[14de349] | 1 | """ |
---|
| 2 | C types wrapper for sasview models. |
---|
| 3 | """ |
---|
[5d4777d] | 4 | import sys |
---|
| 5 | import os |
---|
[14de349] | 6 | import ctypes as ct |
---|
| 7 | from ctypes import c_void_p, c_int, c_double |
---|
| 8 | |
---|
| 9 | import numpy as np |
---|
| 10 | |
---|
| 11 | from . import gen |
---|
| 12 | |
---|
| 13 | from .gen import F32, F64 |
---|
[5d4777d] | 14 | # Compiler platform details |
---|
| 15 | if sys.platform == 'darwin': |
---|
[216a9e1] | 16 | #COMPILE = "gcc-mp-4.7 -shared -fPIC -std=c99 -fopenmp -O2 -Wall %s -o %s -lm -lgomp" |
---|
| 17 | COMPILE = "gcc -shared -fPIC -std=c99 -O2 -Wall %s -o %s -lm" |
---|
[5d4777d] | 18 | elif os.name == 'nt': |
---|
| 19 | COMPILE = "gcc -shared -fPIC -std=c99 -fopenmp -O2 -Wall %s -o %s -lm" |
---|
| 20 | else: |
---|
| 21 | COMPILE = "cc -shared -fPIC -std=c99 -fopenmp -O2 -Wall %s -o %s -lm" |
---|
| 22 | DLL_PATH = "/tmp" |
---|
| 23 | |
---|
| 24 | |
---|
| 25 | def dll_path(info): |
---|
| 26 | """ |
---|
| 27 | Path to the compiled model defined by *info*. |
---|
| 28 | """ |
---|
| 29 | from os.path import join as joinpath, split as splitpath, splitext |
---|
| 30 | basename = splitext(splitpath(info['filename'])[1])[0] |
---|
| 31 | return joinpath(DLL_PATH, basename+'.so') |
---|
| 32 | |
---|
| 33 | |
---|
| 34 | def load_model(kernel_module, dtype=None): |
---|
| 35 | """ |
---|
| 36 | Load the compiled model defined by *kernel_module*. |
---|
| 37 | |
---|
| 38 | Recompile if any files are newer than the model file. |
---|
| 39 | |
---|
| 40 | *dtype* is ignored. Compiled files are always double. |
---|
| 41 | |
---|
| 42 | The DLL is not loaded until the kernel is called so models an |
---|
| 43 | be defined without using too many resources. |
---|
| 44 | """ |
---|
| 45 | import tempfile |
---|
| 46 | |
---|
| 47 | source, info = gen.make(kernel_module) |
---|
| 48 | source_files = gen.sources(info) + [info['filename']] |
---|
| 49 | newest = max(os.path.getmtime(f) for f in source_files) |
---|
| 50 | dllpath = dll_path(info) |
---|
| 51 | if not os.path.exists(dllpath) or os.path.getmtime(dllpath)<newest: |
---|
| 52 | # Replace with a proper temp file |
---|
| 53 | fid, filename = tempfile.mkstemp(suffix=".c",prefix="sas_"+info['name']) |
---|
| 54 | os.fdopen(fid,"w").write(source) |
---|
| 55 | status = os.system(COMPILE%(filename, dllpath)) |
---|
| 56 | if status != 0: |
---|
| 57 | print "compile failed. File is in %r"%filename |
---|
| 58 | else: |
---|
| 59 | ## uncomment the following to keep the generated c file |
---|
| 60 | #os.unlink(filename); print "saving compiled file in %r"%filename |
---|
| 61 | pass |
---|
| 62 | return DllModel(dllpath, info) |
---|
| 63 | |
---|
[14de349] | 64 | |
---|
| 65 | IQ_ARGS = [c_void_p, c_void_p, c_int, c_void_p, c_double] |
---|
| 66 | IQXY_ARGS = [c_void_p, c_void_p, c_void_p, c_int, c_void_p, c_double] |
---|
| 67 | |
---|
| 68 | class DllModel(object): |
---|
| 69 | """ |
---|
| 70 | ctypes wrapper for a single model. |
---|
| 71 | |
---|
[ce27e21] | 72 | *source* and *info* are the model source and interface as returned |
---|
[14de349] | 73 | from :func:`gen.make`. |
---|
| 74 | |
---|
| 75 | *dtype* is the desired model precision. Any numpy dtype for single |
---|
| 76 | or double precision floats will do, such as 'f', 'float32' or 'single' |
---|
| 77 | for single and 'd', 'float64' or 'double' for double. Double precision |
---|
| 78 | is an optional extension which may not be available on all devices. |
---|
[ff7119b] | 79 | |
---|
| 80 | Call :meth:`release` when done with the kernel. |
---|
[14de349] | 81 | """ |
---|
[ce27e21] | 82 | def __init__(self, dllpath, info): |
---|
| 83 | self.info = info |
---|
| 84 | self.dllpath = dllpath |
---|
| 85 | self.dll = None |
---|
[14de349] | 86 | |
---|
[ce27e21] | 87 | def _load_dll(self): |
---|
| 88 | Nfixed1d = len(self.info['partype']['fixed-1d']) |
---|
| 89 | Nfixed2d = len(self.info['partype']['fixed-2d']) |
---|
| 90 | Npd1d = len(self.info['partype']['pd-1d']) |
---|
| 91 | Npd2d = len(self.info['partype']['pd-2d']) |
---|
[14de349] | 92 | |
---|
[ce27e21] | 93 | self.dll = ct.CDLL(self.dllpath) |
---|
[14de349] | 94 | |
---|
[ce27e21] | 95 | self.Iq = self.dll[gen.kernel_name(self.info, False)] |
---|
| 96 | self.Iq.argtypes = IQ_ARGS + [c_double]*Nfixed1d + [c_int]*Npd1d |
---|
| 97 | |
---|
| 98 | self.Iqxy = self.dll[gen.kernel_name(self.info, True)] |
---|
| 99 | self.Iqxy.argtypes = IQXY_ARGS + [c_double]*Nfixed2d + [c_int]*Npd2d |
---|
| 100 | |
---|
| 101 | def __getstate__(self): |
---|
| 102 | return {'info': self.info, 'dllpath': self.dllpath, 'dll': None} |
---|
| 103 | |
---|
| 104 | def __setstate__(self, state): |
---|
| 105 | self.__dict__ = state |
---|
| 106 | |
---|
| 107 | def __call__(self, input): |
---|
| 108 | if self.dll is None: self._load_dll() |
---|
[14de349] | 109 | |
---|
| 110 | kernel = self.Iqxy if input.is_2D else self.Iq |
---|
[ce27e21] | 111 | return DllKernel(kernel, self.info, input) |
---|
[14de349] | 112 | |
---|
| 113 | def make_input(self, q_vectors): |
---|
| 114 | """ |
---|
| 115 | Make q input vectors available to the model. |
---|
| 116 | |
---|
| 117 | This only needs to be done once for all models that operate on the |
---|
| 118 | same input. So for example, if you are adding two different models |
---|
| 119 | together to compare to a data set, then only one model needs to |
---|
| 120 | needs to call make_input, so long as the models have the same dtype. |
---|
| 121 | """ |
---|
| 122 | return DllInput(q_vectors) |
---|
| 123 | |
---|
[ff7119b] | 124 | def release(self): |
---|
| 125 | pass # TODO: should release the dll |
---|
| 126 | |
---|
[14de349] | 127 | |
---|
| 128 | class DllInput(object): |
---|
| 129 | """ |
---|
| 130 | Make q data available to the gpu. |
---|
| 131 | |
---|
| 132 | *q_vectors* is a list of q vectors, which will be *[q]* for 1-D data, |
---|
| 133 | and *[qx, qy]* for 2-D data. Internally, the vectors will be reallocated |
---|
| 134 | to get the best performance on OpenCL, which may involve shifting and |
---|
| 135 | stretching the array to better match the memory architecture. Additional |
---|
| 136 | points will be evaluated with *q=1e-3*. |
---|
| 137 | |
---|
| 138 | *dtype* is the data type for the q vectors. The data type should be |
---|
| 139 | set to match that of the kernel, which is an attribute of |
---|
| 140 | :class:`GpuProgram`. Note that not all kernels support double |
---|
| 141 | precision, so even if the program was created for double precision, |
---|
| 142 | the *GpuProgram.dtype* may be single precision. |
---|
| 143 | |
---|
| 144 | Call :meth:`release` when complete. Even if not called directly, the |
---|
| 145 | buffer will be released when the data object is freed. |
---|
| 146 | """ |
---|
| 147 | def __init__(self, q_vectors): |
---|
| 148 | self.nq = q_vectors[0].size |
---|
| 149 | self.dtype = np.dtype('double') |
---|
| 150 | self.is_2D = (len(q_vectors) == 2) |
---|
| 151 | self.q_vectors = [np.ascontiguousarray(q,self.dtype) for q in q_vectors] |
---|
| 152 | self.q_pointers = [q.ctypes.data for q in q_vectors] |
---|
| 153 | |
---|
| 154 | def release(self): |
---|
| 155 | self.q_vectors = [] |
---|
| 156 | |
---|
| 157 | class DllKernel(object): |
---|
[ff7119b] | 158 | """ |
---|
| 159 | Callable SAS kernel. |
---|
| 160 | |
---|
| 161 | *kernel* is the DllKernel object to call. |
---|
| 162 | |
---|
| 163 | *info* is the module information |
---|
| 164 | |
---|
| 165 | *input* is the DllInput q vectors at which the kernel should be |
---|
| 166 | evaluated. |
---|
| 167 | |
---|
| 168 | The resulting call method takes the *pars*, a list of values for |
---|
| 169 | the fixed parameters to the kernel, and *pd_pars*, a list of (value,weight) |
---|
| 170 | vectors for the polydisperse parameters. *cutoff* determines the |
---|
| 171 | integration limits: any points with combined weight less than *cutoff* |
---|
| 172 | will not be calculated. |
---|
| 173 | |
---|
| 174 | Call :meth:`release` when done with the kernel instance. |
---|
| 175 | """ |
---|
[ce27e21] | 176 | def __init__(self, kernel, info, input): |
---|
[5d4777d] | 177 | self.info = info |
---|
[14de349] | 178 | self.input = input |
---|
| 179 | self.kernel = kernel |
---|
| 180 | self.res = np.empty(input.nq, input.dtype) |
---|
[ce27e21] | 181 | dim = '2d' if input.is_2D else '1d' |
---|
| 182 | self.fixed_pars = info['partype']['fixed-'+dim] |
---|
| 183 | self.pd_pars = info['partype']['pd-'+dim] |
---|
[14de349] | 184 | |
---|
[ce27e21] | 185 | # In dll kernel, but not in opencl kernel |
---|
| 186 | self.p_res = self.res.ctypes.data |
---|
[14de349] | 187 | |
---|
[ce27e21] | 188 | def __call__(self, pars, pd_pars, cutoff): |
---|
[14de349] | 189 | real = np.float32 if self.input.dtype == F32 else np.float64 |
---|
[ce27e21] | 190 | fixed = [real(p) for p in pars] |
---|
| 191 | cutoff = real(cutoff) |
---|
| 192 | loops = np.hstack(pd_pars) |
---|
| 193 | loops = np.ascontiguousarray(loops.T, self.input.dtype).flatten() |
---|
| 194 | loops_N = [np.uint32(len(p[0])) for p in pd_pars] |
---|
[14de349] | 195 | |
---|
[ce27e21] | 196 | nq = c_int(self.input.nq) |
---|
[14de349] | 197 | p_loops = loops.ctypes.data |
---|
[ce27e21] | 198 | args = self.input.q_pointers + [self.p_res, nq, p_loops, cutoff] + fixed + loops_N |
---|
[14de349] | 199 | #print pars |
---|
[ce27e21] | 200 | self.kernel(*args) |
---|
[14de349] | 201 | |
---|
| 202 | return self.res |
---|
| 203 | |
---|
| 204 | def release(self): |
---|
| 205 | pass |
---|