Details are found
29 * in the README & COPYING files - if they are missing, get the
30 * official version at http://www.gromacs.org.
32 * To help us fund GROMACS development, we humbly ask that you cite
33 * the research papers on the package. Check out http://www.gromacs.org.
38 * Implements PmeGpuProgramImpl, which stores permanent PME GPU context-derived data,
39 * such as (compiled) kernel handles.
41 * \author Aleksei Iupinov <a.yupinov@gmail.com>
42 * \ingroup module_ewald
46 #include "gromacs/gpu_utils/gmxopencl.h"
47 #include "gromacs/gpu_utils/ocl_compiler.h"
48 #include "gromacs/utility/stringutil.h"
50 #include "pme-gpu-constants.h"
51 #include "pme-gpu-internal.h" // for GridOrdering enum
52 #include "pme-gpu-program-impl.h"
53 #include "pme-gpu-types-host.h"
56 PmeGpuProgramImpl::PmeGpuProgramImpl(const gmx_device_info_t *deviceInfo)
58 // Context creation (which should happen outside of this class: #2522)
59 cl_platform_id platformId = deviceInfo->ocl_gpu_id.ocl_platform_id;
60 cl_device_id deviceId = deviceInfo->ocl_gpu_id.ocl_device_id;
61 cl_context_properties contextProperties[3];
62 contextProperties[0] = CL_CONTEXT_PLATFORM;
63 contextProperties[1] = reinterpret_cast<cl_context_properties>(platformId);
64 contextProperties[2] = 0; /* Terminates the list of properties */
67 context = clCreateContext(contextProperties, 1, &deviceId, nullptr, nullptr, &clError);
68 if (clError != CL_SUCCESS)
70 const std::string errorString = gmx::formatString("Failed to create context for PME on GPU #%s:\n OpenCL error %d: %s",
71 deviceInfo->device_name, clError, ocl_get_error_string(clError).c_str());
72 GMX_THROW(gmx::InternalError(errorString));
76 warpSize = gmx::ocl::getDeviceWarpSize(context, deviceId);
77 // TODO: for Intel ideally we'd want to set these based on the compiler warp size
78 // but given that we've done no tuning for Intel iGPU, this is as good as anything.
79 spreadWorkGroupSize = std::min(c_spreadMaxWarpsPerBlock * warpSize,
80 deviceInfo->maxWorkGroupSize);
81 solveMaxWorkGroupSize = std::min(c_solveMaxWarpsPerBlock * warpSize,
82 deviceInfo->maxWorkGroupSize);
83 gatherWorkGroupSize = std::min(c_gatherMaxWarpsPerBlock * warpSize,
84 deviceInfo->maxWorkGroupSize);
86 compileKernels(deviceInfo);
89 PmeGpuProgramImpl::~PmeGpuProgramImpl()
91 // TODO: log releasing errors
92 cl_int gmx_used_in_debug stat = 0;
93 stat |= clReleaseKernel(splineAndSpreadKernel);
94 stat |= clReleaseKernel(splineKernel);
95 stat |= clReleaseKernel(spreadKernel);
96 stat |= clReleaseKernel(gatherKernel);
97 stat |= clReleaseKernel(gatherReduceWithInputKernel);
98 stat |= clReleaseKernel(solveXYZKernel);
99 stat |= clReleaseKernel(solveXYZEnergyKernel);
100 stat |= clReleaseKernel(solveYZXKernel);
101 stat |= clReleaseKernel(solveYZXEnergyKernel);
102 stat |= clReleaseContext(context);
103 GMX_ASSERT(stat == CL_SUCCESS, gmx::formatString("Failed to release PME OpenCL resources %d: %s",
104 stat, ocl_get_error_string(stat).c_str()).c_str());
107 /*! \brief Ensure that spread/gather kernels have been compiled to a suitable warp size
109 * On Intel the exec width/warp is decided at compile-time and can be
110 * smaller than the minimum order^2 required in spread/gather ATM which
111 * we need to check for.
113 static void checkRequiredWarpSize(const cl_kernel kernel,
114 const char* kernelName,
115 const gmx_device_info_t *deviceInfo)
117 if (deviceInfo->vendor_e == OCL_VENDOR_INTEL)
119 size_t kernelWarpSize = gmx::ocl::getKernelWarpSize(kernel, deviceInfo->ocl_gpu_id.ocl_device_id);
121 if (kernelWarpSize < c_pmeSpreadGatherMinWarpSize)
123 const std::string errorString = gmx::formatString("PME OpenCL kernels require >=%d execution width, but the %s kernel "
124 "has been compiled for the device %s to a %zu width and therefore it can not execute correctly.",
125 c_pmeSpreadGatherMinWarpSize, kernelName,
126 deviceInfo->device_name, kernelWarpSize);
127 GMX_THROW(gmx::InternalError(errorString));
132 void PmeGpuProgramImpl::compileKernels(const gmx_device_info_t *deviceInfo)
134 // We might consider storing program as a member variable if it's needed later
135 cl_program program = nullptr;
136 /* Need to catch std::bad_alloc here and during compilation string handling. */
139 /* Here we pass macros and static const int variables defined in include
140 * files outside as macros, to avoid including those files
141 * in the JIT compilation that happens at runtime.
143 constexpr int order = 4;
144 const std::string commonDefines = gmx::formatString(
147 "-DPME_SPREADGATHER_ATOMS_PER_WARP=%zd "
148 "-DPME_SPREADGATHER_THREADS_PER_ATOM=%d "
149 // forwarding from pme-grid.h, used for spline computation table sizes only
150 "-Dc_pmeMaxUnitcellShift=%f "
151 // forwarding PME behavior constants from pme-gpu-constants.h
153 "-Dc_skipNeutralAtoms=%d "
154 "-Dc_virialAndEnergyCount=%d "
155 // forwarding kernel work sizes
156 "-Dc_spreadWorkGroupSize=%zd "
157 "-Dc_solveMaxWorkGroupSize=%zd "
158 "-Dc_gatherWorkGroupSize=%zd "
159 // forwarding from vectypes.h
160 "-DDIM=%d -DXX=%d -DYY=%d -DZZ=%d "
161 // decomposition parameter placeholders
162 "-DwrapX=true -DwrapY=true ",
165 warpSize / PME_SPREADGATHER_THREADS_PER_ATOM,
166 PME_SPREADGATHER_THREADS_PER_ATOM,
167 static_cast<float>(c_pmeMaxUnitcellShift),
170 c_virialAndEnergyCount,
172 solveMaxWorkGroupSize,
177 /* TODO when we have a proper MPI-aware logging module,
178 the log output here should be written there */
179 program = gmx::ocl::compileProgram(stderr,
184 deviceInfo->ocl_gpu_id.ocl_device_id,
185 deviceInfo->vendor_e);
187 catch (gmx::GromacsException &e)
189 e.prependContext(gmx::formatString("Failed to compile PME kernels for GPU #%s\n",
190 deviceInfo->device_name));
194 GMX_CATCH_ALL_AND_EXIT_WITH_FATAL_ERROR;
196 constexpr cl_uint expectedKernelCount = 9;
197 // Has to be equal or larger than the number of kernel instances.
198 // If it is not, CL_INVALID_VALUE will be thrown.
199 std::vector<cl_kernel> kernels(expectedKernelCount, nullptr);
200 cl_uint actualKernelCount = 0;
201 cl_int clError = clCreateKernelsInProgram(program, kernels.size(), kernels.data(), &actualKernelCount);
202 if (clError != CL_SUCCESS)
204 const std::string errorString = gmx::formatString("Failed to create kernels for PME on GPU #%s:\n OpenCL error %d: %s",
205 deviceInfo->device_name, clError, ocl_get_error_string(clError).c_str());
206 GMX_THROW(gmx::InternalError(errorString));
208 kernels.resize(actualKernelCount);
210 std::array<char, 100> kernelNamesBuffer;
211 for (const auto &kernel : kernels)
213 clError = clGetKernelInfo(kernel, CL_KERNEL_FUNCTION_NAME,
214 kernelNamesBuffer.size(), kernelNamesBuffer.data(), nullptr);
215 if (clError != CL_SUCCESS)
217 const std::string errorString = gmx::formatString("Failed to parse kernels for PME on GPU #%s:\n OpenCL error %d: %s",
218 deviceInfo->device_name, clError, ocl_get_error_string(clError).c_str());
219 GMX_THROW(gmx::InternalError(errorString));
222 // The names below must correspond to those defined in pme-program.cl
223 // TODO use a map with string key instead?
224 if (!strcmp(kernelNamesBuffer.data(), "pmeSplineKernel"))
226 splineKernel = kernel;
228 else if (!strcmp(kernelNamesBuffer.data(), "pmeSplineAndSpreadKernel"))
230 splineAndSpreadKernel = kernel;
231 checkRequiredWarpSize(splineAndSpreadKernel, kernelNamesBuffer.data(), deviceInfo);
233 else if (!strcmp(kernelNamesBuffer.data(), "pmeSpreadKernel"))
235 spreadKernel = kernel;
236 checkRequiredWarpSize(spreadKernel, kernelNamesBuffer.data(), deviceInfo);
238 else if (!strcmp(kernelNamesBuffer.data(), "pmeGatherKernel"))
240 gatherKernel = kernel;
241 checkRequiredWarpSize(gatherKernel, kernelNamesBuffer.data(), deviceInfo);
243 else if (!strcmp(kernelNamesBuffer.data(), "pmeGatherReduceWithInputKernel"))
245 gatherReduceWithInputKernel = kernel;
246 checkRequiredWarpSize(gatherReduceWithInputKernel, kernelNamesBuffer.data(), deviceInfo);
248 else if (!strcmp(kernelNamesBuffer.data(), "pmeSolveYZXKernel"))
250 solveYZXKernel = kernel;
252 else if (!strcmp(kernelNamesBuffer.data(), "pmeSolveYZXEnergyKernel"))
254 solveYZXEnergyKernel = kernel;
256 else if (!strcmp(kernelNamesBuffer.data(), "pmeSolveXYZKernel"))
258 solveXYZKernel = kernel;
260 else if (!strcmp(kernelNamesBuffer.data(), "pmeSolveXYZEnergyKernel"))
262 solveXYZEnergyKernel = kernel;
265 clReleaseProgram(program);