def __call__(self, *args, **kwargs):
MAX_BLOCK_COUNT = 1024
SMALL_SEQ_COUNT = 4
s1_func = self.stage1_func
s2_func = self.stage2_func
kernel_wrapper = kwargs.get("kernel_wrapper")
if kernel_wrapper is not None:
s1_func = kernel_wrapper(s1_func)
s2_func = kernel_wrapper(s2_func)
stream = kwargs.get("stream")
from gpuarray import empty
f = s1_func
arg_types = self.stage1_arg_types
while True:
invocation_args = []
vectors = []
for arg, arg_tp in zip(args, arg_types):
if arg_tp == "P":
vectors.append(arg)
invocation_args.append(arg.gpudata)
else:
invocation_args.append(arg)
repr_vec = vectors[0]
sz = repr_vec.size
if sz <= self.block_size * SMALL_SEQ_COUNT * MAX_BLOCK_COUNT:
total_block_size = SMALL_SEQ_COUNT * self.block_size
block_count = (sz + total_block_size - 1) // total_block_size
seq_count = SMALL_SEQ_COUNT
else:
block_count = MAX_BLOCK_COUNT
macroblock_size = block_count * self.block_size
seq_count = (sz + macroblock_size - 1) // macroblock_size
if block_count == 1:
result = empty((), self.dtype_out, repr_vec.allocator)
else:
result = empty((block_count, ), self.dtype_out,
repr_vec.allocator)
#print block_count, seq_count, self.block_size
f((block_count, 1), stream,
*([result.gpudata] + invocation_args + [seq_count, sz]))
if block_count == 1:
return result
else:
f = s2_func
arg_types = self.stage2_arg_types
args = [result]
def __call__(self, *args, **kwargs):
MAX_BLOCK_COUNT = 1024
SMALL_SEQ_COUNT = 4
s1_func = self.stage1_func
s2_func = self.stage2_func
kernel_wrapper = kwargs.get("kernel_wrapper")
if kernel_wrapper is not None:
s1_func = kernel_wrapper(s1_func)
s2_func = kernel_wrapper(s2_func)
stream = kwargs.get("stream")
from gpuarray import empty
f = s1_func
arg_types = self.stage1_arg_types
while True:
invocation_args = []
vectors = []
for arg, arg_tp in zip(args, arg_types):
if arg_tp == "P":
vectors.append(arg)
invocation_args.append(arg.gpudata)
else:
invocation_args.append(arg)
repr_vec = vectors[0]
sz = repr_vec.size
if sz <= self.block_size*SMALL_SEQ_COUNT*MAX_BLOCK_COUNT:
total_block_size = SMALL_SEQ_COUNT*self.block_size
block_count = (sz + total_block_size - 1) // total_block_size
seq_count = SMALL_SEQ_COUNT
else:
block_count = MAX_BLOCK_COUNT
macroblock_size = block_count*self.block_size
seq_count = (sz + macroblock_size - 1) // macroblock_size
if block_count == 1:
result = empty((), self.dtype_out, repr_vec.allocator)
else:
result = empty((block_count,), self.dtype_out, repr_vec.allocator)
#print block_count, seq_count, self.block_size
f((block_count, 1), stream,
*([result.gpudata]+invocation_args+[seq_count, sz]))
if block_count == 1:
return result
else:
f = s2_func
arg_types = self.stage2_arg_types
args = [result]
def elemwise2(a,
op,
b,
ary,
odtype=None,
oper=None,
op_tmpl="res[i] = (%(out_t)s)%(a)s %(op)s (%(out_t)s)%(b)s",
broadcast=False):
ndim_extend = True
if not isinstance(a, gpuarray.GpuArray):
a = numpy.asarray(a)
ndim_extend = False
if not isinstance(b, gpuarray.GpuArray):
b = numpy.asarray(b)
ndim_extend = False
if odtype is None:
odtype = get_common_dtype(a, b, True)
a_arg = as_argument(a, 'a')
b_arg = as_argument(b, 'b')
args = [ArrayArg(odtype, 'res'), a_arg, b_arg]
if ndim_extend:
if a.ndim != b.ndim:
nd = max(a.ndim, b.ndim)
if a.ndim < nd:
a = a.reshape(((1, ) * (nd - a.ndim)) + a.shape)
if b.ndim < nd:
b = b.reshape(((1, ) * (nd - b.ndim)) + b.shape)
out_shape = tuple(max(sa, sb) for sa, sb in zip(a.shape, b.shape))
res = gpuarray.empty(out_shape,
dtype=odtype,
context=ary.context,
cls=ary.__class__)
else:
res = ary._empty_like_me(dtype=odtype)
if oper is None:
oper = op_tmpl % {
'a': a_arg.expr(),
'op': op,
'b': b_arg.expr(),
'out_t': dtype_to_ctype(odtype)
}
k = ElemwiseKernel(ary.context, args, oper)
k(res, a, b, broadcast=broadcast)
return res
def __call__(self, *args, **kwargs):
_, nd, dims, strs, offsets, contig = check_args(args,
collapse=False,
broadcast=False)
out = kwargs.pop('out', None)
if len(kwargs) != 0:
raise TypeError('Unexpected keyword argument: %s' %
kwargs.keys()[0])
n = prod(dims)
out_shape = tuple(d for i, d in enumerate(dims) if not self.redux[i])
gs = prod(out_shape)
if gs == 0:
gs = 1
n /= gs
if gs > self.context.maxgsize:
raise ValueError("Array to big to be reduced along the "
"selected axes")
if out is None:
out = gpuarray.empty(out_shape,
context=self.context,
dtype=self.dtype_out)
else:
if out.shape != out_shape or out.dtype != self.dtype_out:
raise TypeError(
"Out array is not of expected type "
"(expected %s %s, got %s %s)" %
(out_shape, self.dtype_out, out.shape, out.dtype))
#Don't compile and cache for nothing for big size
if self.init_local_size < n:
k, _, _, ls = self._get_basic_kernel(self.init_local_size, nd)
else:
k, _, _, ls = self._get_basic_kernel(n, nd)
kargs = [n, out]
kargs.extend(dims)
for i, arg in enumerate(args):
kargs.append(arg)
if isinstance(arg, gpuarray.GpuArray):
kargs.append(offsets[i])
kargs.extend(strs[i])
k(*kargs, ls=ls, gs=gs)
return out
def __call__(self, *args, **kwargs):
_, nd, dims, strs, offsets, contig = check_args(args, collapse=False,
broadcast=False)
out = kwargs.pop('out', None)
if len(kwargs) != 0:
raise TypeError('Unexpected keyword argument: %s' %
kwargs.keys()[0])
n = prod(dims)
out_shape = tuple(d for i, d in enumerate(dims) if not self.redux[i])
gs = prod(out_shape)
if gs == 0:
gs = 1
n /= gs
if gs > self.context.maxgsize:
raise ValueError("Array to big to be reduced along the "
"selected axes")
if out is None:
out = gpuarray.empty(out_shape, context=self.context,
dtype=self.dtype_out)
else:
if out.shape != out_shape or out.dtype != self.dtype_out:
raise TypeError("Out array is not of expected type "
"(expected %s %s, got %s %s)" % (
out_shape, self.dtype_out, out.shape, out.dtype))
#Don't compile and cache for nothing for big size
if self.init_local_size < n:
k, _, _, ls = self._get_basic_kernel(self.init_local_size, nd)
else:
k, _, _, ls = self._get_basic_kernel(n, nd)
kargs = [n, out]
kargs.extend(dims)
for i, arg in enumerate(args):
kargs.append(arg)
if isinstance(arg, gpuarray.GpuArray):
kargs.append(offsets[i])
kargs.extend(strs[i])
k(*kargs, ls=ls, gs=gs)
return out
def elemwise2(a, op, b, ary, odtype=None, oper=None,
op_tmpl="res[i] = (%(out_t)s)%(a)s %(op)s (%(out_t)s)%(b)s",
broadcast=False):
ndim_extend = True
if not isinstance(a, gpuarray.GpuArray):
a = numpy.asarray(a)
ndim_extend = False
if not isinstance(b, gpuarray.GpuArray):
b = numpy.asarray(b)
ndim_extend = False
if odtype is None:
odtype = get_common_dtype(a, b, True)
a_arg = as_argument(a, 'a')
b_arg = as_argument(b, 'b')
args = [ArrayArg(odtype, 'res'), a_arg, b_arg]
if ndim_extend:
if a.ndim != b.ndim:
nd = max(a.ndim, b.ndim)
if a.ndim < nd:
a = a.reshape(((1,) * (nd - a.ndim))+a.shape)
if b.ndim < nd:
b = b.reshape(((1,) * (nd - b.ndim))+b.shape)
out_shape = tuple(max(sa, sb) for sa, sb in zip(a.shape, b.shape))
res = gpuarray.empty(out_shape, dtype=odtype, context=ary.context,
cls=ary.__class__)
else:
res = ary._empty_like_me(dtype=odtype)
if oper is None:
oper = op_tmpl % {'a': a_arg.expr(), 'op': op, 'b': b_arg.expr(),
'out_t': dtype_to_ctype(odtype)}
k = ElemwiseKernel(ary.context, args, oper)
k(res, a, b, broadcast=broadcast)
return res
def __call__(self, *args, **kwargs):
MAX_BLOCK_COUNT = 1024
SMALL_SEQ_COUNT = 4
s1_func = self.stage1_func
s2_func = self.stage2_func
kernel_wrapper = kwargs.get("kernel_wrapper")
if kernel_wrapper is not None:
s1_func = kernel_wrapper(s1_func)
s2_func = kernel_wrapper(s2_func)
stream = kwargs.get("stream")
from gpuarray import empty
f = s1_func
arg_types = self.stage1_arg_types
stage1_args = args
while True:
invocation_args = []
vectors = []
for arg, arg_tp in zip(args, arg_types):
if arg_tp == "P":
if not arg.flags.forc:
raise RuntimeError("ReductionKernel cannot "
"deal with non-contiguous arrays")
vectors.append(arg)
invocation_args.append(arg.gpudata)
else:
invocation_args.append(arg)
repr_vec = vectors[0]
sz = repr_vec.size
if sz <= self.block_size*SMALL_SEQ_COUNT*MAX_BLOCK_COUNT:
total_block_size = SMALL_SEQ_COUNT*self.block_size
block_count = (sz + total_block_size - 1) // total_block_size
seq_count = SMALL_SEQ_COUNT
else:
block_count = MAX_BLOCK_COUNT
macroblock_size = block_count*self.block_size
seq_count = (sz + macroblock_size - 1) // macroblock_size
if block_count == 1:
result = empty((), self.dtype_out, repr_vec.allocator)
else:
result = empty((block_count,), self.dtype_out, repr_vec.allocator)
kwargs = dict(shared_size=self.block_size*self.dtype_out.itemsize)
#print block_count, seq_count, self.block_size, sz
f((block_count, 1), (self.block_size, 1, 1), stream,
*([result.gpudata]+invocation_args+[seq_count, sz]),
**kwargs)
if block_count == 1:
return result
else:
f = s2_func
arg_types = self.stage2_arg_types
args = (result,) + stage1_args