source: sasmodels/sasmodels/kernelcl.py @ bde38b5

core_shell_microgelscostrafo411magnetic_modelrelease_v0.94release_v0.95ticket-1257-vesicle-productticket_1156ticket_1265_superballticket_822_more_unit_tests
Last change on this file since bde38b5 was bde38b5, checked in by Paul Kienzle <pkienzle@…>, 6 years ago

simplify kernel calling

  • Property mode set to 100644
File size: 22.5 KB
Line 
1"""
2GPU driver for C kernels
3
4There should be a single GPU environment running on the system.  This
5environment is constructed on the first call to :func:`env`, and the
6same environment is returned on each call.
7
8After retrieving the environment, the next step is to create the kernel.
9This is done with a call to :meth:`GpuEnvironment.make_kernel`, which
10returns the type of data used by the kernel.
11
12Next a :class:`GpuData` object should be created with the correct kind
13of data.  This data object can be used by multiple kernels, for example,
14if the target model is a weighted sum of multiple kernels.  The data
15should include any extra evaluation points required to compute the proper
16data smearing.  This need not match the square grid for 2D data if there
17is an index saying which q points are active.
18
19Together the GpuData, the program, and a device form a :class:`GpuKernel`.
20This kernel is used during fitting, receiving new sets of parameters and
21evaluating them.  The output value is stored in an output buffer on the
22devices, where it can be combined with other structure factors and form
23factors and have instrumental resolution effects applied.
24
25In order to use OpenCL for your models, you will need OpenCL drivers for
26your machine.  These should be available from your graphics card vendor.
27Intel provides OpenCL drivers for CPUs as well as their integrated HD
28graphics chipsets.  AMD also provides drivers for Intel CPUs, but as of
29this writing the performance is lacking compared to the Intel drivers.
30NVidia combines drivers for CUDA and OpenCL in one package.  The result
31is a bit messy if you have multiple drivers installed.  You can see which
32drivers are available by starting python and running:
33
34    import pyopencl as cl
35    cl.create_some_context(interactive=True)
36
37Once you have done that, it will show the available drivers which you
38can select.  It will then tell you that you can use these drivers
39automatically by setting the PYOPENCL_CTX environment variable.
40
41Some graphics cards have multiple devices on the same card.  You cannot
42yet use both of them concurrently to evaluate models, but you can run
43the program twice using a different device for each session.
44
45OpenCL kernels are compiled when needed by the device driver.  Some
46drivers produce compiler output even when there is no error.  You
47can see the output by setting PYOPENCL_COMPILER_OUTPUT=1.  It should be
48harmless, albeit annoying.
49"""
50from __future__ import print_function
51
52import os
53import warnings
54import logging
55
56import numpy as np  # type: ignore
57
58try:
59    #raise NotImplementedError("OpenCL not yet implemented for new kernel template")
60    import pyopencl as cl  # type: ignore
61    # Ask OpenCL for the default context so that we know that one exists
62    cl.create_some_context(interactive=False)
63except Exception as exc:
64    warnings.warn("OpenCL startup failed with ***"
65                  + str(exc) + "***; using C compiler instead")
66    raise RuntimeError("OpenCL not available")
67
68from pyopencl import mem_flags as mf
69from pyopencl.characterize import get_fast_inaccurate_build_options
70
71from . import generate
72from .kernel import KernelModel, Kernel
73
74try:
75    from typing import Tuple, Callable, Any
76    from .modelinfo import ModelInfo
77    from .details import CallDetails
78except ImportError:
79    pass
80
81# CRUFT: pyopencl < 2017.1  (as of June 2016 needs quotes around include path)
82def quote_path(v):
83    """
84    Quote the path if it is not already quoted.
85
86    If v starts with '-', then assume that it is a -I option or similar
87    and do not quote it.  This is fragile:  -Ipath with space needs to
88    be quoted.
89    """
90    return '"'+v+'"' if v and ' ' in v and not v[0] in "\"'-" else v
91
92def fix_pyopencl_include():
93    """
94    Monkey patch pyopencl to allow spaces in include file path.
95    """
96    import pyopencl as cl
97    if hasattr(cl, '_DEFAULT_INCLUDE_OPTIONS'):
98        cl._DEFAULT_INCLUDE_OPTIONS = [quote_path(v) for v in cl._DEFAULT_INCLUDE_OPTIONS]
99
100fix_pyopencl_include()
101
102
103# The max loops number is limited by the amount of local memory available
104# on the device.  You don't want to make this value too big because it will
105# waste resources, nor too small because it may interfere with users trying
106# to do their polydispersity calculations.  A value of 1024 should be much
107# larger than necessary given that cost grows as npts^k where k is the number
108# of polydisperse parameters.
109MAX_LOOPS = 2048
110
111
112# Pragmas for enable OpenCL features.  Be sure to protect them so that they
113# still compile even if OpenCL is not present.
114_F16_PRAGMA = """\
115#if defined(__OPENCL_VERSION__) // && !defined(cl_khr_fp16)
116#  pragma OPENCL EXTENSION cl_khr_fp16: enable
117#endif
118"""
119
120_F64_PRAGMA = """\
121#if defined(__OPENCL_VERSION__) // && !defined(cl_khr_fp64)
122#  pragma OPENCL EXTENSION cl_khr_fp64: enable
123#endif
124"""
125
126
127ENV = None
128def environment():
129    # type: () -> "GpuEnvironment"
130    """
131    Returns a singleton :class:`GpuEnvironment`.
132
133    This provides an OpenCL context and one queue per device.
134    """
135    global ENV
136    if ENV is None:
137        ENV = GpuEnvironment()
138    return ENV
139
140def has_type(device, dtype):
141    # type: (cl.Device, np.dtype) -> bool
142    """
143    Return true if device supports the requested precision.
144    """
145    if dtype == generate.F32:
146        return True
147    elif dtype == generate.F64:
148        return "cl_khr_fp64" in device.extensions
149    elif dtype == generate.F16:
150        return "cl_khr_fp16" in device.extensions
151    else:
152        return False
153
154def get_warp(kernel, queue):
155    # type: (cl.Kernel, cl.CommandQueue) -> int
156    """
157    Return the size of an execution batch for *kernel* running on *queue*.
158    """
159    return kernel.get_work_group_info(
160        cl.kernel_work_group_info.PREFERRED_WORK_GROUP_SIZE_MULTIPLE,
161        queue.device)
162
163def _stretch_input(vector, dtype, extra=1e-3, boundary=32):
164    # type: (np.ndarray, np.dtype, float, int) -> np.ndarray
165    """
166    Stretch an input vector to the correct boundary.
167
168    Performance on the kernels can drop by a factor of two or more if the
169    number of values to compute does not fall on a nice power of two
170    boundary.   The trailing additional vector elements are given a
171    value of *extra*, and so f(*extra*) will be computed for each of
172    them.  The returned array will thus be a subset of the computed array.
173
174    *boundary* should be a power of 2 which is at least 32 for good
175    performance on current platforms (as of Jan 2015).  It should
176    probably be the max of get_warp(kernel,queue) and
177    device.min_data_type_align_size//4.
178    """
179    remainder = vector.size % boundary
180    if remainder != 0:
181        size = vector.size + (boundary - remainder)
182        vector = np.hstack((vector, [extra] * (size - vector.size)))
183    return np.ascontiguousarray(vector, dtype=dtype)
184
185
186def 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.  The returned type will
192    be float32 even if the desired type is float64 if any of the
193    devices in the context do not support the cl_khr_fp64 extension.
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
219class GpuEnvironment(object):
220    """
221    GPU context, with possibly many devices, and one queue per device.
222    """
223    def __init__(self):
224        # type: () -> None
225        # find gpu context
226        #self.context = cl.create_some_context()
227
228        self.context = None
229        if 'PYOPENCL_CTX' in os.environ:
230            self._create_some_context()
231
232        if not self.context:
233            self.context = _get_default_context()
234
235        # Byte boundary for data alignment
236        #self.data_boundary = max(d.min_data_type_align_size
237        #                         for d in self.context.devices)
238        self.queues = [cl.CommandQueue(context, context.devices[0])
239                       for context in self.context]
240        self.compiled = {}
241
242    def has_type(self, dtype):
243        # type: (np.dtype) -> bool
244        """
245        Return True if all devices support a given type.
246        """
247        return any(has_type(d, dtype)
248                   for context in self.context
249                   for d in context.devices)
250
251    def get_queue(self, dtype):
252        # type: (np.dtype) -> cl.CommandQueue
253        """
254        Return a command queue for the kernels of type dtype.
255        """
256        for context, queue in zip(self.context, self.queues):
257            if all(has_type(d, dtype) for d in context.devices):
258                return queue
259
260    def get_context(self, dtype):
261        # type: (np.dtype) -> cl.Context
262        """
263        Return a OpenCL context for the kernels of type dtype.
264        """
265        for context in self.context:
266            if all(has_type(d, dtype) for d in context.devices):
267                return context
268
269    def _create_some_context(self):
270        # type: () -> cl.Context
271        """
272        Protected call to cl.create_some_context without interactivity.  Use
273        this if PYOPENCL_CTX is set in the environment.  Sets the *context*
274        attribute.
275        """
276        try:
277            self.context = [cl.create_some_context(interactive=False)]
278        except Exception as exc:
279            warnings.warn(str(exc))
280            warnings.warn("pyopencl.create_some_context() failed")
281            warnings.warn("the environment variable 'PYOPENCL_CTX' might not be set correctly")
282
283    def compile_program(self, name, source, dtype, fast=False):
284        # type: (str, str, np.dtype, bool) -> cl.Program
285        """
286        Compile the program for the device in the given context.
287        """
288        key = "%s-%s%s"%(name, dtype, ("-fast" if fast else ""))
289        if key not in self.compiled:
290            context = self.get_context(dtype)
291            logging.info("building %s for OpenCL %s", key,
292                         context.devices[0].name.strip())
293            program = compile_model(self.get_context(dtype),
294                                    str(source), dtype, fast)
295            self.compiled[key] = program
296        return self.compiled[key]
297
298    def release_program(self, name):
299        # type: (str) -> None
300        """
301        Free memory associated with the program on the device.
302        """
303        if name in self.compiled:
304            self.compiled[name].release()
305            del self.compiled[name]
306
307def _get_default_context():
308    # type: () -> List[cl.Context]
309    """
310    Get an OpenCL context, preferring GPU over CPU, and preferring Intel
311    drivers over AMD drivers.
312    """
313    # Note: on mobile devices there is automatic clock scaling if either the
314    # CPU or the GPU is underutilized; probably doesn't affect us, but we if
315    # it did, it would mean that putting a busy loop on the CPU while the GPU
316    # is running may increase throughput.
317    #
318    # Macbook pro, base install:
319    #     {'Apple': [Intel CPU, NVIDIA GPU]}
320    # Macbook pro, base install:
321    #     {'Apple': [Intel CPU, Intel GPU]}
322    # 2 x nvidia 295 with Intel and NVIDIA opencl drivers installed
323    #     {'Intel': [CPU], 'NVIDIA': [GPU, GPU, GPU, GPU]}
324    gpu, cpu = None, None
325    for platform in cl.get_platforms():
326        # AMD provides a much weaker CPU driver than Intel/Apple, so avoid it.
327        # If someone has bothered to install the AMD/NVIDIA drivers, prefer
328        # them over the integrated graphics driver that may have been supplied
329        # with the CPU chipset.
330        preferred_cpu = (platform.vendor.startswith('Intel')
331                         or platform.vendor.startswith('Apple'))
332        preferred_gpu = (platform.vendor.startswith('Advanced')
333                         or platform.vendor.startswith('NVIDIA'))
334        for device in platform.get_devices():
335            if device.type == cl.device_type.GPU:
336                # If the existing type is not GPU then it will be CUSTOM
337                # or ACCELERATOR so don't override it.
338                if gpu is None or (preferred_gpu and gpu.type == cl.device_type.GPU):
339                    gpu = device
340            elif device.type == cl.device_type.CPU:
341                if cpu is None or preferred_cpu:
342                    cpu = device
343            else:
344                # System has cl.device_type.ACCELERATOR or cl.device_type.CUSTOM
345                # Intel Phi for example registers as an accelerator
346                # Since the user installed a custom device on their system
347                # and went through the pain of sorting out OpenCL drivers for
348                # it, lets assume they really do want to use it as their
349                # primary compute device.
350                gpu = device
351
352    # order the devices by gpu then by cpu; when searching for an available
353    # device by data type they will be checked in this order, which means
354    # that if the gpu supports double then the cpu will never be used (though
355    # we may make it possible to explicitly request the cpu at some point).
356    devices = []
357    if gpu is not None:
358        devices.append(gpu)
359    if cpu is not None:
360        devices.append(cpu)
361    return [cl.Context([d]) for d in devices]
362
363
364class GpuModel(KernelModel):
365    """
366    GPU wrapper for a single model.
367
368    *source* and *model_info* are the model source and interface as returned
369    from :func:`generate.make_source` and :func:`generate.make_model_info`.
370
371    *dtype* is the desired model precision.  Any numpy dtype for single
372    or double precision floats will do, such as 'f', 'float32' or 'single'
373    for single and 'd', 'float64' or 'double' for double.  Double precision
374    is an optional extension which may not be available on all devices.
375    Half precision ('float16','half') may be available on some devices.
376    Fast precision ('fast') is a loose version of single precision, indicating
377    that the compiler is allowed to take shortcuts.
378    """
379    def __init__(self, source, model_info, dtype=generate.F32, fast=False):
380        # type: (Dict[str,str], ModelInfo, np.dtype, bool) -> None
381        self.info = model_info
382        self.source = source
383        self.dtype = dtype
384        self.fast = fast
385        self.program = None # delay program creation
386        self._kernels = None
387
388    def __getstate__(self):
389        # type: () -> Tuple[ModelInfo, str, np.dtype, bool]
390        return self.info, self.source, self.dtype, self.fast
391
392    def __setstate__(self, state):
393        # type: (Tuple[ModelInfo, str, np.dtype, bool]) -> None
394        self.info, self.source, self.dtype, self.fast = state
395        self.program = None
396
397    def make_kernel(self, q_vectors):
398        # type: (List[np.ndarray]) -> "GpuKernel"
399        if self.program is None:
400            compile_program = environment().compile_program
401            self.program = compile_program(
402                self.info.name,
403                self.source['opencl'],
404                self.dtype,
405                self.fast)
406            variants = ['Iq', 'Iqxy', 'Imagnetic']
407            names = [generate.kernel_name(self.info, k) for k in variants]
408            kernels = [getattr(self.program, k) for k in names]
409            self._kernels = dict((k, v) for k, v in zip(variants, kernels))
410        is_2d = len(q_vectors) == 2
411        if is_2d:
412            kernel = [self._kernels['Iqxy'], self._kernels['Imagnetic']]
413        else:
414            kernel = [self._kernels['Iq']]*2
415        return GpuKernel(kernel, self.dtype, self.info, q_vectors)
416
417    def release(self):
418        # type: () -> None
419        """
420        Free the resources associated with the model.
421        """
422        if self.program is not None:
423            environment().release_program(self.info.name)
424            self.program = None
425
426    def __del__(self):
427        # type: () -> None
428        self.release()
429
430# TODO: check that we don't need a destructor for buffers which go out of scope
431class GpuInput(object):
432    """
433    Make q data available to the gpu.
434
435    *q_vectors* is a list of q vectors, which will be *[q]* for 1-D data,
436    and *[qx, qy]* for 2-D data.  Internally, the vectors will be reallocated
437    to get the best performance on OpenCL, which may involve shifting and
438    stretching the array to better match the memory architecture.  Additional
439    points will be evaluated with *q=1e-3*.
440
441    *dtype* is the data type for the q vectors. The data type should be
442    set to match that of the kernel, which is an attribute of
443    :class:`GpuProgram`.  Note that not all kernels support double
444    precision, so even if the program was created for double precision,
445    the *GpuProgram.dtype* may be single precision.
446
447    Call :meth:`release` when complete.  Even if not called directly, the
448    buffer will be released when the data object is freed.
449    """
450    def __init__(self, q_vectors, dtype=generate.F32):
451        # type: (List[np.ndarray], np.dtype) -> None
452        # TODO: do we ever need double precision q?
453        env = environment()
454        self.nq = q_vectors[0].size
455        self.dtype = np.dtype(dtype)
456        self.is_2d = (len(q_vectors) == 2)
457        # TODO: stretch input based on get_warp()
458        # not doing it now since warp depends on kernel, which is not known
459        # at this point, so instead using 32, which is good on the set of
460        # architectures tested so far.
461        if self.is_2d:
462            # Note: 17 rather than 15 because results is 2 elements
463            # longer than input.
464            width = ((self.nq+17)//16)*16
465            self.q = np.empty((width, 2), dtype=dtype)
466            self.q[:self.nq, 0] = q_vectors[0]
467            self.q[:self.nq, 1] = q_vectors[1]
468        else:
469            # Note: 33 rather than 31 because results is 2 elements
470            # longer than input.
471            width = ((self.nq+33)//32)*32
472            self.q = np.empty(width, dtype=dtype)
473            self.q[:self.nq] = q_vectors[0]
474        self.global_size = [self.q.shape[0]]
475        context = env.get_context(self.dtype)
476        #print("creating inputs of size", self.global_size)
477        self.q_b = cl.Buffer(context, mf.READ_ONLY | mf.COPY_HOST_PTR,
478                             hostbuf=self.q)
479
480    def release(self):
481        # type: () -> None
482        """
483        Free the memory.
484        """
485        if self.q_b is not None:
486            self.q_b.release()
487            self.q_b = None
488
489    def __del__(self):
490        # type: () -> None
491        self.release()
492
493class GpuKernel(Kernel):
494    """
495    Callable SAS kernel.
496
497    *kernel* is the GpuKernel object to call
498
499    *model_info* is the module information
500
501    *q_vectors* is the q vectors at which the kernel should be evaluated
502
503    *dtype* is the kernel precision
504
505    The resulting call method takes the *pars*, a list of values for
506    the fixed parameters to the kernel, and *pd_pars*, a list of (value,weight)
507    vectors for the polydisperse parameters.  *cutoff* determines the
508    integration limits: any points with combined weight less than *cutoff*
509    will not be calculated.
510
511    Call :meth:`release` when done with the kernel instance.
512    """
513    def __init__(self, kernel, dtype, model_info, q_vectors):
514        # type: (cl.Kernel, np.dtype, ModelInfo, List[np.ndarray]) -> None
515        q_input = GpuInput(q_vectors, dtype)
516        self.kernel = kernel
517        self.info = model_info
518        self.dtype = dtype
519        self.dim = '2d' if q_input.is_2d else '1d'
520        # plus three for the normalization values
521        self.result = np.empty(q_input.nq+3, dtype)
522
523        # Inputs and outputs for each kernel call
524        # Note: res may be shorter than res_b if global_size != nq
525        env = environment()
526        self.queue = env.get_queue(dtype)
527
528        self.result_b = cl.Buffer(self.queue.context, mf.READ_WRITE,
529                                  q_input.global_size[0] * dtype.itemsize)
530        self.q_input = q_input # allocated by GpuInput above
531
532        self._need_release = [self.result_b, self.q_input]
533        self.real = (np.float32 if dtype == generate.F32
534                     else np.float64 if dtype == generate.F64
535                     else np.float16 if dtype == generate.F16
536                     else np.float32)  # will never get here, so use np.float32
537
538    def __call__(self, call_details, values, cutoff, magnetic):
539        # type: (CallDetails, np.ndarray, np.ndarray, float, bool) -> np.ndarray
540        context = self.queue.context
541        # Arrange data transfer to card
542        details_b = cl.Buffer(context, mf.READ_ONLY | mf.COPY_HOST_PTR,
543                              hostbuf=call_details.buffer)
544        values_b = cl.Buffer(context, mf.READ_ONLY | mf.COPY_HOST_PTR,
545                             hostbuf=values)
546
547        kernel = self.kernel[1 if magnetic else 0]
548        args = [
549            np.uint32(self.q_input.nq), None, None,
550            details_b, values_b, self.q_input.q_b, self.result_b,
551            self.real(cutoff),
552        ]
553        #print("Calling OpenCL")
554        #call_details.show(values)
555        # Call kernel and retrieve results
556        last_call = None
557        step = 100
558        for start in range(0, call_details.num_eval, step):
559            stop = min(start + step, call_details.num_eval)
560            #print("queuing",start,stop)
561            args[1:3] = [np.int32(start), np.int32(stop)]
562            last_call = [kernel(self.queue, self.q_input.global_size,
563                                None, *args, wait_for=last_call)]
564        cl.enqueue_copy(self.queue, self.result, self.result_b)
565        #print("result", self.result)
566
567        # Free buffers
568        for v in (details_b, values_b):
569            if v is not None: v.release()
570
571        scale = values[0]/self.result[self.q_input.nq]
572        background = values[1]
573        #print("scale",scale,values[0],self.result[self.q_input.nq])
574        return scale*self.result[:self.q_input.nq] + background
575        # return self.result[:self.q_input.nq]
576
577    def release(self):
578        # type: () -> None
579        """
580        Release resources associated with the kernel.
581        """
582        for v in self._need_release:
583            v.release()
584        self._need_release = []
585
586    def __del__(self):
587        # type: () -> None
588        self.release()
Note: See TracBrowser for help on using the repository browser.