1 | """ |
---|
2 | GPU support through OpenCL |
---|
3 | |
---|
4 | There should be a single GPU environment running on the system. This |
---|
5 | environment is constructed on the first call to :func:`env`, and the |
---|
6 | same environment is returned on each call. |
---|
7 | |
---|
8 | After retrieving the environment, the next step is to create the kernel. |
---|
9 | This is done with a call to :meth:`GpuEnvironment.make_kernel`, which |
---|
10 | returns the type of data used by the kernel. |
---|
11 | |
---|
12 | Next a :class:`GpuData` object should be created with the correct kind |
---|
13 | of data. This data object can be used by multiple kernels, for example, |
---|
14 | if the target model is a weighted sum of multiple kernels. The data |
---|
15 | should include any extra evaluation points required to compute the proper |
---|
16 | data smearing. This need not match the square grid for 2D data if there |
---|
17 | is an index saying which q points are active. |
---|
18 | |
---|
19 | Together the GpuData, the program, and a device form a :class:`GpuKernel`. |
---|
20 | This kernel is used during fitting, receiving new sets of parameters and |
---|
21 | evaluating them. The output value is stored in an output buffer on the |
---|
22 | devices, where it can be combined with other structure factors and form |
---|
23 | factors and have instrumental resolution effects applied. |
---|
24 | |
---|
25 | |
---|
26 | """ |
---|
27 | import numpy as np |
---|
28 | import pyopencl as cl |
---|
29 | from pyopencl import mem_flags as mf |
---|
30 | |
---|
31 | from . import gen |
---|
32 | |
---|
33 | F32_DEFS = """\ |
---|
34 | #define REAL(x) (x##f) |
---|
35 | #define real float |
---|
36 | """ |
---|
37 | |
---|
38 | F64_DEFS = """\ |
---|
39 | #pragma OPENCL EXTENSION cl_khr_fp64: enable |
---|
40 | #define REAL(x) (x) |
---|
41 | #define real double |
---|
42 | """ |
---|
43 | |
---|
44 | # The max loops number is limited by the amount of local memory available |
---|
45 | # on the device. You don't want to make this value too big because it will |
---|
46 | # waste resources, nor too small because it may interfere with users trying |
---|
47 | # to do their polydispersity calculations. A value of 1024 should be much |
---|
48 | # larger than necessary given that cost grows as npts^k where k is the number |
---|
49 | # of polydisperse parameters. |
---|
50 | MAX_LOOPS = 2048 |
---|
51 | |
---|
52 | def load_model(kernel_module, dtype="single"): |
---|
53 | """ |
---|
54 | Load the OpenCL model defined by *kernel_module*. |
---|
55 | |
---|
56 | Access to the OpenCL device is delayed until the kernel is called |
---|
57 | so models can be defined without using too many resources. |
---|
58 | """ |
---|
59 | source, info = gen.make(kernel_module) |
---|
60 | ## for debugging, save source to a .cl file, edit it, and reload as model |
---|
61 | open(info['name']+'.cl','w').write(source) |
---|
62 | #source = open(info['name']+'.cl','r').read() |
---|
63 | return GpuModel(source, info, dtype) |
---|
64 | |
---|
65 | ENV = None |
---|
66 | def environment(): |
---|
67 | """ |
---|
68 | Returns a singleton :class:`GpuEnvironment`. |
---|
69 | |
---|
70 | This provides an OpenCL context and one queue per device. |
---|
71 | """ |
---|
72 | global ENV |
---|
73 | if ENV is None: |
---|
74 | ENV = GpuEnvironment() |
---|
75 | return ENV |
---|
76 | |
---|
77 | def has_double(device): |
---|
78 | """ |
---|
79 | Return true if device supports double precision. |
---|
80 | """ |
---|
81 | return "cl_khr_fp64" in device.extensions |
---|
82 | |
---|
83 | |
---|
84 | def _stretch_input(vector, dtype, extra=1e-3, boundary=128): |
---|
85 | """ |
---|
86 | Stretch an input vector to the correct boundary. |
---|
87 | |
---|
88 | Performance on the kernels can drop by a factor of two or more if the |
---|
89 | number of values to compute does not fall on a nice power of two |
---|
90 | boundary. A good choice for the boundary value is the |
---|
91 | min_data_type_align_size property of the OpenCL device. The usual |
---|
92 | value of 128 gives a working size as a multiple of 32. The trailing |
---|
93 | additional vector elements are given a value of *extra*, and so |
---|
94 | f(*extra*) will be computed for each of them. The returned array |
---|
95 | will thus be a subset of the computed array. |
---|
96 | """ |
---|
97 | boundary // dtype.itemsize |
---|
98 | remainder = vector.size%boundary |
---|
99 | size = vector.size + (boundary - remainder if remainder != 0 else 0) |
---|
100 | if size != vector.size: |
---|
101 | vector = np.hstack((vector, [extra]*(size-vector.size))) |
---|
102 | return np.ascontiguousarray(vector, dtype=dtype) |
---|
103 | |
---|
104 | |
---|
105 | def compile_model(context, source, dtype): |
---|
106 | """ |
---|
107 | Build a model to run on the gpu. |
---|
108 | |
---|
109 | Returns the compiled program and its type. The returned type will |
---|
110 | be float32 even if the desired type is float64 if any of the |
---|
111 | devices in the context do not support the cl_khr_fp64 extension. |
---|
112 | """ |
---|
113 | dtype = np.dtype(dtype) |
---|
114 | if dtype==gen.F64 and not all(has_double(d) for d in context.devices): |
---|
115 | raise RuntimeError("Double precision not supported for devices") |
---|
116 | |
---|
117 | header = F64_DEFS if dtype == gen.F64 else F32_DEFS |
---|
118 | # Note: USE_SINCOS makes the intel cpu slower under opencl |
---|
119 | if context.devices[0].type == cl.device_type.GPU: |
---|
120 | header += "#define USE_SINCOS\n" |
---|
121 | program = cl.Program(context, header+source).build() |
---|
122 | return program |
---|
123 | |
---|
124 | |
---|
125 | def make_result(self, size): |
---|
126 | self.res = np.empty(size, dtype=self.dtype) |
---|
127 | self.res_b = cl.Buffer(self.program.context, mf.READ_WRITE, self.res.nbytes) |
---|
128 | return self.res, self.res_b |
---|
129 | |
---|
130 | |
---|
131 | # for now, this returns one device in the context |
---|
132 | # TODO: create a context that contains all devices on all platforms |
---|
133 | class GpuEnvironment(object): |
---|
134 | """ |
---|
135 | GPU context, with possibly many devices, and one queue per device. |
---|
136 | """ |
---|
137 | def __init__(self): |
---|
138 | self.context = cl.create_some_context() |
---|
139 | self.queues = [cl.CommandQueue(self.context, d) |
---|
140 | for d in self.context.devices] |
---|
141 | self.boundary = max(d.min_data_type_align_size |
---|
142 | for d in self.context.devices) |
---|
143 | self.has_double = all(has_double(d) for d in self.context.devices) |
---|
144 | self.compiled = {} |
---|
145 | |
---|
146 | def compile_program(self, name, source, dtype): |
---|
147 | if name not in self.compiled: |
---|
148 | #print "compiling",name |
---|
149 | self.compiled[name] = compile_model(self.context, source, dtype) |
---|
150 | return self.compiled[name] |
---|
151 | |
---|
152 | def release_program(self, name): |
---|
153 | if name in self.compiled: |
---|
154 | self.compiled[name].release() |
---|
155 | del self.compiled[name] |
---|
156 | |
---|
157 | class GpuModel(object): |
---|
158 | """ |
---|
159 | GPU wrapper for a single model. |
---|
160 | |
---|
161 | *source* and *info* are the model source and interface as returned |
---|
162 | from :func:`gen.make`. |
---|
163 | |
---|
164 | *dtype* is the desired model precision. Any numpy dtype for single |
---|
165 | or double precision floats will do, such as 'f', 'float32' or 'single' |
---|
166 | for single and 'd', 'float64' or 'double' for double. Double precision |
---|
167 | is an optional extension which may not be available on all devices. |
---|
168 | """ |
---|
169 | def __init__(self, source, info, dtype=gen.F32): |
---|
170 | self.info = info |
---|
171 | self.source = source |
---|
172 | self.dtype = dtype |
---|
173 | self.program = None # delay program creation |
---|
174 | |
---|
175 | def __getstate__(self): |
---|
176 | state = self.__dict__.copy() |
---|
177 | state['program'] = None |
---|
178 | return state |
---|
179 | |
---|
180 | def __setstate__(self, state): |
---|
181 | self.__dict__ = state.copy() |
---|
182 | |
---|
183 | def __call__(self, input): |
---|
184 | if self.dtype != input.dtype: |
---|
185 | raise TypeError("data and kernel have different types") |
---|
186 | if self.program is None: |
---|
187 | self.program = environment().compile_program(self.info['name'],self.source, self.dtype) |
---|
188 | kernel_name = gen.kernel_name(self.info, input.is_2D) |
---|
189 | kernel = getattr(self.program, kernel_name) |
---|
190 | return GpuKernel(kernel, self.info, input) |
---|
191 | |
---|
192 | def release(self): |
---|
193 | if self.program is not None: |
---|
194 | environment().release_program(self.info['name']) |
---|
195 | self.program = None |
---|
196 | |
---|
197 | def make_input(self, q_vectors): |
---|
198 | """ |
---|
199 | Make q input vectors available to the model. |
---|
200 | |
---|
201 | This only needs to be done once for all models that operate on the |
---|
202 | same input. So for example, if you are adding two different models |
---|
203 | together to compare to a data set, then only one model needs to |
---|
204 | needs to call make_input, so long as the models have the same dtype. |
---|
205 | """ |
---|
206 | # Note: the weird interface, where make_input doesn't care which |
---|
207 | # model calls it, allows us to ask the model to define the data |
---|
208 | # and the caller does not need to know if it is opencl or ctypes. |
---|
209 | # The returned data object is opaque. |
---|
210 | return GpuInput(q_vectors, dtype=self.dtype) |
---|
211 | |
---|
212 | # TODO: check that we don't need a destructor for buffers which go out of scope |
---|
213 | class GpuInput(object): |
---|
214 | """ |
---|
215 | Make q data available to the gpu. |
---|
216 | |
---|
217 | *q_vectors* is a list of q vectors, which will be *[q]* for 1-D data, |
---|
218 | and *[qx, qy]* for 2-D data. Internally, the vectors will be reallocated |
---|
219 | to get the best performance on OpenCL, which may involve shifting and |
---|
220 | stretching the array to better match the memory architecture. Additional |
---|
221 | points will be evaluated with *q=1e-3*. |
---|
222 | |
---|
223 | *dtype* is the data type for the q vectors. The data type should be |
---|
224 | set to match that of the kernel, which is an attribute of |
---|
225 | :class:`GpuProgram`. Note that not all kernels support double |
---|
226 | precision, so even if the program was created for double precision, |
---|
227 | the *GpuProgram.dtype* may be single precision. |
---|
228 | |
---|
229 | Call :meth:`release` when complete. Even if not called directly, the |
---|
230 | buffer will be released when the data object is freed. |
---|
231 | """ |
---|
232 | def __init__(self, q_vectors, dtype=gen.F32): |
---|
233 | env = environment() |
---|
234 | self.nq = q_vectors[0].size |
---|
235 | self.dtype = np.dtype(dtype) |
---|
236 | self.is_2D = (len(q_vectors) == 2) |
---|
237 | self.q_vectors = [ |
---|
238 | _stretch_input(q, self.dtype, boundary=env.boundary) |
---|
239 | for q in q_vectors |
---|
240 | ] |
---|
241 | self.q_buffers = [ |
---|
242 | cl.Buffer(env.context, mf.READ_ONLY | mf.COPY_HOST_PTR, hostbuf=q) |
---|
243 | for q in self.q_vectors |
---|
244 | ] |
---|
245 | self.global_size = [self.q_vectors[0].size] |
---|
246 | |
---|
247 | def release(self): |
---|
248 | for b in self.q_buffers: |
---|
249 | b.release() |
---|
250 | self.q_buffers = [] |
---|
251 | |
---|
252 | class GpuKernel(object): |
---|
253 | """ |
---|
254 | Callable SAS kernel. |
---|
255 | |
---|
256 | *kernel* is the GpuKernel object to call. |
---|
257 | |
---|
258 | *info* is the module information |
---|
259 | |
---|
260 | *input* is the DllInput q vectors at which the kernel should be |
---|
261 | evaluated. |
---|
262 | |
---|
263 | The resulting call method takes the *pars*, a list of values for |
---|
264 | the fixed parameters to the kernel, and *pd_pars*, a list of (value,weight) |
---|
265 | vectors for the polydisperse parameters. *cutoff* determines the |
---|
266 | integration limits: any points with combined weight less than *cutoff* |
---|
267 | will not be calculated. |
---|
268 | |
---|
269 | Call :meth:`release` when done with the kernel instance. |
---|
270 | """ |
---|
271 | def __init__(self, kernel, info, input): |
---|
272 | self.input = input |
---|
273 | self.kernel = kernel |
---|
274 | self.info = info |
---|
275 | self.res = np.empty(input.nq, input.dtype) |
---|
276 | dim = '2d' if input.is_2D else '1d' |
---|
277 | self.fixed_pars = info['partype']['fixed-'+dim] |
---|
278 | self.pd_pars = info['partype']['pd-'+dim] |
---|
279 | |
---|
280 | # Inputs and outputs for each kernel call |
---|
281 | # Note: res may be shorter than res_b if global_size != nq |
---|
282 | env = environment() |
---|
283 | self.loops_b = [cl.Buffer(env.context, mf.READ_WRITE, |
---|
284 | 2*MAX_LOOPS*input.dtype.itemsize) |
---|
285 | for _ in env.queues] |
---|
286 | self.res_b = [cl.Buffer(env.context, mf.READ_WRITE, |
---|
287 | input.global_size[0]*input.dtype.itemsize) |
---|
288 | for _ in env.queues] |
---|
289 | |
---|
290 | |
---|
291 | def __call__(self, pars, pd_pars, cutoff=1e-5): |
---|
292 | real = np.float32 if self.input.dtype == gen.F32 else np.float64 |
---|
293 | fixed = [real(p) for p in pars] |
---|
294 | cutoff = real(cutoff) |
---|
295 | loops = np.hstack(pd_pars) |
---|
296 | loops = np.ascontiguousarray(loops.T, self.input.dtype).flatten() |
---|
297 | Nloops = [np.uint32(len(p[0])) for p in pd_pars] |
---|
298 | #print "loops",Nloops, loops |
---|
299 | |
---|
300 | #import sys; print >>sys.stderr,"opencl eval",pars |
---|
301 | #print "opencl eval",pars |
---|
302 | if len(loops) > 2*MAX_LOOPS: |
---|
303 | raise ValueError("too many polydispersity points") |
---|
304 | device_num = 0 |
---|
305 | res_bi = self.res_b[device_num] |
---|
306 | queuei = environment().queues[device_num] |
---|
307 | loops_bi = self.loops_b[device_num] |
---|
308 | loops_l = cl.LocalMemory(len(loops.data)) |
---|
309 | cl.enqueue_copy(queuei, loops_bi, loops) |
---|
310 | #ctx = environment().context |
---|
311 | #loops_bi = cl.Buffer(ctx, mf.READ_ONLY | mf.COPY_HOST_PTR, hostbuf=loops) |
---|
312 | args = self.input.q_buffers + [res_bi,loops_bi,loops_l,cutoff] + fixed + Nloops |
---|
313 | self.kernel(queuei, self.input.global_size, None, *args) |
---|
314 | cl.enqueue_copy(queuei, self.res, res_bi) |
---|
315 | |
---|
316 | return self.res |
---|
317 | |
---|
318 | def release(self): |
---|
319 | for b in self.loops_b: |
---|
320 | b.release() |
---|
321 | self.loops_b = [] |
---|
322 | for b in self.res_b: |
---|
323 | b.release() |
---|
324 | self.res_b = [] |
---|
325 | |
---|
326 | def __del__(self): |
---|
327 | self.release() |
---|