File: nvd_kernel.h

package info (click to toggle)
lammps 20220106.git7586adbb6a%2Bds1-2
  • links: PTS, VCS
  • area: main
  • in suites: bookworm
  • size: 348,064 kB
  • sloc: cpp: 831,421; python: 24,896; xml: 14,949; f90: 10,845; ansic: 7,967; sh: 4,226; perl: 4,064; fortran: 2,424; makefile: 1,501; objc: 238; lisp: 163; csh: 16; awk: 14; tcl: 6
file content (421 lines) | stat: -rw-r--r-- 14,756 bytes parent folder | download
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
164
165
166
167
168
169
170
171
172
173
174
175
176
177
178
179
180
181
182
183
184
185
186
187
188
189
190
191
192
193
194
195
196
197
198
199
200
201
202
203
204
205
206
207
208
209
210
211
212
213
214
215
216
217
218
219
220
221
222
223
224
225
226
227
228
229
230
231
232
233
234
235
236
237
238
239
240
241
242
243
244
245
246
247
248
249
250
251
252
253
254
255
256
257
258
259
260
261
262
263
264
265
266
267
268
269
270
271
272
273
274
275
276
277
278
279
280
281
282
283
284
285
286
287
288
289
290
291
292
293
294
295
296
297
298
299
300
301
302
303
304
305
306
307
308
309
310
311
312
313
314
315
316
317
318
319
320
321
322
323
324
325
326
327
328
329
330
331
332
333
334
335
336
337
338
339
340
341
342
343
344
345
346
347
348
349
350
351
352
353
354
355
356
357
358
359
360
361
362
363
364
365
366
367
368
369
370
371
372
373
374
375
376
377
378
379
380
381
382
383
384
385
386
387
388
389
390
391
392
393
394
395
396
397
398
399
400
401
402
403
404
405
406
407
408
409
410
411
412
413
414
415
416
417
418
419
420
421
/***************************************************************************
                                nvd_kernel.h
                             -------------------
                               W. Michael Brown

  Utilities for dealing with CUDA Driver kernels

 __________________________________________________________________________
    This file is part of the Geryon Unified Coprocessor Library (UCL)
 __________________________________________________________________________

    begin                : Tue Feb 9 2010
    copyright            : (C) 2010 by W. Michael Brown
    email                : brownw@ornl.gov
 ***************************************************************************/

/* -----------------------------------------------------------------------
   Copyright (2010) Sandia Corporation.  Under the terms of Contract
   DE-AC04-94AL85000 with Sandia Corporation, the U.S. Government retains
   certain rights in this software.  This software is distributed under
   the Simplified BSD License.
   ----------------------------------------------------------------------- */

#ifndef NVD_KERNEL
#define NVD_KERNEL

#include "nvd_device.h"
#include <fstream>
#include <cstdio>

namespace ucl_cudadr {

class UCL_Texture;
template <class numtyp> class UCL_D_Vec;
template <class numtyp> class UCL_D_Mat;
template <class hosttype, class devtype> class UCL_Vector;
template <class hosttype, class devtype> class UCL_Matrix;
#define UCL_MAX_KERNEL_ARGS 256

/// Class storing 1 or more kernel functions from a single string or file
class UCL_Program {
 public:
  inline UCL_Program(UCL_Device &device) { _cq=device.cq(); }
  inline UCL_Program(UCL_Device &device, const void *program,
                     const char *flags="", std::string *log=nullptr) {
    _cq=device.cq();
    init(device);
    load_string(program,flags,log);
  }

  inline ~UCL_Program() {}

  /// Initialize the program with a device
  inline void init(UCL_Device &device) { _cq=device.cq(); }

  /// Clear any data associated with program
  /** \note Must call init() after each clear **/
  inline void clear() { }

  /// Load a program from a file and compile with flags
  inline int load(const char *filename, const char *flags="",
                  std::string *log=nullptr) {
    std::ifstream in(filename);
    if (!in || in.is_open()==false) {
      #ifndef UCL_NO_EXIT
      std::cerr << "UCL Error: Could not open kernel file: "
                << filename << std::endl;
      UCL_GERYON_EXIT;
      #endif
      return UCL_FILE_NOT_FOUND;
    }

    std::string program((std::istreambuf_iterator<char>(in)),
                        std::istreambuf_iterator<char>());
    in.close();
    return load_string(program.c_str(),flags,log);
  }

  /// Load a program from a string and compile with flags
  inline int load_string(const void *program, const char *flags="",
                         std::string *log=nullptr, FILE* foutput=nullptr) {
    if (std::string(flags)=="BINARY")
      return load_binary((const char *)program);
    const unsigned int num_opts=2;
    CUjit_option options[num_opts];
    void *values[num_opts];

    // set up size of compilation log buffer
    options[0] = CU_JIT_INFO_LOG_BUFFER_SIZE_BYTES;
    values[0] = (void *)(int)10240;
    // set up pointer to the compilation log buffer
    options[1] = CU_JIT_INFO_LOG_BUFFER;
    char clog[10240];
    values[1] = clog;

    CUresult err=cuModuleLoadDataEx(&_module,program,num_opts,
                                    options,(void **)values);

    if (log!=nullptr)
      *log=std::string(clog);

    if (err != CUDA_SUCCESS) {
      #ifndef UCL_NO_EXIT
      std::cerr << std::endl << std::endl
                << "----------------------------------------------------------\n"
                << " UCL Error: Error compiling PTX Program...\n"
                << "----------------------------------------------------------\n";
      std::cerr << log << std::endl
                << "----------------------------------------------------------\n\n";
      #endif
      if (foutput != nullptr) {
        fprintf(foutput,"\n\n");
        fprintf(foutput, "----------------------------------------------------------\n");
        fprintf(foutput, " UCL Error: Error compiling PTX Program...\n");
        fprintf(foutput, "----------------------------------------------------------\n");
        fprintf(foutput, "%s\n",log->c_str());
        fprintf(foutput, "----------------------------------------------------------\n");
        fprintf(foutput,"\n\n");
      }
      return UCL_COMPILE_ERROR;
    }

    return UCL_SUCCESS;
  }

  /// Load a precompiled program from a file
  inline int load_binary(const char *filename) {
    CUmodule _module;
    CUresult err = cuModuleLoad(&_module,filename);
    if (err==301) {
      #ifndef UCL_NO_EXIT
      std::cerr << "UCL Error: Could not open binary kernel file: "
                << filename << std::endl;
      UCL_GERYON_EXIT;
      #endif
      return UCL_FILE_NOT_FOUND;
    } else if (err!=CUDA_SUCCESS) {
      #ifndef UCL_NO_EXIT
      std::cerr << "UCL Error: Error loading binary kernel file: "
                << filename << std::endl;
      UCL_GERYON_EXIT;
      #endif
      return UCL_FILE_NOT_FOUND;
    }
    //int ucl_error=UCL_SUCCESS;
    //if (err==301)
    //  return UCL_FILE_NOT_FOUND;
    //else if (err!=CUDA_SUCCESS)
    //  return UCL_ERROR;
    return UCL_SUCCESS;
  }

  /// Return the default command queue/stream associated with this data
  inline command_queue & cq() { return _cq; }

  friend class UCL_Kernel;
 private:
  CUmodule _module;
  CUstream _cq;
  friend class UCL_Texture;
  friend class UCL_Const;
};

/// Class for dealing with CUDA Driver kernels
class UCL_Kernel {
 public:
  UCL_Kernel() : _dimensions(1), _num_args(0) {
    #if CUDA_VERSION < 4000
    _param_size=0;
    #endif
    _num_blocks[0]=0;
  }

  UCL_Kernel(UCL_Program &program, const char *function) :
    _dimensions(1), _num_args(0) {
    #if CUDA_VERSION < 4000
    _param_size=0;
    #endif
    _num_blocks[0]=0;
    set_function(program,function);
    _cq=program._cq;
  }

  ~UCL_Kernel() {}

  /// Clear any function associated with the kernel
  inline void clear() { }

  /// Get the kernel function from a program
  /** \ret UCL_ERROR_FLAG (UCL_SUCCESS, UCL_FILE_NOT_FOUND, UCL_ERROR) **/
  inline int set_function(UCL_Program &program, const char *function) {
    CUresult err=cuModuleGetFunction(&_kernel,program._module,function);
    if (err!=CUDA_SUCCESS) {
      #ifndef UCL_NO_EXIT
      std::cerr << "UCL Error: Could not find function: " << function
                << " in program.\n";
      UCL_GERYON_EXIT;
      #endif
      return UCL_FUNCTION_NOT_FOUND;
    }
    _cq=program._cq;
    return UCL_SUCCESS;
  }

  /// Set the kernel argument.
  /** If not a device pointer, this must be repeated each time the argument
    * changes
    * \note To set kernel parameter i (i>0), parameter i-1 must be set **/
  template <class dtype>
  inline void set_arg(const unsigned index, const dtype * const arg) {
    if (index==_num_args)
      add_arg(arg);
    else if (index<_num_args)
      #if CUDA_VERSION >= 4000
      _kernel_args[index]=arg;
      #else
      CU_SAFE_CALL(cuParamSetv(_kernel, _offsets[index], arg, sizeof(dtype)));
      #endif
    else
      assert(0==1); // Must add kernel parameters in sequential order
  }

  /// Set a geryon container as a kernel argument.
  template <class numtyp>
  inline void set_arg(const UCL_D_Vec<numtyp> * const arg)
    { set_arg(&arg->begin()); }

  /// Set a geryon container as a kernel argument.
  template <class numtyp>
  inline void set_arg(const UCL_D_Mat<numtyp> * const arg)
    { set_arg(&arg->begin()); }

  /// Set a geryon container as a kernel argument.
  template <class hosttype, class devtype>
  inline void set_arg(const UCL_Vector<hosttype, devtype> * const arg)
    { set_arg(&arg->device.begin()); }

  /// Set a geryon container as a kernel argument.
  template <class hosttype, class devtype>
  inline void set_arg(const UCL_Matrix<hosttype, devtype> * const arg)
    { set_arg(&arg->device.begin()); }

  /// Add a kernel argument.
  inline void add_arg(const CUdeviceptr* const arg) {
    #if CUDA_VERSION >= 4000
    _kernel_args[_num_args]=(void *)arg;
    #else
    void* ptr = (void*)(size_t)(*arg);
    _param_size = (_param_size + __alignof(ptr) - 1) & ~(__alignof(ptr) - 1);
    CU_SAFE_CALL(cuParamSetv(_kernel, _param_size, &ptr, sizeof(ptr)));
    _offsets.push_back(_param_size);
    _param_size+=sizeof(ptr);
    #endif
    _num_args++;
    if (_num_args>UCL_MAX_KERNEL_ARGS) assert(0==1);
  }

  /// Add a kernel argument.
  template <class dtype>
  inline void add_arg(const dtype* const arg) {
    #if CUDA_VERSION >= 4000
    _kernel_args[_num_args]=const_cast<dtype *>(arg);
    #else
    _param_size = (_param_size+__alignof(dtype)-1) & ~(__alignof(dtype)-1);
    CU_SAFE_CALL(cuParamSetv(_kernel,_param_size,(void*)arg,sizeof(dtype)));
    _offsets.push_back(_param_size);
    _param_size+=sizeof(dtype);
    #endif
    _num_args++;
    if (_num_args>UCL_MAX_KERNEL_ARGS) assert(0==1);
  }

  /// Add a geryon container as a kernel argument.
  template <class numtyp>
  inline void add_arg(const UCL_D_Vec<numtyp> * const arg)
    { add_arg(&arg->begin()); }

  /// Add a geryon container as a kernel argument.
  template <class numtyp>
  inline void add_arg(const UCL_D_Mat<numtyp> * const arg)
    { add_arg(&arg->begin()); }

  /// Add a geryon container as a kernel argument.
  template <class hosttype, class devtype>
  inline void add_arg(const UCL_Vector<hosttype, devtype> * const arg)
    { add_arg(&arg->device.begin()); }

  /// Add a geryon container as a kernel argument.
  template <class hosttype, class devtype>
  inline void add_arg(const UCL_Matrix<hosttype, devtype> * const arg)
    { add_arg(&arg->device.begin()); }

  /// Set the number of thread blocks and the number of threads in each block
  /** \note This should be called before any arguments have been added
      \note The default command queue is used for the kernel execution **/
  inline void set_size(const size_t num_blocks, const size_t block_size) {
    _dimensions=1;
    _num_blocks[0]=num_blocks;
    _num_blocks[1]=1;
    _num_blocks[2]=1;
    #if CUDA_VERSION >= 4000
    _block_size[0]=block_size;
    _block_size[1]=1;
    _block_size[2]=1;
    #else
    CU_SAFE_CALL(cuFuncSetBlockShape(_kernel,block_size,1,1));
    #endif
  }

  /// Set the number of thread blocks and the number of threads in each block
  /** \note This should be called before any arguments have been added
      \note The default command queue for the kernel is changed to cq **/
  inline void set_size(const size_t num_blocks, const size_t block_size,
                       command_queue &cq)
    { _cq=cq; set_size(num_blocks,block_size); }

  /// Set the number of thread blocks and the number of threads in each block
  /** \note This should be called before any arguments have been added
      \note The default command queue is used for the kernel execution **/
  inline void set_size(const size_t num_blocks_x, const size_t num_blocks_y,
                       const size_t block_size_x, const size_t block_size_y) {
    _dimensions=2;
    _num_blocks[0]=num_blocks_x;
    _num_blocks[1]=num_blocks_y;
    _num_blocks[2]=1;
    #if CUDA_VERSION >= 4000
    _block_size[0]=block_size_x;
    _block_size[1]=block_size_y;
    _block_size[2]=1;
    #else
    CU_SAFE_CALL(cuFuncSetBlockShape(_kernel,block_size_x,block_size_y,1));
    #endif
  }

  /// Set the number of thread blocks and the number of threads in each block
  /** \note This should be called before any arguments have been added
      \note The default command queue for the kernel is changed to cq **/
  inline void set_size(const size_t num_blocks_x, const size_t num_blocks_y,
                       const size_t block_size_x, const size_t block_size_y,
                       command_queue &cq)
    {_cq=cq; set_size(num_blocks_x, num_blocks_y, block_size_x, block_size_y);}

  /// Set the number of thread blocks and the number of threads in each block
  /** \note This should be called before any arguments have been added
      \note The default command queue is used for the kernel execution **/
  inline void set_size(const size_t num_blocks_x, const size_t num_blocks_y,
                       const size_t block_size_x,
                       const size_t block_size_y, const size_t block_size_z) {
    _dimensions=2;
    _num_blocks[0]=num_blocks_x;
    _num_blocks[1]=num_blocks_y;
    _num_blocks[2]=1;
    #if CUDA_VERSION >= 4000
    _block_size[0]=block_size_x;
    _block_size[1]=block_size_y;
    _block_size[2]=block_size_z;
    #else
    CU_SAFE_CALL(cuFuncSetBlockShape(_kernel,block_size_x,block_size_y,
                                     block_size_z));
    #endif
  }

  /// Set the number of thread blocks and the number of threads in each block
  /** \note This should be called before any arguments have been added
      \note The default command queue is used for the kernel execution **/
  inline void set_size(const size_t num_blocks_x, const size_t num_blocks_y,
                       const size_t block_size_x, const size_t block_size_y,
                       const size_t block_size_z, command_queue &cq) {
    _cq=cq;
    set_size(num_blocks_x, num_blocks_y, block_size_x, block_size_y,
             block_size_z);
  }

  /// Run the kernel in the default command queue
  inline void run() {
    #if CUDA_VERSION >= 4000
    CU_SAFE_CALL(cuLaunchKernel(_kernel,_num_blocks[0],_num_blocks[1],
                                _num_blocks[2],_block_size[0],_block_size[1],
                                _block_size[2],0,_cq,_kernel_args,nullptr));
    #else
    CU_SAFE_CALL(cuParamSetSize(_kernel,_param_size));
    CU_SAFE_CALL(cuLaunchGridAsync(_kernel,_num_blocks[0],_num_blocks[1],_cq));
    #endif
  }

  /// Clear any arguments associated with the kernel
  inline void clear_args() {
    _num_args=0;
    #if CUDA_VERSION < 4000
    _offsets.clear();
    _param_size=0;
    #endif
  }

  /// Return the default command queue/stream associated with this data
  inline command_queue & cq() { return _cq; }
  /// Change the default command queue associated with matrix
  inline void cq(command_queue &cq_in) { _cq=cq_in; }
  #include "ucl_arg_kludge.h"

 private:
  CUfunction _kernel;
  CUstream _cq;
  unsigned _dimensions;
  unsigned _num_blocks[3];
  unsigned _num_args;
  friend class UCL_Texture;

  #if CUDA_VERSION >= 4000
  unsigned _block_size[3];
  void * _kernel_args[UCL_MAX_KERNEL_ARGS];
  #else
  std::vector<unsigned> _offsets;
  unsigned _param_size;
  #endif
};

} // namespace

#endif