src/gromacs/gpu_utils/tests/device_buffer.cpp

   1 /*
   2  * This file is part of the GROMACS molecular simulation package.
   3  *
   4  * Copyright (c) 2020,2021, 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 /*! \internal \file
  36  * \brief Tests for device buffer
  37  *
  38  * \author Artem Zhmurov <zhmurov@gmail.com>
  39  *
  40  * \ingroup module_gpu_utils
  41  */
  42 #include "gmxpre.h"
  43
  44 #include "config.h"
  45
  46 #if GMX_GPU
  47 #    include <numeric>
  48
  49 #    include <gmock/gmock.h>
  50 #    include <gtest/gtest.h>
  51
  52 #    include "gromacs/gpu_utils/device_context.h"
  53 #    include "gromacs/gpu_utils/device_stream.h"
  54 #    include "gromacs/gpu_utils/devicebuffer.h"
  55 #    include "gromacs/gpu_utils/hostallocator.h"
  56
  57 #    include "testutils/test_hardware_environment.h"
  58 #    include "testutils/testasserts.h"
  59
  60 namespace gmx
  61 {
  62
  63 template<typename ValueType>
  64 BasicVector<ValueType>& operator++(BasicVector<ValueType>& in)
  65 {
  66     in[XX]++;
  67     in[YY]++;
  68     in[ZZ]++;
  69     return in;
  70 }
  71
  72 template<typename ValueType>
  73 BasicVector<ValueType>& operator++(BasicVector<ValueType>& in, int /* n */)
  74 {
  75     BasicVector<ValueType> temp = *in;
  76     ++*in;
  77     return temp;
  78 }
  79
  80 template<typename ValueType>
  81 inline bool operator==(const BasicVector<ValueType>& lhs, const BasicVector<ValueType>& rhs)
  82 {
  83     return lhs[XX] == rhs[XX] && lhs[YY] == rhs[YY] && lhs[ZZ] == rhs[ZZ];
  84 }
  85
  86 namespace test
  87 {
  88
  89 namespace
  90 {
  91
  92 using testing::Eq;
  93 using testing::Pointwise;
  94
  95 //! Test fixture (needed for typed tests)
  96 template<typename T>
  97 class DeviceBufferTest : public ::testing::Test
  98 {
  99 };
 100
 101 using TypeParamList = testing::Types<short, int, float, double, gmx::RVec>;
 102 TYPED_TEST_CASE(DeviceBufferTest, TypeParamList);
 103
 104 TYPED_TEST(DeviceBufferTest, CanAllocateAndFreeDeviceBuffer)
 105 {
 106     for (const auto& testDevice : getTestHardwareEnvironment()->getTestDeviceList())
 107     {
 108         const DeviceContext& deviceContext = testDevice->deviceContext();
 109         setActiveDevice(testDevice->deviceInfo());
 110
 111         DeviceBuffer<TypeParam> buffer;
 112         int                     numValues = 123;
 113         allocateDeviceBuffer(&buffer, numValues, deviceContext);
 114         freeDeviceBuffer(&buffer);
 115     }
 116 }
 117
 118 TYPED_TEST(DeviceBufferTest, CanReallocateAndFreeDeviceBuffer)
 119 {
 120     for (const auto& testDevice : getTestHardwareEnvironment()->getTestDeviceList())
 121     {
 122         const DeviceContext& deviceContext = testDevice->deviceContext();
 123         setActiveDevice(testDevice->deviceInfo());
 124
 125         DeviceBuffer<TypeParam> buffer;
 126         int                     currentNumValues    = 456;
 127         int                     newNumValues        = 789;
 128         int                     currentMaxNumValues = 0;
 129         allocateDeviceBuffer(&buffer, currentNumValues, deviceContext);
 130         reallocateDeviceBuffer(&buffer, newNumValues, &currentNumValues, &currentMaxNumValues, deviceContext);
 131         freeDeviceBuffer(&buffer);
 132     }
 133 }
 134
 135 //! Initial value to fill the buffer of the scalar type
 136 template<typename T>
 137 const T c_initialValue = static_cast<T>(1);
 138
 139 //! Initial value to fill the buffer of the vector type
 140 template<>
 141 const gmx::RVec c_initialValue<gmx::RVec> = { 1, -2, 3 };
 142
 143
 144 TYPED_TEST(DeviceBufferTest, CanCopyToAndFromDevice)
 145 {
 146     for (auto transferKind : { GpuApiCallBehavior::Sync, GpuApiCallBehavior::Async })
 147     {
 148         PinningPolicy pinningPolicy = (transferKind == GpuApiCallBehavior::Async)
 149                                               ? PinningPolicy::PinnedIfSupported
 150                                               : PinningPolicy::CannotBePinned;
 151         for (const auto& testDevice : getTestHardwareEnvironment()->getTestDeviceList())
 152         {
 153             const DeviceContext& deviceContext = testDevice->deviceContext();
 154             const DeviceStream&  deviceStream  = testDevice->deviceStream();
 155             setActiveDevice(testDevice->deviceInfo());
 156
 157             DeviceBuffer<TypeParam> buffer;
 158             int                     numValues = 123;
 159             allocateDeviceBuffer(&buffer, numValues, deviceContext);
 160             HostVector<TypeParam> valuesIn(numValues, { pinningPolicy });
 161             HostVector<TypeParam> valuesOut(numValues, { pinningPolicy });
 162
 163             std::iota(valuesIn.begin(), valuesIn.end(), c_initialValue<TypeParam>);
 164
 165             copyToDeviceBuffer(&buffer, valuesIn.data(), 0, numValues, deviceStream, transferKind, nullptr);
 166             copyFromDeviceBuffer(
 167                     valuesOut.data(), &buffer, 0, numValues, deviceStream, transferKind, nullptr);
 168             if (transferKind == GpuApiCallBehavior::Async)
 169             {
 170                 deviceStream.synchronize();
 171             }
 172             EXPECT_THAT(valuesOut, Pointwise(Eq(), valuesIn))
 173                     << "Changed after H2D and D2H " << enumValueToString(transferKind) << " copy.";
 174             freeDeviceBuffer(&buffer);
 175         }
 176     }
 177 }
 178
 179 TYPED_TEST(DeviceBufferTest, CanCopyToAndFromDeviceWithOffset)
 180 {
 181     for (auto transferKind : { GpuApiCallBehavior::Sync, GpuApiCallBehavior::Async })
 182     {
 183         PinningPolicy pinningPolicy = (transferKind == GpuApiCallBehavior::Async)
 184                                               ? PinningPolicy::PinnedIfSupported
 185                                               : PinningPolicy::CannotBePinned;
 186         for (const auto& testDevice : getTestHardwareEnvironment()->getTestDeviceList())
 187         {
 188             const DeviceContext& deviceContext = testDevice->deviceContext();
 189             const DeviceStream&  deviceStream  = testDevice->deviceStream();
 190             setActiveDevice(testDevice->deviceInfo());
 191
 192             DeviceBuffer<TypeParam> buffer;
 193             int                     numValues = 123;
 194             allocateDeviceBuffer(&buffer, 2 * numValues, deviceContext);
 195             HostVector<TypeParam> valuesIn(numValues, { pinningPolicy });
 196             HostVector<TypeParam> valuesOut(2 * numValues, { pinningPolicy });
 197
 198             std::iota(valuesIn.begin(), valuesIn.end(), c_initialValue<TypeParam>);
 199
 200             // Fill the buffer with two copies of valuesIn, one after the other.
 201             copyToDeviceBuffer(&buffer, valuesIn.data(), 0, numValues, deviceStream, transferKind, nullptr);
 202             copyToDeviceBuffer(
 203                     &buffer, valuesIn.data(), numValues, numValues, deviceStream, transferKind, nullptr);
 204             // Wait until GPU is done andd o the same copying on the CPU, so we can test it works correctly.
 205             if (transferKind == GpuApiCallBehavior::Async)
 206             {
 207                 deviceStream.synchronize();
 208             }
 209             valuesIn.insert(valuesIn.end(), valuesIn.begin(), valuesIn.end());
 210
 211             copyFromDeviceBuffer(
 212                     valuesOut.data(), &buffer, 0, 2 * numValues, deviceStream, transferKind, nullptr);
 213             if (transferKind == GpuApiCallBehavior::Async)
 214             {
 215                 deviceStream.synchronize();
 216             }
 217             EXPECT_THAT(valuesOut, Pointwise(Eq(), valuesIn))
 218                     << "Changed after H2D and D2H " << enumValueToString(transferKind) << " copy.";
 219
 220             SCOPED_TRACE("Checking the copy respects the output range");
 221
 222             // Remove the first element, and push another copy of the last
 223             // element, so we can check that a copy of all of the data
 224             // skipping the first element correctly over-writes exactly
 225             // all but one of the old values.
 226             valuesIn.erase(valuesIn.begin());
 227             valuesIn.push_back(valuesIn.back());
 228             copyFromDeviceBuffer(
 229                     valuesOut.data(), &buffer, 1, 2 * numValues - 1, deviceStream, transferKind, nullptr);
 230             if (transferKind == GpuApiCallBehavior::Async)
 231             {
 232                 deviceStream.synchronize();
 233             }
 234             EXPECT_THAT(valuesOut, Pointwise(Eq(), valuesIn))
 235                     << "Changed after H2D and D2H " << enumValueToString(transferKind) << " copy.";
 236         }
 237     }
 238 }
 239
 240 #    if GMX_GPU_CUDA
 241
 242 TYPED_TEST(DeviceBufferTest, CanCopyBetweenDeviceBuffers)
 243 {
 244     for (auto transferKind : { GpuApiCallBehavior::Sync, GpuApiCallBehavior::Async })
 245     {
 246         PinningPolicy pinningPolicy = (transferKind == GpuApiCallBehavior::Async)
 247                                               ? PinningPolicy::PinnedIfSupported
 248                                               : PinningPolicy::CannotBePinned;
 249         for (const auto& testDeviceIn : getTestHardwareEnvironment()->getTestDeviceList())
 250         {
 251             for (const auto& testDeviceOut : getTestHardwareEnvironment()->getTestDeviceList())
 252             {
 253                 int                   numValues = 321;
 254                 HostVector<TypeParam> valuesIn(numValues, { pinningPolicy });
 255                 HostVector<TypeParam> valuesOut(numValues, { pinningPolicy });
 256
 257                 std::iota(valuesIn.begin(), valuesIn.end(), c_initialValue<TypeParam>);
 258
 259                 const DeviceContext& deviceContextIn = testDeviceIn->deviceContext();
 260                 const DeviceStream&  deviceStreamIn  = testDeviceIn->deviceStream();
 261                 setActiveDevice(testDeviceIn->deviceInfo());
 262                 DeviceBuffer<TypeParam> bufferIn;
 263                 allocateDeviceBuffer(&bufferIn, numValues, deviceContextIn);
 264
 265                 const DeviceContext& deviceContextOut = testDeviceOut->deviceContext();
 266                 const DeviceStream&  deviceStreamOut  = testDeviceOut->deviceStream();
 267                 setActiveDevice(testDeviceOut->deviceInfo());
 268                 DeviceBuffer<TypeParam> bufferOut;
 269                 allocateDeviceBuffer(&bufferOut, numValues, deviceContextOut);
 270
 271                 copyToDeviceBuffer(
 272                         &bufferIn, valuesIn.data(), 0, numValues, deviceStreamIn, transferKind, nullptr);
 273                 copyBetweenDeviceBuffers(
 274                         &bufferOut, &bufferIn, numValues, deviceStreamIn, transferKind, nullptr);
 275                 if (transferKind == GpuApiCallBehavior::Async)
 276                 {
 277                     deviceStreamIn.synchronize();
 278                 }
 279                 copyFromDeviceBuffer(
 280                         valuesOut.data(), &bufferOut, 0, numValues, deviceStreamOut, transferKind, nullptr);
 281                 if (transferKind == GpuApiCallBehavior::Async)
 282                 {
 283                     deviceStreamOut.synchronize();
 284                 }
 285                 EXPECT_THAT(valuesOut, Pointwise(Eq(), valuesIn))
 286                         << "Changed after H2D, D2D and D2H " << enumValueToString(transferKind)
 287                         << " copy.";
 288                 freeDeviceBuffer(&bufferIn);
 289                 freeDeviceBuffer(&bufferOut);
 290             }
 291         }
 292     }
 293 }
 294
 295 #    endif // GMX_GPU_CUDA
 296
 297
 298 } // namespace
 299 } // namespace test
 300 } // namespace gmx
 301
 302 #endif // GMX_GPU