from weakref import WeakKeyDictionary
from pyfr.backends.base import (BaseKernelProvider, BaseOrderedMetaKernel,
BasePointwiseKernelProvider,
BaseUnorderedMetaKernel, Kernel)
from pyfr.backends.cuda.generator import CUDAKernelGenerator
from pyfr.backends.cuda.compiler import SourceModule
from pyfr.util import memoize
def get_grid_for_block(block, nrow, ncol=1):
return (-(-nrow // block[0]), -(-ncol // block[1]), 1)
class CUDAKernel(Kernel):
def __init__(self, *args, **kwargs):
super().__init__(*args, **kwargs)
if hasattr(self, 'bind') and hasattr(self, 'add_to_graph'):
self.gnodes = WeakKeyDictionary()
class CUDAOrderedMetaKernel(BaseOrderedMetaKernel):
def add_to_graph(self, graph, dnodes):
for k in self.kernels:
dnodes = [k.add_to_graph(graph, dnodes)]
return dnodes[0]
class CUDAUnorderedMetaKernel(BaseUnorderedMetaKernel):
def add_to_graph(self, graph, dnodes):
nodes = [k.add_to_graph(graph, dnodes) for k in self.kernels]
return graph.graph.add_empty(nodes)
class CUDAKernelProvider(BaseKernelProvider):
@memoize
def _build_kernel(self, name, src, argtypes, argn=[]):
return SourceModule(self.backend, src).get_function(name, argtypes)
def _benchmark(self, kfunc, nbench=4, nwarmup=1):
stream = self.backend.cuda.create_stream()
start_evt = self.backend.cuda.create_event(timing=True)
stop_evt = self.backend.cuda.create_event(timing=True)
for i in range(nbench + nwarmup):
if i == nwarmup:
start_evt.record(stream)
kfunc(stream)
stop_evt.record(stream)
stream.synchronize()
return stop_evt.elapsed_time(start_evt) / nbench
[docs]
class CUDAPointwiseKernelProvider(CUDAKernelProvider,
BasePointwiseKernelProvider):
def __init__(self, *args, **kwargs):
super().__init__(*args, **kwargs)
self._block1d = (64, 1, 1)
self._block2d = (32, 8, 1)
# Pass these block sizes to the generator
class KernelGenerator(CUDAKernelGenerator):
block1d = self._block1d
block2d = self._block2d
self.kernel_generator_cls = KernelGenerator
[docs]
def _instantiate_kernel(self, dims, fun, arglst, argm, argv):
rtargs = []
block = self._block1d if len(dims) == 1 else self._block2d
grid = get_grid_for_block(block, dims[-1])
# Set shared memory carveout locally for kernel
fun.set_shared_size(carveout=25 if fun.shared_mem else 0)
params = fun.make_params(grid, block)
# Process the arguments
for i, k in enumerate(arglst):
if isinstance(k, str):
rtargs.append((i, k))
else:
params.set_arg(i, k)
class PointwiseKernel(CUDAKernel):
if rtargs:
def bind(self, **kwargs):
for i, k in rtargs:
params.set_arg(i, kwargs[k])
# Notify any graphs we're in about our new parameters
for graph, gnode in self.gnodes.items():
graph.stale_kparams[gnode] = params
def add_to_graph(self, graph, deps):
gnode = graph.graph.add_kernel(params, deps)
# If our parameters can change then we need to keep a
# (weak) reference to the graph so we can notify it
if rtargs:
self.gnodes[graph] = gnode
return gnode
def run(self, stream):
fun.exec_async(stream, params)
return PointwiseKernel(argm, argv)