source: sasmodels/sasmodels/kernelcl.py @ 5464d68

core_shell_microgelscostrafo411magnetic_modelrelease_v0.94release_v0.95ticket-1257-vesicle-productticket_1156ticket_1265_superballticket_822_more_unit_tests
Last change on this file since 5464d68 was 5464d68, checked in by Paul Kienzle <pkienzle@…>, 8 years ago

move opencl pragmas into kernelcl

  • Property mode set to 100644
File size: 19.2 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"""
50import os
51import warnings
52
53import numpy as np
54
55try:
56    import pyopencl as cl
57    # Ask OpenCL for the default context so that we know that one exists
58    cl.create_some_context(interactive=False)
59except Exception as exc:
60    warnings.warn(str(exc))
61    raise RuntimeError("OpenCL not available")
62
63from pyopencl import mem_flags as mf
64from pyopencl.characterize import get_fast_inaccurate_build_options
65
66from . import generate
67
68# The max loops number is limited by the amount of local memory available
69# on the device.  You don't want to make this value too big because it will
70# waste resources, nor too small because it may interfere with users trying
71# to do their polydispersity calculations.  A value of 1024 should be much
72# larger than necessary given that cost grows as npts^k where k is the number
73# of polydisperse parameters.
74MAX_LOOPS = 2048
75
76
77# Pragmas for enable OpenCL features.  Be sure to protect them so that they
78# still compile even if OpenCL is not present.
79_F16_PRAGMA = """\
80#if defined(__OPENCL_VERSION__) // && !defined(cl_khr_fp16)
81#  pragma OPENCL EXTENSION cl_khr_fp16: enable
82#endif
83"""
84
85_F64_PRAGMA = """\
86#if defined(__OPENCL_VERSION__) // && !defined(cl_khr_fp64)
87#  pragma OPENCL EXTENSION cl_khr_fp64: enable
88#endif
89"""
90
91
92ENV = None
93def environment():
94    """
95    Returns a singleton :class:`GpuEnvironment`.
96
97    This provides an OpenCL context and one queue per device.
98    """
99    global ENV
100    if ENV is None:
101        ENV = GpuEnvironment()
102    return ENV
103
104def has_type(device, dtype):
105    """
106    Return true if device supports the requested precision.
107    """
108    if dtype == generate.F32:
109        return True
110    elif dtype == generate.F64:
111        return "cl_khr_fp64" in device.extensions
112    elif dtype == generate.F16:
113        return "cl_khr_fp16" in device.extensions
114    else:
115        return False
116
117def get_warp(kernel, queue):
118    """
119    Return the size of an execution batch for *kernel* running on *queue*.
120    """
121    return kernel.get_work_group_info(
122        cl.kernel_work_group_info.PREFERRED_WORK_GROUP_SIZE_MULTIPLE,
123        queue.device)
124
125def _stretch_input(vector, dtype, extra=1e-3, boundary=32):
126    """
127    Stretch an input vector to the correct boundary.
128
129    Performance on the kernels can drop by a factor of two or more if the
130    number of values to compute does not fall on a nice power of two
131    boundary.   The trailing additional vector elements are given a
132    value of *extra*, and so f(*extra*) will be computed for each of
133    them.  The returned array will thus be a subset of the computed array.
134
135    *boundary* should be a power of 2 which is at least 32 for good
136    performance on current platforms (as of Jan 2015).  It should
137    probably be the max of get_warp(kernel,queue) and
138    device.min_data_type_align_size//4.
139    """
140    remainder = vector.size % boundary
141    if remainder != 0:
142        size = vector.size + (boundary - remainder)
143        vector = np.hstack((vector, [extra] * (size - vector.size)))
144    return np.ascontiguousarray(vector, dtype=dtype)
145
146
147def compile_model(context, source, dtype, fast=False):
148    """
149    Build a model to run on the gpu.
150
151    Returns the compiled program and its type.  The returned type will
152    be float32 even if the desired type is float64 if any of the
153    devices in the context do not support the cl_khr_fp64 extension.
154    """
155    dtype = np.dtype(dtype)
156    if not all(has_type(d, dtype) for d in context.devices):
157        raise RuntimeError("%s not supported for devices"%dtype)
158
159    source_list = [generate.convert_type(source, dtype)]
160
161    if dtype == generate.F16:
162        source_list.insert(0, _F16_PRAGMA)
163    elif dtype == generate.F64:
164        source_list.insert(0, _F64_PRAGMA)
165
166
167    # Note: USE_SINCOS makes the intel cpu slower under opencl
168    if context.devices[0].type == cl.device_type.GPU:
169        source_list.insert(0, "#define USE_SINCOS\n")
170    options = (get_fast_inaccurate_build_options(context.devices[0])
171               if fast else [])
172    source = "\n".join(source)
173    program = cl.Program(context, source).build(options=options)
174    return program
175
176
177# for now, this returns one device in the context
178# TODO: create a context that contains all devices on all platforms
179class GpuEnvironment(object):
180    """
181    GPU context, with possibly many devices, and one queue per device.
182    """
183    def __init__(self):
184        # find gpu context
185        #self.context = cl.create_some_context()
186
187        self.context = None
188        if 'PYOPENCL_CTX' in os.environ:
189            self._create_some_context()
190
191        if not self.context:
192            self.context = _get_default_context()
193
194        # Byte boundary for data alignment
195        #self.data_boundary = max(d.min_data_type_align_size
196        #                         for d in self.context.devices)
197        self.queues = [cl.CommandQueue(context, context.devices[0])
198                       for context in self.context]
199        self.compiled = {}
200
201    def has_type(self, dtype):
202        """
203        Return True if all devices support a given type.
204        """
205        dtype = generate.F32 if dtype == 'fast' else np.dtype(dtype)
206        return any(has_type(d, dtype)
207                   for context in self.context
208                   for d in context.devices)
209
210    def get_queue(self, dtype):
211        """
212        Return a command queue for the kernels of type dtype.
213        """
214        for context, queue in zip(self.context, self.queues):
215            if all(has_type(d, dtype) for d in context.devices):
216                return queue
217
218    def get_context(self, dtype):
219        """
220        Return a OpenCL context for the kernels of type dtype.
221        """
222        for context, queue in zip(self.context, self.queues):
223            if all(has_type(d, dtype) for d in context.devices):
224                return context
225
226    def _create_some_context(self):
227        """
228        Protected call to cl.create_some_context without interactivity.  Use
229        this if PYOPENCL_CTX is set in the environment.  Sets the *context*
230        attribute.
231        """
232        try:
233            self.context = [cl.create_some_context(interactive=False)]
234        except Exception as exc:
235            warnings.warn(str(exc))
236            warnings.warn("pyopencl.create_some_context() failed")
237            warnings.warn("the environment variable 'PYOPENCL_CTX' might not be set correctly")
238
239    def compile_program(self, name, source, dtype, fast=False):
240        """
241        Compile the program for the device in the given context.
242        """
243        key = "%s-%s-%s"%(name, dtype, fast)
244        if key not in self.compiled:
245            #print("compiling",name)
246            dtype = np.dtype(dtype)
247            program = compile_model(self.get_context(dtype), source, dtype, fast)
248            self.compiled[key] = program
249        return self.compiled[key]
250
251    def release_program(self, name):
252        """
253        Free memory associated with the program on the device.
254        """
255        if name in self.compiled:
256            self.compiled[name].release()
257            del self.compiled[name]
258
259def _get_default_context():
260    """
261    Get an OpenCL context, preferring GPU over CPU, and preferring Intel
262    drivers over AMD drivers.
263    """
264    # Note: on mobile devices there is automatic clock scaling if either the
265    # CPU or the GPU is underutilized; probably doesn't affect us, but we if
266    # it did, it would mean that putting a busy loop on the CPU while the GPU
267    # is running may increase throughput.
268    #
269    # Macbook pro, base install:
270    #     {'Apple': [Intel CPU, NVIDIA GPU]}
271    # Macbook pro, base install:
272    #     {'Apple': [Intel CPU, Intel GPU]}
273    # 2 x nvidia 295 with Intel and NVIDIA opencl drivers installed
274    #     {'Intel': [CPU], 'NVIDIA': [GPU, GPU, GPU, GPU]}
275    gpu, cpu = None, None
276    for platform in cl.get_platforms():
277        # AMD provides a much weaker CPU driver than Intel/Apple, so avoid it.
278        # If someone has bothered to install the AMD/NVIDIA drivers, prefer them over the integrated
279        # graphics driver that may have been supplied with the CPU chipset.
280        preferred_cpu = platform.vendor.startswith('Intel') or platform.vendor.startswith('Apple')
281        preferred_gpu = platform.vendor.startswith('Advanced') or platform.vendor.startswith('NVIDIA')
282        for device in platform.get_devices():
283            if device.type == cl.device_type.GPU:
284                # If the existing type is not GPU then it will be CUSTOM or ACCELERATOR,
285                # so don't override it.
286                if gpu is None or (preferred_gpu and gpu.type == cl.device_type.GPU):
287                    gpu = device
288            elif device.type == cl.device_type.CPU:
289                if cpu is None or preferred_cpu:
290                    cpu = device
291            else:
292                # System has cl.device_type.ACCELERATOR or cl.device_type.CUSTOM
293                # Intel Phi for example registers as an accelerator
294                # Since the user installed a custom device on their system and went through the
295                # pain of sorting out OpenCL drivers for it, lets assume they really do want to
296                # use it as their primary compute device.
297                gpu = device
298
299    # order the devices by gpu then by cpu; when searching for an available device by data type they
300    # will be checked in this order, which means that if the gpu supports double then the cpu will never
301    # be used (though we may make it possible to explicitly request the cpu at some point).
302    devices = []
303    if gpu is not None:
304        devices.append(gpu)
305    if cpu is not None:
306        devices.append(cpu)
307    return [cl.Context([d]) for d in devices]
308
309
310class GpuModel(object):
311    """
312    GPU wrapper for a single model.
313
314    *source* and *model_info* are the model source and interface as returned
315    from :func:`generate.make_source` and :func:`generate.make_model_info`.
316
317    *dtype* is the desired model precision.  Any numpy dtype for single
318    or double precision floats will do, such as 'f', 'float32' or 'single'
319    for single and 'd', 'float64' or 'double' for double.  Double precision
320    is an optional extension which may not be available on all devices.
321    Half precision ('float16','half') may be available on some devices.
322    Fast precision ('fast') is a loose version of single precision, indicating
323    that the compiler is allowed to take shortcuts.
324    """
325    def __init__(self, source, model_info, dtype=generate.F32):
326        self.info = model_info
327        self.source = source
328        self.dtype = generate.F32 if dtype == 'fast' else np.dtype(dtype)
329        self.fast = (dtype == 'fast')
330        self.program = None # delay program creation
331
332    def __getstate__(self):
333        return self.info, self.source, self.dtype, self.fast
334
335    def __setstate__(self, state):
336        self.info, self.source, self.dtype, self.fast = state
337        self.program = None
338
339    def __call__(self, q_vectors):
340        if self.program is None:
341            compiler = environment().compile_program
342            self.program = compiler(self.info['name'], self.source, self.dtype,
343                                    self.fast)
344        is_2d = len(q_vectors) == 2
345        kernel_name = generate.kernel_name(self.info, is_2d)
346        kernel = getattr(self.program, kernel_name)
347        return GpuKernel(kernel, self.info, q_vectors, self.dtype)
348
349    def release(self):
350        """
351        Free the resources associated with the model.
352        """
353        if self.program is not None:
354            environment().release_program(self.info['name'])
355            self.program = None
356
357    def __del__(self):
358        self.release()
359
360# TODO: check that we don't need a destructor for buffers which go out of scope
361class GpuInput(object):
362    """
363    Make q data available to the gpu.
364
365    *q_vectors* is a list of q vectors, which will be *[q]* for 1-D data,
366    and *[qx, qy]* for 2-D data.  Internally, the vectors will be reallocated
367    to get the best performance on OpenCL, which may involve shifting and
368    stretching the array to better match the memory architecture.  Additional
369    points will be evaluated with *q=1e-3*.
370
371    *dtype* is the data type for the q vectors. The data type should be
372    set to match that of the kernel, which is an attribute of
373    :class:`GpuProgram`.  Note that not all kernels support double
374    precision, so even if the program was created for double precision,
375    the *GpuProgram.dtype* may be single precision.
376
377    Call :meth:`release` when complete.  Even if not called directly, the
378    buffer will be released when the data object is freed.
379    """
380    def __init__(self, q_vectors, dtype=generate.F32):
381        # TODO: do we ever need double precision q?
382        env = environment()
383        self.nq = q_vectors[0].size
384        self.dtype = np.dtype(dtype)
385        self.is_2d = (len(q_vectors) == 2)
386        # TODO: stretch input based on get_warp()
387        # not doing it now since warp depends on kernel, which is not known
388        # at this point, so instead using 32, which is good on the set of
389        # architectures tested so far.
390        self.q_vectors = [_stretch_input(q, self.dtype, 32) for q in q_vectors]
391        context = env.get_context(self.dtype)
392        self.global_size = [self.q_vectors[0].size]
393        #print("creating inputs of size", self.global_size)
394        self.q_buffers = [
395            cl.Buffer(context, mf.READ_ONLY | mf.COPY_HOST_PTR, hostbuf=q)
396            for q in self.q_vectors
397        ]
398
399    def release(self):
400        """
401        Free the memory.
402        """
403        for b in self.q_buffers:
404            b.release()
405        self.q_buffers = []
406
407    def __del__(self):
408        self.release()
409
410class GpuKernel(object):
411    """
412    Callable SAS kernel.
413
414    *kernel* is the GpuKernel object to call
415
416    *model_info* is the module information
417
418    *q_vectors* is the q vectors at which the kernel should be evaluated
419
420    *dtype* is the kernel precision
421
422    The resulting call method takes the *pars*, a list of values for
423    the fixed parameters to the kernel, and *pd_pars*, a list of (value,weight)
424    vectors for the polydisperse parameters.  *cutoff* determines the
425    integration limits: any points with combined weight less than *cutoff*
426    will not be calculated.
427
428    Call :meth:`release` when done with the kernel instance.
429    """
430    def __init__(self, kernel, model_info, q_vectors, dtype):
431        q_input = GpuInput(q_vectors, dtype)
432        self.kernel = kernel
433        self.info = model_info
434        self.res = np.empty(q_input.nq, q_input.dtype)
435        dim = '2d' if q_input.is_2d else '1d'
436        self.fixed_pars = model_info['partype']['fixed-' + dim]
437        self.pd_pars = model_info['partype']['pd-' + dim]
438
439        # Inputs and outputs for each kernel call
440        # Note: res may be shorter than res_b if global_size != nq
441        env = environment()
442        self.queue = env.get_queue(dtype)
443        self.loops_b = cl.Buffer(self.queue.context, mf.READ_WRITE,
444                                 2 * MAX_LOOPS * q_input.dtype.itemsize)
445        self.res_b = cl.Buffer(self.queue.context, mf.READ_WRITE,
446                               q_input.global_size[0] * q_input.dtype.itemsize)
447        self.q_input = q_input
448
449        self._need_release = [self.loops_b, self.res_b, self.q_input]
450
451    def __call__(self, fixed_pars, pd_pars, cutoff=1e-5):
452        real = (np.float32 if self.q_input.dtype == generate.F32
453                else np.float64 if self.q_input.dtype == generate.F64
454                else np.float16 if self.q_input.dtype == generate.F16
455                else np.float32)  # will never get here, so use np.float32
456
457        #print "pars", fixed_pars, pd_pars
458        res_bi = self.res_b
459        nq = np.uint32(self.q_input.nq)
460        if pd_pars:
461            cutoff = real(cutoff)
462            loops_N = [np.uint32(len(p[0])) for p in pd_pars]
463            loops = np.hstack(pd_pars) \
464                if pd_pars else np.empty(0, dtype=self.q_input.dtype)
465            loops = np.ascontiguousarray(loops.T, self.q_input.dtype).flatten()
466            #print("loops",Nloops, loops)
467
468            #import sys; print("opencl eval",pars)
469            #print("opencl eval",pars)
470            if len(loops) > 2 * MAX_LOOPS:
471                raise ValueError("too many polydispersity points")
472
473            loops_bi = self.loops_b
474            cl.enqueue_copy(self.queue, loops_bi, loops)
475            loops_l = cl.LocalMemory(len(loops.data))
476            #ctx = environment().context
477            #loops_bi = cl.Buffer(ctx, mf.READ_ONLY|mf.COPY_HOST_PTR, hostbuf=loops)
478            dispersed = [loops_bi, loops_l, cutoff] + loops_N
479        else:
480            dispersed = []
481        fixed = [real(p) for p in fixed_pars]
482        args = self.q_input.q_buffers + [res_bi, nq] + dispersed + fixed
483        self.kernel(self.queue, self.q_input.global_size, None, *args)
484        cl.enqueue_copy(self.queue, self.res, res_bi)
485
486        return self.res
487
488    def release(self):
489        """
490        Release resources associated with the kernel.
491        """
492        for v in self._need_release:
493            v.release()
494        self._need_release = []
495
496    def __del__(self):
497        self.release()
Note: See TracBrowser for help on using the repository browser.