From 9ca753fed36d7af22715b758ef86c2ae92c1839a Mon Sep 17 00:00:00 2001
From: Andreas Kloeckner <inform@tiker.net>
Date: Thu, 6 Feb 2014 17:59:44 -0600
Subject: [PATCH] Don't store the target CL device in the kernel object
---
doc/reference.rst | 2 +
doc/tutorial.rst | 31 ++++----
examples/hello-loopy.py | 2 +-
loopy/__init__.py | 4 +-
loopy/auto_test.py | 15 ++--
loopy/check.py | 23 +++---
loopy/codegen/__init__.py | 9 ++-
loopy/compiled.py | 10 +--
loopy/kernel/__init__.py | 58 ++++++++++-----
loopy/kernel/creation.py | 5 +-
loopy/preprocess.py | 30 +++++---
loopy/schedule.py | 12 ++--
test/test_dg.py | 4 +-
test/test_linalg.py | 28 ++++----
test/test_loopy.py | 147 ++++++++++++++++++--------------------
test/test_nbody.py | 20 +++---
test/test_sem_reagan.py | 20 +++---
17 files changed, 236 insertions(+), 184 deletions(-)
diff --git a/doc/reference.rst b/doc/reference.rst
index 164a44b5e..fab13029c 100644
--- a/doc/reference.rst
+++ b/doc/reference.rst
@@ -426,6 +426,8 @@ Finishing up
.. autofunction:: generate_loop_schedules
+.. autofunction:: get_one_scheduled_kernel
+
.. autofunction:: generate_code
Running
diff --git a/doc/tutorial.rst b/doc/tutorial.rst
index 72c6c5b0c..7785a4259 100644
--- a/doc/tutorial.rst
+++ b/doc/tutorial.rst
@@ -52,7 +52,7 @@ one vector, doubles it, and writes it to another.
.. doctest::
- >>> knl = lp.make_kernel(ctx.devices[0],
+ >>> knl = lp.make_kernel(
... "{ [i]: 0<=i<n }",
... "out[i] = 2*a[i]")
@@ -80,9 +80,6 @@ The parts that you see here are the two main components of a loopy kernel:
See :ref:`expression-syntax` for a full list of allowed constructs in the
left- and right-hand side expression of an assignment.
-Loopy also needs to know which OpenCL device to target. ``ctx.devices[0]``
-specifies the first device in our OpenCL context.
-
As you create and transform kernels, it's useful to know that you can
always see loopy's view of a kernel by printing it.
@@ -234,6 +231,8 @@ call :func:`loopy.generate_code`:
.. doctest::
>>> typed_knl = lp.add_dtypes(knl, dict(a=np.float32))
+ >>> typed_knl = lp.preprocess_kernel(typed_knl, device=ctx.devices[0])
+ >>> typed_knl = lp.get_one_scheduled_kernel(typed_knl)
>>> code, _ = lp.generate_code(typed_knl)
>>> print code
<BLANKLINE>
@@ -257,7 +256,7 @@ argument:
.. doctest::
>>> # WARNING: Incorrect.
- >>> knl = lp.make_kernel(ctx.devices[0],
+ >>> knl = lp.make_kernel(
... "{ [i,j]: 0<=i,j<n }",
... """
... out[j,i] = a[i,j]
@@ -284,7 +283,7 @@ an explicit dependency:
.. doctest::
>>> # WARNING: Incorrect.
- >>> knl = lp.make_kernel(ctx.devices[0],
+ >>> knl = lp.make_kernel(
... "{ [i,j]: 0<=i,j<n }",
... """
... out[j,i] = a[i,j] {id=transpose}
@@ -384,7 +383,7 @@ with identical bounds, for the use of the transpose:
.. doctest::
- >>> knl = lp.make_kernel(ctx.devices[0],
+ >>> knl = lp.make_kernel(
... "{ [i,j,ii,jj]: 0<=i,j,ii,jj<n }",
... """
... out[j,i] = a[i,j] {id=transpose}
@@ -429,7 +428,7 @@ zero-fill kernel?
.. doctest::
- >>> knl = lp.make_kernel(ctx.devices[0],
+ >>> knl = lp.make_kernel(
... "{ [i,j]: 0<=i,j<n }",
... """
... a[i,j] = 0
@@ -514,7 +513,7 @@ Consider this example:
.. doctest::
- >>> knl = lp.make_kernel(ctx.devices[0],
+ >>> knl = lp.make_kernel(
... "{ [i]: 0<=i<n }",
... "a[i] = 0", assumptions="n>=0")
>>> knl = lp.split_iname(knl, "i", 16)
@@ -564,7 +563,7 @@ commonly called 'loop tiling':
.. doctest::
- >>> knl = lp.make_kernel(ctx.devices[0],
+ >>> knl = lp.make_kernel(
... "{ [i,j]: 0<=i,j<n }",
... "out[i,j] = a[j,i]",
... assumptions="n mod 16 = 0 and n >= 1")
@@ -604,7 +603,7 @@ loop's tag to ``"unr"``:
.. doctest::
- >>> knl = lp.make_kernel(ctx.devices[0],
+ >>> knl = lp.make_kernel(
... "{ [i]: 0<=i<n }",
... "a[i] = 0", assumptions="n>=0 and n mod 4 = 0")
>>> orig_knl = knl
@@ -679,7 +678,7 @@ Let's try this out on our vector fill kernel by creating workgroups of size
.. doctest::
- >>> knl = lp.make_kernel(ctx.devices[0],
+ >>> knl = lp.make_kernel(
... "{ [i]: 0<=i<n }",
... "a[i] = 0", assumptions="n>=0")
>>> knl = lp.split_iname(knl, "i", 128,
@@ -724,7 +723,7 @@ assumption:
.. doctest::
- >>> knl = lp.make_kernel(ctx.devices[0],
+ >>> knl = lp.make_kernel(
... "{ [i]: 0<=i<n }",
... "a[i] = 0", assumptions="n>=0")
>>> orig_knl = knl
@@ -821,7 +820,7 @@ Attempting to create this kernel results in an error:
.. doctest::
- >>> lp.make_kernel(ctx.devices[0],
+ >>> lp.make_kernel(
... "{ [i]: 0<=i<n }",
... """
... out[i] = 5
@@ -848,7 +847,7 @@ be told in order for the error to disappear--note the *assumptions* argument:
.. doctest::
- >>> knl = lp.make_kernel(ctx.devices[0],
+ >>> knl = lp.make_kernel(
... "{ [i]: 0<=i<n }",
... """
... out[i] = 5
@@ -868,7 +867,7 @@ This kernel performs a simple transposition of an input matrix:
.. doctest::
- >>> knl = lp.make_kernel(ctx.devices[0],
+ >>> knl = lp.make_kernel(
... "{ [i,j]: 0<=i,j<n }",
... """
... out[j,i] = a[i,j]
diff --git a/examples/hello-loopy.py b/examples/hello-loopy.py
index c7c7ade30..efdbf1315 100644
--- a/examples/hello-loopy.py
+++ b/examples/hello-loopy.py
@@ -13,7 +13,7 @@ a = cl.array.arange(queue, n, dtype=np.float32)
# create
# ------
-knl = lp.make_kernel(ctx.devices[0],
+knl = lp.make_kernel(
"{ [i]: 0<=i<n }",
"out[i] = 2*a[i]")
diff --git a/loopy/__init__.py b/loopy/__init__.py
index 51925f9b6..8934aebc6 100644
--- a/loopy/__init__.py
+++ b/loopy/__init__.py
@@ -60,7 +60,7 @@ from loopy.padding import (split_arg_axis, find_padding_multiple,
add_padding)
from loopy.preprocess import (preprocess_kernel, realize_reduction,
infer_unknown_types)
-from loopy.schedule import generate_loop_schedules
+from loopy.schedule import generate_loop_schedules, get_one_scheduled_kernel
from loopy.codegen import generate_code
from loopy.compiled import CompiledKernel
from loopy.options import Options
@@ -94,7 +94,7 @@ __all__ = [
"infer_argument_dtypes", "add_and_infer_dtypes",
"preprocess_kernel", "realize_reduction", "infer_unknown_types",
- "generate_loop_schedules",
+ "generate_loop_schedules", "get_one_scheduled_kernel",
"generate_code",
"CompiledKernel",
diff --git a/loopy/auto_test.py b/loopy/auto_test.py
index 62797d00a..46c21ccad 100644
--- a/loopy/auto_test.py
+++ b/loopy/auto_test.py
@@ -28,6 +28,7 @@ import numpy as np
import pyopencl as cl
import pyopencl.array as cl_array
+import loopy as lp
AUTO_TEST_SKIP_RUN = False
@@ -36,7 +37,6 @@ import logging
logger = logging.getLogger(__name__)
-
# {{{ create random argument arrays for testing
def fill_rand(ary):
@@ -398,8 +398,9 @@ def auto_test_vs_ref(
ref_queue = cl.CommandQueue(ref_ctx,
properties=cl.command_queue_properties.PROFILING_ENABLE)
- import loopy as lp
- for knl in lp.generate_loop_schedules(ref_knl):
+ pp_ref_knl = lp.preprocess_kernel(ref_knl, device=dev)
+
+ for knl in lp.generate_loop_schedules(pp_ref_knl):
ref_sched_kernel = knl
break
@@ -487,6 +488,12 @@ def auto_test_vs_ref(
test_kernels = test_knl
else:
+ from loopy.kernel import kernel_state
+ if test_knl.state not in [
+ kernel_state.PREPROCESSED,
+ kernel_state.SCHEDULED]:
+ test_knl = lp.preprocess_kernel(test_knl, device=ctx.devices[0])
+
if not test_knl.schedule:
test_kernels = lp.generate_loop_schedules(test_knl)
else:
@@ -604,7 +611,7 @@ def auto_test_vs_ref(
print("elapsed: %g s event, %s s marker-event %g s wall "
"(%d rounds)%s" % (
- elapsed, elapsed_evt_2, elapsed_wall, timing_rounds, rates))
+ elapsed, elapsed_evt_2, elapsed_wall, timing_rounds, rates))
if do_check:
ref_rates = ""
diff --git a/loopy/check.py b/loopy/check.py
index 70d16c2d3..df8a61b18 100644
--- a/loopy/check.py
+++ b/loopy/check.py
@@ -373,7 +373,7 @@ def pre_schedule_checks(kernel):
# {{{ pre-code-generation checks
-def check_sizes(kernel):
+def check_sizes(kernel, device):
import loopy as lp
from loopy.diagnostic import LoopyAdvisory
@@ -386,7 +386,7 @@ def check_sizes(kernel):
glens, llens = kernel.get_grid_sizes_as_exprs()
if (max(len(glens), len(llens))
- > kernel.device.max_work_item_dimensions):
+ > device.max_work_item_dimensions):
raise LoopyError("too many work item dimensions")
from pymbolic import evaluate
@@ -401,15 +401,15 @@ def check_sizes(kernel):
% name, LoopyAdvisory)
else:
for i in range(len(llens)):
- if llens[i] > kernel.device.max_work_item_sizes[i]:
+ if llens[i] > device.max_work_item_sizes[i]:
raise LoopyError("group axis %d too big" % i)
from pytools import product
- if product(llens) > kernel.device.max_work_group_size:
+ if product(llens) > device.max_work_group_size:
raise LoopyError("work group too big")
from pyopencl.characterize import usable_local_mem_size
- if kernel.local_mem_use() > usable_local_mem_size(kernel.device):
+ if kernel.local_mem_use() > usable_local_mem_size(device):
raise LoopyError("using too much local memory")
from loopy.kernel.data import ConstantArg
@@ -417,7 +417,7 @@ def check_sizes(kernel):
1 for arg in kernel.args
if isinstance(arg, ConstantArg))
- if const_arg_count > kernel.device.max_constant_args:
+ if const_arg_count > device.max_constant_args:
raise LoopyError("too many constant arguments")
@@ -457,11 +457,18 @@ def check_that_shapes_and_strides_are_arguments(kernel):
arg.name, ", ".join(deps-integer_arg_names)))
-def pre_codegen_checks(kernel):
+def pre_codegen_checks(kernel, device=None):
try:
logger.info("pre-codegen check %s: start" % kernel.name)
- check_sizes(kernel)
+ if device is not None:
+ check_sizes(kernel, device)
+ else:
+ from loopy.diagnostic import warn
+ warn(kernel, "no_device_in_pre_codegen_checks",
+ "No device parameter was passed to loopy.pre_codegen_checks. "
+ "Perhaps you want to pass a device argument to generate_code.")
+
check_that_shapes_and_strides_are_arguments(kernel)
logger.info("pre-codegen check %s: done" % kernel.name)
diff --git a/loopy/codegen/__init__.py b/loopy/codegen/__init__.py
index afd8ed762..c33f336cb 100644
--- a/loopy/codegen/__init__.py
+++ b/loopy/codegen/__init__.py
@@ -23,6 +23,7 @@ THE SOFTWARE.
"""
+from loopy.diagnostic import LoopyError
from pytools import Record
import islpy as isl
@@ -287,16 +288,20 @@ class ImplementedDataInfo(Record):
# {{{ main code generation entrypoint
-def generate_code(kernel):
+def generate_code(kernel, device=None):
if kernel.schedule is None:
from loopy.schedule import get_one_scheduled_kernel
kernel = get_one_scheduled_kernel(kernel)
+ from loopy.kernel import kernel_state
+ if kernel.state != kernel_state.SCHEDULED:
+ raise LoopyError("cannot generate code for a kernel that has not been "
+ "scheduled")
from loopy.preprocess import infer_unknown_types
kernel = infer_unknown_types(kernel, expect_completion=True)
from loopy.check import pre_codegen_checks
- pre_codegen_checks(kernel)
+ pre_codegen_checks(kernel, device=device)
from cgen import (FunctionBody, FunctionDeclaration,
Value, Module, Block,
diff --git a/loopy/compiled.py b/loopy/compiled.py
index 47bcac8b8..bfed0c5c8 100644
--- a/loopy/compiled.py
+++ b/loopy/compiled.py
@@ -632,7 +632,7 @@ class _CLKernelInfo(Record):
class CompiledKernel:
- def __init__(self, context, kernel, codegen_kwargs={}):
+ def __init__(self, context, kernel):
"""
:arg kernel: may be a loopy.LoopKernel, a generator returning kernels
(a warning will be issued if more than one is returned). If the
@@ -641,7 +641,6 @@ class CompiledKernel:
"""
self.context = context
- self.codegen_kwargs = codegen_kwargs
self.kernel = kernel
self.packing_controller = SeparateArrayPackingController(kernel)
@@ -676,6 +675,9 @@ class CompiledKernel:
kernel = infer_unknown_types(kernel, expect_completion=True)
if kernel.schedule is None:
+ from loopy.preprocess import preprocess_kernel
+ kernel = preprocess_kernel(kernel, self.context.devices[0])
+
from loopy.schedule import get_one_scheduled_kernel
kernel = get_one_scheduled_kernel(kernel)
@@ -686,7 +688,7 @@ class CompiledKernel:
kernel = self.get_typed_and_scheduled_kernel(arg_to_dtype_set)
from loopy.codegen import generate_code
- code, impl_arg_info = generate_code(kernel, **self.codegen_kwargs)
+ code, impl_arg_info = generate_code(kernel, device=self.context.devices[0])
if self.kernel.options.write_cl:
output = code
@@ -724,7 +726,7 @@ class CompiledKernel:
kernel = self.get_typed_and_scheduled_kernel(arg_to_dtype)
from loopy.codegen import generate_code
- code, arg_info = generate_code(kernel, **self.codegen_kwargs)
+ code, arg_info = generate_code(kernel, device=self.context.devices[0])
return code
def get_highlighted_code(self, arg_to_dtype=None):
diff --git a/loopy/kernel/__init__.py b/loopy/kernel/__init__.py
index c6f3abc14..81a64484c 100644
--- a/loopy/kernel/__init__.py
+++ b/loopy/kernel/__init__.py
@@ -26,12 +26,12 @@ THE SOFTWARE.
import numpy as np
-from pytools import Record, memoize_method
+from pytools import RecordWithoutPickling, memoize_method
import islpy as isl
from islpy import dim_type
import re
-from pytools import UniqueNameGenerator, generate_unique_possibilities
+from pytools import UniqueNameGenerator, generate_unique_names
from loopy.library.function import (
default_function_mangler,
@@ -79,13 +79,15 @@ class _UniqueVarNameGenerator(UniqueNameGenerator):
# {{{ loop kernel object
-class LoopKernel(Record):
- """These correspond more or less directly to arguments of
- :func:`loopy.make_kernel`.
+class kernel_state:
+ INITIAL = 0
+ PREPROCESSED = 1
+ SCHEDULED = 2
- .. attribute:: device
- :class:`pyopencl.Device`
+class LoopKernel(RecordWithoutPickling):
+ """These correspond more or less directly to arguments of
+ :func:`loopy.make_kernel`.
.. attribute:: domains
@@ -137,11 +139,15 @@ class LoopKernel(Record):
.. attribute:: options
An instance of :class:`loopy.Options`
+
+ .. attribute:: state
+
+ A value from :class:`kernel_state`.
"""
# {{{ constructor
- def __init__(self, device, domains, instructions, args=[], schedule=None,
+ def __init__(self, domains, instructions, args=[], schedule=None,
name="loopy_kernel",
preambles=[],
preamble_generators=[default_preamble_generator],
@@ -167,6 +173,8 @@ class LoopKernel(Record):
isl_context=None,
options=None,
+ state=kernel_state.INITIAL,
+
# When kernels get intersected in slab decomposition,
# their grid sizes shouldn't change. This provides
# a way to forward sub-kernel grid size requests.
@@ -247,8 +255,15 @@ class LoopKernel(Record):
# overwrites method down below
self.get_grid_sizes = get_grid_sizes
- Record.__init__(self,
- device=device, domains=domains,
+ if state not in [
+ kernel_state.INITIAL,
+ kernel_state.PREPROCESSED,
+ kernel_state.SCHEDULED,
+ ]:
+ raise ValueError("invalid value for 'state'")
+
+ RecordWithoutPickling.__init__(self,
+ domains=domains,
instructions=instructions,
args=args,
schedule=schedule,
@@ -269,7 +284,8 @@ class LoopKernel(Record):
symbol_manglers=symbol_manglers,
index_dtype=index_dtype,
isl_context=isl_context,
- options=options)
+ options=options,
+ state=state)
# }}}
@@ -310,7 +326,7 @@ class LoopKernel(Record):
used_ids = set(insn.id for insn in insns) | extra_used_ids
- for id_str in generate_unique_possibilities(based_on):
+ for id_str in generate_unique_names(based_on):
if id_str not in used_ids:
return id_str
@@ -745,7 +761,7 @@ class LoopKernel(Record):
dom_intersect_assumptions, iname_idx)
.coalesce())
- class BoundsRecord(Record):
+ class BoundsRecord(RecordWithoutPickling):
pass
size = (upper_bound_pw_aff - lower_bound_pw_aff + 1)
@@ -814,8 +830,6 @@ class LoopKernel(Record):
tgt_dict[tag.axis] = size
- max_dims = self.device.max_work_item_dimensions
-
def to_dim_tuple(size_dict, which, forced_sizes={}):
forced_sizes = forced_sizes.copy()
@@ -840,10 +854,6 @@ class LoopKernel(Record):
size_list.append(size_dict[cur_axis])
- if len(size_list) > max_dims:
- raise ValueError("more %s dimensions assigned than supported "
- "by hardware (%d > %d)" % (which, len(size_list), max_dims))
-
return tuple(size_list)
return (to_dim_tuple(global_sizes, "global"),
@@ -1016,6 +1026,16 @@ class LoopKernel(Record):
# }}}
+ def __getinitargs__(self):
+ result = dict(
+ (key, getattr(self, key))
+ for key in self.__class__.fields
+ if hasattr(self, key))
+
+ result.pop("cache_manager", None)
+
+ return result
+
# }}}
# vim: foldmethod=marker
diff --git a/loopy/kernel/creation.py b/loopy/kernel/creation.py
index 1d5c3d668..e3f7ab14e 100644
--- a/loopy/kernel/creation.py
+++ b/loopy/kernel/creation.py
@@ -935,10 +935,9 @@ def resolve_wildcard_deps(knl):
# {{{ kernel creation top-level
-def make_kernel(device, domains, instructions, kernel_data=["..."], **kwargs):
+def make_kernel(domains, instructions, kernel_data=["..."], **kwargs):
"""User-facing kernel creation entrypoint.
- :arg device: :class:`pyopencl.Device`
:arg domains: :class:`islpy.BasicSet`
:arg instructions:
:arg kernel_data:
@@ -1101,7 +1100,7 @@ def make_kernel(device, domains, instructions, kernel_data=["..."], **kwargs):
kernel_args = arg_guesser.guess_kernel_args_if_requested(kernel_args)
from loopy.kernel import LoopKernel
- knl = LoopKernel(device, domains, instructions, kernel_args,
+ knl = LoopKernel(domains, instructions, kernel_args,
temporary_variables=temporary_variables,
silenced_warnings=silenced_warnings,
options=options,
diff --git a/loopy/preprocess.py b/loopy/preprocess.py
index ec1e750d3..60643f559 100644
--- a/loopy/preprocess.py
+++ b/loopy/preprocess.py
@@ -985,12 +985,12 @@ def assign_automatic_axes(kernel, axis=0, local_size=None):
# {{{ temp storage adjust for bank conflict
-def adjust_local_temp_var_storage(kernel):
+def adjust_local_temp_var_storage(kernel, device):
logger.debug("%s: adjust temp var storage" % kernel.name)
new_temp_vars = {}
- lmem_size = cl_char.usable_local_mem_size(kernel.device)
+ lmem_size = cl_char.usable_local_mem_size(device)
for temp_var in kernel.temporary_variables.itervalues():
if not temp_var.is_local:
new_temp_vars[temp_var.name] = \
@@ -1013,12 +1013,12 @@ def adjust_local_temp_var_storage(kernel):
# below to avoid bank conflicts
from pytools import product
- if kernel.device.local_mem_type == cl.device_local_mem_type.GLOBAL:
+ if device.local_mem_type == cl.device_local_mem_type.GLOBAL:
# FIXME: could try to avoid cache associativity disasters
new_storage_shape = storage_shape
- elif kernel.device.local_mem_type == cl.device_local_mem_type.LOCAL:
- min_mult = cl_char.local_memory_bank_count(kernel.device)
+ elif device.local_mem_type == cl.device_local_mem_type.LOCAL:
+ min_mult = cl_char.local_memory_bank_count(device)
good_incr = None
new_storage_shape = storage_shape
min_why_not = None
@@ -1028,7 +1028,7 @@ def adjust_local_temp_var_storage(kernel):
test_storage_shape = storage_shape[:]
test_storage_shape[-1] = test_storage_shape[-1] + increment
new_mult, why_not = cl_char.why_not_local_access_conflict_free(
- kernel.device, temp_var.dtype.itemsize,
+ device, temp_var.dtype.itemsize,
temp_var.shape, test_storage_shape)
# will choose smallest increment 'automatically'
@@ -1062,7 +1062,12 @@ def adjust_local_temp_var_storage(kernel):
# }}}
-def preprocess_kernel(kernel):
+def preprocess_kernel(kernel, device=None):
+ from loopy.kernel import kernel_state
+ if kernel.state != kernel_state.INITIAL:
+ raise LoopyError("cannot re-preprocess an already preprocessed "
+ "kernel")
+
logger.info("%s: preprocess start" % kernel.name)
from loopy.subst import expand_subst
@@ -1096,11 +1101,18 @@ def preprocess_kernel(kernel):
kernel = assign_automatic_axes(kernel)
kernel = find_boostability(kernel)
kernel = limit_boostability(kernel)
- kernel = adjust_local_temp_var_storage(kernel)
+
+ if device is not None:
+ kernel = adjust_local_temp_var_storage(kernel, device)
+ else:
+ from loopy.diagnostic import warn
+ warn(kernel, "no_device_in_preprocess",
+ "no device parameter was passed to loopy.preprocess")
logger.info("%s: preprocess done" % kernel.name)
- return kernel
+ return kernel.copy(
+ state=kernel_state.PREPROCESSED)
diff --git a/loopy/schedule.py b/loopy/schedule.py
index 697322f11..419672a96 100644
--- a/loopy/schedule.py
+++ b/loopy/schedule.py
@@ -999,10 +999,12 @@ def insert_barriers(kernel, schedule, reverse, kind, level=0):
# {{{ main scheduling entrypoint
def generate_loop_schedules(kernel, debug_args={}):
- loop_priority = kernel.loop_priority
+ from loopy.kernel import kernel_state
+ if kernel.state != kernel_state.PREPROCESSED:
+ raise LoopyError("cannot schedule a kernel that has not been "
+ "preprocessed")
- from loopy.preprocess import preprocess_kernel
- kernel = preprocess_kernel(kernel)
+ loop_priority = kernel.loop_priority
from loopy.check import pre_schedule_checks
pre_schedule_checks(kernel)
@@ -1049,7 +1051,9 @@ def generate_loop_schedules(kernel, debug_args={}):
reverse=False, kind="local")
debug.stop()
- yield kernel.copy(schedule=gen_sched)
+ yield kernel.copy(
+ schedule=gen_sched,
+ state=kernel_state.SCHEDULED)
debug.start()
schedule_count += 1
diff --git a/test/test_dg.py b/test/test_dg.py
index 956bee2d7..291da4484 100644
--- a/test/test_dg.py
+++ b/test/test_dg.py
@@ -47,7 +47,7 @@ def test_dg_volume(ctx_factory):
K = 10000
- knl = lp.make_kernel(ctx.devices[0], [
+ knl = lp.make_kernel([
"{[n,m,k]: 0<= n,m < Np and 0<= k < K}",
],
"""
@@ -175,7 +175,7 @@ def no_test_dg_surface(ctx_factory):
K = 10000
- knl = lp.make_kernel(ctx.devices[0],
+ knl = lp.make_kernel(
[
"{[m,n,k]: 0<= m < NfpNfaces and 0<= n < Np and 0<= k < K }"
],
diff --git a/test/test_linalg.py b/test/test_linalg.py
index ef0d0c0ca..47c7600d6 100644
--- a/test/test_linalg.py
+++ b/test/test_linalg.py
@@ -73,7 +73,7 @@ def test_axpy(ctx_factory):
vec.make_float4(1, 2, 3, 4), vec.make_float4(6, 7, 8, 9)),
(np.float32, None, 5, 7),
]:
- knl = lp.make_kernel(ctx.devices[0],
+ knl = lp.make_kernel(
"[n] -> {[i]: 0<=i<n}",
[
"z[i] = a*x[i]+b*y[i]"
@@ -121,7 +121,7 @@ def test_transpose(ctx_factory):
n = get_suitable_size(ctx)
- knl = lp.make_kernel(ctx.devices[0],
+ knl = lp.make_kernel(
"{[i,j]: 0<=i,j<%d}" % n,
[
"b[i, j] = a[j, i]"
@@ -155,7 +155,7 @@ def test_plain_matrix_mul(ctx_factory):
(cl_array.vec.float4, check_float4, 4),
(np.float32, None, 1),
]:
- knl = lp.make_kernel(ctx.devices[0],
+ knl = lp.make_kernel(
"{[i,j,k]: 0<=i,j,k<%d}" % n,
[
"c[i, j] = sum(k, a[i, k]*b[k, j])"
@@ -189,7 +189,7 @@ def test_variable_size_matrix_mul(ctx_factory):
n = get_suitable_size(ctx)
- knl = lp.make_kernel(ctx.devices[0],
+ knl = lp.make_kernel(
"[n] -> {[i,j,k]: 0<=i,j,k<n}",
[
"c[i, j] = sum(k, a[i, k]*b[k, j]) {id=labl}"
@@ -226,7 +226,7 @@ def test_rank_one(ctx_factory):
#n = int(get_suitable_size(ctx)**(2.7/2))
n = 16**3
- knl = lp.make_kernel(ctx.devices[0],
+ knl = lp.make_kernel(
"[n] -> {[i,j]: 0<=i,j<n}",
[
"c[i, j] = a[i]*b[j] {id=mylabel, priority =5}"
@@ -303,7 +303,7 @@ def test_troublesome_premagma_fermi_matrix_mul(ctx_factory):
n = 6*16*2
- knl = lp.make_kernel(ctx.devices[0],
+ knl = lp.make_kernel(
"{[i,j,k]: 0<=i,j,k<%d}" % n,
[
"c[i, j] = sum(k, a[i, k]*b[k, j])"
@@ -340,7 +340,7 @@ def test_intel_matrix_mul(ctx_factory):
n = 128+32
- knl = lp.make_kernel(ctx.devices[0],
+ knl = lp.make_kernel(
"{[i,j,k]: 0<=i,j,k<%d}" % n,
[
"c[i, j] = sum(k, a[i, k]*b[k, j])"
@@ -392,7 +392,7 @@ def test_magma_fermi_matrix_mul(ctx_factory):
ctx, cl.mem_flags.READ_ONLY, cl.mem_object_type.IMAGE2D):
pytest.skip("image format not supported")
- knl = lp.make_kernel(ctx.devices[0],
+ knl = lp.make_kernel(
"{[i,j,k]: 0<=i,j,k<%d}" % n,
[
"c[i, j] = sum(k, a[i, k]*b[k, j])"
@@ -439,7 +439,7 @@ def test_image_matrix_mul(ctx_factory):
ctx, cl.mem_flags.READ_ONLY, cl.mem_object_type.IMAGE2D):
pytest.skip("image format not supported")
- knl = lp.make_kernel(ctx.devices[0],
+ knl = lp.make_kernel(
"{[i,j,k]: 0<=i,j,k<%d}" % n,
[
"c[i, j] = sum(k, a[i, k]*b[k, j])"
@@ -477,7 +477,7 @@ def test_image_matrix_mul_ilp(ctx_factory):
n = get_suitable_size(ctx)
- knl = lp.make_kernel(ctx.devices[0],
+ knl = lp.make_kernel(
"{[i,j,k]: 0<=i,j,k<%d}" % n,
[
"c[i, j] = sum(k, a[i, k]*b[k, j])"
@@ -510,12 +510,11 @@ def test_image_matrix_mul_ilp(ctx_factory):
@pytest.mark.skipif("sys.version_info < (2,6)")
def test_ilp_race_matmul(ctx_factory):
dtype = np.float32
- ctx = ctx_factory()
order = "C"
n = 9
- knl = lp.make_kernel(ctx.devices[0],
+ knl = lp.make_kernel(
"{[i,j,k]: 0<=i,j,k<%d}" % n,
[
"c[i, j] = sum(k, a[i, k]*b[k, j])"
@@ -534,6 +533,7 @@ def test_ilp_race_matmul(ctx_factory):
from loopy.diagnostic import WriteRaceConditionWarning
from warnings import catch_warnings
with catch_warnings(record=True) as warn_list:
+ knl = lp.preprocess_kernel(knl)
list(lp.generate_loop_schedules(knl))
assert any(isinstance(w.message, WriteRaceConditionWarning)
@@ -548,7 +548,7 @@ def test_fancy_matrix_mul(ctx_factory):
n = get_suitable_size(ctx)
- knl = lp.make_kernel(ctx.devices[0],
+ knl = lp.make_kernel(
"[n] -> {[i,j,k]: 0<=i,j,k<n }",
[
"c[i, j] = sum(k, a[i, k]*b[k, j])"
@@ -582,7 +582,7 @@ def test_small_batched_matvec(ctx_factory):
K = 9997
Np = 36
- knl = lp.make_kernel(ctx.devices[0],
+ knl = lp.make_kernel(
"[K] -> {[i,j,k]: 0<=k<K and 0<= i,j < %d}" % Np,
[
"result[k, i] = sum(j, d[i, j]*f[k, j])"
diff --git a/test/test_loopy.py b/test/test_loopy.py
index b19df76b4..07a737900 100644
--- a/test/test_loopy.py
+++ b/test/test_loopy.py
@@ -52,7 +52,7 @@ __all__ = [
def test_complicated_subst(ctx_factory):
ctx = ctx_factory()
- knl = lp.make_kernel(ctx.devices[0],
+ knl = lp.make_kernel(
"{[i]: 0<=i<n}",
"""
f(x) := x*a[x]
@@ -84,7 +84,7 @@ def test_complicated_subst(ctx_factory):
def test_type_inference_no_artificial_doubles(ctx_factory):
ctx = ctx_factory()
- knl = lp.make_kernel(ctx.devices[0],
+ knl = lp.make_kernel(
"{[i]: 0<=i<n}",
"""
<> bb = a[i] - b[i]
@@ -98,6 +98,7 @@ def test_type_inference_no_artificial_doubles(ctx_factory):
],
assumptions="n>=1")
+ knl = lp.preprocess_kernel(knl, ctx.devices[0])
for k in lp.generate_loop_schedules(knl):
code = lp.generate_code(k)
assert "double" not in code
@@ -106,7 +107,7 @@ def test_type_inference_no_artificial_doubles(ctx_factory):
def test_sized_and_complex_literals(ctx_factory):
ctx = ctx_factory()
- knl = lp.make_kernel(ctx.devices[0],
+ knl = lp.make_kernel(
"{[i]: 0<=i<n}",
"""
<> aa = 5jf
@@ -129,7 +130,7 @@ def test_sized_and_complex_literals(ctx_factory):
def test_simple_side_effect(ctx_factory):
ctx = ctx_factory()
- knl = lp.make_kernel(ctx.devices[0],
+ knl = lp.make_kernel(
"{[i,j]: 0<=i,j<100}",
"""
a[i] = a[i] + 1
@@ -137,6 +138,7 @@ def test_simple_side_effect(ctx_factory):
[lp.GlobalArg("a", np.float32, shape=(100,))]
)
+ knl = lp.preprocess_kernel(knl, ctx.devices[0])
kernel_gen = lp.generate_loop_schedules(knl)
for gen_knl in kernel_gen:
@@ -148,7 +150,7 @@ def test_simple_side_effect(ctx_factory):
def test_nonsense_reduction(ctx_factory):
ctx = ctx_factory()
- knl = lp.make_kernel(ctx.devices[0],
+ knl = lp.make_kernel(
"{[i]: 0<=i<100}",
"""
a[i] = sum(i, 2)
@@ -158,13 +160,13 @@ def test_nonsense_reduction(ctx_factory):
import pytest
with pytest.raises(RuntimeError):
- list(lp.generate_loop_schedules(knl))
+ knl = lp.preprocess_kernel(knl, ctx.devices[0])
def test_owed_barriers(ctx_factory):
ctx = ctx_factory()
- knl = lp.make_kernel(ctx.devices[0],
+ knl = lp.make_kernel(
"{[i]: 0<=i<100}",
[
"<float32> z[i] = a[i]"
@@ -174,6 +176,7 @@ def test_owed_barriers(ctx_factory):
knl = lp.tag_inames(knl, dict(i="l.0"))
+ knl = lp.preprocess_kernel(knl, ctx.devices[0])
kernel_gen = lp.generate_loop_schedules(knl)
for gen_knl in kernel_gen:
@@ -184,7 +187,7 @@ def test_owed_barriers(ctx_factory):
def test_wg_too_small(ctx_factory):
ctx = ctx_factory()
- knl = lp.make_kernel(ctx.devices[0],
+ knl = lp.make_kernel(
"{[i]: 0<=i<100}",
[
"<float32> z[i] = a[i] {id=copy}"
@@ -194,6 +197,7 @@ def test_wg_too_small(ctx_factory):
knl = lp.tag_inames(knl, dict(i="l.0"))
+ knl = lp.preprocess_kernel(knl, ctx.devices[0])
kernel_gen = lp.generate_loop_schedules(knl)
import pytest
@@ -205,7 +209,7 @@ def test_wg_too_small(ctx_factory):
def test_join_inames(ctx_factory):
ctx = ctx_factory()
- knl = lp.make_kernel(ctx.devices[0],
+ knl = lp.make_kernel(
"{[i,j]: 0<=i,j<16}",
[
"b[i,j] = 2*a[i,j]"
@@ -227,7 +231,7 @@ def test_join_inames(ctx_factory):
def test_divisibility_assumption(ctx_factory):
ctx = ctx_factory()
- knl = lp.make_kernel(ctx.devices[0],
+ knl = lp.make_kernel(
"[n] -> {[i]: 0<=i<n}",
[
"b[i] = 2*a[i]"
@@ -243,6 +247,7 @@ def test_divisibility_assumption(ctx_factory):
knl = lp.split_iname(knl, "i", 16)
+ knl = lp.preprocess_kernel(knl, ctx.devices[0])
for k in lp.generate_loop_schedules(knl):
code = lp.generate_code(k)
assert "if" not in code
@@ -254,7 +259,7 @@ def test_divisibility_assumption(ctx_factory):
def test_multi_cse(ctx_factory):
ctx = ctx_factory()
- knl = lp.make_kernel(ctx.devices[0],
+ knl = lp.make_kernel(
"{[i]: 0<=i<100}",
[
"<float32> z[i] = a[i] + a[i]**2"
@@ -265,6 +270,7 @@ def test_multi_cse(ctx_factory):
knl = lp.split_iname(knl, "i", 16, inner_tag="l.0")
knl = lp.add_prefetch(knl, "a", [])
+ knl = lp.preprocess_kernel(knl, ctx.devices[0])
kernel_gen = lp.generate_loop_schedules(knl)
for gen_knl in kernel_gen:
@@ -279,7 +285,7 @@ def test_stencil(ctx_factory):
# non-unifiable, two-constant-segments PwAff as the base index)
n = 256
- knl = lp.make_kernel(ctx.devices[0],
+ knl = lp.make_kernel(
"{[i,j]: 0<= i,j < %d}" % n,
[
"a_offset(ii, jj) := a[ii+1, jj+1]",
@@ -320,7 +326,7 @@ def test_stencil(ctx_factory):
def test_stencil_with_overfetch(ctx_factory):
ctx = ctx_factory()
- knl = lp.make_kernel(ctx.devices[0],
+ knl = lp.make_kernel(
"{[i,j]: 0<= i,j < n}",
[
"a_offset(ii, jj) := a[ii+2, jj+2]",
@@ -361,7 +367,7 @@ def test_stencil_with_overfetch(ctx_factory):
def test_eq_constraint(ctx_factory):
ctx = ctx_factory()
- knl = lp.make_kernel(ctx.devices[0],
+ knl = lp.make_kernel(
"{[i,j]: 0<= i,j < 32}",
[
"a[i] = b[i]"
@@ -374,6 +380,7 @@ def test_eq_constraint(ctx_factory):
knl = lp.split_iname(knl, "i", 16, outer_tag="g.0")
knl = lp.split_iname(knl, "i_inner", 16, outer_tag=None, inner_tag="l.0")
+ knl = lp.preprocess_kernel(knl, ctx.devices[0])
kernel_gen = lp.generate_loop_schedules(knl)
for knl in kernel_gen:
@@ -388,7 +395,7 @@ def test_argmax(ctx_factory):
n = 10000
- knl = lp.make_kernel(ctx.devices[0],
+ knl = lp.make_kernel(
"{[i]: 0<=i<%d}" % n,
[
"<> result = argmax(i, fabs(a[i]))",
@@ -489,7 +496,7 @@ def test_fuzz_code_generator(ctx_factory):
else:
return np.float64
- knl = lp.make_kernel(ctx.devices[0], "{ : }",
+ knl = lp.make_kernel("{ : }",
[lp.ExpressionInstruction("value", expr)],
[lp.GlobalArg("value", np.complex128, shape=())]
+ [
@@ -523,7 +530,7 @@ def test_empty_reduction(ctx_factory):
ctx = ctx_factory()
queue = cl.CommandQueue(ctx)
- knl = lp.make_kernel(ctx.devices[0],
+ knl = lp.make_kernel(
[
"{[i]: 0<=i<20}",
"[i] -> {[j]: 0<=j<0}"
@@ -546,7 +553,7 @@ def test_nested_dependent_reduction(ctx_factory):
ctx = ctx_factory()
queue = cl.CommandQueue(ctx)
- knl = lp.make_kernel(ctx.devices[0],
+ knl = lp.make_kernel(
[
"{[i]: 0<=i<n}",
"{[j]: 0<=j<i+sumlen}"
@@ -575,7 +582,7 @@ def test_multi_nested_dependent_reduction(ctx_factory):
dtype = np.dtype(np.int32)
ctx = ctx_factory()
- knl = lp.make_kernel(ctx.devices[0],
+ knl = lp.make_kernel(
[
"{[itgt]: 0 <= itgt < ntgts}",
"{[isrc_box]: 0 <= isrc_box < nboxes}",
@@ -603,7 +610,7 @@ def test_recursive_nested_dependent_reduction(ctx_factory):
dtype = np.dtype(np.int32)
ctx = ctx_factory()
- knl = lp.make_kernel(ctx.devices[0],
+ knl = lp.make_kernel(
[
"{[itgt]: 0 <= itgt < ntgts}",
"{[isrc_box]: 0 <= isrc_box < nboxes}",
@@ -632,7 +639,7 @@ def test_dependent_loop_bounds(ctx_factory):
dtype = np.dtype(np.float32)
ctx = ctx_factory()
- knl = lp.make_kernel(ctx.devices[0],
+ knl = lp.make_kernel(
[
"{[i]: 0<=i<n}",
"{[jj]: 0<=jj<row_len}",
@@ -660,7 +667,7 @@ def test_dependent_loop_bounds_2(ctx_factory):
dtype = np.dtype(np.float32)
ctx = ctx_factory()
- knl = lp.make_kernel(ctx.devices[0],
+ knl = lp.make_kernel(
[
"{[i]: 0<=i<n}",
"{[jj]: 0<=jj<row_len}",
@@ -696,7 +703,7 @@ def test_dependent_loop_bounds_3(ctx_factory):
dtype = np.dtype(np.float32)
ctx = ctx_factory()
- knl = lp.make_kernel(ctx.devices[0],
+ knl = lp.make_kernel(
[
"{[i]: 0<=i<n}",
"{[jj]: 0<=jj<row_len}",
@@ -724,6 +731,8 @@ def test_dependent_loop_bounds_3(ctx_factory):
knl_bad = lp.split_iname(knl, "jj", 128, outer_tag="g.1",
inner_tag="l.1")
+ knl = lp.preprocess_kernel(knl, ctx.devices[0])
+
import pytest
with pytest.raises(RuntimeError):
list(lp.generate_loop_schedules(knl_bad))
@@ -734,7 +743,7 @@ def test_independent_multi_domain(ctx_factory):
ctx = ctx_factory()
queue = cl.CommandQueue(ctx)
- knl = lp.make_kernel(ctx.devices[0],
+ knl = lp.make_kernel(
[
"{[i]: 0<=i<n}",
"{[j]: 0<=j<n}",
@@ -770,7 +779,7 @@ def test_bare_data_dependency(ctx_factory):
ctx = ctx_factory()
queue = cl.CommandQueue(ctx)
- knl = lp.make_kernel(ctx.devices[0],
+ knl = lp.make_kernel(
[
"[znirp] -> {[i]: 0<=i<znirp}",
],
@@ -799,7 +808,7 @@ def test_equality_constraints(ctx_factory):
n = 10
- knl = lp.make_kernel(ctx.devices[0], [
+ knl = lp.make_kernel([
"[n] -> {[i,j]: 0<=i,j<n }",
"{[k]: k =i+5 and k < n}",
],
@@ -833,7 +842,7 @@ def test_stride(ctx_factory):
n = 10
- knl = lp.make_kernel(ctx.devices[0], [
+ knl = lp.make_kernel([
"{[i]: 0<=i<n and (exists l: i = 2*l)}",
],
[
@@ -859,7 +868,7 @@ def test_domain_dependency_via_existentially_quantified_variable(ctx_factory):
n = 10
- knl = lp.make_kernel(ctx.devices[0], [
+ knl = lp.make_kernel([
"{[i]: 0<=i<n }",
"{[k]: k=i and (exists l: k = 2*l) }",
],
@@ -886,9 +895,8 @@ def test_double_sum(ctx_factory):
n = 20
- knl = lp.make_kernel(ctx.devices[0], [
+ knl = lp.make_kernel(
"{[i,j]: 0<=i,j<n }",
- ],
[
"a = sum((i,j), i*j)",
"b = sum(i, sum(j, i*j))",
@@ -910,9 +918,8 @@ def test_double_sum(ctx_factory):
def test_ilp_write_race_detection_global(ctx_factory):
ctx = ctx_factory()
- knl = lp.make_kernel(ctx.devices[0], [
+ knl = lp.make_kernel(
"[n] -> {[i,j]: 0<=i,j<n }",
- ],
[
"a[i] = 5+i+j",
],
@@ -924,6 +931,8 @@ def test_ilp_write_race_detection_global(ctx_factory):
knl = lp.tag_inames(knl, dict(j="ilp"))
+ knl = lp.preprocess_kernel(knl, ctx.devices[0])
+
from loopy.diagnostic import WriteRaceConditionWarning
from warnings import catch_warnings
with catch_warnings(record=True) as warn_list:
@@ -936,7 +945,7 @@ def test_ilp_write_race_detection_global(ctx_factory):
def test_ilp_write_race_avoidance_local(ctx_factory):
ctx = ctx_factory()
- knl = lp.make_kernel(ctx.devices[0],
+ knl = lp.make_kernel(
"{[i,j]: 0<=i<16 and 0<=j<17 }",
[
"<> a[i] = 5+i+j",
@@ -945,6 +954,7 @@ def test_ilp_write_race_avoidance_local(ctx_factory):
knl = lp.tag_inames(knl, dict(i="l.0", j="ilp"))
+ knl = lp.preprocess_kernel(knl, ctx.devices[0])
for k in lp.generate_loop_schedules(knl):
assert k.temporary_variables["a"].shape == (16, 17)
@@ -952,7 +962,7 @@ def test_ilp_write_race_avoidance_local(ctx_factory):
def test_ilp_write_race_avoidance_private(ctx_factory):
ctx = ctx_factory()
- knl = lp.make_kernel(ctx.devices[0],
+ knl = lp.make_kernel(
"{[j]: 0<=j<16 }",
[
"<> a = 5+j",
@@ -961,6 +971,7 @@ def test_ilp_write_race_avoidance_private(ctx_factory):
knl = lp.tag_inames(knl, dict(j="ilp"))
+ knl = lp.preprocess_kernel(knl, ctx.devices[0])
for k in lp.generate_loop_schedules(knl):
assert k.temporary_variables["a"].shape == (16,)
@@ -971,9 +982,8 @@ def test_write_parameter(ctx_factory):
dtype = np.float32
ctx = ctx_factory()
- knl = lp.make_kernel(ctx.devices[0], [
+ knl = lp.make_kernel(
"{[i,j]: 0<=i,j<n }",
- ],
"""
a = sum((i,j), i*j)
b = sum(i, sum(j, i*j))
@@ -996,9 +1006,8 @@ def test_write_parameter(ctx_factory):
def test_arg_shape_guessing(ctx_factory):
ctx = ctx_factory()
- knl = lp.make_kernel(ctx.devices[0], [
+ knl = lp.make_kernel(
"{[i,j]: 0<=i,j<n }",
- ],
"""
a = 1.5 + sum((i,j), i*j)
b[i, j] = i*j
@@ -1019,9 +1028,8 @@ def test_arg_shape_guessing(ctx_factory):
def test_arg_guessing(ctx_factory):
ctx = ctx_factory()
- knl = lp.make_kernel(ctx.devices[0], [
+ knl = lp.make_kernel(
"{[i,j]: 0<=i,j<n }",
- ],
"""
a = 1.5 + sum((i,j), i*j)
b[i, j] = i*j
@@ -1037,9 +1045,8 @@ def test_arg_guessing_with_reduction(ctx_factory):
#logging.basicConfig(level=logging.DEBUG)
ctx = ctx_factory()
- knl = lp.make_kernel(ctx.devices[0], [
+ knl = lp.make_kernel(
"{[i,j]: 0<=i,j<n }",
- ],
"""
a = 1.5 + sum((i,j), i*j)
d = 1.5 + sum((i,j), b[i,j])
@@ -1057,9 +1064,8 @@ def test_arg_guessing_with_reduction(ctx_factory):
def test_nonlinear_index(ctx_factory):
ctx = ctx_factory()
- knl = lp.make_kernel(ctx.devices[0], [
+ knl = lp.make_kernel(
"{[i,j]: 0<=i,j<n }",
- ],
"""
a[i*i] = 17
""",
@@ -1076,9 +1082,8 @@ def test_nonlinear_index(ctx_factory):
def test_triangle_domain(ctx_factory):
ctx = ctx_factory()
- knl = lp.make_kernel(ctx.devices[0], [
+ knl = lp.make_kernel(
"{[i,j]: 0<=i,j<n and i <= j}",
- ],
"a[i,j] = 17",
assumptions="n>=1")
@@ -1092,9 +1097,8 @@ def test_offsets_and_slicing(ctx_factory):
n = 20
- knl = lp.make_kernel(ctx.devices[0], [
+ knl = lp.make_kernel(
"{[i,j]: 0<=i<n and 0<=j<m }",
- ],
"""
b[i,j] = 2*a[i,j]
""",
@@ -1128,7 +1132,7 @@ def test_offsets_and_slicing(ctx_factory):
def test_vector_ilp_with_prefetch(ctx_factory):
ctx = ctx_factory()
- knl = lp.make_kernel(ctx.devices[0],
+ knl = lp.make_kernel(
"{ [i]: 0<=i<n }",
"out[i] = 2*a[i]",
[
@@ -1157,7 +1161,7 @@ def test_convolution(ctx_factory):
dtype = np.float32
- knl = lp.make_kernel(ctx.devices[0],
+ knl = lp.make_kernel(
"{ [iimg, ifeat, icolor, im_x, im_y, f_x, f_y]: \
-f_w <= f_x,f_y <= f_w \
and 0 <= im_x < im_w and 0 <= im_y < im_h \
@@ -1224,7 +1228,7 @@ def test_convolution_with_nonzero_base(ctx_factory):
dtype = np.float32
- knl = lp.make_kernel(ctx.devices[0],
+ knl = lp.make_kernel(
"{ [iimg, ifeat, icolor, im_x, im_y, f_x, f_y]: \
-f_w <= f_x,f_y <= f_w \
and f_w <= im_x < im_w-f_w and f_w <= im_y < im_h-f_w \
@@ -1276,9 +1280,8 @@ def test_c_instruction(ctx_factory):
#logging.basicConfig(level=logging.DEBUG)
ctx = ctx_factory()
- knl = lp.make_kernel(ctx.devices[0], [
+ knl = lp.make_kernel(
"{[i,j]: 0<=i,j<n }",
- ],
[
lp.CInstruction("i", """
x = sin((float) i);
@@ -1301,7 +1304,7 @@ def test_c_instruction(ctx_factory):
def test_dependent_domain_insn_iname_finding(ctx_factory):
ctx = ctx_factory()
- knl = lp.make_kernel(ctx.devices[0], [
+ knl = lp.make_kernel([
"{[isrc_box]: 0<=isrc_box<nsrc_boxes}",
"{[isrc,idim]: isrc_start<=isrc<isrc_end and 0<=idim<dim}",
],
@@ -1332,11 +1335,8 @@ def test_dependent_domain_insn_iname_finding(ctx_factory):
def test_inames_deps_from_write_subscript(ctx_factory):
- ctx = ctx_factory()
-
- knl = lp.make_kernel(ctx.devices[0], [
+ knl = lp.make_kernel(
"{[i,j]: 0<=i,j<n}",
- ],
"""
<> src_ibox = source_boxes[i]
<int32> something = 5
@@ -1352,11 +1352,8 @@ def test_inames_deps_from_write_subscript(ctx_factory):
def test_split_reduction(ctx_factory):
- ctx = ctx_factory()
-
- knl = lp.make_kernel(ctx.devices[0], [
+ knl = lp.make_kernel(
"{[i,j,k]: 0<=i,j,k<n}",
- ],
"""
b = sum((i,j,k), a[i,j,k])
""",
@@ -1372,9 +1369,8 @@ def test_split_reduction(ctx_factory):
def test_modulo_indexing(ctx_factory):
ctx = ctx_factory()
- knl = lp.make_kernel(ctx.devices[0], [
+ knl = lp.make_kernel(
"{[i,j]: 0<=i<n and 0<=j<5}",
- ],
"""
b[i] = sum(j, a[(i+j)%n])
""",
@@ -1396,7 +1392,7 @@ def test_rob_stroud_bernstein(ctx_factory):
# NOTE: tmp would have to be zero-filled beforehand
- knl = lp.make_kernel(ctx.devices[0],
+ knl = lp.make_kernel(
"{[el, i2, alpha1,alpha2]: \
0 <= el < nels and \
0 <= i2 < nqp1d and \
@@ -1448,7 +1444,7 @@ def test_rob_stroud_bernstein_full(ctx_factory):
# NOTE: result would have to be zero-filled beforehand
- knl = lp.make_kernel(ctx.devices[0],
+ knl = lp.make_kernel(
"{[el, i2, alpha1,alpha2, i1_2, alpha1_2, i2_2]: \
0 <= el < nels and \
0 <= i2 < nqp1d and \
@@ -1517,7 +1513,7 @@ def test_rob_stroud_bernstein_full(ctx_factory):
def test_vector_types(ctx_factory, vec_len):
ctx = ctx_factory()
- knl = lp.make_kernel(ctx.devices[0],
+ knl = lp.make_kernel(
"{ [i,j]: 0<=i<n and 0<=j<vec_len }",
"out[i,j] = 2*a[i,j]",
[
@@ -1547,7 +1543,6 @@ def test_conditional(ctx_factory):
ctx = ctx_factory()
knl = lp.make_kernel(
- ctx.devices[0],
"{ [i,j]: 0<=i,j<n }",
"""
<> my_a = a[i,j] {id=read_a}
@@ -1577,7 +1572,7 @@ def test_ilp_loop_bound(ctx_factory):
# throughout. In ILP'd loops, not so much.
ctx = ctx_factory()
- knl = lp.make_kernel(ctx.devices[0],
+ knl = lp.make_kernel(
"{ [i,j,k]: 0<=i,j,k<n }",
"""
out[i,k] = sum(j, a[i,j]*b[j,k])
@@ -1604,9 +1599,7 @@ def test_arg_shape_uses_assumptions(ctx_factory):
# static shape for out, which is at least 1 x 1 in size, but otherwise of
# size n x n.
- ctx = ctx_factory()
-
- lp.make_kernel(ctx.devices[0],
+ lp.make_kernel(
"{ [i,j]: 0<=i,j<n }",
"""
out[i,j] = 2*a[i,j]
@@ -1618,7 +1611,7 @@ def test_slab_decomposition_does_not_double_execute(ctx_factory):
ctx = ctx_factory()
queue = cl.CommandQueue(ctx)
- knl = lp.make_kernel(ctx.devices[0],
+ knl = lp.make_kernel(
"{ [i]: 0<=i<n }",
"a[i] = 2*a[i]",
assumptions="n>=1")
@@ -1651,7 +1644,7 @@ def test_multiple_writes_to_local_temporary(ctx_factory):
ctx = ctx_factory()
- knl = lp.make_kernel(ctx.devices[0],
+ knl = lp.make_kernel(
"{[i,e]: 0<=i<5 and 0<=e<nelements}",
"""
<> temp[i, 0] = 17
@@ -1659,8 +1652,10 @@ def test_multiple_writes_to_local_temporary(ctx_factory):
""")
knl = lp.tag_inames(knl, dict(i="l.0"))
- code, _ = lp.generate_code(knl)
- print code
+ knl = lp.preprocess_kernel(knl, ctx.devices[0])
+ for k in lp.generate_loop_schedules(knl):
+ code, _ = lp.generate_code(k)
+ print code
if __name__ == "__main__":
diff --git a/test/test_nbody.py b/test/test_nbody.py
index 7ec973156..65e5658b5 100644
--- a/test/test_nbody.py
+++ b/test/test_nbody.py
@@ -40,17 +40,17 @@ def test_nbody(ctx_factory):
dtype = np.float32
ctx = ctx_factory()
- knl = lp.make_kernel(ctx.devices[0],
+ knl = lp.make_kernel(
"[N] -> {[i,j,k]: 0<=i,j<N and 0<=k<3 }",
- [
- "axdist(k) := x[i,k]-x[j,k]",
- "invdist := rsqrt(sum_float32(k, axdist(k)**2))",
- "pot[i] = sum_float32(j, if(i != j, invdist, 0))",
- ], [
- lp.GlobalArg("x", dtype, shape="N,3", order="C"),
- lp.GlobalArg("pot", dtype, shape="N", order="C"),
- lp.ValueArg("N", np.int32),
- ], name="nbody", assumptions="N>=1")
+ [
+ "axdist(k) := x[i,k]-x[j,k]",
+ "invdist := rsqrt(sum_float32(k, axdist(k)**2))",
+ "pot[i] = sum_float32(j, if(i != j, invdist, 0))",
+ ], [
+ lp.GlobalArg("x", dtype, shape="N,3", order="C"),
+ lp.GlobalArg("pot", dtype, shape="N", order="C"),
+ lp.ValueArg("N", np.int32),
+ ], name="nbody", assumptions="N>=1")
seq_knl = knl
diff --git a/test/test_sem_reagan.py b/test/test_sem_reagan.py
index 2de1db43b..cfa23f35d 100644
--- a/test/test_sem_reagan.py
+++ b/test/test_sem_reagan.py
@@ -44,19 +44,19 @@ def test_tim2d(ctx_factory):
field_shape = (K_sym, n, n)
# K - run-time symbolic
- knl = lp.make_kernel(ctx.devices[0],
+ knl = lp.make_kernel(
"[K] -> {[i,j,e,m,o,gi]: 0<=i,j,m,o<%d and 0<=e<K and 0<=gi<3}" % n,
- [
- "ur(a,b) := sum(o, D[a,o]*u[e,o,b])",
- "us(a,b) := sum(o, D[b,o]*u[e,a,o])",
+ [
+ "ur(a,b) := sum(o, D[a,o]*u[e,o,b])",
+ "us(a,b) := sum(o, D[b,o]*u[e,a,o])",
- #"Gu(mat_entry,a,b) := G[mat_entry,e,m,j]*ur(m,j)",
+ #"Gu(mat_entry,a,b) := G[mat_entry,e,m,j]*ur(m,j)",
- "Gux(a,b) := G$x[0,e,a,b]*ur(a,b)+G$x[1,e,a,b]*us(a,b)",
- "Guy(a,b) := G$y[1,e,a,b]*ur(a,b)+G$y[2,e,a,b]*us(a,b)",
- "lap[e,i,j] = "
- " sum(m, D[m,i]*Gux(m,j))"
- "+ sum(m, D[m,j]*Guy(i,m))"
+ "Gux(a,b) := G$x[0,e,a,b]*ur(a,b)+G$x[1,e,a,b]*us(a,b)",
+ "Guy(a,b) := G$y[1,e,a,b]*ur(a,b)+G$y[2,e,a,b]*us(a,b)",
+ "lap[e,i,j] = "
+ " sum(m, D[m,i]*Gux(m,j))"
+ "+ sum(m, D[m,j]*Guy(i,m))"
],
[
--
GitLab