From 3c173a0adf64653429ad66ff6adb84b8af041323 Mon Sep 17 00:00:00 2001
From: Andreas Kloeckner <inform@tiker.net>
Date: Tue, 11 Jun 2013 00:36:05 -0400
Subject: [PATCH] Add set_loop_priority, remove check_kernels, auto-schedule in
auto_test
---
doc/reference.rst | 4 +-
loopy/__init__.py | 11 +++-
loopy/auto_test.py | 30 ++++++----
loopy/check.py | 124 ++++++++++++++++-----------------------
loopy/compiled.py | 14 -----
loopy/kernel/__init__.py | 73 ++++++++++++++---------
loopy/schedule.py | 16 +++--
test/test_dg.py | 9 +--
test/test_linalg.py | 62 +++++---------------
test/test_loopy.py | 48 ++++-----------
test/test_nbody.py | 44 ++++++--------
test/test_sem_reagan.py | 32 ++++------
12 files changed, 193 insertions(+), 274 deletions(-)
diff --git a/doc/reference.rst b/doc/reference.rst
index 975ef80a9..c8d041079 100644
--- a/doc/reference.rst
+++ b/doc/reference.rst
@@ -185,6 +185,8 @@ Wrangling inames
.. autofunction:: remove_unused_inames
+.. autofunction:: set_loop_priority
+
Dealing with Substitution Rules
^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
@@ -238,8 +240,6 @@ Finishing up
.. autofunction:: generate_loop_schedules
-.. autofunction:: check_kernels
-
.. autofunction:: generate_code
Automatic Testing
diff --git a/loopy/__init__.py b/loopy/__init__.py
index 0d868cb47..7851f8a32 100644
--- a/loopy/__init__.py
+++ b/loopy/__init__.py
@@ -70,7 +70,6 @@ from loopy.schedule import generate_loop_schedules
from loopy.codegen import generate_code
from loopy.compiled import CompiledKernel
from loopy.auto_test import auto_test_vs_ref
-from loopy.check import check_kernels
__all__ = [
"auto",
@@ -89,7 +88,7 @@ __all__ = [
"preprocess_kernel", "realize_reduction", "infer_unknown_types",
"generate_loop_schedules",
"generate_code",
- "CompiledKernel", "auto_test_vs_ref", "check_kernels",
+ "CompiledKernel", "auto_test_vs_ref",
"make_kernel",
"split_iname", "join_inames", "tag_inames", "duplicate_inames",
@@ -757,6 +756,14 @@ def remove_unused_inames(knl, inames=None):
# }}}
+# {{{ set loop priority
+
+def set_loop_priority(kernel, loop_priority):
+ return kernel.copy(loop_priority=loop_priority)
+
+# }}}
+
+
# {{{ convenience: add_prefetch
# {{{ process footprint_subscripts
diff --git a/loopy/auto_test.py b/loopy/auto_test.py
index ea172db2f..c28583695 100644
--- a/loopy/auto_test.py
+++ b/loopy/auto_test.py
@@ -325,14 +325,14 @@ def _enumerate_cl_devices_for_ref_test():
# {{{ main automatic testing entrypoint
def auto_test_vs_ref(
- ref_knl, ctx, kernel_gen, op_count=[], op_label=[], parameters={},
+ ref_knl, ctx, test_knl, op_count=[], op_label=[], parameters={},
print_ref_code=False, print_code=True, warmup_rounds=2,
code_op=None, dump_binary=False, codegen_kwargs={},
options=[],
fills_entire_output=True, do_check=True, check_result=None
):
- """Compare results of `ref_knl` to the kernels generated by the generator
- `kernel_gen`.
+ """Compare results of `ref_knl` to the kernels generated by
+ scheduling *test_knl*.
:arg check_result: a callable with :class:`numpy.ndarray` arguments
*(result, reference_result)* returning a a tuple (class:`bool`,
@@ -385,8 +385,7 @@ def auto_test_vs_ref(
properties=cl.command_queue_properties.PROFILING_ENABLE)
import loopy as lp
- ref_kernel_gen = lp.generate_loop_schedules(ref_knl)
- for knl in lp.check_kernels(ref_kernel_gen, parameters):
+ for knl in lp.generate_loop_schedules(ref_knl):
ref_sched_kernel = knl
break
@@ -460,7 +459,22 @@ def auto_test_vs_ref(
properties=cl.command_queue_properties.PROFILING_ENABLE)
args = None
- for i, kernel in enumerate(kernel_gen):
+ from loopy.kernel import LoopKernel
+ if not isinstance(test_knl, LoopKernel):
+ from warnings import warn
+ warn("Passing an iterable of kernels to auto_test_vs_ref "
+ "is deprecated--just pass the kernel instead. "
+ "Scheduling will be performed in auto_test_vs_ref.",
+ DeprecationWarning, stacklevel=2)
+
+ test_kernels = test_knl
+ else:
+ if not test_knl.schedule:
+ test_kernels = lp.generate_loop_schedules(test_knl)
+ else:
+ test_kernels = [test_knl]
+
+ for i, kernel in enumerate(test_kernels):
compiled = CompiledKernel(ctx, kernel, options=options,
codegen_kwargs=codegen_kwargs)
@@ -575,8 +589,4 @@ def auto_test_vs_ref(
# }}}
-from pytools import MovedFunctionDeprecationWrapper
-
-auto_test_vs_seq = MovedFunctionDeprecationWrapper(auto_test_vs_ref)
-
# vim: foldmethod=marker
diff --git a/loopy/check.py b/loopy/check.py
index e2de6405e..ae1d2a925 100644
--- a/loopy/check.py
+++ b/loopy/check.py
@@ -33,6 +33,54 @@ logger = logging.getLogger(__name__)
# {{{ sanity checks run during scheduling
+def check_sizes(kernel):
+ import loopy as lp
+
+ from loopy import LoopyAdvisory
+
+ parameters = {}
+ for arg in kernel.args:
+ if isinstance(arg, lp.ValueArg) and arg.approximately is not None:
+ parameters[arg.name] = arg.approximately
+
+ glens, llens = kernel.get_grid_sizes_as_exprs()
+
+ if (max(len(glens), len(llens))
+ > kernel.device.max_work_item_dimensions):
+ raise RuntimeError("too many work item dimensions")
+
+ from pymbolic import evaluate
+ from pymbolic.mapper.evaluator import UnknownVariableError
+ try:
+ glens = evaluate(glens, parameters)
+ llens = evaluate(llens, parameters)
+ except UnknownVariableError, name:
+ from warnings import warn
+ warn("could not check axis bounds because no value "
+ "for variable '%s' was passed to check_kernels()"
+ % name, LoopyAdvisory)
+ else:
+ for i in range(len(llens)):
+ if llens[i] > kernel.device.max_work_item_sizes[i]:
+ raise RuntimeError("group axis %d too big" % i)
+
+ from pytools import product
+ if product(llens) > kernel.device.max_work_group_size:
+ raise RuntimeError("work group too big")
+
+ from pyopencl.characterize import usable_local_mem_size
+ if kernel.local_mem_use() > usable_local_mem_size(kernel.device):
+ raise RuntimeError(5, "using too much local memory")
+
+ from loopy.kernel.data import ConstantArg
+ const_arg_count = sum(
+ 1 for arg in kernel.args
+ if isinstance(arg, ConstantArg))
+
+ if const_arg_count > kernel.device.max_constant_args:
+ raise RuntimeError("too many constant arguments")
+
+
def check_for_unused_hw_axes_in_insns(kernel):
group_size, local_size = kernel.get_grid_sizes_as_exprs()
@@ -319,6 +367,7 @@ def run_automatic_checks(kernel):
try:
logger.info("sanity-check %s: start" % kernel.name)
+ check_sizes(kernel)
check_for_orphaned_user_hardware_axes(kernel)
check_for_double_use_of_hw_axes(kernel)
check_for_unused_hw_axes_in_insns(kernel)
@@ -419,79 +468,4 @@ def check_implemented_domains(kernel, implemented_domains, code=None):
# }}}
-# {{{ user-invoked checks
-
-def get_problems(kernel, parameters):
- """
- :return: *(max_severity, list of (severity, msg))*, where *severity*
- ranges from 1-5. '5' means 'will certainly not run'.
- """
- msgs = []
-
- def msg(severity, s):
- msgs.append((severity, s))
-
- glens, llens = kernel.get_grid_sizes_as_exprs()
-
- if (max(len(glens), len(llens))
- > kernel.device.max_work_item_dimensions):
- msg(5, "too many work item dimensions")
-
- from pymbolic import evaluate
- from pymbolic.mapper.evaluator import UnknownVariableError
- try:
- glens = evaluate(glens, parameters)
- llens = evaluate(llens, parameters)
- except UnknownVariableError, name:
- msg(1, "could not check axis bounds because no value "
- "for variable '%s' was passed to check_kernels()"
- % name)
- else:
- for i in range(len(llens)):
- if llens[i] > kernel.device.max_work_item_sizes[i]:
- msg(5, "group axis %d too big" % i)
-
- from pytools import product
- if product(llens) > kernel.device.max_work_group_size:
- msg(5, "work group too big")
-
- import pyopencl as cl
- from pyopencl.characterize import usable_local_mem_size
- if kernel.local_mem_use() > usable_local_mem_size(kernel.device):
- if kernel.device.local_mem_type == cl.device_local_mem_type.LOCAL:
- msg(5, "using too much local memory")
- else:
- msg(4, "using more local memory than available--"
- "possibly OK due to cache nature")
-
- from loopy.kernel.data import ConstantArg
- const_arg_count = sum(
- 1 for arg in kernel.args
- if isinstance(arg, ConstantArg))
-
- if const_arg_count > kernel.device.max_constant_args:
- msg(5, "too many constant arguments")
-
- max_severity = 0
- for sev, msg in msgs:
- max_severity = max(sev, max_severity)
- return max_severity, msgs
-
-
-def check_kernels(kernel_gen, parameters={}, kill_level_min=5,
- warn_level_min=1):
- for kernel in kernel_gen:
- max_severity, msgs = get_problems(kernel, parameters)
-
- for severity, msg in msgs:
- if severity >= warn_level_min:
- from warnings import warn
- from loopy import LoopyAdvisory
- warn(msg, LoopyAdvisory)
-
- if max_severity < kill_level_min:
- yield kernel
-
-# }}}
-
# vim: foldmethod=marker
diff --git a/loopy/compiled.py b/loopy/compiled.py
index ee8702335..e2746b381 100644
--- a/loopy/compiled.py
+++ b/loopy/compiled.py
@@ -266,20 +266,6 @@ class CompiledKernel:
specific arguments.
"""
- import loopy as lp
-
- # {{{ do scheduling, if not yet done
-
- if not isinstance(kernel, lp.LoopKernel):
- # someone threw us an iterable of kernels
-
- kernel = _get_kernel_from_iterable(kernel)
-
- # Whether we need to call check_kernels. Since we don't have parameter
- # values now, we'll do that on first invocation.
-
- # }}}
-
self.context = context
self.kernel = kernel
self.codegen_kwargs = codegen_kwargs
diff --git a/loopy/kernel/__init__.py b/loopy/kernel/__init__.py
index 46b79186c..66bb1f56d 100644
--- a/loopy/kernel/__init__.py
+++ b/loopy/kernel/__init__.py
@@ -87,33 +87,55 @@ class LoopKernel(Record):
"""These correspond more or less directly to arguments of
:func:`loopy.make_kernel`.
- :ivar device: :class:`pyopencl.Device`
- :ivar domains: a list of :class:`islpy.BasicSet` instances
- :ivar instructions:
- :ivar args:
- :ivar schedule:
- :ivar name:
- :ivar preambles:
- :ivar preamble_generators:
- :ivar assumptions:
- :ivar local_sizes:
- :ivar temporary_variables:
- :ivar iname_to_tag:
- :ivar function_manglers:
- :ivar symbol_manglers:
-
- The following arguments are not user-facing:
-
- :ivar substitutions: a mapping from substitution names to
+ .. attribute:: device
+
+ :class:`pyopencl.Device`
+
+ .. attribute:: domains
+
+ a list of :class:`islpy.BasicSet` instances
+
+ .. attribute:: instructions
+ .. attribute:: args
+ .. attribute:: schedule
+
+ *None* or a list of :class:`loopy.schedule.ScheduleItem`
+
+ .. attribute:: name
+ .. attribute:: preambles
+ .. attribute:: preamble_generators
+ .. attribute:: assumptions
+ .. attribute:: local_sizes
+ .. attribute:: temporary_variables
+ .. attribute:: iname_to_tag
+ .. attribute:: function_manglers
+ .. attribute:: symbol_manglers
+
+ .. attribute:: substitutions
+
+ a mapping from substitution names to
:class:`SubstitutionRule` objects
- :ivar iname_slab_increments: a dictionary mapping inames to (lower_incr,
+
+ .. attribute:: iname_slab_increments
+
+ a dictionary mapping inames to (lower_incr,
upper_incr) tuples that will be separated out in the execution to generate
'bulk' slabs with fewer conditionals.
- :ivar applied_iname_rewrites: A list of past substitution dictionaries that
+
+ .. attribute:: loop_priority
+
+ A list of inames. The earlier in the list the iname occurs, the earlier
+ it will be scheduled. (This applies to inames with non-parallel
+ implementation tags.)
+
+ .. attribute:: applied_iname_rewrites
+
+ A list of past substitution dictionaries that
were applied to the kernel. These are stored so that they may be repeated
on expressions the user specifies later.
- :ivar cache_manager:
- :ivar isl_context:
+
+ .. attribute:: cache_manager
+ .. attribute:: isl_context
"""
# {{{ constructor
@@ -134,8 +156,8 @@ class LoopKernel(Record):
],
symbol_manglers=[opencl_symbol_mangler],
- # non-user-facing
iname_slab_increments={},
+ loop_priority=[],
applied_iname_rewrites=[],
cache_manager=None,
index_dtype=np.int32,
@@ -145,10 +167,6 @@ class LoopKernel(Record):
# their grid sizes shouldn't change. This provides
# a way to forward sub-kernel grid size requests.
get_grid_sizes=None):
- """
- :arg domain: a :class:`islpy.BasicSet`, or a string parseable to
- a basic set by the isl. Example: "{[i,j]: 0<=i < 10 and 0<= j < 9}"
- """
if cache_manager is None:
from loopy.kernel.tools import SetOperationCacheManager
@@ -235,6 +253,7 @@ class LoopKernel(Record):
preamble_generators=preamble_generators,
assumptions=assumptions,
iname_slab_increments=iname_slab_increments,
+ loop_priority=loop_priority,
temporary_variables=temporary_variables,
local_sizes=local_sizes,
iname_to_tag=iname_to_tag,
diff --git a/loopy/schedule.py b/loopy/schedule.py
index f1e06a37e..4d5c0314a 100644
--- a/loopy/schedule.py
+++ b/loopy/schedule.py
@@ -33,19 +33,23 @@ logger = logging.getLogger(__name__)
# {{{ schedule items
-class EnterLoop(Record):
+class ScheduleItem(Record):
+ __slots__ = []
+
+
+class EnterLoop(ScheduleItem):
__slots__ = ["iname"]
-class LeaveLoop(Record):
+class LeaveLoop(ScheduleItem):
__slots__ = ["iname"]
-class RunInstruction(Record):
+class RunInstruction(ScheduleItem):
__slots__ = ["insn_id"]
-class Barrier(Record):
+class Barrier(ScheduleItem):
__slots__ = ["comment"]
# }}}
@@ -785,7 +789,9 @@ def insert_barriers(kernel, schedule, level=0):
# {{{ main scheduling entrypoint
-def generate_loop_schedules(kernel, loop_priority=[], debug_args={}):
+def generate_loop_schedules(kernel, debug_args={}):
+ loop_priority = kernel.loop_priority
+
from loopy.preprocess import preprocess_kernel
kernel = preprocess_kernel(kernel)
diff --git a/test/test_dg.py b/test/test_dg.py
index 1b54a8328..8de1c8eec 100644
--- a/test/test_dg.py
+++ b/test/test_dg.py
@@ -151,11 +151,8 @@ def test_dg_volume(ctx_factory):
variant_simple_padding,
variant_fancy_padding
]:
- kernel_gen = lp.generate_loop_schedules(variant(knl))
- kernel_gen = lp.check_kernels(kernel_gen, parameters_dict)
-
lp.auto_test_vs_ref(
- seq_knl, ctx, kernel_gen, parameters=parameters_dict,
+ seq_knl, ctx, variant(knl), parameters=parameters_dict,
#codegen_kwargs=dict(with_annotation=True)
)
@@ -225,10 +222,8 @@ def no_test_dg_surface(ctx_factory):
for variant in [
variant_basic,
]:
- kernel_gen = lp.generate_loop_schedules(variant(knl))
- kernel_gen = lp.check_kernels(kernel_gen, parameters_dict)
- lp.auto_test_vs_ref(seq_knl, ctx, kernel_gen, parameters=parameters_dict)
+ lp.auto_test_vs_ref(seq_knl, ctx, variant(knl), parameters=parameters_dict)
if __name__ == "__main__":
diff --git a/test/test_linalg.py b/test/test_linalg.py
index 9bf916221..c9605e5ee 100644
--- a/test/test_linalg.py
+++ b/test/test_linalg.py
@@ -106,10 +106,7 @@ def test_axpy(ctx_factory):
for variant in [variant_cpu, variant_gpu]:
#for variant in [ variant_gpu]:
- kernel_gen = lp.generate_loop_schedules(variant(knl))
- kernel_gen = lp.check_kernels(kernel_gen, dict(n=n))
-
- lp.auto_test_vs_ref(seq_knl, ctx, kernel_gen,
+ lp.auto_test_vs_ref(seq_knl, ctx, variant(knl),
op_count=[np.dtype(dtype).itemsize*n*3/1e9],
op_label=["GBytes"],
parameters={"a": a, "b": b, "n": n}, check_result=check)
@@ -141,10 +138,7 @@ def test_transpose(ctx_factory):
outer_tag="g.1", inner_tag="l.0")
knl = lp.add_prefetch(knl, 'a', ["i_inner", "j_inner"])
- kernel_gen = lp.generate_loop_schedules(knl)
- kernel_gen = lp.check_kernels(kernel_gen, {})
-
- lp.auto_test_vs_ref(seq_knl, ctx, kernel_gen,
+ lp.auto_test_vs_ref(seq_knl, ctx, knl,
op_count=[dtype.itemsize*n**2*2/1e9], op_label=["GByte"],
parameters={})
@@ -181,10 +175,7 @@ def test_plain_matrix_mul(ctx_factory):
knl = lp.add_prefetch(knl, "a", ["k_inner", "i_inner"])
knl = lp.add_prefetch(knl, "b", ["j_inner", "k_inner", ])
- kernel_gen = lp.generate_loop_schedules(knl)
- kernel_gen = lp.check_kernels(kernel_gen, {})
-
- lp.auto_test_vs_ref(ref_knl, ctx, kernel_gen,
+ lp.auto_test_vs_ref(ref_knl, ctx, knl,
op_count=[vec_size*2*n**3/1e9], op_label=["GFlops"],
parameters={"n": n}, check_result=check)
@@ -220,10 +211,7 @@ def test_variable_size_matrix_mul(ctx_factory):
knl = lp.add_prefetch(knl, "a", ["k_inner", "i_inner"])
knl = lp.add_prefetch(knl, "b", ["j_inner", "k_inner"])
- kernel_gen = lp.generate_loop_schedules(knl)
- kernel_gen = lp.check_kernels(kernel_gen, dict(n=n))
-
- lp.auto_test_vs_ref(ref_knl, ctx, kernel_gen,
+ lp.auto_test_vs_ref(ref_knl, ctx, knl,
op_count=[2*n**3/1e9], op_label=["GFlops"],
parameters={"n": n})
@@ -253,6 +241,7 @@ def test_rank_one(ctx_factory):
def variant_1(knl):
knl = lp.add_prefetch(knl, "a")
knl = lp.add_prefetch(knl, "b")
+ knl = knl.set_loop_priority(knl, ["i", "j"])
return knl
def variant_2(knl):
@@ -261,6 +250,7 @@ def test_rank_one(ctx_factory):
knl = lp.split_iname(knl, "j", 16,
outer_tag="g.1", inner_tag="l.1")
+ knl = knl.set_loop_priority(knl, ["i", "j"])
knl = lp.add_prefetch(knl, "a")
knl = lp.add_prefetch(knl, "b")
return knl
@@ -299,11 +289,7 @@ def test_rank_one(ctx_factory):
#for variant in [variant_1, variant_2, variant_3, variant_4]:
for variant in [variant_4]:
- kernel_gen = lp.generate_loop_schedules(variant(knl),
- loop_priority=["i", "j"])
- kernel_gen = lp.check_kernels(kernel_gen, dict(n=n))
-
- lp.auto_test_vs_ref(seq_knl, ctx, kernel_gen,
+ lp.auto_test_vs_ref(seq_knl, ctx, variant(knl),
op_count=[np.dtype(dtype).itemsize*n**2/1e9], op_label=["GBytes"],
parameters={"n": n})
@@ -340,10 +326,7 @@ def test_troublesome_premagma_fermi_matrix_mul(ctx_factory):
knl = lp.split_iname(knl, "k", 16)
knl = lp.add_prefetch(knl, 'a', ["k_inner", "i_inner_inner", "i_inner_outer"])
- kernel_gen = lp.generate_loop_schedules(knl)
- kernel_gen = lp.check_kernels(kernel_gen, dict(n=n))
-
- lp.auto_test_vs_ref(seq_knl, ctx, kernel_gen,
+ lp.auto_test_vs_ref(seq_knl, ctx, knl,
op_count=[2*n**3/1e9], op_label=["GFlops"],
parameters={})
@@ -388,11 +371,9 @@ def test_intel_matrix_mul(ctx_factory):
#knl = lp.add_prefetch(knl, 'b',
# ["k_inner", ("j_inner_inner", "j_inner_outer"),])
- kernel_gen = lp.generate_loop_schedules(knl)
#hints=["k_outer", "k_inner_outer", "k_inner_inner"]
- kernel_gen = lp.check_kernels(kernel_gen, dict(n=n))
- lp.auto_test_vs_ref(seq_knl, ctx, kernel_gen,
+ lp.auto_test_vs_ref(seq_knl, ctx, knl,
op_count=[2*n**3/1e9], op_label=["GFlops"],
parameters={})
@@ -434,10 +415,7 @@ def test_magma_fermi_matrix_mul(ctx_factory):
#knl = lp.add_prefetch(knl, 'b',
# ["k_inner", ("j_inner_inner", "j_inner_outer"),])
- kernel_gen = lp.generate_loop_schedules(knl)
- kernel_gen = lp.check_kernels(kernel_gen, dict(n=n))
-
- lp.auto_test_vs_ref(seq_knl, ctx, kernel_gen,
+ lp.auto_test_vs_ref(seq_knl, ctx, knl,
op_count=[2*n**3/1e9], op_label=["GFlops"],
parameters={})
@@ -470,10 +448,7 @@ def test_image_matrix_mul(ctx_factory):
knl = lp.add_prefetch(knl, 'a', ["i_inner", "k_inner"])
knl = lp.add_prefetch(knl, 'b', ["j_inner", "k_inner"])
- kernel_gen = lp.generate_loop_schedules(knl)
- kernel_gen = lp.check_kernels(kernel_gen, dict(n=n))
-
- lp.auto_test_vs_ref(seq_knl, ctx, kernel_gen,
+ lp.auto_test_vs_ref(seq_knl, ctx, knl,
op_count=[2*n**3/1e9], op_label=["GFlops"],
parameters={}, print_ref_code=True)
@@ -510,10 +485,7 @@ def test_image_matrix_mul_ilp(ctx_factory):
knl = lp.add_prefetch(knl, 'a', ["i_inner", "k_inner"])
knl = lp.add_prefetch(knl, 'b', ["j_inner_outer", "j_inner_inner", "k_inner"])
- kernel_gen = lp.generate_loop_schedules(knl)
- kernel_gen = lp.check_kernels(kernel_gen, dict(n=n))
-
- lp.auto_test_vs_ref(seq_knl, ctx, kernel_gen,
+ lp.auto_test_vs_ref(seq_knl, ctx, knl,
op_count=[2*n**3/1e9], op_label=["GFlops"],
parameters={})
@@ -575,10 +547,7 @@ def test_fancy_matrix_mul(ctx_factory):
knl = lp.add_prefetch(knl, 'a', ["i_inner", "k_inner"])
knl = lp.add_prefetch(knl, 'b', ["k_inner", "j_inner"])
- kernel_gen = lp.generate_loop_schedules(knl)
- kernel_gen = lp.check_kernels(kernel_gen, dict(n=n))
-
- lp.auto_test_vs_ref(seq_knl, ctx, kernel_gen,
+ lp.auto_test_vs_ref(seq_knl, ctx, knl,
op_count=[2*n**3/1e9], op_label=["GFlops"],
parameters=dict(n=n))
@@ -612,10 +581,7 @@ def test_small_batched_matvec(ctx_factory):
knl = lp.split_arg_axis(knl, ("f", 0), pad_mult)
knl = lp.add_padding(knl, "f", 0, align_bytes)
- kernel_gen = lp.generate_loop_schedules(knl)
- kernel_gen = lp.check_kernels(kernel_gen, dict(K=K))
-
- lp.auto_test_vs_ref(seq_knl, ctx, kernel_gen,
+ lp.auto_test_vs_ref(seq_knl, ctx, knl,
op_count=[K*2*Np**2/1e9], op_label=["GFlops"],
parameters=dict(K=K))
diff --git a/test/test_loopy.py b/test/test_loopy.py
index 5875912ba..ba695a5df 100644
--- a/test/test_loopy.py
+++ b/test/test_loopy.py
@@ -114,8 +114,7 @@ def test_sized_and_complex_literals(ctx_factory):
],
assumptions="n>=1")
- lp.auto_test_vs_ref(knl, ctx, lp.generate_loop_schedules(knl),
- parameters=dict(n=5))
+ lp.auto_test_vs_ref(knl, ctx, knl, parameters=dict(n=5))
def test_simple_side_effect(ctx_factory):
@@ -130,7 +129,6 @@ def test_simple_side_effect(ctx_factory):
)
kernel_gen = lp.generate_loop_schedules(knl)
- kernel_gen = lp.check_kernels(kernel_gen)
for gen_knl in kernel_gen:
print gen_knl
@@ -168,7 +166,6 @@ def test_owed_barriers(ctx_factory):
knl = lp.tag_inames(knl, dict(i="l.0"))
kernel_gen = lp.generate_loop_schedules(knl)
- kernel_gen = lp.check_kernels(kernel_gen)
for gen_knl in kernel_gen:
compiled = lp.CompiledKernel(ctx, gen_knl)
@@ -189,7 +186,6 @@ def test_wg_too_small(ctx_factory):
knl = lp.tag_inames(knl, dict(i="l.0"))
kernel_gen = lp.generate_loop_schedules(knl)
- kernel_gen = lp.check_kernels(kernel_gen)
import pytest
for gen_knl in kernel_gen:
@@ -216,10 +212,7 @@ def test_join_inames(ctx_factory):
knl = lp.add_prefetch(knl, "a", sweep_inames=["i", "j"])
knl = lp.join_inames(knl, ["a_dim_0", "a_dim_1"])
- kernel_gen = lp.generate_loop_schedules(knl)
- kernel_gen = lp.check_kernels(kernel_gen)
-
- lp.auto_test_vs_ref(ref_knl, ctx, kernel_gen)
+ lp.auto_test_vs_ref(ref_knl, ctx, knl)
def test_divisibility_assumption(ctx_factory):
@@ -245,10 +238,7 @@ def test_divisibility_assumption(ctx_factory):
code = lp.generate_code(k)
assert "if" not in code
- kernel_gen = lp.generate_loop_schedules(knl)
- kernel_gen = lp.check_kernels(kernel_gen)
-
- lp.auto_test_vs_ref(ref_knl, ctx, kernel_gen,
+ lp.auto_test_vs_ref(ref_knl, ctx, knl,
parameters={"n": 16**3})
@@ -267,7 +257,6 @@ def test_multi_cse(ctx_factory):
knl = lp.add_prefetch(knl, "a", [])
kernel_gen = lp.generate_loop_schedules(knl)
- kernel_gen = lp.check_kernels(kernel_gen)
for gen_knl in kernel_gen:
compiled = lp.CompiledKernel(ctx, gen_knl)
@@ -302,6 +291,7 @@ def test_stencil(ctx_factory):
knl = lp.split_iname(knl, "i", 16, outer_tag="g.1", inner_tag="l.1")
knl = lp.split_iname(knl, "j", 16, outer_tag="g.0", inner_tag="l.0")
knl = lp.add_prefetch(knl, "a", ["i_inner", "j_inner"])
+ knl = lp.set_loop_priority(knl, ["i_outer", "i_inner_0", "j_0"])
return knl
def variant_2(knl):
@@ -309,14 +299,11 @@ def test_stencil(ctx_factory):
knl = lp.split_iname(knl, "j", 16, outer_tag="g.0", inner_tag="l.0")
knl = lp.add_prefetch(knl, "a", ["i_inner", "j_inner"],
fetch_bounding_box=True)
+ knl = lp.set_loop_priority(knl, ["i_outer", "i_inner_0", "j_0"])
return knl
for variant in [variant_1, variant_2]:
- kernel_gen = lp.generate_loop_schedules(variant(knl),
- loop_priority=["i_outer", "i_inner_0", "j_0"])
- kernel_gen = lp.check_kernels(kernel_gen)
-
- lp.auto_test_vs_ref(ref_knl, ctx, kernel_gen,
+ lp.auto_test_vs_ref(ref_knl, ctx, variant(knl),
fills_entire_output=False, print_ref_code=False,
op_count=[n*n], op_label=["cells"])
@@ -352,15 +339,12 @@ def test_stencil_with_overfetch(ctx_factory):
slabs=(1, 1))
knl = lp.add_prefetch(knl, "a", ["i_inner", "j_inner"],
fetch_bounding_box=True)
+ knl = lp.set_loop_priority(knl, ["i_outer", "i_inner_0", "j_0"])
return knl
for variant in [variant_overfetch]:
- kernel_gen = lp.generate_loop_schedules(variant(knl),
- loop_priority=["i_outer", "i_inner_0", "j_0"])
- kernel_gen = lp.check_kernels(kernel_gen)
-
n = 200
- lp.auto_test_vs_ref(ref_knl, ctx, kernel_gen,
+ lp.auto_test_vs_ref(ref_knl, ctx, variant(knl),
fills_entire_output=False, print_ref_code=False,
op_count=[n*n], parameters=dict(n=n), op_label=["cells"])
@@ -382,7 +366,6 @@ def test_eq_constraint(ctx_factory):
knl = lp.split_iname(knl, "i_inner", 16, outer_tag=None, inner_tag="l.0")
kernel_gen = lp.generate_loop_schedules(knl)
- kernel_gen = lp.check_kernels(kernel_gen)
for knl in kernel_gen:
print lp.generate_code(knl)
@@ -828,10 +811,7 @@ def test_equality_constraints(ctx_factory):
#print knl
#print knl.domains[0].detect_equalities()
- kernel_gen = lp.generate_loop_schedules(knl)
- kernel_gen = lp.check_kernels(kernel_gen, dict(n=n))
-
- lp.auto_test_vs_ref(seq_knl, ctx, kernel_gen,
+ lp.auto_test_vs_ref(seq_knl, ctx, knl,
parameters=dict(n=n), print_ref_code=True)
@@ -857,10 +837,7 @@ def test_stride(ctx_factory):
seq_knl = knl
- kernel_gen = lp.generate_loop_schedules(knl)
- kernel_gen = lp.check_kernels(kernel_gen, dict(n=n))
-
- lp.auto_test_vs_ref(seq_knl, ctx, kernel_gen,
+ lp.auto_test_vs_ref(seq_knl, ctx, knl,
parameters=dict(n=n), fills_entire_output=False)
@@ -888,10 +865,7 @@ def test_domain_dependency_via_existentially_quantified_variable(ctx_factory):
seq_knl = knl
- kernel_gen = lp.generate_loop_schedules(knl)
- kernel_gen = lp.check_kernels(kernel_gen, dict(n=n))
-
- lp.auto_test_vs_ref(seq_knl, ctx, kernel_gen,
+ lp.auto_test_vs_ref(seq_knl, ctx, knl,
parameters=dict(n=n), )
diff --git a/test/test_nbody.py b/test/test_nbody.py
index 71830b758..2d80c1383 100644
--- a/test/test_nbody.py
+++ b/test/test_nbody.py
@@ -23,15 +23,11 @@ THE SOFTWARE.
"""
-
-
import numpy as np
-import pyopencl as cl
import loopy as lp
-from pyopencl.tools import pytest_generate_tests_for_pyopencl \
- as pytest_generate_tests
-
+from pyopencl.tools import ( # noqa
+ pytest_generate_tests_for_pyopencl as pytest_generate_tests)
def test_nbody(ctx_factory):
@@ -46,54 +42,50 @@ def test_nbody(ctx_factory):
"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")
+ 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
def variant_1(knl):
knl = lp.split_iname(knl, "i", 256,
outer_tag="g.0", inner_tag="l.0",
- slabs=(0,1))
- knl = lp.split_iname(knl, "j", 256, slabs=(0,1))
- return knl, []
+ slabs=(0, 1))
+ knl = lp.split_iname(knl, "j", 256, slabs=(0, 1))
+ return knl
def variant_cpu(knl):
knl = lp.expand_subst(knl)
knl = lp.split_iname(knl, "i", 1024,
- outer_tag="g.0", slabs=(0,1))
+ outer_tag="g.0", slabs=(0, 1))
knl = lp.add_prefetch(knl, "x[i,k]", ["k"], default_tag=None)
- return knl, []
+ return knl
def variant_gpu(knl):
knl = lp.expand_subst(knl)
knl = lp.split_iname(knl, "i", 256,
- outer_tag="g.0", inner_tag="l.0", slabs=(0,1))
- knl = lp.split_iname(knl, "j", 256, slabs=(0,1))
+ outer_tag="g.0", inner_tag="l.0", slabs=(0, 1))
+ knl = lp.split_iname(knl, "j", 256, slabs=(0, 1))
knl = lp.add_prefetch(knl, "x[j,k]", ["j_inner", "k"],
["x_fetch_j", "x_fetch_k"])
knl = lp.add_prefetch(knl, "x[i,k]", ["k"], default_tag=None)
knl = lp.tag_inames(knl, dict(x_fetch_k="unr"))
- return knl, ["j_outer", "j_inner"]
+ knl = lp.set_loop_priority(knl, ["j_outer", "j_inner"])
+ return knl
n = 3000
for variant in [variant_1, variant_cpu, variant_gpu]:
- variant_knl, loop_prio = variant(knl)
- kernel_gen = lp.generate_loop_schedules(variant_knl,
- loop_priority=loop_prio)
- kernel_gen = lp.check_kernels(kernel_gen, dict(N=n))
+ variant_knl = variant(knl)
- lp.auto_test_vs_ref(seq_knl, ctx, kernel_gen,
+ lp.auto_test_vs_ref(seq_knl, ctx, variant_knl,
op_count=[n**2*1e-6], op_label=["M particle pairs"],
parameters={"N": n})
-
-
if __name__ == "__main__":
import sys
if len(sys.argv) > 1:
diff --git a/test/test_sem_reagan.py b/test/test_sem_reagan.py
index 8ff5bf8ac..819de845d 100644
--- a/test/test_sem_reagan.py
+++ b/test/test_sem_reagan.py
@@ -23,16 +23,11 @@ THE SOFTWARE.
"""
-
-
import numpy as np
-import pyopencl as cl
import loopy as lp
-from pyopencl.tools import pytest_generate_tests_for_pyopencl \
- as pytest_generate_tests
-
-
+from pyopencl.tools import ( # noqa
+ pytest_generate_tests_for_pyopencl as pytest_generate_tests)
def test_tim2d(ctx_factory):
@@ -64,14 +59,14 @@ def test_tim2d(ctx_factory):
],
[
- lp.GlobalArg("u", dtype, shape=field_shape, order=order),
- lp.GlobalArg("lap", dtype, shape=field_shape, order=order),
- lp.GlobalArg("G", dtype, shape=(3,)+field_shape, order=order),
- # lp.ConstantArrayArg("D", dtype, shape=(n, n), order=order),
- lp.GlobalArg("D", dtype, shape=(n, n), order=order),
- # lp.ImageArg("D", dtype, shape=(n, n)),
- lp.ValueArg("K", np.int32, approximately=1000),
- ],
+ lp.GlobalArg("u", dtype, shape=field_shape, order=order),
+ lp.GlobalArg("lap", dtype, shape=field_shape, order=order),
+ lp.GlobalArg("G", dtype, shape=(3,)+field_shape, order=order),
+ # lp.ConstantArrayArg("D", dtype, shape=(n, n), order=order),
+ lp.GlobalArg("D", dtype, shape=(n, n), order=order),
+ # lp.ImageArg("D", dtype, shape=(n, n)),
+ lp.ValueArg("K", np.int32, approximately=1000),
+ ],
name="semlap2D", assumptions="K>=1")
knl = lp.duplicate_inames(knl, "o", within="ur")
@@ -103,18 +98,13 @@ def test_tim2d(ctx_factory):
return knl
for variant in [variant_orig]:
- kernel_gen = lp.generate_loop_schedules(variant(knl))
- kernel_gen = lp.check_kernels(kernel_gen, dict(K=1000))
-
K = 1000
- lp.auto_test_vs_ref(seq_knl, ctx, kernel_gen,
+ lp.auto_test_vs_ref(seq_knl, ctx, variant(knl),
op_count=[K*(n*n*n*2*2 + n*n*2*3 + n**3 * 2*2)/1e9],
op_label=["GFlops"],
parameters={"K": K})
-
-
if __name__ == "__main__":
import sys
if len(sys.argv) > 1:
--
GitLab