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_is_read_only_scalar(&prog
->array
[i
]))
232 if (!prog
->array
[i
].accessed
)
234 p
= isl_printer_start_line(p
);
235 p
= isl_printer_print_str(p
, "cl_mem dev_");
236 p
= isl_printer_print_str(p
, prog
->array
[i
].name
);
237 p
= isl_printer_print_str(p
, ";");
238 p
= isl_printer_end_line(p
);
240 p
= isl_printer_start_line(p
);
241 p
= isl_printer_end_line(p
);
245 /* Given an array, check whether its positive size guard expression is
248 static int is_array_positive_size_guard_trivial(struct gpu_array_info
*array
)
253 guard
= gpu_array_positive_size_guard(array
);
254 is_trivial
= isl_set_plain_is_universe(guard
);
259 /* Allocate a device array for "array'.
261 * Emit a max-expression to ensure the device array can contain at least one
262 * element if the array's positive size guard expression is not trivial.
264 static __isl_give isl_printer
*allocate_device_array(__isl_take isl_printer
*p
,
265 struct gpu_array_info
*array
)
267 int need_lower_bound
;
269 p
= ppcg_start_block(p
);
271 p
= isl_printer_start_line(p
);
272 p
= isl_printer_print_str(p
, "dev_");
273 p
= isl_printer_print_str(p
, array
->name
);
274 p
= isl_printer_print_str(p
, " = clCreateBuffer(context, ");
275 p
= isl_printer_print_str(p
, "CL_MEM_READ_WRITE, ");
277 need_lower_bound
= !is_array_positive_size_guard_trivial(array
);
278 if (need_lower_bound
) {
279 p
= isl_printer_print_str(p
, "max(sizeof(");
280 p
= isl_printer_print_str(p
, array
->type
);
281 p
= isl_printer_print_str(p
, "), ");
283 p
= gpu_array_info_print_size(p
, array
);
284 if (need_lower_bound
)
285 p
= isl_printer_print_str(p
, ")");
287 p
= isl_printer_print_str(p
, ", NULL, &err);");
288 p
= isl_printer_end_line(p
);
289 p
= isl_printer_start_line(p
);
290 p
= isl_printer_print_str(p
, "openclCheckReturn(err);");
291 p
= isl_printer_end_line(p
);
293 p
= ppcg_end_block(p
);
298 /* Allocate accessed device arrays.
300 static __isl_give isl_printer
*opencl_allocate_device_arrays(
301 __isl_take isl_printer
*p
, struct gpu_prog
*prog
)
305 for (i
= 0; i
< prog
->n_array
; ++i
) {
306 struct gpu_array_info
*array
= &prog
->array
[i
];
308 if (gpu_array_is_read_only_scalar(array
))
310 if (!array
->accessed
)
313 p
= allocate_device_array(p
, array
);
315 p
= isl_printer_start_line(p
);
316 p
= isl_printer_end_line(p
);
320 /* Print a call to the OpenCL clSetKernelArg() function which sets
321 * the arguments of the kernel. arg_name and arg_index are the name and the
322 * index of the kernel argument. The index of the leftmost argument of
323 * the kernel is 0 whereas the index of the rightmost argument of the kernel
324 * is n - 1, where n is the total number of the kernel arguments.
325 * read_only_scalar is a boolean that indicates whether the argument is a read
328 static __isl_give isl_printer
*opencl_set_kernel_argument(
329 __isl_take isl_printer
*p
, int kernel_id
,
330 const char *arg_name
, int arg_index
, int read_only_scalar
)
332 p
= isl_printer_start_line(p
);
333 p
= isl_printer_print_str(p
,
334 "openclCheckReturn(clSetKernelArg(kernel");
335 p
= isl_printer_print_int(p
, kernel_id
);
336 p
= isl_printer_print_str(p
, ", ");
337 p
= isl_printer_print_int(p
, arg_index
);
338 p
= isl_printer_print_str(p
, ", sizeof(");
340 if (read_only_scalar
) {
341 p
= isl_printer_print_str(p
, arg_name
);
342 p
= isl_printer_print_str(p
, "), &");
344 p
= isl_printer_print_str(p
, "cl_mem), (void *) &dev_");
346 p
= isl_printer_print_str(p
, arg_name
);
347 p
= isl_printer_print_str(p
, "));");
348 p
= isl_printer_end_line(p
);
353 /* Print the block sizes as a list of the sizes in each
356 static __isl_give isl_printer
*opencl_print_block_sizes(
357 __isl_take isl_printer
*p
, struct ppcg_kernel
*kernel
)
361 if (kernel
->n_block
> 0)
362 for (i
= 0; i
< kernel
->n_block
; ++i
) {
364 p
= isl_printer_print_str(p
, ", ");
365 p
= isl_printer_print_int(p
, kernel
->block_dim
[i
]);
368 p
= isl_printer_print_str(p
, "1");
373 /* Set the arguments of the OpenCL kernel by printing a call to the OpenCL
374 * clSetKernelArg() function for each kernel argument.
376 static __isl_give isl_printer
*opencl_set_kernel_arguments(
377 __isl_take isl_printer
*p
, struct gpu_prog
*prog
,
378 struct ppcg_kernel
*kernel
)
385 for (i
= 0; i
< prog
->n_array
; ++i
) {
388 required
= ppcg_kernel_requires_array_argument(kernel
, i
);
390 return isl_printer_free(p
);
393 ro
= gpu_array_is_read_only_scalar(&prog
->array
[i
]);
394 opencl_set_kernel_argument(p
, kernel
->id
, prog
->array
[i
].name
,
399 space
= isl_union_set_get_space(kernel
->arrays
);
400 nparam
= isl_space_dim(space
, isl_dim_param
);
401 for (i
= 0; i
< nparam
; ++i
) {
404 name
= isl_space_get_dim_name(space
, isl_dim_param
, i
);
405 opencl_set_kernel_argument(p
, kernel
->id
, name
, arg_index
, 1);
408 isl_space_free(space
);
410 n
= isl_space_dim(kernel
->space
, isl_dim_set
);
411 for (i
= 0; i
< n
; ++i
) {
414 name
= isl_space_get_dim_name(kernel
->space
, isl_dim_set
, i
);
415 opencl_set_kernel_argument(p
, kernel
->id
, name
, arg_index
, 1);
422 /* Print the arguments to a kernel declaration or call. If "types" is set,
423 * then print a declaration (including the types of the arguments).
425 * The arguments are printed in the following order
426 * - the arrays accessed by the kernel
428 * - the host loop iterators
430 static __isl_give isl_printer
*opencl_print_kernel_arguments(
431 __isl_take isl_printer
*p
, struct gpu_prog
*prog
,
432 struct ppcg_kernel
*kernel
, int types
)
440 for (i
= 0; i
< prog
->n_array
; ++i
) {
443 required
= ppcg_kernel_requires_array_argument(kernel
, i
);
445 return isl_printer_free(p
);
450 p
= isl_printer_print_str(p
, ", ");
453 p
= gpu_array_info_print_declaration_argument(p
,
454 &prog
->array
[i
], "__global");
456 p
= gpu_array_info_print_call_argument(p
,
462 space
= isl_union_set_get_space(kernel
->arrays
);
463 nparam
= isl_space_dim(space
, isl_dim_param
);
464 for (i
= 0; i
< nparam
; ++i
) {
467 name
= isl_space_get_dim_name(space
, isl_dim_param
, i
);
470 p
= isl_printer_print_str(p
, ", ");
472 p
= isl_printer_print_str(p
, "int ");
473 p
= isl_printer_print_str(p
, name
);
477 isl_space_free(space
);
479 n
= isl_space_dim(kernel
->space
, isl_dim_set
);
480 type
= isl_options_get_ast_iterator_type(prog
->ctx
);
481 for (i
= 0; i
< n
; ++i
) {
485 p
= isl_printer_print_str(p
, ", ");
486 name
= isl_space_get_dim_name(kernel
->space
, isl_dim_set
, i
);
488 p
= isl_printer_print_str(p
, type
);
489 p
= isl_printer_print_str(p
, " ");
491 p
= isl_printer_print_str(p
, name
);
499 /* Print the header of the given kernel.
501 static __isl_give isl_printer
*opencl_print_kernel_header(
502 __isl_take isl_printer
*p
, struct gpu_prog
*prog
,
503 struct ppcg_kernel
*kernel
)
505 p
= isl_printer_start_line(p
);
506 p
= isl_printer_print_str(p
, "__kernel void kernel");
507 p
= isl_printer_print_int(p
, kernel
->id
);
508 p
= isl_printer_print_str(p
, "(");
509 p
= opencl_print_kernel_arguments(p
, prog
, kernel
, 1);
510 p
= isl_printer_print_str(p
, ")");
511 p
= isl_printer_end_line(p
);
516 /* Print a list of iterators of type "type" with names "ids" to "p".
517 * Each iterator is assigned the corresponding opencl identifier returned
518 * by the function "opencl_id".
519 * Unlike the equivalent function in the CUDA backend which prints iterators
520 * in reverse order to promote coalescing, this function does not print
521 * iterators in reverse order. The OpenCL backend currently does not take
522 * into account any coalescing considerations.
524 static __isl_give isl_printer
*print_iterators(__isl_take isl_printer
*p
,
525 const char *type
, __isl_keep isl_id_list
*ids
, const char *opencl_id
)
529 n
= isl_id_list_n_id(ids
);
532 p
= isl_printer_start_line(p
);
533 p
= isl_printer_print_str(p
, type
);
534 p
= isl_printer_print_str(p
, " ");
535 for (i
= 0; i
< n
; ++i
) {
539 p
= isl_printer_print_str(p
, ", ");
540 id
= isl_id_list_get_id(ids
, i
);
541 p
= isl_printer_print_id(p
, id
);
543 p
= isl_printer_print_str(p
, " = ");
544 p
= isl_printer_print_str(p
, opencl_id
);
545 p
= isl_printer_print_str(p
, "(");
546 p
= isl_printer_print_int(p
, i
);
547 p
= isl_printer_print_str(p
, ")");
549 p
= isl_printer_print_str(p
, ";");
550 p
= isl_printer_end_line(p
);
555 static __isl_give isl_printer
*opencl_print_kernel_iterators(
556 __isl_take isl_printer
*p
, struct ppcg_kernel
*kernel
)
558 isl_ctx
*ctx
= isl_ast_node_get_ctx(kernel
->tree
);
561 type
= isl_options_get_ast_iterator_type(ctx
);
563 p
= print_iterators(p
, type
, kernel
->block_ids
, "get_group_id");
564 p
= print_iterators(p
, type
, kernel
->thread_ids
, "get_local_id");
569 static __isl_give isl_printer
*opencl_print_kernel_var(
570 __isl_take isl_printer
*p
, struct ppcg_kernel_var
*var
)
575 p
= isl_printer_start_line(p
);
576 if (var
->type
== ppcg_access_shared
)
577 p
= isl_printer_print_str(p
, "__local ");
578 p
= isl_printer_print_str(p
, var
->array
->type
);
579 p
= isl_printer_print_str(p
, " ");
580 p
= isl_printer_print_str(p
, var
->name
);
581 for (j
= 0; j
< var
->array
->n_index
; ++j
) {
582 p
= isl_printer_print_str(p
, "[");
583 v
= isl_vec_get_element_val(var
->size
, j
);
584 p
= isl_printer_print_val(p
, v
);
585 p
= isl_printer_print_str(p
, "]");
588 p
= isl_printer_print_str(p
, ";");
589 p
= isl_printer_end_line(p
);
594 static __isl_give isl_printer
*opencl_print_kernel_vars(
595 __isl_take isl_printer
*p
, struct ppcg_kernel
*kernel
)
599 for (i
= 0; i
< kernel
->n_var
; ++i
)
600 p
= opencl_print_kernel_var(p
, &kernel
->var
[i
]);
605 /* Print a call to barrier() which is a sync statement.
606 * All work-items in a work-group executing the kernel on a processor must
607 * execute the barrier() function before any are allowed to continue execution
608 * beyond the barrier.
609 * The flag CLK_LOCAL_MEM_FENCE makes the barrier function either flush any
610 * variables stored in local memory or queue a memory fence to ensure correct
611 * ordering of memory operations to local memory.
612 * The flag CLK_GLOBAL_MEM_FENCE makes the barrier function queue a memory
613 * fence to ensure correct ordering of memory operations to global memory.
615 static __isl_give isl_printer
*opencl_print_sync(__isl_take isl_printer
*p
,
616 struct ppcg_kernel_stmt
*stmt
)
618 p
= isl_printer_start_line(p
);
619 p
= isl_printer_print_str(p
,
620 "barrier(CLK_LOCAL_MEM_FENCE | CLK_GLOBAL_MEM_FENCE);");
621 p
= isl_printer_end_line(p
);
626 /* Data structure containing function names for which the calls
627 * should be changed from
633 * opencl_name((type) (arg))
635 static struct ppcg_opencl_fn
{
637 const char *opencl_name
;
640 { "expf", "exp", "float" },
641 { "powf", "pow", "float" },
642 { "sqrtf", "sqrt", "float" },
645 #define ARRAY_SIZE(array) (sizeof(array)/sizeof(*array))
647 /* If the name of function called by "expr" matches any of those
648 * in ppcg_opencl_fn, then replace the call by a cast to the corresponding
649 * type in ppcg_opencl_fn and a call to corresponding OpenCL function.
651 static __isl_give pet_expr
*map_opencl_call(__isl_take pet_expr
*expr
,
657 name
= pet_expr_call_get_name(expr
);
658 for (i
= 0; i
< ARRAY_SIZE(opencl_fn
); ++i
) {
661 if (strcmp(name
, opencl_fn
[i
].name
))
663 expr
= pet_expr_call_set_name(expr
, opencl_fn
[i
].opencl_name
);
664 arg
= pet_expr_get_arg(expr
, 0);
665 arg
= pet_expr_new_cast(opencl_fn
[i
].type
, arg
);
666 expr
= pet_expr_set_arg(expr
, 0, arg
);
671 /* Print the body of a statement from the input program,
672 * for use in OpenCL code.
674 * Before calling ppcg_kernel_print_domain to print the actual statement body,
675 * we first modify this body to take into account that the output code
676 * is OpenCL code. In particular, if the statement calls any function
677 * with a "f" suffix, then it needs to be replaced by a call to
678 * the corresponding function without suffix after casting the argument
681 static __isl_give isl_printer
*print_opencl_kernel_domain(
682 __isl_take isl_printer
*p
, struct ppcg_kernel_stmt
*stmt
)
687 ps
= stmt
->u
.d
.stmt
->stmt
;
688 tree
= pet_tree_copy(ps
->body
);
689 ps
->body
= pet_tree_map_call_expr(ps
->body
, &map_opencl_call
, NULL
);
690 p
= ppcg_kernel_print_domain(p
, stmt
);
691 pet_tree_free(ps
->body
);
697 /* This function is called for each user statement in the AST,
698 * i.e., for each kernel body statement, copy statement or sync statement.
700 static __isl_give isl_printer
*opencl_print_kernel_stmt(
701 __isl_take isl_printer
*p
,
702 __isl_take isl_ast_print_options
*print_options
,
703 __isl_keep isl_ast_node
*node
, void *user
)
706 struct ppcg_kernel_stmt
*stmt
;
708 id
= isl_ast_node_get_annotation(node
);
709 stmt
= isl_id_get_user(id
);
712 isl_ast_print_options_free(print_options
);
714 switch (stmt
->type
) {
715 case ppcg_kernel_copy
:
716 return ppcg_kernel_print_copy(p
, stmt
);
717 case ppcg_kernel_sync
:
718 return opencl_print_sync(p
, stmt
);
719 case ppcg_kernel_domain
:
720 return print_opencl_kernel_domain(p
, stmt
);
726 /* Return true if there is a double array in prog->array or
727 * if any of the types in prog->scop involve any doubles.
728 * To check the latter condition, we simply search for the string "double"
729 * in the type definitions, which may result in false positives.
731 static __isl_give
int any_double_elements(struct gpu_prog
*prog
)
735 for (i
= 0; i
< prog
->n_array
; ++i
)
736 if (strcmp(prog
->array
[i
].type
, "double") == 0)
739 for (i
= 0; i
< prog
->scop
->pet
->n_type
; ++i
) {
740 struct pet_type
*type
= prog
->scop
->pet
->types
[i
];
742 if (strstr(type
->definition
, "double"))
749 /* Prints a #pragma to enable support for double floating-point
750 * precision. OpenCL 1.0 adds support for double precision floating-point as
751 * an optional extension. An application that wants to use double will need to
752 * include the #pragma OPENCL EXTENSION cl_khr_fp64 : enable directive before
753 * any double precision data type is declared in the kernel code.
755 static __isl_give isl_printer
*opencl_enable_double_support(
756 __isl_take isl_printer
*p
)
758 p
= isl_printer_start_line(p
);
759 p
= isl_printer_print_str(p
, "#pragma OPENCL EXTENSION cl_khr_fp64 :"
761 p
= isl_printer_end_line(p
);
762 p
= isl_printer_start_line(p
);
763 p
= isl_printer_end_line(p
);
768 static __isl_give isl_printer
*opencl_print_kernel(struct gpu_prog
*prog
,
769 struct ppcg_kernel
*kernel
, __isl_take isl_printer
*p
)
771 isl_ctx
*ctx
= isl_ast_node_get_ctx(kernel
->tree
);
772 isl_ast_print_options
*print_options
;
774 print_options
= isl_ast_print_options_alloc(ctx
);
775 print_options
= isl_ast_print_options_set_print_user(print_options
,
776 &opencl_print_kernel_stmt
, NULL
);
778 p
= isl_printer_set_output_format(p
, ISL_FORMAT_C
);
779 p
= opencl_print_kernel_header(p
, prog
, kernel
);
780 p
= isl_printer_print_str(p
, "{");
781 p
= isl_printer_end_line(p
);
782 p
= isl_printer_indent(p
, 4);
783 p
= opencl_print_kernel_iterators(p
, kernel
);
784 p
= opencl_print_kernel_vars(p
, kernel
);
785 p
= isl_printer_end_line(p
);
786 p
= gpu_print_macros(p
, kernel
->tree
);
787 p
= isl_ast_node_print(kernel
->tree
, p
, print_options
);
788 p
= isl_printer_indent(p
, -4);
789 p
= isl_printer_start_line(p
);
790 p
= isl_printer_print_str(p
, "}");
791 p
= isl_printer_end_line(p
);
796 struct print_host_user_data_opencl
{
797 struct opencl_info
*opencl
;
798 struct gpu_prog
*prog
;
801 /* This function prints the i'th block size multiplied by the i'th grid size,
802 * where i (a parameter to this function) is one of the possible dimensions of
803 * grid sizes and block sizes.
804 * If the dimension of block sizes is not equal to the dimension of grid sizes
805 * the output is calculated as follows:
808 * block_sizes[dim1] is the list of blocks sizes and it contains dim1 elements.
809 * grid_sizes[dim2] is the list of grid sizes and it contains dim2 elements.
812 * If (i > dim2) then the output is block_sizes[i]
813 * If (i > dim1) then the output is grid_sizes[i]
815 static __isl_give isl_printer
*opencl_print_total_number_of_work_items_for_dim(
816 __isl_take isl_printer
*p
, struct ppcg_kernel
*kernel
, int i
)
818 int grid_dim
, block_dim
;
819 isl_pw_aff
*bound_grid
;
821 grid_dim
= isl_multi_pw_aff_dim(kernel
->grid_size
, isl_dim_set
);
822 block_dim
= kernel
->n_block
;
824 if (i
< min(grid_dim
, block_dim
)) {
825 bound_grid
= isl_multi_pw_aff_get_pw_aff(kernel
->grid_size
, i
);
826 p
= isl_printer_print_str(p
, "(");
827 p
= isl_printer_print_pw_aff(p
, bound_grid
);
828 p
= isl_printer_print_str(p
, ") * ");
829 p
= isl_printer_print_int(p
, kernel
->block_dim
[i
]);
830 isl_pw_aff_free(bound_grid
);
831 } else if (i
>= grid_dim
)
832 p
= isl_printer_print_int(p
, kernel
->block_dim
[i
]);
834 bound_grid
= isl_multi_pw_aff_get_pw_aff(kernel
->grid_size
, i
);
835 p
= isl_printer_print_pw_aff(p
, bound_grid
);
836 isl_pw_aff_free(bound_grid
);
842 /* Print a list that represents the total number of work items. The list is
843 * constructed by performing an element-wise multiplication of the block sizes
844 * and the grid sizes. To explain how the list is constructed, suppose that:
845 * block_sizes[dim1] is the list of blocks sizes and it contains dim1 elements.
846 * grid_sizes[dim2] is the list of grid sizes and it contains dim2 elements.
848 * The output of this function is constructed as follows:
849 * If (dim1 > dim2) then the output is the following list:
850 * grid_sizes[0]*block_sizes[0], ..., grid_sizes[dim2-1]*block_sizes[dim2-1],
851 * block_sizes[dim2], ..., block_sizes[dim1-2], block_sizes[dim1-1].
853 * If (dim2 > dim1) then the output is the following list:
854 * grid_sizes[0]*block_sizes[0], ..., grid_sizes[dim1-1] * block_sizes[dim1-1],
855 * grid_sizes[dim1], grid_sizes[dim2-2], grid_sizes[dim2-1].
857 * To calculate the total number of work items out of the list constructed by
858 * this function, the user should multiply the elements of the list.
860 static __isl_give isl_printer
*opencl_print_total_number_of_work_items_as_list(
861 __isl_take isl_printer
*p
, struct ppcg_kernel
*kernel
)
864 int grid_dim
, block_dim
;
866 grid_dim
= isl_multi_pw_aff_dim(kernel
->grid_size
, isl_dim_set
);
867 block_dim
= kernel
->n_block
;
869 if ((grid_dim
<= 0) || (block_dim
<= 0)) {
870 p
= isl_printer_print_str(p
, "1");
874 for (i
= 0; i
<= max(grid_dim
, block_dim
) - 1; i
++) {
876 p
= isl_printer_print_str(p
, ", ");
878 p
= opencl_print_total_number_of_work_items_for_dim(p
,
885 /* Copy "array" from the host to the device (to_host = 0) or
886 * back from the device to the host (to_host = 1).
888 static __isl_give isl_printer
*copy_array(__isl_take isl_printer
*p
,
889 struct gpu_array_info
*array
, int to_host
)
891 p
= isl_printer_start_line(p
);
892 p
= isl_printer_print_str(p
, "openclCheckReturn(");
894 p
= isl_printer_print_str(p
, "clEnqueueReadBuffer");
896 p
= isl_printer_print_str(p
, "clEnqueueWriteBuffer");
897 p
= isl_printer_print_str(p
, "(queue, dev_");
898 p
= isl_printer_print_str(p
, array
->name
);
899 p
= isl_printer_print_str(p
, ", CL_TRUE, 0, ");
900 p
= gpu_array_info_print_size(p
, array
);
902 if (gpu_array_is_scalar(array
))
903 p
= isl_printer_print_str(p
, ", &");
905 p
= isl_printer_print_str(p
, ", ");
906 p
= isl_printer_print_str(p
, array
->name
);
907 p
= isl_printer_print_str(p
, ", 0, NULL, NULL));");
908 p
= isl_printer_end_line(p
);
913 /* Print a statement for copying an array to or from the device.
914 * The statement identifier is called "to_device_<array name>" or
915 * "from_device_<array name>" and its user pointer points
916 * to the gpu_array_info of the array that needs to be copied.
918 * Extract the array from the identifier and call
919 * copy_array_to_device or copy_array_from_device.
921 static __isl_give isl_printer
*print_to_from_device(__isl_take isl_printer
*p
,
922 __isl_keep isl_ast_node
*node
, struct gpu_prog
*prog
)
924 isl_ast_expr
*expr
, *arg
;
927 struct gpu_array_info
*array
;
929 expr
= isl_ast_node_user_get_expr(node
);
930 arg
= isl_ast_expr_get_op_arg(expr
, 0);
931 id
= isl_ast_expr_get_id(arg
);
932 name
= isl_id_get_name(id
);
933 array
= isl_id_get_user(id
);
935 isl_ast_expr_free(arg
);
936 isl_ast_expr_free(expr
);
941 return isl_printer_free(p
);
943 if (!prefixcmp(name
, "to_device"))
944 return copy_array(p
, array
, 0);
946 return copy_array(p
, array
, 1);
949 /* Print the user statement of the host code to "p".
951 * The host code may contain original user statements, kernel launches and
952 * statements that copy data to/from the device.
953 * The original user statements and the kernel launches have
954 * an associated annotation, while the data copy statements do not.
955 * The latter are handled by print_to_from_device.
956 * The annotation on the user statements is called "user".
958 * In case of a kernel launch, print a block of statements that
959 * defines the grid and the work group and then launches the kernel.
961 * A grid is composed of many work groups (blocks), each work group holds
962 * many work-items (threads).
964 * global_work_size[kernel->n_block] represents the total number of work
965 * items. It points to an array of kernel->n_block unsigned
966 * values that describe the total number of work-items that will execute
967 * the kernel. The total number of work-items is computed as:
968 * global_work_size[0] *...* global_work_size[kernel->n_block - 1].
970 * The size of each work group (i.e. the number of work-items in each work
971 * group) is described using block_size[kernel->n_block]. The total
972 * number of work-items in a block (work-group) is computed as:
973 * block_size[0] *... * block_size[kernel->n_block - 1].
975 * For more information check:
976 * http://www.khronos.org/registry/cl/sdk/1.0/docs/man/xhtml/clEnqueueNDRangeKernel.html
978 static __isl_give isl_printer
*opencl_print_host_user(
979 __isl_take isl_printer
*p
,
980 __isl_take isl_ast_print_options
*print_options
,
981 __isl_keep isl_ast_node
*node
, void *user
)
985 struct ppcg_kernel
*kernel
;
986 struct ppcg_kernel_stmt
*stmt
;
987 struct print_host_user_data_opencl
*data
;
989 isl_ast_print_options_free(print_options
);
991 data
= (struct print_host_user_data_opencl
*) user
;
993 id
= isl_ast_node_get_annotation(node
);
995 return print_to_from_device(p
, node
, data
->prog
);
997 is_user
= !strcmp(isl_id_get_name(id
), "user");
998 kernel
= is_user
? NULL
: isl_id_get_user(id
);
999 stmt
= is_user
? isl_id_get_user(id
) : NULL
;
1003 return ppcg_kernel_print_domain(p
, stmt
);
1005 p
= isl_printer_start_line(p
);
1006 p
= isl_printer_print_str(p
, "{");
1007 p
= isl_printer_end_line(p
);
1008 p
= isl_printer_indent(p
, 2);
1010 p
= isl_printer_start_line(p
);
1011 p
= isl_printer_print_str(p
, "size_t global_work_size[");
1013 if (kernel
->n_block
> 0)
1014 p
= isl_printer_print_int(p
, kernel
->n_block
);
1016 p
= isl_printer_print_int(p
, 1);
1018 p
= isl_printer_print_str(p
, "] = {");
1019 p
= opencl_print_total_number_of_work_items_as_list(p
, kernel
);
1020 p
= isl_printer_print_str(p
, "};");
1021 p
= isl_printer_end_line(p
);
1023 p
= isl_printer_start_line(p
);
1024 p
= isl_printer_print_str(p
, "size_t block_size[");
1026 if (kernel
->n_block
> 0)
1027 p
= isl_printer_print_int(p
, kernel
->n_block
);
1029 p
= isl_printer_print_int(p
, 1);
1031 p
= isl_printer_print_str(p
, "] = {");
1032 p
= opencl_print_block_sizes(p
, kernel
);
1033 p
= isl_printer_print_str(p
, "};");
1034 p
= isl_printer_end_line(p
);
1036 p
= isl_printer_start_line(p
);
1037 p
= isl_printer_print_str(p
, "cl_kernel kernel");
1038 p
= isl_printer_print_int(p
, kernel
->id
);
1039 p
= isl_printer_print_str(p
, " = clCreateKernel(program, \"kernel");
1040 p
= isl_printer_print_int(p
, kernel
->id
);
1041 p
= isl_printer_print_str(p
, "\", &err);");
1042 p
= isl_printer_end_line(p
);
1043 p
= isl_printer_start_line(p
);
1044 p
= isl_printer_print_str(p
, "openclCheckReturn(err);");
1045 p
= isl_printer_end_line(p
);
1047 opencl_set_kernel_arguments(p
, data
->prog
, kernel
);
1049 p
= isl_printer_start_line(p
);
1050 p
= isl_printer_print_str(p
, "openclCheckReturn(clEnqueueNDRangeKernel"
1052 p
= isl_printer_print_int(p
, kernel
->id
);
1053 p
= isl_printer_print_str(p
, ", ");
1054 if (kernel
->n_block
> 0)
1055 p
= isl_printer_print_int(p
, kernel
->n_block
);
1057 p
= isl_printer_print_int(p
, 1);
1059 p
= isl_printer_print_str(p
, ", NULL, global_work_size, "
1061 "0, NULL, NULL));");
1062 p
= isl_printer_end_line(p
);
1063 p
= isl_printer_start_line(p
);
1064 p
= isl_printer_print_str(p
, "openclCheckReturn("
1065 "clReleaseKernel(kernel");
1066 p
= isl_printer_print_int(p
, kernel
->id
);
1067 p
= isl_printer_print_str(p
, "));");
1068 p
= isl_printer_end_line(p
);
1069 p
= isl_printer_start_line(p
);
1070 p
= isl_printer_print_str(p
, "clFinish(queue);");
1071 p
= isl_printer_end_line(p
);
1072 p
= isl_printer_indent(p
, -2);
1073 p
= isl_printer_start_line(p
);
1074 p
= isl_printer_print_str(p
, "}");
1075 p
= isl_printer_end_line(p
);
1077 p
= isl_printer_start_line(p
);
1078 p
= isl_printer_end_line(p
);
1080 data
->opencl
->kprinter
= opencl_print_kernel(data
->prog
, kernel
,
1081 data
->opencl
->kprinter
);
1086 static __isl_give isl_printer
*opencl_print_host_code(
1087 __isl_take isl_printer
*p
, struct gpu_prog
*prog
,
1088 __isl_keep isl_ast_node
*tree
, struct opencl_info
*opencl
)
1090 isl_ast_print_options
*print_options
;
1091 isl_ctx
*ctx
= isl_ast_node_get_ctx(tree
);
1092 struct print_host_user_data_opencl data
= { opencl
, prog
};
1094 print_options
= isl_ast_print_options_alloc(ctx
);
1095 print_options
= isl_ast_print_options_set_print_user(print_options
,
1096 &opencl_print_host_user
, &data
);
1098 p
= gpu_print_macros(p
, tree
);
1099 p
= isl_ast_node_print(tree
, p
, print_options
);
1104 /* Create an OpenCL device, context, command queue and build the kernel.
1105 * input is the name of the input file provided to ppcg.
1107 static __isl_give isl_printer
*opencl_setup(__isl_take isl_printer
*p
,
1108 const char *input
, struct opencl_info
*info
)
1110 p
= isl_printer_start_line(p
);
1111 p
= isl_printer_print_str(p
, "cl_device_id device;");
1112 p
= isl_printer_end_line(p
);
1113 p
= isl_printer_start_line(p
);
1114 p
= isl_printer_print_str(p
, "cl_context context;");
1115 p
= isl_printer_end_line(p
);
1116 p
= isl_printer_start_line(p
);
1117 p
= isl_printer_print_str(p
, "cl_program program;");
1118 p
= isl_printer_end_line(p
);
1119 p
= isl_printer_start_line(p
);
1120 p
= isl_printer_print_str(p
, "cl_command_queue queue;");
1121 p
= isl_printer_end_line(p
);
1122 p
= isl_printer_start_line(p
);
1123 p
= isl_printer_print_str(p
, "cl_int err;");
1124 p
= isl_printer_end_line(p
);
1125 p
= isl_printer_start_line(p
);
1126 p
= isl_printer_print_str(p
, "device = opencl_create_device(");
1127 p
= isl_printer_print_int(p
, info
->options
->opencl_use_gpu
);
1128 p
= isl_printer_print_str(p
, ");");
1129 p
= isl_printer_end_line(p
);
1130 p
= isl_printer_start_line(p
);
1131 p
= isl_printer_print_str(p
, "context = clCreateContext(NULL, 1, "
1132 "&device, NULL, NULL, &err);");
1133 p
= isl_printer_end_line(p
);
1134 p
= isl_printer_start_line(p
);
1135 p
= isl_printer_print_str(p
, "openclCheckReturn(err);");
1136 p
= isl_printer_end_line(p
);
1137 p
= isl_printer_start_line(p
);
1138 p
= isl_printer_print_str(p
, "queue = clCreateCommandQueue"
1139 "(context, device, 0, &err);");
1140 p
= isl_printer_end_line(p
);
1141 p
= isl_printer_start_line(p
);
1142 p
= isl_printer_print_str(p
, "openclCheckReturn(err);");
1143 p
= isl_printer_end_line(p
);
1145 p
= isl_printer_start_line(p
);
1146 p
= isl_printer_print_str(p
, "program = ");
1148 if (info
->options
->opencl_embed_kernel_code
) {
1149 p
= isl_printer_print_str(p
, "opencl_build_program_from_string("
1150 "context, device, kernel_code, "
1151 "sizeof(kernel_code), \"");
1153 p
= isl_printer_print_str(p
, "opencl_build_program_from_file("
1154 "context, device, \"");
1155 p
= isl_printer_print_str(p
, info
->kernel_c_name
);
1156 p
= isl_printer_print_str(p
, "\", \"");
1159 if (info
->options
->opencl_compiler_options
)
1160 p
= isl_printer_print_str(p
,
1161 info
->options
->opencl_compiler_options
);
1163 p
= isl_printer_print_str(p
, "\");");
1164 p
= isl_printer_end_line(p
);
1165 p
= isl_printer_start_line(p
);
1166 p
= isl_printer_end_line(p
);
1171 static __isl_give isl_printer
*opencl_release_cl_objects(
1172 __isl_take isl_printer
*p
, struct opencl_info
*info
)
1174 p
= isl_printer_start_line(p
);
1175 p
= isl_printer_print_str(p
, "openclCheckReturn(clReleaseCommandQueue"
1177 p
= isl_printer_end_line(p
);
1178 p
= isl_printer_start_line(p
);
1179 p
= isl_printer_print_str(p
, "openclCheckReturn(clReleaseProgram"
1181 p
= isl_printer_end_line(p
);
1182 p
= isl_printer_start_line(p
);
1183 p
= isl_printer_print_str(p
, "openclCheckReturn(clReleaseContext"
1185 p
= isl_printer_end_line(p
);
1190 /* Free the device array corresponding to "array"
1192 static __isl_give isl_printer
*release_device_array(__isl_take isl_printer
*p
,
1193 struct gpu_array_info
*array
)
1195 p
= isl_printer_start_line(p
);
1196 p
= isl_printer_print_str(p
, "openclCheckReturn("
1197 "clReleaseMemObject(dev_");
1198 p
= isl_printer_print_str(p
, array
->name
);
1199 p
= isl_printer_print_str(p
, "));");
1200 p
= isl_printer_end_line(p
);
1205 /* Free the accessed device arrays.
1207 static __isl_give isl_printer
*opencl_release_device_arrays(
1208 __isl_take isl_printer
*p
, struct gpu_prog
*prog
)
1212 for (i
= 0; i
< prog
->n_array
; ++i
) {
1213 struct gpu_array_info
*array
= &prog
->array
[i
];
1214 if (gpu_array_is_read_only_scalar(array
))
1216 if (!array
->accessed
)
1219 p
= release_device_array(p
, array
);
1224 /* Given a gpu_prog "prog" and the corresponding transformed AST
1225 * "tree", print the entire OpenCL code to "p".
1227 static __isl_give isl_printer
*print_opencl(__isl_take isl_printer
*p
,
1228 struct gpu_prog
*prog
, __isl_keep isl_ast_node
*tree
,
1229 struct gpu_types
*types
, void *user
)
1231 struct opencl_info
*opencl
= user
;
1233 opencl
->kprinter
= isl_printer_set_output_format(opencl
->kprinter
,
1235 if (any_double_elements(prog
))
1236 opencl
->kprinter
= opencl_enable_double_support(
1238 if (opencl
->options
->opencl_print_kernel_types
)
1239 opencl
->kprinter
= gpu_print_types(opencl
->kprinter
, types
,
1242 if (!opencl
->kprinter
)
1243 return isl_printer_free(p
);
1245 p
= ppcg_start_block(p
);
1247 p
= opencl_print_host_macros(p
);
1249 p
= gpu_print_local_declarations(p
, prog
);
1250 p
= opencl_declare_device_arrays(p
, prog
);
1251 p
= opencl_setup(p
, opencl
->input
, opencl
);
1252 p
= opencl_allocate_device_arrays(p
, prog
);
1254 p
= opencl_print_host_code(p
, prog
, tree
, opencl
);
1256 p
= opencl_release_device_arrays(p
, prog
);
1257 p
= opencl_release_cl_objects(p
, opencl
);
1259 p
= ppcg_end_block(p
);
1264 /* Transform the code in the file called "input" by replacing
1265 * all scops by corresponding OpenCL code.
1266 * The host code is written to "output" or a name derived from
1267 * "input" if "output" is NULL.
1268 * The kernel code is placed in separate files with names
1269 * derived from "output" or "input".
1271 * We let generate_gpu do all the hard work and then let it call
1272 * us back for printing the AST in print_opencl.
1274 * To prepare for this printing, we first open the output files
1275 * and we close them after generate_gpu has finished.
1277 int generate_opencl(isl_ctx
*ctx
, struct ppcg_options
*options
,
1278 const char *input
, const char *output
)
1280 struct opencl_info opencl
= { options
, input
, output
};
1283 opencl
.kprinter
= isl_printer_to_str(ctx
);
1284 r
= opencl_open_files(&opencl
);
1287 r
= generate_gpu(ctx
, input
, opencl
.host_c
, options
,
1288 &print_opencl
, &opencl
);
1290 if (opencl_close_files(&opencl
) < 0)
1292 isl_printer_free(opencl
.kprinter
);