src/gromacs/ewald/pme-gpu-program-impl-ocl.cpp

   1 /*
   2  * This file is part of the GROMACS molecular simulation package.
   3  *
   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.
   8  *
   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.
  13  *
  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.
  18  *
  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.
  23  *
  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.
  31  *
  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.
  34  */
  35
  36 /*! \internal \file
  37  * \brief
  38  * Implements PmeGpuProgramImpl, which stores permanent PME GPU context-derived data,
  39  * such as (compiled) kernel handles.
  40  *
  41  * \author Aleksei Iupinov <a.yupinov@gmail.com>
  42  * \ingroup module_ewald
  43  */
  44 #include "gmxpre.h"
  45
  46 #include "gromacs/gpu_utils/gmxopencl.h"
  47 #include "gromacs/gpu_utils/ocl_compiler.h"
  48 #include "gromacs/utility/stringutil.h"
  49
  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"
  54 #include "pme-grid.h"
  55
  56 PmeGpuProgramImpl::PmeGpuProgramImpl(const gmx_device_info_t *deviceInfo)
  57 {
  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 */
  65
  66     cl_int  clError;
  67     context = clCreateContext(contextProperties, 1, &deviceId, nullptr, nullptr, &clError);
  68     if (clError != CL_SUCCESS)
  69     {
  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));
  73     }
  74
  75     // kernel parameters
  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);
  85
  86     compileKernels(deviceInfo);
  87 }
  88
  89 PmeGpuProgramImpl::~PmeGpuProgramImpl()
  90 {
  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());
 105 }
 106
 107 /*! \brief Ensure that spread/gather kernels have been compiled to a suitable warp size
 108  *
 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.
 112  */
 113 static void checkRequiredWarpSize(const cl_kernel          kernel,
 114                                   const char*              kernelName,
 115                                   const gmx_device_info_t *deviceInfo)
 116 {
 117     if (deviceInfo->vendor_e == OCL_VENDOR_INTEL)
 118     {
 119         size_t kernelWarpSize = gmx::ocl::getKernelWarpSize(kernel, deviceInfo->ocl_gpu_id.ocl_device_id);
 120
 121         if (kernelWarpSize < c_pmeSpreadGatherMinWarpSize)
 122         {
 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));
 128         }
 129     }
 130 }
 131
 132 void PmeGpuProgramImpl::compileKernels(const gmx_device_info_t *deviceInfo)
 133 {
 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. */
 137     try
 138     {
 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.
 142          */
 143         const std::string commonDefines = gmx::formatString(
 144                     "-Dwarp_size=%zd "
 145                     "-Dorder=%d "
 146                     "-DatomsPerWarp=%zd "
 147                     "-DthreadsPerAtom=%d "
 148                     // forwarding from pme-grid.h, used for spline computation table sizes only
 149                     "-Dc_pmeMaxUnitcellShift=%f "
 150                     // forwarding PME behavior constants from pme-gpu-constants.h
 151                     "-Dc_usePadding=%d "
 152                     "-Dc_skipNeutralAtoms=%d "
 153                     "-Dc_virialAndEnergyCount=%d "
 154                     // forwarding kernel work sizes
 155                     "-Dc_spreadWorkGroupSize=%zd "
 156                     "-Dc_solveMaxWorkGroupSize=%zd "
 157                     "-Dc_gatherWorkGroupSize=%zd "
 158                     // forwarding from vectypes.h
 159                     "-DDIM=%d -DXX=%d -DYY=%d -DZZ=%d "
 160                     // decomposition parameter placeholders
 161                     "-DwrapX=true -DwrapY=true ",
 162                     warpSize,
 163                     c_pmeGpuOrder,
 164                     warpSize / c_pmeSpreadGatherThreadsPerAtom,
 165                     c_pmeSpreadGatherThreadsPerAtom,
 166                     static_cast<float>(c_pmeMaxUnitcellShift),
 167                     c_usePadding,
 168                     c_skipNeutralAtoms,
 169                     c_virialAndEnergyCount,
 170                     spreadWorkGroupSize,
 171                     solveMaxWorkGroupSize,
 172                     gatherWorkGroupSize,
 173                     DIM, XX, YY, ZZ);
 174         try
 175         {
 176             /* TODO when we have a proper MPI-aware logging module,
 177                the log output here should be written there */
 178             program = gmx::ocl::compileProgram(stderr,
 179                                                "gromacs/ewald",
 180                                                "pme-program.cl",
 181                                                commonDefines,
 182                                                context,
 183                                                deviceInfo->ocl_gpu_id.ocl_device_id,
 184                                                deviceInfo->vendor_e);
 185         }
 186         catch (gmx::GromacsException &e)
 187         {
 188             e.prependContext(gmx::formatString("Failed to compile PME kernels for GPU #%s\n",
 189                                                deviceInfo->device_name));
 190             throw;
 191         }
 192     }
 193     GMX_CATCH_ALL_AND_EXIT_WITH_FATAL_ERROR;
 194
 195     constexpr cl_uint      expectedKernelCount = 9;
 196     // Has to be equal or larger than the number of kernel instances.
 197     // If it is not, CL_INVALID_VALUE will be thrown.
 198     std::vector<cl_kernel> kernels(expectedKernelCount, nullptr);
 199     cl_uint                actualKernelCount = 0;
 200     cl_int                 clError           = clCreateKernelsInProgram(program, kernels.size(), kernels.data(), &actualKernelCount);
 201     if (clError != CL_SUCCESS)
 202     {
 203         const std::string errorString = gmx::formatString("Failed to create kernels for PME on GPU #%s:\n OpenCL error %d: %s",
 204                                                           deviceInfo->device_name, clError, ocl_get_error_string(clError).c_str());
 205         GMX_THROW(gmx::InternalError(errorString));
 206     }
 207     kernels.resize(actualKernelCount);
 208
 209     std::array<char, 100> kernelNamesBuffer;
 210     for (const auto &kernel : kernels)
 211     {
 212         clError = clGetKernelInfo(kernel, CL_KERNEL_FUNCTION_NAME,
 213                                   kernelNamesBuffer.size(), kernelNamesBuffer.data(), nullptr);
 214         if (clError != CL_SUCCESS)
 215         {
 216             const std::string errorString = gmx::formatString("Failed to parse kernels for PME on GPU #%s:\n OpenCL error %d: %s",
 217                                                               deviceInfo->device_name, clError, ocl_get_error_string(clError).c_str());
 218             GMX_THROW(gmx::InternalError(errorString));
 219         }
 220
 221         // The names below must correspond to those defined in pme-program.cl
 222         // TODO use a map with string key instead?
 223         if (!strcmp(kernelNamesBuffer.data(), "pmeSplineKernel"))
 224         {
 225             splineKernel = kernel;
 226         }
 227         else if (!strcmp(kernelNamesBuffer.data(), "pmeSplineAndSpreadKernel"))
 228         {
 229             splineAndSpreadKernel = kernel;
 230             checkRequiredWarpSize(splineAndSpreadKernel, kernelNamesBuffer.data(), deviceInfo);
 231         }
 232         else if (!strcmp(kernelNamesBuffer.data(), "pmeSpreadKernel"))
 233         {
 234             spreadKernel = kernel;
 235             checkRequiredWarpSize(spreadKernel, kernelNamesBuffer.data(), deviceInfo);
 236         }
 237         else if (!strcmp(kernelNamesBuffer.data(), "pmeGatherKernel"))
 238         {
 239             gatherKernel = kernel;
 240             checkRequiredWarpSize(gatherKernel, kernelNamesBuffer.data(), deviceInfo);
 241         }
 242         else if (!strcmp(kernelNamesBuffer.data(), "pmeGatherReduceWithInputKernel"))
 243         {
 244             gatherReduceWithInputKernel = kernel;
 245             checkRequiredWarpSize(gatherReduceWithInputKernel, kernelNamesBuffer.data(), deviceInfo);
 246         }
 247         else if (!strcmp(kernelNamesBuffer.data(), "pmeSolveYZXKernel"))
 248         {
 249             solveYZXKernel = kernel;
 250         }
 251         else if (!strcmp(kernelNamesBuffer.data(), "pmeSolveYZXEnergyKernel"))
 252         {
 253             solveYZXEnergyKernel = kernel;
 254         }
 255         else if (!strcmp(kernelNamesBuffer.data(), "pmeSolveXYZKernel"))
 256         {
 257             solveXYZKernel = kernel;
 258         }
 259         else if (!strcmp(kernelNamesBuffer.data(), "pmeSolveXYZEnergyKernel"))
 260         {
 261             solveXYZEnergyKernel = kernel;
 262         }
 263     }
 264     clReleaseProgram(program);
 265 }