Changes in / [b26d4c8:119bd3d] in sasmodels
- Files:
-
- 3 edited
Legend:
- Unmodified
- Added
- Removed
-
doc/developer/index.rst
rb85be2d r56fc97a 3 3 *************************** 4 4 5 .. toctree::6 :numbered: 47 :maxdepth: 48 5 9 calculator.rst -
sasmodels/core.py
rcf52f9c r35647ab 216 216 with an error of about 1%, which is usually less than the measurement 217 217 uncertainty. 218 219 *mono* is True if polydispersity should be set to none on all parameters. 220 """ 218 """ 219 print pars 221 220 fixed_pars = [pars.get(name, kernel.info['defaults'][name]) 222 221 for name in kernel.fixed_pars] -
sasmodels/kernel_iq.c
r4b2972c r119bd3d 12 12 */ 13 13 14 /* 15 The environment needs to provide the following #defines: 16 17 USE_OPENCL is defined if running in opencl 18 KERNEL declares a function to be available externally 19 KERNEL_NAME is the name of the function being declared 20 NPARS is the number of parameters in the kernel 21 PARAMETER_DECL is the declaration of the parameters to the kernel. 22 23 Cylinder: 24 25 #define PARAMETER_DECL \ 26 double length; \ 27 double radius; \ 28 double sld; \ 29 double sld_solvent 30 31 Note: scale and background are not included 32 33 Multi-shell cylinder (10 shell max): 34 35 #define PARAMETER_DECL \ 36 double num_shells; \ 37 double length; \ 38 double radius[10]; \ 39 double sld[10]; \ 40 double sld_solvent 41 42 CALL_IQ(q, nq, i, pars) is the declaration of a call to the kernel. 43 44 Cylinder: 45 46 #define CALL_IQ(q, nq, i, var) \ 47 Iq(q[i], \ 48 var.length, \ 49 var.radius, \ 50 var.sld, \ 51 var.sld_solvent) 52 53 Multi-shell cylinder: 54 #define CALL_IQ(q, nq, i, var) \ 55 Iq(q[i], \ 56 var.num_shells, \ 57 var.length, \ 58 var.radius, \ 59 var.sld, \ 60 var.sld_solvent) 61 62 CALL_VOLUME(var) is similar, but for calling the form volume. 63 64 INVALID is a test for model parameters in the correct range 65 66 Cylinder: 67 68 #define INVALID(var) 0 69 70 BarBell: 71 72 #define INVALID(var) (var.bell_radius > var.radius) 73 74 Model with complicated constraints: 75 76 inline bool constrained(p1, p2, p3) { return expression; } 77 #define INVALID(var) constrained(var.p1, var.p2, var.p3) 78 79 IQ_FUNC could be Iq or Iqxy 80 IQ_PARS could be q[i] or q[2*i],q[2*i+1] 81 82 Our design supports a limited number of polydispersity loops, wherein 83 we need to cycle through the values of the polydispersity, calculate 84 the I(q, p) for each combination of parameters, and perform a normalized 85 weighted sum across all the weights. Parameters may be passed to the 86 underlying calculation engine as scalars or vectors, but the polydispersity 87 calculator treats the parameter set as one long vector. 88 89 Let's assume we have 6 parameters in the model, with two polydisperse:: 90 91 0: scale {scl = constant} 92 1: background {bkg = constant} 93 5: length {l = vector of 30pts} 94 4: radius {r = vector of 10pts} 95 3: sld {s = constant/(radius**2*length)} 96 2: sld_solvent {s2 = constant} 97 98 This generates the following call to the kernel (where x stands for an 99 arbitrary value that is not used by the kernel evaluator): 100 101 NPARS = 4 // scale and background are in all models 102 problem { 103 pd_par = {5, 4, x, x} // parameters *radius* and *length* vary 104 pd_length = {30, 10, 0, 0} // *length* has more, so it is first 105 pd_offset = {10, 0, x, x} // *length* starts at index 10 in weights 106 pd_stride = {1, 30, 300, 300} // cumulative product of pd length 107 pd_isvol = {1, 1, x, x} // true if weight is a volume weight 108 par_offset = {2, 3, 303, 313} // parameter offsets 109 par_coord = {0, 3, 2, 1} // bitmap of parameter dependencies 110 fast_coord_index = {5, 3, x, x} 111 fast_coord_count = 2 // two parameters vary with *length* distribution 112 theta_var = -1 // no spherical correction 113 fast_theta = 0 // spherical correction angle is not pd 1 114 } 115 116 weight = { l0, .., l29, r0, .., r9} 117 pars = { scl, bkg, l0, ..., l29, r0, r1, ..., r9, 118 s[l0,r0], ... s[l0,r9], s[l1,r0], ... s[l29,r9] , s2} 119 120 nq = 130 121 q = { q0, q1, ..., q130, x, x } # pad to 8 element boundary 122 result = {r1, ..., r130, norm, vol, vol_norm, x, x, x, x, x, x, x} 123 124 125 The polydisperse parameters are stored in as an array of parameter 126 indices, one for each polydisperse parameter, stored in pd_par[n]. 127 Non-polydisperse parameters do not appear in this array. Each polydisperse 128 parameter has a weight vector whose length is stored in pd_length[n], 129 The weights are stored in a contiguous vector of weights for all 130 parameters, with the starting position for the each parameter stored 131 in pd_offset[n]. The values corresponding to the weights are stored 132 together in a separate weights[] vector, with offset stored in 133 par_offset[pd_par[n]]. Polydisperse parameters should be stored in 134 decreasing order of length for highest efficiency. 135 136 We limit the number of polydisperse dimensions to MAX_PD (currently 4). 137 This cuts the size of the structure in half compared to allowing a 138 separate polydispersity for each parameter. This will help a little 139 bit for models with large numbers of parameters, such as the onion model. 140 141 Parameters may be coordinated. That is, we may have the value of one 142 parameter depend on a set of other parameters, some of which may be 143 polydisperse. For example, if sld is inversely proportional to the 144 volume of a cylinder, and the length and radius are independently 145 polydisperse, then for each combination of length and radius we need a 146 separate value for the sld. The caller must provide a coordination table 147 for each parameter containing the value for each parameter given the 148 value of the polydisperse parameters v1, v2, etc. The tables for each 149 parameter are arranged contiguously in a vector, with offset[k] giving the 150 starting location of parameter k in the vector. Each parameter defines 151 coord[k] as a bit mask indicating which polydispersity parameters the 152 parameter depends upon. Usually this is zero, indicating that the parameter 153 is independent, but for the cylinder example given, the bits for the 154 radius and length polydispersity parameters would both be set, the result 155 being a (#radius x #length) table, or maybe a (#length x #radius) table 156 if length comes first in the polydispersity table. 157 158 NB: If we can guarantee that a compiler and OpenCL driver are available, 159 we could instead create the coordination function on the fly for each 160 parameter, saving memory and transfer time, but requiring a C compiler 161 as part of the environment. 162 163 In ordering the polydisperse parameters by decreasing length we can 164 iterate over the longest dispersion weight vector first. All parameters 165 coordinated with this weight vector (the 'fast' parameters), can be 166 updated with a simple increment to the next position in the parameter 167 value table. The indices of these parameters is stored in fast_coord_index[], 168 with fast_coord_count being the number of fast parameters. A total 169 of NPARS slots is allocated to allow for the case that all parameters 170 are coordinated with the fast index, though this will likely be mostly 171 empty. When the fast increment count reaches the end of the weight 172 vector, then the index of the second polydisperse parameter must be 173 incremented, and all of its coordinated parameters updated. Because this 174 operation is not in the inner loop, a slower algorithm can be used. 175 176 If there is no polydispersity we pretend that it is polydisperisty with one 177 parameter, pd_start=0 and pd_stop=1. We may or may not short circuit the 178 calculation in this case, depending on how much time it saves. 179 180 The problem details structure can be allocated and sent in as an integer 181 array using the read-only flag. This allows us to copy it once per fit 182 along with the weights vector, since features such as the number of 183 polydisperity elements per pd parameter or the coordinated won't change 184 between function evaluations. A new parameter vector is sent for 185 each I(q) evaluation. 186 187 To protect against expensive evaluations taking all the GPU resource 188 on large fits, the entire polydispersity will not be computed at once. 189 Instead, a start and stop location will be sent, indicating where in the 190 polydispersity loop the calculation should start and where it should 191 stop. We can do this for arbitrary start/stop points since we have 192 unwound the nested loop. Instead, we use the same technique as array 193 index translation, using div and mod to figure out the i,j,k,... 194 indices in the virtual nested loop. 195 196 The results array will be initialized to zero for polydispersity loop 197 entry zero, and preserved between calls to [start, stop] so that the 198 results accumulate by the time the loop has completed. Background and 199 scale will be applied when the loop reaches the end. This does require 200 that the results array be allocated read-write, which is less efficient 201 for the GPU, but it makes the calling sequence much more manageable. 202 203 Scale and background cannot be coordinated with other polydisperse parameters 204 205 Oriented objects in 2-D need a spherical correction on the angular variation 206 in order to preserve the 'surface area' of the weight distribution. 207 208 Cutoff specifies the integration area by discarding regions in polydisperisty 209 hypercubue that has no parameters defined 210 */ 14 211 15 212 #define MAX_PD 4 // MAX_PD is the max number of polydisperse parameters … … 34 231 } ParameterBlock; 35 232 36 #define FULL_KERNEL_NAME KERNEL_NAME ## _ ## IQ_FUNC 37 KERNEL 233 #define KERNEL_NAME test_Iq 234 #define FULL_KERNEL_NAME test_Iq 235 #define IQ_FUNC Iq 236 237 38 238 void FULL_KERNEL_NAME( 39 239 int nq, // number of q values … … 204 404 } 205 405 } 406 } 407 }
Note: See TracChangeset
for help on using the changeset viewer.