def get_typed_and_scheduled_kernel(self, var_to_dtype_set):
kernel = self.kernel
from loopy.kernel.tools import add_dtypes
if var_to_dtype_set:
var_to_dtype = {}
for var, dtype in var_to_dtype_set:
try:
dest_name = kernel.impl_arg_to_arg[var].name
except KeyError:
dest_name = var
try:
var_to_dtype[dest_name] = dtype
except KeyError:
raise LoopyError("cannot set type for '%s': "
"no known variable/argument with that name"
% var)
kernel = add_dtypes(kernel, var_to_dtype)
from loopy.preprocess import infer_unknown_types
kernel = infer_unknown_types(kernel, expect_completion=True)
if kernel.schedule is None:
from loopy.preprocess import preprocess_kernel
kernel = preprocess_kernel(kernel)
from loopy.schedule import get_one_scheduled_kernel
kernel = get_one_scheduled_kernel(kernel)
return kernel
def get_typed_and_scheduled_kernel(self, var_to_dtype_set):
kernel = self.kernel
from loopy.kernel.tools import add_dtypes
if var_to_dtype_set:
var_to_dtype = {}
for var, dtype in var_to_dtype_set:
try:
dest_name = kernel.impl_arg_to_arg[var].name
except KeyError:
dest_name = var
try:
var_to_dtype[dest_name] = dtype
except KeyError:
raise LoopyError("cannot set type for '%s': "
"no known variable/argument with that name"
% var)
kernel = add_dtypes(kernel, var_to_dtype)
from loopy.preprocess import infer_unknown_types
kernel = infer_unknown_types(kernel, expect_completion=True)
if kernel.schedule is None:
from loopy.preprocess import preprocess_kernel
kernel = preprocess_kernel(kernel)
from loopy.schedule import get_one_scheduled_kernel
kernel = get_one_scheduled_kernel(kernel)
return kernel
def get_typed_and_scheduled_translation_unit_uncached(
self, entrypoint, arg_to_dtype_set):
from loopy.kernel.tools import add_dtypes
from loopy.kernel import KernelState
from loopy.translation_unit import resolve_callables
program = resolve_callables(self.program)
if arg_to_dtype_set:
var_to_dtype = {}
entry_knl = program[entrypoint]
for var, dtype in arg_to_dtype_set:
if var in entry_knl.impl_arg_to_arg:
dest_name = entry_knl.impl_arg_to_arg[var].name
else:
dest_name = var
var_to_dtype[dest_name] = dtype
program = program.with_kernel(add_dtypes(entry_knl, var_to_dtype))
from loopy.type_inference import infer_unknown_types
program = infer_unknown_types(program, expect_completion=True)
if program.state < KernelState.LINEARIZED:
from loopy.preprocess import preprocess_program
program = preprocess_program(program)
from loopy.schedule import get_one_linearized_kernel
for e in program.entrypoints:
program = program.with_kernel(
get_one_linearized_kernel(program[e], program.callables_table))
return program
import pyopencl as cl
import pyopencl.array
knl = lp.make_kernel(
"{ [i,k]: 0<=i<n and 0<=k<3 }", """
c[k,i] = a[k, i + 1]
out[k,i] = c[k,i]
""")
# transform
knl = lp.split_iname(knl, "i", 128, outer_tag="g.0", inner_tag="l.0")
from loopy.kernel.tools import add_dtypes
knl = add_dtypes(knl, {
"a": np.float32,
"c": np.float32,
"out": np.float32,
"n": np.int32
})
# schedule
from loopy.preprocess import preprocess_kernel
knl = preprocess_kernel(knl)
from loopy.schedule import get_one_scheduled_kernel
knl = get_one_scheduled_kernel(knl)
# map schedule onto host or device
print(knl)
import numpy as np
import loopy as lp
import pyopencl as cl
import pyopencl.array
knl = lp.make_kernel(
"{ [i,k]: 0<=i<n and 0<=k<3 }",
"""
c[k,i] = a[k, i + 1]
out[k,i] = c[k,i]
""")
# transform
knl = lp.split_iname(knl, "i", 128, outer_tag="g.0", inner_tag="l.0")
from loopy.kernel.tools import add_dtypes
knl = add_dtypes(knl,
{"a": np.float32, "c": np.float32, "out": np.float32, "n": np.int32})
# schedule
from loopy.preprocess import preprocess_kernel
knl = preprocess_kernel(knl)
from loopy.schedule import get_one_scheduled_kernel
knl = get_one_scheduled_kernel(knl)
# map schedule onto host or device
print(knl)
cgr = lp.generate_code_v2(knl)
print(cgr.device_code())
print(cgr.host_code())
