1 | r""" |
---|
2 | DLL driver for C kernels |
---|
3 | |
---|
4 | The global attribute *ALLOW_SINGLE_PRECISION_DLLS* should be set to *True* if |
---|
5 | you wish to allow single precision floating point evaluation for the compiled |
---|
6 | models, otherwise it defaults to *False*. |
---|
7 | |
---|
8 | The compiler command line is stored in the attribute *COMPILE*, with string |
---|
9 | substitutions for %(source)s and %(output)s indicating what to compile and |
---|
10 | where to store it. The actual command is system dependent. |
---|
11 | |
---|
12 | On windows systems, you have a choice of compilers. *MinGW* is the GNU |
---|
13 | compiler toolchain, available in packages such as anaconda and PythonXY, |
---|
14 | or available stand alone. This toolchain has had difficulties on some |
---|
15 | systems, and may or may not work for you. In order to build DLLs, *gcc* |
---|
16 | must be on your path. If the environment variable *SAS_OPENMP* is given |
---|
17 | then -fopenmp is added to the compiler flags. This requires a version |
---|
18 | of MinGW compiled with OpenMP support. |
---|
19 | |
---|
20 | An alternative toolchain uses the Microsoft Visual C++ compiler, available |
---|
21 | free from microsoft: |
---|
22 | |
---|
23 | `<http://www.microsoft.com/en-us/download/details.aspx?id=44266>`_ |
---|
24 | |
---|
25 | Again, this requires that the compiler is available on your path. This is |
---|
26 | done by running vcvarsall.bat in a windows terminal. Install locations are |
---|
27 | system dependent, such as: |
---|
28 | |
---|
29 | C:\Program Files (x86)\Common Files\Microsoft\Visual C++ for Python\9.0\vcvarsall.bat |
---|
30 | |
---|
31 | or maybe |
---|
32 | |
---|
33 | C:\Users\yourname\AppData\Local\Programs\Common\Microsoft\Visual C++ for Python\9.0\vcvarsall.bat |
---|
34 | |
---|
35 | And again, the environment variable *SAS_OPENMP* controls whether OpenMP is |
---|
36 | used to compile the C code. This requires the Microsoft vcomp90.dll library, |
---|
37 | which doesn't seem to be included with the compiler, nor does there appear |
---|
38 | to be a public download location. There may be one on your machine already |
---|
39 | in a location such as: |
---|
40 | |
---|
41 | C:\Windows\winsxs\x86_microsoft.vc90.openmp*\vcomp90.dll |
---|
42 | |
---|
43 | If you copy this onto your path, such as the python directory or the install |
---|
44 | directory for this application, then OpenMP should be supported. |
---|
45 | """ |
---|
46 | from __future__ import print_function |
---|
47 | |
---|
48 | import sys |
---|
49 | import os |
---|
50 | from os.path import join as joinpath, split as splitpath, splitext |
---|
51 | import tempfile |
---|
52 | import ctypes as ct |
---|
53 | from ctypes import c_void_p, c_int, c_longdouble, c_double, c_float |
---|
54 | import logging |
---|
55 | |
---|
56 | import numpy as np |
---|
57 | |
---|
58 | from . import generate |
---|
59 | from .kernelpy import PyInput, PyModel |
---|
60 | from .exception import annotate_exception |
---|
61 | |
---|
62 | if os.name == 'nt': |
---|
63 | # Windows compiler; check if TinyCC is available |
---|
64 | try: |
---|
65 | from tinycc import TCC |
---|
66 | except ImportError: |
---|
67 | TCC = None |
---|
68 | # call vcvarsall.bat before compiling to set path, headers, libs, etc. |
---|
69 | if "VCINSTALLDIR" in os.environ: |
---|
70 | # MSVC compiler is available, so use it. OpenMP requires a copy of |
---|
71 | # vcomp90.dll on the path. One may be found here: |
---|
72 | # C:/Windows/winsxs/x86_microsoft.vc90.openmp*/vcomp90.dll |
---|
73 | # Copy this to the python directory and uncomment the OpenMP COMPILE |
---|
74 | # TODO: remove intermediate OBJ file created in the directory |
---|
75 | # TODO: maybe don't use randomized name for the c file |
---|
76 | # TODO: maybe ask distutils to find MSVC |
---|
77 | CC = "cl /nologo /Ox /MD /W3 /GS- /DNDEBUG /Tp%(source)s " |
---|
78 | LN = "/link /DLL /INCREMENTAL:NO /MANIFEST /OUT:%(output)s" |
---|
79 | if "SAS_OPENMP" in os.environ: |
---|
80 | COMPILE = " ".join((CC, "/openmp", LN)) |
---|
81 | else: |
---|
82 | COMPILE = " ".join((CC, LN)) |
---|
83 | elif TCC: |
---|
84 | # TinyCC compiler. |
---|
85 | COMPILE = TCC + " -shared -rdynamic -Wall %(source)s -o %(output)s" |
---|
86 | else: |
---|
87 | # MinGW compiler. |
---|
88 | COMPILE = "gcc -shared -std=c99 -O2 -Wall %(source)s -o %(output)s -lm" |
---|
89 | if "SAS_OPENMP" in os.environ: |
---|
90 | COMPILE += " -fopenmp" |
---|
91 | else: |
---|
92 | # Generic unix compile |
---|
93 | # On mac users will need the X code command line tools installed |
---|
94 | COMPILE = "cc -shared -fPIC -std=c99 -O2 -Wall %(source)s -o %(output)s -lm" |
---|
95 | |
---|
96 | # add openmp support if not running on a mac |
---|
97 | if sys.platform != 'darwin': |
---|
98 | #COMPILE = "gcc-mp-4.7 -shared -fPIC -std=c99 -fopenmp -O2 -Wall %s -o %s -lm -lgomp" |
---|
99 | COMPILE += " -fopenmp" |
---|
100 | |
---|
101 | # Windows-specific solution |
---|
102 | if os.name == 'nt': |
---|
103 | # Assume the default location of module DLLs is in .sasmodels/compiled_models. |
---|
104 | DLL_PATH = os.path.join(os.path.expanduser("~"), ".sasmodels", "compiled_models") |
---|
105 | if not os.path.exists(DLL_PATH): |
---|
106 | os.makedirs(DLL_PATH) |
---|
107 | else: |
---|
108 | # Set up the default path for compiled modules. |
---|
109 | DLL_PATH = tempfile.gettempdir() |
---|
110 | |
---|
111 | ALLOW_SINGLE_PRECISION_DLLS = True |
---|
112 | |
---|
113 | |
---|
114 | def dll_path(model_info, dtype="double"): |
---|
115 | """ |
---|
116 | Path to the compiled model defined by *model_info*. |
---|
117 | """ |
---|
118 | basename = splitext(splitpath(model_info['filename'])[1])[0] |
---|
119 | if np.dtype(dtype) == generate.F32: |
---|
120 | basename += "32" |
---|
121 | elif np.dtype(dtype) == generate.F64: |
---|
122 | basename += "64" |
---|
123 | else: |
---|
124 | basename += "128" |
---|
125 | |
---|
126 | # Hack to find precompiled dlls |
---|
127 | path = joinpath(generate.DATA_PATH, '..', 'compiled_models', basename+'.so') |
---|
128 | if os.path.exists(path): |
---|
129 | return path |
---|
130 | |
---|
131 | return joinpath(DLL_PATH, basename+'.so') |
---|
132 | |
---|
133 | def make_dll(source, model_info, dtype="double"): |
---|
134 | """ |
---|
135 | Load the compiled model defined by *kernel_module*. |
---|
136 | |
---|
137 | Recompile if any files are newer than the model file. |
---|
138 | |
---|
139 | *dtype* is a numpy floating point precision specifier indicating whether |
---|
140 | the model should be single or double precision. The default is double |
---|
141 | precision. |
---|
142 | |
---|
143 | The DLL is not loaded until the kernel is called so models can |
---|
144 | be defined without using too many resources. |
---|
145 | |
---|
146 | Set *sasmodels.kerneldll.DLL_PATH* to the compiled dll output path. |
---|
147 | The default is the system temporary directory. |
---|
148 | |
---|
149 | Set *sasmodels.ALLOW_SINGLE_PRECISION_DLLS* to True if single precision |
---|
150 | models are allowed as DLLs. |
---|
151 | """ |
---|
152 | if callable(model_info.get('Iq', None)): |
---|
153 | return PyModel(model_info) |
---|
154 | |
---|
155 | dtype = np.dtype(dtype) |
---|
156 | if dtype == generate.F16: |
---|
157 | raise ValueError("16 bit floats not supported") |
---|
158 | if dtype == generate.F32 and not ALLOW_SINGLE_PRECISION_DLLS: |
---|
159 | dtype = generate.F64 # Force 64-bit dll |
---|
160 | |
---|
161 | if dtype == generate.F32: # 32-bit dll |
---|
162 | tempfile_prefix = 'sas_' + model_info['name'] + '32_' |
---|
163 | elif dtype == generate.F64: |
---|
164 | tempfile_prefix = 'sas_' + model_info['name'] + '64_' |
---|
165 | else: |
---|
166 | tempfile_prefix = 'sas_' + model_info['name'] + '128_' |
---|
167 | |
---|
168 | dll = dll_path(model_info, dtype) |
---|
169 | |
---|
170 | if not os.path.exists(dll): |
---|
171 | need_recompile = True |
---|
172 | elif getattr(sys, 'frozen', None) is not None: |
---|
173 | # TODO: don't suppress time stamp |
---|
174 | # Currently suppressing recompile when running in a frozen environment |
---|
175 | need_recompile = False |
---|
176 | else: |
---|
177 | dll_time = os.path.getmtime(dll) |
---|
178 | source_files = generate.model_sources(model_info) + [model_info['filename']] |
---|
179 | newest_source = max(os.path.getmtime(f) for f in source_files) |
---|
180 | need_recompile = dll_time < newest_source |
---|
181 | if need_recompile: |
---|
182 | fid, filename = tempfile.mkstemp(suffix=".c", prefix=tempfile_prefix) |
---|
183 | source = generate.convert_type(source, dtype) |
---|
184 | os.fdopen(fid, "w").write(source) |
---|
185 | command = COMPILE%{"source":filename, "output":dll} |
---|
186 | logging.info(command) |
---|
187 | status = os.system(command) |
---|
188 | if status != 0 or not os.path.exists(dll): |
---|
189 | raise RuntimeError("compile failed. File is in %r"%filename) |
---|
190 | else: |
---|
191 | ## comment the following to keep the generated c file |
---|
192 | os.unlink(filename) |
---|
193 | #print("saving compiled file in %r"%filename) |
---|
194 | return dll |
---|
195 | |
---|
196 | |
---|
197 | def load_dll(source, model_info, dtype="double"): |
---|
198 | """ |
---|
199 | Create and load a dll corresponding to the source, info pair returned |
---|
200 | from :func:`sasmodels.generate.make` compiled for the target precision. |
---|
201 | |
---|
202 | See :func:`make_dll` for details on controlling the dll path and the |
---|
203 | allowed floating point precision. |
---|
204 | """ |
---|
205 | filename = make_dll(source, model_info, dtype=dtype) |
---|
206 | return DllModel(filename, model_info, dtype=dtype) |
---|
207 | |
---|
208 | |
---|
209 | IQ_ARGS = [c_void_p, c_void_p, c_int] |
---|
210 | IQXY_ARGS = [c_void_p, c_void_p, c_void_p, c_int] |
---|
211 | |
---|
212 | class DllModel(object): |
---|
213 | """ |
---|
214 | ctypes wrapper for a single model. |
---|
215 | |
---|
216 | *source* and *model_info* are the model source and interface as returned |
---|
217 | from :func:`gen.make`. |
---|
218 | |
---|
219 | *dtype* is the desired model precision. Any numpy dtype for single |
---|
220 | or double precision floats will do, such as 'f', 'float32' or 'single' |
---|
221 | for single and 'd', 'float64' or 'double' for double. Double precision |
---|
222 | is an optional extension which may not be available on all devices. |
---|
223 | |
---|
224 | Call :meth:`release` when done with the kernel. |
---|
225 | """ |
---|
226 | |
---|
227 | def __init__(self, dllpath, model_info, dtype=generate.F32): |
---|
228 | self.info = model_info |
---|
229 | self.dllpath = dllpath |
---|
230 | self.dll = None |
---|
231 | self.dtype = np.dtype(dtype) |
---|
232 | |
---|
233 | def _load_dll(self): |
---|
234 | Nfixed1d = len(self.info['partype']['fixed-1d']) |
---|
235 | Nfixed2d = len(self.info['partype']['fixed-2d']) |
---|
236 | Npd1d = len(self.info['partype']['pd-1d']) |
---|
237 | Npd2d = len(self.info['partype']['pd-2d']) |
---|
238 | |
---|
239 | #print("dll", self.dllpath) |
---|
240 | try: |
---|
241 | self.dll = ct.CDLL(self.dllpath) |
---|
242 | except: |
---|
243 | annotate_exception("while loading "+self.dllpath) |
---|
244 | raise |
---|
245 | |
---|
246 | fp = (c_float if self.dtype == generate.F32 |
---|
247 | else c_double if self.dtype == generate.F64 |
---|
248 | else c_longdouble) |
---|
249 | pd_args_1d = [c_void_p, fp] + [c_int]*Npd1d if Npd1d else [] |
---|
250 | pd_args_2d = [c_void_p, fp] + [c_int]*Npd2d if Npd2d else [] |
---|
251 | self.Iq = self.dll[generate.kernel_name(self.info, False)] |
---|
252 | self.Iq.argtypes = IQ_ARGS + pd_args_1d + [fp]*Nfixed1d |
---|
253 | |
---|
254 | self.Iqxy = self.dll[generate.kernel_name(self.info, True)] |
---|
255 | self.Iqxy.argtypes = IQXY_ARGS + pd_args_2d + [fp]*Nfixed2d |
---|
256 | |
---|
257 | self.release() |
---|
258 | |
---|
259 | def __getstate__(self): |
---|
260 | return self.info, self.dllpath |
---|
261 | |
---|
262 | def __setstate__(self, state): |
---|
263 | self.info, self.dllpath = state |
---|
264 | self.dll = None |
---|
265 | |
---|
266 | def make_kernel(self, q_vectors): |
---|
267 | q_input = PyInput(q_vectors, self.dtype) |
---|
268 | if self.dll is None: self._load_dll() |
---|
269 | kernel = self.Iqxy if q_input.is_2d else self.Iq |
---|
270 | return DllKernel(kernel, self.info, q_input) |
---|
271 | |
---|
272 | def release(self): |
---|
273 | """ |
---|
274 | Release any resources associated with the model. |
---|
275 | """ |
---|
276 | if os.name == 'nt': |
---|
277 | #dll = ct.cdll.LoadLibrary(self.dllpath) |
---|
278 | dll = ct.CDLL(self.dllpath) |
---|
279 | libHandle = dll._handle |
---|
280 | #libHandle = ct.c_void_p(dll._handle) |
---|
281 | del dll, self.dll |
---|
282 | self.dll = None |
---|
283 | ct.windll.kernel32.FreeLibrary(libHandle) |
---|
284 | else: |
---|
285 | pass |
---|
286 | |
---|
287 | |
---|
288 | class DllKernel(object): |
---|
289 | """ |
---|
290 | Callable SAS kernel. |
---|
291 | |
---|
292 | *kernel* is the c function to call. |
---|
293 | |
---|
294 | *model_info* is the module information |
---|
295 | |
---|
296 | *q_input* is the DllInput q vectors at which the kernel should be |
---|
297 | evaluated. |
---|
298 | |
---|
299 | The resulting call method takes the *pars*, a list of values for |
---|
300 | the fixed parameters to the kernel, and *pd_pars*, a list of (value, weight) |
---|
301 | vectors for the polydisperse parameters. *cutoff* determines the |
---|
302 | integration limits: any points with combined weight less than *cutoff* |
---|
303 | will not be calculated. |
---|
304 | |
---|
305 | Call :meth:`release` when done with the kernel instance. |
---|
306 | """ |
---|
307 | def __init__(self, kernel, model_info, q_input): |
---|
308 | self.info = model_info |
---|
309 | self.q_input = q_input |
---|
310 | self.kernel = kernel |
---|
311 | self.res = np.empty(q_input.nq, q_input.dtype) |
---|
312 | dim = '2d' if q_input.is_2d else '1d' |
---|
313 | self.fixed_pars = model_info['partype']['fixed-' + dim] |
---|
314 | self.pd_pars = model_info['partype']['pd-' + dim] |
---|
315 | |
---|
316 | # In dll kernel, but not in opencl kernel |
---|
317 | self.p_res = self.res.ctypes.data |
---|
318 | |
---|
319 | def __call__(self, fixed_pars, pd_pars, cutoff): |
---|
320 | real = (np.float32 if self.q_input.dtype == generate.F32 |
---|
321 | else np.float64 if self.q_input.dtype == generate.F64 |
---|
322 | else np.float128) |
---|
323 | |
---|
324 | nq = c_int(self.q_input.nq) |
---|
325 | if pd_pars: |
---|
326 | cutoff = real(cutoff) |
---|
327 | loops_N = [np.uint32(len(p[0])) for p in pd_pars] |
---|
328 | loops = np.hstack(pd_pars) |
---|
329 | loops = np.ascontiguousarray(loops.T, self.q_input.dtype).flatten() |
---|
330 | p_loops = loops.ctypes.data |
---|
331 | dispersed = [p_loops, cutoff] + loops_N |
---|
332 | else: |
---|
333 | dispersed = [] |
---|
334 | fixed = [real(p) for p in fixed_pars] |
---|
335 | args = self.q_input.q_pointers + [self.p_res, nq] + dispersed + fixed |
---|
336 | #print(pars) |
---|
337 | self.kernel(*args) |
---|
338 | |
---|
339 | return self.res |
---|
340 | |
---|
341 | def release(self): |
---|
342 | """ |
---|
343 | Release any resources associated with the kernel. |
---|
344 | """ |
---|
345 | pass |
---|