| 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 KOKKOSBATCHED_HOSTLEVEL_GEMM_HANDLE_DECL_HPP |
| 18 |
#define KOKKOSBATCHED_HOSTLEVEL_GEMM_HANDLE_DECL_HPP |
| 19 |
|
| 20 |
#include "KokkosBatched_Kernel_Handle.hpp" |
| 21 |
|
| 22 |
namespace KokkosBatched { |
| 23 |
|
| 24 |
/// \brief Tpl algorithm types. See BatchedGemmHandle for details. |
| 25 |
namespace GemmTplAlgos { |
| 26 |
enum GEMM_TPL_ALGOS : int { CUBLAS = N_BASE_ALGOS, MAGMA, N }; |
| 27 |
} |
| 28 |
|
| 29 |
/// \brief KokkosBatched algorithm types. See BatchedGemmHandle for details. |
| 30 |
namespace GemmKokkosBatchedAlgos { |
| 31 |
enum GEMM_KOKKOS_BATCHED_ALGOS : int { |
| 32 |
KK_TEAM = GemmTplAlgos::N, |
| 33 |
KK_TEAMVECTOR, |
| 34 |
KK_SERIALSIMD, |
| 35 |
KK_TEAMSIMD, |
| 36 |
KK_SERIAL_RANK0, |
| 37 |
KK_SERIAL_SHMEM, |
| 38 |
KK_DBLBUF, |
| 39 |
N |
| 40 |
}; |
| 41 |
} |
| 42 |
|
| 43 |
#define GEMM_ALGO_STRS \ |
| 44 |
"GemmTplAlgos::CUBLAS", "GemmTplAlgos::MAGMA", "GemmKokkosBatchedAlgos::KK_TEAM", \ |
| 45 |
"GemmKokkosBatchedAlgos::KK_TEAMVECTOR", "GemmKokkosBatchedAlgos::KK_SERIALSIMD", \ |
| 46 |
"GemmKokkosBatchedAlgos::KK_TEAMSIMD", "GemmKokkosBatchedAlgos::KK_SERIAL_RANK0", \ |
| 47 |
"GemmKokkosBatchedAlgos::KK_SERIAL_SHMEM", "GemmKokkosBatchedAlgos::KK_DBLBUF" |
| 48 |
// clang-format off |
| 49 |
/// \brief Handle for selecting runtime behavior of the BatchedGemm interface. |
| 50 |
/// |
| 51 |
/// \param kernelAlgoType Specifies which algorithm to use for invocation (default, SQUARE). |
| 52 |
/// |
| 53 |
/// Specifies whether to select optimal invocations based on inputs and |
| 54 |
/// heuristics: |
| 55 |
/// SQUARE select invocations based on square matrix heuristics where M=N |
| 56 |
/// TALL select invocations based on tall matrix heuristics where M>N |
| 57 |
/// WIDE select invocations based on wide matrix heuristics where M<N |
| 58 |
/// |
| 59 |
/// Specifies which cmake-enabled TPL algorithm to invoke: |
| 60 |
/// ARMPL Invoke the ArmPL TPL interface (Currently UNSUPPORTED) |
| 61 |
/// MKL Invoke the MKL TPL interface (Currently UNSUPPORTED) |
| 62 |
/// CUBLAS Invoke the CuBLAS TPL interface (Currently UNSUPPORTED) |
| 63 |
/// MAGMA Invoke the Magma TPL interface (Currently UNSUPPORTED) |
| 64 |
/// Note: Requires that input views for A, B, and C reside on either host |
| 65 |
/// or device depending on the TPL selected. |
| 66 |
/// Note: If the user selects a TPL, an error will be thrown if: |
| 67 |
/// 1. The TPL is not enabled via cmake |
| 68 |
/// 2. The input views do not reside on the host/device as needed |
| 69 |
/// |
| 70 |
/// Specifies which kokkos-kernels (KK) algorithm to invoke: |
| 71 |
/// KK_SERIAL Invoke SerialGemm via RangePolicy(BatchSz) |
| 72 |
/// KK_TEAM Invoke TeamGemm via TeamPolicy(BatchSz) |
| 73 |
/// KK_TEAMVECTOR Invoke TeamVectorGemm via TeamPolicy(BatchSz) |
| 74 |
/// KK_SERIALSIMD Invoke SerialGemm via TeamPolicy(BatchSz) |
| 75 |
/// KK_TEAMSIMD Invoke TeamGemm via TeamPolicy(BatchSz) |
| 76 |
/// KK_SERIAL_RANK0 Invoke SerialGemm via RangePolicy(BatchSz*N*M) |
| 77 |
/// Each thread computes one element of C. |
| 78 |
/// KK_SERIAL_SHMEM Invoke SerialGemm via TeamPolicy(BatchSz) |
| 79 |
/// Copies A and B to shared memory before GEMM. |
| 80 |
/// Each vector lane solves one element of C via SerialGemm. |
| 81 |
/// KK_DBLBUF Solve GEMM via TeamPolicy(BatchSz*TILES) |
| 82 |
/// Uses a tuned functor with tiling and double buffering |
| 83 |
/// via shared memory and register buffers. |
| 84 |
/// KK_DBLBUF generally performs better on GPUs when M, N >= 24. |
| 85 |
/// \param teamSz Specifies the team size that will affect any KK algorithm which uses |
| 86 |
/// TeamPolicy (default, Kokkos::AUTO). |
| 87 |
/// Note: Only applied if useAlgo_type == KK_* |
| 88 |
/// \param vecLen Specifies the vector length that will affect any KK algorithm which |
| 89 |
/// uses TeamPolicy and Kokkos::ThreadVectorRange or Kokkos::TeamVectorRange |
| 90 |
/// (default, Kokkos::AUTO). |
| 91 |
/// Note: Only applied if useAlgo_type == KK_* |
| 92 |
// clang-format on |
| 93 |
class BatchedGemmHandle : public BatchedKernelHandle { |
| 94 |
public: |
| 95 |
BatchedGemmHandle(int kernelAlgoType = BaseHeuristicAlgos::SQUARE, int teamSize = 0, int vecLength = 0) |
| 96 |
: BatchedKernelHandle(kernelAlgoType, teamSize, vecLength) { |
| 97 |
#if defined(KOKKOSKERNELS_ENABLE_TPL_CUBLAS) |
| 98 |
if (!_tplParamsSet && kernelAlgoType == GemmTplAlgos::CUBLAS) { |
| 99 |
static cublasHandle_t cublas_handle; |
| 100 |
_tplParamsSingleton.cublas_handle = &cublas_handle; |
| 101 |
_tplParamsSet = true; |
| 102 |
} |
| 103 |
#endif // CUBLAS |
| 104 |
|
| 105 |
#if defined(KOKKOSKERNELS_ENABLE_TPL_MAGMA) |
| 106 |
if (!_tplParamsSet && kernelAlgoType == GemmTplAlgos::MAGMA) { |
| 107 |
static magma_queue_t magma_queue; |
| 108 |
_tplParamsSingleton.magma_queue = &magma_queue; |
| 109 |
_tplParamsSet = true; |
| 110 |
} |
| 111 |
#endif // MAGMA |
| 112 |
}; |
| 113 |
|
| 114 |
BatchedGemmHandle(bool tplParamsSet, int kernelAlgoType = BaseHeuristicAlgos::SQUARE, int teamSize = 0, |
| 115 |
int vecLength = 0) |
| 116 |
: BatchedKernelHandle(kernelAlgoType, teamSize, vecLength) { |
| 117 |
_tplParamsSet = tplParamsSet; |
| 118 |
}; |
| 119 |
|
| 120 |
#if defined(KOKKOSKERNELS_ENABLE_TPL_CUBLAS) |
| 121 |
BatchedGemmHandle(cublasHandle_t &cublas_handle, int kernelAlgoType = BaseHeuristicAlgos::SQUARE, int teamSize = 0, |
| 122 |
int vecLength = 0) |
| 123 |
: BatchedGemmHandle(true, kernelAlgoType, teamSize, vecLength) { |
| 124 |
_tplParamsSingleton.cublas_handle = &cublas_handle; |
| 125 |
}; |
| 126 |
#endif // CUBLAS |
| 127 |
|
| 128 |
#if defined(KOKKOSKERNELS_ENABLE_TPL_MAGMA) |
| 129 |
BatchedGemmHandle(magma_queue_t &magma_queue, int kernelAlgoType = BaseHeuristicAlgos::SQUARE, int teamSize = 0, |
| 130 |
int vecLength = 0) |
| 131 |
: BatchedGemmHandle(true, kernelAlgoType, teamSize, vecLength) { |
| 132 |
_tplParamsSingleton.magma_queue = &magma_queue; |
| 133 |
}; |
| 134 |
#endif // MAGMA |
| 135 |
|
| 136 |
decltype(auto) get_tpl_params() { |
| 137 |
#if _kernelAlgoType == CUBLAS && defined(KOKKOSKERNELS_ENABLE_TPL_CUBLAS) |
| 138 |
return _tplParamsSingleton.cublas_handle; |
| 139 |
#elif _kernelAlgoType == MAGMA && defined(KOKKOSKERNELS_ENABLE_TPL_MAGMA) |
| 140 |
return _tplParamsSingleton.magma_queue; |
| 141 |
#else |
| 142 |
return this->BatchedKernelHandle::get_tpl_params(); |
| 143 |
#endif |
| 144 |
} |
| 145 |
|
| 146 |
std::string get_kernel_algo_type_str() const { return gemm_algo_type_strs[_kernelAlgoType]; } |
| 147 |
|
| 148 |
private: |
| 149 |
const char *gemm_algo_type_strs[GemmKokkosBatchedAlgos::N] = {BASE_ALGO_STRS, GEMM_ALGO_STRS}; |
| 150 |
}; |
| 151 |
|
| 152 |
} // namespace KokkosBatched |
| 153 |
|
| 154 |
#endif // KOKKOSBATCHED_HOSTLEVEL_GEMM_HANDLE_DECL_HPP |