Halide  20.0.0
Halide compiler and libraries
HalideRuntime.h
Go to the documentation of this file.
1 #ifndef HALIDE_HALIDERUNTIME_H
2 #define HALIDE_HALIDERUNTIME_H
3 
4 #ifndef COMPILING_HALIDE_RUNTIME
5 #ifdef __cplusplus
6 #include <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 20
27 #define HALIDE_VERSION_MINOR 0
28 #define HALIDE_VERSION_PATCH 0
29 
30 #ifdef __cplusplus
31 // Forward declare type to allow naming typed handles.
32 // See Type.h for documentation.
33 template<typename T>
35 #endif
36 
37 #ifdef __cplusplus
38 extern "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 // ASAN builds can cause linker errors for Float16, so sniff for that and
90 // don't enable it by default.
91 #if defined(__has_feature)
92 #if __has_feature(address_sanitizer)
93 #define HALIDE_RUNTIME_ASAN_DETECTED
94 #endif
95 #endif
96 
97 #if defined(__SANITIZE_ADDRESS__) && !defined(HALIDE_RUNTIME_ASAN_DETECTED)
98 #define HALIDE_RUNTIME_ASAN_DETECTED
99 #endif
100 
101 #if !defined(HALIDE_RUNTIME_ASAN_DETECTED)
102 
103 // clang had _Float16 added as a reserved name in clang 8, but
104 // doesn't actually support it on most platforms until clang 15.
105 // Ideally there would be a better way to detect if the type
106 // is supported, even in a compiler independent fashion, but
107 // coming up with one has proven elusive.
108 #if defined(__clang__) && (__clang_major__ >= 15) && !defined(__EMSCRIPTEN__) && !defined(__i386__)
109 #if defined(__is_identifier)
110 #if !__is_identifier(_Float16)
111 #define HALIDE_CPP_COMPILER_HAS_FLOAT16
112 #endif
113 #endif
114 #endif
115 
116 // Similarly, detecting _Float16 for gcc is problematic.
117 // For now, we say that if >= v12, and compiling on x86 or arm,
118 // we assume support. This may need revision.
119 #if defined(__GNUC__) && (__GNUC__ >= 12)
120 #if defined(__x86_64__) || (defined(__i386__) && (__GNUC__ >= 14) && defined(__SSE2__)) || ((defined(__arm__) || defined(__aarch64__)) && (__GNUC__ >= 13))
121 #define HALIDE_CPP_COMPILER_HAS_FLOAT16
122 #endif
123 #endif
124 
125 #endif // !HALIDE_RUNTIME_ASAN_DETECTED
126 
127 #endif // !COMPILING_HALIDE_RUNTIME
128 
129 /** \file
130  *
131  * This file declares the routines used by Halide internally in its
132  * runtime. On platforms that support weak linking, these can be
133  * replaced with user-defined versions by defining an extern "C"
134  * function with the same name and signature.
135  *
136  * When doing Just In Time (JIT) compilation members of
137  * some_pipeline_or_func.jit_handlers() must be replaced instead. The
138  * corresponding methods are documented below.
139  *
140  * All of these functions take a "void *user_context" parameter as their
141  * first argument; if the Halide kernel that calls back to any of these
142  * functions has been compiled with the UserContext feature set on its Target,
143  * then the value of that pointer passed from the code that calls the
144  * Halide kernel is piped through to the function.
145  *
146  * Some of these are also useful to call when using the default
147  * implementation. E.g. halide_shutdown_thread_pool.
148  *
149  * Note that even on platforms with weak linking, some linker setups
150  * may not respect the override you provide. E.g. if the override is
151  * in a shared library and the halide object files are linked directly
152  * into the output, the builtin versions of the runtime functions will
153  * be called. See your linker documentation for more details. On
154  * Linux, LD_DYNAMIC_WEAK=1 may help.
155  *
156  */
157 
158 // Forward-declare to suppress warnings if compiling as C.
159 struct halide_buffer_t;
160 
161 /** Print a message to stderr. Main use is to support tracing
162  * functionality, print, and print_when calls. Also called by the default
163  * halide_error. This function can be replaced in JITed code by using
164  * halide_custom_print and providing an implementation of halide_print
165  * in AOT code. See Func::set_custom_print.
166  */
167 // @{
168 extern void halide_print(void *user_context, const char *);
169 extern void halide_default_print(void *user_context, const char *);
170 typedef void (*halide_print_t)(void *, const char *);
172 // @}
173 
174 /** Halide calls this function on runtime errors (for example bounds
175  * checking failures). This function can be replaced in JITed code by
176  * using Func::set_error_handler, or in AOT code by calling
177  * halide_set_error_handler. In AOT code on platforms that support
178  * weak linking (i.e. not Windows), you can also override it by simply
179  * defining your own halide_error.
180  */
181 // @{
182 extern void halide_error(void *user_context, const char *);
183 extern void halide_default_error(void *user_context, const char *);
184 typedef void (*halide_error_handler_t)(void *, const char *);
186 // @}
187 
188 /** Cross-platform mutex. Must be initialized with zero and implementation
189  * must treat zero as an unlocked mutex with no waiters, etc.
190  */
191 struct halide_mutex {
193 };
194 
195 /** Cross platform condition variable. Must be initialized to 0. */
196 struct halide_cond {
198 };
199 
200 /** A basic set of mutex and condition variable functions, which call
201  * platform specific code for mutual exclusion. Equivalent to posix
202  * calls. */
203 //@{
204 extern void halide_mutex_lock(struct halide_mutex *mutex);
205 extern void halide_mutex_unlock(struct halide_mutex *mutex);
206 extern void halide_cond_signal(struct halide_cond *cond);
207 extern void halide_cond_broadcast(struct halide_cond *cond);
208 extern void halide_cond_wait(struct halide_cond *cond, struct halide_mutex *mutex);
209 //@}
210 
211 /** Functions for constructing/destroying/locking/unlocking arrays of mutexes. */
212 struct halide_mutex_array;
213 //@{
215 extern void halide_mutex_array_destroy(void *user_context, void *array);
216 extern int halide_mutex_array_lock(struct halide_mutex_array *array, int entry);
217 extern int halide_mutex_array_unlock(struct halide_mutex_array *array, int entry);
218 //@}
219 
220 /** Define halide_do_par_for to replace the default thread pool
221  * implementation. halide_shutdown_thread_pool can also be called to
222  * release resources used by the default thread pool on platforms
223  * where it makes sense. See Func::set_custom_do_task and
224  * Func::set_custom_do_par_for. Should return zero if all the jobs
225  * return zero, or an arbitrarily chosen return value from one of the
226  * jobs otherwise.
227  */
228 //@{
229 typedef int (*halide_task_t)(void *user_context, int task_number, uint8_t *closure);
230 extern int halide_do_par_for(void *user_context,
231  halide_task_t task,
232  int min, int size, uint8_t *closure);
233 extern void halide_shutdown_thread_pool(void);
234 //@}
235 
236 /** Set a custom method for performing a parallel for loop. Returns
237  * the old do_par_for handler. */
238 typedef int (*halide_do_par_for_t)(void *, halide_task_t, int, int, uint8_t *);
240 
241 /** An opaque struct representing a semaphore. Used by the task system for async tasks. */
244 };
245 
246 /** A struct representing a semaphore and a number of items that must
247  * be acquired from it. Used in halide_parallel_task_t below. */
250  int count;
251 };
252 extern int halide_semaphore_init(struct halide_semaphore_t *, int n);
253 extern int halide_semaphore_release(struct halide_semaphore_t *, int n);
254 extern bool halide_semaphore_try_acquire(struct halide_semaphore_t *, int n);
255 typedef int (*halide_semaphore_init_t)(struct halide_semaphore_t *, int);
256 typedef int (*halide_semaphore_release_t)(struct halide_semaphore_t *, int);
257 typedef bool (*halide_semaphore_try_acquire_t)(struct halide_semaphore_t *, int);
258 
259 /** A task representing a serial for loop evaluated over some range.
260  * Note that task_parent is a pass through argument that should be
261  * passed to any dependent taks that are invoked using halide_do_parallel_tasks
262  * underneath this call. */
263 typedef int (*halide_loop_task_t)(void *user_context, int min, int extent,
264  uint8_t *closure, void *task_parent);
265 
266 /** A parallel task to be passed to halide_do_parallel_tasks. This
267  * task may recursively call halide_do_parallel_tasks, and there may
268  * be complex dependencies between seemingly unrelated tasks expressed
269  * using semaphores. If you are using a custom task system, care must
270  * be taken to avoid potential deadlock. This can be done by carefully
271  * respecting the static metadata at the end of the task struct.*/
273  // The function to call. It takes a user context, a min and
274  // extent, a closure, and a task system pass through argument.
276 
277  // The closure to pass it
279 
280  // The name of the function to be called. For debugging purposes only.
281  const char *name;
282 
283  // An array of semaphores that must be acquired before the
284  // function is called. Must be reacquired for every call made.
287 
288  // The entire range the function should be called over. This range
289  // may be sliced up and the function called multiple times.
290  int min, extent;
291 
292  // A parallel task provides several pieces of metadata to prevent
293  // unbounded resource usage or deadlock.
294 
295  // The first is the minimum number of execution contexts (call
296  // stacks or threads) necessary for the function to run to
297  // completion. This may be greater than one when there is nested
298  // parallelism with internal producer-consumer relationships
299  // (calling the function recursively spawns and blocks on parallel
300  // sub-tasks that communicate with each other via semaphores). If
301  // a parallel runtime calls the function when fewer than this many
302  // threads are idle, it may need to create more threads to
303  // complete the task, or else risk deadlock due to committing all
304  // threads to tasks that cannot complete without more.
305  //
306  // FIXME: Note that extern stages are assumed to only require a
307  // single thread to complete. If the extern stage is itself a
308  // Halide pipeline, this may be an underestimate.
310 
311  // The calls to the function should be in serial order from min to min+extent-1, with only
312  // one executing at a time. If false, any order is fine, and
313  // concurrency is fine.
314  bool serial;
315 };
316 
317 /** Enqueue some number of the tasks described above and wait for them
318  * to complete. While waiting, the calling threads assists with either
319  * the tasks enqueued, or other non-blocking tasks in the task
320  * system. Note that task_parent should be NULL for top-level calls
321  * and the pass through argument if this call is being made from
322  * another task. */
323 extern int halide_do_parallel_tasks(void *user_context, int num_tasks,
324  struct halide_parallel_task_t *tasks,
325  void *task_parent);
326 
327 /** If you use the default do_par_for, you can still set a custom
328  * handler to perform each individual task. Returns the old handler. */
329 //@{
330 typedef int (*halide_do_task_t)(void *, halide_task_t, int, uint8_t *);
332 extern int halide_do_task(void *user_context, halide_task_t f, int idx,
333  uint8_t *closure);
334 //@}
335 
336 /** The version of do_task called for loop tasks. By default calls the
337  * loop task with the same arguments. */
338 // @{
339 typedef int (*halide_do_loop_task_t)(void *, halide_loop_task_t, int, int, uint8_t *, void *);
341 extern int halide_do_loop_task(void *user_context, halide_loop_task_t f, int min, int extent,
342  uint8_t *closure, void *task_parent);
343 //@}
344 
345 /** Provide an entire custom tasking runtime via function
346  * pointers. Note that do_task and semaphore_try_acquire are only ever
347  * called by halide_default_do_par_for and
348  * halide_default_do_parallel_tasks, so it's only necessary to provide
349  * those if you are mixing in the default implementations of
350  * do_par_for and do_parallel_tasks. */
351 // @{
352 typedef int (*halide_do_parallel_tasks_t)(void *, int, struct halide_parallel_task_t *,
353  void *task_parent);
362 // @}
363 
364 /** The default versions of the parallel runtime functions. */
365 // @{
366 extern int halide_default_do_par_for(void *user_context,
367  halide_task_t task,
368  int min, int size, uint8_t *closure);
370  int num_tasks,
371  struct halide_parallel_task_t *tasks,
372  void *task_parent);
373 extern int halide_default_do_task(void *user_context, halide_task_t f, int idx,
374  uint8_t *closure);
376  int min, int extent,
377  uint8_t *closure, void *task_parent);
378 extern int halide_default_semaphore_init(struct halide_semaphore_t *, int n);
379 extern int halide_default_semaphore_release(struct halide_semaphore_t *, int n);
380 extern bool halide_default_semaphore_try_acquire(struct halide_semaphore_t *, int n);
381 // @}
382 
383 struct halide_thread;
384 
385 /** Spawn a thread. Returns a handle to the thread for the purposes of
386  * joining it. The thread must be joined in order to clean up any
387  * resources associated with it. */
388 extern struct halide_thread *halide_spawn_thread(void (*f)(void *), void *closure);
389 
390 /** Join a thread. */
391 extern void halide_join_thread(struct halide_thread *);
392 
393 /** Get or set the number of threads used by Halide's thread pool. Set returns
394  * the old number.
395  *
396  * n < 0 : error condition
397  * n == 0 : use a reasonable system default (typically, number of cpus online).
398  * n == 1 : use exactly one thread; this will always enforce serial execution
399  * n > 1 : use a pool of exactly n threads.
400  *
401  * (Note that this is only guaranteed when using the default implementations
402  * of halide_do_par_for(); custom implementations may completely ignore values
403  * passed to halide_set_num_threads().)
404  */
405 // @{
406 extern int halide_get_num_threads();
407 extern int halide_set_num_threads(int n);
408 // @}
409 
410 /** Halide calls these functions to allocate and free memory. To
411  * replace in AOT code, use the halide_set_custom_malloc and
412  * halide_set_custom_free, or (on platforms that support weak
413  * linking), simply define these functions yourself. In JIT-compiled
414  * code use Func::set_custom_allocator.
415  *
416  * If you override them, and find yourself wanting to call the default
417  * implementation from within your override, use
418  * halide_default_malloc/free.
419  *
420  * Note that halide_malloc must return a pointer aligned to the
421  * maximum meaningful alignment for the platform for the purpose of
422  * vector loads and stores, *and* with an allocated size that is (at least)
423  * an integral multiple of that same alignment. The default implementation
424  * uses 32-byte alignment on arm and 64-byte alignment on x86. Additionally,
425  * it must be safe to read at least 8 bytes before the start and beyond the end.
426  */
427 //@{
428 extern void *halide_malloc(void *user_context, size_t x);
429 extern void halide_free(void *user_context, void *ptr);
430 extern void *halide_default_malloc(void *user_context, size_t x);
431 extern void halide_default_free(void *user_context, void *ptr);
432 typedef void *(*halide_malloc_t)(void *, size_t);
433 typedef void (*halide_free_t)(void *, void *);
436 //@}
437 
438 /** Halide calls these functions to interact with the underlying
439  * system runtime functions. To replace in AOT code on platforms that
440  * support weak linking, define these functions yourself, or use
441  * the halide_set_custom_load_library() and halide_set_custom_get_library_symbol()
442  * functions. In JIT-compiled code, use JITSharedRuntime::set_default_handlers().
443  *
444  * halide_load_library and halide_get_library_symbol are equivalent to
445  * dlopen and dlsym. halide_get_symbol(sym) is equivalent to
446  * dlsym(RTLD_DEFAULT, sym).
447  */
448 //@{
449 extern void *halide_get_symbol(const char *name);
450 extern void *halide_load_library(const char *name);
451 extern void *halide_get_library_symbol(void *lib, const char *name);
452 extern void *halide_default_get_symbol(const char *name);
453 extern void *halide_default_load_library(const char *name);
454 extern void *halide_default_get_library_symbol(void *lib, const char *name);
455 typedef void *(*halide_get_symbol_t)(const char *name);
456 typedef void *(*halide_load_library_t)(const char *name);
457 typedef void *(*halide_get_library_symbol_t)(void *lib, const char *name);
461 //@}
462 
463 /** Called when debug_to_file is used inside %Halide code. See
464  * Func::debug_to_file for how this is called
465  *
466  * Cannot be replaced in JITted code at present.
467  */
468 extern int32_t halide_debug_to_file(void *user_context, const char *filename,
469  struct halide_buffer_t *buf);
470 
471 /** Types in the halide type system. They can be ints, unsigned ints,
472  * or floats (of various bit-widths), or a handle (which is always 64-bits).
473  * Note that the int/uint/float values do not imply a specific bit width
474  * (the bit width is expected to be encoded in a separate value).
475  */
476 typedef enum halide_type_code_t
477 #if (__cplusplus >= 201103L || _MSVC_LANG >= 201103L)
478  : uint8_t
479 #endif
480 {
481  halide_type_int = 0, ///< signed integers
482  halide_type_uint = 1, ///< unsigned integers
483  halide_type_float = 2, ///< IEEE floating point numbers
484  halide_type_handle = 3, ///< opaque pointer type (void *)
485  halide_type_bfloat = 4, ///< floating point numbers in the bfloat format
487 
488 // Note that while __attribute__ can go before or after the declaration,
489 // __declspec apparently is only allowed before.
490 #ifndef HALIDE_ATTRIBUTE_ALIGN
491 #ifdef _MSC_VER
492 #define HALIDE_ATTRIBUTE_ALIGN(x) __declspec(align(x))
493 #else
494 #define HALIDE_ATTRIBUTE_ALIGN(x) __attribute__((aligned(x)))
495 #endif
496 #endif
497 
498 /** A runtime tag for a type in the halide type system. Can be ints,
499  * unsigned ints, or floats of various bit-widths (the 'bits'
500  * field). Can also be vectors of the same (by setting the 'lanes'
501  * field to something larger than one). This struct should be
502  * exactly 32-bits in size. */
504  /** The basic type code: signed integer, unsigned integer, or floating point. */
505 #if (__cplusplus >= 201103L || _MSVC_LANG >= 201103L)
507  halide_type_code_t code; // halide_type_code_t
508 #else
510  uint8_t code; // halide_type_code_t
511 #endif
512 
513  /** The number of bits of precision of a single scalar value of this type. */
516 
517  /** How many elements in a vector. This is 1 for scalar types. */
520 
521 #if (__cplusplus >= 201103L || _MSVC_LANG >= 201103L)
522  /** Construct a runtime representation of a Halide type from:
523  * code: The fundamental type from an enum.
524  * bits: The bit size of one element.
525  * lanes: The number of vector elements in the type. */
527  : code(code), bits(bits), lanes(lanes) {
528  }
529 
530  /** Default constructor is required e.g. to declare halide_trace_event
531  * instances. */
533  : code((halide_type_code_t)0), bits(0), lanes(0) {
534  }
535 
536  HALIDE_ALWAYS_INLINE constexpr halide_type_t with_lanes(uint16_t new_lanes) const {
537  return halide_type_t((halide_type_code_t)code, bits, new_lanes);
538  }
539 
540  HALIDE_ALWAYS_INLINE constexpr halide_type_t element_of() const {
541  return with_lanes(1);
542  }
543  /** Compare two types for equality. */
544  HALIDE_ALWAYS_INLINE constexpr bool operator==(const halide_type_t &other) const {
545  return as_u32() == other.as_u32();
546  }
547 
548  HALIDE_ALWAYS_INLINE constexpr bool operator!=(const halide_type_t &other) const {
549  return !(*this == other);
550  }
551 
552  HALIDE_ALWAYS_INLINE constexpr bool operator<(const halide_type_t &other) const {
553  return as_u32() < other.as_u32();
554  }
555 
556  /** Size in bytes for a single element, even if width is not 1, of this type. */
557  HALIDE_ALWAYS_INLINE constexpr int bytes() const {
558  return (bits + 7) / 8;
559  }
560 
561  HALIDE_ALWAYS_INLINE constexpr uint32_t as_u32() const {
562  // Note that this produces a result that is identical to memcpy'ing 'this'
563  // into a u32 (on a little-endian machine, anyway), and at -O1 or greater
564  // on Clang, the compiler knows this and optimizes this into a single 32-bit move.
565  // (At -O0 it will look awful.)
566  return static_cast<uint8_t>(code) |
567  (static_cast<uint16_t>(bits) << 8) |
568  (static_cast<uint32_t>(lanes) << 16);
569  }
570 #endif
571 };
572 
573 #if (__cplusplus >= 201103L || _MSVC_LANG >= 201103L)
574 static_assert(sizeof(halide_type_t) == sizeof(uint32_t), "size mismatch in halide_type_t");
575 #endif
576 
588 
590  /** The name of the Func or Pipeline that this event refers to */
591  const char *func;
592 
593  /** If the event type is a load or a store, this points to the
594  * value being loaded or stored. Use the type field to safely cast
595  * this to a concrete pointer type and retrieve it. For other
596  * events this is null. */
597  void *value;
598 
599  /** For loads and stores, an array which contains the location
600  * being accessed. For vector loads or stores it is an array of
601  * vectors of coordinates (the vector dimension is innermost).
602  *
603  * For realization or production-related events, this will contain
604  * the mins and extents of the region being accessed, in the order
605  * min0, extent0, min1, extent1, ...
606  *
607  * For pipeline-related events, this will be null.
608  */
610 
611  /** For halide_trace_tag, this points to a read-only null-terminated string
612  * of arbitrary text. For all other events, this will be null.
613  */
614  const char *trace_tag;
615 
616  /** If the event type is a load or a store, this is the type of
617  * the data. Otherwise, the value is meaningless. */
618  struct halide_type_t type;
619 
620  /** The type of event */
622 
623  /* The ID of the parent event (see below for an explanation of
624  * event ancestry). */
626 
627  /** If this was a load or store of a Tuple-valued Func, this is
628  * which tuple element was accessed. */
630 
631  /** The length of the coordinates array */
633 };
634 
635 /** Called when Funcs are marked as trace_load, trace_store, or
636  * trace_realization. See Func::set_custom_trace. The default
637  * implementation either prints events via halide_print, or if
638  * HL_TRACE_FILE is defined, dumps the trace to that file in a
639  * sequence of trace packets. The header for a trace packet is defined
640  * below. If the trace is going to be large, you may want to make the
641  * file a named pipe, and then read from that pipe into gzip.
642  *
643  * halide_trace returns a unique ID which will be passed to future
644  * events that "belong" to the earlier event as the parent id. The
645  * ownership hierarchy looks like:
646  *
647  * begin_pipeline
648  * +--trace_tag (if any)
649  * +--trace_tag (if any)
650  * ...
651  * +--begin_realization
652  * | +--produce
653  * | | +--load/store
654  * | | +--end_produce
655  * | +--consume
656  * | | +--load
657  * | | +--end_consume
658  * | +--end_realization
659  * +--end_pipeline
660  *
661  * Threading means that ownership cannot be inferred from the ordering
662  * of events. There can be many active realizations of a given
663  * function, or many active productions for a single
664  * realization. Within a single production, the ordering of events is
665  * meaningful.
666  *
667  * Note that all trace_tag events (if any) will occur just after the begin_pipeline
668  * event, but before any begin_realization events. All trace_tags for a given Func
669  * will be emitted in the order added.
670  */
671 // @}
672 extern int32_t halide_trace(void *user_context, const struct halide_trace_event_t *event);
674 typedef int32_t (*halide_trace_t)(void *user_context, const struct halide_trace_event_t *);
676 // @}
677 
678 /** The header of a packet in a binary trace. All fields are 32-bit. */
680  /** The total size of this packet in bytes. Always a multiple of
681  * four. Equivalently, the number of bytes until the next
682  * packet. */
684 
685  /** The id of this packet (for the purpose of parent_id). */
687 
688  /** The remaining fields are equivalent to those in halide_trace_event_t */
689  // @{
690  struct halide_type_t type;
695  // @}
696 
697 #if (__cplusplus >= 201103L || _MSVC_LANG >= 201103L)
698  /** Get the coordinates array, assuming this packet is laid out in
699  * memory as it was written. The coordinates array comes
700  * immediately after the packet header. */
701  HALIDE_ALWAYS_INLINE const int *coordinates() const {
702  return (const int *)(this + 1);
703  }
704 
705  HALIDE_ALWAYS_INLINE int *coordinates() {
706  return (int *)(this + 1);
707  }
708 
709  /** Get the value, assuming this packet is laid out in memory as
710  * it was written. The packet comes immediately after the coordinates
711  * array. */
712  HALIDE_ALWAYS_INLINE const void *value() const {
713  return (const void *)(coordinates() + dimensions);
714  }
715 
716  HALIDE_ALWAYS_INLINE void *value() {
717  return (void *)(coordinates() + dimensions);
718  }
719 
720  /** Get the func name, assuming this packet is laid out in memory
721  * as it was written. It comes after the value. */
722  HALIDE_ALWAYS_INLINE const char *func() const {
723  return (const char *)value() + type.lanes * type.bytes();
724  }
725 
726  HALIDE_ALWAYS_INLINE char *func() {
727  return (char *)value() + type.lanes * type.bytes();
728  }
729 
730  /** Get the trace_tag (if any), assuming this packet is laid out in memory
731  * as it was written. It comes after the func name. If there is no trace_tag,
732  * this will return a pointer to an empty string. */
733  HALIDE_ALWAYS_INLINE const char *trace_tag() const {
734  const char *f = func();
735  // strlen may not be available here
736  while (*f++) {
737  // nothing
738  }
739  return f;
740  }
741 
742  HALIDE_ALWAYS_INLINE char *trace_tag() {
743  char *f = func();
744  // strlen may not be available here
745  while (*f++) {
746  // nothing
747  }
748  return f;
749  }
750 #endif
751 };
752 
753 /** Set the file descriptor that Halide should write binary trace
754  * events to. If called with 0 as the argument, Halide outputs trace
755  * information to stdout in a human-readable format. If never called,
756  * Halide checks the for existence of an environment variable called
757  * HL_TRACE_FILE and opens that file. If HL_TRACE_FILE is not defined,
758  * it outputs trace information to stdout in a human-readable
759  * format. */
760 extern void halide_set_trace_file(int fd);
761 
762 /** Halide calls this to retrieve the file descriptor to write binary
763  * trace events to. The default implementation returns the value set
764  * by halide_set_trace_file. Implement it yourself if you wish to use
765  * a custom file descriptor per user_context. Return zero from your
766  * implementation to tell Halide to print human-readable trace
767  * information to stdout. */
769 
770 /** If tracing is writing to a file. This call closes that file
771  * (flushing the trace). Returns zero on success. */
772 extern int halide_shutdown_trace(void);
773 
774 /** All Halide GPU or device backend implementations provide an
775  * interface to be used with halide_device_malloc, etc. This is
776  * accessed via the functions below.
777  */
778 
779 /** An opaque struct containing per-GPU API implementations of the
780  * device functions. */
782 
783 /** Each GPU API provides a halide_device_interface_t struct pointing
784  * to the code that manages device allocations. You can access these
785  * functions directly from the struct member function pointers, or by
786  * calling the functions declared below. Note that the global
787  * functions are not available when using Halide as a JIT compiler.
788  * If you are using raw halide_buffer_t in that context you must use
789  * the function pointers in the device_interface struct.
790  *
791  * The function pointers below are currently the same for every GPU
792  * API; only the impl field varies. These top-level functions do the
793  * bookkeeping that is common across all GPU APIs, and then dispatch
794  * to more API-specific functions via another set of function pointers
795  * hidden inside the impl field.
796  */
798  int (*device_malloc)(void *user_context, struct halide_buffer_t *buf,
799  const struct halide_device_interface_t *device_interface);
800  int (*device_free)(void *user_context, struct halide_buffer_t *buf);
801  int (*device_sync)(void *user_context, struct halide_buffer_t *buf);
803  const struct halide_device_interface_t *device_interface);
804  int (*copy_to_host)(void *user_context, struct halide_buffer_t *buf);
805  int (*copy_to_device)(void *user_context, struct halide_buffer_t *buf,
806  const struct halide_device_interface_t *device_interface);
808  const struct halide_device_interface_t *device_interface);
810  int (*buffer_copy)(void *user_context, struct halide_buffer_t *src,
811  const struct halide_device_interface_t *dst_device_interface, struct halide_buffer_t *dst);
812  int (*device_crop)(void *user_context, const struct halide_buffer_t *src,
813  struct halide_buffer_t *dst);
814  int (*device_slice)(void *user_context, const struct halide_buffer_t *src,
815  int slice_dim, int slice_pos, struct halide_buffer_t *dst);
817  int (*wrap_native)(void *user_context, struct halide_buffer_t *buf, uint64_t handle,
818  const struct halide_device_interface_t *device_interface);
819  int (*detach_native)(void *user_context, struct halide_buffer_t *buf);
820  int (*compute_capability)(void *user_context, int *major, int *minor);
822 };
823 
824 /** Release all data associated with the given device interface, in
825  * particular all resources (memory, texture, context handles)
826  * allocated by Halide. Must be called explicitly when using AOT
827  * compilation. This is *not* thread-safe with respect to actively
828  * running Halide code. Ensure all pipelines are finished before
829  * calling this. */
831  const struct halide_device_interface_t *device_interface);
832 
833 /** Copy image data from device memory to host memory. This must be called
834  * explicitly to copy back the results of a GPU-based filter. */
835 extern int halide_copy_to_host(void *user_context, struct halide_buffer_t *buf);
836 
837 /** Copy image data from host memory to device memory. This should not
838  * be called directly; Halide handles copying to the device
839  * automatically. If interface is NULL and the buf has a non-zero dev
840  * field, the device associated with the dev handle will be
841  * used. Otherwise if the dev field is 0 and interface is NULL, an
842  * error is returned. */
843 extern int halide_copy_to_device(void *user_context, struct halide_buffer_t *buf,
844  const struct halide_device_interface_t *device_interface);
845 
846 /** Copy data from one buffer to another. The buffers may have
847  * different shapes and sizes, but the destination buffer's shape must
848  * be contained within the source buffer's shape. That is, for each
849  * dimension, the min on the destination buffer must be greater than
850  * or equal to the min on the source buffer, and min+extent on the
851  * destination buffer must be less that or equal to min+extent on the
852  * source buffer. The source data is pulled from either device or
853  * host memory on the source, depending on the dirty flags. host is
854  * preferred if both are valid. The dst_device_interface parameter
855  * controls the destination memory space. NULL means host memory. */
856 extern int halide_buffer_copy(void *user_context, struct halide_buffer_t *src,
857  const struct halide_device_interface_t *dst_device_interface,
858  struct halide_buffer_t *dst);
859 
860 /** Give the destination buffer a device allocation which is an alias
861  * for the same coordinate range in the source buffer. Modifies the
862  * device, device_interface, and the device_dirty flag only. Only
863  * supported by some device APIs (others will return
864  * halide_error_code_device_crop_unsupported). Call
865  * halide_device_release_crop instead of halide_device_free to clean
866  * up resources associated with the cropped view. Do not free the
867  * device allocation on the source buffer while the destination buffer
868  * still lives. Note that the two buffers do not share dirty flags, so
869  * care must be taken to update them together as needed. Note that src
870  * and dst are required to have the same number of dimensions.
871  *
872  * Note also that (in theory) device interfaces which support cropping may
873  * still not support cropping a crop (instead, create a new crop of the parent
874  * buffer); in practice, no known implementation has this limitation, although
875  * it is possible that some future implementations may require it. */
877  const struct halide_buffer_t *src,
878  struct halide_buffer_t *dst);
879 
880 /** Give the destination buffer a device allocation which is an alias
881  * for a similar coordinate range in the source buffer, but with one dimension
882  * sliced away in the dst. Modifies the device, device_interface, and the
883  * device_dirty flag only. Only supported by some device APIs (others will return
884  * halide_error_code_device_crop_unsupported). Call
885  * halide_device_release_crop instead of halide_device_free to clean
886  * up resources associated with the sliced view. Do not free the
887  * device allocation on the source buffer while the destination buffer
888  * still lives. Note that the two buffers do not share dirty flags, so
889  * care must be taken to update them together as needed. Note that the dst buffer
890  * must have exactly one fewer dimension than the src buffer, and that slice_dim
891  * and slice_pos must be valid within src. */
893  const struct halide_buffer_t *src,
894  int slice_dim, int slice_pos,
895  struct halide_buffer_t *dst);
896 
897 /** Release any resources associated with a cropped/sliced view of another
898  * buffer. */
900  struct halide_buffer_t *buf);
901 
902 /** Wait for current GPU operations to complete. Calling this explicitly
903  * should rarely be necessary, except maybe for profiling. */
904 extern int halide_device_sync(void *user_context, struct halide_buffer_t *buf);
905 
906 /**
907  * Wait for current GPU operations to complete. Calling this explicitly
908  * should rarely be necessary, except maybe for profiling.
909  * This variation of the synchronizing is useful when a synchronization is desirable
910  * without specifying any buffer to synchronize on.
911  * Calling this with a null device_interface is always illegal.
912  */
914  const struct halide_device_interface_t *device_interface);
915 
916 /** Allocate device memory to back a halide_buffer_t. */
917 extern int halide_device_malloc(void *user_context, struct halide_buffer_t *buf,
918  const struct halide_device_interface_t *device_interface);
919 
920 /** Free device memory. */
921 extern int halide_device_free(void *user_context, struct halide_buffer_t *buf);
922 
923 /** Wrap or detach a native device handle, setting the device field
924  * and device_interface field as appropriate for the given GPU
925  * API. The meaning of the opaque handle is specific to the device
926  * interface, so if you know the device interface in use, call the
927  * more specific functions in the runtime headers for your specific
928  * device API instead (e.g. HalideRuntimeCuda.h). */
929 // @{
931  struct halide_buffer_t *buf,
932  uint64_t handle,
933  const struct halide_device_interface_t *device_interface);
935 // @}
936 
937 /** Selects which gpu device to use. 0 is usually the display
938  * device. If never called, Halide uses the environment variable
939  * HL_GPU_DEVICE. If that variable is unset, Halide uses the last
940  * device. Set this to -1 to use the last device. */
941 extern void halide_set_gpu_device(int n);
942 
943 /** Halide calls this to get the desired halide gpu device
944  * setting. Implement this yourself to use a different gpu device per
945  * user_context. The default implementation returns the value set by
946  * halide_set_gpu_device, or the environment variable
947  * HL_GPU_DEVICE. */
948 extern int halide_get_gpu_device(void *user_context);
949 
950 /** Set the soft maximum amount of memory, in bytes, that the LRU
951  * cache will use to memoize Func results. This is not a strict
952  * maximum in that concurrency and simultaneous use of memoized
953  * reults larger than the cache size can both cause it to
954  * temporariliy be larger than the size specified here.
955  */
957 
958 /** Given a cache key for a memoized result, currently constructed
959  * from the Func name and top-level Func name plus the arguments of
960  * the computation, determine if the result is in the cache and
961  * return it if so. (The internals of the cache key should be
962  * considered opaque by this function.) If this routine returns true,
963  * it is a cache miss. Otherwise, it will return false and the
964  * buffers passed in will be filled, via copying, with memoized
965  * data. The last argument is a list if halide_buffer_t pointers which
966  * represents the outputs of the memoized Func. If the Func does not
967  * return a Tuple, there will only be one halide_buffer_t in the list. The
968  * tuple_count parameters determines the length of the list.
969  *
970  * The return values are:
971  * -1: Signals an error.
972  * 0: Success and cache hit.
973  * 1: Success and cache miss.
974  */
975 extern int halide_memoization_cache_lookup(void *user_context, const uint8_t *cache_key, int32_t size,
976  struct halide_buffer_t *realized_bounds,
977  int32_t tuple_count, struct halide_buffer_t **tuple_buffers);
978 
979 /** Given a cache key for a memoized result, currently constructed
980  * from the Func name and top-level Func name plus the arguments of
981  * the computation, store the result in the cache for futre access by
982  * halide_memoization_cache_lookup. (The internals of the cache key
983  * should be considered opaque by this function.) Data is copied out
984  * from the inputs and inputs are unmodified. The last argument is a
985  * list if halide_buffer_t pointers which represents the outputs of the
986  * memoized Func. If the Func does not return a Tuple, there will
987  * only be one halide_buffer_t in the list. The tuple_count parameters
988  * determines the length of the list.
989  *
990  * If there is a memory allocation failure, the store does not store
991  * the data into the cache.
992  *
993  * If has_eviction_key is true, the entry is marked with eviction_key to
994  * allow removing the key with halide_memoization_cache_evict.
995  */
996 extern int halide_memoization_cache_store(void *user_context, const uint8_t *cache_key, int32_t size,
997  struct halide_buffer_t *realized_bounds,
998  int32_t tuple_count,
999  struct halide_buffer_t **tuple_buffers,
1000  bool has_eviction_key, uint64_t eviction_key);
1001 
1002 /** Evict all cache entries that were tagged with the given
1003  * eviction_key in the memoize scheduling directive.
1004  */
1005 extern void halide_memoization_cache_evict(void *user_context, uint64_t eviction_key);
1006 
1007 /** If halide_memoization_cache_lookup succeeds,
1008  * halide_memoization_cache_release must be called to signal the
1009  * storage is no longer being used by the caller. It will be passed
1010  * the host pointer of one the buffers returned by
1011  * halide_memoization_cache_lookup. That is
1012  * halide_memoization_cache_release will be called multiple times for
1013  * the case where halide_memoization_cache_lookup is handling multiple
1014  * buffers. (This corresponds to memoizing a Tuple in Halide.) Note
1015  * that the host pointer must be sufficient to get to all information
1016  * the release operation needs. The default Halide cache impleemntation
1017  * accomplishes this by storing extra data before the start of the user
1018  * modifiable host storage.
1019  *
1020  * This call is like free and does not have a failure return.
1021  */
1022 extern void halide_memoization_cache_release(void *user_context, void *host);
1023 
1024 /** Free all memory and resources associated with the memoization cache.
1025  * Must be called at a time when no other threads are accessing the cache.
1026  */
1028 
1029 /** Verify that a given range of memory has been initialized; only used when Target::MSAN is enabled.
1030  *
1031  * The default implementation simply calls the LLVM-provided __msan_check_mem_is_initialized() function.
1032  *
1033  * The return value should always be zero.
1034  */
1035 extern int halide_msan_check_memory_is_initialized(void *user_context, const void *ptr, uint64_t len, const char *name);
1036 
1037 /** Verify that the data pointed to by the halide_buffer_t is initialized (but *not* the halide_buffer_t itself),
1038  * using halide_msan_check_memory_is_initialized() for checking.
1039  *
1040  * The default implementation takes pains to only check the active memory ranges
1041  * (skipping padding), and sorting into ranges to always check the smallest number of
1042  * ranges, in monotonically increasing memory order.
1043  *
1044  * Most client code should never need to replace the default implementation.
1045  *
1046  * The return value should always be zero.
1047  */
1048 extern int halide_msan_check_buffer_is_initialized(void *user_context, struct halide_buffer_t *buffer, const char *buf_name);
1049 
1050 /** Annotate that a given range of memory has been initialized;
1051  * only used when Target::MSAN is enabled.
1052  *
1053  * The default implementation simply calls the LLVM-provided __msan_unpoison() function.
1054  *
1055  * The return value should always be zero.
1056  */
1057 extern int halide_msan_annotate_memory_is_initialized(void *user_context, const void *ptr, uint64_t len);
1058 
1059 /** Mark the data pointed to by the halide_buffer_t as initialized (but *not* the halide_buffer_t itself),
1060  * using halide_msan_annotate_memory_is_initialized() for marking.
1061  *
1062  * The default implementation takes pains to only mark the active memory ranges
1063  * (skipping padding), and sorting into ranges to always mark the smallest number of
1064  * ranges, in monotonically increasing memory order.
1065  *
1066  * Most client code should never need to replace the default implementation.
1067  *
1068  * The return value should always be zero.
1069  */
1072 
1073 /** The error codes that may be returned by a Halide pipeline. */
1075  /** There was no error. This is the value returned by Halide on success. */
1077 
1078  /** An uncategorized error occurred. Refer to the string passed to halide_error. */
1080 
1081  /** A Func was given an explicit bound via Func::bound, but this
1082  * was not large enough to encompass the region that is used of
1083  * the Func by the rest of the pipeline. */
1085 
1086  /** The elem_size field of a halide_buffer_t does not match the size in
1087  * bytes of the type of that ImageParam. Probable type mismatch. */
1089 
1090  /** A pipeline would access memory outside of the halide_buffer_t passed
1091  * in. */
1093 
1094  /** A halide_buffer_t was given that spans more than 2GB of memory. */
1096 
1097  /** A halide_buffer_t was given with extents that multiply to a number
1098  * greater than 2^31-1 */
1100 
1101  /** Applying explicit constraints on the size of an input or
1102  * output buffer shrank the size of that buffer below what will be
1103  * accessed by the pipeline. */
1105 
1106  /** A constraint on a size or stride of an input or output buffer
1107  * was not met by the halide_buffer_t passed in. */
1109 
1110  /** A scalar parameter passed in was smaller than its minimum
1111  * declared value. */
1113 
1114  /** A scalar parameter passed in was greater than its minimum
1115  * declared value. */
1117 
1118  /** A call to halide_malloc returned NULL. */
1120 
1121  /** A halide_buffer_t pointer passed in was NULL. */
1123 
1124  /** debug_to_file failed to open or write to the specified
1125  * file. */
1127 
1128  /** The Halide runtime encountered an error while trying to copy
1129  * from device to host. Turn on -debug in your target string to
1130  * see more details. */
1132 
1133  /** The Halide runtime encountered an error while trying to copy
1134  * from host to device. Turn on -debug in your target string to
1135  * see more details. */
1137 
1138  /** The Halide runtime encountered an error while trying to
1139  * allocate memory on device. Turn on -debug in your target string
1140  * to see more details. */
1142 
1143  /** The Halide runtime encountered an error while trying to
1144  * synchronize with a device. Turn on -debug in your target string
1145  * to see more details. */
1147 
1148  /** The Halide runtime encountered an error while trying to free a
1149  * device allocation. Turn on -debug in your target string to see
1150  * more details. */
1152 
1153  /** Buffer has a non-zero device but no device interface, which
1154  * violates a Halide invariant. */
1156 
1157  /** This part of the Halide runtime is unimplemented on this platform. */
1159 
1160  /** A runtime symbol could not be loaded. */
1162 
1163  /** There is a bug in the Halide compiler. */
1165 
1166  /** The Halide runtime encountered an error while trying to launch
1167  * a GPU kernel. Turn on -debug in your target string to see more
1168  * details. */
1170 
1171  /** The Halide runtime encountered a host pointer that violated
1172  * the alignment set for it by way of a call to
1173  * set_host_alignment */
1175 
1176  /** A fold_storage directive was used on a dimension that is not
1177  * accessed in a monotonically increasing or decreasing fashion. */
1179 
1180  /** A fold_storage directive was used with a fold factor that was
1181  * too small to store all the values of a producer needed by the
1182  * consumer. */
1184 
1185  /** User-specified require() expression was not satisfied. */
1187 
1188  /** At least one of the buffer's extents are negative. */
1190 
1191  /** Call(s) to a GPU backend API failed. */
1193 
1194  /** Failure recording trace packets for one of the halide_target_feature_trace features. */
1196 
1197  /** A specialize_fail() schedule branch was selected at runtime. */
1199 
1200  /** The Halide runtime encountered an error while trying to wrap a
1201  * native device handle. Turn on -debug in your target string to
1202  * see more details. */
1204 
1205  /** The Halide runtime encountered an error while trying to detach
1206  * a native device handle. Turn on -debug in your target string
1207  * to see more details. */
1209 
1210  /** The host field on an input or output was null, the device
1211  * field was not zero, and the pipeline tries to use the buffer on
1212  * the host. You may be passing a GPU-only buffer to a pipeline
1213  * which is scheduled to use it on the CPU. */
1215 
1216  /** A folded buffer was passed to an extern stage, but the region
1217  * touched wraps around the fold boundary. */
1219 
1220  /** Buffer has a non-null device_interface but device is 0, which
1221  * violates a Halide invariant. */
1223 
1224  /** Buffer has both host and device dirty bits set, which violates
1225  * a Halide invariant. */
1227 
1228  /** The halide_buffer_t * passed to a halide runtime routine is
1229  * nullptr and this is not allowed. */
1231 
1232  /** The Halide runtime encountered an error while trying to copy
1233  * from one buffer to another. Turn on -debug in your target
1234  * string to see more details. */
1236 
1237  /** Attempted to make cropped/sliced alias of a buffer with a device
1238  * field, but the device_interface does not support cropping. */
1240 
1241  /** Cropping/slicing a buffer failed for some other reason. Turn on -debug
1242  * in your target string. */
1244 
1245  /** An operation on a buffer required an allocation on a
1246  * particular device interface, but a device allocation already
1247  * existed on a different device interface. Free the old one
1248  * first. */
1250 
1251  /** The dimensions field of a halide_buffer_t does not match the dimensions of that ImageParam. */
1253 
1254  /** A buffer with the device_dirty flag set was passed to a
1255  * pipeline compiled with no device backends enabled, so it
1256  * doesn't know how to copy the data back from device memory to
1257  * host memory. Either call copy_to_host before calling the Halide
1258  * pipeline, or enable the appropriate device backend. */
1260 
1261  /** An explicit storage bound provided is too small to store
1262  * all the values produced by the function. */
1264 
1265  /** A factor used to split a loop was discovered to be zero or negative at
1266  * runtime. */
1268 
1269  /** "vscale" value of Scalable Vector detected in runtime does not match
1270  * the vscale value used in compilation. */
1272 
1273  /** Profiling failed for a pipeline invocation. */
1275 };
1276 
1277 /** Halide calls the functions below on various error conditions. The
1278  * default implementations construct an error message, call
1279  * halide_error, then return the matching error code above. On
1280  * platforms that support weak linking, you can override these to
1281  * catch the errors individually. */
1282 
1283 /** A call into an extern stage for the purposes of bounds inference
1284  * failed. Returns the error code given by the extern stage. */
1285 extern int halide_error_bounds_inference_call_failed(void *user_context, const char *extern_stage_name, int result);
1286 
1287 /** A call to an extern stage failed. Returned the error code given by
1288  * the extern stage. */
1289 extern int halide_error_extern_stage_failed(void *user_context, const char *extern_stage_name, int result);
1290 
1291 /** Various other error conditions. See the enum above for a
1292  * description of each. */
1293 // @{
1294 extern int halide_error_explicit_bounds_too_small(void *user_context, const char *func_name, const char *var_name,
1295  int min_bound, int max_bound, int min_required, int max_required);
1296 extern int halide_error_bad_type(void *user_context, const char *func_name,
1297  uint32_t type_given, uint32_t correct_type); // N.B. The last two args are the bit representation of a halide_type_t
1298 extern int halide_error_bad_dimensions(void *user_context, const char *func_name,
1299  int32_t dimensions_given, int32_t correct_dimensions);
1300 extern int halide_error_access_out_of_bounds(void *user_context, const char *func_name,
1301  int dimension, int min_touched, int max_touched,
1302  int min_valid, int max_valid);
1303 extern int halide_error_buffer_allocation_too_large(void *user_context, const char *buffer_name,
1304  uint64_t allocation_size, uint64_t max_size);
1305 extern int halide_error_buffer_extents_negative(void *user_context, const char *buffer_name, int dimension, int extent);
1306 extern int halide_error_buffer_extents_too_large(void *user_context, const char *buffer_name,
1307  int64_t actual_size, int64_t max_size);
1309  int dimension,
1310  int constrained_min, int constrained_extent,
1311  int required_min, int required_extent);
1312 extern int halide_error_constraint_violated(void *user_context, const char *var, int val,
1313  const char *constrained_var, int constrained_val);
1314 extern int halide_error_param_too_small_i64(void *user_context, const char *param_name,
1315  int64_t val, int64_t min_val);
1316 extern int halide_error_param_too_small_u64(void *user_context, const char *param_name,
1317  uint64_t val, uint64_t min_val);
1318 extern int halide_error_param_too_small_f64(void *user_context, const char *param_name,
1319  double val, double min_val);
1320 extern int halide_error_param_too_large_i64(void *user_context, const char *param_name,
1321  int64_t val, int64_t max_val);
1322 extern int halide_error_param_too_large_u64(void *user_context, const char *param_name,
1323  uint64_t val, uint64_t max_val);
1324 extern int halide_error_param_too_large_f64(void *user_context, const char *param_name,
1325  double val, double max_val);
1327 extern int halide_error_buffer_argument_is_null(void *user_context, const char *buffer_name);
1328 extern int halide_error_debug_to_file_failed(void *user_context, const char *func,
1329  const char *filename, int error_code);
1330 extern int halide_error_unaligned_host_ptr(void *user_context, const char *func_name, int alignment);
1331 extern int halide_error_host_is_null(void *user_context, const char *func_name);
1332 extern int halide_error_bad_fold(void *user_context, const char *func_name, const char *var_name,
1333  const char *loop_name);
1334 extern int halide_error_bad_extern_fold(void *user_context, const char *func_name,
1335  int dim, int min, int extent, int valid_min, int fold_factor);
1336 
1337 extern int halide_error_fold_factor_too_small(void *user_context, const char *func_name, const char *var_name,
1338  int fold_factor, const char *loop_name, int required_extent);
1339 extern int halide_error_requirement_failed(void *user_context, const char *condition, const char *message);
1340 extern int halide_error_specialize_fail(void *user_context, const char *message);
1344 extern int halide_error_buffer_is_null(void *user_context, const char *routine);
1345 extern int halide_error_device_dirty_with_no_device_support(void *user_context, const char *buffer_name);
1346 extern int halide_error_storage_bound_too_small(void *user_context, const char *func_name, const char *var_name,
1347  int provided_size, int required_size);
1349 extern int halide_error_split_factor_not_positive(void *user_context, const char *func_name, const char *orig, const char *outer, const char *inner, const char *factor_str, int factor);
1350 extern int halide_error_vscale_invalid(void *user_context, const char *func_name, int runtime_vscale, int compiletime_vscale);
1351 // @}
1352 
1353 /** Optional features a compilation Target can have.
1354  * Be sure to keep this in sync with the Feature enum in Target.h and the implementation of
1355  * get_runtime_compatible_target in Target.cpp if you add a new feature.
1356  */
1358  halide_target_feature_jit = 0, ///< Generate code that will run immediately inside the calling process.
1359  halide_target_feature_debug, ///< Turn on debug info and output for runtime code.
1360  halide_target_feature_no_asserts, ///< Disable all runtime checks, for slightly tighter code.
1361  halide_target_feature_no_bounds_query, ///< Disable the bounds querying functionality.
1362 
1363  halide_target_feature_sse41, ///< Use SSE 4.1 and earlier instructions. Only relevant on x86.
1364  halide_target_feature_avx, ///< Use AVX 1 instructions. Only relevant on x86.
1365  halide_target_feature_avx2, ///< Use AVX 2 instructions. Only relevant on x86.
1366  halide_target_feature_fma, ///< Enable x86 FMA instruction
1367  halide_target_feature_fma4, ///< Enable x86 (AMD) FMA4 instruction set
1368  halide_target_feature_f16c, ///< Enable x86 16-bit float support
1369 
1370  halide_target_feature_armv7s, ///< Generate code for ARMv7s. Only relevant for 32-bit ARM.
1371  halide_target_feature_no_neon, ///< Avoid using NEON instructions. Only relevant for 32-bit ARM.
1372 
1373  halide_target_feature_vsx, ///< Use VSX instructions. Only relevant on POWERPC.
1374  halide_target_feature_power_arch_2_07, ///< Use POWER ISA 2.07 new instructions. Only relevant on POWERPC.
1375 
1376  halide_target_feature_cuda, ///< Enable the CUDA runtime. Defaults to compute capability 2.0 (Fermi)
1377  halide_target_feature_cuda_capability30, ///< Enable CUDA compute capability 3.0 (Kepler)
1378  halide_target_feature_cuda_capability32, ///< Enable CUDA compute capability 3.2 (Tegra K1)
1379  halide_target_feature_cuda_capability35, ///< Enable CUDA compute capability 3.5 (Kepler)
1380  halide_target_feature_cuda_capability50, ///< Enable CUDA compute capability 5.0 (Maxwell)
1381  halide_target_feature_cuda_capability61, ///< Enable CUDA compute capability 6.1 (Pascal)
1382  halide_target_feature_cuda_capability70, ///< Enable CUDA compute capability 7.0 (Volta)
1383  halide_target_feature_cuda_capability75, ///< Enable CUDA compute capability 7.5 (Turing)
1384  halide_target_feature_cuda_capability80, ///< Enable CUDA compute capability 8.0 (Ampere)
1385  halide_target_feature_cuda_capability86, ///< Enable CUDA compute capability 8.6 (Ampere)
1386 
1387  halide_target_feature_opencl, ///< Enable the OpenCL runtime.
1388  halide_target_feature_cl_doubles, ///< Enable double support on OpenCL targets
1389  halide_target_feature_cl_atomic64, ///< Enable 64-bit atomics operations on OpenCL targets
1390 
1391  halide_target_feature_user_context, ///< Generated code takes a user_context pointer as first argument
1392 
1393  halide_target_feature_profile, ///< Launch a sampling profiler alongside the Halide pipeline that monitors and reports the runtime used by each Func
1394  halide_target_feature_no_runtime, ///< Do not include a copy of the Halide runtime in any generated object file or assembly
1395 
1396  halide_target_feature_metal, ///< Enable the (Apple) Metal runtime.
1397 
1398  halide_target_feature_c_plus_plus_mangling, ///< Generate C++ mangled names for result function, et al
1399 
1400  halide_target_feature_large_buffers, ///< Enable 64-bit buffer indexing to support buffers > 2GB. Ignored if bits != 64.
1401 
1402  halide_target_feature_hvx_128, ///< Enable HVX 128 byte mode.
1403  halide_target_feature_hvx_v62, ///< Enable Hexagon v62 architecture.
1404  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.
1405  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.
1406  halide_target_feature_msan, ///< Enable hooks for MSAN support.
1407  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.
1408  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.
1409  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.
1410  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.
1411  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.
1412  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.
1413  halide_target_feature_trace_loads, ///< Trace all loads done by the pipeline. Equivalent to calling Func::trace_loads on every non-inlined Func.
1414  halide_target_feature_trace_stores, ///< Trace all stores done by the pipeline. Equivalent to calling Func::trace_stores on every non-inlined Func.
1415  halide_target_feature_trace_realizations, ///< Trace all realizations done by the pipeline. Equivalent to calling Func::trace_realizations on every non-inlined Func.
1416  halide_target_feature_trace_pipeline, ///< Trace the pipeline.
1417  halide_target_feature_hvx_v65, ///< Enable Hexagon v65 architecture.
1418  halide_target_feature_hvx_v66, ///< Enable Hexagon v66 architecture.
1419  halide_target_feature_hvx_v68, ///< Enable Hexagon v68 architecture.
1420  halide_target_feature_cl_half, ///< Enable half support on OpenCL targets
1421  halide_target_feature_strict_float, ///< Turn off all non-IEEE floating-point optimization. Currently applies only to LLVM targets.
1422  halide_target_feature_tsan, ///< Enable hooks for TSAN support.
1423  halide_target_feature_asan, ///< Enable hooks for ASAN support.
1424  halide_target_feature_d3d12compute, ///< Enable Direct3D 12 Compute runtime.
1425  halide_target_feature_check_unsafe_promises, ///< Insert assertions for promises.
1426  halide_target_feature_hexagon_dma, ///< Enable Hexagon DMA buffers.
1427  halide_target_feature_embed_bitcode, ///< Emulate clang -fembed-bitcode flag.
1428  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.)
1429  halide_target_feature_wasm_mvponly, ///< Disable all extensions to WebAssembly codegen (including +sign-ext and +nontrapping-fptoint, which are on by default).
1430  halide_target_feature_wasm_simd128, ///< Enable +simd128 instructions for WebAssembly codegen.
1431  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.
1432  halide_target_feature_wasm_bulk_memory, ///< Enable +bulk-memory instructions for WebAssembly codegen.
1433  halide_target_feature_webgpu, ///< Enable the WebGPU runtime.
1434  halide_target_feature_sve, ///< Enable ARM Scalable Vector Extensions
1435  halide_target_feature_sve2, ///< Enable ARM Scalable Vector Extensions v2
1436  halide_target_feature_egl, ///< Force use of EGL support.
1437  halide_target_feature_arm_dot_prod, ///< Enable ARMv8.2-a dotprod extension (i.e. udot and sdot instructions)
1438  halide_target_feature_arm_fp16, ///< Enable ARMv8.2-a half-precision floating point data processing
1439  halide_llvm_large_code_model, ///< Use the LLVM large code model to compile
1440  halide_target_feature_rvv, ///< Enable RISCV "V" Vector Extension
1441  halide_target_feature_armv8a, ///< Enable ARMv8a instructions
1442  halide_target_feature_armv81a, ///< Enable ARMv8.1a instructions
1443  halide_target_feature_armv82a, ///< Enable ARMv8.2a instructions
1444  halide_target_feature_armv83a, ///< Enable ARMv8.3a instructions
1445  halide_target_feature_armv84a, ///< Enable ARMv8.4a instructions
1446  halide_target_feature_armv85a, ///< Enable ARMv8.5a instructions
1447  halide_target_feature_armv86a, ///< Enable ARMv8.6a instructions
1448  halide_target_feature_armv87a, ///< Enable ARMv8.7a instructions
1449  halide_target_feature_armv88a, ///< Enable ARMv8.8a instructions
1450  halide_target_feature_armv89a, ///< Enable ARMv8.9a instructions
1451  halide_target_feature_sanitizer_coverage, ///< Enable hooks for SanitizerCoverage support.
1452  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.
1453  halide_target_feature_spirv, ///< Enable SPIR-V code generation support.
1454  halide_target_feature_vulkan, ///< Enable Vulkan runtime support.
1455  halide_target_feature_vulkan_int8, ///< Enable Vulkan 8-bit integer support.
1456  halide_target_feature_vulkan_int16, ///< Enable Vulkan 16-bit integer support.
1457  halide_target_feature_vulkan_int64, ///< Enable Vulkan 64-bit integer support.
1458  halide_target_feature_vulkan_float16, ///< Enable Vulkan 16-bit float support.
1459  halide_target_feature_vulkan_float64, ///< Enable Vulkan 64-bit float support.
1460  halide_target_feature_vulkan_version10, ///< Enable Vulkan v1.0 runtime target support.
1461  halide_target_feature_vulkan_version12, ///< Enable Vulkan v1.2 runtime target support.
1462  halide_target_feature_vulkan_version13, ///< Enable Vulkan v1.3 runtime target support.
1463  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.
1464  halide_target_feature_avx10_1, ///< Intel AVX10 version 1 support. vector_bits is used to indicate width.
1465  halide_target_feature_x86_apx, ///< Intel x86 APX support. Covers initial set of features released as APX: egpr,push2pop2,ppx,ndd .
1466  halide_target_feature_end ///< A sentinel. Every target is considered to have this feature, and setting this feature does nothing.
1468 
1469 /** This function is called internally by Halide in some situations to determine
1470  * if the current execution environment can support the given set of
1471  * halide_target_feature_t flags. The implementation must do the following:
1472  *
1473  * -- If there are flags set in features that the function knows *cannot* be supported, return 0.
1474  * -- Otherwise, return 1.
1475  * -- Note that any flags set in features that the function doesn't know how to test should be ignored;
1476  * this implies that a return value of 1 means "not known to be bad" rather than "known to be good".
1477  *
1478  * In other words: a return value of 0 means "It is not safe to use code compiled with these features",
1479  * while a return value of 1 means "It is not obviously unsafe to use code compiled with these features".
1480  *
1481  * The default implementation simply calls halide_default_can_use_target_features.
1482  *
1483  * Note that `features` points to an array of `count` uint64_t; this array must contain enough
1484  * bits to represent all the currently known features. Any excess bits must be set to zero.
1485  */
1486 // @{
1487 extern int halide_can_use_target_features(int count, const uint64_t *features);
1488 typedef int (*halide_can_use_target_features_t)(int count, const uint64_t *features);
1490 // @}
1491 
1492 /**
1493  * This is the default implementation of halide_can_use_target_features; it is provided
1494  * for convenience of user code that may wish to extend halide_can_use_target_features
1495  * but continue providing existing support, e.g.
1496  *
1497  * int halide_can_use_target_features(int count, const uint64_t *features) {
1498  * if (features[halide_target_somefeature >> 6] & (1LL << (halide_target_somefeature & 63))) {
1499  * if (!can_use_somefeature()) {
1500  * return 0;
1501  * }
1502  * }
1503  * return halide_default_can_use_target_features(count, features);
1504  * }
1505  */
1506 extern int halide_default_can_use_target_features(int count, const uint64_t *features);
1507 
1508 typedef struct halide_dimension_t {
1509 #if (__cplusplus >= 201103L || _MSVC_LANG >= 201103L)
1510  int32_t min = 0, extent = 0, stride = 0;
1511 
1512  // Per-dimension flags. None are defined yet (This is reserved for future use).
1513  uint32_t flags = 0;
1514 
1517  : min(m), extent(e), stride(s), flags(f) {
1518  }
1519 
1520  HALIDE_ALWAYS_INLINE bool operator==(const halide_dimension_t &other) const {
1521  return (min == other.min) &&
1522  (extent == other.extent) &&
1523  (stride == other.stride) &&
1524  (flags == other.flags);
1525  }
1526 
1527  HALIDE_ALWAYS_INLINE bool operator!=(const halide_dimension_t &other) const {
1528  return !(*this == other);
1529  }
1530 #else
1532 
1533  // Per-dimension flags. None are defined yet (This is reserved for future use).
1535 #endif
1537 
1538 #ifdef __cplusplus
1539 } // extern "C"
1540 #endif
1541 
1544 
1545 /**
1546  * The raw representation of an image passed around by generated
1547  * Halide code. It includes some stuff to track whether the image is
1548  * not actually in main memory, but instead on a device (like a
1549  * GPU). For a more convenient C++ wrapper, use Halide::Buffer<T>. */
1550 typedef struct halide_buffer_t {
1551  /** A device-handle for e.g. GPU memory used to back this buffer. */
1553 
1554  /** The interface used to interpret the above handle. */
1556 
1557  /** A pointer to the start of the data in main memory. In terms of
1558  * the Halide coordinate system, this is the address of the min
1559  * coordinates (defined below). */
1561 
1562  /** flags with various meanings. */
1564 
1565  /** The type of each buffer element. */
1566  struct halide_type_t type;
1567 
1568  /** The dimensionality of the buffer. */
1570 
1571  /** The shape of the buffer. Halide does not own this array - you
1572  * must manage the memory for it yourself. */
1574 
1575  /** Pads the buffer up to a multiple of 8 bytes */
1576  void *padding;
1577 
1578 #if (__cplusplus >= 201103L || _MSVC_LANG >= 201103L)
1579  /** Convenience methods for accessing the flags */
1580  // @{
1581  HALIDE_ALWAYS_INLINE bool get_flag(halide_buffer_flags flag) const {
1582  return (flags & flag) != 0;
1583  }
1584 
1585  HALIDE_ALWAYS_INLINE void set_flag(halide_buffer_flags flag, bool value) {
1586  if (value) {
1587  flags |= flag;
1588  } else {
1589  flags &= ~uint64_t(flag);
1590  }
1591  }
1592 
1593  HALIDE_MUST_USE_RESULT HALIDE_ALWAYS_INLINE bool host_dirty() const {
1594  return get_flag(halide_buffer_flag_host_dirty);
1595  }
1596 
1597  HALIDE_MUST_USE_RESULT HALIDE_ALWAYS_INLINE bool device_dirty() const {
1598  return get_flag(halide_buffer_flag_device_dirty);
1599  }
1600 
1601  HALIDE_ALWAYS_INLINE void set_host_dirty(bool v = true) {
1602  set_flag(halide_buffer_flag_host_dirty, v);
1603  }
1604 
1605  HALIDE_ALWAYS_INLINE void set_device_dirty(bool v = true) {
1606  set_flag(halide_buffer_flag_device_dirty, v);
1607  }
1608  // @}
1609 
1610  /** The total number of elements this buffer represents. Equal to
1611  * the product of the extents */
1612  HALIDE_ALWAYS_INLINE size_t number_of_elements() const {
1613  size_t s = 1;
1614  for (int i = 0; i < dimensions; i++) {
1615  s *= dim[i].extent;
1616  }
1617  return s;
1618  }
1619 
1620  /** Offset to the element with the lowest address.
1621  * If all strides are positive, equal to zero.
1622  * Offset is in elements, not bytes.
1623  * Unlike begin(), this is ok to call on an unallocated buffer. */
1624  HALIDE_ALWAYS_INLINE ptrdiff_t begin_offset() const {
1625  ptrdiff_t index = 0;
1626  for (int i = 0; i < dimensions; i++) {
1627  const int stride = dim[i].stride;
1628  if (stride < 0) {
1629  index += stride * (ptrdiff_t)(dim[i].extent - 1);
1630  }
1631  }
1632  return index;
1633  }
1634 
1635  /** An offset to one beyond the element with the highest address.
1636  * Offset is in elements, not bytes.
1637  * Unlike end(), this is ok to call on an unallocated buffer. */
1638  HALIDE_ALWAYS_INLINE ptrdiff_t end_offset() const {
1639  ptrdiff_t index = 0;
1640  for (int i = 0; i < dimensions; i++) {
1641  const int stride = dim[i].stride;
1642  if (stride > 0) {
1643  index += stride * (ptrdiff_t)(dim[i].extent - 1);
1644  }
1645  }
1646  index += 1;
1647  return index;
1648  }
1649 
1650  /** A pointer to the element with the lowest address.
1651  * If all strides are positive, equal to the host pointer.
1652  * Illegal to call on an unallocated buffer. */
1654  return host + begin_offset() * type.bytes();
1655  }
1656 
1657  /** A pointer to one beyond the element with the highest address.
1658  * Illegal to call on an unallocated buffer. */
1659  HALIDE_ALWAYS_INLINE uint8_t *end() const {
1660  return host + end_offset() * type.bytes();
1661  }
1662 
1663  /** The total number of bytes spanned by the data in memory. */
1664  HALIDE_ALWAYS_INLINE size_t size_in_bytes() const {
1665  return (size_t)(end_offset() - begin_offset()) * type.bytes();
1666  }
1667 
1668  /** A pointer to the element at the given location. */
1669  HALIDE_ALWAYS_INLINE uint8_t *address_of(const int *pos) const {
1670  ptrdiff_t index = 0;
1671  for (int i = 0; i < dimensions; i++) {
1672  index += (ptrdiff_t)dim[i].stride * (pos[i] - dim[i].min);
1673  }
1674  return host + index * type.bytes();
1675  }
1676 
1677  /** Attempt to call device_sync for the buffer. If the buffer
1678  * has no device_interface (or no device_sync), this is a quiet no-op.
1679  * Calling this explicitly should rarely be necessary, except for profiling. */
1680  HALIDE_ALWAYS_INLINE int device_sync(void *ctx = nullptr) {
1682  return device_interface->device_sync(ctx, this);
1683  }
1684  return 0;
1685  }
1686 
1687  /** Check if an input buffer passed extern stage is a querying
1688  * bounds. Compared to doing the host pointer check directly,
1689  * this both adds clarity to code and will facilitate moving to
1690  * another representation for bounds query arguments. */
1691  HALIDE_ALWAYS_INLINE bool is_bounds_query() const {
1692  return host == nullptr && device == 0;
1693  }
1694 
1695 #endif
1697 
1698 #ifdef __cplusplus
1699 extern "C" {
1700 #endif
1701 
1702 #ifndef HALIDE_ATTRIBUTE_DEPRECATED
1703 #ifdef HALIDE_ALLOW_DEPRECATED
1704 #define HALIDE_ATTRIBUTE_DEPRECATED(x)
1705 #else
1706 #ifdef _MSC_VER
1707 #define HALIDE_ATTRIBUTE_DEPRECATED(x) __declspec(deprecated(x))
1708 #else
1709 #define HALIDE_ATTRIBUTE_DEPRECATED(x) __attribute__((deprecated(x)))
1710 #endif
1711 #endif
1712 #endif
1713 
1714 /** halide_scalar_value_t is a simple union able to represent all the well-known
1715  * scalar values in a filter argument. Note that it isn't tagged with a type;
1716  * you must ensure you know the proper type before accessing. Most user
1717  * code will never need to create instances of this struct; its primary use
1718  * is to hold def/min/max values in a halide_filter_argument_t. (Note that
1719  * this is conceptually just a union; it's wrapped in a struct to ensure
1720  * that it doesn't get anonymized by LLVM.)
1721  */
1723  union {
1724  bool b;
1733  float f32;
1734  double f64;
1735  void *handle;
1736  } u;
1737 #ifdef __cplusplus
1739  u.u64 = 0;
1740  }
1741 #endif
1742 };
1743 
1748 };
1749 
1750 /*
1751  These structs must be robust across different compilers and settings; when
1752  modifying them, strive for the following rules:
1753 
1754  1) All fields are explicitly sized. I.e. must use int32_t and not "int"
1755  2) All fields must land on an alignment boundary that is the same as their size
1756  3) Explicit padding is added to make that so
1757  4) The sizeof the struct is padded out to a multiple of the largest natural size thing in the struct
1758  5) don't forget that 32 and 64 bit pointers are different sizes
1759 */
1760 
1761 /**
1762  * Obsolete version of halide_filter_argument_t; only present in
1763  * code that wrote halide_filter_metadata_t version 0.
1764  */
1766  const char *name;
1769  struct halide_type_t type;
1770  const struct halide_scalar_value_t *def, *min, *max;
1771 };
1772 
1773 /**
1774  * halide_filter_argument_t is essentially a plain-C-struct equivalent to
1775  * Halide::Argument; most user code will never need to create one.
1776  */
1778  const char *name; // name of the argument; will never be null or empty.
1779  int32_t kind; // actually halide_argument_kind_t
1780  int32_t dimensions; // always zero for scalar arguments
1781  struct halide_type_t type;
1782  // These pointers should always be null for buffer arguments,
1783  // and *may* be null for scalar arguments. (A null value means
1784  // there is no def/min/max/estimate specified for this argument.)
1786  // This pointer should always be null for scalar arguments,
1787  // and *may* be null for buffer arguments. If not null, it should always
1788  // point to an array of dimensions*2 pointers, which will be the (min, extent)
1789  // estimates for each dimension of the buffer. (Note that any of the pointers
1790  // may be null as well.)
1791  int64_t const *const *buffer_estimates;
1792 };
1793 
1795 #ifdef __cplusplus
1796  static const int32_t VERSION = 1;
1797 #endif
1798 
1799  /** version of this metadata; currently always 1. */
1801 
1802  /** The number of entries in the arguments field. This is always >= 1. */
1804 
1805  /** An array of the filters input and output arguments; this will never be
1806  * null. The order of arguments is not guaranteed (input and output arguments
1807  * may come in any order); however, it is guaranteed that all arguments
1808  * will have a unique name within a given filter. */
1810 
1811  /** The Target for which the filter was compiled. This is always
1812  * a canonical Target string (ie a product of Target::to_string). */
1813  const char *target;
1814 
1815  /** The function name of the filter. */
1816  const char *name;
1817 };
1818 
1819 /** halide_register_argv_and_metadata() is a **user-defined** function that
1820  * must be provided in order to use the registration.cc files produced
1821  * by Generators when the 'registration' output is requested. Each registration.cc
1822  * file provides a static initializer that calls this function with the given
1823  * filter's argv-call variant, its metadata, and (optionally) and additional
1824  * textual data that the build system chooses to tack on for its own purposes.
1825  * Note that this will be called at static-initializer time (i.e., before
1826  * main() is called), and in an unpredictable order. Note that extra_key_value_pairs
1827  * may be nullptr; if it's not null, it's expected to be a null-terminated list
1828  * of strings, with an even number of entries. */
1830  int (*filter_argv_call)(void **),
1831  const struct halide_filter_metadata_t *filter_metadata,
1832  const char *const *extra_key_value_pairs);
1833 
1834 /** The functions below here are relevant for pipelines compiled with
1835  * the -profile target flag, which runs a sampling profiler thread
1836  * alongside the pipeline. */
1837 
1838 /** Per-Func state tracked by the sampling profiler. */
1839 struct HALIDE_ATTRIBUTE_ALIGN(8) halide_profiler_func_stats {
1840  /** Total time taken evaluating this Func (in nanoseconds). */
1841  uint64_t time;
1842 
1843  /** The current memory allocation of this Func. */
1844  uint64_t memory_current;
1845 
1846  /** The peak memory allocation of this Func. */
1847  uint64_t memory_peak;
1848 
1849  /** The total memory allocation of this Func. */
1850  uint64_t memory_total;
1851 
1852  /** The peak stack allocation of this Func's threads. */
1853  uint64_t stack_peak;
1854 
1855  /** The average number of thread pool worker threads active while computing this Func. */
1856  uint64_t active_threads_numerator, active_threads_denominator;
1857 
1858  /** The name of this Func. A global constant string. */
1859  const char *name;
1860 
1861  /** The total number of memory allocation of this Func. */
1862  int num_allocs;
1863 };
1864 
1865 /** Per-pipeline state tracked by the sampling profiler. These exist
1866  * in a linked list. */
1867 struct HALIDE_ATTRIBUTE_ALIGN(8) halide_profiler_pipeline_stats {
1868  /** Total time spent in this pipeline (in nanoseconds) */
1869  uint64_t time;
1870 
1871  /** The current memory allocation of funcs in this pipeline. */
1872  uint64_t memory_current;
1873 
1874  /** The peak memory allocation of funcs in this pipeline. */
1875  uint64_t memory_peak;
1876 
1877  /** The total memory allocation of funcs in this pipeline. */
1878  uint64_t memory_total;
1879 
1880  /** The average number of thread pool worker threads doing useful
1881  * work while computing this pipeline. */
1882  uint64_t active_threads_numerator, active_threads_denominator;
1883 
1884  /** The name of this pipeline. A global constant string. */
1885  const char *name;
1886 
1887  /** An array containing states for each Func in this pipeline. */
1888  struct halide_profiler_func_stats *funcs;
1889 
1890  /** The next pipeline_stats pointer. It's a void * because types
1891  * in the Halide runtime may not currently be recursive. */
1892  void *next;
1893 
1894  /** The number of funcs in this pipeline. */
1895  int num_funcs;
1896 
1897  /** The number of times this pipeline has been run. */
1898  int runs;
1899 
1900  /** The total number of samples taken inside of this pipeline. */
1901  int samples;
1902 
1903  /** The total number of memory allocation of funcs in this pipeline. */
1904  int num_allocs;
1905 };
1906 
1907 /** Per-invocation-of-a-pipeline state. Lives on the stack of the Halide
1908  * code. Exists in a doubly-linked list to that it can be cleanly
1909  * removed. */
1910 struct HALIDE_ATTRIBUTE_ALIGN(8) halide_profiler_instance_state {
1911  /** Time billed to funcs in this instance by the sampling thread. */
1912  uint64_t billed_time;
1913 
1914  /** Wall clock time of the start of the instance. */
1915  uint64_t start_time;
1916 
1917  /** The current memory allocation of funcs in this instance. */
1918  uint64_t memory_current;
1919 
1920  /** The peak memory allocation of funcs in this instance. */
1921  uint64_t memory_peak;
1922 
1923  /** The total memory allocation of funcs in this instance. */
1924  uint64_t memory_total;
1925 
1926  /** The average number of thread pool worker threads doing useful
1927  * work while computing this instance. */
1928  uint64_t active_threads_numerator, active_threads_denominator;
1929 
1930  /** A pointer to the next running instance, so that the running instances
1931  * can exist in a linked list. */
1932  struct halide_profiler_instance_state *next;
1933 
1934  /** A pointer to the address of the next pointer of the previous instance,
1935  * so that this can be removed from the linked list when the instance
1936  * terminates. */
1937  struct halide_profiler_instance_state **prev_next;
1938 
1939  /** Information shared across all instances. The stats above are merged into
1940  * it when the instance is retired. */
1941  struct halide_profiler_pipeline_stats *pipeline_stats;
1942 
1943  /** An array containing states for each Func in this instance of this pipeline. */
1944  struct halide_profiler_func_stats *funcs;
1945 
1946  /** The id of the current running Func. Set by the pipeline, read
1947  * periodically by the profiler thread. */
1948  int current_func;
1949 
1950  /** The number of threads currently doing work on this pipeline instance. */
1951  int active_threads;
1952 
1953  /** The number of samples taken by this instance. */
1954  int samples;
1955 
1956  /** The total number of memory allocation of funcs in this instance. */
1957  int num_allocs;
1958 
1959  /** Whether or not this instance should count towards pipeline
1960  * statistics. */
1961  int should_collect_statistics;
1962 };
1963 
1964 /** The global state of the profiler. */
1966  /** Guards access to the fields below. If not locked, the sampling
1967  * profiler thread is free to modify things below (including
1968  * reordering the linked list of pipeline stats). */
1969  struct halide_mutex lock;
1970 
1971  /** A linked list of stats gathered for each pipeline. */
1972  struct halide_profiler_pipeline_stats *pipelines;
1973 
1974  /** Retrieve remote profiler state. Used so that the sampling
1975  * profiler can follow along with execution that occurs elsewhere,
1976  * e.g. on a DSP. If null, it reads from the int above instead. */
1977 
1978  /** Sampling thread reference to be joined at shutdown. */
1979  struct halide_thread *sampling_thread;
1980 
1981  /** The running instances of Halide pipelines. */
1982  struct halide_profiler_instance_state *instances;
1983 
1984  /** If this callback is defined, the profiler asserts that there is a single
1985  * live instance, and then uses it to get the current func and number of
1986  * active threads insted of reading the fields in the instance. This is used
1987  * so that the profiler can follow along with execution that occurs
1988  * elsewhere (e.g. on an accelerator). */
1989  void (*get_remote_profiler_state)(int *func, int *active_workers);
1990 
1991  /** The amount of time the profiler thread sleeps between samples in
1992  * microseconds. Defaults to 1000. To change it call
1993  * halide_profiler_get_state and mutate this field. */
1995 
1996  /** Set to 1 when you want the profiler to wait for all running instances to
1997  * finish and then stop gracefully. */
1999 };
2000 
2001 /** Get a pointer to the global profiler state for programmatic
2002  * inspection. Lock it before using to pause the profiler. */
2004 
2005 /** Get a pointer to the pipeline state associated with pipeline_name.
2006  * This function grabs the global profiler state's lock on entry. */
2007 extern struct halide_profiler_pipeline_stats *halide_profiler_get_pipeline_state(const char *pipeline_name);
2008 
2009 /** Collects profiling information. Intended to be called from a timer
2010  * interrupt handler if timer based profiling is being used.
2011  * State argument is acquired via halide_profiler_get_pipeline_state.
2012  * prev_t argument is the previous time and can be used to set a more
2013  * accurate time interval if desired. */
2015 
2016 /** Reset profiler state cheaply. May leave threads running or some memory
2017  * allocated but all accumulated statistics are reset. Blocks until all running
2018  * profiled Halide pipelines exit. */
2019 extern void halide_profiler_reset(void);
2020 
2021 /** Reset all profiler state. Blocks until all running profiled Halide
2022  * pipelines exit. */
2023 extern void halide_profiler_shutdown(void);
2024 
2025 /** Print out timing statistics for everything run since the last
2026  * reset. Also happens at process exit. */
2028 
2029 /** These routines are called to temporarily disable and then reenable
2030  * the profiler. */
2031 //@{
2034 //@}
2035 
2036 /// \name "Float16" functions
2037 /// These functions operate of bits (``uint16_t``) representing a half
2038 /// precision floating point number (IEEE-754 2008 binary16).
2039 //{@
2040 
2041 /** Read bits representing a half precision floating point number and return
2042  * the float that represents the same value */
2044 
2045 /** Read bits representing a half precision floating point number and return
2046  * the double that represents the same value */
2048 
2049 // TODO: Conversion functions to half
2050 
2051 //@}
2052 
2053 // Allocating and freeing device memory is often very slow. The
2054 // methods below give Halide's runtime permission to hold onto device
2055 // memory to service future requests instead of returning it to the
2056 // underlying device API. The API does not manage an allocation pool,
2057 // all it does is provide access to a shared counter that acts as a
2058 // limit on the unused memory not yet returned to the underlying
2059 // device API. It makes callbacks to participants when memory needs to
2060 // be released because the limit is about to be exceeded (either
2061 // because the limit has been reduced, or because the memory owned by
2062 // some participant becomes unused).
2063 
2064 /** Tell Halide whether or not it is permitted to hold onto device
2065  * allocations to service future requests instead of returning them
2066  * eagerly to the underlying device API. Many device allocators are
2067  * quite slow, so it can be beneficial to set this to true. The
2068  * default value for now is false.
2069  *
2070  * Note that if enabled, the eviction policy is very simplistic. The
2071  * 32 most-recently used allocations are preserved, regardless of
2072  * their size. Additionally, if a call to cuMalloc results in an
2073  * out-of-memory error, the entire cache is flushed and the allocation
2074  * is retried. See https://github.com/halide/Halide/issues/4093
2075  *
2076  * If set to false, releases all unused device allocations back to the
2077  * underlying device APIs. For finer-grained control, see specific
2078  * methods in each device api runtime.
2079  *
2080  * Note that if the flag is set to true, this call *must* succeed and return
2081  * a value of halide_error_code_success (i.e., zero); if you replace
2082  * the implementation of this call in the runtime, you must honor this contract.
2083  * */
2085 
2086 /** Determines whether on device_free the memory is returned
2087  * immediately to the device API, or placed on a free list for future
2088  * use. Override and switch based on the user_context for
2089  * finer-grained control. By default just returns the value most
2090  * recently set by the method above. */
2092 
2096 };
2097 
2098 /** Register a callback to be informed when
2099  * halide_reuse_device_allocations(false) is called, and all unused
2100  * device allocations must be released. The object passed should have
2101  * global lifetime, and its next field will be clobbered. */
2103 
2104 #ifdef __cplusplus
2105 } // End extern "C"
2106 #endif
2107 
2108 #if (__cplusplus >= 201103L || _MSVC_LANG >= 201103L)
2109 
2110 namespace {
2111 
2112 template<typename T>
2113 struct check_is_pointer {
2114  static constexpr bool value = false;
2115 };
2116 
2117 template<typename T>
2118 struct check_is_pointer<T *> {
2119  static constexpr bool value = true;
2120 };
2121 
2122 } // namespace
2123 
2124 /** Construct the halide equivalent of a C type */
2125 template<typename T>
2126 HALIDE_ALWAYS_INLINE constexpr halide_type_t halide_type_of() {
2127  // Create a compile-time error if T is not a pointer (without
2128  // using any includes - this code goes into the runtime).
2129  // (Note that we can't have uninitialized variables in constexpr functions,
2130  // even if those variables aren't used.)
2131  static_assert(check_is_pointer<T>::value, "Expected a pointer type here");
2132  return halide_type_t(halide_type_handle, 64);
2133 }
2134 
2135 #ifdef HALIDE_CPP_COMPILER_HAS_FLOAT16
2136 template<>
2137 HALIDE_ALWAYS_INLINE constexpr halide_type_t halide_type_of<_Float16>() {
2138  return halide_type_t(halide_type_float, 16);
2139 }
2140 #endif
2141 
2142 template<>
2143 HALIDE_ALWAYS_INLINE constexpr halide_type_t halide_type_of<float>() {
2144  return halide_type_t(halide_type_float, 32);
2145 }
2146 
2147 template<>
2148 HALIDE_ALWAYS_INLINE constexpr halide_type_t halide_type_of<double>() {
2149  return halide_type_t(halide_type_float, 64);
2150 }
2151 
2152 template<>
2153 HALIDE_ALWAYS_INLINE constexpr halide_type_t halide_type_of<bool>() {
2154  return halide_type_t(halide_type_uint, 1);
2155 }
2156 
2157 template<>
2158 HALIDE_ALWAYS_INLINE constexpr halide_type_t halide_type_of<uint8_t>() {
2159  return halide_type_t(halide_type_uint, 8);
2160 }
2161 
2162 template<>
2163 HALIDE_ALWAYS_INLINE constexpr halide_type_t halide_type_of<uint16_t>() {
2164  return halide_type_t(halide_type_uint, 16);
2165 }
2166 
2167 template<>
2168 HALIDE_ALWAYS_INLINE constexpr halide_type_t halide_type_of<uint32_t>() {
2169  return halide_type_t(halide_type_uint, 32);
2170 }
2171 
2172 template<>
2173 HALIDE_ALWAYS_INLINE constexpr halide_type_t halide_type_of<uint64_t>() {
2174  return halide_type_t(halide_type_uint, 64);
2175 }
2176 
2177 template<>
2178 HALIDE_ALWAYS_INLINE constexpr halide_type_t halide_type_of<int8_t>() {
2179  return halide_type_t(halide_type_int, 8);
2180 }
2181 
2182 template<>
2183 HALIDE_ALWAYS_INLINE constexpr halide_type_t halide_type_of<int16_t>() {
2184  return halide_type_t(halide_type_int, 16);
2185 }
2186 
2187 template<>
2188 HALIDE_ALWAYS_INLINE constexpr halide_type_t halide_type_of<int32_t>() {
2189  return halide_type_t(halide_type_int, 32);
2190 }
2191 
2192 template<>
2193 HALIDE_ALWAYS_INLINE constexpr halide_type_t halide_type_of<int64_t>() {
2194  return halide_type_t(halide_type_int, 64);
2195 }
2196 
2197 #ifndef COMPILING_HALIDE_RUNTIME
2198 
2199 // These structures are used by `function_info_header` files
2200 // (generated by passing `-e function_info_header` to a Generator).
2201 // The generated files contain documentation on the proper usage.
2202 namespace HalideFunctionInfo {
2203 
2204 enum ArgumentKind { InputScalar = 0,
2205  InputBuffer = 1,
2206  OutputBuffer = 2 };
2207 
2208 struct ArgumentInfo {
2209  std::string_view name;
2210  ArgumentKind kind;
2211  int32_t dimensions; // always zero for scalar arguments
2212  halide_type_t type;
2213 };
2214 
2215 } // namespace HalideFunctionInfo
2216 
2217 #endif // COMPILING_HALIDE_RUNTIME
2218 
2219 #endif // (__cplusplus >= 201103L || _MSVC_LANG >= 201103L)
2220 
2221 #endif // HALIDE_HALIDERUNTIME_H
halide_error_handler_t halide_set_error_handler(halide_error_handler_t handler)
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)
void *(* halide_get_library_symbol_t)(void *lib, const char *name)
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 ...
void *(* halide_malloc_t)(void *, size_t)
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)
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_profiler_lock(struct halide_profiler_state *)
These routines are called to temporarily disable and then reenable the profiler.
void halide_cond_signal(struct halide_cond *cond)
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_default_get_symbol(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_get_symbol(const char *name)
Halide calls these functions to interact with the underlying system runtime functions.
void halide_default_print(void *user_context, const char *)
void * halide_get_library_symbol(void *lib, const char *name)
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_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_armv86a
Enable ARMv8.6a instructions.
@ 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_armv89a
Enable ARMv8.9a instructions.
@ 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_armv88a
Enable ARMv8.8a instructions.
@ 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_hvx_v68
Enable Hexagon v68 architecture.
@ halide_target_feature_avx10_1
Intel AVX10 version 1 support. vector_bits is used to indicate width.
@ 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_armv8a
Enable ARMv8a instructions.
@ halide_target_feature_asan
Enable hooks for ASAN support.
@ halide_target_feature_armv87a
Enable ARMv8.7a instructions.
@ 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_armv82a
Enable ARMv8.2a instructions.
@ halide_target_feature_hvx_v62
Enable Hexagon v62 architecture.
@ halide_target_feature_armv84a
Enable ARMv8.4a instructions.
@ halide_target_feature_cuda
Enable the CUDA runtime. Defaults to compute capability 2.0 (Fermi)
@ halide_target_feature_armv81a
Enable ARMv8.1a 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_armv85a
Enable ARMv8.5a instructions.
@ 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_armv83a
Enable ARMv8.3a instructions.
@ 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_x86_apx
Intel x86 APX support. Covers initial set of features released as APX: egpr,push2pop2,...
@ halide_target_feature_hvx_v66
Enable Hexagon v66 architecture.
void halide_free(void *user_context, void *ptr)
bool halide_semaphore_try_acquire(struct halide_semaphore_t *, int n)
halide_buffer_flags
@ halide_buffer_flag_device_dirty
@ halide_buffer_flag_host_dirty
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)
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)
int halide_copy_to_host(void *user_context, struct halide_buffer_t *buf)
Copy image data from device memory to host memory.
int halide_default_do_par_for(void *user_context, halide_task_t task, int min, int size, uint8_t *closure)
The default versions of the parallel runtime functions.
int halide_msan_annotate_memory_is_initialized(void *user_context, const void *ptr, uint64_t len)
Annotate that a given range of memory has been initialized; only used when Target::MSAN is enabled.
halide_print_t halide_set_custom_print(halide_print_t print)
int halide_error_bad_dimensions(void *user_context, const char *func_name, int32_t dimensions_given, int32_t correct_dimensions)
int halide_mutex_array_unlock(struct halide_mutex_array *array, int entry)
int halide_error_constraint_violated(void *user_context, const char *var, int val, const char *constrained_var, int constrained_val)
int halide_default_do_loop_task(void *user_context, halide_loop_task_t f, int min, int extent, uint8_t *closure, void *task_parent)
void halide_profiler_reset(void)
Reset profiler state cheaply.
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_default_load_library(const char *name)
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
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_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_load_library(const char *name)
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_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)
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)
int halide_error_out_of_memory(void *user_context)
int32_t halide_debug_to_file(void *user_context, const char *filename, struct halide_buffer_t *buf)
Called when debug_to_file is used inside Halide code.
struct halide_mutex_array * halide_mutex_array_create(uint64_t sz)
int halide_error_no_device_interface(void *user_context)
void *(* halide_load_library_t)(const char *name)
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)
struct halide_thread * halide_spawn_thread(void(*f)(void *), void *closure)
Spawn a thread.
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 ...
struct halide_profiler_state * halide_profiler_get_state(void)
Get a pointer to the global profiler state for programmatic inspection.
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)
void * halide_default_malloc(void *user_context, size_t x)
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_vscale_invalid(void *user_context, const char *func_name, int runtime_vscale, int compiletime_vscale)
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)
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.
void halide_profiler_shutdown(void)
Reset all profiler state.
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)
void *(* halide_get_symbol_t)(const char *name)
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_malloc(void *user_context, size_t x)
Halide calls these functions to allocate and free memory.
void * halide_default_get_library_symbol(void *lib, 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)
int halide_get_num_threads()
Get or set the number of threads used by Halide's thread pool.
void halide_default_free(void *user_context, void *ptr)
void halide_shutdown_thread_pool(void)
#define HALIDE_MUST_USE_RESULT
Definition: HalideRuntime.h:65
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_split_factor_not_positive
A factor used to split a loop was discovered to be zero or negative at runtime.
@ 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_vscale_invalid
"vscale" value of Scalable Vector detected in runtime does not match the vscale value used in compila...
@ 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_cannot_profile_pipeline
Profiling failed for a pipeline invocation.
@ 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 ...
void halide_memoization_cache_set_size(int64_t size)
Set the soft maximum amount of memory, in bytes, that the LRU cache will use to memoize Func results.
#define HALIDE_ALWAYS_INLINE
Definition: HalideRuntime.h:49
void halide_cond_broadcast(struct halide_cond *cond)
int halide_device_free(void *user_context, struct halide_buffer_t *buf)
Free device memory.
void halide_profiler_unlock(struct halide_profiler_state *)
void halide_memoization_cache_cleanup(void)
Free all memory and resources associated with the memoization cache.
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 *)
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)
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_split_factor_not_positive(void *user_context, const char *func_name, const char *orig, const char *outer, const char *inner, const char *factor_str, int factor)
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.
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.
int halide_shutdown_trace(void)
If tracing is writing to a file.
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.
auto begin(reverse_adaptor< T > i)
Definition: Util.h:467
ConstantInterval min(const ConstantInterval &a, const ConstantInterval &b)
bool operator<(const ConstantInterval &a, const ConstantInterval &b)
Expr with_lanes(const Expr &x, int lanes)
Rewrite the expression x to have lanes lanes.
auto end(reverse_adaptor< T > i)
Definition: Util.h:472
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:1130
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:1143
Expr print(const std::vector< Expr > &values)
Create an Expr that prints out its value whenever it is evaluated.
unsigned __INT64_TYPE__ uint64_t
signed __INT64_TYPE__ int64_t
__UINTPTR_TYPE__ uintptr_t
signed __INT32_TYPE__ int32_t
unsigned __INT8_TYPE__ uint8_t
__PTRDIFF_TYPE__ ptrdiff_t
unsigned __INT16_TYPE__ uint16_t
__SIZE_TYPE__ size_t
unsigned __INT32_TYPE__ uint32_t
signed __INT16_TYPE__ int16_t
signed __INT8_TYPE__ int8_t
The raw representation of an image passed around by generated Halide code.
void * padding
Pads the buffer up to a multiple of 8 bytes.
int32_t dimensions
The dimensionality of the buffer.
halide_dimension_t * dim
The shape of the buffer.
uint64_t device
A device-handle for e.g.
uint8_t * host
A pointer to the start of the data in main memory.
struct halide_type_t type
The type of each buffer element.
const struct halide_device_interface_t * device_interface
The interface used to interpret the above handle.
uint64_t flags
flags with various meanings.
Cross platform condition variable.
uintptr_t _private[1]
struct halide_device_allocation_pool * next
int(* release_unused)(void *user_context)
Each GPU API provides a halide_device_interface_t struct pointing to the code that manages device all...
int(* device_slice)(void *user_context, const struct halide_buffer_t *src, int slice_dim, int slice_pos, struct halide_buffer_t *dst)
int(* device_and_host_malloc)(void *user_context, struct halide_buffer_t *buf, const struct halide_device_interface_t *device_interface)
const struct halide_device_interface_impl_t * impl
int(* wrap_native)(void *user_context, struct halide_buffer_t *buf, uint64_t handle, const struct halide_device_interface_t *device_interface)
int(* compute_capability)(void *user_context, int *major, int *minor)
int(* device_release_crop)(void *user_context, struct halide_buffer_t *buf)
int(* device_crop)(void *user_context, const struct halide_buffer_t *src, struct halide_buffer_t *dst)
void(* device_release)(void *user_context, const struct halide_device_interface_t *device_interface)
int(* copy_to_host)(void *user_context, struct halide_buffer_t *buf)
int(* copy_to_device)(void *user_context, struct halide_buffer_t *buf, const struct halide_device_interface_t *device_interface)
int(* device_free)(void *user_context, struct halide_buffer_t *buf)
int(* device_sync)(void *user_context, struct halide_buffer_t *buf)
int(* detach_native)(void *user_context, struct halide_buffer_t *buf)
int(* device_and_host_free)(void *user_context, struct halide_buffer_t *buf)
int(* device_malloc)(void *user_context, struct halide_buffer_t *buf, const struct halide_device_interface_t *device_interface)
int(* buffer_copy)(void *user_context, struct halide_buffer_t *src, const struct halide_device_interface_t *dst_device_interface, struct halide_buffer_t *dst)
Obsolete version of halide_filter_argument_t; only present in code that wrote halide_filter_metadata_...
const struct halide_scalar_value_t * min
const struct halide_scalar_value_t * def
const struct halide_scalar_value_t * max
struct halide_type_t type
halide_filter_argument_t is essentially a plain-C-struct equivalent to Halide::Argument; most user co...
const struct halide_scalar_value_t * scalar_estimate
const struct halide_scalar_value_t * scalar_max
int64_t const *const * buffer_estimates
const struct halide_scalar_value_t * scalar_def
struct halide_type_t type
const struct halide_scalar_value_t * scalar_min
const char * name
The function name of the filter.
int32_t version
version of this metadata; currently always 1.
const struct halide_filter_argument_t * arguments
An array of the filters input and output arguments; this will never be null.
int32_t num_arguments
The number of entries in the arguments field.
const char * target
The Target for which the filter was compiled.
A type traits template to provide a halide_handle_cplusplus_type value from a C++ type.
Definition: Type.h:256
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)
If this callback is defined, the profiler asserts that there is a single live instance,...
struct halide_thread * sampling_thread
Retrieve remote profiler state.
int sleep_time
The amount of time the profiler thread sleeps between samples in microseconds.
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.
struct halide_profiler_instance_state * instances
The running instances of Halide pipelines.
int shutdown
Set to 1 when you want the profiler to wait for all running instances to finish and then stop gracefu...
halide_scalar_value_t is a simple union able to represent all the well-known scalar values in a filte...
union halide_scalar_value_t::@3 u
A struct representing a semaphore and a number of items that must be acquired from it.
struct halide_semaphore_t * semaphore
An opaque struct representing a semaphore.
uint64_t _private[2]
void * value
If the event type is a load or a store, this points to the value being loaded or stored.
int32_t * coordinates
For loads and stores, an array which contains the location being accessed.
const char * func
The name of the Func or Pipeline that this event refers to.
const char * trace_tag
For halide_trace_tag, this points to a read-only null-terminated string of arbitrary text.
struct halide_type_t type
If the event type is a load or a store, this is the type of the data.
int32_t value_index
If this was a load or store of a Tuple-valued Func, this is which tuple element was accessed.
enum halide_trace_event_code_t event
The type of event.
int32_t dimensions
The length of the coordinates array.
The header of a packet in a binary trace.
uint32_t size
The total size of this packet in bytes.
int32_t id
The id of this packet (for the purpose of parent_id).
enum halide_trace_event_code_t event
struct halide_type_t type
The remaining fields are equivalent to those in halide_trace_event_t.
A runtime tag for a type in the halide type system.
uint8_t bits
The number of bits of precision of a single scalar value of this type.
uint16_t lanes
How many elements in a vector.
uint8_t code
The basic type code: signed integer, unsigned integer, or floating point.
int error_code
void * user_context