group.hpp

Go to the documentation of this file.

//==-------------- group.hpp --- SYCL work group ---------------------------==//
//
// Part of the LLVM Project, under the Apache License v2.0 with LLVM Exceptions.
// See https://llvm.org/LICENSE.txt for license information.
// SPDX-License-Identifier: Apache-2.0 WITH LLVM-exception
//
//===----------------------------------------------------------------------===//
 
#pragma once
 
#include <CL/__spirv/spirv_ops.hpp>
#include <CL/__spirv/spirv_types.hpp>
#include <CL/__spirv/spirv_vars.hpp>
#include <CL/sycl/detail/common.hpp>
#include <CL/sycl/detail/generic_type_traits.hpp>
#include <CL/sycl/detail/helpers.hpp>
#include <CL/sycl/detail/spirv.hpp>
#include <CL/sycl/device_event.hpp>
#include <CL/sycl/h_item.hpp>
#include <CL/sycl/id.hpp>
#include <CL/sycl/memory_enums.hpp>
#include <CL/sycl/pointers.hpp>
#include <CL/sycl/range.hpp>
#include <stdexcept>
#include <type_traits>
 
__SYCL_INLINE_NAMESPACE(cl) {
namespace sycl {
namespace detail {
class Builder;
 
// Implements a barrier accross work items within a work group.
static inline void workGroupBarrier() {
#ifdef __SYCL_DEVICE_ONLY__
  constexpr uint32_t flags =
      static_cast<uint32_t>(
          __spv::MemorySemanticsMask::SequentiallyConsistent) |
      static_cast<uint32_t>(__spv::MemorySemanticsMask::WorkgroupMemory);
  __spirv_ControlBarrier(__spv::Scope::Workgroup, __spv::Scope::Workgroup,
                         flags);
#endif // __SYCL_DEVICE_ONLY__
}
 
} // namespace detail
 
// SYCL 1.2.1rev5, section "4.8.5.3 Parallel For hierarchical invoke":
// Quote:
//   ... To guarantee use of private per-work-item memory, the private_memory
//   class can be used to wrap the data. This class very simply constructs
//   private data for a given group across the entire group.The id of the
//   current work-item is passed to any access to grab the correct data.
template <typename T, int Dimensions = 1> class private_memory {
public:
  // Construct based directly off the number of work-items
  private_memory(const group<Dimensions> &G) {
#ifndef __SYCL_DEVICE_ONLY__
    // serial host => one instance per work-group - allocate space for each WI
    // in the group:
    Val.reset(new T[G.get_local_range().size()]);
#endif // __SYCL_DEVICE_ONLY__
    (void)G;
  }
 
  // Access the instance for the current work-item
  T &operator()(const h_item<Dimensions> &Id) {
#ifndef __SYCL_DEVICE_ONLY__
    // Calculate the linear index of current WI and return reference to the
    // corresponding spot in the value array:
    size_t Ind = Id.get_physical_local().get_linear_id();
    return Val.get()[Ind];
#else
    (void)Id;
    return Val;
#endif // __SYCL_DEVICE_ONLY__
  }
 
private:
#ifdef __SYCL_DEVICE_ONLY__
  // On SYCL device private_memory<T> instance is created per physical WI, so
  // there is 1:1 correspondence betwen this class instances and per-WI memory.
  T Val;
#else
  // On serial host there is one private_memory<T> instance per work group, so
  // it must have space to hold separate value per WI in the group.
  std::unique_ptr<T[]> Val;
#endif // #ifdef __SYCL_DEVICE_ONLY__
};
 
template <int Dimensions = 1> class group {
public:
#ifndef __DISABLE_SYCL_INTEL_GROUP_ALGORITHMS__
  using id_type = id<Dimensions>;
  using range_type = range<Dimensions>;
  using linear_id_type = size_t;
  static constexpr int dimensions = Dimensions;
#endif // __DISABLE_SYCL_INTEL_GROUP_ALGORITHMS__
 
  static constexpr sycl::memory_scope fence_scope =
      sycl::memory_scope::work_group;
 
  group() = delete;
 
  id<Dimensions> get_id() const { return index; }
 
  size_t get_id(int dimension) const { return index[dimension]; }
 
  range<Dimensions> get_global_range() const { return globalRange; }
 
  size_t get_global_range(int dimension) const {
    return globalRange[dimension];
  }
 
  range<Dimensions> get_local_range() const { return localRange; }
 
  size_t get_local_range(int dimension) const { return localRange[dimension]; }
 
  range<Dimensions> get_group_range() const { return groupRange; }
 
  size_t get_group_range(int dimension) const {
    return get_group_range()[dimension];
  }
 
  size_t operator[](int dimension) const { return index[dimension]; }
 
  template <int dims = Dimensions>
  typename detail::enable_if_t<(dims == 1), size_t> get_linear_id() const {
    return index[0];
  }
 
  template <int dims = Dimensions>
  typename detail::enable_if_t<(dims == 2), size_t> get_linear_id() const {
    return index[0] * groupRange[1] + index[1];
  }
 
  // SYCL specification 1.2.1rev5, section 4.7.6.5 "Buffer accessor":
  //    Whenever a multi-dimensional index is passed to a SYCL accessor the
  //    linear index is calculated based on the index {id1, id2, id3} provided
  //    and the range of the SYCL accessor {r1, r2, r3} according to row-major
  //    ordering as follows:
  //      id3 + (id2 · r3) + (id1 · r3 · r2)            (4.3)
  // section 4.8.1.8 "group class":
  //    size_t get_linear_id()const
  //    Get a linearized version of the work-group id. Calculating a linear
  //    work-group id from a multi-dimensional index follows the equation 4.3.
  template <int dims = Dimensions>
  typename detail::enable_if_t<(dims == 3), size_t> get_linear_id() const {
    return (index[0] * groupRange[1] * groupRange[2]) +
           (index[1] * groupRange[2]) + index[2];
  }
 
  template <typename WorkItemFunctionT>
  void parallel_for_work_item(WorkItemFunctionT Func) const {
    // need barriers to enforce SYCL semantics for the work item loop -
    // compilers are expected to optimize when possible
    detail::workGroupBarrier();
#ifdef __SYCL_DEVICE_ONLY__
    range<Dimensions> GlobalSize{
        __spirv::initGlobalSize<Dimensions, range<Dimensions>>()};
    range<Dimensions> LocalSize{
        __spirv::initWorkgroupSize<Dimensions, range<Dimensions>>()};
    id<Dimensions> GlobalId{
        __spirv::initGlobalInvocationId<Dimensions, id<Dimensions>>()};
    id<Dimensions> LocalId{
        __spirv::initLocalInvocationId<Dimensions, id<Dimensions>>()};
 
    // no 'iterate' in the device code variant, because
    // (1) this code is already invoked by each work item as a part of the
    //     enclosing parallel_for_work_group kernel
    // (2) the range this pfwi iterates over matches work group size exactly
    item<Dimensions, false> GlobalItem =
        detail::Builder::createItem<Dimensions, false>(GlobalSize, GlobalId);
    item<Dimensions, false> LocalItem =
        detail::Builder::createItem<Dimensions, false>(LocalSize, LocalId);
    h_item<Dimensions> HItem =
        detail::Builder::createHItem<Dimensions>(GlobalItem, LocalItem);
 
    Func(HItem);
#else
    id<Dimensions> GroupStartID = index * localRange;
 
    // ... host variant needs explicit 'iterate' because it is serial
    detail::NDLoop<Dimensions>::iterate(
        localRange, [&](const id<Dimensions> &LocalID) {
          item<Dimensions, false> GlobalItem =
              detail::Builder::createItem<Dimensions, false>(
                  globalRange, GroupStartID + LocalID);
          item<Dimensions, false> LocalItem =
              detail::Builder::createItem<Dimensions, false>(localRange,
                                                             LocalID);
          h_item<Dimensions> HItem =
              detail::Builder::createHItem<Dimensions>(GlobalItem, LocalItem);
          Func(HItem);
        });
#endif // __SYCL_DEVICE_ONLY__
    // Need both barriers here - before and after the parallel_for_work_item
    // (PFWI). There can be work group scope code after the PFWI which reads
    // work group local data written within this PFWI. Back Ends are expected to
    // optimize away unneeded barriers (e.g. two barriers in a row).
    detail::workGroupBarrier();
  }
 
  template <typename WorkItemFunctionT>
  void parallel_for_work_item(range<Dimensions> flexibleRange,
                              WorkItemFunctionT Func) const {
    detail::workGroupBarrier();
#ifdef __SYCL_DEVICE_ONLY__
    range<Dimensions> GlobalSize{
        __spirv::initGlobalSize<Dimensions, range<Dimensions>>()};
    range<Dimensions> LocalSize{
        __spirv::initWorkgroupSize<Dimensions, range<Dimensions>>()};
    id<Dimensions> GlobalId{
        __spirv::initGlobalInvocationId<Dimensions, id<Dimensions>>()};
    id<Dimensions> LocalId{
        __spirv::initLocalInvocationId<Dimensions, id<Dimensions>>()};
 
    item<Dimensions, false> GlobalItem =
        detail::Builder::createItem<Dimensions, false>(GlobalSize, GlobalId);
    item<Dimensions, false> LocalItem =
        detail::Builder::createItem<Dimensions, false>(LocalSize, LocalId);
    h_item<Dimensions> HItem = detail::Builder::createHItem<Dimensions>(
        GlobalItem, LocalItem, flexibleRange);
 
    // iterate over flexible range with work group size stride; each item
    // performs flexibleRange/LocalSize iterations (if the former is divisible
    // by the latter)
    detail::NDLoop<Dimensions>::iterate(
        LocalId, LocalSize, flexibleRange,
        [&](const id<Dimensions> &LogicalLocalID) {
          HItem.setLogicalLocalID(LogicalLocalID);
          Func(HItem);
        });
#else
    id<Dimensions> GroupStartID = index * localRange;
 
    detail::NDLoop<Dimensions>::iterate(
        localRange, [&](const id<Dimensions> &LocalID) {
          item<Dimensions, false> GlobalItem =
              detail::Builder::createItem<Dimensions, false>(
                  globalRange, GroupStartID + LocalID);
          item<Dimensions, false> LocalItem =
              detail::Builder::createItem<Dimensions, false>(localRange,
                                                             LocalID);
          h_item<Dimensions> HItem = detail::Builder::createHItem<Dimensions>(
              GlobalItem, LocalItem, flexibleRange);
 
          detail::NDLoop<Dimensions>::iterate(
              LocalID, localRange, flexibleRange,
              [&](const id<Dimensions> &LogicalLocalID) {
                HItem.setLogicalLocalID(LogicalLocalID);
                Func(HItem);
              });
        });
#endif // __SYCL_DEVICE_ONLY__
    detail::workGroupBarrier();
  }
 
  template <access::mode accessMode = access::mode::read_write>
  void mem_fence(
      typename detail::enable_if_t<accessMode == access::mode::read ||
                                       accessMode == access::mode::write ||
                                       accessMode == access::mode::read_write,
                                   access::fence_space>
          accessSpace = access::fence_space::global_and_local) const {
    uint32_t flags = detail::getSPIRVMemorySemanticsMask(accessSpace);
    // TODO: currently, there is no good way in SPIR-V to set the memory
    // barrier only for load operations or only for store operations.
    // The full read-and-write barrier is used and the template parameter
    // 'accessMode' is ignored for now. Either SPIR-V or SYCL spec may be
    // changed to address this discrepancy between SPIR-V and SYCL,
    // or if we decide that 'accessMode' is the important feature then
    // we can fix this later, for example, by using OpenCL 1.2 functions
    // read_mem_fence() and write_mem_fence().
    __spirv_MemoryBarrier(__spv::Scope::Workgroup, flags);
  }
 
  template <typename dataT>
  detail::enable_if_t<!detail::is_bool<dataT>::value, device_event>
  async_work_group_copy(local_ptr<dataT> dest, global_ptr<dataT> src,
                        size_t numElements, size_t srcStride) const {
    using DestT = detail::ConvertToOpenCLType_t<decltype(dest)>;
    using SrcT = detail::ConvertToOpenCLType_t<decltype(src)>;
 
    __ocl_event_t E = __SYCL_OpGroupAsyncCopyGlobalToLocal(
        __spv::Scope::Workgroup, DestT(dest.get()), SrcT(src.get()),
        numElements, srcStride, 0);
    return device_event(&E);
  }
 
  template <typename dataT>
  detail::enable_if_t<!detail::is_bool<dataT>::value, device_event>
  async_work_group_copy(global_ptr<dataT> dest, local_ptr<dataT> src,
                        size_t numElements, size_t destStride) const {
    using DestT = detail::ConvertToOpenCLType_t<decltype(dest)>;
    using SrcT = detail::ConvertToOpenCLType_t<decltype(src)>;
 
    __ocl_event_t E = __SYCL_OpGroupAsyncCopyLocalToGlobal(
        __spv::Scope::Workgroup, DestT(dest.get()), SrcT(src.get()),
        numElements, destStride, 0);
    return device_event(&E);
  }
 
  template <typename T, access::address_space DestS, access::address_space SrcS>
  detail::enable_if_t<detail::is_scalar_bool<T>::value, device_event>
  async_work_group_copy(multi_ptr<T, DestS> Dest, multi_ptr<T, SrcS> Src,
                        size_t NumElements, size_t Stride) const {
    static_assert(sizeof(bool) == sizeof(uint8_t),
                  "Async copy to/from bool memory is not supported.");
    auto DestP =
        multi_ptr<uint8_t, DestS>(reinterpret_cast<uint8_t *>(Dest.get()));
    auto SrcP =
        multi_ptr<uint8_t, SrcS>(reinterpret_cast<uint8_t *>(Src.get()));
    return async_work_group_copy(DestP, SrcP, NumElements, Stride);
  }
 
  template <typename T, access::address_space DestS, access::address_space SrcS>
  detail::enable_if_t<detail::is_vector_bool<T>::value, device_event>
  async_work_group_copy(multi_ptr<T, DestS> Dest, multi_ptr<T, SrcS> Src,
                        size_t NumElements, size_t Stride) const {
    static_assert(sizeof(bool) == sizeof(uint8_t),
                  "Async copy to/from bool memory is not supported.");
    using VecT = detail::change_base_type_t<T, uint8_t>;
    auto DestP = multi_ptr<VecT, DestS>(reinterpret_cast<VecT *>(Dest.get()));
    auto SrcP = multi_ptr<VecT, SrcS>(reinterpret_cast<VecT *>(Src.get()));
    return async_work_group_copy(DestP, SrcP, NumElements, Stride);
  }
 
  template <typename dataT>
  device_event async_work_group_copy(local_ptr<dataT> dest,
                                     global_ptr<dataT> src,
                                     size_t numElements) const {
    return async_work_group_copy(dest, src, numElements, 1);
  }
 
  template <typename dataT>
  device_event async_work_group_copy(global_ptr<dataT> dest,
                                     local_ptr<dataT> src,
                                     size_t numElements) const {
    return async_work_group_copy(dest, src, numElements, 1);
  }
 
  template <typename... eventTN>
  void wait_for(eventTN... Events) const {
    waitForHelper(Events...);
  }
 
  bool operator==(const group<Dimensions> &rhs) const {
    bool Result = (rhs.globalRange == globalRange) &&
                  (rhs.localRange == localRange) && (rhs.index == index);
    __SYCL_ASSERT(rhs.groupRange == groupRange &&
                  "inconsistent group class fields");
    return Result;
  }
 
  bool operator!=(const group<Dimensions> &rhs) const {
    return !((*this) == rhs);
  }
 
private:
  range<Dimensions> globalRange;
  range<Dimensions> localRange;
  range<Dimensions> groupRange;
  id<Dimensions> index;
 
  void waitForHelper() const {}
 
  void waitForHelper(device_event Event) const {
    Event.wait();
  }
 
  template <typename T, typename... Ts>
  void waitForHelper(T E, Ts... Es) const {
    waitForHelper(E);
    waitForHelper(Es...);
  }
 
protected:
  friend class detail::Builder;
  group(const range<Dimensions> &G, const range<Dimensions> &L,
        const range<Dimensions> GroupRange, const id<Dimensions> &I)
      : globalRange(G), localRange(L), groupRange(GroupRange), index(I) {
    // Make sure local range divides global without remainder:
    __SYCL_ASSERT(((G % L).size() == 0) &&
                  "global range is not multiple of local");
    __SYCL_ASSERT((((G / L) - GroupRange).size() == 0) &&
                  "inconsistent group constructor arguments");
  }
};
 
namespace detail {
template <int Dims> group<Dims> store_group(const group<Dims> *g) {
  return get_or_store(g);
}
} // namespace detail
 
template <int Dims>
__SYCL_DEPRECATED("use sycl::ext::oneapi::experimental::this_group() instead")
group<Dims> this_group() {
#ifdef __SYCL_DEVICE_ONLY__
  return detail::Builder::getElement(detail::declptr<group<Dims>>());
#else
  return detail::store_group<Dims>(nullptr);
#endif
}
 
namespace ext {
namespace oneapi {
namespace experimental {
template <int Dims> group<Dims> this_group() {
#ifdef __SYCL_DEVICE_ONLY__
  return sycl::detail::Builder::getElement(detail::declptr<group<Dims>>());
#else
  return sycl::detail::store_group<Dims>(nullptr);
#endif
}
} // namespace experimental
} // namespace oneapi
} // namespace ext
} // namespace sycl
} // __SYCL_INLINE_NAMESPACE(cl)