diff --git a/hysop/backend/device/codegen/base/codegen.py b/hysop/backend/device/codegen/base/codegen.py
index a4b397e59dfb87fb8add171d7c12554c47ffbb06..0d313e3e82a795fbdabe2f52122bfaa259b6c9a2 100644
--- a/hysop/backend/device/codegen/base/codegen.py
+++ b/hysop/backend/device/codegen/base/codegen.py
@@ -2,12 +2,11 @@
from contextlib import contextmanager
from subprocess import call
-import sys, os, string, tempfile, operator
-import itertools as it
-
-import pyopencl as cl
-
+from hysop.tools.types import check_instance
+from hysop.deps import it, sys, os, string, tempfile, operator
+from hysop.backend.device.opencl import cl
from hysop.backend.device.codegen.base.utils import WriteOnceDict, VarDict
+from hysop.backend.device.codegen.base.variables import CodegenVariable
class CodeGenerator(object):
@@ -249,13 +248,25 @@ class CodeGenerator(object):
else:
self.append(code)
- def decl_vars(self, *variables):
+ def decl_vars(self, *variables, **kargs):
assert len(set(var.base_ctype() for var in variables))==1
base= variables[0].base_ctype()
svars=[]
for var in variables:
- svars.append(var.declare(multidecl=True))
- return '{} {};'.format(base, ', '.join(svars))
+ check_instance(var, CodegenVariable)
+ svars.append(var.declare(multidecl=True, **kargs))
+ decl = '{} {};'.format(base, ', '.join(svars))
+ self.append(decl)
+
+ def decl_aligned_vars(self, *variables, **kargs):
+ assert len(variables)>0
+ jmp = kargs.pop('jmp', True)
+ with self._align_() as al:
+ for var in variables:
+ check_instance(var, CodegenVariable)
+ var.declare(codegen=al, align=True, **kargs)
+ if jmp:
+ al.jumpline()
class VarBlock(object):
@@ -335,7 +346,9 @@ class CodeGenerator(object):
def code(self):
if self._parts_count is None:
- raise RuntimeError('Call at least one append() before closing an _align_!')
+ msg='Call at least one append() before closing an _align_!'
+ msg+='\n got: {}'.format(self._lines)
+ raise RuntimeError(msg)
maxlen = lambda i: max([len(line[i]) for line in self._lines if len(line)>1])
line_str = ''
diff --git a/hysop/backend/device/codegen/base/variables.py b/hysop/backend/device/codegen/base/variables.py
index 430840bc2c357cd9b6eb7d39ba4a62769d1fb2ab..3d454ea62c9eaa21bccdf9487c5e9c83f66f89b2 100644
--- a/hysop/backend/device/codegen/base/variables.py
+++ b/hysop/backend/device/codegen/base/variables.py
@@ -82,11 +82,6 @@ class CodegenVariable(object):
check_instance(add_impl_const, bool)
check_instance(nl, bool, allow_none=True)
- if const and add_impl_const:
- msg='Variable {} is const and add_impl_const has been specified!'
- msg=msg.format(name)
- raise ValueError(msg)
-
self.name = name
self.ctype = ctype
self.typegen = typegen
@@ -158,6 +153,20 @@ class CodegenVariable(object):
self.value = value
self.svalue = svalue
self.init=init
+
+ # check
+ if add_impl_const:
+ if (not is_ptr):
+ if const:
+ msg='Variable {} is const and add_impl_const has been specified!'
+ msg=msg.format(name)
+ raise ValueError(msg)
+ else:
+ if ptr_const[-1]:
+ msg='Variable {} has ptr_const[-1]=True and add_impl_const has been specified!'
+ msg=msg.format(name)
+ raise ValueError(msg)
+
def newvar(name, nl=False,
storage=None, value=None, svalue=None, init=None,
@@ -219,7 +228,8 @@ class CodegenVariable(object):
def base_ctype(self, storage=None, ctype=None,
const=None, volatile=False,
- impl=True, align=False):
+ impl=True, align=False,
+ add_impl_const=None):
storage = self.storage if (storage is None) else storage
ctype = self.ctype if (ctype is None) else ctype
volatile = self.volatile if (volatile is None) else volatile
@@ -227,9 +237,9 @@ class CodegenVariable(object):
if (const is None):
const = self.const
if impl and (not self.is_ptr) and (not const):
- const = self.add_impl_const
+ const = self.add_impl_const if (add_impl_const is None) else add_impl_const
- base_ctype = '{storage}${ctype}${const}${volatile}'.format(
+ base_ctype = '{storage}${const}${volatile}${ctype}'.format(
storage='{} '.format(storage) if (storage is not None) else '',
const='const ' if const else '',
volatile='volatile ' if volatile else '',
@@ -237,20 +247,15 @@ class CodegenVariable(object):
if not align:
base_ctype = base_ctype.replace('$','')
return base_ctype.strip()
-
- def full_ctype(self, storage=None, ctype=None, const=None, volatile=False,
- impl=True, multidecl=False, align=False, cast=False):
- if multidecl:
- base_ctype = ''
- else:
- base_ctype = self.base_ctype(storage,ctype,const,volatile,impl,align)
-
+
+ def ptr_ctype(self, impl=True, add_impl_const=None, cast=False):
if self.is_ptr:
ptrs=[]
+ add_impl_const = self.add_impl_const if (add_impl_const is None) else add_impl_const
for i, (c,v,r) in enumerate(zip(self.ptr_const, self.ptr_volatile, self.ptr_restrict)):
if i==self.ptr_level-1:
- c = c or (impl and self.add_impl_const)
+ c = c or (impl and add_impl_const)
ptr=' $*{const}${volatile}${restrict}'.format(
const = 'const ' if (c and not cast) else '',
volatile = 'volatile ' if (v and not cast) else '',
@@ -259,6 +264,22 @@ class CodegenVariable(object):
ptr_ctype = ''.join(ptrs)
else:
ptr_ctype=''
+ return ptr_ctype
+
+ def full_ctype(self, storage=None, ctype=None, const=None, volatile=False,
+ impl=True, multidecl=False, align=False, cast=False,
+ add_impl_const=None):
+
+ if multidecl:
+ base_ctype = ''
+ else:
+ base_ctype = self.base_ctype(storage,ctype,const,volatile,impl,align,
+ add_impl_const=add_impl_const)
+ if len(base_ctype)==0:
+ msg= 'Failed to get base ctype in {}.'.format(self.__class__)
+ raise RuntimeError(msg)
+
+ ptr_ctype = self.ptr_ctype(impl=impl, add_impl_const=add_impl_const, cast=cast)
full_ctype = '{}{}'.format(base_ctype, ptr_ctype)
if not align:
@@ -316,27 +337,39 @@ class CodegenVariable(object):
acc += '->' if self.is_ptr else '.'
return acc
+ def decl_name(self):
+ return self.name
+
def declare(self, codegen=None, align=False,
- multidecl=False, const=None, init=None):
-
- ctype = self.full_ctype(align=align,multidecl=multidecl,const=const)
+ multidecl=False, const=None, init=None,
+ compact=False):
+ # const means add_impl_const, ie. declare current variable as constant (not pointed types)
+ ctype = self.full_ctype(align=align,multidecl=multidecl,add_impl_const=const)
+ if (not multidecl) and len(ctype)==0:
+ msg= 'Failed to get full ctype in {}.'.format(self.__class__)
+ raise RuntimeError(msg)
# static array ctype needs to be split
- split = ctype.split('[',1)
- ctype = split[0]
- is_static_array = len(split)==2
- name_suffix = ('['+split[1]) if is_static_array else ''
- name = self.name+name_suffix
+ name = self.decl_name()
init = init if (init is not None) else self.init
- code = '{} ${}'.format(ctype, name)
+ if (len(ctype)>0) and ctype[-1]=='*':
+ code = '{}${}'.format(ctype, name)
+ else:
+ code = '{} ${}'.format(ctype, name)
if (init is not None):
- code = '{} $= {}'.format(code,init)
+ if compact:
+ code = '{}={}'.format(code,init)
+ else:
+ code = '{} $= {}'.format(code,init)
elif self.known():
self.force_symbolic(False)
sval = self.sval()
- code = '{} $= {}'.format(code,sval)
+ if compact:
+ code = '{}={}'.format(code,sval)
+ else:
+ code = '{} $= {}'.format(code,sval)
if not multidecl:
code+=';'
@@ -347,7 +380,7 @@ class CodegenVariable(object):
code = code.replace('$','')
if codegen is not None:
codegen.append(code)
- return code
+ return code.strip()
def __getitem__(self,ss):
if self.is_ptr:
@@ -499,6 +532,46 @@ class CodegenArray(CodegenVariable):
symbolic_mode=symbolic_mode, struct_var=struct_var)
self.shape = shape
self.sshape = sshape
+
+ def decl_name(self):
+ if self.shape:
+ static_array = ['[{}]'.format(val) for val in self.shape]
+ elif self.shape:
+ static_array = ['[{}]'.format(val) for val in self.sshape]
+ else:
+ static_array = []
+ return '{}{}'.format(self.name, ''.join(static_array))
+
+ def array_dim(self):
+ if (self.shape is not None):
+ return self.shape.size
+ if (self.sshape is not None):
+ return self.sshape.size
+ msg='unknown array dim.'
+ raise RuntimeError(msg)
+
+ def ptr_ctype(self, impl=True, add_impl_const=None, cast=False):
+ if self.is_ptr:
+ add_impl_const = self.add_impl_const if (add_impl_const is None) else add_impl_const
+
+ dim = self.array_dim()
+ ptr_const = self.ptr_const[dim:]
+ ptr_volatile = self.ptr_volatile[dim:]
+ ptr_restrict = self.ptr_restrict[dim:]
+
+ ptrs=[]
+ for i, (c,v,r) in enumerate(zip(ptr_const, ptr_volatile, ptr_restrict)):
+ if i==self.ptr_level-1:
+ c = c or (impl and add_impl_const)
+ ptr=' $*{const}${volatile}${restrict}'.format(
+ const = 'const ' if (c and not cast) else '',
+ volatile = 'volatile ' if (v and not cast) else '',
+ restrict = 'restrict ' if (r and not cast) else '')
+ ptrs.append(ptr)
+ ptr_ctype = ''.join(ptrs)
+ else:
+ ptr_ctype=''
+ return ptr_ctype
@@ -663,14 +736,14 @@ class CodegenVectorClBuiltin(CodegenVector):
else:
raise TypeError, 'Invalid key type!'
- def declare(self,codegen=None,align=False,const=None,init=None):
+ def declare(self, init=None, **kargs):
if isinstance(init,int):
init = ','.join([self.typegen.dump(init) for _ in xrange(self.dim)])
init = '({})({})'.format(self.ctype,init)
elif init.__class__ in [list,tuple,np.ndarray]:
init = ','.join([self.typegen.dump(init[i]) for i in xrange(self.dim)])
init = '({})({})'.format(self.ctype,init)
- return super(CodegenVectorClBuiltin,self).declare(codegen=codegen,align=align,const=const,init=init)
+ return super(CodegenVectorClBuiltin,self).declare(init=init, **kargs)
class CodegenVectorClBuiltinFunc(CodegenVectorClBuiltin):
def __init__(self,fname,name,btype,dim,typegen,
diff --git a/hysop/backend/device/codegen/kernels/directional_remesh.py b/hysop/backend/device/codegen/kernels/directional_remesh.py
index 2aee3c01630d859542006c119d4474de7591d9d6..2199743e92c9df8a3f48b77aef871999744de610 100644
--- a/hysop/backend/device/codegen/kernels/directional_remesh.py
+++ b/hysop/backend/device/codegen/kernels/directional_remesh.py
@@ -43,27 +43,26 @@ from hysop.constants import DirectionLabels
class DirectionalRemeshKernel(KernelCodeGenerator):
@staticmethod
- def codegen_name(work_dim, direction,
+ def codegen_name(work_dim,
remesh_kernel, ftype,
nparticles, nscalars,
remesh_criteria_eps,
use_atomics, is_inplace):
inplace = 'inplace_' if is_inplace else ''
atomic = 'atomic_' if use_atomics else ''
- criteria = '{}eps__' if (remesh_criteria_eps is not None) else ''
- return 'directional_{}{}remesh_{}d__lambda_{}_{}__{}__{}p__{}s__{}{}'.format(
+ criteria = '{}eps__' if (remesh_criteria_eps is not None) else 'full'
+ return 'directional_{}{}remesh_{}d__lambda_{}_{}__{}__{}p__{}s__{}'.format(
inplace, atomic, work_dim,
remesh_kernel.n, remesh_kernel.r, ftype,
nparticles, nscalars,
- criteria,
- DirectionLabels[direction])
+ criteria)
@staticmethod
def cache_ghosts(scalar_cfl, remesh_kernel):
assert remesh_kernel.n % 2 == 0
return int(1+math.ceil(scalar_cfl)+remesh_kernel.n/2)
- def __init__(self, typegen, work_dim, direction, ftype,
+ def __init__(self, typegen, work_dim, ftype,
nparticles, nscalars, sboundary, is_inplace,
scalar_cfl, remesh_kernel,
remesh_criteria_eps=None,
@@ -73,7 +72,6 @@ class DirectionalRemeshKernel(KernelCodeGenerator):
known_vars = None):
assert work_dim>0 and work_dim<=3
- assert direction>=0 and direction<work_dim
assert nscalars>0
assert nparticles in [1,2,4,8,16]
check_instance(sboundary[0],BoundaryCondition)
@@ -92,7 +90,7 @@ class DirectionalRemeshKernel(KernelCodeGenerator):
vftype = tg.vtype(ftype, nparticles)
vitype = tg.vtype(itype, nparticles)
- name = DirectionalRemeshKernel.codegen_name(work_dim, direction,
+ name = DirectionalRemeshKernel.codegen_name(work_dim,
remesh_kernel, ftype,
nparticles,nscalars, remesh_criteria_eps,
use_atomics, is_inplace)
@@ -122,7 +120,6 @@ class DirectionalRemeshKernel(KernelCodeGenerator):
self.vitype = vitype
self.vftype = vftype
self.work_dim = work_dim
- self.direction = direction
self.sboundary = sboundary
self.nparticles = nparticles
self.nscalars = nscalars
@@ -154,22 +151,22 @@ class DirectionalRemeshKernel(KernelCodeGenerator):
kargs = ArgDict()
self.position = OpenClArrayBackend.build_codegen_argument(kargs, name='position',
storage=self._global, ctype=ftype, typegen=typegen,
- ptr_restrict=True, ptr_const=True)
+ ptr_restrict=True, const=True)
if is_inplace:
self.scalars_in = tuple(
- OpenClArrayBackend.build_codegen_argument(kargs, name=' s{}_in'.format(i),
+ OpenClArrayBackend.build_codegen_argument(kargs, name='S{}_inout'.format(i),
storage=self._global, ctype=ftype, typegen=typegen,
- ptr_restrict=True, ptr_const=False) for i in xrange(nscalars))
+ const=False, ptr_restrict=True) for i in xrange(nscalars))
self.scalars_out = self.scalars_in
else:
self.scalars_in = tuple(
- OpenClArrayBackend.build_codegen_argument(kargs, name='s{}_in'.format(i),
+ OpenClArrayBackend.build_codegen_argument(kargs, name='S{}_in'.format(i),
storage=self._global, ctype=ftype, typegen=typegen,
- ptr_restrict=True, ptr_const=True) for i in xrange(nscalars))
+ const=True, ptr_restrict=True) for i in xrange(nscalars))
self.scalars_out = tuple(
- OpenClArrayBackend.build_codegen_argument(kargs, name='s{}_out'.format(i),
+ OpenClArrayBackend.build_codegen_argument(kargs, name='S{}_out'.format(i),
storage=self._global, ctype=ftype, typegen=typegen,
- ptr_restrict=True, ptr_const=False) for i in xrange(nscalars))
+ const=False, ptr_restrict=True) for i in xrange(nscalars))
if debug_mode:
kargs['dbg0'] = CodegenVariable(storage=self._global,name='dbg0',ctype=itype,
@@ -233,7 +230,7 @@ class DirectionalRemeshKernel(KernelCodeGenerator):
scalars_out = s.scalars_out
compute_grid_size = position_mesh_info['local_mesh']['compute_resolution'].view(
- 'compute_grid_size',slice(None,work_dim))
+ 'compute_grid_size',slice(None,work_dim),const=True)
position_grid_ghosts = position_mesh_info['ghosts'].view(
'pos_grid_ghosts', slice(0,work_dim), const=True)
@@ -333,59 +330,36 @@ class DirectionalRemeshKernel(KernelCodeGenerator):
yield ctx
except:
raise
-
- nested_loops = [_work_iterate_(i) for i in xrange(dim-1,-1,-1)]
-
+
+
with s._kernel_():
s.jumpline()
- with s._align_() as al:
- position.declare(al,const=True,align=True)
- for i in xrange(nscalars):
- scalars_in[i].declare(al,const=True,align=True)
- if not is_inplace:
- for i in xrange(nscalars):
- scalars_out[i].declare(al,const=False,align=True)
- s.jumpline()
- with s._align_() as al:
- local_id.declare(al,align=True,const=True)
- global_size.declare(al,align=True,const=True)
- local_size.declare(al,align=True,const=True)
- s.jumpline()
- with s._align_() as al:
- xmin.declare(al,align=True)
- inv_dx.declare(al,align=True)
- dx.declare(al,align=True)
- s.jumpline()
+ if is_inplace:
+ s.decl_aligned_vars(*((position,)+scalars_in))
+ else:
+ s.decl_aligned_vars(*((position,)+scalars_in+scalars_out))
- with s._align_() as al:
- npart.declare(al, const=True, align=True)
- cache_ghosts.declare(al,const=True, align=True)
- cache_width.declare(al,const=True, align=True)
- local_work.declare(al,const=True, align=True)
- s.jumpline()
+ s.decl_aligned_vars(local_id, global_size, local_size, const=True)
- compute_grid_size.declare(s,const=True)
- s.jumpline()
- with s._align_() as al:
- position_grid_ghosts.declare(al,align=True)
- for sgg in scalars_grid_ghosts:
- sgg.declare(al,align=True)
- s.jumpline()
+ s.decl_aligned_vars(xmin, inv_dx, dx, const=True)
- with s._align_() as al:
- position_global_id.declare(al,align=True)
- for sgid in scalars_global_id:
- sgid.declare(al,align=True)
- s.jumpline()
+ s.decl_aligned_vars(npart, cache_ghosts, cache_width, local_work, const=True)
- #with s._align_() as al:
- #for var in cached_scalars:
- #var.declare(al,align=True);
- s.decl_vars(*cached_scalars)
+ s.decl_aligned_vars(compute_grid_size, position_grid_ghosts, *scalars_grid_ghosts)
+
+ s.decl_vars(position_global_id, *scalars_global_id)
s.jumpline()
-
+
+ if local_size_known:
+ s.decl_vars(*cached_scalars)
+ s.jumpline()
+ else:
+ s.decl_aligned_vars(*cached_scalars)
+
+
+ nested_loops = [_work_iterate_(i) for i in xrange(dim-1,-1,-1)]
with contextlib.nested(*nested_loops):
# s.barrier(_local=True)
# s.jumpline()
@@ -411,7 +385,6 @@ if __name__ == '__main__':
dak = DirectionalRemeshKernel(typegen=tg, ftype=tg.fbtype,
work_dim=work_dim,
- direction=2,
nparticles=4,
nscalars=2,
remesh_kernel=kernel,
@@ -424,7 +397,7 @@ if __name__ == '__main__':
s0_mesh_info=smesh_info,
s1_mesh_info=smesh_info,
position_mesh_info=pmesh_info,
- local_size=local_size[:work_dim],
+ #local_size=local_size[:work_dim],
global_size=global_size[:work_dim]
)
)
diff --git a/hysop/backend/device/opencl/opencl_array_backend.py b/hysop/backend/device/opencl/opencl_array_backend.py
index d53ee94989ebfc977f568dfb307a6affa5b187d4..47bfa8fa26e81a38e70427151ab42cfb7d2d799c 100644
--- a/hysop/backend/device/opencl/opencl_array_backend.py
+++ b/hysop/backend/device/opencl/opencl_array_backend.py
@@ -2993,8 +2993,8 @@ class OpenClArrayBackend(ArrayBackend):
typegen=typegen, ctype=itype,
add_impl_const=True, nl=True)
- char_alias = args[base].full_ctype(ctype='char', cast=True)
- ctype_alias = args[base].full_ctype(cast=True)
+ char_alias = args[base].full_ctype(ctype='char', cast=True, align=True)
+ ctype_alias = args[base].full_ctype(cast=True, align=True)
init = '({})(({})({})+{})'.format(ctype_alias, char_alias, base, offset)
var = CodegenVariable(name=name, typegen=typegen,
diff --git a/hysop/deps.py b/hysop/deps.py
index 6f41dbbcbeb98116fc14ca4f2e3d88d7612c0982..342a42e5c43b05a2bfa3ec8a21ae3a90b705e820 100644
--- a/hysop/deps.py
+++ b/hysop/deps.py
@@ -21,7 +21,7 @@ except ImportError as e:
print(msg)
import sys, os, subprocess, platform
-import resource, psutil
+import resource, psutil, tempfile
import inspect, functools, operator
import hashlib, gzip, copy, types, string
import math, re, contextlib