| 1 | """ |
|---|
| 2 | C types wrapper for sasview models. |
|---|
| 3 | """ |
|---|
| 4 | |
|---|
| 5 | import ctypes as ct |
|---|
| 6 | from ctypes import c_void_p, c_int, c_double |
|---|
| 7 | |
|---|
| 8 | import numpy as np |
|---|
| 9 | |
|---|
| 10 | from . import gen |
|---|
| 11 | |
|---|
| 12 | from .gen import F32, F64 |
|---|
| 13 | |
|---|
| 14 | IQ_ARGS = [c_void_p, c_void_p, c_int, c_void_p, c_double] |
|---|
| 15 | IQXY_ARGS = [c_void_p, c_void_p, c_void_p, c_int, c_void_p, c_double] |
|---|
| 16 | |
|---|
| 17 | class DllModel(object): |
|---|
| 18 | """ |
|---|
| 19 | ctypes wrapper for a single model. |
|---|
| 20 | |
|---|
| 21 | *source* and *meta* are the model source and interface as returned |
|---|
| 22 | from :func:`gen.make`. |
|---|
| 23 | |
|---|
| 24 | *dtype* is the desired model precision. Any numpy dtype for single |
|---|
| 25 | or double precision floats will do, such as 'f', 'float32' or 'single' |
|---|
| 26 | for single and 'd', 'float64' or 'double' for double. Double precision |
|---|
| 27 | is an optional extension which may not be available on all devices. |
|---|
| 28 | """ |
|---|
| 29 | def __init__(self, dllpath, meta): |
|---|
| 30 | self.meta = meta |
|---|
| 31 | self.dll = ct.CDLL(dllpath) |
|---|
| 32 | self.Iq = self.dll[gen.kernel_name(self.meta, False)] |
|---|
| 33 | self.Iqxy = self.dll[gen.kernel_name(self.meta, True)] |
|---|
| 34 | |
|---|
| 35 | |
|---|
| 36 | self.PARS = dict((p[0],p[2]) for p in meta['parameters']) |
|---|
| 37 | self.PD_PARS = [p[0] for p in meta['parameters'] if p[4] != ""] |
|---|
| 38 | |
|---|
| 39 | # Determine the set of fixed and polydisperse parameters |
|---|
| 40 | Nfixed = len([p[0] for p in meta['parameters'] if p[4] == ""]) |
|---|
| 41 | N1D = len([p for p in meta['parameters'] if p[4]=="volume"]) |
|---|
| 42 | N2D = len([p for p in meta['parameters'] if p[4]!=""]) |
|---|
| 43 | self.Iq.argtypes = IQ_ARGS + [c_double]*Nfixed + [c_int]*N1D |
|---|
| 44 | self.Iqxy.argtypes = IQXY_ARGS + [c_double]*Nfixed + [c_int]*N2D |
|---|
| 45 | |
|---|
| 46 | def __call__(self, input, cutoff=1e-5): |
|---|
| 47 | kernel = self.Iqxy if input.is_2D else self.Iq |
|---|
| 48 | return DllKernel(kernel, self.meta, input, cutoff) |
|---|
| 49 | |
|---|
| 50 | def make_input(self, q_vectors): |
|---|
| 51 | """ |
|---|
| 52 | Make q input vectors available to the model. |
|---|
| 53 | |
|---|
| 54 | This only needs to be done once for all models that operate on the |
|---|
| 55 | same input. So for example, if you are adding two different models |
|---|
| 56 | together to compare to a data set, then only one model needs to |
|---|
| 57 | needs to call make_input, so long as the models have the same dtype. |
|---|
| 58 | """ |
|---|
| 59 | return DllInput(q_vectors) |
|---|
| 60 | |
|---|
| 61 | |
|---|
| 62 | class DllInput(object): |
|---|
| 63 | """ |
|---|
| 64 | Make q data available to the gpu. |
|---|
| 65 | |
|---|
| 66 | *q_vectors* is a list of q vectors, which will be *[q]* for 1-D data, |
|---|
| 67 | and *[qx, qy]* for 2-D data. Internally, the vectors will be reallocated |
|---|
| 68 | to get the best performance on OpenCL, which may involve shifting and |
|---|
| 69 | stretching the array to better match the memory architecture. Additional |
|---|
| 70 | points will be evaluated with *q=1e-3*. |
|---|
| 71 | |
|---|
| 72 | *dtype* is the data type for the q vectors. The data type should be |
|---|
| 73 | set to match that of the kernel, which is an attribute of |
|---|
| 74 | :class:`GpuProgram`. Note that not all kernels support double |
|---|
| 75 | precision, so even if the program was created for double precision, |
|---|
| 76 | the *GpuProgram.dtype* may be single precision. |
|---|
| 77 | |
|---|
| 78 | Call :meth:`release` when complete. Even if not called directly, the |
|---|
| 79 | buffer will be released when the data object is freed. |
|---|
| 80 | """ |
|---|
| 81 | def __init__(self, q_vectors): |
|---|
| 82 | self.nq = q_vectors[0].size |
|---|
| 83 | self.dtype = np.dtype('double') |
|---|
| 84 | self.is_2D = (len(q_vectors) == 2) |
|---|
| 85 | self.q_vectors = [np.ascontiguousarray(q,self.dtype) for q in q_vectors] |
|---|
| 86 | self.q_pointers = [q.ctypes.data for q in q_vectors] |
|---|
| 87 | |
|---|
| 88 | def release(self): |
|---|
| 89 | self.q_vectors = [] |
|---|
| 90 | |
|---|
| 91 | class DllKernel(object): |
|---|
| 92 | def __init__(self, kernel, meta, input, cutoff): |
|---|
| 93 | self.cutoff = cutoff |
|---|
| 94 | self.input = input |
|---|
| 95 | self.kernel = kernel |
|---|
| 96 | self.meta = meta |
|---|
| 97 | |
|---|
| 98 | self.res = np.empty(input.nq, input.dtype) |
|---|
| 99 | self.p_res = self.res.ctypes.data |
|---|
| 100 | |
|---|
| 101 | # Determine the set of fixed and polydisperse parameters |
|---|
| 102 | self.fixed_pars = [p[0] for p in meta['parameters'] if p[4] == ""] |
|---|
| 103 | self.pd_pars = [p for p in meta['parameters'] |
|---|
| 104 | if p[4]=="volume" or (p[4]=="orientation" and input.is_2D)] |
|---|
| 105 | |
|---|
| 106 | def eval(self, pars): |
|---|
| 107 | fixed, loops, loop_n = \ |
|---|
| 108 | gen.kernel_pars(pars, self.meta, self.input.is_2D, dtype=self.input.dtype) |
|---|
| 109 | real = np.float32 if self.input.dtype == F32 else np.float64 |
|---|
| 110 | nq = c_int(self.input.nq) |
|---|
| 111 | cutoff = real(self.cutoff) |
|---|
| 112 | |
|---|
| 113 | p_loops = loops.ctypes.data |
|---|
| 114 | pars = self.input.q_pointers + [self.p_res, nq, p_loops, cutoff] + fixed + loop_n |
|---|
| 115 | #print pars |
|---|
| 116 | self.kernel(*pars) |
|---|
| 117 | |
|---|
| 118 | return self.res |
|---|
| 119 | |
|---|
| 120 | def release(self): |
|---|
| 121 | pass |
|---|
| 122 | |
|---|
| 123 | def __del__(self): |
|---|
| 124 | self.release() |
|---|
