From a6c9db115e97ffaf59eb48389be342fb1f14c699 Mon Sep 17 00:00:00 2001
From: Jean-Baptiste Keck <jean-baptiste.keck@imag.fr>
Date: Sun, 1 Oct 2017 13:27:58 +0200
Subject: [PATCH] custom hashes for autotunable kernels
---
hysop/__init__.py | 2 +-
hysop/backend/device/autotunable_kernel.py | 48 ++++++++++++++++++-
hysop/backend/device/kernel_autotuner.py | 11 +++--
.../autotunable_kernels/advection_dir.py | 10 ++--
.../opencl/autotunable_kernels/remesh_dir.py | 10 ++--
.../opencl/autotunable_kernels/transpose.py | 4 +-
.../opencl/opencl_autotunable_kernel.py | 5 +-
hysop/backend/device/opencl/opencl_env.py | 14 +++---
.../operator/directional/advection_dir.py | 2 +-
hysop/operator/base/redistribute_operator.py | 4 +-
10 files changed, 81 insertions(+), 29 deletions(-)
diff --git a/hysop/__init__.py b/hysop/__init__.py
index 279e86abe..d877538ec 100644
--- a/hysop/__init__.py
+++ b/hysop/__init__.py
@@ -20,7 +20,7 @@ __VERBOSE__ = False
__DEBUG__ = False
__TRACE__ = False
__TRACE_WARNINGS__ = False
-__KERNEL_DEBUG__ = True
+__KERNEL_DEBUG__ = False
__PROFILE__ = True
__ENABLE_LONG_TESTS__ = "OFF" is "ON"
diff --git a/hysop/backend/device/autotunable_kernel.py b/hysop/backend/device/autotunable_kernel.py
index 13403e47d..12916e206 100644
--- a/hysop/backend/device/autotunable_kernel.py
+++ b/hysop/backend/device/autotunable_kernel.py
@@ -21,9 +21,55 @@ class AutotunableKernel(object):
self.dump_src = first_not_None(dump_src, autotuner_config.debug)
self.symbolic_mode = first_not_None(symbolic_mode, autotuner_config.debug)
+
+ def custom_hash(self, *args, **kwds):
+ assert args or kwds, 'no arguments to be hashed.'
+ def _hash_arg(a):
+ if isinstance(a, list):
+ return hash(tuple(_hash_arg(x) for x in a))
+ elif isinstance(a, set):
+ return hash(tuple(_hash_arg(x) for x in a))
+ elif isinstance(a, dict):
+ return hash(tuple((_hash_arg(k), _hash_arg(v)) for (k,v) in a.items()))
+ elif isinstance(a, npw.ndarray):
+ assert a.ndim == 1
+ assert a.size < 17, 'Only parameters up to size 16 are allowed.'
+ return hash(tuple(a.tolist()))
+ else:
+ return hash(a)
+ def _hash_karg(k,v):
+ if k == 'mesh_info_vars':
+ # for mesh infos we just hash the code generated constants that
+ # may alter the code branching.
+ from hysop.backend.device.codegen.base.variables import CodegenStruct
+ check_instance(v, dict, keys=str, values=CodegenStruct)
+ mesh_infos = tuple(str(v[k]) for k in sorted(v.keys()))
+ h = hash(mesh_infos)
+ return h
+ else:
+ msg='Unknown custom hash key \'\'.'.format(k)
+ raise KeyError(msg)
+
+ h = None
+ if args:
+ h = _hash_arg(args[0])
+ for arg in args[1:]:
+ h ^= _hash_arg(arg)
+ if kwds:
+ items = kwds.items()
+ if h is None:
+ h = _hash_karg(*items[0])
+ else:
+ h ^= _hash_karg(*items[0])
+ for it in items[1:]:
+ h ^= _hash_karg(*it)
+ return h
+
@abstractmethod
- def autotune(self, name, kernel_args, **extra_kwds):
+ def autotune(self, name, kernel_args,
+ force_verbose=False, force_debug=False,
+ **extra_kwds):
"""Autotune this kernel with given name and extra_kwds."""
pass
diff --git a/hysop/backend/device/kernel_autotuner.py b/hysop/backend/device/kernel_autotuner.py
index 58ba6ef68..4ce75bb7d 100644
--- a/hysop/backend/device/kernel_autotuner.py
+++ b/hysop/backend/device/kernel_autotuner.py
@@ -150,14 +150,16 @@ class KernelAutotuner(object):
src_hash = hasher.hexdigest()
if (kernel_name != cached_kernel_name):
- msg='\nCached kernel name did not match the benched one:\n {}\n {}'
- msg=msg.format(kernel_name, cached_kernel_name)
+ msg='\nCached kernel name did not match the benched one:\n {}\n {}\n'
+ msg+='\nThis might be due to a faulty implementation of {}.hash_extra_kwds().'
+ msg=msg.format(kernel_name, cached_kernel_name, type(tkernel).__name__)
warnings.warn(msg, CodeGeneratorWarning)
return None
if (src_hash != cached_src_hash):
msg='\nCached kernel source hash did not match the benched one.\n {}\n {}'
- msg=msg.format(src_hash, cached_src_hash)
+ msg+='\nThis might be due to a faulty implementation of {}.hash_extra_kwds().'
+ msg=msg.format(src_hash, cached_src_hash, type(tkernel).__name__)
warnings.warn(msg, CodeGeneratorWarning)
return None
@@ -235,7 +237,8 @@ class KernelAutotuner(object):
(cache_src_hash, cache_stats) = results[run_key]
if cache_src_hash != src_hash:
msg='\nCached parameters candidate did not match the benched one.\n {}\n {}'
- msg=msg.format(src_hash, cache_src_hash)
+ msg+='\nThis might be due to a faulty implementation of {}.hash_extra_kwds().'
+ msg=msg.format(src_hash, cached_src_hash, type(tunable_kernel).__name__)
warnings.warn(msg, CodeGeneratorWarning)
old_stats = None
else:
diff --git a/hysop/backend/device/opencl/autotunable_kernels/advection_dir.py b/hysop/backend/device/opencl/autotunable_kernels/advection_dir.py
index 274e3fdf8..99e43a432 100644
--- a/hysop/backend/device/opencl/autotunable_kernels/advection_dir.py
+++ b/hysop/backend/device/opencl/autotunable_kernels/advection_dir.py
@@ -13,7 +13,7 @@ class OpenClAutotunableDirectionalAdvectionKernel(OpenClAutotunableKernel):
"""Autotunable interface for directional advection kernel code generators."""
def autotune(self, direction, time_integrator, velocity_cfl,
- velocity, position, precision):
+ velocity, position, precision, **kwds):
"""Autotune this kernel with specified configuration."""
dim = velocity.dim
@@ -67,7 +67,7 @@ class OpenClAutotunableDirectionalAdvectionKernel(OpenClAutotunableKernel):
rk_scheme=time_integrator, kernel_args=kernel_args, cache_ghosts=cache_ghosts,
vboundaries=vboundaries, precision=precision, ftype=ftype,
mesh_info_vars=mesh_info_vars, work_dim=dim,
- work_size=position.compute_resolution)
+ work_size=position.compute_resolution, **kwds)
def compute_args_mapping(self, extra_kwds, extra_parameters):
@@ -161,7 +161,7 @@ class OpenClAutotunableDirectionalAdvectionKernel(OpenClAutotunableKernel):
def hash_extra_kwds(self, extra_kwds):
"""Hash extra_kwds dictionnary for caching purposes."""
kwds = ('rk_scheme', 'ftype', 'work_dim',
- 'vboundaries', 'cache_ghosts')
- return hash(tuple(extra_kwds[kwd] for kwd in kwds)) ^ \
- hash(tuple(extra_kwds['work_size'].tolist()))
+ 'vboundaries', 'cache_ghosts', 'work_size')
+ return self.custom_hash(*tuple(extra_kwds[kwd] for kwd in kwds),
+ mesh_info_vars=extra_kwds['mesh_info_vars'])
diff --git a/hysop/backend/device/opencl/autotunable_kernels/remesh_dir.py b/hysop/backend/device/opencl/autotunable_kernels/remesh_dir.py
index f0ea8a914..8e55ca03d 100644
--- a/hysop/backend/device/opencl/autotunable_kernels/remesh_dir.py
+++ b/hysop/backend/device/opencl/autotunable_kernels/remesh_dir.py
@@ -20,7 +20,7 @@ class OpenClAutotunableDirectionalRemeshKernel(OpenClAutotunableKernel):
def autotune(self, precision, direction, scalar_cfl,
position, scalars_in, scalars_out, is_inplace,
remesh_kernel, remesh_criteria_eps,
- force_atomics, relax_min_particles):
+ force_atomics, relax_min_particles, **kwds):
"""Autotune this kernel with specified configuration."""
check_instance(scalars_in, tuple, values=CartesianDiscreteFieldView)
check_instance(scalars_out, tuple, values=CartesianDiscreteFieldView)
@@ -161,7 +161,7 @@ class OpenClAutotunableDirectionalRemeshKernel(OpenClAutotunableKernel):
remesh_kernel=remesh_kernel, remesh_criteria_eps=remesh_criteria_eps,
force_atomics=force_atomics, min_nparticles=min_nparticles, ftype=ftype,
scalar_cfl=scalar_cfl, kernel_args=kernel_args, mesh_info_vars=mesh_info_vars,
- work_dim=work_dim, work_size=work_size, min_wg_size=min_wg_size)
+ work_dim=work_dim, work_size=work_size, min_wg_size=min_wg_size, **kwds)
def compute_args_mapping(self, extra_kwds, extra_parameters):
@@ -339,7 +339,7 @@ class OpenClAutotunableDirectionalRemeshKernel(OpenClAutotunableKernel):
def hash_extra_kwds(self, extra_kwds):
"""Hash extra_kwds dictionnary for caching purposes."""
kwds = ('remesh_criteria_eps', 'nscalars', 'ftype',
- 'is_inplace', 'remesh_kernel')
- return hash(tuple(extra_kwds[kwd] for kwd in kwds)) ^ \
- hash(tuple(extra_kwds['work_size'].tolist()))
+ 'is_inplace', 'remesh_kernel', 'work_size')
+ return self.custom_hash(*tuple(extra_kwds[kwd] for kwd in kwds),
+ mesh_info_vars=extra_kwds['mesh_info_vars'])
diff --git a/hysop/backend/device/opencl/autotunable_kernels/transpose.py b/hysop/backend/device/opencl/autotunable_kernels/transpose.py
index 14dea8a47..78d282253 100644
--- a/hysop/backend/device/opencl/autotunable_kernels/transpose.py
+++ b/hysop/backend/device/opencl/autotunable_kernels/transpose.py
@@ -31,7 +31,7 @@ class OpenClAutotunableTransposeKernel(OpenClAutotunableKernel):
def autotune(self, is_inplace,
input_field, output_field,
- axes, name=None):
+ axes, name=None, **kwds):
"""Autotune this kernel with specified axes, inputs and outputs."""
check_instance(axes, tuple, values=int)
@@ -89,7 +89,7 @@ class OpenClAutotunableTransposeKernel(OpenClAutotunableKernel):
tile_indices=tile_indices,
work_dim=work_dim,
work_size=work_shape,
- last_axe_permuted=last_axe_permuted)
+ last_axe_permuted=last_axe_permuted, **kwds)
def compute_parameters(self, extra_kwds):
"""Register extra parameters to optimize."""
diff --git a/hysop/backend/device/opencl/opencl_autotunable_kernel.py b/hysop/backend/device/opencl/opencl_autotunable_kernel.py
index 56d1974c3..a261662ca 100644
--- a/hysop/backend/device/opencl/opencl_autotunable_kernel.py
+++ b/hysop/backend/device/opencl/opencl_autotunable_kernel.py
@@ -28,11 +28,12 @@ class OpenClAutotunableKernel(AutotunableKernel):
self.cl_env = cl_env
self.usable_cache_bytes_per_wg = clCharacterize.usable_local_mem_size(cl_env.device)
- def autotune(self, name, **extra_kwds):
+ def autotune(self, name, force_verbose=False, force_debug=False, **extra_kwds):
from hysop.backend.device.opencl.opencl_kernel_autotuner import OpenClKernelAutotuner
autotuner = OpenClKernelAutotuner(name=name, tunable_kernel=self)
- best_candidate_results = autotuner.autotune(extra_kwds=extra_kwds)
+ best_candidate_results = autotuner.autotune(extra_kwds=extra_kwds,
+ force_verbose=force_verbose, force_debug=force_debug)
check_instance(best_candidate_results, dict)
return self.format_best_candidate(name=name, extra_kwds=extra_kwds,
diff --git a/hysop/backend/device/opencl/opencl_env.py b/hysop/backend/device/opencl/opencl_env.py
index d9324eb3e..bd0921b08 100644
--- a/hysop/backend/device/opencl/opencl_env.py
+++ b/hysop/backend/device/opencl/opencl_env.py
@@ -587,6 +587,14 @@ Dumped OpenCL Kernel '{}'
dump_folder=IO.default_path()+'/'+OPENCL_KERNEL_DUMP_FOLDER
if not os.path.exists(dump_folder):
os.makedirs(dump_folder)
+
+ if DEBUG:
+ # dump kernel source while in debug mode
+ dump_file=dump_folder+'/'+'{}_dump.cl'.format(kernel_name)
+ print('Dumping kernel src at \'{}\'.'.format(dump_file))
+ with open(dump_file, 'w+') as f:
+ f.write(gpu_src)
+ #build_opts += ' '+' '.join(['-g', '-s "{}"'.format(dump_file)])
# Build OpenCL program
try:
@@ -607,12 +615,6 @@ Dumped OpenCL Kernel '{}'
build.get_build_info(self.device, cl.program_build_info.STATUS))
vprint('Compiler log: ',
build.get_build_info(self.device, cl.program_build_info.LOG))
- if DEBUG:
- # dump kernel source while in debug mode
- dump_file=dump_folder+'/'+'{}_dump.cl'.format(kernel_name)
- print('Dumping kernel src at \'{}\'.'.format(dump_file))
- with open(dump_file, 'w+') as f:
- f.write(gpu_src)
if VERBOSE:
print("===\n")
diff --git a/hysop/backend/device/opencl/operator/directional/advection_dir.py b/hysop/backend/device/opencl/operator/directional/advection_dir.py
index 854c1c784..4b9950563 100644
--- a/hysop/backend/device/opencl/operator/directional/advection_dir.py
+++ b/hysop/backend/device/opencl/operator/directional/advection_dir.py
@@ -107,7 +107,7 @@ class OpenClDirectionalAdvection(DirectionalAdvectionBase, OpenClDirectionalOper
kwds['velocity_cfl'] = self.velocity_cfl
kwds['time_integrator'] = self.time_integrator
- (advec_kernel, args_dict) = kernel.autotune(**kwds)
+ (advec_kernel, args_dict) = kernel.autotune(force_debug=True, **kwds)
args_dict.pop('dt')
advec_launcher = advec_kernel.build_launcher(**args_dict)
diff --git a/hysop/operator/base/redistribute_operator.py b/hysop/operator/base/redistribute_operator.py
index 9bf6d91d8..4f6360bd6 100644
--- a/hysop/operator/base/redistribute_operator.py
+++ b/hysop/operator/base/redistribute_operator.py
@@ -69,12 +69,12 @@ class RedistributeOperatorBase(ComputationalGraphOperator):
for field in self.input_vars:
_, req = reqs.get_input_requirement(field)
- req.transposition_states = None
+ req.axes = None
req.basis = None
for field in self.output_vars:
_, req = reqs.get_output_requirement(field)
- req.transposition_states = None
+ req.axes = None
req.basis = None
return reqs
--
GitLab