1 | """ |
---|
2 | Kernel Call Details |
---|
3 | =================== |
---|
4 | |
---|
5 | When calling sas computational kernels with polydispersity there are a |
---|
6 | number of details that need to be sent to the caller. This includes the |
---|
7 | list of polydisperse parameters, the number of points in the polydispersity |
---|
8 | weight distribution, and which parameter is the "theta" parameter for |
---|
9 | polar coordinate integration. The :class:`CallDetails` object maintains |
---|
10 | this data. Use :func:`build_details` to build a *details* object which |
---|
11 | can be passed to one of the computational kernels. |
---|
12 | """ |
---|
13 | |
---|
14 | from __future__ import print_function |
---|
15 | |
---|
16 | import numpy as np # type: ignore |
---|
17 | from numpy import cos, sin, radians |
---|
18 | |
---|
19 | try: |
---|
20 | np.meshgrid([]) |
---|
21 | meshgrid = np.meshgrid |
---|
22 | except Exception: |
---|
23 | # CRUFT: np.meshgrid requires multiple vectors |
---|
24 | def meshgrid(*args): |
---|
25 | """See docs from a recent version of numpy""" |
---|
26 | if len(args) > 1: |
---|
27 | return np.meshgrid(*args) |
---|
28 | else: |
---|
29 | return [np.asarray(v) for v in args] |
---|
30 | |
---|
31 | # pylint: disable=unused-import |
---|
32 | try: |
---|
33 | from typing import List, Tuple, Sequence |
---|
34 | except ImportError: |
---|
35 | pass |
---|
36 | else: |
---|
37 | from .modelinfo import ModelInfo |
---|
38 | from .modelinfo import ParameterTable |
---|
39 | # pylint: enable=unused-import |
---|
40 | |
---|
41 | |
---|
42 | class CallDetails(object): |
---|
43 | """ |
---|
44 | Manage the polydispersity information for the kernel call. |
---|
45 | |
---|
46 | Conceptually, a polydispersity calculation is an integral over a mesh |
---|
47 | in n-D space where n is the number of polydisperse parameters. In order |
---|
48 | to keep the program responsive, and not crash the GPU, only a portion |
---|
49 | of the mesh is computed at a time. Meshes with a large number of points |
---|
50 | will therefore require many calls to the polydispersity loop. Restarting |
---|
51 | a nested loop in the middle requires that the indices of the individual |
---|
52 | mesh dimensions can be computed for the current loop location. This |
---|
53 | is handled by the *pd_stride* vector, with n//stride giving the loop |
---|
54 | index and n%stride giving the position in the sub loops. |
---|
55 | |
---|
56 | One of the parameters may be the latitude. When integrating in polar |
---|
57 | coordinates, the total circumference decreases as latitude varies from |
---|
58 | pi r^2 at the equator to 0 at the pole, and the weight associated |
---|
59 | with a range of latitude values needs to be scaled by this circumference. |
---|
60 | This scale factor needs to be updated each time the theta value |
---|
61 | changes. *theta_par* indicates which of the values in the parameter |
---|
62 | vector is the latitude parameter, or -1 if there is no latitude |
---|
63 | parameter in the model. In practice, the normalization term cancels |
---|
64 | if the latitude is not a polydisperse parameter. |
---|
65 | """ |
---|
66 | parts = None # type: List["CallDetails"] |
---|
67 | def __init__(self, model_info): |
---|
68 | # type: (ModelInfo) -> None |
---|
69 | parameters = model_info.parameters |
---|
70 | max_pd = parameters.max_pd |
---|
71 | |
---|
72 | # Structure of the call details buffer: |
---|
73 | # pd_par[max_pd] pd params in order of length |
---|
74 | # pd_length[max_pd] length of each pd param |
---|
75 | # pd_offset[max_pd] offset of pd values in parameter array |
---|
76 | # pd_stride[max_pd] index of pd value in loop = n//stride[k] |
---|
77 | # num_eval total length of pd loop |
---|
78 | # num_weights total length of the weight vector |
---|
79 | # num_active number of pd params |
---|
80 | # theta_par parameter number for theta parameter |
---|
81 | self.buffer = np.empty(4*max_pd + 4, 'i4') |
---|
82 | |
---|
83 | # generate views on different parts of the array |
---|
84 | self._pd_par = self.buffer[0 * max_pd:1 * max_pd] |
---|
85 | self._pd_length = self.buffer[1 * max_pd:2 * max_pd] |
---|
86 | self._pd_offset = self.buffer[2 * max_pd:3 * max_pd] |
---|
87 | self._pd_stride = self.buffer[3 * max_pd:4 * max_pd] |
---|
88 | |
---|
89 | # theta_par is fixed |
---|
90 | self.theta_par = parameters.theta_offset |
---|
91 | |
---|
92 | # offset and length are for all parameters, not just pd parameters |
---|
93 | # They are not sent to the kernel function, though they could be. |
---|
94 | # They are used by the composite models (sum and product) to |
---|
95 | # figure out offsets into the combined value list. |
---|
96 | self.offset = None # type: np.ndarray |
---|
97 | self.length = None # type: np.ndarray |
---|
98 | |
---|
99 | # keep hold of ifno show() so we can break a values vector |
---|
100 | # into the individual components |
---|
101 | self.info = model_info |
---|
102 | |
---|
103 | @property |
---|
104 | def pd_par(self): |
---|
105 | """List of polydisperse parameters""" |
---|
106 | return self._pd_par |
---|
107 | |
---|
108 | @property |
---|
109 | def pd_length(self): |
---|
110 | """Number of weights for each polydisperse parameter""" |
---|
111 | return self._pd_length |
---|
112 | |
---|
113 | @property |
---|
114 | def pd_offset(self): |
---|
115 | """Offsets for the individual weight vectors in the set of weights""" |
---|
116 | return self._pd_offset |
---|
117 | |
---|
118 | @property |
---|
119 | def pd_stride(self): |
---|
120 | """Stride in the pd mesh for each pd dimension""" |
---|
121 | return self._pd_stride |
---|
122 | |
---|
123 | @property |
---|
124 | def num_eval(self): |
---|
125 | """Total size of the pd mesh""" |
---|
126 | return self.buffer[-4] |
---|
127 | |
---|
128 | @num_eval.setter |
---|
129 | def num_eval(self, v): |
---|
130 | """Total size of the pd mesh""" |
---|
131 | self.buffer[-4] = v |
---|
132 | |
---|
133 | @property |
---|
134 | def num_weights(self): |
---|
135 | """Total length of all the weight vectors""" |
---|
136 | return self.buffer[-3] |
---|
137 | |
---|
138 | @num_weights.setter |
---|
139 | def num_weights(self, v): |
---|
140 | """Total length of all the weight vectors""" |
---|
141 | self.buffer[-3] = v |
---|
142 | |
---|
143 | @property |
---|
144 | def num_active(self): |
---|
145 | """Number of active polydispersity loops""" |
---|
146 | return self.buffer[-2] |
---|
147 | |
---|
148 | @num_active.setter |
---|
149 | def num_active(self, v): |
---|
150 | """Number of active polydispersity loops""" |
---|
151 | self.buffer[-2] = v |
---|
152 | |
---|
153 | @property |
---|
154 | def theta_par(self): |
---|
155 | """Location of the theta parameter in the parameter vector""" |
---|
156 | return self.buffer[-1] |
---|
157 | |
---|
158 | @theta_par.setter |
---|
159 | def theta_par(self, v): |
---|
160 | """Location of the theta parameter in the parameter vector""" |
---|
161 | self.buffer[-1] = v |
---|
162 | |
---|
163 | def show(self, values=None): |
---|
164 | """Print the polydispersity call details to the console""" |
---|
165 | print("===== %s details ===="%self.info.name) |
---|
166 | print("num_active:%d num_eval:%d num_weights:%d theta=%d" |
---|
167 | % (self.num_active, self.num_eval, self.num_weights, self.theta_par)) |
---|
168 | if self.pd_par.size: |
---|
169 | print("pd_par", self.pd_par) |
---|
170 | print("pd_length", self.pd_length) |
---|
171 | print("pd_offset", self.pd_offset) |
---|
172 | print("pd_stride", self.pd_stride) |
---|
173 | if values is not None: |
---|
174 | nvalues = self.info.parameters.nvalues |
---|
175 | print("scale, background", values[:2]) |
---|
176 | print("val", values[2:nvalues]) |
---|
177 | print("pd", values[nvalues:nvalues+self.num_weights]) |
---|
178 | print("wt", values[nvalues+self.num_weights:nvalues+2*self.num_weights]) |
---|
179 | print("offsets", self.offset) |
---|
180 | |
---|
181 | |
---|
182 | def make_details(model_info, length, offset, num_weights): |
---|
183 | # type: (ModelInfo, np.ndarray, np.ndarray, int) -> CallDetails |
---|
184 | """ |
---|
185 | Return a :class:`CallDetails` object for a polydisperse calculation |
---|
186 | of the model defined by *model_info*. Polydispersity is defined by |
---|
187 | the *length* of the polydispersity distribution for each parameter |
---|
188 | and the *offset* of the distribution in the polydispersity array. |
---|
189 | Monodisperse parameters should use a polydispersity length of one |
---|
190 | with weight 1.0. *num_weights* is the total length of the polydispersity |
---|
191 | array. |
---|
192 | """ |
---|
193 | #pars = model_info.parameters.call_parameters[2:model_info.parameters.npars+2] |
---|
194 | #print(", ".join(str(i)+"-"+p.id for i,p in enumerate(pars))) |
---|
195 | #print("len:",length) |
---|
196 | #print("off:",offset) |
---|
197 | |
---|
198 | # Check that we aren't using too many polydispersity loops |
---|
199 | num_active = np.sum(length > 1) |
---|
200 | max_pd = model_info.parameters.max_pd |
---|
201 | if num_active > max_pd: |
---|
202 | raise ValueError("Too many polydisperse parameters") |
---|
203 | |
---|
204 | # Decreasing list of polydpersity lengths |
---|
205 | # Note: the reversing view, x[::-1], does not require a copy |
---|
206 | idx = np.argsort(length)[::-1][:max_pd] |
---|
207 | pd_stride = np.cumprod(np.hstack((1, length[idx]))) |
---|
208 | |
---|
209 | call_details = CallDetails(model_info) |
---|
210 | call_details.pd_par[:max_pd] = idx |
---|
211 | call_details.pd_length[:max_pd] = length[idx] |
---|
212 | call_details.pd_offset[:max_pd] = offset[idx] |
---|
213 | call_details.pd_stride[:max_pd] = pd_stride[:-1] |
---|
214 | call_details.num_eval = pd_stride[-1] |
---|
215 | call_details.num_weights = num_weights |
---|
216 | call_details.num_active = num_active |
---|
217 | call_details.length = length |
---|
218 | call_details.offset = offset |
---|
219 | #call_details.show() |
---|
220 | return call_details |
---|
221 | |
---|
222 | |
---|
223 | ZEROS = tuple([0.]*31) |
---|
224 | def make_kernel_args(kernel, # type: Kernel |
---|
225 | mesh # type: Tuple[List[np.ndarray], List[np.ndarray]] |
---|
226 | ): |
---|
227 | # type: (...) -> Tuple[CallDetails, np.ndarray, bool] |
---|
228 | """ |
---|
229 | Converts (value, dispersity, weight) for each parameter into kernel pars. |
---|
230 | |
---|
231 | Returns a CallDetails object indicating the polydispersity, a data object |
---|
232 | containing the different values, and the magnetic flag indicating whether |
---|
233 | any magnetic magnitudes are non-zero. Magnetic vectors (M0, phi, theta) are |
---|
234 | converted to rectangular coordinates (mx, my, mz). |
---|
235 | """ |
---|
236 | npars = kernel.info.parameters.npars |
---|
237 | nvalues = kernel.info.parameters.nvalues |
---|
238 | scalars = [value for value, _dispersity, _weight in mesh] |
---|
239 | # skipping scale and background when building values and weights |
---|
240 | _values, dispersity, weights = zip(*mesh[2:npars+2]) if npars else ((), (), ()) |
---|
241 | #weights = correct_theta_weights(kernel.info.parameters, dispersity, weights) |
---|
242 | length = np.array([len(w) for w in weights]) |
---|
243 | offset = np.cumsum(np.hstack((0, length))) |
---|
244 | call_details = make_details(kernel.info, length, offset[:-1], offset[-1]) |
---|
245 | # Pad value array to a 32 value boundary |
---|
246 | data_len = nvalues + 2*sum(len(v) for v in dispersity) |
---|
247 | extra = (32 - data_len%32)%32 |
---|
248 | data = np.hstack((scalars,) + dispersity + weights + ZEROS[:extra]) |
---|
249 | data = data.astype(kernel.dtype) |
---|
250 | is_magnetic = convert_magnetism(kernel.info.parameters, data) |
---|
251 | #call_details.show() |
---|
252 | #print("data", data) |
---|
253 | return call_details, data, is_magnetic |
---|
254 | |
---|
255 | def correct_theta_weights(parameters, # type: ParameterTable |
---|
256 | dispersity, # type: Sequence[np.ndarray] |
---|
257 | weights # type: Sequence[np.ndarray] |
---|
258 | ): |
---|
259 | # type: (...) -> Sequence[np.ndarray] |
---|
260 | """ |
---|
261 | **Deprecated** Theta weights will be computed in the kernel wrapper if |
---|
262 | they are needed. |
---|
263 | |
---|
264 | If there is a theta parameter, update the weights of that parameter so that |
---|
265 | the cosine weighting required for polar integration is preserved. |
---|
266 | |
---|
267 | Avoid evaluation strictly at the pole, which would otherwise send the |
---|
268 | weight to zero. This is probably not a problem in practice (if dispersity |
---|
269 | is +/- 90, then you probably should be using a 1-D model of the circular |
---|
270 | average). |
---|
271 | |
---|
272 | Note: scale and background parameters are not include in the tuples for |
---|
273 | dispersity and weights, so index is parameters.theta_offset, not |
---|
274 | parameters.theta_offset+2 |
---|
275 | |
---|
276 | Returns updated weights vectors |
---|
277 | """ |
---|
278 | # Apparently the parameters.theta_offset similarly skips scale and |
---|
279 | # and background, so the indexing works out, but they are still shipped |
---|
280 | # to the kernel, so we need to add two there. |
---|
281 | if parameters.theta_offset >= 0: |
---|
282 | index = parameters.theta_offset |
---|
283 | theta = dispersity[index] |
---|
284 | theta_weight = abs(cos(radians(theta))) |
---|
285 | weights = tuple(theta_weight*w if k == index else w |
---|
286 | for k, w in enumerate(weights)) |
---|
287 | return weights |
---|
288 | |
---|
289 | |
---|
290 | def convert_magnetism(parameters, values): |
---|
291 | # type: (ParameterTable, Sequence[np.ndarray]) -> bool |
---|
292 | """ |
---|
293 | Convert magnetism values from polar to rectangular coordinates. |
---|
294 | |
---|
295 | Returns True if any magnetism is present. |
---|
296 | """ |
---|
297 | mag = values[parameters.nvalues-3*parameters.nmagnetic:parameters.nvalues] |
---|
298 | mag = mag.reshape(-1, 3) |
---|
299 | if np.any(mag[:, 0] != 0.0): |
---|
300 | M0 = mag[:, 0].copy() |
---|
301 | theta, phi = radians(mag[:, 1]), radians(mag[:, 2]) |
---|
302 | mag[:, 0] = +M0*cos(theta)*cos(phi) # mx |
---|
303 | mag[:, 1] = +M0*sin(theta) # my |
---|
304 | mag[:, 2] = -M0*cos(theta)*sin(phi) # mz |
---|
305 | return True |
---|
306 | else: |
---|
307 | return False |
---|
308 | |
---|
309 | |
---|
310 | def dispersion_mesh(model_info, mesh): |
---|
311 | # type: (ModelInfo) -> Tuple[List[np.ndarray], List[np.ndarray]] |
---|
312 | """ |
---|
313 | Create a mesh grid of dispersion parameters and weights. |
---|
314 | |
---|
315 | *mesh* is a list of (value, dispersity, weights), where the values |
---|
316 | are the individual parameter values, and (dispersity, weights) is |
---|
317 | the distribution of parameter values. |
---|
318 | |
---|
319 | Only the volume parameters should be included in this list. Orientation |
---|
320 | parameters do not affect the calculation of effective radius or volume |
---|
321 | ratio. This is convenient since it avoids the distinction between |
---|
322 | value and dispersity that is present in orientation parameters but not |
---|
323 | shape parameters. |
---|
324 | |
---|
325 | Returns [p1,p2,...],w where pj is a vector of values for parameter j |
---|
326 | and w is a vector containing the products for weights for each |
---|
327 | parameter set in the vector. |
---|
328 | """ |
---|
329 | _, dispersity, weight = zip(*mesh) |
---|
330 | #weight = [w if len(w)>0 else [1.] for w in weight] |
---|
331 | weight = np.vstack([v.flatten() for v in meshgrid(*weight)]) |
---|
332 | weight = np.prod(weight, axis=0) |
---|
333 | dispersity = [v.flatten() for v in meshgrid(*dispersity)] |
---|
334 | lengths = [par.length for par in model_info.parameters.kernel_parameters |
---|
335 | if par.type == 'volume'] |
---|
336 | if any(n > 1 for n in lengths): |
---|
337 | pars = [] |
---|
338 | offset = 0 |
---|
339 | for n in lengths: |
---|
340 | pars.append(np.vstack(dispersity[offset:offset+n]) |
---|
341 | if n > 1 else dispersity[offset]) |
---|
342 | offset += n |
---|
343 | dispersity = pars |
---|
344 | return dispersity, weight |
---|