1 // Copyright (c) 2017-2022, Lawrence Livermore National Security, LLC and other CEED contributors. 2 // All Rights Reserved. See the top-level LICENSE and NOTICE files for details. 3 // 4 // SPDX-License-Identifier: BSD-2-Clause 5 // 6 // This file is part of CEED: http://github.com/ceed 7 8 // magma functions specific to ceed 9 #ifndef CEED_MAGMA_H 10 #define CEED_MAGMA_H 11 12 #include <ceed.h> 13 #include <ceed/backend.h> 14 #include <magma_v2.h> 15 16 #define MAGMA_MAXTHREADS_1D 128 17 #define MAGMA_MAXTHREADS_2D 128 18 #define MAGMA_MAXTHREADS_3D 64 19 20 // Define macro for determining number of threads in y-direction for basis kernels 21 #define MAGMA_BASIS_NTCOL(x, maxt) (((maxt) < (x)) ? 1 : ((maxt) / (x))) 22 23 // Define macros for non-tensor kernel instances 24 #define MAGMA_NONTENSOR_CUSTOM_KERNEL_MAX_P 40 25 #define MAGMA_NONTENSOR_CUSTOM_KERNEL_MAX_Q 40 26 #define MAGMA_NONTENSOR_KERNEL_INSTANCES 5 27 #define MAGMA_NONTENSOR_KERNEL_N_VALUES 10240, 51200, 102400, 512000, 1024000 28 29 #ifdef CEED_MAGMA_USE_HIP 30 typedef hipModule_t CeedMagmaModule; 31 typedef hipFunction_t CeedMagmaFunction; 32 #define CeedCompileMagma CeedCompile_Hip 33 #define CeedGetKernelMagma CeedGetKernel_Hip 34 #define CeedRunKernelMagma CeedRunKernel_Hip 35 #define CeedRunKernelDimMagma CeedRunKernelDim_Hip 36 #define CeedRunKernelDimSharedMagma CeedRunKernelDimShared_Hip 37 #else 38 typedef CUmodule CeedMagmaModule; 39 typedef CUfunction CeedMagmaFunction; 40 #define CeedCompileMagma CeedCompile_Cuda 41 #define CeedGetKernelMagma CeedGetKernel_Cuda 42 #define CeedRunKernelMagma CeedRunKernel_Cuda 43 #define CeedRunKernelDimMagma CeedRunKernelDim_Cuda 44 #define CeedRunKernelDimSharedMagma CeedRunKernelDimShared_Cuda 45 #endif 46 47 typedef struct { 48 CeedMagmaModule module; 49 CeedMagmaFunction Interp; 50 CeedMagmaFunction InterpTranspose; 51 CeedMagmaFunction Grad; 52 CeedMagmaFunction GradTranspose; 53 CeedMagmaFunction Weight; 54 CeedScalar *d_interp_1d; 55 CeedScalar *d_grad_1d; 56 CeedScalar *d_q_weight_1d; 57 } CeedBasis_Magma; 58 59 typedef struct { 60 CeedMagmaModule module_weight, module_interp[MAGMA_NONTENSOR_KERNEL_INSTANCES]; 61 CeedMagmaFunction Interp[MAGMA_NONTENSOR_KERNEL_INSTANCES]; 62 CeedMagmaFunction InterpTranspose[MAGMA_NONTENSOR_KERNEL_INSTANCES]; 63 CeedMagmaFunction Grad[MAGMA_NONTENSOR_KERNEL_INSTANCES]; 64 CeedMagmaFunction GradTranspose[MAGMA_NONTENSOR_KERNEL_INSTANCES]; 65 CeedMagmaFunction Weight; 66 CeedInt NB_interp[MAGMA_NONTENSOR_KERNEL_INSTANCES], NB_interp_t[MAGMA_NONTENSOR_KERNEL_INSTANCES]; 67 CeedInt NB_grad[MAGMA_NONTENSOR_KERNEL_INSTANCES], NB_grad_t[MAGMA_NONTENSOR_KERNEL_INSTANCES]; 68 CeedScalar *d_interp; 69 CeedScalar *d_grad; 70 CeedScalar *d_q_weight; 71 } CeedBasisNonTensor_Magma; 72 73 CEED_INTERN int CeedBasisCreateTensorH1_Magma(CeedInt dim, CeedInt P_1d, CeedInt Q_1d, const CeedScalar *interp_1d, const CeedScalar *grad_1d, 74 const CeedScalar *q_ref_1d, const CeedScalar *q_weight_1d, CeedBasis basis); 75 76 CEED_INTERN int CeedBasisCreateH1_Magma(CeedElemTopology topo, CeedInt dim, CeedInt num_nodes, CeedInt num_qpts, const CeedScalar *interp, 77 const CeedScalar *grad, const CeedScalar *q_ref, const CeedScalar *q_weight, CeedBasis basis); 78 79 CEED_INTERN magma_int_t magma_isdevptr(const void *); 80 81 // Comment the line below to use the default magma_is_devptr function 82 #define magma_is_devptr magma_isdevptr 83 84 // If magma and cuda/ref are using the null stream, then ceed_magma_queue_sync should do nothing 85 #define ceed_magma_queue_sync(...) 86 87 #endif // CEED_MAGMA_H 88