1 | """ |
---|
2 | GPU driver for C kernels |
---|
3 | |
---|
4 | TODO: docs are out of date |
---|
5 | |
---|
6 | There should be a single GPU environment running on the system. This |
---|
7 | environment is constructed on the first call to :func:`env`, and the |
---|
8 | same environment is returned on each call. |
---|
9 | |
---|
10 | After retrieving the environment, the next step is to create the kernel. |
---|
11 | This is done with a call to :meth:`GpuEnvironment.make_kernel`, which |
---|
12 | returns the type of data used by the kernel. |
---|
13 | |
---|
14 | Next a :class:`GpuData` object should be created with the correct kind |
---|
15 | of data. This data object can be used by multiple kernels, for example, |
---|
16 | if the target model is a weighted sum of multiple kernels. The data |
---|
17 | should include any extra evaluation points required to compute the proper |
---|
18 | data smearing. This need not match the square grid for 2D data if there |
---|
19 | is an index saying which q points are active. |
---|
20 | |
---|
21 | Together the GpuData, the program, and a device form a :class:`GpuKernel`. |
---|
22 | This kernel is used during fitting, receiving new sets of parameters and |
---|
23 | evaluating them. The output value is stored in an output buffer on the |
---|
24 | devices, where it can be combined with other structure factors and form |
---|
25 | factors and have instrumental resolution effects applied. |
---|
26 | |
---|
27 | In order to use OpenCL for your models, you will need OpenCL drivers for |
---|
28 | your machine. These should be available from your graphics card vendor. |
---|
29 | Intel provides OpenCL drivers for CPUs as well as their integrated HD |
---|
30 | graphics chipsets. AMD also provides drivers for Intel CPUs, but as of |
---|
31 | this writing the performance is lacking compared to the Intel drivers. |
---|
32 | NVidia combines drivers for CUDA and OpenCL in one package. The result |
---|
33 | is a bit messy if you have multiple drivers installed. You can see which |
---|
34 | drivers are available by starting python and running: |
---|
35 | |
---|
36 | import pyopencl as cl |
---|
37 | cl.create_some_context(interactive=True) |
---|
38 | |
---|
39 | Once you have done that, it will show the available drivers which you |
---|
40 | can select. It will then tell you that you can use these drivers |
---|
41 | automatically by setting the SAS_OPENCL environment variable, which is |
---|
42 | PYOPENCL_CTX equivalent but not conflicting with other pyopnecl programs. |
---|
43 | |
---|
44 | Some graphics cards have multiple devices on the same card. You cannot |
---|
45 | yet use both of them concurrently to evaluate models, but you can run |
---|
46 | the program twice using a different device for each session. |
---|
47 | |
---|
48 | OpenCL kernels are compiled when needed by the device driver. Some |
---|
49 | drivers produce compiler output even when there is no error. You |
---|
50 | can see the output by setting PYOPENCL_COMPILER_OUTPUT=1. It should be |
---|
51 | harmless, albeit annoying. |
---|
52 | """ |
---|
53 | from __future__ import print_function |
---|
54 | |
---|
55 | import os |
---|
56 | import warnings |
---|
57 | import logging |
---|
58 | import time |
---|
59 | |
---|
60 | import numpy as np # type: ignore |
---|
61 | |
---|
62 | |
---|
63 | # Attempt to setup opencl. This may fail if the pyopencl package is not |
---|
64 | # installed or if it is installed but there are no devices available. |
---|
65 | try: |
---|
66 | import pyopencl as cl # type: ignore |
---|
67 | from pyopencl import mem_flags as mf |
---|
68 | from pyopencl.characterize import get_fast_inaccurate_build_options |
---|
69 | # Ask OpenCL for the default context so that we know that one exists |
---|
70 | cl.create_some_context(interactive=False) |
---|
71 | HAVE_OPENCL = True |
---|
72 | OPENCL_ERROR = "" |
---|
73 | except Exception as exc: |
---|
74 | HAVE_OPENCL = False |
---|
75 | OPENCL_ERROR = str(exc) |
---|
76 | |
---|
77 | from . import generate |
---|
78 | from .generate import F32, F64 |
---|
79 | from .kernel import KernelModel, Kernel |
---|
80 | |
---|
81 | # pylint: disable=unused-import |
---|
82 | try: |
---|
83 | from typing import Tuple, Callable, Any |
---|
84 | from .modelinfo import ModelInfo |
---|
85 | from .details import CallDetails |
---|
86 | except ImportError: |
---|
87 | pass |
---|
88 | # pylint: enable=unused-import |
---|
89 | |
---|
90 | # CRUFT: pyopencl < 2017.1 (as of June 2016 needs quotes around include path) |
---|
91 | def quote_path(v): |
---|
92 | """ |
---|
93 | Quote the path if it is not already quoted. |
---|
94 | |
---|
95 | If v starts with '-', then assume that it is a -I option or similar |
---|
96 | and do not quote it. This is fragile: -Ipath with space needs to |
---|
97 | be quoted. |
---|
98 | """ |
---|
99 | return '"'+v+'"' if v and ' ' in v and not v[0] in "\"'-" else v |
---|
100 | |
---|
101 | def fix_pyopencl_include(): |
---|
102 | """ |
---|
103 | Monkey patch pyopencl to allow spaces in include file path. |
---|
104 | """ |
---|
105 | import pyopencl as cl |
---|
106 | if hasattr(cl, '_DEFAULT_INCLUDE_OPTIONS'): |
---|
107 | cl._DEFAULT_INCLUDE_OPTIONS = [quote_path(v) for v in cl._DEFAULT_INCLUDE_OPTIONS] |
---|
108 | |
---|
109 | if HAVE_OPENCL: |
---|
110 | fix_pyopencl_include() |
---|
111 | |
---|
112 | # The max loops number is limited by the amount of local memory available |
---|
113 | # on the device. You don't want to make this value too big because it will |
---|
114 | # waste resources, nor too small because it may interfere with users trying |
---|
115 | # to do their polydispersity calculations. A value of 1024 should be much |
---|
116 | # larger than necessary given that cost grows as npts^k where k is the number |
---|
117 | # of polydisperse parameters. |
---|
118 | MAX_LOOPS = 2048 |
---|
119 | |
---|
120 | |
---|
121 | # Pragmas for enable OpenCL features. Be sure to protect them so that they |
---|
122 | # still compile even if OpenCL is not present. |
---|
123 | _F16_PRAGMA = """\ |
---|
124 | #if defined(__OPENCL_VERSION__) // && !defined(cl_khr_fp16) |
---|
125 | # pragma OPENCL EXTENSION cl_khr_fp16: enable |
---|
126 | #endif |
---|
127 | """ |
---|
128 | |
---|
129 | _F64_PRAGMA = """\ |
---|
130 | #if defined(__OPENCL_VERSION__) // && !defined(cl_khr_fp64) |
---|
131 | # pragma OPENCL EXTENSION cl_khr_fp64: enable |
---|
132 | #endif |
---|
133 | """ |
---|
134 | |
---|
135 | def use_opencl(): |
---|
136 | sas_opencl = os.environ.get("SAS_OPENCL", "OpenCL").lower() |
---|
137 | return HAVE_OPENCL and sas_opencl != "none" and not sas_opencl.startswith("cuda") |
---|
138 | |
---|
139 | ENV = None |
---|
140 | def reset_environment(): |
---|
141 | """ |
---|
142 | Call to create a new OpenCL context, such as after a change to SAS_OPENCL. |
---|
143 | """ |
---|
144 | global ENV |
---|
145 | ENV = GpuEnvironment() if use_opencl() else None |
---|
146 | |
---|
147 | def environment(): |
---|
148 | # type: () -> "GpuEnvironment" |
---|
149 | """ |
---|
150 | Returns a singleton :class:`GpuEnvironment`. |
---|
151 | |
---|
152 | This provides an OpenCL context and one queue per device. |
---|
153 | """ |
---|
154 | if ENV is None: |
---|
155 | if not HAVE_OPENCL: |
---|
156 | raise RuntimeError("OpenCL startup failed with ***" |
---|
157 | + OPENCL_ERROR + "***; using C compiler instead") |
---|
158 | reset_environment() |
---|
159 | if ENV is None: |
---|
160 | raise RuntimeError("SAS_OPENCL=None in environment") |
---|
161 | return ENV |
---|
162 | |
---|
163 | def has_type(device, dtype): |
---|
164 | # type: (cl.Device, np.dtype) -> bool |
---|
165 | """ |
---|
166 | Return true if device supports the requested precision. |
---|
167 | """ |
---|
168 | if dtype == F32: |
---|
169 | return True |
---|
170 | elif dtype == generate.F64: |
---|
171 | return "cl_khr_fp64" in device.extensions |
---|
172 | elif dtype == generate.F16: |
---|
173 | return "cl_khr_fp16" in device.extensions |
---|
174 | else: |
---|
175 | return False |
---|
176 | |
---|
177 | def get_warp(kernel, queue): |
---|
178 | # type: (cl.Kernel, cl.CommandQueue) -> int |
---|
179 | """ |
---|
180 | Return the size of an execution batch for *kernel* running on *queue*. |
---|
181 | """ |
---|
182 | return kernel.get_work_group_info( |
---|
183 | cl.kernel_work_group_info.PREFERRED_WORK_GROUP_SIZE_MULTIPLE, |
---|
184 | queue.device) |
---|
185 | |
---|
186 | def compile_model(context, source, dtype, fast=False): |
---|
187 | # type: (cl.Context, str, np.dtype, bool) -> cl.Program |
---|
188 | """ |
---|
189 | Build a model to run on the gpu. |
---|
190 | |
---|
191 | Returns the compiled program and its type. |
---|
192 | |
---|
193 | Raises an error if the desired precision is not available. |
---|
194 | """ |
---|
195 | dtype = np.dtype(dtype) |
---|
196 | if not all(has_type(d, dtype) for d in context.devices): |
---|
197 | raise RuntimeError("%s not supported for devices"%dtype) |
---|
198 | |
---|
199 | source_list = [generate.convert_type(source, dtype)] |
---|
200 | |
---|
201 | if dtype == generate.F16: |
---|
202 | source_list.insert(0, _F16_PRAGMA) |
---|
203 | elif dtype == generate.F64: |
---|
204 | source_list.insert(0, _F64_PRAGMA) |
---|
205 | |
---|
206 | # Note: USE_SINCOS makes the intel cpu slower under opencl |
---|
207 | if context.devices[0].type == cl.device_type.GPU: |
---|
208 | source_list.insert(0, "#define USE_SINCOS\n") |
---|
209 | options = (get_fast_inaccurate_build_options(context.devices[0]) |
---|
210 | if fast else []) |
---|
211 | source = "\n".join(source_list) |
---|
212 | program = cl.Program(context, source).build(options=options) |
---|
213 | #print("done with "+program) |
---|
214 | return program |
---|
215 | |
---|
216 | |
---|
217 | # for now, this returns one device in the context |
---|
218 | # TODO: create a context that contains all devices on all platforms |
---|
219 | class GpuEnvironment(object): |
---|
220 | """ |
---|
221 | GPU context, with possibly many devices, and one queue per device. |
---|
222 | |
---|
223 | Because the environment can be reset during a live program (e.g., if the |
---|
224 | user changes the active GPU device in the GUI), everything associated |
---|
225 | with the device context must be cached in the environment and recreated |
---|
226 | if the environment changes. The *cache* attribute is a simple dictionary |
---|
227 | which holds keys and references to objects, such as compiled kernels and |
---|
228 | allocated buffers. The running program should check in the cache for |
---|
229 | long lived objects and create them if they are not there. The program |
---|
230 | should not hold onto cached objects, but instead only keep them active |
---|
231 | for the duration of a function call. When the environment is destroyed |
---|
232 | then the *release* method for each active cache item is called before |
---|
233 | the environment is freed. This means that each cl buffer should be |
---|
234 | in its own cache entry. |
---|
235 | """ |
---|
236 | def __init__(self): |
---|
237 | # type: () -> None |
---|
238 | # find gpu context |
---|
239 | context_list = _create_some_context() |
---|
240 | |
---|
241 | # Find a context for F32 and for F64 (maybe the same one). |
---|
242 | # F16 isn't good enough. |
---|
243 | self.context = {} |
---|
244 | for dtype in (F32, F64): |
---|
245 | for context in context_list: |
---|
246 | if has_type(context.devices[0], dtype): |
---|
247 | self.context[dtype] = context |
---|
248 | break |
---|
249 | else: |
---|
250 | self.context[dtype] = None |
---|
251 | |
---|
252 | # Build a queue for each context |
---|
253 | self.queue = {} |
---|
254 | context = self.context[F32] |
---|
255 | self.queue[F32] = cl.CommandQueue(context, context.devices[0]) |
---|
256 | if self.context[F64] == self.context[F32]: |
---|
257 | self.queue[F64] = self.queue[F32] |
---|
258 | else: |
---|
259 | context = self.context[F64] |
---|
260 | self.queue[F64] = cl.CommandQueue(context, context.devices[0]) |
---|
261 | |
---|
262 | # Byte boundary for data alignment |
---|
263 | #self.data_boundary = max(context.devices[0].min_data_type_align_size |
---|
264 | # for context in self.context.values()) |
---|
265 | |
---|
266 | # Cache for compiled programs, and for items in context |
---|
267 | self.compiled = {} |
---|
268 | self.cache = {} |
---|
269 | |
---|
270 | def has_type(self, dtype): |
---|
271 | # type: (np.dtype) -> bool |
---|
272 | """ |
---|
273 | Return True if all devices support a given type. |
---|
274 | """ |
---|
275 | return self.context.get(dtype, None) is not None |
---|
276 | |
---|
277 | def compile_program(self, name, source, dtype, fast, timestamp): |
---|
278 | # type: (str, str, np.dtype, bool, float) -> cl.Program |
---|
279 | """ |
---|
280 | Compile the program for the device in the given context. |
---|
281 | """ |
---|
282 | # Note: PyOpenCL caches based on md5 hash of source, options and device |
---|
283 | # so we don't really need to cache things for ourselves. I'll do so |
---|
284 | # anyway just to save some data munging time. |
---|
285 | tag = generate.tag_source(source) |
---|
286 | key = "%s-%s-%s%s"%(name, dtype, tag, ("-fast" if fast else "")) |
---|
287 | # Check timestamp on program |
---|
288 | program, program_timestamp = self.compiled.get(key, (None, np.inf)) |
---|
289 | if program_timestamp < timestamp: |
---|
290 | del self.compiled[key] |
---|
291 | if key not in self.compiled: |
---|
292 | context = self.context[dtype] |
---|
293 | logging.info("building %s for OpenCL %s", key, |
---|
294 | context.devices[0].name.strip()) |
---|
295 | program = compile_model(self.context[dtype], |
---|
296 | str(source), dtype, fast) |
---|
297 | self.compiled[key] = (program, timestamp) |
---|
298 | return program |
---|
299 | |
---|
300 | def free_buffer(self, key): |
---|
301 | if key in self.cache: |
---|
302 | self.cache[key].release() |
---|
303 | del self.cache[key] |
---|
304 | |
---|
305 | def __del__(self): |
---|
306 | for v in self.cache.values(): |
---|
307 | release = getattr(v, 'release', lambda: None) |
---|
308 | release() |
---|
309 | self.cache = {} |
---|
310 | |
---|
311 | _CURRENT_ID = 0 |
---|
312 | def unique_id(): |
---|
313 | global _CURRENT_ID |
---|
314 | _CURRENT_ID += 1 |
---|
315 | return _CURRENT_ID |
---|
316 | |
---|
317 | def _create_some_context(): |
---|
318 | # type: () -> cl.Context |
---|
319 | """ |
---|
320 | Protected call to cl.create_some_context without interactivity. |
---|
321 | |
---|
322 | Uses SAS_OPENCL or PYOPENCL_CTX if they are set in the environment, |
---|
323 | otherwise scans for the most appropriate device using |
---|
324 | :func:`_get_default_context`. Ignore *SAS_OPENCL=OpenCL*, which |
---|
325 | indicates that an OpenCL device should be used without specifying |
---|
326 | which one (and not a CUDA device, or no GPU). |
---|
327 | """ |
---|
328 | # Assume we do not get here if SAS_OPENCL is None or CUDA |
---|
329 | sas_opencl = os.environ.get('SAS_OPENCL', 'opencl') |
---|
330 | if sas_opencl.lower() != 'opencl': |
---|
331 | # Setting PYOPENCL_CTX as a SAS_OPENCL to create cl context |
---|
332 | os.environ["PYOPENCL_CTX"] = sas_opencl |
---|
333 | |
---|
334 | if 'PYOPENCL_CTX' in os.environ: |
---|
335 | try: |
---|
336 | return [cl.create_some_context(interactive=False)] |
---|
337 | except Exception as exc: |
---|
338 | warnings.warn(str(exc)) |
---|
339 | warnings.warn("pyopencl.create_some_context() failed") |
---|
340 | warnings.warn("the environment variable 'SAS_OPENCL' or 'PYOPENCL_CTX' might not be set correctly") |
---|
341 | |
---|
342 | return _get_default_context() |
---|
343 | |
---|
344 | def _get_default_context(): |
---|
345 | # type: () -> List[cl.Context] |
---|
346 | """ |
---|
347 | Get an OpenCL context, preferring GPU over CPU, and preferring Intel |
---|
348 | drivers over AMD drivers. |
---|
349 | """ |
---|
350 | # Note: on mobile devices there is automatic clock scaling if either the |
---|
351 | # CPU or the GPU is underutilized; probably doesn't affect us, but we if |
---|
352 | # it did, it would mean that putting a busy loop on the CPU while the GPU |
---|
353 | # is running may increase throughput. |
---|
354 | # |
---|
355 | # Macbook pro, base install: |
---|
356 | # {'Apple': [Intel CPU, NVIDIA GPU]} |
---|
357 | # Macbook pro, base install: |
---|
358 | # {'Apple': [Intel CPU, Intel GPU]} |
---|
359 | # 2 x nvidia 295 with Intel and NVIDIA opencl drivers installed |
---|
360 | # {'Intel': [CPU], 'NVIDIA': [GPU, GPU, GPU, GPU]} |
---|
361 | gpu, cpu = None, None |
---|
362 | for platform in cl.get_platforms(): |
---|
363 | # AMD provides a much weaker CPU driver than Intel/Apple, so avoid it. |
---|
364 | # If someone has bothered to install the AMD/NVIDIA drivers, prefer |
---|
365 | # them over the integrated graphics driver that may have been supplied |
---|
366 | # with the CPU chipset. |
---|
367 | preferred_cpu = (platform.vendor.startswith('Intel') |
---|
368 | or platform.vendor.startswith('Apple')) |
---|
369 | preferred_gpu = (platform.vendor.startswith('Advanced') |
---|
370 | or platform.vendor.startswith('NVIDIA')) |
---|
371 | for device in platform.get_devices(): |
---|
372 | if device.type == cl.device_type.GPU: |
---|
373 | # If the existing type is not GPU then it will be CUSTOM |
---|
374 | # or ACCELERATOR so don't override it. |
---|
375 | if gpu is None or (preferred_gpu and gpu.type == cl.device_type.GPU): |
---|
376 | gpu = device |
---|
377 | elif device.type == cl.device_type.CPU: |
---|
378 | if cpu is None or preferred_cpu: |
---|
379 | cpu = device |
---|
380 | else: |
---|
381 | # System has cl.device_type.ACCELERATOR or cl.device_type.CUSTOM |
---|
382 | # Intel Phi for example registers as an accelerator |
---|
383 | # Since the user installed a custom device on their system |
---|
384 | # and went through the pain of sorting out OpenCL drivers for |
---|
385 | # it, lets assume they really do want to use it as their |
---|
386 | # primary compute device. |
---|
387 | gpu = device |
---|
388 | |
---|
389 | # order the devices by gpu then by cpu; when searching for an available |
---|
390 | # device by data type they will be checked in this order, which means |
---|
391 | # that if the gpu supports double then the cpu will never be used (though |
---|
392 | # we may make it possible to explicitly request the cpu at some point). |
---|
393 | devices = [] |
---|
394 | if gpu is not None: |
---|
395 | devices.append(gpu) |
---|
396 | if cpu is not None: |
---|
397 | devices.append(cpu) |
---|
398 | return [cl.Context([d]) for d in devices] |
---|
399 | |
---|
400 | |
---|
401 | class GpuModel(KernelModel): |
---|
402 | """ |
---|
403 | GPU wrapper for a single model. |
---|
404 | |
---|
405 | *source* and *model_info* are the model source and interface as returned |
---|
406 | from :func:`generate.make_source` and :func:`generate.make_model_info`. |
---|
407 | |
---|
408 | *dtype* is the desired model precision. Any numpy dtype for single |
---|
409 | or double precision floats will do, such as 'f', 'float32' or 'single' |
---|
410 | for single and 'd', 'float64' or 'double' for double. Double precision |
---|
411 | is an optional extension which may not be available on all devices. |
---|
412 | Half precision ('float16','half') may be available on some devices. |
---|
413 | Fast precision ('fast') is a loose version of single precision, indicating |
---|
414 | that the compiler is allowed to take shortcuts. |
---|
415 | """ |
---|
416 | def __init__(self, source, model_info, dtype=generate.F32, fast=False): |
---|
417 | # type: (Dict[str,str], ModelInfo, np.dtype, bool) -> None |
---|
418 | self.info = model_info |
---|
419 | self.source = source |
---|
420 | self.dtype = dtype |
---|
421 | self.fast = fast |
---|
422 | self.timestamp = generate.ocl_timestamp(self.info) |
---|
423 | self._cache_key = unique_id() |
---|
424 | |
---|
425 | def __getstate__(self): |
---|
426 | # type: () -> Tuple[ModelInfo, str, np.dtype, bool] |
---|
427 | return self.info, self.source, self.dtype, self.fast |
---|
428 | |
---|
429 | def __setstate__(self, state): |
---|
430 | # type: (Tuple[ModelInfo, str, np.dtype, bool]) -> None |
---|
431 | self.info, self.source, self.dtype, self.fast = state |
---|
432 | |
---|
433 | def make_kernel(self, q_vectors): |
---|
434 | # type: (List[np.ndarray]) -> "GpuKernel" |
---|
435 | return GpuKernel(self, q_vectors) |
---|
436 | |
---|
437 | @property |
---|
438 | def Iq(self): |
---|
439 | return self._fetch_kernel('Iq') |
---|
440 | |
---|
441 | def fetch_kernel(self, name): |
---|
442 | # type: (str) -> cl.Kernel |
---|
443 | """ |
---|
444 | Fetch the kernel from the environment by name, compiling it if it |
---|
445 | does not already exist. |
---|
446 | """ |
---|
447 | gpu = environment() |
---|
448 | key = self._cache_key |
---|
449 | if key not in gpu.cache: |
---|
450 | program = gpu.compile_program( |
---|
451 | self.info.name, |
---|
452 | self.source['opencl'], |
---|
453 | self.dtype, |
---|
454 | self.fast, |
---|
455 | self.timestamp) |
---|
456 | variants = ['Iq', 'Iqxy', 'Imagnetic'] |
---|
457 | names = [generate.kernel_name(self.info, k) for k in variants] |
---|
458 | kernels = [getattr(program, k) for k in names] |
---|
459 | data = dict((k, v) for k, v in zip(variants, kernels)) |
---|
460 | # keep a handle to program so GC doesn't collect |
---|
461 | data['program'] = program |
---|
462 | gpu.cache[key] = data |
---|
463 | else: |
---|
464 | data = gpu.cache[key] |
---|
465 | return data[name] |
---|
466 | |
---|
467 | # TODO: check that we don't need a destructor for buffers which go out of scope |
---|
468 | class GpuInput(object): |
---|
469 | """ |
---|
470 | Make q data available to the gpu. |
---|
471 | |
---|
472 | *q_vectors* is a list of q vectors, which will be *[q]* for 1-D data, |
---|
473 | and *[qx, qy]* for 2-D data. Internally, the vectors will be reallocated |
---|
474 | to get the best performance on OpenCL, which may involve shifting and |
---|
475 | stretching the array to better match the memory architecture. Additional |
---|
476 | points will be evaluated with *q=1e-3*. |
---|
477 | |
---|
478 | *dtype* is the data type for the q vectors. The data type should be |
---|
479 | set to match that of the kernel, which is an attribute of |
---|
480 | :class:`GpuProgram`. Note that not all kernels support double |
---|
481 | precision, so even if the program was created for double precision, |
---|
482 | the *GpuProgram.dtype* may be single precision. |
---|
483 | |
---|
484 | Call :meth:`release` when complete. Even if not called directly, the |
---|
485 | buffer will be released when the data object is freed. |
---|
486 | """ |
---|
487 | def __init__(self, q_vectors, dtype=generate.F32): |
---|
488 | # type: (List[np.ndarray], np.dtype) -> None |
---|
489 | # TODO: do we ever need double precision q? |
---|
490 | self.nq = q_vectors[0].size |
---|
491 | self.dtype = np.dtype(dtype) |
---|
492 | self.is_2d = (len(q_vectors) == 2) |
---|
493 | # TODO: stretch input based on get_warp() |
---|
494 | # not doing it now since warp depends on kernel, which is not known |
---|
495 | # at this point, so instead using 32, which is good on the set of |
---|
496 | # architectures tested so far. |
---|
497 | if self.is_2d: |
---|
498 | width = ((self.nq+15)//16)*16 |
---|
499 | self.q = np.empty((width, 2), dtype=dtype) |
---|
500 | self.q[:self.nq, 0] = q_vectors[0] |
---|
501 | self.q[:self.nq, 1] = q_vectors[1] |
---|
502 | else: |
---|
503 | width = ((self.nq+31)//32)*32 |
---|
504 | self.q = np.empty(width, dtype=dtype) |
---|
505 | self.q[:self.nq] = q_vectors[0] |
---|
506 | self.global_size = [self.q.shape[0]] |
---|
507 | self._cache_key = unique_id() |
---|
508 | |
---|
509 | @property |
---|
510 | def q_b(self): |
---|
511 | """Lazy creation of q buffer so it can survive context reset""" |
---|
512 | env = environment() |
---|
513 | key = self._cache_key |
---|
514 | if key not in env.cache: |
---|
515 | context = env.context[self.dtype] |
---|
516 | #print("creating inputs of size", self.global_size) |
---|
517 | buffer = cl.Buffer(context, mf.READ_ONLY | mf.COPY_HOST_PTR, |
---|
518 | hostbuf=self.q) |
---|
519 | env.cache[key] = buffer |
---|
520 | return env.cache[key] |
---|
521 | |
---|
522 | def release(self): |
---|
523 | # type: () -> None |
---|
524 | """ |
---|
525 | Free the buffer associated with the q value |
---|
526 | """ |
---|
527 | environment().free_buffer(id(self)) |
---|
528 | |
---|
529 | def __del__(self): |
---|
530 | # type: () -> None |
---|
531 | self.release() |
---|
532 | |
---|
533 | class GpuKernel(Kernel): |
---|
534 | """ |
---|
535 | Callable SAS kernel. |
---|
536 | |
---|
537 | *model* is the GpuModel object to call |
---|
538 | |
---|
539 | The following attributes are defined: |
---|
540 | |
---|
541 | *info* is the module information |
---|
542 | |
---|
543 | *dtype* is the kernel precision |
---|
544 | |
---|
545 | *dim* is '1d' or '2d' |
---|
546 | |
---|
547 | *result* is a vector to contain the results of the call |
---|
548 | |
---|
549 | The resulting call method takes the *pars*, a list of values for |
---|
550 | the fixed parameters to the kernel, and *pd_pars*, a list of (value,weight) |
---|
551 | vectors for the polydisperse parameters. *cutoff* determines the |
---|
552 | integration limits: any points with combined weight less than *cutoff* |
---|
553 | will not be calculated. |
---|
554 | |
---|
555 | Call :meth:`release` when done with the kernel instance. |
---|
556 | """ |
---|
557 | def __init__(self, model, q_vectors): |
---|
558 | # type: (cl.Kernel, np.dtype, ModelInfo, List[np.ndarray]) -> None |
---|
559 | dtype = model.dtype |
---|
560 | self.q_input = GpuInput(q_vectors, dtype) |
---|
561 | self._model = model |
---|
562 | self._as_dtype = (np.float32 if dtype == generate.F32 |
---|
563 | else np.float64 if dtype == generate.F64 |
---|
564 | else np.float16 if dtype == generate.F16 |
---|
565 | else np.float32) # will never get here, so use np.float32 |
---|
566 | self._cache_key = unique_id() |
---|
567 | |
---|
568 | # attributes accessed from the outside |
---|
569 | self.dim = '2d' if self.q_input.is_2d else '1d' |
---|
570 | self.info = model.info |
---|
571 | self.dtype = model.dtype |
---|
572 | |
---|
573 | # holding place for the returned value |
---|
574 | nout = 2 if self.info.have_Fq and self.dim == '1d' else 1 |
---|
575 | extra_q = 4 # total weight, form volume, shell volume and R_eff |
---|
576 | self.result = np.empty(self.q_input.nq*nout+extra_q, dtype) |
---|
577 | |
---|
578 | @property |
---|
579 | def _result_b(self): |
---|
580 | """Lazy creation of result buffer so it can survive context reset""" |
---|
581 | env = environment() |
---|
582 | key = self._cache_key |
---|
583 | if key not in env.cache: |
---|
584 | context = env.context[self.dtype] |
---|
585 | width = ((self.result.size+31)//32)*32 * self.dtype.itemsize |
---|
586 | buffer = cl.Buffer(context, mf.READ_WRITE, width) |
---|
587 | env.cache[key] = buffer |
---|
588 | return env.cache[key] |
---|
589 | |
---|
590 | def _call_kernel(self, call_details, values, cutoff, magnetic, effective_radius_type): |
---|
591 | # type: (CallDetails, np.ndarray, np.ndarray, float, bool) -> np.ndarray |
---|
592 | env = environment() |
---|
593 | queue = env.queue[self._model.dtype] |
---|
594 | context = queue.context |
---|
595 | |
---|
596 | # Arrange data transfer to/from card |
---|
597 | q_b = self.q_input.q_b |
---|
598 | result_b = self._result_b |
---|
599 | details_b = cl.Buffer(context, mf.READ_ONLY | mf.COPY_HOST_PTR, |
---|
600 | hostbuf=call_details.buffer) |
---|
601 | values_b = cl.Buffer(context, mf.READ_ONLY | mf.COPY_HOST_PTR, |
---|
602 | hostbuf=values) |
---|
603 | |
---|
604 | name = 'Iq' if self.dim == '1d' else 'Imagnetic' if magnetic else 'Iqxy' |
---|
605 | kernel = self._model.fetch_kernel(name) |
---|
606 | kernel_args = [ |
---|
607 | np.uint32(self.q_input.nq), None, None, |
---|
608 | details_b, values_b, q_b, result_b, |
---|
609 | self._as_dtype(cutoff), |
---|
610 | np.uint32(effective_radius_type), |
---|
611 | ] |
---|
612 | #print("Calling OpenCL") |
---|
613 | #call_details.show(values) |
---|
614 | #Call kernel and retrieve results |
---|
615 | wait_for = None |
---|
616 | last_nap = time.clock() |
---|
617 | step = 1000000//self.q_input.nq + 1 |
---|
618 | for start in range(0, call_details.num_eval, step): |
---|
619 | stop = min(start + step, call_details.num_eval) |
---|
620 | #print("queuing",start,stop) |
---|
621 | kernel_args[1:3] = [np.int32(start), np.int32(stop)] |
---|
622 | wait_for = [kernel(queue, self.q_input.global_size, None, |
---|
623 | *kernel_args, wait_for=wait_for)] |
---|
624 | if stop < call_details.num_eval: |
---|
625 | # Allow other processes to run |
---|
626 | wait_for[0].wait() |
---|
627 | current_time = time.clock() |
---|
628 | if current_time - last_nap > 0.5: |
---|
629 | time.sleep(0.001) |
---|
630 | last_nap = current_time |
---|
631 | cl.enqueue_copy(queue, self.result, result_b, wait_for=wait_for) |
---|
632 | #print("result", self.result) |
---|
633 | |
---|
634 | # Free buffers |
---|
635 | for v in (details_b, values_b): |
---|
636 | if v is not None: |
---|
637 | v.release() |
---|
638 | |
---|
639 | def release(self): |
---|
640 | # type: () -> None |
---|
641 | """ |
---|
642 | Release resources associated with the kernel. |
---|
643 | """ |
---|
644 | environment().free_buffer(id(self)) |
---|
645 | self.q_input.release() |
---|
646 | |
---|
647 | def __del__(self): |
---|
648 | # type: () -> None |
---|
649 | self.release() |
---|