HalideRuntime.h
#ifndef HALIDE_HALIDERUNTIME_H
#define HALIDE_HALIDERUNTIME_H
#ifndef COMPILING_HALIDE_RUNTIME
#include <stddef.h>
#include <stdint.h>
#include <stdbool.h>
#else
#include "runtime_internal.h"
#endif
#ifdef __cplusplus
// Forward declare type to allow naming typed handles.
// See Type.h for documentation.
template<typename T> struct halide_handle_traits;
#endif
#ifdef __cplusplus
extern "C" {
#endif
/** \file
*
* This file declares the routines used by Halide internally in its
* runtime. On platforms that support weak linking, these can be
* replaced with user-defined versions by defining an extern "C"
* function with the same name and signature.
*
* When doing Just In Time (JIT) compilation methods on the Func being
* compiled must be called instead. The corresponding methods are
* documented below.
*
* All of these functions take a "void *user_context" parameter as their
* first argument; if the Halide kernel that calls back to any of these
* functions has been compiled with the UserContext feature set on its Target,
* then the value of that pointer passed from the code that calls the
* Halide kernel is piped through to the function.
*
* Some of these are also useful to call when using the default
* implementation. E.g. halide_shutdown_thread_pool.
*
* Note that even on platforms with weak linking, some linker setups
* may not respect the override you provide. E.g. if the override is
* in a shared library and the halide object files are linked directly
* into the output, the builtin versions of the runtime functions will
* be called. See your linker documentation for more details. On
* Linux, LD_DYNAMIC_WEAK=1 may help.
*
*/
// Forward-declare to suppress warnings if compiling as C.
#ifndef BUFFER_T_DEFINED
struct buffer_t;
#endif
/** Print a message to stderr. Main use is to support HL_TRACE
* functionality, print, and print_when calls. Also called by the default
* halide_error. This function can be replaced in JITed code by using
* halide_custom_print and providing an implementation of halide_print
* in AOT code. See Func::set_custom_print.
*/
// @{
extern void halide_print(void *user_context, const char *);
typedef void (*halide_print_t)(void *, const char *);
extern halide_print_t halide_set_custom_print(halide_print_t print);
// @}
/** Halide calls this function on runtime errors (for example bounds
* checking failures). This function can be replaced in JITed code by
* using Func::set_error_handler, or in AOT code by calling
* halide_set_error_handler. In AOT code on platforms that support
* weak linking (i.e. not Windows), you can also override it by simply
* defining your own halide_error.
*/
// @{
extern void halide_error(void *user_context, const char *);
typedef void (*halide_error_handler_t)(void *, const char *);
extern halide_error_handler_t halide_set_error_handler(halide_error_handler_t handler);
// @}
/** Cross-platform mutex. These are allocated statically inside the
* runtime, hence the fixed size. They must be initialized with
* zero. The first time halide_mutex_lock is called, the lock must be
* initialized in a thread safe manner. This incurs a small overhead
* for a once mechanism, but makes the lock reliably easy to setup and
* use without depending on e.g. C++ constructor logic.
*/
struct halide_mutex {
uint64_t _private[8];
};
/** A basic set of mutex and condition variable functions, which call
* platform specific code for mutual exclusion. Equivalent to posix
* calls. Mutexes should initially be set to zero'd memory. Any
* resources required are created on first lock. Calling destroy
* re-zeros the memory.
*/
//@{
extern void halide_mutex_lock(struct halide_mutex *mutex);
extern void halide_mutex_unlock(struct halide_mutex *mutex);
extern void halide_mutex_destroy(struct halide_mutex *mutex);
//@}
/** Define halide_do_par_for to replace the default thread pool
* implementation. halide_shutdown_thread_pool can also be called to
* release resources used by the default thread pool on platforms
* where it makes sense. (E.g. On Mac OS, Grand Central Dispatch is
* used so %Halide does not own the threads backing the pool and they
* cannot be released.) See Func::set_custom_do_task and
* Func::set_custom_do_par_for. Should return zero if all the jobs
* return zero, or an arbitrarily chosen return value from one of the
* jobs otherwise.
*/
//@{
typedef int (*halide_task_t)(void *user_context, int task_number, uint8_t *closure);
extern int halide_do_par_for(void *user_context,
halide_task_t task,
int min, int size, uint8_t *closure);
extern void halide_shutdown_thread_pool();
//@}
/** Set a custom method for performing a parallel for loop. Returns
* the old do_par_for handler. */
typedef int (*halide_do_par_for_t)(void *, halide_task_t, int, int, uint8_t*);
extern halide_do_par_for_t halide_set_custom_do_par_for(halide_do_par_for_t do_par_for);
/** If you use the default do_par_for, you can still set a custom
* handler to perform each individual task. Returns the old handler. */
//@{
typedef int (*halide_do_task_t)(void *, halide_task_t, int, uint8_t *);
extern halide_do_task_t halide_set_custom_do_task(halide_do_task_t do_task);
extern int halide_do_task(void *user_context, halide_task_t f, int idx,
uint8_t *closure);
//@}
struct halide_thread;
/** Spawn a thread. Returns a handle to the thread for the purposes of
* joining it. The thread must be joined in order to clean up any
* resources associated with it. */
extern struct halide_thread *halide_spawn_thread(void (*f)(void *), void *closure);
/** Join a thread. */
extern void halide_join_thread(struct halide_thread *);
/** Set the number of threads used by Halide's thread pool. Returns
* the old number.
*
* n < 0 : error condition
* n == 0 : use a reasonable system default (typically, number of cpus online).
* n == 1 : use exactly one thread; this will always enforce serial execution
* n > 1 : use a pool of exactly n threads.
*
* Note that the default iOS and OSX behavior will treat n > 1 like n == 0;
* that is, any positive value other than 1 will use a system-determined number
* of threads.
*
* (Note that this is only guaranteed when using the default implementations
* of halide_do_par_for(); custom implementations may completely ignore values
* passed to halide_set_num_threads().)
*/
extern int halide_set_num_threads(int n);
/** Halide calls these functions to allocate and free memory. To
* replace in AOT code, use the halide_set_custom_malloc and
* halide_set_custom_free, or (on platforms that support weak
* linking), simply define these functions yourself. In JIT-compiled
* code use Func::set_custom_allocator.
*
* Note that halide_malloc must return a pointer aligned to the
* maximum meaningful alignment for the platform for the purpose of
* vector loads and stores. The default implementation uses 32-byte
* alignment, which is safe for arm and x86. Additionally, it must be
* safe to read at least 8 bytes before the start and beyond the
* end.
*/
//@{
extern void *halide_malloc(void *user_context, size_t x);
extern void halide_free(void *user_context, void *ptr);
typedef void *(*halide_malloc_t)(void *, size_t);
typedef void (*halide_free_t)(void *, void *);
extern halide_malloc_t halide_set_custom_malloc(halide_malloc_t user_malloc);
extern halide_free_t halide_set_custom_free(halide_free_t user_free);
//@}
/** Halide calls these functions to interact with the underlying
* system runtime functions. To replace in AOT code on platforms that
* support weak linking, define these functions yourself.
*
* halide_load_library and halide_get_library_symbol are equivalent to
* dlopen and dlsym. halide_get_symbol(sym) is equivalent to
* dlsym(RTLD_DEFAULT, sym).
*/
//@{
extern void *halide_get_symbol(const char *name);
extern void *halide_load_library(const char *name);
extern void *halide_get_library_symbol(void *lib, const char *name);
//@}
/** Called when debug_to_file is used inside %Halide code. See
* Func::debug_to_file for how this is called
*
* Cannot be replaced in JITted code at present.
*/
extern int32_t halide_debug_to_file(void *user_context, const char *filename,
int32_t type_code,
struct buffer_t *buf);
/** Types in the halide type system. They can be ints, unsigned ints,
* or floats (of various bit-widths), or a handle (which is always 64-bits).
* Note that the int/uint/float values do not imply a specific bit width
* (the bit width is expected to be encoded in a separate value).
*/
typedef enum halide_type_code_t
#if __cplusplus >= 201103L
: uint8_t
#endif
{
halide_type_int = 0, //!< signed integers
halide_type_uint = 1, //!< unsigned integers
halide_type_float = 2, //!< floating point numbers
halide_type_handle = 3 //!< opaque pointer type (void *)
} halide_type_code_t;
// Note that while __attribute__ can go before or after the declaration,
// __declspec apparently is only allowed before.
#ifndef HALIDE_ATTRIBUTE_ALIGN
#ifdef _MSC_VER
#define HALIDE_ATTRIBUTE_ALIGN(x) __declspec(align(x))
#else
#define HALIDE_ATTRIBUTE_ALIGN(x) __attribute__((aligned(x)))
#endif
#endif
/** A runtime tag for a type in the halide type system. Can be ints,
* unsigned ints, or floats of various bit-widths (the 'bits'
* field). Can also be vectors of the same (by setting the 'lanes'
* field to something larger than one). This struct should be
* exactly 32-bits in size. */
struct halide_type_t {
/** The basic type code: signed integer, unsigned integer, or floating point. */
#if __cplusplus >= 201103L
HALIDE_ATTRIBUTE_ALIGN(1) halide_type_code_t code; // halide_type_code_t
#else
HALIDE_ATTRIBUTE_ALIGN(1) uint8_t code; // halide_type_code_t
#endif
/** The number of bits of precision of a single scalar value of this type. */
HALIDE_ATTRIBUTE_ALIGN(1) uint8_t bits;
/** How many elements in a vector. This is 1 for scalar types. */
HALIDE_ATTRIBUTE_ALIGN(2) uint16_t lanes;
#ifdef __cplusplus
/** Construct a runtime representation of a Halide type from:
* code: The fundamental type from an enum.
* bits: The bit size of one element.
* lanes: The number of vector elements in the type. */
halide_type_t(halide_type_code_t code, uint8_t bits, uint16_t lanes = 1)
: code(code), bits(bits), lanes(lanes) {
}
/** Default constructor is required e.g. to declare halide_trace_event
* instances. */
halide_type_t() : code((halide_type_code_t)0), bits(0), lanes(0) {}
/** Compare two types for equality. */
bool operator==(const halide_type_t &other) const {
return (code == other.code &&
bits == other.bits &&
lanes == other.lanes);
}
bool operator!=(const halide_type_t &other) const {
return !(*this == other);
}
/** Size in bytes for a single element, even if width is not 1, of this type. */
size_t bytes() const { return (bits + 7) / 8; }
#endif
};
enum halide_trace_event_code {halide_trace_load = 0,
halide_trace_store = 1,
halide_trace_begin_realization = 2,
halide_trace_end_realization = 3,
halide_trace_produce = 4,
halide_trace_consume = 5,
halide_trace_end_consume = 6,
halide_trace_begin_pipeline = 7,
halide_trace_end_pipeline = 8};
#pragma pack(push, 1)
struct halide_trace_event {
const char *func;
enum halide_trace_event_code event;
int32_t parent_id;
struct halide_type_t type;
int32_t value_index;
void *value;
int32_t dimensions;
int32_t *coordinates;
};
#pragma pack(pop)
/** Called when Funcs are marked as trace_load, trace_store, or
* trace_realization. See Func::set_custom_trace. The default
* implementation either prints events via halide_printf, or if
* HL_TRACE_FILE is defined, dumps the trace to that file in a
* yet-to-be-documented binary format (see src/runtime/tracing.cpp to
* reverse engineer the format). If the trace is going to be large,
* you may want to make the file a named pipe, and then read from that
* pipe into gzip.
*
* halide_trace returns a unique ID which will be passed to future
* events that "belong" to the earlier event as the parent id. The
* ownership hierarchy looks like:
*
* begin_realization
* produce
* store
* update
* load/store
* consume
* load
* end_consume
* end_realization
*
* Threading means that ownership cannot be inferred from the ordering
* of events. There can be many active realizations of a given
* function, or many active productions for a single
* realization. Within a single production, the ordering of events is
* meaningful.
*/
// @}
extern int32_t halide_trace(void *user_context, const struct halide_trace_event *event);
typedef int32_t (*halide_trace_t)(void *user_context, const struct halide_trace_event *);
extern halide_trace_t halide_set_custom_trace(halide_trace_t trace);
// @}
/** Set the file descriptor that Halide should write binary trace
* events to. If called with 0 as the argument, Halide outputs trace
* information to stdout in a human-readable format. If never called,
* Halide checks the for existence of an environment variable called
* HL_TRACE_FILE and opens that file. If HL_TRACE_FILE is not defined,
* it outputs trace information to stdout in a human-readable
* format. */
extern void halide_set_trace_file(int fd);
/** Halide calls this to retrieve the file descriptor to write binary
* trace events to. The default implementation returns the value set
* by halide_set_trace_file. Implement it yourself if you wish to use
* a custom file descriptor per user_context. Return zero from your
* implementation to tell Halide to print human-readable trace
* information to stdout. */
extern int halide_get_trace_file(void *user_context);
/** If tracing is writing to a file. This call closes that file
* (flushing the trace). Returns zero on success. */
extern int halide_shutdown_trace();
/** All Halide GPU or device backend implementations much provide an interface
* to be used with halide_device_malloc, etc.
*/
struct halide_device_interface_t;
/** Release all data associated with the current GPU backend, in particular
* all resources (memory, texture, context handles) allocated by Halide. Must
* be called explicitly when using AOT compilation. */
extern void halide_device_release(void *user_context, const struct halide_device_interface_t *device_interface);
/** Copy image data from device memory to host memory. This must be called
* explicitly to copy back the results of a GPU-based filter. */
extern int halide_copy_to_host(void *user_context, struct buffer_t *buf);
/** Copy image data from host memory to device memory. This should not
* be called directly; Halide handles copying to the device
* automatically. If interface is NULL and the bug has a non-zero dev
* field, the device associated with the dev handle will be
* used. Otherwise if the dev field is 0 and interface is NULL, an
* error is returned. */
extern int halide_copy_to_device(void *user_context, struct buffer_t *buf,
const struct halide_device_interface_t *device_interface);
/** Wait for current GPU operations to complete. Calling this explicitly
* should rarely be necessary, except maybe for profiling. */
extern int halide_device_sync(void *user_context, struct buffer_t *buf);
/** Allocate device memory to back a buffer_t. */
extern int halide_device_malloc(void *user_context, struct buffer_t *buf, const struct halide_device_interface_t *device_interface);
/** Free device memory. */
extern int halide_device_free(void *user_context, struct buffer_t *buf);
/** Get a pointer to halide_device_free if a Halide runtime has been
* linked in. Returns null if it has not. This requires a different
* mechanism on different platforms. */
typedef int (*halide_device_free_t)(void *, struct buffer_t *);
#ifdef _MSC_VER
extern const __declspec(selectany) void *halide_dummy_device_free = NULL;
extern int halide_weak_device_free(void *user_context, struct buffer_t *buf);
// The following pragma tells the windows linker to make
// halide_device_free_weak the same symbol as halide_dummy_device_free
// if it can't resolve halide_weak_device_free normally
#ifdef _WIN64
#pragma comment(linker, "/alternatename:halide_weak_device_free=halide_dummy_device_free")
#else
#pragma comment(linker, "/alternatename:_halide_weak_device_free=_halide_dummy_device_free")
#endif
inline halide_device_free_t halide_get_device_free_fn() {
if ((const void **)(&halide_weak_device_free) == &halide_dummy_device_free) {
return NULL;
} else {
return &halide_weak_device_free;
}
};
#elif __MINGW32__
inline halide_device_free_t halide_get_device_free_fn() {
// There is no workable mechanism for doing this that we know of on mingw.
return &halide_device_free;
}
#else
extern __attribute__((weak)) int halide_weak_device_free(void *user_context, struct buffer_t *buf);
inline halide_device_free_t halide_get_device_free_fn() {
return &halide_weak_device_free;
}
#endif
/** Selects which gpu device to use. 0 is usually the display
* device. If never called, Halide uses the environment variable
* HL_GPU_DEVICE. If that variable is unset, Halide uses the last
* device. Set this to -1 to use the last device. */
extern void halide_set_gpu_device(int n);
/** Halide calls this to get the desired halide gpu device
* setting. Implement this yourself to use a different gpu device per
* user_context. The default implementation returns the value set by
* halide_set_gpu_device, or the environment variable
* HL_GPU_DEVICE. */
extern int halide_get_gpu_device(void *user_context);
/** Set the soft maximum amount of memory, in bytes, that the LRU
* cache will use to memoize Func results. This is not a strict
* maximum in that concurrency and simultaneous use of memoized
* reults larger than the cache size can both cause it to
* temporariliy be larger than the size specified here.
*/
extern void halide_memoization_cache_set_size(int64_t size);
/** Given a cache key for a memoized result, currently constructed
* from the Func name and top-level Func name plus the arguments of
* the computation, determine if the result is in the cache and
* return it if so. (The internals of the cache key should be
* considered opaque by this function.) If this routine returns true,
* it is a cache miss. Otherwise, it will return false and the
* buffers passed in will be filled, via copying, with memoized
* data. The last argument is a list if buffer_t pointers which
* represents the outputs of the memoized Func. If the Func does not
* return a Tuple, there will only be one buffer_t in the list. The
* tuple_count parameters determines the length of the list.
*
* The return values are:
* -1: Signals an error.
* 0: Success and cache hit.
* 1: Success and cache miss.
*/
extern int halide_memoization_cache_lookup(void *user_context, const uint8_t *cache_key, int32_t size,
struct buffer_t *realized_bounds, int32_t tuple_count,
struct buffer_t **tuple_buffers);
/** Given a cache key for a memoized result, currently constructed
* from the Func name and top-level Func name plus the arguments of
* the computation, store the result in the cache for futre access by
* halide_memoization_cache_lookup. (The internals of the cache key
* should be considered opaque by this function.) Data is copied out
* from the inputs and inputs are unmodified. The last argument is a
* list if buffer_t pointers which represents the outputs of the
* memoized Func. If the Func does not return a Tuple, there will
* only be one buffer_t in the list. The tuple_count parameters
* determines the length of the list.
*
* If there is a memory allocation failure, the store does not store
* the data into the cache.
*/
extern int halide_memoization_cache_store(void *user_context, const uint8_t *cache_key, int32_t size,
struct buffer_t *realized_bounds, int32_t tuple_count,
struct buffer_t **tuple_buffers);
/** If halide_memoization_cache_lookup succeeds,
* halide_memoization_cache_release must be called to signal the
* storage is no longer being used by the caller. It will be passed
* the host pointer of one the buffers returned by
* halide_memoization_cache_lookup. That is
* halide_memoization_cache_release will be called multiple times for
* the case where halide_memoization_cache_lookup is handling multiple
* buffers. (This corresponds to memoizing a Tuple in Halide.) Note
* that the host pointer must be sufficient to get to all information
* the relase operation needs. The default Halide cache impleemntation
* accomplishes this by storing extra data before the start of the user
* modifiable host storage.
*
* This call is like free and does not have a failure return.
*/
extern void halide_memoization_cache_release(void *user_context, void *host);
/** Free all memory and resources associated with the memoization cache.
* Must be called at a time when no other threads are accessing the cache.
*/
extern void halide_memoization_cache_cleanup();
/** Create a unique file with a name of the form prefixXXXXXsuffix in an arbitrary
* (but writable) directory; this is typically $TMP or /tmp, but the specific
* location is not guaranteed. (Note that the exact form of the file name
* may vary; in particular, the suffix may be ignored on non-Posix systems.)
* The file is created (but not opened), thus this can be called from
* different threads (or processes, e.g. when building with parallel make)
* without risking collision. Note that the caller is always responsible
* for deleting this file. Returns nonzero value if an error occurs.
*/
extern int halide_create_temp_file(void *user_context,
const char *prefix, const char *suffix,
char *path_buf, size_t path_buf_size);
/** Annotate that a given range of memory has been initialized;
* only used when Target::MSAN is enabled.
*
* The default implementation uses the LLVM-provided AnnotateMemoryIsInitialized() function.
*/
extern void halide_msan_annotate_memory_is_initialized(void *user_context, const void *ptr, uint64_t len);
/** Mark the data pointed to by the buffer_t as initialized (but *not* the buffer_t itself),
* using halide_msan_annotate_memory_is_initialized() for marking.
*
* The default implementation takes pains to only mark the active memory ranges
* (skipping padding), and sorting into ranges to always mark the smallest number of
* ranges, in monotonically increasing memory order.
*
* Most client code should never need to replace the default implementation.
*/
extern void halide_msan_annotate_buffer_is_initialized(void *user_context, struct buffer_t *buffer);
extern void halide_msan_annotate_buffer_is_initialized_as_destructor(void *user_context, void *buffer);
/** The error codes that may be returned by a Halide pipeline. */
enum halide_error_code_t {
/** There was no error. This is the value returned by Halide on success. */
halide_error_code_success = 0,
/** An uncategorized error occurred. Refer to the string passed to halide_error. */
halide_error_code_generic_error = -1,
/** A Func was given an explicit bound via Func::bound, but this
* was not large enough to encompass the region that is used of
* the Func by the rest of the pipeline. */
halide_error_code_explicit_bounds_too_small = -2,
/** The elem_size field of a buffer_t does not match the size in
* bytes of the type of that ImageParam. Probable type mismatch. */
halide_error_code_bad_elem_size = -3,
/** A pipeline would access memory outside of the buffer_t passed
* in. */
halide_error_code_access_out_of_bounds = -4,
/** A buffer_t was given that spans more than 2GB of memory. */
halide_error_code_buffer_allocation_too_large = -5,
/** A buffer_t was given with extents that multiply to a number
* greater than 2^31-1 */
halide_error_code_buffer_extents_too_large = -6,
/** Applying explicit constraints on the size of an input or
* output buffer shrank the size of that buffer below what will be
* accessed by the pipeline. */
halide_error_code_constraints_make_required_region_smaller = -7,
/** A constraint on a size or stride of an input or output buffer
* was not met by the buffer_t passed in. */
halide_error_code_constraint_violated = -8,
/** A scalar parameter passed in was smaller than its minimum
* declared value. */
halide_error_code_param_too_small = -9,
/** A scalar parameter passed in was greater than its minimum
* declared value. */
halide_error_code_param_too_large = -10,
/** A call to halide_malloc returned NULL. */
halide_error_code_out_of_memory = -11,
/** A buffer_t pointer passed in was NULL. */
halide_error_code_buffer_argument_is_null = -12,
/** debug_to_file failed to open or write to the specified
* file. */
halide_error_code_debug_to_file_failed = -13,
/** The Halide runtime encountered an error while trying to copy
* from device to host. Turn on -debug in your target string to
* see more details. */
halide_error_code_copy_to_host_failed = -14,
/** The Halide runtime encountered an error while trying to copy
* from host to device. Turn on -debug in your target string to
* see more details. */
halide_error_code_copy_to_device_failed = -15,
/** The Halide runtime encountered an error while trying to
* allocate memory on device. Turn on -debug in your target string
* to see more details. */
halide_error_code_device_malloc_failed = -16,
/** The Halide runtime encountered an error while trying to
* synchronize with a device. Turn on -debug in your target string
* to see more details. */
halide_error_code_device_sync_failed = -17,
/** The Halide runtime encountered an error while trying to free a
* device allocation. Turn on -debug in your target string to see
* more details. */
halide_error_code_device_free_failed = -18,
/** A device operation was attempted on a buffer with no device
* interface. */
halide_error_code_no_device_interface = -19,
/** An error occurred when attempting to initialize the Matlab
* runtime. */
halide_error_code_matlab_init_failed = -20,
/** The type of an mxArray did not match the expected type. */
halide_error_code_matlab_bad_param_type = -21,
/** There is a bug in the Halide compiler. */
halide_error_code_internal_error = -22,
/** The Halide runtime encountered an error while trying to launch
* a GPU kernel. Turn on -debug in your target string to see more
* details. */
halide_error_code_device_run_failed = -23,
/** The Halide runtime encountered a host pointer that violated
* the alignment set for it by way of a call to
* set_host_alignment */
halide_error_code_unaligned_host_ptr = -24,
/** A fold_storage directive was used on a dimension that is not
* accessed in a monotonically increasing or decreasing fashion. */
halide_error_code_bad_fold = -25,
/** A fold_storage directive was used with a fold factor that was
* too small to store all the values of a producer needed by the
* consumer. */
halide_error_code_fold_factor_too_small = -26,
/** User-specified require() expression was not satisfied. */
halide_error_code_requirement_failed = -27,
};
/** Halide calls the functions below on various error conditions. The
* default implementations construct an error message, call
* halide_error, then return the matching error code above. On
* platforms that support weak linking, you can override these to
* catch the errors individually. */
/** A call into an extern stage for the purposes of bounds inference
* failed. Returns the error code given by the extern stage. */
extern int halide_error_bounds_inference_call_failed(void *user_context, const char *extern_stage_name, int result);
/** A call to an extern stage failed. Returned the error code given by
* the extern stage. */
extern int halide_error_extern_stage_failed(void *user_context, const char *extern_stage_name, int result);
/** Various other error conditions. See the enum above for a
* description of each. */
// @{
extern 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);
extern int halide_error_bad_elem_size(void *user_context, const char *func_name,
const char *type_name, int elem_size_given, int correct_elem_size);
extern 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);
extern int halide_error_buffer_allocation_too_large(void *user_context, const char *buffer_name,
uint64_t allocation_size, uint64_t max_size);
extern int halide_error_buffer_extents_too_large(void *user_context, const char *buffer_name,
int64_t actual_size, int64_t max_size);
extern 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);
extern int halide_error_constraint_violated(void *user_context, const char *var, int val,
const char *constrained_var, int constrained_val);
extern int halide_error_param_too_small_i64(void *user_context, const char *param_name,
int64_t val, int64_t min_val);
extern int halide_error_param_too_small_u64(void *user_context, const char *param_name,
uint64_t val, uint64_t min_val);
extern int halide_error_param_too_small_f64(void *user_context, const char *param_name,
double val, double min_val);
extern int halide_error_param_too_large_i64(void *user_context, const char *param_name,
int64_t val, int64_t max_val);
extern int halide_error_param_too_large_u64(void *user_context, const char *param_name,
uint64_t val, uint64_t max_val);
extern int halide_error_param_too_large_f64(void *user_context, const char *param_name,
double val, double max_val);
extern int halide_error_out_of_memory(void *user_context);
extern int halide_error_buffer_argument_is_null(void *user_context, const char *buffer_name);
extern int halide_error_debug_to_file_failed(void *user_context, const char *func,
const char *filename, int error_code);
extern int halide_error_unaligned_host_ptr(void *user_context, const char *func_name, int alignment);
extern int halide_error_bad_fold(void *user_context, const char *func_name, const char *var_name,
const char *loop_name);
extern 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);
extern int halide_error_requirement_failed(void *user_context, const char *condition, const char *message);
// @}
/** Optional features a compilation Target can have.
*/
typedef enum halide_target_feature_t {
halide_target_feature_jit = 0, ///< Generate code that will run immediately inside the calling process.
halide_target_feature_debug = 1, ///< Turn on debug info and output for runtime code.
halide_target_feature_no_asserts = 2, ///< Disable all runtime checks, for slightly tighter code.
halide_target_feature_no_bounds_query = 3, ///< Disable the bounds querying functionality.
halide_target_feature_sse41 = 4, ///< Use SSE 4.1 and earlier instructions. Only relevant on x86.
halide_target_feature_avx = 5, ///< Use AVX 1 instructions. Only relevant on x86.
halide_target_feature_avx2 = 6, ///< Use AVX 2 instructions. Only relevant on x86.
halide_target_feature_fma = 7, ///< Enable x86 FMA instruction
halide_target_feature_fma4 = 8, ///< Enable x86 (AMD) FMA4 instruction set
halide_target_feature_f16c = 9, ///< Enable x86 16-bit float support
halide_target_feature_armv7s = 10, ///< Generate code for ARMv7s. Only relevant for 32-bit ARM.
halide_target_feature_no_neon = 11, ///< Avoid using NEON instructions. Only relevant for 32-bit ARM.
halide_target_feature_vsx = 12, ///< Use VSX instructions. Only relevant on POWERPC.
halide_target_feature_power_arch_2_07 = 13, ///< Use POWER ISA 2.07 new instructions. Only relevant on POWERPC.
halide_target_feature_cuda = 14, ///< Enable the CUDA runtime. Defaults to compute capability 2.0 (Fermi)
halide_target_feature_cuda_capability30 = 15, ///< Enable CUDA compute capability 3.0 (Kepler)
halide_target_feature_cuda_capability32 = 16, ///< Enable CUDA compute capability 3.2 (Tegra K1)
halide_target_feature_cuda_capability35 = 17, ///< Enable CUDA compute capability 3.5 (Kepler)
halide_target_feature_cuda_capability50 = 18, ///< Enable CUDA compute capability 5.0 (Maxwell)
halide_target_feature_opencl = 19, ///< Enable the OpenCL runtime.
halide_target_feature_cl_doubles = 20, ///< Enable double support on OpenCL targets
halide_target_feature_opengl = 21, ///< Enable the OpenGL runtime.
halide_target_feature_openglcompute = 22, ///< Enable OpenGL Compute runtime.
halide_target_feature_unused_23 = 23, ///< Unused. (Formerly: Enable the RenderScript runtime.)
halide_target_feature_user_context = 24, ///< Generated code takes a user_context pointer as first argument
halide_target_feature_matlab = 25, ///< Generate a mexFunction compatible with Matlab mex libraries. See tools/mex_halide.m.
halide_target_feature_profile = 26, ///< Launch a sampling profiler alongside the Halide pipeline that monitors and reports the runtime used by each Func
halide_target_feature_no_runtime = 27, ///< Do not include a copy of the Halide runtime in any generated object file or assembly
halide_target_feature_metal = 28, ///< Enable the (Apple) Metal runtime.
halide_target_feature_mingw = 29, ///< For Windows compile to MinGW toolset rather then Visual Studio
halide_target_feature_c_plus_plus_mangling = 30, ///< Generate C++ mangled names for result function, et al
halide_target_feature_large_buffers = 31, ///< Enable 64-bit buffer indexing to support buffers > 2GB.
halide_target_feature_hvx_64 = 32, ///< Enable HVX 64 byte mode.
halide_target_feature_hvx_128 = 33, ///< Enable HVX 128 byte mode.
halide_target_feature_hvx_v62 = 34, ///< Enable Hexagon v62 architecture.
halide_target_feature_fuzz_float_stores = 35, ///< 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.
halide_target_feature_soft_float_abi = 36, ///< Enable soft float ABI. This only enables the soft float ABI calling convention, which does not necessarily use soft floats.
halide_target_feature_msan = 37, ///< Enable hooks for MSAN support.
halide_target_feature_avx512 = 38, ///< 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.
halide_target_feature_avx512_knl = 39, ///< 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.
halide_target_feature_avx512_skylake = 40, ///< 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.
halide_target_feature_avx512_cannonlake = 41, ///< 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.
halide_target_feature_end = 42 ///< A sentinel. Every target is considered to have this feature, and setting this feature does nothing.
} halide_target_feature_t;
/** This function is called internally by Halide in some situations to determine
* if the current execution environment can support the given set of
* halide_target_feature_t flags. The implementation must do the following:
*
* -- If there are flags set in features that the function knows *cannot* be supported, return 0.
* -- Otherwise, return 1.
* -- Note that any flags set in features that the function doesn't know how to test should be ignored;
* this implies that a return value of 1 means "not known to be bad" rather than "known to be good".
*
* In other words: a return value of 0 means "It is not safe to use code compiled with these features",
* while a return value of 1 means "It is not obviously unsafe to use code compiled with these features".
*
* The default implementation simply calls halide_default_can_use_target_features.
*/
// @{
extern int halide_can_use_target_features(uint64_t features);
typedef int (*halide_can_use_target_features_t)(uint64_t);
extern halide_can_use_target_features_t halide_set_custom_can_use_target_features(halide_can_use_target_features_t);
// @}
/**
* This is the default implementation of halide_can_use_target_features; it is provided
* for convenience of user code that may wish to extend halide_can_use_target_features
* but continue providing existing support, e.g.
*
* int halide_can_use_target_features(uint64_t features) {
* if (features & halide_target_somefeature) {
* if (!can_use_somefeature()) {
* return 0;
* }
* }
* return halide_default_can_use_target_features(features);
* }
*/
extern int halide_default_can_use_target_features(uint64_t features);
#ifndef HALIDE_ATTRIBUTE_DEPRECATED
#ifdef HALIDE_ALLOW_DEPRECATED
#define HALIDE_ATTRIBUTE_DEPRECATED(x)
#else
#ifdef _MSC_VER
#define HALIDE_ATTRIBUTE_DEPRECATED(x) __declspec(deprecated(x))
#else
#define HALIDE_ATTRIBUTE_DEPRECATED(x) __attribute__((deprecated(x)))
#endif
#endif
#endif
#ifndef BUFFER_T_DEFINED
#define BUFFER_T_DEFINED
/**
* The raw representation of an image passed around by generated
* Halide code. It includes some stuff to track whether the image is
* not actually in main memory, but instead on a device (like a
* GPU). */
typedef struct buffer_t {
/** A device-handle for e.g. GPU memory used to back this buffer. */
uint64_t dev;
/** A pointer to the start of the data in main memory. In terms of
* the Halide coordinate system, this is the address of the min
* coordinates (defined below). */
uint8_t* host;
/** The size of the buffer in each dimension. */
int32_t extent[4];
/** Gives the spacing in memory between adjacent elements in the
* given dimension. The correct memory address for a load from
* this buffer at position x, y, z, w is:
* host + elem_size * ((x - min[0]) * stride[0] +
* (y - min[1]) * stride[1] +
* (z - min[2]) * stride[2] +
* (w - min[3]) * stride[3])
* By manipulating the strides and extents you can lazily crop,
* transpose, and even flip buffers without modifying the data.
*/
int32_t stride[4];
/** Buffers often represent evaluation of a Func over some
* domain. The min field encodes the top left corner of the
* domain. */
int32_t min[4];
/** How many bytes does each buffer element take. This may be
* replaced with a more general type code in the future. */
int32_t elem_size;
/** This should be true if there is an existing device allocation
* mirroring this buffer, and the data has been modified on the
* host side. */
HALIDE_ATTRIBUTE_ALIGN(1) bool host_dirty;
/** This should be true if there is an existing device allocation
mirroring this buffer, and the data has been modified on the
device side. */
HALIDE_ATTRIBUTE_ALIGN(1) bool dev_dirty;
// Some compilers will add extra padding at the end to ensure
// the size is a multiple of 8; we'll do that explicitly so that
// there is no ambiguity.
HALIDE_ATTRIBUTE_ALIGN(1) uint8_t _padding[10 - sizeof(void *)];
} buffer_t;
#endif
/** halide_scalar_value_t is a simple union able to represent all the well-known
* scalar values in a filter argument. Note that it isn't tagged with a type;
* you must ensure you know the proper type before accessing. Most user
* code will never need to create instances of this struct; its primary use
* is to hold def/min/max values in a halide_filter_argument_t. (Note that
* this is conceptually just a union; it's wrapped in a struct to ensure
* that it doesn't get anonymized by LLVM.)
*/
struct halide_scalar_value_t {
union {
bool b;
int8_t i8;
int16_t i16;
int32_t i32;
int64_t i64;
uint8_t u8;
uint16_t u16;
uint32_t u32;
uint64_t u64;
float f32;
double f64;
void *handle;
} u;
};
enum halide_argument_kind_t {
halide_argument_kind_input_scalar = 0,
halide_argument_kind_input_buffer = 1,
halide_argument_kind_output_buffer = 2
};
/*
These structs must be robust across different compilers and settings; when
modifying them, strive for the following rules:
1) All fields are explicitly sized. I.e. must use int32_t and not "int"
2) All fields must land on an alignment boundary that is the same as their size
3) Explicit padding is added to make that so
4) The sizeof the struct is padded out to a multiple of the largest natural size thing in the struct
5) don't forget that 32 and 64 bit pointers are different sizes
*/
/**
* halide_filter_argument_t is essentially a plain-C-struct equivalent to
* Halide::Argument; most user code will never need to create one.
*/
struct halide_filter_argument_t {
const char *name; // name of the argument; will never be null or empty.
int32_t kind; // actually halide_argument_kind_t
int32_t dimensions; // always zero for scalar arguments
struct halide_type_t type;
// These pointers should always be null for buffer arguments,
// and *may* be null for scalar arguments. (A null value means
// there is no def/min/max specified for this argument.)
const struct halide_scalar_value_t *def;
const struct halide_scalar_value_t *min;
const struct halide_scalar_value_t *max;
};
struct halide_filter_metadata_t {
/** version of this metadata; currently always 0. */
int32_t version;
/** The number of entries in the arguments field. This is always >= 1. */
int32_t num_arguments;
/** An array of the filters input and output arguments; this will never be
* null. The order of arguments is not guaranteed (input and output arguments
* may come in any order); however, it is guaranteed that all arguments
* will have a unique name within a given filter. */
const struct halide_filter_argument_t* arguments;
/** The Target for which the filter was compiled. This is always
* a canonical Target string (ie a product of Target::to_string). */
const char* target;
/** The function name of the filter. */
const char* name;
};
/** The functions below here are relevant for pipelines compiled with
* the -profile target flag, which runs a sampling profiler thread
* alongside the pipeline. */
/** Per-Func state tracked by the sampling profiler. */
struct halide_profiler_func_stats {
/** Total time taken evaluating this Func (in nanoseconds). */
uint64_t time;
/** The current memory allocation of this Func. */
uint64_t memory_current;
/** The peak memory allocation of this Func. */
uint64_t memory_peak;
/** The total memory allocation of this Func. */
uint64_t memory_total;
/** The peak stack allocation of this Func's threads. */
uint64_t stack_peak;
/** The average number of thread pool worker threads active while computing this Func. */
uint64_t active_threads_numerator, active_threads_denominator;
/** The name of this Func. A global constant string. */
const char *name;
/** The total number of memory allocation of this Func. */
int num_allocs;
};
/** Per-pipeline state tracked by the sampling profiler. These exist
* in a linked list. */
struct halide_profiler_pipeline_stats {
/** Total time spent inside this pipeline (in nanoseconds) */
uint64_t time;
/** The current memory allocation of funcs in this pipeline. */
uint64_t memory_current;
/** The peak memory allocation of funcs in this pipeline. */
uint64_t memory_peak;
/** The total memory allocation of funcs in this pipeline. */
uint64_t memory_total;
/** The average number of thread pool worker threads doing useful
* work while computing this pipeline. */
uint64_t active_threads_numerator, active_threads_denominator;
/** The name of this pipeline. A global constant string. */
const char *name;
/** An array containing states for each Func in this pipeline. */
struct halide_profiler_func_stats *funcs;
/** The next pipeline_stats pointer. It's a void * because types
* in the Halide runtime may not currently be recursive. */
void *next;
/** The number of funcs in this pipeline. */
int num_funcs;
/** An internal base id used to identify the funcs in this pipeline. */
int first_func_id;
/** The number of times this pipeline has been run. */
int runs;
/** The total number of samples taken inside of this pipeline. */
int samples;
/** The total number of memory allocation of funcs in this pipeline. */
int num_allocs;
};
/** The global state of the profiler. */
struct halide_profiler_state {
/** Guards access to the fields below. If not locked, the sampling
* profiler thread is free to modify things below (including
* reordering the linked list of pipeline stats). */
struct halide_mutex lock;
/** The amount of time the profiler thread sleeps between samples
* in milliseconds. Defaults to 1 */
int sleep_time;
/** An internal id used for bookkeeping. */
int first_free_id;
/** The id of the current running Func. Set by the pipeline, read
* periodically by the profiler thread. */
int current_func;
/** The number of threads currently doing work. */
int active_threads;
/** A linked list of stats gathered for each pipeline. */
struct halide_profiler_pipeline_stats *pipelines;
/** Retrieve remote profiler state. Used so that the sampling
* profiler can follow along with execution that occurs elsewhere,
* e.g. on a DSP. If null, it reads from the int above instead. */
void (*get_remote_profiler_state)(int *func, int *active_workers);
/** Is the profiler thread running. */
bool started;
};
/** Profiler func ids with special meanings. */
enum {
/// current_func takes on this value when not inside Halide code
halide_profiler_outside_of_halide = -1,
/// Set current_func to this value to tell the profiling thread to
/// halt. It will start up again next time you run a pipeline with
/// profiling enabled.
halide_profiler_please_stop = -2
};
/** Get a pointer to the global profiler state for programmatic
* inspection. Lock it before using to pause the profiler. */
extern struct halide_profiler_state *halide_profiler_get_state();
/** Get a pointer to the pipeline state associated with pipeline_name.
* This function grabs the global profiler state's lock on entry. */
extern struct halide_profiler_pipeline_stats *halide_profiler_get_pipeline_state(const char *pipeline_name);
/** Reset all profiler state.
* WARNING: Do NOT call this method while any halide pipeline is
* running; halide_profiler_memory_allocate/free and
* halide_profiler_stack_peak_update update the profiler pipeline's
* state without grabbing the global profiler state's lock. */
extern void halide_profiler_reset();
/** Print out timing statistics for everything run since the last
* reset. Also happens at process exit. */
extern void halide_profiler_report(void *user_context);
/// \name "Float16" functions
/// These functions operate of bits (``uint16_t``) representing a half
/// precision floating point number (IEEE-754 2008 binary16).
//{@
/** Read bits representing a half precision floating point number and return
* the float that represents the same value */
extern float halide_float16_bits_to_float(uint16_t);
/** Read bits representing a half precision floating point number and return
* the double that represents the same value */
extern double halide_float16_bits_to_double(uint16_t);
// TODO: Conversion functions to half
//@}
#ifdef __cplusplus
} // End extern "C"
#endif
#ifdef __cplusplus
namespace {
template<typename T>
struct halide_type_of_helper;
template<typename T>
struct halide_type_of_helper<T *> {
operator halide_type_t() {
return halide_type_t(halide_type_handle, 64);
}
};
template<typename T>
struct halide_type_of_helper<T &> {
operator halide_type_t() {
return halide_type_t(halide_type_handle, 64);
}
};
// Halide runtime does not require C++11
#if __cplusplus > 199711L
template<typename T>
struct halide_type_of_helper<T &&> {
operator halide_type_t() {
return halide_type_t(halide_type_handle, 64);
}
};
#endif
template<>
struct halide_type_of_helper<float> {
operator halide_type_t() { return halide_type_t(halide_type_float, 32); }
};
template<>
struct halide_type_of_helper<double> {
operator halide_type_t() { return halide_type_t(halide_type_float, 64); }
};
template<>
struct halide_type_of_helper<uint8_t> {
operator halide_type_t() { return halide_type_t(halide_type_uint, 8); }
};
template<>
struct halide_type_of_helper<uint16_t> {
operator halide_type_t() { return halide_type_t(halide_type_uint, 16); }
};
template<>
struct halide_type_of_helper<uint32_t> {
operator halide_type_t() { return halide_type_t(halide_type_uint, 32); }
};
template<>
struct halide_type_of_helper<uint64_t> {
operator halide_type_t() { return halide_type_t(halide_type_uint, 64); }
};
template<>
struct halide_type_of_helper<int8_t> {
operator halide_type_t() { return halide_type_t(halide_type_int, 8); }
};
template<>
struct halide_type_of_helper<int16_t> {
operator halide_type_t() { return halide_type_t(halide_type_int, 16); }
};
template<>
struct halide_type_of_helper<int32_t> {
operator halide_type_t() { return halide_type_t(halide_type_int, 32); }
};
template<>
struct halide_type_of_helper<int64_t> {
operator halide_type_t() { return halide_type_t(halide_type_int, 64); }
};
template<>
struct halide_type_of_helper<bool> {
operator halide_type_t() { return halide_type_t(halide_type_uint, 1); }
};
}
/** Construct the halide equivalent of a C type */
template<typename T> halide_type_t halide_type_of() {
return halide_type_of_helper<T>();
}
#endif
#endif // HALIDE_HALIDERUNTIME_H
