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 array accessed at all by the program? */
  39         int accessed;
  40
  41         /* Is this a scalar that is read-only within the entire program? */
  42         int read_only_scalar;
  43
  44         /* Are the elements of the array structures? */
  45         int has_compound_element;
  46
  47         /* Is the array local to the scop? */
  48         int local;
  49
  50         /* Should the array be linearized? */
  51         int linearize;
  52
  53         /* Order dependences on this array.
  54          * Only used if live_range_reordering option is set.
  55          * It is set to NULL otherwise.
  56          */
  57         isl_union_map *dep_order;
  58         /* Should the array (scalar) be forcibly mapped to a register? */
  59         int force_private;
  60 };
  61
  62 /* For each index i with 0 <= i < n_index,
  63  * bound[i] is equal to array->bound[i] specialized to the current kernel.
  64  */
  65 struct gpu_local_array_info {
  66         unsigned n_index;
  67         isl_pw_aff_list *bound;
  68 };
  69
  70 __isl_give isl_ast_expr *gpu_local_array_info_linearize_index(
  71         struct gpu_local_array_info *array, __isl_take isl_ast_expr *expr);
  72
  73 /* A sequence of "n" names of types.
  74  */
  75 struct gpu_types {
  76         int n;
  77         char **name;
  78 };
  79
  80 /* "read" and "write" contain the original access relations, possibly
  81  * involving member accesses.
  82  *
  83  * The elements of "array", as well as the ranges of "copy_in" and "copy_out"
  84  * only refer to the outer arrays of any possible member accesses.
  85  */
  86 struct gpu_prog {
  87         isl_ctx *ctx;
  88
  89         struct ppcg_scop *scop;
  90
  91         /* Set of parameter values */
  92         isl_set *context;
  93
  94         /* All potential read accesses in the entire program */
  95         isl_union_map *read;
  96
  97         /* All potential write accesses in the entire program */
  98         isl_union_map *may_write;
  99         /* All definite write accesses in the entire program */
 100         isl_union_map *must_write;
 101         /* All tagged definite kills in the entire program */
 102         isl_union_map *tagged_must_kill;
 103
 104         /* The set of inner array elements that may be preserved. */
 105         isl_union_set *may_persist;
 106
 107         /* Set of outer array elements that need to be copied in. */
 108         isl_union_set *copy_in;
 109         /* Set of outer array elements that need to be copied out. */
 110         isl_union_set *copy_out;
 111
 112         /* A mapping from all innermost arrays to their outer arrays. */
 113         isl_union_map *to_outer;
 114         /* A mapping from the outer arrays to all corresponding inner arrays. */
 115         isl_union_map *to_inner;
 116         /* A mapping from all intermediate arrays to their outer arrays,
 117          * including an identity mapping from the anoymous 1D space to itself.
 118          */
 119         isl_union_map *any_to_outer;
 120
 121         /* Order dependences on non-scalars. */
 122         isl_union_map *array_order;
 123
 124         /* Array of statements */
 125         int n_stmts;
 126         struct gpu_stmt *stmts;
 127
 128         int n_array;
 129         struct gpu_array_info *array;
 130 };
 131
 132 enum ppcg_kernel_access_type {
 133         ppcg_access_global,
 134         ppcg_access_shared,
 135         ppcg_access_private
 136 };
 137
 138 enum ppcg_kernel_stmt_type {
 139         ppcg_kernel_copy,
 140         ppcg_kernel_domain,
 141         ppcg_kernel_sync
 142 };
 143
 144 /* Representation of special statements, in particular copy statements
 145  * and __syncthreads statements, inside a kernel.
 146  *
 147  * type represents the kind of statement
 148  *
 149  *
 150  * for ppcg_kernel_copy statements we have
 151  *
 152  * read is set if the statement should copy data from global memory
 153  * to shared memory or registers.
 154  *
 155  * index expresses an access to the array element that needs to be copied
 156  * local_index expresses the corresponding element in the tile
 157  *
 158  * array refers to the original array being copied
 159  * local_array is a pointer to the appropriate element in the "array"
 160  *      array of the ppcg_kernel to which this copy access belongs
 161  *
 162  *
 163  * for ppcg_kernel_domain statements we have
 164  *
 165  * stmt is the corresponding input statement
 166  *
 167  * n_access is the number of accesses in stmt
 168  * access is an array of local information about the accesses
 169  */
 170 struct ppcg_kernel_stmt {
 171         enum ppcg_kernel_stmt_type type;
 172
 173         union {
 174                 struct {
 175                         int read;
 176                         isl_ast_expr *index;
 177                         isl_ast_expr *local_index;
 178                         struct gpu_array_info *array;
 179                         struct gpu_local_array_info *local_array;
 180                 } c;
 181                 struct {
 182                         struct gpu_stmt *stmt;
 183                         isl_id_to_ast_expr *ref2expr;
 184                 } d;
 185         } u;
 186 };
 187
 188 /* Representation of a local variable in a kernel.
 189  */
 190 struct ppcg_kernel_var {
 191         struct gpu_array_info *array;
 192         enum ppcg_kernel_access_type type;
 193         char *name;
 194         isl_vec *size;
 195 };
 196
 197 /* Representation of a kernel.
 198  *
 199  * id is the sequence number of the kernel.
 200  *
 201  * block_ids contains the list of block identifiers for this kernel.
 202  * thread_ids contains the list of thread identifiers for this kernel.
 203  *
 204  * the first n_block elements of block_dim represent the effective size
 205  * of the block.
 206  *
 207  * grid_size reflects the effective grid size.
 208  *
 209  * context is a parametric set containing the values of the parameters
 210  * for which this kernel may be run.
 211  *
 212  * arrays is the set of possibly accessed outer array elements.
 213  *
 214  * space is the schedule space of the AST context.  That is, it represents
 215  * the loops of the generated host code containing the kernel launch.
 216  *
 217  * n_array is the total number of arrays in the input program and also
 218  * the number of element in the array array.
 219  * array contains information about each array that is local
 220  * to the current kernel.  If an array is not used in a kernel,
 221  * then the corresponding entry does not contain any information.
 222  */
 223 struct ppcg_kernel {
 224         int id;
 225
 226         isl_id_list *block_ids;
 227         isl_id_list *thread_ids;
 228
 229         int n_block;
 230         int block_dim[3];
 231
 232         isl_multi_pw_aff *grid_size;
 233         isl_set *context;
 234
 235         isl_union_set *arrays;
 236
 237         isl_space *space;
 238
 239         int n_array;
 240         struct gpu_local_array_info *array;
 241
 242         int n_var;
 243         struct ppcg_kernel_var *var;
 244
 245         isl_ast_node *tree;
 246 };
 247
 248 int gpu_array_is_scalar(struct gpu_array_info *array);
 249 int gpu_array_is_read_only_scalar(struct gpu_array_info *array);
 250 __isl_give isl_set *gpu_array_positive_size_guard(struct gpu_array_info *array);
 251
 252 struct gpu_prog *gpu_prog_alloc(isl_ctx *ctx, struct ppcg_scop *scop);
 253 void *gpu_prog_free(struct gpu_prog *prog);
 254
 255 int generate_gpu(isl_ctx *ctx, const char *input, FILE *out,
 256         struct ppcg_options *options,
 257         __isl_give isl_printer *(*print)(__isl_take isl_printer *p,
 258                 struct gpu_prog *prog, __isl_keep isl_ast_node *tree,
 259                 struct gpu_types *types, void *user), void *user);
 260
 261 #endif