Halide 17.0.2
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// Note that the canonical Halide version is considered to be defined here
22// (rather than in the build system); we redundantly define the value in
23// our CMake build, so that we ensure that the in-build metadata (eg soversion)
24// matches, but keeping the canonical version here makes it easier to keep
25// downstream build systems (eg Blaze/Bazel) properly in sync with the source.
26#define HALIDE_VERSION_MAJOR 17
27#define HALIDE_VERSION_MINOR 0
28#define HALIDE_VERSION_PATCH 2
29
30#ifdef __cplusplus
31// Forward declare type to allow naming typed handles.
32// See Type.h for documentation.
33template<typename T>
35#endif
36
37#ifdef __cplusplus
38extern "C" {
39#endif
40
41#ifdef _MSC_VER
42// Note that (for MSVC) you should not use "inline" along with HALIDE_ALWAYS_INLINE;
43// it is not necessary, and may produce warnings for some build configurations.
44#define HALIDE_ALWAYS_INLINE __forceinline
45#define HALIDE_NEVER_INLINE __declspec(noinline)
46#else
47// Note that (for Posixy compilers) you should always use "inline" along with HALIDE_ALWAYS_INLINE;
48// otherwise some corner-case scenarios may erroneously report link errors.
49#define HALIDE_ALWAYS_INLINE inline __attribute__((always_inline))
50#define HALIDE_NEVER_INLINE __attribute__((noinline))
51#endif
52
53#ifndef HALIDE_MUST_USE_RESULT
54#ifdef __has_attribute
55#if __has_attribute(nodiscard)
56// C++17 or later
57#define HALIDE_MUST_USE_RESULT [[nodiscard]]
58#elif __has_attribute(warn_unused_result)
59// Clang/GCC
60#define HALIDE_MUST_USE_RESULT __attribute__((warn_unused_result))
61#else
62#define HALIDE_MUST_USE_RESULT
63#endif
64#else
65#define HALIDE_MUST_USE_RESULT
66#endif
67#endif
68
69// Annotation for AOT and JIT calls -- if undefined, use no annotation.
70// To ensure that all results are checked, do something like
71//
72// -DHALIDE_FUNCTION_ATTRS=HALIDE_MUST_USE_RESULT
73//
74// in your C++ compiler options
75#ifndef HALIDE_FUNCTION_ATTRS
76#define HALIDE_FUNCTION_ATTRS
77#endif
78
79#ifndef HALIDE_EXPORT_SYMBOL
80#ifdef _MSC_VER
81#define HALIDE_EXPORT_SYMBOL __declspec(dllexport)
82#else
83#define HALIDE_EXPORT_SYMBOL __attribute__((visibility("default")))
84#endif
85#endif
86
87#ifndef COMPILING_HALIDE_RUNTIME
88
89// clang had _Float16 added as a reserved name in clang 8, but
90// doesn't actually support it on most platforms until clang 15.
91// Ideally there would be a better way to detect if the type
92// is supported, even in a compiler independent fashion, but
93// coming up with one has proven elusive.
94#if defined(__clang__) && (__clang_major__ >= 16) && !defined(__EMSCRIPTEN__)
95#if defined(__is_identifier)
96#if !__is_identifier(_Float16)
97#define HALIDE_CPP_COMPILER_HAS_FLOAT16
98#endif
99#endif
100#endif
101
102// Similarly, detecting _Float16 for gcc is problematic.
103// For now, we say that if >= v12, and compiling on x86 or arm,
104// we assume support. This may need revision.
105#if defined(__GNUC__) && (__GNUC__ >= 12)
106#if defined(__x86_64__) || defined(__i386__) || defined(__arm__) || defined(__aarch64__)
107#define HALIDE_CPP_COMPILER_HAS_FLOAT16
108#endif
109#endif
110
111#endif // !COMPILING_HALIDE_RUNTIME
112
113/** \file
114 *
115 * This file declares the routines used by Halide internally in its
116 * runtime. On platforms that support weak linking, these can be
117 * replaced with user-defined versions by defining an extern "C"
118 * function with the same name and signature.
119 *
120 * When doing Just In Time (JIT) compilation members of
121 * some_pipeline_or_func.jit_handlers() must be replaced instead. The
122 * corresponding methods are documented below.
123 *
124 * All of these functions take a "void *user_context" parameter as their
125 * first argument; if the Halide kernel that calls back to any of these
126 * functions has been compiled with the UserContext feature set on its Target,
127 * then the value of that pointer passed from the code that calls the
128 * Halide kernel is piped through to the function.
129 *
130 * Some of these are also useful to call when using the default
131 * implementation. E.g. halide_shutdown_thread_pool.
132 *
133 * Note that even on platforms with weak linking, some linker setups
134 * may not respect the override you provide. E.g. if the override is
135 * in a shared library and the halide object files are linked directly
136 * into the output, the builtin versions of the runtime functions will
137 * be called. See your linker documentation for more details. On
138 * Linux, LD_DYNAMIC_WEAK=1 may help.
139 *
140 */
141
142// Forward-declare to suppress warnings if compiling as C.
143struct halide_buffer_t;
144
145/** Print a message to stderr. Main use is to support tracing
146 * functionality, print, and print_when calls. Also called by the default
147 * halide_error. This function can be replaced in JITed code by using
148 * halide_custom_print and providing an implementation of halide_print
149 * in AOT code. See Func::set_custom_print.
150 */
151// @{
152extern void halide_print(void *user_context, const char *);
153extern void halide_default_print(void *user_context, const char *);
154typedef void (*halide_print_t)(void *, const char *);
156// @}
157
158/** Halide calls this function on runtime errors (for example bounds
159 * checking failures). This function can be replaced in JITed code by
160 * using Func::set_error_handler, or in AOT code by calling
161 * halide_set_error_handler. In AOT code on platforms that support
162 * weak linking (i.e. not Windows), you can also override it by simply
163 * defining your own halide_error.
164 */
165// @{
166extern void halide_error(void *user_context, const char *);
167extern void halide_default_error(void *user_context, const char *);
168typedef void (*halide_error_handler_t)(void *, const char *);
170// @}
171
172/** Cross-platform mutex. Must be initialized with zero and implementation
173 * must treat zero as an unlocked mutex with no waiters, etc.
174 */
178
179/** Cross platform condition variable. Must be initialized to 0. */
183
184/** A basic set of mutex and condition variable functions, which call
185 * platform specific code for mutual exclusion. Equivalent to posix
186 * calls. */
187//@{
188extern void halide_mutex_lock(struct halide_mutex *mutex);
189extern void halide_mutex_unlock(struct halide_mutex *mutex);
190extern void halide_cond_signal(struct halide_cond *cond);
191extern void halide_cond_broadcast(struct halide_cond *cond);
192extern void halide_cond_wait(struct halide_cond *cond, struct halide_mutex *mutex);
193//@}
194
195/** Functions for constructing/destroying/locking/unlocking arrays of mutexes. */
196struct halide_mutex_array;
197//@{
198extern struct halide_mutex_array *halide_mutex_array_create(int sz);
199extern void halide_mutex_array_destroy(void *user_context, void *array);
200extern int halide_mutex_array_lock(struct halide_mutex_array *array, int entry);
201extern int halide_mutex_array_unlock(struct halide_mutex_array *array, int entry);
202//@}
203
204/** Define halide_do_par_for to replace the default thread pool
205 * implementation. halide_shutdown_thread_pool can also be called to
206 * release resources used by the default thread pool on platforms
207 * where it makes sense. See Func::set_custom_do_task and
208 * Func::set_custom_do_par_for. Should return zero if all the jobs
209 * return zero, or an arbitrarily chosen return value from one of the
210 * jobs otherwise.
211 */
212//@{
213typedef int (*halide_task_t)(void *user_context, int task_number, uint8_t *closure);
214extern int halide_do_par_for(void *user_context,
215 halide_task_t task,
216 int min, int size, uint8_t *closure);
217extern void halide_shutdown_thread_pool();
218//@}
219
220/** Set a custom method for performing a parallel for loop. Returns
221 * the old do_par_for handler. */
222typedef int (*halide_do_par_for_t)(void *, halide_task_t, int, int, uint8_t *);
224
225/** An opaque struct representing a semaphore. Used by the task system for async tasks. */
229
230/** A struct representing a semaphore and a number of items that must
231 * be acquired from it. Used in halide_parallel_task_t below. */
236extern int halide_semaphore_init(struct halide_semaphore_t *, int n);
237extern int halide_semaphore_release(struct halide_semaphore_t *, int n);
238extern bool halide_semaphore_try_acquire(struct halide_semaphore_t *, int n);
239typedef int (*halide_semaphore_init_t)(struct halide_semaphore_t *, int);
240typedef int (*halide_semaphore_release_t)(struct halide_semaphore_t *, int);
242
243/** A task representing a serial for loop evaluated over some range.
244 * Note that task_parent is a pass through argument that should be
245 * passed to any dependent taks that are invoked using halide_do_parallel_tasks
246 * underneath this call. */
247typedef int (*halide_loop_task_t)(void *user_context, int min, int extent,
248 uint8_t *closure, void *task_parent);
249
250/** A parallel task to be passed to halide_do_parallel_tasks. This
251 * task may recursively call halide_do_parallel_tasks, and there may
252 * be complex dependencies between seemingly unrelated tasks expressed
253 * using semaphores. If you are using a custom task system, care must
254 * be taken to avoid potential deadlock. This can be done by carefully
255 * respecting the static metadata at the end of the task struct.*/
257 // The function to call. It takes a user context, a min and
258 // extent, a closure, and a task system pass through argument.
260
261 // The closure to pass it
263
264 // The name of the function to be called. For debugging purposes only.
265 const char *name;
266
267 // An array of semaphores that must be acquired before the
268 // function is called. Must be reacquired for every call made.
271
272 // The entire range the function should be called over. This range
273 // may be sliced up and the function called multiple times.
275
276 // A parallel task provides several pieces of metadata to prevent
277 // unbounded resource usage or deadlock.
278
279 // The first is the minimum number of execution contexts (call
280 // stacks or threads) necessary for the function to run to
281 // completion. This may be greater than one when there is nested
282 // parallelism with internal producer-consumer relationships
283 // (calling the function recursively spawns and blocks on parallel
284 // sub-tasks that communicate with each other via semaphores). If
285 // a parallel runtime calls the function when fewer than this many
286 // threads are idle, it may need to create more threads to
287 // complete the task, or else risk deadlock due to committing all
288 // threads to tasks that cannot complete without more.
289 //
290 // FIXME: Note that extern stages are assumed to only require a
291 // single thread to complete. If the extern stage is itself a
292 // Halide pipeline, this may be an underestimate.
294
295 // The calls to the function should be in serial order from min to min+extent-1, with only
296 // one executing at a time. If false, any order is fine, and
297 // concurrency is fine.
298 bool serial;
299};
300
301/** Enqueue some number of the tasks described above and wait for them
302 * to complete. While waiting, the calling threads assists with either
303 * the tasks enqueued, or other non-blocking tasks in the task
304 * system. Note that task_parent should be NULL for top-level calls
305 * and the pass through argument if this call is being made from
306 * another task. */
307extern int halide_do_parallel_tasks(void *user_context, int num_tasks,
308 struct halide_parallel_task_t *tasks,
309 void *task_parent);
310
311/** If you use the default do_par_for, you can still set a custom
312 * handler to perform each individual task. Returns the old handler. */
313//@{
314typedef int (*halide_do_task_t)(void *, halide_task_t, int, uint8_t *);
316extern int halide_do_task(void *user_context, halide_task_t f, int idx,
317 uint8_t *closure);
318//@}
319
320/** The version of do_task called for loop tasks. By default calls the
321 * loop task with the same arguments. */
322// @{
323typedef int (*halide_do_loop_task_t)(void *, halide_loop_task_t, int, int, uint8_t *, void *);
325extern int halide_do_loop_task(void *user_context, halide_loop_task_t f, int min, int extent,
326 uint8_t *closure, void *task_parent);
327//@}
328
329/** Provide an entire custom tasking runtime via function
330 * pointers. Note that do_task and semaphore_try_acquire are only ever
331 * called by halide_default_do_par_for and
332 * halide_default_do_parallel_tasks, so it's only necessary to provide
333 * those if you are mixing in the default implementations of
334 * do_par_for and do_parallel_tasks. */
335// @{
336typedef int (*halide_do_parallel_tasks_t)(void *, int, struct halide_parallel_task_t *,
337 void *task_parent);
346// @}
347
348/** The default versions of the parallel runtime functions. */
349// @{
350extern int halide_default_do_par_for(void *user_context,
351 halide_task_t task,
352 int min, int size, uint8_t *closure);
353extern int halide_default_do_parallel_tasks(void *user_context,
354 int num_tasks,
355 struct halide_parallel_task_t *tasks,
356 void *task_parent);
357extern int halide_default_do_task(void *user_context, halide_task_t f, int idx,
359extern int halide_default_do_loop_task(void *user_context, halide_loop_task_t f,
360 int min, int extent,
361 uint8_t *closure, void *task_parent);
362extern int halide_default_semaphore_init(struct halide_semaphore_t *, int n);
363extern int halide_default_semaphore_release(struct halide_semaphore_t *, int n);
365// @}
366
367struct halide_thread;
368
369/** Spawn a thread. Returns a handle to the thread for the purposes of
370 * joining it. The thread must be joined in order to clean up any
371 * resources associated with it. */
372extern struct halide_thread *halide_spawn_thread(void (*f)(void *), void *closure);
373
374/** Join a thread. */
375extern void halide_join_thread(struct halide_thread *);
376
377/** Set the number of threads used by Halide's thread pool. Returns
378 * the old number.
379 *
380 * n < 0 : error condition
381 * n == 0 : use a reasonable system default (typically, number of cpus online).
382 * n == 1 : use exactly one thread; this will always enforce serial execution
383 * n > 1 : use a pool of exactly n threads.
384 *
385 * (Note that this is only guaranteed when using the default implementations
386 * of halide_do_par_for(); custom implementations may completely ignore values
387 * passed to halide_set_num_threads().)
388 */
389extern int halide_set_num_threads(int n);
390
391/** Halide calls these functions to allocate and free memory. To
392 * replace in AOT code, use the halide_set_custom_malloc and
393 * halide_set_custom_free, or (on platforms that support weak
394 * linking), simply define these functions yourself. In JIT-compiled
395 * code use Func::set_custom_allocator.
396 *
397 * If you override them, and find yourself wanting to call the default
398 * implementation from within your override, use
399 * halide_default_malloc/free.
400 *
401 * Note that halide_malloc must return a pointer aligned to the
402 * maximum meaningful alignment for the platform for the purpose of
403 * vector loads and stores, *and* with an allocated size that is (at least)
404 * an integral multiple of that same alignment. The default implementation
405 * uses 32-byte alignment on arm and 64-byte alignment on x86. Additionally,
406 * it must be safe to read at least 8 bytes before the start and beyond the end.
407 */
408//@{
409extern void *halide_malloc(void *user_context, size_t x);
410extern void halide_free(void *user_context, void *ptr);
411extern void *halide_default_malloc(void *user_context, size_t x);
412extern void halide_default_free(void *user_context, void *ptr);
413typedef void *(*halide_malloc_t)(void *, size_t);
414typedef void (*halide_free_t)(void *, void *);
417//@}
418
419/** Halide calls these functions to interact with the underlying
420 * system runtime functions. To replace in AOT code on platforms that
421 * support weak linking, define these functions yourself, or use
422 * the halide_set_custom_load_library() and halide_set_custom_get_library_symbol()
423 * functions. In JIT-compiled code, use JITSharedRuntime::set_default_handlers().
424 *
425 * halide_load_library and halide_get_library_symbol are equivalent to
426 * dlopen and dlsym. halide_get_symbol(sym) is equivalent to
427 * dlsym(RTLD_DEFAULT, sym).
428 */
429//@{
430extern void *halide_get_symbol(const char *name);
431extern void *halide_load_library(const char *name);
432extern void *halide_get_library_symbol(void *lib, const char *name);
433extern void *halide_default_get_symbol(const char *name);
434extern void *halide_default_load_library(const char *name);
435extern void *halide_default_get_library_symbol(void *lib, const char *name);
436typedef void *(*halide_get_symbol_t)(const char *name);
437typedef void *(*halide_load_library_t)(const char *name);
438typedef void *(*halide_get_library_symbol_t)(void *lib, const char *name);
442//@}
443
444/** Called when debug_to_file is used inside %Halide code. See
445 * Func::debug_to_file for how this is called
446 *
447 * Cannot be replaced in JITted code at present.
448 */
449extern int32_t halide_debug_to_file(void *user_context, const char *filename,
450 int32_t type_code,
451 struct halide_buffer_t *buf);
452
453/** Types in the halide type system. They can be ints, unsigned ints,
454 * or floats (of various bit-widths), or a handle (which is always 64-bits).
455 * Note that the int/uint/float values do not imply a specific bit width
456 * (the bit width is expected to be encoded in a separate value).
457 */
459#if (__cplusplus >= 201103L || _MSVC_LANG >= 201103L)
460 : uint8_t
461#endif
462{
463 halide_type_int = 0, ///< signed integers
464 halide_type_uint = 1, ///< unsigned integers
465 halide_type_float = 2, ///< IEEE floating point numbers
466 halide_type_handle = 3, ///< opaque pointer type (void *)
467 halide_type_bfloat = 4, ///< floating point numbers in the bfloat format
469
470// Note that while __attribute__ can go before or after the declaration,
471// __declspec apparently is only allowed before.
472#ifndef HALIDE_ATTRIBUTE_ALIGN
473#ifdef _MSC_VER
474#define HALIDE_ATTRIBUTE_ALIGN(x) __declspec(align(x))
475#else
476#define HALIDE_ATTRIBUTE_ALIGN(x) __attribute__((aligned(x)))
477#endif
478#endif
479
480/** A runtime tag for a type in the halide type system. Can be ints,
481 * unsigned ints, or floats of various bit-widths (the 'bits'
482 * field). Can also be vectors of the same (by setting the 'lanes'
483 * field to something larger than one). This struct should be
484 * exactly 32-bits in size. */
486 /** The basic type code: signed integer, unsigned integer, or floating point. */
487#if (__cplusplus >= 201103L || _MSVC_LANG >= 201103L)
489 halide_type_code_t code; // halide_type_code_t
490#else
492 uint8_t code; // halide_type_code_t
493#endif
494
495 /** The number of bits of precision of a single scalar value of this type. */
498
499 /** How many elements in a vector. This is 1 for scalar types. */
502
503#if (__cplusplus >= 201103L || _MSVC_LANG >= 201103L)
504 /** Construct a runtime representation of a Halide type from:
505 * code: The fundamental type from an enum.
506 * bits: The bit size of one element.
507 * lanes: The number of vector elements in the type. */
509 : code(code), bits(bits), lanes(lanes) {
510 }
511
512 /** Default constructor is required e.g. to declare halide_trace_event
513 * instances. */
515 : code((halide_type_code_t)0), bits(0), lanes(0) {
516 }
517
518 HALIDE_ALWAYS_INLINE constexpr halide_type_t with_lanes(uint16_t new_lanes) const {
519 return halide_type_t((halide_type_code_t)code, bits, new_lanes);
520 }
521
522 HALIDE_ALWAYS_INLINE constexpr halide_type_t element_of() const {
523 return with_lanes(1);
524 }
525 /** Compare two types for equality. */
526 HALIDE_ALWAYS_INLINE constexpr bool operator==(const halide_type_t &other) const {
527 return as_u32() == other.as_u32();
528 }
529
530 HALIDE_ALWAYS_INLINE constexpr bool operator!=(const halide_type_t &other) const {
531 return !(*this == other);
532 }
533
534 HALIDE_ALWAYS_INLINE constexpr bool operator<(const halide_type_t &other) const {
535 return as_u32() < other.as_u32();
536 }
537
538 /** Size in bytes for a single element, even if width is not 1, of this type. */
539 HALIDE_ALWAYS_INLINE constexpr int bytes() const {
540 return (bits + 7) / 8;
541 }
542
543 HALIDE_ALWAYS_INLINE constexpr uint32_t as_u32() const {
544 // Note that this produces a result that is identical to memcpy'ing 'this'
545 // into a u32 (on a little-endian machine, anyway), and at -O1 or greater
546 // on Clang, the compiler knows this and optimizes this into a single 32-bit move.
547 // (At -O0 it will look awful.)
548 return static_cast<uint8_t>(code) |
549 (static_cast<uint16_t>(bits) << 8) |
550 (static_cast<uint32_t>(lanes) << 16);
551 }
552#endif
553};
554
555#if (__cplusplus >= 201103L || _MSVC_LANG >= 201103L)
556static_assert(sizeof(halide_type_t) == sizeof(uint32_t), "size mismatch in halide_type_t");
557#endif
558
570
572 /** The name of the Func or Pipeline that this event refers to */
573 const char *func;
574
575 /** If the event type is a load or a store, this points to the
576 * value being loaded or stored. Use the type field to safely cast
577 * this to a concrete pointer type and retrieve it. For other
578 * events this is null. */
579 void *value;
580
581 /** For loads and stores, an array which contains the location
582 * being accessed. For vector loads or stores it is an array of
583 * vectors of coordinates (the vector dimension is innermost).
584 *
585 * For realization or production-related events, this will contain
586 * the mins and extents of the region being accessed, in the order
587 * min0, extent0, min1, extent1, ...
588 *
589 * For pipeline-related events, this will be null.
590 */
592
593 /** For halide_trace_tag, this points to a read-only null-terminated string
594 * of arbitrary text. For all other events, this will be null.
595 */
596 const char *trace_tag;
597
598 /** If the event type is a load or a store, this is the type of
599 * the data. Otherwise, the value is meaningless. */
601
602 /** The type of event */
604
605 /* The ID of the parent event (see below for an explanation of
606 * event ancestry). */
608
609 /** If this was a load or store of a Tuple-valued Func, this is
610 * which tuple element was accessed. */
612
613 /** The length of the coordinates array */
615};
616
617/** Called when Funcs are marked as trace_load, trace_store, or
618 * trace_realization. See Func::set_custom_trace. The default
619 * implementation either prints events via halide_print, or if
620 * HL_TRACE_FILE is defined, dumps the trace to that file in a
621 * sequence of trace packets. The header for a trace packet is defined
622 * below. If the trace is going to be large, you may want to make the
623 * file a named pipe, and then read from that pipe into gzip.
624 *
625 * halide_trace returns a unique ID which will be passed to future
626 * events that "belong" to the earlier event as the parent id. The
627 * ownership hierarchy looks like:
628 *
629 * begin_pipeline
630 * +--trace_tag (if any)
631 * +--trace_tag (if any)
632 * ...
633 * +--begin_realization
634 * | +--produce
635 * | | +--load/store
636 * | | +--end_produce
637 * | +--consume
638 * | | +--load
639 * | | +--end_consume
640 * | +--end_realization
641 * +--end_pipeline
642 *
643 * Threading means that ownership cannot be inferred from the ordering
644 * of events. There can be many active realizations of a given
645 * function, or many active productions for a single
646 * realization. Within a single production, the ordering of events is
647 * meaningful.
648 *
649 * Note that all trace_tag events (if any) will occur just after the begin_pipeline
650 * event, but before any begin_realization events. All trace_tags for a given Func
651 * will be emitted in the order added.
652 */
653// @}
654extern int32_t halide_trace(void *user_context, const struct halide_trace_event_t *event);
655extern int32_t halide_default_trace(void *user_context, const struct halide_trace_event_t *event);
656typedef int32_t (*halide_trace_t)(void *user_context, const struct halide_trace_event_t *);
658// @}
659
660/** The header of a packet in a binary trace. All fields are 32-bit. */
662 /** The total size of this packet in bytes. Always a multiple of
663 * four. Equivalently, the number of bytes until the next
664 * packet. */
666
667 /** The id of this packet (for the purpose of parent_id). */
669
670 /** The remaining fields are equivalent to those in halide_trace_event_t */
671 // @{
677 // @}
678
679#if (__cplusplus >= 201103L || _MSVC_LANG >= 201103L)
680 /** Get the coordinates array, assuming this packet is laid out in
681 * memory as it was written. The coordinates array comes
682 * immediately after the packet header. */
683 HALIDE_ALWAYS_INLINE const int *coordinates() const {
684 return (const int *)(this + 1);
685 }
686
687 HALIDE_ALWAYS_INLINE int *coordinates() {
688 return (int *)(this + 1);
689 }
690
691 /** Get the value, assuming this packet is laid out in memory as
692 * it was written. The packet comes immediately after the coordinates
693 * array. */
694 HALIDE_ALWAYS_INLINE const void *value() const {
695 return (const void *)(coordinates() + dimensions);
696 }
697
698 HALIDE_ALWAYS_INLINE void *value() {
699 return (void *)(coordinates() + dimensions);
700 }
701
702 /** Get the func name, assuming this packet is laid out in memory
703 * as it was written. It comes after the value. */
704 HALIDE_ALWAYS_INLINE const char *func() const {
705 return (const char *)value() + type.lanes * type.bytes();
706 }
707
708 HALIDE_ALWAYS_INLINE char *func() {
709 return (char *)value() + type.lanes * type.bytes();
710 }
711
712 /** Get the trace_tag (if any), assuming this packet is laid out in memory
713 * as it was written. It comes after the func name. If there is no trace_tag,
714 * this will return a pointer to an empty string. */
715 HALIDE_ALWAYS_INLINE const char *trace_tag() const {
716 const char *f = func();
717 // strlen may not be available here
718 while (*f++) {
719 // nothing
720 }
721 return f;
722 }
723
724 HALIDE_ALWAYS_INLINE char *trace_tag() {
725 char *f = func();
726 // strlen may not be available here
727 while (*f++) {
728 // nothing
729 }
730 return f;
731 }
732#endif
733};
734
735/** Set the file descriptor that Halide should write binary trace
736 * events to. If called with 0 as the argument, Halide outputs trace
737 * information to stdout in a human-readable format. If never called,
738 * Halide checks the for existence of an environment variable called
739 * HL_TRACE_FILE and opens that file. If HL_TRACE_FILE is not defined,
740 * it outputs trace information to stdout in a human-readable
741 * format. */
742extern void halide_set_trace_file(int fd);
743
744/** Halide calls this to retrieve the file descriptor to write binary
745 * trace events to. The default implementation returns the value set
746 * by halide_set_trace_file. Implement it yourself if you wish to use
747 * a custom file descriptor per user_context. Return zero from your
748 * implementation to tell Halide to print human-readable trace
749 * information to stdout. */
750extern int halide_get_trace_file(void *user_context);
751
752/** If tracing is writing to a file. This call closes that file
753 * (flushing the trace). Returns zero on success. */
755
756/** All Halide GPU or device backend implementations provide an
757 * interface to be used with halide_device_malloc, etc. This is
758 * accessed via the functions below.
759 */
760
761/** An opaque struct containing per-GPU API implementations of the
762 * device functions. */
764
765/** Each GPU API provides a halide_device_interface_t struct pointing
766 * to the code that manages device allocations. You can access these
767 * functions directly from the struct member function pointers, or by
768 * calling the functions declared below. Note that the global
769 * functions are not available when using Halide as a JIT compiler.
770 * If you are using raw halide_buffer_t in that context you must use
771 * the function pointers in the device_interface struct.
772 *
773 * The function pointers below are currently the same for every GPU
774 * API; only the impl field varies. These top-level functions do the
775 * bookkeeping that is common across all GPU APIs, and then dispatch
776 * to more API-specific functions via another set of function pointers
777 * hidden inside the impl field.
778 */
780 int (*device_malloc)(void *user_context, struct halide_buffer_t *buf,
781 const struct halide_device_interface_t *device_interface);
782 int (*device_free)(void *user_context, struct halide_buffer_t *buf);
783 int (*device_sync)(void *user_context, struct halide_buffer_t *buf);
784 void (*device_release)(void *user_context,
785 const struct halide_device_interface_t *device_interface);
786 int (*copy_to_host)(void *user_context, struct halide_buffer_t *buf);
787 int (*copy_to_device)(void *user_context, struct halide_buffer_t *buf,
788 const struct halide_device_interface_t *device_interface);
789 int (*device_and_host_malloc)(void *user_context, struct halide_buffer_t *buf,
790 const struct halide_device_interface_t *device_interface);
791 int (*device_and_host_free)(void *user_context, struct halide_buffer_t *buf);
792 int (*buffer_copy)(void *user_context, struct halide_buffer_t *src,
793 const struct halide_device_interface_t *dst_device_interface, struct halide_buffer_t *dst);
794 int (*device_crop)(void *user_context, const struct halide_buffer_t *src,
795 struct halide_buffer_t *dst);
796 int (*device_slice)(void *user_context, const struct halide_buffer_t *src,
797 int slice_dim, int slice_pos, struct halide_buffer_t *dst);
798 int (*device_release_crop)(void *user_context, struct halide_buffer_t *buf);
799 int (*wrap_native)(void *user_context, struct halide_buffer_t *buf, uint64_t handle,
800 const struct halide_device_interface_t *device_interface);
801 int (*detach_native)(void *user_context, struct halide_buffer_t *buf);
802 int (*compute_capability)(void *user_context, int *major, int *minor);
804};
805
806/** Release all data associated with the given device interface, in
807 * particular all resources (memory, texture, context handles)
808 * allocated by Halide. Must be called explicitly when using AOT
809 * compilation. This is *not* thread-safe with respect to actively
810 * running Halide code. Ensure all pipelines are finished before
811 * calling this. */
812extern void halide_device_release(void *user_context,
813 const struct halide_device_interface_t *device_interface);
814
815/** Copy image data from device memory to host memory. This must be called
816 * explicitly to copy back the results of a GPU-based filter. */
817extern int halide_copy_to_host(void *user_context, struct halide_buffer_t *buf);
818
819/** Copy image data from host memory to device memory. This should not
820 * be called directly; Halide handles copying to the device
821 * automatically. If interface is NULL and the buf has a non-zero dev
822 * field, the device associated with the dev handle will be
823 * used. Otherwise if the dev field is 0 and interface is NULL, an
824 * error is returned. */
825extern int halide_copy_to_device(void *user_context, struct halide_buffer_t *buf,
826 const struct halide_device_interface_t *device_interface);
827
828/** Copy data from one buffer to another. The buffers may have
829 * different shapes and sizes, but the destination buffer's shape must
830 * be contained within the source buffer's shape. That is, for each
831 * dimension, the min on the destination buffer must be greater than
832 * or equal to the min on the source buffer, and min+extent on the
833 * destination buffer must be less that or equal to min+extent on the
834 * source buffer. The source data is pulled from either device or
835 * host memory on the source, depending on the dirty flags. host is
836 * preferred if both are valid. The dst_device_interface parameter
837 * controls the destination memory space. NULL means host memory. */
838extern int halide_buffer_copy(void *user_context, struct halide_buffer_t *src,
839 const struct halide_device_interface_t *dst_device_interface,
840 struct halide_buffer_t *dst);
841
842/** Give the destination buffer a device allocation which is an alias
843 * for the same coordinate range in the source buffer. Modifies the
844 * device, device_interface, and the device_dirty flag only. Only
845 * supported by some device APIs (others will return
846 * halide_error_code_device_crop_unsupported). Call
847 * halide_device_release_crop instead of halide_device_free to clean
848 * up resources associated with the cropped view. Do not free the
849 * device allocation on the source buffer while the destination buffer
850 * still lives. Note that the two buffers do not share dirty flags, so
851 * care must be taken to update them together as needed. Note that src
852 * and dst are required to have the same number of dimensions.
853 *
854 * Note also that (in theory) device interfaces which support cropping may
855 * still not support cropping a crop (instead, create a new crop of the parent
856 * buffer); in practice, no known implementation has this limitation, although
857 * it is possible that some future implementations may require it. */
858extern int halide_device_crop(void *user_context,
859 const struct halide_buffer_t *src,
860 struct halide_buffer_t *dst);
861
862/** Give the destination buffer a device allocation which is an alias
863 * for a similar coordinate range in the source buffer, but with one dimension
864 * sliced away in the dst. Modifies the device, device_interface, and the
865 * device_dirty flag only. Only supported by some device APIs (others will return
866 * halide_error_code_device_crop_unsupported). Call
867 * halide_device_release_crop instead of halide_device_free to clean
868 * up resources associated with the sliced view. Do not free the
869 * device allocation on the source buffer while the destination buffer
870 * still lives. Note that the two buffers do not share dirty flags, so
871 * care must be taken to update them together as needed. Note that the dst buffer
872 * must have exactly one fewer dimension than the src buffer, and that slice_dim
873 * and slice_pos must be valid within src. */
874extern int halide_device_slice(void *user_context,
875 const struct halide_buffer_t *src,
876 int slice_dim, int slice_pos,
877 struct halide_buffer_t *dst);
878
879/** Release any resources associated with a cropped/sliced view of another
880 * buffer. */
881extern int halide_device_release_crop(void *user_context,
882 struct halide_buffer_t *buf);
883
884/** Wait for current GPU operations to complete. Calling this explicitly
885 * should rarely be necessary, except maybe for profiling. */
886extern int halide_device_sync(void *user_context, struct halide_buffer_t *buf);
887
888/**
889 * Wait for current GPU operations to complete. Calling this explicitly
890 * should rarely be necessary, except maybe for profiling.
891 * This variation of the synchronizing is useful when a synchronization is desirable
892 * without specifying any buffer to synchronize on.
893 * Calling this with a null device_interface is always illegal.
894 */
895extern int halide_device_sync_global(void *user_context, const struct halide_device_interface_t *device_interface);
896
897/** Allocate device memory to back a halide_buffer_t. */
898extern int halide_device_malloc(void *user_context, struct halide_buffer_t *buf,
899 const struct halide_device_interface_t *device_interface);
900
901/** Free device memory. */
902extern int halide_device_free(void *user_context, struct halide_buffer_t *buf);
903
904/** Wrap or detach a native device handle, setting the device field
905 * and device_interface field as appropriate for the given GPU
906 * API. The meaning of the opaque handle is specific to the device
907 * interface, so if you know the device interface in use, call the
908 * more specific functions in the runtime headers for your specific
909 * device API instead (e.g. HalideRuntimeCuda.h). */
910// @{
911extern int halide_device_wrap_native(void *user_context,
912 struct halide_buffer_t *buf,
913 uint64_t handle,
914 const struct halide_device_interface_t *device_interface);
915extern int halide_device_detach_native(void *user_context, struct halide_buffer_t *buf);
916// @}
917
918/** Selects which gpu device to use. 0 is usually the display
919 * device. If never called, Halide uses the environment variable
920 * HL_GPU_DEVICE. If that variable is unset, Halide uses the last
921 * device. Set this to -1 to use the last device. */
922extern void halide_set_gpu_device(int n);
923
924/** Halide calls this to get the desired halide gpu device
925 * setting. Implement this yourself to use a different gpu device per
926 * user_context. The default implementation returns the value set by
927 * halide_set_gpu_device, or the environment variable
928 * HL_GPU_DEVICE. */
929extern int halide_get_gpu_device(void *user_context);
930
931/** Set the soft maximum amount of memory, in bytes, that the LRU
932 * cache will use to memoize Func results. This is not a strict
933 * maximum in that concurrency and simultaneous use of memoized
934 * reults larger than the cache size can both cause it to
935 * temporariliy be larger than the size specified here.
936 */
938
939/** Given a cache key for a memoized result, currently constructed
940 * from the Func name and top-level Func name plus the arguments of
941 * the computation, determine if the result is in the cache and
942 * return it if so. (The internals of the cache key should be
943 * considered opaque by this function.) If this routine returns true,
944 * it is a cache miss. Otherwise, it will return false and the
945 * buffers passed in will be filled, via copying, with memoized
946 * data. The last argument is a list if halide_buffer_t pointers which
947 * represents the outputs of the memoized Func. If the Func does not
948 * return a Tuple, there will only be one halide_buffer_t in the list. The
949 * tuple_count parameters determines the length of the list.
950 *
951 * The return values are:
952 * -1: Signals an error.
953 * 0: Success and cache hit.
954 * 1: Success and cache miss.
955 */
956extern int halide_memoization_cache_lookup(void *user_context, const uint8_t *cache_key, int32_t size,
957 struct halide_buffer_t *realized_bounds,
958 int32_t tuple_count, struct halide_buffer_t **tuple_buffers);
959
960/** Given a cache key for a memoized result, currently constructed
961 * from the Func name and top-level Func name plus the arguments of
962 * the computation, store the result in the cache for futre access by
963 * halide_memoization_cache_lookup. (The internals of the cache key
964 * should be considered opaque by this function.) Data is copied out
965 * from the inputs and inputs are unmodified. The last argument is a
966 * list if halide_buffer_t pointers which represents the outputs of the
967 * memoized Func. If the Func does not return a Tuple, there will
968 * only be one halide_buffer_t in the list. The tuple_count parameters
969 * determines the length of the list.
970 *
971 * If there is a memory allocation failure, the store does not store
972 * the data into the cache.
973 *
974 * If has_eviction_key is true, the entry is marked with eviction_key to
975 * allow removing the key with halide_memoization_cache_evict.
976 */
977extern int halide_memoization_cache_store(void *user_context, const uint8_t *cache_key, int32_t size,
978 struct halide_buffer_t *realized_bounds,
979 int32_t tuple_count,
980 struct halide_buffer_t **tuple_buffers,
981 bool has_eviction_key, uint64_t eviction_key);
982
983/** Evict all cache entries that were tagged with the given
984 * eviction_key in the memoize scheduling directive.
985 */
986extern void halide_memoization_cache_evict(void *user_context, uint64_t eviction_key);
987
988/** If halide_memoization_cache_lookup succeeds,
989 * halide_memoization_cache_release must be called to signal the
990 * storage is no longer being used by the caller. It will be passed
991 * the host pointer of one the buffers returned by
992 * halide_memoization_cache_lookup. That is
993 * halide_memoization_cache_release will be called multiple times for
994 * the case where halide_memoization_cache_lookup is handling multiple
995 * buffers. (This corresponds to memoizing a Tuple in Halide.) Note
996 * that the host pointer must be sufficient to get to all information
997 * the release operation needs. The default Halide cache impleemntation
998 * accomplishes this by storing extra data before the start of the user
999 * modifiable host storage.
1000 *
1001 * This call is like free and does not have a failure return.
1002 */
1003extern void halide_memoization_cache_release(void *user_context, void *host);
1004
1005/** Free all memory and resources associated with the memoization cache.
1006 * Must be called at a time when no other threads are accessing the cache.
1007 */
1009
1010/** Verify that a given range of memory has been initialized; only used when Target::MSAN is enabled.
1011 *
1012 * The default implementation simply calls the LLVM-provided __msan_check_mem_is_initialized() function.
1013 *
1014 * The return value should always be zero.
1015 */
1016extern int halide_msan_check_memory_is_initialized(void *user_context, const void *ptr, uint64_t len, const char *name);
1017
1018/** Verify that the data pointed to by the halide_buffer_t is initialized (but *not* the halide_buffer_t itself),
1019 * using halide_msan_check_memory_is_initialized() for checking.
1020 *
1021 * The default implementation takes pains to only check the active memory ranges
1022 * (skipping padding), and sorting into ranges to always check the smallest number of
1023 * ranges, in monotonically increasing memory order.
1024 *
1025 * Most client code should never need to replace the default implementation.
1026 *
1027 * The return value should always be zero.
1028 */
1029extern int halide_msan_check_buffer_is_initialized(void *user_context, struct halide_buffer_t *buffer, const char *buf_name);
1030
1031/** Annotate that a given range of memory has been initialized;
1032 * only used when Target::MSAN is enabled.
1033 *
1034 * The default implementation simply calls the LLVM-provided __msan_unpoison() function.
1035 *
1036 * The return value should always be zero.
1037 */
1038extern int halide_msan_annotate_memory_is_initialized(void *user_context, const void *ptr, uint64_t len);
1039
1040/** Mark the data pointed to by the halide_buffer_t as initialized (but *not* the halide_buffer_t itself),
1041 * using halide_msan_annotate_memory_is_initialized() for marking.
1042 *
1043 * The default implementation takes pains to only mark the active memory ranges
1044 * (skipping padding), and sorting into ranges to always mark the smallest number of
1045 * ranges, in monotonically increasing memory order.
1046 *
1047 * Most client code should never need to replace the default implementation.
1048 *
1049 * The return value should always be zero.
1050 */
1051extern int halide_msan_annotate_buffer_is_initialized(void *user_context, struct halide_buffer_t *buffer);
1052extern void halide_msan_annotate_buffer_is_initialized_as_destructor(void *user_context, void *buffer);
1053
1054/** The error codes that may be returned by a Halide pipeline. */
1056 /** There was no error. This is the value returned by Halide on success. */
1058
1059 /** An uncategorized error occurred. Refer to the string passed to halide_error. */
1061
1062 /** A Func was given an explicit bound via Func::bound, but this
1063 * was not large enough to encompass the region that is used of
1064 * the Func by the rest of the pipeline. */
1066
1067 /** The elem_size field of a halide_buffer_t does not match the size in
1068 * bytes of the type of that ImageParam. Probable type mismatch. */
1070
1071 /** A pipeline would access memory outside of the halide_buffer_t passed
1072 * in. */
1074
1075 /** A halide_buffer_t was given that spans more than 2GB of memory. */
1077
1078 /** A halide_buffer_t was given with extents that multiply to a number
1079 * greater than 2^31-1 */
1081
1082 /** Applying explicit constraints on the size of an input or
1083 * output buffer shrank the size of that buffer below what will be
1084 * accessed by the pipeline. */
1086
1087 /** A constraint on a size or stride of an input or output buffer
1088 * was not met by the halide_buffer_t passed in. */
1090
1091 /** A scalar parameter passed in was smaller than its minimum
1092 * declared value. */
1094
1095 /** A scalar parameter passed in was greater than its minimum
1096 * declared value. */
1098
1099 /** A call to halide_malloc returned NULL. */
1101
1102 /** A halide_buffer_t pointer passed in was NULL. */
1104
1105 /** debug_to_file failed to open or write to the specified
1106 * file. */
1108
1109 /** The Halide runtime encountered an error while trying to copy
1110 * from device to host. Turn on -debug in your target string to
1111 * see more details. */
1113
1114 /** The Halide runtime encountered an error while trying to copy
1115 * from host to device. Turn on -debug in your target string to
1116 * see more details. */
1118
1119 /** The Halide runtime encountered an error while trying to
1120 * allocate memory on device. Turn on -debug in your target string
1121 * to see more details. */
1123
1124 /** The Halide runtime encountered an error while trying to
1125 * synchronize with a device. Turn on -debug in your target string
1126 * to see more details. */
1128
1129 /** The Halide runtime encountered an error while trying to free a
1130 * device allocation. Turn on -debug in your target string to see
1131 * more details. */
1133
1134 /** Buffer has a non-zero device but no device interface, which
1135 * violates a Halide invariant. */
1137
1138 /** This part of the Halide runtime is unimplemented on this platform. */
1140
1141 /** A runtime symbol could not be loaded. */
1143
1144 /** There is a bug in the Halide compiler. */
1146
1147 /** The Halide runtime encountered an error while trying to launch
1148 * a GPU kernel. Turn on -debug in your target string to see more
1149 * details. */
1151
1152 /** The Halide runtime encountered a host pointer that violated
1153 * the alignment set for it by way of a call to
1154 * set_host_alignment */
1156
1157 /** A fold_storage directive was used on a dimension that is not
1158 * accessed in a monotonically increasing or decreasing fashion. */
1160
1161 /** A fold_storage directive was used with a fold factor that was
1162 * too small to store all the values of a producer needed by the
1163 * consumer. */
1165
1166 /** User-specified require() expression was not satisfied. */
1168
1169 /** At least one of the buffer's extents are negative. */
1171
1172 /** Call(s) to a GPU backend API failed. */
1174
1175 /** Failure recording trace packets for one of the halide_target_feature_trace features. */
1177
1178 /** A specialize_fail() schedule branch was selected at runtime. */
1180
1181 /** The Halide runtime encountered an error while trying to wrap a
1182 * native device handle. Turn on -debug in your target string to
1183 * see more details. */
1185
1186 /** The Halide runtime encountered an error while trying to detach
1187 * a native device handle. Turn on -debug in your target string
1188 * to see more details. */
1190
1191 /** The host field on an input or output was null, the device
1192 * field was not zero, and the pipeline tries to use the buffer on
1193 * the host. You may be passing a GPU-only buffer to a pipeline
1194 * which is scheduled to use it on the CPU. */
1196
1197 /** A folded buffer was passed to an extern stage, but the region
1198 * touched wraps around the fold boundary. */
1200
1201 /** Buffer has a non-null device_interface but device is 0, which
1202 * violates a Halide invariant. */
1204
1205 /** Buffer has both host and device dirty bits set, which violates
1206 * a Halide invariant. */
1208
1209 /** The halide_buffer_t * passed to a halide runtime routine is
1210 * nullptr and this is not allowed. */
1212
1213 /** The Halide runtime encountered an error while trying to copy
1214 * from one buffer to another. Turn on -debug in your target
1215 * string to see more details. */
1217
1218 /** Attempted to make cropped/sliced alias of a buffer with a device
1219 * field, but the device_interface does not support cropping. */
1221
1222 /** Cropping/slicing a buffer failed for some other reason. Turn on -debug
1223 * in your target string. */
1225
1226 /** An operation on a buffer required an allocation on a
1227 * particular device interface, but a device allocation already
1228 * existed on a different device interface. Free the old one
1229 * first. */
1231
1232 /** The dimensions field of a halide_buffer_t does not match the dimensions of that ImageParam. */
1234
1235 /** A buffer with the device_dirty flag set was passed to a
1236 * pipeline compiled with no device backends enabled, so it
1237 * doesn't know how to copy the data back from device memory to
1238 * host memory. Either call copy_to_host before calling the Halide
1239 * pipeline, or enable the appropriate device backend. */
1241
1242 /** An explicit storage bound provided is too small to store
1243 * all the values produced by the function. */
1245};
1246
1247/** Halide calls the functions below on various error conditions. The
1248 * default implementations construct an error message, call
1249 * halide_error, then return the matching error code above. On
1250 * platforms that support weak linking, you can override these to
1251 * catch the errors individually. */
1252
1253/** A call into an extern stage for the purposes of bounds inference
1254 * failed. Returns the error code given by the extern stage. */
1255extern int halide_error_bounds_inference_call_failed(void *user_context, const char *extern_stage_name, int result);
1256
1257/** A call to an extern stage failed. Returned the error code given by
1258 * the extern stage. */
1259extern int halide_error_extern_stage_failed(void *user_context, const char *extern_stage_name, int result);
1260
1261/** Various other error conditions. See the enum above for a
1262 * description of each. */
1263// @{
1264extern int halide_error_explicit_bounds_too_small(void *user_context, const char *func_name, const char *var_name,
1265 int min_bound, int max_bound, int min_required, int max_required);
1266extern int halide_error_bad_type(void *user_context, const char *func_name,
1267 uint32_t type_given, uint32_t correct_type); // N.B. The last two args are the bit representation of a halide_type_t
1268extern int halide_error_bad_dimensions(void *user_context, const char *func_name,
1269 int32_t dimensions_given, int32_t correct_dimensions);
1270extern int halide_error_access_out_of_bounds(void *user_context, const char *func_name,
1271 int dimension, int min_touched, int max_touched,
1272 int min_valid, int max_valid);
1273extern int halide_error_buffer_allocation_too_large(void *user_context, const char *buffer_name,
1274 uint64_t allocation_size, uint64_t max_size);
1275extern int halide_error_buffer_extents_negative(void *user_context, const char *buffer_name, int dimension, int extent);
1276extern int halide_error_buffer_extents_too_large(void *user_context, const char *buffer_name,
1277 int64_t actual_size, int64_t max_size);
1278extern int halide_error_constraints_make_required_region_smaller(void *user_context, const char *buffer_name,
1279 int dimension,
1280 int constrained_min, int constrained_extent,
1281 int required_min, int required_extent);
1282extern int halide_error_constraint_violated(void *user_context, const char *var, int val,
1283 const char *constrained_var, int constrained_val);
1284extern int halide_error_param_too_small_i64(void *user_context, const char *param_name,
1285 int64_t val, int64_t min_val);
1286extern int halide_error_param_too_small_u64(void *user_context, const char *param_name,
1287 uint64_t val, uint64_t min_val);
1288extern int halide_error_param_too_small_f64(void *user_context, const char *param_name,
1289 double val, double min_val);
1290extern int halide_error_param_too_large_i64(void *user_context, const char *param_name,
1291 int64_t val, int64_t max_val);
1292extern int halide_error_param_too_large_u64(void *user_context, const char *param_name,
1293 uint64_t val, uint64_t max_val);
1294extern int halide_error_param_too_large_f64(void *user_context, const char *param_name,
1295 double val, double max_val);
1296extern int halide_error_out_of_memory(void *user_context);
1297extern int halide_error_buffer_argument_is_null(void *user_context, const char *buffer_name);
1298extern int halide_error_debug_to_file_failed(void *user_context, const char *func,
1299 const char *filename, int error_code);
1300extern int halide_error_unaligned_host_ptr(void *user_context, const char *func_name, int alignment);
1301extern int halide_error_host_is_null(void *user_context, const char *func_name);
1302extern int halide_error_bad_fold(void *user_context, const char *func_name, const char *var_name,
1303 const char *loop_name);
1304extern int halide_error_bad_extern_fold(void *user_context, const char *func_name,
1305 int dim, int min, int extent, int valid_min, int fold_factor);
1306
1307extern int halide_error_fold_factor_too_small(void *user_context, const char *func_name, const char *var_name,
1308 int fold_factor, const char *loop_name, int required_extent);
1309extern int halide_error_requirement_failed(void *user_context, const char *condition, const char *message);
1310extern int halide_error_specialize_fail(void *user_context, const char *message);
1311extern int halide_error_no_device_interface(void *user_context);
1312extern int halide_error_device_interface_no_device(void *user_context);
1313extern int halide_error_host_and_device_dirty(void *user_context);
1314extern int halide_error_buffer_is_null(void *user_context, const char *routine);
1315extern int halide_error_device_dirty_with_no_device_support(void *user_context, const char *buffer_name);
1316extern int halide_error_storage_bound_too_small(void *user_context, const char *func_name, const char *var_name,
1317 int provided_size, int required_size);
1318extern int halide_error_device_crop_failed(void *user_context);
1319// @}
1320
1321/** Optional features a compilation Target can have.
1322 * Be sure to keep this in sync with the Feature enum in Target.h and the implementation of
1323 * get_runtime_compatible_target in Target.cpp if you add a new feature.
1324 */
1326 halide_target_feature_jit = 0, ///< Generate code that will run immediately inside the calling process.
1327 halide_target_feature_debug, ///< Turn on debug info and output for runtime code.
1328 halide_target_feature_no_asserts, ///< Disable all runtime checks, for slightly tighter code.
1329 halide_target_feature_no_bounds_query, ///< Disable the bounds querying functionality.
1330
1331 halide_target_feature_sse41, ///< Use SSE 4.1 and earlier instructions. Only relevant on x86.
1332 halide_target_feature_avx, ///< Use AVX 1 instructions. Only relevant on x86.
1333 halide_target_feature_avx2, ///< Use AVX 2 instructions. Only relevant on x86.
1334 halide_target_feature_fma, ///< Enable x86 FMA instruction
1335 halide_target_feature_fma4, ///< Enable x86 (AMD) FMA4 instruction set
1336 halide_target_feature_f16c, ///< Enable x86 16-bit float support
1337
1338 halide_target_feature_armv7s, ///< Generate code for ARMv7s. Only relevant for 32-bit ARM.
1339 halide_target_feature_no_neon, ///< Avoid using NEON instructions. Only relevant for 32-bit ARM.
1340
1341 halide_target_feature_vsx, ///< Use VSX instructions. Only relevant on POWERPC.
1342 halide_target_feature_power_arch_2_07, ///< Use POWER ISA 2.07 new instructions. Only relevant on POWERPC.
1343
1344 halide_target_feature_cuda, ///< Enable the CUDA runtime. Defaults to compute capability 2.0 (Fermi)
1345 halide_target_feature_cuda_capability30, ///< Enable CUDA compute capability 3.0 (Kepler)
1346 halide_target_feature_cuda_capability32, ///< Enable CUDA compute capability 3.2 (Tegra K1)
1347 halide_target_feature_cuda_capability35, ///< Enable CUDA compute capability 3.5 (Kepler)
1348 halide_target_feature_cuda_capability50, ///< Enable CUDA compute capability 5.0 (Maxwell)
1349 halide_target_feature_cuda_capability61, ///< Enable CUDA compute capability 6.1 (Pascal)
1350 halide_target_feature_cuda_capability70, ///< Enable CUDA compute capability 7.0 (Volta)
1351 halide_target_feature_cuda_capability75, ///< Enable CUDA compute capability 7.5 (Turing)
1352 halide_target_feature_cuda_capability80, ///< Enable CUDA compute capability 8.0 (Ampere)
1353 halide_target_feature_cuda_capability86, ///< Enable CUDA compute capability 8.6 (Ampere)
1354
1355 halide_target_feature_opencl, ///< Enable the OpenCL runtime.
1356 halide_target_feature_cl_doubles, ///< Enable double support on OpenCL targets
1357 halide_target_feature_cl_atomic64, ///< Enable 64-bit atomics operations on OpenCL targets
1358
1359 halide_target_feature_openglcompute, ///< Enable OpenGL Compute runtime. NOTE: This feature is deprecated and will be removed in Halide 17.
1360
1361 halide_target_feature_user_context, ///< Generated code takes a user_context pointer as first argument
1362
1363 halide_target_feature_profile, ///< Launch a sampling profiler alongside the Halide pipeline that monitors and reports the runtime used by each Func
1364 halide_target_feature_no_runtime, ///< Do not include a copy of the Halide runtime in any generated object file or assembly
1365
1366 halide_target_feature_metal, ///< Enable the (Apple) Metal runtime.
1367
1368 halide_target_feature_c_plus_plus_mangling, ///< Generate C++ mangled names for result function, et al
1369
1370 halide_target_feature_large_buffers, ///< Enable 64-bit buffer indexing to support buffers > 2GB. Ignored if bits != 64.
1371
1372 halide_target_feature_hvx_128, ///< Enable HVX 128 byte mode.
1373 halide_target_feature_hvx_v62, ///< Enable Hexagon v62 architecture.
1374 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.
1375 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.
1376 halide_target_feature_msan, ///< Enable hooks for MSAN support.
1377 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.
1378 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.
1379 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.
1380 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.
1381 halide_target_feature_avx512_zen4, ///< Enable the AVX512 features supported by Zen4 processors. This include all of the Cannonlake features, plus AVX512-VNNI, AVX512-BF16, and more.
1382 halide_target_feature_avx512_sapphirerapids, ///< Enable the AVX512 features supported by Sapphire Rapids processors. This include all of the Zen4 features, plus AVX-VNNI and AMX instructions.
1383 halide_target_feature_trace_loads, ///< Trace all loads done by the pipeline. Equivalent to calling Func::trace_loads on every non-inlined Func.
1384 halide_target_feature_trace_stores, ///< Trace all stores done by the pipeline. Equivalent to calling Func::trace_stores on every non-inlined Func.
1385 halide_target_feature_trace_realizations, ///< Trace all realizations done by the pipeline. Equivalent to calling Func::trace_realizations on every non-inlined Func.
1387 halide_target_feature_hvx_v65, ///< Enable Hexagon v65 architecture.
1388 halide_target_feature_hvx_v66, ///< Enable Hexagon v66 architecture.
1389 halide_target_feature_cl_half, ///< Enable half support on OpenCL targets
1390 halide_target_feature_strict_float, ///< Turn off all non-IEEE floating-point optimization. Currently applies only to LLVM targets.
1391 halide_target_feature_tsan, ///< Enable hooks for TSAN support.
1392 halide_target_feature_asan, ///< Enable hooks for ASAN support.
1393 halide_target_feature_d3d12compute, ///< Enable Direct3D 12 Compute runtime.
1394 halide_target_feature_check_unsafe_promises, ///< Insert assertions for promises.
1395 halide_target_feature_hexagon_dma, ///< Enable Hexagon DMA buffers.
1396 halide_target_feature_embed_bitcode, ///< Emulate clang -fembed-bitcode flag.
1397 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.)
1398 halide_target_feature_wasm_mvponly, ///< Disable all extensions to WebAssembly codegen (including +sign-ext and +nontrapping-fptoint, which are on by default).
1399 halide_target_feature_wasm_simd128, ///< Enable +simd128 instructions for WebAssembly codegen.
1400 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.
1401 halide_target_feature_wasm_bulk_memory, ///< Enable +bulk-memory instructions for WebAssembly codegen.
1402 halide_target_feature_webgpu, ///< Enable the WebGPU runtime.
1403 halide_target_feature_sve, ///< Enable ARM Scalable Vector Extensions
1404 halide_target_feature_sve2, ///< Enable ARM Scalable Vector Extensions v2
1405 halide_target_feature_egl, ///< Force use of EGL support.
1406 halide_target_feature_arm_dot_prod, ///< Enable ARMv8.2-a dotprod extension (i.e. udot and sdot instructions)
1407 halide_target_feature_arm_fp16, ///< Enable ARMv8.2-a half-precision floating point data processing
1408 halide_llvm_large_code_model, ///< Use the LLVM large code model to compile
1409 halide_target_feature_rvv, ///< Enable RISCV "V" Vector Extension
1410 halide_target_feature_armv81a, ///< Enable ARMv8.1-a instructions
1411 halide_target_feature_sanitizer_coverage, ///< Enable hooks for SanitizerCoverage support.
1412 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.
1413 halide_target_feature_spirv, ///< Enable SPIR-V code generation support.
1414 halide_target_feature_vulkan, ///< Enable Vulkan runtime support.
1415 halide_target_feature_vulkan_int8, ///< Enable Vulkan 8-bit integer support.
1416 halide_target_feature_vulkan_int16, ///< Enable Vulkan 16-bit integer support.
1417 halide_target_feature_vulkan_int64, ///< Enable Vulkan 64-bit integer support.
1418 halide_target_feature_vulkan_float16, ///< Enable Vulkan 16-bit float support.
1419 halide_target_feature_vulkan_float64, ///< Enable Vulkan 64-bit float support.
1420 halide_target_feature_vulkan_version10, ///< Enable Vulkan v1.0 runtime target support.
1421 halide_target_feature_vulkan_version12, ///< Enable Vulkan v1.2 runtime target support.
1422 halide_target_feature_vulkan_version13, ///< Enable Vulkan v1.3 runtime target support.
1423 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.
1424 halide_target_feature_end ///< A sentinel. Every target is considered to have this feature, and setting this feature does nothing.
1426
1427/** This function is called internally by Halide in some situations to determine
1428 * if the current execution environment can support the given set of
1429 * halide_target_feature_t flags. The implementation must do the following:
1430 *
1431 * -- If there are flags set in features that the function knows *cannot* be supported, return 0.
1432 * -- Otherwise, return 1.
1433 * -- Note that any flags set in features that the function doesn't know how to test should be ignored;
1434 * this implies that a return value of 1 means "not known to be bad" rather than "known to be good".
1435 *
1436 * In other words: a return value of 0 means "It is not safe to use code compiled with these features",
1437 * while a return value of 1 means "It is not obviously unsafe to use code compiled with these features".
1438 *
1439 * The default implementation simply calls halide_default_can_use_target_features.
1440 *
1441 * Note that `features` points to an array of `count` uint64_t; this array must contain enough
1442 * bits to represent all the currently known features. Any excess bits must be set to zero.
1443 */
1444// @{
1445extern int halide_can_use_target_features(int count, const uint64_t *features);
1446typedef int (*halide_can_use_target_features_t)(int count, const uint64_t *features);
1448// @}
1449
1450/**
1451 * This is the default implementation of halide_can_use_target_features; it is provided
1452 * for convenience of user code that may wish to extend halide_can_use_target_features
1453 * but continue providing existing support, e.g.
1454 *
1455 * int halide_can_use_target_features(int count, const uint64_t *features) {
1456 * if (features[halide_target_somefeature >> 6] & (1LL << (halide_target_somefeature & 63))) {
1457 * if (!can_use_somefeature()) {
1458 * return 0;
1459 * }
1460 * }
1461 * return halide_default_can_use_target_features(count, features);
1462 * }
1463 */
1464extern int halide_default_can_use_target_features(int count, const uint64_t *features);
1465
1466typedef struct halide_dimension_t {
1467#if (__cplusplus >= 201103L || _MSVC_LANG >= 201103L)
1468 int32_t min = 0, extent = 0, stride = 0;
1469
1470 // Per-dimension flags. None are defined yet (This is reserved for future use).
1471 uint32_t flags = 0;
1472
1475 : min(m), extent(e), stride(s), flags(f) {
1476 }
1477
1478 HALIDE_ALWAYS_INLINE bool operator==(const halide_dimension_t &other) const {
1479 return (min == other.min) &&
1480 (extent == other.extent) &&
1481 (stride == other.stride) &&
1482 (flags == other.flags);
1483 }
1484
1485 HALIDE_ALWAYS_INLINE bool operator!=(const halide_dimension_t &other) const {
1486 return !(*this == other);
1487 }
1488#else
1490
1491 // Per-dimension flags. None are defined yet (This is reserved for future use).
1493#endif
1495
1496#ifdef __cplusplus
1497} // extern "C"
1498#endif
1499
1502
1503/**
1504 * The raw representation of an image passed around by generated
1505 * Halide code. It includes some stuff to track whether the image is
1506 * not actually in main memory, but instead on a device (like a
1507 * GPU). For a more convenient C++ wrapper, use Halide::Buffer<T>. */
1508typedef struct halide_buffer_t {
1509 /** A device-handle for e.g. GPU memory used to back this buffer. */
1511
1512 /** The interface used to interpret the above handle. */
1514
1515 /** A pointer to the start of the data in main memory. In terms of
1516 * the Halide coordinate system, this is the address of the min
1517 * coordinates (defined below). */
1519
1520 /** flags with various meanings. */
1522
1523 /** The type of each buffer element. */
1525
1526 /** The dimensionality of the buffer. */
1528
1529 /** The shape of the buffer. Halide does not own this array - you
1530 * must manage the memory for it yourself. */
1532
1533 /** Pads the buffer up to a multiple of 8 bytes */
1534 void *padding;
1535
1536#if (__cplusplus >= 201103L || _MSVC_LANG >= 201103L)
1537 /** Convenience methods for accessing the flags */
1538 // @{
1539 HALIDE_ALWAYS_INLINE bool get_flag(halide_buffer_flags flag) const {
1540 return (flags & flag) != 0;
1541 }
1542
1543 HALIDE_ALWAYS_INLINE void set_flag(halide_buffer_flags flag, bool value) {
1544 if (value) {
1545 flags |= flag;
1546 } else {
1547 flags &= ~uint64_t(flag);
1548 }
1549 }
1550
1551 HALIDE_ALWAYS_INLINE bool host_dirty() const {
1552 return get_flag(halide_buffer_flag_host_dirty);
1553 }
1554
1555 HALIDE_ALWAYS_INLINE bool device_dirty() const {
1556 return get_flag(halide_buffer_flag_device_dirty);
1557 }
1558
1559 HALIDE_ALWAYS_INLINE void set_host_dirty(bool v = true) {
1560 set_flag(halide_buffer_flag_host_dirty, v);
1561 }
1562
1563 HALIDE_ALWAYS_INLINE void set_device_dirty(bool v = true) {
1565 }
1566 // @}
1567
1568 /** The total number of elements this buffer represents. Equal to
1569 * the product of the extents */
1570 HALIDE_ALWAYS_INLINE size_t number_of_elements() const {
1571 size_t s = 1;
1572 for (int i = 0; i < dimensions; i++) {
1573 s *= dim[i].extent;
1574 }
1575 return s;
1576 }
1577
1578 /** Offset to the element with the lowest address.
1579 * If all strides are positive, equal to zero.
1580 * Offset is in elements, not bytes.
1581 * Unlike begin(), this is ok to call on an unallocated buffer. */
1582 HALIDE_ALWAYS_INLINE ptrdiff_t begin_offset() const {
1583 ptrdiff_t index = 0;
1584 for (int i = 0; i < dimensions; i++) {
1585 const int stride = dim[i].stride;
1586 if (stride < 0) {
1587 index += stride * (ptrdiff_t)(dim[i].extent - 1);
1588 }
1589 }
1590 return index;
1591 }
1592
1593 /** An offset to one beyond the element with the highest address.
1594 * Offset is in elements, not bytes.
1595 * Unlike end(), this is ok to call on an unallocated buffer. */
1596 HALIDE_ALWAYS_INLINE ptrdiff_t end_offset() const {
1597 ptrdiff_t index = 0;
1598 for (int i = 0; i < dimensions; i++) {
1599 const int stride = dim[i].stride;
1600 if (stride > 0) {
1601 index += stride * (ptrdiff_t)(dim[i].extent - 1);
1602 }
1603 }
1604 index += 1;
1605 return index;
1606 }
1607
1608 /** A pointer to the element with the lowest address.
1609 * If all strides are positive, equal to the host pointer.
1610 * Illegal to call on an unallocated buffer. */
1611 HALIDE_ALWAYS_INLINE uint8_t *begin() const {
1612 return host + begin_offset() * type.bytes();
1613 }
1614
1615 /** A pointer to one beyond the element with the highest address.
1616 * Illegal to call on an unallocated buffer. */
1617 HALIDE_ALWAYS_INLINE uint8_t *end() const {
1618 return host + end_offset() * type.bytes();
1619 }
1620
1621 /** The total number of bytes spanned by the data in memory. */
1622 HALIDE_ALWAYS_INLINE size_t size_in_bytes() const {
1623 return (size_t)(end_offset() - begin_offset()) * type.bytes();
1624 }
1625
1626 /** A pointer to the element at the given location. */
1627 HALIDE_ALWAYS_INLINE uint8_t *address_of(const int *pos) const {
1628 ptrdiff_t index = 0;
1629 for (int i = 0; i < dimensions; i++) {
1630 index += (ptrdiff_t)dim[i].stride * (pos[i] - dim[i].min);
1631 }
1632 return host + index * type.bytes();
1633 }
1634
1635 /** Attempt to call device_sync for the buffer. If the buffer
1636 * has no device_interface (or no device_sync), this is a quiet no-op.
1637 * Calling this explicitly should rarely be necessary, except for profiling. */
1638 HALIDE_ALWAYS_INLINE int device_sync(void *ctx = nullptr) {
1640 return device_interface->device_sync(ctx, this);
1641 }
1642 return 0;
1643 }
1644
1645 /** Check if an input buffer passed extern stage is a querying
1646 * bounds. Compared to doing the host pointer check directly,
1647 * this both adds clarity to code and will facilitate moving to
1648 * another representation for bounds query arguments. */
1649 HALIDE_ALWAYS_INLINE bool is_bounds_query() const {
1650 return host == nullptr && device == 0;
1651 }
1652
1653#endif
1655
1656#ifdef __cplusplus
1657extern "C" {
1658#endif
1659
1660#ifndef HALIDE_ATTRIBUTE_DEPRECATED
1661#ifdef HALIDE_ALLOW_DEPRECATED
1662#define HALIDE_ATTRIBUTE_DEPRECATED(x)
1663#else
1664#ifdef _MSC_VER
1665#define HALIDE_ATTRIBUTE_DEPRECATED(x) __declspec(deprecated(x))
1666#else
1667#define HALIDE_ATTRIBUTE_DEPRECATED(x) __attribute__((deprecated(x)))
1668#endif
1669#endif
1670#endif
1671
1672/** halide_scalar_value_t is a simple union able to represent all the well-known
1673 * scalar values in a filter argument. Note that it isn't tagged with a type;
1674 * you must ensure you know the proper type before accessing. Most user
1675 * code will never need to create instances of this struct; its primary use
1676 * is to hold def/min/max values in a halide_filter_argument_t. (Note that
1677 * this is conceptually just a union; it's wrapped in a struct to ensure
1678 * that it doesn't get anonymized by LLVM.)
1679 */
1681 union {
1682 bool b;
1691 float f32;
1692 double f64;
1693 void *handle;
1694 } u;
1695#ifdef __cplusplus
1697 u.u64 = 0;
1698 }
1699#endif
1700};
1701
1707
1708/*
1709 These structs must be robust across different compilers and settings; when
1710 modifying them, strive for the following rules:
1711
1712 1) All fields are explicitly sized. I.e. must use int32_t and not "int"
1713 2) All fields must land on an alignment boundary that is the same as their size
1714 3) Explicit padding is added to make that so
1715 4) The sizeof the struct is padded out to a multiple of the largest natural size thing in the struct
1716 5) don't forget that 32 and 64 bit pointers are different sizes
1717*/
1718
1719/**
1720 * Obsolete version of halide_filter_argument_t; only present in
1721 * code that wrote halide_filter_metadata_t version 0.
1722 */
1730
1731/**
1732 * halide_filter_argument_t is essentially a plain-C-struct equivalent to
1733 * Halide::Argument; most user code will never need to create one.
1734 */
1736 const char *name; // name of the argument; will never be null or empty.
1737 int32_t kind; // actually halide_argument_kind_t
1738 int32_t dimensions; // always zero for scalar arguments
1740 // These pointers should always be null for buffer arguments,
1741 // and *may* be null for scalar arguments. (A null value means
1742 // there is no def/min/max/estimate specified for this argument.)
1744 // This pointer should always be null for scalar arguments,
1745 // and *may* be null for buffer arguments. If not null, it should always
1746 // point to an array of dimensions*2 pointers, which will be the (min, extent)
1747 // estimates for each dimension of the buffer. (Note that any of the pointers
1748 // may be null as well.)
1750};
1751
1753#ifdef __cplusplus
1754 static const int32_t VERSION = 1;
1755#endif
1756
1757 /** version of this metadata; currently always 1. */
1759
1760 /** The number of entries in the arguments field. This is always >= 1. */
1762
1763 /** An array of the filters input and output arguments; this will never be
1764 * null. The order of arguments is not guaranteed (input and output arguments
1765 * may come in any order); however, it is guaranteed that all arguments
1766 * will have a unique name within a given filter. */
1768
1769 /** The Target for which the filter was compiled. This is always
1770 * a canonical Target string (ie a product of Target::to_string). */
1771 const char *target;
1772
1773 /** The function name of the filter. */
1774 const char *name;
1775};
1776
1777/** halide_register_argv_and_metadata() is a **user-defined** function that
1778 * must be provided in order to use the registration.cc files produced
1779 * by Generators when the 'registration' output is requested. Each registration.cc
1780 * file provides a static initializer that calls this function with the given
1781 * filter's argv-call variant, its metadata, and (optionally) and additional
1782 * textual data that the build system chooses to tack on for its own purposes.
1783 * Note that this will be called at static-initializer time (i.e., before
1784 * main() is called), and in an unpredictable order. Note that extra_key_value_pairs
1785 * may be nullptr; if it's not null, it's expected to be a null-terminated list
1786 * of strings, with an even number of entries. */
1788 int (*filter_argv_call)(void **),
1789 const struct halide_filter_metadata_t *filter_metadata,
1790 const char *const *extra_key_value_pairs);
1791
1792/** The functions below here are relevant for pipelines compiled with
1793 * the -profile target flag, which runs a sampling profiler thread
1794 * alongside the pipeline. */
1795
1796/** Per-Func state tracked by the sampling profiler. */
1797struct HALIDE_ATTRIBUTE_ALIGN(8) halide_profiler_func_stats {
1798 /** Total time taken evaluating this Func (in nanoseconds). */
1799 uint64_t time;
1800
1801 /** The current memory allocation of this Func. */
1802 uint64_t memory_current;
1803
1804 /** The peak memory allocation of this Func. */
1805 uint64_t memory_peak;
1806
1807 /** The total memory allocation of this Func. */
1808 uint64_t memory_total;
1809
1810 /** The peak stack allocation of this Func's threads. */
1811 uint64_t stack_peak;
1812
1813 /** The average number of thread pool worker threads active while computing this Func. */
1814 uint64_t active_threads_numerator, active_threads_denominator;
1815
1816 /** The name of this Func. A global constant string. */
1817 const char *name;
1818
1819 /** The total number of memory allocation of this Func. */
1820 int num_allocs;
1821};
1822
1823/** Per-pipeline state tracked by the sampling profiler. These exist
1824 * in a linked list. */
1825struct HALIDE_ATTRIBUTE_ALIGN(8) halide_profiler_pipeline_stats {
1826 /** Total time spent inside this pipeline (in nanoseconds) */
1827 uint64_t time;
1828
1829 /** The current memory allocation of funcs in this pipeline. */
1830 uint64_t memory_current;
1831
1832 /** The peak memory allocation of funcs in this pipeline. */
1833 uint64_t memory_peak;
1834
1835 /** The total memory allocation of funcs in this pipeline. */
1836 uint64_t memory_total;
1837
1838 /** The average number of thread pool worker threads doing useful
1839 * work while computing this pipeline. */
1840 uint64_t active_threads_numerator, active_threads_denominator;
1841
1842 /** The name of this pipeline. A global constant string. */
1843 const char *name;
1844
1845 /** An array containing states for each Func in this pipeline. */
1846 struct halide_profiler_func_stats *funcs;
1847
1848 /** The next pipeline_stats pointer. It's a void * because types
1849 * in the Halide runtime may not currently be recursive. */
1850 void *next;
1851
1852 /** The number of funcs in this pipeline. */
1853 int num_funcs;
1854
1855 /** An internal base id used to identify the funcs in this pipeline. */
1856 int first_func_id;
1857
1858 /** The number of times this pipeline has been run. */
1859 int runs;
1860
1861 /** The total number of samples taken inside of this pipeline. */
1862 int samples;
1863
1864 /** The total number of memory allocation of funcs in this pipeline. */
1865 int num_allocs;
1866};
1867
1868/** The global state of the profiler. */
1869
1871 /** Guards access to the fields below. If not locked, the sampling
1872 * profiler thread is free to modify things below (including
1873 * reordering the linked list of pipeline stats). */
1875
1876 /** The amount of time the profiler thread sleeps between samples
1877 * in milliseconds. Defaults to 1 */
1879
1880 /** An internal id used for bookkeeping. */
1882
1883 /** The id of the current running Func. Set by the pipeline, read
1884 * periodically by the profiler thread. */
1886
1887 /** The number of threads currently doing work. */
1889
1890 /** A linked list of stats gathered for each pipeline. */
1891 struct halide_profiler_pipeline_stats *pipelines;
1892
1893 /** Retrieve remote profiler state. Used so that the sampling
1894 * profiler can follow along with execution that occurs elsewhere,
1895 * e.g. on a DSP. If null, it reads from the int above instead. */
1896 void (*get_remote_profiler_state)(int *func, int *active_workers);
1897
1898 /** Sampling thread reference to be joined at shutdown. */
1899 struct halide_thread *sampling_thread;
1900};
1901
1902/** Profiler func ids with special meanings. */
1903enum {
1904 /// current_func takes on this value when not inside Halide code
1906 /// Set current_func to this value to tell the profiling thread to
1907 /// halt. It will start up again next time you run a pipeline with
1908 /// profiling enabled.
1911
1912/** Get a pointer to the global profiler state for programmatic
1913 * inspection. Lock it before using to pause the profiler. */
1915
1916/** Get a pointer to the pipeline state associated with pipeline_name.
1917 * This function grabs the global profiler state's lock on entry. */
1918extern struct halide_profiler_pipeline_stats *halide_profiler_get_pipeline_state(const char *pipeline_name);
1919
1920/** Collects profiling information. Intended to be called from a timer
1921 * interrupt handler if timer based profiling is being used.
1922 * State argument is acquired via halide_profiler_get_pipeline_state.
1923 * prev_t argument is the previous time and can be used to set a more
1924 * accurate time interval if desired. */
1926
1927/** Reset profiler state cheaply. May leave threads running or some
1928 * memory allocated but all accumluated statistics are reset.
1929 * WARNING: Do NOT call this method while any halide pipeline is
1930 * running; halide_profiler_memory_allocate/free and
1931 * halide_profiler_stack_peak_update update the profiler pipeline's
1932 * state without grabbing the global profiler state's lock. */
1934
1935/** Reset all profiler state.
1936 * WARNING: Do NOT call this method while any halide pipeline is
1937 * running; halide_profiler_memory_allocate/free and
1938 * halide_profiler_stack_peak_update update the profiler pipeline's
1939 * state without grabbing the global profiler state's lock. */
1941
1942/** Print out timing statistics for everything run since the last
1943 * reset. Also happens at process exit. */
1944extern void halide_profiler_report(void *user_context);
1945
1946/** For timer based profiling, this routine starts the timer chain running.
1947 * halide_get_profiler_state can be called to get the current timer interval.
1948 */
1950/** These routines are called to temporarily disable and then reenable
1951 * timer interuppts for profiling */
1952//@{
1955//@}
1956
1957/// \name "Float16" functions
1958/// These functions operate of bits (``uint16_t``) representing a half
1959/// precision floating point number (IEEE-754 2008 binary16).
1960//{@
1961
1962/** Read bits representing a half precision floating point number and return
1963 * the float that represents the same value */
1965
1966/** Read bits representing a half precision floating point number and return
1967 * the double that represents the same value */
1969
1970// TODO: Conversion functions to half
1971
1972//@}
1973
1974// Allocating and freeing device memory is often very slow. The
1975// methods below give Halide's runtime permission to hold onto device
1976// memory to service future requests instead of returning it to the
1977// underlying device API. The API does not manage an allocation pool,
1978// all it does is provide access to a shared counter that acts as a
1979// limit on the unused memory not yet returned to the underlying
1980// device API. It makes callbacks to participants when memory needs to
1981// be released because the limit is about to be exceeded (either
1982// because the limit has been reduced, or because the memory owned by
1983// some participant becomes unused).
1984
1985/** Tell Halide whether or not it is permitted to hold onto device
1986 * allocations to service future requests instead of returning them
1987 * eagerly to the underlying device API. Many device allocators are
1988 * quite slow, so it can be beneficial to set this to true. The
1989 * default value for now is false.
1990 *
1991 * Note that if enabled, the eviction policy is very simplistic. The
1992 * 32 most-recently used allocations are preserved, regardless of
1993 * their size. Additionally, if a call to cuMalloc results in an
1994 * out-of-memory error, the entire cache is flushed and the allocation
1995 * is retried. See https://github.com/halide/Halide/issues/4093
1996 *
1997 * If set to false, releases all unused device allocations back to the
1998 * underlying device APIs. For finer-grained control, see specific
1999 * methods in each device api runtime.
2000 *
2001 * Note that if the flag is set to true, this call *must* succeed and return
2002 * a value of halide_error_code_success (i.e., zero); if you replace
2003 * the implementation of this call in the runtime, you must honor this contract.
2004 * */
2005extern int halide_reuse_device_allocations(void *user_context, bool);
2006
2007/** Determines whether on device_free the memory is returned
2008 * immediately to the device API, or placed on a free list for future
2009 * use. Override and switch based on the user_context for
2010 * finer-grained control. By default just returns the value most
2011 * recently set by the method above. */
2012extern bool halide_can_reuse_device_allocations(void *user_context);
2013
2018
2019/** Register a callback to be informed when
2020 * halide_reuse_device_allocations(false) is called, and all unused
2021 * device allocations must be released. The object passed should have
2022 * global lifetime, and its next field will be clobbered. */
2024
2025#ifdef __cplusplus
2026} // End extern "C"
2027#endif
2028
2029#if (__cplusplus >= 201103L || _MSVC_LANG >= 201103L)
2030
2031namespace {
2032
2033template<typename T>
2034struct check_is_pointer {
2035 static constexpr bool value = false;
2036};
2037
2038template<typename T>
2039struct check_is_pointer<T *> {
2040 static constexpr bool value = true;
2041};
2042
2043} // namespace
2044
2045/** Construct the halide equivalent of a C type */
2046template<typename T>
2047HALIDE_ALWAYS_INLINE constexpr halide_type_t halide_type_of() {
2048 // Create a compile-time error if T is not a pointer (without
2049 // using any includes - this code goes into the runtime).
2050 // (Note that we can't have uninitialized variables in constexpr functions,
2051 // even if those variables aren't used.)
2052 static_assert(check_is_pointer<T>::value, "Expected a pointer type here");
2054}
2055
2056#ifdef HALIDE_CPP_COMPILER_HAS_FLOAT16
2057template<>
2058HALIDE_ALWAYS_INLINE constexpr halide_type_t halide_type_of<_Float16>() {
2059 return halide_type_t(halide_type_float, 16);
2060}
2061#endif
2062
2063template<>
2064HALIDE_ALWAYS_INLINE constexpr halide_type_t halide_type_of<float>() {
2065 return halide_type_t(halide_type_float, 32);
2066}
2067
2068template<>
2069HALIDE_ALWAYS_INLINE constexpr halide_type_t halide_type_of<double>() {
2070 return halide_type_t(halide_type_float, 64);
2071}
2072
2073template<>
2074HALIDE_ALWAYS_INLINE constexpr halide_type_t halide_type_of<bool>() {
2076}
2077
2078template<>
2079HALIDE_ALWAYS_INLINE constexpr halide_type_t halide_type_of<uint8_t>() {
2081}
2082
2083template<>
2084HALIDE_ALWAYS_INLINE constexpr halide_type_t halide_type_of<uint16_t>() {
2085 return halide_type_t(halide_type_uint, 16);
2086}
2087
2088template<>
2089HALIDE_ALWAYS_INLINE constexpr halide_type_t halide_type_of<uint32_t>() {
2090 return halide_type_t(halide_type_uint, 32);
2091}
2092
2093template<>
2094HALIDE_ALWAYS_INLINE constexpr halide_type_t halide_type_of<uint64_t>() {
2095 return halide_type_t(halide_type_uint, 64);
2096}
2097
2098template<>
2099HALIDE_ALWAYS_INLINE constexpr halide_type_t halide_type_of<int8_t>() {
2100 return halide_type_t(halide_type_int, 8);
2101}
2102
2103template<>
2104HALIDE_ALWAYS_INLINE constexpr halide_type_t halide_type_of<int16_t>() {
2105 return halide_type_t(halide_type_int, 16);
2106}
2107
2108template<>
2109HALIDE_ALWAYS_INLINE constexpr halide_type_t halide_type_of<int32_t>() {
2110 return halide_type_t(halide_type_int, 32);
2111}
2112
2113template<>
2114HALIDE_ALWAYS_INLINE constexpr halide_type_t halide_type_of<int64_t>() {
2115 return halide_type_t(halide_type_int, 64);
2116}
2117
2118#ifndef COMPILING_HALIDE_RUNTIME
2119
2120// These structures are used by `function_info_header` files
2121// (generated by passing `-e function_info_header` to a Generator).
2122// The generated files contain documentation on the proper usage.
2123namespace HalideFunctionInfo {
2124
2125enum ArgumentKind { InputScalar = 0,
2126 InputBuffer = 1,
2127 OutputBuffer = 2 };
2128
2129struct ArgumentInfo {
2130 std::string_view name;
2131 ArgumentKind kind;
2132 int32_t dimensions; // always zero for scalar arguments
2133 halide_type_t type;
2134};
2135
2136} // namespace HalideFunctionInfo
2137
2138#endif // COMPILING_HALIDE_RUNTIME
2139
2140#endif // (__cplusplus >= 201103L || _MSVC_LANG >= 201103L)
2141
2142#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_avx512_zen4
Enable the AVX512 features supported by Zen4 processors. This include all of the Cannonlake features,...
@ 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_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 Zen4 feat...
@ 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_wasm_mvponly
Disable all extensions to WebAssembly codegen (including +sign-ext and +nontrapping-fptoint,...
@ 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_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)
struct halide_dimension_t halide_dimension_t
int halide_error_requirement_failed(void *user_context, const char *condition, const char *message)
void halide_memoization_cache_release(void *user_context, void *host)
If halide_memoization_cache_lookup succeeds, halide_memoization_cache_release must be called to signa...
int(* halide_can_use_target_features_t)(int count, const uint64_t *features)
void halide_register_argv_and_metadata(int(*filter_argv_call)(void **), const struct halide_filter_metadata_t *filter_metadata, const char *const *extra_key_value_pairs)
halide_register_argv_and_metadata() is a user-defined function that must be provided in order to use ...
void *(* halide_get_symbol_t)(const char *name)
int halide_error_param_too_large_f64(void *user_context, const char *param_name, double val, double max_val)
int halide_msan_check_memory_is_initialized(void *user_context, const void *ptr, uint64_t len, const char *name)
Verify that a given range of memory has been initialized; only used when Target::MSAN is enabled.
int(* halide_do_loop_task_t)(void *, halide_loop_task_t, int, int, uint8_t *, void *)
The version of do_task called for loop tasks.
int halide_error_buffer_extents_too_large(void *user_context, const char *buffer_name, int64_t actual_size, int64_t max_size)
int32_t halide_trace(void *user_context, const struct halide_trace_event_t *event)
Called when Funcs are marked as trace_load, trace_store, or trace_realization.
int halide_error_extern_stage_failed(void *user_context, const char *extern_stage_name, int result)
A call to an extern stage failed.
halide_can_use_target_features_t halide_set_custom_can_use_target_features(halide_can_use_target_features_t)
int halide_error_host_is_null(void *user_context, const char *func_name)
void halide_set_trace_file(int fd)
Set the file descriptor that Halide should write binary trace events to.
int halide_memoization_cache_store(void *user_context, const uint8_t *cache_key, int32_t size, struct halide_buffer_t *realized_bounds, int32_t tuple_count, struct halide_buffer_t **tuple_buffers, bool has_eviction_key, uint64_t eviction_key)
Given a cache key for a memoized result, currently constructed from the Func name and top-level Func ...
int halide_buffer_copy(void *user_context, struct halide_buffer_t *src, const struct halide_device_interface_t *dst_device_interface, struct halide_buffer_t *dst)
Copy data from one buffer to another.
int halide_error_fold_factor_too_small(void *user_context, const char *func_name, const char *var_name, int fold_factor, const char *loop_name, int required_extent)
int halide_error_param_too_small_f64(void *user_context, const char *param_name, double val, double min_val)
int halide_error_param_too_large_i64(void *user_context, const char *param_name, int64_t val, int64_t max_val)
void halide_enable_timer_interrupt()
int halide_error_param_too_small_u64(void *user_context, const char *param_name, uint64_t val, uint64_t min_val)
void(* halide_print_t)(void *, const char *)
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_device_sync_global(void *user_context, const struct halide_device_interface_t *device_interface)
Wait for current GPU operations to complete.
int halide_error_param_too_large_u64(void *user_context, const char *param_name, uint64_t val, uint64_t max_val)
int32_t(* halide_trace_t)(void *user_context, const struct halide_trace_event_t *)
struct halide_profiler_state * halide_profiler_get_state()
Get a pointer to the global profiler state for programmatic inspection.
int(* halide_do_task_t)(void *, halide_task_t, int, uint8_t *)
If you use the default do_par_for, you can still set a custom handler to perform each individual task...
halide_free_t halide_set_custom_free(halide_free_t user_free)
int halide_default_do_parallel_tasks(void *user_context, int num_tasks, struct halide_parallel_task_t *tasks, void *task_parent)
int halide_get_gpu_device(void *user_context)
Halide calls this to get the desired halide gpu device setting.
int(* halide_semaphore_init_t)(struct halide_semaphore_t *, int)
halide_do_loop_task_t halide_set_custom_do_loop_task(halide_do_loop_task_t do_task)
int halide_device_detach_native(void *user_context, struct halide_buffer_t *buf)
void * halide_default_get_symbol(const char *name)
int halide_copy_to_device(void *user_context, struct halide_buffer_t *buf, const struct halide_device_interface_t *device_interface)
Copy image data from host memory to device memory.
halide_get_library_symbol_t halide_set_custom_get_library_symbol(halide_get_library_symbol_t user_get_library_symbol)
void halide_error(void *user_context, const char *)
Halide calls this function on runtime errors (for example bounds checking failures).
int halide_device_release_crop(void *user_context, struct halide_buffer_t *buf)
Release any resources associated with a cropped/sliced view of another buffer.
int halide_can_use_target_features(int count, const uint64_t *features)
This function is called internally by Halide in some situations to determine if the current execution...
int halide_error_bad_type(void *user_context, const char *func_name, uint32_t type_given, uint32_t correct_type)
halide_do_task_t halide_set_custom_do_task(halide_do_task_t do_task)
int halide_do_parallel_tasks(void *user_context, int num_tasks, struct halide_parallel_task_t *tasks, void *task_parent)
Enqueue some number of the tasks described above and wait for them to complete.
void *(* halide_malloc_t)(void *, size_t)
int halide_device_wrap_native(void *user_context, struct halide_buffer_t *buf, uint64_t handle, const struct halide_device_interface_t *device_interface)
Wrap or detach a native device handle, setting the device field and device_interface field as appropr...
int halide_get_trace_file(void *user_context)
Halide calls this to retrieve the file descriptor to write binary trace events to.
float halide_float16_bits_to_float(uint16_t)
Read bits representing a half precision floating point number and return the float that represents th...
int halide_error_unaligned_host_ptr(void *user_context, const char *func_name, int alignment)
int halide_profiler_sample(struct halide_profiler_state *s, uint64_t *prev_t)
Collects profiling information.
void halide_register_device_allocation_pool(struct halide_device_allocation_pool *)
Register a callback to be informed when halide_reuse_device_allocations(false) is called,...
#define HALIDE_ATTRIBUTE_ALIGN(x)
int halide_msan_check_buffer_is_initialized(void *user_context, struct halide_buffer_t *buffer, const char *buf_name)
Verify that the data pointed to by the halide_buffer_t is initialized (but not the halide_buffer_t it...
int(* halide_do_parallel_tasks_t)(void *, int, struct halide_parallel_task_t *, void *task_parent)
Provide an entire custom tasking runtime via function pointers.
halide_argument_kind_t
@ halide_argument_kind_output_buffer
@ halide_argument_kind_input_scalar
@ halide_argument_kind_input_buffer
int halide_semaphore_release(struct halide_semaphore_t *, int n)
struct halide_buffer_t halide_buffer_t
The raw representation of an image passed around by generated Halide code.
Expr with_lanes(const Expr &x, int lanes)
Rewrite the expression x to have lanes lanes.
auto operator==(const Other &a, const GeneratorParam< T > &b) -> decltype(a==(T) b)
Equality comparison between GeneratorParam<T> and any type that supports operator== with T.
Definition Generator.h:1131
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:1092
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:1144
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.