def get_target(lang, device=None, compiler=None):
"""
Parameters
----------
lang : str
One of the supported languages, {'c', 'cuda', 'opencl'}
device : :class:`pyopencl.Device`
If supplied, and lang is 'opencl', passed to the
:class:`loopy.PyOpenCLTarget`
compiler: str
If supplied, the C-compiler to use
Returns
-------
The correct loopy target type
"""
utils.check_lang(lang)
# set target
if lang == 'opencl':
return lp.PyOpenCLTarget(device=device)
elif lang == 'c':
return lp.ExecutableCTarget(compiler=compiler)
elif lang == 'cuda':
return lp.CudaTarget()
elif lang == 'ispc':
return lp.ISPCTarget()
def test_ispc_streaming_stores():
stream_dtype = np.float32
index_dtype = np.int32
knl = lp.make_kernel(
"{[i]: 0<=i<n}",
"a[i] = b[i] + scalar * c[i]",
target=lp.ISPCTarget(), index_dtype=index_dtype,
name="stream_triad")
vars = ["a", "b", "c", "scalar"]
knl = lp.assume(knl, "n>0")
knl = lp.split_iname(
knl, "i", 2**18, outer_tag="g.0", slabs=(0, 1))
knl = lp.split_iname(knl, "i_inner", 8, inner_tag="l.0")
knl = lp.tag_instructions(knl, "!streaming_store")
knl = lp.add_and_infer_dtypes(knl, {
var: stream_dtype
for var in vars
})
knl = lp.set_argument_order(knl, vars + ["n"])
knl = lp.preprocess_kernel(knl)
knl = lp.get_one_scheduled_kernel(knl)
lp.generate_code_v2(knl).all_code()
def make_knl(name, insn, vars):
knl = lp.make_kernel(
"{[i]: 0<=i<n}",
insn,
target=lp.ISPCTarget(), index_dtype=INDEX_DTYPE,
name="stream_"+name+"_tasks")
knl = transform(knl, vars, STREAM_DTYPE)
return knl
