[14de349] | 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 | """ |
---|
[250fa25] | 25 | import os |
---|
| 26 | import warnings |
---|
| 27 | |
---|
[14de349] | 28 | import numpy as np |
---|
[b3f6bc3] | 29 | |
---|
[250fa25] | 30 | try: |
---|
| 31 | import pyopencl as cl |
---|
| 32 | except ImportError,exc: |
---|
| 33 | warnings.warn(str(exc)) |
---|
| 34 | raise RuntimeError("OpenCL not available") |
---|
| 35 | |
---|
[14de349] | 36 | from pyopencl import mem_flags as mf |
---|
| 37 | |
---|
[cb6ecf4] | 38 | from . import generate |
---|
[f786ff3] | 39 | from .kernelpy import PyInput, PyKernel |
---|
[14de349] | 40 | |
---|
| 41 | F64_DEFS = """\ |
---|
[d087487b] | 42 | #ifdef cl_khr_fp64 |
---|
| 43 | # pragma OPENCL EXTENSION cl_khr_fp64: enable |
---|
| 44 | #endif |
---|
[14de349] | 45 | """ |
---|
| 46 | |
---|
[ce27e21] | 47 | # The max loops number is limited by the amount of local memory available |
---|
| 48 | # on the device. You don't want to make this value too big because it will |
---|
| 49 | # waste resources, nor too small because it may interfere with users trying |
---|
| 50 | # to do their polydispersity calculations. A value of 1024 should be much |
---|
| 51 | # larger than necessary given that cost grows as npts^k where k is the number |
---|
| 52 | # of polydisperse parameters. |
---|
[5d4777d] | 53 | MAX_LOOPS = 2048 |
---|
| 54 | |
---|
| 55 | def load_model(kernel_module, dtype="single"): |
---|
| 56 | """ |
---|
| 57 | Load the OpenCL model defined by *kernel_module*. |
---|
| 58 | |
---|
| 59 | Access to the OpenCL device is delayed until the kernel is called |
---|
| 60 | so models can be defined without using too many resources. |
---|
| 61 | """ |
---|
[cb6ecf4] | 62 | source, info = generate.make(kernel_module) |
---|
[5d4777d] | 63 | ## for debugging, save source to a .cl file, edit it, and reload as model |
---|
[f4cf580] | 64 | #open(info['name']+'.cl','w').write(source) |
---|
[5d4777d] | 65 | #source = open(info['name']+'.cl','r').read() |
---|
| 66 | return GpuModel(source, info, dtype) |
---|
[ce27e21] | 67 | |
---|
[14de349] | 68 | ENV = None |
---|
| 69 | def environment(): |
---|
| 70 | """ |
---|
| 71 | Returns a singleton :class:`GpuEnvironment`. |
---|
| 72 | |
---|
| 73 | This provides an OpenCL context and one queue per device. |
---|
| 74 | """ |
---|
| 75 | global ENV |
---|
| 76 | if ENV is None: |
---|
| 77 | ENV = GpuEnvironment() |
---|
| 78 | return ENV |
---|
| 79 | |
---|
| 80 | def has_double(device): |
---|
| 81 | """ |
---|
| 82 | Return true if device supports double precision. |
---|
| 83 | """ |
---|
| 84 | return "cl_khr_fp64" in device.extensions |
---|
| 85 | |
---|
[f5b9a6b] | 86 | def get_warp(kernel, queue): |
---|
| 87 | """ |
---|
| 88 | Return the size of an execution batch for *kernel* running on *queue*. |
---|
| 89 | """ |
---|
| 90 | return kernel.get_work_group_info(cl.kernel_work_group_info.PREFERRED_WORK_GROUP_SIZE_MULTIPLE, |
---|
| 91 | queue.device) |
---|
[14de349] | 92 | |
---|
[f5b9a6b] | 93 | def _stretch_input(vector, dtype, extra=1e-3, boundary=32): |
---|
[14de349] | 94 | """ |
---|
| 95 | Stretch an input vector to the correct boundary. |
---|
| 96 | |
---|
| 97 | Performance on the kernels can drop by a factor of two or more if the |
---|
| 98 | number of values to compute does not fall on a nice power of two |
---|
[f5b9a6b] | 99 | boundary. The trailing additional vector elements are given a |
---|
| 100 | value of *extra*, and so f(*extra*) will be computed for each of |
---|
| 101 | them. The returned array will thus be a subset of the computed array. |
---|
| 102 | |
---|
| 103 | *boundary* should be a power of 2 which is at least 32 for good |
---|
| 104 | performance on current platforms (as of Jan 2015). It should |
---|
| 105 | probably be the max of get_warp(kernel,queue) and |
---|
| 106 | device.min_data_type_align_size//4. |
---|
| 107 | """ |
---|
[14de349] | 108 | remainder = vector.size%boundary |
---|
[f5b9a6b] | 109 | if remainder != 0: |
---|
| 110 | size = vector.size + (boundary - remainder) |
---|
[14de349] | 111 | vector = np.hstack((vector, [extra]*(size-vector.size))) |
---|
| 112 | return np.ascontiguousarray(vector, dtype=dtype) |
---|
| 113 | |
---|
| 114 | |
---|
| 115 | def compile_model(context, source, dtype): |
---|
| 116 | """ |
---|
| 117 | Build a model to run on the gpu. |
---|
| 118 | |
---|
| 119 | Returns the compiled program and its type. The returned type will |
---|
| 120 | be float32 even if the desired type is float64 if any of the |
---|
| 121 | devices in the context do not support the cl_khr_fp64 extension. |
---|
| 122 | """ |
---|
| 123 | dtype = np.dtype(dtype) |
---|
[cb6ecf4] | 124 | if dtype==generate.F64 and not all(has_double(d) for d in context.devices): |
---|
[ce27e21] | 125 | raise RuntimeError("Double precision not supported for devices") |
---|
[14de349] | 126 | |
---|
[cb6ecf4] | 127 | header = F64_DEFS if dtype == generate.F64 else "" |
---|
| 128 | if dtype == generate.F32: |
---|
| 129 | source = generate.use_single(source) |
---|
[14de349] | 130 | # Note: USE_SINCOS makes the intel cpu slower under opencl |
---|
| 131 | if context.devices[0].type == cl.device_type.GPU: |
---|
| 132 | header += "#define USE_SINCOS\n" |
---|
| 133 | program = cl.Program(context, header+source).build() |
---|
[ce27e21] | 134 | return program |
---|
[14de349] | 135 | |
---|
| 136 | |
---|
| 137 | def make_result(self, size): |
---|
| 138 | self.res = np.empty(size, dtype=self.dtype) |
---|
| 139 | self.res_b = cl.Buffer(self.program.context, mf.READ_WRITE, self.res.nbytes) |
---|
| 140 | return self.res, self.res_b |
---|
| 141 | |
---|
| 142 | |
---|
| 143 | # for now, this returns one device in the context |
---|
| 144 | # TODO: create a context that contains all devices on all platforms |
---|
| 145 | class GpuEnvironment(object): |
---|
| 146 | """ |
---|
| 147 | GPU context, with possibly many devices, and one queue per device. |
---|
| 148 | """ |
---|
| 149 | def __init__(self): |
---|
[250fa25] | 150 | # find gpu context |
---|
| 151 | #self.context = cl.create_some_context() |
---|
| 152 | |
---|
| 153 | self.context = None |
---|
| 154 | if 'PYOPENCL_CTX' in os.environ: |
---|
| 155 | self._create_some_context() |
---|
| 156 | |
---|
| 157 | if not self.context: |
---|
| 158 | self.context = self._find_context() |
---|
| 159 | |
---|
[f5b9a6b] | 160 | # Byte boundary for data alignment |
---|
| 161 | #self.data_boundary = max(d.min_data_type_align_size |
---|
| 162 | # for d in self.context.devices) |
---|
[250fa25] | 163 | self.queues = [cl.CommandQueue(self.context, d) |
---|
| 164 | for d in self.context.devices] |
---|
[ce27e21] | 165 | self.has_double = all(has_double(d) for d in self.context.devices) |
---|
| 166 | self.compiled = {} |
---|
| 167 | |
---|
[250fa25] | 168 | def _create_some_context(self): |
---|
| 169 | try: |
---|
| 170 | self.context = cl.create_some_context(interactive=False) |
---|
| 171 | except Exception,exc: |
---|
| 172 | warnings.warn(str(exc)) |
---|
| 173 | warnings.warn("pyopencl.create_some_context() failed") |
---|
| 174 | warnings.warn("the environment variable 'PYOPENCL_CTX' might not be set correctly") |
---|
| 175 | |
---|
| 176 | def _find_context(self): |
---|
| 177 | default = None |
---|
| 178 | for platform in cl.get_platforms(): |
---|
| 179 | for device in platform.get_devices(): |
---|
| 180 | if device.type == cl.device_type.GPU: |
---|
| 181 | return cl.Context([device]) |
---|
| 182 | if default is None: |
---|
| 183 | default = device |
---|
| 184 | |
---|
| 185 | if not default: |
---|
| 186 | raise RuntimeError("OpenCL device not found") |
---|
| 187 | |
---|
| 188 | return cl.Context([default]) |
---|
| 189 | |
---|
[ce27e21] | 190 | def compile_program(self, name, source, dtype): |
---|
| 191 | if name not in self.compiled: |
---|
| 192 | #print "compiling",name |
---|
| 193 | self.compiled[name] = compile_model(self.context, source, dtype) |
---|
| 194 | return self.compiled[name] |
---|
| 195 | |
---|
| 196 | def release_program(self, name): |
---|
| 197 | if name in self.compiled: |
---|
| 198 | self.compiled[name].release() |
---|
| 199 | del self.compiled[name] |
---|
[14de349] | 200 | |
---|
[250fa25] | 201 | |
---|
[14de349] | 202 | class GpuModel(object): |
---|
| 203 | """ |
---|
| 204 | GPU wrapper for a single model. |
---|
| 205 | |
---|
[ce27e21] | 206 | *source* and *info* are the model source and interface as returned |
---|
[14de349] | 207 | from :func:`gen.make`. |
---|
| 208 | |
---|
| 209 | *dtype* is the desired model precision. Any numpy dtype for single |
---|
| 210 | or double precision floats will do, such as 'f', 'float32' or 'single' |
---|
| 211 | for single and 'd', 'float64' or 'double' for double. Double precision |
---|
| 212 | is an optional extension which may not be available on all devices. |
---|
| 213 | """ |
---|
[cb6ecf4] | 214 | def __init__(self, source, info, dtype=generate.F32): |
---|
[ce27e21] | 215 | self.info = info |
---|
| 216 | self.source = source |
---|
| 217 | self.dtype = dtype |
---|
| 218 | self.program = None # delay program creation |
---|
[14de349] | 219 | |
---|
[ce27e21] | 220 | def __getstate__(self): |
---|
| 221 | state = self.__dict__.copy() |
---|
| 222 | state['program'] = None |
---|
| 223 | return state |
---|
[14de349] | 224 | |
---|
[ce27e21] | 225 | def __setstate__(self, state): |
---|
| 226 | self.__dict__ = state.copy() |
---|
| 227 | |
---|
| 228 | def __call__(self, input): |
---|
[b3f6bc3] | 229 | # Support pure python kernel call |
---|
| 230 | if input.is_2D and callable(self.info['Iqxy']): |
---|
| 231 | return PyKernel(self.info['Iqxy'], self.info, input) |
---|
| 232 | elif not input.is_2D and callable(self.info['Iq']): |
---|
| 233 | return PyKernel(self.info['Iq'], self.info, input) |
---|
| 234 | |
---|
[14de349] | 235 | if self.dtype != input.dtype: |
---|
| 236 | raise TypeError("data and kernel have different types") |
---|
[ce27e21] | 237 | if self.program is None: |
---|
| 238 | self.program = environment().compile_program(self.info['name'],self.source, self.dtype) |
---|
[cb6ecf4] | 239 | kernel_name = generate.kernel_name(self.info, input.is_2D) |
---|
[14de349] | 240 | kernel = getattr(self.program, kernel_name) |
---|
[ce27e21] | 241 | return GpuKernel(kernel, self.info, input) |
---|
| 242 | |
---|
| 243 | def release(self): |
---|
| 244 | if self.program is not None: |
---|
| 245 | environment().release_program(self.info['name']) |
---|
| 246 | self.program = None |
---|
[14de349] | 247 | |
---|
| 248 | def make_input(self, q_vectors): |
---|
| 249 | """ |
---|
| 250 | Make q input vectors available to the model. |
---|
| 251 | |
---|
[b3f6bc3] | 252 | Note that each model needs its own q vector even if the case of |
---|
| 253 | mixture models because some models may be OpenCL, some may be |
---|
| 254 | ctypes and some may be pure python. |
---|
[14de349] | 255 | """ |
---|
[b3f6bc3] | 256 | # Support pure python kernel call |
---|
| 257 | if len(q_vectors) == 1 and callable(self.info['Iq']): |
---|
| 258 | return PyInput(q_vectors, dtype=self.dtype) |
---|
| 259 | elif callable(self.info['Iqxy']): |
---|
| 260 | return PyInput(q_vectors, dtype=self.dtype) |
---|
| 261 | else: |
---|
| 262 | return GpuInput(q_vectors, dtype=self.dtype) |
---|
[14de349] | 263 | |
---|
| 264 | # TODO: check that we don't need a destructor for buffers which go out of scope |
---|
| 265 | class GpuInput(object): |
---|
| 266 | """ |
---|
| 267 | Make q data available to the gpu. |
---|
| 268 | |
---|
| 269 | *q_vectors* is a list of q vectors, which will be *[q]* for 1-D data, |
---|
| 270 | and *[qx, qy]* for 2-D data. Internally, the vectors will be reallocated |
---|
| 271 | to get the best performance on OpenCL, which may involve shifting and |
---|
| 272 | stretching the array to better match the memory architecture. Additional |
---|
| 273 | points will be evaluated with *q=1e-3*. |
---|
| 274 | |
---|
| 275 | *dtype* is the data type for the q vectors. The data type should be |
---|
| 276 | set to match that of the kernel, which is an attribute of |
---|
| 277 | :class:`GpuProgram`. Note that not all kernels support double |
---|
| 278 | precision, so even if the program was created for double precision, |
---|
| 279 | the *GpuProgram.dtype* may be single precision. |
---|
| 280 | |
---|
| 281 | Call :meth:`release` when complete. Even if not called directly, the |
---|
| 282 | buffer will be released when the data object is freed. |
---|
| 283 | """ |
---|
[cb6ecf4] | 284 | def __init__(self, q_vectors, dtype=generate.F32): |
---|
[14de349] | 285 | env = environment() |
---|
| 286 | self.nq = q_vectors[0].size |
---|
| 287 | self.dtype = np.dtype(dtype) |
---|
| 288 | self.is_2D = (len(q_vectors) == 2) |
---|
[f5b9a6b] | 289 | # TODO: stretch input based on get_warp() |
---|
| 290 | # not doing it now since warp depends on kernel, which is not known |
---|
| 291 | # at this point, so instead using 32, which is good on the set of |
---|
| 292 | # architectures tested so far. |
---|
| 293 | self.q_vectors = [_stretch_input(q, self.dtype, 32) for q in q_vectors] |
---|
[14de349] | 294 | self.q_buffers = [ |
---|
| 295 | cl.Buffer(env.context, mf.READ_ONLY | mf.COPY_HOST_PTR, hostbuf=q) |
---|
| 296 | for q in self.q_vectors |
---|
| 297 | ] |
---|
| 298 | self.global_size = [self.q_vectors[0].size] |
---|
| 299 | |
---|
| 300 | def release(self): |
---|
| 301 | for b in self.q_buffers: |
---|
| 302 | b.release() |
---|
| 303 | self.q_buffers = [] |
---|
| 304 | |
---|
| 305 | class GpuKernel(object): |
---|
[ff7119b] | 306 | """ |
---|
| 307 | Callable SAS kernel. |
---|
| 308 | |
---|
| 309 | *kernel* is the GpuKernel object to call. |
---|
| 310 | |
---|
| 311 | *info* is the module information |
---|
| 312 | |
---|
| 313 | *input* is the DllInput q vectors at which the kernel should be |
---|
| 314 | evaluated. |
---|
| 315 | |
---|
| 316 | The resulting call method takes the *pars*, a list of values for |
---|
| 317 | the fixed parameters to the kernel, and *pd_pars*, a list of (value,weight) |
---|
| 318 | vectors for the polydisperse parameters. *cutoff* determines the |
---|
| 319 | integration limits: any points with combined weight less than *cutoff* |
---|
| 320 | will not be calculated. |
---|
| 321 | |
---|
| 322 | Call :meth:`release` when done with the kernel instance. |
---|
| 323 | """ |
---|
[ce27e21] | 324 | def __init__(self, kernel, info, input): |
---|
[14de349] | 325 | self.input = input |
---|
| 326 | self.kernel = kernel |
---|
[ce27e21] | 327 | self.info = info |
---|
| 328 | self.res = np.empty(input.nq, input.dtype) |
---|
| 329 | dim = '2d' if input.is_2D else '1d' |
---|
| 330 | self.fixed_pars = info['partype']['fixed-'+dim] |
---|
| 331 | self.pd_pars = info['partype']['pd-'+dim] |
---|
[14de349] | 332 | |
---|
| 333 | # Inputs and outputs for each kernel call |
---|
[ce27e21] | 334 | # Note: res may be shorter than res_b if global_size != nq |
---|
| 335 | env = environment() |
---|
[14de349] | 336 | self.loops_b = [cl.Buffer(env.context, mf.READ_WRITE, |
---|
[5d4777d] | 337 | 2*MAX_LOOPS*input.dtype.itemsize) |
---|
[14de349] | 338 | for _ in env.queues] |
---|
| 339 | self.res_b = [cl.Buffer(env.context, mf.READ_WRITE, |
---|
| 340 | input.global_size[0]*input.dtype.itemsize) |
---|
| 341 | for _ in env.queues] |
---|
| 342 | |
---|
| 343 | |
---|
[ce27e21] | 344 | def __call__(self, pars, pd_pars, cutoff=1e-5): |
---|
[cb6ecf4] | 345 | real = np.float32 if self.input.dtype == generate.F32 else np.float64 |
---|
[ce27e21] | 346 | fixed = [real(p) for p in pars] |
---|
| 347 | cutoff = real(cutoff) |
---|
| 348 | loops = np.hstack(pd_pars) |
---|
| 349 | loops = np.ascontiguousarray(loops.T, self.input.dtype).flatten() |
---|
[5d4777d] | 350 | Nloops = [np.uint32(len(p[0])) for p in pd_pars] |
---|
| 351 | #print "loops",Nloops, loops |
---|
[ce27e21] | 352 | |
---|
[1780d59] | 353 | #import sys; print >>sys.stderr,"opencl eval",pars |
---|
[ce27e21] | 354 | #print "opencl eval",pars |
---|
[5d4777d] | 355 | if len(loops) > 2*MAX_LOOPS: |
---|
[ce27e21] | 356 | raise ValueError("too many polydispersity points") |
---|
[14de349] | 357 | device_num = 0 |
---|
| 358 | res_bi = self.res_b[device_num] |
---|
| 359 | queuei = environment().queues[device_num] |
---|
| 360 | loops_bi = self.loops_b[device_num] |
---|
[ce27e21] | 361 | loops_l = cl.LocalMemory(len(loops.data)) |
---|
[14de349] | 362 | cl.enqueue_copy(queuei, loops_bi, loops) |
---|
| 363 | #ctx = environment().context |
---|
| 364 | #loops_bi = cl.Buffer(ctx, mf.READ_ONLY | mf.COPY_HOST_PTR, hostbuf=loops) |
---|
[5d4777d] | 365 | args = self.input.q_buffers + [res_bi,loops_bi,loops_l,cutoff] + fixed + Nloops |
---|
[ce27e21] | 366 | self.kernel(queuei, self.input.global_size, None, *args) |
---|
[14de349] | 367 | cl.enqueue_copy(queuei, self.res, res_bi) |
---|
| 368 | |
---|
| 369 | return self.res |
---|
| 370 | |
---|
| 371 | def release(self): |
---|
| 372 | for b in self.loops_b: |
---|
| 373 | b.release() |
---|
| 374 | self.loops_b = [] |
---|
| 375 | for b in self.res_b: |
---|
| 376 | b.release() |
---|
| 377 | self.res_b = [] |
---|
| 378 | |
---|
| 379 | def __del__(self): |
---|
| 380 | self.release() |
---|