2 * Copyright 2013 Ecole Normale Superieure
4 * Use of this software is governed by the MIT license
6 * Written by Sven Verdoolaege and Riyadh Baghdadi,
7 * Ecole Normale Superieure, 45 rue d’Ulm, 75230 Paris, France
18 #include "gpu_print.h"
25 #define min(a, b) (((a) < (b)) ? (a) : (b))
26 #define max(a, b) (((a) > (b)) ? (a) : (b))
28 /* options are the global options passed to generate_opencl.
29 * input is the name of the input file.
30 * output is the user-specified output file name and may be NULL
31 * if not specified by the user.
32 * kernel_c_name is the name of the kernel_c file.
33 * kprinter is an isl_printer for the kernel file.
34 * host_c is the generated source file for the host code. kernel_c is
35 * the generated source file for the kernel.
38 struct ppcg_options
*options
;
41 char kernel_c_name
[PATH_MAX
];
43 isl_printer
*kprinter
;
49 /* Open the file called "name" for writing or print an error message.
51 static FILE *open_or_croak(const char *name
)
55 file
= fopen(name
, "w");
57 fprintf(stderr
, "Failed to open \"%s\" for writing\n", name
);
61 /* Open the host .c file and the kernel .h and .cl files for writing.
62 * Their names are derived from info->output (or info->input if
63 * the user did not specify an output file name).
64 * Add the necessary includes to these files, including those specified
67 * Return 0 on success and -1 on failure.
69 static int opencl_open_files(struct opencl_info
*info
)
78 ext
= strrchr(info
->output
, '.');
79 len
= ext
? ext
- info
->output
: strlen(info
->output
);
80 memcpy(name
, info
->output
, len
);
82 info
->host_c
= open_or_croak(info
->output
);
84 len
= ppcg_extract_base_name(name
, info
->input
);
86 strcpy(name
+ len
, "_host.c");
87 info
->host_c
= open_or_croak(name
);
90 memcpy(info
->kernel_c_name
, name
, len
);
91 strcpy(info
->kernel_c_name
+ len
, "_kernel.cl");
92 info
->kernel_c
= open_or_croak(info
->kernel_c_name
);
94 if (!info
->host_c
|| !info
->kernel_c
)
97 fprintf(info
->host_c
, "#include <assert.h>\n");
98 fprintf(info
->host_c
, "#include <stdio.h>\n");
99 fprintf(info
->host_c
, "#include \"ocl_utilities.h\"\n");
100 if (info
->options
->opencl_embed_kernel_code
) {
101 fprintf(info
->host_c
, "#include \"%s\"\n\n",
102 info
->kernel_c_name
);
105 for (i
= 0; i
< info
->options
->opencl_n_include_file
; ++i
) {
106 info
->kprinter
= isl_printer_print_str(info
->kprinter
,
108 info
->kprinter
= isl_printer_print_str(info
->kprinter
,
109 info
->options
->opencl_include_files
[i
]);
110 info
->kprinter
= isl_printer_print_str(info
->kprinter
, ">\n");
116 /* Write text to a file and escape some special characters that would break a
119 static void opencl_print_escaped(const char *str
, const char *end
, FILE *file
)
121 const char *prev
= str
;
123 while ((str
= strpbrk(prev
, "\"\\")) && str
< end
) {
124 fwrite(prev
, 1, str
- prev
, file
);
125 fprintf(file
, "\\%c", *str
);
130 fwrite(prev
, 1, end
- prev
, file
);
133 /* Write text to a file as a C string literal.
135 * This function also prints any characters after the last newline, although
136 * normally the input string should end with a newline.
138 static void opencl_print_as_c_string(const char *str
, FILE *file
)
140 const char *prev
= str
;
142 while ((str
= strchr(prev
, '\n'))) {
143 fprintf(file
, "\n\"");
144 opencl_print_escaped(prev
, str
, file
);
145 fprintf(file
, "\\n\"");
151 fprintf(file
, "\n\"");
152 opencl_print_escaped(prev
, prev
+ strlen(prev
), file
);
157 /* Write the code that we have accumulated in the kernel isl_printer to the
158 * kernel.cl file. If the opencl_embed_kernel_code option has been set, print
159 * the code as a C string literal. Start that string literal with an empty
160 * line, such that line numbers reported by the OpenCL C compiler match those
161 * of the kernel file.
163 * Return 0 on success and -1 on failure.
165 static int opencl_write_kernel_file(struct opencl_info
*opencl
)
167 char *raw
= isl_printer_get_str(opencl
->kprinter
);
172 if (opencl
->options
->opencl_embed_kernel_code
) {
173 fprintf(opencl
->kernel_c
,
174 "static const char kernel_code[] = \"\\n\"");
175 opencl_print_as_c_string(raw
, opencl
->kernel_c
);
176 fprintf(opencl
->kernel_c
, ";\n");
178 fprintf(opencl
->kernel_c
, "%s", raw
);
185 /* Close all output files. Write the kernel contents to the kernel file before
188 * Return 0 on success and -1 on failure.
190 static int opencl_close_files(struct opencl_info
*info
)
194 if (info
->kernel_c
) {
195 r
= opencl_write_kernel_file(info
);
196 fclose(info
->kernel_c
);
199 fclose(info
->host_c
);
204 static __isl_give isl_printer
*opencl_print_host_macros(
205 __isl_take isl_printer
*p
)
208 "#define openclCheckReturn(ret) \\\n"
209 " if (ret != CL_SUCCESS) {\\\n"
210 " fprintf(stderr, \"OpenCL error: %s\\n\", "
211 "opencl_error_string(ret)); \\\n"
212 " fflush(stderr); \\\n"
213 " assert(ret == CL_SUCCESS);\\\n }\n";
215 p
= isl_printer_start_line(p
);
216 p
= isl_printer_print_str(p
, macros
);
217 p
= isl_printer_end_line(p
);
219 p
= isl_ast_op_type_print_macro(isl_ast_op_max
, p
);
224 static __isl_give isl_printer
*opencl_declare_device_arrays(
225 __isl_take isl_printer
*p
, struct gpu_prog
*prog
)
229 for (i
= 0; i
< prog
->n_array
; ++i
) {
230 if (!gpu_array_requires_device_allocation(&prog
->array
[i
]))
232 p
= isl_printer_start_line(p
);
233 p
= isl_printer_print_str(p
, "cl_mem dev_");
234 p
= isl_printer_print_str(p
, prog
->array
[i
].name
);
235 p
= isl_printer_print_str(p
, ";");
236 p
= isl_printer_end_line(p
);
238 p
= isl_printer_start_line(p
);
239 p
= isl_printer_end_line(p
);
243 /* Given an array, check whether its positive size guard expression is
246 static int is_array_positive_size_guard_trivial(struct gpu_array_info
*array
)
251 guard
= gpu_array_positive_size_guard(array
);
252 is_trivial
= isl_set_plain_is_universe(guard
);
257 /* Allocate a device array for "array'.
259 * Emit a max-expression to ensure the device array can contain at least one
260 * element if the array's positive size guard expression is not trivial.
262 static __isl_give isl_printer
*allocate_device_array(__isl_take isl_printer
*p
,
263 struct gpu_array_info
*array
)
265 int need_lower_bound
;
267 p
= ppcg_start_block(p
);
269 p
= isl_printer_start_line(p
);
270 p
= isl_printer_print_str(p
, "dev_");
271 p
= isl_printer_print_str(p
, array
->name
);
272 p
= isl_printer_print_str(p
, " = clCreateBuffer(context, ");
273 p
= isl_printer_print_str(p
, "CL_MEM_READ_WRITE, ");
275 need_lower_bound
= !is_array_positive_size_guard_trivial(array
);
276 if (need_lower_bound
) {
277 p
= isl_printer_print_str(p
, "max(sizeof(");
278 p
= isl_printer_print_str(p
, array
->type
);
279 p
= isl_printer_print_str(p
, "), ");
281 p
= gpu_array_info_print_size(p
, array
);
282 if (need_lower_bound
)
283 p
= isl_printer_print_str(p
, ")");
285 p
= isl_printer_print_str(p
, ", NULL, &err);");
286 p
= isl_printer_end_line(p
);
287 p
= isl_printer_start_line(p
);
288 p
= isl_printer_print_str(p
, "openclCheckReturn(err);");
289 p
= isl_printer_end_line(p
);
291 p
= ppcg_end_block(p
);
296 /* Allocate accessed device arrays.
298 static __isl_give isl_printer
*opencl_allocate_device_arrays(
299 __isl_take isl_printer
*p
, struct gpu_prog
*prog
)
303 for (i
= 0; i
< prog
->n_array
; ++i
) {
304 struct gpu_array_info
*array
= &prog
->array
[i
];
306 if (!gpu_array_requires_device_allocation(array
))
309 p
= allocate_device_array(p
, array
);
311 p
= isl_printer_start_line(p
);
312 p
= isl_printer_end_line(p
);
316 /* Print a call to the OpenCL clSetKernelArg() function which sets
317 * the arguments of the kernel. arg_name and arg_index are the name and the
318 * index of the kernel argument. The index of the leftmost argument of
319 * the kernel is 0 whereas the index of the rightmost argument of the kernel
320 * is n - 1, where n is the total number of the kernel arguments.
321 * read_only_scalar is a boolean that indicates whether the argument is a read
324 static __isl_give isl_printer
*opencl_set_kernel_argument(
325 __isl_take isl_printer
*p
, int kernel_id
,
326 const char *arg_name
, int arg_index
, int read_only_scalar
)
328 p
= isl_printer_start_line(p
);
329 p
= isl_printer_print_str(p
,
330 "openclCheckReturn(clSetKernelArg(kernel");
331 p
= isl_printer_print_int(p
, kernel_id
);
332 p
= isl_printer_print_str(p
, ", ");
333 p
= isl_printer_print_int(p
, arg_index
);
334 p
= isl_printer_print_str(p
, ", sizeof(");
336 if (read_only_scalar
) {
337 p
= isl_printer_print_str(p
, arg_name
);
338 p
= isl_printer_print_str(p
, "), &");
340 p
= isl_printer_print_str(p
, "cl_mem), (void *) &dev_");
342 p
= isl_printer_print_str(p
, arg_name
);
343 p
= isl_printer_print_str(p
, "));");
344 p
= isl_printer_end_line(p
);
349 /* Print the block sizes as a list of the sizes in each
352 static __isl_give isl_printer
*opencl_print_block_sizes(
353 __isl_take isl_printer
*p
, struct ppcg_kernel
*kernel
)
357 if (kernel
->n_block
> 0)
358 for (i
= 0; i
< kernel
->n_block
; ++i
) {
360 p
= isl_printer_print_str(p
, ", ");
361 p
= isl_printer_print_int(p
, kernel
->block_dim
[i
]);
364 p
= isl_printer_print_str(p
, "1");
369 /* Set the arguments of the OpenCL kernel by printing a call to the OpenCL
370 * clSetKernelArg() function for each kernel argument.
372 static __isl_give isl_printer
*opencl_set_kernel_arguments(
373 __isl_take isl_printer
*p
, struct gpu_prog
*prog
,
374 struct ppcg_kernel
*kernel
)
381 for (i
= 0; i
< prog
->n_array
; ++i
) {
384 required
= ppcg_kernel_requires_array_argument(kernel
, i
);
386 return isl_printer_free(p
);
389 ro
= gpu_array_is_read_only_scalar(&prog
->array
[i
]);
390 opencl_set_kernel_argument(p
, kernel
->id
, prog
->array
[i
].name
,
395 space
= isl_union_set_get_space(kernel
->arrays
);
396 nparam
= isl_space_dim(space
, isl_dim_param
);
397 for (i
= 0; i
< nparam
; ++i
) {
400 name
= isl_space_get_dim_name(space
, isl_dim_param
, i
);
401 opencl_set_kernel_argument(p
, kernel
->id
, name
, arg_index
, 1);
404 isl_space_free(space
);
406 n
= isl_space_dim(kernel
->space
, isl_dim_set
);
407 for (i
= 0; i
< n
; ++i
) {
410 name
= isl_space_get_dim_name(kernel
->space
, isl_dim_set
, i
);
411 opencl_set_kernel_argument(p
, kernel
->id
, name
, arg_index
, 1);
418 /* Print the arguments to a kernel declaration or call. If "types" is set,
419 * then print a declaration (including the types of the arguments).
421 * The arguments are printed in the following order
422 * - the arrays accessed by the kernel
424 * - the host loop iterators
426 static __isl_give isl_printer
*opencl_print_kernel_arguments(
427 __isl_take isl_printer
*p
, struct gpu_prog
*prog
,
428 struct ppcg_kernel
*kernel
, int types
)
436 for (i
= 0; i
< prog
->n_array
; ++i
) {
439 required
= ppcg_kernel_requires_array_argument(kernel
, i
);
441 return isl_printer_free(p
);
446 p
= isl_printer_print_str(p
, ", ");
449 p
= gpu_array_info_print_declaration_argument(p
,
450 &prog
->array
[i
], "__global");
452 p
= gpu_array_info_print_call_argument(p
,
458 space
= isl_union_set_get_space(kernel
->arrays
);
459 nparam
= isl_space_dim(space
, isl_dim_param
);
460 for (i
= 0; i
< nparam
; ++i
) {
463 name
= isl_space_get_dim_name(space
, isl_dim_param
, i
);
466 p
= isl_printer_print_str(p
, ", ");
468 p
= isl_printer_print_str(p
, "int ");
469 p
= isl_printer_print_str(p
, name
);
473 isl_space_free(space
);
475 n
= isl_space_dim(kernel
->space
, isl_dim_set
);
476 type
= isl_options_get_ast_iterator_type(prog
->ctx
);
477 for (i
= 0; i
< n
; ++i
) {
481 p
= isl_printer_print_str(p
, ", ");
482 name
= isl_space_get_dim_name(kernel
->space
, isl_dim_set
, i
);
484 p
= isl_printer_print_str(p
, type
);
485 p
= isl_printer_print_str(p
, " ");
487 p
= isl_printer_print_str(p
, name
);
495 /* Print the header of the given kernel.
497 static __isl_give isl_printer
*opencl_print_kernel_header(
498 __isl_take isl_printer
*p
, struct gpu_prog
*prog
,
499 struct ppcg_kernel
*kernel
)
501 p
= isl_printer_start_line(p
);
502 p
= isl_printer_print_str(p
, "__kernel void kernel");
503 p
= isl_printer_print_int(p
, kernel
->id
);
504 p
= isl_printer_print_str(p
, "(");
505 p
= opencl_print_kernel_arguments(p
, prog
, kernel
, 1);
506 p
= isl_printer_print_str(p
, ")");
507 p
= isl_printer_end_line(p
);
512 /* Print a list of iterators of type "type" with names "ids" to "p".
513 * Each iterator is assigned the corresponding opencl identifier returned
514 * by the function "opencl_id".
515 * Unlike the equivalent function in the CUDA backend which prints iterators
516 * in reverse order to promote coalescing, this function does not print
517 * iterators in reverse order. The OpenCL backend currently does not take
518 * into account any coalescing considerations.
520 static __isl_give isl_printer
*print_iterators(__isl_take isl_printer
*p
,
521 const char *type
, __isl_keep isl_id_list
*ids
, const char *opencl_id
)
525 n
= isl_id_list_n_id(ids
);
528 p
= isl_printer_start_line(p
);
529 p
= isl_printer_print_str(p
, type
);
530 p
= isl_printer_print_str(p
, " ");
531 for (i
= 0; i
< n
; ++i
) {
535 p
= isl_printer_print_str(p
, ", ");
536 id
= isl_id_list_get_id(ids
, i
);
537 p
= isl_printer_print_id(p
, id
);
539 p
= isl_printer_print_str(p
, " = ");
540 p
= isl_printer_print_str(p
, opencl_id
);
541 p
= isl_printer_print_str(p
, "(");
542 p
= isl_printer_print_int(p
, i
);
543 p
= isl_printer_print_str(p
, ")");
545 p
= isl_printer_print_str(p
, ";");
546 p
= isl_printer_end_line(p
);
551 static __isl_give isl_printer
*opencl_print_kernel_iterators(
552 __isl_take isl_printer
*p
, struct ppcg_kernel
*kernel
)
554 isl_ctx
*ctx
= isl_ast_node_get_ctx(kernel
->tree
);
557 type
= isl_options_get_ast_iterator_type(ctx
);
559 p
= print_iterators(p
, type
, kernel
->block_ids
, "get_group_id");
560 p
= print_iterators(p
, type
, kernel
->thread_ids
, "get_local_id");
565 static __isl_give isl_printer
*opencl_print_kernel_var(
566 __isl_take isl_printer
*p
, struct ppcg_kernel_var
*var
)
571 p
= isl_printer_start_line(p
);
572 if (var
->type
== ppcg_access_shared
)
573 p
= isl_printer_print_str(p
, "__local ");
574 p
= isl_printer_print_str(p
, var
->array
->type
);
575 p
= isl_printer_print_str(p
, " ");
576 p
= isl_printer_print_str(p
, var
->name
);
577 for (j
= 0; j
< var
->array
->n_index
; ++j
) {
578 p
= isl_printer_print_str(p
, "[");
579 v
= isl_vec_get_element_val(var
->size
, j
);
580 p
= isl_printer_print_val(p
, v
);
581 p
= isl_printer_print_str(p
, "]");
584 p
= isl_printer_print_str(p
, ";");
585 p
= isl_printer_end_line(p
);
590 static __isl_give isl_printer
*opencl_print_kernel_vars(
591 __isl_take isl_printer
*p
, struct ppcg_kernel
*kernel
)
595 for (i
= 0; i
< kernel
->n_var
; ++i
)
596 p
= opencl_print_kernel_var(p
, &kernel
->var
[i
]);
601 /* Print a call to barrier() which is a sync statement.
602 * All work-items in a work-group executing the kernel on a processor must
603 * execute the barrier() function before any are allowed to continue execution
604 * beyond the barrier.
605 * The flag CLK_LOCAL_MEM_FENCE makes the barrier function either flush any
606 * variables stored in local memory or queue a memory fence to ensure correct
607 * ordering of memory operations to local memory.
608 * The flag CLK_GLOBAL_MEM_FENCE makes the barrier function queue a memory
609 * fence to ensure correct ordering of memory operations to global memory.
611 static __isl_give isl_printer
*opencl_print_sync(__isl_take isl_printer
*p
,
612 struct ppcg_kernel_stmt
*stmt
)
614 p
= isl_printer_start_line(p
);
615 p
= isl_printer_print_str(p
,
616 "barrier(CLK_LOCAL_MEM_FENCE | CLK_GLOBAL_MEM_FENCE);");
617 p
= isl_printer_end_line(p
);
622 /* Data structure containing function names for which the calls
623 * should be changed from
629 * opencl_name((type) (arg))
631 static struct ppcg_opencl_fn
{
633 const char *opencl_name
;
636 { "expf", "exp", "float" },
637 { "powf", "pow", "float" },
638 { "sqrtf", "sqrt", "float" },
641 #define ARRAY_SIZE(array) (sizeof(array)/sizeof(*array))
643 /* If the name of function called by "expr" matches any of those
644 * in ppcg_opencl_fn, then replace the call by a cast to the corresponding
645 * type in ppcg_opencl_fn and a call to corresponding OpenCL function.
647 static __isl_give pet_expr
*map_opencl_call(__isl_take pet_expr
*expr
,
653 name
= pet_expr_call_get_name(expr
);
654 for (i
= 0; i
< ARRAY_SIZE(opencl_fn
); ++i
) {
657 if (strcmp(name
, opencl_fn
[i
].name
))
659 expr
= pet_expr_call_set_name(expr
, opencl_fn
[i
].opencl_name
);
660 arg
= pet_expr_get_arg(expr
, 0);
661 arg
= pet_expr_new_cast(opencl_fn
[i
].type
, arg
);
662 expr
= pet_expr_set_arg(expr
, 0, arg
);
667 /* Print the body of a statement from the input program,
668 * for use in OpenCL code.
670 * Before calling ppcg_kernel_print_domain to print the actual statement body,
671 * we first modify this body to take into account that the output code
672 * is OpenCL code. In particular, if the statement calls any function
673 * with a "f" suffix, then it needs to be replaced by a call to
674 * the corresponding function without suffix after casting the argument
677 static __isl_give isl_printer
*print_opencl_kernel_domain(
678 __isl_take isl_printer
*p
, struct ppcg_kernel_stmt
*stmt
)
683 ps
= stmt
->u
.d
.stmt
->stmt
;
684 tree
= pet_tree_copy(ps
->body
);
685 ps
->body
= pet_tree_map_call_expr(ps
->body
, &map_opencl_call
, NULL
);
686 p
= ppcg_kernel_print_domain(p
, stmt
);
687 pet_tree_free(ps
->body
);
693 /* This function is called for each user statement in the AST,
694 * i.e., for each kernel body statement, copy statement or sync statement.
696 static __isl_give isl_printer
*opencl_print_kernel_stmt(
697 __isl_take isl_printer
*p
,
698 __isl_take isl_ast_print_options
*print_options
,
699 __isl_keep isl_ast_node
*node
, void *user
)
702 struct ppcg_kernel_stmt
*stmt
;
704 id
= isl_ast_node_get_annotation(node
);
705 stmt
= isl_id_get_user(id
);
708 isl_ast_print_options_free(print_options
);
710 switch (stmt
->type
) {
711 case ppcg_kernel_copy
:
712 return ppcg_kernel_print_copy(p
, stmt
);
713 case ppcg_kernel_sync
:
714 return opencl_print_sync(p
, stmt
);
715 case ppcg_kernel_domain
:
716 return print_opencl_kernel_domain(p
, stmt
);
722 /* Return true if there is a double array in prog->array or
723 * if any of the types in prog->scop involve any doubles.
724 * To check the latter condition, we simply search for the string "double"
725 * in the type definitions, which may result in false positives.
727 static __isl_give
int any_double_elements(struct gpu_prog
*prog
)
731 for (i
= 0; i
< prog
->n_array
; ++i
)
732 if (strcmp(prog
->array
[i
].type
, "double") == 0)
735 for (i
= 0; i
< prog
->scop
->pet
->n_type
; ++i
) {
736 struct pet_type
*type
= prog
->scop
->pet
->types
[i
];
738 if (strstr(type
->definition
, "double"))
745 /* Prints a #pragma to enable support for double floating-point
746 * precision. OpenCL 1.0 adds support for double precision floating-point as
747 * an optional extension. An application that wants to use double will need to
748 * include the #pragma OPENCL EXTENSION cl_khr_fp64 : enable directive before
749 * any double precision data type is declared in the kernel code.
751 static __isl_give isl_printer
*opencl_enable_double_support(
752 __isl_take isl_printer
*p
)
754 p
= isl_printer_start_line(p
);
755 p
= isl_printer_print_str(p
, "#pragma OPENCL EXTENSION cl_khr_fp64 :"
757 p
= isl_printer_end_line(p
);
758 p
= isl_printer_start_line(p
);
759 p
= isl_printer_end_line(p
);
764 static __isl_give isl_printer
*opencl_print_kernel(struct gpu_prog
*prog
,
765 struct ppcg_kernel
*kernel
, __isl_take isl_printer
*p
)
767 isl_ctx
*ctx
= isl_ast_node_get_ctx(kernel
->tree
);
768 isl_ast_print_options
*print_options
;
770 print_options
= isl_ast_print_options_alloc(ctx
);
771 print_options
= isl_ast_print_options_set_print_user(print_options
,
772 &opencl_print_kernel_stmt
, NULL
);
774 p
= isl_printer_set_output_format(p
, ISL_FORMAT_C
);
775 p
= opencl_print_kernel_header(p
, prog
, kernel
);
776 p
= isl_printer_print_str(p
, "{");
777 p
= isl_printer_end_line(p
);
778 p
= isl_printer_indent(p
, 4);
779 p
= opencl_print_kernel_iterators(p
, kernel
);
780 p
= opencl_print_kernel_vars(p
, kernel
);
781 p
= isl_printer_end_line(p
);
782 p
= isl_ast_op_type_print_macro(isl_ast_op_fdiv_q
, p
);
783 p
= ppcg_print_macros(p
, kernel
->tree
);
784 p
= isl_ast_node_print(kernel
->tree
, p
, print_options
);
785 p
= isl_printer_indent(p
, -4);
786 p
= isl_printer_start_line(p
);
787 p
= isl_printer_print_str(p
, "}");
788 p
= isl_printer_end_line(p
);
793 struct print_host_user_data_opencl
{
794 struct opencl_info
*opencl
;
795 struct gpu_prog
*prog
;
798 /* This function prints the i'th block size multiplied by the i'th grid size,
799 * where i (a parameter to this function) is one of the possible dimensions of
800 * grid sizes and block sizes.
801 * If the dimension of block sizes is not equal to the dimension of grid sizes
802 * the output is calculated as follows:
805 * block_sizes[dim1] is the list of blocks sizes and it contains dim1 elements.
806 * grid_sizes[dim2] is the list of grid sizes and it contains dim2 elements.
809 * If (i > dim2) then the output is block_sizes[i]
810 * If (i > dim1) then the output is grid_sizes[i]
812 static __isl_give isl_printer
*opencl_print_total_number_of_work_items_for_dim(
813 __isl_take isl_printer
*p
, struct ppcg_kernel
*kernel
, int i
)
815 int grid_dim
, block_dim
;
816 isl_pw_aff
*bound_grid
;
818 grid_dim
= isl_multi_pw_aff_dim(kernel
->grid_size
, isl_dim_set
);
819 block_dim
= kernel
->n_block
;
821 if (i
< min(grid_dim
, block_dim
)) {
822 bound_grid
= isl_multi_pw_aff_get_pw_aff(kernel
->grid_size
, i
);
823 p
= isl_printer_print_str(p
, "(");
824 p
= isl_printer_print_pw_aff(p
, bound_grid
);
825 p
= isl_printer_print_str(p
, ") * ");
826 p
= isl_printer_print_int(p
, kernel
->block_dim
[i
]);
827 isl_pw_aff_free(bound_grid
);
828 } else if (i
>= grid_dim
)
829 p
= isl_printer_print_int(p
, kernel
->block_dim
[i
]);
831 bound_grid
= isl_multi_pw_aff_get_pw_aff(kernel
->grid_size
, i
);
832 p
= isl_printer_print_pw_aff(p
, bound_grid
);
833 isl_pw_aff_free(bound_grid
);
839 /* Print a list that represents the total number of work items. The list is
840 * constructed by performing an element-wise multiplication of the block sizes
841 * and the grid sizes. To explain how the list is constructed, suppose that:
842 * block_sizes[dim1] is the list of blocks sizes and it contains dim1 elements.
843 * grid_sizes[dim2] is the list of grid sizes and it contains dim2 elements.
845 * The output of this function is constructed as follows:
846 * If (dim1 > dim2) then the output is the following list:
847 * grid_sizes[0]*block_sizes[0], ..., grid_sizes[dim2-1]*block_sizes[dim2-1],
848 * block_sizes[dim2], ..., block_sizes[dim1-2], block_sizes[dim1-1].
850 * If (dim2 > dim1) then the output is the following list:
851 * grid_sizes[0]*block_sizes[0], ..., grid_sizes[dim1-1] * block_sizes[dim1-1],
852 * grid_sizes[dim1], grid_sizes[dim2-2], grid_sizes[dim2-1].
854 * To calculate the total number of work items out of the list constructed by
855 * this function, the user should multiply the elements of the list.
857 static __isl_give isl_printer
*opencl_print_total_number_of_work_items_as_list(
858 __isl_take isl_printer
*p
, struct ppcg_kernel
*kernel
)
861 int grid_dim
, block_dim
;
863 grid_dim
= isl_multi_pw_aff_dim(kernel
->grid_size
, isl_dim_set
);
864 block_dim
= kernel
->n_block
;
866 if ((grid_dim
<= 0) || (block_dim
<= 0)) {
867 p
= isl_printer_print_str(p
, "1");
871 for (i
= 0; i
<= max(grid_dim
, block_dim
) - 1; i
++) {
873 p
= isl_printer_print_str(p
, ", ");
875 p
= opencl_print_total_number_of_work_items_for_dim(p
,
882 /* Copy "array" from the host to the device (to_host = 0) or
883 * back from the device to the host (to_host = 1).
885 static __isl_give isl_printer
*copy_array(__isl_take isl_printer
*p
,
886 struct gpu_array_info
*array
, int to_host
)
888 p
= isl_printer_start_line(p
);
889 p
= isl_printer_print_str(p
, "openclCheckReturn(");
891 p
= isl_printer_print_str(p
, "clEnqueueReadBuffer");
893 p
= isl_printer_print_str(p
, "clEnqueueWriteBuffer");
894 p
= isl_printer_print_str(p
, "(queue, dev_");
895 p
= isl_printer_print_str(p
, array
->name
);
896 p
= isl_printer_print_str(p
, ", CL_TRUE, 0, ");
897 p
= gpu_array_info_print_size(p
, array
);
899 if (gpu_array_is_scalar(array
))
900 p
= isl_printer_print_str(p
, ", &");
902 p
= isl_printer_print_str(p
, ", ");
903 p
= isl_printer_print_str(p
, array
->name
);
904 p
= isl_printer_print_str(p
, ", 0, NULL, NULL));");
905 p
= isl_printer_end_line(p
);
910 /* Print a statement for copying an array to or from the device.
911 * The statement identifier is called "to_device_<array name>" or
912 * "from_device_<array name>" and its user pointer points
913 * to the gpu_array_info of the array that needs to be copied.
915 * Extract the array from the identifier and call
916 * copy_array_to_device or copy_array_from_device.
918 static __isl_give isl_printer
*print_to_from_device(__isl_take isl_printer
*p
,
919 __isl_keep isl_ast_node
*node
, struct gpu_prog
*prog
)
921 isl_ast_expr
*expr
, *arg
;
924 struct gpu_array_info
*array
;
926 expr
= isl_ast_node_user_get_expr(node
);
927 arg
= isl_ast_expr_get_op_arg(expr
, 0);
928 id
= isl_ast_expr_get_id(arg
);
929 name
= isl_id_get_name(id
);
930 array
= isl_id_get_user(id
);
932 isl_ast_expr_free(arg
);
933 isl_ast_expr_free(expr
);
938 return isl_printer_free(p
);
940 if (!prefixcmp(name
, "to_device"))
941 return copy_array(p
, array
, 0);
943 return copy_array(p
, array
, 1);
946 /* Print the user statement of the host code to "p".
948 * The host code may contain original user statements, kernel launches and
949 * statements that copy data to/from the device.
950 * The original user statements and the kernel launches have
951 * an associated annotation, while the data copy statements do not.
952 * The latter are handled by print_to_from_device.
953 * The annotation on the user statements is called "user".
955 * In case of a kernel launch, print a block of statements that
956 * defines the grid and the work group and then launches the kernel.
958 * A grid is composed of many work groups (blocks), each work group holds
959 * many work-items (threads).
961 * global_work_size[kernel->n_block] represents the total number of work
962 * items. It points to an array of kernel->n_block unsigned
963 * values that describe the total number of work-items that will execute
964 * the kernel. The total number of work-items is computed as:
965 * global_work_size[0] *...* global_work_size[kernel->n_block - 1].
967 * The size of each work group (i.e. the number of work-items in each work
968 * group) is described using block_size[kernel->n_block]. The total
969 * number of work-items in a block (work-group) is computed as:
970 * block_size[0] *... * block_size[kernel->n_block - 1].
972 * For more information check:
973 * http://www.khronos.org/registry/cl/sdk/1.0/docs/man/xhtml/clEnqueueNDRangeKernel.html
975 static __isl_give isl_printer
*opencl_print_host_user(
976 __isl_take isl_printer
*p
,
977 __isl_take isl_ast_print_options
*print_options
,
978 __isl_keep isl_ast_node
*node
, void *user
)
982 struct ppcg_kernel
*kernel
;
983 struct ppcg_kernel_stmt
*stmt
;
984 struct print_host_user_data_opencl
*data
;
986 isl_ast_print_options_free(print_options
);
988 data
= (struct print_host_user_data_opencl
*) user
;
990 id
= isl_ast_node_get_annotation(node
);
992 return print_to_from_device(p
, node
, data
->prog
);
994 is_user
= !strcmp(isl_id_get_name(id
), "user");
995 kernel
= is_user
? NULL
: isl_id_get_user(id
);
996 stmt
= is_user
? isl_id_get_user(id
) : NULL
;
1000 return ppcg_kernel_print_domain(p
, stmt
);
1002 p
= isl_printer_start_line(p
);
1003 p
= isl_printer_print_str(p
, "{");
1004 p
= isl_printer_end_line(p
);
1005 p
= isl_printer_indent(p
, 2);
1007 p
= isl_printer_start_line(p
);
1008 p
= isl_printer_print_str(p
, "size_t global_work_size[");
1010 if (kernel
->n_block
> 0)
1011 p
= isl_printer_print_int(p
, kernel
->n_block
);
1013 p
= isl_printer_print_int(p
, 1);
1015 p
= isl_printer_print_str(p
, "] = {");
1016 p
= opencl_print_total_number_of_work_items_as_list(p
, kernel
);
1017 p
= isl_printer_print_str(p
, "};");
1018 p
= isl_printer_end_line(p
);
1020 p
= isl_printer_start_line(p
);
1021 p
= isl_printer_print_str(p
, "size_t block_size[");
1023 if (kernel
->n_block
> 0)
1024 p
= isl_printer_print_int(p
, kernel
->n_block
);
1026 p
= isl_printer_print_int(p
, 1);
1028 p
= isl_printer_print_str(p
, "] = {");
1029 p
= opencl_print_block_sizes(p
, kernel
);
1030 p
= isl_printer_print_str(p
, "};");
1031 p
= isl_printer_end_line(p
);
1033 p
= isl_printer_start_line(p
);
1034 p
= isl_printer_print_str(p
, "cl_kernel kernel");
1035 p
= isl_printer_print_int(p
, kernel
->id
);
1036 p
= isl_printer_print_str(p
, " = clCreateKernel(program, \"kernel");
1037 p
= isl_printer_print_int(p
, kernel
->id
);
1038 p
= isl_printer_print_str(p
, "\", &err);");
1039 p
= isl_printer_end_line(p
);
1040 p
= isl_printer_start_line(p
);
1041 p
= isl_printer_print_str(p
, "openclCheckReturn(err);");
1042 p
= isl_printer_end_line(p
);
1044 opencl_set_kernel_arguments(p
, data
->prog
, kernel
);
1046 p
= isl_printer_start_line(p
);
1047 p
= isl_printer_print_str(p
, "openclCheckReturn(clEnqueueNDRangeKernel"
1049 p
= isl_printer_print_int(p
, kernel
->id
);
1050 p
= isl_printer_print_str(p
, ", ");
1051 if (kernel
->n_block
> 0)
1052 p
= isl_printer_print_int(p
, kernel
->n_block
);
1054 p
= isl_printer_print_int(p
, 1);
1056 p
= isl_printer_print_str(p
, ", NULL, global_work_size, "
1058 "0, NULL, NULL));");
1059 p
= isl_printer_end_line(p
);
1060 p
= isl_printer_start_line(p
);
1061 p
= isl_printer_print_str(p
, "openclCheckReturn("
1062 "clReleaseKernel(kernel");
1063 p
= isl_printer_print_int(p
, kernel
->id
);
1064 p
= isl_printer_print_str(p
, "));");
1065 p
= isl_printer_end_line(p
);
1066 p
= isl_printer_start_line(p
);
1067 p
= isl_printer_print_str(p
, "clFinish(queue);");
1068 p
= isl_printer_end_line(p
);
1069 p
= isl_printer_indent(p
, -2);
1070 p
= isl_printer_start_line(p
);
1071 p
= isl_printer_print_str(p
, "}");
1072 p
= isl_printer_end_line(p
);
1074 p
= isl_printer_start_line(p
);
1075 p
= isl_printer_end_line(p
);
1077 data
->opencl
->kprinter
= opencl_print_kernel(data
->prog
, kernel
,
1078 data
->opencl
->kprinter
);
1083 static __isl_give isl_printer
*opencl_print_host_code(
1084 __isl_take isl_printer
*p
, struct gpu_prog
*prog
,
1085 __isl_keep isl_ast_node
*tree
, struct opencl_info
*opencl
)
1087 isl_ast_print_options
*print_options
;
1088 isl_ctx
*ctx
= isl_ast_node_get_ctx(tree
);
1089 struct print_host_user_data_opencl data
= { opencl
, prog
};
1091 print_options
= isl_ast_print_options_alloc(ctx
);
1092 print_options
= isl_ast_print_options_set_print_user(print_options
,
1093 &opencl_print_host_user
, &data
);
1095 p
= ppcg_print_macros(p
, tree
);
1096 p
= isl_ast_node_print(tree
, p
, print_options
);
1101 /* Create an OpenCL device, context, command queue and build the kernel.
1102 * input is the name of the input file provided to ppcg.
1104 static __isl_give isl_printer
*opencl_setup(__isl_take isl_printer
*p
,
1105 const char *input
, struct opencl_info
*info
)
1107 p
= isl_printer_start_line(p
);
1108 p
= isl_printer_print_str(p
, "cl_device_id device;");
1109 p
= isl_printer_end_line(p
);
1110 p
= isl_printer_start_line(p
);
1111 p
= isl_printer_print_str(p
, "cl_context context;");
1112 p
= isl_printer_end_line(p
);
1113 p
= isl_printer_start_line(p
);
1114 p
= isl_printer_print_str(p
, "cl_program program;");
1115 p
= isl_printer_end_line(p
);
1116 p
= isl_printer_start_line(p
);
1117 p
= isl_printer_print_str(p
, "cl_command_queue queue;");
1118 p
= isl_printer_end_line(p
);
1119 p
= isl_printer_start_line(p
);
1120 p
= isl_printer_print_str(p
, "cl_int err;");
1121 p
= isl_printer_end_line(p
);
1122 p
= isl_printer_start_line(p
);
1123 p
= isl_printer_print_str(p
, "device = opencl_create_device(");
1124 p
= isl_printer_print_int(p
, info
->options
->opencl_use_gpu
);
1125 p
= isl_printer_print_str(p
, ");");
1126 p
= isl_printer_end_line(p
);
1127 p
= isl_printer_start_line(p
);
1128 p
= isl_printer_print_str(p
, "context = clCreateContext(NULL, 1, "
1129 "&device, NULL, NULL, &err);");
1130 p
= isl_printer_end_line(p
);
1131 p
= isl_printer_start_line(p
);
1132 p
= isl_printer_print_str(p
, "openclCheckReturn(err);");
1133 p
= isl_printer_end_line(p
);
1134 p
= isl_printer_start_line(p
);
1135 p
= isl_printer_print_str(p
, "queue = clCreateCommandQueue"
1136 "(context, device, 0, &err);");
1137 p
= isl_printer_end_line(p
);
1138 p
= isl_printer_start_line(p
);
1139 p
= isl_printer_print_str(p
, "openclCheckReturn(err);");
1140 p
= isl_printer_end_line(p
);
1142 p
= isl_printer_start_line(p
);
1143 p
= isl_printer_print_str(p
, "program = ");
1145 if (info
->options
->opencl_embed_kernel_code
) {
1146 p
= isl_printer_print_str(p
, "opencl_build_program_from_string("
1147 "context, device, kernel_code, "
1148 "sizeof(kernel_code), \"");
1150 p
= isl_printer_print_str(p
, "opencl_build_program_from_file("
1151 "context, device, \"");
1152 p
= isl_printer_print_str(p
, info
->kernel_c_name
);
1153 p
= isl_printer_print_str(p
, "\", \"");
1156 if (info
->options
->opencl_compiler_options
)
1157 p
= isl_printer_print_str(p
,
1158 info
->options
->opencl_compiler_options
);
1160 p
= isl_printer_print_str(p
, "\");");
1161 p
= isl_printer_end_line(p
);
1162 p
= isl_printer_start_line(p
);
1163 p
= isl_printer_end_line(p
);
1168 static __isl_give isl_printer
*opencl_release_cl_objects(
1169 __isl_take isl_printer
*p
, struct opencl_info
*info
)
1171 p
= isl_printer_start_line(p
);
1172 p
= isl_printer_print_str(p
, "openclCheckReturn(clReleaseCommandQueue"
1174 p
= isl_printer_end_line(p
);
1175 p
= isl_printer_start_line(p
);
1176 p
= isl_printer_print_str(p
, "openclCheckReturn(clReleaseProgram"
1178 p
= isl_printer_end_line(p
);
1179 p
= isl_printer_start_line(p
);
1180 p
= isl_printer_print_str(p
, "openclCheckReturn(clReleaseContext"
1182 p
= isl_printer_end_line(p
);
1187 /* Free the device array corresponding to "array"
1189 static __isl_give isl_printer
*release_device_array(__isl_take isl_printer
*p
,
1190 struct gpu_array_info
*array
)
1192 p
= isl_printer_start_line(p
);
1193 p
= isl_printer_print_str(p
, "openclCheckReturn("
1194 "clReleaseMemObject(dev_");
1195 p
= isl_printer_print_str(p
, array
->name
);
1196 p
= isl_printer_print_str(p
, "));");
1197 p
= isl_printer_end_line(p
);
1202 /* Free the accessed device arrays.
1204 static __isl_give isl_printer
*opencl_release_device_arrays(
1205 __isl_take isl_printer
*p
, struct gpu_prog
*prog
)
1209 for (i
= 0; i
< prog
->n_array
; ++i
) {
1210 struct gpu_array_info
*array
= &prog
->array
[i
];
1211 if (!gpu_array_requires_device_allocation(array
))
1214 p
= release_device_array(p
, array
);
1219 /* Given a gpu_prog "prog" and the corresponding transformed AST
1220 * "tree", print the entire OpenCL code to "p".
1222 static __isl_give isl_printer
*print_opencl(__isl_take isl_printer
*p
,
1223 struct gpu_prog
*prog
, __isl_keep isl_ast_node
*tree
,
1224 struct gpu_types
*types
, void *user
)
1226 struct opencl_info
*opencl
= user
;
1228 opencl
->kprinter
= isl_printer_set_output_format(opencl
->kprinter
,
1230 if (any_double_elements(prog
))
1231 opencl
->kprinter
= opencl_enable_double_support(
1233 if (opencl
->options
->opencl_print_kernel_types
)
1234 opencl
->kprinter
= gpu_print_types(opencl
->kprinter
, types
,
1237 if (!opencl
->kprinter
)
1238 return isl_printer_free(p
);
1240 p
= ppcg_start_block(p
);
1242 p
= opencl_print_host_macros(p
);
1244 p
= gpu_print_local_declarations(p
, prog
);
1245 p
= opencl_declare_device_arrays(p
, prog
);
1246 p
= opencl_setup(p
, opencl
->input
, opencl
);
1247 p
= opencl_allocate_device_arrays(p
, prog
);
1249 p
= opencl_print_host_code(p
, prog
, tree
, opencl
);
1251 p
= opencl_release_device_arrays(p
, prog
);
1252 p
= opencl_release_cl_objects(p
, opencl
);
1254 p
= ppcg_end_block(p
);
1259 /* Transform the code in the file called "input" by replacing
1260 * all scops by corresponding OpenCL code.
1261 * The host code is written to "output" or a name derived from
1262 * "input" if "output" is NULL.
1263 * The kernel code is placed in separate files with names
1264 * derived from "output" or "input".
1266 * We let generate_gpu do all the hard work and then let it call
1267 * us back for printing the AST in print_opencl.
1269 * To prepare for this printing, we first open the output files
1270 * and we close them after generate_gpu has finished.
1272 int generate_opencl(isl_ctx
*ctx
, struct ppcg_options
*options
,
1273 const char *input
, const char *output
)
1275 struct opencl_info opencl
= { options
, input
, output
};
1278 opencl
.kprinter
= isl_printer_to_str(ctx
);
1279 r
= opencl_open_files(&opencl
);
1282 r
= generate_gpu(ctx
, input
, opencl
.host_c
, options
,
1283 &print_opencl
, &opencl
);
1285 if (opencl_close_files(&opencl
) < 0)
1287 isl_printer_free(opencl
.kprinter
);