Source code for arrayfire.device

#######################################################
# Copyright (c) 2015, ArrayFire
# All rights reserved.
#
# This file is distributed under 3-clause BSD license.
# The complete license agreement can be obtained at:
# http://arrayfire.com/licenses/BSD-3-Clause
########################################################
"""
Functions to handle the available devices in the backend.
"""

from .library import *
from .util import (safe_call, to_str, get_version)

[docs]def init(): """ Note ----- This function may need to be called when interoperating with other libraries """ safe_call(backend.get().af_init())
[docs]def info(): """ Displays the information about the following: - ArrayFire build and version number. - The number of devices available. - The names of the devices. - The current device being used. """ safe_call(backend.get().af_info())
[docs]def device_info(): """ Returns a map with the following fields: - 'device': Name of the current device. - 'backend': The current backend being used. - 'toolkit': The toolkit version for the backend. - 'compute': The compute version of the device. """ c_char_256 = c_char_t * 256 device_name = c_char_256() backend_name = c_char_256() toolkit = c_char_256() compute = c_char_256() safe_call(backend.get().af_device_info(c_pointer(device_name), c_pointer(backend_name), c_pointer(toolkit), c_pointer(compute))) dev_info = {} dev_info['device'] = to_str(device_name) dev_info['backend'] = to_str(backend_name) dev_info['toolkit'] = to_str(toolkit) dev_info['compute'] = to_str(compute) return dev_info
[docs]def get_device_count(): """ Returns the number of devices available. """ c_num = c_int_t(0) safe_call(backend.get().af_get_device_count(c_pointer(c_num))) return c_num.value
[docs]def get_device(): """ Returns the id of the current device. """ c_dev = c_int_t(0) safe_call(backend.get().af_get_device(c_pointer(c_dev))) return c_dev.value
[docs]def set_device(num): """ Change the active device to the specified id. Parameters ----------- num: int. id of the desired device. """ safe_call(backend.get().af_set_device(num))
[docs]def info_str(verbose = False): """ Returns information about the following as a string: - ArrayFire version number. - The number of devices available. - The names of the devices. - The current device being used. """ import platform res_str = 'ArrayFire' major, minor, patch = get_version() dev_info = device_info() backend_str = dev_info['backend'] res_str += ' v' + str(major) + '.' + str(minor) + '.' + str(patch) res_str += ' (' + backend_str + ' ' + platform.architecture()[0] + ')\n' num_devices = get_device_count() curr_device_id = get_device() for n in range(num_devices): # To suppress warnings on CPU if (n != curr_device_id): set_device(n) if (n == curr_device_id): res_str += '[%d] ' % n else: res_str += '-%d- ' % n dev_info = device_info() if (backend_str.lower() == 'opencl'): res_str += dev_info['toolkit'] res_str += ': ' + dev_info['device'] if (backend_str.lower() != 'cpu'): res_str += ' (Compute ' + dev_info['compute'] + ')' res_str += '\n' # To suppress warnings on CPU if (curr_device_id != get_device()): set_device(curr_device_id) return res_str
[docs]def is_dbl_supported(device=None): """ Check if double precision is supported on specified device. Parameters ----------- device: optional: int. default: None. id of the desired device. Returns -------- - True if double precision supported. - False if double precision not supported. """ dev = device if device is not None else get_device() res = c_bool_t(False) safe_call(backend.get().af_get_dbl_support(c_pointer(res), dev)) return res.value
[docs]def is_half_supported(device=None): """ Check if half precision is supported on specified device. Parameters ----------- device: optional: int. default: None. id of the desired device. Returns -------- - True if half precision supported. - False if half precision not supported. """ dev = device if device is not None else get_device() res = c_bool_t(False) safe_call(backend.get().af_get_half_support(c_pointer(res), dev)) return res.value
[docs]def sync(device=None): """ Block until all the functions on the device have completed execution. Parameters ----------- device: optional: int. default: None. id of the desired device. """ dev = device if device is not None else get_device() safe_call(backend.get().af_sync(dev))
def __eval(*args): nargs = len(args) if (nargs == 1): safe_call(backend.get().af_eval(args[0].arr)) else: c_void_p_n = c_void_ptr_t * nargs arrs = c_void_p_n() for n in range(nargs): arrs[n] = args[n].arr safe_call(backend.get().af_eval_multiple(c_int_t(nargs), c_pointer(arrs))) return
[docs]def eval(*args): """ Evaluate one or more inputs together Parameters ----------- args : arguments to be evaluated Note ----- All the input arrays to this function should be of the same size. Examples -------- >>> a = af.constant(1, 3, 3) >>> b = af.constant(2, 3, 3) >>> c = a + b >>> d = a - b >>> af.eval(c, d) # A single kernel is launched here >>> c arrayfire.Array() Type: float [3 3 1 1] 3.0000 3.0000 3.0000 3.0000 3.0000 3.0000 3.0000 3.0000 3.0000 >>> d arrayfire.Array() Type: float [3 3 1 1] -1.0000 -1.0000 -1.0000 -1.0000 -1.0000 -1.0000 -1.0000 -1.0000 -1.0000 """ for arg in args: if not isinstance(arg, Array): raise RuntimeError("All inputs to eval must be of type arrayfire.Array") __eval(*args)
[docs]def set_manual_eval_flag(flag): """ Tells the backend JIT engine to disable heuristics for determining when to evaluate a JIT tree. Parameters ---------- flag : optional: bool. - Specifies if the heuristic evaluation of the JIT tree needs to be disabled. Note ---- This does not affect the evaluation that occurs when a non JIT function forces the evaluation. """ safe_call(backend.get().af_set_manual_eval_flag(flag))
[docs]def get_manual_eval_flag(): """ Query the backend JIT engine to see if the user disabled heuristic evaluation of the JIT tree. Note ---- This does not affect the evaluation that occurs when a non JIT function forces the evaluation. """ res = c_bool_t(False) safe_call(backend.get().af_get_manual_eval_flag(c_pointer(res))) return res.value
[docs]def device_mem_info(): """ Returns a map with the following fields: - 'alloc': Contains the map of the following - 'buffers' : Total number of buffers allocated by memory manager. - 'bytes' : Total number of bytes allocated by memory manager. - 'lock': Contains the map of the following - 'buffers' : Total number of buffers currently in scope. - 'bytes' : Total number of bytes currently in scope. Note ----- ArrayFire does not free memory when array goes out of scope. The memory is marked for reuse. - The difference between alloc buffers and lock buffers equals the number of free buffers. - The difference between alloc bytes and lock bytes equals the number of free bytes. """ alloc_bytes = c_size_t(0) alloc_buffers = c_size_t(0) lock_bytes = c_size_t(0) lock_buffers = c_size_t(0) safe_call(backend.get().af_device_mem_info(c_pointer(alloc_bytes), c_pointer(alloc_buffers), c_pointer(lock_bytes), c_pointer(lock_buffers))) mem_info = {} mem_info['alloc'] = {'buffers' : alloc_buffers.value, 'bytes' : alloc_bytes.value} mem_info['lock'] = {'buffers' : lock_buffers.value, 'bytes' : lock_bytes.value} return mem_info
[docs]def device_gc(): """ Ask the garbage collector to free all unlocked memory """ safe_call(backend.get().af_device_gc())
[docs]def get_device_ptr(a): """ Get the raw device pointer of an array Parameters ---------- a: af.Array - A multi dimensional arrayfire array. Returns ------- - internal device pointer held by a Note ----- - The device pointer of `a` is not freed by memory manager until `unlock_device_ptr()` is called. - This function enables the user to interoperate arrayfire with other CUDA/OpenCL/C libraries. """ ptr = c_void_ptr_t(0) safe_call(backend.get().af_get_device_ptr(c_pointer(ptr), a.arr)) return ptr
[docs]def lock_device_ptr(a): """ This functions is deprecated. Please use lock_array instead. """ import warnings warnings.warn("This function is deprecated. Use lock_array instead.", DeprecationWarning) lock_array(a)
[docs]def lock_array(a): """ Ask arrayfire to not perform garbage collection on raw data held by an array. Parameters ---------- a: af.Array - A multi dimensional arrayfire array. Note ----- - The device pointer of `a` is not freed by memory manager until `unlock_array()` is called. """ safe_call(backend.get().af_lock_array(a.arr))
[docs]def is_locked_array(a): """ Check if the input array is locked by the user. Parameters ---------- a: af.Array - A multi dimensional arrayfire array. Returns ----------- A bool specifying if the input array is locked. """ res = c_bool_t(False) safe_call(backend.get().af_is_locked_array(c_pointer(res), a.arr)) return res.value
[docs]def unlock_device_ptr(a): """ This functions is deprecated. Please use unlock_array instead. """ import warnings warnings.warn("This function is deprecated. Use unlock_array instead.", DeprecationWarning) unlock_array(a)
[docs]def unlock_array(a): """ Tell arrayfire to resume garbage collection on raw data held by an array. Parameters ---------- a: af.Array - A multi dimensional arrayfire array. """ safe_call(backend.get().af_unlock_array(a.arr))
[docs]def alloc_device(num_bytes): """ Allocate a buffer on the device with specified number of bytes. """ ptr = c_void_ptr_t(0) c_num_bytes = c_dim_t(num_bytes) safe_call(backend.get().af_alloc_device(c_pointer(ptr), c_num_bytes)) return ptr.value
[docs]def alloc_host(num_bytes): """ Allocate a buffer on the host with specified number of bytes. """ ptr = c_void_ptr_t(0) c_num_bytes = c_dim_t(num_bytes) safe_call(backend.get().af_alloc_host(c_pointer(ptr), c_num_bytes)) return ptr.value
[docs]def alloc_pinned(num_bytes): """ Allocate a buffer on the host using pinned memory with specified number of bytes. """ ptr = c_void_ptr_t(0) c_num_bytes = c_dim_t(num_bytes) safe_call(backend.get().af_alloc_pinned(c_pointer(ptr), c_num_bytes)) return ptr.value
[docs]def free_device(ptr): """ Free the device memory allocated by alloc_device """ cptr = c_void_ptr_t(ptr) safe_call(backend.get().af_free_device(cptr))
[docs]def free_host(ptr): """ Free the host memory allocated by alloc_host """ cptr = c_void_ptr_t(ptr) safe_call(backend.get().af_free_host(cptr))
[docs]def free_pinned(ptr): """ Free the pinned memory allocated by alloc_pinned """ cptr = c_void_ptr_t(ptr) safe_call(backend.get().af_free_pinned(cptr))
from .array import Array