cuda.c

   1 /*
   2  * Copyright 2012      Ecole Normale Superieure
   3  *
   4  * Use of this software is governed by the MIT license
   5  *
   6  * Written by Sven Verdoolaege,
   7  * Ecole Normale Superieure, 45 rue d’Ulm, 75230 Paris, France
   8  */
   9
  10 #include <isl/aff.h>
  11 #include <isl/ast.h>
  12
  13 #include "cuda_common.h"
  14 #include "cuda.h"
  15 #include "gpu.h"
  16 #include "gpu_print.h"
  17 #include "print.h"
  18
  19 static __isl_give isl_printer *print_cuda_macros(__isl_take isl_printer *p)
  20 {
  21         const char *macros =
  22                 "#define cudaCheckReturn(ret) \\\n"
  23                 "  do { \\\n"
  24                 "    cudaError_t cudaCheckReturn_e = (ret); \\\n"
  25                 "    if (cudaCheckReturn_e != cudaSuccess) { \\\n"
  26                 "      fprintf(stderr, \"CUDA error: %s\\n\", "
  27                 "cudaGetErrorString(cudaCheckReturn_e)); \\\n"
  28                 "      fflush(stderr); \\\n"
  29                 "    } \\\n"
  30                 "    assert(cudaCheckReturn_e == cudaSuccess); \\\n"
  31                 "  } while(0)\n"
  32                 "#define cudaCheckKernel() \\\n"
  33                 "  do { \\\n"
  34                 "    cudaCheckReturn(cudaGetLastError()); \\\n"
  35                 "  } while(0)\n\n";
  36
  37         p = isl_printer_print_str(p, macros);
  38         return p;
  39 }
  40
  41 /* Print a declaration for the device array corresponding to "array" on "p".
  42  */
  43 static __isl_give isl_printer *declare_device_array(__isl_take isl_printer *p,
  44         struct gpu_array_info *array)
  45 {
  46         int i;
  47
  48         p = isl_printer_start_line(p);
  49         p = isl_printer_print_str(p, array->type);
  50         p = isl_printer_print_str(p, " ");
  51         if (!array->linearize && array->n_index > 1)
  52                 p = isl_printer_print_str(p, "(");
  53         p = isl_printer_print_str(p, "*dev_");
  54         p = isl_printer_print_str(p, array->name);
  55         if (!array->linearize && array->n_index > 1) {
  56                 p = isl_printer_print_str(p, ")");
  57                 for (i = 1; i < array->n_index; i++) {
  58                         p = isl_printer_print_str(p, "[");
  59                         p = isl_printer_print_pw_aff(p, array->bound[i]);
  60                         p = isl_printer_print_str(p, "]");
  61                 }
  62         }
  63         p = isl_printer_print_str(p, ";");
  64         p = isl_printer_end_line(p);
  65
  66         return p;
  67 }
  68
  69 static __isl_give isl_printer *declare_device_arrays(__isl_take isl_printer *p,
  70         struct gpu_prog *prog)
  71 {
  72         int i;
  73
  74         for (i = 0; i < prog->n_array; ++i) {
  75                 if (gpu_array_is_read_only_scalar(&prog->array[i]))
  76                         continue;
  77
  78                 p = declare_device_array(p, &prog->array[i]);
  79         }
  80         p = isl_printer_start_line(p);
  81         p = isl_printer_end_line(p);
  82         return p;
  83 }
  84
  85 static __isl_give isl_printer *allocate_device_arrays(
  86         __isl_take isl_printer *p, struct gpu_prog *prog)
  87 {
  88         int i;
  89
  90         for (i = 0; i < prog->n_array; ++i) {
  91                 if (gpu_array_is_read_only_scalar(&prog->array[i]))
  92                         continue;
  93                 p = isl_printer_start_line(p);
  94                 p = isl_printer_print_str(p,
  95                         "cudaCheckReturn(cudaMalloc((void **) &dev_");
  96                 p = isl_printer_print_str(p, prog->array[i].name);
  97                 p = isl_printer_print_str(p, ", ");
  98                 p = gpu_array_info_print_size(p, &prog->array[i]);
  99                 p = isl_printer_print_str(p, "));");
 100                 p = isl_printer_end_line(p);
 101         }
 102         p = isl_printer_start_line(p);
 103         p = isl_printer_end_line(p);
 104         return p;
 105 }
 106
 107 static __isl_give isl_printer *copy_arrays_to_device(__isl_take isl_printer *p,
 108         struct gpu_prog *prog)
 109 {
 110         int i;
 111
 112         for (i = 0; i < prog->n_array; ++i) {
 113                 isl_space *dim;
 114                 isl_set *read_i;
 115                 int empty;
 116
 117                 if (gpu_array_is_read_only_scalar(&prog->array[i]))
 118                         continue;
 119
 120                 dim = isl_space_copy(prog->array[i].space);
 121                 read_i = isl_union_set_extract_set(prog->copy_in, dim);
 122                 empty = isl_set_plain_is_empty(read_i);
 123                 isl_set_free(read_i);
 124                 if (empty)
 125                         continue;
 126
 127                 p = isl_printer_start_line(p);
 128                 p = isl_printer_print_str(p, "cudaCheckReturn(cudaMemcpy(dev_");
 129                 p = isl_printer_print_str(p, prog->array[i].name);
 130                 p = isl_printer_print_str(p, ", ");
 131
 132                 if (gpu_array_is_scalar(&prog->array[i]))
 133                         p = isl_printer_print_str(p, "&");
 134                 p = isl_printer_print_str(p, prog->array[i].name);
 135                 p = isl_printer_print_str(p, ", ");
 136
 137                 p = gpu_array_info_print_size(p, &prog->array[i]);
 138                 p = isl_printer_print_str(p, ", cudaMemcpyHostToDevice));");
 139                 p = isl_printer_end_line(p);
 140         }
 141         p = isl_printer_start_line(p);
 142         p = isl_printer_end_line(p);
 143         return p;
 144 }
 145
 146 static void print_reverse_list(FILE *out, int len, int *list)
 147 {
 148         int i;
 149
 150         if (len == 0)
 151                 return;
 152
 153         fprintf(out, "(");
 154         for (i = 0; i < len; ++i) {
 155                 if (i)
 156                         fprintf(out, ", ");
 157                 fprintf(out, "%d", list[len - 1 - i]);
 158         }
 159         fprintf(out, ")");
 160 }
 161
 162 /* Print the effective grid size as a list of the sizes in each
 163  * dimension, from innermost to outermost.
 164  */
 165 static __isl_give isl_printer *print_grid_size(__isl_take isl_printer *p,
 166         struct ppcg_kernel *kernel)
 167 {
 168         int i;
 169         int dim;
 170
 171         dim = isl_multi_pw_aff_dim(kernel->grid_size, isl_dim_set);
 172         if (dim == 0)
 173                 return p;
 174
 175         p = isl_printer_print_str(p, "(");
 176         for (i = dim - 1; i >= 0; --i) {
 177                 isl_pw_aff *bound;
 178
 179                 bound = isl_multi_pw_aff_get_pw_aff(kernel->grid_size, i);
 180                 p = isl_printer_print_pw_aff(p, bound);
 181                 isl_pw_aff_free(bound);
 182
 183                 if (i > 0)
 184                         p = isl_printer_print_str(p, ", ");
 185         }
 186
 187         p = isl_printer_print_str(p, ")");
 188
 189         return p;
 190 }
 191
 192 /* Print the grid definition.
 193  */
 194 static __isl_give isl_printer *print_grid(__isl_take isl_printer *p,
 195         struct ppcg_kernel *kernel)
 196 {
 197         p = isl_printer_start_line(p);
 198         p = isl_printer_print_str(p, "dim3 k");
 199         p = isl_printer_print_int(p, kernel->id);
 200         p = isl_printer_print_str(p, "_dimGrid");
 201         p = print_grid_size(p, kernel);
 202         p = isl_printer_print_str(p, ";");
 203         p = isl_printer_end_line(p);
 204
 205         return p;
 206 }
 207
 208 /* Print the arguments to a kernel declaration or call.  If "types" is set,
 209  * then print a declaration (including the types of the arguments).
 210  *
 211  * The arguments are printed in the following order
 212  * - the arrays accessed by the kernel
 213  * - the parameters
 214  * - the host loop iterators
 215  */
 216 static __isl_give isl_printer *print_kernel_arguments(__isl_take isl_printer *p,
 217         struct gpu_prog *prog, struct ppcg_kernel *kernel, int types)
 218 {
 219         int i, n;
 220         int first = 1;
 221         unsigned nparam;
 222         isl_space *space;
 223         const char *type;
 224
 225         for (i = 0; i < prog->n_array; ++i) {
 226                 isl_set *arr;
 227                 int empty;
 228
 229                 space = isl_space_copy(prog->array[i].space);
 230                 arr = isl_union_set_extract_set(kernel->arrays, space);
 231                 empty = isl_set_plain_is_empty(arr);
 232                 isl_set_free(arr);
 233                 if (empty)
 234                         continue;
 235
 236                 if (!first)
 237                         p = isl_printer_print_str(p, ", ");
 238
 239                 if (types)
 240                         p = gpu_array_info_print_declaration_argument(p,
 241                                 &prog->array[i], NULL);
 242                 else
 243                         p = gpu_array_info_print_call_argument(p,
 244                                 &prog->array[i]);
 245
 246                 first = 0;
 247         }
 248
 249         space = isl_union_set_get_space(kernel->arrays);
 250         nparam = isl_space_dim(space, isl_dim_param);
 251         for (i = 0; i < nparam; ++i) {
 252                 const char *name;
 253
 254                 name = isl_space_get_dim_name(space, isl_dim_param, i);
 255
 256                 if (!first)
 257                         p = isl_printer_print_str(p, ", ");
 258                 if (types)
 259                         p = isl_printer_print_str(p, "int ");
 260                 p = isl_printer_print_str(p, name);
 261
 262                 first = 0;
 263         }
 264         isl_space_free(space);
 265
 266         n = isl_space_dim(kernel->space, isl_dim_set);
 267         type = isl_options_get_ast_iterator_type(prog->ctx);
 268         for (i = 0; i < n; ++i) {
 269                 const char *name;
 270
 271                 if (!first)
 272                         p = isl_printer_print_str(p, ", ");
 273                 name = isl_space_get_dim_name(kernel->space, isl_dim_set, i);
 274                 if (types) {
 275                         p = isl_printer_print_str(p, type);
 276                         p = isl_printer_print_str(p, " ");
 277                 }
 278                 p = isl_printer_print_str(p, name);
 279
 280                 first = 0;
 281         }
 282
 283         return p;
 284 }
 285
 286 /* Print the header of the given kernel.
 287  */
 288 static __isl_give isl_printer *print_kernel_header(__isl_take isl_printer *p,
 289         struct gpu_prog *prog, struct ppcg_kernel *kernel)
 290 {
 291         p = isl_printer_start_line(p);
 292         p = isl_printer_print_str(p, "__global__ void kernel");
 293         p = isl_printer_print_int(p, kernel->id);
 294         p = isl_printer_print_str(p, "(");
 295         p = print_kernel_arguments(p, prog, kernel, 1);
 296         p = isl_printer_print_str(p, ")");
 297
 298         return p;
 299 }
 300
 301 /* Print the header of the given kernel to both gen->cuda.kernel_h
 302  * and gen->cuda.kernel_c.
 303  */
 304 static void print_kernel_headers(struct gpu_prog *prog,
 305         struct ppcg_kernel *kernel, struct cuda_info *cuda)
 306 {
 307         isl_printer *p;
 308
 309         p = isl_printer_to_file(prog->ctx, cuda->kernel_h);
 310         p = isl_printer_set_output_format(p, ISL_FORMAT_C);
 311         p = print_kernel_header(p, prog, kernel);
 312         p = isl_printer_print_str(p, ";");
 313         p = isl_printer_end_line(p);
 314         isl_printer_free(p);
 315
 316         p = isl_printer_to_file(prog->ctx, cuda->kernel_c);
 317         p = isl_printer_set_output_format(p, ISL_FORMAT_C);
 318         p = print_kernel_header(p, prog, kernel);
 319         p = isl_printer_end_line(p);
 320         isl_printer_free(p);
 321 }
 322
 323 static void print_indent(FILE *dst, int indent)
 324 {
 325         fprintf(dst, "%*s", indent, "");
 326 }
 327
 328 static void print_kernel_iterators(FILE *out, struct ppcg_kernel *kernel)
 329 {
 330         int i, n_grid;
 331         isl_ctx *ctx = isl_ast_node_get_ctx(kernel->tree);
 332         const char *type;
 333         const char *block_dims[] = { "blockIdx.x", "blockIdx.y" };
 334         const char *thread_dims[] = { "threadIdx.x", "threadIdx.y",
 335                                         "threadIdx.z" };
 336
 337         type = isl_options_get_ast_iterator_type(ctx);
 338
 339         n_grid = isl_multi_pw_aff_dim(kernel->grid_size, isl_dim_set);
 340         if (n_grid > 0) {
 341                 print_indent(out, 4);
 342                 fprintf(out, "%s ", type);
 343                 for (i = 0; i < n_grid; ++i) {
 344                         if (i)
 345                                 fprintf(out, ", ");
 346                         fprintf(out, "b%d = %s",
 347                                 i, block_dims[n_grid - 1 - i]);
 348                 }
 349                 fprintf(out, ";\n");
 350         }
 351
 352         if (kernel->n_block > 0) {
 353                 print_indent(out, 4);
 354                 fprintf(out, "%s ", type);
 355                 for (i = 0; i < kernel->n_block; ++i) {
 356                         if (i)
 357                                 fprintf(out, ", ");
 358                         fprintf(out, "t%d = %s",
 359                                 i, thread_dims[kernel->n_block - 1 - i]);
 360                 }
 361                 fprintf(out, ";\n");
 362         }
 363 }
 364
 365 static __isl_give isl_printer *print_kernel_var(__isl_take isl_printer *p,
 366         struct ppcg_kernel_var *var)
 367 {
 368         int j;
 369
 370         p = isl_printer_start_line(p);
 371         if (var->type == ppcg_access_shared)
 372                 p = isl_printer_print_str(p, "__shared__ ");
 373         p = isl_printer_print_str(p, var->array->type);
 374         p = isl_printer_print_str(p, " ");
 375         p = isl_printer_print_str(p,  var->name);
 376         for (j = 0; j < var->array->n_index; ++j) {
 377                 isl_val *v;
 378
 379                 p = isl_printer_print_str(p, "[");
 380                 v = isl_vec_get_element_val(var->size, j);
 381                 p = isl_printer_print_val(p, v);
 382                 isl_val_free(v);
 383                 p = isl_printer_print_str(p, "]");
 384         }
 385         p = isl_printer_print_str(p, ";");
 386         p = isl_printer_end_line(p);
 387
 388         return p;
 389 }
 390
 391 static __isl_give isl_printer *print_kernel_vars(__isl_take isl_printer *p,
 392         struct ppcg_kernel *kernel)
 393 {
 394         int i;
 395
 396         for (i = 0; i < kernel->n_var; ++i)
 397                 p = print_kernel_var(p, &kernel->var[i]);
 398
 399         return p;
 400 }
 401
 402 /* Print a sync statement.
 403  */
 404 static __isl_give isl_printer *print_sync(__isl_take isl_printer *p,
 405         struct ppcg_kernel_stmt *stmt)
 406 {
 407         p = isl_printer_start_line(p);
 408         p = isl_printer_print_str(p, "__syncthreads();");
 409         p = isl_printer_end_line(p);
 410
 411         return p;
 412 }
 413
 414 /* This function is called for each user statement in the AST,
 415  * i.e., for each kernel body statement, copy statement or sync statement.
 416  */
 417 static __isl_give isl_printer *print_kernel_stmt(__isl_take isl_printer *p,
 418         __isl_take isl_ast_print_options *print_options,
 419         __isl_keep isl_ast_node *node, void *user)
 420 {
 421         isl_id *id;
 422         struct ppcg_kernel_stmt *stmt;
 423
 424         id = isl_ast_node_get_annotation(node);
 425         stmt = isl_id_get_user(id);
 426         isl_id_free(id);
 427
 428         isl_ast_print_options_free(print_options);
 429
 430         switch (stmt->type) {
 431         case ppcg_kernel_copy:
 432                 return ppcg_kernel_print_copy(p, stmt);
 433         case ppcg_kernel_sync:
 434                 return print_sync(p, stmt);
 435         case ppcg_kernel_domain:
 436                 return ppcg_kernel_print_domain(p, stmt);
 437         }
 438
 439         return p;
 440 }
 441
 442 static void print_kernel(struct gpu_prog *prog, struct ppcg_kernel *kernel,
 443         struct cuda_info *cuda)
 444 {
 445         isl_ctx *ctx = isl_ast_node_get_ctx(kernel->tree);
 446         isl_ast_print_options *print_options;
 447         isl_printer *p;
 448
 449         print_kernel_headers(prog, kernel, cuda);
 450         fprintf(cuda->kernel_c, "{\n");
 451         print_kernel_iterators(cuda->kernel_c, kernel);
 452
 453         p = isl_printer_to_file(ctx, cuda->kernel_c);
 454         p = isl_printer_set_output_format(p, ISL_FORMAT_C);
 455         p = isl_printer_indent(p, 4);
 456
 457         p = print_kernel_vars(p, kernel);
 458         p = isl_printer_end_line(p);
 459         p = gpu_print_macros(p, kernel->tree);
 460
 461         print_options = isl_ast_print_options_alloc(ctx);
 462         print_options = isl_ast_print_options_set_print_user(print_options,
 463                                                     &print_kernel_stmt, NULL);
 464         p = isl_ast_node_print(kernel->tree, p, print_options);
 465         isl_printer_free(p);
 466
 467         fprintf(cuda->kernel_c, "}\n");
 468 }
 469
 470 struct print_host_user_data {
 471         struct cuda_info *cuda;
 472         struct gpu_prog *prog;
 473 };
 474
 475 /* Print the user statement of the host code to "p".
 476  *
 477  * In particular, print a block of statements that defines the grid
 478  * and the block and then launches the kernel.
 479  */
 480 static __isl_give isl_printer *print_host_user(__isl_take isl_printer *p,
 481         __isl_take isl_ast_print_options *print_options,
 482         __isl_keep isl_ast_node *node, void *user)
 483 {
 484         isl_id *id;
 485         struct ppcg_kernel *kernel;
 486         struct print_host_user_data *data;
 487
 488         id = isl_ast_node_get_annotation(node);
 489         kernel = isl_id_get_user(id);
 490         isl_id_free(id);
 491
 492         data = (struct print_host_user_data *) user;
 493
 494         p = isl_printer_start_line(p);
 495         p = isl_printer_print_str(p, "{");
 496         p = isl_printer_end_line(p);
 497         p = isl_printer_indent(p, 2);
 498
 499         p = isl_printer_start_line(p);
 500         p = isl_printer_print_str(p, "dim3 k");
 501         p = isl_printer_print_int(p, kernel->id);
 502         p = isl_printer_print_str(p, "_dimBlock");
 503         print_reverse_list(isl_printer_get_file(p),
 504                                 kernel->n_block, kernel->block_dim);
 505         p = isl_printer_print_str(p, ";");
 506         p = isl_printer_end_line(p);
 507
 508         p = print_grid(p, kernel);
 509
 510         p = isl_printer_start_line(p);
 511         p = isl_printer_print_str(p, "kernel");
 512         p = isl_printer_print_int(p, kernel->id);
 513         p = isl_printer_print_str(p, " <<<k");
 514         p = isl_printer_print_int(p, kernel->id);
 515         p = isl_printer_print_str(p, "_dimGrid, k");
 516         p = isl_printer_print_int(p, kernel->id);
 517         p = isl_printer_print_str(p, "_dimBlock>>> (");
 518         p = print_kernel_arguments(p, data->prog, kernel, 0);
 519         p = isl_printer_print_str(p, ");");
 520         p = isl_printer_end_line(p);
 521
 522         p = isl_printer_start_line(p);
 523         p = isl_printer_print_str(p, "cudaCheckKernel();");
 524         p = isl_printer_end_line(p);
 525
 526         p = isl_printer_indent(p, -2);
 527         p = isl_printer_start_line(p);
 528         p = isl_printer_print_str(p, "}");
 529         p = isl_printer_end_line(p);
 530
 531         p = isl_printer_start_line(p);
 532         p = isl_printer_end_line(p);
 533
 534         print_kernel(data->prog, kernel, data->cuda);
 535
 536         isl_ast_print_options_free(print_options);
 537
 538         return p;
 539 }
 540
 541 static __isl_give isl_printer *print_host_code(__isl_take isl_printer *p,
 542         struct gpu_prog *prog, __isl_keep isl_ast_node *tree,
 543         struct cuda_info *cuda)
 544 {
 545         isl_ast_print_options *print_options;
 546         isl_ctx *ctx = isl_ast_node_get_ctx(tree);
 547         struct print_host_user_data data = { cuda, prog };
 548
 549         print_options = isl_ast_print_options_alloc(ctx);
 550         print_options = isl_ast_print_options_set_print_user(print_options,
 551                                                 &print_host_user, &data);
 552
 553         p = gpu_print_macros(p, tree);
 554         p = isl_ast_node_print(tree, p, print_options);
 555
 556         return p;
 557 }
 558
 559 /* For each array that needs to be copied out (based on prog->copy_out),
 560  * copy the contents back from the GPU to the host.
 561  *
 562  * If any element of a given array appears in prog->copy_out, then its
 563  * entire extent is in prog->copy_out.  The bounds on this extent have
 564  * been precomputed in extract_array_info and are used in
 565  * gpu_array_info_print_size.
 566  */
 567 static __isl_give isl_printer *copy_arrays_from_device(
 568         __isl_take isl_printer *p, struct gpu_prog *prog)
 569 {
 570         int i;
 571         isl_union_set *copy_out;
 572         copy_out = isl_union_set_copy(prog->copy_out);
 573
 574         for (i = 0; i < prog->n_array; ++i) {
 575                 isl_space *dim;
 576                 isl_set *copy_out_i;
 577                 int empty;
 578
 579                 dim = isl_space_copy(prog->array[i].space);
 580                 copy_out_i = isl_union_set_extract_set(copy_out, dim);
 581                 empty = isl_set_plain_is_empty(copy_out_i);
 582                 isl_set_free(copy_out_i);
 583                 if (empty)
 584                         continue;
 585
 586                 p = isl_printer_start_line(p);
 587                 p = isl_printer_print_str(p, "cudaCheckReturn(cudaMemcpy(");
 588                 if (gpu_array_is_scalar(&prog->array[i]))
 589                         p = isl_printer_print_str(p, "&");
 590                 p = isl_printer_print_str(p, prog->array[i].name);
 591                 p = isl_printer_print_str(p, ", dev_");
 592                 p = isl_printer_print_str(p, prog->array[i].name);
 593                 p = isl_printer_print_str(p, ", ");
 594                 p = gpu_array_info_print_size(p, &prog->array[i]);
 595                 p = isl_printer_print_str(p, ", cudaMemcpyDeviceToHost));");
 596                 p = isl_printer_end_line(p);
 597         }
 598
 599         isl_union_set_free(copy_out);
 600         p = isl_printer_start_line(p);
 601         p = isl_printer_end_line(p);
 602         return p;
 603 }
 604
 605 static __isl_give isl_printer *free_device_arrays(__isl_take isl_printer *p,
 606         struct gpu_prog *prog)
 607 {
 608         int i;
 609
 610         for (i = 0; i < prog->n_array; ++i) {
 611                 if (gpu_array_is_read_only_scalar(&prog->array[i]))
 612                         continue;
 613                 p = isl_printer_start_line(p);
 614                 p = isl_printer_print_str(p, "cudaCheckReturn(cudaFree(dev_");
 615                 p = isl_printer_print_str(p, prog->array[i].name);
 616                 p = isl_printer_print_str(p, "));");
 617                 p = isl_printer_end_line(p);
 618         }
 619
 620         return p;
 621 }
 622
 623 /* Given a gpu_prog "prog" and the corresponding transformed AST
 624  * "tree", print the entire CUDA code to "p".
 625  * "types" collects the types for which a definition has already
 626  * been printed.
 627  */
 628 static __isl_give isl_printer *print_cuda(__isl_take isl_printer *p,
 629         struct gpu_prog *prog, __isl_keep isl_ast_node *tree,
 630         struct gpu_types *types, void *user)
 631 {
 632         struct cuda_info *cuda = user;
 633         isl_printer *kernel;
 634
 635         kernel = isl_printer_to_file(isl_printer_get_ctx(p), cuda->kernel_c);
 636         kernel = isl_printer_set_output_format(kernel, ISL_FORMAT_C);
 637         kernel = gpu_print_types(kernel, types, prog);
 638         isl_printer_free(kernel);
 639
 640         if (!kernel)
 641                 return isl_printer_free(p);
 642
 643         p = ppcg_start_block(p);
 644
 645         p = print_cuda_macros(p);
 646
 647         p = declare_device_arrays(p, prog);
 648         p = allocate_device_arrays(p, prog);
 649         p = copy_arrays_to_device(p, prog);
 650
 651         p = print_host_code(p, prog, tree, cuda);
 652
 653         p = copy_arrays_from_device(p, prog);
 654         p = free_device_arrays(p, prog);
 655
 656         p = ppcg_end_block(p);
 657
 658         return p;
 659 }
 660
 661 /* Transform the code in the file called "input" by replacing
 662  * all scops by corresponding CUDA code.
 663  * The names of the output files are derived from "input".
 664  *
 665  * We let generate_gpu do all the hard work and then let it call
 666  * us back for printing the AST in print_cuda.
 667  *
 668  * To prepare for this printing, we first open the output files
 669  * and we close them after generate_gpu has finished.
 670  */
 671 int generate_cuda(isl_ctx *ctx, struct ppcg_options *options,
 672         const char *input)
 673 {
 674         struct cuda_info cuda;
 675         int r;
 676
 677         cuda_open_files(&cuda, input);
 678
 679         r = generate_gpu(ctx, input, cuda.host_c, options, &print_cuda, &cuda);
 680
 681         cuda_close_files(&cuda);
 682
 683         return r;
 684 }