1 | """ |
---|
2 | Core model handling routines. |
---|
3 | """ |
---|
4 | from __future__ import print_function |
---|
5 | |
---|
6 | __all__ = [ |
---|
7 | "list_models", "load_model", "load_model_info", |
---|
8 | "build_model", "precompile_dlls", |
---|
9 | ] |
---|
10 | |
---|
11 | import os |
---|
12 | from os.path import basename, join as joinpath |
---|
13 | from glob import glob |
---|
14 | import re |
---|
15 | |
---|
16 | import numpy as np # type: ignore |
---|
17 | |
---|
18 | from . import generate |
---|
19 | from . import modelinfo |
---|
20 | from . import product |
---|
21 | from . import mixture |
---|
22 | from . import kernelpy |
---|
23 | from . import kernelcl |
---|
24 | from . import kerneldll |
---|
25 | from . import custom |
---|
26 | |
---|
27 | # pylint: disable=unused-import |
---|
28 | try: |
---|
29 | from typing import List, Union, Optional, Any |
---|
30 | from .kernel import KernelModel |
---|
31 | from .modelinfo import ModelInfo |
---|
32 | except ImportError: |
---|
33 | pass |
---|
34 | # pylint: enable=unused-import |
---|
35 | |
---|
36 | CUSTOM_MODEL_PATH = os.environ.get('SAS_MODELPATH', "") |
---|
37 | if CUSTOM_MODEL_PATH == "": |
---|
38 | CUSTOM_MODEL_PATH = joinpath(os.path.expanduser("~"), ".sasmodels", "custom_models") |
---|
39 | #if not os.path.isdir(CUSTOM_MODEL_PATH): |
---|
40 | # os.makedirs(CUSTOM_MODEL_PATH) |
---|
41 | |
---|
42 | # TODO: refactor composite model support |
---|
43 | # The current load_model_info/build_model does not reuse existing model |
---|
44 | # definitions when loading a composite model, instead reloading and |
---|
45 | # rebuilding the kernel for each component model in the expression. This |
---|
46 | # is fine in a scripting environment where the model is built when the script |
---|
47 | # starts and is thrown away when the script ends, but may not be the best |
---|
48 | # solution in a long-lived application. This affects the following functions: |
---|
49 | # |
---|
50 | # load_model |
---|
51 | # load_model_info |
---|
52 | # build_model |
---|
53 | |
---|
54 | KINDS = ("all", "py", "c", "double", "single", "opencl", "1d", "2d", |
---|
55 | "nonmagnetic", "magnetic") |
---|
56 | def list_models(kind=None): |
---|
57 | # type: (str) -> List[str] |
---|
58 | """ |
---|
59 | Return the list of available models on the model path. |
---|
60 | |
---|
61 | *kind* can be one of the following: |
---|
62 | |
---|
63 | * all: all models |
---|
64 | * py: python models only |
---|
65 | * c: compiled models only |
---|
66 | * single: models which support single precision |
---|
67 | * double: models which require double precision |
---|
68 | * opencl: controls if OpenCL is supperessed |
---|
69 | * 1d: models which are 1D only, or 2D using abs(q) |
---|
70 | * 2d: models which can be 2D |
---|
71 | * magnetic: models with an sld |
---|
72 | * nommagnetic: models without an sld |
---|
73 | |
---|
74 | For multiple conditions, combine with plus. For example, *c+single+2d* |
---|
75 | would return all oriented models implemented in C which can be computed |
---|
76 | accurately with single precision arithmetic. |
---|
77 | """ |
---|
78 | if kind and any(k not in KINDS for k in kind.split('+')): |
---|
79 | raise ValueError("kind not in " + ", ".join(KINDS)) |
---|
80 | files = sorted(glob(joinpath(generate.MODEL_PATH, "[a-zA-Z]*.py"))) |
---|
81 | available_models = [basename(f)[:-3] for f in files] |
---|
82 | if kind and '+' in kind: |
---|
83 | all_kinds = kind.split('+') |
---|
84 | condition = lambda name: all(_matches(name, k) for k in all_kinds) |
---|
85 | else: |
---|
86 | condition = lambda name: _matches(name, kind) |
---|
87 | selected = [name for name in available_models if condition(name)] |
---|
88 | |
---|
89 | return selected |
---|
90 | |
---|
91 | def _matches(name, kind): |
---|
92 | if kind is None or kind == "all": |
---|
93 | return True |
---|
94 | info = load_model_info(name) |
---|
95 | pars = info.parameters.kernel_parameters |
---|
96 | if kind == "py" and callable(info.Iq): |
---|
97 | return True |
---|
98 | elif kind == "c" and not callable(info.Iq): |
---|
99 | return True |
---|
100 | elif kind == "double" and not info.single: |
---|
101 | return True |
---|
102 | elif kind == "single" and info.single: |
---|
103 | return True |
---|
104 | elif kind == "opencl" and info.opencl: |
---|
105 | return True |
---|
106 | elif kind == "2d" and any(p.type == 'orientation' for p in pars): |
---|
107 | return True |
---|
108 | elif kind == "1d" and all(p.type != 'orientation' for p in pars): |
---|
109 | return True |
---|
110 | elif kind == "magnetic" and any(p.type == 'sld' for p in pars): |
---|
111 | return True |
---|
112 | elif kind == "nonmagnetic" and any(p.type != 'sld' for p in pars): |
---|
113 | return True |
---|
114 | return False |
---|
115 | |
---|
116 | def load_model(model_name, dtype=None, platform='ocl'): |
---|
117 | # type: (str, str, str) -> KernelModel |
---|
118 | """ |
---|
119 | Load model info and build model. |
---|
120 | |
---|
121 | *model_name* is the name of the model, or perhaps a model expression |
---|
122 | such as sphere*hardsphere or sphere+cylinder. |
---|
123 | |
---|
124 | *dtype* and *platform* are given by :func:`build_model`. |
---|
125 | """ |
---|
126 | return build_model(load_model_info(model_name), |
---|
127 | dtype=dtype, platform=platform) |
---|
128 | |
---|
129 | def load_model_info(model_string): |
---|
130 | # type: (str) -> modelinfo.ModelInfo |
---|
131 | """ |
---|
132 | Load a model definition given the model name. |
---|
133 | |
---|
134 | *model_string* is the name of the model, or perhaps a model expression |
---|
135 | such as sphere*cylinder or sphere+cylinder. Use '@' for a structure |
---|
136 | factor product, e.g. sphere@hardsphere. Custom models can be specified by |
---|
137 | prefixing the model name with 'custom.', e.g. 'custom.MyModel+sphere'. |
---|
138 | |
---|
139 | This returns a handle to the module defining the model. This can be |
---|
140 | used with functions in generate to build the docs or extract model info. |
---|
141 | """ |
---|
142 | if "+" in model_string: |
---|
143 | parts = [load_model_info(part) |
---|
144 | for part in model_string.split("+")] |
---|
145 | return mixture.make_mixture_info(parts, operation='+') |
---|
146 | elif "*" in model_string: |
---|
147 | parts = [load_model_info(part) |
---|
148 | for part in model_string.split("*")] |
---|
149 | return mixture.make_mixture_info(parts, operation='*') |
---|
150 | elif "@" in model_string: |
---|
151 | p_info, q_info = [load_model_info(part) |
---|
152 | for part in model_string.split("@")] |
---|
153 | return product.make_product_info(p_info, q_info) |
---|
154 | # We are now dealing with a pure model |
---|
155 | elif "custom." in model_string: |
---|
156 | pattern = "custom.([A-Za-z0-9_-]+)" |
---|
157 | result = re.match(pattern, model_string) |
---|
158 | if result is None: |
---|
159 | raise ValueError("Model name in invalid format: " + model_string) |
---|
160 | model_name = result.group(1) |
---|
161 | # Use ModelName to find the path to the custom model file |
---|
162 | model_path = joinpath(CUSTOM_MODEL_PATH, model_name + ".py") |
---|
163 | if not os.path.isfile(model_path): |
---|
164 | raise ValueError("The model file {} doesn't exist".format(model_path)) |
---|
165 | kernel_module = custom.load_custom_kernel_module(model_path) |
---|
166 | return modelinfo.make_model_info(kernel_module) |
---|
167 | kernel_module = generate.load_kernel_module(model_string) |
---|
168 | return modelinfo.make_model_info(kernel_module) |
---|
169 | |
---|
170 | |
---|
171 | def build_model(model_info, dtype=None, platform="ocl"): |
---|
172 | # type: (modelinfo.ModelInfo, str, str) -> KernelModel |
---|
173 | """ |
---|
174 | Prepare the model for the default execution platform. |
---|
175 | |
---|
176 | This will return an OpenCL model, a DLL model or a python model depending |
---|
177 | on the model and the computing platform. |
---|
178 | |
---|
179 | *model_info* is the model definition structure returned from |
---|
180 | :func:`load_model_info`. |
---|
181 | |
---|
182 | *dtype* indicates whether the model should use single or double precision |
---|
183 | for the calculation. Choices are 'single', 'double', 'quad', 'half', |
---|
184 | or 'fast'. If *dtype* ends with '!', then force the use of the DLL rather |
---|
185 | than OpenCL for the calculation. |
---|
186 | |
---|
187 | *platform* should be "dll" to force the dll to be used for C models, |
---|
188 | otherwise it uses the default "ocl". |
---|
189 | """ |
---|
190 | composition = model_info.composition |
---|
191 | if composition is not None: |
---|
192 | composition_type, parts = composition |
---|
193 | models = [build_model(p, dtype=dtype, platform=platform) for p in parts] |
---|
194 | if composition_type == 'mixture': |
---|
195 | return mixture.MixtureModel(model_info, models) |
---|
196 | elif composition_type == 'product': |
---|
197 | P, S = models |
---|
198 | return product.ProductModel(model_info, P, S) |
---|
199 | else: |
---|
200 | raise ValueError('unknown mixture type %s'%composition_type) |
---|
201 | |
---|
202 | # If it is a python model, return it immediately |
---|
203 | if callable(model_info.Iq): |
---|
204 | return kernelpy.PyModel(model_info) |
---|
205 | |
---|
206 | numpy_dtype, fast, platform = parse_dtype(model_info, dtype, platform) |
---|
207 | |
---|
208 | source = generate.make_source(model_info) |
---|
209 | if platform == "dll": |
---|
210 | #print("building dll", numpy_dtype) |
---|
211 | return kerneldll.load_dll(source['dll'], model_info, numpy_dtype) |
---|
212 | else: |
---|
213 | #print("building ocl", numpy_dtype) |
---|
214 | return kernelcl.GpuModel(source, model_info, numpy_dtype, fast=fast) |
---|
215 | |
---|
216 | def precompile_dlls(path, dtype="double"): |
---|
217 | # type: (str, str) -> List[str] |
---|
218 | """ |
---|
219 | Precompile the dlls for all builtin models, returning a list of dll paths. |
---|
220 | |
---|
221 | *path* is the directory in which to save the dlls. It will be created if |
---|
222 | it does not already exist. |
---|
223 | |
---|
224 | This can be used when build the windows distribution of sasmodels |
---|
225 | which may be missing the OpenCL driver and the dll compiler. |
---|
226 | """ |
---|
227 | numpy_dtype = np.dtype(dtype) |
---|
228 | if not os.path.exists(path): |
---|
229 | os.makedirs(path) |
---|
230 | compiled_dlls = [] |
---|
231 | for model_name in list_models(): |
---|
232 | model_info = load_model_info(model_name) |
---|
233 | if not callable(model_info.Iq): |
---|
234 | source = generate.make_source(model_info)['dll'] |
---|
235 | old_path = kerneldll.SAS_DLL_PATH |
---|
236 | try: |
---|
237 | kerneldll.SAS_DLL_PATH = path |
---|
238 | dll = kerneldll.make_dll(source, model_info, dtype=numpy_dtype) |
---|
239 | finally: |
---|
240 | kerneldll.SAS_DLL_PATH = old_path |
---|
241 | compiled_dlls.append(dll) |
---|
242 | return compiled_dlls |
---|
243 | |
---|
244 | def parse_dtype(model_info, dtype=None, platform=None): |
---|
245 | # type: (ModelInfo, str, str) -> (np.dtype, bool, str) |
---|
246 | """ |
---|
247 | Interpret dtype string, returning np.dtype and fast flag. |
---|
248 | |
---|
249 | Possible types include 'half', 'single', 'double' and 'quad'. If the |
---|
250 | type is 'fast', then this is equivalent to dtype 'single' but using |
---|
251 | fast native functions rather than those with the precision level |
---|
252 | guaranteed by the OpenCL standard. 'default' will choose the appropriate |
---|
253 | default for the model and platform. |
---|
254 | |
---|
255 | Platform preference can be specfied ("ocl" vs "dll"), with the default |
---|
256 | being OpenCL if it is availabe. If the dtype name ends with '!' then |
---|
257 | platform is forced to be DLL rather than OpenCL. |
---|
258 | |
---|
259 | This routine ignores the preferences within the model definition. This |
---|
260 | is by design. It allows us to test models in single precision even when |
---|
261 | we have flagged them as requiring double precision so we can easily check |
---|
262 | the performance on different platforms without having to change the model |
---|
263 | definition. |
---|
264 | """ |
---|
265 | # Assign default platform, overriding ocl with dll if OpenCL is unavailable |
---|
266 | # If opencl=False OpenCL is switched off |
---|
267 | |
---|
268 | if platform is None: |
---|
269 | platform = "ocl" |
---|
270 | if not kernelcl.use_opencl() or not model_info.opencl: |
---|
271 | platform = "dll" |
---|
272 | |
---|
273 | # Check if type indicates dll regardless of which platform is given |
---|
274 | if dtype is not None and dtype.endswith('!'): |
---|
275 | platform = "dll" |
---|
276 | dtype = dtype[:-1] |
---|
277 | |
---|
278 | # Convert special type names "half", "fast", and "quad" |
---|
279 | fast = (dtype == "fast") |
---|
280 | if fast: |
---|
281 | dtype = "single" |
---|
282 | elif dtype == "quad": |
---|
283 | dtype = "longdouble" |
---|
284 | elif dtype == "half": |
---|
285 | dtype = "float16" |
---|
286 | |
---|
287 | # Convert dtype string to numpy dtype. |
---|
288 | if dtype is None or dtype == "default": |
---|
289 | numpy_dtype = (generate.F32 if platform == "ocl" and model_info.single |
---|
290 | else generate.F64) |
---|
291 | else: |
---|
292 | numpy_dtype = np.dtype(dtype) |
---|
293 | |
---|
294 | # Make sure that the type is supported by opencl, otherwise use dll |
---|
295 | if platform == "ocl": |
---|
296 | env = kernelcl.environment() |
---|
297 | if not env.has_type(numpy_dtype): |
---|
298 | platform = "dll" |
---|
299 | if dtype is None: |
---|
300 | numpy_dtype = generate.F64 |
---|
301 | |
---|
302 | return numpy_dtype, fast, platform |
---|
303 | |
---|
304 | def list_models_main(): |
---|
305 | # type: () -> None |
---|
306 | """ |
---|
307 | Run list_models as a main program. See :func:`list_models` for the |
---|
308 | kinds of models that can be requested on the command line. |
---|
309 | """ |
---|
310 | import sys |
---|
311 | kind = sys.argv[1] if len(sys.argv) > 1 else "all" |
---|
312 | print("\n".join(list_models(kind))) |
---|
313 | |
---|
314 | def test_composite_order(): |
---|
315 | def test_models(fst, snd): |
---|
316 | """Confirm that two models produce the same parameters""" |
---|
317 | fst = load_model(fst) |
---|
318 | snd = load_model(snd) |
---|
319 | # Un-disambiguate parameter names so that we can check if the same |
---|
320 | # parameters are in a pair of composite models. Since each parameter in |
---|
321 | # the mixture model is tagged as e.g., A_sld, we ought to use a |
---|
322 | # regex subsitution s/^[A-Z]+_/_/, but removing all uppercase letters |
---|
323 | # is good enough. |
---|
324 | fst = [[x for x in p.name if x == x.lower()] for p in fst.info.parameters.kernel_parameters] |
---|
325 | snd = [[x for x in p.name if x == x.lower()] for p in snd.info.parameters.kernel_parameters] |
---|
326 | assert sorted(fst) == sorted(snd), "{} != {}".format(fst, snd) |
---|
327 | |
---|
328 | def build_test(first, second): |
---|
329 | test = lambda description: test_models(first, second) |
---|
330 | description = first + " vs. " + second |
---|
331 | return test, description |
---|
332 | |
---|
333 | yield build_test( |
---|
334 | "cylinder+sphere", |
---|
335 | "sphere+cylinder") |
---|
336 | yield build_test( |
---|
337 | "cylinder*sphere", |
---|
338 | "sphere*cylinder") |
---|
339 | yield build_test( |
---|
340 | "cylinder@hardsphere*sphere", |
---|
341 | "sphere*cylinder@hardsphere") |
---|
342 | yield build_test( |
---|
343 | "barbell+sphere*cylinder@hardsphere", |
---|
344 | "sphere*cylinder@hardsphere+barbell") |
---|
345 | yield build_test( |
---|
346 | "barbell+cylinder@hardsphere*sphere", |
---|
347 | "cylinder@hardsphere*sphere+barbell") |
---|
348 | yield build_test( |
---|
349 | "barbell+sphere*cylinder@hardsphere", |
---|
350 | "barbell+cylinder@hardsphere*sphere") |
---|
351 | yield build_test( |
---|
352 | "sphere*cylinder@hardsphere+barbell", |
---|
353 | "cylinder@hardsphere*sphere+barbell") |
---|
354 | yield build_test( |
---|
355 | "barbell+sphere*cylinder@hardsphere", |
---|
356 | "cylinder@hardsphere*sphere+barbell") |
---|
357 | yield build_test( |
---|
358 | "barbell+cylinder@hardsphere*sphere", |
---|
359 | "sphere*cylinder@hardsphere+barbell") |
---|
360 | |
---|
361 | def test_composite(): |
---|
362 | # type: () -> None |
---|
363 | """Check that model load works""" |
---|
364 | #Test the the model produces the parameters that we would expect |
---|
365 | model = load_model("cylinder@hardsphere*sphere") |
---|
366 | actual = [p.name for p in model.info.parameters.kernel_parameters] |
---|
367 | target = ("sld sld_solvent radius length theta phi volfraction" |
---|
368 | " A_sld A_sld_solvent A_radius").split() |
---|
369 | assert target == actual, "%s != %s"%(target, actual) |
---|
370 | |
---|
371 | if __name__ == "__main__": |
---|
372 | list_models_main() |
---|