2 * This file is part of the GROMACS molecular simulation package.
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.
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.
35 #ifndef GMX_GPU_UTILS_DEVICEBUFFER_SYCL_H
36 #define GMX_GPU_UTILS_DEVICEBUFFER_SYCL_H
38 /*! \libinternal \file
39 * \brief Implements the DeviceBuffer type and routines for SYCL.
40 * Should only be included directly by the main DeviceBuffer file devicebuffer.h.
41 * TODO: the intent is for DeviceBuffer to become a class.
43 * \author Artem Zhmurov <zhmurov@gmail.com>
44 * \author Erik Lindahl <erik.lindahl@gmail.com>
45 * \author Andrey Alekseenko <al42and@gmail.com>
52 #include "gromacs/gpu_utils/device_context.h"
53 #include "gromacs/gpu_utils/device_stream.h"
54 #include "gromacs/gpu_utils/devicebuffer_datatype.h"
55 #include "gromacs/gpu_utils/gmxsycl.h"
56 #include "gromacs/gpu_utils/gpu_utils.h" //only for GpuApiCallBehavior
57 #include "gromacs/gpu_utils/gputraits_sycl.h"
58 #include "gromacs/utility/fatalerror.h"
59 #include "gromacs/utility/gmxassert.h"
60 #include "gromacs/utility/stringutil.h"
64 class DeviceBuffer<T>::ClSyclBufferWrapper : public cl::sycl::buffer<T, 1>
66 using cl::sycl::buffer<T, 1>::buffer; // Get all the constructors
70 using ClSyclBufferWrapper = typename DeviceBuffer<T>::ClSyclBufferWrapper;
74 DeviceBuffer<T>::DeviceBuffer() : buffer_(nullptr)
80 DeviceBuffer<T>::~DeviceBuffer() = default;
82 //! Copy constructor (references the same underlying SYCL buffer).
84 DeviceBuffer<T>::DeviceBuffer(DeviceBuffer<T> const& src)
88 buffer_ = std::make_unique<ClSyclBufferWrapper>(*src.buffer_);
98 DeviceBuffer<T>::DeviceBuffer(DeviceBuffer<T>&& src) noexcept = default;
100 //! Copy assignment (references the same underlying SYCL buffer).
102 DeviceBuffer<T>& DeviceBuffer<T>::operator=(DeviceBuffer<T> const& src)
106 buffer_ = std::make_unique<ClSyclBufferWrapper>(*src.buffer_);
110 buffer_.reset(nullptr);
117 DeviceBuffer<T>& DeviceBuffer<T>::operator=(DeviceBuffer<T>&& src) noexcept = default;
119 /*! \brief Dummy assignment operator to allow compilation of some cross-platform code.
121 * A hacky way to make SYCL implementation of DeviceBuffer compatible with details of CUDA and
122 * OpenCL implementations.
124 * \todo Should be removed after DeviceBuffer refactoring.
126 * \tparam T Type of buffer content.
127 * \param nullPtr \c std::nullptr. Not possible to assign any other pointers.
130 DeviceBuffer<T>& DeviceBuffer<T>::operator=(std::nullptr_t nullPtr)
132 buffer_.reset(nullPtr);
137 namespace gmx::internal
139 //! Shorthand alias to create a placeholder SYCL accessor with chosen data type and access mode.
140 template<class T, cl::sycl::access::mode mode>
141 using PlaceholderAccessor =
142 cl::sycl::accessor<T, 1, mode, cl::sycl::access::target::global_buffer, cl::sycl::access::placeholder::true_t>;
143 } // namespace gmx::internal
146 * Thin wrapper around placeholder accessor that allows implicit construction from \c DeviceBuffer.
148 * "Placeholder accessor" is an indicator of the intent to create an accessor for certain buffer
149 * of a certain type, that is not yet bound to a specific command group handler (device). Such
150 * accessors can be created outside SYCL kernels, which is helpful if we want to pass them as
151 * function arguments.
153 * \tparam T Type of buffer content.
154 * \tparam mode Access mode.
156 template<class T, cl::sycl::access::mode mode>
157 class DeviceAccessor : public gmx::internal::PlaceholderAccessor<T, mode>
160 // Inherit all the constructors
161 using gmx::internal::PlaceholderAccessor<T, mode>::PlaceholderAccessor;
162 //! Construct Accessor from DeviceBuffer (must be initialized)
163 DeviceAccessor(DeviceBuffer<T>& buffer) :
164 gmx::internal::PlaceholderAccessor<T, mode>(getSyclBuffer(buffer))
167 //! Construct read-only Accessor from a const DeviceBuffer (must be initialized)
168 DeviceAccessor(const DeviceBuffer<T>& buffer) :
169 gmx::internal::PlaceholderAccessor<T, mode>(getSyclBuffer(const_cast<DeviceBuffer<T>&>(buffer)))
171 /* There were some discussions about making it possible to create read-only sycl::accessor
172 * from a const sycl::buffer (https://github.com/KhronosGroup/SYCL-Docs/issues/10), but
173 * it did not make it into the SYCL2020 standard. So, we have to use const_cast above */
174 /* Using static_assert to ensure that only mode::read accessors can be created from a
175 * const DeviceBuffer. static_assert provides better error messages than std::enable_if. */
176 static_assert(mode == cl::sycl::access::mode::read,
177 "Can not create non-read-only accessor from a const DeviceBuffer");
179 void bind(cl::sycl::handler& cgh) { cgh.require(*this); }
182 //! Helper function to get sycl:buffer object from DeviceBuffer wrapper, with a sanity check.
183 static inline cl::sycl::buffer<T, 1>& getSyclBuffer(DeviceBuffer<T>& buffer)
185 GMX_ASSERT(bool(buffer), "Trying to construct accessor from an uninitialized buffer");
186 return *buffer.buffer_;
190 namespace gmx::internal
192 //! A non-functional class that can be used instead of real accessors
196 NullAccessor(const DeviceBuffer<T>& /*buffer*/) {}
197 //! Allow casting to nullptr
198 constexpr operator std::nullptr_t() const { return nullptr; }
199 //! Placeholder implementation of \c cl::sycl::accessor::get_pointer.
200 T* get_pointer() const noexcept { return nullptr; }
201 void bind(cl::sycl::handler& /*cgh*/) { assert(false); }
203 } // namespace gmx::internal
206 * Helper class to be used as function argument. Will either correspond to a device accessor, or an empty class.
210 template <bool doFoo>
211 void getBarKernel(handler& cgh, OptionalAccessor<float, mode::read, doFoo> a_fooPrms)
214 cgh.require(a_fooPrms);
215 // Can only use a_fooPrms if doFoo == true
218 template <bool doFoo>
219 void callBar(DeviceBuffer<float> b_fooPrms)
221 // If doFoo is false, b_fooPrms will be ignored (can be not initialized).
222 // Otherwise, an accessor will be built (b_fooPrms must be a valid buffer).
223 auto kernel = getBarKernel<doFoo>(b_fooPrms);
224 // If the accessor in not enabled, anything can be passed as its ctor argument.
225 auto kernel2 = getBarKernel<false>(nullptr_t);
229 * \tparam T Data type of the underlying buffer
230 * \tparam mode Access mode of the accessor
231 * \tparam enabled Compile-time flag indicating whether we want to actually create an accessor.
233 template<class T, cl::sycl::access::mode mode, bool enabled>
234 using OptionalAccessor =
235 std::conditional_t<enabled, DeviceAccessor<T, mode>, gmx::internal::NullAccessor<T>>;
237 #endif // #ifndef DOXYGEN
239 /*! \brief Check the validity of the device buffer.
241 * Checks if the buffer is valid and if its allocation is big enough.
243 * \param[in] buffer Device buffer to be checked.
244 * \param[in] requiredSize Number of elements that the buffer will have to accommodate.
246 * \returns Whether the device buffer exists and has enough capacity.
249 static gmx_unused bool checkDeviceBuffer(const DeviceBuffer<T>& buffer, int requiredSize)
251 return buffer.buffer_ && (static_cast<int>(buffer.buffer_->get_count()) >= requiredSize);
254 /*! \libinternal \brief
255 * Allocates a device-side buffer.
256 * It is currently a caller's responsibility to call it only on not-yet allocated buffers.
258 * \tparam ValueType Raw value type of the \p buffer.
259 * \param[in,out] buffer Pointer to the device-side buffer.
260 * \param[in] numValues Number of values to accommodate.
261 * \param[in] deviceContext The buffer's device context-to-be.
263 template<typename ValueType>
264 void allocateDeviceBuffer(DeviceBuffer<ValueType>* buffer, size_t numValues, const DeviceContext& deviceContext)
266 /* SYCL does not require binding buffer to a specific context or device. The ::context_bound
267 * property only enforces the use of only given context, and possibly offers some optimizations */
268 const cl::sycl::property_list bufferProperties{ cl::sycl::property::buffer::context_bound(
269 deviceContext.context()) };
270 buffer->buffer_.reset(
271 new ClSyclBufferWrapper<ValueType>(cl::sycl::range<1>(numValues), bufferProperties));
275 * Frees a device-side buffer.
276 * This does not reset separately stored size/capacity integers,
277 * as this is planned to be a destructor of DeviceBuffer as a proper class,
278 * and no calls on \p buffer should be made afterwards.
280 * \param[in] buffer Pointer to the buffer to free.
282 template<typename ValueType>
283 void freeDeviceBuffer(DeviceBuffer<ValueType>* buffer)
285 buffer->buffer_.reset(nullptr);
289 * Performs the host-to-device data copy, synchronous or asynchronously on request.
291 * Unlike in CUDA and OpenCL, synchronous call does not guarantee that all previously
292 * submitted operations are complete, only the ones that are required for \p buffer consistency.
294 * \tparam ValueType Raw value type of the \p buffer.
295 * \param[in,out] buffer Pointer to the device-side buffer.
296 * \param[in] hostBuffer Pointer to the raw host-side memory, also typed \p ValueType.
297 * \param[in] startingOffset Offset (in values) at the device-side buffer to copy into.
298 * \param[in] numValues Number of values to copy.
299 * \param[in] deviceStream GPU stream to perform asynchronous copy in.
300 * \param[in] transferKind Copy type: synchronous or asynchronous.
301 * \param[out] timingEvent A pointer to the H2D copy timing event to be filled in.
304 template<typename ValueType>
305 void copyToDeviceBuffer(DeviceBuffer<ValueType>* buffer,
306 const ValueType* hostBuffer,
307 size_t startingOffset,
309 const DeviceStream& deviceStream,
310 GpuApiCallBehavior transferKind,
311 CommandEvent* gmx_unused timingEvent)
315 return; // such calls are actually made with empty domains
317 GMX_ASSERT(buffer, "needs a buffer pointer");
318 GMX_ASSERT(hostBuffer, "needs a host buffer pointer");
320 GMX_ASSERT(checkDeviceBuffer(*buffer, startingOffset + numValues),
321 "buffer too small or not initialized");
323 cl::sycl::buffer<ValueType>& syclBuffer = *buffer->buffer_;
325 cl::sycl::event ev = deviceStream.stream().submit([&](cl::sycl::handler& cgh) {
326 /* Here and elsewhere in this file, accessor constructor is user instead of a more common
327 * buffer::get_access, since the compiler (icpx 2021.1-beta09) occasionally gets confused
328 * by all the overloads */
329 auto d_bufferAccessor = cl::sycl::accessor<ValueType, 1, cl::sycl::access::mode::discard_write>{
330 syclBuffer, cgh, cl::sycl::range(numValues), cl::sycl::id(startingOffset)
332 cgh.copy(hostBuffer, d_bufferAccessor);
334 if (transferKind == GpuApiCallBehavior::Sync)
341 * Performs the device-to-host data copy, synchronous or asynchronously on request.
343 * Unlike in CUDA and OpenCL, synchronous call does not guarantee that all previously
344 * submitted operations are complete, only the ones that are required for \p buffer consistency.
346 * \tparam ValueType Raw value type of the \p buffer.
347 * \param[in,out] hostBuffer Pointer to the raw host-side memory, also typed \p ValueType
348 * \param[in] buffer Pointer to the device-side buffer.
349 * \param[in] startingOffset Offset (in values) at the device-side buffer to copy from.
350 * \param[in] numValues Number of values to copy.
351 * \param[in] deviceStream GPU stream to perform asynchronous copy in.
352 * \param[in] transferKind Copy type: synchronous or asynchronous.
353 * \param[out] timingEvent A pointer to the H2D copy timing event to be filled in.
356 template<typename ValueType>
357 void copyFromDeviceBuffer(ValueType* hostBuffer,
358 DeviceBuffer<ValueType>* buffer,
359 size_t startingOffset,
361 const DeviceStream& deviceStream,
362 GpuApiCallBehavior transferKind,
363 CommandEvent* gmx_unused timingEvent)
367 return; // such calls are actually made with empty domains
369 GMX_ASSERT(buffer, "needs a buffer pointer");
370 GMX_ASSERT(hostBuffer, "needs a host buffer pointer");
372 GMX_ASSERT(checkDeviceBuffer(*buffer, startingOffset + numValues),
373 "buffer too small or not initialized");
375 cl::sycl::buffer<ValueType>& syclBuffer = *buffer->buffer_;
377 cl::sycl::event ev = deviceStream.stream().submit([&](cl::sycl::handler& cgh) {
378 const auto d_bufferAccessor = cl::sycl::accessor<ValueType, 1, cl::sycl::access::mode::read>{
379 syclBuffer, cgh, cl::sycl::range(numValues), cl::sycl::id(startingOffset)
381 cgh.copy(d_bufferAccessor, hostBuffer);
383 if (transferKind == GpuApiCallBehavior::Sync)
390 * Performs the device-to-device data copy, synchronous or asynchronously on request.
392 * \tparam ValueType Raw value type of the \p buffer.
394 template<typename ValueType>
395 void copyBetweenDeviceBuffers(DeviceBuffer<ValueType>* /* destinationDeviceBuffer */,
396 DeviceBuffer<ValueType>* /* sourceDeviceBuffer */,
397 size_t /* numValues */,
398 const DeviceStream& /* deviceStream */,
399 GpuApiCallBehavior /* transferKind */,
400 CommandEvent* /*timingEvent*/)
403 gmx_fatal(FARGS, "D2D copy stub was called. Not yet implemented in SYCL.");
407 namespace gmx::internal
409 /*! \brief Helper function to clear device buffer.
411 * Not applicable to GROMACS's Float3 (a.k.a. gmx::RVec) and other custom types.
412 * From SYCL specs: "T must be a scalar value or a SYCL vector type."
414 template<typename ValueType>
415 cl::sycl::event fillSyclBufferWithNull(cl::sycl::buffer<ValueType, 1>& buffer,
416 size_t startingOffset,
418 cl::sycl::queue queue)
420 using cl::sycl::access::mode;
421 const cl::sycl::range<1> range(numValues);
422 const cl::sycl::id<1> offset(startingOffset);
423 const ValueType pattern = ValueType(0); // SYCL vectors support initialization by scalar
425 return queue.submit([&](cl::sycl::handler& cgh) {
426 auto d_bufferAccessor =
427 cl::sycl::accessor<ValueType, 1, mode::discard_write>{ buffer, cgh, range, offset };
428 cgh.fill(d_bufferAccessor, pattern);
432 //! \brief Helper function to clear device buffer of type Float3.
434 inline cl::sycl::event fillSyclBufferWithNull(cl::sycl::buffer<Float3, 1>& buffer,
435 size_t startingOffset,
437 cl::sycl::queue queue)
439 constexpr bool usingHipSycl =
446 if constexpr (usingHipSycl)
448 // hipSYCL does not support reinterpret but allows using Float3 directly.
449 using cl::sycl::access::mode;
450 const cl::sycl::range<1> range(numValues);
451 const cl::sycl::id<1> offset(startingOffset);
452 const Float3 pattern{ 0, 0, 0 };
454 return queue.submit([&](cl::sycl::handler& cgh) {
455 auto d_bufferAccessor =
456 cl::sycl::accessor<Float3, 1, mode::discard_write>{ buffer, cgh, range, offset };
457 cgh.fill(d_bufferAccessor, pattern);
460 else // When not using hipSYCL, reinterpret as a flat float array
463 cl::sycl::buffer<float, 1> bufferAsFloat = buffer.reinterpret<float, 1>(buffer.get_count() * DIM);
464 return fillSyclBufferWithNull<float>(
465 bufferAsFloat, startingOffset * DIM, numValues * DIM, std::move(queue));
470 } // namespace gmx::internal
473 * Clears the device buffer asynchronously.
475 * \tparam ValueType Raw value type of the \p buffer.
476 * \param[in,out] buffer Pointer to the device-side buffer.
477 * \param[in] startingOffset Offset (in values) at the device-side buffer to start clearing at.
478 * \param[in] numValues Number of values to clear.
479 * \param[in] deviceStream GPU stream.
481 template<typename ValueType>
482 void clearDeviceBufferAsync(DeviceBuffer<ValueType>* buffer,
483 size_t startingOffset,
485 const DeviceStream& deviceStream)
491 GMX_ASSERT(buffer, "needs a buffer pointer");
493 GMX_ASSERT(checkDeviceBuffer(*buffer, startingOffset + numValues),
494 "buffer too small or not initialized");
496 cl::sycl::buffer<ValueType>& syclBuffer = *(buffer->buffer_);
498 gmx::internal::fillSyclBufferWithNull<ValueType>(
499 syclBuffer, startingOffset, numValues, deviceStream.stream());
502 /*! \brief Create a texture object for an array of type ValueType.
504 * Creates the device buffer and copies read-only data for an array of type ValueType.
505 * Like OpenCL, does not really do anything with textures, simply creates a buffer
506 * and initializes it.
508 * \tparam ValueType Raw data type.
510 * \param[out] deviceBuffer Device buffer to store data in.
511 * \param[in] hostBuffer Host buffer to get date from.
512 * \param[in] numValues Number of elements in the buffer.
513 * \param[in] deviceContext GPU device context.
515 template<typename ValueType>
516 void initParamLookupTable(DeviceBuffer<ValueType>* deviceBuffer,
517 DeviceTexture* /* deviceTexture */,
518 const ValueType* hostBuffer,
520 const DeviceContext& deviceContext)
522 GMX_ASSERT(hostBuffer, "Host buffer should be specified.");
523 GMX_ASSERT(deviceBuffer, "Device buffer should be specified.");
525 /* Constructing buffer with cl::sycl::buffer(T* data, size_t size) will take ownership
526 * of this memory region making it unusable, which might lead to side-effects.
527 * On the other hand, cl::sycl::buffer(InputIterator<T> begin, InputIterator<T> end) will
528 * initialize the buffer without affecting ownership of the memory, although
529 * it will consume extra memory on host. */
530 const cl::sycl::property_list bufferProperties{ cl::sycl::property::buffer::context_bound(
531 deviceContext.context()) };
532 deviceBuffer->buffer_.reset(new ClSyclBufferWrapper<ValueType>(
533 hostBuffer, hostBuffer + numValues, bufferProperties));
536 /*! \brief Release the OpenCL device buffer.
538 * \tparam ValueType Raw data type.
540 * \param[in,out] deviceBuffer Device buffer to store data in.
542 template<typename ValueType>
543 void destroyParamLookupTable(DeviceBuffer<ValueType>* deviceBuffer, DeviceTexture* /* deviceTexture */)
545 freeDeviceBuffer(deviceBuffer);
548 #endif // GMX_GPU_UTILS_DEVICEBUFFER_SYCL_H