gpu.h

   1 #ifndef _GPU_H
   2 #define _GPU_H
   3
   4 #include <pet.h>
   5 #include "ppcg_options.h"
   6
   7 struct gpu_array_info {
   8         isl_space *dim;
   9         /* Element type. */
  10         char *type;
  11         /* Element size. */
  12         int size;
  13         /* Name of the array. */
  14         char *name;
  15         /* Number of indices. */
  16         unsigned n_index;
  17         /* For each index, a bound on the array in that direction. */
  18         isl_pw_aff **bound;
  19
  20         /* All references to this array; point to elements of a linked list. */
  21         int n_ref;
  22         struct gpu_stmt_access **refs;
  23
  24         /* The reference groups associated to this array. */
  25         int n_group;
  26         struct gpu_array_ref_group **groups;
  27
  28         /* For scalars, is this scalar read-only within the entire program? */
  29         int read_only;
  30 };
  31
  32 /* For each index i, array->bound[i] specialized to the current kernel. */
  33 struct gpu_local_array_info {
  34         isl_pw_aff_list *bound;
  35 };
  36
  37 struct gpu_prog {
  38         isl_ctx *ctx;
  39
  40         struct pet_scop *scop;
  41
  42         /* Set of parameter values */
  43         isl_set *context;
  44
  45         /* All read accesses in the entire program */
  46         isl_union_map *read;
  47
  48         /* All write accesses in the entire program */
  49         isl_union_map *write;
  50
  51         /* Uninitialized data elements (or an overapproximation) */
  52         isl_union_set *copy_in;
  53
  54         /* Array of statements */
  55         int n_stmts;
  56         struct gpu_stmt *stmts;
  57
  58         int n_array;
  59         struct gpu_array_info *array;
  60 };
  61
  62 enum ppcg_kernel_access_type {
  63         ppcg_access_global,
  64         ppcg_access_shared,
  65         ppcg_access_private
  66 };
  67
  68 enum ppcg_kernel_stmt_type {
  69         ppcg_kernel_copy,
  70         ppcg_kernel_domain,
  71         ppcg_kernel_sync
  72 };
  73
  74 /* Instance specific information about an access inside a kernel statement.
  75  *
  76  * type indicates whether it is a global, shared or private access
  77  * array is the original array information and may be NULL in case
  78  *      of an affine expression
  79  * local_array is a pointer to the appropriate element in the "array"
  80  *      array of the ppcg_kernel to which this access belongs.  It is
  81  *      NULL whenever array is NULL.
  82  * local_name is the name of the array or its local copy
  83  * index is the sequence of local index expressions
  84  */
  85 struct ppcg_kernel_access {
  86         enum ppcg_kernel_access_type type;
  87         struct gpu_array_info *array;
  88         struct gpu_local_array_info *local_array;
  89         char *local_name;
  90         isl_ast_expr_list *index;
  91 };
  92
  93 /* Representation of special statements, in particular copy statements
  94  * and __syncthreads statements, inside a kernel.
  95  *
  96  * type represents the kind of statement
  97  *
  98  *
  99  * for ppcg_kernel_copy statements we have
 100  *
 101  * read is set if the statement should copy data from global memory
 102  * to shared memory or registers.
 103  *
 104  * domain is the (parametric) domain of index and local_index
 105  *
 106  * index expresses the array element that needs to be copied in terms
 107  * of parameters
 108  * local_index expresses the corresponding element in the tile
 109  *
 110  * array refers to the original array being copied
 111  * local_array is a pointer to the appropriate element in the "array"
 112  *      array of the ppcg_kernel to which this copy access belongs
 113  *
 114  *
 115  * for ppcg_kernel_domain statements we have
 116  *
 117  * stmt is the corresponding input statement
 118  *
 119  * n_access is the number of accesses in stmt
 120  * access is an array of local information about the accesses
 121  */
 122 struct ppcg_kernel_stmt {
 123         enum ppcg_kernel_stmt_type type;
 124
 125         union {
 126                 struct {
 127                         int read;
 128                         isl_set *domain;
 129                         isl_pw_multi_aff *index;
 130                         isl_pw_multi_aff *local_index;
 131                         struct gpu_array_info *array;
 132                         struct gpu_local_array_info *local_array;
 133                 } c;
 134                 struct {
 135                         struct gpu_stmt *stmt;
 136
 137                         int n_access;
 138                         struct ppcg_kernel_access *access;
 139                 } d;
 140         } u;
 141 };
 142
 143 /* Representation of a local variable in a kernel.
 144  */
 145 struct ppcg_kernel_var {
 146         struct gpu_array_info *array;
 147         enum ppcg_kernel_access_type type;
 148         char *name;
 149         isl_vec *size;
 150 };
 151
 152 /* Representation of a kernel.
 153  *
 154  * id is the sequence number of the kernel.
 155  *
 156  * the first n_block elements of block_dim represent the size of the block.
 157  *
 158  * grid contains the values of the block ids.
 159  *
 160  * context is a parametric set containing the values of the parameters
 161  * for which this kernel may be run.
 162  *
 163  * arrays is the set of accessed array elements.
 164  *
 165  * space is the schedule space of the AST context.  That is, it represents
 166  * the loops of the generated host code containing the kernel launch.
 167  *
 168  * n_array is the total number of arrays in the input program and also
 169  * the number of element in the array array.
 170  * array contains information about each array that is local
 171  * to the current kernel.  If an array is not ussed in a kernel,
 172  * then the corresponding entry does not contain any information.
 173  */
 174 struct ppcg_kernel {
 175         int id;
 176
 177         int n_grid;
 178         int grid_dim[2];
 179
 180         int n_block;
 181         int block_dim[3];
 182
 183         isl_set *grid;
 184         isl_set *context;
 185
 186         isl_union_set *arrays;
 187
 188         isl_space *space;
 189
 190         int n_array;
 191         struct gpu_local_array_info *array;
 192
 193         int n_var;
 194         struct ppcg_kernel_var *var;
 195
 196         isl_ast_node *tree;
 197 };
 198
 199 int gpu_array_is_scalar(struct gpu_array_info *array);
 200 int gpu_array_is_read_only_scalar(struct gpu_array_info *array);
 201
 202 struct gpu_prog *gpu_prog_alloc(isl_ctx *ctx, struct pet_scop *scop);
 203 void gpu_prog_free(struct gpu_prog *prog);
 204
 205 __isl_give isl_set *add_context_from_str(__isl_take isl_set *set,
 206         const char *str);
 207
 208 __isl_give isl_ast_node *generate_gpu(isl_ctx *ctx, struct gpu_prog *prog,
 209         struct ppcg_options *options);
 210
 211 #endif