| 1 |
//@HEADER |
| 2 |
// ************************************************************************ |
| 3 |
// |
| 4 |
// Kokkos v. 4.0 |
| 5 |
// Copyright (2022) National Technology & Engineering |
| 6 |
// Solutions of Sandia, LLC (NTESS). |
| 7 |
// |
| 8 |
// Under the terms of Contract DE-NA0003525 with NTESS, |
| 9 |
// the U.S. Government retains certain rights in this software. |
| 10 |
// |
| 11 |
// Part of Kokkos, under the Apache License v2.0 with LLVM Exceptions. |
| 12 |
// See https://kokkos.org/LICENSE for license information. |
| 13 |
// SPDX-License-Identifier: Apache-2.0 WITH LLVM-exception |
| 14 |
// |
| 15 |
//@HEADER |
| 16 |
|
| 17 |
#ifndef KOKKOSKERNELS_KOKKOSBATCHED_KERNEL_HEADER_HPP |
| 18 |
#define KOKKOSKERNELS_KOKKOSBATCHED_KERNEL_HEADER_HPP |
| 19 |
|
| 20 |
#include <sstream> |
| 21 |
#include "KokkosKernels_Error.hpp" |
| 22 |
|
| 23 |
#if defined(KOKKOSKERNELS_ENABLE_TPL_MKL) |
| 24 |
#include <mkl.h> |
| 25 |
#endif // KOKKOSKERNELS_ENABLE_TPL_MKL |
| 26 |
|
| 27 |
#if defined(KOKKOSKERNELS_ENABLE_TPL_ARMPL) |
| 28 |
#include "armpl.h" |
| 29 |
#endif // KOKKOSKERNELS_ENABLE_TPL_ARMPL |
| 30 |
|
| 31 |
#if defined(KOKKOSKERNELS_ENABLE_TPL_CUBLAS) |
| 32 |
#include "cuda_runtime.h" |
| 33 |
#include "cublas_v2.h" |
| 34 |
#endif // KOKKOSKERNELS_ENABLE_TPL_CUBLAS |
| 35 |
|
| 36 |
#if defined(KOKKOSKERNELS_ENABLE_TPL_MAGMA) |
| 37 |
#include <magma_v2.h> |
| 38 |
#include <magma_batched.h> |
| 39 |
#endif // KOKKOSKERNELS_ENABLE_TPL_MAGMA |
| 40 |
|
| 41 |
namespace KokkosBatched { |
| 42 |
|
| 43 |
/// \brief Heuristic algorithm types. See BatchedKernelHandle for details. |
| 44 |
namespace BaseHeuristicAlgos { |
| 45 |
enum BASE_HEURISTIC_ALGOS : int { SQUARE = 0, TALL, WIDE, N }; |
| 46 |
} |
| 47 |
|
| 48 |
/// \brief Tpl algorithm types. See BatchedKernelHandle for details. |
| 49 |
namespace BaseTplAlgos { |
| 50 |
enum BASE_TPL_ALGOS : int { ARMPL = BaseHeuristicAlgos::N, MKL, N }; |
| 51 |
} |
| 52 |
|
| 53 |
/// \brief KokkosBatched algorithm types. See BatchedKernelHandle for details. |
| 54 |
namespace BaseKokkosBatchedAlgos { |
| 55 |
enum BASE_KOKKOS_BATCHED_ALGOS : int { KK_SERIAL = BaseTplAlgos::N, N }; |
| 56 |
} |
| 57 |
|
| 58 |
#define N_BASE_ALGOS BaseKokkosBatchedAlgos::N |
| 59 |
#define BASE_ALGO_STRS \ |
| 60 |
"BaseHeuristicAlgos::SQUARE", "BaseHeuristicAlgos::TALL", "BaseHeuristicAlgos::WIDE", "BaseTplAlgos::ARMPL", \ |
| 61 |
"BaseTplAlgosMKL", "BaseKokkosBatchedAlgos::KK_SERIAL" |
| 62 |
|
| 63 |
/// \brief TplParams abstracts underlying handle or execution queue type. |
| 64 |
struct TplParams { |
| 65 |
union { |
| 66 |
#if defined(KOKKOSKERNELS_ENABLE_TPL_MKL) |
| 67 |
// queue mkl_queue; |
| 68 |
// TODO: Add queue header? Cannot find any declarations in intel-18, let |
| 69 |
// alone oneAPI 2021 |
| 70 |
#endif // KOKKOSKERNELS_ENABLE_TPL_MKL |
| 71 |
|
| 72 |
#if defined(KOKKOSKERNELS_ENABLE_TPL_ARMPL) |
| 73 |
armpl_int_t ninter = 1; |
| 74 |
#endif // KOKKOSKERNELS_ENABLE_TPL_ARMPL |
| 75 |
|
| 76 |
#if defined(KOKKOSKERNELS_ENABLE_TPL_CUBLAS) |
| 77 |
cublasHandle_t *cublas_handle; |
| 78 |
#endif // KOKKOSKERNELS_ENABLE_TPL_CUBLAS |
| 79 |
|
| 80 |
#if defined(KOKKOSKERNELS_ENABLE_TPL_MAGMA) |
| 81 |
magma_queue_t *magma_queue; |
| 82 |
#endif // KOKKOSKERNELS_ENABLE_TPL_MAGMA |
| 83 |
|
| 84 |
uint8_t empty_union; // suppress empty union build warning |
| 85 |
}; |
| 86 |
}; |
| 87 |
|
| 88 |
// clang-format off |
| 89 |
/// \brief Handle for selecting runtime behavior of the BatchedGemm interface. |
| 90 |
/// |
| 91 |
/// \param kernelAlgoType Specifies which algorithm to use for invocation (default, SQUARE). |
| 92 |
/// |
| 93 |
/// Specifies whether to select optimal invocations based on inputs and |
| 94 |
/// heuristics: |
| 95 |
/// SQUARE select invocations based on square matrix heuristics where M=N |
| 96 |
/// TALL select invocations based on tall matrix heuristics where M>N |
| 97 |
/// WIDE select invocations based on wide matrix heuristics where M<N |
| 98 |
/// Note: If the heuristics indicate SIMD views are required for optimal |
| 99 |
/// performance, notify the user that SIMD views are required for |
| 100 |
/// optimal performance. |
| 101 |
/// |
| 102 |
/// Specifies which cmake-enabled TPL algorithm to invoke: |
| 103 |
/// TPL algorithms currently SUPPORTED by all batched routines: |
| 104 |
/// |
| 105 |
/// TPL algorithms currently UNSUPPORTED by all batched routines: |
| 106 |
/// ARMPL Invoke the ArmPL TPL interface |
| 107 |
/// MKL Invoke the MKL TPL interface |
| 108 |
/// CUBLAS Invoke the CuBLAS TPL interface |
| 109 |
/// MAGMA Invoke the Magma TPL interface |
| 110 |
/// Note: See each routine's handle for details about which TPLs are supported. |
| 111 |
/// Note: Requires that input views for A, B, and C reside on either host |
| 112 |
/// or device depending on the TPL selected. |
| 113 |
/// Note: If the user selects a TPL, an error will be thrown if: |
| 114 |
/// 1. The TPL is not enabled via cmake |
| 115 |
/// 2. The input views do not reside on the host/device as needed |
| 116 |
/// |
| 117 |
/// Specifies which kokkos-kernels (KK) algorithm to invoke: |
| 118 |
/// KK algorithms currently SUPPORTED by all KK batched routines: |
| 119 |
/// KK_SERIAL Invoke SerialFUNC via RangePolicy(BatchSz) |
| 120 |
/// KK algorithms currently UNSUPPORTED by all KK batched routines: |
| 121 |
/// KK_TEAM Invoke TeamFUNC via TeamPolicy(BatchSz) |
| 122 |
/// KK_TEAMVECTOR Invoke TeamVectorFUNC via TeamPolicy(BatchSz) |
| 123 |
/// KK_SERIALSIMD Invoke SerialFUNC via TeamPolicy(BatchSz) |
| 124 |
/// KK_TEAMSIMD Invoke TeamFUNC via TeamPolicy(BatchSz) |
| 125 |
/// KK_SERIAL_RANK0 Invoke SerialFUNC via RangePolicy(BatchSz*LHS_N*LHS_M) |
| 126 |
/// KK_SERIAL_SHMEM Invoke SerialFUNC via TeamPolicy(BatchSz) |
| 127 |
/// KK_DBLBUF Solve FUNC via TeamPolicy(BatchSz*TILES) |
| 128 |
/// Uses a tuned functor with tiling and double buffering |
| 129 |
/// via shared memory and register buffers. |
| 130 |
/// Note: See each routine's handle for details about which KK algorithms are |
| 131 |
/// supported. |
| 132 |
/// \param teamSz Specifies the team size that will affect any KK algorithm which uses |
| 133 |
/// TeamPolicy (default, Kokkos::AUTO). |
| 134 |
/// Note: Only applied if useAlgo_type == KK_* |
| 135 |
/// \param vecLen Specifies the vector length that will affect any KK algorithm which |
| 136 |
/// uses TeamPolicy and Kokkos::ThreadVectorRange or Kokkos::TeamVectorRange |
| 137 |
/// (default, Kokkos::AUTO). |
| 138 |
/// Note: Only applied if useAlgo_type == KK_* |
| 139 |
/// \param enabledDebug toggle debug messages. |
| 140 |
// clang-format on |
| 141 |
class BatchedKernelHandle { |
| 142 |
public: |
| 143 |
int teamSz = 0; |
| 144 |
int vecLen = 0; |
| 145 |
bool enableDebug = false; |
| 146 |
|
| 147 |
BatchedKernelHandle(int kernelAlgoType = BaseHeuristicAlgos::SQUARE, int teamSize = 0, int vecLength = 0) |
| 148 |
: teamSz(teamSize), vecLen(vecLength), _kernelAlgoType(kernelAlgoType) { |
| 149 |
#if !defined(KOKKOSKERNELS_ENABLE_TPL_ARMPL) || ARMPL_BUILD < 1058 |
| 150 |
if (_kernelAlgoType == BaseTplAlgos::ARMPL) { |
| 151 |
std::ostringstream os; |
| 152 |
os << "KokkosBatched::BatchedKernelHandle requires " |
| 153 |
"KOKKOSKERNELS_ENABLE_TPL_ARMPL and armpl version 21.0.0+" |
| 154 |
<< std::endl; |
| 155 |
KokkosKernels::Impl::throw_runtime_exception(os.str()); |
| 156 |
} |
| 157 |
#endif // !defined(KOKKOSKERNELS_ENABLE_TPL_ARMPL) |
| 158 |
}; |
| 159 |
|
| 160 |
int get_kernel_algo_type() const { return _kernelAlgoType; } |
| 161 |
|
| 162 |
std::string get_kernel_algo_type_str() const { return algo_type_strs[_kernelAlgoType]; } |
| 163 |
|
| 164 |
decltype(auto) get_tpl_params() const { |
| 165 |
#if _kernelAlgoType == ARMPL && defined(KOKKOSKERNELS_ENABLE_TPL_ARMPL) |
| 166 |
return &_tplParamsSingleton.ninter; |
| 167 |
#elif _kernelAlgoType == MKL && defined(KOKKOSKERNELS_ENABLE_TPL_MKL) |
| 168 |
return "BaseTplAlgos::MKL does not support any tpl parameters"; |
| 169 |
#else |
| 170 |
return "Unsupported kernelAlgoType:" + get_kernel_algo_type_str() + "."; |
| 171 |
#endif |
| 172 |
} |
| 173 |
|
| 174 |
// clang-format off |
| 175 |
/// kernelAlgoType: Specifies which algorithm to use for invocation (default, SQUARE). |
| 176 |
/// _tplParams: a handle or queue specific to the TPL API. |
| 177 |
/// managed internally unless provided by user via |
| 178 |
/// constructor overload |
| 179 |
// clang-format on |
| 180 |
protected: |
| 181 |
// Define TPL params singleton as static class method variable |
| 182 |
// Only one instance of tplParamsGlobalStorage may exist per process. |
| 183 |
static TplParams &_get_tpl_params_singleton() { |
| 184 |
static TplParams tplParamsGlobalStorage; |
| 185 |
return tplParamsGlobalStorage; |
| 186 |
} |
| 187 |
|
| 188 |
int _kernelAlgoType = BaseHeuristicAlgos::SQUARE; |
| 189 |
TplParams &_tplParamsSingleton = _get_tpl_params_singleton(); |
| 190 |
bool _tplParamsSet = false; |
| 191 |
|
| 192 |
private: |
| 193 |
const char *algo_type_strs[N_BASE_ALGOS] = {BASE_ALGO_STRS}; |
| 194 |
}; |
| 195 |
|
| 196 |
} // namespace KokkosBatched |
| 197 |
|
| 198 |
#endif // KOKKOSKERNELS_KOKKOSBATCHED_KERNEL_HEADER_HPP |