Halide 14.0.0
Halide compiler and libraries
HalideRuntime.h
Go to the documentation of this file.
1#ifndef HALIDE_HALIDERUNTIME_H
2#define HALIDE_HALIDERUNTIME_H
3
4#ifndef COMPILING_HALIDE_RUNTIME
5#ifdef __cplusplus
6#include <cstddef>
7#include <cstdint>
8#include <cstring>
9#else
10#include <stdbool.h>
11#include <stddef.h>
12#include <stdint.h>
13#include <string.h>
14#endif
15#else
16#include "runtime_internal.h"
17#endif
18
19#ifdef __cplusplus
20// Forward declare type to allow naming typed handles.
21// See Type.h for documentation.
22template<typename T>
24#endif
25
26#ifdef __cplusplus
27extern "C" {
28#endif
29
30#ifdef _MSC_VER
31// Note that (for MSVC) you should not use "inline" along with HALIDE_ALWAYS_INLINE;
32// it is not necessary, and may produce warnings for some build configurations.
33#define HALIDE_ALWAYS_INLINE __forceinline
34#define HALIDE_NEVER_INLINE __declspec(noinline)
35#else
36// Note that (for Posixy compilers) you should always use "inline" along with HALIDE_ALWAYS_INLINE;
37// otherwise some corner-case scenarios may erroneously report link errors.
38#define HALIDE_ALWAYS_INLINE inline __attribute__((always_inline))
39#define HALIDE_NEVER_INLINE __attribute__((noinline))
40#endif
41
42#ifndef HALIDE_MUST_USE_RESULT
43#ifdef __has_attribute
44#if __has_attribute(nodiscard)
45// C++17 or later
46#define HALIDE_MUST_USE_RESULT [[nodiscard]]
47#elif __has_attribute(warn_unused_result)
48// Clang/GCC
49#define HALIDE_MUST_USE_RESULT __attribute__((warn_unused_result))
50#else
51#define HALIDE_MUST_USE_RESULT
52#endif
53#else
54#define HALIDE_MUST_USE_RESULT
55#endif
56#endif
57
58/** \file
59 *
60 * This file declares the routines used by Halide internally in its
61 * runtime. On platforms that support weak linking, these can be
62 * replaced with user-defined versions by defining an extern "C"
63 * function with the same name and signature.
64 *
65 * When doing Just In Time (JIT) compilation members of
66 * some_pipeline_or_func.jit_handlers() must be replaced instead. The
67 * corresponding methods are documented below.
68 *
69 * All of these functions take a "void *user_context" parameter as their
70 * first argument; if the Halide kernel that calls back to any of these
71 * functions has been compiled with the UserContext feature set on its Target,
72 * then the value of that pointer passed from the code that calls the
73 * Halide kernel is piped through to the function.
74 *
75 * Some of these are also useful to call when using the default
76 * implementation. E.g. halide_shutdown_thread_pool.
77 *
78 * Note that even on platforms with weak linking, some linker setups
79 * may not respect the override you provide. E.g. if the override is
80 * in a shared library and the halide object files are linked directly
81 * into the output, the builtin versions of the runtime functions will
82 * be called. See your linker documentation for more details. On
83 * Linux, LD_DYNAMIC_WEAK=1 may help.
84 *
85 */
86
87// Forward-declare to suppress warnings if compiling as C.
88struct halide_buffer_t;
89
90/** Print a message to stderr. Main use is to support tracing
91 * functionality, print, and print_when calls. Also called by the default
92 * halide_error. This function can be replaced in JITed code by using
93 * halide_custom_print and providing an implementation of halide_print
94 * in AOT code. See Func::set_custom_print.
95 */
96// @{
97extern void halide_print(void *user_context, const char *);
98extern void halide_default_print(void *user_context, const char *);
99typedef void (*halide_print_t)(void *, const char *);
101// @}
102
103/** Halide calls this function on runtime errors (for example bounds
104 * checking failures). This function can be replaced in JITed code by
105 * using Func::set_error_handler, or in AOT code by calling
106 * halide_set_error_handler. In AOT code on platforms that support
107 * weak linking (i.e. not Windows), you can also override it by simply
108 * defining your own halide_error.
109 */
110// @{
111extern void halide_error(void *user_context, const char *);
112extern void halide_default_error(void *user_context, const char *);
113typedef void (*halide_error_handler_t)(void *, const char *);
115// @}
116
117/** Cross-platform mutex. Must be initialized with zero and implementation
118 * must treat zero as an unlocked mutex with no waiters, etc.
119 */
121 uintptr_t _private[1];
122};
123
124/** Cross platform condition variable. Must be initialized to 0. */
126 uintptr_t _private[1];
127};
128
129/** A basic set of mutex and condition variable functions, which call
130 * platform specific code for mutual exclusion. Equivalent to posix
131 * calls. */
132//@{
133extern void halide_mutex_lock(struct halide_mutex *mutex);
134extern void halide_mutex_unlock(struct halide_mutex *mutex);
135extern void halide_cond_signal(struct halide_cond *cond);
136extern void halide_cond_broadcast(struct halide_cond *cond);
137extern void halide_cond_wait(struct halide_cond *cond, struct halide_mutex *mutex);
138//@}
139
140/** Functions for constructing/destroying/locking/unlocking arrays of mutexes. */
141struct halide_mutex_array;
142//@{
143extern struct halide_mutex_array *halide_mutex_array_create(int sz);
144extern void halide_mutex_array_destroy(void *user_context, void *array);
145extern int halide_mutex_array_lock(struct halide_mutex_array *array, int entry);
146extern int halide_mutex_array_unlock(struct halide_mutex_array *array, int entry);
147//@}
148
149/** Define halide_do_par_for to replace the default thread pool
150 * implementation. halide_shutdown_thread_pool can also be called to
151 * release resources used by the default thread pool on platforms
152 * where it makes sense. See Func::set_custom_do_task and
153 * Func::set_custom_do_par_for. Should return zero if all the jobs
154 * return zero, or an arbitrarily chosen return value from one of the
155 * jobs otherwise.
156 */
157//@{
158typedef int (*halide_task_t)(void *user_context, int task_number, uint8_t *closure);
159extern int halide_do_par_for(void *user_context,
160 halide_task_t task,
161 int min, int size, uint8_t *closure);
162extern void halide_shutdown_thread_pool();
163//@}
164
165/** Set a custom method for performing a parallel for loop. Returns
166 * the old do_par_for handler. */
167typedef int (*halide_do_par_for_t)(void *, halide_task_t, int, int, uint8_t *);
169
170/** An opaque struct representing a semaphore. Used by the task system for async tasks. */
173};
174
175/** A struct representing a semaphore and a number of items that must
176 * be acquired from it. Used in halide_parallel_task_t below. */
179 int count;
180};
181extern int halide_semaphore_init(struct halide_semaphore_t *, int n);
182extern int halide_semaphore_release(struct halide_semaphore_t *, int n);
183extern bool halide_semaphore_try_acquire(struct halide_semaphore_t *, int n);
184typedef int (*halide_semaphore_init_t)(struct halide_semaphore_t *, int);
185typedef int (*halide_semaphore_release_t)(struct halide_semaphore_t *, int);
187
188/** A task representing a serial for loop evaluated over some range.
189 * Note that task_parent is a pass through argument that should be
190 * passed to any dependent taks that are invoked using halide_do_parallel_tasks
191 * underneath this call. */
192typedef int (*halide_loop_task_t)(void *user_context, int min, int extent,
193 uint8_t *closure, void *task_parent);
194
195/** A parallel task to be passed to halide_do_parallel_tasks. This
196 * task may recursively call halide_do_parallel_tasks, and there may
197 * be complex dependencies between seemingly unrelated tasks expressed
198 * using semaphores. If you are using a custom task system, care must
199 * be taken to avoid potential deadlock. This can be done by carefully
200 * respecting the static metadata at the end of the task struct.*/
202 // The function to call. It takes a user context, a min and
203 // extent, a closure, and a task system pass through argument.
205
206 // The closure to pass it
208
209 // The name of the function to be called. For debugging purposes only.
210 const char *name;
211
212 // An array of semaphores that must be acquired before the
213 // function is called. Must be reacquired for every call made.
216
217 // The entire range the function should be called over. This range
218 // may be sliced up and the function called multiple times.
220
221 // A parallel task provides several pieces of metadata to prevent
222 // unbounded resource usage or deadlock.
223
224 // The first is the minimum number of execution contexts (call
225 // stacks or threads) necessary for the function to run to
226 // completion. This may be greater than one when there is nested
227 // parallelism with internal producer-consumer relationships
228 // (calling the function recursively spawns and blocks on parallel
229 // sub-tasks that communicate with each other via semaphores). If
230 // a parallel runtime calls the function when fewer than this many
231 // threads are idle, it may need to create more threads to
232 // complete the task, or else risk deadlock due to committing all
233 // threads to tasks that cannot complete without more.
234 //
235 // FIXME: Note that extern stages are assumed to only require a
236 // single thread to complete. If the extern stage is itself a
237 // Halide pipeline, this may be an underestimate.
239
240 // The calls to the function should be in serial order from min to min+extent-1, with only
241 // one executing at a time. If false, any order is fine, and
242 // concurrency is fine.
243 bool serial;
244};
245
246/** Enqueue some number of the tasks described above and wait for them
247 * to complete. While waiting, the calling threads assists with either
248 * the tasks enqueued, or other non-blocking tasks in the task
249 * system. Note that task_parent should be NULL for top-level calls
250 * and the pass through argument if this call is being made from
251 * another task. */
252extern int halide_do_parallel_tasks(void *user_context, int num_tasks,
253 struct halide_parallel_task_t *tasks,
254 void *task_parent);
255
256/** If you use the default do_par_for, you can still set a custom
257 * handler to perform each individual task. Returns the old handler. */
258//@{
259typedef int (*halide_do_task_t)(void *, halide_task_t, int, uint8_t *);
261extern int halide_do_task(void *user_context, halide_task_t f, int idx,
262 uint8_t *closure);
263//@}
264
265/** The version of do_task called for loop tasks. By default calls the
266 * loop task with the same arguments. */
267// @{
268typedef int (*halide_do_loop_task_t)(void *, halide_loop_task_t, int, int, uint8_t *, void *);
270extern int halide_do_loop_task(void *user_context, halide_loop_task_t f, int min, int extent,
271 uint8_t *closure, void *task_parent);
272//@}
273
274/** Provide an entire custom tasking runtime via function
275 * pointers. Note that do_task and semaphore_try_acquire are only ever
276 * called by halide_default_do_par_for and
277 * halide_default_do_parallel_tasks, so it's only necessary to provide
278 * those if you are mixing in the default implementations of
279 * do_par_for and do_parallel_tasks. */
280// @{
281typedef int (*halide_do_parallel_tasks_t)(void *, int, struct halide_parallel_task_t *,
282 void *task_parent);
291// @}
292
293/** The default versions of the parallel runtime functions. */
294// @{
295extern int halide_default_do_par_for(void *user_context,
296 halide_task_t task,
297 int min, int size, uint8_t *closure);
298extern int halide_default_do_parallel_tasks(void *user_context,
299 int num_tasks,
300 struct halide_parallel_task_t *tasks,
301 void *task_parent);
302extern int halide_default_do_task(void *user_context, halide_task_t f, int idx,
304extern int halide_default_do_loop_task(void *user_context, halide_loop_task_t f,
305 int min, int extent,
306 uint8_t *closure, void *task_parent);
307extern int halide_default_semaphore_init(struct halide_semaphore_t *, int n);
308extern int halide_default_semaphore_release(struct halide_semaphore_t *, int n);
310// @}
311
312struct halide_thread;
313
314/** Spawn a thread. Returns a handle to the thread for the purposes of
315 * joining it. The thread must be joined in order to clean up any
316 * resources associated with it. */
317extern struct halide_thread *halide_spawn_thread(void (*f)(void *), void *closure);
318
319/** Join a thread. */
320extern void halide_join_thread(struct halide_thread *);
321
322/** Set the number of threads used by Halide's thread pool. Returns
323 * the old number.
324 *
325 * n < 0 : error condition
326 * n == 0 : use a reasonable system default (typically, number of cpus online).
327 * n == 1 : use exactly one thread; this will always enforce serial execution
328 * n > 1 : use a pool of exactly n threads.
329 *
330 * (Note that this is only guaranteed when using the default implementations
331 * of halide_do_par_for(); custom implementations may completely ignore values
332 * passed to halide_set_num_threads().)
333 */
334extern int halide_set_num_threads(int n);
335
336/** Halide calls these functions to allocate and free memory. To
337 * replace in AOT code, use the halide_set_custom_malloc and
338 * halide_set_custom_free, or (on platforms that support weak
339 * linking), simply define these functions yourself. In JIT-compiled
340 * code use Func::set_custom_allocator.
341 *
342 * If you override them, and find yourself wanting to call the default
343 * implementation from within your override, use
344 * halide_default_malloc/free.
345 *
346 * Note that halide_malloc must return a pointer aligned to the
347 * maximum meaningful alignment for the platform for the purpose of
348 * vector loads and stores. The default implementation uses 32-byte
349 * alignment, which is safe for arm and x86. Additionally, it must be
350 * safe to read at least 8 bytes before the start and beyond the
351 * end.
352 */
353//@{
354extern void *halide_malloc(void *user_context, size_t x);
355extern void halide_free(void *user_context, void *ptr);
356extern void *halide_default_malloc(void *user_context, size_t x);
357extern void halide_default_free(void *user_context, void *ptr);
358typedef void *(*halide_malloc_t)(void *, size_t);
359typedef void (*halide_free_t)(void *, void *);
362//@}
363
364/** Halide calls these functions to interact with the underlying
365 * system runtime functions. To replace in AOT code on platforms that
366 * support weak linking, define these functions yourself, or use
367 * the halide_set_custom_load_library() and halide_set_custom_get_library_symbol()
368 * functions. In JIT-compiled code, use JITSharedRuntime::set_default_handlers().
369 *
370 * halide_load_library and halide_get_library_symbol are equivalent to
371 * dlopen and dlsym. halide_get_symbol(sym) is equivalent to
372 * dlsym(RTLD_DEFAULT, sym).
373 */
374//@{
375extern void *halide_get_symbol(const char *name);
376extern void *halide_load_library(const char *name);
377extern void *halide_get_library_symbol(void *lib, const char *name);
378extern void *halide_default_get_symbol(const char *name);
379extern void *halide_default_load_library(const char *name);
380extern void *halide_default_get_library_symbol(void *lib, const char *name);
381typedef void *(*halide_get_symbol_t)(const char *name);
382typedef void *(*halide_load_library_t)(const char *name);
383typedef void *(*halide_get_library_symbol_t)(void *lib, const char *name);
387//@}
388
389/** Called when debug_to_file is used inside %Halide code. See
390 * Func::debug_to_file for how this is called
391 *
392 * Cannot be replaced in JITted code at present.
393 */
394extern int32_t halide_debug_to_file(void *user_context, const char *filename,
395 int32_t type_code,
396 struct halide_buffer_t *buf);
397
398/** Types in the halide type system. They can be ints, unsigned ints,
399 * or floats (of various bit-widths), or a handle (which is always 64-bits).
400 * Note that the int/uint/float values do not imply a specific bit width
401 * (the bit width is expected to be encoded in a separate value).
402 */
404#if (__cplusplus >= 201103L || _MSVC_LANG >= 201103L)
405 : uint8_t
406#endif
407{
408 halide_type_int = 0, ///< signed integers
409 halide_type_uint = 1, ///< unsigned integers
410 halide_type_float = 2, ///< IEEE floating point numbers
411 halide_type_handle = 3, ///< opaque pointer type (void *)
412 halide_type_bfloat = 4, ///< floating point numbers in the bfloat format
414
415// Note that while __attribute__ can go before or after the declaration,
416// __declspec apparently is only allowed before.
417#ifndef HALIDE_ATTRIBUTE_ALIGN
418#ifdef _MSC_VER
419#define HALIDE_ATTRIBUTE_ALIGN(x) __declspec(align(x))
420#else
421#define HALIDE_ATTRIBUTE_ALIGN(x) __attribute__((aligned(x)))
422#endif
423#endif
424
425/** A runtime tag for a type in the halide type system. Can be ints,
426 * unsigned ints, or floats of various bit-widths (the 'bits'
427 * field). Can also be vectors of the same (by setting the 'lanes'
428 * field to something larger than one). This struct should be
429 * exactly 32-bits in size. */
431 /** The basic type code: signed integer, unsigned integer, or floating point. */
432#if (__cplusplus >= 201103L || _MSVC_LANG >= 201103L)
434 halide_type_code_t code; // halide_type_code_t
435#else
437 uint8_t code; // halide_type_code_t
438#endif
439
440 /** The number of bits of precision of a single scalar value of this type. */
443
444 /** How many elements in a vector. This is 1 for scalar types. */
447
448#if (__cplusplus >= 201103L || _MSVC_LANG >= 201103L)
449 /** Construct a runtime representation of a Halide type from:
450 * code: The fundamental type from an enum.
451 * bits: The bit size of one element.
452 * lanes: The number of vector elements in the type. */
454 : code(code), bits(bits), lanes(lanes) {
455 }
456
457 /** Default constructor is required e.g. to declare halide_trace_event
458 * instances. */
460 : code((halide_type_code_t)0), bits(0), lanes(0) {
461 }
462
463 HALIDE_ALWAYS_INLINE constexpr halide_type_t with_lanes(uint16_t new_lanes) const {
464 return halide_type_t((halide_type_code_t)code, bits, new_lanes);
465 }
466
467 HALIDE_ALWAYS_INLINE constexpr halide_type_t element_of() const {
468 return with_lanes(1);
469 }
470 /** Compare two types for equality. */
471 HALIDE_ALWAYS_INLINE constexpr bool operator==(const halide_type_t &other) const {
472 return as_u32() == other.as_u32();
473 }
474
475 HALIDE_ALWAYS_INLINE constexpr bool operator!=(const halide_type_t &other) const {
476 return !(*this == other);
477 }
478
479 HALIDE_ALWAYS_INLINE constexpr bool operator<(const halide_type_t &other) const {
480 return as_u32() < other.as_u32();
481 }
482
483 /** Size in bytes for a single element, even if width is not 1, of this type. */
484 HALIDE_ALWAYS_INLINE constexpr int bytes() const {
485 return (bits + 7) / 8;
486 }
487
488 HALIDE_ALWAYS_INLINE constexpr uint32_t as_u32() const {
489 // Note that this produces a result that is identical to memcpy'ing 'this'
490 // into a u32 (on a little-endian machine, anyway), and at -O1 or greater
491 // on Clang, the compiler knows this and optimizes this into a single 32-bit move.
492 // (At -O0 it will look awful.)
493 return static_cast<uint8_t>(code) |
494 (static_cast<uint16_t>(bits) << 8) |
495 (static_cast<uint32_t>(lanes) << 16);
496 }
497#endif
498};
499
500#if (__cplusplus >= 201103L || _MSVC_LANG >= 201103L)
501static_assert(sizeof(halide_type_t) == sizeof(uint32_t), "size mismatch in halide_type_t");
502#endif
503
515
517 /** The name of the Func or Pipeline that this event refers to */
518 const char *func;
519
520 /** If the event type is a load or a store, this points to the
521 * value being loaded or stored. Use the type field to safely cast
522 * this to a concrete pointer type and retrieve it. For other
523 * events this is null. */
524 void *value;
525
526 /** For loads and stores, an array which contains the location
527 * being accessed. For vector loads or stores it is an array of
528 * vectors of coordinates (the vector dimension is innermost).
529 *
530 * For realization or production-related events, this will contain
531 * the mins and extents of the region being accessed, in the order
532 * min0, extent0, min1, extent1, ...
533 *
534 * For pipeline-related events, this will be null.
535 */
537
538 /** For halide_trace_tag, this points to a read-only null-terminated string
539 * of arbitrary text. For all other events, this will be null.
540 */
541 const char *trace_tag;
542
543 /** If the event type is a load or a store, this is the type of
544 * the data. Otherwise, the value is meaningless. */
546
547 /** The type of event */
549
550 /* The ID of the parent event (see below for an explanation of
551 * event ancestry). */
553
554 /** If this was a load or store of a Tuple-valued Func, this is
555 * which tuple element was accessed. */
557
558 /** The length of the coordinates array */
560};
561
562/** Called when Funcs are marked as trace_load, trace_store, or
563 * trace_realization. See Func::set_custom_trace. The default
564 * implementation either prints events via halide_print, or if
565 * HL_TRACE_FILE is defined, dumps the trace to that file in a
566 * sequence of trace packets. The header for a trace packet is defined
567 * below. If the trace is going to be large, you may want to make the
568 * file a named pipe, and then read from that pipe into gzip.
569 *
570 * halide_trace returns a unique ID which will be passed to future
571 * events that "belong" to the earlier event as the parent id. The
572 * ownership hierarchy looks like:
573 *
574 * begin_pipeline
575 * +--trace_tag (if any)
576 * +--trace_tag (if any)
577 * ...
578 * +--begin_realization
579 * | +--produce
580 * | | +--load/store
581 * | | +--end_produce
582 * | +--consume
583 * | | +--load
584 * | | +--end_consume
585 * | +--end_realization
586 * +--end_pipeline
587 *
588 * Threading means that ownership cannot be inferred from the ordering
589 * of events. There can be many active realizations of a given
590 * function, or many active productions for a single
591 * realization. Within a single production, the ordering of events is
592 * meaningful.
593 *
594 * Note that all trace_tag events (if any) will occur just after the begin_pipeline
595 * event, but before any begin_realization events. All trace_tags for a given Func
596 * will be emitted in the order added.
597 */
598// @}
599extern int32_t halide_trace(void *user_context, const struct halide_trace_event_t *event);
600extern int32_t halide_default_trace(void *user_context, const struct halide_trace_event_t *event);
601typedef int32_t (*halide_trace_t)(void *user_context, const struct halide_trace_event_t *);
603// @}
604
605/** The header of a packet in a binary trace. All fields are 32-bit. */
607 /** The total size of this packet in bytes. Always a multiple of
608 * four. Equivalently, the number of bytes until the next
609 * packet. */
611
612 /** The id of this packet (for the purpose of parent_id). */
614
615 /** The remaining fields are equivalent to those in halide_trace_event_t */
616 // @{
622 // @}
623
624#if (__cplusplus >= 201103L || _MSVC_LANG >= 201103L)
625 /** Get the coordinates array, assuming this packet is laid out in
626 * memory as it was written. The coordinates array comes
627 * immediately after the packet header. */
628 HALIDE_ALWAYS_INLINE const int *coordinates() const {
629 return (const int *)(this + 1);
630 }
631
632 HALIDE_ALWAYS_INLINE int *coordinates() {
633 return (int *)(this + 1);
634 }
635
636 /** Get the value, assuming this packet is laid out in memory as
637 * it was written. The packet comes immediately after the coordinates
638 * array. */
639 HALIDE_ALWAYS_INLINE const void *value() const {
640 return (const void *)(coordinates() + dimensions);
641 }
642
643 HALIDE_ALWAYS_INLINE void *value() {
644 return (void *)(coordinates() + dimensions);
645 }
646
647 /** Get the func name, assuming this packet is laid out in memory
648 * as it was written. It comes after the value. */
649 HALIDE_ALWAYS_INLINE const char *func() const {
650 return (const char *)value() + type.lanes * type.bytes();
651 }
652
653 HALIDE_ALWAYS_INLINE char *func() {
654 return (char *)value() + type.lanes * type.bytes();
655 }
656
657 /** Get the trace_tag (if any), assuming this packet is laid out in memory
658 * as it was written. It comes after the func name. If there is no trace_tag,
659 * this will return a pointer to an empty string. */
660 HALIDE_ALWAYS_INLINE const char *trace_tag() const {
661 const char *f = func();
662 // strlen may not be available here
663 while (*f++) {
664 // nothing
665 }
666 return f;
667 }
668
669 HALIDE_ALWAYS_INLINE char *trace_tag() {
670 char *f = func();
671 // strlen may not be available here
672 while (*f++) {
673 // nothing
674 }
675 return f;
676 }
677#endif
678};
679
680/** Set the file descriptor that Halide should write binary trace
681 * events to. If called with 0 as the argument, Halide outputs trace
682 * information to stdout in a human-readable format. If never called,
683 * Halide checks the for existence of an environment variable called
684 * HL_TRACE_FILE and opens that file. If HL_TRACE_FILE is not defined,
685 * it outputs trace information to stdout in a human-readable
686 * format. */
687extern void halide_set_trace_file(int fd);
688
689/** Halide calls this to retrieve the file descriptor to write binary
690 * trace events to. The default implementation returns the value set
691 * by halide_set_trace_file. Implement it yourself if you wish to use
692 * a custom file descriptor per user_context. Return zero from your
693 * implementation to tell Halide to print human-readable trace
694 * information to stdout. */
695extern int halide_get_trace_file(void *user_context);
696
697/** If tracing is writing to a file. This call closes that file
698 * (flushing the trace). Returns zero on success. */
700
701/** All Halide GPU or device backend implementations provide an
702 * interface to be used with halide_device_malloc, etc. This is
703 * accessed via the functions below.
704 */
705
706/** An opaque struct containing per-GPU API implementations of the
707 * device functions. */
709
710/** Each GPU API provides a halide_device_interface_t struct pointing
711 * to the code that manages device allocations. You can access these
712 * functions directly from the struct member function pointers, or by
713 * calling the functions declared below. Note that the global
714 * functions are not available when using Halide as a JIT compiler.
715 * If you are using raw halide_buffer_t in that context you must use
716 * the function pointers in the device_interface struct.
717 *
718 * The function pointers below are currently the same for every GPU
719 * API; only the impl field varies. These top-level functions do the
720 * bookkeeping that is common across all GPU APIs, and then dispatch
721 * to more API-specific functions via another set of function pointers
722 * hidden inside the impl field.
723 */
725 int (*device_malloc)(void *user_context, struct halide_buffer_t *buf,
726 const struct halide_device_interface_t *device_interface);
727 int (*device_free)(void *user_context, struct halide_buffer_t *buf);
728 int (*device_sync)(void *user_context, struct halide_buffer_t *buf);
729 void (*device_release)(void *user_context,
730 const struct halide_device_interface_t *device_interface);
731 int (*copy_to_host)(void *user_context, struct halide_buffer_t *buf);
732 int (*copy_to_device)(void *user_context, struct halide_buffer_t *buf,
733 const struct halide_device_interface_t *device_interface);
734 int (*device_and_host_malloc)(void *user_context, struct halide_buffer_t *buf,
735 const struct halide_device_interface_t *device_interface);
736 int (*device_and_host_free)(void *user_context, struct halide_buffer_t *buf);
737 int (*buffer_copy)(void *user_context, struct halide_buffer_t *src,
738 const struct halide_device_interface_t *dst_device_interface, struct halide_buffer_t *dst);
739 int (*device_crop)(void *user_context, const struct halide_buffer_t *src,
740 struct halide_buffer_t *dst);
741 int (*device_slice)(void *user_context, const struct halide_buffer_t *src,
742 int slice_dim, int slice_pos, struct halide_buffer_t *dst);
743 int (*device_release_crop)(void *user_context, struct halide_buffer_t *buf);
744 int (*wrap_native)(void *user_context, struct halide_buffer_t *buf, uint64_t handle,
745 const struct halide_device_interface_t *device_interface);
746 int (*detach_native)(void *user_context, struct halide_buffer_t *buf);
747 int (*compute_capability)(void *user_context, int *major, int *minor);
749};
750
751/** Release all data associated with the given device interface, in
752 * particular all resources (memory, texture, context handles)
753 * allocated by Halide. Must be called explicitly when using AOT
754 * compilation. This is *not* thread-safe with respect to actively
755 * running Halide code. Ensure all pipelines are finished before
756 * calling this. */
757extern void halide_device_release(void *user_context,
758 const struct halide_device_interface_t *device_interface);
759
760/** Copy image data from device memory to host memory. This must be called
761 * explicitly to copy back the results of a GPU-based filter. */
762extern int halide_copy_to_host(void *user_context, struct halide_buffer_t *buf);
763
764/** Copy image data from host memory to device memory. This should not
765 * be called directly; Halide handles copying to the device
766 * automatically. If interface is NULL and the buf has a non-zero dev
767 * field, the device associated with the dev handle will be
768 * used. Otherwise if the dev field is 0 and interface is NULL, an
769 * error is returned. */
770extern int halide_copy_to_device(void *user_context, struct halide_buffer_t *buf,
771 const struct halide_device_interface_t *device_interface);
772
773/** Copy data from one buffer to another. The buffers may have
774 * different shapes and sizes, but the destination buffer's shape must
775 * be contained within the source buffer's shape. That is, for each
776 * dimension, the min on the destination buffer must be greater than
777 * or equal to the min on the source buffer, and min+extent on the
778 * destination buffer must be less that or equal to min+extent on the
779 * source buffer. The source data is pulled from either device or
780 * host memory on the source, depending on the dirty flags. host is
781 * preferred if both are valid. The dst_device_interface parameter
782 * controls the destination memory space. NULL means host memory. */
783extern int halide_buffer_copy(void *user_context, struct halide_buffer_t *src,
784 const struct halide_device_interface_t *dst_device_interface,
785 struct halide_buffer_t *dst);
786
787/** Give the destination buffer a device allocation which is an alias
788 * for the same coordinate range in the source buffer. Modifies the
789 * device, device_interface, and the device_dirty flag only. Only
790 * supported by some device APIs (others will return
791 * halide_error_code_device_crop_unsupported). Call
792 * halide_device_release_crop instead of halide_device_free to clean
793 * up resources associated with the cropped view. Do not free the
794 * device allocation on the source buffer while the destination buffer
795 * still lives. Note that the two buffers do not share dirty flags, so
796 * care must be taken to update them together as needed. Note that src
797 * and dst are required to have the same number of dimensions.
798 *
799 * Note also that (in theory) device interfaces which support cropping may
800 * still not support cropping a crop (instead, create a new crop of the parent
801 * buffer); in practice, no known implementation has this limitation, although
802 * it is possible that some future implementations may require it. */
803extern int halide_device_crop(void *user_context,
804 const struct halide_buffer_t *src,
805 struct halide_buffer_t *dst);
806
807/** Give the destination buffer a device allocation which is an alias
808 * for a similar coordinate range in the source buffer, but with one dimension
809 * sliced away in the dst. Modifies the device, device_interface, and the
810 * device_dirty flag only. Only supported by some device APIs (others will return
811 * halide_error_code_device_crop_unsupported). Call
812 * halide_device_release_crop instead of halide_device_free to clean
813 * up resources associated with the sliced view. Do not free the
814 * device allocation on the source buffer while the destination buffer
815 * still lives. Note that the two buffers do not share dirty flags, so
816 * care must be taken to update them together as needed. Note that the dst buffer
817 * must have exactly one fewer dimension than the src buffer, and that slice_dim
818 * and slice_pos must be valid within src. */
819extern int halide_device_slice(void *user_context,
820 const struct halide_buffer_t *src,
821 int slice_dim, int slice_pos,
822 struct halide_buffer_t *dst);
823
824/** Release any resources associated with a cropped/sliced view of another
825 * buffer. */
826extern int halide_device_release_crop(void *user_context,
827 struct halide_buffer_t *buf);
828
829/** Wait for current GPU operations to complete. Calling this explicitly
830 * should rarely be necessary, except maybe for profiling. */
831extern int halide_device_sync(void *user_context, struct halide_buffer_t *buf);
832
833/** Allocate device memory to back a halide_buffer_t. */
834extern int halide_device_malloc(void *user_context, struct halide_buffer_t *buf,
835 const struct halide_device_interface_t *device_interface);
836
837/** Free device memory. */
838extern int halide_device_free(void *user_context, struct halide_buffer_t *buf);
839
840/** Wrap or detach a native device handle, setting the device field
841 * and device_interface field as appropriate for the given GPU
842 * API. The meaning of the opaque handle is specific to the device
843 * interface, so if you know the device interface in use, call the
844 * more specific functions in the runtime headers for your specific
845 * device API instead (e.g. HalideRuntimeCuda.h). */
846// @{
847extern int halide_device_wrap_native(void *user_context,
848 struct halide_buffer_t *buf,
849 uint64_t handle,
850 const struct halide_device_interface_t *device_interface);
851extern int halide_device_detach_native(void *user_context, struct halide_buffer_t *buf);
852// @}
853
854/** Selects which gpu device to use. 0 is usually the display
855 * device. If never called, Halide uses the environment variable
856 * HL_GPU_DEVICE. If that variable is unset, Halide uses the last
857 * device. Set this to -1 to use the last device. */
858extern void halide_set_gpu_device(int n);
859
860/** Halide calls this to get the desired halide gpu device
861 * setting. Implement this yourself to use a different gpu device per
862 * user_context. The default implementation returns the value set by
863 * halide_set_gpu_device, or the environment variable
864 * HL_GPU_DEVICE. */
865extern int halide_get_gpu_device(void *user_context);
866
867/** Set the soft maximum amount of memory, in bytes, that the LRU
868 * cache will use to memoize Func results. This is not a strict
869 * maximum in that concurrency and simultaneous use of memoized
870 * reults larger than the cache size can both cause it to
871 * temporariliy be larger than the size specified here.
872 */
874
875/** Given a cache key for a memoized result, currently constructed
876 * from the Func name and top-level Func name plus the arguments of
877 * the computation, determine if the result is in the cache and
878 * return it if so. (The internals of the cache key should be
879 * considered opaque by this function.) If this routine returns true,
880 * it is a cache miss. Otherwise, it will return false and the
881 * buffers passed in will be filled, via copying, with memoized
882 * data. The last argument is a list if halide_buffer_t pointers which
883 * represents the outputs of the memoized Func. If the Func does not
884 * return a Tuple, there will only be one halide_buffer_t in the list. The
885 * tuple_count parameters determines the length of the list.
886 *
887 * The return values are:
888 * -1: Signals an error.
889 * 0: Success and cache hit.
890 * 1: Success and cache miss.
891 */
892extern int halide_memoization_cache_lookup(void *user_context, const uint8_t *cache_key, int32_t size,
893 struct halide_buffer_t *realized_bounds,
894 int32_t tuple_count, struct halide_buffer_t **tuple_buffers);
895
896/** Given a cache key for a memoized result, currently constructed
897 * from the Func name and top-level Func name plus the arguments of
898 * the computation, store the result in the cache for futre access by
899 * halide_memoization_cache_lookup. (The internals of the cache key
900 * should be considered opaque by this function.) Data is copied out
901 * from the inputs and inputs are unmodified. The last argument is a
902 * list if halide_buffer_t pointers which represents the outputs of the
903 * memoized Func. If the Func does not return a Tuple, there will
904 * only be one halide_buffer_t in the list. The tuple_count parameters
905 * determines the length of the list.
906 *
907 * If there is a memory allocation failure, the store does not store
908 * the data into the cache.
909 *
910 * If has_eviction_key is true, the entry is marked with eviction_key to
911 * allow removing the key with halide_memoization_cache_evict.
912 */
913extern int halide_memoization_cache_store(void *user_context, const uint8_t *cache_key, int32_t size,
914 struct halide_buffer_t *realized_bounds,
915 int32_t tuple_count,
916 struct halide_buffer_t **tuple_buffers,
917 bool has_eviction_key, uint64_t eviction_key);
918
919/** Evict all cache entries that were tagged with the given
920 * eviction_key in the memoize scheduling directive.
921 */
922extern void halide_memoization_cache_evict(void *user_context, uint64_t eviction_key);
923
924/** If halide_memoization_cache_lookup succeeds,
925 * halide_memoization_cache_release must be called to signal the
926 * storage is no longer being used by the caller. It will be passed
927 * the host pointer of one the buffers returned by
928 * halide_memoization_cache_lookup. That is
929 * halide_memoization_cache_release will be called multiple times for
930 * the case where halide_memoization_cache_lookup is handling multiple
931 * buffers. (This corresponds to memoizing a Tuple in Halide.) Note
932 * that the host pointer must be sufficient to get to all information
933 * the release operation needs. The default Halide cache impleemntation
934 * accomplishes this by storing extra data before the start of the user
935 * modifiable host storage.
936 *
937 * This call is like free and does not have a failure return.
938 */
939extern void halide_memoization_cache_release(void *user_context, void *host);
940
941/** Free all memory and resources associated with the memoization cache.
942 * Must be called at a time when no other threads are accessing the cache.
943 */
945
946/** Verify that a given range of memory has been initialized; only used when Target::MSAN is enabled.
947 *
948 * The default implementation simply calls the LLVM-provided __msan_check_mem_is_initialized() function.
949 *
950 * The return value should always be zero.
951 */
952extern int halide_msan_check_memory_is_initialized(void *user_context, const void *ptr, uint64_t len, const char *name);
953
954/** Verify that the data pointed to by the halide_buffer_t is initialized (but *not* the halide_buffer_t itself),
955 * using halide_msan_check_memory_is_initialized() for checking.
956 *
957 * The default implementation takes pains to only check the active memory ranges
958 * (skipping padding), and sorting into ranges to always check the smallest number of
959 * ranges, in monotonically increasing memory order.
960 *
961 * Most client code should never need to replace the default implementation.
962 *
963 * The return value should always be zero.
964 */
965extern int halide_msan_check_buffer_is_initialized(void *user_context, struct halide_buffer_t *buffer, const char *buf_name);
966
967/** Annotate that a given range of memory has been initialized;
968 * only used when Target::MSAN is enabled.
969 *
970 * The default implementation simply calls the LLVM-provided __msan_unpoison() function.
971 *
972 * The return value should always be zero.
973 */
974extern int halide_msan_annotate_memory_is_initialized(void *user_context, const void *ptr, uint64_t len);
975
976/** Mark the data pointed to by the halide_buffer_t as initialized (but *not* the halide_buffer_t itself),
977 * using halide_msan_annotate_memory_is_initialized() for marking.
978 *
979 * The default implementation takes pains to only mark the active memory ranges
980 * (skipping padding), and sorting into ranges to always mark the smallest number of
981 * ranges, in monotonically increasing memory order.
982 *
983 * Most client code should never need to replace the default implementation.
984 *
985 * The return value should always be zero.
986 */
987extern int halide_msan_annotate_buffer_is_initialized(void *user_context, struct halide_buffer_t *buffer);
988extern void halide_msan_annotate_buffer_is_initialized_as_destructor(void *user_context, void *buffer);
989
990/** The error codes that may be returned by a Halide pipeline. */
992 /** There was no error. This is the value returned by Halide on success. */
994
995 /** An uncategorized error occurred. Refer to the string passed to halide_error. */
997
998 /** A Func was given an explicit bound via Func::bound, but this
999 * was not large enough to encompass the region that is used of
1000 * the Func by the rest of the pipeline. */
1002
1003 /** The elem_size field of a halide_buffer_t does not match the size in
1004 * bytes of the type of that ImageParam. Probable type mismatch. */
1006
1007 /** A pipeline would access memory outside of the halide_buffer_t passed
1008 * in. */
1010
1011 /** A halide_buffer_t was given that spans more than 2GB of memory. */
1013
1014 /** A halide_buffer_t was given with extents that multiply to a number
1015 * greater than 2^31-1 */
1017
1018 /** Applying explicit constraints on the size of an input or
1019 * output buffer shrank the size of that buffer below what will be
1020 * accessed by the pipeline. */
1022
1023 /** A constraint on a size or stride of an input or output buffer
1024 * was not met by the halide_buffer_t passed in. */
1026
1027 /** A scalar parameter passed in was smaller than its minimum
1028 * declared value. */
1030
1031 /** A scalar parameter passed in was greater than its minimum
1032 * declared value. */
1034
1035 /** A call to halide_malloc returned NULL. */
1037
1038 /** A halide_buffer_t pointer passed in was NULL. */
1040
1041 /** debug_to_file failed to open or write to the specified
1042 * file. */
1044
1045 /** The Halide runtime encountered an error while trying to copy
1046 * from device to host. Turn on -debug in your target string to
1047 * see more details. */
1049
1050 /** The Halide runtime encountered an error while trying to copy
1051 * from host to device. Turn on -debug in your target string to
1052 * see more details. */
1054
1055 /** The Halide runtime encountered an error while trying to
1056 * allocate memory on device. Turn on -debug in your target string
1057 * to see more details. */
1059
1060 /** The Halide runtime encountered an error while trying to
1061 * synchronize with a device. Turn on -debug in your target string
1062 * to see more details. */
1064
1065 /** The Halide runtime encountered an error while trying to free a
1066 * device allocation. Turn on -debug in your target string to see
1067 * more details. */
1069
1070 /** Buffer has a non-zero device but no device interface, which
1071 * violates a Halide invariant. */
1073
1074 /** An error occurred when attempting to initialize the Matlab
1075 * runtime. */
1077
1078 /** The type of an mxArray did not match the expected type. */
1080
1081 /** There is a bug in the Halide compiler. */
1083
1084 /** The Halide runtime encountered an error while trying to launch
1085 * a GPU kernel. Turn on -debug in your target string to see more
1086 * details. */
1088
1089 /** The Halide runtime encountered a host pointer that violated
1090 * the alignment set for it by way of a call to
1091 * set_host_alignment */
1093
1094 /** A fold_storage directive was used on a dimension that is not
1095 * accessed in a monotonically increasing or decreasing fashion. */
1097
1098 /** A fold_storage directive was used with a fold factor that was
1099 * too small to store all the values of a producer needed by the
1100 * consumer. */
1102
1103 /** User-specified require() expression was not satisfied. */
1105
1106 /** At least one of the buffer's extents are negative. */
1108
1110
1112
1113 /** A specialize_fail() schedule branch was selected at runtime. */
1115
1116 /** The Halide runtime encountered an error while trying to wrap a
1117 * native device handle. Turn on -debug in your target string to
1118 * see more details. */
1120
1121 /** The Halide runtime encountered an error while trying to detach
1122 * a native device handle. Turn on -debug in your target string
1123 * to see more details. */
1125
1126 /** The host field on an input or output was null, the device
1127 * field was not zero, and the pipeline tries to use the buffer on
1128 * the host. You may be passing a GPU-only buffer to a pipeline
1129 * which is scheduled to use it on the CPU. */
1131
1132 /** A folded buffer was passed to an extern stage, but the region
1133 * touched wraps around the fold boundary. */
1135
1136 /** Buffer has a non-null device_interface but device is 0, which
1137 * violates a Halide invariant. */
1139
1140 /** Buffer has both host and device dirty bits set, which violates
1141 * a Halide invariant. */
1143
1144 /** The halide_buffer_t * passed to a halide runtime routine is
1145 * nullptr and this is not allowed. */
1147
1148 /** The Halide runtime encountered an error while trying to copy
1149 * from one buffer to another. Turn on -debug in your target
1150 * string to see more details. */
1152
1153 /** Attempted to make cropped/sliced alias of a buffer with a device
1154 * field, but the device_interface does not support cropping. */
1156
1157 /** Cropping/slicing a buffer failed for some other reason. Turn on -debug
1158 * in your target string. */
1160
1161 /** An operation on a buffer required an allocation on a
1162 * particular device interface, but a device allocation already
1163 * existed on a different device interface. Free the old one
1164 * first. */
1166
1167 /** The dimensions field of a halide_buffer_t does not match the dimensions of that ImageParam. */
1169
1170 /** A buffer with the device_dirty flag set was passed to a
1171 * pipeline compiled with no device backends enabled, so it
1172 * doesn't know how to copy the data back from device memory to
1173 * host memory. Either call copy_to_host before calling the Halide
1174 * pipeline, or enable the appropriate device backend. */
1176
1177 /** An explicit storage bound provided is too small to store
1178 * all the values produced by the function. */
1180};
1181
1182/** Halide calls the functions below on various error conditions. The
1183 * default implementations construct an error message, call
1184 * halide_error, then return the matching error code above. On
1185 * platforms that support weak linking, you can override these to
1186 * catch the errors individually. */
1187
1188/** A call into an extern stage for the purposes of bounds inference
1189 * failed. Returns the error code given by the extern stage. */
1190extern int halide_error_bounds_inference_call_failed(void *user_context, const char *extern_stage_name, int result);
1191
1192/** A call to an extern stage failed. Returned the error code given by
1193 * the extern stage. */
1194extern int halide_error_extern_stage_failed(void *user_context, const char *extern_stage_name, int result);
1195
1196/** Various other error conditions. See the enum above for a
1197 * description of each. */
1198// @{
1199extern int halide_error_explicit_bounds_too_small(void *user_context, const char *func_name, const char *var_name,
1200 int min_bound, int max_bound, int min_required, int max_required);
1201extern int halide_error_bad_type(void *user_context, const char *func_name,
1202 uint32_t type_given, uint32_t correct_type); // N.B. The last two args are the bit representation of a halide_type_t
1203extern int halide_error_bad_dimensions(void *user_context, const char *func_name,
1204 int32_t dimensions_given, int32_t correct_dimensions);
1205extern int halide_error_access_out_of_bounds(void *user_context, const char *func_name,
1206 int dimension, int min_touched, int max_touched,
1207 int min_valid, int max_valid);
1208extern int halide_error_buffer_allocation_too_large(void *user_context, const char *buffer_name,
1209 uint64_t allocation_size, uint64_t max_size);
1210extern int halide_error_buffer_extents_negative(void *user_context, const char *buffer_name, int dimension, int extent);
1211extern int halide_error_buffer_extents_too_large(void *user_context, const char *buffer_name,
1212 int64_t actual_size, int64_t max_size);
1213extern int halide_error_constraints_make_required_region_smaller(void *user_context, const char *buffer_name,
1214 int dimension,
1215 int constrained_min, int constrained_extent,
1216 int required_min, int required_extent);
1217extern int halide_error_constraint_violated(void *user_context, const char *var, int val,
1218 const char *constrained_var, int constrained_val);
1219extern int halide_error_param_too_small_i64(void *user_context, const char *param_name,
1220 int64_t val, int64_t min_val);
1221extern int halide_error_param_too_small_u64(void *user_context, const char *param_name,
1222 uint64_t val, uint64_t min_val);
1223extern int halide_error_param_too_small_f64(void *user_context, const char *param_name,
1224 double val, double min_val);
1225extern int halide_error_param_too_large_i64(void *user_context, const char *param_name,
1226 int64_t val, int64_t max_val);
1227extern int halide_error_param_too_large_u64(void *user_context, const char *param_name,
1228 uint64_t val, uint64_t max_val);
1229extern int halide_error_param_too_large_f64(void *user_context, const char *param_name,
1230 double val, double max_val);
1231extern int halide_error_out_of_memory(void *user_context);
1232extern int halide_error_buffer_argument_is_null(void *user_context, const char *buffer_name);
1233extern int halide_error_debug_to_file_failed(void *user_context, const char *func,
1234 const char *filename, int error_code);
1235extern int halide_error_unaligned_host_ptr(void *user_context, const char *func_name, int alignment);
1236extern int halide_error_host_is_null(void *user_context, const char *func_name);
1237extern int halide_error_bad_fold(void *user_context, const char *func_name, const char *var_name,
1238 const char *loop_name);
1239extern int halide_error_bad_extern_fold(void *user_context, const char *func_name,
1240 int dim, int min, int extent, int valid_min, int fold_factor);
1241
1242extern int halide_error_fold_factor_too_small(void *user_context, const char *func_name, const char *var_name,
1243 int fold_factor, const char *loop_name, int required_extent);
1244extern int halide_error_requirement_failed(void *user_context, const char *condition, const char *message);
1245extern int halide_error_specialize_fail(void *user_context, const char *message);
1246extern int halide_error_no_device_interface(void *user_context);
1247extern int halide_error_device_interface_no_device(void *user_context);
1248extern int halide_error_host_and_device_dirty(void *user_context);
1249extern int halide_error_buffer_is_null(void *user_context, const char *routine);
1250extern int halide_error_device_dirty_with_no_device_support(void *user_context, const char *buffer_name);
1251extern int halide_error_storage_bound_too_small(void *user_context, const char *func_name, const char *var_name,
1252 int provided_size, int required_size);
1253extern int halide_error_device_crop_failed(void *user_context);
1254// @}
1255
1256/** Optional features a compilation Target can have.
1257 * Be sure to keep this in sync with the Feature enum in Target.h and the implementation of
1258 * get_runtime_compatible_target in Target.cpp if you add a new feature.
1259 */
1261 halide_target_feature_jit = 0, ///< Generate code that will run immediately inside the calling process.
1262 halide_target_feature_debug, ///< Turn on debug info and output for runtime code.
1263 halide_target_feature_no_asserts, ///< Disable all runtime checks, for slightly tighter code.
1264 halide_target_feature_no_bounds_query, ///< Disable the bounds querying functionality.
1265
1266 halide_target_feature_sse41, ///< Use SSE 4.1 and earlier instructions. Only relevant on x86.
1267 halide_target_feature_avx, ///< Use AVX 1 instructions. Only relevant on x86.
1268 halide_target_feature_avx2, ///< Use AVX 2 instructions. Only relevant on x86.
1269 halide_target_feature_fma, ///< Enable x86 FMA instruction
1270 halide_target_feature_fma4, ///< Enable x86 (AMD) FMA4 instruction set
1271 halide_target_feature_f16c, ///< Enable x86 16-bit float support
1272
1273 halide_target_feature_armv7s, ///< Generate code for ARMv7s. Only relevant for 32-bit ARM.
1274 halide_target_feature_no_neon, ///< Avoid using NEON instructions. Only relevant for 32-bit ARM.
1275
1276 halide_target_feature_vsx, ///< Use VSX instructions. Only relevant on POWERPC.
1277 halide_target_feature_power_arch_2_07, ///< Use POWER ISA 2.07 new instructions. Only relevant on POWERPC.
1278
1279 halide_target_feature_cuda, ///< Enable the CUDA runtime. Defaults to compute capability 2.0 (Fermi)
1280 halide_target_feature_cuda_capability30, ///< Enable CUDA compute capability 3.0 (Kepler)
1281 halide_target_feature_cuda_capability32, ///< Enable CUDA compute capability 3.2 (Tegra K1)
1282 halide_target_feature_cuda_capability35, ///< Enable CUDA compute capability 3.5 (Kepler)
1283 halide_target_feature_cuda_capability50, ///< Enable CUDA compute capability 5.0 (Maxwell)
1284 halide_target_feature_cuda_capability61, ///< Enable CUDA compute capability 6.1 (Pascal)
1285 halide_target_feature_cuda_capability70, ///< Enable CUDA compute capability 7.0 (Volta)
1286 halide_target_feature_cuda_capability75, ///< Enable CUDA compute capability 7.5 (Turing)
1287 halide_target_feature_cuda_capability80, ///< Enable CUDA compute capability 8.0 (Ampere)
1288 halide_target_feature_cuda_capability86, ///< Enable CUDA compute capability 8.6 (Ampere)
1289
1290 halide_target_feature_opencl, ///< Enable the OpenCL runtime.
1291 halide_target_feature_cl_doubles, ///< Enable double support on OpenCL targets
1292 halide_target_feature_cl_atomic64, ///< Enable 64-bit atomics operations on OpenCL targets
1293
1294 halide_target_feature_openglcompute, ///< Enable OpenGL Compute runtime.
1295
1296 halide_target_feature_user_context, ///< Generated code takes a user_context pointer as first argument
1297
1298 halide_target_feature_matlab, ///< Generate a mexFunction compatible with Matlab mex libraries. See tools/mex_halide.m.
1299
1300 halide_target_feature_profile, ///< Launch a sampling profiler alongside the Halide pipeline that monitors and reports the runtime used by each Func
1301 halide_target_feature_no_runtime, ///< Do not include a copy of the Halide runtime in any generated object file or assembly
1302
1303 halide_target_feature_metal, ///< Enable the (Apple) Metal runtime.
1304
1305 halide_target_feature_c_plus_plus_mangling, ///< Generate C++ mangled names for result function, et al
1306
1307 halide_target_feature_large_buffers, ///< Enable 64-bit buffer indexing to support buffers > 2GB. Ignored if bits != 64.
1308
1309 halide_target_feature_hvx_128, ///< Enable HVX 128 byte mode.
1310 halide_target_feature_hvx_v62, ///< Enable Hexagon v62 architecture.
1311 halide_target_feature_fuzz_float_stores, ///< On every floating point store, set the last bit of the mantissa to zero. Pipelines for which the output is very different with this feature enabled may also produce very different output on different processors.
1312 halide_target_feature_soft_float_abi, ///< Enable soft float ABI. This only enables the soft float ABI calling convention, which does not necessarily use soft floats.
1313 halide_target_feature_msan, ///< Enable hooks for MSAN support.
1314 halide_target_feature_avx512, ///< Enable the base AVX512 subset supported by all AVX512 architectures. The specific feature sets are AVX-512F and AVX512-CD. See https://en.wikipedia.org/wiki/AVX-512 for a description of each AVX subset.
1315 halide_target_feature_avx512_knl, ///< Enable the AVX512 features supported by Knight's Landing chips, such as the Xeon Phi x200. This includes the base AVX512 set, and also AVX512-CD and AVX512-ER.
1316 halide_target_feature_avx512_skylake, ///< Enable the AVX512 features supported by Skylake Xeon server processors. This adds AVX512-VL, AVX512-BW, and AVX512-DQ to the base set. The main difference from the base AVX512 set is better support for small integer ops. Note that this does not include the Knight's Landing features. Note also that these features are not available on Skylake desktop and mobile processors.
1317 halide_target_feature_avx512_cannonlake, ///< Enable the AVX512 features expected to be supported by future Cannonlake processors. This includes all of the Skylake features, plus AVX512-IFMA and AVX512-VBMI.
1318 halide_target_feature_avx512_sapphirerapids, ///< Enable the AVX512 features supported by Sapphire Rapids processors. This include all of the Cannonlake features, plus AVX512-VNNI and AVX512-BF16.
1320 halide_target_feature_trace_loads, ///< Trace all loads done by the pipeline. Equivalent to calling Func::trace_loads on every non-inlined Func.
1321 halide_target_feature_trace_stores, ///< Trace all stores done by the pipeline. Equivalent to calling Func::trace_stores on every non-inlined Func.
1322 halide_target_feature_trace_realizations, ///< Trace all realizations done by the pipeline. Equivalent to calling Func::trace_realizations on every non-inlined Func.
1324 halide_target_feature_hvx_v65, ///< Enable Hexagon v65 architecture.
1325 halide_target_feature_hvx_v66, ///< Enable Hexagon v66 architecture.
1326 halide_target_feature_cl_half, ///< Enable half support on OpenCL targets
1327 halide_target_feature_strict_float, ///< Turn off all non-IEEE floating-point optimization. Currently applies only to LLVM targets.
1328 halide_target_feature_tsan, ///< Enable hooks for TSAN support.
1329 halide_target_feature_asan, ///< Enable hooks for ASAN support.
1330 halide_target_feature_d3d12compute, ///< Enable Direct3D 12 Compute runtime.
1331 halide_target_feature_check_unsafe_promises, ///< Insert assertions for promises.
1332 halide_target_feature_hexagon_dma, ///< Enable Hexagon DMA buffers.
1333 halide_target_feature_embed_bitcode, ///< Emulate clang -fembed-bitcode flag.
1334 halide_target_feature_enable_llvm_loop_opt, ///< Enable loop vectorization + unrolling in LLVM. Overrides halide_target_feature_disable_llvm_loop_opt. (Ignored for non-LLVM targets.)
1335 halide_target_feature_disable_llvm_loop_opt, ///< Disable loop vectorization + unrolling in LLVM. (Ignored for non-LLVM targets.)
1336 halide_target_feature_wasm_simd128, ///< Enable +simd128 instructions for WebAssembly codegen.
1337 halide_target_feature_wasm_signext, ///< Enable +sign-ext instructions for WebAssembly codegen.
1338 halide_target_feature_wasm_sat_float_to_int, ///< Enable saturating (nontrapping) float-to-int instructions for WebAssembly codegen.
1339 halide_target_feature_wasm_threads, ///< Enable use of threads in WebAssembly codegen. Requires the use of a wasm runtime that provides pthread-compatible wrappers (typically, Emscripten with the -pthreads flag). Unsupported under WASI.
1340 halide_target_feature_wasm_bulk_memory, ///< Enable +bulk-memory instructions for WebAssembly codegen.
1341 halide_target_feature_sve, ///< Enable ARM Scalable Vector Extensions
1342 halide_target_feature_sve2, ///< Enable ARM Scalable Vector Extensions v2
1343 halide_target_feature_egl, ///< Force use of EGL support.
1344 halide_target_feature_arm_dot_prod, ///< Enable ARMv8.2-a dotprod extension (i.e. udot and sdot instructions)
1345 halide_target_feature_arm_fp16, ///< Enable ARMv8.2-a half-precision floating point data processing
1346 halide_llvm_large_code_model, ///< Use the LLVM large code model to compile
1347 halide_target_feature_rvv, ///< Enable RISCV "V" Vector Extension
1348 halide_target_feature_armv81a, ///< Enable ARMv8.1-a instructions
1349 halide_target_feature_sanitizer_coverage, ///< Enable hooks for SanitizerCoverage support.
1350 halide_target_feature_profile_by_timer, ///< Alternative to halide_target_feature_profile using timer interrupt for systems without threads or applicartions that need to avoid them.
1351 halide_target_feature_end ///< A sentinel. Every target is considered to have this feature, and setting this feature does nothing.
1353
1354/** This function is called internally by Halide in some situations to determine
1355 * if the current execution environment can support the given set of
1356 * halide_target_feature_t flags. The implementation must do the following:
1357 *
1358 * -- If there are flags set in features that the function knows *cannot* be supported, return 0.
1359 * -- Otherwise, return 1.
1360 * -- Note that any flags set in features that the function doesn't know how to test should be ignored;
1361 * this implies that a return value of 1 means "not known to be bad" rather than "known to be good".
1362 *
1363 * In other words: a return value of 0 means "It is not safe to use code compiled with these features",
1364 * while a return value of 1 means "It is not obviously unsafe to use code compiled with these features".
1365 *
1366 * The default implementation simply calls halide_default_can_use_target_features.
1367 *
1368 * Note that `features` points to an array of `count` uint64_t; this array must contain enough
1369 * bits to represent all the currently known features. Any excess bits must be set to zero.
1370 */
1371// @{
1372extern int halide_can_use_target_features(int count, const uint64_t *features);
1373typedef int (*halide_can_use_target_features_t)(int count, const uint64_t *features);
1375// @}
1376
1377/**
1378 * This is the default implementation of halide_can_use_target_features; it is provided
1379 * for convenience of user code that may wish to extend halide_can_use_target_features
1380 * but continue providing existing support, e.g.
1381 *
1382 * int halide_can_use_target_features(int count, const uint64_t *features) {
1383 * if (features[halide_target_somefeature >> 6] & (1LL << (halide_target_somefeature & 63))) {
1384 * if (!can_use_somefeature()) {
1385 * return 0;
1386 * }
1387 * }
1388 * return halide_default_can_use_target_features(count, features);
1389 * }
1390 */
1391extern int halide_default_can_use_target_features(int count, const uint64_t *features);
1392
1393typedef struct halide_dimension_t {
1394#if (__cplusplus >= 201103L || _MSVC_LANG >= 201103L)
1395 int32_t min = 0, extent = 0, stride = 0;
1396
1397 // Per-dimension flags. None are defined yet (This is reserved for future use).
1398 uint32_t flags = 0;
1399
1402 : min(m), extent(e), stride(s), flags(f) {
1403 }
1404
1405 HALIDE_ALWAYS_INLINE bool operator==(const halide_dimension_t &other) const {
1406 return (min == other.min) &&
1407 (extent == other.extent) &&
1408 (stride == other.stride) &&
1409 (flags == other.flags);
1410 }
1411
1412 HALIDE_ALWAYS_INLINE bool operator!=(const halide_dimension_t &other) const {
1413 return !(*this == other);
1414 }
1415#else
1417
1418 // Per-dimension flags. None are defined yet (This is reserved for future use).
1420#endif
1422
1423#ifdef __cplusplus
1424} // extern "C"
1425#endif
1426
1429
1430/**
1431 * The raw representation of an image passed around by generated
1432 * Halide code. It includes some stuff to track whether the image is
1433 * not actually in main memory, but instead on a device (like a
1434 * GPU). For a more convenient C++ wrapper, use Halide::Buffer<T>. */
1435typedef struct halide_buffer_t {
1436 /** A device-handle for e.g. GPU memory used to back this buffer. */
1438
1439 /** The interface used to interpret the above handle. */
1441
1442 /** A pointer to the start of the data in main memory. In terms of
1443 * the Halide coordinate system, this is the address of the min
1444 * coordinates (defined below). */
1446
1447 /** flags with various meanings. */
1449
1450 /** The type of each buffer element. */
1452
1453 /** The dimensionality of the buffer. */
1455
1456 /** The shape of the buffer. Halide does not own this array - you
1457 * must manage the memory for it yourself. */
1459
1460 /** Pads the buffer up to a multiple of 8 bytes */
1461 void *padding;
1462
1463#if (__cplusplus >= 201103L || _MSVC_LANG >= 201103L)
1464 /** Convenience methods for accessing the flags */
1465 // @{
1466 HALIDE_ALWAYS_INLINE bool get_flag(halide_buffer_flags flag) const {
1467 return (flags & flag) != 0;
1468 }
1469
1470 HALIDE_ALWAYS_INLINE void set_flag(halide_buffer_flags flag, bool value) {
1471 if (value) {
1472 flags |= flag;
1473 } else {
1474 flags &= ~uint64_t(flag);
1475 }
1476 }
1477
1478 HALIDE_ALWAYS_INLINE bool host_dirty() const {
1479 return get_flag(halide_buffer_flag_host_dirty);
1480 }
1481
1482 HALIDE_ALWAYS_INLINE bool device_dirty() const {
1483 return get_flag(halide_buffer_flag_device_dirty);
1484 }
1485
1486 HALIDE_ALWAYS_INLINE void set_host_dirty(bool v = true) {
1487 set_flag(halide_buffer_flag_host_dirty, v);
1488 }
1489
1490 HALIDE_ALWAYS_INLINE void set_device_dirty(bool v = true) {
1492 }
1493 // @}
1494
1495 /** The total number of elements this buffer represents. Equal to
1496 * the product of the extents */
1497 HALIDE_ALWAYS_INLINE size_t number_of_elements() const {
1498 size_t s = 1;
1499 for (int i = 0; i < dimensions; i++) {
1500 s *= dim[i].extent;
1501 }
1502 return s;
1503 }
1504
1505 /** Offset to the element with the lowest address.
1506 * If all strides are positive, equal to zero.
1507 * Offset is in elements, not bytes.
1508 * Unlike begin(), this is ok to call on an unallocated buffer. */
1509 HALIDE_ALWAYS_INLINE ptrdiff_t begin_offset() const {
1510 ptrdiff_t index = 0;
1511 for (int i = 0; i < dimensions; i++) {
1512 const int stride = dim[i].stride;
1513 if (stride < 0) {
1514 index += stride * (ptrdiff_t)(dim[i].extent - 1);
1515 }
1516 }
1517 return index;
1518 }
1519
1520 /** An offset to one beyond the element with the highest address.
1521 * Offset is in elements, not bytes.
1522 * Unlike end(), this is ok to call on an unallocated buffer. */
1523 HALIDE_ALWAYS_INLINE ptrdiff_t end_offset() const {
1524 ptrdiff_t index = 0;
1525 for (int i = 0; i < dimensions; i++) {
1526 const int stride = dim[i].stride;
1527 if (stride > 0) {
1528 index += stride * (ptrdiff_t)(dim[i].extent - 1);
1529 }
1530 }
1531 index += 1;
1532 return index;
1533 }
1534
1535 /** A pointer to the element with the lowest address.
1536 * If all strides are positive, equal to the host pointer.
1537 * Illegal to call on an unallocated buffer. */
1538 HALIDE_ALWAYS_INLINE uint8_t *begin() const {
1539 return host + begin_offset() * type.bytes();
1540 }
1541
1542 /** A pointer to one beyond the element with the highest address.
1543 * Illegal to call on an unallocated buffer. */
1544 HALIDE_ALWAYS_INLINE uint8_t *end() const {
1545 return host + end_offset() * type.bytes();
1546 }
1547
1548 /** The total number of bytes spanned by the data in memory. */
1549 HALIDE_ALWAYS_INLINE size_t size_in_bytes() const {
1550 return (size_t)(end_offset() - begin_offset()) * type.bytes();
1551 }
1552
1553 /** A pointer to the element at the given location. */
1554 HALIDE_ALWAYS_INLINE uint8_t *address_of(const int *pos) const {
1555 ptrdiff_t index = 0;
1556 for (int i = 0; i < dimensions; i++) {
1557 index += (ptrdiff_t)dim[i].stride * (pos[i] - dim[i].min);
1558 }
1559 return host + index * type.bytes();
1560 }
1561
1562 /** Attempt to call device_sync for the buffer. If the buffer
1563 * has no device_interface (or no device_sync), this is a quiet no-op.
1564 * Calling this explicitly should rarely be necessary, except for profiling. */
1565 HALIDE_ALWAYS_INLINE int device_sync(void *ctx = nullptr) {
1567 return device_interface->device_sync(ctx, this);
1568 }
1569 return 0;
1570 }
1571
1572 /** Check if an input buffer passed extern stage is a querying
1573 * bounds. Compared to doing the host pointer check directly,
1574 * this both adds clarity to code and will facilitate moving to
1575 * another representation for bounds query arguments. */
1576 HALIDE_ALWAYS_INLINE bool is_bounds_query() const {
1577 return host == nullptr && device == 0;
1578 }
1579
1580#endif
1582
1583#ifdef __cplusplus
1584extern "C" {
1585#endif
1586
1587#ifndef HALIDE_ATTRIBUTE_DEPRECATED
1588#ifdef HALIDE_ALLOW_DEPRECATED
1589#define HALIDE_ATTRIBUTE_DEPRECATED(x)
1590#else
1591#ifdef _MSC_VER
1592#define HALIDE_ATTRIBUTE_DEPRECATED(x) __declspec(deprecated(x))
1593#else
1594#define HALIDE_ATTRIBUTE_DEPRECATED(x) __attribute__((deprecated(x)))
1595#endif
1596#endif
1597#endif
1598
1599/** halide_scalar_value_t is a simple union able to represent all the well-known
1600 * scalar values in a filter argument. Note that it isn't tagged with a type;
1601 * you must ensure you know the proper type before accessing. Most user
1602 * code will never need to create instances of this struct; its primary use
1603 * is to hold def/min/max values in a halide_filter_argument_t. (Note that
1604 * this is conceptually just a union; it's wrapped in a struct to ensure
1605 * that it doesn't get anonymized by LLVM.)
1606 */
1608 union {
1609 bool b;
1618 float f32;
1619 double f64;
1620 void *handle;
1621 } u;
1622#ifdef __cplusplus
1624 u.u64 = 0;
1625 }
1626#endif
1627};
1628
1634
1635/*
1636 These structs must be robust across different compilers and settings; when
1637 modifying them, strive for the following rules:
1638
1639 1) All fields are explicitly sized. I.e. must use int32_t and not "int"
1640 2) All fields must land on an alignment boundary that is the same as their size
1641 3) Explicit padding is added to make that so
1642 4) The sizeof the struct is padded out to a multiple of the largest natural size thing in the struct
1643 5) don't forget that 32 and 64 bit pointers are different sizes
1644*/
1645
1646/**
1647 * Obsolete version of halide_filter_argument_t; only present in
1648 * code that wrote halide_filter_metadata_t version 0.
1649 */
1651 const char *name;
1656};
1657
1658/**
1659 * halide_filter_argument_t is essentially a plain-C-struct equivalent to
1660 * Halide::Argument; most user code will never need to create one.
1661 */
1663 const char *name; // name of the argument; will never be null or empty.
1664 int32_t kind; // actually halide_argument_kind_t
1665 int32_t dimensions; // always zero for scalar arguments
1667 // These pointers should always be null for buffer arguments,
1668 // and *may* be null for scalar arguments. (A null value means
1669 // there is no def/min/max/estimate specified for this argument.)
1671 // This pointer should always be null for scalar arguments,
1672 // and *may* be null for buffer arguments. If not null, it should always
1673 // point to an array of dimensions*2 pointers, which will be the (min, extent)
1674 // estimates for each dimension of the buffer. (Note that any of the pointers
1675 // may be null as well.)
1677};
1678
1680#ifdef __cplusplus
1681 static const int32_t VERSION = 1;
1682#endif
1683
1684 /** version of this metadata; currently always 1. */
1686
1687 /** The number of entries in the arguments field. This is always >= 1. */
1689
1690 /** An array of the filters input and output arguments; this will never be
1691 * null. The order of arguments is not guaranteed (input and output arguments
1692 * may come in any order); however, it is guaranteed that all arguments
1693 * will have a unique name within a given filter. */
1695
1696 /** The Target for which the filter was compiled. This is always
1697 * a canonical Target string (ie a product of Target::to_string). */
1698 const char *target;
1699
1700 /** The function name of the filter. */
1701 const char *name;
1702};
1703
1704/** halide_register_argv_and_metadata() is a **user-defined** function that
1705 * must be provided in order to use the registration.cc files produced
1706 * by Generators when the 'registration' output is requested. Each registration.cc
1707 * file provides a static initializer that calls this function with the given
1708 * filter's argv-call variant, its metadata, and (optionally) and additional
1709 * textual data that the build system chooses to tack on for its own purposes.
1710 * Note that this will be called at static-initializer time (i.e., before
1711 * main() is called), and in an unpredictable order. Note that extra_key_value_pairs
1712 * may be nullptr; if it's not null, it's expected to be a null-terminated list
1713 * of strings, with an even number of entries. */
1715 int (*filter_argv_call)(void **),
1716 const struct halide_filter_metadata_t *filter_metadata,
1717 const char *const *extra_key_value_pairs);
1718
1719/** The functions below here are relevant for pipelines compiled with
1720 * the -profile target flag, which runs a sampling profiler thread
1721 * alongside the pipeline. */
1722
1723/** Per-Func state tracked by the sampling profiler. */
1725 /** Total time taken evaluating this Func (in nanoseconds). */
1727
1728 /** The current memory allocation of this Func. */
1730
1731 /** The peak memory allocation of this Func. */
1733
1734 /** The total memory allocation of this Func. */
1736
1737 /** The peak stack allocation of this Func's threads. */
1739
1740 /** The average number of thread pool worker threads active while computing this Func. */
1742
1743 /** The name of this Func. A global constant string. */
1744 const char *name;
1745
1746 /** The total number of memory allocation of this Func. */
1748};
1749
1750/** Per-pipeline state tracked by the sampling profiler. These exist
1751 * in a linked list. */
1753 /** Total time spent inside this pipeline (in nanoseconds) */
1755
1756 /** The current memory allocation of funcs in this pipeline. */
1758
1759 /** The peak memory allocation of funcs in this pipeline. */
1761
1762 /** The total memory allocation of funcs in this pipeline. */
1764
1765 /** The average number of thread pool worker threads doing useful
1766 * work while computing this pipeline. */
1768
1769 /** The name of this pipeline. A global constant string. */
1770 const char *name;
1771
1772 /** An array containing states for each Func in this pipeline. */
1774
1775 /** The next pipeline_stats pointer. It's a void * because types
1776 * in the Halide runtime may not currently be recursive. */
1777 void *next;
1778
1779 /** The number of funcs in this pipeline. */
1781
1782 /** An internal base id used to identify the funcs in this pipeline. */
1784
1785 /** The number of times this pipeline has been run. */
1786 int runs;
1787
1788 /** The total number of samples taken inside of this pipeline. */
1790
1791 /** The total number of memory allocation of funcs in this pipeline. */
1793};
1794
1795/** The global state of the profiler. */
1796
1798 /** Guards access to the fields below. If not locked, the sampling
1799 * profiler thread is free to modify things below (including
1800 * reordering the linked list of pipeline stats). */
1802
1803 /** The amount of time the profiler thread sleeps between samples
1804 * in milliseconds. Defaults to 1 */
1806
1807 /** An internal id used for bookkeeping. */
1809
1810 /** The id of the current running Func. Set by the pipeline, read
1811 * periodically by the profiler thread. */
1813
1814 /** The number of threads currently doing work. */
1816
1817 /** A linked list of stats gathered for each pipeline. */
1819
1820 /** Retrieve remote profiler state. Used so that the sampling
1821 * profiler can follow along with execution that occurs elsewhere,
1822 * e.g. on a DSP. If null, it reads from the int above instead. */
1823 void (*get_remote_profiler_state)(int *func, int *active_workers);
1824
1825 /** Sampling thread reference to be joined at shutdown. */
1826 struct halide_thread *sampling_thread;
1827};
1828
1829/** Profiler func ids with special meanings. */
1830enum {
1831 /// current_func takes on this value when not inside Halide code
1833 /// Set current_func to this value to tell the profiling thread to
1834 /// halt. It will start up again next time you run a pipeline with
1835 /// profiling enabled.
1838
1839/** Get a pointer to the global profiler state for programmatic
1840 * inspection. Lock it before using to pause the profiler. */
1842
1843/** Get a pointer to the pipeline state associated with pipeline_name.
1844 * This function grabs the global profiler state's lock on entry. */
1846
1847/** Collects profiling information. Intended to be called from a timer
1848 * interrupt handler if timer based profiling is being used.
1849 * State argument is acquired via halide_profiler_get_pipeline_state.
1850 * prev_t argument is the previous time and can be used to set a more
1851 * accurate time interval if desired. */
1853
1854/** Reset profiler state cheaply. May leave threads running or some
1855 * memory allocated but all accumluated statistics are reset.
1856 * WARNING: Do NOT call this method while any halide pipeline is
1857 * running; halide_profiler_memory_allocate/free and
1858 * halide_profiler_stack_peak_update update the profiler pipeline's
1859 * state without grabbing the global profiler state's lock. */
1861
1862/** Reset all profiler state.
1863 * WARNING: Do NOT call this method while any halide pipeline is
1864 * running; halide_profiler_memory_allocate/free and
1865 * halide_profiler_stack_peak_update update the profiler pipeline's
1866 * state without grabbing the global profiler state's lock. */
1868
1869/** Print out timing statistics for everything run since the last
1870 * reset. Also happens at process exit. */
1871extern void halide_profiler_report(void *user_context);
1872
1873/** For timer based profiling, this routine starts the timer chain running.
1874 * halide_get_profiler_state can be called to get the current timer interval.
1875 */
1877/** These routines are called to temporarily disable and then reenable
1878 * timer interuppts for profiling */
1879//@{
1882//@}
1883
1884/// \name "Float16" functions
1885/// These functions operate of bits (``uint16_t``) representing a half
1886/// precision floating point number (IEEE-754 2008 binary16).
1887//{@
1888
1889/** Read bits representing a half precision floating point number and return
1890 * the float that represents the same value */
1892
1893/** Read bits representing a half precision floating point number and return
1894 * the double that represents the same value */
1896
1897// TODO: Conversion functions to half
1898
1899//@}
1900
1901// Allocating and freeing device memory is often very slow. The
1902// methods below give Halide's runtime permission to hold onto device
1903// memory to service future requests instead of returning it to the
1904// underlying device API. The API does not manage an allocation pool,
1905// all it does is provide access to a shared counter that acts as a
1906// limit on the unused memory not yet returned to the underlying
1907// device API. It makes callbacks to participants when memory needs to
1908// be released because the limit is about to be exceeded (either
1909// because the limit has been reduced, or because the memory owned by
1910// some participant becomes unused).
1911
1912/** Tell Halide whether or not it is permitted to hold onto device
1913 * allocations to service future requests instead of returning them
1914 * eagerly to the underlying device API. Many device allocators are
1915 * quite slow, so it can be beneficial to set this to true. The
1916 * default value for now is false.
1917 *
1918 * Note that if enabled, the eviction policy is very simplistic. The
1919 * 32 most-recently used allocations are preserved, regardless of
1920 * their size. Additionally, if a call to cuMalloc results in an
1921 * out-of-memory error, the entire cache is flushed and the allocation
1922 * is retried. See https://github.com/halide/Halide/issues/4093
1923 *
1924 * If set to false, releases all unused device allocations back to the
1925 * underlying device APIs. For finer-grained control, see specific
1926 * methods in each device api runtime. */
1927extern int halide_reuse_device_allocations(void *user_context, bool);
1928
1929/** Determines whether on device_free the memory is returned
1930 * immediately to the device API, or placed on a free list for future
1931 * use. Override and switch based on the user_context for
1932 * finer-grained control. By default just returns the value most
1933 * recently set by the method above. */
1934extern bool halide_can_reuse_device_allocations(void *user_context);
1935
1937 int (*release_unused)(void *user_context);
1939};
1940
1941/** Register a callback to be informed when
1942 * halide_reuse_device_allocations(false) is called, and all unused
1943 * device allocations must be released. The object passed should have
1944 * global lifetime, and its next field will be clobbered. */
1946
1947#ifdef __cplusplus
1948} // End extern "C"
1949#endif
1950
1951#if (__cplusplus >= 201103L || _MSVC_LANG >= 201103L)
1952
1953namespace {
1954
1955template<typename T>
1956struct check_is_pointer {
1957 static constexpr bool value = false;
1958};
1959
1960template<typename T>
1961struct check_is_pointer<T *> {
1962 static constexpr bool value = true;
1963};
1964
1965} // namespace
1966
1967/** Construct the halide equivalent of a C type */
1968template<typename T>
1969HALIDE_ALWAYS_INLINE constexpr halide_type_t halide_type_of() {
1970 // Create a compile-time error if T is not a pointer (without
1971 // using any includes - this code goes into the runtime).
1972 // (Note that we can't have uninitialized variables in constexpr functions,
1973 // even if those variables aren't used.)
1974 static_assert(check_is_pointer<T>::value, "Expected a pointer type here");
1976}
1977
1978template<>
1979HALIDE_ALWAYS_INLINE constexpr halide_type_t halide_type_of<float>() {
1980 return halide_type_t(halide_type_float, 32);
1981}
1982
1983template<>
1984HALIDE_ALWAYS_INLINE constexpr halide_type_t halide_type_of<double>() {
1985 return halide_type_t(halide_type_float, 64);
1986}
1987
1988template<>
1989HALIDE_ALWAYS_INLINE constexpr halide_type_t halide_type_of<bool>() {
1991}
1992
1993template<>
1994HALIDE_ALWAYS_INLINE constexpr halide_type_t halide_type_of<uint8_t>() {
1996}
1997
1998template<>
1999HALIDE_ALWAYS_INLINE constexpr halide_type_t halide_type_of<uint16_t>() {
2000 return halide_type_t(halide_type_uint, 16);
2001}
2002
2003template<>
2004HALIDE_ALWAYS_INLINE constexpr halide_type_t halide_type_of<uint32_t>() {
2005 return halide_type_t(halide_type_uint, 32);
2006}
2007
2008template<>
2009HALIDE_ALWAYS_INLINE constexpr halide_type_t halide_type_of<uint64_t>() {
2010 return halide_type_t(halide_type_uint, 64);
2011}
2012
2013template<>
2014HALIDE_ALWAYS_INLINE constexpr halide_type_t halide_type_of<int8_t>() {
2015 return halide_type_t(halide_type_int, 8);
2016}
2017
2018template<>
2019HALIDE_ALWAYS_INLINE constexpr halide_type_t halide_type_of<int16_t>() {
2020 return halide_type_t(halide_type_int, 16);
2021}
2022
2023template<>
2024HALIDE_ALWAYS_INLINE constexpr halide_type_t halide_type_of<int32_t>() {
2025 return halide_type_t(halide_type_int, 32);
2026}
2027
2028template<>
2029HALIDE_ALWAYS_INLINE constexpr halide_type_t halide_type_of<int64_t>() {
2030 return halide_type_t(halide_type_int, 64);
2031}
2032
2033#endif // (__cplusplus >= 201103L || _MSVC_LANG >= 201103L)
2034
2035#endif // HALIDE_HALIDERUNTIME_H
halide_error_handler_t halide_set_error_handler(halide_error_handler_t handler)
void halide_profiler_reset()
Reset profiler state cheaply.
void halide_set_custom_parallel_runtime(halide_do_par_for_t, halide_do_task_t, halide_do_loop_task_t, halide_do_parallel_tasks_t, halide_semaphore_init_t, halide_semaphore_try_acquire_t, halide_semaphore_release_t)
int halide_memoization_cache_lookup(void *user_context, const uint8_t *cache_key, int32_t size, struct halide_buffer_t *realized_bounds, int32_t tuple_count, struct halide_buffer_t **tuple_buffers)
Given a cache key for a memoized result, currently constructed from the Func name and top-level Func ...
int halide_error_bad_extern_fold(void *user_context, const char *func_name, int dim, int min, int extent, int valid_min, int fold_factor)
int halide_device_sync(void *user_context, struct halide_buffer_t *buf)
Wait for current GPU operations to complete.
int halide_default_do_task(void *user_context, halide_task_t f, int idx, uint8_t *closure)
int halide_do_par_for(void *user_context, halide_task_t task, int min, int size, uint8_t *closure)
void *(* halide_load_library_t)(const char *name)
int halide_error_bad_fold(void *user_context, const char *func_name, const char *var_name, const char *loop_name)
int(* halide_semaphore_release_t)(struct halide_semaphore_t *, int)
void halide_cond_signal(struct halide_cond *cond)
void * halide_default_get_library_symbol(void *lib, const char *name)
int halide_do_loop_task(void *user_context, halide_loop_task_t f, int min, int extent, uint8_t *closure, void *task_parent)
halide_load_library_t halide_set_custom_load_library(halide_load_library_t user_load_library)
int halide_device_crop(void *user_context, const struct halide_buffer_t *src, struct halide_buffer_t *dst)
Give the destination buffer a device allocation which is an alias for the same coordinate range in th...
int halide_semaphore_init(struct halide_semaphore_t *, int n)
halide_get_symbol_t halide_set_custom_get_symbol(halide_get_symbol_t user_get_symbol)
double halide_float16_bits_to_double(uint16_t)
Read bits representing a half precision floating point number and return the double that represents t...
void * halide_malloc(void *user_context, size_t x)
Halide calls these functions to allocate and free memory.
void *(* halide_get_library_symbol_t)(void *lib, const char *name)
int halide_msan_annotate_buffer_is_initialized(void *user_context, struct halide_buffer_t *buffer)
Mark the data pointed to by the halide_buffer_t as initialized (but not the halide_buffer_t itself),...
void halide_default_print(void *user_context, const char *)
halide_target_feature_t
Optional features a compilation Target can have.
@ halide_target_feature_large_buffers
Enable 64-bit buffer indexing to support buffers > 2GB. Ignored if bits != 64.
@ halide_target_feature_fma
Enable x86 FMA instruction.
@ halide_target_feature_wasm_bulk_memory
Enable +bulk-memory instructions for WebAssembly codegen.
@ halide_target_feature_tsan
Enable hooks for TSAN support.
@ halide_target_feature_msan
Enable hooks for MSAN support.
@ halide_target_feature_wasm_threads
Enable use of threads in WebAssembly codegen. Requires the use of a wasm runtime that provides pthrea...
@ halide_target_feature_trace_loads
Trace all loads done by the pipeline. Equivalent to calling Func::trace_loads on every non-inlined Fu...
@ halide_target_feature_enable_llvm_loop_opt
Enable loop vectorization + unrolling in LLVM. Overrides halide_target_feature_disable_llvm_loop_opt....
@ halide_target_feature_no_asserts
Disable all runtime checks, for slightly tighter code.
@ halide_target_feature_cl_doubles
Enable double support on OpenCL targets.
@ halide_target_feature_rvv
Enable RISCV "V" Vector Extension.
@ halide_target_feature_openglcompute
Enable OpenGL Compute runtime.
@ halide_target_feature_avx2
Use AVX 2 instructions. Only relevant on x86.
@ halide_target_feature_trace_realizations
Trace all realizations done by the pipeline. Equivalent to calling Func::trace_realizations on every ...
@ halide_target_feature_c_plus_plus_mangling
Generate C++ mangled names for result function, et al.
@ halide_target_feature_no_runtime
Do not include a copy of the Halide runtime in any generated object file or assembly.
@ halide_target_feature_hvx_v65
Enable Hexagon v65 architecture.
@ halide_target_feature_debug
Turn on debug info and output for runtime code.
@ halide_target_feature_embed_bitcode
Emulate clang -fembed-bitcode flag.
@ halide_target_feature_wasm_simd128
Enable +simd128 instructions for WebAssembly codegen.
@ halide_target_feature_end
A sentinel. Every target is considered to have this feature, and setting this feature does nothing.
@ halide_llvm_large_code_model
Use the LLVM large code model to compile.
@ halide_target_feature_profile_by_timer
Alternative to halide_target_feature_profile using timer interrupt for systems without threads or app...
@ halide_target_feature_soft_float_abi
Enable soft float ABI. This only enables the soft float ABI calling convention, which does not necess...
@ halide_target_feature_sve2
Enable ARM Scalable Vector Extensions v2.
@ halide_target_feature_disable_llvm_loop_opt
Disable loop vectorization + unrolling in LLVM. (Ignored for non-LLVM targets.)
@ halide_target_feature_d3d12compute
Enable Direct3D 12 Compute runtime.
@ halide_target_feature_cuda_capability86
Enable CUDA compute capability 8.6 (Ampere)
@ halide_target_feature_matlab
Generate a mexFunction compatible with Matlab mex libraries. See tools/mex_halide....
@ halide_target_feature_avx512_skylake
Enable the AVX512 features supported by Skylake Xeon server processors. This adds AVX512-VL,...
@ halide_target_feature_avx512_cannonlake
Enable the AVX512 features expected to be supported by future Cannonlake processors....
@ halide_target_feature_metal
Enable the (Apple) Metal runtime.
@ halide_target_feature_hvx_128
Enable HVX 128 byte mode.
@ halide_target_feature_cuda_capability70
Enable CUDA compute capability 7.0 (Volta)
@ halide_target_feature_fma4
Enable x86 (AMD) FMA4 instruction set.
@ halide_target_feature_wasm_sat_float_to_int
Enable saturating (nontrapping) float-to-int instructions for WebAssembly codegen.
@ halide_target_feature_cuda_capability30
Enable CUDA compute capability 3.0 (Kepler)
@ halide_target_feature_no_neon
Avoid using NEON instructions. Only relevant for 32-bit ARM.
@ halide_target_feature_cuda_capability61
Enable CUDA compute capability 6.1 (Pascal)
@ halide_target_feature_armv7s
Generate code for ARMv7s. Only relevant for 32-bit ARM.
@ halide_target_feature_trace_pipeline
Trace the pipeline.
@ halide_target_feature_cl_atomic64
Enable 64-bit atomics operations on OpenCL targets.
@ halide_target_feature_egl
Force use of EGL support.
@ halide_target_feature_profile
Launch a sampling profiler alongside the Halide pipeline that monitors and reports the runtime used b...
@ halide_target_feature_strict_float
Turn off all non-IEEE floating-point optimization. Currently applies only to LLVM targets.
@ halide_target_feature_cuda_capability35
Enable CUDA compute capability 3.5 (Kepler)
@ halide_target_feature_asan
Enable hooks for ASAN support.
@ halide_target_feature_cl_half
Enable half support on OpenCL targets.
@ halide_target_feature_arm_dot_prod
Enable ARMv8.2-a dotprod extension (i.e. udot and sdot instructions)
@ halide_target_feature_avx512_sapphirerapids
Enable the AVX512 features supported by Sapphire Rapids processors. This include all of the Cannonlak...
@ halide_target_feature_sse41
Use SSE 4.1 and earlier instructions. Only relevant on x86.
@ halide_target_feature_power_arch_2_07
Use POWER ISA 2.07 new instructions. Only relevant on POWERPC.
@ halide_target_feature_opencl
Enable the OpenCL runtime.
@ halide_target_feature_trace_stores
Trace all stores done by the pipeline. Equivalent to calling Func::trace_stores on every non-inlined ...
@ halide_target_feature_hexagon_dma
Enable Hexagon DMA buffers.
@ halide_target_feature_avx512
Enable the base AVX512 subset supported by all AVX512 architectures. The specific feature sets are AV...
@ halide_target_feature_avx512_knl
Enable the AVX512 features supported by Knight's Landing chips, such as the Xeon Phi x200....
@ halide_target_feature_cuda_capability50
Enable CUDA compute capability 5.0 (Maxwell)
@ halide_target_feature_arm_fp16
Enable ARMv8.2-a half-precision floating point data processing.
@ halide_target_feature_hvx_v62
Enable Hexagon v62 architecture.
@ halide_target_feature_hvx_use_shared_object
Deprecated.
@ halide_target_feature_cuda
Enable the CUDA runtime. Defaults to compute capability 2.0 (Fermi)
@ halide_target_feature_armv81a
Enable ARMv8.1-a instructions.
@ halide_target_feature_sanitizer_coverage
Enable hooks for SanitizerCoverage support.
@ halide_target_feature_cuda_capability80
Enable CUDA compute capability 8.0 (Ampere)
@ halide_target_feature_f16c
Enable x86 16-bit float support.
@ halide_target_feature_cuda_capability32
Enable CUDA compute capability 3.2 (Tegra K1)
@ halide_target_feature_jit
Generate code that will run immediately inside the calling process.
@ halide_target_feature_wasm_signext
Enable +sign-ext instructions for WebAssembly codegen.
@ halide_target_feature_avx
Use AVX 1 instructions. Only relevant on x86.
@ halide_target_feature_cuda_capability75
Enable CUDA compute capability 7.5 (Turing)
@ halide_target_feature_check_unsafe_promises
Insert assertions for promises.
@ halide_target_feature_vsx
Use VSX instructions. Only relevant on POWERPC.
@ halide_target_feature_user_context
Generated code takes a user_context pointer as first argument.
@ halide_target_feature_no_bounds_query
Disable the bounds querying functionality.
@ halide_target_feature_fuzz_float_stores
On every floating point store, set the last bit of the mantissa to zero. Pipelines for which the outp...
@ halide_target_feature_sve
Enable ARM Scalable Vector Extensions.
@ halide_target_feature_hvx_v66
Enable Hexagon v66 architecture.
void halide_free(void *user_context, void *ptr)
void halide_memoization_cache_cleanup()
Free all memory and resources associated with the memoization cache.
bool halide_semaphore_try_acquire(struct halide_semaphore_t *, int n)
halide_buffer_flags
@ halide_buffer_flag_device_dirty
@ halide_buffer_flag_host_dirty
void halide_profiler_shutdown()
Reset all profiler state.
bool halide_default_semaphore_try_acquire(struct halide_semaphore_t *, int n)
int halide_error_buffer_allocation_too_large(void *user_context, const char *buffer_name, uint64_t allocation_size, uint64_t max_size)
struct halide_mutex_array * halide_mutex_array_create(int sz)
void halide_start_timer_chain()
For timer based profiling, this routine starts the timer chain running.
void halide_cond_wait(struct halide_cond *cond, struct halide_mutex *mutex)
int(* halide_do_par_for_t)(void *, halide_task_t, int, int, uint8_t *)
Set a custom method for performing a parallel for loop.
int halide_set_num_threads(int n)
Set the number of threads used by Halide's thread pool.
int halide_copy_to_host(void *user_context, struct halide_buffer_t *buf)
Copy image data from device memory to host memory.
int halide_default_do_par_for(void *user_context, halide_task_t task, int min, int size, uint8_t *closure)
The default versions of the parallel runtime functions.
int halide_msan_annotate_memory_is_initialized(void *user_context, const void *ptr, uint64_t len)
Annotate that a given range of memory has been initialized; only used when Target::MSAN is enabled.
halide_print_t halide_set_custom_print(halide_print_t print)
int halide_error_bad_dimensions(void *user_context, const char *func_name, int32_t dimensions_given, int32_t correct_dimensions)
int halide_mutex_array_unlock(struct halide_mutex_array *array, int entry)
int halide_error_constraint_violated(void *user_context, const char *var, int val, const char *constrained_var, int constrained_val)
int halide_default_do_loop_task(void *user_context, halide_loop_task_t f, int min, int extent, uint8_t *closure, void *task_parent)
int(* halide_task_t)(void *user_context, int task_number, uint8_t *closure)
Define halide_do_par_for to replace the default thread pool implementation.
void halide_mutex_lock(struct halide_mutex *mutex)
A basic set of mutex and condition variable functions, which call platform specific code for mutual e...
halide_trace_event_code_t
@ halide_trace_consume
@ halide_trace_load
@ halide_trace_tag
@ halide_trace_store
@ halide_trace_begin_pipeline
@ halide_trace_end_pipeline
@ halide_trace_end_produce
@ halide_trace_produce
@ halide_trace_end_consume
@ halide_trace_end_realization
@ halide_trace_begin_realization
struct halide_profiler_pipeline_stats * halide_profiler_get_pipeline_state(const char *pipeline_name)
Get a pointer to the pipeline state associated with pipeline_name.
int halide_do_task(void *user_context, halide_task_t f, int idx, uint8_t *closure)
halide_malloc_t halide_set_custom_malloc(halide_malloc_t user_malloc)
int halide_error_device_dirty_with_no_device_support(void *user_context, const char *buffer_name)
int halide_default_semaphore_init(struct halide_semaphore_t *, int n)
void halide_msan_annotate_buffer_is_initialized_as_destructor(void *user_context, void *buffer)
void * halide_get_library_symbol(void *lib, const char *name)
void(* halide_error_handler_t)(void *, const char *)
void halide_device_release(void *user_context, const struct halide_device_interface_t *device_interface)
Release all data associated with the given device interface, in particular all resources (memory,...
int halide_error_bounds_inference_call_failed(void *user_context, const char *extern_stage_name, int result)
Halide calls the functions below on various error conditions.
void halide_disable_timer_interrupt()
These routines are called to temporarily disable and then reenable timer interuppts for profiling.
int halide_error_buffer_extents_negative(void *user_context, const char *buffer_name, int dimension, int extent)
int halide_error_buffer_is_null(void *user_context, const char *routine)
void halide_mutex_unlock(struct halide_mutex *mutex)
void halide_shutdown_thread_pool()
int halide_error_constraints_make_required_region_smaller(void *user_context, const char *buffer_name, int dimension, int constrained_min, int constrained_extent, int required_min, int required_extent)
int32_t halide_debug_to_file(void *user_context, const char *filename, int32_t type_code, struct halide_buffer_t *buf)
Called when debug_to_file is used inside Halide code.
int halide_shutdown_trace()
If tracing is writing to a file.
int halide_error_out_of_memory(void *user_context)
void * halide_get_symbol(const char *name)
Halide calls these functions to interact with the underlying system runtime functions.
int halide_error_no_device_interface(void *user_context)
struct halide_thread * halide_spawn_thread(void(*f)(void *), void *closure)
Spawn a thread.
int halide_error_debug_to_file_failed(void *user_context, const char *func, const char *filename, int error_code)
struct halide_dimension_t halide_dimension_t
int halide_error_requirement_failed(void *user_context, const char *condition, const char *message)
void halide_memoization_cache_release(void *user_context, void *host)
If halide_memoization_cache_lookup succeeds, halide_memoization_cache_release must be called to signa...
int(* halide_can_use_target_features_t)(int count, const uint64_t *features)
void halide_register_argv_and_metadata(int(*filter_argv_call)(void **), const struct halide_filter_metadata_t *filter_metadata, const char *const *extra_key_value_pairs)
halide_register_argv_and_metadata() is a user-defined function that must be provided in order to use ...
void *(* halide_get_symbol_t)(const char *name)
int halide_error_param_too_large_f64(void *user_context, const char *param_name, double val, double max_val)
int halide_msan_check_memory_is_initialized(void *user_context, const void *ptr, uint64_t len, const char *name)
Verify that a given range of memory has been initialized; only used when Target::MSAN is enabled.
int(* halide_do_loop_task_t)(void *, halide_loop_task_t, int, int, uint8_t *, void *)
The version of do_task called for loop tasks.
int halide_error_buffer_extents_too_large(void *user_context, const char *buffer_name, int64_t actual_size, int64_t max_size)
int32_t halide_trace(void *user_context, const struct halide_trace_event_t *event)
Called when Funcs are marked as trace_load, trace_store, or trace_realization.
int halide_error_extern_stage_failed(void *user_context, const char *extern_stage_name, int result)
A call to an extern stage failed.
halide_can_use_target_features_t halide_set_custom_can_use_target_features(halide_can_use_target_features_t)
int halide_error_host_is_null(void *user_context, const char *func_name)
void halide_set_trace_file(int fd)
Set the file descriptor that Halide should write binary trace events to.
int halide_memoization_cache_store(void *user_context, const uint8_t *cache_key, int32_t size, struct halide_buffer_t *realized_bounds, int32_t tuple_count, struct halide_buffer_t **tuple_buffers, bool has_eviction_key, uint64_t eviction_key)
Given a cache key for a memoized result, currently constructed from the Func name and top-level Func ...
int halide_buffer_copy(void *user_context, struct halide_buffer_t *src, const struct halide_device_interface_t *dst_device_interface, struct halide_buffer_t *dst)
Copy data from one buffer to another.
int halide_error_fold_factor_too_small(void *user_context, const char *func_name, const char *var_name, int fold_factor, const char *loop_name, int required_extent)
int halide_error_param_too_small_f64(void *user_context, const char *param_name, double val, double min_val)
int halide_error_param_too_large_i64(void *user_context, const char *param_name, int64_t val, int64_t max_val)
void halide_enable_timer_interrupt()
int halide_error_param_too_small_u64(void *user_context, const char *param_name, uint64_t val, uint64_t min_val)
void(* halide_print_t)(void *, const char *)
Definition: HalideRuntime.h:99
halide_trace_t halide_set_custom_trace(halide_trace_t trace)
void halide_print(void *user_context, const char *)
Print a message to stderr.
bool(* halide_semaphore_try_acquire_t)(struct halide_semaphore_t *, int)
void halide_profiler_report(void *user_context)
Print out timing statistics for everything run since the last reset.
void halide_set_gpu_device(int n)
Selects which gpu device to use.
void halide_mutex_array_destroy(void *user_context, void *array)
int32_t halide_default_trace(void *user_context, const struct halide_trace_event_t *event)
@ halide_profiler_please_stop
Set current_func to this value to tell the profiling thread to halt.
@ halide_profiler_outside_of_halide
current_func takes on this value when not inside Halide code
int halide_default_can_use_target_features(int count, const uint64_t *features)
This is the default implementation of halide_can_use_target_features; it is provided for convenience ...
int halide_reuse_device_allocations(void *user_context, bool)
Tell Halide whether or not it is permitted to hold onto device allocations to service future requests...
int halide_error_param_too_small_i64(void *user_context, const char *param_name, int64_t val, int64_t min_val)
halide_type_code_t
Types in the halide type system.
@ halide_type_float
IEEE floating point numbers.
@ halide_type_handle
opaque pointer type (void *)
@ halide_type_bfloat
floating point numbers in the bfloat format
@ halide_type_int
signed integers
@ halide_type_uint
unsigned integers
int halide_device_malloc(void *user_context, struct halide_buffer_t *buf, const struct halide_device_interface_t *device_interface)
Allocate device memory to back a halide_buffer_t.
bool halide_can_reuse_device_allocations(void *user_context)
Determines whether on device_free the memory is returned immediately to the device API,...
int halide_mutex_array_lock(struct halide_mutex_array *array, int entry)
void(* halide_free_t)(void *, void *)
int halide_error_storage_bound_too_small(void *user_context, const char *func_name, const char *var_name, int provided_size, int required_size)
int(* halide_loop_task_t)(void *user_context, int min, int extent, uint8_t *closure, void *task_parent)
A task representing a serial for loop evaluated over some range.
int halide_error_specialize_fail(void *user_context, const char *message)
int halide_error_device_interface_no_device(void *user_context)
int halide_error_host_and_device_dirty(void *user_context)
int halide_error_access_out_of_bounds(void *user_context, const char *func_name, int dimension, int min_touched, int max_touched, int min_valid, int max_valid)
int halide_error_explicit_bounds_too_small(void *user_context, const char *func_name, const char *var_name, int min_bound, int max_bound, int min_required, int max_required)
Various other error conditions.
void * halide_default_malloc(void *user_context, size_t x)
void * halide_default_load_library(const char *name)
int halide_error_buffer_argument_is_null(void *user_context, const char *buffer_name)
int halide_default_semaphore_release(struct halide_semaphore_t *, int n)
void halide_default_error(void *user_context, const char *)
int halide_device_slice(void *user_context, const struct halide_buffer_t *src, int slice_dim, int slice_pos, struct halide_buffer_t *dst)
Give the destination buffer a device allocation which is an alias for a similar coordinate range in t...
int halide_error_device_crop_failed(void *user_context)
void halide_default_free(void *user_context, void *ptr)
void halide_memoization_cache_evict(void *user_context, uint64_t eviction_key)
Evict all cache entries that were tagged with the given eviction_key in the memoize scheduling direct...
halide_do_par_for_t halide_set_custom_do_par_for(halide_do_par_for_t do_par_for)
void halide_join_thread(struct halide_thread *)
Join a thread.
halide_error_code_t
The error codes that may be returned by a Halide pipeline.
@ halide_error_code_no_device_interface
Buffer has a non-zero device but no device interface, which violates a Halide invariant.
@ halide_error_code_bad_fold
A fold_storage directive was used on a dimension that is not accessed in a monotonically increasing o...
@ halide_error_code_fold_factor_too_small
A fold_storage directive was used with a fold factor that was too small to store all the values of a ...
@ halide_error_code_device_interface_no_device
Buffer has a non-null device_interface but device is 0, which violates a Halide invariant.
@ halide_error_code_param_too_large
A scalar parameter passed in was greater than its minimum declared value.
@ halide_error_code_param_too_small
A scalar parameter passed in was smaller than its minimum declared value.
@ halide_error_code_access_out_of_bounds
A pipeline would access memory outside of the halide_buffer_t passed in.
@ halide_error_code_specialize_fail
A specialize_fail() schedule branch was selected at runtime.
@ halide_error_code_requirement_failed
User-specified require() expression was not satisfied.
@ halide_error_code_bad_extern_fold
A folded buffer was passed to an extern stage, but the region touched wraps around the fold boundary.
@ halide_error_code_incompatible_device_interface
An operation on a buffer required an allocation on a particular device interface, but a device alloca...
@ halide_error_code_internal_error
There is a bug in the Halide compiler.
@ halide_error_code_buffer_extents_negative
At least one of the buffer's extents are negative.
@ halide_error_code_constraints_make_required_region_smaller
Applying explicit constraints on the size of an input or output buffer shrank the size of that buffer...
@ halide_error_code_unused_29
@ halide_error_code_copy_to_device_failed
The Halide runtime encountered an error while trying to copy from host to device.
@ halide_error_code_generic_error
An uncategorized error occurred.
@ halide_error_code_device_crop_failed
Cropping/slicing a buffer failed for some other reason.
@ halide_error_code_success
There was no error.
@ halide_error_code_copy_to_host_failed
The Halide runtime encountered an error while trying to copy from device to host.
@ halide_error_code_matlab_init_failed
An error occurred when attempting to initialize the Matlab runtime.
@ halide_error_code_device_sync_failed
The Halide runtime encountered an error while trying to synchronize with a device.
@ halide_error_code_buffer_argument_is_null
A halide_buffer_t pointer passed in was NULL.
@ halide_error_code_bad_dimensions
The dimensions field of a halide_buffer_t does not match the dimensions of that ImageParam.
@ halide_error_code_device_malloc_failed
The Halide runtime encountered an error while trying to allocate memory on device.
@ halide_error_code_host_and_device_dirty
Buffer has both host and device dirty bits set, which violates a Halide invariant.
@ halide_error_code_debug_to_file_failed
debug_to_file failed to open or write to the specified file.
@ halide_error_code_unused_30
@ halide_error_code_buffer_is_null
The halide_buffer_t * passed to a halide runtime routine is nullptr and this is not allowed.
@ halide_error_code_device_crop_unsupported
Attempted to make cropped/sliced alias of a buffer with a device field, but the device_interface does...
@ halide_error_code_device_buffer_copy_failed
The Halide runtime encountered an error while trying to copy from one buffer to another.
@ halide_error_code_device_free_failed
The Halide runtime encountered an error while trying to free a device allocation.
@ halide_error_code_buffer_allocation_too_large
A halide_buffer_t was given that spans more than 2GB of memory.
@ halide_error_code_bad_type
The elem_size field of a halide_buffer_t does not match the size in bytes of the type of that ImagePa...
@ halide_error_code_device_run_failed
The Halide runtime encountered an error while trying to launch a GPU kernel.
@ halide_error_code_device_dirty_with_no_device_support
A buffer with the device_dirty flag set was passed to a pipeline compiled with no device backends ena...
@ halide_error_code_explicit_bounds_too_small
A Func was given an explicit bound via Func::bound, but this was not large enough to encompass the re...
@ halide_error_code_buffer_extents_too_large
A halide_buffer_t was given with extents that multiply to a number greater than 2^31-1.
@ halide_error_code_device_detach_native_failed
The Halide runtime encountered an error while trying to detach a native device handle.
@ halide_error_code_storage_bound_too_small
An explicit storage bound provided is too small to store all the values produced by the function.
@ halide_error_code_out_of_memory
A call to halide_malloc returned NULL.
@ halide_error_code_matlab_bad_param_type
The type of an mxArray did not match the expected type.
@ halide_error_code_device_wrap_native_failed
The Halide runtime encountered an error while trying to wrap a native device handle.
@ halide_error_code_constraint_violated
A constraint on a size or stride of an input or output buffer was not met by the halide_buffer_t pass...
@ halide_error_code_unaligned_host_ptr
The Halide runtime encountered a host pointer that violated the alignment set for it by way of a call...
@ halide_error_code_host_is_null
The host field on an input or output was null, the device field was not zero, and the pipeline tries ...
void * halide_load_library(const char *name)
void halide_memoization_cache_set_size(int64_t size)
Set the soft maximum amount of memory, in bytes, that the LRU cache will use to memoize Func results.
#define HALIDE_ALWAYS_INLINE
Definition: HalideRuntime.h:38
void halide_cond_broadcast(struct halide_cond *cond)
int halide_device_free(void *user_context, struct halide_buffer_t *buf)
Free device memory.
int halide_error_param_too_large_u64(void *user_context, const char *param_name, uint64_t val, uint64_t max_val)
int32_t(* halide_trace_t)(void *user_context, const struct halide_trace_event_t *)
struct halide_profiler_state * halide_profiler_get_state()
Get a pointer to the global profiler state for programmatic inspection.
int(* halide_do_task_t)(void *, halide_task_t, int, uint8_t *)
If you use the default do_par_for, you can still set a custom handler to perform each individual task...
halide_free_t halide_set_custom_free(halide_free_t user_free)
int halide_default_do_parallel_tasks(void *user_context, int num_tasks, struct halide_parallel_task_t *tasks, void *task_parent)
int halide_get_gpu_device(void *user_context)
Halide calls this to get the desired halide gpu device setting.
int(* halide_semaphore_init_t)(struct halide_semaphore_t *, int)
halide_do_loop_task_t halide_set_custom_do_loop_task(halide_do_loop_task_t do_task)
int halide_device_detach_native(void *user_context, struct halide_buffer_t *buf)
void * halide_default_get_symbol(const char *name)
int halide_copy_to_device(void *user_context, struct halide_buffer_t *buf, const struct halide_device_interface_t *device_interface)
Copy image data from host memory to device memory.
halide_get_library_symbol_t halide_set_custom_get_library_symbol(halide_get_library_symbol_t user_get_library_symbol)
void halide_error(void *user_context, const char *)
Halide calls this function on runtime errors (for example bounds checking failures).
int halide_device_release_crop(void *user_context, struct halide_buffer_t *buf)
Release any resources associated with a cropped/sliced view of another buffer.
int halide_can_use_target_features(int count, const uint64_t *features)
This function is called internally by Halide in some situations to determine if the current execution...
int halide_error_bad_type(void *user_context, const char *func_name, uint32_t type_given, uint32_t correct_type)
halide_do_task_t halide_set_custom_do_task(halide_do_task_t do_task)
int halide_do_parallel_tasks(void *user_context, int num_tasks, struct halide_parallel_task_t *tasks, void *task_parent)
Enqueue some number of the tasks described above and wait for them to complete.
void *(* halide_malloc_t)(void *, size_t)
int halide_device_wrap_native(void *user_context, struct halide_buffer_t *buf, uint64_t handle, const struct halide_device_interface_t *device_interface)
Wrap or detach a native device handle, setting the device field and device_interface field as appropr...
int halide_get_trace_file(void *user_context)
Halide calls this to retrieve the file descriptor to write binary trace events to.
float halide_float16_bits_to_float(uint16_t)
Read bits representing a half precision floating point number and return the float that represents th...
int halide_error_unaligned_host_ptr(void *user_context, const char *func_name, int alignment)
int halide_profiler_sample(struct halide_profiler_state *s, uint64_t *prev_t)
Collects profiling information.
void halide_register_device_allocation_pool(struct halide_device_allocation_pool *)
Register a callback to be informed when halide_reuse_device_allocations(false) is called,...
#define HALIDE_ATTRIBUTE_ALIGN(x)
int halide_msan_check_buffer_is_initialized(void *user_context, struct halide_buffer_t *buffer, const char *buf_name)
Verify that the data pointed to by the halide_buffer_t is initialized (but not the halide_buffer_t it...
int(* halide_do_parallel_tasks_t)(void *, int, struct halide_parallel_task_t *, void *task_parent)
Provide an entire custom tasking runtime via function pointers.
halide_argument_kind_t
@ halide_argument_kind_output_buffer
@ halide_argument_kind_input_scalar
@ halide_argument_kind_input_buffer
int halide_semaphore_release(struct halide_semaphore_t *, int n)
struct halide_buffer_t halide_buffer_t
The raw representation of an image passed around by generated Halide code.
Expr with_lanes(const Expr &x, int lanes)
Rewrite the expression x to have lanes lanes.
auto operator==(const Other &a, const GeneratorParam< T > &b) -> decltype(a==(T) b)
Equality comparison between GeneratorParam<T> and any type that supports operator== with T.
Definition: Generator.h:1102
auto operator<(const Other &a, const GeneratorParam< T > &b) -> decltype(a<(T) b)
Less than comparison between GeneratorParam<T> and any type that supports operator< with T.
Definition: Generator.h:1063
Expr min(const FuncRef &a, const FuncRef &b)
Explicit overloads of min and max for FuncRef.
Definition: Func.h:600
auto operator!=(const Other &a, const GeneratorParam< T > &b) -> decltype(a !=(T) b)
Inequality comparison between between GeneratorParam<T> and any type that supports operator!...
Definition: Generator.h:1115
Expr print(const std::vector< Expr > &values)
Create an Expr that prints out its value whenever it is evaluated.
unsigned __INT64_TYPE__ uint64_t
signed __INT64_TYPE__ int64_t
signed __INT32_TYPE__ int32_t
unsigned __INT8_TYPE__ uint8_t
__PTRDIFF_TYPE__ ptrdiff_t
unsigned __INT16_TYPE__ uint16_t
__SIZE_TYPE__ size_t
unsigned __INT32_TYPE__ uint32_t
signed __INT16_TYPE__ int16_t
signed __INT8_TYPE__ int8_t
The raw representation of an image passed around by generated Halide code.
void * padding
Pads the buffer up to a multiple of 8 bytes.
int32_t dimensions
The dimensionality of the buffer.
halide_dimension_t * dim
The shape of the buffer.
uint64_t device
A device-handle for e.g.
uint8_t * host
A pointer to the start of the data in main memory.
struct halide_type_t type
The type of each buffer element.
const struct halide_device_interface_t * device_interface
The interface used to interpret the above handle.
uint64_t flags
flags with various meanings.
Cross platform condition variable.
uintptr_t _private[1]
struct halide_device_allocation_pool * next
int(* release_unused)(void *user_context)
Each GPU API provides a halide_device_interface_t struct pointing to the code that manages device all...
int(* device_slice)(void *user_context, const struct halide_buffer_t *src, int slice_dim, int slice_pos, struct halide_buffer_t *dst)
int(* device_and_host_malloc)(void *user_context, struct halide_buffer_t *buf, const struct halide_device_interface_t *device_interface)
const struct halide_device_interface_impl_t * impl
int(* wrap_native)(void *user_context, struct halide_buffer_t *buf, uint64_t handle, const struct halide_device_interface_t *device_interface)
int(* compute_capability)(void *user_context, int *major, int *minor)
int(* device_release_crop)(void *user_context, struct halide_buffer_t *buf)
int(* device_crop)(void *user_context, const struct halide_buffer_t *src, struct halide_buffer_t *dst)
void(* device_release)(void *user_context, const struct halide_device_interface_t *device_interface)
int(* copy_to_host)(void *user_context, struct halide_buffer_t *buf)
int(* copy_to_device)(void *user_context, struct halide_buffer_t *buf, const struct halide_device_interface_t *device_interface)
int(* device_free)(void *user_context, struct halide_buffer_t *buf)
int(* device_sync)(void *user_context, struct halide_buffer_t *buf)
int(* detach_native)(void *user_context, struct halide_buffer_t *buf)
int(* device_and_host_free)(void *user_context, struct halide_buffer_t *buf)
int(* device_malloc)(void *user_context, struct halide_buffer_t *buf, const struct halide_device_interface_t *device_interface)
int(* buffer_copy)(void *user_context, struct halide_buffer_t *src, const struct halide_device_interface_t *dst_device_interface, struct halide_buffer_t *dst)
Obsolete version of halide_filter_argument_t; only present in code that wrote halide_filter_metadata_...
const struct halide_scalar_value_t * min
const struct halide_scalar_value_t * def
const struct halide_scalar_value_t * max
struct halide_type_t type
halide_filter_argument_t is essentially a plain-C-struct equivalent to Halide::Argument; most user co...
const struct halide_scalar_value_t * scalar_estimate
const struct halide_scalar_value_t * scalar_max
int64_t const *const * buffer_estimates
const struct halide_scalar_value_t * scalar_def
struct halide_type_t type
const struct halide_scalar_value_t * scalar_min
const char * name
The function name of the filter.
int32_t version
version of this metadata; currently always 1.
const struct halide_filter_argument_t * arguments
An array of the filters input and output arguments; this will never be null.
int32_t num_arguments
The number of entries in the arguments field.
const char * target
The Target for which the filter was compiled.
A type traits template to provide a halide_handle_cplusplus_type value from a C++ type.
Definition: Type.h:243
struct halide_mutex * array
Cross-platform mutex.
uintptr_t _private[1]
A parallel task to be passed to halide_do_parallel_tasks.
struct halide_semaphore_acquire_t * semaphores
halide_loop_task_t fn
The functions below here are relevant for pipelines compiled with the -profile target flag,...
uint64_t memory_peak
The peak memory allocation of this Func.
const char * name
The name of this Func.
uint64_t stack_peak
The peak stack allocation of this Func's threads.
int num_allocs
The total number of memory allocation of this Func.
uint64_t active_threads_numerator
The average number of thread pool worker threads active while computing this Func.
uint64_t memory_total
The total memory allocation of this Func.
uint64_t time
Total time taken evaluating this Func (in nanoseconds).
uint64_t memory_current
The current memory allocation of this Func.
Per-pipeline state tracked by the sampling profiler.
int samples
The total number of samples taken inside of this pipeline.
uint64_t time
Total time spent inside this pipeline (in nanoseconds)
const char * name
The name of this pipeline.
int first_func_id
An internal base id used to identify the funcs in this pipeline.
int num_allocs
The total number of memory allocation of funcs in this pipeline.
void * next
The next pipeline_stats pointer.
uint64_t memory_current
The current memory allocation of funcs in this pipeline.
int num_funcs
The number of funcs in this pipeline.
uint64_t memory_peak
The peak memory allocation of funcs in this pipeline.
struct halide_profiler_func_stats * funcs
An array containing states for each Func in this pipeline.
int runs
The number of times this pipeline has been run.
uint64_t memory_total
The total memory allocation of funcs in this pipeline.
uint64_t active_threads_numerator
The average number of thread pool worker threads doing useful work while computing this pipeline.
The global state of the profiler.
void(* get_remote_profiler_state)(int *func, int *active_workers)
Retrieve remote profiler state.
struct halide_thread * sampling_thread
Sampling thread reference to be joined at shutdown.
int sleep_time
The amount of time the profiler thread sleeps between samples in milliseconds.
int current_func
The id of the current running Func.
int first_free_id
An internal id used for bookkeeping.
struct halide_profiler_pipeline_stats * pipelines
A linked list of stats gathered for each pipeline.
struct halide_mutex lock
Guards access to the fields below.
int active_threads
The number of threads currently doing work.
halide_scalar_value_t is a simple union able to represent all the well-known scalar values in a filte...
union halide_scalar_value_t::@3 u
A struct representing a semaphore and a number of items that must be acquired from it.
struct halide_semaphore_t * semaphore
An opaque struct representing a semaphore.
uint64_t _private[2]
void * value
If the event type is a load or a store, this points to the value being loaded or stored.
int32_t * coordinates
For loads and stores, an array which contains the location being accessed.
const char * func
The name of the Func or Pipeline that this event refers to.
const char * trace_tag
For halide_trace_tag, this points to a read-only null-terminated string of arbitrary text.
struct halide_type_t type
If the event type is a load or a store, this is the type of the data.
int32_t value_index
If this was a load or store of a Tuple-valued Func, this is which tuple element was accessed.
enum halide_trace_event_code_t event
The type of event.
int32_t dimensions
The length of the coordinates array.
The header of a packet in a binary trace.
uint32_t size
The total size of this packet in bytes.
int32_t id
The id of this packet (for the purpose of parent_id).
enum halide_trace_event_code_t event
struct halide_type_t type
The remaining fields are equivalent to those in halide_trace_event_t.
A runtime tag for a type in the halide type system.
uint8_t bits
The number of bits of precision of a single scalar value of this type.
uint16_t lanes
How many elements in a vector.
uint8_t code
The basic type code: signed integer, unsigned integer, or floating point.