2 * Copyright 2013 Ecole Normale Superieure
4 * Use of this software is governed by the GNU LGPLv2.1 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"
24 #define min(a, b) (((a) < (b)) ? (a) : (b))
25 #define max(a, b) (((a) > (b)) ? (a) : (b))
27 /* options are the global options passed to generate_opencl.
28 * input is the name of the input file.
29 * output is the user-specified output file name and may be NULL
30 * if not specified by the user.
31 * kernel_c_name is the name of the kernel_c file.
32 * host_c is the generated source file for the host code. kernel_c is
33 * the generated source file for the kernel. kernel_h is the generated
34 * header file for the kernel.
37 struct ppcg_options
*options
;
40 char kernel_c_name
[PATH_MAX
];
47 /* Open the file called "name" for writing or print an error message.
49 static FILE *open_or_croak(const char *name
)
53 file
= fopen(name
, "w");
55 fprintf(stderr
, "Failed to open \"%s\" for writing\n", name
);
59 /* Open the host .c file and the kernel .h and .cl files for writing.
60 * Their names are derived from info->output (or info->input if
61 * the user did not specify an output file name).
62 * Add the necessary includes to these files.
64 * Return 0 on success and -1 on failure.
66 static int opencl_open_files(struct opencl_info
*info
)
74 ext
= strrchr(info
->output
, '.');
75 len
= ext
? ext
- info
->output
: strlen(info
->output
);
76 memcpy(name
, info
->output
, len
);
78 info
->host_c
= open_or_croak(info
->output
);
80 len
= ppcg_extract_base_name(name
, info
->input
);
82 strcpy(name
+ len
, "_host.c");
83 info
->host_c
= open_or_croak(name
);
86 memcpy(info
->kernel_c_name
, name
, len
);
87 strcpy(info
->kernel_c_name
+ len
, "_kernel.cl");
88 info
->kernel_c
= open_or_croak(info
->kernel_c_name
);
90 strcpy(name
+ len
, "_kernel.h");
91 info
->kernel_h
= open_or_croak(name
);
93 if (!info
->host_c
|| !info
->kernel_c
|| !info
->host_c
)
96 fprintf(info
->host_c
, "#include <assert.h>\n");
97 fprintf(info
->host_c
, "#include <stdio.h>\n");
98 fprintf(info
->host_c
, "#include \"%s\"\n\n", ppcg_base_name(name
));
99 fprintf(info
->kernel_h
, "#if defined(__APPLE__)\n");
100 fprintf(info
->kernel_h
, "#include <OpenCL/opencl.h>\n");
101 fprintf(info
->kernel_h
, "#else\n");
102 fprintf(info
->kernel_h
, "#include <CL/opencl.h>\n");
103 fprintf(info
->kernel_h
, "#endif\n\n");
104 fprintf(info
->kernel_h
, "cl_device_id opencl_create_device("
106 fprintf(info
->kernel_h
, "cl_program opencl_build_program("
108 "cl_device_id dev, const char *filename, "
109 "const char *opencl_options);\n");
110 fprintf(info
->kernel_h
,
111 "const char *opencl_error_string(cl_int error);\n");
116 /* Close all output files.
118 static void opencl_close_files(struct opencl_info
*info
)
121 fclose(info
->kernel_c
);
123 fclose(info
->kernel_h
);
125 fclose(info
->host_c
);
128 static __isl_give isl_printer
*opencl_print_host_macros(__isl_take isl_printer
*p
)
131 "#define openclCheckReturn(ret) \\\n"
132 " if (ret != CL_SUCCESS) {\\\n"
133 " fprintf(stderr, \"OpenCL error: %s\\n\", "
134 "opencl_error_string(ret)); \\\n"
135 " fflush(stderr); \\\n"
136 " assert(ret == CL_SUCCESS);\\\n }\n";
138 p
= isl_printer_start_line(p
);
139 p
= isl_printer_print_str(p
, macros
);
140 p
= isl_printer_end_line(p
);
142 p
= isl_ast_op_type_print_macro(isl_ast_op_max
, p
);
147 static __isl_give isl_printer
*opencl_declare_device_arrays(
148 __isl_take isl_printer
*p
, struct gpu_prog
*prog
)
152 for (i
= 0; i
< prog
->n_array
; ++i
) {
153 if (gpu_array_is_read_only_scalar(&prog
->array
[i
]))
155 p
= isl_printer_start_line(p
);
156 p
= isl_printer_print_str(p
, "cl_mem dev_");
157 p
= isl_printer_print_str(p
, prog
->array
[i
].name
);
158 p
= isl_printer_print_str(p
, ";");
159 p
= isl_printer_end_line(p
);
161 p
= isl_printer_start_line(p
);
162 p
= isl_printer_end_line(p
);
166 /* Given an array, check whether its positive size guard expression is
169 static int is_array_positive_size_guard_trivial(struct gpu_array_info
*array
)
174 guard
= gpu_array_positive_size_guard(array
);
175 is_trivial
= isl_set_plain_is_universe(guard
);
180 /* Allocate a device array for array and copy the contents to the device
183 * Emit a max-expression to ensure the device array can contain at least one
184 * element if the array's positive size guard expression is not trivial.
186 static __isl_give isl_printer
*allocate_device_array(__isl_take isl_printer
*p
,
187 struct gpu_array_info
*array
, int copy
)
189 int need_lower_bound
;
191 p
= ppcg_start_block(p
);
193 p
= isl_printer_start_line(p
);
194 p
= isl_printer_print_str(p
, "dev_");
195 p
= isl_printer_print_str(p
, array
->name
);
196 p
= isl_printer_print_str(p
, " = clCreateBuffer(context, ");
197 p
= isl_printer_print_str(p
, "CL_MEM_READ_WRITE");
200 p
= isl_printer_print_str(p
, ", ");
202 p
= isl_printer_print_str(p
, " | CL_MEM_COPY_HOST_PTR, ");
204 need_lower_bound
= !is_array_positive_size_guard_trivial(array
);
205 if (need_lower_bound
) {
206 p
= isl_printer_print_str(p
, "max(sizeof(");
207 p
= isl_printer_print_str(p
, array
->type
);
208 p
= isl_printer_print_str(p
, "), ");
210 p
= gpu_array_info_print_size(p
, array
);
211 if (need_lower_bound
)
212 p
= isl_printer_print_str(p
, ")");
215 p
= isl_printer_print_str(p
, ", NULL");
216 else if (gpu_array_is_scalar(array
)) {
217 p
= isl_printer_print_str(p
, ", &");
218 p
= isl_printer_print_str(p
, array
->name
);
220 p
= isl_printer_print_str(p
, ", ");
221 p
= isl_printer_print_str(p
, array
->name
);
224 p
= isl_printer_print_str(p
, ", &err);");
225 p
= isl_printer_end_line(p
);
226 p
= isl_printer_start_line(p
);
227 p
= isl_printer_print_str(p
, "openclCheckReturn(err);");
228 p
= isl_printer_end_line(p
);
230 p
= ppcg_end_block(p
);
235 /* Allocate device arrays and copy the contents of copy_in arrays into device.
237 static __isl_give isl_printer
*opencl_allocate_device_arrays(
238 __isl_take isl_printer
*p
, struct gpu_prog
*prog
)
242 for (i
= 0; i
< prog
->n_array
; ++i
) {
243 struct gpu_array_info
*array
= &prog
->array
[i
];
248 if (gpu_array_is_read_only_scalar(array
))
251 space
= isl_space_copy(array
->space
);
252 read_i
= isl_union_set_extract_set(prog
->copy_in
, space
);
253 empty
= isl_set_plain_is_empty(read_i
);
254 isl_set_free(read_i
);
256 p
= allocate_device_array(p
, array
, !empty
);
258 p
= isl_printer_start_line(p
);
259 p
= isl_printer_end_line(p
);
263 /* Print a call to the OpenCL clSetKernelArg() function which sets
264 * the arguments of the kernel. arg_name and arg_index are the name and the
265 * index of the kernel argument. The index of the leftmost argument of
266 * the kernel is 0 whereas the index of the rightmost argument of the kernel
267 * is n - 1, where n is the total number of the kernel arguments.
268 * read_only_scalar is a boolean that indicates whether the argument is a read
271 static __isl_give isl_printer
*opencl_set_kernel_argument(
272 __isl_take isl_printer
*p
, int kernel_id
,
273 const char *arg_name
, int arg_index
, int read_only_scalar
)
275 p
= isl_printer_start_line(p
);
276 p
= isl_printer_print_str(p
,
277 "openclCheckReturn(clSetKernelArg(kernel");
278 p
= isl_printer_print_int(p
, kernel_id
);
279 p
= isl_printer_print_str(p
, ", ");
280 p
= isl_printer_print_int(p
, arg_index
);
281 p
= isl_printer_print_str(p
, ", sizeof(");
283 if (read_only_scalar
) {
284 p
= isl_printer_print_str(p
, arg_name
);
285 p
= isl_printer_print_str(p
, "), &");
287 p
= isl_printer_print_str(p
, "cl_mem), (void *) &dev_");
289 p
= isl_printer_print_str(p
, arg_name
);
290 p
= isl_printer_print_str(p
, "));");
291 p
= isl_printer_end_line(p
);
296 /* Print the block sizes as a list of the sizes in each
299 static __isl_give isl_printer
*opencl_print_block_sizes(
300 __isl_take isl_printer
*p
, struct ppcg_kernel
*kernel
)
304 if (kernel
->n_block
> 0)
305 for (i
= 0; i
< kernel
->n_block
; ++i
) {
307 p
= isl_printer_print_str(p
, ", ");
308 p
= isl_printer_print_int(p
, kernel
->block_dim
[i
]);
311 p
= isl_printer_print_str(p
, "1");
316 /* Set the arguments of the OpenCL kernel by printing a call to the OpenCL
317 * clSetKernelArg() function for each kernel argument.
319 static __isl_give isl_printer
*opencl_set_kernel_arguments(
320 __isl_take isl_printer
*p
, struct gpu_prog
*prog
,
321 struct ppcg_kernel
*kernel
)
329 for (i
= 0; i
< prog
->n_array
; ++i
) {
333 space
= isl_space_copy(prog
->array
[i
].space
);
334 arr
= isl_union_set_extract_set(kernel
->arrays
, space
);
335 empty
= isl_set_plain_is_empty(arr
);
339 ro
= gpu_array_is_read_only_scalar(&prog
->array
[i
]);
340 opencl_set_kernel_argument(p
, kernel
->id
, prog
->array
[i
].name
,
345 space
= isl_union_set_get_space(kernel
->arrays
);
346 nparam
= isl_space_dim(space
, isl_dim_param
);
347 for (i
= 0; i
< nparam
; ++i
) {
350 name
= isl_space_get_dim_name(space
, isl_dim_param
, i
);
351 opencl_set_kernel_argument(p
, kernel
->id
, name
, arg_index
, 1);
354 isl_space_free(space
);
356 n
= isl_space_dim(kernel
->space
, isl_dim_set
);
357 for (i
= 0; i
< n
; ++i
) {
361 name
= isl_space_get_dim_name(kernel
->space
, isl_dim_set
, i
);
362 opencl_set_kernel_argument(p
, kernel
->id
, name
, arg_index
, 1);
369 /* Print the arguments to a kernel declaration or call. If "types" is set,
370 * then print a declaration (including the types of the arguments).
372 * The arguments are printed in the following order
373 * - the arrays accessed by the kernel
375 * - the host loop iterators
377 static __isl_give isl_printer
*opencl_print_kernel_arguments(
378 __isl_take isl_printer
*p
, struct gpu_prog
*prog
,
379 struct ppcg_kernel
*kernel
, int types
)
387 for (i
= 0; i
< prog
->n_array
; ++i
) {
391 space
= isl_space_copy(prog
->array
[i
].space
);
392 arr
= isl_union_set_extract_set(kernel
->arrays
, space
);
393 empty
= isl_set_plain_is_empty(arr
);
399 p
= isl_printer_print_str(p
, ", ");
402 p
= gpu_array_info_print_declaration_argument(p
,
403 &prog
->array
[i
], "__global");
405 p
= gpu_array_info_print_call_argument(p
,
411 space
= isl_union_set_get_space(kernel
->arrays
);
412 nparam
= isl_space_dim(space
, isl_dim_param
);
413 for (i
= 0; i
< nparam
; ++i
) {
416 name
= isl_space_get_dim_name(space
, isl_dim_param
, i
);
419 p
= isl_printer_print_str(p
, ", ");
421 p
= isl_printer_print_str(p
, "int ");
422 p
= isl_printer_print_str(p
, name
);
426 isl_space_free(space
);
428 n
= isl_space_dim(kernel
->space
, isl_dim_set
);
429 type
= isl_options_get_ast_iterator_type(prog
->ctx
);
430 for (i
= 0; i
< n
; ++i
) {
435 p
= isl_printer_print_str(p
, ", ");
436 name
= isl_space_get_dim_name(kernel
->space
, isl_dim_set
, i
);
438 p
= isl_printer_print_str(p
, type
);
439 p
= isl_printer_print_str(p
, " ");
441 p
= isl_printer_print_str(p
, name
);
449 /* Print the header of the given kernel.
451 static __isl_give isl_printer
*opencl_print_kernel_header(
452 __isl_take isl_printer
*p
, struct gpu_prog
*prog
,
453 struct ppcg_kernel
*kernel
)
455 p
= isl_printer_start_line(p
);
456 p
= isl_printer_print_str(p
, "__kernel void kernel");
457 p
= isl_printer_print_int(p
, kernel
->id
);
458 p
= isl_printer_print_str(p
, "(");
459 p
= opencl_print_kernel_arguments(p
, prog
, kernel
, 1);
460 p
= isl_printer_print_str(p
, ")");
461 p
= isl_printer_end_line(p
);
466 /* Unlike the equivalent function in the CUDA backend which prints iterators
467 * in reverse order to promote coalescing, this function does not print
468 * iterators in reverse order. The OpenCL backend currently does not take
469 * into account any coalescing considerations.
471 static __isl_give isl_printer
*opencl_print_kernel_iterators(
472 __isl_take isl_printer
*p
, struct ppcg_kernel
*kernel
)
475 isl_ctx
*ctx
= isl_ast_node_get_ctx(kernel
->tree
);
478 type
= isl_options_get_ast_iterator_type(ctx
);
480 n_grid
= isl_multi_pw_aff_dim(kernel
->grid_size
, isl_dim_set
);
482 p
= isl_printer_start_line(p
);
483 p
= isl_printer_print_str(p
, type
);
484 p
= isl_printer_print_str(p
, " ");
485 for (i
= 0; i
< n_grid
; ++i
) {
487 p
= isl_printer_print_str(p
, ", ");
488 p
= isl_printer_print_str(p
, "b");
489 p
= isl_printer_print_int(p
, i
);
490 p
= isl_printer_print_str(p
, " = get_group_id(");
491 p
= isl_printer_print_int(p
, i
);
492 p
= isl_printer_print_str(p
, ")");
494 p
= isl_printer_print_str(p
, ";");
495 p
= isl_printer_end_line(p
);
498 if (kernel
->n_block
> 0) {
499 p
= isl_printer_start_line(p
);
500 p
= isl_printer_print_str(p
, type
);
501 p
= isl_printer_print_str(p
, " ");
502 for (i
= 0; i
< kernel
->n_block
; ++i
) {
504 p
= isl_printer_print_str(p
, ", ");
505 p
= isl_printer_print_str(p
, "t");
506 p
= isl_printer_print_int(p
, i
);
507 p
= isl_printer_print_str(p
, " = get_local_id(");
508 p
= isl_printer_print_int(p
, i
);
509 p
= isl_printer_print_str(p
, ")");
511 p
= isl_printer_print_str(p
, ";");
512 p
= isl_printer_end_line(p
);
518 static __isl_give isl_printer
*opencl_print_kernel_var(
519 __isl_take isl_printer
*p
, struct ppcg_kernel_var
*var
)
524 p
= isl_printer_start_line(p
);
525 if (var
->type
== ppcg_access_shared
)
526 p
= isl_printer_print_str(p
, "__local ");
527 p
= isl_printer_print_str(p
, var
->array
->type
);
528 p
= isl_printer_print_str(p
, " ");
529 p
= isl_printer_print_str(p
, var
->name
);
530 for (j
= 0; j
< var
->array
->n_index
; ++j
) {
531 p
= isl_printer_print_str(p
, "[");
532 v
= isl_vec_get_element_val(var
->size
, j
);
533 p
= isl_printer_print_val(p
, v
);
534 p
= isl_printer_print_str(p
, "]");
537 p
= isl_printer_print_str(p
, ";");
538 p
= isl_printer_end_line(p
);
543 static __isl_give isl_printer
*opencl_print_kernel_vars(
544 __isl_take isl_printer
*p
, struct ppcg_kernel
*kernel
)
548 for (i
= 0; i
< kernel
->n_var
; ++i
)
549 p
= opencl_print_kernel_var(p
, &kernel
->var
[i
]);
554 /* Print a call to barrier() which is a sync statement.
555 * All work-items in a work-group executing the kernel on a processor must
556 * execute the barrier() function before any are allowed to continue execution
557 * beyond the barrier.
558 * The flag CLK_LOCAL_MEM_FENCE makes the barrier function either flush any
559 * variables stored in local memory or queue a memory fence to ensure correct
560 * ordering of memory operations to local memory.
561 * The flag CLK_GLOBAL_MEM_FENCE makes the barrier function queue a memory
562 * fence to ensure correct ordering of memory operations to global memory.
564 static __isl_give isl_printer
*opencl_print_sync(__isl_take isl_printer
*p
,
565 struct ppcg_kernel_stmt
*stmt
)
567 p
= isl_printer_start_line(p
);
568 p
= isl_printer_print_str(p
,
569 "barrier(CLK_LOCAL_MEM_FENCE | CLK_GLOBAL_MEM_FENCE);");
570 p
= isl_printer_end_line(p
);
575 /* This function is called for each user statement in the AST,
576 * i.e., for each kernel body statement, copy statement or sync statement.
578 static __isl_give isl_printer
*opencl_print_kernel_stmt(
579 __isl_take isl_printer
*p
,
580 __isl_take isl_ast_print_options
*print_options
,
581 __isl_keep isl_ast_node
*node
, void *user
)
584 struct ppcg_kernel_stmt
*stmt
;
586 id
= isl_ast_node_get_annotation(node
);
587 stmt
= isl_id_get_user(id
);
590 isl_ast_print_options_free(print_options
);
592 switch (stmt
->type
) {
593 case ppcg_kernel_copy
:
594 return ppcg_kernel_print_copy(p
, stmt
);
595 case ppcg_kernel_sync
:
596 return opencl_print_sync(p
, stmt
);
597 case ppcg_kernel_domain
:
598 return ppcg_kernel_print_domain(p
, stmt
);
604 /* Return true if there is a double array in prog->array or
605 * if any of the types in prog->scop involve any doubles.
606 * To check the latter condition, we simply search for the string "double"
607 * in the type definitions, which may result in false positives.
609 static __isl_give
int any_double_elements(struct gpu_prog
*prog
)
613 for (i
= 0; i
< prog
->n_array
; ++i
)
614 if (strcmp(prog
->array
[i
].type
, "double") == 0)
617 for (i
= 0; i
< prog
->scop
->n_type
; ++i
) {
618 struct pet_type
*type
= prog
->scop
->types
[i
];
620 if (strstr(type
->definition
, "double"))
627 /* Prints a #pragma to enable support for double floating-point
628 * precision. OpenCL 1.0 adds support for double precision floating-point as
629 * an optional extension. An application that wants to use double will need to
630 * include the #pragma OPENCL EXTENSION cl_khr_fp64 : enable directive before
631 * any double precision data type is declared in the kernel code.
633 static __isl_give isl_printer
*opencl_enable_double_support(
634 __isl_take isl_printer
*p
)
638 p
= isl_printer_start_line(p
);
639 p
= isl_printer_print_str(p
, "#pragma OPENCL EXTENSION cl_khr_fp64 :"
641 p
= isl_printer_end_line(p
);
642 p
= isl_printer_start_line(p
);
643 p
= isl_printer_end_line(p
);
648 static void opencl_print_kernel(struct gpu_prog
*prog
,
649 struct ppcg_kernel
*kernel
, struct opencl_info
*opencl
)
651 isl_ctx
*ctx
= isl_ast_node_get_ctx(kernel
->tree
);
652 isl_ast_print_options
*print_options
;
655 p
= isl_printer_to_file(ctx
, opencl
->kernel_c
);
656 print_options
= isl_ast_print_options_alloc(ctx
);
657 print_options
= isl_ast_print_options_set_print_user(print_options
,
658 &opencl_print_kernel_stmt
, NULL
);
660 p
= isl_printer_set_output_format(p
, ISL_FORMAT_C
);
661 p
= opencl_print_kernel_header(p
, prog
, kernel
);
662 p
= isl_printer_print_str(p
, "{");
663 p
= isl_printer_end_line(p
);
664 p
= isl_printer_indent(p
, 4);
665 p
= opencl_print_kernel_iterators(p
, kernel
);
666 p
= opencl_print_kernel_vars(p
, kernel
);
667 p
= isl_printer_end_line(p
);
668 p
= gpu_print_macros(p
, kernel
->tree
);
669 p
= isl_ast_node_print(kernel
->tree
, p
, print_options
);
670 p
= isl_printer_print_str(p
, "}");
671 p
= isl_printer_end_line(p
);
675 struct print_host_user_data_opencl
{
676 struct opencl_info
*opencl
;
677 struct gpu_prog
*prog
;
680 /* This function prints the i'th block size multiplied by the i'th grid size,
681 * where i (a parameter to this function) is one of the possible dimensions of
682 * grid sizes and block sizes.
683 * If the dimension of block sizes is not equal to the dimension of grid sizes
684 * the output is calculated as follows:
687 * block_sizes[dim1] is the list of blocks sizes and it contains dim1 elements.
688 * grid_sizes[dim2] is the list of grid sizes and it contains dim2 elements.
691 * If (i > dim2) then the output is block_sizes[i]
692 * If (i > dim1) then the output is grid_sizes[i]
694 static __isl_give isl_printer
*opencl_print_total_number_of_work_items_for_dim(
695 __isl_take isl_printer
*p
, struct ppcg_kernel
*kernel
, int i
)
697 int grid_dim
, block_dim
;
699 grid_dim
= isl_multi_pw_aff_dim(kernel
->grid_size
, isl_dim_set
);
700 block_dim
= kernel
->n_block
;
702 isl_pw_aff
*bound_grid
;
704 if (i
< min(grid_dim
, block_dim
)) {
705 bound_grid
= isl_multi_pw_aff_get_pw_aff(kernel
->grid_size
, i
);
706 p
= isl_printer_print_str(p
, "(");
707 p
= isl_printer_print_pw_aff(p
, bound_grid
);
708 p
= isl_printer_print_str(p
, ") * ");
709 p
= isl_printer_print_int(p
, kernel
->block_dim
[i
]);
710 isl_pw_aff_free(bound_grid
);
711 } else if (i
>= grid_dim
)
712 p
= isl_printer_print_int(p
, kernel
->block_dim
[i
]);
714 bound_grid
= isl_multi_pw_aff_get_pw_aff(kernel
->grid_size
, i
);
715 p
= isl_printer_print_pw_aff(p
, bound_grid
);
716 isl_pw_aff_free(bound_grid
);
722 /* Print a list that represents the total number of work items. The list is
723 * constructed by performing an element-wise multiplication of the block sizes
724 * and the grid sizes. To explain how the list is constructed, suppose that:
725 * block_sizes[dim1] is the list of blocks sizes and it contains dim1 elements.
726 * grid_sizes[dim2] is the list of grid sizes and it contains dim2 elements.
728 * The output of this function is constructed as follows:
729 * If (dim1 > dim2) then the output is the following list:
730 * grid_sizes[0]*block_sizes[0], ..., grid_sizes[dim2-1]*block_sizes[dim2-1],
731 * block_sizes[dim2], ..., block_sizes[dim1-2], block_sizes[dim1-1].
733 * If (dim2 > dim1) then the output is the following list:
734 * grid_sizes[0]*block_sizes[0], ..., grid_sizes[dim1-1] * block_sizes[dim1-1],
735 * grid_sizes[dim1], grid_sizes[dim2-2], grid_sizes[dim2-1].
737 * To calculate the total number of work items out of the list constructed by
738 * this function, the user should multiply the elements of the list.
740 static __isl_give isl_printer
*opencl_print_total_number_of_work_items_as_list(
741 __isl_take isl_printer
*p
, struct ppcg_kernel
*kernel
)
744 int grid_dim
, block_dim
;
746 grid_dim
= isl_multi_pw_aff_dim(kernel
->grid_size
, isl_dim_set
);
747 block_dim
= kernel
->n_block
;
749 if ((grid_dim
<= 0) || (block_dim
<= 0)) {
750 p
= isl_printer_print_str(p
, "1");
754 for (i
= 0; i
<= max(grid_dim
, block_dim
) - 1; i
++) {
756 p
= isl_printer_print_str(p
, ", ");
758 p
= opencl_print_total_number_of_work_items_for_dim(p
,
765 /* Print the user statement of the host code to "p".
767 * In particular, print a block of statements that defines the grid
768 * and the work group and then launches the kernel.
770 * A grid is composed of many work groups (blocks), each work group holds
771 * many work-items (threads).
773 * global_work_size[kernel->n_block] represents the total number of work
774 * items. It points to an array of kernel->n_block unsigned
775 * values that describe the total number of work-items that will execute
776 * the kernel. The total number of work-items is computed as:
777 * global_work_size[0] *...* global_work_size[kernel->n_block - 1].
779 * The size of each work group (i.e. the number of work-items in each work
780 * group) is described using block_size[kernel->n_block]. The total
781 * number of work-items in a block (work-group) is computed as:
782 * block_size[0] *... * block_size[kernel->n_block - 1].
784 * For more information check:
785 * http://www.khronos.org/registry/cl/sdk/1.0/docs/man/xhtml/clEnqueueNDRangeKernel.html
787 static __isl_give isl_printer
*opencl_print_host_user(
788 __isl_take isl_printer
*p
,
789 __isl_take isl_ast_print_options
*print_options
,
790 __isl_keep isl_ast_node
*node
, void *user
)
793 struct ppcg_kernel
*kernel
;
794 struct print_host_user_data_opencl
*data
;
797 id
= isl_ast_node_get_annotation(node
);
798 kernel
= isl_id_get_user(id
);
801 data
= (struct print_host_user_data_opencl
*) user
;
803 p
= isl_printer_start_line(p
);
804 p
= isl_printer_print_str(p
, "{");
805 p
= isl_printer_end_line(p
);
806 p
= isl_printer_indent(p
, 2);
808 p
= isl_printer_start_line(p
);
809 p
= isl_printer_print_str(p
, "size_t global_work_size[");
811 if (kernel
->n_block
> 0)
812 p
= isl_printer_print_int(p
, kernel
->n_block
);
814 p
= isl_printer_print_int(p
, 1);
816 p
= isl_printer_print_str(p
, "] = {");
817 p
= opencl_print_total_number_of_work_items_as_list(p
, kernel
);
818 p
= isl_printer_print_str(p
, "};");
819 p
= isl_printer_end_line(p
);
821 p
= isl_printer_start_line(p
);
822 p
= isl_printer_print_str(p
, "size_t block_size[");
824 if (kernel
->n_block
> 0)
825 p
= isl_printer_print_int(p
, kernel
->n_block
);
827 p
= isl_printer_print_int(p
, 1);
829 p
= isl_printer_print_str(p
, "] = {");
830 p
= opencl_print_block_sizes(p
, kernel
);
831 p
= isl_printer_print_str(p
, "};");
832 p
= isl_printer_end_line(p
);
834 p
= isl_printer_start_line(p
);
835 p
= isl_printer_print_str(p
, "cl_kernel kernel");
836 p
= isl_printer_print_int(p
, kernel
->id
);
837 p
= isl_printer_print_str(p
, " = clCreateKernel(program, \"kernel");
838 p
= isl_printer_print_int(p
, kernel
->id
);
839 p
= isl_printer_print_str(p
, "\", &err);");
840 p
= isl_printer_end_line(p
);
841 p
= isl_printer_start_line(p
);
842 p
= isl_printer_print_str(p
, "openclCheckReturn(err);");
843 p
= isl_printer_end_line(p
);
845 opencl_set_kernel_arguments(p
, data
->prog
, kernel
);
847 p
= isl_printer_start_line(p
);
848 p
= isl_printer_print_str(p
, "openclCheckReturn(clEnqueueNDRangeKernel"
850 p
= isl_printer_print_int(p
, kernel
->id
);
851 p
= isl_printer_print_str(p
, ", ");
852 if (kernel
->n_block
> 0)
853 p
= isl_printer_print_int(p
, kernel
->n_block
);
855 p
= isl_printer_print_int(p
, 1);
857 p
= isl_printer_print_str(p
, ", NULL, global_work_size, "
860 p
= isl_printer_end_line(p
);
861 p
= isl_printer_start_line(p
);
862 p
= isl_printer_print_str(p
, "openclCheckReturn("
863 "clReleaseKernel(kernel");
864 p
= isl_printer_print_int(p
, kernel
->id
);
865 p
= isl_printer_print_str(p
, "));");
866 p
= isl_printer_end_line(p
);
867 p
= isl_printer_start_line(p
);
868 p
= isl_printer_print_str(p
, "clFinish(queue);");
869 p
= isl_printer_end_line(p
);
870 p
= isl_printer_indent(p
, -2);
871 p
= isl_printer_start_line(p
);
872 p
= isl_printer_print_str(p
, "}");
873 p
= isl_printer_end_line(p
);
875 p
= isl_printer_start_line(p
);
876 p
= isl_printer_end_line(p
);
878 opencl_print_kernel(data
->prog
, kernel
, data
->opencl
);
880 isl_ast_print_options_free(print_options
);
885 static __isl_give isl_printer
*opencl_print_host_code(
886 __isl_take isl_printer
*p
, struct gpu_prog
*prog
,
887 __isl_keep isl_ast_node
*tree
, struct opencl_info
*opencl
)
889 isl_ast_print_options
*print_options
;
890 isl_ctx
*ctx
= isl_ast_node_get_ctx(tree
);
891 struct print_host_user_data_opencl data
= { opencl
, prog
};
893 print_options
= isl_ast_print_options_alloc(ctx
);
894 print_options
= isl_ast_print_options_set_print_user(print_options
,
895 &opencl_print_host_user
, &data
);
897 p
= gpu_print_macros(p
, tree
);
898 p
= isl_ast_node_print(tree
, p
, print_options
);
903 /* Copy "array" back from the GPU to the host.
905 static __isl_give isl_printer
*copy_array_from_device(__isl_take isl_printer
*p
,
908 struct gpu_array_info
*array
= user
;
910 p
= isl_printer_start_line(p
);
911 p
= isl_printer_print_str(p
, "openclCheckReturn("
912 "clEnqueueReadBuffer(queue,"
914 p
= isl_printer_print_str(p
, array
->name
);
915 p
= isl_printer_print_str(p
, ", CL_TRUE, 0, ");
916 p
= gpu_array_info_print_size(p
, array
);
918 if (gpu_array_is_scalar(array
))
919 p
= isl_printer_print_str(p
, ", &");
921 p
= isl_printer_print_str(p
, ", ");
922 p
= isl_printer_print_str(p
, array
->name
);
923 p
= isl_printer_print_str(p
, ", 0, NULL, NULL));");
924 p
= isl_printer_end_line(p
);
929 /* Copy copy_out arrays back from the GPU to the host.
931 * Only perform the copying for arrays with strictly positive size.
933 static __isl_give isl_printer
*opencl_copy_arrays_from_device(
934 __isl_take isl_printer
*p
, struct gpu_prog
*prog
)
937 isl_union_set
*copy_out
;
938 copy_out
= isl_union_set_copy(prog
->copy_out
);
940 for (i
= 0; i
< prog
->n_array
; ++i
) {
941 struct gpu_array_info
*array
= &prog
->array
[i
];
947 space
= isl_space_copy(array
->space
);
948 copy_out_i
= isl_union_set_extract_set(copy_out
, space
);
949 empty
= isl_set_plain_is_empty(copy_out_i
);
950 isl_set_free(copy_out_i
);
954 guard
= gpu_array_positive_size_guard(array
);
955 p
= ppcg_print_guarded(p
, guard
, isl_set_copy(prog
->context
),
956 ©_array_from_device
, array
);
959 isl_union_set_free(copy_out
);
960 p
= isl_printer_start_line(p
);
961 p
= isl_printer_end_line(p
);
965 /* Create an OpenCL device, context, command queue and build the kernel.
966 * input is the name of the input file provided to ppcg.
968 static __isl_give isl_printer
*opencl_setup(__isl_take isl_printer
*p
,
969 const char *input
, struct opencl_info
*info
)
973 p
= isl_printer_start_line(p
);
974 p
= isl_printer_print_str(p
, "cl_device_id device;");
975 p
= isl_printer_end_line(p
);
976 p
= isl_printer_start_line(p
);
977 p
= isl_printer_print_str(p
, "cl_context context;");
978 p
= isl_printer_end_line(p
);
979 p
= isl_printer_start_line(p
);
980 p
= isl_printer_print_str(p
, "cl_program program;");
981 p
= isl_printer_end_line(p
);
982 p
= isl_printer_start_line(p
);
983 p
= isl_printer_print_str(p
, "cl_command_queue queue;");
984 p
= isl_printer_end_line(p
);
985 p
= isl_printer_start_line(p
);
986 p
= isl_printer_print_str(p
, "cl_int err;");
987 p
= isl_printer_end_line(p
);
988 p
= isl_printer_start_line(p
);
989 p
= isl_printer_print_str(p
, "device = opencl_create_device(");
990 p
= isl_printer_print_int(p
, info
->options
->opencl_use_gpu
);
991 p
= isl_printer_print_str(p
, ");");
992 p
= isl_printer_end_line(p
);
993 p
= isl_printer_start_line(p
);
994 p
= isl_printer_print_str(p
, "context = clCreateContext(NULL, 1, "
995 "&device, NULL, NULL, &err);");
996 p
= isl_printer_end_line(p
);
997 p
= isl_printer_start_line(p
);
998 p
= isl_printer_print_str(p
, "openclCheckReturn(err);");
999 p
= isl_printer_end_line(p
);
1000 p
= isl_printer_start_line(p
);
1001 p
= isl_printer_print_str(p
, "queue = clCreateCommandQueue"
1002 "(context, device, 0, &err);");
1003 p
= isl_printer_end_line(p
);
1004 p
= isl_printer_start_line(p
);
1005 p
= isl_printer_print_str(p
, "openclCheckReturn(err);");
1006 p
= isl_printer_end_line(p
);
1008 p
= isl_printer_start_line(p
);
1009 p
= isl_printer_print_str(p
, "program = opencl_build_program("
1010 "context, device, \"");
1011 p
= isl_printer_print_str(p
, info
->kernel_c_name
);
1012 p
= isl_printer_print_str(p
, "\", \"");
1014 if (info
->options
->opencl_compiler_options
)
1015 p
= isl_printer_print_str(p
,
1016 info
->options
->opencl_compiler_options
);
1018 p
= isl_printer_print_str(p
, "\");");
1019 p
= isl_printer_end_line(p
);
1020 p
= isl_printer_start_line(p
);
1021 p
= isl_printer_end_line(p
);
1026 static __isl_give isl_printer
*opencl_release_cl_objects(
1027 __isl_take isl_printer
*p
, struct opencl_info
*info
)
1029 p
= isl_printer_start_line(p
);
1030 p
= isl_printer_print_str(p
, "openclCheckReturn(clReleaseCommandQueue"
1032 p
= isl_printer_end_line(p
);
1033 p
= isl_printer_start_line(p
);
1034 p
= isl_printer_print_str(p
, "openclCheckReturn(clReleaseProgram"
1036 p
= isl_printer_end_line(p
);
1037 p
= isl_printer_start_line(p
);
1038 p
= isl_printer_print_str(p
, "openclCheckReturn(clReleaseContext"
1040 p
= isl_printer_end_line(p
);
1045 /* Free the device array corresponding to "array"
1047 static __isl_give isl_printer
*release_device_array(__isl_take isl_printer
*p
,
1048 struct gpu_array_info
*array
)
1050 p
= isl_printer_start_line(p
);
1051 p
= isl_printer_print_str(p
, "openclCheckReturn("
1052 "clReleaseMemObject(dev_");
1053 p
= isl_printer_print_str(p
, array
->name
);
1054 p
= isl_printer_print_str(p
, "));");
1055 p
= isl_printer_end_line(p
);
1060 /* Free the device arrays.
1062 static __isl_give isl_printer
*opencl_release_device_arrays(
1063 __isl_take isl_printer
*p
, struct gpu_prog
*prog
)
1067 for (i
= 0; i
< prog
->n_array
; ++i
) {
1068 struct gpu_array_info
*array
= &prog
->array
[i
];
1069 if (gpu_array_is_read_only_scalar(array
))
1072 p
= release_device_array(p
, array
);
1077 /* Given a gpu_prog "prog" and the corresponding transformed AST
1078 * "tree", print the entire OpenCL code to "p".
1080 static __isl_give isl_printer
*print_opencl(__isl_take isl_printer
*p
,
1081 struct gpu_prog
*prog
, __isl_keep isl_ast_node
*tree
,
1082 struct gpu_types
*types
, void *user
)
1084 struct opencl_info
*opencl
= user
;
1085 isl_printer
*kernel
;
1087 kernel
= isl_printer_to_file(isl_printer_get_ctx(p
), opencl
->kernel_c
);
1088 kernel
= isl_printer_set_output_format(kernel
, ISL_FORMAT_C
);
1089 if (any_double_elements(prog
))
1090 kernel
= opencl_enable_double_support(kernel
);
1091 kernel
= gpu_print_types(kernel
, types
, prog
);
1092 isl_printer_free(kernel
);
1095 return isl_printer_free(p
);
1097 p
= ppcg_start_block(p
);
1099 p
= opencl_print_host_macros(p
);
1101 p
= opencl_declare_device_arrays(p
, prog
);
1102 p
= opencl_setup(p
, opencl
->input
, opencl
);
1103 p
= opencl_allocate_device_arrays(p
, prog
);
1105 p
= opencl_print_host_code(p
, prog
, tree
, opencl
);
1107 p
= opencl_copy_arrays_from_device(p
, prog
);
1108 p
= opencl_release_device_arrays(p
, prog
);
1109 p
= opencl_release_cl_objects(p
, opencl
);
1111 p
= ppcg_end_block(p
);
1116 /* Transform the code in the file called "input" by replacing
1117 * all scops by corresponding OpenCL code.
1118 * The host code is written to "output" or a name derived from
1119 * "input" if "output" is NULL.
1120 * The kernel code is placed in separate files with names
1121 * derived from "output" or "input".
1123 * We let generate_gpu do all the hard work and then let it call
1124 * us back for printing the AST in print_cuda.
1126 * To prepare for this printing, we first open the output files
1127 * and we close them after generate_gpu has finished.
1129 int generate_opencl(isl_ctx
*ctx
, struct ppcg_options
*options
,
1130 const char *input
, const char *output
)
1132 struct opencl_info opencl
= { options
, input
, output
};
1135 r
= opencl_open_files(&opencl
);
1138 r
= generate_gpu(ctx
, input
, opencl
.host_c
, options
,
1139 &print_opencl
, &opencl
);
1141 opencl_close_files(&opencl
);