[0db7dbd] | 1 | """ |
---|
| 2 | GPU driver for C kernels |
---|
| 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 | In order to use OpenCL for your models, you will need OpenCL drivers for |
---|
| 26 | your machine. These should be available from your graphics card vendor. |
---|
| 27 | Intel provides OpenCL drivers for CPUs as well as their integrated HD |
---|
| 28 | graphics chipsets. AMD also provides drivers for Intel CPUs, but as of |
---|
| 29 | this writing the performance is lacking compared to the Intel drivers. |
---|
| 30 | NVidia combines drivers for CUDA and OpenCL in one package. The result |
---|
| 31 | is a bit messy if you have multiple drivers installed. You can see which |
---|
| 32 | drivers are available by starting python and running: |
---|
| 33 | |
---|
| 34 | import pyopencl as cl |
---|
| 35 | cl.create_some_context(interactive=True) |
---|
| 36 | |
---|
| 37 | Once you have done that, it will show the available drivers which you |
---|
| 38 | can select. It will then tell you that you can use these drivers |
---|
| 39 | automatically by setting the SAS_OPENCL environment variable, which is |
---|
| 40 | PYOPENCL_CTX equivalent but not conflicting with other pyopnecl programs. |
---|
| 41 | |
---|
| 42 | Some graphics cards have multiple devices on the same card. You cannot |
---|
| 43 | yet use both of them concurrently to evaluate models, but you can run |
---|
| 44 | the program twice using a different device for each session. |
---|
| 45 | |
---|
| 46 | OpenCL kernels are compiled when needed by the device driver. Some |
---|
| 47 | drivers produce compiler output even when there is no error. You |
---|
| 48 | can see the output by setting PYOPENCL_COMPILER_OUTPUT=1. It should be |
---|
| 49 | harmless, albeit annoying. |
---|
| 50 | """ |
---|
| 51 | from __future__ import print_function |
---|
| 52 | |
---|
| 53 | import os |
---|
| 54 | import warnings |
---|
| 55 | import logging |
---|
| 56 | import time |
---|
| 57 | |
---|
| 58 | import numpy as np # type: ignore |
---|
| 59 | |
---|
| 60 | |
---|
| 61 | # Attempt to setup opencl. This may fail if the opencl package is not |
---|
| 62 | # installed or if it is installed but there are no devices available. |
---|
| 63 | try: |
---|
| 64 | import pycuda.autoinit |
---|
| 65 | import pycuda.driver as cuda # type: ignore |
---|
| 66 | from pycuda.compiler import SourceModule |
---|
| 67 | HAVE_CUDA = True |
---|
| 68 | CUDA_ERROR = "" |
---|
| 69 | except Exception as exc: |
---|
| 70 | HAVE_CUDA = False |
---|
| 71 | CUDA_ERROR = str(exc) |
---|
| 72 | |
---|
| 73 | from . import generate |
---|
| 74 | from .kernel import KernelModel, Kernel |
---|
| 75 | |
---|
| 76 | # pylint: disable=unused-import |
---|
| 77 | try: |
---|
| 78 | from typing import Tuple, Callable, Any |
---|
| 79 | from .modelinfo import ModelInfo |
---|
| 80 | from .details import CallDetails |
---|
| 81 | except ImportError: |
---|
| 82 | pass |
---|
| 83 | # pylint: enable=unused-import |
---|
| 84 | |
---|
| 85 | # The max loops number is limited by the amount of local memory available |
---|
| 86 | # on the device. You don't want to make this value too big because it will |
---|
| 87 | # waste resources, nor too small because it may interfere with users trying |
---|
| 88 | # to do their polydispersity calculations. A value of 1024 should be much |
---|
| 89 | # larger than necessary given that cost grows as npts^k where k is the number |
---|
| 90 | # of polydisperse parameters. |
---|
| 91 | MAX_LOOPS = 2048 |
---|
| 92 | |
---|
| 93 | |
---|
| 94 | # Pragmas for enable OpenCL features. Be sure to protect them so that they |
---|
| 95 | # still compile even if OpenCL is not present. |
---|
| 96 | _F16_PRAGMA = """\ |
---|
| 97 | #if defined(__OPENCL_VERSION__) // && !defined(cl_khr_fp16) |
---|
| 98 | # pragma OPENCL EXTENSION cl_khr_fp16: enable |
---|
| 99 | #endif |
---|
| 100 | """ |
---|
| 101 | |
---|
| 102 | _F64_PRAGMA = """\ |
---|
| 103 | #if defined(__OPENCL_VERSION__) // && !defined(cl_khr_fp64) |
---|
| 104 | # pragma OPENCL EXTENSION cl_khr_fp64: enable |
---|
| 105 | #endif |
---|
| 106 | """ |
---|
| 107 | |
---|
| 108 | def use_cuda(): |
---|
| 109 | return HAVE_CUDA |
---|
| 110 | |
---|
| 111 | ENV = None |
---|
| 112 | def reset_environment(): |
---|
| 113 | """ |
---|
| 114 | Call to create a new OpenCL context, such as after a change to SAS_OPENCL. |
---|
| 115 | """ |
---|
| 116 | global ENV |
---|
| 117 | ENV = GpuEnvironment() if use_cuda() else None |
---|
| 118 | |
---|
| 119 | def environment(): |
---|
| 120 | # type: () -> "GpuEnvironment" |
---|
| 121 | """ |
---|
| 122 | Returns a singleton :class:`GpuEnvironment`. |
---|
| 123 | |
---|
| 124 | This provides an OpenCL context and one queue per device. |
---|
| 125 | """ |
---|
| 126 | if ENV is None: |
---|
| 127 | if not HAVE_CUDA: |
---|
| 128 | raise RuntimeError("OpenCL startup failed with ***" |
---|
| 129 | + CUDA_ERROR + "***; using C compiler instead") |
---|
| 130 | reset_environment() |
---|
| 131 | if ENV is None: |
---|
| 132 | raise RuntimeError("SAS_OPENCL=None in environment") |
---|
| 133 | return ENV |
---|
| 134 | |
---|
| 135 | def _stretch_input(vector, dtype, extra=1e-3, boundary=32): |
---|
| 136 | # type: (np.ndarray, np.dtype, float, int) -> np.ndarray |
---|
| 137 | """ |
---|
| 138 | Stretch an input vector to the correct boundary. |
---|
| 139 | |
---|
| 140 | Performance on the kernels can drop by a factor of two or more if the |
---|
| 141 | number of values to compute does not fall on a nice power of two |
---|
| 142 | boundary. The trailing additional vector elements are given a |
---|
| 143 | value of *extra*, and so f(*extra*) will be computed for each of |
---|
| 144 | them. The returned array will thus be a subset of the computed array. |
---|
| 145 | |
---|
| 146 | *boundary* should be a power of 2 which is at least 32 for good |
---|
| 147 | performance on current platforms (as of Jan 2015). It should |
---|
| 148 | probably be the max of get_warp(kernel,queue) and |
---|
| 149 | device.min_data_type_align_size//4. |
---|
| 150 | """ |
---|
| 151 | remainder = vector.size % boundary |
---|
| 152 | if remainder != 0: |
---|
| 153 | size = vector.size + (boundary - remainder) |
---|
| 154 | vector = np.hstack((vector, [extra] * (size - vector.size))) |
---|
| 155 | return np.ascontiguousarray(vector, dtype=dtype) |
---|
| 156 | |
---|
| 157 | |
---|
| 158 | def compile_model(source, dtype, fast=False): |
---|
| 159 | # type: (str, np.dtype, bool) -> cl.Program |
---|
| 160 | """ |
---|
| 161 | Build a model to run on the gpu. |
---|
| 162 | |
---|
| 163 | Returns the compiled program and its type. The returned type will |
---|
| 164 | be float32 even if the desired type is float64 if any of the |
---|
| 165 | devices in the context do not support the cl_khr_fp64 extension. |
---|
| 166 | """ |
---|
| 167 | source_list = [generate.convert_type(source, dtype)] |
---|
| 168 | |
---|
| 169 | if dtype == generate.F16: |
---|
| 170 | source_list.insert(0, _F16_PRAGMA) |
---|
| 171 | elif dtype == generate.F64: |
---|
| 172 | source_list.insert(0, _F64_PRAGMA) |
---|
| 173 | |
---|
| 174 | source_list.insert(0, "#define USE_SINCOS\n") |
---|
| 175 | source = "\n".join(source_list) |
---|
| 176 | program = SourceModule(source) # no_extern_c=True, include_dirs=[...] |
---|
| 177 | return program |
---|
| 178 | |
---|
| 179 | |
---|
| 180 | # for now, this returns one device in the context |
---|
| 181 | # TODO: create a context that contains all devices on all platforms |
---|
| 182 | class GpuEnvironment(object): |
---|
| 183 | """ |
---|
| 184 | GPU context, with possibly many devices, and one queue per device. |
---|
| 185 | """ |
---|
| 186 | def __init__(self, devnum=0): |
---|
| 187 | # type: () -> None |
---|
| 188 | # Byte boundary for data alignment |
---|
| 189 | #self.data_boundary = max(d.min_data_type_align_size |
---|
| 190 | # for d in self.context.devices) |
---|
| 191 | self.compiled = {} |
---|
| 192 | #self.device = cuda.Device(devnum) |
---|
| 193 | #self.context = self.device.make_context() |
---|
| 194 | |
---|
| 195 | def has_type(self, dtype): |
---|
| 196 | # type: (np.dtype) -> bool |
---|
| 197 | """ |
---|
| 198 | Return True if all devices support a given type. |
---|
| 199 | """ |
---|
| 200 | return True |
---|
| 201 | |
---|
| 202 | def compile_program(self, name, source, dtype, fast, timestamp): |
---|
| 203 | # type: (str, str, np.dtype, bool, float) -> cl.Program |
---|
| 204 | """ |
---|
| 205 | Compile the program for the device in the given context. |
---|
| 206 | """ |
---|
| 207 | # Note: PyOpenCL caches based on md5 hash of source, options and device |
---|
| 208 | # so we don't really need to cache things for ourselves. I'll do so |
---|
| 209 | # anyway just to save some data munging time. |
---|
| 210 | tag = generate.tag_source(source) |
---|
| 211 | key = "%s-%s-%s%s"%(name, dtype, tag, ("-fast" if fast else "")) |
---|
| 212 | # Check timestamp on program |
---|
| 213 | program, program_timestamp = self.compiled.get(key, (None, np.inf)) |
---|
| 214 | if program_timestamp < timestamp: |
---|
| 215 | del self.compiled[key] |
---|
| 216 | if key not in self.compiled: |
---|
| 217 | logging.info("building %s for CUDA", key) |
---|
| 218 | program = compile_model(str(source), dtype, fast) |
---|
| 219 | self.compiled[key] = (program, timestamp) |
---|
| 220 | return program |
---|
| 221 | |
---|
| 222 | class GpuModel(KernelModel): |
---|
| 223 | """ |
---|
| 224 | GPU wrapper for a single model. |
---|
| 225 | |
---|
| 226 | *source* and *model_info* are the model source and interface as returned |
---|
| 227 | from :func:`generate.make_source` and :func:`generate.make_model_info`. |
---|
| 228 | |
---|
| 229 | *dtype* is the desired model precision. Any numpy dtype for single |
---|
| 230 | or double precision floats will do, such as 'f', 'float32' or 'single' |
---|
| 231 | for single and 'd', 'float64' or 'double' for double. Double precision |
---|
| 232 | is an optional extension which may not be available on all devices. |
---|
| 233 | Half precision ('float16','half') may be available on some devices. |
---|
| 234 | Fast precision ('fast') is a loose version of single precision, indicating |
---|
| 235 | that the compiler is allowed to take shortcuts. |
---|
| 236 | """ |
---|
| 237 | def __init__(self, source, model_info, dtype=generate.F32, fast=False): |
---|
| 238 | # type: (Dict[str,str], ModelInfo, np.dtype, bool) -> None |
---|
| 239 | self.info = model_info |
---|
| 240 | self.source = source |
---|
| 241 | self.dtype = dtype |
---|
| 242 | self.fast = fast |
---|
| 243 | self.program = None # delay program creation |
---|
| 244 | self._kernels = None |
---|
| 245 | |
---|
| 246 | def __getstate__(self): |
---|
| 247 | # type: () -> Tuple[ModelInfo, str, np.dtype, bool] |
---|
| 248 | return self.info, self.source, self.dtype, self.fast |
---|
| 249 | |
---|
| 250 | def __setstate__(self, state): |
---|
| 251 | # type: (Tuple[ModelInfo, str, np.dtype, bool]) -> None |
---|
| 252 | self.info, self.source, self.dtype, self.fast = state |
---|
| 253 | self.program = None |
---|
| 254 | |
---|
| 255 | def make_kernel(self, q_vectors): |
---|
| 256 | # type: (List[np.ndarray]) -> "GpuKernel" |
---|
| 257 | if self.program is None: |
---|
| 258 | compile_program = environment().compile_program |
---|
| 259 | timestamp = generate.ocl_timestamp(self.info) |
---|
| 260 | self.program = compile_program( |
---|
| 261 | self.info.name, |
---|
| 262 | self.source['opencl'], |
---|
| 263 | self.dtype, |
---|
| 264 | self.fast, |
---|
| 265 | timestamp) |
---|
| 266 | variants = ['Iq', 'Iqxy', 'Imagnetic'] |
---|
| 267 | names = [generate.kernel_name(self.info, k) for k in variants] |
---|
| 268 | kernels = [self.program.get_function(k) for k in names] |
---|
| 269 | self._kernels = dict((k, v) for k, v in zip(variants, kernels)) |
---|
| 270 | is_2d = len(q_vectors) == 2 |
---|
| 271 | if is_2d: |
---|
| 272 | kernel = [self._kernels['Iqxy'], self._kernels['Imagnetic']] |
---|
| 273 | else: |
---|
| 274 | kernel = [self._kernels['Iq']]*2 |
---|
| 275 | return GpuKernel(kernel, self.dtype, self.info, q_vectors) |
---|
| 276 | |
---|
| 277 | def release(self): |
---|
| 278 | # type: () -> None |
---|
| 279 | """ |
---|
| 280 | Free the resources associated with the model. |
---|
| 281 | """ |
---|
| 282 | if self.program is not None: |
---|
| 283 | self.program = None |
---|
| 284 | |
---|
| 285 | def __del__(self): |
---|
| 286 | # type: () -> None |
---|
| 287 | self.release() |
---|
| 288 | |
---|
| 289 | # TODO: check that we don't need a destructor for buffers which go out of scope |
---|
| 290 | class GpuInput(object): |
---|
| 291 | """ |
---|
| 292 | Make q data available to the gpu. |
---|
| 293 | |
---|
| 294 | *q_vectors* is a list of q vectors, which will be *[q]* for 1-D data, |
---|
| 295 | and *[qx, qy]* for 2-D data. Internally, the vectors will be reallocated |
---|
| 296 | to get the best performance on OpenCL, which may involve shifting and |
---|
| 297 | stretching the array to better match the memory architecture. Additional |
---|
| 298 | points will be evaluated with *q=1e-3*. |
---|
| 299 | |
---|
| 300 | *dtype* is the data type for the q vectors. The data type should be |
---|
| 301 | set to match that of the kernel, which is an attribute of |
---|
| 302 | :class:`GpuProgram`. Note that not all kernels support double |
---|
| 303 | precision, so even if the program was created for double precision, |
---|
| 304 | the *GpuProgram.dtype* may be single precision. |
---|
| 305 | |
---|
| 306 | Call :meth:`release` when complete. Even if not called directly, the |
---|
| 307 | buffer will be released when the data object is freed. |
---|
| 308 | """ |
---|
| 309 | def __init__(self, q_vectors, dtype=generate.F32): |
---|
| 310 | # type: (List[np.ndarray], np.dtype) -> None |
---|
| 311 | # TODO: do we ever need double precision q? |
---|
| 312 | self.nq = q_vectors[0].size |
---|
| 313 | self.dtype = np.dtype(dtype) |
---|
| 314 | self.is_2d = (len(q_vectors) == 2) |
---|
| 315 | # TODO: stretch input based on get_warp() |
---|
| 316 | # not doing it now since warp depends on kernel, which is not known |
---|
| 317 | # at this point, so instead using 32, which is good on the set of |
---|
| 318 | # architectures tested so far. |
---|
| 319 | if self.is_2d: |
---|
| 320 | # Note: 16 rather than 15 because result is 1 longer than input. |
---|
| 321 | width = ((self.nq+16)//16)*16 |
---|
| 322 | self.q = np.empty((width, 2), dtype=dtype) |
---|
| 323 | self.q[:self.nq, 0] = q_vectors[0] |
---|
| 324 | self.q[:self.nq, 1] = q_vectors[1] |
---|
| 325 | else: |
---|
| 326 | # Note: 32 rather than 31 because result is 1 longer than input. |
---|
| 327 | width = ((self.nq+32)//32)*32 |
---|
| 328 | self.q = np.empty(width, dtype=dtype) |
---|
| 329 | self.q[:self.nq] = q_vectors[0] |
---|
| 330 | self.global_size = [self.q.shape[0]] |
---|
| 331 | #print("creating inputs of size", self.global_size) |
---|
| 332 | self.q_b = cuda.to_device(self.q) |
---|
| 333 | |
---|
| 334 | def release(self): |
---|
| 335 | # type: () -> None |
---|
| 336 | """ |
---|
| 337 | Free the memory. |
---|
| 338 | """ |
---|
| 339 | if self.q_b is not None: |
---|
| 340 | self.q_b.free() |
---|
| 341 | self.q_b = None |
---|
| 342 | |
---|
| 343 | def __del__(self): |
---|
| 344 | # type: () -> None |
---|
| 345 | self.release() |
---|
| 346 | |
---|
| 347 | class GpuKernel(Kernel): |
---|
| 348 | """ |
---|
| 349 | Callable SAS kernel. |
---|
| 350 | |
---|
| 351 | *kernel* is the GpuKernel object to call |
---|
| 352 | |
---|
| 353 | *model_info* is the module information |
---|
| 354 | |
---|
| 355 | *q_vectors* is the q vectors at which the kernel should be evaluated |
---|
| 356 | |
---|
| 357 | *dtype* is the kernel precision |
---|
| 358 | |
---|
| 359 | The resulting call method takes the *pars*, a list of values for |
---|
| 360 | the fixed parameters to the kernel, and *pd_pars*, a list of (value,weight) |
---|
| 361 | vectors for the polydisperse parameters. *cutoff* determines the |
---|
| 362 | integration limits: any points with combined weight less than *cutoff* |
---|
| 363 | will not be calculated. |
---|
| 364 | |
---|
| 365 | Call :meth:`release` when done with the kernel instance. |
---|
| 366 | """ |
---|
| 367 | def __init__(self, kernel, dtype, model_info, q_vectors): |
---|
| 368 | # type: (cl.Kernel, np.dtype, ModelInfo, List[np.ndarray]) -> None |
---|
| 369 | q_input = GpuInput(q_vectors, dtype) |
---|
| 370 | self.kernel = kernel |
---|
| 371 | self.info = model_info |
---|
| 372 | self.dtype = dtype |
---|
| 373 | self.dim = '2d' if q_input.is_2d else '1d' |
---|
| 374 | # plus three for the normalization values |
---|
| 375 | self.result = np.empty(q_input.nq+1, dtype) |
---|
| 376 | |
---|
| 377 | # Inputs and outputs for each kernel call |
---|
| 378 | # Note: res may be shorter than res_b if global_size != nq |
---|
| 379 | self.result_b = cuda.mem_alloc(q_input.global_size[0] * dtype.itemsize) |
---|
| 380 | self.q_input = q_input # allocated by GpuInput above |
---|
| 381 | |
---|
| 382 | self._need_release = [self.result_b, self.q_input] |
---|
| 383 | self.real = (np.float32 if dtype == generate.F32 |
---|
| 384 | else np.float64 if dtype == generate.F64 |
---|
| 385 | else np.float16 if dtype == generate.F16 |
---|
| 386 | else np.float32) # will never get here, so use np.float32 |
---|
| 387 | |
---|
| 388 | def __call__(self, call_details, values, cutoff, magnetic): |
---|
| 389 | # type: (CallDetails, np.ndarray, np.ndarray, float, bool) -> np.ndarray |
---|
| 390 | # Arrange data transfer to card |
---|
| 391 | details_b = cuda.to_device(call_details.buffer) |
---|
| 392 | values_b = cuda.to_device(values) |
---|
| 393 | |
---|
| 394 | kernel = self.kernel[1 if magnetic else 0] |
---|
| 395 | args = [ |
---|
| 396 | np.uint32(self.q_input.nq), None, None, |
---|
| 397 | details_b, values_b, self.q_input.q_b, self.result_b, |
---|
| 398 | self.real(cutoff), |
---|
| 399 | ] |
---|
| 400 | grid = partition(self.q_input.nq) |
---|
| 401 | #print("Calling OpenCL") |
---|
| 402 | #call_details.show(values) |
---|
| 403 | # Call kernel and retrieve results |
---|
| 404 | last_nap = time.clock() |
---|
| 405 | step = 1000000//self.q_input.nq + 1 |
---|
| 406 | #step = 1000000000 |
---|
| 407 | for start in range(0, call_details.num_eval, step): |
---|
| 408 | stop = min(start + step, call_details.num_eval) |
---|
| 409 | #print("queuing",start,stop) |
---|
| 410 | args[1:3] = [np.int32(start), np.int32(stop)] |
---|
| 411 | kernel(*args, **grid) |
---|
| 412 | if stop < call_details.num_eval: |
---|
| 413 | sync() |
---|
| 414 | # Allow other processes to run |
---|
| 415 | current_time = time.clock() |
---|
| 416 | if current_time - last_nap > 0.5: |
---|
| 417 | time.sleep(0.05) |
---|
| 418 | last_nap = current_time |
---|
| 419 | sync() |
---|
| 420 | details_b.free() |
---|
| 421 | values_b.free() |
---|
| 422 | cuda.memcpy_dtoh(self.result, self.result_b) |
---|
| 423 | #print("result", self.result) |
---|
| 424 | |
---|
| 425 | pd_norm = self.result[self.q_input.nq] |
---|
| 426 | scale = values[0]/(pd_norm if pd_norm != 0.0 else 1.0) |
---|
| 427 | background = values[1] |
---|
| 428 | #print("scale",scale,values[0],self.result[self.q_input.nq],background) |
---|
| 429 | return scale*self.result[:self.q_input.nq] + background |
---|
| 430 | # return self.result[:self.q_input.nq] |
---|
| 431 | |
---|
| 432 | def release(self): |
---|
| 433 | # type: () -> None |
---|
| 434 | """ |
---|
| 435 | Release resources associated with the kernel. |
---|
| 436 | """ |
---|
| 437 | if self.result_b is not None: |
---|
| 438 | self.result_b.free() |
---|
| 439 | self.result_b = None |
---|
| 440 | |
---|
| 441 | def __del__(self): |
---|
| 442 | # type: () -> None |
---|
| 443 | self.release() |
---|
| 444 | |
---|
| 445 | |
---|
| 446 | def sync(): |
---|
| 447 | """ |
---|
| 448 | Overview: |
---|
| 449 | Waits for operation in the current context to complete. |
---|
| 450 | |
---|
| 451 | Note: Maybe context.synchronize() is sufficient. |
---|
| 452 | """ |
---|
| 453 | #return # The following works in C++; don't know what pycuda is doing |
---|
| 454 | # Create an event with which to synchronize |
---|
| 455 | done = cuda.Event() |
---|
| 456 | |
---|
| 457 | # Schedule an event trigger on the GPU. |
---|
| 458 | done.record() |
---|
| 459 | |
---|
| 460 | #line added to not hog resources |
---|
| 461 | while not done.query(): time.sleep(0.01) |
---|
| 462 | |
---|
| 463 | # Block until the GPU executes the kernel. |
---|
| 464 | done.synchronize() |
---|
| 465 | # Clean up the event; I don't think they can be reused. |
---|
| 466 | del done |
---|
| 467 | |
---|
| 468 | |
---|
| 469 | def partition(n): |
---|
| 470 | ''' |
---|
| 471 | Overview: |
---|
| 472 | Auto grids the thread blocks to achieve some level of calculation |
---|
| 473 | efficiency. |
---|
| 474 | ''' |
---|
| 475 | max_gx,max_gy = 65535,65535 |
---|
| 476 | blocksize = 32 |
---|
| 477 | #max_gx,max_gy = 5,65536 |
---|
| 478 | #blocksize = 3 |
---|
| 479 | block = (blocksize,1,1) |
---|
| 480 | num_blocks = int((n+blocksize-1)/blocksize) |
---|
| 481 | if num_blocks < max_gx: |
---|
| 482 | grid = (num_blocks,1) |
---|
| 483 | else: |
---|
| 484 | gx = max_gx |
---|
| 485 | gy = (num_blocks + max_gx - 1) / max_gx |
---|
| 486 | if gy >= max_gy: raise ValueError("vector is too large") |
---|
| 487 | grid = (gx,gy) |
---|
| 488 | #print "block",block,"grid",grid |
---|
| 489 | #print "waste",block[0]*block[1]*block[2]*grid[0]*grid[1] - n |
---|
| 490 | return dict(block=block,grid=grid) |
---|
| 491 | |
---|