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 * kprinter is an isl_printer for the kernel file.
33 * host_c is the generated source file for the host code. kernel_c is
34 * the generated source file for the kernel. kernel_h is the generated
35 * header file for the kernel.
38 struct ppcg_options
*options
;
41 char kernel_c_name
[PATH_MAX
];
43 isl_printer
*kprinter
;
50 /* Open the file called "name" for writing or print an error message.
52 static FILE *open_or_croak(const char *name
)
56 file
= fopen(name
, "w");
58 fprintf(stderr
, "Failed to open \"%s\" for writing\n", name
);
62 /* Open the host .c file and the kernel .h and .cl files for writing.
63 * Their names are derived from info->output (or info->input if
64 * the user did not specify an output file name).
65 * Add the necessary includes to these files, including those specified
68 * Return 0 on success and -1 on failure.
70 static int opencl_open_files(struct opencl_info
*info
)
79 ext
= strrchr(info
->output
, '.');
80 len
= ext
? ext
- info
->output
: strlen(info
->output
);
81 memcpy(name
, info
->output
, len
);
83 info
->host_c
= open_or_croak(info
->output
);
85 len
= ppcg_extract_base_name(name
, info
->input
);
87 strcpy(name
+ len
, "_host.c");
88 info
->host_c
= open_or_croak(name
);
91 memcpy(info
->kernel_c_name
, name
, len
);
92 strcpy(info
->kernel_c_name
+ len
, "_kernel.cl");
93 info
->kernel_c
= open_or_croak(info
->kernel_c_name
);
95 strcpy(name
+ len
, "_kernel.h");
96 info
->kernel_h
= open_or_croak(name
);
98 if (!info
->host_c
|| !info
->kernel_c
|| !info
->host_c
)
101 fprintf(info
->host_c
, "#include <assert.h>\n");
102 fprintf(info
->host_c
, "#include <stdio.h>\n");
103 fprintf(info
->host_c
, "#include \"%s\"\n\n", ppcg_base_name(name
));
104 if (info
->options
->opencl_embed_kernel_code
) {
105 fprintf(info
->host_c
, "#include \"%s\"\n\n",
106 info
->kernel_c_name
);
109 fprintf(info
->kernel_h
, "#if defined(__APPLE__)\n");
110 fprintf(info
->kernel_h
, "#include <OpenCL/opencl.h>\n");
111 fprintf(info
->kernel_h
, "#else\n");
112 fprintf(info
->kernel_h
, "#include <CL/opencl.h>\n");
113 fprintf(info
->kernel_h
, "#endif\n\n");
114 fprintf(info
->kernel_h
, "cl_device_id opencl_create_device("
116 fprintf(info
->kernel_h
, "cl_program opencl_build_program_from_string("
118 "cl_device_id dev, const char *program_source, "
119 "size_t program_size, "
120 "const char *opencl_options);\n");
121 fprintf(info
->kernel_h
, "cl_program opencl_build_program_from_file("
123 "cl_device_id dev, const char *filename, "
124 "const char *opencl_options);\n");
125 fprintf(info
->kernel_h
,
126 "const char *opencl_error_string(cl_int error);\n");
127 for (i
= 0; i
< info
->options
->opencl_n_include_file
; ++i
) {
128 info
->kprinter
= isl_printer_print_str(info
->kprinter
,
130 info
->kprinter
= isl_printer_print_str(info
->kprinter
,
131 info
->options
->opencl_include_files
[i
]);
132 info
->kprinter
= isl_printer_print_str(info
->kprinter
, ">\n");
138 /* Write text to a file and escape some special characters that would break a
141 static void opencl_print_escaped(const char *str
, const char *end
, FILE *file
)
143 const char *prev
= str
;
145 while ((str
= strpbrk(prev
, "\"\\")) && str
< end
) {
146 fwrite(prev
, 1, str
- prev
, file
);
147 fprintf(file
, "\\%c", *str
);
152 fwrite(prev
, 1, end
- prev
, file
);
155 /* Write text to a file as a C string literal.
157 * This function also prints any characters after the last newline, although
158 * normally the input string should end with a newline.
160 static void opencl_print_as_c_string(const char *str
, FILE *file
)
162 const char *prev
= str
;
164 while ((str
= strchr(prev
, '\n'))) {
165 fprintf(file
, "\n\"");
166 opencl_print_escaped(prev
, str
, file
);
167 fprintf(file
, "\\n\"");
173 fprintf(file
, "\n\"");
174 opencl_print_escaped(prev
, prev
+ strlen(prev
), file
);
179 /* Write the code that we have accumulated in the kernel isl_printer to the
180 * kernel.cl file. If the opencl_embed_kernel_code option has been set, print
181 * the code as a C string literal. Start that string literal with an empty
182 * line, such that line numbers reported by the OpenCL C compiler match those
183 * of the kernel file.
185 static void opencl_write_kernel_file(struct opencl_info
*opencl
)
187 char *raw
= isl_printer_get_str(opencl
->kprinter
);
189 if (opencl
->options
->opencl_embed_kernel_code
) {
190 fprintf(opencl
->kernel_c
,
191 "static const char kernel_code[] = \"\\n\"");
192 opencl_print_as_c_string(raw
, opencl
->kernel_c
);
193 fprintf(opencl
->kernel_c
, ";\n");
195 fprintf(opencl
->kernel_c
, "%s", raw
);
200 /* Close all output files. Write the kernel contents to the kernel file before
203 static void opencl_close_files(struct opencl_info
*info
)
205 if (info
->kernel_c
) {
206 opencl_write_kernel_file(info
);
207 fclose(info
->kernel_c
);
210 fclose(info
->kernel_h
);
212 fclose(info
->host_c
);
215 static __isl_give isl_printer
*opencl_print_host_macros(
216 __isl_take isl_printer
*p
)
219 "#define openclCheckReturn(ret) \\\n"
220 " if (ret != CL_SUCCESS) {\\\n"
221 " fprintf(stderr, \"OpenCL error: %s\\n\", "
222 "opencl_error_string(ret)); \\\n"
223 " fflush(stderr); \\\n"
224 " assert(ret == CL_SUCCESS);\\\n }\n";
226 p
= isl_printer_start_line(p
);
227 p
= isl_printer_print_str(p
, macros
);
228 p
= isl_printer_end_line(p
);
230 p
= isl_ast_op_type_print_macro(isl_ast_op_max
, p
);
235 static __isl_give isl_printer
*opencl_declare_device_arrays(
236 __isl_take isl_printer
*p
, struct gpu_prog
*prog
)
240 for (i
= 0; i
< prog
->n_array
; ++i
) {
241 if (gpu_array_is_read_only_scalar(&prog
->array
[i
]))
243 p
= isl_printer_start_line(p
);
244 p
= isl_printer_print_str(p
, "cl_mem dev_");
245 p
= isl_printer_print_str(p
, prog
->array
[i
].name
);
246 p
= isl_printer_print_str(p
, ";");
247 p
= isl_printer_end_line(p
);
249 p
= isl_printer_start_line(p
);
250 p
= isl_printer_end_line(p
);
254 /* Given an array, check whether its positive size guard expression is
257 static int is_array_positive_size_guard_trivial(struct gpu_array_info
*array
)
262 guard
= gpu_array_positive_size_guard(array
);
263 is_trivial
= isl_set_plain_is_universe(guard
);
268 /* Allocate a device array for "array'.
270 * Emit a max-expression to ensure the device array can contain at least one
271 * element if the array's positive size guard expression is not trivial.
273 static __isl_give isl_printer
*allocate_device_array(__isl_take isl_printer
*p
,
274 struct gpu_array_info
*array
)
276 int need_lower_bound
;
278 p
= ppcg_start_block(p
);
280 p
= isl_printer_start_line(p
);
281 p
= isl_printer_print_str(p
, "dev_");
282 p
= isl_printer_print_str(p
, array
->name
);
283 p
= isl_printer_print_str(p
, " = clCreateBuffer(context, ");
284 p
= isl_printer_print_str(p
, "CL_MEM_READ_WRITE, ");
286 need_lower_bound
= !is_array_positive_size_guard_trivial(array
);
287 if (need_lower_bound
) {
288 p
= isl_printer_print_str(p
, "max(sizeof(");
289 p
= isl_printer_print_str(p
, array
->type
);
290 p
= isl_printer_print_str(p
, "), ");
292 p
= gpu_array_info_print_size(p
, array
);
293 if (need_lower_bound
)
294 p
= isl_printer_print_str(p
, ")");
296 p
= isl_printer_print_str(p
, ", NULL, &err);");
297 p
= isl_printer_end_line(p
);
298 p
= isl_printer_start_line(p
);
299 p
= isl_printer_print_str(p
, "openclCheckReturn(err);");
300 p
= isl_printer_end_line(p
);
302 p
= ppcg_end_block(p
);
307 /* Allocate device arrays.
309 static __isl_give isl_printer
*opencl_allocate_device_arrays(
310 __isl_take isl_printer
*p
, struct gpu_prog
*prog
)
314 for (i
= 0; i
< prog
->n_array
; ++i
) {
315 struct gpu_array_info
*array
= &prog
->array
[i
];
317 if (gpu_array_is_read_only_scalar(array
))
320 p
= allocate_device_array(p
, array
);
322 p
= isl_printer_start_line(p
);
323 p
= isl_printer_end_line(p
);
327 /* Print a call to the OpenCL clSetKernelArg() function which sets
328 * the arguments of the kernel. arg_name and arg_index are the name and the
329 * index of the kernel argument. The index of the leftmost argument of
330 * the kernel is 0 whereas the index of the rightmost argument of the kernel
331 * is n - 1, where n is the total number of the kernel arguments.
332 * read_only_scalar is a boolean that indicates whether the argument is a read
335 static __isl_give isl_printer
*opencl_set_kernel_argument(
336 __isl_take isl_printer
*p
, int kernel_id
,
337 const char *arg_name
, int arg_index
, int read_only_scalar
)
339 p
= isl_printer_start_line(p
);
340 p
= isl_printer_print_str(p
,
341 "openclCheckReturn(clSetKernelArg(kernel");
342 p
= isl_printer_print_int(p
, kernel_id
);
343 p
= isl_printer_print_str(p
, ", ");
344 p
= isl_printer_print_int(p
, arg_index
);
345 p
= isl_printer_print_str(p
, ", sizeof(");
347 if (read_only_scalar
) {
348 p
= isl_printer_print_str(p
, arg_name
);
349 p
= isl_printer_print_str(p
, "), &");
351 p
= isl_printer_print_str(p
, "cl_mem), (void *) &dev_");
353 p
= isl_printer_print_str(p
, arg_name
);
354 p
= isl_printer_print_str(p
, "));");
355 p
= isl_printer_end_line(p
);
360 /* Print the block sizes as a list of the sizes in each
363 static __isl_give isl_printer
*opencl_print_block_sizes(
364 __isl_take isl_printer
*p
, struct ppcg_kernel
*kernel
)
368 if (kernel
->n_block
> 0)
369 for (i
= 0; i
< kernel
->n_block
; ++i
) {
371 p
= isl_printer_print_str(p
, ", ");
372 p
= isl_printer_print_int(p
, kernel
->block_dim
[i
]);
375 p
= isl_printer_print_str(p
, "1");
380 /* Set the arguments of the OpenCL kernel by printing a call to the OpenCL
381 * clSetKernelArg() function for each kernel argument.
383 static __isl_give isl_printer
*opencl_set_kernel_arguments(
384 __isl_take isl_printer
*p
, struct gpu_prog
*prog
,
385 struct ppcg_kernel
*kernel
)
393 for (i
= 0; i
< prog
->n_array
; ++i
) {
397 space
= isl_space_copy(prog
->array
[i
].space
);
398 arr
= isl_union_set_extract_set(kernel
->arrays
, space
);
399 empty
= isl_set_plain_is_empty(arr
);
403 ro
= gpu_array_is_read_only_scalar(&prog
->array
[i
]);
404 opencl_set_kernel_argument(p
, kernel
->id
, prog
->array
[i
].name
,
409 space
= isl_union_set_get_space(kernel
->arrays
);
410 nparam
= isl_space_dim(space
, isl_dim_param
);
411 for (i
= 0; i
< nparam
; ++i
) {
414 name
= isl_space_get_dim_name(space
, isl_dim_param
, i
);
415 opencl_set_kernel_argument(p
, kernel
->id
, name
, arg_index
, 1);
418 isl_space_free(space
);
420 n
= isl_space_dim(kernel
->space
, isl_dim_set
);
421 for (i
= 0; i
< n
; ++i
) {
425 name
= isl_space_get_dim_name(kernel
->space
, isl_dim_set
, i
);
426 opencl_set_kernel_argument(p
, kernel
->id
, name
, arg_index
, 1);
433 /* Print the arguments to a kernel declaration or call. If "types" is set,
434 * then print a declaration (including the types of the arguments).
436 * The arguments are printed in the following order
437 * - the arrays accessed by the kernel
439 * - the host loop iterators
441 static __isl_give isl_printer
*opencl_print_kernel_arguments(
442 __isl_take isl_printer
*p
, struct gpu_prog
*prog
,
443 struct ppcg_kernel
*kernel
, int types
)
451 for (i
= 0; i
< prog
->n_array
; ++i
) {
455 space
= isl_space_copy(prog
->array
[i
].space
);
456 arr
= isl_union_set_extract_set(kernel
->arrays
, space
);
457 empty
= isl_set_plain_is_empty(arr
);
463 p
= isl_printer_print_str(p
, ", ");
466 p
= gpu_array_info_print_declaration_argument(p
,
467 &prog
->array
[i
], "__global");
469 p
= gpu_array_info_print_call_argument(p
,
475 space
= isl_union_set_get_space(kernel
->arrays
);
476 nparam
= isl_space_dim(space
, isl_dim_param
);
477 for (i
= 0; i
< nparam
; ++i
) {
480 name
= isl_space_get_dim_name(space
, isl_dim_param
, i
);
483 p
= isl_printer_print_str(p
, ", ");
485 p
= isl_printer_print_str(p
, "int ");
486 p
= isl_printer_print_str(p
, name
);
490 isl_space_free(space
);
492 n
= isl_space_dim(kernel
->space
, isl_dim_set
);
493 type
= isl_options_get_ast_iterator_type(prog
->ctx
);
494 for (i
= 0; i
< n
; ++i
) {
499 p
= isl_printer_print_str(p
, ", ");
500 name
= isl_space_get_dim_name(kernel
->space
, isl_dim_set
, i
);
502 p
= isl_printer_print_str(p
, type
);
503 p
= isl_printer_print_str(p
, " ");
505 p
= isl_printer_print_str(p
, name
);
513 /* Print the header of the given kernel.
515 static __isl_give isl_printer
*opencl_print_kernel_header(
516 __isl_take isl_printer
*p
, struct gpu_prog
*prog
,
517 struct ppcg_kernel
*kernel
)
519 p
= isl_printer_start_line(p
);
520 p
= isl_printer_print_str(p
, "__kernel void kernel");
521 p
= isl_printer_print_int(p
, kernel
->id
);
522 p
= isl_printer_print_str(p
, "(");
523 p
= opencl_print_kernel_arguments(p
, prog
, kernel
, 1);
524 p
= isl_printer_print_str(p
, ")");
525 p
= isl_printer_end_line(p
);
530 /* Unlike the equivalent function in the CUDA backend which prints iterators
531 * in reverse order to promote coalescing, this function does not print
532 * iterators in reverse order. The OpenCL backend currently does not take
533 * into account any coalescing considerations.
535 static __isl_give isl_printer
*opencl_print_kernel_iterators(
536 __isl_take isl_printer
*p
, struct ppcg_kernel
*kernel
)
539 isl_ctx
*ctx
= isl_ast_node_get_ctx(kernel
->tree
);
542 type
= isl_options_get_ast_iterator_type(ctx
);
544 n_grid
= isl_multi_pw_aff_dim(kernel
->grid_size
, isl_dim_set
);
546 p
= isl_printer_start_line(p
);
547 p
= isl_printer_print_str(p
, type
);
548 p
= isl_printer_print_str(p
, " ");
549 for (i
= 0; i
< n_grid
; ++i
) {
551 p
= isl_printer_print_str(p
, ", ");
552 p
= isl_printer_print_str(p
, "b");
553 p
= isl_printer_print_int(p
, i
);
554 p
= isl_printer_print_str(p
, " = get_group_id(");
555 p
= isl_printer_print_int(p
, i
);
556 p
= isl_printer_print_str(p
, ")");
558 p
= isl_printer_print_str(p
, ";");
559 p
= isl_printer_end_line(p
);
562 if (kernel
->n_block
> 0) {
563 p
= isl_printer_start_line(p
);
564 p
= isl_printer_print_str(p
, type
);
565 p
= isl_printer_print_str(p
, " ");
566 for (i
= 0; i
< kernel
->n_block
; ++i
) {
568 p
= isl_printer_print_str(p
, ", ");
569 p
= isl_printer_print_str(p
, "t");
570 p
= isl_printer_print_int(p
, i
);
571 p
= isl_printer_print_str(p
, " = get_local_id(");
572 p
= isl_printer_print_int(p
, i
);
573 p
= isl_printer_print_str(p
, ")");
575 p
= isl_printer_print_str(p
, ";");
576 p
= isl_printer_end_line(p
);
582 static __isl_give isl_printer
*opencl_print_kernel_var(
583 __isl_take isl_printer
*p
, struct ppcg_kernel_var
*var
)
588 p
= isl_printer_start_line(p
);
589 if (var
->type
== ppcg_access_shared
)
590 p
= isl_printer_print_str(p
, "__local ");
591 p
= isl_printer_print_str(p
, var
->array
->type
);
592 p
= isl_printer_print_str(p
, " ");
593 p
= isl_printer_print_str(p
, var
->name
);
594 for (j
= 0; j
< var
->array
->n_index
; ++j
) {
595 p
= isl_printer_print_str(p
, "[");
596 v
= isl_vec_get_element_val(var
->size
, j
);
597 p
= isl_printer_print_val(p
, v
);
598 p
= isl_printer_print_str(p
, "]");
601 p
= isl_printer_print_str(p
, ";");
602 p
= isl_printer_end_line(p
);
607 static __isl_give isl_printer
*opencl_print_kernel_vars(
608 __isl_take isl_printer
*p
, struct ppcg_kernel
*kernel
)
612 for (i
= 0; i
< kernel
->n_var
; ++i
)
613 p
= opencl_print_kernel_var(p
, &kernel
->var
[i
]);
618 /* Print a call to barrier() which is a sync statement.
619 * All work-items in a work-group executing the kernel on a processor must
620 * execute the barrier() function before any are allowed to continue execution
621 * beyond the barrier.
622 * The flag CLK_LOCAL_MEM_FENCE makes the barrier function either flush any
623 * variables stored in local memory or queue a memory fence to ensure correct
624 * ordering of memory operations to local memory.
625 * The flag CLK_GLOBAL_MEM_FENCE makes the barrier function queue a memory
626 * fence to ensure correct ordering of memory operations to global memory.
628 static __isl_give isl_printer
*opencl_print_sync(__isl_take isl_printer
*p
,
629 struct ppcg_kernel_stmt
*stmt
)
631 p
= isl_printer_start_line(p
);
632 p
= isl_printer_print_str(p
,
633 "barrier(CLK_LOCAL_MEM_FENCE | CLK_GLOBAL_MEM_FENCE);");
634 p
= isl_printer_end_line(p
);
639 /* This function is called for each user statement in the AST,
640 * i.e., for each kernel body statement, copy statement or sync statement.
642 static __isl_give isl_printer
*opencl_print_kernel_stmt(
643 __isl_take isl_printer
*p
,
644 __isl_take isl_ast_print_options
*print_options
,
645 __isl_keep isl_ast_node
*node
, void *user
)
648 struct ppcg_kernel_stmt
*stmt
;
650 id
= isl_ast_node_get_annotation(node
);
651 stmt
= isl_id_get_user(id
);
654 isl_ast_print_options_free(print_options
);
656 switch (stmt
->type
) {
657 case ppcg_kernel_copy
:
658 return ppcg_kernel_print_copy(p
, stmt
);
659 case ppcg_kernel_sync
:
660 return opencl_print_sync(p
, stmt
);
661 case ppcg_kernel_domain
:
662 return ppcg_kernel_print_domain(p
, stmt
);
668 /* Return true if there is a double array in prog->array or
669 * if any of the types in prog->scop involve any doubles.
670 * To check the latter condition, we simply search for the string "double"
671 * in the type definitions, which may result in false positives.
673 static __isl_give
int any_double_elements(struct gpu_prog
*prog
)
677 for (i
= 0; i
< prog
->n_array
; ++i
)
678 if (strcmp(prog
->array
[i
].type
, "double") == 0)
681 for (i
= 0; i
< prog
->scop
->pet
->n_type
; ++i
) {
682 struct pet_type
*type
= prog
->scop
->pet
->types
[i
];
684 if (strstr(type
->definition
, "double"))
691 /* Prints a #pragma to enable support for double floating-point
692 * precision. OpenCL 1.0 adds support for double precision floating-point as
693 * an optional extension. An application that wants to use double will need to
694 * include the #pragma OPENCL EXTENSION cl_khr_fp64 : enable directive before
695 * any double precision data type is declared in the kernel code.
697 static __isl_give isl_printer
*opencl_enable_double_support(
698 __isl_take isl_printer
*p
)
702 p
= isl_printer_start_line(p
);
703 p
= isl_printer_print_str(p
, "#pragma OPENCL EXTENSION cl_khr_fp64 :"
705 p
= isl_printer_end_line(p
);
706 p
= isl_printer_start_line(p
);
707 p
= isl_printer_end_line(p
);
712 static __isl_give isl_printer
*opencl_print_kernel(struct gpu_prog
*prog
,
713 struct ppcg_kernel
*kernel
, __isl_take isl_printer
*p
)
715 isl_ctx
*ctx
= isl_ast_node_get_ctx(kernel
->tree
);
716 isl_ast_print_options
*print_options
;
718 print_options
= isl_ast_print_options_alloc(ctx
);
719 print_options
= isl_ast_print_options_set_print_user(print_options
,
720 &opencl_print_kernel_stmt
, NULL
);
722 p
= isl_printer_set_output_format(p
, ISL_FORMAT_C
);
723 p
= opencl_print_kernel_header(p
, prog
, kernel
);
724 p
= isl_printer_print_str(p
, "{");
725 p
= isl_printer_end_line(p
);
726 p
= isl_printer_indent(p
, 4);
727 p
= opencl_print_kernel_iterators(p
, kernel
);
728 p
= opencl_print_kernel_vars(p
, kernel
);
729 p
= isl_printer_end_line(p
);
730 p
= gpu_print_macros(p
, kernel
->tree
);
731 p
= isl_ast_node_print(kernel
->tree
, p
, print_options
);
732 p
= isl_printer_indent(p
, -4);
733 p
= isl_printer_start_line(p
);
734 p
= isl_printer_print_str(p
, "}");
735 p
= isl_printer_end_line(p
);
740 struct print_host_user_data_opencl
{
741 struct opencl_info
*opencl
;
742 struct gpu_prog
*prog
;
745 /* This function prints the i'th block size multiplied by the i'th grid size,
746 * where i (a parameter to this function) is one of the possible dimensions of
747 * grid sizes and block sizes.
748 * If the dimension of block sizes is not equal to the dimension of grid sizes
749 * the output is calculated as follows:
752 * block_sizes[dim1] is the list of blocks sizes and it contains dim1 elements.
753 * grid_sizes[dim2] is the list of grid sizes and it contains dim2 elements.
756 * If (i > dim2) then the output is block_sizes[i]
757 * If (i > dim1) then the output is grid_sizes[i]
759 static __isl_give isl_printer
*opencl_print_total_number_of_work_items_for_dim(
760 __isl_take isl_printer
*p
, struct ppcg_kernel
*kernel
, int i
)
762 int grid_dim
, block_dim
;
763 isl_pw_aff
*bound_grid
;
765 grid_dim
= isl_multi_pw_aff_dim(kernel
->grid_size
, isl_dim_set
);
766 block_dim
= kernel
->n_block
;
768 if (i
< min(grid_dim
, block_dim
)) {
769 bound_grid
= isl_multi_pw_aff_get_pw_aff(kernel
->grid_size
, i
);
770 p
= isl_printer_print_str(p
, "(");
771 p
= isl_printer_print_pw_aff(p
, bound_grid
);
772 p
= isl_printer_print_str(p
, ") * ");
773 p
= isl_printer_print_int(p
, kernel
->block_dim
[i
]);
774 isl_pw_aff_free(bound_grid
);
775 } else if (i
>= grid_dim
)
776 p
= isl_printer_print_int(p
, kernel
->block_dim
[i
]);
778 bound_grid
= isl_multi_pw_aff_get_pw_aff(kernel
->grid_size
, i
);
779 p
= isl_printer_print_pw_aff(p
, bound_grid
);
780 isl_pw_aff_free(bound_grid
);
786 /* Print a list that represents the total number of work items. The list is
787 * constructed by performing an element-wise multiplication of the block sizes
788 * and the grid sizes. To explain how the list is constructed, suppose that:
789 * block_sizes[dim1] is the list of blocks sizes and it contains dim1 elements.
790 * grid_sizes[dim2] is the list of grid sizes and it contains dim2 elements.
792 * The output of this function is constructed as follows:
793 * If (dim1 > dim2) then the output is the following list:
794 * grid_sizes[0]*block_sizes[0], ..., grid_sizes[dim2-1]*block_sizes[dim2-1],
795 * block_sizes[dim2], ..., block_sizes[dim1-2], block_sizes[dim1-1].
797 * If (dim2 > dim1) then the output is the following list:
798 * grid_sizes[0]*block_sizes[0], ..., grid_sizes[dim1-1] * block_sizes[dim1-1],
799 * grid_sizes[dim1], grid_sizes[dim2-2], grid_sizes[dim2-1].
801 * To calculate the total number of work items out of the list constructed by
802 * this function, the user should multiply the elements of the list.
804 static __isl_give isl_printer
*opencl_print_total_number_of_work_items_as_list(
805 __isl_take isl_printer
*p
, struct ppcg_kernel
*kernel
)
808 int grid_dim
, block_dim
;
810 grid_dim
= isl_multi_pw_aff_dim(kernel
->grid_size
, isl_dim_set
);
811 block_dim
= kernel
->n_block
;
813 if ((grid_dim
<= 0) || (block_dim
<= 0)) {
814 p
= isl_printer_print_str(p
, "1");
818 for (i
= 0; i
<= max(grid_dim
, block_dim
) - 1; i
++) {
820 p
= isl_printer_print_str(p
, ", ");
822 p
= opencl_print_total_number_of_work_items_for_dim(p
,
829 /* Copy "array" from the host to the device (to_host = 0) or
830 * back from the device to the host (to_host = 1).
832 static __isl_give isl_printer
*copy_array(__isl_take isl_printer
*p
,
833 struct gpu_array_info
*array
, int to_host
)
835 p
= isl_printer_start_line(p
);
836 p
= isl_printer_print_str(p
, "openclCheckReturn(");
838 p
= isl_printer_print_str(p
, "clEnqueueReadBuffer");
840 p
= isl_printer_print_str(p
, "clEnqueueWriteBuffer");
841 p
= isl_printer_print_str(p
, "(queue, dev_");
842 p
= isl_printer_print_str(p
, array
->name
);
843 p
= isl_printer_print_str(p
, ", CL_TRUE, 0, ");
844 p
= gpu_array_info_print_size(p
, array
);
846 if (gpu_array_is_scalar(array
))
847 p
= isl_printer_print_str(p
, ", &");
849 p
= isl_printer_print_str(p
, ", ");
850 p
= isl_printer_print_str(p
, array
->name
);
851 p
= isl_printer_print_str(p
, ", 0, NULL, NULL));");
852 p
= isl_printer_end_line(p
);
857 /* Copy "array" from the host to the device.
859 static __isl_give isl_printer
*copy_array_to_device(__isl_take isl_printer
*p
,
862 struct gpu_array_info
*array
= user
;
864 return copy_array(p
, array
, 0);
867 /* Copy "array" back from the device to the host.
869 static __isl_give isl_printer
*copy_array_from_device(__isl_take isl_printer
*p
,
872 struct gpu_array_info
*array
= user
;
874 return copy_array(p
, array
, 1);
877 /* Copy the "copy" arrays from the host to the device (to_host = 0) or
878 * back from the device to the host (to_host = 1).
880 * Only perform the copying for arrays with strictly positive size.
882 static __isl_give isl_printer
*opencl_copy_arrays(__isl_take isl_printer
*p
,
883 struct gpu_prog
*prog
, __isl_keep isl_union_set
*copy
, int to_host
)
887 for (i
= 0; i
< prog
->n_array
; ++i
) {
888 struct gpu_array_info
*array
= &prog
->array
[i
];
894 if (gpu_array_is_read_only_scalar(array
))
897 space
= isl_space_copy(array
->space
);
898 copy_i
= isl_union_set_extract_set(copy
, space
);
899 empty
= isl_set_plain_is_empty(copy_i
);
900 isl_set_free(copy_i
);
904 guard
= gpu_array_positive_size_guard(array
);
905 p
= ppcg_print_guarded(p
, guard
, isl_set_copy(prog
->context
),
906 to_host
? ©_array_from_device
:
907 ©_array_to_device
, array
);
910 p
= isl_printer_start_line(p
);
911 p
= isl_printer_end_line(p
);
915 /* Copy the prog->copy_in arrays from the host to the device.
917 static __isl_give isl_printer
*opencl_copy_arrays_to_device(
918 __isl_take isl_printer
*p
, struct gpu_prog
*prog
)
920 return opencl_copy_arrays(p
, prog
, prog
->copy_in
, 0);
923 /* Copy the prog->copy_out arrays back from the device to the host.
925 static __isl_give isl_printer
*opencl_copy_arrays_from_device(
926 __isl_take isl_printer
*p
, struct gpu_prog
*prog
)
928 return opencl_copy_arrays(p
, prog
, prog
->copy_out
, 1);
931 /* Print the user statement of the host code to "p".
933 * In particular, print a block of statements that defines the grid
934 * and the work group and then launches the kernel.
936 * A grid is composed of many work groups (blocks), each work group holds
937 * many work-items (threads).
939 * global_work_size[kernel->n_block] represents the total number of work
940 * items. It points to an array of kernel->n_block unsigned
941 * values that describe the total number of work-items that will execute
942 * the kernel. The total number of work-items is computed as:
943 * global_work_size[0] *...* global_work_size[kernel->n_block - 1].
945 * The size of each work group (i.e. the number of work-items in each work
946 * group) is described using block_size[kernel->n_block]. The total
947 * number of work-items in a block (work-group) is computed as:
948 * block_size[0] *... * block_size[kernel->n_block - 1].
950 * For more information check:
951 * http://www.khronos.org/registry/cl/sdk/1.0/docs/man/xhtml/clEnqueueNDRangeKernel.html
953 static __isl_give isl_printer
*opencl_print_host_user(
954 __isl_take isl_printer
*p
,
955 __isl_take isl_ast_print_options
*print_options
,
956 __isl_keep isl_ast_node
*node
, void *user
)
959 struct ppcg_kernel
*kernel
;
960 struct print_host_user_data_opencl
*data
;
963 id
= isl_ast_node_get_annotation(node
);
964 kernel
= isl_id_get_user(id
);
967 data
= (struct print_host_user_data_opencl
*) user
;
969 p
= isl_printer_start_line(p
);
970 p
= isl_printer_print_str(p
, "{");
971 p
= isl_printer_end_line(p
);
972 p
= isl_printer_indent(p
, 2);
974 p
= isl_printer_start_line(p
);
975 p
= isl_printer_print_str(p
, "size_t global_work_size[");
977 if (kernel
->n_block
> 0)
978 p
= isl_printer_print_int(p
, kernel
->n_block
);
980 p
= isl_printer_print_int(p
, 1);
982 p
= isl_printer_print_str(p
, "] = {");
983 p
= opencl_print_total_number_of_work_items_as_list(p
, kernel
);
984 p
= isl_printer_print_str(p
, "};");
985 p
= isl_printer_end_line(p
);
987 p
= isl_printer_start_line(p
);
988 p
= isl_printer_print_str(p
, "size_t block_size[");
990 if (kernel
->n_block
> 0)
991 p
= isl_printer_print_int(p
, kernel
->n_block
);
993 p
= isl_printer_print_int(p
, 1);
995 p
= isl_printer_print_str(p
, "] = {");
996 p
= opencl_print_block_sizes(p
, kernel
);
997 p
= isl_printer_print_str(p
, "};");
998 p
= isl_printer_end_line(p
);
1000 p
= isl_printer_start_line(p
);
1001 p
= isl_printer_print_str(p
, "cl_kernel kernel");
1002 p
= isl_printer_print_int(p
, kernel
->id
);
1003 p
= isl_printer_print_str(p
, " = clCreateKernel(program, \"kernel");
1004 p
= isl_printer_print_int(p
, kernel
->id
);
1005 p
= isl_printer_print_str(p
, "\", &err);");
1006 p
= isl_printer_end_line(p
);
1007 p
= isl_printer_start_line(p
);
1008 p
= isl_printer_print_str(p
, "openclCheckReturn(err);");
1009 p
= isl_printer_end_line(p
);
1011 opencl_set_kernel_arguments(p
, data
->prog
, kernel
);
1013 p
= isl_printer_start_line(p
);
1014 p
= isl_printer_print_str(p
, "openclCheckReturn(clEnqueueNDRangeKernel"
1016 p
= isl_printer_print_int(p
, kernel
->id
);
1017 p
= isl_printer_print_str(p
, ", ");
1018 if (kernel
->n_block
> 0)
1019 p
= isl_printer_print_int(p
, kernel
->n_block
);
1021 p
= isl_printer_print_int(p
, 1);
1023 p
= isl_printer_print_str(p
, ", NULL, global_work_size, "
1025 "0, NULL, NULL));");
1026 p
= isl_printer_end_line(p
);
1027 p
= isl_printer_start_line(p
);
1028 p
= isl_printer_print_str(p
, "openclCheckReturn("
1029 "clReleaseKernel(kernel");
1030 p
= isl_printer_print_int(p
, kernel
->id
);
1031 p
= isl_printer_print_str(p
, "));");
1032 p
= isl_printer_end_line(p
);
1033 p
= isl_printer_start_line(p
);
1034 p
= isl_printer_print_str(p
, "clFinish(queue);");
1035 p
= isl_printer_end_line(p
);
1036 p
= isl_printer_indent(p
, -2);
1037 p
= isl_printer_start_line(p
);
1038 p
= isl_printer_print_str(p
, "}");
1039 p
= isl_printer_end_line(p
);
1041 p
= isl_printer_start_line(p
);
1042 p
= isl_printer_end_line(p
);
1044 data
->opencl
->kprinter
= opencl_print_kernel(data
->prog
, kernel
,
1045 data
->opencl
->kprinter
);
1047 isl_ast_print_options_free(print_options
);
1052 static __isl_give isl_printer
*opencl_print_host_code(
1053 __isl_take isl_printer
*p
, struct gpu_prog
*prog
,
1054 __isl_keep isl_ast_node
*tree
, struct opencl_info
*opencl
)
1056 isl_ast_print_options
*print_options
;
1057 isl_ctx
*ctx
= isl_ast_node_get_ctx(tree
);
1058 struct print_host_user_data_opencl data
= { opencl
, prog
};
1060 print_options
= isl_ast_print_options_alloc(ctx
);
1061 print_options
= isl_ast_print_options_set_print_user(print_options
,
1062 &opencl_print_host_user
, &data
);
1064 p
= gpu_print_macros(p
, tree
);
1065 p
= isl_ast_node_print(tree
, p
, print_options
);
1070 /* Create an OpenCL device, context, command queue and build the kernel.
1071 * input is the name of the input file provided to ppcg.
1073 static __isl_give isl_printer
*opencl_setup(__isl_take isl_printer
*p
,
1074 const char *input
, struct opencl_info
*info
)
1078 p
= isl_printer_start_line(p
);
1079 p
= isl_printer_print_str(p
, "cl_device_id device;");
1080 p
= isl_printer_end_line(p
);
1081 p
= isl_printer_start_line(p
);
1082 p
= isl_printer_print_str(p
, "cl_context context;");
1083 p
= isl_printer_end_line(p
);
1084 p
= isl_printer_start_line(p
);
1085 p
= isl_printer_print_str(p
, "cl_program program;");
1086 p
= isl_printer_end_line(p
);
1087 p
= isl_printer_start_line(p
);
1088 p
= isl_printer_print_str(p
, "cl_command_queue queue;");
1089 p
= isl_printer_end_line(p
);
1090 p
= isl_printer_start_line(p
);
1091 p
= isl_printer_print_str(p
, "cl_int err;");
1092 p
= isl_printer_end_line(p
);
1093 p
= isl_printer_start_line(p
);
1094 p
= isl_printer_print_str(p
, "device = opencl_create_device(");
1095 p
= isl_printer_print_int(p
, info
->options
->opencl_use_gpu
);
1096 p
= isl_printer_print_str(p
, ");");
1097 p
= isl_printer_end_line(p
);
1098 p
= isl_printer_start_line(p
);
1099 p
= isl_printer_print_str(p
, "context = clCreateContext(NULL, 1, "
1100 "&device, NULL, NULL, &err);");
1101 p
= isl_printer_end_line(p
);
1102 p
= isl_printer_start_line(p
);
1103 p
= isl_printer_print_str(p
, "openclCheckReturn(err);");
1104 p
= isl_printer_end_line(p
);
1105 p
= isl_printer_start_line(p
);
1106 p
= isl_printer_print_str(p
, "queue = clCreateCommandQueue"
1107 "(context, device, 0, &err);");
1108 p
= isl_printer_end_line(p
);
1109 p
= isl_printer_start_line(p
);
1110 p
= isl_printer_print_str(p
, "openclCheckReturn(err);");
1111 p
= isl_printer_end_line(p
);
1113 p
= isl_printer_start_line(p
);
1114 p
= isl_printer_print_str(p
, "program = ");
1116 if (info
->options
->opencl_embed_kernel_code
) {
1117 p
= isl_printer_print_str(p
, "opencl_build_program_from_string("
1118 "context, device, kernel_code, "
1119 "sizeof(kernel_code), \"");
1121 p
= isl_printer_print_str(p
, "opencl_build_program_from_file("
1122 "context, device, \"");
1123 p
= isl_printer_print_str(p
, info
->kernel_c_name
);
1124 p
= isl_printer_print_str(p
, "\", \"");
1127 if (info
->options
->opencl_compiler_options
)
1128 p
= isl_printer_print_str(p
,
1129 info
->options
->opencl_compiler_options
);
1131 p
= isl_printer_print_str(p
, "\");");
1132 p
= isl_printer_end_line(p
);
1133 p
= isl_printer_start_line(p
);
1134 p
= isl_printer_end_line(p
);
1139 static __isl_give isl_printer
*opencl_release_cl_objects(
1140 __isl_take isl_printer
*p
, struct opencl_info
*info
)
1142 p
= isl_printer_start_line(p
);
1143 p
= isl_printer_print_str(p
, "openclCheckReturn(clReleaseCommandQueue"
1145 p
= isl_printer_end_line(p
);
1146 p
= isl_printer_start_line(p
);
1147 p
= isl_printer_print_str(p
, "openclCheckReturn(clReleaseProgram"
1149 p
= isl_printer_end_line(p
);
1150 p
= isl_printer_start_line(p
);
1151 p
= isl_printer_print_str(p
, "openclCheckReturn(clReleaseContext"
1153 p
= isl_printer_end_line(p
);
1158 /* Free the device array corresponding to "array"
1160 static __isl_give isl_printer
*release_device_array(__isl_take isl_printer
*p
,
1161 struct gpu_array_info
*array
)
1163 p
= isl_printer_start_line(p
);
1164 p
= isl_printer_print_str(p
, "openclCheckReturn("
1165 "clReleaseMemObject(dev_");
1166 p
= isl_printer_print_str(p
, array
->name
);
1167 p
= isl_printer_print_str(p
, "));");
1168 p
= isl_printer_end_line(p
);
1173 /* Free the device arrays.
1175 static __isl_give isl_printer
*opencl_release_device_arrays(
1176 __isl_take isl_printer
*p
, struct gpu_prog
*prog
)
1180 for (i
= 0; i
< prog
->n_array
; ++i
) {
1181 struct gpu_array_info
*array
= &prog
->array
[i
];
1182 if (gpu_array_is_read_only_scalar(array
))
1185 p
= release_device_array(p
, array
);
1190 /* Given a gpu_prog "prog" and the corresponding transformed AST
1191 * "tree", print the entire OpenCL code to "p".
1193 static __isl_give isl_printer
*print_opencl(__isl_take isl_printer
*p
,
1194 struct gpu_prog
*prog
, __isl_keep isl_ast_node
*tree
,
1195 struct gpu_types
*types
, void *user
)
1197 struct opencl_info
*opencl
= user
;
1199 opencl
->kprinter
= isl_printer_set_output_format(opencl
->kprinter
,
1201 if (any_double_elements(prog
))
1202 opencl
->kprinter
= opencl_enable_double_support(
1204 opencl
->kprinter
= gpu_print_types(opencl
->kprinter
, types
, prog
);
1206 if (!opencl
->kprinter
)
1207 return isl_printer_free(p
);
1209 p
= ppcg_start_block(p
);
1211 p
= opencl_print_host_macros(p
);
1213 p
= opencl_declare_device_arrays(p
, prog
);
1214 p
= opencl_setup(p
, opencl
->input
, opencl
);
1215 p
= opencl_allocate_device_arrays(p
, prog
);
1216 p
= opencl_copy_arrays_to_device(p
, prog
);
1218 p
= opencl_print_host_code(p
, prog
, tree
, opencl
);
1220 p
= opencl_copy_arrays_from_device(p
, prog
);
1221 p
= opencl_release_device_arrays(p
, prog
);
1222 p
= opencl_release_cl_objects(p
, opencl
);
1224 p
= ppcg_end_block(p
);
1229 /* Transform the code in the file called "input" by replacing
1230 * all scops by corresponding OpenCL code.
1231 * The host code is written to "output" or a name derived from
1232 * "input" if "output" is NULL.
1233 * The kernel code is placed in separate files with names
1234 * derived from "output" or "input".
1236 * We let generate_gpu do all the hard work and then let it call
1237 * us back for printing the AST in print_cuda.
1239 * To prepare for this printing, we first open the output files
1240 * and we close them after generate_gpu has finished.
1242 int generate_opencl(isl_ctx
*ctx
, struct ppcg_options
*options
,
1243 const char *input
, const char *output
)
1245 struct opencl_info opencl
= { options
, input
, output
};
1248 opencl
.kprinter
= isl_printer_to_str(ctx
);
1249 r
= opencl_open_files(&opencl
);
1252 r
= generate_gpu(ctx
, input
, opencl
.host_c
, options
,
1253 &print_opencl
, &opencl
);
1255 opencl_close_files(&opencl
);
1256 isl_printer_free(opencl
.kprinter
);