5 #include <isl/id_to_ast_expr.h>
8 #include "ppcg_options.h"
10 /* Represents an outer array possibly accessed by a gpu_prog.
12 struct gpu_array_info
{
13 /* The array data space. */
19 /* Name of the array. */
21 /* Extent of the array that needs to be copied. */
23 /* Number of indices. */
25 /* For each index, a bound on "extent" in that direction. */
28 /* All references to this array; point to elements of a linked list. */
30 struct gpu_stmt_access
**refs
;
32 /* Is this array accessed at all by the program? */
35 /* Is this a scalar that is read-only within the entire program? */
38 /* Are the elements of the array structures? */
39 int has_compound_element
;
41 /* Is the array local to the scop? */
44 /* Should the array be linearized? */
47 /* Order dependences on this array.
48 * Only used if live_range_reordering option is set.
49 * It is set to NULL otherwise.
51 isl_union_map
*dep_order
;
54 /* Represents an outer array accessed by a ppcg_kernel, localized
55 * to the context of this kernel.
57 * "array" points to the corresponding array in the gpu_prog.
58 * The "n_group" "groups" are the reference groups associated to the array.
59 * If the outer array represented by the gpu_local_array_info
60 * contains structures, then the references are not
61 * collected and the reference groups are not computed.
62 * If "force_private" is set, then the array (in practice a scalar)
63 * must be mapped to a register.
64 * For each index i with 0 <= i < n_index,
65 * bound[i] is equal to array->bound[i] specialized to the current kernel.
67 struct gpu_local_array_info
{
68 struct gpu_array_info
*array
;
71 struct gpu_array_ref_group
**groups
;
76 isl_pw_aff_list
*bound
;
79 __isl_give isl_ast_expr
*gpu_local_array_info_linearize_index(
80 struct gpu_local_array_info
*array
, __isl_take isl_ast_expr
*expr
);
82 /* A sequence of "n" names of types.
89 /* "read" and "write" contain the original access relations, possibly
90 * involving member accesses.
92 * The elements of "array", as well as the ranges of "copy_in" and "copy_out"
93 * only refer to the outer arrays of any possible member accesses.
98 struct ppcg_scop
*scop
;
100 /* Set of parameter values */
103 /* All potential read accesses in the entire program */
106 /* All potential write accesses in the entire program */
107 isl_union_map
*may_write
;
108 /* All definite write accesses in the entire program */
109 isl_union_map
*must_write
;
110 /* All tagged definite kills in the entire program */
111 isl_union_map
*tagged_must_kill
;
113 /* The set of inner array elements that may be preserved. */
114 isl_union_set
*may_persist
;
116 /* Set of outer array elements that need to be copied in. */
117 isl_union_set
*copy_in
;
118 /* Set of outer array elements that need to be copied out. */
119 isl_union_set
*copy_out
;
121 /* A mapping from all innermost arrays to their outer arrays. */
122 isl_union_map
*to_outer
;
123 /* A mapping from the outer arrays to all corresponding inner arrays. */
124 isl_union_map
*to_inner
;
125 /* A mapping from all intermediate arrays to their outer arrays,
126 * including an identity mapping from the anoymous 1D space to itself.
128 isl_union_map
*any_to_outer
;
130 /* Order dependences on non-scalars. */
131 isl_union_map
*array_order
;
133 /* Array of statements */
135 struct gpu_stmt
*stmts
;
138 struct gpu_array_info
*array
;
143 struct ppcg_options
*options
;
145 /* Callback for printing of AST in appropriate format. */
146 __isl_give isl_printer
*(*print
)(__isl_take isl_printer
*p
,
147 struct gpu_prog
*prog
, __isl_keep isl_ast_node
*tree
,
148 struct gpu_types
*types
, void *user
);
151 struct gpu_prog
*prog
;
152 /* The generated AST. */
155 /* The sequence of types for which a definition has been printed. */
156 struct gpu_types types
;
158 /* User specified tile, grid and block sizes for each kernel */
159 isl_union_map
*sizes
;
161 /* Effectively used tile, grid and block sizes for each kernel */
162 isl_union_map
*used_sizes
;
164 /* Identifier of current kernel. */
166 /* Pointer to the current kernel. */
167 struct ppcg_kernel
*kernel
;
168 /* Does the computed schedule exhibit any parallelism? */
171 /* First tile dimension. */
173 /* Number of tile dimensions. */
175 /* Number of initial parallel loops among tile dimensions. */
178 /* Number of dimensions determining shared memory. */
181 /* Number of rows in the untiled schedule. */
183 /* Number of rows in the tiled schedule. */
185 /* Number of rows in schedule after tiling/wrapping over threads. */
186 int thread_tiled_len
;
188 /* A schedule tree corresponding to the host code. */
189 isl_schedule
*host_schedule
;
190 /* Global untiled schedule. */
191 isl_union_map
*sched
;
192 /* Local (per kernel launch) tiled schedule. */
193 isl_union_map
*tiled_sched
;
194 /* Local schedule per shared memory tile loop iteration. */
195 isl_union_map
*local_sched
;
197 /* Local tiled schedule projected onto the shared tile loops and
198 * the loops that will be wrapped over the threads,
199 * with all shared tile loops parametrized.
201 isl_union_map
*shared_sched
;
202 /* Projects out the loops that will be wrapped over the threads
205 isl_union_map
*shared_proj
;
207 /* A map that takes the range of shared_sched as input,
208 * wraps the appropriate loops over the threads and then projects
211 isl_map
*privatization
;
213 /* The array reference group corresponding to copy_sched. */
214 struct gpu_array_ref_group
*copy_group
;
216 /* First loop to unroll (or -1 if none) in the current part of the
223 /* Note: in the input file, the sizes of the grid and the blocks
224 * are specified in the order x, y, z, but internally, the sizes
225 * are stored in reverse order, so that the last element always
226 * refers to the x dimension.
233 enum ppcg_kernel_access_type
{
239 enum ppcg_kernel_stmt_type
{
245 /* Representation of special statements, in particular copy statements
246 * and __syncthreads statements, inside a kernel.
248 * type represents the kind of statement
251 * for ppcg_kernel_copy statements we have
253 * read is set if the statement should copy data from global memory
254 * to shared memory or registers.
256 * index expresses an access to the array element that needs to be copied
257 * local_index expresses the corresponding element in the tile
259 * array refers to the original array being copied
260 * local_array is a pointer to the appropriate element in the "array"
261 * array of the ppcg_kernel to which this copy access belongs
264 * for ppcg_kernel_domain statements we have
266 * stmt is the corresponding input statement
268 * n_access is the number of accesses in stmt
269 * access is an array of local information about the accesses
271 struct ppcg_kernel_stmt
{
272 enum ppcg_kernel_stmt_type type
;
278 isl_ast_expr
*local_index
;
279 struct gpu_array_info
*array
;
280 struct gpu_local_array_info
*local_array
;
283 struct gpu_stmt
*stmt
;
284 isl_id_to_ast_expr
*ref2expr
;
289 /* Representation of a local variable in a kernel.
291 struct ppcg_kernel_var
{
292 struct gpu_array_info
*array
;
293 enum ppcg_kernel_access_type type
;
298 /* Representation of a kernel.
300 * id is the sequence number of the kernel.
302 * block_ids contains the list of block identifiers for this kernel.
303 * thread_ids contains the list of thread identifiers for this kernel.
305 * the first n_block elements of block_dim represent the effective size
308 * grid_size reflects the effective grid size.
310 * context is a parametric set containing the values of the parameters
311 * for which this kernel may be run.
313 * arrays is the set of possibly accessed outer array elements.
315 * space is the schedule space of the AST context. That is, it represents
316 * the loops of the generated host code containing the kernel launch.
318 * n_array is the total number of arrays in the input program and also
319 * the number of element in the array array.
320 * array contains information about each array that is local
321 * to the current kernel. If an array is not used in a kernel,
322 * then the corresponding entry does not contain any information.
324 * any_force_private is set if any array in the kernel is marked force_private
328 struct ppcg_options
*options
;
332 isl_id_list
*block_ids
;
333 isl_id_list
*thread_ids
;
338 isl_multi_pw_aff
*grid_size
;
341 isl_union_set
*arrays
;
346 struct gpu_local_array_info
*array
;
349 struct ppcg_kernel_var
*var
;
351 int any_force_private
;
356 int gpu_array_is_scalar(struct gpu_array_info
*array
);
357 int gpu_array_is_read_only_scalar(struct gpu_array_info
*array
);
358 __isl_give isl_set
*gpu_array_positive_size_guard(struct gpu_array_info
*array
);
360 struct gpu_prog
*gpu_prog_alloc(isl_ctx
*ctx
, struct ppcg_scop
*scop
);
361 void *gpu_prog_free(struct gpu_prog
*prog
);
363 int generate_gpu(isl_ctx
*ctx
, const char *input
, FILE *out
,
364 struct ppcg_options
*options
,
365 __isl_give isl_printer
*(*print
)(__isl_take isl_printer
*p
,
366 struct gpu_prog
*prog
, __isl_keep isl_ast_node
*tree
,
367 struct gpu_types
*types
, void *user
), void *user
);