gpu.h

   1 #ifndef _GPU_H
   2 #define _GPU_H
   3
   4 #include <isl/ast.h>
   5 #include <isl/id_to_ast_expr.h>
   6
   7 #include "ppcg.h"
   8 #include "ppcg_options.h"
   9
  10 struct gpu_array_info {
  11         /* The array data space. */
  12         isl_space *space;
  13         /* Element type. */
  14         char *type;
  15         /* Element size. */
  16         int size;
  17         /* Name of the array. */
  18         char *name;
  19         /* Extent of the array that needs to be copied. */
  20         isl_set *extent;
  21         /* Number of indices. */
  22         unsigned n_index;
  23         /* For each index, a bound on "extent" in that direction. */
  24         isl_pw_aff **bound;
  25
  26         /* All references to this array; point to elements of a linked list. */
  27         int n_ref;
  28         struct gpu_stmt_access **refs;
  29
  30         /* The reference groups associated to this array. */
  31         int n_group;
  32         struct gpu_array_ref_group **groups;
  33
  34         /* Is this a scalar that is read-only within the entire program? */
  35         int read_only_scalar;
  36
  37         /* Is the array local to the scop? */
  38         int local;
  39 };
  40
  41 /* For each index i, array->bound[i] specialized to the current kernel. */
  42 struct gpu_local_array_info {
  43         isl_pw_aff_list *bound;
  44 };
  45
  46 __isl_give isl_ast_expr *gpu_local_array_info_linearize_index(
  47         struct gpu_local_array_info *array, __isl_take isl_ast_expr *expr);
  48
  49 struct gpu_prog {
  50         isl_ctx *ctx;
  51
  52         struct ppcg_scop *scop;
  53
  54         /* Set of parameter values */
  55         isl_set *context;
  56
  57         /* All read accesses in the entire program */
  58         isl_union_map *read;
  59
  60         /* All write accesses in the entire program */
  61         isl_union_map *write;
  62
  63         /* Set of array elements that need to be copied in. */
  64         isl_union_set *copy_in;
  65         /* Set of array elements that need to be copied out. */
  66         isl_union_set *copy_out;
  67
  68         /* Array of statements */
  69         int n_stmts;
  70         struct gpu_stmt *stmts;
  71
  72         int n_array;
  73         struct gpu_array_info *array;
  74 };
  75
  76 enum ppcg_kernel_access_type {
  77         ppcg_access_global,
  78         ppcg_access_shared,
  79         ppcg_access_private
  80 };
  81
  82 enum ppcg_kernel_stmt_type {
  83         ppcg_kernel_copy,
  84         ppcg_kernel_domain,
  85         ppcg_kernel_sync
  86 };
  87
  88 /* Representation of special statements, in particular copy statements
  89  * and __syncthreads statements, inside a kernel.
  90  *
  91  * type represents the kind of statement
  92  *
  93  *
  94  * for ppcg_kernel_copy statements we have
  95  *
  96  * read is set if the statement should copy data from global memory
  97  * to shared memory or registers.
  98  *
  99  * index expresses an access to the array element that needs to be copied
 100  * local_index expresses the corresponding element in the tile
 101  *
 102  * array refers to the original array being copied
 103  * local_array is a pointer to the appropriate element in the "array"
 104  *      array of the ppcg_kernel to which this copy access belongs
 105  *
 106  *
 107  * for ppcg_kernel_domain statements we have
 108  *
 109  * stmt is the corresponding input statement
 110  *
 111  * n_access is the number of accesses in stmt
 112  * access is an array of local information about the accesses
 113  */
 114 struct ppcg_kernel_stmt {
 115         enum ppcg_kernel_stmt_type type;
 116
 117         union {
 118                 struct {
 119                         int read;
 120                         isl_ast_expr *index;
 121                         isl_ast_expr *local_index;
 122                         struct gpu_array_info *array;
 123                         struct gpu_local_array_info *local_array;
 124                 } c;
 125                 struct {
 126                         struct gpu_stmt *stmt;
 127                         isl_id_to_ast_expr *ref2expr;
 128                 } d;
 129         } u;
 130 };
 131
 132 /* Representation of a local variable in a kernel.
 133  */
 134 struct ppcg_kernel_var {
 135         struct gpu_array_info *array;
 136         enum ppcg_kernel_access_type type;
 137         char *name;
 138         isl_vec *size;
 139 };
 140
 141 /* Representation of a kernel.
 142  *
 143  * id is the sequence number of the kernel.
 144  *
 145  * the first n_block elements of block_dim represent the effective size
 146  * of the block.
 147  *
 148  * grid_size reflects the effective grid size.
 149  *
 150  * context is a parametric set containing the values of the parameters
 151  * for which this kernel may be run.
 152  *
 153  * arrays is the set of accessed array elements.
 154  *
 155  * space is the schedule space of the AST context.  That is, it represents
 156  * the loops of the generated host code containing the kernel launch.
 157  *
 158  * n_array is the total number of arrays in the input program and also
 159  * the number of element in the array array.
 160  * array contains information about each array that is local
 161  * to the current kernel.  If an array is not used in a kernel,
 162  * then the corresponding entry does not contain any information.
 163  */
 164 struct ppcg_kernel {
 165         int id;
 166
 167         int n_block;
 168         int block_dim[3];
 169
 170         isl_multi_pw_aff *grid_size;
 171         isl_set *context;
 172
 173         isl_union_set *arrays;
 174
 175         isl_space *space;
 176
 177         int n_array;
 178         struct gpu_local_array_info *array;
 179
 180         int n_var;
 181         struct ppcg_kernel_var *var;
 182
 183         isl_ast_node *tree;
 184 };
 185
 186 int gpu_array_is_scalar(struct gpu_array_info *array);
 187 int gpu_array_is_read_only_scalar(struct gpu_array_info *array);
 188
 189 struct gpu_prog *gpu_prog_alloc(isl_ctx *ctx, struct ppcg_scop *scop);
 190 void *gpu_prog_free(struct gpu_prog *prog);
 191
 192 int generate_gpu(isl_ctx *ctx, const char *input, FILE *out,
 193         struct ppcg_options *options,
 194         __isl_give isl_printer *(*print)(__isl_take isl_printer *p,
 195                 struct gpu_prog *prog, __isl_keep isl_ast_node *tree,
 196                 void *user), void *user);
 197
 198 #endif