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 /* Represents an outer array possibly accessed by a gpu_prog.
  11  * If this outer array contains structures, then the references are not
  12  * collected and the reference groups are not computed.
  13  */
  14 struct gpu_array_info {
  15         /* The array data space. */
  16         isl_space *space;
  17         /* Element type. */
  18         char *type;
  19         /* Element size. */
  20         int size;
  21         /* Name of the array. */
  22         char *name;
  23         /* Extent of the array that needs to be copied. */
  24         isl_set *extent;
  25         /* Number of indices. */
  26         unsigned n_index;
  27         /* For each index, a bound on "extent" in that direction. */
  28         isl_pw_aff **bound;
  29
  30         /* All references to this array; point to elements of a linked list. */
  31         int n_ref;
  32         struct gpu_stmt_access **refs;
  33
  34         /* The reference groups associated to this array. */
  35         int n_group;
  36         struct gpu_array_ref_group **groups;
  37
  38         /* Is this a scalar that is read-only within the entire program? */
  39         int read_only_scalar;
  40
  41         /* Are the elements of the array structures? */
  42         int has_compound_element;
  43
  44         /* Is the array local to the scop? */
  45         int local;
  46
  47         /* Should the array be linearized? */
  48         int linearize;
  49
  50         /* Order dependences on this array.
  51          * Only used if live_range_reordering option is set.
  52          * It is set to NULL otherwise.
  53          */
  54         isl_union_map *dep_order;
  55         /* Should the array (scalar) be forcibly mapped to a register? */
  56         int force_private;
  57 };
  58
  59 /* For each index i with 0 <= i < n_index,
  60  * bound[i] is equal to array->bound[i] specialized to the current kernel.
  61  */
  62 struct gpu_local_array_info {
  63         unsigned n_index;
  64         isl_pw_aff_list *bound;
  65 };
  66
  67 __isl_give isl_ast_expr *gpu_local_array_info_linearize_index(
  68         struct gpu_local_array_info *array, __isl_take isl_ast_expr *expr);
  69
  70 /* A sequence of "n" names of types.
  71  */
  72 struct gpu_types {
  73         int n;
  74         char **name;
  75 };
  76
  77 /* "read" and "write" contain the original access relations, possibly
  78  * involving member accesses.
  79  *
  80  * The elements of "array", as well as the ranges of "copy_in" and "copy_out"
  81  * only refer to the outer arrays of any possible member accesses.
  82  */
  83 struct gpu_prog {
  84         isl_ctx *ctx;
  85
  86         struct ppcg_scop *scop;
  87
  88         /* Set of parameter values */
  89         isl_set *context;
  90
  91         /* All potential read accesses in the entire program */
  92         isl_union_map *read;
  93
  94         /* All potential write accesses in the entire program */
  95         isl_union_map *may_write;
  96         /* All definite write accesses in the entire program */
  97         isl_union_map *must_write;
  98
  99         /* Set of outer array elements that need to be copied in. */
 100         isl_union_set *copy_in;
 101         /* Set of outer array elements that need to be copied out. */
 102         isl_union_set *copy_out;
 103
 104         /* A mapping from all innermost arrays to their outer arrays. */
 105         isl_union_map *to_outer;
 106         /* A mapping from the outer arrays to all corresponding inner arrays. */
 107         isl_union_map *to_inner;
 108
 109         /* Order dependences on non-scalars. */
 110         isl_union_map *array_order;
 111
 112         /* Array of statements */
 113         int n_stmts;
 114         struct gpu_stmt *stmts;
 115
 116         int n_array;
 117         struct gpu_array_info *array;
 118 };
 119
 120 enum ppcg_kernel_access_type {
 121         ppcg_access_global,
 122         ppcg_access_shared,
 123         ppcg_access_private
 124 };
 125
 126 enum ppcg_kernel_stmt_type {
 127         ppcg_kernel_copy,
 128         ppcg_kernel_domain,
 129         ppcg_kernel_sync
 130 };
 131
 132 /* Representation of special statements, in particular copy statements
 133  * and __syncthreads statements, inside a kernel.
 134  *
 135  * type represents the kind of statement
 136  *
 137  *
 138  * for ppcg_kernel_copy statements we have
 139  *
 140  * read is set if the statement should copy data from global memory
 141  * to shared memory or registers.
 142  *
 143  * index expresses an access to the array element that needs to be copied
 144  * local_index expresses the corresponding element in the tile
 145  *
 146  * array refers to the original array being copied
 147  * local_array is a pointer to the appropriate element in the "array"
 148  *      array of the ppcg_kernel to which this copy access belongs
 149  *
 150  *
 151  * for ppcg_kernel_domain statements we have
 152  *
 153  * stmt is the corresponding input statement
 154  *
 155  * n_access is the number of accesses in stmt
 156  * access is an array of local information about the accesses
 157  */
 158 struct ppcg_kernel_stmt {
 159         enum ppcg_kernel_stmt_type type;
 160
 161         union {
 162                 struct {
 163                         int read;
 164                         isl_ast_expr *index;
 165                         isl_ast_expr *local_index;
 166                         struct gpu_array_info *array;
 167                         struct gpu_local_array_info *local_array;
 168                 } c;
 169                 struct {
 170                         struct gpu_stmt *stmt;
 171                         isl_id_to_ast_expr *ref2expr;
 172                 } d;
 173         } u;
 174 };
 175
 176 /* Representation of a local variable in a kernel.
 177  */
 178 struct ppcg_kernel_var {
 179         struct gpu_array_info *array;
 180         enum ppcg_kernel_access_type type;
 181         char *name;
 182         isl_vec *size;
 183 };
 184
 185 /* Representation of a kernel.
 186  *
 187  * id is the sequence number of the kernel.
 188  *
 189  * the first n_block elements of block_dim represent the effective size
 190  * of the block.
 191  *
 192  * grid_size reflects the effective grid size.
 193  *
 194  * context is a parametric set containing the values of the parameters
 195  * for which this kernel may be run.
 196  *
 197  * arrays is the set of possibly accessed outer array elements.
 198  *
 199  * space is the schedule space of the AST context.  That is, it represents
 200  * the loops of the generated host code containing the kernel launch.
 201  *
 202  * n_array is the total number of arrays in the input program and also
 203  * the number of element in the array array.
 204  * array contains information about each array that is local
 205  * to the current kernel.  If an array is not used in a kernel,
 206  * then the corresponding entry does not contain any information.
 207  */
 208 struct ppcg_kernel {
 209         int id;
 210
 211         int n_block;
 212         int block_dim[3];
 213
 214         isl_multi_pw_aff *grid_size;
 215         isl_set *context;
 216
 217         isl_union_set *arrays;
 218
 219         isl_space *space;
 220
 221         int n_array;
 222         struct gpu_local_array_info *array;
 223
 224         int n_var;
 225         struct ppcg_kernel_var *var;
 226
 227         isl_ast_node *tree;
 228 };
 229
 230 int gpu_array_is_scalar(struct gpu_array_info *array);
 231 int gpu_array_is_read_only_scalar(struct gpu_array_info *array);
 232 __isl_give isl_set *gpu_array_positive_size_guard(struct gpu_array_info *array);
 233
 234 struct gpu_prog *gpu_prog_alloc(isl_ctx *ctx, struct ppcg_scop *scop);
 235 void *gpu_prog_free(struct gpu_prog *prog);
 236
 237 int generate_gpu(isl_ctx *ctx, const char *input, FILE *out,
 238         struct ppcg_options *options,
 239         __isl_give isl_printer *(*print)(__isl_take isl_printer *p,
 240                 struct gpu_prog *prog, __isl_keep isl_ast_node *tree,
 241                 struct gpu_types *types, void *user), void *user);
 242
 243 #endif