/home/docs/checkouts/readthedocs.org/user_builds/advanced-micro-devices-composable-kernel/checkouts/docs-6.4.3/include/ck/tensor_operation/gpu/grid/gridwise_put_element_1d.hpp Source File

• docs-6.4.3
• include
• ck
• tensor_operation
• gpu
• grid

gridwise_put_element_1d.hpp

Go to the documentation of this file.

1 // SPDX-License-Identifier: MIT

2 // Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.

3 

4 #pragma once

5 

6 #include "ck/utility/data_type.hpp"

7 #include "ck/tensor_operation/gpu/thread/threadwise_tensor_slice_transfer.hpp"

8 

9 namespace ck {

10 

11 template <typename GridwisePutElementwise1dFunctor,

12  typename InGrid1dDesc,

13  typename InDataType,

14  typename IndexDataType,

15  typename OutDataType,

16  typename ElementwiseOperation>

17 __global__ void kernel_put_element_1d(const InGrid1dDesc in_grid_1d_desc,

18  const InDataType* __restrict__ p_in_global,

19  const IndexDataType* __restrict__ p_indices_global,

20  OutDataType* __restrict__ p_out_global,

21  const ElementwiseOperation elementwise_op)

22 {

23  GridwisePutElementwise1dFunctor::Run(

24  in_grid_1d_desc, p_in_global, p_indices_global, p_out_global, elementwise_op);

25 }

26 

27 // output[indices] = input

28 template <typename InGrid1dDesc,

29  typename InDataType,

30  typename IndexDataType,

31  typename OutDataType,

32  typename ElementwiseOperation,

33  InMemoryDataOperationEnum MemOp,

34  index_t InVectorSize>

35 struct GridwisePutElement_1D

36 {

37  static constexpr auto I0 = Number<0>{};

38 

39  static constexpr auto thread_buffer_desc_m =

40  make_naive_tensor_descriptor_packed(make_tuple(Number<InVectorSize>{}));

41 

42  __device__ static void Run(const InGrid1dDesc& in_grid_1d_desc,

43  const InDataType* __restrict__ p_in_global,

44  const IndexDataType* __restrict__ p_indices_global,

45  OutDataType* __restrict__ p_out_global,

46  const ElementwiseOperation& elementwise_op)

47  {

48  // Global Memory

49  const auto in_global_buf = make_dynamic_buffer<AddressSpaceEnum::Global>(

50  p_in_global, in_grid_1d_desc.GetElementSpaceSize());

51 

52  const auto indices_global_buf =

53  make_dynamic_buffer<AddressSpaceEnum::Global>(p_indices_global,

54  in_grid_1d_desc.GetElementSpaceSize(),

55  NumericLimits<IndexDataType>::Lowest());

56 

57  // VGPR

58  StaticBuffer<AddressSpaceEnum::Vgpr, InDataType, InVectorSize, true> in_thread_buf;

59  StaticBuffer<AddressSpaceEnum::Vgpr, IndexDataType, InVectorSize, true> indices_thread_buf;

60 

61  // Thread id, Block id and index

62  const index_t thread_global_id = get_thread_global_1d_id();

63  const auto thread_global_offset = make_multi_index(thread_global_id * InVectorSize);

64  const index_t blockSize = get_block_size();

65  const index_t blockPerGrid = get_grid_size();

66  const auto M = in_grid_1d_desc.GetLength(I0);

67  const index_t loop_step = blockPerGrid * blockSize * InVectorSize;

68  const auto loop_step_index = make_multi_index(loop_step);

69 

70  auto in_global_load =

71  ThreadwiseTensorSliceTransfer_v2<InDataType,

72  InDataType,

73  decltype(in_grid_1d_desc),

74  decltype(thread_buffer_desc_m),

75  Sequence<InVectorSize>, // SliceLengths

76  Sequence<0>, // DimAccessOrder

77  0, // SrcVectorDim

78  InVectorSize, // ScalarPerVector

79  1, // SrcScalarStrideInVector

80  false>{in_grid_1d_desc, thread_global_offset};

81 

82  auto indices_global_load =

83  ThreadwiseTensorSliceTransfer_v2<IndexDataType,

84  IndexDataType,

85  decltype(in_grid_1d_desc),

86  decltype(thread_buffer_desc_m),

87  Sequence<InVectorSize>, // SliceLengths

88  Sequence<0>, // DimAccessOrder

89  0, // SrcVectorDim

90  InVectorSize, // ScalarPerVector

91  1, // SrcScalarStrideInVector

92  false>{in_grid_1d_desc, thread_global_offset};

93 

94  index_t num_iter = M / loop_step;

95  do

96  {

97  in_global_load.Run(in_grid_1d_desc,

98  in_global_buf,

99  thread_buffer_desc_m,

100  make_tuple(I0),

101  in_thread_buf);

102 

103  in_global_load.MoveSrcSliceWindow(in_grid_1d_desc, loop_step_index);

104 

105  static_for<0, InVectorSize, 1>{}(

106  [&](auto iM) { elementwise_op(in_thread_buf(iM), in_thread_buf[iM]); });

107 

108  indices_global_load.Run(in_grid_1d_desc,

109  indices_global_buf,

110  thread_buffer_desc_m,

111  make_tuple(I0),

112  indices_thread_buf);

113 

114  indices_global_load.MoveSrcSliceWindow(in_grid_1d_desc, loop_step_index);

115 

116  static_for<0, InVectorSize, 1>{}([&](auto iM) {

117  if(indices_thread_buf[iM] >= 0)

118  {

119  if constexpr(MemOp == InMemoryDataOperationEnum::Set)

120  {

121  // User should guarantee each index in p_indices_global is different

122  *(p_out_global + indices_thread_buf[iM]) =

123  ck::type_convert<OutDataType>(in_thread_buf[iM]);

124  }

125  else if constexpr(MemOp == InMemoryDataOperationEnum::AtomicAdd)

126  {

127  atomic_add<OutDataType>(p_out_global + indices_thread_buf[iM],

128  ck::type_convert<OutDataType>(in_thread_buf[iM]));

129  }

130  else if constexpr(MemOp == InMemoryDataOperationEnum::AtomicMax)

131  {

132  atomic_max<OutDataType>(p_out_global + indices_thread_buf[iM],

133  ck::type_convert<OutDataType>(in_thread_buf[iM]));

134  }

135  else if constexpr(MemOp == InMemoryDataOperationEnum::Add)

136  {

137  // User should guarantee each index in p_indices_global is different

138  *(p_out_global + indices_thread_buf[iM]) +=

139  ck::type_convert<OutDataType>(in_thread_buf[iM]);

140  }

141  else

142  {

143  static_assert(MemOp == InMemoryDataOperationEnum::Set ||

144  MemOp == InMemoryDataOperationEnum::AtomicAdd ||

145  MemOp == InMemoryDataOperationEnum::AtomicMax ||

146  MemOp == InMemoryDataOperationEnum::Add);

147  }

148  }

149  });

150 

151  } while(--num_iter);

152  }

153 };

154 

155 } // namespace ck

data_type.hpp

ck

Definition: ck.hpp:264

ck::make_multi_index

__host__ constexpr __device__ auto make_multi_index(Xs &&... xs)

Definition: array_multi_index.hpp:15

ck::get_grid_size

__device__ index_t get_grid_size()

Definition: get_id.hpp:24

ck::InMemoryDataOperationEnum

InMemoryDataOperationEnum

Definition: ck.hpp:267

ck::InMemoryDataOperationEnum::Set

@ Set

ck::InMemoryDataOperationEnum::AtomicMax

@ AtomicMax

ck::InMemoryDataOperationEnum::AtomicAdd

@ AtomicAdd

ck::InMemoryDataOperationEnum::Add

@ Add

ck::make_naive_tensor_descriptor_packed

__host__ constexpr __device__ auto make_naive_tensor_descriptor_packed(const Tuple< Lengths... > &lengths)

Definition: tensor_descriptor_helper.hpp:101

ck::get_block_size

__device__ index_t get_block_size()

Definition: get_id.hpp:26

ck::get_thread_global_1d_id

__device__ index_t get_thread_global_1d_id()

Definition: get_id.hpp:18

ck::kernel_put_element_1d

__global__ void kernel_put_element_1d(const InGrid1dDesc in_grid_1d_desc, const InDataType *__restrict__ p_in_global, const IndexDataType *__restrict__ p_indices_global, OutDataType *__restrict__ p_out_global, const ElementwiseOperation elementwise_op)

Definition: gridwise_put_element_1d.hpp:17

ck::make_tuple

__host__ constexpr __device__ auto make_tuple(Xs &&... xs)

Definition: tuple.hpp:211

ck::index_t

int32_t index_t

Definition: ck.hpp:289

ck::GridwisePutElement_1D

Definition: gridwise_put_element_1d.hpp:36

ck::GridwisePutElement_1D::Run

static __device__ void Run(const InGrid1dDesc &in_grid_1d_desc, const InDataType *__restrict__ p_in_global, const IndexDataType *__restrict__ p_indices_global, OutDataType *__restrict__ p_out_global, const ElementwiseOperation &elementwise_op)

Definition: gridwise_put_element_1d.hpp:42

ck::GridwisePutElement_1D::thread_buffer_desc_m

static constexpr auto thread_buffer_desc_m

Definition: gridwise_put_element_1d.hpp:39

ck::GridwisePutElement_1D::I0

static constexpr auto I0

Definition: gridwise_put_element_1d.hpp:37

ck::NumericLimits

Definition: data_type.hpp:2831

ck::Sequence

Definition: sequence.hpp:43

ck::StaticBuffer

Definition: static_buffer.hpp:16

ck::ThreadwiseTensorSliceTransfer_v2

Definition: threadwise_tensor_slice_transfer.hpp:214

ck::integral_constant

Definition: integral_constant.hpp:10

ck::static_for

Definition: functional2.hpp:31

threadwise_tensor_slice_transfer.hpp