/home/docs/checkouts/readthedocs.org/user_builds/advanced-micro-devices-composable-kernel/checkouts/develop/include/ck_tile/host/reference/reference_batched_elementwise.hpp Source File

• develop
• include
• ck_tile
• host
• reference

reference_batched_elementwise.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_tile/core.hpp"

7 #include "ck_tile/host/host_tensor.hpp"

8 #include <thread>

9 

10 namespace ck_tile {

11 

12 template <typename ADataType,

13  typename BDataType,

14  typename AccDataType,

15  typename CDataType,

16  typename AElementOp = ck_tile::identity,

17  typename BElementOp = ck_tile::identity,

18  typename BinaryElementOp = ck_tile::plus<AccDataType>>

19 CK_TILE_HOST void reference_batched_elementwise(const HostTensor<ADataType>& a_b_m_n,

20  const HostTensor<BDataType>& b_b_m_n,

21  HostTensor<CDataType>& c_b_m_n,

22  const AElementOp& a_element_op = {},

23  const BElementOp& b_element_op = {},

24  const BinaryElementOp& binary_element_op = {})

25 {

26  const ck_tile::index_t N = c_b_m_n.mDesc.get_lengths()[2];

27 

28  const bool broadcast_a_dim_b = (a_b_m_n.get_lengths()[0] == 1);

29  const bool broadcast_a_dim_m = (a_b_m_n.get_lengths()[1] == 1);

30  const bool broadcast_a_dim_n = (a_b_m_n.get_lengths()[2] == 1);

31 

32  const bool broadcast_b_dim_b = (b_b_m_n.get_lengths()[0] == 1);

33  const bool broadcast_b_dim_m = (b_b_m_n.get_lengths()[1] == 1);

34  const bool broadcast_b_dim_n = (b_b_m_n.get_lengths()[2] == 1);

35 

36  auto f = [&](auto batch, auto m) {

37  for(ck_tile::index_t n = 0; n < N; ++n)

38  {

39  AccDataType v_a{};

40  {

41  ck_tile::index_t i_b = (broadcast_a_dim_b ? 0 : batch);

42  ck_tile::index_t i_m = (broadcast_a_dim_m ? 0 : m);

43  ck_tile::index_t i_n = (broadcast_a_dim_n ? 0 : n);

44 

45  v_a = ck_tile::type_convert<AccDataType>(a_element_op(a_b_m_n(i_b, i_m, i_n)));

46  }

47 

48  AccDataType v_b{};

49  {

50  ck_tile::index_t i_b = (broadcast_b_dim_b ? 0 : batch);

51  ck_tile::index_t i_m = (broadcast_b_dim_m ? 0 : m);

52  ck_tile::index_t i_n = (broadcast_b_dim_n ? 0 : n);

53 

54  v_b = ck_tile::type_convert<AccDataType>(b_element_op(b_b_m_n(i_b, i_m, i_n)));

55  }

56 

57  c_b_m_n(batch, m, n) = ck_tile::type_convert<CDataType>(binary_element_op(v_a, v_b));

58  }

59  };

60 

61  make_ParallelTensorFunctor(f, c_b_m_n.mDesc.get_lengths()[0], c_b_m_n.mDesc.get_lengths()[1])(

62  std::thread::hardware_concurrency());

63 }

64 } // namespace ck_tile

CK_TILE_HOST

#define CK_TILE_HOST

Definition: config.hpp:40

core.hpp

ck_tile

Definition: cluster_descriptor.hpp:13

ck_tile::make_ParallelTensorFunctor

CK_TILE_HOST auto make_ParallelTensorFunctor(F f, Xs... xs)

Definition: host_tensor.hpp:329

ck_tile::index_t

int32_t index_t

Definition: integer.hpp:9

ck_tile::reference_batched_elementwise

CK_TILE_HOST void reference_batched_elementwise(const HostTensor< ADataType > &a_b_m_n, const HostTensor< BDataType > &b_b_m_n, HostTensor< CDataType > &c_b_m_n, const AElementOp &a_element_op={}, const BElementOp &b_element_op={}, const BinaryElementOp &binary_element_op={})

Definition: reference_batched_elementwise.hpp:19

ck_tile::HostTensorDescriptor::get_lengths

const std::vector< std::size_t > & get_lengths() const

Definition: host_tensor.hpp:198

ck_tile::HostTensor

Definition: host_tensor.hpp:336

ck_tile::HostTensor::get_lengths

decltype(auto) get_lengths() const

Definition: host_tensor.hpp:390

ck_tile::HostTensor::mDesc

Descriptor mDesc

Definition: host_tensor.hpp:800

ck_tile::identity

Definition: functional.hpp:86

ck_tile::plus

Definition: math.hpp:50

host_tensor.hpp