Halide 16.0.0
Halide compiler and libraries
Loading...
Searching...
No Matches
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 <array>
7#include <cstddef>
8#include <cstdint>
9#include <cstring>
10#include <string_view>
11#else
12#include <stdbool.h>
13#include <stddef.h>
14#include <stdint.h>
15#include <string.h>
16#endif
17#else
18#include "runtime_internal.h"
19#endif
20
21#ifdef __cplusplus
22// Forward declare type to allow naming typed handles.
23// See Type.h for documentation.
24template<typename T>
26#endif
27
28#ifdef __cplusplus
29extern "C" {
30#endif
31
32#ifdef _MSC_VER
33// Note that (for MSVC) you should not use "inline" along with HALIDE_ALWAYS_INLINE;
34// it is not necessary, and may produce warnings for some build configurations.
35#define HALIDE_ALWAYS_INLINE __forceinline
36#define HALIDE_NEVER_INLINE __declspec(noinline)
37#else
38// Note that (for Posixy compilers) you should always use "inline" along with HALIDE_ALWAYS_INLINE;
39// otherwise some corner-case scenarios may erroneously report link errors.
40#define HALIDE_ALWAYS_INLINE inline __attribute__((always_inline))
41#define HALIDE_NEVER_INLINE __attribute__((noinline))
42#endif
43
44#ifndef HALIDE_MUST_USE_RESULT
45#ifdef __has_attribute
46#if __has_attribute(nodiscard)
47// C++17 or later
48#define HALIDE_MUST_USE_RESULT [[nodiscard]]
49#elif __has_attribute(warn_unused_result)
50// Clang/GCC
51#define HALIDE_MUST_USE_RESULT __attribute__((warn_unused_result))
52#else
53#define HALIDE_MUST_USE_RESULT
54#endif
55#else
56#define HALIDE_MUST_USE_RESULT
57#endif
58#endif
59
60// Annotation for AOT and JIT calls -- if undefined, use no annotation.
61// To ensure that all results are checked, do something like
62//
63// -DHALIDE_FUNCTION_ATTRS=HALIDE_MUST_USE_RESULT
64//
65// in your C++ compiler options
66#ifndef HALIDE_FUNCTION_ATTRS
67#define HALIDE_FUNCTION_ATTRS
68#endif
69
70#ifndef HALIDE_EXPORT_SYMBOL
71#ifdef _MSC_VER
72#define HALIDE_EXPORT_SYMBOL __declspec(dllexport)
73#else
74#define HALIDE_EXPORT_SYMBOL __attribute__((visibility("default")))
75#endif
76#endif
77
78#ifndef COMPILING_HALIDE_RUNTIME
79
80// clang had _Float16 added as a reserved name in clang 8, but
81// doesn't actually support it on most platforms until clang 15.
82// Ideally there would be a better way to detect if the type
83// is supported, even in a compiler independent fashion, but
84// coming up with one has proven elusive.
85#if defined(__clang__) && (__clang_major__ >= 16) && !defined(__EMSCRIPTEN__) && !defined(__i386__)
86#if defined(__is_identifier)
87#if !__is_identifier(_Float16)
88#define HALIDE_CPP_COMPILER_HAS_FLOAT16
89#endif
90#endif
91#endif
92
93// Similarly, detecting _Float16 for gcc is problematic.
94// For now, we say that if >= v12, and compiling on x86 or arm,
95// we assume support. This may need revision.
96#if defined(__GNUC__) && (__GNUC__ >= 12)
97#if defined(__x86_64__) || defined(__i386__) || defined(__arm__) || defined(__aarch64__)
98#define HALIDE_CPP_COMPILER_HAS_FLOAT16
99#endif
100#endif
101
102#endif // !COMPILING_HALIDE_RUNTIME
103
104/** \file
105 *
106 * This file declares the routines used by Halide internally in its
107 * runtime. On platforms that support weak linking, these can be
108 * replaced with user-defined versions by defining an extern "C"
109 * function with the same name and signature.
110 *
111 * When doing Just In Time (JIT) compilation members of
112 * some_pipeline_or_func.jit_handlers() must be replaced instead. The
113 * corresponding methods are documented below.
114 *
115 * All of these functions take a "void *user_context" parameter as their
116 * first argument; if the Halide kernel that calls back to any of these
117 * functions has been compiled with the UserContext feature set on its Target,
118 * then the value of that pointer passed from the code that calls the
119 * Halide kernel is piped through to the function.
120 *
121 * Some of these are also useful to call when using the default
122 * implementation. E.g. halide_shutdown_thread_pool.
123 *
124 * Note that even on platforms with weak linking, some linker setups
125 * may not respect the override you provide. E.g. if the override is
126 * in a shared library and the halide object files are linked directly
127 * into the output, the builtin versions of the runtime functions will
128 * be called. See your linker documentation for more details. On
129 * Linux, LD_DYNAMIC_WEAK=1 may help.
130 *
131 */
132
133// Forward-declare to suppress warnings if compiling as C.
134struct halide_buffer_t;
135
136/** Print a message to stderr. Main use is to support tracing
137 * functionality, print, and print_when calls. Also called by the default
138 * halide_error. This function can be replaced in JITed code by using
139 * halide_custom_print and providing an implementation of halide_print
140 * in AOT code. See Func::set_custom_print.
141 */
142// @{
143extern void halide_print(void *user_context, const char *);
144extern void halide_default_print(void *user_context, const char *);
145typedef void (*halide_print_t)(void *, const char *);
147// @}
148
149/** Halide calls this function on runtime errors (for example bounds
150 * checking failures). This function can be replaced in JITed code by
151 * using Func::set_error_handler, or in AOT code by calling
152 * halide_set_error_handler. In AOT code on platforms that support
153 * weak linking (i.e. not Windows), you can also override it by simply
154 * defining your own halide_error.
155 */
156// @{
157extern void halide_error(void *user_context, const char *);
158extern void halide_default_error(void *user_context, const char *);
159typedef void (*halide_error_handler_t)(void *, const char *);
161// @}
162
163/** Cross-platform mutex. Must be initialized with zero and implementation
164 * must treat zero as an unlocked mutex with no waiters, etc.
165 */
169
170/** Cross platform condition variable. Must be initialized to 0. */
174
175/** A basic set of mutex and condition variable functions, which call
176 * platform specific code for mutual exclusion. Equivalent to posix
177 * calls. */
178//@{
179extern void halide_mutex_lock(struct halide_mutex *mutex);
180extern void halide_mutex_unlock(struct halide_mutex *mutex);
181extern void halide_cond_signal(struct halide_cond *cond);
182extern void halide_cond_broadcast(struct halide_cond *cond);
183extern void halide_cond_wait(struct halide_cond *cond, struct halide_mutex *mutex);
184//@}
185
186/** Functions for constructing/destroying/locking/unlocking arrays of mutexes. */
187struct halide_mutex_array;
188//@{
189extern struct halide_mutex_array *halide_mutex_array_create(int sz);
190extern void halide_mutex_array_destroy(void *user_context, void *array);
191extern int halide_mutex_array_lock(struct halide_mutex_array *array, int entry);
192extern int halide_mutex_array_unlock(struct halide_mutex_array *array, int entry);
193//@}
194
195/** Define halide_do_par_for to replace the default thread pool
196 * implementation. halide_shutdown_thread_pool can also be called to
197 * release resources used by the default thread pool on platforms
198 * where it makes sense. See Func::set_custom_do_task and
199 * Func::set_custom_do_par_for. Should return zero if all the jobs
200 * return zero, or an arbitrarily chosen return value from one of the
201 * jobs otherwise.
202 */
203//@{
204typedef int (*halide_task_t)(void *user_context, int task_number, uint8_t *closure);
205extern int halide_do_par_for(void *user_context,
206 halide_task_t task,
207 int min, int size, uint8_t *closure);
208extern void halide_shutdown_thread_pool();
209//@}
210
211/** Set a custom method for performing a parallel for loop. Returns
212 * the old do_par_for handler. */
213typedef int (*halide_do_par_for_t)(void *, halide_task_t, int, int, uint8_t *);
215
216/** An opaque struct representing a semaphore. Used by the task system for async tasks. */
220
221/** A struct representing a semaphore and a number of items that must
222 * be acquired from it. Used in halide_parallel_task_t below. */
227extern int halide_semaphore_init(struct halide_semaphore_t *, int n);
228extern int halide_semaphore_release(struct halide_semaphore_t *, int n);
229extern bool halide_semaphore_try_acquire(struct halide_semaphore_t *, int n);
230typedef int (*halide_semaphore_init_t)(struct halide_semaphore_t *, int);
231typedef int (*halide_semaphore_release_t)(struct halide_semaphore_t *, int);
233
234/** A task representing a serial for loop evaluated over some range.
235 * Note that task_parent is a pass through argument that should be
236 * passed to any dependent taks that are invoked using halide_do_parallel_tasks
237 * underneath this call. */
238typedef int (*halide_loop_task_t)(void *user_context, int min, int extent,
239 uint8_t *closure, void *task_parent);
240
241/** A parallel task to be passed to halide_do_parallel_tasks. This
242 * task may recursively call halide_do_parallel_tasks, and there may
243 * be complex dependencies between seemingly unrelated tasks expressed
244 * using semaphores. If you are using a custom task system, care must
245 * be taken to avoid potential deadlock. This can be done by carefully
246 * respecting the static metadata at the end of the task struct.*/
248 // The function to call. It takes a user context, a min and
249 // extent, a closure, and a task system pass through argument.
251
252 // The closure to pass it
254
255 // The name of the function to be called. For debugging purposes only.
256 const char *name;
257
258 // An array of semaphores that must be acquired before the
259 // function is called. Must be reacquired for every call made.
262
263 // The entire range the function should be called over. This range
264 // may be sliced up and the function called multiple times.
266
267 // A parallel task provides several pieces of metadata to prevent
268 // unbounded resource usage or deadlock.
269
270 // The first is the minimum number of execution contexts (call
271 // stacks or threads) necessary for the function to run to
272 // completion. This may be greater than one when there is nested
273 // parallelism with internal producer-consumer relationships
274 // (calling the function recursively spawns and blocks on parallel
275 // sub-tasks that communicate with each other via semaphores). If
276 // a parallel runtime calls the function when fewer than this many
277 // threads are idle, it may need to create more threads to
278 // complete the task, or else risk deadlock due to committing all
279 // threads to tasks that cannot complete without more.
280 //
281 // FIXME: Note that extern stages are assumed to only require a
282 // single thread to complete. If the extern stage is itself a
283 // Halide pipeline, this may be an underestimate.
285
286 // The calls to the function should be in serial order from min to min+extent-1, with only
287 // one executing at a time. If false, any order is fine, and
288 // concurrency is fine.
289 bool serial;
290};
291
292/** Enqueue some number of the tasks described above and wait for them
293 * to complete. While waiting, the calling threads assists with either
294 * the tasks enqueued, or other non-blocking tasks in the task
295 * system. Note that task_parent should be NULL for top-level calls
296 * and the pass through argument if this call is being made from
297 * another task. */
298extern int halide_do_parallel_tasks(void *user_context, int num_tasks,
299 struct halide_parallel_task_t *tasks,
300 void *task_parent);
301
302/** If you use the default do_par_for, you can still set a custom
303 * handler to perform each individual task. Returns the old handler. */
304//@{
305typedef int (*halide_do_task_t)(void *, halide_task_t, int, uint8_t *);
307extern int halide_do_task(void *user_context, halide_task_t f, int idx,
308 uint8_t *closure);
309//@}
310
311/** The version of do_task called for loop tasks. By default calls the
312 * loop task with the same arguments. */
313// @{
314typedef int (*halide_do_loop_task_t)(void *, halide_loop_task_t, int, int, uint8_t *, void *);
316extern int halide_do_loop_task(void *user_context, halide_loop_task_t f, int min, int extent,
317 uint8_t *closure, void *task_parent);
318//@}
319
320/** Provide an entire custom tasking runtime via function
321 * pointers. Note that do_task and semaphore_try_acquire are only ever
322 * called by halide_default_do_par_for and
323 * halide_default_do_parallel_tasks, so it's only necessary to provide
324 * those if you are mixing in the default implementations of
325 * do_par_for and do_parallel_tasks. */
326// @{
327typedef int (*halide_do_parallel_tasks_t)(void *, int, struct halide_parallel_task_t *,
328 void *task_parent);
337// @}
338
339/** The default versions of the parallel runtime functions. */
340// @{
341extern int halide_default_do_par_for(void *user_context,
342 halide_task_t task,
343 int min, int size, uint8_t *closure);
344extern int halide_default_do_parallel_tasks(void *user_context,
345 int num_tasks,
346 struct halide_parallel_task_t *tasks,
347 void *task_parent);
348extern int halide_default_do_task(void *user_context, halide_task_t f, int idx,
350extern int halide_default_do_loop_task(void *user_context, halide_loop_task_t f,
351 int min, int extent,
352 uint8_t *closure, void *task_parent);
353extern int halide_default_semaphore_init(struct halide_semaphore_t *, int n);
354extern int halide_default_semaphore_release(struct halide_semaphore_t *, int n);
356// @}
357
358struct halide_thread;
359
360/** Spawn a thread. Returns a handle to the thread for the purposes of
361 * joining it. The thread must be joined in order to clean up any
362 * resources associated with it. */
363extern struct halide_thread *halide_spawn_thread(void (*f)(void *), void *closure);
364
365/** Join a thread. */
366extern void halide_join_thread(struct halide_thread *);
367
368/** Set the number of threads used by Halide's thread pool. Returns
369 * the old number.
370 *
371 * n < 0 : error condition
372 * n == 0 : use a reasonable system default (typically, number of cpus online).
373 * n == 1 : use exactly one thread; this will always enforce serial execution
374 * n > 1 : use a pool of exactly n threads.
375 *
376 * (Note that this is only guaranteed when using the default implementations
377 * of halide_do_par_for(); custom implementations may completely ignore values
378 * passed to halide_set_num_threads().)
379 */
380extern int halide_set_num_threads(int n);
381
382/** Halide calls these functions to allocate and free memory. To
383 * replace in AOT code, use the halide_set_custom_malloc and
384 * halide_set_custom_free, or (on platforms that support weak
385 * linking), simply define these functions yourself. In JIT-compiled
386 * code use Func::set_custom_allocator.
387 *
388 * If you override them, and find yourself wanting to call the default
389 * implementation from within your override, use
390 * halide_default_malloc/free.
391 *
392 * Note that halide_malloc must return a pointer aligned to the
393 * maximum meaningful alignment for the platform for the purpose of
394 * vector loads and stores, *and* with an allocated size that is (at least)
395 * an integral multiple of that same alignment. The default implementation
396 * uses 32-byte alignment on arm and 64-byte alignment on x86. Additionally,
397 * it must be safe to read at least 8 bytes before the start and beyond the end.
398 */
399//@{
400extern void *halide_malloc(void *user_context, size_t x);
401extern void halide_free(void *user_context, void *ptr);
402extern void *halide_default_malloc(void *user_context, size_t x);
403extern void halide_default_free(void *user_context, void *ptr);
404typedef void *(*halide_malloc_t)(void *, size_t);
405typedef void (*halide_free_t)(void *, void *);
408//@}
409
410/** Halide calls these functions to interact with the underlying
411 * system runtime functions. To replace in AOT code on platforms that
412 * support weak linking, define these functions yourself, or use
413 * the halide_set_custom_load_library() and halide_set_custom_get_library_symbol()
414 * functions. In JIT-compiled code, use JITSharedRuntime::set_default_handlers().
415 *
416 * halide_load_library and halide_get_library_symbol are equivalent to
417 * dlopen and dlsym. halide_get_symbol(sym) is equivalent to
418 * dlsym(RTLD_DEFAULT, sym).
419 */
420//@{
421extern void *halide_get_symbol(const char *name);
422extern void *halide_load_library(const char *name);
423extern void *halide_get_library_symbol(void *lib, const char *name);
424extern void *halide_default_get_symbol(const char *name);
425extern void *halide_default_load_library(const char *name);
426extern void *halide_default_get_library_symbol(void *lib, const char *name);
427typedef void *(*halide_get_symbol_t)(const char *name);
428typedef void *(*halide_load_library_t)(const char *name);
429typedef void *(*halide_get_library_symbol_t)(void *lib, const char *name);
433//@}
434
435/** Called when debug_to_file is used inside %Halide code. See
436 * Func::debug_to_file for how this is called
437 *
438 * Cannot be replaced in JITted code at present.
439 */
440extern int32_t halide_debug_to_file(void *user_context, const char *filename,
441 int32_t type_code,
442 struct halide_buffer_t *buf);
443
444/** Types in the halide type system. They can be ints, unsigned ints,
445 * or floats (of various bit-widths), or a handle (which is always 64-bits).
446 * Note that the int/uint/float values do not imply a specific bit width
447 * (the bit width is expected to be encoded in a separate value).
448 */
450#if (__cplusplus >= 201103L || _MSVC_LANG >= 201103L)
451 : uint8_t
452#endif
453{
454 halide_type_int = 0, ///< signed integers
455 halide_type_uint = 1, ///< unsigned integers
456 halide_type_float = 2, ///< IEEE floating point numbers
457 halide_type_handle = 3, ///< opaque pointer type (void *)
458 halide_type_bfloat = 4, ///< floating point numbers in the bfloat format
460
461// Note that while __attribute__ can go before or after the declaration,
462// __declspec apparently is only allowed before.
463#ifndef HALIDE_ATTRIBUTE_ALIGN
464#ifdef _MSC_VER
465#define HALIDE_ATTRIBUTE_ALIGN(x) __declspec(align(x))
466#else
467#define HALIDE_ATTRIBUTE_ALIGN(x) __attribute__((aligned(x)))
468#endif
469#endif
470
471/** A runtime tag for a type in the halide type system. Can be ints,
472 * unsigned ints, or floats of various bit-widths (the 'bits'
473 * field). Can also be vectors of the same (by setting the 'lanes'
474 * field to something larger than one). This struct should be
475 * exactly 32-bits in size. */
477 /** The basic type code: signed integer, unsigned integer, or floating point. */
478#if (__cplusplus >= 201103L || _MSVC_LANG >= 201103L)
480 halide_type_code_t code; // halide_type_code_t
481#else
483 uint8_t code; // halide_type_code_t
484#endif
485
486 /** The number of bits of precision of a single scalar value of this type. */
489
490 /** How many elements in a vector. This is 1 for scalar types. */
493
494#if (__cplusplus >= 201103L || _MSVC_LANG >= 201103L)
495 /** Construct a runtime representation of a Halide type from:
496 * code: The fundamental type from an enum.
497 * bits: The bit size of one element.
498 * lanes: The number of vector elements in the type. */
500 : code(code), bits(bits), lanes(lanes) {
501 }
502
503 /** Default constructor is required e.g. to declare halide_trace_event
504 * instances. */
506 : code((halide_type_code_t)0), bits(0), lanes(0) {
507 }
508
509 HALIDE_ALWAYS_INLINE constexpr halide_type_t with_lanes(uint16_t new_lanes) const {
510 return halide_type_t((halide_type_code_t)code, bits, new_lanes);
511 }
512
513 HALIDE_ALWAYS_INLINE constexpr halide_type_t element_of() const {
514 return with_lanes(1);
515 }
516 /** Compare two types for equality. */
517 HALIDE_ALWAYS_INLINE constexpr bool operator==(const halide_type_t &other) const {
518 return as_u32() == other.as_u32();
519 }
520
521 HALIDE_ALWAYS_INLINE constexpr bool operator!=(const halide_type_t &other) const {
522 return !(*this == other);
523 }
524
525 HALIDE_ALWAYS_INLINE constexpr bool operator<(const halide_type_t &other) const {
526 return as_u32() < other.as_u32();
527 }
528
529 /** Size in bytes for a single element, even if width is not 1, of this type. */
530 HALIDE_ALWAYS_INLINE constexpr int bytes() const {
531 return (bits + 7) / 8;
532 }
533
534 HALIDE_ALWAYS_INLINE constexpr uint32_t as_u32() const {
535 // Note that this produces a result that is identical to memcpy'ing 'this'
536 // into a u32 (on a little-endian machine, anyway), and at -O1 or greater
537 // on Clang, the compiler knows this and optimizes this into a single 32-bit move.
538 // (At -O0 it will look awful.)
539 return static_cast<uint8_t>(code) |
540 (static_cast<uint16_t>(bits) << 8) |
541 (static_cast<uint32_t>(lanes) << 16);
542 }
543#endif
544};
545
546#if (__cplusplus >= 201103L || _MSVC_LANG >= 201103L)
547static_assert(sizeof(halide_type_t) == sizeof(uint32_t), "size mismatch in halide_type_t");
548#endif
549
561
563 /** The name of the Func or Pipeline that this event refers to */
564 const char *func;
565
566 /** If the event type is a load or a store, this points to the
567 * value being loaded or stored. Use the type field to safely cast
568 * this to a concrete pointer type and retrieve it. For other
569 * events this is null. */
570 void *value;
571
572 /** For loads and stores, an array which contains the location
573 * being accessed. For vector loads or stores it is an array of
574 * vectors of coordinates (the vector dimension is innermost).
575 *
576 * For realization or production-related events, this will contain
577 * the mins and extents of the region being accessed, in the order
578 * min0, extent0, min1, extent1, ...
579 *
580 * For pipeline-related events, this will be null.
581 */
583
584 /** For halide_trace_tag, this points to a read-only null-terminated string
585 * of arbitrary text. For all other events, this will be null.
586 */
587 const char *trace_tag;
588
589 /** If the event type is a load or a store, this is the type of
590 * the data. Otherwise, the value is meaningless. */
592
593 /** The type of event */
595
596 /* The ID of the parent event (see below for an explanation of
597 * event ancestry). */
599
600 /** If this was a load or store of a Tuple-valued Func, this is
601 * which tuple element was accessed. */
603
604 /** The length of the coordinates array */
606};
607
608/** Called when Funcs are marked as trace_load, trace_store, or
609 * trace_realization. See Func::set_custom_trace. The default
610 * implementation either prints events via halide_print, or if
611 * HL_TRACE_FILE is defined, dumps the trace to that file in a
612 * sequence of trace packets. The header for a trace packet is defined
613 * below. If the trace is going to be large, you may want to make the
614 * file a named pipe, and then read from that pipe into gzip.
615 *
616 * halide_trace returns a unique ID which will be passed to future
617 * events that "belong" to the earlier event as the parent id. The
618 * ownership hierarchy looks like:
619 *
620 * begin_pipeline
621 * +--trace_tag (if any)
622 * +--trace_tag (if any)
623 * ...
624 * +--begin_realization
625 * | +--produce
626 * | | +--load/store
627 * | | +--end_produce
628 * | +--consume
629 * | | +--load
630 * | | +--end_consume
631 * | +--end_realization
632 * +--end_pipeline
633 *
634 * Threading means that ownership cannot be inferred from the ordering
635 * of events. There can be many active realizations of a given
636 * function, or many active productions for a single
637 * realization. Within a single production, the ordering of events is
638 * meaningful.
639 *
640 * Note that all trace_tag events (if any) will occur just after the begin_pipeline
641 * event, but before any begin_realization events. All trace_tags for a given Func
642 * will be emitted in the order added.
643 */
644// @}
645extern int32_t halide_trace(void *user_context, const struct halide_trace_event_t *event);
646extern int32_t halide_default_trace(void *user_context, const struct halide_trace_event_t *event);
647typedef int32_t (*halide_trace_t)(void *user_context, const struct halide_trace_event_t *);
649// @}
650
651/** The header of a packet in a binary trace. All fields are 32-bit. */
653 /** The total size of this packet in bytes. Always a multiple of
654 * four. Equivalently, the number of bytes until the next
655 * packet. */
657
658 /** The id of this packet (for the purpose of parent_id). */
660
661 /** The remaining fields are equivalent to those in halide_trace_event_t */
662 // @{
668 // @}
669
670#if (__cplusplus >= 201103L || _MSVC_LANG >= 201103L)
671 /** Get the coordinates array, assuming this packet is laid out in
672 * memory as it was written. The coordinates array comes
673 * immediately after the packet header. */
674 HALIDE_ALWAYS_INLINE const int *coordinates() const {
675 return (const int *)(this + 1);
676 }
677
678 HALIDE_ALWAYS_INLINE int *coordinates() {
679 return (int *)(this + 1);
680 }
681
682 /** Get the value, assuming this packet is laid out in memory as
683 * it was written. The packet comes immediately after the coordinates
684 * array. */
685 HALIDE_ALWAYS_INLINE const void *value() const {
686 return (const void *)(coordinates() + dimensions);
687 }
688
689 HALIDE_ALWAYS_INLINE void *value() {
690 return (void *)(coordinates() + dimensions);
691 }
692
693 /** Get the func name, assuming this packet is laid out in memory
694 * as it was written. It comes after the value. */
695 HALIDE_ALWAYS_INLINE const char *func() const {
696 return (const char *)value() + type.lanes * type.bytes();
697 }
698
699 HALIDE_ALWAYS_INLINE char *func() {
700 return (char *)value() + type.lanes * type.bytes();
701 }
702
703 /** Get the trace_tag (if any), assuming this packet is laid out in memory
704 * as it was written. It comes after the func name. If there is no trace_tag,
705 * this will return a pointer to an empty string. */
706 HALIDE_ALWAYS_INLINE const char *trace_tag() const {
707 const char *f = func();
708 // strlen may not be available here
709 while (*f++) {
710 // nothing
711 }
712 return f;
713 }
714
715 HALIDE_ALWAYS_INLINE char *trace_tag() {
716 char *f = func();
717 // strlen may not be available here
718 while (*f++) {
719 // nothing
720 }
721 return f;
722 }
723#endif
724};
725
726/** Set the file descriptor that Halide should write binary trace
727 * events to. If called with 0 as the argument, Halide outputs trace
728 * information to stdout in a human-readable format. If never called,
729 * Halide checks the for existence of an environment variable called
730 * HL_TRACE_FILE and opens that file. If HL_TRACE_FILE is not defined,
731 * it outputs trace information to stdout in a human-readable
732 * format. */
733extern void halide_set_trace_file(int fd);
734
735/** Halide calls this to retrieve the file descriptor to write binary
736 * trace events to. The default implementation returns the value set
737 * by halide_set_trace_file. Implement it yourself if you wish to use
738 * a custom file descriptor per user_context. Return zero from your
739 * implementation to tell Halide to print human-readable trace
740 * information to stdout. */
741extern int halide_get_trace_file(void *user_context);
742
743/** If tracing is writing to a file. This call closes that file
744 * (flushing the trace). Returns zero on success. */
746
747/** All Halide GPU or device backend implementations provide an
748 * interface to be used with halide_device_malloc, etc. This is
749 * accessed via the functions below.
750 */
751
752/** An opaque struct containing per-GPU API implementations of the
753 * device functions. */
755
756/** Each GPU API provides a halide_device_interface_t struct pointing
757 * to the code that manages device allocations. You can access these
758 * functions directly from the struct member function pointers, or by
759 * calling the functions declared below. Note that the global
760 * functions are not available when using Halide as a JIT compiler.
761 * If you are using raw halide_buffer_t in that context you must use
762 * the function pointers in the device_interface struct.
763 *
764 * The function pointers below are currently the same for every GPU
765 * API; only the impl field varies. These top-level functions do the
766 * bookkeeping that is common across all GPU APIs, and then dispatch
767 * to more API-specific functions via another set of function pointers
768 * hidden inside the impl field.
769 */
771 int (*device_malloc)(void *user_context, struct halide_buffer_t *buf,
772 const struct halide_device_interface_t *device_interface);
773 int (*device_free)(void *user_context, struct halide_buffer_t *buf);
774 int (*device_sync)(void *user_context, struct halide_buffer_t *buf);
775 void (*device_release)(void *user_context,
776 const struct halide_device_interface_t *device_interface);
777 int (*copy_to_host)(void *user_context, struct halide_buffer_t *buf);
778 int (*copy_to_device)(void *user_context, struct halide_buffer_t *buf,
779 const struct halide_device_interface_t *device_interface);
780 int (*device_and_host_malloc)(void *user_context, struct halide_buffer_t *buf,
781 const struct halide_device_interface_t *device_interface);
782 int (*device_and_host_free)(void *user_context, struct halide_buffer_t *buf);
783 int (*buffer_copy)(void *user_context, struct halide_buffer_t *src,
784 const struct halide_device_interface_t *dst_device_interface, struct halide_buffer_t *dst);
785 int (*device_crop)(void *user_context, const struct halide_buffer_t *src,
786 struct halide_buffer_t *dst);
787 int (*device_slice)(void *user_context, const struct halide_buffer_t *src,
788 int slice_dim, int slice_pos, struct halide_buffer_t *dst);
789 int (*device_release_crop)(void *user_context, struct halide_buffer_t *buf);
790 int (*wrap_native)(void *user_context, struct halide_buffer_t *buf, uint64_t handle,
791 const struct halide_device_interface_t *device_interface);
792 int (*detach_native)(void *user_context, struct halide_buffer_t *buf);
793 int (*compute_capability)(void *user_context, int *major, int *minor);
795};
796
797/** Release all data associated with the given device interface, in
798 * particular all resources (memory, texture, context handles)
799 * allocated by Halide. Must be called explicitly when using AOT
800 * compilation. This is *not* thread-safe with respect to actively
801 * running Halide code. Ensure all pipelines are finished before
802 * calling this. */
803extern void halide_device_release(void *user_context,
804 const struct halide_device_interface_t *device_interface);
805
806/** Copy image data from device memory to host memory. This must be called
807 * explicitly to copy back the results of a GPU-based filter. */
808extern int halide_copy_to_host(void *user_context, struct halide_buffer_t *buf);
809
810/** Copy image data from host memory to device memory. This should not
811 * be called directly; Halide handles copying to the device
812 * automatically. If interface is NULL and the buf has a non-zero dev
813 * field, the device associated with the dev handle will be
814 * used. Otherwise if the dev field is 0 and interface is NULL, an
815 * error is returned. */
816extern int halide_copy_to_device(void *user_context, struct halide_buffer_t *buf,
817 const struct halide_device_interface_t *device_interface);
818
819/** Copy data from one buffer to another. The buffers may have
820 * different shapes and sizes, but the destination buffer's shape must
821 * be contained within the source buffer's shape. That is, for each
822 * dimension, the min on the destination buffer must be greater than
823 * or equal to the min on the source buffer, and min+extent on the
824 * destination buffer must be less that or equal to min+extent on the
825 * source buffer. The source data is pulled from either device or
826 * host memory on the source, depending on the dirty flags. host is
827 * preferred if both are valid. The dst_device_interface parameter
828 * controls the destination memory space. NULL means host memory. */
829extern int halide_buffer_copy(void *user_context, struct halide_buffer_t *src,
830 const struct halide_device_interface_t *dst_device_interface,
831 struct halide_buffer_t *dst);
832
833/** Give the destination buffer a device allocation which is an alias
834 * for the same coordinate range in the source buffer. Modifies the
835 * device, device_interface, and the device_dirty flag only. Only
836 * supported by some device APIs (others will return
837 * halide_error_code_device_crop_unsupported). Call
838 * halide_device_release_crop instead of halide_device_free to clean
839 * up resources associated with the cropped view. Do not free the
840 * device allocation on the source buffer while the destination buffer
841 * still lives. Note that the two buffers do not share dirty flags, so
842 * care must be taken to update them together as needed. Note that src
843 * and dst are required to have the same number of dimensions.
844 *
845 * Note also that (in theory) device interfaces which support cropping may
846 * still not support cropping a crop (instead, create a new crop of the parent
847 * buffer); in practice, no known implementation has this limitation, although
848 * it is possible that some future implementations may require it. */
849extern int halide_device_crop(void *user_context,
850 const struct halide_buffer_t *src,
851 struct halide_buffer_t *dst);
852
853/** Give the destination buffer a device allocation which is an alias
854 * for a similar coordinate range in the source buffer, but with one dimension
855 * sliced away in the dst. Modifies the device, device_interface, and the
856 * device_dirty flag only. Only supported by some device APIs (others will return
857 * halide_error_code_device_crop_unsupported). Call
858 * halide_device_release_crop instead of halide_device_free to clean
859 * up resources associated with the sliced view. Do not free the
860 * device allocation on the source buffer while the destination buffer
861 * still lives. Note that the two buffers do not share dirty flags, so
862 * care must be taken to update them together as needed. Note that the dst buffer
863 * must have exactly one fewer dimension than the src buffer, and that slice_dim
864 * and slice_pos must be valid within src. */
865extern int halide_device_slice(void *user_context,
866 const struct halide_buffer_t *src,
867 int slice_dim, int slice_pos,
868 struct halide_buffer_t *dst);
869
870/** Release any resources associated with a cropped/sliced view of another
871 * buffer. */
872extern int halide_device_release_crop(void *user_context,
873 struct halide_buffer_t *buf);
874
875/** Wait for current GPU operations to complete. Calling this explicitly
876 * should rarely be necessary, except maybe for profiling. */
877extern int halide_device_sync(void *user_context, struct halide_buffer_t *buf);
878
879/** Allocate device memory to back a halide_buffer_t. */
880extern int halide_device_malloc(void *user_context, struct halide_buffer_t *buf,
881 const struct halide_device_interface_t *device_interface);
882
883/** Free device memory. */
884extern int halide_device_free(void *user_context, struct halide_buffer_t *buf);
885
886/** Wrap or detach a native device handle, setting the device field
887 * and device_interface field as appropriate for the given GPU
888 * API. The meaning of the opaque handle is specific to the device
889 * interface, so if you know the device interface in use, call the
890 * more specific functions in the runtime headers for your specific
891 * device API instead (e.g. HalideRuntimeCuda.h). */
892// @{
893extern int halide_device_wrap_native(void *user_context,
894 struct halide_buffer_t *buf,
895 uint64_t handle,
896 const struct halide_device_interface_t *device_interface);
897extern int halide_device_detach_native(void *user_context, struct halide_buffer_t *buf);
898// @}
899
900/** Selects which gpu device to use. 0 is usually the display
901 * device. If never called, Halide uses the environment variable
902 * HL_GPU_DEVICE. If that variable is unset, Halide uses the last
903 * device. Set this to -1 to use the last device. */
904extern void halide_set_gpu_device(int n);
905
906/** Halide calls this to get the desired halide gpu device
907 * setting. Implement this yourself to use a different gpu device per
908 * user_context. The default implementation returns the value set by
909 * halide_set_gpu_device, or the environment variable
910 * HL_GPU_DEVICE. */
911extern int halide_get_gpu_device(void *user_context);
912
913/** Set the soft maximum amount of memory, in bytes, that the LRU
914 * cache will use to memoize Func results. This is not a strict
915 * maximum in that concurrency and simultaneous use of memoized
916 * reults larger than the cache size can both cause it to
917 * temporariliy be larger than the size specified here.
918 */
920
921/** Given a cache key for a memoized result, currently constructed
922 * from the Func name and top-level Func name plus the arguments of
923 * the computation, determine if the result is in the cache and
924 * return it if so. (The internals of the cache key should be
925 * considered opaque by this function.) If this routine returns true,
926 * it is a cache miss. Otherwise, it will return false and the
927 * buffers passed in will be filled, via copying, with memoized
928 * data. The last argument is a list if halide_buffer_t pointers which
929 * represents the outputs of the memoized Func. If the Func does not
930 * return a Tuple, there will only be one halide_buffer_t in the list. The
931 * tuple_count parameters determines the length of the list.
932 *
933 * The return values are:
934 * -1: Signals an error.
935 * 0: Success and cache hit.
936 * 1: Success and cache miss.
937 */
938extern int halide_memoization_cache_lookup(void *user_context, const uint8_t *cache_key, int32_t size,
939 struct halide_buffer_t *realized_bounds,
940 int32_t tuple_count, struct halide_buffer_t **tuple_buffers);
941
942/** Given a cache key for a memoized result, currently constructed
943 * from the Func name and top-level Func name plus the arguments of
944 * the computation, store the result in the cache for futre access by
945 * halide_memoization_cache_lookup. (The internals of the cache key
946 * should be considered opaque by this function.) Data is copied out
947 * from the inputs and inputs are unmodified. The last argument is a
948 * list if halide_buffer_t pointers which represents the outputs of the
949 * memoized Func. If the Func does not return a Tuple, there will
950 * only be one halide_buffer_t in the list. The tuple_count parameters
951 * determines the length of the list.
952 *
953 * If there is a memory allocation failure, the store does not store
954 * the data into the cache.
955 *
956 * If has_eviction_key is true, the entry is marked with eviction_key to
957 * allow removing the key with halide_memoization_cache_evict.
958 */
959extern int halide_memoization_cache_store(void *user_context, const uint8_t *cache_key, int32_t size,
960 struct halide_buffer_t *realized_bounds,
961 int32_t tuple_count,
962 struct halide_buffer_t **tuple_buffers,
963 bool has_eviction_key, uint64_t eviction_key);
964
965/** Evict all cache entries that were tagged with the given
966 * eviction_key in the memoize scheduling directive.
967 */
968extern void halide_memoization_cache_evict(void *user_context, uint64_t eviction_key);
969
970/** If halide_memoization_cache_lookup succeeds,
971 * halide_memoization_cache_release must be called to signal the
972 * storage is no longer being used by the caller. It will be passed
973 * the host pointer of one the buffers returned by
974 * halide_memoization_cache_lookup. That is
975 * halide_memoization_cache_release will be called multiple times for
976 * the case where halide_memoization_cache_lookup is handling multiple
977 * buffers. (This corresponds to memoizing a Tuple in Halide.) Note
978 * that the host pointer must be sufficient to get to all information
979 * the release operation needs. The default Halide cache impleemntation
980 * accomplishes this by storing extra data before the start of the user
981 * modifiable host storage.
982 *
983 * This call is like free and does not have a failure return.
984 */
985extern void halide_memoization_cache_release(void *user_context, void *host);
986
987/** Free all memory and resources associated with the memoization cache.
988 * Must be called at a time when no other threads are accessing the cache.
989 */
991
992/** Verify that a given range of memory has been initialized; only used when Target::MSAN is enabled.
993 *
994 * The default implementation simply calls the LLVM-provided __msan_check_mem_is_initialized() function.
995 *
996 * The return value should always be zero.
997 */
998extern int halide_msan_check_memory_is_initialized(void *user_context, const void *ptr, uint64_t len, const char *name);
999
1000/** Verify that the data pointed to by the halide_buffer_t is initialized (but *not* the halide_buffer_t itself),
1001 * using halide_msan_check_memory_is_initialized() for checking.
1002 *
1003 * The default implementation takes pains to only check the active memory ranges
1004 * (skipping padding), and sorting into ranges to always check the smallest number of
1005 * ranges, in monotonically increasing memory order.
1006 *
1007 * Most client code should never need to replace the default implementation.
1008 *
1009 * The return value should always be zero.
1010 */
1011extern int halide_msan_check_buffer_is_initialized(void *user_context, struct halide_buffer_t *buffer, const char *buf_name);
1012
1013/** Annotate that a given range of memory has been initialized;
1014 * only used when Target::MSAN is enabled.
1015 *
1016 * The default implementation simply calls the LLVM-provided __msan_unpoison() function.
1017 *
1018 * The return value should always be zero.
1019 */
1020extern int halide_msan_annotate_memory_is_initialized(void *user_context, const void *ptr, uint64_t len);
1021
1022/** Mark the data pointed to by the halide_buffer_t as initialized (but *not* the halide_buffer_t itself),
1023 * using halide_msan_annotate_memory_is_initialized() for marking.
1024 *
1025 * The default implementation takes pains to only mark the active memory ranges
1026 * (skipping padding), and sorting into ranges to always mark the smallest number of
1027 * ranges, in monotonically increasing memory order.
1028 *
1029 * Most client code should never need to replace the default implementation.
1030 *
1031 * The return value should always be zero.
1032 */
1033extern int halide_msan_annotate_buffer_is_initialized(void *user_context, struct halide_buffer_t *buffer);
1034extern void halide_msan_annotate_buffer_is_initialized_as_destructor(void *user_context, void *buffer);
1035
1036/** The error codes that may be returned by a Halide pipeline. */
1038 /** There was no error. This is the value returned by Halide on success. */
1040
1041 /** An uncategorized error occurred. Refer to the string passed to halide_error. */
1043
1044 /** A Func was given an explicit bound via Func::bound, but this
1045 * was not large enough to encompass the region that is used of
1046 * the Func by the rest of the pipeline. */
1048
1049 /** The elem_size field of a halide_buffer_t does not match the size in
1050 * bytes of the type of that ImageParam. Probable type mismatch. */
1052
1053 /** A pipeline would access memory outside of the halide_buffer_t passed
1054 * in. */
1056
1057 /** A halide_buffer_t was given that spans more than 2GB of memory. */
1059
1060 /** A halide_buffer_t was given with extents that multiply to a number
1061 * greater than 2^31-1 */
1063
1064 /** Applying explicit constraints on the size of an input or
1065 * output buffer shrank the size of that buffer below what will be
1066 * accessed by the pipeline. */
1068
1069 /** A constraint on a size or stride of an input or output buffer
1070 * was not met by the halide_buffer_t passed in. */
1072
1073 /** A scalar parameter passed in was smaller than its minimum
1074 * declared value. */
1076
1077 /** A scalar parameter passed in was greater than its minimum
1078 * declared value. */
1080
1081 /** A call to halide_malloc returned NULL. */
1083
1084 /** A halide_buffer_t pointer passed in was NULL. */
1086
1087 /** debug_to_file failed to open or write to the specified
1088 * file. */
1090
1091 /** The Halide runtime encountered an error while trying to copy
1092 * from device to host. Turn on -debug in your target string to
1093 * see more details. */
1095
1096 /** The Halide runtime encountered an error while trying to copy
1097 * from host to device. Turn on -debug in your target string to
1098 * see more details. */
1100
1101 /** The Halide runtime encountered an error while trying to
1102 * allocate memory on device. Turn on -debug in your target string
1103 * to see more details. */
1105
1106 /** The Halide runtime encountered an error while trying to
1107 * synchronize with a device. Turn on -debug in your target string
1108 * to see more details. */
1110
1111 /** The Halide runtime encountered an error while trying to free a
1112 * device allocation. Turn on -debug in your target string to see
1113 * more details. */
1115
1116 /** Buffer has a non-zero device but no device interface, which
1117 * violates a Halide invariant. */
1119
1120 /** This part of the Halide runtime is unimplemented on this platform. */
1122
1123 /** A runtime symbol could not be loaded. */
1125
1126 /** There is a bug in the Halide compiler. */
1128
1129 /** The Halide runtime encountered an error while trying to launch
1130 * a GPU kernel. Turn on -debug in your target string to see more
1131 * details. */
1133
1134 /** The Halide runtime encountered a host pointer that violated
1135 * the alignment set for it by way of a call to
1136 * set_host_alignment */
1138
1139 /** A fold_storage directive was used on a dimension that is not
1140 * accessed in a monotonically increasing or decreasing fashion. */
1142
1143 /** A fold_storage directive was used with a fold factor that was
1144 * too small to store all the values of a producer needed by the
1145 * consumer. */
1147
1148 /** User-specified require() expression was not satisfied. */
1150
1151 /** At least one of the buffer's extents are negative. */
1153
1154 /** Call(s) to a GPU backend API failed. */
1156
1157 /** Failure recording trace packets for one of the halide_target_feature_trace features. */
1159
1160 /** A specialize_fail() schedule branch was selected at runtime. */
1162
1163 /** The Halide runtime encountered an error while trying to wrap a
1164 * native device handle. Turn on -debug in your target string to
1165 * see more details. */
1167
1168 /** The Halide runtime encountered an error while trying to detach
1169 * a native device handle. Turn on -debug in your target string
1170 * to see more details. */
1172
1173 /** The host field on an input or output was null, the device
1174 * field was not zero, and the pipeline tries to use the buffer on
1175 * the host. You may be passing a GPU-only buffer to a pipeline
1176 * which is scheduled to use it on the CPU. */
1178
1179 /** A folded buffer was passed to an extern stage, but the region
1180 * touched wraps around the fold boundary. */
1182
1183 /** Buffer has a non-null device_interface but device is 0, which
1184 * violates a Halide invariant. */
1186
1187 /** Buffer has both host and device dirty bits set, which violates
1188 * a Halide invariant. */
1190
1191 /** The halide_buffer_t * passed to a halide runtime routine is
1192 * nullptr and this is not allowed. */
1194
1195 /** The Halide runtime encountered an error while trying to copy
1196 * from one buffer to another. Turn on -debug in your target
1197 * string to see more details. */
1199
1200 /** Attempted to make cropped/sliced alias of a buffer with a device
1201 * field, but the device_interface does not support cropping. */
1203
1204 /** Cropping/slicing a buffer failed for some other reason. Turn on -debug
1205 * in your target string. */
1207
1208 /** An operation on a buffer required an allocation on a
1209 * particular device interface, but a device allocation already
1210 * existed on a different device interface. Free the old one
1211 * first. */
1213
1214 /** The dimensions field of a halide_buffer_t does not match the dimensions of that ImageParam. */
1216
1217 /** A buffer with the device_dirty flag set was passed to a
1218 * pipeline compiled with no device backends enabled, so it
1219 * doesn't know how to copy the data back from device memory to
1220 * host memory. Either call copy_to_host before calling the Halide
1221 * pipeline, or enable the appropriate device backend. */
1223
1224 /** An explicit storage bound provided is too small to store
1225 * all the values produced by the function. */
1227};
1228
1229/** Halide calls the functions below on various error conditions. The
1230 * default implementations construct an error message, call
1231 * halide_error, then return the matching error code above. On
1232 * platforms that support weak linking, you can override these to
1233 * catch the errors individually. */
1234
1235/** A call into an extern stage for the purposes of bounds inference
1236 * failed. Returns the error code given by the extern stage. */
1237extern int halide_error_bounds_inference_call_failed(void *user_context, const char *extern_stage_name, int result);
1238
1239/** A call to an extern stage failed. Returned the error code given by
1240 * the extern stage. */
1241extern int halide_error_extern_stage_failed(void *user_context, const char *extern_stage_name, int result);
1242
1243/** Various other error conditions. See the enum above for a
1244 * description of each. */
1245// @{
1246extern int halide_error_explicit_bounds_too_small(void *user_context, const char *func_name, const char *var_name,
1247 int min_bound, int max_bound, int min_required, int max_required);
1248extern int halide_error_bad_type(void *user_context, const char *func_name,
1249 uint32_t type_given, uint32_t correct_type); // N.B. The last two args are the bit representation of a halide_type_t
1250extern int halide_error_bad_dimensions(void *user_context, const char *func_name,
1251 int32_t dimensions_given, int32_t correct_dimensions);
1252extern int halide_error_access_out_of_bounds(void *user_context, const char *func_name,
1253 int dimension, int min_touched, int max_touched,
1254 int min_valid, int max_valid);
1255extern int halide_error_buffer_allocation_too_large(void *user_context, const char *buffer_name,
1256 uint64_t allocation_size, uint64_t max_size);
1257extern int halide_error_buffer_extents_negative(void *user_context, const char *buffer_name, int dimension, int extent);
1258extern int halide_error_buffer_extents_too_large(void *user_context, const char *buffer_name,
1259 int64_t actual_size, int64_t max_size);
1260extern int halide_error_constraints_make_required_region_smaller(void *user_context, const char *buffer_name,
1261 int dimension,
1262 int constrained_min, int constrained_extent,
1263 int required_min, int required_extent);
1264extern int halide_error_constraint_violated(void *user_context, const char *var, int val,
1265 const char *constrained_var, int constrained_val);
1266extern int halide_error_param_too_small_i64(void *user_context, const char *param_name,
1267 int64_t val, int64_t min_val);
1268extern int halide_error_param_too_small_u64(void *user_context, const char *param_name,
1269 uint64_t val, uint64_t min_val);
1270extern int halide_error_param_too_small_f64(void *user_context, const char *param_name,
1271 double val, double min_val);
1272extern int halide_error_param_too_large_i64(void *user_context, const char *param_name,
1273 int64_t val, int64_t max_val);
1274extern int halide_error_param_too_large_u64(void *user_context, const char *param_name,
1275 uint64_t val, uint64_t max_val);
1276extern int halide_error_param_too_large_f64(void *user_context, const char *param_name,
1277 double val, double max_val);
1278extern int halide_error_out_of_memory(void *user_context);
1279extern int halide_error_buffer_argument_is_null(void *user_context, const char *buffer_name);
1280extern int halide_error_debug_to_file_failed(void *user_context, const char *func,
1281 const char *filename, int error_code);
1282extern int halide_error_unaligned_host_ptr(void *user_context, const char *func_name, int alignment);
1283extern int halide_error_host_is_null(void *user_context, const char *func_name);
1284extern int halide_error_bad_fold(void *user_context, const char *func_name, const char *var_name,
1285 const char *loop_name);
1286extern int halide_error_bad_extern_fold(void *user_context, const char *func_name,
1287 int dim, int min, int extent, int valid_min, int fold_factor);
1288
1289extern int halide_error_fold_factor_too_small(void *user_context, const char *func_name, const char *var_name,
1290 int fold_factor, const char *loop_name, int required_extent);
1291extern int halide_error_requirement_failed(void *user_context, const char *condition, const char *message);
1292extern int halide_error_specialize_fail(void *user_context, const char *message);
1293extern int halide_error_no_device_interface(void *user_context);
1294extern int halide_error_device_interface_no_device(void *user_context);
1295extern int halide_error_host_and_device_dirty(void *user_context);
1296extern int halide_error_buffer_is_null(void *user_context, const char *routine);
1297extern int halide_error_device_dirty_with_no_device_support(void *user_context, const char *buffer_name);
1298extern int halide_error_storage_bound_too_small(void *user_context, const char *func_name, const char *var_name,
1299 int provided_size, int required_size);
1300extern int halide_error_device_crop_failed(void *user_context);
1301// @}
1302
1303/** Optional features a compilation Target can have.
1304 * Be sure to keep this in sync with the Feature enum in Target.h and the implementation of
1305 * get_runtime_compatible_target in Target.cpp if you add a new feature.
1306 */
1308 halide_target_feature_jit = 0, ///< Generate code that will run immediately inside the calling process.
1309 halide_target_feature_debug, ///< Turn on debug info and output for runtime code.
1310 halide_target_feature_no_asserts, ///< Disable all runtime checks, for slightly tighter code.
1311 halide_target_feature_no_bounds_query, ///< Disable the bounds querying functionality.
1312
1313 halide_target_feature_sse41, ///< Use SSE 4.1 and earlier instructions. Only relevant on x86.
1314 halide_target_feature_avx, ///< Use AVX 1 instructions. Only relevant on x86.
1315 halide_target_feature_avx2, ///< Use AVX 2 instructions. Only relevant on x86.
1316 halide_target_feature_fma, ///< Enable x86 FMA instruction
1317 halide_target_feature_fma4, ///< Enable x86 (AMD) FMA4 instruction set
1318 halide_target_feature_f16c, ///< Enable x86 16-bit float support
1319
1320 halide_target_feature_armv7s, ///< Generate code for ARMv7s. Only relevant for 32-bit ARM.
1321 halide_target_feature_no_neon, ///< Avoid using NEON instructions. Only relevant for 32-bit ARM.
1322
1323 halide_target_feature_vsx, ///< Use VSX instructions. Only relevant on POWERPC.
1324 halide_target_feature_power_arch_2_07, ///< Use POWER ISA 2.07 new instructions. Only relevant on POWERPC.
1325
1326 halide_target_feature_cuda, ///< Enable the CUDA runtime. Defaults to compute capability 2.0 (Fermi)
1327 halide_target_feature_cuda_capability30, ///< Enable CUDA compute capability 3.0 (Kepler)
1328 halide_target_feature_cuda_capability32, ///< Enable CUDA compute capability 3.2 (Tegra K1)
1329 halide_target_feature_cuda_capability35, ///< Enable CUDA compute capability 3.5 (Kepler)
1330 halide_target_feature_cuda_capability50, ///< Enable CUDA compute capability 5.0 (Maxwell)
1331 halide_target_feature_cuda_capability61, ///< Enable CUDA compute capability 6.1 (Pascal)
1332 halide_target_feature_cuda_capability70, ///< Enable CUDA compute capability 7.0 (Volta)
1333 halide_target_feature_cuda_capability75, ///< Enable CUDA compute capability 7.5 (Turing)
1334 halide_target_feature_cuda_capability80, ///< Enable CUDA compute capability 8.0 (Ampere)
1335 halide_target_feature_cuda_capability86, ///< Enable CUDA compute capability 8.6 (Ampere)
1336
1337 halide_target_feature_opencl, ///< Enable the OpenCL runtime.
1338 halide_target_feature_cl_doubles, ///< Enable double support on OpenCL targets
1339 halide_target_feature_cl_atomic64, ///< Enable 64-bit atomics operations on OpenCL targets
1340
1341 halide_target_feature_openglcompute, ///< Enable OpenGL Compute runtime. NOTE: This feature is deprecated and will be removed in Halide 17.
1342
1343 halide_target_feature_user_context, ///< Generated code takes a user_context pointer as first argument
1344
1345 halide_target_feature_profile, ///< Launch a sampling profiler alongside the Halide pipeline that monitors and reports the runtime used by each Func
1346 halide_target_feature_no_runtime, ///< Do not include a copy of the Halide runtime in any generated object file or assembly
1347
1348 halide_target_feature_metal, ///< Enable the (Apple) Metal runtime.
1349
1350 halide_target_feature_c_plus_plus_mangling, ///< Generate C++ mangled names for result function, et al
1351
1352 halide_target_feature_large_buffers, ///< Enable 64-bit buffer indexing to support buffers > 2GB. Ignored if bits != 64.
1353
1354 halide_target_feature_hvx_128, ///< Enable HVX 128 byte mode.
1355 halide_target_feature_hvx_v62, ///< Enable Hexagon v62 architecture.
1356 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.
1357 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.
1358 halide_target_feature_msan, ///< Enable hooks for MSAN support.
1359 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.
1360 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.
1361 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.
1362 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.
1363 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.
1364 halide_target_feature_trace_loads, ///< Trace all loads done by the pipeline. Equivalent to calling Func::trace_loads on every non-inlined Func.
1365 halide_target_feature_trace_stores, ///< Trace all stores done by the pipeline. Equivalent to calling Func::trace_stores on every non-inlined Func.
1366 halide_target_feature_trace_realizations, ///< Trace all realizations done by the pipeline. Equivalent to calling Func::trace_realizations on every non-inlined Func.
1368 halide_target_feature_hvx_v65, ///< Enable Hexagon v65 architecture.
1369 halide_target_feature_hvx_v66, ///< Enable Hexagon v66 architecture.
1370 halide_target_feature_cl_half, ///< Enable half support on OpenCL targets
1371 halide_target_feature_strict_float, ///< Turn off all non-IEEE floating-point optimization. Currently applies only to LLVM targets.
1372 halide_target_feature_tsan, ///< Enable hooks for TSAN support.
1373 halide_target_feature_asan, ///< Enable hooks for ASAN support.
1374 halide_target_feature_d3d12compute, ///< Enable Direct3D 12 Compute runtime.
1375 halide_target_feature_check_unsafe_promises, ///< Insert assertions for promises.
1376 halide_target_feature_hexagon_dma, ///< Enable Hexagon DMA buffers.
1377 halide_target_feature_embed_bitcode, ///< Emulate clang -fembed-bitcode flag.
1378 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.)
1379 halide_target_feature_wasm_simd128, ///< Enable +simd128 instructions for WebAssembly codegen.
1380 halide_target_feature_wasm_signext, ///< Enable +sign-ext instructions for WebAssembly codegen.
1381 halide_target_feature_wasm_sat_float_to_int, ///< Enable saturating (nontrapping) float-to-int instructions for WebAssembly codegen.
1382 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.
1383 halide_target_feature_wasm_bulk_memory, ///< Enable +bulk-memory instructions for WebAssembly codegen.
1384 halide_target_feature_webgpu, ///< Enable the WebGPU runtime.
1385 halide_target_feature_sve, ///< Enable ARM Scalable Vector Extensions
1386 halide_target_feature_sve2, ///< Enable ARM Scalable Vector Extensions v2
1387 halide_target_feature_egl, ///< Force use of EGL support.
1388 halide_target_feature_arm_dot_prod, ///< Enable ARMv8.2-a dotprod extension (i.e. udot and sdot instructions)
1389 halide_target_feature_arm_fp16, ///< Enable ARMv8.2-a half-precision floating point data processing
1390 halide_llvm_large_code_model, ///< Use the LLVM large code model to compile
1391 halide_target_feature_rvv, ///< Enable RISCV "V" Vector Extension
1392 halide_target_feature_armv81a, ///< Enable ARMv8.1-a instructions
1393 halide_target_feature_sanitizer_coverage, ///< Enable hooks for SanitizerCoverage support.
1394 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.
1395 halide_target_feature_spirv, ///< Enable SPIR-V code generation support.
1396 halide_target_feature_vulkan, ///< Enable Vulkan runtime support.
1397 halide_target_feature_vulkan_int8, ///< Enable Vulkan 8-bit integer support.
1398 halide_target_feature_vulkan_int16, ///< Enable Vulkan 16-bit integer support.
1399 halide_target_feature_vulkan_int64, ///< Enable Vulkan 64-bit integer support.
1400 halide_target_feature_vulkan_float16, ///< Enable Vulkan 16-bit float support.
1401 halide_target_feature_vulkan_float64, ///< Enable Vulkan 64-bit float support.
1402 halide_target_feature_vulkan_version10, ///< Enable Vulkan v1.0 runtime target support.
1403 halide_target_feature_vulkan_version12, ///< Enable Vulkan v1.2 runtime target support.
1404 halide_target_feature_vulkan_version13, ///< Enable Vulkan v1.3 runtime target support.
1405 halide_target_feature_semihosting, ///< Used together with Target::NoOS for the baremetal target built with semihosting library and run with semihosting mode where minimum I/O communication with a host PC is available.
1406 halide_target_feature_end ///< A sentinel. Every target is considered to have this feature, and setting this feature does nothing.
1408
1409/** This function is called internally by Halide in some situations to determine
1410 * if the current execution environment can support the given set of
1411 * halide_target_feature_t flags. The implementation must do the following:
1412 *
1413 * -- If there are flags set in features that the function knows *cannot* be supported, return 0.
1414 * -- Otherwise, return 1.
1415 * -- Note that any flags set in features that the function doesn't know how to test should be ignored;
1416 * this implies that a return value of 1 means "not known to be bad" rather than "known to be good".
1417 *
1418 * In other words: a return value of 0 means "It is not safe to use code compiled with these features",
1419 * while a return value of 1 means "It is not obviously unsafe to use code compiled with these features".
1420 *
1421 * The default implementation simply calls halide_default_can_use_target_features.
1422 *
1423 * Note that `features` points to an array of `count` uint64_t; this array must contain enough
1424 * bits to represent all the currently known features. Any excess bits must be set to zero.
1425 */
1426// @{
1427extern int halide_can_use_target_features(int count, const uint64_t *features);
1428typedef int (*halide_can_use_target_features_t)(int count, const uint64_t *features);
1430// @}
1431
1432/**
1433 * This is the default implementation of halide_can_use_target_features; it is provided
1434 * for convenience of user code that may wish to extend halide_can_use_target_features
1435 * but continue providing existing support, e.g.
1436 *
1437 * int halide_can_use_target_features(int count, const uint64_t *features) {
1438 * if (features[halide_target_somefeature >> 6] & (1LL << (halide_target_somefeature & 63))) {
1439 * if (!can_use_somefeature()) {
1440 * return 0;
1441 * }
1442 * }
1443 * return halide_default_can_use_target_features(count, features);
1444 * }
1445 */
1446extern int halide_default_can_use_target_features(int count, const uint64_t *features);
1447
1448typedef struct halide_dimension_t {
1449#if (__cplusplus >= 201103L || _MSVC_LANG >= 201103L)
1450 int32_t min = 0, extent = 0, stride = 0;
1451
1452 // Per-dimension flags. None are defined yet (This is reserved for future use).
1453 uint32_t flags = 0;
1454
1457 : min(m), extent(e), stride(s), flags(f) {
1458 }
1459
1460 HALIDE_ALWAYS_INLINE bool operator==(const halide_dimension_t &other) const {
1461 return (min == other.min) &&
1462 (extent == other.extent) &&
1463 (stride == other.stride) &&
1464 (flags == other.flags);
1465 }
1466
1467 HALIDE_ALWAYS_INLINE bool operator!=(const halide_dimension_t &other) const {
1468 return !(*this == other);
1469 }
1470#else
1472
1473 // Per-dimension flags. None are defined yet (This is reserved for future use).
1475#endif
1477
1478#ifdef __cplusplus
1479} // extern "C"
1480#endif
1481
1484
1485/**
1486 * The raw representation of an image passed around by generated
1487 * Halide code. It includes some stuff to track whether the image is
1488 * not actually in main memory, but instead on a device (like a
1489 * GPU). For a more convenient C++ wrapper, use Halide::Buffer<T>. */
1490typedef struct halide_buffer_t {
1491 /** A device-handle for e.g. GPU memory used to back this buffer. */
1493
1494 /** The interface used to interpret the above handle. */
1496
1497 /** A pointer to the start of the data in main memory. In terms of
1498 * the Halide coordinate system, this is the address of the min
1499 * coordinates (defined below). */
1501
1502 /** flags with various meanings. */
1504
1505 /** The type of each buffer element. */
1507
1508 /** The dimensionality of the buffer. */
1510
1511 /** The shape of the buffer. Halide does not own this array - you
1512 * must manage the memory for it yourself. */
1514
1515 /** Pads the buffer up to a multiple of 8 bytes */
1516 void *padding;
1517
1518#if (__cplusplus >= 201103L || _MSVC_LANG >= 201103L)
1519 /** Convenience methods for accessing the flags */
1520 // @{
1521 HALIDE_ALWAYS_INLINE bool get_flag(halide_buffer_flags flag) const {
1522 return (flags & flag) != 0;
1523 }
1524
1525 HALIDE_ALWAYS_INLINE void set_flag(halide_buffer_flags flag, bool value) {
1526 if (value) {
1527 flags |= flag;
1528 } else {
1529 flags &= ~uint64_t(flag);
1530 }
1531 }
1532
1533 HALIDE_ALWAYS_INLINE bool host_dirty() const {
1534 return get_flag(halide_buffer_flag_host_dirty);
1535 }
1536
1537 HALIDE_ALWAYS_INLINE bool device_dirty() const {
1538 return get_flag(halide_buffer_flag_device_dirty);
1539 }
1540
1541 HALIDE_ALWAYS_INLINE void set_host_dirty(bool v = true) {
1542 set_flag(halide_buffer_flag_host_dirty, v);
1543 }
1544
1545 HALIDE_ALWAYS_INLINE void set_device_dirty(bool v = true) {
1547 }
1548 // @}
1549
1550 /** The total number of elements this buffer represents. Equal to
1551 * the product of the extents */
1552 HALIDE_ALWAYS_INLINE size_t number_of_elements() const {
1553 size_t s = 1;
1554 for (int i = 0; i < dimensions; i++) {
1555 s *= dim[i].extent;
1556 }
1557 return s;
1558 }
1559
1560 /** Offset to the element with the lowest address.
1561 * If all strides are positive, equal to zero.
1562 * Offset is in elements, not bytes.
1563 * Unlike begin(), this is ok to call on an unallocated buffer. */
1564 HALIDE_ALWAYS_INLINE ptrdiff_t begin_offset() const {
1565 ptrdiff_t index = 0;
1566 for (int i = 0; i < dimensions; i++) {
1567 const int stride = dim[i].stride;
1568 if (stride < 0) {
1569 index += stride * (ptrdiff_t)(dim[i].extent - 1);
1570 }
1571 }
1572 return index;
1573 }
1574
1575 /** An offset to one beyond the element with the highest address.
1576 * Offset is in elements, not bytes.
1577 * Unlike end(), this is ok to call on an unallocated buffer. */
1578 HALIDE_ALWAYS_INLINE ptrdiff_t end_offset() const {
1579 ptrdiff_t index = 0;
1580 for (int i = 0; i < dimensions; i++) {
1581 const int stride = dim[i].stride;
1582 if (stride > 0) {
1583 index += stride * (ptrdiff_t)(dim[i].extent - 1);
1584 }
1585 }
1586 index += 1;
1587 return index;
1588 }
1589
1590 /** A pointer to the element with the lowest address.
1591 * If all strides are positive, equal to the host pointer.
1592 * Illegal to call on an unallocated buffer. */
1593 HALIDE_ALWAYS_INLINE uint8_t *begin() const {
1594 return host + begin_offset() * type.bytes();
1595 }
1596
1597 /** A pointer to one beyond the element with the highest address.
1598 * Illegal to call on an unallocated buffer. */
1599 HALIDE_ALWAYS_INLINE uint8_t *end() const {
1600 return host + end_offset() * type.bytes();
1601 }
1602
1603 /** The total number of bytes spanned by the data in memory. */
1604 HALIDE_ALWAYS_INLINE size_t size_in_bytes() const {
1605 return (size_t)(end_offset() - begin_offset()) * type.bytes();
1606 }
1607
1608 /** A pointer to the element at the given location. */
1609 HALIDE_ALWAYS_INLINE uint8_t *address_of(const int *pos) const {
1610 ptrdiff_t index = 0;
1611 for (int i = 0; i < dimensions; i++) {
1612 index += (ptrdiff_t)dim[i].stride * (pos[i] - dim[i].min);
1613 }
1614 return host + index * type.bytes();
1615 }
1616
1617 /** Attempt to call device_sync for the buffer. If the buffer
1618 * has no device_interface (or no device_sync), this is a quiet no-op.
1619 * Calling this explicitly should rarely be necessary, except for profiling. */
1620 HALIDE_ALWAYS_INLINE int device_sync(void *ctx = nullptr) {
1622 return device_interface->device_sync(ctx, this);
1623 }
1624 return 0;
1625 }
1626
1627 /** Check if an input buffer passed extern stage is a querying
1628 * bounds. Compared to doing the host pointer check directly,
1629 * this both adds clarity to code and will facilitate moving to
1630 * another representation for bounds query arguments. */
1631 HALIDE_ALWAYS_INLINE bool is_bounds_query() const {
1632 return host == nullptr && device == 0;
1633 }
1634
1635#endif
1637
1638#ifdef __cplusplus
1639extern "C" {
1640#endif
1641
1642#ifndef HALIDE_ATTRIBUTE_DEPRECATED
1643#ifdef HALIDE_ALLOW_DEPRECATED
1644#define HALIDE_ATTRIBUTE_DEPRECATED(x)
1645#else
1646#ifdef _MSC_VER
1647#define HALIDE_ATTRIBUTE_DEPRECATED(x) __declspec(deprecated(x))
1648#else
1649#define HALIDE_ATTRIBUTE_DEPRECATED(x) __attribute__((deprecated(x)))
1650#endif
1651#endif
1652#endif
1653
1654/** halide_scalar_value_t is a simple union able to represent all the well-known
1655 * scalar values in a filter argument. Note that it isn't tagged with a type;
1656 * you must ensure you know the proper type before accessing. Most user
1657 * code will never need to create instances of this struct; its primary use
1658 * is to hold def/min/max values in a halide_filter_argument_t. (Note that
1659 * this is conceptually just a union; it's wrapped in a struct to ensure
1660 * that it doesn't get anonymized by LLVM.)
1661 */
1663 union {
1664 bool b;
1673 float f32;
1674 double f64;
1675 void *handle;
1676 } u;
1677#ifdef __cplusplus
1679 u.u64 = 0;
1680 }
1681#endif
1682};
1683
1689
1690/*
1691 These structs must be robust across different compilers and settings; when
1692 modifying them, strive for the following rules:
1693
1694 1) All fields are explicitly sized. I.e. must use int32_t and not "int"
1695 2) All fields must land on an alignment boundary that is the same as their size
1696 3) Explicit padding is added to make that so
1697 4) The sizeof the struct is padded out to a multiple of the largest natural size thing in the struct
1698 5) don't forget that 32 and 64 bit pointers are different sizes
1699*/
1700
1701/**
1702 * Obsolete version of halide_filter_argument_t; only present in
1703 * code that wrote halide_filter_metadata_t version 0.
1704 */
1712
1713/**
1714 * halide_filter_argument_t is essentially a plain-C-struct equivalent to
1715 * Halide::Argument; most user code will never need to create one.
1716 */
1718 const char *name; // name of the argument; will never be null or empty.
1719 int32_t kind; // actually halide_argument_kind_t
1720 int32_t dimensions; // always zero for scalar arguments
1722 // These pointers should always be null for buffer arguments,
1723 // and *may* be null for scalar arguments. (A null value means
1724 // there is no def/min/max/estimate specified for this argument.)
1726 // This pointer should always be null for scalar arguments,
1727 // and *may* be null for buffer arguments. If not null, it should always
1728 // point to an array of dimensions*2 pointers, which will be the (min, extent)
1729 // estimates for each dimension of the buffer. (Note that any of the pointers
1730 // may be null as well.)
1732};
1733
1735#ifdef __cplusplus
1736 static const int32_t VERSION = 1;
1737#endif
1738
1739 /** version of this metadata; currently always 1. */
1741
1742 /** The number of entries in the arguments field. This is always >= 1. */
1744
1745 /** An array of the filters input and output arguments; this will never be
1746 * null. The order of arguments is not guaranteed (input and output arguments
1747 * may come in any order); however, it is guaranteed that all arguments
1748 * will have a unique name within a given filter. */
1750
1751 /** The Target for which the filter was compiled. This is always
1752 * a canonical Target string (ie a product of Target::to_string). */
1753 const char *target;
1754
1755 /** The function name of the filter. */
1756 const char *name;
1757};
1758
1759/** halide_register_argv_and_metadata() is a **user-defined** function that
1760 * must be provided in order to use the registration.cc files produced
1761 * by Generators when the 'registration' output is requested. Each registration.cc
1762 * file provides a static initializer that calls this function with the given
1763 * filter's argv-call variant, its metadata, and (optionally) and additional
1764 * textual data that the build system chooses to tack on for its own purposes.
1765 * Note that this will be called at static-initializer time (i.e., before
1766 * main() is called), and in an unpredictable order. Note that extra_key_value_pairs
1767 * may be nullptr; if it's not null, it's expected to be a null-terminated list
1768 * of strings, with an even number of entries. */
1770 int (*filter_argv_call)(void **),
1771 const struct halide_filter_metadata_t *filter_metadata,
1772 const char *const *extra_key_value_pairs);
1773
1774/** The functions below here are relevant for pipelines compiled with
1775 * the -profile target flag, which runs a sampling profiler thread
1776 * alongside the pipeline. */
1777
1778/** Per-Func state tracked by the sampling profiler. */
1779struct HALIDE_ATTRIBUTE_ALIGN(8) halide_profiler_func_stats {
1780 /** Total time taken evaluating this Func (in nanoseconds). */
1781 uint64_t time;
1782
1783 /** The current memory allocation of this Func. */
1784 uint64_t memory_current;
1785
1786 /** The peak memory allocation of this Func. */
1787 uint64_t memory_peak;
1788
1789 /** The total memory allocation of this Func. */
1790 uint64_t memory_total;
1791
1792 /** The peak stack allocation of this Func's threads. */
1793 uint64_t stack_peak;
1794
1795 /** The average number of thread pool worker threads active while computing this Func. */
1796 uint64_t active_threads_numerator, active_threads_denominator;
1797
1798 /** The name of this Func. A global constant string. */
1799 const char *name;
1800
1801 /** The total number of memory allocation of this Func. */
1802 int num_allocs;
1803};
1804
1805/** Per-pipeline state tracked by the sampling profiler. These exist
1806 * in a linked list. */
1807struct HALIDE_ATTRIBUTE_ALIGN(8) halide_profiler_pipeline_stats {
1808 /** Total time spent inside this pipeline (in nanoseconds) */
1809 uint64_t time;
1810
1811 /** The current memory allocation of funcs in this pipeline. */
1812 uint64_t memory_current;
1813
1814 /** The peak memory allocation of funcs in this pipeline. */
1815 uint64_t memory_peak;
1816
1817 /** The total memory allocation of funcs in this pipeline. */
1818 uint64_t memory_total;
1819
1820 /** The average number of thread pool worker threads doing useful
1821 * work while computing this pipeline. */
1822 uint64_t active_threads_numerator, active_threads_denominator;
1823
1824 /** The name of this pipeline. A global constant string. */
1825 const char *name;
1826
1827 /** An array containing states for each Func in this pipeline. */
1828 struct halide_profiler_func_stats *funcs;
1829
1830 /** The next pipeline_stats pointer. It's a void * because types
1831 * in the Halide runtime may not currently be recursive. */
1832 void *next;
1833
1834 /** The number of funcs in this pipeline. */
1835 int num_funcs;
1836
1837 /** An internal base id used to identify the funcs in this pipeline. */
1838 int first_func_id;
1839
1840 /** The number of times this pipeline has been run. */
1841 int runs;
1842
1843 /** The total number of samples taken inside of this pipeline. */
1844 int samples;
1845
1846 /** The total number of memory allocation of funcs in this pipeline. */
1847 int num_allocs;
1848};
1849
1850/** The global state of the profiler. */
1851
1853 /** Guards access to the fields below. If not locked, the sampling
1854 * profiler thread is free to modify things below (including
1855 * reordering the linked list of pipeline stats). */
1857
1858 /** The amount of time the profiler thread sleeps between samples
1859 * in milliseconds. Defaults to 1 */
1861
1862 /** An internal id used for bookkeeping. */
1864
1865 /** The id of the current running Func. Set by the pipeline, read
1866 * periodically by the profiler thread. */
1868
1869 /** The number of threads currently doing work. */
1871
1872 /** A linked list of stats gathered for each pipeline. */
1873 struct halide_profiler_pipeline_stats *pipelines;
1874
1875 /** Retrieve remote profiler state. Used so that the sampling
1876 * profiler can follow along with execution that occurs elsewhere,
1877 * e.g. on a DSP. If null, it reads from the int above instead. */
1878 void (*get_remote_profiler_state)(int *func, int *active_workers);
1879
1880 /** Sampling thread reference to be joined at shutdown. */
1881 struct halide_thread *sampling_thread;
1882};
1883
1884/** Profiler func ids with special meanings. */
1885enum {
1886 /// current_func takes on this value when not inside Halide code
1888 /// Set current_func to this value to tell the profiling thread to
1889 /// halt. It will start up again next time you run a pipeline with
1890 /// profiling enabled.
1893
1894/** Get a pointer to the global profiler state for programmatic
1895 * inspection. Lock it before using to pause the profiler. */
1897
1898/** Get a pointer to the pipeline state associated with pipeline_name.
1899 * This function grabs the global profiler state's lock on entry. */
1900extern struct halide_profiler_pipeline_stats *halide_profiler_get_pipeline_state(const char *pipeline_name);
1901
1902/** Collects profiling information. Intended to be called from a timer
1903 * interrupt handler if timer based profiling is being used.
1904 * State argument is acquired via halide_profiler_get_pipeline_state.
1905 * prev_t argument is the previous time and can be used to set a more
1906 * accurate time interval if desired. */
1908
1909/** Reset profiler state cheaply. May leave threads running or some
1910 * memory allocated but all accumluated statistics are reset.
1911 * WARNING: Do NOT call this method while any halide pipeline is
1912 * running; halide_profiler_memory_allocate/free and
1913 * halide_profiler_stack_peak_update update the profiler pipeline's
1914 * state without grabbing the global profiler state's lock. */
1916
1917/** Reset all profiler state.
1918 * WARNING: Do NOT call this method while any halide pipeline is
1919 * running; halide_profiler_memory_allocate/free and
1920 * halide_profiler_stack_peak_update update the profiler pipeline's
1921 * state without grabbing the global profiler state's lock. */
1923
1924/** Print out timing statistics for everything run since the last
1925 * reset. Also happens at process exit. */
1926extern void halide_profiler_report(void *user_context);
1927
1928/** For timer based profiling, this routine starts the timer chain running.
1929 * halide_get_profiler_state can be called to get the current timer interval.
1930 */
1932/** These routines are called to temporarily disable and then reenable
1933 * timer interuppts for profiling */
1934//@{
1937//@}
1938
1939/// \name "Float16" functions
1940/// These functions operate of bits (``uint16_t``) representing a half
1941/// precision floating point number (IEEE-754 2008 binary16).
1942//{@
1943
1944/** Read bits representing a half precision floating point number and return
1945 * the float that represents the same value */
1947
1948/** Read bits representing a half precision floating point number and return
1949 * the double that represents the same value */
1951
1952// TODO: Conversion functions to half
1953
1954//@}
1955
1956// Allocating and freeing device memory is often very slow. The
1957// methods below give Halide's runtime permission to hold onto device
1958// memory to service future requests instead of returning it to the
1959// underlying device API. The API does not manage an allocation pool,
1960// all it does is provide access to a shared counter that acts as a
1961// limit on the unused memory not yet returned to the underlying
1962// device API. It makes callbacks to participants when memory needs to
1963// be released because the limit is about to be exceeded (either
1964// because the limit has been reduced, or because the memory owned by
1965// some participant becomes unused).
1966
1967/** Tell Halide whether or not it is permitted to hold onto device
1968 * allocations to service future requests instead of returning them
1969 * eagerly to the underlying device API. Many device allocators are
1970 * quite slow, so it can be beneficial to set this to true. The
1971 * default value for now is false.
1972 *
1973 * Note that if enabled, the eviction policy is very simplistic. The
1974 * 32 most-recently used allocations are preserved, regardless of
1975 * their size. Additionally, if a call to cuMalloc results in an
1976 * out-of-memory error, the entire cache is flushed and the allocation
1977 * is retried. See https://github.com/halide/Halide/issues/4093
1978 *
1979 * If set to false, releases all unused device allocations back to the
1980 * underlying device APIs. For finer-grained control, see specific
1981 * methods in each device api runtime.
1982 *
1983 * Note that if the flag is set to true, this call *must* succeed and return
1984 * a value of halide_error_code_success (i.e., zero); if you replace
1985 * the implementation of this call in the runtime, you must honor this contract.
1986 * */
1987extern int halide_reuse_device_allocations(void *user_context, bool);
1988
1989/** Determines whether on device_free the memory is returned
1990 * immediately to the device API, or placed on a free list for future
1991 * use. Override and switch based on the user_context for
1992 * finer-grained control. By default just returns the value most
1993 * recently set by the method above. */
1994extern bool halide_can_reuse_device_allocations(void *user_context);
1995
2000
2001/** Register a callback to be informed when
2002 * halide_reuse_device_allocations(false) is called, and all unused
2003 * device allocations must be released. The object passed should have
2004 * global lifetime, and its next field will be clobbered. */
2006
2007#ifdef __cplusplus
2008} // End extern "C"
2009#endif
2010
2011#if (__cplusplus >= 201103L || _MSVC_LANG >= 201103L)
2012
2013namespace {
2014
2015template<typename T>
2016struct check_is_pointer {
2017 static constexpr bool value = false;
2018};
2019
2020template<typename T>
2021struct check_is_pointer<T *> {
2022 static constexpr bool value = true;
2023};
2024
2025} // namespace
2026
2027/** Construct the halide equivalent of a C type */
2028template<typename T>
2029HALIDE_ALWAYS_INLINE constexpr halide_type_t halide_type_of() {
2030 // Create a compile-time error if T is not a pointer (without
2031 // using any includes - this code goes into the runtime).
2032 // (Note that we can't have uninitialized variables in constexpr functions,
2033 // even if those variables aren't used.)
2034 static_assert(check_is_pointer<T>::value, "Expected a pointer type here");
2036}
2037
2038#ifdef HALIDE_CPP_COMPILER_HAS_FLOAT16
2039template<>
2040HALIDE_ALWAYS_INLINE constexpr halide_type_t halide_type_of<_Float16>() {
2041 return halide_type_t(halide_type_float, 16);
2042}
2043#endif
2044
2045template<>
2046HALIDE_ALWAYS_INLINE constexpr halide_type_t halide_type_of<float>() {
2047 return halide_type_t(halide_type_float, 32);
2048}
2049
2050template<>
2051HALIDE_ALWAYS_INLINE constexpr halide_type_t halide_type_of<double>() {
2052 return halide_type_t(halide_type_float, 64);
2053}
2054
2055template<>
2056HALIDE_ALWAYS_INLINE constexpr halide_type_t halide_type_of<bool>() {
2058}
2059
2060template<>
2061HALIDE_ALWAYS_INLINE constexpr halide_type_t halide_type_of<uint8_t>() {
2063}
2064
2065template<>
2066HALIDE_ALWAYS_INLINE constexpr halide_type_t halide_type_of<uint16_t>() {
2067 return halide_type_t(halide_type_uint, 16);
2068}
2069
2070template<>
2071HALIDE_ALWAYS_INLINE constexpr halide_type_t halide_type_of<uint32_t>() {
2072 return halide_type_t(halide_type_uint, 32);
2073}
2074
2075template<>
2076HALIDE_ALWAYS_INLINE constexpr halide_type_t halide_type_of<uint64_t>() {
2077 return halide_type_t(halide_type_uint, 64);
2078}
2079
2080template<>
2081HALIDE_ALWAYS_INLINE constexpr halide_type_t halide_type_of<int8_t>() {
2082 return halide_type_t(halide_type_int, 8);
2083}
2084
2085template<>
2086HALIDE_ALWAYS_INLINE constexpr halide_type_t halide_type_of<int16_t>() {
2087 return halide_type_t(halide_type_int, 16);
2088}
2089
2090template<>
2091HALIDE_ALWAYS_INLINE constexpr halide_type_t halide_type_of<int32_t>() {
2092 return halide_type_t(halide_type_int, 32);
2093}
2094
2095template<>
2096HALIDE_ALWAYS_INLINE constexpr halide_type_t halide_type_of<int64_t>() {
2097 return halide_type_t(halide_type_int, 64);
2098}
2099
2100#ifndef COMPILING_HALIDE_RUNTIME
2101
2102// These structures are used by `function_info_header` files
2103// (generated by passing `-e function_info_header` to a Generator).
2104// The generated files contain documentation on the proper usage.
2105namespace HalideFunctionInfo {
2106
2107enum ArgumentKind { InputScalar = 0,
2108 InputBuffer = 1,
2109 OutputBuffer = 2 };
2110
2111struct ArgumentInfo {
2112 std::string_view name;
2113 ArgumentKind kind;
2114 int32_t dimensions; // always zero for scalar arguments
2115 halide_type_t type;
2116};
2117
2118} // namespace HalideFunctionInfo
2119
2120#endif // COMPILING_HALIDE_RUNTIME
2121
2122#endif // (__cplusplus >= 201103L || _MSVC_LANG >= 201103L)
2123
2124#endif // HALIDE_HALIDERUNTIME_H
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. NOTE: This feature is deprecated and will be removed in Halide 17.
@ 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_vulkan_float16
Enable Vulkan 16-bit float support.
@ 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_vulkan
Enable Vulkan runtime support.
@ 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_semihosting
Used together with Target::NoOS for the baremetal target built with semihosting library and run with ...
@ 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_d3d12compute
Enable Direct3D 12 Compute runtime.
@ halide_target_feature_cuda_capability86
Enable CUDA compute capability 8.6 (Ampere)
@ 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_spirv
Enable SPIR-V code generation support.
@ 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_vulkan_float64
Enable Vulkan 64-bit float support.
@ 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_vulkan_version13
Enable Vulkan v1.3 runtime target support.
@ halide_target_feature_vulkan_version12
Enable Vulkan v1.2 runtime target support.
@ 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_cuda
Enable the CUDA runtime. Defaults to compute capability 2.0 (Fermi)
@ halide_target_feature_armv81a
Enable ARMv8.1-a instructions.
@ halide_target_feature_webgpu
Enable the WebGPU runtime.
@ 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_vulkan_int16
Enable Vulkan 16-bit integer 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_vulkan_int8
Enable Vulkan 8-bit integer support.
@ halide_target_feature_vulkan_int64
Enable Vulkan 64-bit integer support.
@ 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_vulkan_version10
Enable Vulkan v1.0 runtime target support.
@ 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.
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)
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 *)
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)
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_symbol_not_found
A runtime symbol could not be loaded.
@ 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_unimplemented
This part of the Halide runtime is unimplemented on this platform.
@ 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_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_trace_failed
Failure recording trace packets for one of the halide_target_feature_trace features.
@ 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_gpu_device_error
Call(s) to a GPU backend API failed.
@ 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_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 ...
@ 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
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
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)
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:1127
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:1088
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:1140
unsigned __INT64_TYPE__ uint64_t
signed __INT64_TYPE__ int64_t
__UINTPTR_TYPE__ uintptr_t
signed __INT32_TYPE__ int32_t
unsigned __INT8_TYPE__ uint8_t
__PTRDIFF_TYPE__ ptrdiff_t
WEAK void(*)(void *, const char *) halide_set_custom_print(void(*print)(void *, const char *))
unsigned __INT16_TYPE__ uint16_t
__SIZE_TYPE__ size_t
unsigned __INT32_TYPE__ uint32_t
signed __INT16_TYPE__ int16_t
signed __INT8_TYPE__ int8_t
WEAK void(*)(void *, const char *) halide_set_error_handler(void(*handler)(void *, const char *))
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:253
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 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::@4 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.
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.