/*
* This file is part of the GROMACS molecular simulation package.
*
- * Copyright (c) 2020, by the GROMACS development team, led by
+ * Copyright (c) 2020,2021, by the GROMACS development team, led by
* Mark Abraham, David van der Spoel, Berk Hess, and Erik Lindahl,
* and including many others, as listed in the AUTHORS file in the
* top-level source directory and at http://www.gromacs.org.
* TODO: the intent is for DeviceBuffer to become a class.
*
* \author Artem Zhmurov <zhmurov@gmail.com>
+ * \author Erik Lindahl <erik.lindahl@gmail.com>
+ * \author Andrey Alekseenko <al42and@gmail.com>
*
* \inlibraryapi
*/
+#include <utility>
+
#include "gromacs/gpu_utils/device_context.h"
#include "gromacs/gpu_utils/device_stream.h"
#include "gromacs/gpu_utils/devicebuffer_datatype.h"
+#include "gromacs/gpu_utils/gmxsycl.h"
#include "gromacs/gpu_utils/gpu_utils.h" //only for GpuApiCallBehavior
-#include "gromacs/gpu_utils/gputraits_ocl.h"
+#include "gromacs/gpu_utils/gputraits_sycl.h"
+#include "gromacs/utility/fatalerror.h"
#include "gromacs/utility/gmxassert.h"
#include "gromacs/utility/stringutil.h"
+#ifndef DOXYGEN
+template<typename T>
+class DeviceBuffer<T>::ClSyclBufferWrapper : public cl::sycl::buffer<T, 1>
+{
+ using cl::sycl::buffer<T, 1>::buffer; // Get all the constructors
+};
+
+template<typename T>
+using ClSyclBufferWrapper = typename DeviceBuffer<T>::ClSyclBufferWrapper;
+
+//! Constructor.
+template<typename T>
+DeviceBuffer<T>::DeviceBuffer() : buffer_(nullptr)
+{
+}
+
+//! Destructor.
+template<typename T>
+DeviceBuffer<T>::~DeviceBuffer() = default;
+
+//! Copy constructor (references the same underlying SYCL buffer).
+template<typename T>
+DeviceBuffer<T>::DeviceBuffer(DeviceBuffer<T> const& src)
+{
+ if (src.buffer_)
+ {
+ buffer_ = std::make_unique<ClSyclBufferWrapper>(*src.buffer_);
+ }
+ else
+ {
+ buffer_ = nullptr;
+ }
+}
+
+//! Move constructor.
+template<typename T>
+DeviceBuffer<T>::DeviceBuffer(DeviceBuffer<T>&& src) noexcept = default;
+
+//! Copy assignment (references the same underlying SYCL buffer).
+template<typename T>
+DeviceBuffer<T>& DeviceBuffer<T>::operator=(DeviceBuffer<T> const& src)
+{
+ if (src.buffer_)
+ {
+ buffer_ = std::make_unique<ClSyclBufferWrapper>(*src.buffer_);
+ }
+ else
+ {
+ buffer_.reset(nullptr);
+ }
+ return *this;
+}
+
+//! Move assignment.
+template<typename T>
+DeviceBuffer<T>& DeviceBuffer<T>::operator=(DeviceBuffer<T>&& src) noexcept = default;
+
+/*! \brief Dummy assignment operator to allow compilation of some cross-platform code.
+ *
+ * A hacky way to make SYCL implementation of DeviceBuffer compatible with details of CUDA and
+ * OpenCL implementations.
+ *
+ * \todo Should be removed after DeviceBuffer refactoring.
+ *
+ * \tparam T Type of buffer content.
+ * \param nullPtr \c std::nullptr. Not possible to assign any other pointers.
+ */
+template<typename T>
+DeviceBuffer<T>& DeviceBuffer<T>::operator=(std::nullptr_t nullPtr)
+{
+ buffer_.reset(nullPtr);
+ return *this;
+}
+
+
+namespace gmx::internal
+{
+//! Shorthand alias to create a placeholder SYCL accessor with chosen data type and access mode.
+template<class T, cl::sycl::access::mode mode>
+using PlaceholderAccessor =
+ cl::sycl::accessor<T, 1, mode, cl::sycl::access::target::global_buffer, cl::sycl::access::placeholder::true_t>;
+} // namespace gmx::internal
+
+/** \brief
+ * Thin wrapper around placeholder accessor that allows implicit construction from \c DeviceBuffer.
+ *
+ * "Placeholder accessor" is an indicator of the intent to create an accessor for certain buffer
+ * of a certain type, that is not yet bound to a specific command group handler (device). Such
+ * accessors can be created outside SYCL kernels, which is helpful if we want to pass them as
+ * function arguments.
+ *
+ * \tparam T Type of buffer content.
+ * \tparam mode Access mode.
+ */
+template<class T, cl::sycl::access::mode mode>
+class DeviceAccessor : public gmx::internal::PlaceholderAccessor<T, mode>
+{
+public:
+ // Inherit all the constructors
+ using gmx::internal::PlaceholderAccessor<T, mode>::PlaceholderAccessor;
+ //! Construct Accessor from DeviceBuffer (must be initialized)
+ DeviceAccessor(DeviceBuffer<T>& buffer) :
+ gmx::internal::PlaceholderAccessor<T, mode>(getSyclBuffer(buffer))
+ {
+ }
+ //! Construct read-only Accessor from a const DeviceBuffer (must be initialized)
+ DeviceAccessor(const DeviceBuffer<T>& buffer) :
+ gmx::internal::PlaceholderAccessor<T, mode>(getSyclBuffer(const_cast<DeviceBuffer<T>&>(buffer)))
+ {
+ /* There were some discussions about making it possible to create read-only sycl::accessor
+ * from a const sycl::buffer (https://github.com/KhronosGroup/SYCL-Docs/issues/10), but
+ * it did not make it into the SYCL2020 standard. So, we have to use const_cast above */
+ /* Using static_assert to ensure that only mode::read accessors can be created from a
+ * const DeviceBuffer. static_assert provides better error messages than std::enable_if. */
+ static_assert(mode == cl::sycl::access::mode::read,
+ "Can not create non-read-only accessor from a const DeviceBuffer");
+ }
+ void bind(cl::sycl::handler& cgh) { cgh.require(*this); }
+
+private:
+ //! Helper function to get sycl:buffer object from DeviceBuffer wrapper, with a sanity check.
+ static inline cl::sycl::buffer<T, 1>& getSyclBuffer(DeviceBuffer<T>& buffer)
+ {
+ GMX_ASSERT(bool(buffer), "Trying to construct accessor from an uninitialized buffer");
+ return *buffer.buffer_;
+ }
+};
+
+namespace gmx::internal
+{
+//! A non-functional class that can be used instead of real accessors
+template<class T>
+struct NullAccessor
+{
+ NullAccessor(const DeviceBuffer<T>& /*buffer*/) {}
+ //! Allow casting to nullptr
+ constexpr operator std::nullptr_t() const { return nullptr; }
+ //! Placeholder implementation of \c cl::sycl::accessor::get_pointer.
+ T* get_pointer() const noexcept { return nullptr; }
+ void bind(cl::sycl::handler& /*cgh*/) { assert(false); }
+};
+} // namespace gmx::internal
+
+/** \brief
+ * Helper class to be used as function argument. Will either correspond to a device accessor, or an empty class.
+ *
+ * Example usage:
+ * \code
+ template <bool doFoo>
+ void getBarKernel(handler& cgh, OptionalAccessor<float, mode::read, doFoo> a_fooPrms)
+ {
+ if constexpr (doFoo)
+ cgh.require(a_fooPrms);
+ // Can only use a_fooPrms if doFoo == true
+ }
+
+ template <bool doFoo>
+ void callBar(DeviceBuffer<float> b_fooPrms)
+ {
+ // If doFoo is false, b_fooPrms will be ignored (can be not initialized).
+ // Otherwise, an accessor will be built (b_fooPrms must be a valid buffer).
+ auto kernel = getBarKernel<doFoo>(b_fooPrms);
+ // If the accessor in not enabled, anything can be passed as its ctor argument.
+ auto kernel2 = getBarKernel<false>(nullptr_t);
+ }
+ * \endcode
+ *
+ * \tparam T Data type of the underlying buffer
+ * \tparam mode Access mode of the accessor
+ * \tparam enabled Compile-time flag indicating whether we want to actually create an accessor.
+ */
+template<class T, cl::sycl::access::mode mode, bool enabled>
+using OptionalAccessor =
+ std::conditional_t<enabled, DeviceAccessor<T, mode>, gmx::internal::NullAccessor<T>>;
+
+#endif // #ifndef DOXYGEN
+
+/*! \brief Check the validity of the device buffer.
+ *
+ * Checks if the buffer is valid and if its allocation is big enough.
+ *
+ * \param[in] buffer Device buffer to be checked.
+ * \param[in] requiredSize Number of elements that the buffer will have to accommodate.
+ *
+ * \returns Whether the device buffer exists and has enough capacity.
+ */
+template<typename T>
+static gmx_unused bool checkDeviceBuffer(const DeviceBuffer<T>& buffer, int requiredSize)
+{
+ return buffer.buffer_ && (static_cast<int>(buffer.buffer_->get_count()) >= requiredSize);
+}
+
/*! \libinternal \brief
* Allocates a device-side buffer.
* It is currently a caller's responsibility to call it only on not-yet allocated buffers.
*
* \tparam ValueType Raw value type of the \p buffer.
* \param[in,out] buffer Pointer to the device-side buffer.
- * \param[in] numValues Number of values to accomodate.
+ * \param[in] numValues Number of values to accommodate.
* \param[in] deviceContext The buffer's device context-to-be.
*/
template<typename ValueType>
-void allocateDeviceBuffer(DeviceBuffer<ValueType>* gmx_unused buffer,
- size_t gmx_unused numValues,
- const DeviceContext& gmx_unused deviceContext)
+void allocateDeviceBuffer(DeviceBuffer<ValueType>* buffer, size_t numValues, const DeviceContext& deviceContext)
{
- // SYCL-TODO
+ /* SYCL does not require binding buffer to a specific context or device. The ::context_bound
+ * property only enforces the use of only given context, and possibly offers some optimizations */
+ const cl::sycl::property_list bufferProperties{ cl::sycl::property::buffer::context_bound(
+ deviceContext.context()) };
+ buffer->buffer_.reset(
+ new ClSyclBufferWrapper<ValueType>(cl::sycl::range<1>(numValues), bufferProperties));
}
/*! \brief
*
* \param[in] buffer Pointer to the buffer to free.
*/
-template<typename DeviceBuffer>
-void freeDeviceBuffer(DeviceBuffer* gmx_unused buffer)
+template<typename ValueType>
+void freeDeviceBuffer(DeviceBuffer<ValueType>* buffer)
{
- // SYCL-TODO
+ buffer->buffer_.reset(nullptr);
}
/*! \brief
* Performs the host-to-device data copy, synchronous or asynchronously on request.
*
- * Note that synchronous copy will not synchronize the stream in case of zero \p numValues
- * because of the early return.
+ * Unlike in CUDA and OpenCL, synchronous call does not guarantee that all previously
+ * submitted operations are complete, only the ones that are required for \p buffer consistency.
*
* \tparam ValueType Raw value type of the \p buffer.
- * \param[in,out] buffer Pointer to the device-side buffer
- * \param[in] hostBuffer Pointer to the raw host-side memory, also typed \p ValueType
+ * \param[in,out] buffer Pointer to the device-side buffer.
+ * \param[in] hostBuffer Pointer to the raw host-side memory, also typed \p ValueType.
* \param[in] startingOffset Offset (in values) at the device-side buffer to copy into.
* \param[in] numValues Number of values to copy.
* \param[in] deviceStream GPU stream to perform asynchronous copy in.
* \param[in] transferKind Copy type: synchronous or asynchronous.
* \param[out] timingEvent A pointer to the H2D copy timing event to be filled in.
- * If the pointer is not null, the event can further be used
- * to queue a wait for this operation or to query profiling information.
+ * Ignored in SYCL.
*/
template<typename ValueType>
-void copyToDeviceBuffer(DeviceBuffer<ValueType>* gmx_unused buffer,
- const ValueType* gmx_unused hostBuffer,
- size_t gmx_unused startingOffset,
- size_t gmx_unused numValues,
- const DeviceStream& gmx_unused deviceStream,
- GpuApiCallBehavior gmx_unused transferKind,
+void copyToDeviceBuffer(DeviceBuffer<ValueType>* buffer,
+ const ValueType* hostBuffer,
+ size_t startingOffset,
+ size_t numValues,
+ const DeviceStream& deviceStream,
+ GpuApiCallBehavior transferKind,
CommandEvent* gmx_unused timingEvent)
{
- // SYCL-TODO
+ if (numValues == 0)
+ {
+ return; // such calls are actually made with empty domains
+ }
+ GMX_ASSERT(buffer, "needs a buffer pointer");
+ GMX_ASSERT(hostBuffer, "needs a host buffer pointer");
+
+ GMX_ASSERT(checkDeviceBuffer(*buffer, startingOffset + numValues),
+ "buffer too small or not initialized");
+
+ cl::sycl::buffer<ValueType>& syclBuffer = *buffer->buffer_;
+
+ cl::sycl::event ev = deviceStream.stream().submit([&](cl::sycl::handler& cgh) {
+ /* Here and elsewhere in this file, accessor constructor is user instead of a more common
+ * buffer::get_access, since the compiler (icpx 2021.1-beta09) occasionally gets confused
+ * by all the overloads */
+ auto d_bufferAccessor = cl::sycl::accessor<ValueType, 1, cl::sycl::access::mode::discard_write>{
+ syclBuffer, cgh, cl::sycl::range(numValues), cl::sycl::id(startingOffset)
+ };
+ cgh.copy(hostBuffer, d_bufferAccessor);
+ });
+ if (transferKind == GpuApiCallBehavior::Sync)
+ {
+ ev.wait_and_throw();
+ }
}
/*! \brief
* Performs the device-to-host data copy, synchronous or asynchronously on request.
*
- * Note that synchronous copy will not synchronize the stream in case of zero \p numValues
- * because of the early return.
+ * Unlike in CUDA and OpenCL, synchronous call does not guarantee that all previously
+ * submitted operations are complete, only the ones that are required for \p buffer consistency.
*
* \tparam ValueType Raw value type of the \p buffer.
* \param[in,out] hostBuffer Pointer to the raw host-side memory, also typed \p ValueType
- * \param[in] buffer Pointer to the device-side buffer
+ * \param[in] buffer Pointer to the device-side buffer.
* \param[in] startingOffset Offset (in values) at the device-side buffer to copy from.
* \param[in] numValues Number of values to copy.
* \param[in] deviceStream GPU stream to perform asynchronous copy in.
* \param[in] transferKind Copy type: synchronous or asynchronous.
* \param[out] timingEvent A pointer to the H2D copy timing event to be filled in.
- * If the pointer is not null, the event can further be used
- * to queue a wait for this operation or to query profiling information.
+ * Ignored in SYCL.
*/
template<typename ValueType>
-void copyFromDeviceBuffer(ValueType* gmx_unused hostBuffer,
- DeviceBuffer<ValueType>* gmx_unused buffer,
- size_t gmx_unused startingOffset,
- size_t gmx_unused numValues,
- const DeviceStream& gmx_unused deviceStream,
- GpuApiCallBehavior gmx_unused transferKind,
+void copyFromDeviceBuffer(ValueType* hostBuffer,
+ DeviceBuffer<ValueType>* buffer,
+ size_t startingOffset,
+ size_t numValues,
+ const DeviceStream& deviceStream,
+ GpuApiCallBehavior transferKind,
CommandEvent* gmx_unused timingEvent)
{
- // SYCL-TODO
+ if (numValues == 0)
+ {
+ return; // such calls are actually made with empty domains
+ }
+ GMX_ASSERT(buffer, "needs a buffer pointer");
+ GMX_ASSERT(hostBuffer, "needs a host buffer pointer");
+
+ GMX_ASSERT(checkDeviceBuffer(*buffer, startingOffset + numValues),
+ "buffer too small or not initialized");
+
+ cl::sycl::buffer<ValueType>& syclBuffer = *buffer->buffer_;
+
+ cl::sycl::event ev = deviceStream.stream().submit([&](cl::sycl::handler& cgh) {
+ const auto d_bufferAccessor = cl::sycl::accessor<ValueType, 1, cl::sycl::access::mode::read>{
+ syclBuffer, cgh, cl::sycl::range(numValues), cl::sycl::id(startingOffset)
+ };
+ cgh.copy(d_bufferAccessor, hostBuffer);
+ });
+ if (transferKind == GpuApiCallBehavior::Sync)
+ {
+ ev.wait_and_throw();
+ }
}
/*! \brief
- * Clears the device buffer asynchronously.
+ * Performs the device-to-device data copy, synchronous or asynchronously on request.
*
- * \tparam ValueType Raw value type of the \p buffer.
- * \param[in,out] buffer Pointer to the device-side buffer
- * \param[in] startingOffset Offset (in values) at the device-side buffer to start clearing at.
- * \param[in] numValues Number of values to clear.
- * \param[in] deviceStream GPU stream.
+ * \tparam ValueType Raw value type of the \p buffer.
*/
template<typename ValueType>
-void clearDeviceBufferAsync(DeviceBuffer<ValueType>* gmx_unused buffer,
- size_t gmx_unused startingOffset,
- size_t gmx_unused numValues,
- const DeviceStream& gmx_unused deviceStream)
+void copyBetweenDeviceBuffers(DeviceBuffer<ValueType>* /* destinationDeviceBuffer */,
+ DeviceBuffer<ValueType>* /* sourceDeviceBuffer */,
+ size_t /* numValues */,
+ const DeviceStream& /* deviceStream */,
+ GpuApiCallBehavior /* transferKind */,
+ CommandEvent* /*timingEvent*/)
{
// SYCL-TODO
+ gmx_fatal(FARGS, "D2D copy stub was called. Not yet implemented in SYCL.");
}
-/*! \brief Check the validity of the device buffer.
- *
- * Checks if the buffer is not nullptr and if its allocation is big enough.
- *
- * \param[in] buffer Device buffer to be checked.
- * \param[in] requiredSize Number of elements that the buffer will have to accommodate.
+
+namespace gmx::internal
+{
+/*! \brief Helper function to clear device buffer.
*
- * \returns Whether the device buffer can be set.
+ * Not applicable to GROMACS's Float3 (a.k.a. gmx::RVec) and other custom types.
+ * From SYCL specs: "T must be a scalar value or a SYCL vector type."
*/
-template<typename T>
-static bool gmx_unused checkDeviceBuffer(DeviceBuffer<T> gmx_unused buffer, int gmx_unused requiredSize)
+template<typename ValueType>
+cl::sycl::event fillSyclBufferWithNull(cl::sycl::buffer<ValueType, 1>& buffer,
+ size_t startingOffset,
+ size_t numValues,
+ cl::sycl::queue queue)
{
- // SYCL-TODO
+ using cl::sycl::access::mode;
+ const cl::sycl::range<1> range(numValues);
+ const cl::sycl::id<1> offset(startingOffset);
+ const ValueType pattern = ValueType(0); // SYCL vectors support initialization by scalar
+
+ return queue.submit([&](cl::sycl::handler& cgh) {
+ auto d_bufferAccessor =
+ cl::sycl::accessor<ValueType, 1, mode::discard_write>{ buffer, cgh, range, offset };
+ cgh.fill(d_bufferAccessor, pattern);
+ });
+}
+
+//! \brief Helper function to clear device buffer of type Float3.
+template<>
+inline cl::sycl::event fillSyclBufferWithNull(cl::sycl::buffer<Float3, 1>& buffer,
+ size_t startingOffset,
+ size_t numValues,
+ cl::sycl::queue queue)
+{
+ constexpr bool usingHipSycl =
+#ifdef __HIPSYCL__
+ true;
+#else
+ false;
+#endif
+
+ if constexpr (usingHipSycl)
+ {
+ // hipSYCL does not support reinterpret but allows using Float3 directly.
+ using cl::sycl::access::mode;
+ const cl::sycl::range<1> range(numValues);
+ const cl::sycl::id<1> offset(startingOffset);
+ const Float3 pattern{ 0, 0, 0 };
+
+ return queue.submit([&](cl::sycl::handler& cgh) {
+ auto d_bufferAccessor =
+ cl::sycl::accessor<Float3, 1, mode::discard_write>{ buffer, cgh, range, offset };
+ cgh.fill(d_bufferAccessor, pattern);
+ });
+ }
+ else // When not using hipSYCL, reinterpret as a flat float array
+ {
+#ifndef __HIPSYCL__
+ cl::sycl::buffer<float, 1> bufferAsFloat = buffer.reinterpret<float, 1>(buffer.get_count() * DIM);
+ return fillSyclBufferWithNull<float>(
+ bufferAsFloat, startingOffset * DIM, numValues * DIM, std::move(queue));
+#endif
+ }
}
-//! Device texture wrapper.
-using DeviceTexture = void*;
+} // namespace gmx::internal
+
+/*! \brief
+ * Clears the device buffer asynchronously.
+ *
+ * \tparam ValueType Raw value type of the \p buffer.
+ * \param[in,out] buffer Pointer to the device-side buffer.
+ * \param[in] startingOffset Offset (in values) at the device-side buffer to start clearing at.
+ * \param[in] numValues Number of values to clear.
+ * \param[in] deviceStream GPU stream.
+ */
+template<typename ValueType>
+void clearDeviceBufferAsync(DeviceBuffer<ValueType>* buffer,
+ size_t startingOffset,
+ size_t numValues,
+ const DeviceStream& deviceStream)
+{
+ if (numValues == 0)
+ {
+ return;
+ }
+ GMX_ASSERT(buffer, "needs a buffer pointer");
+
+ GMX_ASSERT(checkDeviceBuffer(*buffer, startingOffset + numValues),
+ "buffer too small or not initialized");
+
+ cl::sycl::buffer<ValueType>& syclBuffer = *(buffer->buffer_);
+
+ gmx::internal::fillSyclBufferWithNull<ValueType>(
+ syclBuffer, startingOffset, numValues, deviceStream.stream());
+}
/*! \brief Create a texture object for an array of type ValueType.
*
* Creates the device buffer and copies read-only data for an array of type ValueType.
- *
- * \todo Decide if using image2d is most efficient.
+ * Like OpenCL, does not really do anything with textures, simply creates a buffer
+ * and initializes it.
*
* \tparam ValueType Raw data type.
*
* \param[out] deviceBuffer Device buffer to store data in.
- * \param[out] deviceTexture New texture object
* \param[in] hostBuffer Host buffer to get date from.
* \param[in] numValues Number of elements in the buffer.
* \param[in] deviceContext GPU device context.
*/
template<typename ValueType>
-void initParamLookupTable(DeviceBuffer<ValueType>* gmx_unused deviceBuffer,
- DeviceTexture* gmx_unused deviceTexture,
- const ValueType* gmx_unused hostBuffer,
- int gmx_unused numValues,
- const DeviceContext& gmx_unused deviceContext)
+void initParamLookupTable(DeviceBuffer<ValueType>* deviceBuffer,
+ DeviceTexture* /* deviceTexture */,
+ const ValueType* hostBuffer,
+ int numValues,
+ const DeviceContext& deviceContext)
{
- // SYCL-TODO
+ GMX_ASSERT(hostBuffer, "Host buffer should be specified.");
+ GMX_ASSERT(deviceBuffer, "Device buffer should be specified.");
+
+ /* Constructing buffer with cl::sycl::buffer(T* data, size_t size) will take ownership
+ * of this memory region making it unusable, which might lead to side-effects.
+ * On the other hand, cl::sycl::buffer(InputIterator<T> begin, InputIterator<T> end) will
+ * initialize the buffer without affecting ownership of the memory, although
+ * it will consume extra memory on host. */
+ const cl::sycl::property_list bufferProperties{ cl::sycl::property::buffer::context_bound(
+ deviceContext.context()) };
+ deviceBuffer->buffer_.reset(new ClSyclBufferWrapper<ValueType>(
+ hostBuffer, hostBuffer + numValues, bufferProperties));
}
/*! \brief Release the OpenCL device buffer.
* \tparam ValueType Raw data type.
*
* \param[in,out] deviceBuffer Device buffer to store data in.
- * \param[in] deviceTexture Reference to texture object
*/
template<typename ValueType>
-void destroyParamLookupTable(DeviceBuffer<ValueType>* gmx_unused deviceBuffer,
- DeviceTexture& gmx_unused deviceTexture)
+void destroyParamLookupTable(DeviceBuffer<ValueType>* deviceBuffer, DeviceTexture* /* deviceTexture */)
{
- // SYCL-TODO
+ freeDeviceBuffer(deviceBuffer);
}
#endif // GMX_GPU_UTILS_DEVICEBUFFER_SYCL_H
biod.pnpi.spb.ru / alexxy / gromacs.git / blobdiff