2 * Copyright 2010-2011 INRIA Saclay
3 * Copyright 2012-2013 Ecole Normale Superieure
5 * Use of this software is governed by the MIT license
7 * Written by Sven Verdoolaege, INRIA Saclay - Ile-de-France,
8 * Parc Club Orsay Universite, ZAC des vignes, 4 rue Jacques Monod,
10 * and Ecole Normale Superieure, 45 rue d’Ulm, 75230 Paris, France
17 #include <isl/polynomial.h>
18 #include <isl/union_set.h>
23 #include <isl/schedule.h>
24 #include <isl/options.h>
25 #include <isl/ast_build.h>
30 #include "ppcg_options.h"
33 /* The fields stride, shift and shift_map only contain valid information
35 * If so, they express that current index is such that if you add shift,
36 * then the result is always a multiple of stride.
37 * shift_map contains the mapping
39 * i -> (i + shift)/stride
41 * Let D represent the initial shared_len dimensions of the computed schedule.
42 * The spaces of "lb" and "shift" are of the form
46 * "shift_map" is of the form
48 * [D -> i] -> [D -> (i + shift(D))/stride]
50 struct gpu_array_bound
{
56 isl_basic_map
*shift_map
;
59 /* A tile of an array.
61 * n is the dimension of the array.
62 * bound is an array of size "n" representing the lower bound
63 * and size for each index.
65 * tiling maps a tile in the global array to the corresponding
66 * shared/private memory tile and is of the form
68 * { [D[i] -> A[a]] -> T[(a + shift(i))/stride - lb(i)] }
70 * where D represents the initial shared_len dimensions
71 * of the computed schedule.
73 struct gpu_array_tile
{
75 struct gpu_array_bound
*bound
;
76 isl_multi_aff
*tiling
;
79 struct gpu_array_info
;
81 /* A group of array references in a kernel that should be handled together.
82 * If private_tile is not NULL, then it is mapped to registers.
83 * Otherwise, if shared_tile is not NULL, it is mapped to shared memory.
84 * Otherwise, it is accessed from global memory.
86 struct gpu_array_ref_group
{
87 /* The references in this group access this array. */
88 struct gpu_array_info
*array
;
89 /* Position of this group in the list of reference groups of array. */
92 /* The following fields are use during the construction of the groups.
93 * access is the combined access relation relative to the shared
94 * memory tiling. In particular, the domain of the map corresponds
95 * to the first shared_len dimensions of the computed schedule.
96 * write is set if any access in the group is a write.
97 * exact_write is set if all writes are definite writes.
103 /* The shared memory tile, NULL if none. */
104 struct gpu_array_tile
*shared_tile
;
106 /* The private memory tile, NULL if none. */
107 struct gpu_array_tile
*private_tile
;
109 /* References in this group; point to elements of a linked list. */
111 struct gpu_stmt_access
**refs
;
113 /* Last shared memory tile dimension that affects tile of this group. */
119 struct ppcg_options
*options
;
121 /* Callback for printing of AST in appropriate format. */
122 __isl_give isl_printer
*(*print
)(__isl_take isl_printer
*p
,
123 struct gpu_prog
*prog
, __isl_keep isl_ast_node
*tree
,
124 struct gpu_types
*types
, void *user
);
127 struct gpu_prog
*prog
;
128 /* The generated AST. */
131 /* The sequence of types for which a definition has been printed. */
132 struct gpu_types types
;
134 /* tile, grid and block sizes for each kernel */
135 isl_union_map
*sizes
;
137 /* Identifier of current kernel. */
139 /* Pointer to the current kernel. */
140 struct ppcg_kernel
*kernel
;
141 /* Does the computed schedule exhibit any parallelism? */
144 /* First tile dimension. */
146 /* Number of tile dimensions. */
148 /* Number of initial parallel loops among tile dimensions. */
151 /* Number of dimensions determining shared memory. */
154 /* Number of rows in the untiled schedule. */
156 /* Number of rows in the tiled schedule. */
158 /* Number of rows in schedule after tiling/wrapping over threads. */
159 int thread_tiled_len
;
161 /* Global untiled schedule. */
162 isl_union_map
*sched
;
163 /* Local (per kernel launch) tiled schedule. */
164 isl_union_map
*tiled_sched
;
165 /* Local schedule per shared memory tile loop iteration. */
166 isl_union_map
*local_sched
;
168 /* Local tiled schedule projected onto the shared tile loops and
169 * the loops that will be wrapped over the threads,
170 * with all shared tile loops parametrized.
172 isl_union_map
*shared_sched
;
173 /* Projects out the loops that will be wrapped over the threads
176 isl_union_map
*shared_proj
;
178 /* A map that takes the range of shared_sched as input,
179 * wraps the appropriate loops over the threads and then projects
182 isl_map
*privatization
;
184 /* A map from the shared memory tile loops and the thread indices
185 * (as parameters) to the set of accessed memory elements that
186 * will be accessed through private copies.
188 isl_union_map
*private_access
;
190 /* The schedule for the current private/shared access
191 * (within print_private_access or print_shared_access).
194 /* The array reference group corresponding to copy_sched. */
195 struct gpu_array_ref_group
*copy_group
;
197 /* First loop to unroll (or -1 if none) in the current part of the
204 /* Note: in the input file, the sizes of the grid and the blocks
205 * are specified in the order x, y, z, but internally, the sizes
206 * are stored in reverse order, so that the last element always
207 * refers to the x dimension.
214 /* Print the name of the local copy of a given group of array references.
216 static __isl_give isl_printer
*print_array_name(__isl_take isl_printer
*p
,
217 struct gpu_array_ref_group
*group
)
221 if (group
->private_tile
)
222 p
= isl_printer_print_str(p
, "private_");
223 else if (group
->shared_tile
)
224 p
= isl_printer_print_str(p
, "shared_");
227 p
= isl_printer_print_str(p
, group
->array
->name
);
228 if (!global
&& group
->array
->n_group
> 1) {
229 p
= isl_printer_print_str(p
, "_");
230 p
= isl_printer_print_int(p
, group
->nr
);
236 /* Collect all references to the given array and store pointers to them
239 * If the array contains structures, then there is no need to collect
240 * the references since we will not be computing any reference groups.
242 static void collect_references(struct gpu_prog
*prog
,
243 struct gpu_array_info
*array
)
248 if (array
->has_compound_element
)
252 for (i
= 0; i
< prog
->n_stmts
; ++i
) {
253 struct gpu_stmt
*stmt
= &prog
->stmts
[i
];
254 struct gpu_stmt_access
*access
;
256 for (access
= stmt
->accesses
; access
; access
= access
->next
) {
258 name
= isl_map_get_tuple_name(access
->access
,
260 if (name
&& !strcmp(array
->name
, name
))
266 array
->refs
= isl_alloc_array(prog
->ctx
, struct gpu_stmt_access
*, n
);
270 for (i
= 0; i
< prog
->n_stmts
; ++i
) {
271 struct gpu_stmt
*stmt
= &prog
->stmts
[i
];
272 struct gpu_stmt_access
*access
;
274 for (access
= stmt
->accesses
; access
; access
= access
->next
) {
276 name
= isl_map_get_tuple_name(access
->access
,
278 if (!name
|| strcmp(array
->name
, name
))
281 array
->refs
[n
++] = access
;
286 /* Create a gpu_array_tile for an array of dimension "n_index".
288 static struct gpu_array_tile
*create_tile(isl_ctx
*ctx
, int n_index
)
291 struct gpu_array_tile
*tile
;
293 tile
= isl_calloc_type(ctx
, struct gpu_array_tile
);
298 tile
->bound
= isl_alloc_array(ctx
, struct gpu_array_bound
, n_index
);
301 for (i
= 0; i
< n_index
; ++i
) {
302 tile
->bound
[i
].size
= NULL
;
303 tile
->bound
[i
].lb
= NULL
;
304 tile
->bound
[i
].stride
= NULL
;
305 tile
->bound
[i
].shift
= NULL
;
306 tile
->bound
[i
].shift_map
= NULL
;
312 static void *free_tile(struct gpu_array_tile
*tile
)
319 for (j
= 0; j
< tile
->n
; ++j
) {
320 isl_val_free(tile
->bound
[j
].size
);
321 isl_val_free(tile
->bound
[j
].stride
);
322 isl_aff_free(tile
->bound
[j
].lb
);
323 isl_aff_free(tile
->bound
[j
].shift
);
324 isl_basic_map_free(tile
->bound
[j
].shift_map
);
327 isl_multi_aff_free(tile
->tiling
);
333 static struct pet_array
*find_array(struct ppcg_scop
*scop
,
334 __isl_keep isl_set
*accessed
)
339 id
= isl_set_get_tuple_id(accessed
);
341 for (i
= 0; i
< scop
->n_array
; ++i
) {
344 id_i
= isl_set_get_tuple_id(scop
->arrays
[i
]->extent
);
351 return i
< scop
->n_array
? scop
->arrays
[i
] : NULL
;
354 /* Compute and return the extent of "array", taking into account the set of
357 * In particular, the extent in the outer dimension is taken
358 * from "accessed", while then extent in the remaing dimensions
359 * are taken from array->extent.
361 * The extent in the outer dimension cannot be taken from array->extent
362 * because that may be unbounded. Furthermore, even if it is bounded,
363 * it may be larger than the piece of the array that is being accessed.
365 static __isl_give isl_set
*compute_extent(struct pet_array
*array
,
366 __isl_keep isl_set
*accessed
)
373 extent
= isl_set_copy(array
->extent
);
375 n_index
= isl_set_dim(accessed
, isl_dim_set
);
379 extent
= isl_set_project_out(extent
, isl_dim_set
, 0, 1);
380 outer
= isl_set_copy(accessed
);
381 outer
= isl_set_project_out(outer
, isl_dim_set
, 1, n_index
- 1);
382 extent
= isl_set_flat_product(outer
, extent
);
383 id
= isl_set_get_tuple_id(accessed
);
384 extent
= isl_set_set_tuple_id(extent
, id
);
389 /* Is the array "array" being extracted a read-only scalar?
391 * That is, is "array" a scalar that is never possibly written to.
392 * An array containing structures is never considered to be a scalar.
394 static int is_read_only_scalar(struct gpu_array_info
*array
,
395 struct gpu_prog
*prog
)
398 isl_union_map
*write
;
401 if (array
->has_compound_element
)
403 if (array
->n_index
!= 0)
406 write
= isl_union_map_copy(prog
->may_write
);
407 space
= isl_set_universe(isl_space_copy(array
->space
));
408 write
= isl_union_map_intersect_range(write
,
409 isl_union_set_from_set(space
));
410 empty
= isl_union_map_is_empty(write
);
411 isl_union_map_free(write
);
416 /* Compute bounds on the host arrays based on the accessed elements
417 * and collect all references to the array.
419 * If the array is zero-dimensional and does not contain structures,
420 * i.e., if the array is a scalar, we check whether it is read-only.
422 static int extract_array_info(__isl_take isl_set
*array
, void *user
)
425 struct gpu_prog
*prog
= (struct gpu_prog
*)user
;
429 struct pet_array
*pa
;
430 struct gpu_array_info
*info
;
433 info
= &prog
->array
[prog
->n_array
];
436 n_index
= isl_set_dim(array
, isl_dim_set
);
437 name
= isl_set_get_tuple_name(array
);
438 bounds
= isl_alloc_array(isl_set_get_ctx(array
),
439 isl_pw_aff
*, n_index
);
443 info
->space
= isl_set_get_space(array
);
444 info
->name
= strdup(name
);
445 info
->n_index
= n_index
;
446 info
->bound
= bounds
;
447 info
->linearize
= prog
->scop
->options
->linearize_device_arrays
;
449 pa
= find_array(prog
->scop
, array
);
451 isl_die(isl_set_get_ctx(array
), isl_error_internal
,
452 "unable to find array in scop", goto error
);
454 info
->type
= strdup(pa
->element_type
);
455 info
->size
= pa
->element_size
;
456 info
->local
= pa
->declared
&& !pa
->exposed
;
457 info
->has_compound_element
= pa
->element_is_record
;
458 info
->read_only_scalar
= is_read_only_scalar(info
, prog
);
460 extent
= compute_extent(pa
, array
);
461 for (i
= 0; i
< n_index
; ++i
) {
467 bound
= isl_set_dim_max(isl_set_copy(extent
), i
);
469 dom
= isl_pw_aff_domain(isl_pw_aff_copy(bound
));
470 ls
= isl_local_space_from_space(isl_set_get_space(dom
));
471 one
= isl_aff_zero_on_domain(ls
);
472 one
= isl_aff_add_constant_si(one
, 1);
473 bound
= isl_pw_aff_add(bound
, isl_pw_aff_alloc(dom
, one
));
474 bound
= isl_pw_aff_gist(bound
, isl_set_copy(prog
->context
));
477 if (!isl_pw_aff_is_cst(bound
))
480 info
->extent
= extent
;
482 collect_references(prog
, info
);
491 /* Compute a mapping from all outer arrays (of structs) in scop
492 * to their innermost arrays.
494 * In particular, for each array of a primitive type, the result
495 * contains the identity mapping on that array.
496 * For each array involving member accesses, the result
497 * contains a mapping from the elements of the outer array of structs
498 * to all corresponding elements of the innermost nested arrays.
500 static __isl_give isl_union_map
*compute_to_inner(struct ppcg_scop
*scop
)
503 isl_union_map
*to_inner
;
505 to_inner
= isl_union_map_empty(isl_set_get_space(scop
->context
));
507 for (i
= 0; i
< scop
->n_array
; ++i
) {
508 struct pet_array
*array
= scop
->arrays
[i
];
512 if (array
->element_is_record
)
515 set
= isl_set_copy(array
->extent
);
516 map
= isl_set_identity(isl_set_copy(set
));
518 while (set
&& isl_set_is_wrapping(set
)) {
522 id
= isl_set_get_tuple_id(set
);
523 wrapped
= isl_set_unwrap(set
);
524 wrapped
= isl_map_domain_map(wrapped
);
525 wrapped
= isl_map_set_tuple_id(wrapped
, isl_dim_in
, id
);
526 map
= isl_map_apply_domain(map
, wrapped
);
527 set
= isl_map_domain(isl_map_copy(map
));
530 map
= isl_map_gist_domain(map
, set
);
532 to_inner
= isl_union_map_add_map(to_inner
, map
);
538 /* Construct a gpu_array_info for each array possibly accessed by "prog" and
539 * collect them in prog->array.
541 * If there are any member accesses involved, then they are first mapped
542 * to the outer arrays of structs.
544 static int collect_array_info(struct gpu_prog
*prog
)
547 isl_union_set
*arrays
;
549 arrays
= isl_union_map_range(isl_union_map_copy(prog
->read
));
550 arrays
= isl_union_set_union(arrays
,
551 isl_union_map_range(isl_union_map_copy(prog
->may_write
)));
553 arrays
= isl_union_set_apply(arrays
,
554 isl_union_map_copy(prog
->to_outer
));
556 arrays
= isl_union_set_coalesce(arrays
);
558 prog
->n_array
= isl_union_set_n_set(arrays
);
559 prog
->array
= isl_calloc_array(prog
->ctx
,
560 struct gpu_array_info
, prog
->n_array
);
563 r
= isl_union_set_foreach_set(arrays
, &extract_array_info
, prog
);
564 isl_union_set_free(arrays
);
569 static void free_array_info(struct gpu_prog
*prog
)
573 for (i
= 0; i
< prog
->n_array
; ++i
) {
574 int n_index
= prog
->array
[i
].n_index
;
575 free(prog
->array
[i
].type
);
576 free(prog
->array
[i
].name
);
577 for (j
= 0; j
< n_index
; ++j
)
578 isl_pw_aff_free(prog
->array
[i
].bound
[j
]);
579 isl_space_free(prog
->array
[i
].space
);
580 isl_set_free(prog
->array
[i
].extent
);
581 free(prog
->array
[i
].bound
);
582 free(prog
->array
[i
].refs
);
587 /* Check if a gpu array is a scalar. A scalar is a value that is not stored
588 * as an array or through a pointer reference, but as a single data element.
589 * At the moment, scalars are represented as zero-dimensional arrays.
590 * A zero-dimensional array containing structures is not considered
593 int gpu_array_is_scalar(struct gpu_array_info
*array
)
595 return !array
->has_compound_element
&& array
->n_index
== 0;
598 /* Is "array" a read-only scalar?
600 int gpu_array_is_read_only_scalar(struct gpu_array_info
*array
)
602 return array
->read_only_scalar
;
605 /* Internal data structure for extract_size_of_type.
606 * "type" specifies the name of the space that we want to extract.
607 * "res" is used to store the subset of that space.
609 struct ppcg_extract_size_data
{
614 /* This function is called for each set in a union_set.
615 * If the name of the set matches data->type, we store the
618 static int extract_size_of_type(__isl_take isl_set
*size
, void *user
)
620 struct ppcg_extract_size_data
*data
= user
;
623 name
= isl_set_get_tuple_name(size
);
624 if (name
&& !strcmp(name
, data
->type
)) {
633 /* Given a union map { kernel[i] -> *[...] },
634 * return the range in the space called "type" for the kernel with
635 * sequence number "id".
637 static __isl_give isl_set
*extract_sizes(__isl_keep isl_union_map
*sizes
,
638 const char *type
, int id
)
642 isl_union_set
*local_sizes
;
643 struct ppcg_extract_size_data data
= { type
, NULL
};
648 space
= isl_union_map_get_space(sizes
);
649 space
= isl_space_set_from_params(space
);
650 space
= isl_space_add_dims(space
, isl_dim_set
, 1);
651 space
= isl_space_set_tuple_name(space
, isl_dim_set
, "kernel");
652 dom
= isl_set_universe(space
);
653 dom
= isl_set_fix_si(dom
, isl_dim_set
, 0, id
);
655 local_sizes
= isl_union_set_apply(isl_union_set_from_set(dom
),
656 isl_union_map_copy(sizes
));
657 isl_union_set_foreach_set(local_sizes
, &extract_size_of_type
, &data
);
658 isl_union_set_free(local_sizes
);
662 /* Given a singleton set, extract the first (at most *len) elements
663 * of the single integer tuple into *sizes and update *len if needed.
665 static void read_sizes_from_set(__isl_take isl_set
*set
, int *sizes
, int *len
)
673 dim
= isl_set_dim(set
, isl_dim_set
);
677 for (i
= 0; i
< *len
; ++i
) {
680 v
= isl_set_plain_get_val_if_fixed(set
, isl_dim_set
, i
);
683 sizes
[i
] = isl_val_get_num_si(v
);
690 /* Extract user specified "tile" sizes from the "sizes" command line option,
691 * defaulting to option->tile_size in each dimension.
693 static void read_tile_sizes(struct gpu_gen
*gen
)
698 gen
->tile_size
= isl_alloc_array(gen
->ctx
, int, gen
->tile_len
);
699 assert(gen
->tile_size
);
700 for (n
= 0; n
< gen
->tile_len
; ++n
)
701 gen
->tile_size
[n
] = gen
->options
->tile_size
;
703 size
= extract_sizes(gen
->sizes
, "tile", gen
->kernel_id
);
704 read_sizes_from_set(size
, gen
->tile_size
, &gen
->tile_len
);
706 if (gen
->n_parallel
> gen
->tile_len
)
707 gen
->n_parallel
= gen
->tile_len
;
710 /* Extract user specified "block" sizes from the "sizes" command line option,
711 * after filling in some potentially useful defaults.
713 static void read_block_sizes(struct gpu_gen
*gen
)
719 gen
->n_block
= (n
<= 3) ? n
: 3;
720 switch (gen
->n_block
) {
722 gen
->block_dim
[0] = 512;
725 gen
->block_dim
[0] = 32;
726 gen
->block_dim
[1] = 16;
729 gen
->block_dim
[0] = 32;
730 gen
->block_dim
[1] = 4;
731 gen
->block_dim
[2] = 4;
735 size
= extract_sizes(gen
->sizes
, "block", gen
->kernel_id
);
736 read_sizes_from_set(size
, gen
->block_dim
, &gen
->n_block
);
739 /* Extract user specified "grid" sizes from the "sizes" command line option,
740 * after filling in some potentially useful defaults.
742 static void read_grid_sizes(struct gpu_gen
*gen
)
744 int n
= gen
->n_parallel
;
747 gen
->n_grid
= (n
<= 2) ? n
: 2;
748 switch (gen
->n_grid
) {
750 gen
->grid_dim
[0] = 32768;
753 gen
->grid_dim
[0] = 256;
754 gen
->grid_dim
[1] = 256;
758 size
= extract_sizes(gen
->sizes
, "grid", gen
->kernel_id
);
759 read_sizes_from_set(size
, gen
->grid_dim
, &gen
->n_grid
);
762 /* Extract user specified sizes from the "sizes" command line option
763 * after filling in some potentially useful defaults.
765 static void read_sizes(struct gpu_gen
*gen
)
767 read_tile_sizes(gen
);
768 read_block_sizes(gen
);
769 read_grid_sizes(gen
);
772 static void *free_stmts(struct gpu_stmt
*stmts
, int n
)
779 for (i
= 0; i
< n
; ++i
) {
780 struct gpu_stmt_access
*access
, *next
;
782 for (access
= stmts
[i
].accesses
; access
; access
= next
) {
784 isl_id_free(access
->ref_id
);
785 isl_map_free(access
->access
);
786 isl_map_free(access
->tagged_access
);
790 isl_id_free(stmts
[i
].id
);
797 /* Construct a map from a domain of dimensionality "len"
798 * to a domain of dimensionality "len" + "tile_len" that tiles
799 * the "tile_len" coordinates starting at "first".
800 * In particular, [s_i] -> [s_i / tile_size[i], s_i % tile_size[i]].
801 * "dim" prescribes the parameters.
803 static __isl_give isl_map
*tile(__isl_take isl_space
*dim
, int len
,
804 int first
, int tile_len
, int *tile_size
)
811 dim
= isl_space_add_dims(dim
, isl_dim_in
, len
);
812 dim
= isl_space_add_dims(dim
, isl_dim_out
, len
+ tile_len
);
813 bmap
= isl_basic_map_universe(isl_space_copy(dim
));
814 ls
= isl_local_space_from_space(dim
);
816 for (i
= 0; i
< len
- tile_len
; ++i
) {
817 int j
= i
< first
? i
: i
+ tile_len
;
818 int k
= i
< first
? i
: i
+ 2 * tile_len
;
820 c
= isl_equality_alloc(isl_local_space_copy(ls
));
821 c
= isl_constraint_set_coefficient_si(c
, isl_dim_in
, j
, -1);
822 c
= isl_constraint_set_coefficient_si(c
, isl_dim_out
, k
, 1);
823 bmap
= isl_basic_map_add_constraint(bmap
, c
);
826 for (i
= 0; i
< tile_len
; ++i
) {
827 c
= isl_equality_alloc(isl_local_space_copy(ls
));
828 c
= isl_constraint_set_coefficient_si(c
, isl_dim_in
,
830 c
= isl_constraint_set_coefficient_si(c
, isl_dim_out
,
831 first
+ i
, tile_size
[i
]);
832 c
= isl_constraint_set_coefficient_si(c
, isl_dim_out
,
833 first
+ i
+ tile_len
, 1);
834 bmap
= isl_basic_map_add_constraint(bmap
, c
);
836 c
= isl_inequality_alloc(isl_local_space_copy(ls
));
837 c
= isl_constraint_set_coefficient_si(c
, isl_dim_out
,
838 first
+ i
+ tile_len
, 1);
839 bmap
= isl_basic_map_add_constraint(bmap
, c
);
841 c
= isl_inequality_alloc(isl_local_space_copy(ls
));
842 c
= isl_constraint_set_coefficient_si(c
, isl_dim_out
,
843 first
+ i
+ tile_len
, -1);
844 c
= isl_constraint_set_constant_si(c
, tile_size
[i
] - 1);
845 bmap
= isl_basic_map_add_constraint(bmap
, c
);
848 isl_local_space_free(ls
);
850 return isl_map_from_basic_map(bmap
);
853 /* Construct a map from a domain of dimensionality "len"
854 * to a domain of dimensionality "len" + "wrap_len" that "wraps"
855 * the "wrap_len" coordinates starting at "first" according to "wrap_size".
856 * In particular, [s_i] -> [s_i, s_i % wrap_size[i]].
857 * To do so, we need extra variables corresponding to [s_i / wrap_size[i]],
858 * that are projected out at the end.
859 * "dim" prescribes the parameters.
861 static __isl_give isl_map
*wrap(__isl_take isl_space
*dim
, int len
,
862 int first
, int wrap_len
, int *wrap_size
)
869 dim
= isl_space_add_dims(dim
, isl_dim_in
, len
);
870 dim
= isl_space_add_dims(dim
, isl_dim_out
, len
+ 2 * wrap_len
);
871 bmap
= isl_basic_map_universe(isl_space_copy(dim
));
872 ls
= isl_local_space_from_space(dim
);
874 for (i
= 0; i
< len
; ++i
) {
875 int k
= i
< first
+ wrap_len
? i
: i
+ 2 * wrap_len
;
877 c
= isl_equality_alloc(isl_local_space_copy(ls
));
878 c
= isl_constraint_set_coefficient_si(c
, isl_dim_in
, i
, -1);
879 c
= isl_constraint_set_coefficient_si(c
, isl_dim_out
, k
, 1);
880 bmap
= isl_basic_map_add_constraint(bmap
, c
);
883 for (i
= 0; i
< wrap_len
; ++i
) {
884 c
= isl_equality_alloc(isl_local_space_copy(ls
));
885 c
= isl_constraint_set_coefficient_si(c
, isl_dim_out
,
887 c
= isl_constraint_set_coefficient_si(c
, isl_dim_out
,
888 first
+ wrap_len
+ i
, 1);
889 c
= isl_constraint_set_coefficient_si(c
, isl_dim_out
,
890 first
+ 2 * wrap_len
+ i
, wrap_size
[i
]);
891 bmap
= isl_basic_map_add_constraint(bmap
, c
);
893 c
= isl_inequality_alloc(isl_local_space_copy(ls
));
894 c
= isl_constraint_set_coefficient_si(c
, isl_dim_out
,
895 first
+ wrap_len
+ i
, 1);
896 bmap
= isl_basic_map_add_constraint(bmap
, c
);
898 c
= isl_inequality_alloc(isl_local_space_copy(ls
));
899 c
= isl_constraint_set_coefficient_si(c
, isl_dim_out
,
900 first
+ wrap_len
+ i
, -1);
901 c
= isl_constraint_set_constant_si(c
, wrap_size
[i
] - 1);
902 bmap
= isl_basic_map_add_constraint(bmap
, c
);
905 isl_local_space_free(ls
);
907 bmap
= isl_basic_map_project_out(bmap
, isl_dim_out
,
908 first
+ 2 * wrap_len
, wrap_len
);
910 return isl_map_from_basic_map(bmap
);
913 /* Add "n" parameters named prefix%d.
915 static __isl_give isl_set
*add_params( __isl_take isl_set
*set
,
916 int n
, const char *prefix
)
922 nparam
= isl_set_dim(set
, isl_dim_param
);
923 set
= isl_set_add_dims(set
, isl_dim_param
, n
);
925 for (i
= 0; i
< n
; ++i
) {
926 snprintf(name
, sizeof(name
), "%s%d", prefix
, i
);
927 set
= isl_set_set_dim_name(set
, isl_dim_param
,
934 /* Equate the "n" dimensions of "set" starting at "first" to
935 * freshly created parameters named prefix%d.
937 static __isl_give isl_set
*parametrize(__isl_take isl_set
*set
,
938 int first
, int n
, const char *prefix
)
943 nparam
= isl_set_dim(set
, isl_dim_param
);
945 set
= add_params(set
, n
, prefix
);
947 for (i
= 0; i
< n
; ++i
)
948 set
= isl_set_equate(set
, isl_dim_param
, nparam
+ i
,
949 isl_dim_set
, first
+ i
);
954 /* Given a parameter space "space", create a set of dimension "len"
955 * of which the "n" dimensions starting at "first" are equated to
956 * freshly created parameters named prefix%d.
958 static __isl_give isl_set
*parametrization(__isl_take isl_space
*space
,
959 int len
, int first
, int n
, const char *prefix
)
963 space
= isl_space_set_from_params(space
);
964 space
= isl_space_add_dims(space
, isl_dim_set
, len
);
965 set
= isl_set_universe(space
);
967 return parametrize(set
, first
, n
, prefix
);
970 /* Tile the B loops over the tile sizes and then tile/wrap
971 * the T1 loops over the blocks.
973 static __isl_give isl_union_map
*tile_schedule(struct gpu_gen
*gen
,
974 __isl_take isl_union_map
*sched
)
977 isl_map
*tiling
, *block_tiling
;
979 dim
= isl_union_map_get_space(sched
);
980 tiling
= tile(isl_space_copy(dim
), gen
->untiled_len
,
981 gen
->tile_first
, gen
->tile_len
, gen
->tile_size
);
983 if (gen
->options
->wrap
)
984 block_tiling
= wrap(dim
, gen
->untiled_len
+ gen
->tile_len
,
985 gen
->tile_first
, gen
->n_grid
, gen
->grid_dim
);
987 block_tiling
= tile(dim
, gen
->untiled_len
+ gen
->tile_len
,
988 gen
->tile_first
, gen
->n_grid
, gen
->grid_dim
);
990 gen
->tiled_len
= gen
->untiled_len
+ gen
->tile_len
+ gen
->n_grid
;
992 tiling
= isl_map_apply_range(tiling
, block_tiling
);
994 sched
= isl_union_map_apply_range(sched
,
995 isl_union_map_from_map(tiling
));
997 gen
->shared_len
= gen
->tile_first
+ gen
->tile_len
+ gen
->n_grid
;
1002 /* Equate the "T1P" iterators in the tiled schedule "sched"
1003 * to the block dimensions.
1005 static __isl_give isl_union_map
*parametrize_tiled_schedule(
1006 struct gpu_gen
*gen
, __isl_take isl_union_map
*sched
)
1011 dim
= isl_union_map_get_space(sched
);
1012 par
= parametrization(dim
, gen
->tiled_len
,
1013 gen
->tile_first
+ gen
->n_grid
, gen
->n_grid
, "b");
1014 sched
= isl_union_map_intersect_range(sched
,
1015 isl_union_set_from_set(par
));
1020 /* Tile/wrap the P1 loops over the threads.
1022 static __isl_give isl_union_map
*thread_tile_schedule(struct gpu_gen
*gen
,
1023 __isl_take isl_union_map
*sched
)
1029 dim
= isl_union_map_get_space(sched
);
1031 if (gen
->options
->wrap
)
1032 tiling
= wrap(isl_space_copy(dim
), gen
->tiled_len
,
1033 gen
->shared_len
, gen
->n_block
, gen
->block_dim
);
1035 tiling
= tile(isl_space_copy(dim
), gen
->tiled_len
,
1036 gen
->shared_len
, gen
->n_block
, gen
->block_dim
);
1037 gen
->thread_tiled_len
= gen
->tiled_len
+ gen
->n_block
;
1039 sched
= isl_union_map_apply_range(sched
,
1040 isl_union_map_from_map(tiling
));
1042 par
= parametrization(dim
, gen
->thread_tiled_len
,
1043 gen
->tile_first
+ gen
->tile_len
+ gen
->n_grid
+ gen
->n_block
,
1045 sched
= isl_union_map_intersect_range(sched
,
1046 isl_union_set_from_set(par
));
1048 gen
->shared_len
= gen
->tile_first
+ gen
->tile_len
+ gen
->n_grid
;
1053 /* If the user asked for it, scale the shared memory tile loops
1054 * (T1T and T2) of "sched" by gen->tile_size[i].
1055 * If we are not performing "wrapping", then additionally scale the T1P
1056 * loops by gen->grid_dim[i].
1058 static __isl_give isl_union_map
*scale_tile_loops(struct gpu_gen
*gen
,
1059 __isl_take isl_union_map
*sched
)
1063 isl_basic_map
*scale
;
1065 isl_local_space
*ls
;
1067 if (!gen
->options
->scale_tile_loops
)
1070 dim
= isl_union_map_get_space(sched
);
1071 dim
= isl_space_add_dims(dim
, isl_dim_in
, gen
->tiled_len
);
1072 dim
= isl_space_add_dims(dim
, isl_dim_out
, gen
->tiled_len
);
1073 scale
= isl_basic_map_universe(isl_space_copy(dim
));
1074 ls
= isl_local_space_from_space(dim
);
1076 for (i
= 0; i
< gen
->tiled_len
; ++i
) {
1079 if (i
>= gen
->tile_first
&& i
< gen
->tile_first
+ gen
->n_grid
) {
1080 f
= gen
->tile_size
[i
- gen
->tile_first
];
1081 if (!gen
->options
->wrap
)
1082 f
*= gen
->grid_dim
[i
- gen
->tile_first
];
1083 } else if (i
>= gen
->tile_first
+ gen
->n_grid
&&
1084 i
< gen
->tile_first
+ gen
->n_grid
+ gen
->tile_len
) {
1085 f
= gen
->tile_size
[i
- (gen
->tile_first
+ gen
->n_grid
)];
1088 c
= isl_equality_alloc(isl_local_space_copy(ls
));
1089 c
= isl_constraint_set_coefficient_si(c
, isl_dim_in
, i
, f
);
1090 c
= isl_constraint_set_coefficient_si(c
, isl_dim_out
, i
, -1);
1091 scale
= isl_basic_map_add_constraint(scale
, c
);
1094 isl_local_space_free(ls
);
1096 sched
= isl_union_map_apply_range(sched
,
1097 isl_union_map_from_map(isl_map_from_basic_map(scale
)));
1102 /* If we are not performing "wrapping" and if the user asked for it,
1103 * scale the thread tile loops (P1T) of "sched" by gen->block_dim[i].
1105 static __isl_give isl_union_map
*scale_thread_tile_loops(struct gpu_gen
*gen
,
1106 __isl_take isl_union_map
*sched
)
1110 isl_basic_map
*scale
;
1112 isl_local_space
*ls
;
1114 if (gen
->options
->wrap
)
1116 if (!gen
->options
->scale_tile_loops
)
1119 dim
= isl_union_map_get_space(sched
);
1120 dim
= isl_space_add_dims(dim
, isl_dim_in
, gen
->thread_tiled_len
);
1121 dim
= isl_space_add_dims(dim
, isl_dim_out
, gen
->thread_tiled_len
);
1122 scale
= isl_basic_map_universe(isl_space_copy(dim
));
1123 ls
= isl_local_space_from_space(dim
);
1125 for (i
= 0; i
< gen
->thread_tiled_len
; ++i
) {
1128 if (i
>= gen
->shared_len
&&
1129 i
< gen
->shared_len
+ gen
->n_block
)
1130 f
= gen
->block_dim
[i
- gen
->shared_len
];
1132 c
= isl_equality_alloc(isl_local_space_copy(ls
));
1133 c
= isl_constraint_set_coefficient_si(c
, isl_dim_in
, i
, f
);
1134 c
= isl_constraint_set_coefficient_si(c
, isl_dim_out
, i
, -1);
1135 scale
= isl_basic_map_add_constraint(scale
, c
);
1138 isl_local_space_free(ls
);
1140 sched
= isl_union_map_apply_range(sched
,
1141 isl_union_map_from_map(isl_map_from_basic_map(scale
)));
1146 /* If we are not performing "wrapping" and if the user asked for it,
1147 * scale the "n_tile" loops starting at "first" of "sched" by gen->block_dim[i].
1149 static __isl_give isl_union_map
*scale_access_tile_loops(struct gpu_gen
*gen
,
1150 __isl_take isl_union_map
*sched
, int len
, int first
, int n_tile
)
1154 isl_basic_map
*scale
;
1156 isl_local_space
*ls
;
1158 if (gen
->options
->wrap
)
1160 if (!gen
->options
->scale_tile_loops
)
1163 dim
= isl_union_map_get_space(sched
);
1164 dim
= isl_space_add_dims(dim
, isl_dim_in
, len
);
1165 dim
= isl_space_add_dims(dim
, isl_dim_out
, len
);
1166 scale
= isl_basic_map_universe(isl_space_copy(dim
));
1167 ls
= isl_local_space_from_space(dim
);
1169 for (i
= 0; i
< len
; ++i
) {
1172 if (i
>= first
&& i
< first
+ n_tile
)
1173 f
= gen
->kernel
->block_dim
[i
- first
];
1175 c
= isl_equality_alloc(isl_local_space_copy(ls
));
1176 c
= isl_constraint_set_coefficient_si(c
, isl_dim_in
, i
, f
);
1177 c
= isl_constraint_set_coefficient_si(c
, isl_dim_out
, i
, -1);
1178 scale
= isl_basic_map_add_constraint(scale
, c
);
1181 isl_local_space_free(ls
);
1183 sched
= isl_union_map_apply_range(sched
,
1184 isl_union_map_from_map(isl_map_from_basic_map(scale
)));
1189 /* Add "len" parameters p[i] called prefix%d,
1190 * with bounds to 0 <= p[i] < size[i].
1192 __isl_give isl_set
*add_bounded_parameters(__isl_take isl_set
*set
,
1193 int len
, int *size
, const char *prefix
)
1198 isl_basic_set
*bset
;
1200 isl_local_space
*ls
;
1203 nparam
= isl_set_dim(set
, isl_dim_param
);
1204 set
= isl_set_add_dims(set
, isl_dim_param
, len
);
1206 for (i
= 0; i
< len
; ++i
) {
1207 snprintf(name
, sizeof(name
), "%s%d", prefix
, i
);
1208 set
= isl_set_set_dim_name(set
, isl_dim_param
,
1212 dim
= isl_set_get_space(set
);
1213 bset
= isl_basic_set_universe(isl_space_copy(dim
));
1214 ls
= isl_local_space_from_space(dim
);
1216 for (i
= 0; i
< len
; ++i
) {
1217 c
= isl_inequality_alloc(isl_local_space_copy(ls
));
1218 c
= isl_constraint_set_coefficient_si(c
, isl_dim_param
,
1220 bset
= isl_basic_set_add_constraint(bset
, c
);
1222 c
= isl_inequality_alloc(isl_local_space_copy(ls
));
1223 c
= isl_constraint_set_coefficient_si(c
, isl_dim_param
,
1225 c
= isl_constraint_set_constant_si(c
, size
[i
] - 1);
1226 bset
= isl_basic_set_add_constraint(bset
, c
);
1229 isl_local_space_free(ls
);
1231 return isl_set_intersect(set
, isl_set_from_basic_set(bset
));
1234 /* Add "len" parameters p[i] called prefix%d,
1235 * with bounds to 0 <= p[i] < size[i].
1237 static __isl_give isl_set
*add_bounded_parameters_dynamic(
1238 __isl_take isl_set
*set
, __isl_keep isl_multi_pw_aff
*size
,
1244 isl_local_space
*ls
;
1247 len
= isl_multi_pw_aff_dim(size
, isl_dim_out
);
1248 nparam
= isl_set_dim(set
, isl_dim_param
);
1249 set
= isl_set_add_dims(set
, isl_dim_param
, len
);
1251 for (i
= 0; i
< len
; ++i
) {
1252 snprintf(name
, sizeof(name
), "%s%d", prefix
, i
);
1253 set
= isl_set_set_dim_name(set
, isl_dim_param
,
1257 space
= isl_space_params(isl_set_get_space(set
));
1258 ls
= isl_local_space_from_space(space
);
1259 for (i
= 0; i
< len
; ++i
) {
1260 isl_pw_aff
*param
, *size_i
, *zero
;
1263 param
= isl_pw_aff_var_on_domain(isl_local_space_copy(ls
),
1264 isl_dim_param
, nparam
+ i
);
1266 size_i
= isl_multi_pw_aff_get_pw_aff(size
, i
);
1267 bound
= isl_pw_aff_lt_set(isl_pw_aff_copy(param
), size_i
);
1268 set
= isl_set_intersect_params(set
, bound
);
1270 zero
= isl_pw_aff_zero_on_domain(isl_local_space_copy(ls
));
1271 bound
= isl_pw_aff_ge_set(param
, zero
);
1272 set
= isl_set_intersect_params(set
, bound
);
1274 isl_local_space_free(ls
);
1279 /* Construct a map from an access to group->array to the corresponding
1280 * shared/private memory tile.
1281 * The map is of the form
1283 * { [D[i] -> A[a]] -> T[t] }
1285 * where D represents the initial shared_len dimensions
1286 * of the computed schedule.
1288 static __isl_give isl_map
*shift_access(struct gpu_array_ref_group
*group
)
1290 struct gpu_array_tile
*tile
;
1291 isl_multi_aff
*tiling
;
1293 tile
= group
->private_tile
;
1295 tile
= group
->shared_tile
;
1297 tiling
= isl_multi_aff_copy(tile
->tiling
);
1299 return isl_map_from_multi_aff(tiling
);
1302 /* Does "map" have an obviously fixed value at variable "pos" of "type"?
1304 static int map_plain_is_fixed(isl_map
*map
, enum isl_dim_type type
,
1310 v
= isl_map_plain_get_val_if_fixed(map
, type
, pos
);
1313 fixed
= isl_val_is_int(v
);
1319 /* Given a schedule that iterates over all elements in a piece of an array,
1320 * perform tiling/wrapping over the threads.
1322 * In particular, we tile the final iterators so that the final thread
1323 * dimension runs over the final array dimension.
1324 * However, if those final iterators have only a single iteration,
1325 * we try to tile earlier iterators instead.
1327 static __isl_give isl_map
*tile_access_schedule(struct gpu_gen
*gen
,
1328 __isl_take isl_map
*sched
)
1331 isl_union_map
*usched
;
1334 unsigned nvar
= isl_map_dim(sched
, isl_dim_out
);
1338 n_tile
= gen
->kernel
->n_block
;
1339 if (n_tile
> nvar
) {
1341 sched
= isl_map_insert_dims(sched
,
1342 isl_dim_out
, 0, n_tile
- nvar
);
1343 for (i
= 0; i
< n_tile
- nvar
; ++i
)
1344 sched
= isl_map_fix_si(sched
, isl_dim_out
, i
, 0);
1348 first
= nvar
- n_tile
;
1350 for (; first
> 0; first
--)
1351 if (!map_plain_is_fixed(sched
, isl_dim_out
, first
+ n_tile
- 1))
1354 dim
= isl_map_get_space(sched
);
1355 dim
= isl_space_params(dim
);
1356 if (gen
->options
->wrap
)
1357 tiling
= wrap(isl_space_copy(dim
), nvar
, first
,
1358 n_tile
, gen
->kernel
->block_dim
);
1360 tiling
= tile(isl_space_copy(dim
), nvar
, first
,
1361 n_tile
, gen
->kernel
->block_dim
);
1362 sched
= isl_map_apply_range(sched
, tiling
);
1364 par
= parametrization(dim
, nvar
+ n_tile
, first
+ n_tile
, n_tile
, "t");
1365 sched
= isl_map_intersect_range(sched
, par
);
1367 usched
= isl_union_map_from_map(sched
);
1368 usched
= scale_access_tile_loops(gen
, usched
, nvar
+ n_tile
,
1370 sched
= isl_map_from_union_map(usched
);
1375 /* Return the union of all read (read = 1) and/or write (write = 1)
1376 * access relations in the group.
1378 static __isl_give isl_union_map
*group_access_relation(
1379 struct gpu_array_ref_group
*group
, int read
, int write
)
1382 isl_union_map
*access
;
1384 access
= isl_union_map_empty(isl_map_get_space(group
->access
));
1385 for (i
= 0; i
< group
->n_ref
; ++i
) {
1388 if (!((read
&& group
->refs
[i
]->read
) ||
1389 (write
&& group
->refs
[i
]->write
)))
1391 map_i
= isl_map_copy(group
->refs
[i
]->access
);
1392 access
= isl_union_map_union(access
,
1393 isl_union_map_from_map(map_i
));
1399 /* Return the union of all tagged access relations in the group.
1401 static __isl_give isl_union_map
*group_tagged_access_relation(
1402 struct gpu_array_ref_group
*group
)
1405 isl_union_map
*access
;
1407 access
= isl_union_map_empty(isl_map_get_space(group
->access
));
1408 for (i
= 0; i
< group
->n_ref
; ++i
) {
1411 map_i
= isl_map_copy(group
->refs
[i
]->tagged_access
);
1412 access
= isl_union_map_union(access
,
1413 isl_union_map_from_map(map_i
));
1419 /* Return the extent of "array", recomputed from the bounds.
1420 * The recomputed extent may be simpler than the original extent.
1422 static __isl_give isl_set
*array_extent(struct gpu_array_info
*array
)
1427 isl_local_space
*ls
;
1430 id
= isl_set_get_tuple_id(array
->extent
);
1431 space
= isl_set_get_space(array
->extent
);
1432 extent
= isl_set_universe(isl_space_copy(space
));
1433 ls
= isl_local_space_from_space(space
);
1434 for (i
= 0; i
< array
->n_index
; ++i
) {
1440 extent
= isl_set_lower_bound_si(extent
, isl_dim_set
, i
, 0);
1442 aff
= isl_aff_var_on_domain(isl_local_space_copy(ls
),
1444 index
= isl_pw_aff_from_aff(aff
);
1445 bound
= isl_pw_aff_copy(array
->bound
[i
]);
1446 bound
= isl_pw_aff_from_range(bound
);
1447 bound
= isl_pw_aff_add_dims(bound
, isl_dim_in
, array
->n_index
);
1448 bound
= isl_pw_aff_set_tuple_id(bound
, isl_dim_in
,
1450 lt
= isl_pw_aff_lt_set(index
, bound
);
1451 extent
= isl_set_intersect(extent
, lt
);
1453 isl_local_space_free(ls
);
1459 /* Return a map from the first shared_len dimensions of the computed
1460 * schedule to the array tile in
1461 * global memory that corresponds to the shared memory copy.
1463 * In particular, return a map
1469 * tile_offset(i) <= a <= tile_offset(i) + tile_size - 1 (1)
1473 * 0 <= a <= array_size - 1 (2)
1475 * Note that if some stride has been detected (i.e., when
1476 * group->shared_tile->bound[i].shift is set), then a in (1) refers
1477 * to the shifted and scaled down version.
1479 * Constraints (1) are obtained by mapping the size constraints on the
1480 * shared/private memory tile back to the access relation.
1481 * Constraints (2) are obtained from the (recomputed) extent.
1483 static __isl_give isl_map
*group_tile(struct gpu_array_ref_group
*group
)
1486 int n_index
= group
->array
->n_index
;
1492 space
= isl_multi_aff_get_space(group
->shared_tile
->tiling
);
1493 space
= isl_space_range(space
);
1494 local
= isl_set_universe(space
);
1495 for (i
= 0; i
< n_index
; ++i
) {
1498 local
= isl_set_lower_bound_si(local
, isl_dim_set
, i
, 0);
1499 bound
= isl_val_copy(group
->shared_tile
->bound
[i
].size
);
1500 bound
= isl_val_sub_ui(bound
, 1);
1501 local
= isl_set_upper_bound_val(local
, isl_dim_set
, i
, bound
);
1503 local
= isl_set_preimage_multi_aff(local
,
1504 isl_multi_aff_copy(group
->shared_tile
->tiling
));
1505 tile
= isl_set_unwrap(local
);
1506 extent
= array_extent(group
->array
);
1507 tile
= isl_map_intersect_range(tile
, extent
);
1512 /* Given a mapping "iterator_map" from the AST schedule to a domain,
1513 * return the corresponding mapping from the AST schedule to
1514 * to the first shared_len dimensions of the schedule computed by PPCG.
1516 static __isl_give isl_pw_multi_aff
*compute_sched_to_shared(struct gpu_gen
*gen
,
1517 __isl_take isl_pw_multi_aff
*iterator_map
)
1519 isl_union_map
*umap
;
1521 isl_map
*map
, *sched
;;
1523 space
= isl_space_range(isl_pw_multi_aff_get_space(iterator_map
));
1524 space
= isl_space_from_domain(space
);
1525 space
= isl_space_add_dims(space
, isl_dim_out
, gen
->shared_len
);
1527 umap
= isl_union_map_copy(gen
->shared_sched
);
1528 umap
= isl_union_map_apply_range(umap
,
1529 isl_union_map_copy(gen
->shared_proj
));
1530 map
= isl_union_map_extract_map(umap
, space
);
1531 isl_union_map_free(umap
);
1533 sched
= isl_map_preimage_domain_pw_multi_aff(map
, iterator_map
);
1534 sched
= isl_map_detect_equalities(sched
);
1536 return isl_pw_multi_aff_from_map(sched
);
1539 /* Set unroll[j] if the input dimension j is involved in
1540 * the index expression represented by ma.
1542 static int check_unroll(__isl_take isl_set
*set
, __isl_take isl_multi_aff
*ma
,
1546 int n_in
= isl_multi_aff_dim(ma
, isl_dim_in
);
1547 int n_out
= isl_multi_aff_dim(ma
, isl_dim_out
);
1550 for (i
= 0; i
< n_out
; ++i
) {
1553 aff
= isl_multi_aff_get_aff(ma
, i
);
1554 for (j
= 0; j
< n_in
; ++j
)
1555 if (isl_aff_involves_dims(aff
, isl_dim_in
, j
, 1))
1561 isl_multi_aff_free(ma
);
1565 /* Given an array pos mapping input dimensions to the corresponding
1566 * output dimension, construct the corresponding map.
1568 static __isl_give isl_map
*permutation(__isl_take isl_space
*dim
,
1573 isl_basic_map
*bmap
;
1574 isl_local_space
*ls
;
1576 dim
= isl_space_add_dims(dim
, isl_dim_in
, len
);
1577 dim
= isl_space_add_dims(dim
, isl_dim_out
, len
);
1578 bmap
= isl_basic_map_universe(isl_space_copy(dim
));
1579 ls
= isl_local_space_from_space(dim
);
1581 for (i
= 0; i
< len
; ++i
) {
1582 c
= isl_equality_alloc(isl_local_space_copy(ls
));
1583 c
= isl_constraint_set_coefficient_si(c
, isl_dim_in
, i
,
1585 c
= isl_constraint_set_coefficient_si(c
, isl_dim_out
, pos
[i
],
1587 bmap
= isl_basic_map_add_constraint(bmap
, c
);
1589 isl_local_space_free(ls
);
1591 return isl_map_from_basic_map(bmap
);
1594 /* Find all loops involved in any of the index expressions for any of
1595 * the private accesses, move them innermost and then mark them as
1596 * requiring unrolling by setting gen->first_unroll.
1597 * The loops involved should all be parallel because of the checks
1598 * we performed in check_private_group_access. Moving them innermost
1599 * is therefore a valid transformation.
1601 * Loops up to gen->shared_len are generated before the mapping to
1602 * threads is applied. They should therefore be ignored.
1604 * We compute the hidden equalities of the schedule first
1605 * since we will need them in our calls to isl_pw_multi_aff_from_map
1606 * and because we want to make sure that the same equalities
1607 * are also available to the code generator.
1609 static __isl_give isl_union_map
*interchange_for_unroll(struct gpu_gen
*gen
,
1610 __isl_take isl_union_map
*sched
)
1613 int unroll
[gen
->thread_tiled_len
];
1614 int perm
[gen
->thread_tiled_len
];
1617 int len
= gen
->shared_len
+ gen
->n_parallel
+ gen
->n_block
;
1619 gen
->first_unroll
= -1;
1621 sched
= isl_union_map_detect_equalities(sched
);
1622 for (i
= 0; i
< gen
->thread_tiled_len
; ++i
)
1624 for (i
= 0; i
< gen
->prog
->n_array
; ++i
) {
1625 struct gpu_array_info
*array
= &gen
->prog
->array
[i
];
1627 for (j
= 0; j
< array
->n_group
; ++j
) {
1628 isl_union_map
*access
;
1630 isl_pw_multi_aff
*pma
;
1632 if (!array
->groups
[j
]->private_tile
)
1635 access
= group_access_relation(array
->groups
[j
], 1, 1);
1636 access
= isl_union_map_apply_domain(access
,
1637 isl_union_map_copy(sched
));
1639 acc
= isl_map_from_union_map(access
);
1640 pma
= isl_pw_multi_aff_from_map(acc
);
1641 isl_pw_multi_aff_foreach_piece(pma
,
1642 &check_unroll
, unroll
);
1644 isl_pw_multi_aff_free(pma
);
1648 for (i
= gen
->shared_len
; i
< len
; ++i
)
1655 for (i
= len
; i
< gen
->thread_tiled_len
; ++i
)
1660 for (i
= 0; i
< gen
->shared_len
; ++i
)
1662 for (i
= gen
->shared_len
; i
< gen
->thread_tiled_len
; ++i
)
1665 gen
->first_unroll
= j
- gen
->shared_len
;
1666 for (i
= gen
->shared_len
; i
< len
; ++i
)
1670 dim
= isl_union_map_get_space(sched
);
1671 permute
= permutation(dim
, perm
, gen
->thread_tiled_len
);
1672 sched
= isl_union_map_apply_range(sched
,
1673 isl_union_map_from_map(permute
));
1678 /* Given a constraint
1680 * a(p,i) + j = g f(e)
1682 * or -a(p,i) - j = g f(e) if sign < 0,
1683 * store a(p,i) in bound->shift and g (stride) in bound->stride.
1684 * a(p,i) is assumed to be an expression in only the parameters
1685 * and the input dimensions.
1687 static void extract_stride(__isl_keep isl_constraint
*c
,
1688 struct gpu_array_bound
*bound
, __isl_keep isl_val
*stride
, int sign
)
1697 isl_val_free(bound
->stride
);
1698 bound
->stride
= isl_val_copy(stride
);
1700 space
= isl_constraint_get_space(c
);
1701 space
= isl_space_domain(space
);
1703 nparam
= isl_space_dim(space
, isl_dim_param
);
1704 nvar
= isl_space_dim(space
, isl_dim_set
);
1706 v
= isl_constraint_get_constant_val(c
);
1709 aff
= isl_aff_zero_on_domain(isl_local_space_from_space(space
));
1710 aff
= isl_aff_set_constant_val(aff
, v
);
1712 for (i
= 0; i
< nparam
; ++i
) {
1713 if (!isl_constraint_involves_dims(c
, isl_dim_param
, i
, 1))
1715 v
= isl_constraint_get_coefficient_val(c
, isl_dim_param
, i
);
1718 aff
= isl_aff_add_coefficient_val(aff
, isl_dim_param
, i
, v
);
1721 for (i
= 0; i
< nvar
; ++i
) {
1722 if (!isl_constraint_involves_dims(c
, isl_dim_in
, i
, 1))
1724 v
= isl_constraint_get_coefficient_val(c
, isl_dim_in
, i
);
1727 aff
= isl_aff_add_coefficient_val(aff
, isl_dim_in
, i
, v
);
1733 /* Given an equality constraint of a map with a single output dimension j,
1734 * check if the constraint is of the form
1736 * a(p,i) + j = g f(e)
1738 * with a(p,i) an expression in the parameters and input dimensions
1739 * and f(e) an expression in the existentially quantified variables.
1740 * If so, and if g is larger than any such g from a previously considered
1741 * constraint, then call extract_stride to record the stride information
1744 static int check_stride_constraint(__isl_take isl_constraint
*c
, void *user
)
1750 struct gpu_array_bound
*bound
= user
;
1752 ctx
= isl_constraint_get_ctx(c
);
1753 n_div
= isl_constraint_dim(c
, isl_dim_div
);
1754 v
= isl_constraint_get_coefficient_val(c
, isl_dim_out
, 0);
1756 if (n_div
&& (isl_val_is_one(v
) || isl_val_is_negone(v
))) {
1757 int s
= isl_val_sgn(v
);
1758 isl_val
*stride
= isl_val_zero(ctx
);
1761 for (i
= 0; i
< n_div
; ++i
) {
1762 v
= isl_constraint_get_coefficient_val(c
,
1764 stride
= isl_val_gcd(stride
, v
);
1766 if (!isl_val_is_zero(stride
) &&
1767 isl_val_gt(stride
, bound
->stride
))
1768 extract_stride(c
, bound
, stride
, s
);
1770 isl_val_free(stride
);
1774 isl_constraint_free(c
);
1778 /* Given contraints on an array index i, check if we can find
1779 * a shift a(p) and a stride g such that
1781 * a(p) + i = 0 mod g
1783 * If so, record the information in bound and apply the mapping
1784 * i -> (i + a(p))/g to the array index in bounds and return
1785 * the new constraints.
1786 * If not, simply return the original constraints.
1788 * If bounds is a subset of the space
1792 * then the bound recorded in bound->shift is of the form
1796 * with s(D) equal to a(p) above.
1797 * The mapping recorded in bound->shift_map is of the form
1799 * [D -> i] -> [D -> (i + S(D))/g]
1801 * This mapping is computed as follows.
1802 * We first introduce "i" in the domain through precomposition
1803 * with [D -> i] -> D obtaining
1807 * Adding [D -> i] -> i produces
1809 * [D -> i] -> i + s(D)
1811 * and the domain product with [D -> i] -> D yields
1813 * [D -> i] -> [D -> i + s(D)]
1815 * Composition with [D -> i] -> [D -> i/g] gives the desired result.
1817 static __isl_give isl_basic_map
*check_stride(struct gpu_array_bound
*bound
,
1818 __isl_take isl_basic_map
*bounds
)
1821 isl_basic_map
*hull
;
1822 isl_basic_map
*shift
, *id
, *bmap
, *scale
;
1823 isl_basic_set
*bset
;
1826 bound
->stride
= NULL
;
1828 hull
= isl_basic_map_affine_hull(isl_basic_map_copy(bounds
));
1830 isl_basic_map_foreach_constraint(hull
, &check_stride_constraint
, bound
);
1832 isl_basic_map_free(hull
);
1837 shift
= isl_basic_map_from_aff(isl_aff_copy(bound
->shift
));
1838 space
= isl_basic_map_get_space(bounds
);
1839 bmap
= isl_basic_map_domain_map(isl_basic_map_universe(space
));
1840 shift
= isl_basic_map_apply_range(bmap
, shift
);
1841 space
= isl_basic_map_get_space(bounds
);
1842 id
= isl_basic_map_range_map(isl_basic_map_universe(space
));
1843 shift
= isl_basic_map_sum(id
, shift
);
1844 space
= isl_basic_map_get_space(bounds
);
1845 id
= isl_basic_map_domain_map(isl_basic_map_universe(space
));
1846 shift
= isl_basic_map_range_product(id
, shift
);
1848 space
= isl_space_domain(isl_basic_map_get_space(bounds
));
1849 id
= isl_basic_map_identity(isl_space_map_from_set(space
));
1850 space
= isl_space_range(isl_basic_map_get_space(bounds
));
1851 aff
= isl_aff_zero_on_domain(isl_local_space_from_space(space
));
1852 aff
= isl_aff_add_coefficient_si(aff
, isl_dim_in
, 0, 1);
1853 aff
= isl_aff_scale_down_val(aff
, isl_val_copy(bound
->stride
));
1854 scale
= isl_basic_map_from_aff(aff
);
1855 scale
= isl_basic_map_product(id
, scale
);
1857 bound
->shift_map
= isl_basic_map_apply_range(shift
, scale
);
1858 bmap
= isl_basic_map_copy(bound
->shift_map
);
1859 bset
= isl_basic_set_apply(isl_basic_map_wrap(bounds
), bmap
);
1860 bounds
= isl_basic_set_unwrap(bset
);
1865 /* Data used in compute_array_dim_size and compute_size_in_direction.
1867 * pos is the position of the variable representing the array index,
1868 * i.e., the variable for which want to compute the size. This variable
1869 * is also the last variable in the set.
1871 struct gpu_size_info
{
1872 isl_basic_set
*bset
;
1873 struct gpu_array_bound
*bound
;
1877 /* Given a constraint from the basic set describing the bounds on
1878 * an array index, check if it is a lower bound, say m i >= b(x), and,
1879 * if so, check whether the expression "i - ceil(b(x)/m) + 1" has a constant
1880 * upper bound. If so, and if this bound is smaller than any bound
1881 * derived from earlier constraints, set the size to this bound on
1882 * the expression and the lower bound to ceil(b(x)/m).
1884 static int compute_size_in_direction(__isl_take isl_constraint
*c
, void *user
)
1886 struct gpu_size_info
*size
= user
;
1893 nparam
= isl_basic_set_dim(size
->bset
, isl_dim_param
);
1894 n_div
= isl_constraint_dim(c
, isl_dim_div
);
1896 if (isl_constraint_involves_dims(c
, isl_dim_div
, 0, n_div
) ||
1897 !isl_constraint_is_lower_bound(c
, isl_dim_set
, size
->pos
)) {
1898 isl_constraint_free(c
);
1902 aff
= isl_constraint_get_bound(c
, isl_dim_set
, size
->pos
);
1903 aff
= isl_aff_ceil(aff
);
1905 lb
= isl_aff_copy(aff
);
1907 aff
= isl_aff_neg(aff
);
1908 aff
= isl_aff_add_coefficient_si(aff
, isl_dim_in
, size
->pos
, 1);
1910 v
= isl_basic_set_max_val(size
->bset
, aff
);
1913 if (isl_val_is_int(v
)) {
1914 v
= isl_val_add_ui(v
, 1);
1915 if (!size
->bound
->size
|| isl_val_lt(v
, size
->bound
->size
)) {
1916 isl_val_free(size
->bound
->size
);
1917 size
->bound
->size
= isl_val_copy(v
);
1918 lb
= isl_aff_drop_dims(lb
, isl_dim_in
, size
->pos
, 1);
1919 isl_aff_free(size
->bound
->lb
);
1920 size
->bound
->lb
= isl_aff_copy(lb
);
1926 isl_constraint_free(c
);
1931 /* Given a basic map "bounds" that maps parameters and input dimensions
1932 * to a single output dimension, look for an expression in the parameters
1933 * and input dimensions such that the range of the output dimension shifted
1934 * by this expression is a constant.
1936 * In particular, we currently only consider lower bounds on the output
1937 * dimension as candidate expressions.
1939 static int compute_array_dim_size(struct gpu_array_bound
*bound
,
1940 __isl_take isl_basic_map
*bounds
)
1942 struct gpu_size_info size
;
1944 bounds
= isl_basic_map_detect_equalities(bounds
);
1945 bounds
= check_stride(bound
, bounds
);
1951 size
.pos
= isl_basic_map_dim(bounds
, isl_dim_in
);
1952 size
.bset
= isl_basic_map_wrap(bounds
);
1953 size
.bset
= isl_basic_set_flatten(size
.bset
);
1954 size
.bset
= isl_set_simple_hull(isl_basic_set_compute_divs(size
.bset
));
1955 isl_basic_set_foreach_constraint(size
.bset
, &compute_size_in_direction
,
1957 isl_basic_set_free(size
.bset
);
1959 return bound
->size
? 0 : -1;
1962 /* Check if we can find a memory tile for the given array
1963 * based on the given accesses, and if so, put the results in "tile".
1965 * We project the accesses on each index in turn and look for a parametric
1966 * offset such that the size is constant.
1968 static int can_tile(__isl_keep isl_map
*access
, struct gpu_array_tile
*tile
)
1972 for (i
= 0; i
< tile
->n
; ++i
) {
1974 isl_basic_map
*hull
;
1976 access_i
= isl_map_copy(access
);
1977 access_i
= isl_map_project_out(access_i
, isl_dim_out
, 0, i
);
1978 access_i
= isl_map_project_out(access_i
, isl_dim_out
,
1979 1, tile
->n
- (i
+ 1));
1980 access_i
= isl_map_compute_divs(access_i
);
1981 hull
= isl_map_simple_hull(access_i
);
1982 if (compute_array_dim_size(&tile
->bound
[i
], hull
) < 0)
1989 /* Construct a map with input the shared tile loops and the loops that
1990 * will be wrapped around the threads that relates these later loops
1991 * to the thread indices and then projects them out.
1993 static __isl_give isl_map
*compute_privatization(struct gpu_gen
*gen
)
2001 dim
= isl_union_map_get_space(gen
->shared_sched
);
2003 if (gen
->options
->wrap
)
2004 tiling
= wrap(isl_space_copy(dim
), gen
->shared_len
+ gen
->n_block
,
2005 gen
->shared_len
, gen
->n_block
, gen
->block_dim
);
2007 tiling
= tile(isl_space_copy(dim
), gen
->shared_len
+ gen
->n_block
,
2008 gen
->shared_len
, gen
->n_block
, gen
->block_dim
);
2012 par
= parametrization(dim
, gen
->shared_len
+ 2 * gen
->n_block
,
2013 gen
->tile_first
+ gen
->tile_len
+ gen
->n_grid
+ gen
->n_block
,
2016 priv
= isl_map_align_params(priv
, isl_set_get_space(par
));
2017 priv
= isl_map_intersect_range(priv
, par
);
2019 dim
= isl_map_get_space(priv
);
2020 dim
= isl_space_drop_dims(dim
, isl_dim_in
, 0, isl_space_dim(dim
, isl_dim_in
));
2021 dim
= isl_space_drop_dims(dim
, isl_dim_out
, 0, isl_space_dim(dim
, isl_dim_out
));
2022 proj
= projection(dim
, gen
->shared_len
+ 2 * gen
->n_block
,
2025 priv
= isl_map_apply_range(priv
, proj
);
2030 /* Construct a map from domain_dim to domain_dim that increments
2031 * the dimension at position "pos" and leaves all other dimensions
2034 static __isl_give isl_map
*next(__isl_take isl_space
*domain_dim
, int pos
)
2037 int len
= isl_space_dim(domain_dim
, isl_dim_set
);
2039 isl_basic_map
*next
;
2040 isl_local_space
*ls
;
2042 dim
= isl_space_map_from_set(domain_dim
);
2043 next
= isl_basic_map_universe(isl_space_copy(dim
));
2044 ls
= isl_local_space_from_space(dim
);
2046 for (i
= 0; i
< len
; ++i
) {
2049 c
= isl_equality_alloc(isl_local_space_copy(ls
));
2050 c
= isl_constraint_set_coefficient_si(c
, isl_dim_in
, i
, 1);
2051 c
= isl_constraint_set_coefficient_si(c
, isl_dim_out
, i
, -1);
2053 c
= isl_constraint_set_constant_si(c
, 1);
2054 next
= isl_basic_map_add_constraint(next
, c
);
2057 isl_local_space_free(ls
);
2059 return isl_map_from_basic_map(next
);
2062 /* Check if the given access is coalesced.
2063 * That is, check whether incrementing the dimension that will get
2064 * wrapped over the last thread index results in incrementing
2065 * the last array index.
2067 * This function is only called for access relations without reuse.
2069 static int access_is_coalesced(struct gpu_gen
*gen
,
2070 __isl_keep isl_union_map
*access
)
2073 isl_map
*access_map
;
2074 isl_map
*next_thread_x
;
2075 isl_map
*next_element
;
2079 access
= isl_union_map_copy(access
);
2080 access
= isl_union_map_apply_domain(access
,
2081 isl_union_map_copy(gen
->tiled_sched
));
2082 access_map
= isl_map_from_union_map(access
);
2084 dim
= isl_map_get_space(access_map
);
2085 dim
= isl_space_domain(dim
);
2086 next_thread_x
= next(dim
, gen
->shared_len
+ gen
->n_block
- 1);
2088 dim
= isl_map_get_space(access_map
);
2089 dim
= isl_space_range(dim
);
2090 next_element
= next(dim
, isl_space_dim(dim
, isl_dim_set
) - 1);
2092 map
= isl_map_apply_domain(next_thread_x
, isl_map_copy(access_map
));
2093 map
= isl_map_apply_range(map
, access_map
);
2095 coalesced
= isl_map_is_subset(map
, next_element
);
2097 isl_map_free(next_element
);
2103 /* Given an access relation in terms of the first gen->shared_len + gen->n_block
2104 * dimensions of the computed schedule, check if it is bijective for
2105 * fixed values of the first gen->shared_len dimensions.
2106 * We perform this check by equating these dimensions to parameters.
2108 static int access_is_bijective(struct gpu_gen
*gen
, __isl_keep isl_map
*access
)
2114 access
= isl_map_copy(access
);
2115 space
= isl_space_params(isl_map_get_space(access
));
2116 par
= parametrization(space
, gen
->shared_len
+ gen
->n_block
,
2117 0, gen
->shared_len
, "s");
2118 access
= isl_map_intersect_domain(access
, par
);
2119 res
= isl_map_is_bijective(access
);
2120 isl_map_free(access
);
2125 /* Look for the last shared tile loop that affects the offset of "tile"
2126 * and return the result.
2127 * If there is no such loop, then return the index of the loop
2128 * before the first shared tile loop, in particular gen->tile_first - 1.
2130 static int compute_tile_last_shared(struct gpu_gen
*gen
,
2131 struct gpu_array_tile
*tile
)
2135 for (j
= gen
->shared_len
- 1; j
>= gen
->tile_first
; --j
) {
2136 for (i
= 0; i
< tile
->n
; ++i
) {
2140 lb
= tile
->bound
[i
].lb
;
2141 if (isl_aff_involves_dims(lb
, isl_dim_in
, j
, 1))
2144 shift
= tile
->bound
[i
].shift
;
2147 if (isl_aff_involves_dims(shift
, isl_dim_in
, j
, 1))
2157 /* Look for the last shared tile loop that affects the offset of the
2158 * shared or private tile and store the result in group->last_shared.
2159 * If there is no such loop, then group->last_shared is set to a value
2160 * before the first shared tile loop, in particular gen->tile_first - 1.
2161 * If there is no tile defined on the array reference group,
2162 * then set group->last_shared to gen->shared_len - 1.
2164 static void set_last_shared(struct gpu_gen
*gen
,
2165 struct gpu_array_ref_group
*group
)
2167 struct gpu_array_tile
*tile
;
2169 group
->last_shared
= gen
->shared_len
- 1;
2171 tile
= group
->private_tile
;
2173 tile
= group
->shared_tile
;
2177 group
->last_shared
= compute_tile_last_shared(gen
, tile
);
2180 /* Compute a privatized copy of all access relations from reference groups that
2181 * are mapped to private memory and store the result in gen->privatization.
2183 * Read-only scalars and arrays containing structures are not mapped
2184 * to private memory.
2186 static void compute_private_access(struct gpu_gen
*gen
)
2189 isl_union_map
*private;
2191 if (!gen
->options
->use_private_memory
)
2194 private = isl_union_map_empty(isl_union_map_get_space(gen
->shared_sched
));
2196 for (i
= 0; i
< gen
->prog
->n_array
; ++i
) {
2197 struct gpu_array_info
*array
= &gen
->prog
->array
[i
];
2199 if (gpu_array_is_read_only_scalar(array
))
2201 if (array
->has_compound_element
)
2204 for (j
= 0; j
< array
->n_group
; ++j
) {
2205 if (!array
->groups
[j
]->private_tile
)
2208 private = isl_union_map_union(private,
2209 group_access_relation(array
->groups
[j
], 1, 1));
2213 if (isl_union_map_is_empty(private))
2214 isl_union_map_free(private);
2216 isl_union_map
*priv
;
2218 private = isl_union_map_apply_domain(private,
2219 isl_union_map_copy(gen
->shared_sched
));
2220 priv
= isl_union_map_from_map(isl_map_copy(gen
->privatization
));
2221 private = isl_union_map_apply_domain(private, priv
);
2222 gen
->private_access
= private;
2226 /* Compute the size of the tile specified by "tile"
2227 * in number of elements and return the result.
2229 static __isl_give isl_val
*tile_size(isl_ctx
*ctx
, struct gpu_array_tile
*tile
)
2234 size
= isl_val_one(ctx
);
2236 for (i
= 0; i
< tile
->n
; ++i
)
2237 size
= isl_val_mul(size
, isl_val_copy(tile
->bound
[i
].size
));
2242 /* If max_shared_memory is not set to infinity (-1), then make
2243 * sure that the total amount of shared memory required by the
2244 * array reference groups mapped to shared memory is no larger
2245 * than this maximum.
2247 * We apply a greedy approach and discard (keep in global memory)
2248 * those groups that would result in a total memory size that
2249 * is larger than the maximum.
2251 static void check_shared_memory_bound(struct gpu_gen
*gen
)
2254 isl_val
*left
, *size
;
2256 if (gen
->options
->max_shared_memory
< 0)
2259 left
= isl_val_int_from_si(gen
->ctx
, gen
->options
->max_shared_memory
);
2261 for (i
= 0; i
< gen
->prog
->n_array
; ++i
) {
2262 struct gpu_array_info
*array
= &gen
->prog
->array
[i
];
2264 for (j
= 0; j
< array
->n_group
; ++j
) {
2265 struct gpu_array_ref_group
*group
;
2267 group
= array
->groups
[j
];
2268 if (group
->private_tile
)
2270 if (!group
->shared_tile
)
2273 size
= tile_size(gen
->ctx
, group
->shared_tile
);
2274 size
= isl_val_mul_ui(size
, array
->size
);
2276 if (isl_val_le(size
, left
)) {
2277 left
= isl_val_sub(left
, size
);
2282 group
->shared_tile
= free_tile(group
->shared_tile
);
2289 /* Given a description of an array tile "tile" and the "space"
2293 * where D represents the first shared_len schedule dimensions
2294 * and A represents the array, construct an isl_multi_aff
2296 * { [D[i] -> A[a]] -> A'[a'] }
2298 * with A' a scaled down copy of A according to the shifts and strides
2299 * in "tile". In particular,
2301 * a' = (a + shift(i))/stride
2303 * "insert_array" represents
2307 * and is used to insert A into the domain of functions that only
2310 static __isl_give isl_multi_aff
*strided_tile(
2311 struct gpu_array_tile
*tile
, __isl_keep isl_space
*space
,
2312 __isl_keep isl_multi_aff
*insert_array
)
2316 isl_multi_aff
*shift
;
2317 isl_multi_val
*stride
;
2319 isl_local_space
*ls
;
2320 isl_multi_aff
*tiling
;
2322 ctx
= isl_space_get_ctx(space
);
2323 space2
= isl_space_domain(isl_space_copy(space
));
2324 ls
= isl_local_space_from_space(space2
);
2325 space2
= isl_space_range(isl_space_copy(space
));
2326 stride
= isl_multi_val_zero(space2
);
2327 shift
= isl_multi_aff_zero(isl_space_copy(space
));
2329 for (i
= 0; i
< tile
->n
; ++i
) {
2330 struct gpu_array_bound
*bound
= &tile
->bound
[i
];
2334 if (tile
->bound
[i
].shift
) {
2335 stride_i
= isl_val_copy(bound
->stride
);
2336 shift_i
= isl_aff_copy(bound
->shift
);
2338 stride_i
= isl_val_one(ctx
);
2339 shift_i
= isl_aff_zero_on_domain(
2340 isl_local_space_copy(ls
));
2343 stride
= isl_multi_val_set_val(stride
, i
, stride_i
);
2344 shift
= isl_multi_aff_set_aff(shift
, i
, shift_i
);
2346 isl_local_space_free(ls
);
2348 shift
= isl_multi_aff_pullback_multi_aff(shift
,
2349 isl_multi_aff_copy(insert_array
));
2351 tiling
= isl_multi_aff_range_map(isl_space_copy(space
));
2352 tiling
= isl_multi_aff_add(tiling
, shift
);
2353 tiling
= isl_multi_aff_scale_down_multi_val(tiling
, stride
);
2358 /* Compute a tiling for the array reference group "group".
2360 * The tiling is of the form
2362 * { [D[i] -> A[a]] -> T[t] }
2364 * where D represents the first shared_len schedule dimensions,
2365 * A represents the global array and T represents the shared or
2366 * private memory tile. The name of T is the name of the local
2369 * If there is any stride in the accesses, then the mapping is
2371 * t = (a + shift(i))/stride - lb(i)
2373 * otherwise, it is simply
2377 static void compute_group_tiling(struct gpu_array_ref_group
*group
)
2380 struct gpu_array_tile
*tile
;
2381 struct gpu_array_info
*array
= group
->array
;
2383 isl_multi_aff
*tiling
, *lb
, *insert_array
;
2387 tile
= group
->private_tile
;
2389 tile
= group
->shared_tile
;
2393 space
= isl_map_get_space(group
->access
);
2394 insert_array
= isl_multi_aff_domain_map(isl_space_copy(space
));
2396 for (i
= 0; i
< tile
->n
; ++i
)
2397 if (tile
->bound
[i
].shift
)
2401 tiling
= strided_tile(tile
, space
, insert_array
);
2403 tiling
= isl_multi_aff_range_map(isl_space_copy(space
));
2405 lb
= isl_multi_aff_zero(space
);
2406 for (i
= 0; i
< tile
->n
; ++i
) {
2407 isl_aff
*lb_i
= isl_aff_copy(tile
->bound
[i
].lb
);
2408 lb
= isl_multi_aff_set_aff(lb
, i
, lb_i
);
2410 lb
= isl_multi_aff_pullback_multi_aff(lb
, insert_array
);
2412 tiling
= isl_multi_aff_sub(tiling
, lb
);
2414 p
= isl_printer_to_str(isl_multi_aff_get_ctx(tiling
));
2415 p
= print_array_name(p
, group
);
2416 local_name
= isl_printer_get_str(p
);
2417 isl_printer_free(p
);
2418 tiling
= isl_multi_aff_set_tuple_name(tiling
, isl_dim_out
, local_name
);
2421 tile
->tiling
= tiling
;
2424 /* Compute a tiling for all the array reference groups.
2426 static void compute_group_tilings(struct gpu_gen
*gen
)
2430 for (i
= 0; i
< gen
->prog
->n_array
; ++i
) {
2431 struct gpu_array_info
*array
= &gen
->prog
->array
[i
];
2433 for (j
= 0; j
< array
->n_group
; ++j
)
2434 compute_group_tiling(array
->groups
[j
]);
2438 /* Fill up the groups array with singleton groups, i.e., one group
2439 * per reference, initializing the array, access, write, n_ref and refs fields.
2440 * In particular the access field is initialized to the scheduled
2441 * access relation of the array reference.
2443 * Return the number of elements initialized, i.e., the number of
2444 * active references in the current kernel.
2446 static int populate_array_references(struct gpu_array_info
*array
,
2447 __isl_keep isl_union_map
*sched
, struct gpu_array_ref_group
**groups
)
2451 isl_ctx
*ctx
= isl_union_map_get_ctx(sched
);
2454 for (i
= 0; i
< array
->n_ref
; ++i
) {
2455 isl_union_map
*umap
;
2457 struct gpu_array_ref_group
*group
;
2458 struct gpu_stmt_access
*access
= array
->refs
[i
];
2460 map
= isl_map_copy(access
->access
);
2461 umap
= isl_union_map_from_map(map
);
2462 umap
= isl_union_map_apply_domain(umap
,
2463 isl_union_map_copy(sched
));
2465 if (isl_union_map_is_empty(umap
)) {
2466 isl_union_map_free(umap
);
2470 map
= isl_map_from_union_map(umap
);
2471 map
= isl_map_detect_equalities(map
);
2473 group
= isl_calloc_type(ctx
, struct gpu_array_ref_group
);
2475 group
->array
= array
;
2476 group
->access
= map
;
2477 group
->write
= access
->write
;
2478 group
->exact_write
= access
->exact_write
;
2479 group
->refs
= &array
->refs
[i
];
2482 groups
[n
++] = group
;
2488 /* If group->n_ref == 1, then group->refs was set by
2489 * populate_array_references to point directly into
2490 * group->array->refs and should not be freed.
2491 * If group->n_ref > 1, then group->refs was set by join_groups
2492 * to point to a newly allocated array.
2494 static void free_array_ref_group(struct gpu_array_ref_group
*group
)
2498 free_tile(group
->shared_tile
);
2499 free_tile(group
->private_tile
);
2500 isl_map_free(group
->access
);
2501 if (group
->n_ref
> 1)
2506 /* Given a map where the input dimensions represent the tile loops,
2507 * eliminate the innermost of those that have a fixed value
2508 * until we reach one that does not (obviously) have a fixed value.
2510 static __isl_give isl_map
*eliminate_fixed_inner_loops(
2511 __isl_take isl_map
*access
)
2515 n
= isl_map_dim(access
, isl_dim_in
);
2517 for (i
= n
- 1; i
>= 0; --i
) {
2518 if (!map_plain_is_fixed(access
, isl_dim_in
, i
))
2520 access
= isl_map_eliminate(access
, isl_dim_in
, i
, 1);
2525 /* Check if the access relations of group1 and group2 overlap within
2526 * the innermost loop. In particular, ignore any inner dimension
2527 * with a fixed value.
2528 * The copying to and from shared memory will be performed within
2529 * the innermost actual loop so we are only allowed to consider
2530 * the dimensions up to that innermost loop while checking whether
2531 * two access relations overlap.
2533 static int accesses_overlap(struct gpu_array_ref_group
*group1
,
2534 struct gpu_array_ref_group
*group2
)
2537 isl_map
*access1
, *access2
;
2539 access1
= isl_map_copy(group1
->access
);
2540 access1
= eliminate_fixed_inner_loops(access1
);
2541 access2
= isl_map_copy(group2
->access
);
2542 access2
= eliminate_fixed_inner_loops(access2
);
2543 access1
= isl_map_intersect(access1
, access2
);
2544 empty
= isl_map_is_empty(access1
);
2545 isl_map_free(access1
);
2550 /* Combine the given two groups into a single group, containing
2551 * the references of both groups.
2553 static struct gpu_array_ref_group
*join_groups(
2554 struct gpu_array_ref_group
*group1
,
2555 struct gpu_array_ref_group
*group2
)
2559 struct gpu_array_ref_group
*group
;
2561 ctx
= isl_map_get_ctx(group1
->access
);
2562 group
= isl_calloc_type(ctx
, struct gpu_array_ref_group
);
2564 group
->array
= group1
->array
;
2565 group
->access
= isl_map_union(isl_map_copy(group1
->access
),
2566 isl_map_copy(group2
->access
));
2567 group
->write
= group1
->write
|| group2
->write
;
2568 group
->exact_write
= group1
->exact_write
&& group2
->exact_write
;
2569 group
->n_ref
= group1
->n_ref
+ group2
->n_ref
;
2570 group
->refs
= isl_alloc_array(ctx
, struct gpu_stmt_access
*,
2572 assert(group
->refs
);
2573 for (i
= 0; i
< group1
->n_ref
; ++i
)
2574 group
->refs
[i
] = group1
->refs
[i
];
2575 for (i
= 0; i
< group2
->n_ref
; ++i
)
2576 group
->refs
[group1
->n_ref
+ i
] = group2
->refs
[i
];
2581 /* Combine the given two groups into a single group and free
2582 * the original two groups.
2584 static struct gpu_array_ref_group
*join_groups_and_free(
2585 struct gpu_array_ref_group
*group1
,
2586 struct gpu_array_ref_group
*group2
)
2588 struct gpu_array_ref_group
*group
;
2590 group
= join_groups(group1
, group2
);
2591 free_array_ref_group(group1
);
2592 free_array_ref_group(group2
);
2596 /* Compute the private and/or shared memory tiles for the array
2597 * reference group "group" of array "array".
2598 * Return 0 on success and -1 on error.
2600 * If the array is a read-only scalar or if the user requested
2601 * not to use shared or private memory, then we do not need to do anything.
2603 * If the array group involves any may writes (that are not must writes),
2604 * then we would have to make sure that we load the data into shared/private
2605 * memory first in case the data is not written by the kernel
2606 * (but still written back out to global memory).
2607 * Since we don't have any such mechanism at the moment, we don't
2608 * compute shared/private tiles for groups involving may writes.
2610 * We only try to compute a shared memory tile if there is any reuse
2611 * or if the access is not coalesced.
2613 * For computing a private memory tile, we also require that there is
2614 * some reuse. Moreover, we require that the access is private
2615 * to the thread. That is, we check that any given array element
2616 * is only accessed by a single thread.
2617 * We compute an access relation that maps the shared tile loop iterators
2618 * and the shared point loop iterators that will be wrapped over the
2619 * threads to the array elements.
2620 * We actually check that those iterators that will be wrapped
2621 * partition the array space. This check is stricter than necessary
2622 * since several iterations may be mapped onto the same thread
2623 * and then they could be allowed to access the same memory elements,
2624 * but our check does not allow this situation.
2626 * We also check that the index expression only depends on parallel
2627 * loops. That way, we can move those loops innermost and unroll them.
2628 * Again, we use a test that is stricter than necessary.
2629 * We actually check whether the index expression only depends
2630 * on the iterators that are wrapped over the threads.
2631 * These are necessarily parallel, but there may be more parallel loops.
2633 * Combining the injectivity of the first test with the single-valuedness
2634 * of the second test, we simply test for bijectivity.
2636 * If it turns out we can use registers, we compute the private memory
2637 * tile size using can_tile, after introducing a dependence
2638 * on the thread indices.
2640 static int compute_group_bounds_core(struct gpu_gen
*gen
,
2641 struct gpu_array_ref_group
*group
)
2643 isl_ctx
*ctx
= isl_space_get_ctx(group
->array
->space
);
2644 isl_union_map
*access
;
2645 int n_index
= group
->array
->n_index
;
2648 int use_shared
= gen
->options
->use_shared_memory
;
2649 int use_private
= gen
->options
->use_private_memory
;
2651 if (!use_shared
&& !use_private
)
2653 if (gpu_array_is_read_only_scalar(group
->array
))
2655 if (!group
->exact_write
)
2658 access
= group_access_relation(group
, 1, 1);
2659 no_reuse
= isl_union_map_is_injective(access
);
2661 if (use_shared
&& (!no_reuse
|| !access_is_coalesced(gen
, access
))) {
2662 group
->shared_tile
= create_tile(ctx
, group
->array
->n_index
);
2663 if (!can_tile(group
->access
, group
->shared_tile
))
2664 group
->shared_tile
= free_tile(group
->shared_tile
);
2667 if (!use_private
|| no_reuse
) {
2668 isl_union_map_free(access
);
2672 access
= isl_union_map_apply_domain(access
,
2673 isl_union_map_copy(gen
->shared_sched
));
2675 acc
= isl_map_from_union_map(access
);
2677 if (!access_is_bijective(gen
, acc
)) {
2682 group
->private_tile
= create_tile(gen
->ctx
, n_index
);
2683 acc
= isl_map_apply_domain(acc
, isl_map_copy(gen
->privatization
));
2684 if (!can_tile(acc
, group
->private_tile
))
2685 group
->private_tile
= free_tile(group
->private_tile
);
2692 /* Compute the private and/or shared memory tiles for the array
2693 * reference group "group" of array "array" and set last_shared.
2694 * Return 0 on success and -1 on error.
2696 static int compute_group_bounds(struct gpu_gen
*gen
,
2697 struct gpu_array_ref_group
*group
)
2699 if (compute_group_bounds_core(gen
, group
) < 0)
2701 set_last_shared(gen
, group
);
2706 /* If two groups have overlapping access relations (as determined by
2707 * the "overlap" function) and if one of them involves a write,
2708 * then merge the two groups into one.
2709 * If "compute_bounds" is set, then call compute_group_bounds
2710 * on the merged groups.
2712 * Return the updated number of groups.
2713 * Return -1 on error.
2715 static int group_writes(struct gpu_gen
*gen
,
2716 int n
, struct gpu_array_ref_group
**groups
,
2717 int (*overlap
)(struct gpu_array_ref_group
*group1
,
2718 struct gpu_array_ref_group
*group2
), int compute_bounds
)
2722 for (i
= 0; i
< n
; ++i
) {
2723 for (j
= n
- 1; j
> i
; --j
) {
2724 if (!groups
[i
]->write
&& !groups
[j
]->write
)
2727 if (!overlap(groups
[i
], groups
[j
]))
2730 groups
[i
] = join_groups_and_free(groups
[i
], groups
[j
]);
2731 if (compute_bounds
&&
2732 compute_group_bounds(gen
, groups
[i
]) < 0)
2735 groups
[j
] = groups
[n
- 1];
2736 groups
[n
- 1] = NULL
;
2744 /* If two groups have overlapping access relations (within the innermost
2745 * loop) and if one of them involves a write, then merge the two groups
2748 * Return the updated number of groups.
2750 static int group_overlapping_writes(struct gpu_gen
*gen
,
2751 int n
, struct gpu_array_ref_group
**groups
)
2753 return group_writes(gen
, n
, groups
, &accesses_overlap
, 0);
2756 /* Check if the access relations of group1 and group2 overlap within
2757 * the outermost min(group1->last_shared, group2->last_shared) loops.
2759 static int last_shared_accesses_overlap(struct gpu_array_ref_group
*group1
,
2760 struct gpu_array_ref_group
*group2
)
2765 isl_map
*map_i
, *map_j
, *map
;
2767 last_shared
= group1
->last_shared
;
2768 if (group2
->last_shared
< last_shared
)
2769 last_shared
= group2
->last_shared
;
2770 map_i
= isl_map_copy(group1
->access
);
2771 dim
= isl_map_dim(map_i
, isl_dim_in
);
2772 map_i
= isl_map_eliminate(map_i
, isl_dim_in
,
2773 last_shared
+ 1, dim
- (last_shared
+ 1));
2774 map_j
= isl_map_copy(group2
->access
);
2775 map_j
= isl_map_eliminate(map_j
, isl_dim_in
,
2776 last_shared
+ 1, dim
- (last_shared
+ 1));
2777 map
= isl_map_intersect(map_i
, map_j
);
2778 empty
= isl_map_is_empty(map
);
2784 /* If two groups have overlapping access relations (within the outer
2785 * last_shared loops) and if one of them involves a write,
2786 * then merge the two groups into one.
2788 * Return the updated number of groups.
2790 static int group_last_shared_overlapping_writes(struct gpu_gen
*gen
, int n
,
2791 struct gpu_array_ref_group
**groups
)
2793 return group_writes(gen
, n
, groups
, &last_shared_accesses_overlap
, 1);
2796 /* Is the size of the tile specified by "tile" smaller than the sum of
2797 * the sizes of the tiles specified by "tile1" and "tile2"?
2799 static int smaller_tile(isl_ctx
*ctx
, struct gpu_array_tile
*tile
,
2800 struct gpu_array_tile
*tile1
, struct gpu_array_tile
*tile2
)
2803 isl_val
*size
, *size1
, *size2
;
2805 size
= tile_size(ctx
, tile
);
2806 size1
= tile_size(ctx
, tile1
);
2807 size2
= tile_size(ctx
, tile2
);
2809 size
= isl_val_sub(size
, size1
);
2810 size
= isl_val_sub(size
, size2
);
2811 smaller
= isl_val_is_neg(size
);
2818 /* Given an initial grouping of array references and shared memory tiles
2819 * for each group that allows for a shared memory tile, merge two groups
2820 * if both have a shared memory tile, the merged group also has
2821 * a shared memory tile and the size of the tile for the merge group
2822 * is smaller than the sum of the tile sizes of the individual groups.
2824 * If merging two groups decreases the "last_shared" dimension of
2825 * one or both of the two groups, then we need to check for overlapping
2828 * Return the number of groups after merging.
2829 * Return -1 on error.
2831 static int group_common_shared_memory_tile(struct gpu_gen
*gen
,
2832 struct gpu_array_info
*array
, int n
,
2833 struct gpu_array_ref_group
**groups
)
2836 int recompute_overlap
= 0;
2837 isl_ctx
*ctx
= isl_space_get_ctx(array
->space
);
2839 for (i
= 0; i
< n
; ++i
) {
2840 if (!groups
[i
]->shared_tile
)
2842 for (j
= n
- 1; j
> i
; --j
) {
2845 struct gpu_array_ref_group
*group
;
2847 if (!groups
[j
]->shared_tile
)
2850 map
= isl_map_intersect(isl_map_copy(groups
[i
]->access
),
2851 isl_map_copy(groups
[j
]->access
));
2852 empty
= isl_map_is_empty(map
);
2858 group
= join_groups(groups
[i
], groups
[j
]);
2859 if (compute_group_bounds(gen
, group
) < 0) {
2860 free_array_ref_group(group
);
2863 if (!group
->shared_tile
||
2864 !smaller_tile(ctx
, group
->shared_tile
,
2865 groups
[i
]->shared_tile
,
2866 groups
[j
]->shared_tile
)) {
2867 free_array_ref_group(group
);
2871 if (group
->last_shared
< groups
[i
]->last_shared
||
2872 group
->last_shared
< groups
[j
]->last_shared
)
2873 recompute_overlap
= 1;
2874 free_array_ref_group(groups
[i
]);
2875 free_array_ref_group(groups
[j
]);
2878 groups
[j
] = groups
[n
- 1];
2883 if (recompute_overlap
)
2884 n
= group_last_shared_overlapping_writes(gen
, n
, groups
);
2888 /* Set array->n_group and array->groups to n and groups.
2890 * Additionally, set the "nr" field of each group
2891 * and the "group" field of each reference in each group.
2893 static void set_array_groups(struct gpu_array_info
*array
,
2894 int n
, struct gpu_array_ref_group
**groups
)
2899 array
->groups
= groups
;
2901 for (i
= 0; i
< n
; ++i
) {
2904 for (j
= 0; j
< groups
[i
]->n_ref
; ++j
)
2905 groups
[i
]->refs
[j
]->group
= i
;
2909 /* Group array references that should be considered together when
2910 * deciding whether to access them from private, shared or global memory.
2911 * Return -1 on error.
2913 * In particular, if two array references overlap and if one of them
2914 * is a write, then the two references are grouped together.
2915 * We first perform an initial grouping based only on the access relation.
2916 * After computing shared and private memory tiles, we check for
2917 * overlapping writes again, but this time taking into account
2918 * the "last_shared" property.
2920 * Furthermore, if two groups admit a shared memory tile and if the
2921 * combination of the two also admits a shared memory tile, we merge
2924 * If the array contains structures, then there is no need to compute
2925 * reference groups since we do not map such arrays to private or shared
2928 static int group_array_references(struct gpu_gen
*gen
,
2929 struct gpu_array_info
*array
, __isl_keep isl_union_map
*sched
)
2933 isl_ctx
*ctx
= isl_union_map_get_ctx(sched
);
2934 struct gpu_array_ref_group
**groups
;
2936 if (array
->has_compound_element
)
2939 groups
= isl_calloc_array(ctx
, struct gpu_array_ref_group
*,
2944 n
= populate_array_references(array
, sched
, groups
);
2946 n
= group_overlapping_writes(gen
, n
, groups
);
2948 for (i
= 0; i
< n
; ++i
)
2949 if (compute_group_bounds(gen
, groups
[i
]) < 0)
2952 n
= group_last_shared_overlapping_writes(gen
, n
, groups
);
2954 n
= group_common_shared_memory_tile(gen
, array
, n
, groups
);
2956 set_array_groups(array
, n
, groups
);
2961 for (i
= 0; i
< array
->n_ref
; ++i
)
2962 free_array_ref_group(groups
[i
]);
2966 /* Take tiled_sched, project it onto the shared tile loops and
2967 * the loops that will be wrapped over the threads and
2968 * store the result in gen->shared_sched.
2969 * Also compute a projection that projects out the loops that will be
2970 * wrapped over the threads and store this projection in gen->shared_proj.
2972 static void compute_shared_sched(struct gpu_gen
*gen
)
2977 isl_union_map
*sched
;
2979 sched
= isl_union_map_copy(gen
->tiled_sched
);
2981 dim
= isl_union_map_get_space(sched
);
2982 proj
= projection(dim
, gen
->tiled_len
, gen
->shared_len
+ gen
->n_block
);
2983 sched
= isl_union_map_apply_range(sched
, isl_union_map_from_map(proj
));
2985 dim
= isl_union_map_get_space(sched
);
2986 proj
= projection(dim
, gen
->shared_len
+ gen
->n_block
, gen
->shared_len
);
2988 gen
->shared_sched
= sched
;
2989 gen
->shared_proj
= isl_union_map_from_map(proj
);
2992 /* Group references of all arrays in the program.
2994 static int group_references(struct gpu_gen
*gen
)
2998 isl_union_map
*sched
;
3000 sched
= isl_union_map_apply_range(isl_union_map_copy(gen
->shared_sched
),
3001 isl_union_map_copy(gen
->shared_proj
));
3003 for (i
= 0; i
< gen
->prog
->n_array
; ++i
) {
3004 r
= group_array_references(gen
, &gen
->prog
->array
[i
], sched
);
3009 isl_union_map_free(sched
);
3014 /* Free all array information that is local to the current kernel.
3016 static void free_local_array_info(struct gpu_gen
*gen
)
3020 for (i
= 0; i
< gen
->prog
->n_array
; ++i
) {
3021 struct gpu_array_info
*array
= &gen
->prog
->array
[i
];
3023 for (j
= 0; j
< array
->n_group
; ++j
)
3024 free_array_ref_group(array
->groups
[j
]);
3025 free(array
->groups
);
3029 /* Compute the size of a bounding box around the origin and "set",
3030 * where "set" is assumed to contain only non-negative elements.
3031 * In particular, compute the maximal value of "set" in each direction
3034 static __isl_give isl_multi_pw_aff
*extract_size(__isl_take isl_set
*set
,
3035 __isl_keep isl_set
*context
)
3038 isl_multi_pw_aff
*mpa
;
3040 n
= isl_set_dim(set
, isl_dim_set
);
3041 mpa
= isl_multi_pw_aff_zero(isl_set_get_space(set
));
3042 for (i
= 0; i
< n
; ++i
) {
3047 bound
= isl_set_dim_max(isl_set_copy(set
), i
);
3048 bound
= isl_pw_aff_coalesce(bound
);
3049 bound
= isl_pw_aff_gist(bound
, isl_set_copy(context
));
3051 space
= isl_pw_aff_get_domain_space(bound
);
3052 one
= isl_aff_zero_on_domain(isl_local_space_from_space(space
));
3053 one
= isl_aff_add_constant_si(one
, 1);
3054 bound
= isl_pw_aff_add(bound
, isl_pw_aff_from_aff(one
));
3055 mpa
= isl_multi_pw_aff_set_pw_aff(mpa
, i
, bound
);
3062 /* Compute the effective grid size as a list of the sizes in each dimension.
3064 * The grid size specified by the user or set by default
3065 * in read_grid_sizes() and applied in tile_schedule(),
3066 * may be too large for the given code in the sense that
3067 * it may contain blocks that don't need to execute anything.
3068 * We therefore don't return this grid size, but instead the
3069 * smallest grid size that ensures that all blocks that actually
3070 * execute code are included in the grid.
3072 * We first extract a description of the grid, i.e., the possible values
3073 * of the block ids, from gen->tiled_sched.
3074 * The block ids are parameters in gen->tiled_sched.
3075 * We simply need to change them into set dimensions.
3077 * Then, for each block dimension, we compute the maximal value of the block id
3080 static __isl_give isl_multi_pw_aff
*extract_grid_size(struct gpu_gen
*gen
,
3081 struct ppcg_kernel
*kernel
)
3086 grid
= isl_union_map_params(isl_union_map_copy(gen
->tiled_sched
));
3087 grid
= isl_set_from_params(grid
);
3088 grid
= isl_set_add_dims(grid
, isl_dim_set
, gen
->n_grid
);
3089 for (i
= 0; i
< gen
->n_grid
; ++i
) {
3093 snprintf(name
, sizeof(name
), "b%d", i
);
3094 pos
= isl_set_find_dim_by_name(grid
, isl_dim_param
, name
);
3096 grid
= isl_set_equate(grid
, isl_dim_param
, pos
, isl_dim_set
, i
);
3097 grid
= isl_set_project_out(grid
, isl_dim_param
, pos
, 1);
3100 return extract_size(grid
, kernel
->context
);
3103 /* Compute the size of a fixed bounding box around the origin and "set",
3104 * where "set" is assumed to contain only non-negative elements,
3105 * and store the results in "size".
3106 * In particular, compute the maximal value of "set" in each direction
3109 static void extract_fixed_size(__isl_take isl_set
*set
, int *size
)
3112 isl_local_space
*ls
;
3115 n
= isl_set_dim(set
, isl_dim_set
);
3116 ls
= isl_local_space_from_space(isl_set_get_space(set
));
3117 obj
= isl_aff_zero_on_domain(ls
);
3118 for (i
= 0; i
< n
; ++i
) {
3121 obj
= isl_aff_set_coefficient_si(obj
, isl_dim_in
, i
, 1);
3122 max
= isl_set_max_val(set
, obj
);
3123 size
[i
] = isl_val_get_num_si(max
) + 1;
3125 obj
= isl_aff_set_coefficient_si(obj
, isl_dim_in
, i
, 0);
3131 /* Compute the effective block size as a list of the sizes in each dimension
3132 * and store the sizes in kernel->block_dim.
3134 * The block size specified by the user or set by default
3135 * in read_block_sizes() and applied in thread_tile_schedule(),
3136 * may be too large for the given code in the sense that
3137 * it may contain threads that don't need to execute anything.
3138 * We therefore don't store this block size in kernel->block_dim,
3139 * but instead the smallest block size that ensures that all threads
3140 * that actually execute code are included in the block.
3142 * The current implementation eliminates all parameters, ensuring
3143 * that the size is a fixed constant in each dimension.
3144 * In principle we could also compute parametric sizes.
3145 * We would have to make sure to project out all b%d and t%d parameters,
3148 static void extract_block_size(struct gpu_gen
*gen
, struct ppcg_kernel
*kernel
)
3153 isl_multi_pw_aff
*mpa
;
3155 block
= isl_union_map_params(isl_union_map_copy(gen
->local_sched
));
3156 block
= isl_set_from_params(block
);
3157 block
= isl_set_add_dims(block
, isl_dim_set
, gen
->n_block
);
3158 kernel
->n_block
= gen
->n_block
;
3159 for (i
= 0; i
< gen
->n_block
; ++i
) {
3163 snprintf(name
, sizeof(name
), "t%d", i
);
3164 pos
= isl_set_find_dim_by_name(block
, isl_dim_param
, name
);
3166 block
= isl_set_equate(block
, isl_dim_param
, pos
,
3169 nparam
= isl_set_dim(block
, isl_dim_param
);
3170 block
= isl_set_project_out(block
, isl_dim_param
, 0, nparam
);
3172 extract_fixed_size(block
, kernel
->block_dim
);
3175 void ppcg_kernel_free(void *user
)
3177 struct ppcg_kernel
*kernel
= user
;
3183 isl_multi_pw_aff_free(kernel
->grid_size
);
3184 isl_set_free(kernel
->context
);
3185 isl_union_set_free(kernel
->arrays
);
3186 isl_space_free(kernel
->space
);
3187 isl_ast_node_free(kernel
->tree
);
3189 for (i
= 0; i
< kernel
->n_array
; ++i
)
3190 isl_pw_aff_list_free(kernel
->array
[i
].bound
);
3191 free(kernel
->array
);
3193 for (i
= 0; i
< kernel
->n_var
; ++i
) {
3194 free(kernel
->var
[i
].name
);
3195 isl_vec_free(kernel
->var
[i
].size
);
3202 static void create_kernel_var(isl_ctx
*ctx
, struct gpu_array_ref_group
*group
,
3203 struct ppcg_kernel_var
*var
)
3206 struct gpu_array_tile
*tile
;
3210 var
->array
= group
->array
;
3212 tile
= group
->private_tile
;
3213 var
->type
= ppcg_access_private
;
3215 tile
= group
->shared_tile
;
3216 var
->type
= ppcg_access_shared
;
3219 p
= isl_printer_to_str(ctx
);
3220 p
= print_array_name(p
, group
);
3221 var
->name
= isl_printer_get_str(p
);
3222 isl_printer_free(p
);
3224 var
->size
= isl_vec_alloc(ctx
, group
->array
->n_index
);
3226 for (j
= 0; j
< group
->array
->n_index
; ++j
)
3227 var
->size
= isl_vec_set_element_val(var
->size
, j
,
3228 isl_val_copy(tile
->bound
[j
].size
));
3231 static void create_kernel_vars(struct gpu_gen
*gen
, struct ppcg_kernel
*kernel
)
3236 for (i
= 0; i
< gen
->prog
->n_array
; ++i
) {
3237 struct gpu_array_info
*array
= &gen
->prog
->array
[i
];
3239 for (j
= 0; j
< array
->n_group
; ++j
) {
3240 struct gpu_array_ref_group
*group
= array
->groups
[j
];
3241 if (group
->private_tile
|| group
->shared_tile
)
3247 kernel
->var
= isl_calloc_array(gen
->ctx
, struct ppcg_kernel_var
, n
);
3248 assert(kernel
->var
);
3251 for (i
= 0; i
< gen
->prog
->n_array
; ++i
) {
3252 struct gpu_array_info
*array
= &gen
->prog
->array
[i
];
3254 for (j
= 0; j
< array
->n_group
; ++j
) {
3255 struct gpu_array_ref_group
*group
= array
->groups
[j
];
3256 if (!group
->private_tile
&& !group
->shared_tile
)
3258 create_kernel_var(gen
->ctx
, group
, &kernel
->var
[n
]);
3264 /* The sizes of the arrays on the host that have been computed by
3265 * extract_array_info may depend on the parameters. Use the extra
3266 * constraints on the parameters that are valid at "host_domain"
3267 * to simplify these expressions and store the results in kernel->array.
3269 * We only need these localized bounds for arrays that are accessed
3270 * by the current kernel. If we have found at least one reference group
3271 * then the array is accessed by the kernel. If the array has compound
3272 * elements then we skipped the construction of array reference groups.
3274 static void localize_bounds(struct gpu_gen
*gen
, struct ppcg_kernel
*kernel
,
3275 __isl_keep isl_set
*host_domain
)
3280 kernel
->array
= isl_calloc_array(gen
->ctx
,
3281 struct gpu_local_array_info
, gen
->prog
->n_array
);
3282 assert(kernel
->array
);
3283 kernel
->n_array
= gen
->prog
->n_array
;
3285 context
= isl_set_copy(host_domain
);
3286 context
= isl_set_params(context
);
3288 for (i
= 0; i
< gen
->prog
->n_array
; ++i
) {
3289 struct gpu_array_info
*array
= &gen
->prog
->array
[i
];
3290 isl_pw_aff_list
*local
;
3292 if (array
->n_group
== 0 && !array
->has_compound_element
)
3295 local
= isl_pw_aff_list_alloc(gen
->ctx
, array
->n_index
);
3297 for (j
= 0; j
< array
->n_index
; ++j
) {
3300 pwaff
= isl_pw_aff_copy(array
->bound
[j
]);
3301 pwaff
= isl_pw_aff_gist(pwaff
, isl_set_copy(context
));
3302 local
= isl_pw_aff_list_add(local
, pwaff
);
3305 kernel
->array
[i
].bound
= local
;
3307 isl_set_free(context
);
3310 /* Find the element in gen->stmt that has the given "id".
3311 * Return NULL if no such gpu_stmt can be found.
3313 static struct gpu_stmt
*find_stmt(struct gpu_prog
*prog
, __isl_keep isl_id
*id
)
3317 for (i
= 0; i
< prog
->n_stmts
; ++i
) {
3318 if (id
== prog
->stmts
[i
].id
)
3322 return i
< prog
->n_stmts
? &prog
->stmts
[i
] : NULL
;
3325 /* Set gen->tile_len and gen->n_parallel to those of the statement
3326 * affected by the first map (part of the schedule)
3327 * on which this function is called.
3328 * Because of the way the schedule is constructed, the other statements
3329 * in the list, if any, should have the same values for these properties.
3331 static int extract_tile_len(__isl_take isl_map
*map
, void *user
)
3333 struct gpu_gen
*gen
= (struct gpu_gen
*) user
;
3335 struct gpu_stmt
*stmt
;
3337 id
= isl_map_get_tuple_id(map
, isl_dim_in
);
3338 stmt
= find_stmt(gen
->prog
, id
);
3344 isl_die(gen
->ctx
, isl_error_unknown
,
3345 "statement not found", return -1);
3347 gen
->tile_len
= stmt
->tile_len
;
3348 gen
->n_parallel
= stmt
->n_parallel
;
3353 void ppcg_kernel_stmt_free(void *user
)
3356 struct ppcg_kernel_stmt
*stmt
= user
;
3361 switch (stmt
->type
) {
3362 case ppcg_kernel_copy
:
3363 isl_ast_expr_free(stmt
->u
.c
.index
);
3364 isl_ast_expr_free(stmt
->u
.c
.local_index
);
3366 case ppcg_kernel_domain
:
3367 isl_id_to_ast_expr_free(stmt
->u
.d
.ref2expr
);
3369 case ppcg_kernel_sync
:
3376 /* Set the options of "context" to
3378 * { space -> [x] : x >= first }
3380 static __isl_give isl_ast_build
*set_unroll(
3381 __isl_take isl_ast_build
*build
, __isl_take isl_space
*space
,
3388 ctx
= isl_ast_build_get_ctx(build
);
3390 space
= isl_space_from_domain(space
);
3391 space
= isl_space_add_dims(space
, isl_dim_out
, 1);
3392 space
= isl_space_set_tuple_name(space
, isl_dim_out
, "unroll");
3393 unroll
= isl_map_universe(space
);
3394 unroll
= isl_map_lower_bound_si(unroll
, isl_dim_out
, 0, first
);
3395 opt
= isl_union_map_from_map(unroll
);
3397 build
= isl_ast_build_set_options(build
, opt
);
3402 /* Return a list of isl_ids of the form "prefix%d".
3404 static __isl_give isl_id_list
*generate_names(isl_ctx
*ctx
,
3405 int n
, const char *prefix
)
3411 names
= isl_id_list_alloc(ctx
, n
);
3412 for (i
= 0; i
< n
; ++i
) {
3415 snprintf(name
, sizeof(name
), "%s%d", prefix
, i
);
3416 id
= isl_id_alloc(ctx
, name
, NULL
);
3417 names
= isl_id_list_add(names
, id
);
3423 /* Extend the schedule "schedule" with the part of "extension"
3424 * starting at "first" up to "len".
3426 static __isl_give isl_union_map
*extend_schedule(
3427 __isl_take isl_union_map
*schedule
,
3428 __isl_take isl_union_map
*extension
, int first
, int len
)
3432 isl_union_map
*umap
;
3435 space
= isl_union_map_get_space(schedule
);
3436 space
= isl_space_set_from_params(space
);
3437 space
= isl_space_add_dims(space
, isl_dim_set
, len
);
3438 proj
= isl_set_identity(isl_set_universe(space
));
3439 proj
= isl_map_project_out(proj
, isl_dim_out
, 0, first
);
3440 extension
= isl_union_map_apply_range(extension
,
3441 isl_union_map_from_map(proj
));
3443 schedule
= isl_union_map_range_product(schedule
, extension
);
3448 /* Return the gpu_stmt_access in the list "accesses" that corresponds
3451 static struct gpu_stmt_access
*find_access(struct gpu_stmt_access
*accesses
,
3452 __isl_keep isl_id
*ref_id
)
3454 struct gpu_stmt_access
*access
;
3456 for (access
= accesses
; access
; access
= access
->next
)
3457 if (access
->ref_id
== ref_id
)
3463 /* Return the index of the array called "name" in the list of arrays.
3465 static int find_array_index(struct gpu_gen
*gen
, const char *name
)
3469 for (i
= 0; i
< gen
->prog
->n_array
; ++i
)
3470 if (!strcmp(name
, gen
->prog
->array
[i
].name
))
3476 /* Internal data structure for the index and AST expression transformation
3477 * callbacks for pet_stmt_build_ast_exprs.
3479 * "accesses" is the list of gpu_stmt_access in the statement.
3480 * "iterator_map" expresses the statement iterators in terms of
3481 * the AST loop iterators.
3482 * "sched2shared" expresses the first shared_len dimensions of
3483 * the computed schedule in terms of the AST loop iterators.
3485 * The following fields are set in transform_index and used in transform_expr.
3486 * "array" is the array that is being accessed.
3487 * "global" is set if the global array is accessed (rather than
3488 * shared/private memory).
3489 * "local_array" refers to information on the array specialized
3490 * to the current kernel.
3492 struct ppcg_transform_data
{
3493 struct gpu_gen
*gen
;
3494 struct gpu_stmt_access
*accesses
;
3495 isl_pw_multi_aff
*iterator_map
;
3496 isl_pw_multi_aff
*sched2shared
;
3498 struct gpu_array_info
*array
;
3500 struct gpu_local_array_info
*local_array
;
3503 /* Return the name of the outer array (of structs) accessed by "access".
3505 static const char *get_outer_array_name(__isl_keep isl_map
*access
)
3510 space
= isl_space_range(isl_map_get_space(access
));
3511 while (space
&& isl_space_is_wrapping(space
))
3512 space
= isl_space_domain(isl_space_unwrap(space
));
3513 name
= isl_space_get_tuple_name(space
, isl_dim_set
);
3514 isl_space_free(space
);
3519 /* Index transformation callback for pet_stmt_build_ast_exprs.
3521 * "index" expresses the array indices in terms of statement iterators
3523 * We first reformulate "index" in terms of the AST loop iterators.
3524 * Then we check if we are accessing the global array or
3525 * a shared/private copy. In the former case, we simply return
3526 * the updated index. If "index" is an affine expression rather
3527 * than an array access, then we also return the updated index here.
3529 * If no reference groups have been computed for the array,
3530 * then we can only be accessing the global array.
3532 * Otherwise, we apply the tiling to the index.
3533 * This tiling is of the form
3537 * The index is of the form
3541 * We update the tiling to refer to the AST loop iteratos
3545 * and modify index to keep track of those iterators
3549 * Combining these two yields a tiled index expression in terms
3550 * of the AST loop iterators
3554 static __isl_give isl_multi_pw_aff
*transform_index(
3555 __isl_take isl_multi_pw_aff
*index
, __isl_keep isl_id
*ref_id
,
3558 struct ppcg_transform_data
*data
= user
;
3559 struct gpu_stmt_access
*access
;
3560 struct gpu_array_ref_group
*group
;
3561 struct gpu_array_tile
*tile
;
3562 isl_pw_multi_aff
*iterator_map
;
3566 isl_multi_pw_aff
*tiling
;
3567 isl_pw_multi_aff
*pma
;
3568 isl_multi_pw_aff
*mpa
;
3572 iterator_map
= isl_pw_multi_aff_copy(data
->iterator_map
);
3573 index
= isl_multi_pw_aff_pullback_pw_multi_aff(index
, iterator_map
);
3575 access
= find_access(data
->accesses
, ref_id
);
3578 if (!isl_map_has_tuple_name(access
->access
, isl_dim_out
))
3581 name
= get_outer_array_name(access
->access
);
3582 i
= find_array_index(data
->gen
, name
);
3584 isl_die(isl_multi_pw_aff_get_ctx(index
), isl_error_internal
,
3585 "cannot find array",
3586 return isl_multi_pw_aff_free(index
));
3587 data
->array
= &data
->gen
->prog
->array
[i
];
3588 data
->local_array
= &data
->gen
->kernel
->array
[i
];
3590 if (access
->group
< 0) {
3595 group
= data
->array
->groups
[access
->group
];
3596 tile
= group
->private_tile
;
3598 tile
= group
->shared_tile
;
3599 data
->global
= !tile
;
3603 space
= isl_space_range(isl_multi_pw_aff_get_space(index
));
3604 space
= isl_space_map_from_set(space
);
3605 pma
= isl_pw_multi_aff_identity(space
);
3606 pma
= isl_pw_multi_aff_product(
3607 isl_pw_multi_aff_copy(data
->sched2shared
), pma
);
3608 tiling
= isl_multi_pw_aff_from_multi_aff(
3609 isl_multi_aff_copy(tile
->tiling
));
3610 tiling
= isl_multi_pw_aff_pullback_pw_multi_aff(tiling
, pma
);
3612 space
= isl_space_domain(isl_multi_pw_aff_get_space(index
));
3613 space
= isl_space_map_from_set(space
);
3614 mpa
= isl_multi_pw_aff_identity(space
);
3615 index
= isl_multi_pw_aff_range_product(mpa
, index
);
3616 index
= isl_multi_pw_aff_pullback_multi_pw_aff(tiling
, index
);
3621 /* Dereference "expr" by adding an index [0].
3622 * The original "expr" is assumed not to have any indices.
3624 * If "expr" is a member access, then the dereferencing needs
3625 * to be applied to the structure argument of this member access.
3627 static __isl_give isl_ast_expr
*dereference(__isl_take isl_ast_expr
*expr
)
3631 isl_ast_expr_list
*list
;
3633 if (isl_ast_expr_get_op_type(expr
) == isl_ast_op_member
) {
3636 arg
= isl_ast_expr_get_op_arg(expr
, 0);
3637 arg
= dereference(arg
);
3638 expr
= isl_ast_expr_set_op_arg(expr
, 0, arg
);
3643 ctx
= isl_ast_expr_get_ctx(expr
);
3644 res
= isl_ast_expr_from_val(isl_val_zero(ctx
));
3645 list
= isl_ast_expr_list_from_ast_expr(res
);
3646 res
= isl_ast_expr_get_op_arg(expr
, 0);
3647 res
= isl_ast_expr_access(res
, list
);
3648 isl_ast_expr_free(expr
);
3653 /* Linearize the index expression "expr" based on the array bounds
3656 * That is, transform expression
3658 * A[i_0][i_1]...[i_n]
3662 * A[(..((i_0 * b_1 + i_1) ... ) * b_n + i_n]
3664 * where b_0, b_1, ..., b_n are the bounds on the array.
3666 * If the base of "expr" is a member access, then the linearization needs
3667 * to be applied to the structure argument of this member access.
3669 __isl_give isl_ast_expr
*gpu_local_array_info_linearize_index(
3670 struct gpu_local_array_info
*array
, __isl_take isl_ast_expr
*expr
)
3677 isl_ast_expr_list
*list
;
3678 isl_ast_build
*build
;
3680 arg0
= isl_ast_expr_get_op_arg(expr
, 0);
3681 if (isl_ast_expr_get_type(arg0
) == isl_ast_expr_op
&&
3682 isl_ast_expr_get_op_type(arg0
) == isl_ast_op_member
) {
3685 arg
= isl_ast_expr_get_op_arg(arg0
, 0);
3686 arg
= gpu_local_array_info_linearize_index(array
, arg
);
3687 arg0
= isl_ast_expr_set_op_arg(arg0
, 0, arg
);
3688 expr
= isl_ast_expr_set_op_arg(expr
, 0, arg0
);
3692 isl_ast_expr_free(arg0
);
3694 ctx
= isl_ast_expr_get_ctx(expr
);
3695 context
= isl_set_universe(isl_space_params_alloc(ctx
, 0));
3696 build
= isl_ast_build_from_context(context
);
3698 n
= isl_ast_expr_get_op_n_arg(expr
);
3699 res
= isl_ast_expr_get_op_arg(expr
, 1);
3700 for (i
= 2; i
< n
; ++i
) {
3701 isl_pw_aff
*bound_i
;
3702 isl_ast_expr
*expr_i
;
3704 bound_i
= isl_pw_aff_list_get_pw_aff(array
->bound
, i
- 1);
3705 expr_i
= isl_ast_build_expr_from_pw_aff(build
, bound_i
);
3706 res
= isl_ast_expr_mul(res
, expr_i
);
3707 expr_i
= isl_ast_expr_get_op_arg(expr
, i
);
3708 res
= isl_ast_expr_add(res
, expr_i
);
3711 isl_ast_build_free(build
);
3713 list
= isl_ast_expr_list_from_ast_expr(res
);
3714 res
= isl_ast_expr_get_op_arg(expr
, 0);
3715 res
= isl_ast_expr_access(res
, list
);
3717 isl_ast_expr_free(expr
);
3722 /* AST expression transformation callback for pet_stmt_build_ast_exprs.
3724 * If the AST expression refers to a global scalar that is not
3725 * a read-only scalar, then its address was passed to the kernel and
3726 * we need to dereference it.
3728 * If the AST expression refers to an access to a global array,
3729 * then we linearize the access exploiting the bounds in data->local_array.
3731 static __isl_give isl_ast_expr
*transform_expr(__isl_take isl_ast_expr
*expr
,
3732 __isl_keep isl_id
*id
, void *user
)
3734 struct ppcg_transform_data
*data
= user
;
3738 if (gpu_array_is_read_only_scalar(data
->array
))
3742 if (data
->array
->n_index
== 0)
3743 return dereference(expr
);
3744 if (!data
->array
->linearize
)
3747 return gpu_local_array_info_linearize_index(data
->local_array
, expr
);
3750 /* This function is called for each instance of a user statement
3753 * We attach a struct ppcg_kernel_stmt to the "node", containing
3754 * a computed AST expression for each access.
3755 * These AST expressions are computed from iterator_map,
3756 * which expresses the domain
3757 * elements in terms of the generated loops, and sched2shared,
3758 * which expresses the first shared_len dimensions of the schedule
3759 * computed by PPCG in terms of the generated loops.
3761 static __isl_give isl_ast_node
*at_each_domain(__isl_take isl_ast_node
*node
,
3762 __isl_keep isl_ast_build
*build
, void *user
)
3764 struct ppcg_transform_data data
;
3765 struct gpu_gen
*gen
= (struct gpu_gen
*) user
;
3766 struct ppcg_kernel_stmt
*stmt
;
3768 isl_pw_multi_aff
*sched2shared
;
3770 isl_pw_multi_aff
*iterator_map
;
3771 isl_ast_expr
*expr
, *arg
;
3772 isl_union_map
*schedule
;
3774 struct gpu_stmt_access
*access
;
3776 stmt
= isl_calloc_type(gen
->ctx
, struct ppcg_kernel_stmt
);
3778 return isl_ast_node_free(node
);
3780 expr
= isl_ast_node_user_get_expr(node
);
3781 arg
= isl_ast_expr_get_op_arg(expr
, 0);
3782 id
= isl_ast_expr_get_id(arg
);
3784 schedule
= isl_ast_build_get_schedule(build
);
3785 map
= isl_map_reverse(isl_map_from_union_map(schedule
));
3786 iterator_map
= isl_pw_multi_aff_from_map(map
);
3787 sched2shared
= compute_sched_to_shared(gen
,
3788 isl_pw_multi_aff_copy(iterator_map
));
3790 stmt
->type
= ppcg_kernel_domain
;
3791 stmt
->u
.d
.stmt
= find_stmt(gen
->prog
, id
);
3792 if (!stmt
->u
.d
.stmt
)
3796 data
.accesses
= stmt
->u
.d
.stmt
->accesses
;
3797 data
.iterator_map
= iterator_map
;
3798 data
.sched2shared
= sched2shared
;
3799 stmt
->u
.d
.ref2expr
= pet_stmt_build_ast_exprs(stmt
->u
.d
.stmt
->stmt
,
3800 build
, &transform_index
, &data
,
3801 &transform_expr
, &data
);
3804 isl_pw_multi_aff_free(iterator_map
);
3805 isl_pw_multi_aff_free(sched2shared
);
3806 isl_ast_expr_free(arg
);
3807 isl_ast_expr_free(expr
);
3809 id
= isl_id_alloc(gen
->ctx
, NULL
, stmt
);
3810 id
= isl_id_set_free_user(id
, &ppcg_kernel_stmt_free
);
3811 return isl_ast_node_set_annotation(node
, id
);
3814 isl_pw_multi_aff_free(iterator_map
);
3815 ppcg_kernel_stmt_free(stmt
);
3816 isl_pw_multi_aff_free(sched2shared
);
3817 return isl_ast_node_free(node
);
3820 /* This function is called when code has been generated for the shared
3821 * tile loops. The "schedule" refers only to the original statements.
3823 * We extend the schedule with that part of gen->local_sched that hasn't
3824 * been taken into account yet. This introduces parameters referring
3825 * to thread ids in the schedule, so we add them (with the appropriate
3826 * bounds to the context as well).
3827 * Finally, we set the appropriate unrolling options
3828 * if gen->first_unroll is set.
3830 static __isl_give isl_ast_node
*create_domain_leaf(
3831 __isl_take isl_union_map
*schedule
, __isl_take isl_ast_build
*build
,
3834 struct gpu_gen
*gen
= (struct gpu_gen
*) user
;
3836 isl_union_map
*sched
;
3839 isl_id_list
*iterators
;
3842 schedule
= extend_schedule(schedule
,
3843 isl_union_map_copy(gen
->local_sched
),
3844 gen
->shared_len
, gen
->thread_tiled_len
);
3846 space
= isl_ast_build_get_schedule_space(build
);
3847 set
= isl_set_universe(space
);
3848 set
= add_bounded_parameters(set
, gen
->kernel
->n_block
,
3849 gen
->kernel
->block_dim
, "t");
3850 build
= isl_ast_build_restrict(build
, set
);
3852 n
= gen
->thread_tiled_len
- gen
->shared_len
;
3854 if (gen
->first_unroll
>= 0) {
3855 space
= isl_space_set_alloc(gen
->ctx
, 0, n
);
3856 build
= set_unroll(build
, space
, gen
->first_unroll
);
3858 iterators
= generate_names(gen
->ctx
, n
, "c");
3859 build
= isl_ast_build_set_iterators(build
, iterators
);
3860 build
= isl_ast_build_set_at_each_domain(build
, &at_each_domain
, gen
);
3861 tree
= isl_ast_build_ast_from_schedule(build
, schedule
);
3862 isl_ast_build_free(build
);
3867 /* This function is called for each statement node in the AST of the code
3868 * for copying to or from shared/private memory.
3869 * Attach a pointer to a ppcg_kernel_stmt representing the copy
3870 * statement to the node.
3871 * The statement name is "read" or "write", depending on whether we are
3872 * reading from global memory or writing to global memory.
3873 * The name of the T space is {shared,private}_<array>.
3875 * The schedule is of the form
3879 * where A refers to a piece of an array and T to the corresponding
3880 * shifted tile. We split this schedule into mappings L -> A and L -> T
3881 * and store the corresponding expressions in stmt->index and stmt->local_index,
3882 * where stmt points to the ppcg_kernel_stmt that is attached to the node.
3884 static __isl_give isl_ast_node
*attach_copy_stmt(__isl_take isl_ast_node
*node
,
3885 __isl_keep isl_ast_build
*build
, void *user
)
3887 struct gpu_gen
*gen
= (struct gpu_gen
*) user
;
3888 struct ppcg_kernel_stmt
*stmt
;
3892 isl_map
*access
, *local_access
, *map
;
3893 isl_pw_multi_aff
*pma
;
3897 stmt
= isl_calloc_type(gen
->ctx
, struct ppcg_kernel_stmt
);
3899 return isl_ast_node_free(node
);
3901 access
= isl_map_from_union_map(isl_ast_build_get_schedule(build
));
3902 type
= isl_map_get_tuple_name(access
, isl_dim_in
);
3903 stmt
->u
.c
.read
= !strcmp(type
, "read");
3904 access
= isl_map_reverse(access
);
3905 space
= isl_space_unwrap(isl_space_range(isl_map_get_space(access
)));
3906 local_access
= isl_map_copy(access
);
3908 map
= isl_map_domain_map(isl_map_universe(isl_space_copy(space
)));
3909 id
= isl_map_get_tuple_id(access
, isl_dim_out
);
3910 map
= isl_map_set_tuple_id(map
, isl_dim_in
, id
);
3911 access
= isl_map_apply_range(access
, map
);
3912 pma
= isl_pw_multi_aff_from_map(access
);
3913 expr
= isl_ast_build_access_from_pw_multi_aff(build
, pma
);
3914 stmt
->u
.c
.index
= expr
;
3916 map
= isl_map_range_map(isl_map_universe(space
));
3917 id
= isl_map_get_tuple_id(local_access
, isl_dim_out
);
3918 map
= isl_map_set_tuple_id(map
, isl_dim_in
, id
);
3919 local_access
= isl_map_apply_range(local_access
, map
);
3920 pma
= isl_pw_multi_aff_from_map(local_access
);
3921 expr
= isl_ast_build_access_from_pw_multi_aff(build
, pma
);
3922 stmt
->u
.c
.local_index
= expr
;
3924 stmt
->u
.c
.array
= gen
->copy_group
->array
;
3925 array_index
= stmt
->u
.c
.array
- gen
->prog
->array
;
3926 stmt
->u
.c
.local_array
= &gen
->kernel
->array
[array_index
];
3927 stmt
->type
= ppcg_kernel_copy
;
3929 id
= isl_id_alloc(gen
->ctx
, NULL
, stmt
);
3930 id
= isl_id_set_free_user(id
, &ppcg_kernel_stmt_free
);
3931 return isl_ast_node_set_annotation(node
, id
);
3934 /* Given a schedule of the form
3938 * (with S the first shared_len dimensions of the computed schedule,
3939 * A the array and L the schedule correponding to the generated loops),
3940 * indicating where to copy the array elements that need to be copied,
3941 * construct code for performing the copying.
3943 * "group" is the array reference group that is being copied
3944 * "type" is either "read" or "write"
3945 * private is set if copying needs to be performed to/from registers
3947 * We first construct a mapping to a shifted tile of the array,
3949 * [S -> A] -> T(S,A) (1)
3951 * If private is set, then we also use this mapping as a schedule
3952 * (which is already thread-specific and will be completely unrolled).
3953 * Otherwise, we wrap/tile the range over the threads.
3956 * [S -> A] -> T'(S,A)
3958 * Combined with the given schedule, we have
3960 * [S -> A] -> [L -> T'(S,A)] (2)
3962 * From the shifted tile mapping, we construct a mapping
3964 * [S -> A] -> [A -> T(S,A)]
3966 * and apply it to the schedule (2), obtaining
3968 * [A -> T(S(L),A)] -> [L -> T'(S(L),A)]
3970 * Note that we can project out S because it is uniquely defined by L.
3972 static __isl_give isl_ast_node
*copy_access(struct gpu_gen
*gen
,
3973 __isl_take isl_map
*sched
,
3974 const char *type
, struct gpu_array_ref_group
*group
,
3975 __isl_take isl_ast_build
*build
, int private)
3979 isl_map
*schedule
, *shift
, *map
;
3981 isl_id_list
*iterators
;
3984 shift
= shift_access(group
);
3986 schedule
= isl_map_copy(shift
);
3987 schedule
= isl_map_reset_tuple_id(schedule
, isl_dim_out
);
3989 schedule
= tile_access_schedule(gen
, schedule
);
3991 n
= isl_map_dim(schedule
, isl_dim_out
);
3992 set
= isl_set_universe(isl_ast_build_get_schedule_space(build
));
3993 set
= add_bounded_parameters(set
, gen
->kernel
->n_block
,
3994 gen
->kernel
->block_dim
, "t");
3996 schedule
= isl_map_range_product(sched
, schedule
);
3998 space
= isl_space_domain(isl_map_get_space(shift
));
3999 map
= isl_map_range_map(isl_map_universe(isl_space_unwrap(space
)));
4000 map
= isl_map_range_product(map
, shift
);
4002 schedule
= isl_map_apply_domain(schedule
, map
);
4004 schedule
= isl_map_set_tuple_name(schedule
, isl_dim_in
, type
);
4006 build
= isl_ast_build_restrict(build
, set
);
4008 gen
->copy_group
= group
;
4011 space
= isl_space_range(isl_map_get_space(schedule
));
4012 space
= isl_space_range(isl_space_unwrap(space
));
4013 build
= set_unroll(build
, space
, 0);
4015 iterators
= generate_names(gen
->ctx
, n
, "c");
4016 build
= isl_ast_build_set_iterators(build
, iterators
);
4017 build
= isl_ast_build_set_at_each_domain(build
, &attach_copy_stmt
, gen
);
4018 tree
= isl_ast_build_ast_from_schedule(build
,
4019 isl_union_map_from_map(schedule
));
4020 isl_ast_build_free(build
);
4025 /* Return code for reading into or writing from shared memory
4026 * the given array reference group.
4028 * If we are performing a read from global memory to shared memory and
4029 * if the array involved is not a scalar, then we copy
4030 * the entire tile to shared memory. This may result in some extra
4031 * elements getting copied, but it should lead to simpler code
4032 * (which means that fewer registers may be needed) and less divergence.
4034 * Otherwise, we only copy the elements that will be read or have been written
4038 * The input "sched" is of the form.
4042 * with S the first shared_len dimensions of the computed schedule,
4043 * A the array and L the schedule correponding to the generated loops.
4045 * We first drop "type",
4049 * If the above conditions are satisfied, we project out A,
4054 * and then introduce the group tile [S -> T], resulting in
4058 static __isl_give isl_ast_node
*copy_group_shared_accesses(
4059 struct gpu_gen
*gen
, struct gpu_array_ref_group
*group
,
4060 __isl_take isl_map
*sched
, __isl_take isl_ast_build
*build
)
4064 isl_union_map
*access
;
4066 type
= isl_map_get_tuple_name(sched
, isl_dim_in
);
4067 read
= !strcmp(type
, "read");
4069 sched
= isl_map_reset_tuple_id(sched
, isl_dim_in
);
4071 if (read
&& !gpu_array_is_scalar(group
->array
)) {
4075 space
= isl_space_domain(isl_map_get_space(sched
));
4076 space
= isl_space_unwrap(space
);
4077 map
= isl_map_domain_map(isl_map_universe(space
));
4078 sched
= isl_map_apply_domain(sched
, map
);
4080 map
= group_tile(group
);
4081 map
= isl_map_reverse(isl_map_domain_map(map
));
4082 sched
= isl_map_apply_domain(sched
, map
);
4085 return copy_access(gen
, sched
, type
, group
, build
, 0);
4088 /* Return code for reading into or writing from private memory
4089 * the given array reference group.
4091 * Let S be the first shared_len dimensions of the computed schedule,
4092 * D the iteration domains, A the array and L the schedule correponding
4093 * to the generated loops.
4094 * "sched" is of the form
4098 * where type is either "read" or "write".
4099 * We apply the privatization D -> S(t), with t the thread ids,
4100 * to the access relation D -> A to obtain the privatized access relation
4104 * We drop the type from "sched" and intersect with the privatized access
4105 * relation to obtain
4109 static __isl_give isl_ast_node
*copy_group_private_accesses(
4110 struct gpu_gen
*gen
, struct gpu_array_ref_group
*group
,
4111 __isl_take isl_map
*sched
, __isl_take isl_ast_build
*build
)
4115 isl_union_map
*priv
;
4116 isl_union_map
*access
;
4117 isl_map
*access_map
;
4119 type
= isl_map_get_tuple_name(sched
, isl_dim_in
);
4120 read
= !strcmp(type
, "read");
4122 priv
= isl_union_map_from_map(isl_map_copy(gen
->privatization
));
4123 priv
= isl_union_map_apply_range(isl_union_map_copy(gen
->shared_sched
),
4126 access
= group_access_relation(group
, read
, !read
);
4127 access
= isl_union_map_apply_domain(access
, priv
);
4128 access_map
= isl_map_from_union_map(access
);
4130 sched
= isl_map_reset_tuple_id(sched
, isl_dim_in
);
4131 sched
= isl_map_intersect_domain(sched
, isl_map_wrap(access_map
));
4133 return copy_access(gen
, sched
, type
, group
, build
, 1);
4136 /* Return code for reading into or writing from shared or private memory.
4138 * "schedule" is of the form
4142 * with S be the first shared_len dimensions of the computed schedule,
4143 * A the array and L the schedule correponding to the generated loops.
4144 * The array reference group is attached to "type".
4146 static __isl_give isl_ast_node
*create_access_leaf(
4147 struct gpu_gen
*gen
, __isl_take isl_map
*schedule
,
4148 __isl_take isl_ast_build
*build
)
4150 struct gpu_array_ref_group
*group
;
4153 id
= isl_map_get_tuple_id(schedule
, isl_dim_in
);
4154 group
= isl_id_get_user(id
);
4157 if (group
->private_tile
)
4158 return copy_group_private_accesses(gen
, group
, schedule
,
4161 return copy_group_shared_accesses(gen
, group
, schedule
,
4165 /* Create a domain node representing a synchronization.
4167 static __isl_give isl_ast_node
*create_sync_leaf(
4168 struct gpu_gen
*gen
, __isl_take isl_map
*schedule
,
4169 __isl_take isl_ast_build
*build
)
4171 struct ppcg_kernel_stmt
*stmt
;
4177 isl_map_free(schedule
);
4179 stmt
= isl_calloc_type(gen
->ctx
, struct ppcg_kernel_stmt
);
4183 stmt
->type
= ppcg_kernel_sync
;
4185 space
= isl_ast_build_get_schedule_space(build
);
4186 space
= isl_space_from_domain(space
);
4187 space
= isl_space_set_tuple_name(space
, isl_dim_out
, "sync");
4188 expr
= isl_ast_build_call_from_pw_multi_aff(build
,
4189 isl_pw_multi_aff_from_multi_aff(isl_multi_aff_zero(space
)));
4190 node
= isl_ast_node_alloc_user(expr
);
4191 isl_ast_build_free(build
);
4193 id
= isl_id_alloc(gen
->ctx
, NULL
, stmt
);
4194 id
= isl_id_set_free_user(id
, &ppcg_kernel_stmt_free
);
4195 return isl_ast_node_set_annotation(node
, id
);
4198 /* This function is called during the code generation at the point
4199 * where the schedule domain element is completely determined by
4200 * the generated code. The input schedule contains the original
4201 * statements as well as synchronization and copy "statements".
4202 * The latter are scheduled at different points than any of the original
4203 * statements, so they will only arrive here in isolation.
4205 * If the current schedule only refers to a single statement,
4206 * we check if it is a copy or synchronization statement and
4207 * call the appropriate functions.
4208 * Otherwise, we assume we are dealing with the original statements
4209 * and we call create_domain_leaf.
4211 static __isl_give isl_ast_node
*create_kernel_leaf(
4212 __isl_take isl_ast_build
*build
, void *user
)
4214 struct gpu_gen
*gen
= (struct gpu_gen
*) user
;
4216 isl_union_map
*schedule
;
4219 schedule
= isl_ast_build_get_schedule(build
);
4221 if (isl_union_map_n_map(schedule
) != 1)
4222 return create_domain_leaf(schedule
, build
, user
);
4224 map
= isl_map_from_union_map(schedule
);
4225 name
= isl_map_get_tuple_name(map
, isl_dim_in
);
4226 if (!strcmp(name
, "read") || !strcmp(name
, "write"))
4227 return create_access_leaf(gen
, map
, build
);
4228 if (!strcmp(name
, "sync"))
4229 return create_sync_leaf(gen
, map
, build
);
4231 return create_domain_leaf(isl_union_map_from_map(map
), build
, user
);
4234 /* Mark all odd schedule dimensions as "atomic" (when the even dimensions
4235 * have value 0) and all even schedule dimensions as "unroll".
4237 * That is, the options look as follows
4239 * { [0, b, 0, d, ..., 0] -> atomic[i] : exists a : i = 2 a + 1;
4240 * [a, b, c, d, ..., z] -> unroll[i] : exists a : i = 2 a }
4242 * The even positions are used to be able to schedule copying blocks
4243 * and synchronization before or after each level of the shared memory
4244 * tile loops and we want to make sure that code for these is generated
4245 * separately (within each level).
4247 static __isl_give isl_ast_build
*set_atomic_and_unroll(
4248 __isl_take isl_ast_build
*build
,
4249 __isl_take isl_space
*space
, int sched_len
)
4255 isl_local_space
*ls
;
4258 ctx
= isl_ast_build_get_ctx(build
);
4260 space
= isl_space_params(space
);
4261 space
= isl_space_add_dims(space
, isl_dim_set
, sched_len
);
4262 space
= isl_space_from_domain(space
);
4263 space
= isl_space_add_dims(space
, isl_dim_out
, 2);
4264 map
= isl_map_universe(isl_space_copy(space
));
4265 for (i
= 0; i
< sched_len
; i
+= 2)
4266 map
= isl_map_fix_si(map
, isl_dim_in
, i
, 0);
4267 ls
= isl_local_space_from_space(isl_map_get_space(map
));
4268 c
= isl_equality_alloc(ls
);
4269 c
= isl_constraint_set_coefficient_si(c
, isl_dim_out
, 0, 1);
4270 c
= isl_constraint_set_coefficient_si(c
, isl_dim_out
, 1, 2);
4271 c
= isl_constraint_set_constant_si(c
, 1);
4272 map
= isl_map_add_constraint(map
, c
);
4273 map
= isl_map_project_out(map
, isl_dim_out
, 1, 1);
4274 map
= isl_map_set_tuple_name(map
, isl_dim_out
, "atomic");
4275 opt
= isl_union_map_from_map(map
);
4277 map
= isl_map_universe(space
);
4278 ls
= isl_local_space_from_space(isl_map_get_space(map
));
4279 c
= isl_equality_alloc(ls
);
4280 c
= isl_constraint_set_coefficient_si(c
, isl_dim_out
, 0, 1);
4281 c
= isl_constraint_set_coefficient_si(c
, isl_dim_out
, 1, 2);
4282 map
= isl_map_add_constraint(map
, c
);
4283 map
= isl_map_project_out(map
, isl_dim_out
, 1, 1);
4284 map
= isl_map_set_tuple_name(map
, isl_dim_out
, "unroll");
4285 opt
= isl_union_map_add_map(opt
, map
);
4287 build
= isl_ast_build_set_options(build
, opt
);
4292 /* Return a map that maps a space of dimension gen->shared_len
4293 * to its last dimensions starting at gen->tile_first.
4294 * The range is of dimension
4296 * 2 * (gen->shared_len - gen->tile_first) + 1
4298 * The input dimensions are mapped to the odd dimensions in the output,
4299 * while the even dimensions (except 2*pos) are fixed to 0.
4300 * Output dimension 2*pos (if pos >= 0) is fixed to "val".
4301 * If pos >= 0, then only the pos first dimensions starting at gen->tile_first
4302 * are mapped to the output. The remaining input dimensions are projected
4303 * out and the corresponding output dimensions are fixed to 0.
4305 static __isl_give isl_map
*insert_even(struct gpu_gen
*gen
,
4306 __isl_take isl_space
*space
, int pos
, int val
)
4311 space
= isl_space_set_from_params(space
);
4312 space
= isl_space_add_dims(space
, isl_dim_set
, gen
->shared_len
);
4313 space
= isl_space_map_from_set(space
);
4314 proj
= isl_map_identity(space
);
4315 proj
= isl_map_project_out(proj
, isl_dim_out
, 0, gen
->tile_first
);
4316 n
= gen
->shared_len
- gen
->tile_first
;
4317 for (i
= 0; i
<= n
; ++i
) {
4318 proj
= isl_map_insert_dims(proj
, isl_dim_out
, 2 * i
, 1);
4320 proj
= isl_map_fix_si(proj
, isl_dim_out
, 2 * i
, val
);
4322 proj
= isl_map_fix_si(proj
, isl_dim_out
, 2 * i
, 0);
4328 proj
= isl_map_eliminate(proj
, isl_dim_in
, gen
->tile_first
+ pos
,
4329 gen
->shared_len
- (gen
->tile_first
+ pos
));
4330 for (i
= pos
; i
< n
; ++i
)
4331 proj
= isl_map_fix_si(proj
, isl_dim_out
, 2 * i
+ 1, 0);
4336 /* Given the AST context schedule "schedule" and the mapping from
4337 * domains to the shared tile loops "shared_sched", add a schedule
4338 * for a synchronization operation at position "val" of loop level "pos".
4340 * schedule is of the form
4344 * (with D the iteration domains and L the already generated loops),
4345 * while shared_sched is of the form
4349 * We combine them into
4355 * [s_0,...] -> [0,s_{tile_first},0,..., val, 0, 0, ... 0]
4357 * and use the result as a schedule for "sync".
4359 static __isl_give isl_union_map
*add_sync_schedule(struct gpu_gen
*gen
,
4360 __isl_take isl_union_map
*res
, __isl_keep isl_union_map
*schedule
,
4361 __isl_keep isl_union_map
*shared_sched
, int pos
, int val
)
4364 isl_map
*proj
, *map
;
4366 shared_sched
= isl_union_map_copy(shared_sched
);
4367 schedule
= isl_union_map_copy(schedule
);
4369 space
= isl_union_map_get_space(shared_sched
);
4370 schedule
= isl_union_map_apply_domain(shared_sched
, schedule
);
4371 map
= isl_map_from_union_map(schedule
);
4373 proj
= insert_even(gen
, space
, pos
, val
);
4374 map
= isl_map_apply_range(map
, proj
);
4375 map
= isl_map_from_range(isl_map_wrap(map
));
4376 map
= isl_map_set_tuple_name(map
, isl_dim_in
, "sync");
4378 res
= isl_union_map_add_map(res
, map
);
4383 /* Given a set of wrapped references "ref", return the corresponding
4384 * access relations based on the tagged access relations "tagged".
4386 * The elements of "ref" are of the form
4390 * with D an iteration domains and R a reference.
4391 * The elements of "tagged" are of the form
4397 * Extend "tagged" to include the iteration domain in the range, i.e.,
4399 * [D -> R] -> [D -> A]
4401 * apply the result to "ref" and then unwrap the resulting set
4402 * to obtain relations of the form
4406 static __isl_give isl_union_map
*wrapped_reference_to_access(
4407 __isl_take isl_union_set
*ref
, __isl_take isl_union_map
*tagged
)
4409 isl_union_map
*tag2access
;
4411 tag2access
= isl_union_map_copy(tagged
);
4412 tag2access
= isl_union_map_universe(tag2access
);
4413 tag2access
= isl_union_set_unwrap(isl_union_map_domain(tag2access
));
4414 tag2access
= isl_union_map_domain_map(tag2access
);
4415 tag2access
= isl_union_map_range_product(tag2access
, tagged
);
4417 ref
= isl_union_set_coalesce(ref
);
4418 ref
= isl_union_set_apply(ref
, tag2access
);
4420 return isl_union_set_unwrap(ref
);
4423 /* Given an access relation "access" from "group", remove those reads
4424 * if ("read" is 1) or writes (if "read" is 0) that are only needed to
4425 * communicate data within the same iteration of the last_shared dimension
4428 * If the access is a read then it is necessarily an element of
4430 * live_in union (range flow)
4432 * where live_in and flow may be overapproximations.
4433 * If the access is a write then it is necessarily an element of
4435 * live_out union (domain flow)
4437 * In both cases, the access relation is also a subset of
4438 * the group access relation.
4440 * Essentially, we compute the intersection of "access" with either
4442 * live_in union (range non-local-flow)
4446 * live_out union (domain non-local-flow)
4448 * We first construct a relation "local"
4450 * [[D -> R] -> [D' -> R']]
4452 * of pairs of domain iterations accessing the reference group
4453 * and references in the group that are scheduled to the same iteration
4454 * of the last_shared dimension.
4456 * If this relation does not intersect the dataflow dependences,
4457 * then there is nothing we can possibly remove and we simply
4460 * Otherwise, we remove the "local" dataflow dependences from
4461 * the set of all dataflow dependences.
4462 * Note that if the potential dataflow dependences are an overapproximation
4463 * of the actual dataflow dependences, then the result remains an
4464 * overapproximation of the non-local dataflow dependences.
4465 * Copying to/from global memory is only needed for the references
4466 * in the domain/range of the result or for accesses that are live out/in
4467 * for the entire scop.
4469 * We therefore map the domain/range of the "external" relation
4470 * to the corresponding access relation and take the union with
4471 * the live out/in relation.
4473 static __isl_give isl_union_map
*remove_local_accesses(struct gpu_gen
*gen
,
4474 struct gpu_array_ref_group
*group
, __isl_take isl_union_map
*access
,
4478 isl_union_map
*tagger
;
4479 isl_union_set
*domain
;
4481 isl_union_map
*sched
, *local
, *tagged
, *external
;
4482 isl_union_set
*tag_set
;
4485 if (isl_union_map_is_empty(access
))
4488 tagged
= group_tagged_access_relation(group
);
4490 sched
= isl_union_map_copy(gen
->sched
);
4492 space
= isl_union_map_get_space(sched
);
4493 proj
= projection(space
, gen
->untiled_len
, group
->last_shared
+ 1);
4494 sched
= isl_union_map_apply_range(sched
, isl_union_map_from_map(proj
));
4496 tagger
= isl_union_map_copy(gen
->prog
->scop
->tagger
);
4497 domain
= isl_union_map_domain(isl_union_map_copy(tagged
));
4498 tagger
= isl_union_map_intersect_range(tagger
, domain
);
4499 sched
= isl_union_map_apply_domain(sched
, tagger
);
4501 local
= isl_union_map_apply_range(sched
,
4502 isl_union_map_reverse(isl_union_map_copy(sched
)));
4503 local
= isl_union_map_intersect(local
,
4504 isl_union_map_copy(gen
->prog
->scop
->tagged_dep_flow
));
4506 empty
= isl_union_map_is_empty(local
);
4507 if (empty
< 0 || empty
) {
4508 isl_union_map_free(tagged
);
4509 isl_union_map_free(local
);
4511 return isl_union_map_free(access
);
4515 external
= isl_union_map_copy(gen
->prog
->scop
->tagged_dep_flow
);
4516 external
= isl_union_map_intersect_params(external
,
4517 isl_set_copy(gen
->prog
->scop
->context
));
4518 external
= isl_union_map_subtract(external
, local
);
4521 tag_set
= isl_union_map_range(external
);
4522 external
= wrapped_reference_to_access(tag_set
, tagged
);
4523 external
= isl_union_map_union(external
,
4524 isl_union_map_copy(gen
->prog
->scop
->live_in
));
4526 tag_set
= isl_union_map_domain(external
);
4527 external
= wrapped_reference_to_access(tag_set
, tagged
);
4528 external
= isl_union_map_union(external
,
4529 isl_union_map_copy(gen
->prog
->scop
->live_out
));
4532 access
= isl_union_map_intersect(access
, external
);
4537 /* Given the AST context schedule "schedule" and the mapping from
4538 * domains to the shared tile loops "shared_sched", add a schedule
4539 * for copying an array reference group to/from shared/private memory.
4540 * "read" is set if data should be copied from global memory
4541 * to shared/private memory.
4542 * "k" represents the current group
4543 * "s" is the total number of groups
4545 * We schedule an operation before or after the innermost loop
4546 * of "shared_sched" that affects the tile of the array reference group.
4548 * schedule is of the form
4552 * (with D the iteration domains and L the already generated loops),
4553 * while shared_sched is of the form
4557 * We first compute the access relation for the reference group
4561 * and remove from this access relation those reads or writes
4562 * that only needed to communicate data within the same iteration
4563 * of the last_shared dimension of the group.
4564 * We then combine what is left with shared_sched into
4568 * If this results in an empty relation, no copying needs to be performed
4570 * Otherwise, we invert the relation and combine it with "schedule" into
4574 * The actual additional piece of the schedule is obtained from combining
4580 * [s_0,...] -> [0,s_{tile_first},0,..., val, 0, 0, ... 0]
4582 * The position of "val" corresponds to the innermost loop that affects
4583 * the tile and the value indicates where the copying is scheduled
4584 * with respect to the actual kernel code (at value 0).
4585 * Reads are schedule before the code, writes to global memory from
4586 * private memory are scheduled at values 1 to s, writes to global
4587 * memory from shared memory are scheduled at values s + 2 to 2 * s + 1.
4589 * If we are scheduling a read from global memory to shared memory,
4590 * we insert a synchronization before the kernel code (at the innermost
4592 * If we are scheduling a write to global memory, then we add
4593 * a synchronization after all writes (at value 2 *s + 2).
4594 * However, there is no need for a synchronization after the outermost loop.
4595 * A write to global memory from private memory at the innermost level
4596 * does not require a synchronization, because it is covered by
4597 * the synchronization after the kernel inserted by body_schedule.
4599 static __isl_give isl_union_map
*add_group_schedule(struct gpu_gen
*gen
,
4600 __isl_take isl_union_map
*res
, __isl_keep isl_union_map
*schedule
,
4601 __isl_keep isl_union_map
*shared_sched
,
4602 struct gpu_array_ref_group
*group
, int read
, int k
, int s
)
4607 isl_union_map
*access
;
4608 isl_map
*map
, *proj
, *access_map
;
4611 access
= group_access_relation(group
, read
, !read
);
4612 access
= remove_local_accesses(gen
, group
, access
, read
);
4613 access
= isl_union_map_range_product(isl_union_map_copy(shared_sched
),
4616 if (isl_union_map_is_empty(access
)) {
4617 isl_union_map_free(access
);
4621 access
= isl_union_map_reverse(access
);
4622 access
= isl_union_map_apply_range(access
,
4623 isl_union_map_copy(schedule
));
4624 access_map
= isl_map_from_union_map(access
);
4626 space
= isl_space_copy(group
->array
->space
);
4627 space
= isl_space_from_range(space
);
4628 space
= isl_space_add_dims(space
, isl_dim_in
, gen
->shared_len
);
4629 map
= isl_map_domain_map(isl_map_universe(space
));
4631 space
= isl_union_map_get_space(schedule
);
4632 pos
= group
->last_shared
+ 1 - gen
->tile_first
;
4636 else if (group
->private_tile
)
4639 val
= 1 + s
+ 1 + k
;
4640 proj
= insert_even(gen
, space
, pos
, val
);
4641 map
= isl_map_apply_range(map
, proj
);
4643 access_map
= isl_map_range_product(access_map
, map
);
4645 id
= isl_id_alloc(gen
->ctx
, read
? "read" : "write", group
);
4646 access_map
= isl_map_set_tuple_id(access_map
, isl_dim_in
, id
);
4648 res
= isl_union_map_add_map(res
, access_map
);
4650 n
= gen
->shared_len
- gen
->tile_first
;
4652 if (!group
->private_tile
)
4653 res
= add_sync_schedule(gen
, res
, schedule
,
4654 shared_sched
, n
, -1);
4658 if (pos
== n
&& group
->private_tile
)
4660 res
= add_sync_schedule(gen
, res
, schedule
, shared_sched
,
4667 /* Return a schedule for the shared tile loops based on the current
4668 * AST context schedule.
4670 * We create a "shared_sched" that maps the domains to the first
4671 * shared_len dimensions of the computed schedule, project out the
4672 * first tile_first dimensions (as these are already covered by
4673 * the host code) and insert "statement-level" dimensions at even
4674 * positions so that we can schedule copy blocks and synchronization
4675 * before/after each level.
4677 * In particular, copy blocks are inserted inside the innermost
4678 * level that affect the tile. For the copying to global memory,
4679 * those from private memory are scheduled before those from shared
4680 * memory such that synchronization can be inserted between the two
4681 * at the innermost level.
4682 * Synchronization is inserted at the innermost level before the
4683 * actual kernel code if there is any copying from global memory
4684 * to shared memory. It is inserted unconditionally at the innermost
4685 * level after the actual kernel code and the copying to global memory
4686 * from private memory (if any). Finally, it is inserted after
4687 * any copying to global memory, except at the outermost level
4688 * and at the innermost level if there is no copying from shared
4689 * memory. The copying from private memory is covered by the unconditional
4690 * synchronization at the innermost level.
4692 static __isl_give isl_union_map
*body_schedule(struct gpu_gen
*gen
,
4693 __isl_take isl_union_map
*schedule
)
4697 isl_union_map
*shared_sched
;
4698 isl_union_map
*sched
;
4699 isl_map
*proj
, *map
;
4702 shared_sched
= isl_union_map_copy(gen
->tiled_sched
);
4703 proj
= projection(isl_union_map_get_space(shared_sched
),
4704 gen
->tiled_len
, gen
->shared_len
);
4705 shared_sched
= isl_union_map_apply_range(shared_sched
,
4706 isl_union_map_from_map(proj
));
4707 space
= isl_union_map_get_space(shared_sched
);
4708 proj
= insert_even(gen
, space
, -1, 0);
4709 sched
= isl_union_map_apply_range(isl_union_map_copy(shared_sched
),
4710 isl_union_map_from_map(proj
));
4712 res
= isl_union_map_range_product(isl_union_map_copy(schedule
), sched
);
4715 for (i
= 0; i
< gen
->prog
->n_array
; ++i
)
4716 s
+= gen
->prog
->array
[i
].n_group
;
4719 for (i
= 0; i
< gen
->prog
->n_array
; ++i
) {
4720 struct gpu_array_info
*array
= &gen
->prog
->array
[i
];
4722 for (j
= 0; j
< array
->n_group
; ++j
) {
4723 struct gpu_array_ref_group
*group
;
4725 group
= array
->groups
[j
];
4726 if (!group
->private_tile
&& !group
->shared_tile
)
4728 res
= add_group_schedule(gen
, res
, schedule
,
4729 shared_sched
, group
, 0, k
, s
);
4730 res
= add_group_schedule(gen
, res
, schedule
,
4731 shared_sched
, group
, 1, k
, s
);
4736 res
= add_sync_schedule(gen
, res
, schedule
, shared_sched
,
4737 gen
->shared_len
- gen
->tile_first
, 1 + s
);
4739 isl_union_map_free(shared_sched
);
4740 isl_union_map_free(schedule
);
4745 /* Generate code for "kernel" in the given "context".
4747 * We first generate code for the shared tile loops (T1T, T1P and T2)
4748 * in a context that includes the block ids.
4749 * Within each iteration of these loops an additional code generation
4750 * is performed (within create_kernel_leaf) for the rest of the schedule
4751 * in a context that includes the thread ids.
4753 static __isl_give isl_ast_node
*generate_kernel(struct gpu_gen
*gen
,
4754 __isl_keep isl_ast_build
*build
, __isl_keep isl_set
*host_domain
,
4755 __isl_keep isl_multi_pw_aff
*grid_size
)
4759 isl_id_list
*iterators
;
4760 isl_union_map
*schedule
;
4764 schedule
= isl_ast_build_get_schedule(build
);
4766 build
= isl_ast_build_copy(build
);
4767 build
= isl_ast_build_restrict(build
, isl_set_copy(host_domain
));
4768 space
= isl_ast_build_get_schedule_space(build
);
4769 set
= isl_set_universe(isl_space_copy(space
));
4770 set
= add_bounded_parameters_dynamic(set
, grid_size
, "b");
4771 build
= isl_ast_build_restrict(build
, set
);
4773 schedule
= body_schedule(gen
, schedule
);
4775 sched_len
= 2 * (gen
->shared_len
- gen
->tile_first
) + 1;
4777 build
= set_atomic_and_unroll(build
, space
, sched_len
);
4778 iterators
= generate_names(gen
->ctx
, sched_len
, "g");
4779 build
= isl_ast_build_set_iterators(build
, iterators
);
4780 build
= isl_ast_build_set_create_leaf(build
, &create_kernel_leaf
, gen
);
4781 tree
= isl_ast_build_ast_from_schedule(build
, schedule
);
4782 isl_ast_build_free(build
);
4787 /* Attach "id" to the given node.
4789 static __isl_give isl_ast_node
*attach_id(__isl_take isl_ast_node
*node
,
4790 __isl_keep isl_ast_build
*build
, void *user
)
4794 node
= isl_ast_node_set_annotation(node
, id
);
4799 /* Construct an AST node for performing a kernel launch and attach
4800 * the information about the kernel to that node.
4802 * The kernel AST has been constructed in the context of the range
4803 * of "schedule". In particular, the grid size has been computed
4804 * in the context. We therefore still need to make sure that these
4805 * constraints are expressed in the code. We do this by creating a schedule
4807 * kernel[] -> [S -> []]
4809 * where S is the schedule domain, i.e., the range of "schedule".
4810 * The AST generation will then create a single call surrounded by
4811 * all the condition in "S" that have not been expressed yet.
4813 * The kernel information is attached to this node in attach_id.
4815 static __isl_give isl_ast_node
*construct_launch(
4816 __isl_take isl_ast_build
*build
, __isl_take isl_union_map
*schedule
,
4817 __isl_take
struct ppcg_kernel
*kernel
)
4821 isl_union_set
*domain
;
4826 ctx
= isl_ast_build_get_ctx(build
);
4828 id
= isl_id_alloc(ctx
, NULL
, kernel
);
4829 id
= isl_id_set_free_user(id
, &ppcg_kernel_free
);
4831 domain
= isl_union_map_range(schedule
);
4832 set
= isl_set_from_union_set(domain
);
4833 map
= isl_map_from_domain(set
);
4834 map
= isl_map_from_range(isl_map_wrap(map
));
4835 map
= isl_map_set_tuple_name(map
, isl_dim_in
, "kernel");
4836 schedule
= isl_union_map_from_map(map
);
4838 build
= isl_ast_build_set_at_each_domain(build
, &attach_id
, id
);
4839 node
= isl_ast_build_ast_from_schedule(build
, schedule
);
4840 isl_ast_build_free(build
);
4845 /* This function is called for each leaf in the AST of the host code.
4846 * We first specialize the schedule to the site of the leaf, compute
4847 * the size of shared memory and then construct the body of the host code
4848 * and the associated kernel.
4850 * The necessary information for printing the kernel launch is
4851 * stored in a struct ppcg_kernel and attached to the leaf node
4852 * created to represent the launch.
4854 static __isl_give isl_ast_node
*create_host_leaf(
4855 __isl_take isl_ast_build
*build
, void *user
)
4857 struct gpu_gen
*gen
= (struct gpu_gen
*) user
;
4860 struct ppcg_kernel
*kernel
;
4861 isl_set
*host_domain
;
4862 isl_union_map
*schedule
;
4863 isl_union_map
*local_sched
;
4864 isl_union_map
*access
;
4865 isl_union_set
*domain
;
4868 schedule
= isl_ast_build_get_schedule(build
);
4870 isl_union_map_foreach_map(schedule
, &extract_tile_len
, gen
);
4873 domain
= isl_union_map_domain(isl_union_map_copy(schedule
));
4875 local_sched
= isl_union_map_copy(gen
->sched
);
4876 local_sched
= isl_union_map_intersect_domain(local_sched
, domain
);
4877 access
= isl_union_map_union(isl_union_map_copy(gen
->prog
->read
),
4878 isl_union_map_copy(gen
->prog
->may_write
));
4879 access
= isl_union_map_apply_domain(access
,
4880 isl_union_map_copy(local_sched
));
4882 gen
->tiled_sched
= tile_schedule(gen
, local_sched
);
4883 gen
->tiled_sched
= parametrize_tiled_schedule(gen
, gen
->tiled_sched
);
4884 gen
->tiled_sched
= scale_tile_loops(gen
, gen
->tiled_sched
);
4886 gen
->local_sched
= isl_union_map_copy(gen
->tiled_sched
);
4887 gen
->local_sched
= thread_tile_schedule(gen
, gen
->local_sched
);
4888 gen
->local_sched
= scale_thread_tile_loops(gen
, gen
->local_sched
);
4890 kernel
= gen
->kernel
= isl_calloc_type(gen
->ctx
, struct ppcg_kernel
);
4894 kernel
->id
= gen
->kernel_id
++;
4895 kernel
->context
= isl_union_map_params(isl_union_map_copy(schedule
));
4896 kernel
->grid_size
= extract_grid_size(gen
, kernel
);
4897 extract_block_size(gen
, kernel
);
4898 kernel
->arrays
= isl_union_map_range(access
);
4899 kernel
->arrays
= isl_union_set_apply(kernel
->arrays
,
4900 isl_union_map_copy(gen
->prog
->to_outer
));
4901 kernel
->space
= isl_ast_build_get_schedule_space(build
);
4903 gen
->private_access
= NULL
;
4904 compute_shared_sched(gen
);
4905 gen
->privatization
= compute_privatization(gen
);
4906 if (group_references(gen
) < 0)
4907 schedule
= isl_union_map_free(schedule
);
4908 compute_private_access(gen
);
4909 check_shared_memory_bound(gen
);
4910 compute_group_tilings(gen
);
4911 host_domain
= isl_set_from_union_set(isl_union_map_range(
4912 isl_union_map_copy(schedule
)));
4913 localize_bounds(gen
, kernel
, host_domain
);
4915 gen
->local_sched
= interchange_for_unroll(gen
, gen
->local_sched
);
4917 kernel
->tree
= generate_kernel(gen
, build
, host_domain
,
4919 create_kernel_vars(gen
, kernel
);
4921 free_local_array_info(gen
);
4922 isl_map_free(gen
->privatization
);
4923 isl_union_map_free(gen
->private_access
);
4924 isl_union_map_free(gen
->local_sched
);
4925 isl_union_map_free(gen
->tiled_sched
);
4926 isl_union_map_free(gen
->shared_sched
);
4927 isl_union_map_free(gen
->shared_proj
);
4928 isl_set_free(host_domain
);
4929 free(gen
->tile_size
);
4931 node
= construct_launch(build
, schedule
, kernel
);
4935 isl_union_map_free(schedule
);
4939 /* Use isl to generate code for the outer gen->tile_first loops
4940 * of the global schedule in gen->sched, resulting in the host code.
4941 * Within each iteration of this partial schedule, i.e., for each kernel
4942 * launch, create_host_leaf takes care of generating the kernel code.
4944 static __isl_give isl_ast_node
*generate_host_code(struct gpu_gen
*gen
)
4946 isl_ast_build
*build
;
4948 isl_union_map
*sched
;
4950 isl_id_list
*iterators
;
4952 sched
= isl_union_map_copy(gen
->sched
);
4953 proj
= projection(isl_union_map_get_space(sched
),
4954 gen
->untiled_len
, gen
->tile_first
);
4955 sched
= isl_union_map_apply_range(sched
, isl_union_map_from_map(proj
));
4957 isl_options_set_ast_build_group_coscheduled(gen
->ctx
, 1);
4958 build
= isl_ast_build_from_context(isl_set_copy(gen
->prog
->context
));
4959 iterators
= generate_names(gen
->ctx
, gen
->tile_first
, "h");
4960 build
= isl_ast_build_set_iterators(build
, iterators
);
4961 build
= isl_ast_build_set_create_leaf(build
, &create_host_leaf
, gen
);
4962 tree
= isl_ast_build_ast_from_schedule(build
, sched
);
4963 isl_ast_build_free(build
);
4968 __isl_give isl_union_map
*extract_sizes_from_str(isl_ctx
*ctx
, const char *str
)
4972 return isl_union_map_read_from_str(ctx
, str
);
4975 /* Information about the outermost tilable bands in the forest of bands.
4977 * tile_len and n_parallel are only sets on band_info structures
4978 * that correspond to outermost bands. For other bands (in particular,
4979 * ancestors of the outermost bands), n_parallal is set to 0.
4981 * prefix is the (padded) schedule leading up to the outermost tilable bands.
4983 * tile_first is the number of schedule dimensions in prefix.
4985 * suffix is the schedule of the outermost tilable bands and their descendants.
4988 struct gpu_gen
*gen
;
4992 isl_union_map
*prefix
;
4993 isl_union_map
*suffix
;
4996 /* Set tile_len and n_parallel of the statement to that of
4997 * their outermost band, recorded in the band_info.
4999 static int set_stmt_tile_len(__isl_take isl_map
*map
, void *user
)
5001 struct band_info
*info
= user
;
5002 struct gpu_stmt
*stmt
;
5005 id
= isl_map_get_tuple_id(map
, isl_dim_in
);
5006 stmt
= find_stmt(info
->gen
->prog
, id
);
5009 stmt
->tile_len
= info
->tile_len
;
5010 stmt
->n_parallel
= info
->n_parallel
;
5017 static void list_select_outer_band(struct gpu_gen
*gen
,
5018 __isl_take isl_band_list
*list
, int pos
, struct band_info
*list_info
);
5020 /* Check if this band has any parallel loops. If so, take it as
5021 * the outermost tilable band. If not, continue looking for the
5022 * outermost tilable band in the children of the current band.
5024 static void band_select_outer_band(struct gpu_gen
*gen
,
5025 __isl_take isl_band
*band
, int pos
, struct band_info
*info
)
5027 int n
= isl_band_n_member(band
);
5030 for (n_parallel
= 0; n_parallel
< n
; ++n_parallel
)
5031 if (!isl_band_member_is_coincident(band
, n_parallel
))
5034 info
->n_parallel
= n_parallel
;
5036 gen
->any_parallelism
= 1;
5038 info
->tile_first
= pos
;
5040 info
->prefix
= isl_band_get_prefix_schedule(band
);
5041 info
->suffix
= isl_union_map_flat_range_product(
5042 isl_band_get_partial_schedule(band
),
5043 isl_band_get_suffix_schedule(band
));
5044 isl_union_map_foreach_map(info
->prefix
,
5045 &set_stmt_tile_len
, info
);
5046 } else if (isl_band_has_children(band
)) {
5047 isl_band_list
*children
;
5048 children
= isl_band_get_children(band
);
5049 list_select_outer_band(gen
, children
, pos
+ n
, info
);
5052 info
->tile_first
= pos
+ n
;
5054 info
->prefix
= isl_union_map_flat_range_product(
5055 isl_band_get_prefix_schedule(band
),
5056 isl_band_get_partial_schedule(band
));
5057 info
->suffix
= isl_band_get_suffix_schedule(band
);
5058 isl_union_map_foreach_map(info
->prefix
,
5059 &set_stmt_tile_len
, info
);
5062 isl_band_free(band
);
5065 /* Comparison function that returns a non-zero value for band_infos
5066 * with different tile_len fields or different n_parallel fields.
5068 static int cmp_band(const void *p1
, const void *p2
)
5070 const struct band_info
*info1
= p1
;
5071 const struct band_info
*info2
= p2
;
5073 if (info1
->tile_len
!= info2
->tile_len
)
5074 return info1
->tile_len
- info2
->tile_len
;
5076 return info1
->n_parallel
- info2
->n_parallel
;
5079 /* Extend "umap" with coordinates with fixed value "val"
5080 * to a total length of "dst_len", assuming the original dimension is "src_len".
5082 static __isl_give isl_union_map
*extend_range(
5083 __isl_take isl_union_map
*umap
, int src_len
, int dst_len
, int val
)
5089 dim
= isl_union_map_get_space(umap
);
5090 map
= isl_map_reverse(projection(dim
, dst_len
, src_len
));
5091 for (i
= src_len
; i
< dst_len
; ++i
)
5092 map
= isl_map_fix_si(map
, isl_dim_out
, i
, val
);
5094 umap
= isl_union_map_apply_range(umap
, isl_union_map_from_map(map
));
5099 /* Group bands with the same values for tile_len and n_parallel.
5100 * The prefix schedule is then extended with a fixed coordinate that
5101 * is different for each such group.
5102 * Note that the actual values for this coordinate are not important.
5103 * The bands have already been effectively separated at a higher level
5104 * or they are independent and may be executed in parallel.
5105 * The list of band_info has been sorted before this functions is called.
5107 static void separate_bands(struct band_info
*info
, int n
)
5112 for (i
= 0; i
< n
; ++i
) {
5113 int l
= info
[i
].tile_first
;
5116 (info
[i
].tile_len
!= info
[i
- 1].tile_len
||
5117 info
[i
].n_parallel
!= info
[i
- 1].n_parallel
))
5120 info
[i
].prefix
= extend_range(info
[i
].prefix
,
5122 info
[i
].tile_first
= l
+ 1;
5126 /* Select the outermost bands in the elements of the list, align
5127 * their prefix schedules, separate bands with different values
5128 * for tile_len and/or n_parallel and then combine the resulting
5129 * prefix and suffix schedules into a single pair of prefix and
5130 * suffix schedules for the entire list.
5132 static void list_select_outer_band(struct gpu_gen
*gen
,
5133 __isl_take isl_band_list
*list
, int pos
, struct band_info
*list_info
)
5137 int n
= isl_band_list_n_band(list
);
5138 isl_ctx
*ctx
= isl_band_list_get_ctx(list
);
5139 struct band_info
*info
;
5141 isl_union_map
*prefix
;
5142 isl_union_map
*suffix
;
5145 info
= isl_calloc_array(ctx
, struct band_info
, n
);
5149 for (i
= 0; i
< n
; ++i
) {
5150 band
= isl_band_list_get_band(list
, i
);
5151 band_select_outer_band(gen
, band
, pos
, &info
[i
]);
5152 if (info
[i
].tile_first
> max_tile_first
)
5153 max_tile_first
= info
[i
].tile_first
;
5156 for (i
= 0; i
< n
; ++i
) {
5157 if (info
[i
].tile_first
== max_tile_first
)
5159 info
[i
].prefix
= extend_range(info
[i
].prefix
,
5160 info
[i
].tile_first
, max_tile_first
, 0);
5161 info
[i
].tile_first
= max_tile_first
;
5164 qsort(info
, n
, sizeof(struct band_info
), &cmp_band
);
5166 for (i
= 0; i
< n
- 1; ++i
)
5167 if (info
[i
].tile_len
!= info
[i
+ 1].tile_len
||
5168 info
[i
].n_parallel
!= info
[i
+ 1].n_parallel
)
5172 separate_bands(info
, n
);
5174 prefix
= info
[0].prefix
;
5175 suffix
= info
[0].suffix
;
5177 for (i
= 1; i
< n
; ++i
) {
5178 prefix
= isl_union_map_union(prefix
, info
[i
].prefix
);
5179 suffix
= isl_union_map_union(suffix
, info
[i
].suffix
);
5182 list_info
->tile_first
= info
[0].tile_first
;
5183 list_info
->tile_len
= -1;
5184 list_info
->prefix
= prefix
;
5185 list_info
->suffix
= suffix
;
5187 isl_band_list_free(list
);
5191 /* Select the outermost tilable band that (by construction)
5192 * has at least one parallel loop.
5193 * The starting position of the aligned band is stored in the pair
5195 * The sizes and number of parallel loops may be different in different
5196 * parts of the band forest and are therefore stored in the gpu_stmts.
5198 * Return the complete schedule, with the tilable bands aligned
5199 * at gen->tile_first and padded with zero, if needed.
5201 static __isl_give isl_union_map
*select_outer_tilable_band(struct gpu_gen
*gen
,
5202 __isl_keep isl_schedule
*schedule
)
5204 isl_band_list
*list
;
5205 struct band_info info
;
5207 gen
->n_parallel
= 0;
5210 list
= isl_schedule_get_band_forest(schedule
);
5212 if (isl_band_list_n_band(list
) == 0) {
5213 isl_band_list_free(list
);
5214 return isl_schedule_get_map(schedule
);
5217 list_select_outer_band(gen
, list
, 0, &info
);
5219 gen
->tile_first
= info
.tile_first
;
5220 info
.suffix
= align_range(info
.suffix
);
5222 return isl_union_map_flat_range_product(info
.prefix
, info
.suffix
);
5225 /* Set gen->untiled_len to the number of scheduling dimensions
5226 * for the schedule of the first domain.
5227 * We assume here that this number is the same for all domains.
5229 static int set_untiled_len(__isl_take isl_map
*map
, void *user
)
5231 unsigned *untiled_len
= user
;
5233 *untiled_len
= isl_map_dim(map
, isl_dim_out
);
5239 /* Compute an appropriate schedule based on the accesses in
5240 * gen->read and gen->write.
5242 * We use the dependences in gen->prog->scop to compute
5243 * a schedule that has a parallel loop in each tilable band.
5244 * Finally, we select the outermost tilable band.
5246 static void compute_schedule(struct gpu_gen
*gen
)
5248 isl_union_set
*domain
;
5249 isl_union_map
*dep_raw
, *dep
;
5250 isl_union_map
*sched
;
5251 isl_schedule_constraints
*sc
;
5252 isl_schedule
*schedule
;
5254 dep_raw
= isl_union_map_copy(gen
->prog
->scop
->dep_flow
);
5256 dep
= isl_union_map_copy(gen
->prog
->scop
->dep_false
);
5257 dep
= isl_union_map_union(dep
, dep_raw
);
5258 dep
= isl_union_map_coalesce(dep
);
5260 domain
= isl_union_set_copy(gen
->prog
->scop
->domain
);
5261 domain
= isl_union_set_intersect_params(domain
,
5262 isl_set_copy(gen
->prog
->scop
->context
));
5263 sc
= isl_schedule_constraints_on_domain(isl_union_set_copy(domain
));
5264 sc
= isl_schedule_constraints_set_validity(sc
, isl_union_map_copy(dep
));
5265 sc
= isl_schedule_constraints_set_coincidence(sc
,
5266 isl_union_map_copy(dep
));
5267 sc
= isl_schedule_constraints_set_proximity(sc
, dep
);
5269 if (gen
->options
->debug
->dump_schedule_constraints
)
5270 isl_schedule_constraints_dump(sc
);
5271 schedule
= isl_schedule_constraints_compute_schedule(sc
);
5272 if (gen
->options
->debug
->dump_schedule
)
5273 isl_schedule_dump(schedule
);
5275 sched
= select_outer_tilable_band(gen
, schedule
);
5277 isl_union_map_foreach_map(sched
, &set_untiled_len
, &gen
->untiled_len
);
5278 sched
= isl_union_map_intersect_domain(sched
, domain
);
5281 isl_schedule_free(schedule
);
5284 /* Compute the sets of outer array elements that need to be copied in and out.
5286 * In particular, for each array that is possibly written anywhere in
5287 * gen->prog and that is visible outside the corresponding scop,
5288 * we copy out its entire extent.
5290 * Any array elements that is read without first being written needs
5291 * to be copied in. Furthermore, if there are any array elements that
5292 * are copied out, but that may not be written inside gen->prog, then
5293 * they also need to be copied in to ensure that the value after execution
5294 * is the same as the value before execution.
5295 * In case the array elements are structures, we need to take into
5296 * account that all members of the structures need to be written
5297 * by gen->prog before we can avoid copying the data structure in.
5299 * While computing the set of array elements that are copied out but
5300 * not necessarily written, we intersect both sets with the context.
5301 * This helps in those cases where the arrays are declared with a fixed size,
5302 * while the accesses are parametric and the context assigns a fixed value
5303 * to the parameters.
5305 * If an element from a local array is read without first being written,
5306 * then there is no point in copying it in since it cannot have been
5307 * written prior to the scop. Warn about the uninitialized read instead.
5309 static void compute_copy_in_and_out(struct gpu_gen
*gen
)
5312 isl_union_set
*local
;
5313 isl_union_set
*may_write
, *must_write
;
5314 isl_union_set
*copy_in
, *copy_out
;
5315 isl_union_set
*not_written
;
5316 isl_union_map
*uninitialized
;
5317 isl_union_map
*local_uninitialized
;
5319 must_write
= isl_union_map_range(
5320 isl_union_map_copy(gen
->prog
->must_write
));
5321 must_write
= isl_union_set_intersect_params(must_write
,
5322 isl_set_copy(gen
->prog
->context
));
5323 may_write
= isl_union_map_range(
5324 isl_union_map_copy(gen
->prog
->may_write
));
5325 may_write
= isl_union_set_intersect_params(may_write
,
5326 isl_set_copy(gen
->prog
->context
));
5327 may_write
= isl_union_set_universe(may_write
);
5328 may_write
= isl_union_set_apply(may_write
,
5329 isl_union_map_copy(gen
->prog
->to_outer
));
5330 copy_out
= isl_union_set_empty(isl_union_set_get_space(may_write
));
5331 local
= isl_union_set_copy(copy_out
);
5333 for (i
= 0; i
< gen
->prog
->n_array
; ++i
) {
5338 space
= isl_space_copy(gen
->prog
->array
[i
].space
);
5340 if (gen
->prog
->array
[i
].local
) {
5343 set
= isl_set_universe(space
);
5344 local
= isl_union_set_add_set(local
, set
);
5348 write_i
= isl_union_set_extract_set(may_write
, space
);
5349 empty
= isl_set_fast_is_empty(write_i
);
5350 isl_set_free(write_i
);
5354 write_i
= isl_set_copy(gen
->prog
->array
[i
].extent
);
5355 copy_out
= isl_union_set_add_set(copy_out
, write_i
);
5357 isl_union_set_free(may_write
);
5359 copy_out
= isl_union_set_intersect_params(copy_out
,
5360 isl_set_copy(gen
->prog
->context
));
5362 gen
->prog
->copy_out
= isl_union_set_copy(copy_out
);
5364 copy_out
= isl_union_set_apply(copy_out
,
5365 isl_union_map_copy(gen
->prog
->to_inner
));
5366 not_written
= isl_union_set_subtract(copy_out
, must_write
);
5368 uninitialized
= isl_union_map_copy(gen
->prog
->scop
->live_in
);
5369 local_uninitialized
= isl_union_map_copy(uninitialized
);
5371 local
= isl_union_set_apply(local
,
5372 isl_union_map_copy(gen
->prog
->to_inner
));
5373 local_uninitialized
= isl_union_map_intersect_range(local_uninitialized
,
5375 if (!isl_union_map_is_empty(local_uninitialized
)) {
5377 "possibly uninitialized reads (not copied in):\n");
5378 isl_union_map_dump(local_uninitialized
);
5380 uninitialized
= isl_union_map_subtract(uninitialized
,
5381 local_uninitialized
);
5382 copy_in
= isl_union_map_range(uninitialized
);
5383 copy_in
= isl_union_set_union(copy_in
, not_written
);
5384 copy_in
= isl_union_set_apply(copy_in
,
5385 isl_union_map_copy(gen
->prog
->to_outer
));
5387 gen
->prog
->copy_in
= copy_in
;
5390 /* Extract a gpu_stmt_access from "expr", append it to the list
5391 * that ends in *next_access and return the updated end of the list.
5393 static struct gpu_stmt_access
**expr_extract_access(struct pet_expr
*expr
,
5394 struct gpu_stmt_access
**next_access
)
5396 struct gpu_stmt_access
*access
;
5397 isl_ctx
*ctx
= isl_map_get_ctx(expr
->acc
.access
);
5399 access
= isl_alloc_type(ctx
, struct gpu_stmt_access
);
5401 access
->next
= NULL
;
5402 access
->read
= expr
->acc
.read
;
5403 access
->write
= expr
->acc
.write
;
5404 access
->access
= pet_expr_access_get_may_access(expr
);
5405 access
->tagged_access
= pet_expr_access_get_tagged_may_access(expr
);
5406 access
->exact_write
= !expr
->acc
.write
||
5407 isl_map_is_equal(expr
->acc
.access
, access
->access
);
5408 access
->ref_id
= isl_id_copy(expr
->acc
.ref_id
);
5411 *next_access
= access
;
5412 next_access
= &(*next_access
)->next
;
5416 static struct gpu_stmt_access
**expr_extract_accesses(struct pet_expr
*expr
,
5417 struct gpu_stmt_access
**next_access
)
5421 for (i
= 0; i
< expr
->n_arg
; ++i
)
5422 next_access
= expr_extract_accesses(expr
->args
[i
],
5425 if (expr
->type
== pet_expr_access
)
5426 next_access
= expr_extract_access(expr
, next_access
);
5431 static void pet_stmt_extract_accesses(struct gpu_stmt
*stmt
)
5433 struct gpu_stmt_access
**next_access
= &stmt
->accesses
;
5435 stmt
->accesses
= NULL
;
5436 expr_extract_accesses(stmt
->stmt
->body
, next_access
);
5439 /* Return an array of gpu_stmt representing the statements in "scop".
5441 static struct gpu_stmt
*extract_stmts(isl_ctx
*ctx
, struct ppcg_scop
*scop
,
5442 __isl_keep isl_set
*context
)
5445 struct gpu_stmt
*stmts
;
5447 stmts
= isl_calloc_array(ctx
, struct gpu_stmt
, scop
->n_stmt
);
5451 for (i
= 0; i
< scop
->n_stmt
; ++i
) {
5452 struct gpu_stmt
*s
= &stmts
[i
];
5454 s
->id
= isl_set_get_tuple_id(scop
->stmts
[i
]->domain
);
5455 s
->stmt
= scop
->stmts
[i
];
5456 pet_stmt_extract_accesses(s
);
5462 /* Callback for ppcg_print_guarded that calls the callback for generate_gpu.
5464 static __isl_give isl_printer
*print_gpu(__isl_take isl_printer
*p
, void *user
)
5466 struct gpu_gen
*gen
= user
;
5468 return gen
->print(p
, gen
->prog
, gen
->tree
, &gen
->types
,
5472 /* Generate CUDA code for "scop" and print it to "p".
5473 * After generating an AST for the transformed scop as explained below,
5474 * we call "gen->print" to print the AST in the desired output format
5477 * If it turns out that it does not make sense to generate GPU code,
5478 * then we generate CPU code instead.
5480 * The GPU code is generated in a context where at least one
5481 * statement instance is executed. The corresponding guard (if any) is printed
5482 * around the entire generated GPU code, except for the declaration
5483 * of the arrays that are visible outside of the scop and that therefore
5484 * cannot be declared inside the body of any possible guard.
5486 * We first compute a schedule that respects the dependences
5487 * of the original program and select the outermost band
5488 * of tilable dimensions that has at least one parallel loop.
5489 * We then have three blocks of dimensions
5493 * The tilable band "B" is first tiled according to "tile" sizes, resulting
5498 * For each iteration of the T loop and for each array, we compute
5499 * the array elements accessed by that iteration, construct a rectangular
5500 * box around it and shift it to the origin. The result is used
5501 * as shared memory for the array.
5503 * We then split off at most 2 parallel loops from the T loops and
5504 * at most 3 parallel loops from the P loops
5508 * The T1/P1 loops are then tiled or "wrapped" over the blocks/threads,
5509 * according to "grid"/"block" sizes.
5511 * H T1T T1P T2 P1T P1P P2 G
5513 * Finally, the T1P and P1P iterators are equated to the block and
5514 * thread dimensions respectively and so are effectively removed.
5515 * The H loops are run on the host. The T1T, T2, P1T, P2 and G loops
5516 * are run on the GPU.
5518 * Code is generated in three stages. We first generate code for the
5519 * host (the H loops), with iterators h%d. Then, for each leaf node
5520 * of the resulting AST, we generate code for the shared loops (up to
5521 * and including T2), with iterators g%d and after equating the H loops
5522 * to h%d parameters and the T1P loops to the block dimensions.
5523 * Finally, we generate code for the remaining loops in a similar fashion.
5525 static __isl_give isl_printer
*generate(__isl_take isl_printer
*p
,
5526 struct gpu_gen
*gen
, struct ppcg_scop
*scop
,
5527 struct ppcg_options
*options
)
5529 struct gpu_prog
*prog
;
5531 isl_set
*context
, *guard
;
5534 return isl_printer_free(p
);
5536 ctx
= isl_printer_get_ctx(p
);
5537 prog
= gpu_prog_alloc(ctx
, scop
);
5539 return isl_printer_free(p
);
5541 context
= isl_set_copy(prog
->context
);
5542 guard
= isl_union_set_params(isl_union_set_copy(prog
->scop
->domain
));
5543 prog
->context
= isl_set_intersect(prog
->context
, isl_set_copy(guard
));
5546 gen
->any_parallelism
= 0;
5547 compute_schedule(gen
);
5549 if (!gen
->any_parallelism
) {
5550 isl_set_free(context
);
5551 isl_set_free(guard
);
5552 p
= print_cpu(p
, scop
, options
);
5554 compute_copy_in_and_out(gen
);
5555 gen
->tree
= generate_host_code(gen
);
5556 p
= ppcg_print_exposed_declarations(p
, prog
->scop
);
5557 p
= ppcg_print_guarded(p
, guard
, context
, &print_gpu
, gen
);
5558 isl_ast_node_free(gen
->tree
);
5561 isl_union_map_free(gen
->sched
);
5563 gpu_prog_free(prog
);
5568 /* Wrapper around generate for use as a ppcg_transform callback.
5570 static __isl_give isl_printer
*generate_wrap(__isl_take isl_printer
*p
,
5571 struct ppcg_scop
*scop
, void *user
)
5573 struct gpu_gen
*gen
= user
;
5575 return generate(p
, gen
, scop
, gen
->options
);
5578 /* Transform the code in the file called "input" by replacing
5579 * all scops by corresponding GPU code and write the results to "out".
5581 int generate_gpu(isl_ctx
*ctx
, const char *input
, FILE *out
,
5582 struct ppcg_options
*options
,
5583 __isl_give isl_printer
*(*print
)(__isl_take isl_printer
*p
,
5584 struct gpu_prog
*prog
, __isl_keep isl_ast_node
*tree
,
5585 struct gpu_types
*types
, void *user
), void *user
)
5592 gen
.sizes
= extract_sizes_from_str(ctx
, options
->sizes
);
5593 gen
.options
= options
;
5596 gen
.print_user
= user
;
5598 gen
.types
.name
= NULL
;
5600 r
= ppcg_transform(ctx
, input
, out
, options
, &generate_wrap
, &gen
);
5602 isl_union_map_free(gen
.sizes
);
5603 for (i
= 0; i
< gen
.types
.n
; ++i
)
5604 free(gen
.types
.name
[i
]);
5605 free(gen
.types
.name
);
5610 struct gpu_prog
*gpu_prog_alloc(isl_ctx
*ctx
, struct ppcg_scop
*scop
)
5612 struct gpu_prog
*prog
;
5617 prog
= isl_calloc_type(ctx
, struct gpu_prog
);
5622 prog
->context
= isl_set_copy(scop
->context
);
5623 prog
->n_stmts
= scop
->n_stmt
;
5624 prog
->stmts
= extract_stmts(ctx
, scop
, prog
->context
);
5625 prog
->read
= isl_union_map_copy(scop
->reads
);
5626 prog
->may_write
= isl_union_map_copy(scop
->may_writes
);
5627 prog
->must_write
= isl_union_map_copy(scop
->must_writes
);
5628 prog
->to_inner
= compute_to_inner(scop
);
5629 prog
->to_outer
= isl_union_map_copy(prog
->to_inner
);
5630 prog
->to_outer
= isl_union_map_reverse(prog
->to_outer
);
5633 return gpu_prog_free(prog
);
5635 if (collect_array_info(prog
) < 0)
5636 return gpu_prog_free(prog
);
5641 void *gpu_prog_free(struct gpu_prog
*prog
)
5645 free_array_info(prog
);
5646 free_stmts(prog
->stmts
, prog
->n_stmts
);
5647 isl_union_map_free(prog
->to_outer
);
5648 isl_union_map_free(prog
->to_inner
);
5649 isl_union_set_free(prog
->copy_in
);
5650 isl_union_set_free(prog
->copy_out
);
5651 isl_union_map_free(prog
->read
);
5652 isl_union_map_free(prog
->may_write
);
5653 isl_union_map_free(prog
->must_write
);
5654 isl_set_free(prog
->context
);