1 | r""" |
---|
2 | DLL driver for C kernels |
---|
3 | |
---|
4 | If the environment variable *SAS_OPENMP* is set, then sasmodels |
---|
5 | will attempt to compile with OpenMP flags so that the model can use all |
---|
6 | available kernels. This may or may not be available on your compiler |
---|
7 | toolchain. Depending on operating system and environment. |
---|
8 | |
---|
9 | Windows does not have provide a compiler with the operating system. |
---|
10 | Instead, we assume that TinyCC is installed and available. This can |
---|
11 | be done with a simple pip command if it is not already available:: |
---|
12 | |
---|
13 | pip install tinycc |
---|
14 | |
---|
15 | If Microsoft Visual C++ is available (because VCINSTALLDIR is |
---|
16 | defined in the environment), then that will be used instead. |
---|
17 | Microsoft Visual C++ for Python is available from Microsoft: |
---|
18 | |
---|
19 | `<http://www.microsoft.com/en-us/download/details.aspx?id=44266>`_ |
---|
20 | |
---|
21 | If neither compiler is available, sasmodels will check for *MinGW*, |
---|
22 | the GNU compiler toolchain. This available in packages such as Anaconda |
---|
23 | and PythonXY, or available stand alone. This toolchain has had |
---|
24 | difficulties on some systems, and may or may not work for you. |
---|
25 | |
---|
26 | You can control which compiler to use by setting SAS_COMPILER in the |
---|
27 | environment: |
---|
28 | |
---|
29 | - tinycc (Windows): use the TinyCC compiler shipped with SasView |
---|
30 | - msvc (Windows): use the Microsoft Visual C++ compiler |
---|
31 | - mingw (Windows): use the MinGW GNU cc compiler |
---|
32 | - unix (Linux): use the system cc compiler. |
---|
33 | - unix (Mac): use the clang compiler. You will need XCode installed, and |
---|
34 | the XCode command line tools. Mac comes with OpenCL drivers, so generally |
---|
35 | this will not be needed. |
---|
36 | |
---|
37 | Both *msvc* and *mingw* require that the compiler is available on your path. |
---|
38 | For *msvc*, this can done by running vcvarsall.bat in a windows terminal. |
---|
39 | Install locations are system dependent, such as: |
---|
40 | |
---|
41 | C:\Program Files (x86)\Common Files\Microsoft\Visual |
---|
42 | C++ for Python\9.0\vcvarsall.bat |
---|
43 | |
---|
44 | or maybe |
---|
45 | |
---|
46 | C:\Users\yourname\AppData\Local\Programs\Common\Microsoft\Visual |
---|
47 | C++ for Python\9.0\vcvarsall.bat |
---|
48 | |
---|
49 | OpenMP for *msvc* requires the Microsoft vcomp90.dll library, which doesn't |
---|
50 | seem to be included with the compiler, nor does there appear to be a public |
---|
51 | download location. There may be one on your machine already in a location |
---|
52 | such as: |
---|
53 | |
---|
54 | C:\Windows\winsxs\x86_microsoft.vc90.openmp*\vcomp90.dll |
---|
55 | |
---|
56 | If you copy this to somewhere on your path, such as the python directory or |
---|
57 | the install directory for this application, then OpenMP should be supported. |
---|
58 | |
---|
59 | For full control of the compiler, define a function |
---|
60 | *compile_command(source,output)* which takes the name of the source file |
---|
61 | and the name of the output file and returns a compile command that can be |
---|
62 | evaluated in the shell. For even more control, replace the entire |
---|
63 | *compile(source,output)* function. |
---|
64 | |
---|
65 | The global attribute *ALLOW_SINGLE_PRECISION_DLLS* should be set to *False* if |
---|
66 | you wish to prevent single precision floating point evaluation for the compiled |
---|
67 | models, otherwise set it defaults to *True*. |
---|
68 | """ |
---|
69 | from __future__ import print_function |
---|
70 | |
---|
71 | import sys |
---|
72 | import os |
---|
73 | from os.path import join as joinpath, splitext |
---|
74 | import subprocess |
---|
75 | import tempfile |
---|
76 | import ctypes as ct # type: ignore |
---|
77 | import _ctypes as _ct |
---|
78 | import logging |
---|
79 | |
---|
80 | import numpy as np # type: ignore |
---|
81 | |
---|
82 | try: |
---|
83 | import tinycc |
---|
84 | except ImportError: |
---|
85 | tinycc = None |
---|
86 | |
---|
87 | from . import generate |
---|
88 | from .kernel import KernelModel, Kernel |
---|
89 | from .kernelpy import PyInput |
---|
90 | from .exception import annotate_exception |
---|
91 | from .generate import F16, F32, F64 |
---|
92 | |
---|
93 | # pylint: disable=unused-import |
---|
94 | try: |
---|
95 | from typing import Tuple, Callable, Any |
---|
96 | from .modelinfo import ModelInfo |
---|
97 | from .details import CallDetails |
---|
98 | except ImportError: |
---|
99 | pass |
---|
100 | # pylint: enable=unused-import |
---|
101 | |
---|
102 | # Compiler output is a byte stream that needs to be decode in python 3. |
---|
103 | decode = (lambda s: s) if sys.version_info[0] < 3 else (lambda s: s.decode('utf8')) |
---|
104 | |
---|
105 | if "SAS_DLL_PATH" in os.environ: |
---|
106 | SAS_DLL_PATH = os.environ["SAS_DLL_PATH"] |
---|
107 | else: |
---|
108 | # Assume the default location of module DLLs is in .sasmodels/compiled_models. |
---|
109 | SAS_DLL_PATH = os.path.join(os.path.expanduser("~"), ".sasmodels", "compiled_models") |
---|
110 | |
---|
111 | if "SAS_COMPILER" in os.environ: |
---|
112 | COMPILER = os.environ["SAS_COMPILER"] |
---|
113 | elif os.name == 'nt': |
---|
114 | if tinycc is not None: |
---|
115 | COMPILER = "tinycc" |
---|
116 | elif "VCINSTALLDIR" in os.environ: |
---|
117 | # If vcvarsall.bat has been called, then VCINSTALLDIR is in the |
---|
118 | # environment and we can use the MSVC compiler. Otherwise, if |
---|
119 | # tinycc is available then use it. Otherwise, hope that mingw |
---|
120 | # is available. |
---|
121 | COMPILER = "msvc" |
---|
122 | else: |
---|
123 | COMPILER = "mingw" |
---|
124 | else: |
---|
125 | COMPILER = "unix" |
---|
126 | |
---|
127 | ARCH = "" if ct.sizeof(ct.c_void_p) > 4 else "x86" # 4 byte pointers on x86. |
---|
128 | if COMPILER == "unix": |
---|
129 | # Generic unix compile. |
---|
130 | # On Mac users will need the X code command line tools installed. |
---|
131 | #COMPILE = "gcc-mp-4.7 -shared -fPIC -std=c99 -fopenmp -O2 -Wall %s -o %s -lm -lgomp" |
---|
132 | CC = "cc -shared -fPIC -std=c99 -O2 -Wall".split() |
---|
133 | # Add OpenMP support if not running on a Mac. |
---|
134 | if sys.platform != "darwin": |
---|
135 | # OpenMP seems to be broken on gcc 5.4.0 (ubuntu 16.04.9). |
---|
136 | # Shut it off for all unix until we can investigate. |
---|
137 | #CC.append("-fopenmp") |
---|
138 | pass |
---|
139 | def compile_command(source, output): |
---|
140 | """unix compiler command""" |
---|
141 | return CC + [source, "-o", output, "-lm"] |
---|
142 | elif COMPILER == "msvc": |
---|
143 | # Call vcvarsall.bat before compiling to set path, headers, libs, etc. |
---|
144 | # MSVC compiler is available, so use it. OpenMP requires a copy of |
---|
145 | # vcomp90.dll on the path. One may be found here: |
---|
146 | # C:/Windows/winsxs/x86_microsoft.vc90.openmp*/vcomp90.dll |
---|
147 | # Copy this to the python directory and uncomment the OpenMP COMPILE. |
---|
148 | # TODO: Remove intermediate OBJ file created in the directory. |
---|
149 | # TODO: Maybe don't use randomized name for the c file. |
---|
150 | # TODO: Maybe ask distutils to find MSVC. |
---|
151 | CC = "cl /nologo /Ox /MD /W3 /GS- /DNDEBUG".split() |
---|
152 | if "SAS_OPENMP" in os.environ: |
---|
153 | CC.append("/openmp") |
---|
154 | LN = "/link /DLL /INCREMENTAL:NO /MANIFEST".split() |
---|
155 | def compile_command(source, output): |
---|
156 | """MSVC compiler command""" |
---|
157 | return CC + ["/Tp%s"%source] + LN + ["/OUT:%s"%output] |
---|
158 | elif COMPILER == "tinycc": |
---|
159 | # TinyCC compiler. |
---|
160 | CC = [tinycc.TCC] + "-shared -rdynamic -Wall".split() |
---|
161 | def compile_command(source, output): |
---|
162 | """tinycc compiler command""" |
---|
163 | return CC + [source, "-o", output] |
---|
164 | elif COMPILER == "mingw": |
---|
165 | # MinGW compiler. |
---|
166 | CC = "gcc -shared -std=c99 -O2 -Wall".split() |
---|
167 | if "SAS_OPENMP" in os.environ: |
---|
168 | CC.append("-fopenmp") |
---|
169 | def compile_command(source, output): |
---|
170 | """mingw compiler command""" |
---|
171 | return CC + [source, "-o", output, "-lm"] |
---|
172 | |
---|
173 | ALLOW_SINGLE_PRECISION_DLLS = True |
---|
174 | |
---|
175 | |
---|
176 | def compile_model(source, output): |
---|
177 | # type: (str, str) -> None |
---|
178 | """ |
---|
179 | Compile *source* producing *output*. |
---|
180 | |
---|
181 | Raises RuntimeError if the compile failed or the output wasn't produced. |
---|
182 | """ |
---|
183 | command = compile_command(source=source, output=output) |
---|
184 | command_str = " ".join('"%s"'%p if ' ' in p else p for p in command) |
---|
185 | logging.info(command_str) |
---|
186 | try: |
---|
187 | # Need shell=True on windows to keep console box from popping up. |
---|
188 | shell = (os.name == 'nt') |
---|
189 | subprocess.check_output(command, shell=shell, stderr=subprocess.STDOUT) |
---|
190 | except subprocess.CalledProcessError as exc: |
---|
191 | output = decode(exc.output) |
---|
192 | raise RuntimeError("compile failed.\n%s\n%s"%(command_str, output)) |
---|
193 | if not os.path.exists(output): |
---|
194 | raise RuntimeError("compile failed. File is in %r"%source) |
---|
195 | |
---|
196 | |
---|
197 | def dll_name(model_info, dtype): |
---|
198 | # type: (ModelInfo, np.dtype) -> str |
---|
199 | """ |
---|
200 | Name of the dll containing the model. This is the base file name without |
---|
201 | any path or extension, with a form such as 'sas_sphere32'. |
---|
202 | """ |
---|
203 | bits = 8*dtype.itemsize |
---|
204 | basename = "sas%d_%s"%(bits, model_info.id) |
---|
205 | basename += ARCH + ".so" |
---|
206 | |
---|
207 | # Hack to find precompiled dlls. |
---|
208 | path = joinpath(generate.DATA_PATH, '..', 'compiled_models', basename) |
---|
209 | if os.path.exists(path): |
---|
210 | return path |
---|
211 | |
---|
212 | return joinpath(SAS_DLL_PATH, basename) |
---|
213 | |
---|
214 | |
---|
215 | def dll_path(model_info, dtype): |
---|
216 | # type: (ModelInfo, np.dtype) -> str |
---|
217 | """ |
---|
218 | Complete path to the dll for the model. Note that the dll may not |
---|
219 | exist yet if it hasn't been compiled. |
---|
220 | """ |
---|
221 | return os.path.join(SAS_DLL_PATH, dll_name(model_info, dtype)) |
---|
222 | |
---|
223 | |
---|
224 | def make_dll(source, model_info, dtype=F64): |
---|
225 | # type: (str, ModelInfo, np.dtype) -> str |
---|
226 | """ |
---|
227 | Returns the path to the compiled model defined by *kernel_module*. |
---|
228 | |
---|
229 | If the model has not been compiled, or if the source file(s) are newer |
---|
230 | than the dll, then *make_dll* will compile the model before returning. |
---|
231 | This routine does not load the resulting dll. |
---|
232 | |
---|
233 | *dtype* is a numpy floating point precision specifier indicating whether |
---|
234 | the model should be single, double or long double precision. The default |
---|
235 | is double precision, *np.dtype('d')*. |
---|
236 | |
---|
237 | Set *sasmodels.ALLOW_SINGLE_PRECISION_DLLS* to False if single precision |
---|
238 | models are not allowed as DLLs. |
---|
239 | |
---|
240 | Set *sasmodels.kerneldll.SAS_DLL_PATH* to the compiled dll output path. |
---|
241 | Alternatively, set the environment variable *SAS_DLL_PATH*. |
---|
242 | The default is in ~/.sasmodels/compiled_models. |
---|
243 | """ |
---|
244 | if dtype == F16: |
---|
245 | raise ValueError("16 bit floats not supported") |
---|
246 | if dtype == F32 and not ALLOW_SINGLE_PRECISION_DLLS: |
---|
247 | dtype = F64 # Force 64-bit dll. |
---|
248 | # Note: dtype may be F128 for long double precision. |
---|
249 | |
---|
250 | dll = dll_path(model_info, dtype) |
---|
251 | |
---|
252 | if not os.path.exists(dll): |
---|
253 | need_recompile = True |
---|
254 | else: |
---|
255 | dll_time = os.path.getmtime(dll) |
---|
256 | newest_source = generate.dll_timestamp(model_info) |
---|
257 | need_recompile = dll_time < newest_source |
---|
258 | if need_recompile: |
---|
259 | # Make sure the DLL path exists. |
---|
260 | if not os.path.exists(SAS_DLL_PATH): |
---|
261 | os.makedirs(SAS_DLL_PATH) |
---|
262 | basename = splitext(os.path.basename(dll))[0] + "_" |
---|
263 | system_fd, filename = tempfile.mkstemp(suffix=".c", prefix=basename) |
---|
264 | source = generate.convert_type(source, dtype) |
---|
265 | with os.fdopen(system_fd, "w") as file_handle: |
---|
266 | file_handle.write(source) |
---|
267 | compile_model(source=filename, output=dll) |
---|
268 | # Comment the following to keep the generated C file. |
---|
269 | # Note: If there is a syntax error then compile raises an error |
---|
270 | # and the source file will not be deleted. |
---|
271 | os.unlink(filename) |
---|
272 | #print("saving compiled file in %r"%filename) |
---|
273 | return dll |
---|
274 | |
---|
275 | |
---|
276 | def load_dll(source, model_info, dtype=F64): |
---|
277 | # type: (str, ModelInfo, np.dtype) -> "DllModel" |
---|
278 | """ |
---|
279 | Create and load a dll corresponding to the source, info pair returned |
---|
280 | from :func:`sasmodels.generate.make` compiled for the target precision. |
---|
281 | |
---|
282 | See :func:`make_dll` for details on controlling the dll path and the |
---|
283 | allowed floating point precision. |
---|
284 | """ |
---|
285 | filename = make_dll(source, model_info, dtype=dtype) |
---|
286 | return DllModel(filename, model_info, dtype=dtype) |
---|
287 | |
---|
288 | |
---|
289 | class DllModel(KernelModel): |
---|
290 | """ |
---|
291 | ctypes wrapper for a single model. |
---|
292 | |
---|
293 | *source* and *model_info* are the model source and interface as returned |
---|
294 | from :func:`gen.make`. |
---|
295 | |
---|
296 | *dtype* is the desired model precision. Any numpy dtype for single |
---|
297 | or double precision floats will do, such as 'f', 'float32' or 'single' |
---|
298 | for single and 'd', 'float64' or 'double' for double. Double precision |
---|
299 | is an optional extension which may not be available on all devices. |
---|
300 | |
---|
301 | Call :meth:`release` when done with the kernel. |
---|
302 | """ |
---|
303 | def __init__(self, dllpath, model_info, dtype=generate.F32): |
---|
304 | # type: (str, ModelInfo, np.dtype) -> None |
---|
305 | self.info = model_info |
---|
306 | self.dllpath = dllpath |
---|
307 | self._dll = None # type: ct.CDLL |
---|
308 | self._kernels = None # type: List[Callable, Callable] |
---|
309 | self.dtype = np.dtype(dtype) |
---|
310 | |
---|
311 | def _load_dll(self): |
---|
312 | # type: () -> None |
---|
313 | try: |
---|
314 | self._dll = ct.CDLL(self.dllpath) |
---|
315 | except: |
---|
316 | annotate_exception("while loading "+self.dllpath) |
---|
317 | raise |
---|
318 | |
---|
319 | float_type = (ct.c_float if self.dtype == generate.F32 |
---|
320 | else ct.c_double if self.dtype == generate.F64 |
---|
321 | else ct.c_longdouble) |
---|
322 | |
---|
323 | # int, int, int, int*, double*, double*, double*, double*, double |
---|
324 | argtypes = [ct.c_int32]*3 + [ct.c_void_p]*4 + [float_type, ct.c_int32] |
---|
325 | names = [generate.kernel_name(self.info, variant) |
---|
326 | for variant in ("Iq", "Iqxy", "Imagnetic")] |
---|
327 | self._kernels = [self._dll[name] for name in names] |
---|
328 | for k in self._kernels: |
---|
329 | k.argtypes = argtypes |
---|
330 | |
---|
331 | def __getstate__(self): |
---|
332 | # type: () -> Tuple[ModelInfo, str] |
---|
333 | return self.info, self.dllpath |
---|
334 | |
---|
335 | def __setstate__(self, state): |
---|
336 | # type: (Tuple[ModelInfo, str]) -> None |
---|
337 | self.info, self.dllpath = state |
---|
338 | self._dll = None |
---|
339 | |
---|
340 | def make_kernel(self, q_vectors): |
---|
341 | # type: (List[np.ndarray]) -> DllKernel |
---|
342 | q_input = PyInput(q_vectors, self.dtype) |
---|
343 | # Note: DLL is lazy loaded. |
---|
344 | if self._dll is None: |
---|
345 | self._load_dll() |
---|
346 | is_2d = len(q_vectors) == 2 |
---|
347 | kernel = self._kernels[1:3] if is_2d else [self._kernels[0]]*2 |
---|
348 | return DllKernel(kernel, self.info, q_input) |
---|
349 | |
---|
350 | def release(self): |
---|
351 | # type: () -> None |
---|
352 | """ |
---|
353 | Release any resources associated with the model. |
---|
354 | """ |
---|
355 | dll_handle = self._dll._handle |
---|
356 | if os.name == 'nt': |
---|
357 | ct.windll.kernel32.FreeLibrary(dll_handle) |
---|
358 | else: |
---|
359 | _ct.dlclose(dll_handle) |
---|
360 | del self._dll |
---|
361 | self._dll = None |
---|
362 | |
---|
363 | |
---|
364 | class DllKernel(Kernel): |
---|
365 | """ |
---|
366 | Callable SAS kernel. |
---|
367 | |
---|
368 | *kernel* is the c function to call. |
---|
369 | |
---|
370 | *model_info* is the module information |
---|
371 | |
---|
372 | *q_input* is the DllInput q vectors at which the kernel should be |
---|
373 | evaluated. |
---|
374 | |
---|
375 | The resulting call method takes the *pars*, a list of values for |
---|
376 | the fixed parameters to the kernel, and *pd_pars*, a list of (value, weight) |
---|
377 | vectors for the polydisperse parameters. *cutoff* determines the |
---|
378 | integration limits: any points with combined weight less than *cutoff* |
---|
379 | will not be calculated. |
---|
380 | |
---|
381 | Call :meth:`release` when done with the kernel instance. |
---|
382 | """ |
---|
383 | def __init__(self, kernel, model_info, q_input): |
---|
384 | # type: (Callable[[], np.ndarray], ModelInfo, PyInput) -> None |
---|
385 | dtype = q_input.dtype |
---|
386 | self.q_input = q_input |
---|
387 | self.kernel = kernel |
---|
388 | |
---|
389 | # Attributes accessed from the outside. |
---|
390 | self.dim = '2d' if q_input.is_2d else '1d' |
---|
391 | self.info = model_info |
---|
392 | self.dtype = dtype |
---|
393 | |
---|
394 | # Converter to translate input to target type. |
---|
395 | self._as_dtype = (np.float32 if dtype == generate.F32 |
---|
396 | else np.float64 if dtype == generate.F64 |
---|
397 | else np.float128) |
---|
398 | |
---|
399 | # Holding place for the returned value. |
---|
400 | nout = 2 if self.info.have_Fq else 1 |
---|
401 | extra_q = 4 # Total weight, form volume, shell volume and R_eff. |
---|
402 | self.result = np.empty(self.q_input.nq*nout + extra_q, dtype) |
---|
403 | |
---|
404 | def _call_kernel(self, call_details, values, cutoff, magnetic, |
---|
405 | effective_radius_type): |
---|
406 | # type: (CallDetails, np.ndarray, float, bool, int) -> np.ndarray |
---|
407 | |
---|
408 | # Setup kernel function and arguments. |
---|
409 | kernel = self.kernel[1 if magnetic else 0] |
---|
410 | kernel_args = [ |
---|
411 | self.q_input.nq, # Number of inputs. |
---|
412 | None, # Placeholder for pd_start. |
---|
413 | None, # Placeholder for pd_stop. |
---|
414 | call_details.buffer.ctypes.data, # Problem definition. |
---|
415 | values.ctypes.data, # Parameter values. |
---|
416 | self.q_input.q.ctypes.data, # Q values. |
---|
417 | self.result.ctypes.data, # Result storage. |
---|
418 | self._as_dtype(cutoff), # Probability cutoff. |
---|
419 | effective_radius_type, # R_eff mode. |
---|
420 | ] |
---|
421 | |
---|
422 | # Call kernel and retrieve results. |
---|
423 | #print("Calling DLL") |
---|
424 | #call_details.show(values) |
---|
425 | step = 100 |
---|
426 | # TODO: Do we need the explicit sleep like the OpenCL and CUDA loops? |
---|
427 | for start in range(0, call_details.num_eval, step): |
---|
428 | stop = min(start + step, call_details.num_eval) |
---|
429 | kernel_args[1:3] = [start, stop] |
---|
430 | kernel(*kernel_args) # type: ignore |
---|
431 | |
---|
432 | def release(self): |
---|
433 | # type: () -> None |
---|
434 | """ |
---|
435 | Release resources associated with the kernel. |
---|
436 | """ |
---|
437 | # TODO: OpenCL/CUDA allocate q_input in __init__ and free it in release. |
---|
438 | # Should we be doing the same for DLL? |
---|
439 | #self.q_input.release() |
---|
440 | pass |
---|
441 | |
---|
442 | def __del__(self): |
---|
443 | # type: () -> None |
---|
444 | self.release() |
---|