1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
164
165
166
167
168
169
170
171
172
173
174
175
176
177
178
179
180
181
182
183
184
185
186
187
188
189
190
191
192
193
194
195
196
197
198
199
200
201
202
203
204
205
206
207
208
209
210
211
212
213
214
215
216
217
218
219
220
221
222
|
import numpy as np
import pytools
import pycuda.tools
from pycuda import characterize
import pycuda.driver as cuda
import pycuda.compiler
from pycuda import gpuarray as ga
from chroma.cuda import srcdir
# standard nvcc options
cuda_options = ('--use_fast_math',)#, '--ptxas-options=-v']
@pycuda.tools.context_dependent_memoize
def get_cu_module(name, options=None, include_source_directory=True):
"""Returns a pycuda.compiler.SourceModule object from a CUDA source file
located in the chroma cuda directory at cuda/[name]."""
if options is None:
options = []
elif isinstance(options, tuple):
options = list(options)
else:
raise TypeError('`options` must be a tuple.')
if include_source_directory:
options += ['-I' + srcdir]
with open('%s/%s' % (srcdir, name)) as f:
source = f.read()
return pycuda.compiler.SourceModule(source, options=options,
no_extern_c=True)
@pytools.memoize
def get_cu_source(name):
"""Get the source code for a CUDA source file located in the chroma cuda
directory at src/[name]."""
with open('%s/%s' % (srcdir, name)) as f:
source = f.read()
return source
class GPUFuncs(object):
"""Simple container class for GPU functions as attributes."""
def __init__(self, module):
self.module = module
self.funcs = {}
def __getattr__(self, name):
try:
return self.funcs[name]
except KeyError:
f = self.module.get_function(name)
self.funcs[name] = f
return f
init_rng_src = """
#include <curand_kernel.h>
extern "C"
{
__global__ void init_rng(int nthreads, curandState *s, unsigned long long seed, unsigned long long offset)
{
int id = blockIdx.x*blockDim.x + threadIdx.x;
if (id >= nthreads)
return;
curand_init(seed, id, offset, &s[id]);
}
} // extern "C"
"""
def get_rng_states(size, seed=1):
"Return `size` number of CUDA random number generator states."
rng_states = cuda.mem_alloc(size*characterize.sizeof('curandStateXORWOW', '#include <curand_kernel.h>'))
module = pycuda.compiler.SourceModule(init_rng_src, no_extern_c=True)
init_rng = module.get_function('init_rng')
init_rng(np.int32(size), rng_states, np.uint64(seed), np.uint64(0), block=(64,1,1), grid=(size//64+1,1))
return rng_states
def to_float3(arr):
"Returns an pycuda.gpuarray.vec.float3 array from an (N,3) array."
if not arr.flags['C_CONTIGUOUS']:
arr = np.asarray(arr, order='c')
return arr.astype(np.float32).view(ga.vec.float3)[:,0]
def to_uint3(arr):
"Returns a pycuda.gpuarray.vec.uint3 array from an (N,3) array."
if not arr.flags['C_CONTIGUOUS']:
arr = np.asarray(arr, order='c')
return arr.astype(np.uint32).view(ga.vec.uint3)[:,0]
def chunk_iterator(nelements, nthreads_per_block=64, max_blocks=1024):
"""Iterator that yields tuples with the values requried to process
a long array in multiple kernel passes on the GPU.
Each yielded value is of the form,
(first_index, elements_this_iteration, nblocks_this_iteration)
Example:
>>> list(chunk_iterator(300, 32, 2))
[(0, 64, 2), (64, 64, 2), (128, 64, 2), (192, 64, 2), (256, 9, 1)]
"""
first = 0
while first < nelements:
elements_left = nelements - first
blocks = int(elements_left // nthreads_per_block)
if elements_left % nthreads_per_block != 0:
blocks += 1 # Round up only if needed
blocks = min(max_blocks, blocks)
elements_this_round = min(elements_left, blocks * nthreads_per_block)
yield (first, elements_this_round, blocks)
first += elements_this_round
def create_cuda_context(device_id=None):
"""Initialize and return a CUDA context on the specified device.
If device_id is None, the default device is used."""
if device_id is None:
try:
context = pycuda.tools.make_default_context()
except cuda.LogicError:
# initialize cuda
cuda.init()
context = pycuda.tools.make_default_context()
else:
try:
device = cuda.Device(device_id)
except cuda.LogicError:
# initialize cuda
cuda.init()
device = cuda.Device(device_id)
context = device.make_context()
context.set_cache_config(cuda.func_cache.PREFER_L1)
return context
vec_dtypes = set([ x for x in ga.vec.__dict__.values() if type(x) == np.dtype ])
def make_gpu_struct(size, members):
struct = cuda.mem_alloc(size)
i = 0
for member in members:
if isinstance(member, ga.GPUArray):
member = member.gpudata
if isinstance(member, cuda.DeviceAllocation):
if i % 8:
raise Exception('cannot align 64-bit pointer. '
'arrange struct member variables in order of '
'decreasing size.')
cuda.memcpy_htod(int(struct)+i, np.intp(int(member)))
i += 8
elif np.isscalar(member) or (hasattr(member, 'dtype') and member.dtype in vec_dtypes and member.shape == ()):
cuda.memcpy_htod(int(struct)+i, member)
i += member.nbytes
else:
raise TypeError('expected a GPU device pointer or scalar type.')
return struct
def format_size(size):
if size < 1e3:
return '%.1f%s' % (size, ' ')
elif size < 1e6:
return '%.1f%s' % (size/1e3, 'K')
elif size < 1e9:
return '%.1f%s' % (size/1e6, 'M')
else:
return '%.1f%s' % (size/1e9, 'G')
def format_array(name, array):
return '%-15s %6s %6s' % \
(name, format_size(len(array)), format_size(array.nbytes))
def Mapped(array):
'''Analog to pycuda.driver.InOut(), but indicates this array
is memory mapped to the device space and should not be copied.
To simplify coding, Mapped() will pass anything with a gpudata
member, like a gpuarray, through unchanged.
'''
if hasattr(array, 'gpudata'):
return array
else:
return np.intp(array.base.get_device_pointer())
def mapped_alloc(pagelocked_alloc_func, shape, dtype, write_combined):
'''Returns a pagelocked host array mapped into the CUDA device
address space, with a gpudata field set so it just works with CUDA
functions.'''
flags = cuda.host_alloc_flags.DEVICEMAP
if write_combined:
flags |= cuda.host_alloc_flags.WRITECOMBINED
array = pagelocked_alloc_func(shape=shape, dtype=dtype, mem_flags=flags)
return array
def mapped_empty(shape, dtype, write_combined=False):
'''See mapped_alloc()'''
return mapped_alloc(cuda.pagelocked_empty, shape, dtype, write_combined)
def mapped_empty_like(other, write_combined=False):
'''See mapped_alloc()'''
return mapped_alloc(cuda.pagelocked_empty, other.shape, other.dtype,
write_combined)
def mapped_zeros(shape, dtype, write_combined=False):
'''See mapped_alloc()'''
return mapped_alloc(cuda.pagelocked_zeros, shape, dtype, write_combined)
def mapped_zeros_like(other, write_combined=False):
'''See mapped_alloc()'''
return mapped_alloc(cuda.pagelocked_zeros, other.shape, other.dtype,
write_combined)
|
