26 #include <opm/simulators/linalg/bda/opencl/opencl.hpp>
33 using spmv_blocked_kernel_type = cl::KernelFunctor<cl::Buffer&, cl::Buffer&, cl::Buffer&,
const unsigned int,
34 const cl::Buffer&, cl::Buffer&,
const unsigned int, cl::LocalSpaceArg>;
35 using spmv_kernel_type = cl::KernelFunctor<cl::Buffer&, cl::Buffer&, cl::Buffer&,
const unsigned int,
36 const cl::Buffer&, cl::Buffer&, cl::LocalSpaceArg>;
37 using residual_blocked_kernel_type = cl::KernelFunctor<cl::Buffer&, cl::Buffer&, cl::Buffer&,
const unsigned int,
38 cl::Buffer&,
const cl::Buffer&, cl::Buffer&,
const unsigned int, cl::LocalSpaceArg>;
39 using residual_kernel_type = cl::KernelFunctor<cl::Buffer&, cl::Buffer&, cl::Buffer&,
const unsigned int,
40 cl::Buffer&,
const cl::Buffer&, cl::Buffer&, cl::LocalSpaceArg>;
41 using ilu_apply1_kernel_type = cl::KernelFunctor<cl::Buffer&, cl::Buffer&, cl::Buffer&, cl::Buffer&,
const cl::Buffer&,
42 cl::Buffer&, cl::Buffer&,
const unsigned int,
const unsigned int, cl::LocalSpaceArg>;
43 using ilu_apply2_kernel_type = cl::KernelFunctor<cl::Buffer&, cl::Buffer&, cl::Buffer&, cl::Buffer&, cl::Buffer&,
44 cl::Buffer&, cl::Buffer&,
const unsigned int,
const unsigned int, cl::LocalSpaceArg>;
45 using stdwell_apply_kernel_type = cl::KernelFunctor<cl::Buffer&, cl::Buffer&, cl::Buffer&, cl::Buffer&,
46 cl::Buffer&, cl::Buffer&, cl::Buffer&, cl::Buffer&,
47 const unsigned int,
const unsigned int, cl::Buffer&,
48 cl::LocalSpaceArg, cl::LocalSpaceArg, cl::LocalSpaceArg>;
49 using stdwell_apply_no_reorder_kernel_type = cl::KernelFunctor<cl::Buffer&, cl::Buffer&, cl::Buffer&, cl::Buffer&,
50 cl::Buffer&, cl::Buffer&, cl::Buffer&,
51 const unsigned int,
const unsigned int, cl::Buffer&,
52 cl::LocalSpaceArg, cl::LocalSpaceArg, cl::LocalSpaceArg>;
53 using ilu_decomp_kernel_type = cl::KernelFunctor<
const unsigned int,
const unsigned int, cl::Buffer&, cl::Buffer&,
54 cl::Buffer&, cl::Buffer&, cl::Buffer&,
const int, cl::LocalSpaceArg>;
55 using isaiL_kernel_type = cl::KernelFunctor<cl::Buffer&, cl::Buffer&, cl::Buffer&, cl::Buffer&, cl::Buffer&, cl::Buffer&,
56 cl::Buffer&, cl::Buffer&, cl::Buffer&,
const unsigned int>;
57 using isaiU_kernel_type = cl::KernelFunctor<cl::Buffer&, cl::Buffer&, cl::Buffer&, cl::Buffer&, cl::Buffer&, cl::Buffer&,
58 cl::Buffer&, cl::Buffer&, cl::Buffer&, cl::Buffer&, cl::Buffer&,
const unsigned int>;
64 static cl::CommandQueue *queue;
65 static std::vector<double> tmp;
66 static bool initialized;
68 static std::unique_ptr<cl::KernelFunctor<cl::Buffer&, cl::Buffer&, cl::Buffer&, const unsigned int, cl::LocalSpaceArg> > dot_k;
69 static std::unique_ptr<cl::KernelFunctor<cl::Buffer&, cl::Buffer&, const unsigned int, cl::LocalSpaceArg> > norm_k;
70 static std::unique_ptr<cl::KernelFunctor<cl::Buffer&, const double, cl::Buffer&, const unsigned int> > axpy_k;
71 static std::unique_ptr<cl::KernelFunctor<cl::Buffer&, const double, const unsigned int> > scale_k;
72 static std::unique_ptr<cl::KernelFunctor<const double, cl::Buffer&, cl::Buffer&, cl::Buffer&, const unsigned int> > vmul_k;
73 static std::unique_ptr<cl::KernelFunctor<cl::Buffer&, cl::Buffer&, cl::Buffer&, const double, const double, const unsigned int> > custom_k;
74 static std::unique_ptr<cl::KernelFunctor<const cl::Buffer&, cl::Buffer&, cl::Buffer&, const unsigned int> > full_to_pressure_restriction_k;
75 static std::unique_ptr<cl::KernelFunctor<cl::Buffer&, cl::Buffer&, const unsigned int, const unsigned int> > add_coarse_pressure_correction_k;
76 static std::unique_ptr<cl::KernelFunctor<const cl::Buffer&, cl::Buffer&, const cl::Buffer&, const unsigned int> > prolongate_vector_k;
77 static std::unique_ptr<spmv_blocked_kernel_type> spmv_blocked_k;
78 static std::unique_ptr<spmv_blocked_kernel_type> spmv_blocked_add_k;
79 static std::unique_ptr<spmv_kernel_type> spmv_k;
80 static std::unique_ptr<spmv_kernel_type> spmv_noreset_k;
81 static std::unique_ptr<residual_blocked_kernel_type> residual_blocked_k;
82 static std::unique_ptr<residual_kernel_type> residual_k;
83 static std::unique_ptr<ilu_apply1_kernel_type> ILU_apply1_k;
84 static std::unique_ptr<ilu_apply2_kernel_type> ILU_apply2_k;
85 static std::unique_ptr<stdwell_apply_kernel_type> stdwell_apply_k;
86 static std::unique_ptr<stdwell_apply_no_reorder_kernel_type> stdwell_apply_no_reorder_k;
87 static std::unique_ptr<ilu_decomp_kernel_type> ilu_decomp_k;
88 static std::unique_ptr<isaiL_kernel_type> isaiL_k;
89 static std::unique_ptr<isaiU_kernel_type> isaiU_k;
94 static const std::string axpy_str;
95 static const std::string scale_str;
96 static const std::string vmul_str;
97 static const std::string dot_1_str;
98 static const std::string norm_str;
99 static const std::string custom_str;
100 static const std::string full_to_pressure_restriction_str;
101 static const std::string add_coarse_pressure_correction_str;
102 static const std::string prolongate_vector_str;
103 static const std::string spmv_blocked_str;
104 static const std::string spmv_blocked_add_str;
105 static const std::string spmv_str;
106 static const std::string spmv_noreset_str;
107 static const std::string residual_blocked_str;
108 static const std::string residual_str;
110 static const std::string ILU_apply1_str;
111 static const std::string ILU_apply2_str;
113 static const std::string ILU_apply1_fm_str;
114 static const std::string ILU_apply2_fm_str;
116 static const std::string stdwell_apply_str;
117 static const std::string stdwell_apply_no_reorder_str;
118 static const std::string ILU_decomp_str;
119 static const std::string isaiL_str;
120 static const std::string isaiU_str;
122 static void init(cl::Context *context, cl::CommandQueue *queue, std::vector<cl::Device>& devices,
int verbosity);
124 static double dot(cl::Buffer& in1, cl::Buffer& in2, cl::Buffer& out,
int N);
125 static double norm(cl::Buffer& in, cl::Buffer& out,
int N);
126 static void axpy(cl::Buffer& in,
const double a, cl::Buffer& out,
int N);
127 static void scale(cl::Buffer& in,
const double a,
int N);
128 static void vmul(
const double alpha, cl::Buffer& in1, cl::Buffer& in2, cl::Buffer& out,
int N);
129 static void custom(cl::Buffer& p, cl::Buffer& v, cl::Buffer& r,
const double omega,
const double beta,
int N);
130 static void full_to_pressure_restriction(
const cl::Buffer& fine_y, cl::Buffer& weights, cl::Buffer& coarse_y,
int Nb);
131 static void add_coarse_pressure_correction(cl::Buffer& coarse_x, cl::Buffer& fine_x,
int pressure_idx,
int Nb);
132 static void prolongate_vector(
const cl::Buffer& in, cl::Buffer& out,
const cl::Buffer& cols,
int N);
133 static void spmv(cl::Buffer& vals, cl::Buffer& cols, cl::Buffer& rows,
const cl::Buffer& x, cl::Buffer& b,
int Nb,
unsigned int block_size,
bool reset =
true,
bool add =
false);
134 static void residual(cl::Buffer& vals, cl::Buffer& cols, cl::Buffer& rows, cl::Buffer& x,
const cl::Buffer& rhs, cl::Buffer& out,
int Nb,
unsigned int block_size);
136 static void ILU_apply1(cl::Buffer& vals, cl::Buffer& cols, cl::Buffer& rows, cl::Buffer& diagIndex,
137 const cl::Buffer& y, cl::Buffer& x, cl::Buffer& rowsPerColor,
int color,
int Nb,
unsigned int block_size);
139 static void ILU_apply2(cl::Buffer& vals, cl::Buffer& cols, cl::Buffer& rows, cl::Buffer& diagIndex,
140 cl::Buffer& invDiagVals, cl::Buffer& x, cl::Buffer& rowsPerColor,
int color,
int Nb,
unsigned int block_size);
142 static void ILU_decomp(
int firstRow,
int lastRow, cl::Buffer& vals, cl::Buffer& cols, cl::Buffer& rows,
143 cl::Buffer& diagIndex, cl::Buffer& invDiagVals,
int Nb,
unsigned int block_size);
145 static void apply_stdwells_reorder(cl::Buffer& d_Cnnzs_ocl, cl::Buffer &d_Dnnzs_ocl, cl::Buffer &d_Bnnzs_ocl,
146 cl::Buffer &d_Ccols_ocl, cl::Buffer &d_Bcols_ocl, cl::Buffer &d_x, cl::Buffer &d_y,
147 cl::Buffer &d_toOrder,
int dim,
int dim_wells, cl::Buffer &d_val_pointers_ocl,
int num_std_wells);
149 static void apply_stdwells_no_reorder(cl::Buffer& d_Cnnzs_ocl, cl::Buffer &d_Dnnzs_ocl, cl::Buffer &d_Bnnzs_ocl,
150 cl::Buffer &d_Ccols_ocl, cl::Buffer &d_Bcols_ocl, cl::Buffer &d_x, cl::Buffer &d_y,
151 int dim,
int dim_wells, cl::Buffer &d_val_pointers_ocl,
int num_std_wells);
153 static void isaiL(cl::Buffer& diagIndex, cl::Buffer& colPointers, cl::Buffer& mapping, cl::Buffer& nvc,
154 cl::Buffer& luIdxs, cl::Buffer& xxIdxs, cl::Buffer& dxIdxs, cl::Buffer& LUvals, cl::Buffer& invLvals,
unsigned int Nb);
156 static void isaiU(cl::Buffer& diagIndex, cl::Buffer& colPointers, cl::Buffer& rowIndices, cl::Buffer& mapping,
157 cl::Buffer& nvc, cl::Buffer& luIdxs, cl::Buffer& xxIdxs, cl::Buffer& dxIdxs, cl::Buffer& LUvals,
158 cl::Buffer& invDiagVals, cl::Buffer& invUvals,
unsigned int Nb);
Definition: openclKernels.hpp:61
This file contains a set of helper functions used by VFPProd / VFPInj.
Definition: BlackoilPhases.hpp:27