def make_codepy_module(self, toolchain, dtype):
from codepy.libraries import add_codepy
toolchain = toolchain.copy()
add_codepy(toolchain)
from cgen import (Value, Include, Statement,
Typedef, FunctionBody, FunctionDeclaration, Block, Const,
Line, POD, Initializer, CustomLoop)
S = Statement
from codepy.bpl import BoostPythonModule
mod = BoostPythonModule()
mod.add_to_preamble([
Include("vector"),
Include("algorithm"),
Include("hedge/base.hpp"),
Include("hedge/volume_operators.hpp"),
Include("boost/foreach.hpp"),
Include("boost/numeric/ublas/io.hpp"),
]+self.get_cpu_extra_includes())
mod.add_to_module([
S("namespace ublas = boost::numeric::ublas"),
S("using namespace hedge"),
S("using namespace pyublas"),
Line(),
Typedef(POD(dtype, "value_type")),
Line(),
])
mod.add_function(FunctionBody(
FunctionDeclaration(Value("void", "process_elements"), [
Const(Value("uniform_element_ranges", "ers")),
Const(Value("numpy_vector<value_type>", "field")),
Value("numpy_vector<value_type>", "result"),
]+self.get_cpu_extra_parameter_declarators()),
Block([
Typedef(Value("numpy_vector<value_type>::iterator",
"it_type")),
Typedef(Value("numpy_vector<value_type>::const_iterator",
"cit_type")),
Line(),
Initializer(Value("it_type", "result_it"),
"result.begin()"),
Initializer(Value("cit_type", "field_it"),
"field.begin()"),
Line() ]+self.get_cpu_extra_preamble()+[ Line(),
CustomLoop(
"BOOST_FOREACH(const element_range er, ers)",
Block(self.get_cpu_per_element_code())
)
])))
#print mod.generate()
#toolchain = toolchain.copy()
#toolchain.enable_debugging
return mod.compile(toolchain)
def get_function_definition(self, codegen_state, codegen_result,
schedule_index,
function_decl, function_body):
kernel = codegen_state.kernel
from cgen import (
FunctionBody,
# Post-mid-2016 cgens have 'Collection', too.
Module as Collection,
Initializer,
Line)
result = []
from loopy.kernel.data import AddressSpace
from loopy.schedule import CallKernel
# We only need to write declarations for global variables with
# the first device program. `is_first_dev_prog` determines
# whether this is the first device program in the schedule.
is_first_dev_prog = codegen_state.is_generating_device_code
for i in range(schedule_index):
if isinstance(kernel.schedule[i], CallKernel):
is_first_dev_prog = False
break
if is_first_dev_prog:
for tv in sorted(
six.itervalues(kernel.temporary_variables),
key=lambda tv: tv.name):
if tv.address_space == AddressSpace.GLOBAL and (
tv.initializer is not None):
assert tv.read_only
decl_info, = tv.decl_info(self.target,
index_dtype=kernel.index_dtype)
decl = self.wrap_global_constant(
self.get_temporary_decl(
codegen_state, schedule_index, tv,
decl_info))
if tv.initializer is not None:
decl = Initializer(decl, generate_array_literal(
codegen_state, tv, tv.initializer))
result.append(decl)
fbody = FunctionBody(function_decl, function_body)
if not result:
return fbody
else:
return Collection(result+[Line(), fbody])
def test_cgen():
s = Struct(
"yuck",
[
POD(
np.float32,
"h",
),
POD(np.float32, "order"),
POD(np.float32, "face_jacobian"),
ArrayOf(POD(np.float32, "normal"), 17),
POD(np.uint16, "a_base"),
POD(np.uint16, "b_base"),
#CudaGlobal(POD(np.uint8, "a_ilist_number")),
POD(np.uint8, "b_ilist_number"),
POD(np.uint8, "bdry_flux_number"), # 0 if not on boundary
POD(np.uint8, "reserved"),
POD(np.uint32, "b_global_base"),
])
f_decl = FunctionDeclaration(POD(np.uint16, "get_num"), [
POD(np.uint8, "reserved"),
POD(np.uint32, "b_global_base"),
])
f_body = FunctionBody(
f_decl,
Block([
POD(np.uint32, "i"),
For(
"i = 0",
"i < 17",
"++i",
If(
"a > b",
Assign("a", "b"),
Block([
Assign("a", "b-1"),
#Break(),
])),
),
#BlankLine(),
Comment("all done"),
]))
t_decl = Template(
'typename T',
FunctionDeclaration(
Value('CUdeviceptr', 'scan'),
[Value('CUdeviceptr', 'inputPtr'),
Value('int', 'length')]))
print(s)
print(f_body)
print(t_decl)
def make_greet_mod(greeting):
from cgen import FunctionBody, FunctionDeclaration, Block, \
Const, Pointer, Value, Statement
from codepy.bpl import BoostPythonModule
mod = BoostPythonModule()
mod.add_function(
FunctionBody(
FunctionDeclaration(Const(Pointer(Value("char", "greet"))), []),
Block([Statement('return "%s"' % greeting)])))
from codepy.toolchain import guess_toolchain
return mod.compile(guess_toolchain(), wait_on_error=True)
def get_elwise_module_descriptor(arguments, operation, name="kernel"):
from codepy.bpl import BoostPythonModule
from cgen import FunctionBody, FunctionDeclaration, \
Value, POD, Struct, For, Initializer, Include, Statement, \
Line, Block
S = Statement # noqa: N806
mod = BoostPythonModule()
mod.add_to_preamble([
Include("pyublas/numpy.hpp"),
])
mod.add_to_module([
S("namespace ublas = boost::numeric::ublas"),
S("using namespace pyublas"),
Line(),
])
body = Block([
Initializer(
Value(
"numpy_array<{} >::iterator".format(dtype_to_ctype(
varg.dtype)), varg.name), f"args.{varg.name}_ary.begin()")
for varg in arguments if isinstance(varg, VectorArg)
] + [
Initializer(sarg.declarator(), f"args.{sarg.name}")
for sarg in arguments if isinstance(sarg, ScalarArg)
])
body.extend([
Line(),
For("unsigned i = 0", "i < codepy_length", "++i",
Block([S(operation)]))
])
arg_struct = Struct("arg_struct", [arg.declarator() for arg in arguments])
mod.add_struct(arg_struct, "ArgStruct")
mod.add_to_module([Line()])
mod.add_function(
FunctionBody(
FunctionDeclaration(Value("void", name), [
POD(numpy.uintp, "codepy_length"),
Value("arg_struct", "args")
]), body))
return mod
def get_function_definition(self, codegen_state, codegen_result,
schedule_index,
function_decl, function_body):
kernel = codegen_state.kernel
from cgen import (
FunctionBody,
# Post-mid-2016 cgens have 'Collection', too.
Module as Collection,
Initializer,
Line)
result = []
from loopy.kernel.data import temp_var_scope
for tv in sorted(
six.itervalues(kernel.temporary_variables),
key=lambda tv: tv.name):
if tv.scope == temp_var_scope.GLOBAL and tv.initializer is not None:
assert tv.read_only
decl_info, = tv.decl_info(self.target,
index_dtype=kernel.index_dtype)
decl = self.wrap_global_constant(
self.get_temporary_decl(
codegen_state, schedule_index, tv,
decl_info))
if tv.initializer is not None:
decl = Initializer(decl, generate_array_literal(
codegen_state, tv, tv.initializer))
result.append(decl)
fbody = FunctionBody(function_decl, function_body)
if not result:
return fbody
else:
return Collection(result+[Line(), fbody])
from cgen import FunctionBody, \
FunctionDeclaration, Typedef, POD, Value, \
Pointer, Module, Block, Initializer, Assign
from cgen.cuda import CudaGlobal
mod = Module([
FunctionBody(
CudaGlobal(
FunctionDeclaration(Value("void", "add"),
arg_decls=[
Pointer(POD(dtype, name))
for name in ["tgt", "op1", "op2"]
])),
Block([
Initializer(
POD(numpy.int32, "idx"), "threadIdx.x + %d*blockIdx.x" %
(block_size * thread_strides)),
] + [
Assign(
"tgt[idx+%d]" % (o * block_size), "op1[idx+%d] + op2[idx+%d]" %
(o * block_size, o * block_size))
for o in range(thread_strides)
]))
])
mod = SourceModule(mod)
func = mod.get_function("add")
func(c_gpu, a_gpu, b_gpu, block=(block_size, 1, 1), grid=(macroblock_count, 1))
def _cusp_solver(M, parameters):
cache_key = lambda t, p: (t, p['ksp_type'], p['pc_type'], p['ksp_rtol'], p[
'ksp_atol'], p['ksp_max_it'], p['ksp_gmres_restart'], p['ksp_monitor'])
module = _cusp_cache.get(cache_key(M.ctype, parameters))
if module:
return module
import codepy.toolchain
from cgen import FunctionBody, FunctionDeclaration
from cgen import Block, Statement, Include, Value
from codepy.bpl import BoostPythonModule
from codepy.cuda import CudaModule
gcc_toolchain = codepy.toolchain.guess_toolchain()
nvcc_toolchain = codepy.toolchain.guess_nvcc_toolchain()
if 'CUSP_HOME' in os.environ:
nvcc_toolchain.add_library('cusp', [os.environ['CUSP_HOME']], [], [])
host_mod = BoostPythonModule()
nvcc_mod = CudaModule(host_mod)
nvcc_includes = [
'thrust/device_vector.h', 'thrust/fill.h', 'cusp/csr_matrix.h',
'cusp/krylov/cg.h', 'cusp/krylov/bicgstab.h', 'cusp/krylov/gmres.h',
'cusp/precond/diagonal.h', 'cusp/precond/smoothed_aggregation.h',
'cusp/precond/ainv.h', 'string'
]
nvcc_mod.add_to_preamble([Include(s) for s in nvcc_includes])
nvcc_mod.add_to_preamble([Statement('using namespace std')])
# We're translating PETSc preconditioner types to CUSP
diag = Statement(
'cusp::precond::diagonal< ValueType, cusp::device_memory >M(A)')
ainv = Statement(
'cusp::precond::scaled_bridson_ainv< ValueType, cusp::device_memory >M(A)'
)
amg = Statement(
'cusp::precond::smoothed_aggregation< IndexType, ValueType, cusp::device_memory >M(A)'
)
none = Statement(
'cusp::identity_operator< ValueType, cusp::device_memory >M(nrows, ncols)'
)
preconditioners = {
'diagonal': diag,
'jacobi': diag,
'ainv': ainv,
'ainvcusp': ainv,
'amg': amg,
'hypre': amg,
'none': none,
None: none
}
try:
precond_call = preconditioners[parameters['pc_type']]
except KeyError:
raise RuntimeError("Cusp does not support preconditioner type %s" %
parameters['pc_type'])
solvers = {
'cg':
Statement('cusp::krylov::cg(A, x, b, monitor, M)'),
'bicgstab':
Statement('cusp::krylov::bicgstab(A, x, b, monitor, M)'),
'gmres':
Statement(
'cusp::krylov::gmres(A, x, b, %(ksp_gmres_restart)d, monitor, M)' %
parameters)
}
try:
solve_call = solvers[parameters['ksp_type']]
except KeyError:
raise RuntimeError("Cusp does not support solver type %s" %
parameters['ksp_type'])
monitor = 'monitor(b, %(ksp_max_it)d, %(ksp_rtol)g, %(ksp_atol)g)' % parameters
nvcc_function = FunctionBody(
FunctionDeclaration(Value('void', '__cusp_solve'), [
Value('CUdeviceptr', '_rowptr'),
Value('CUdeviceptr', '_colidx'),
Value('CUdeviceptr', '_csrdata'),
Value('CUdeviceptr', '_b'),
Value('CUdeviceptr', '_x'),
Value('int', 'nrows'),
Value('int', 'ncols'),
Value('int', 'nnz')
]),
Block([
Statement('typedef int IndexType'),
Statement('typedef %s ValueType' % M.ctype),
Statement(
'typedef typename cusp::array1d_view< thrust::device_ptr<IndexType> > indices'
),
Statement(
'typedef typename cusp::array1d_view< thrust::device_ptr<ValueType> > values'
),
Statement(
'typedef cusp::csr_matrix_view< indices, indices, values, IndexType, ValueType, cusp::device_memory > matrix'
),
Statement(
'thrust::device_ptr< IndexType > rowptr((IndexType *)_rowptr)'
),
Statement(
'thrust::device_ptr< IndexType > colidx((IndexType *)_colidx)'
),
Statement(
'thrust::device_ptr< ValueType > csrdata((ValueType *)_csrdata)'
),
Statement('thrust::device_ptr< ValueType > d_b((ValueType *)_b)'),
Statement('thrust::device_ptr< ValueType > d_x((ValueType *)_x)'),
Statement('indices row_offsets(rowptr, rowptr + nrows + 1)'),
Statement('indices column_indices(colidx, colidx + nnz)'),
Statement('values matrix_values(csrdata, csrdata + nnz)'),
Statement('values b(d_b, d_b + nrows)'),
Statement('values x(d_x, d_x + ncols)'),
Statement('thrust::fill(x.begin(), x.end(), (ValueType)0)'),
Statement(
'matrix A(nrows, ncols, nnz, row_offsets, column_indices, matrix_values)'
),
Statement('cusp::%s_monitor< ValueType > %s' %
('verbose' if parameters['ksp_monitor'] else 'default',
monitor)), precond_call, solve_call
]))
host_mod.add_to_preamble(
[Include('boost/python/extract.hpp'),
Include('string')])
host_mod.add_to_preamble([Statement('using namespace boost::python')])
host_mod.add_to_preamble([Statement('using namespace std')])
nvcc_mod.add_function(nvcc_function)
host_mod.add_function(
FunctionBody(
FunctionDeclaration(Value('void', 'solve'), [
Value('object', '_rowptr'),
Value('object', '_colidx'),
Value('object', '_csrdata'),
Value('object', '_b'),
Value('object', '_x'),
Value('object', '_nrows'),
Value('object', '_ncols'),
Value('object', '_nnz')
]),
Block([
Statement(
'CUdeviceptr rowptr = extract<CUdeviceptr>(_rowptr.attr("gpudata"))'
),
Statement(
'CUdeviceptr colidx = extract<CUdeviceptr>(_colidx.attr("gpudata"))'
),
Statement(
'CUdeviceptr csrdata = extract<CUdeviceptr>(_csrdata.attr("gpudata"))'
),
Statement(
'CUdeviceptr b = extract<CUdeviceptr>(_b.attr("gpudata"))'
),
Statement(
'CUdeviceptr x = extract<CUdeviceptr>(_x.attr("gpudata"))'
),
Statement('int nrows = extract<int>(_nrows)'),
Statement('int ncols = extract<int>(_ncols)'),
Statement('int nnz = extract<int>(_nnz)'),
Statement(
'__cusp_solve(rowptr, colidx, csrdata, b, x, nrows, ncols, nnz)'
)
])))
nvcc_toolchain.cflags.append('-arch')
nvcc_toolchain.cflags.append('sm_20')
nvcc_toolchain.cflags.append('-O3')
module = nvcc_mod.compile(gcc_toolchain,
nvcc_toolchain,
debug=configuration["debug"])
_cusp_cache[cache_key(M.ctype, parameters)] = module
return module
def make_lift(self, fgroup, with_scale, dtype):
discr = self.discr
from cgen import (FunctionDeclaration, FunctionBody, Typedef, Const,
Reference, Value, POD, Statement, Include, Line,
Block, Initializer, Assign, For, If, Define)
from pytools import to_uncomplex_dtype
from codepy.bpl import BoostPythonModule
mod = BoostPythonModule()
S = Statement
mod.add_to_preamble([
Include("hedge/face_operators.hpp"),
Include("hedge/volume_operators.hpp"),
Include("boost/foreach.hpp"),
])
mod.add_to_module([
S("namespace ublas = boost::numeric::ublas"),
S("using namespace hedge"),
S("using namespace pyublas"),
Line(),
Define("DOFS_PER_EL", fgroup.ldis_loc.node_count()),
Define("FACES_PER_EL", fgroup.ldis_loc.face_count()),
Define("DIMENSIONS", discr.dimensions),
Line(),
Typedef(POD(dtype, "value_type")),
Typedef(POD(to_uncomplex_dtype(dtype), "uncomplex_type")),
])
def if_(cond, result, else_=None):
if cond:
return [result]
else:
if else_ is None:
return []
else:
return [else_]
fdecl = FunctionDeclaration(Value("void", "lift"), [
Const(
Reference(Value("face_group<face_pair<straight_face> >",
"fg"))),
Value("ublas::matrix<uncomplex_type>", "matrix"),
Value("numpy_array<value_type>", "field"),
Value("numpy_array<value_type>", "result")
] + if_(
with_scale,
Const(
Reference(Value("numpy_array<double>",
"elwise_post_scaling")))))
def make_it(name, is_const=True, tpname="value_type"):
if is_const:
const = "const_"
else:
const = ""
return Initializer(
Value("numpy_array<%s>::%siterator" % (tpname, const),
name + "_it"), "%s.begin()" % name)
fbody = Block([
make_it("field"),
make_it("result", is_const=False),
] + if_(with_scale, make_it("elwise_post_scaling", tpname="double")) + [
Line(),
For(
"unsigned fg_el_nr = 0", "fg_el_nr < fg.element_count()",
"++fg_el_nr",
Block([
Initializer(Value("node_number_t", "dest_el_base"),
"fg.local_el_write_base[fg_el_nr]"),
Initializer(Value("node_number_t", "src_el_base"),
"FACES_PER_EL*fg.face_length()*fg_el_nr"),
Line(),
For(
"unsigned i = 0", "i < DOFS_PER_EL", "++i",
Block([
Initializer(Value("value_type", "tmp"), 0),
Line(),
For(
"unsigned j = 0",
"j < FACES_PER_EL*fg.face_length()", "++j",
S("tmp += matrix(i, j)*field_it[src_el_base+j]"
)),
Line(),
] + if_(
with_scale,
Assign(
"result_it[dest_el_base+i]",
"tmp * value_type(*elwise_post_scaling_it)"),
Assign("result_it[dest_el_base+i]", "tmp")))),
] + if_(with_scale, S("elwise_post_scaling_it++"))))
])
mod.add_function(FunctionBody(fdecl, fbody))
#print "----------------------------------------------------------------"
#print FunctionBody(fdecl, fbody)
#raw_input()
return mod.compile(self.discr.toolchain).lift
def get_kernel(self, diff_op, elgroup, for_benchmark=False):
from cgen import \
Pointer, POD, Value, ArrayOf, Const, \
Module, FunctionDeclaration, FunctionBody, Block, \
Comment, Line, Define, Include, \
Initializer, If, For, Statement, Assign
from pycuda.tools import dtype_to_ctype
from cgen.cuda import CudaShared, CudaGlobal
discr = self.discr
d = discr.dimensions
dims = range(d)
plan = self.plan
given = plan.given
elgroup, = discr.element_groups
float_type = given.float_type
f_decl = CudaGlobal(FunctionDeclaration(Value("void", "apply_diff_mat_smem"),
[Pointer(POD(float_type, "debugbuf")), Pointer(POD(float_type, "field")), ]
+ [Pointer(POD(float_type, "drst%d_global" % i)) for i in dims]
))
par = plan.parallelism
cmod = Module([
Include("pycuda-helpers.hpp"),
])
if float_type == numpy.float64:
cmod.append(Value("texture<fp_tex_double, 1, cudaReadModeElementType>",
"diff_rst_mat_tex"))
elif float_type == numpy.float32:
rst_channels = given.devdata.make_valid_tex_channel_count(d)
cmod.append(Value("texture<float%d, 1, cudaReadModeElementType>"
% rst_channels, "diff_rst_mat_tex"))
else:
raise ValueError("unsupported float type: %s" % float_type)
# only preimage size variation is supported here
assert plan.image_dofs_per_el == given.dofs_per_el()
assert plan.aligned_image_dofs_per_microblock == given.microblock.aligned_floats
# FIXME: aligned_image_dofs_per_microblock must be divisible
# by this, therefore hardcoding for now.
chunk_size = 16
cmod.extend([
Line(),
Define("DIMENSIONS", discr.dimensions),
Define("IMAGE_DOFS_PER_EL", plan.image_dofs_per_el),
Define("PREIMAGE_DOFS_PER_EL", plan.preimage_dofs_per_el),
Define("ALIGNED_IMAGE_DOFS_PER_MB", plan.aligned_image_dofs_per_microblock),
Define("ALIGNED_PREIMAGE_DOFS_PER_MB", plan.aligned_preimage_dofs_per_microblock),
Define("ELS_PER_MB", given.microblock.elements),
Define("IMAGE_DOFS_PER_MB", "(IMAGE_DOFS_PER_EL*ELS_PER_MB)"),
Line(),
Define("CHUNK_SIZE", chunk_size),
Define("CHUNK_DOF", "threadIdx.x"),
Define("PAR_MB_NR", "threadIdx.y"),
Define("CHUNK_NR", "threadIdx.z"),
Define("IMAGE_MB_DOF", "(CHUNK_NR*CHUNK_SIZE+CHUNK_DOF)"),
Define("IMAGE_EL_DOF", "(IMAGE_MB_DOF - mb_el*IMAGE_DOFS_PER_EL)"),
Line(),
Define("MACROBLOCK_NR", "blockIdx.x"),
Line(),
Define("PAR_MB_COUNT", par.parallel),
Define("INLINE_MB_COUNT", par.inline),
Define("SEQ_MB_COUNT", par.serial),
Line(),
Define("GLOBAL_MB_NR_BASE",
"(MACROBLOCK_NR*PAR_MB_COUNT*INLINE_MB_COUNT*SEQ_MB_COUNT)"),
Define("GLOBAL_MB_NR",
"(GLOBAL_MB_NR_BASE"
"+ (seq_mb_number*PAR_MB_COUNT + PAR_MB_NR)*INLINE_MB_COUNT)"),
Define("GLOBAL_MB_IMAGE_DOF_BASE", "(GLOBAL_MB_NR*ALIGNED_IMAGE_DOFS_PER_MB)"),
Define("GLOBAL_MB_PREIMAGE_DOF_BASE", "(GLOBAL_MB_NR*ALIGNED_PREIMAGE_DOFS_PER_MB)"),
Line(),
CudaShared(
ArrayOf(
ArrayOf(
ArrayOf(
POD(float_type, "smem_field"),
"PAR_MB_COUNT"),
"INLINE_MB_COUNT"),
"ALIGNED_PREIMAGE_DOFS_PER_MB")),
Line(),
])
S = Statement
f_body = Block([
Initializer(Const(POD(numpy.uint16, "mb_el")),
"IMAGE_MB_DOF / IMAGE_DOFS_PER_EL"),
Line(),
])
# ---------------------------------------------------------------------
def get_load_code():
mb_img_dofs = plan.aligned_image_dofs_per_microblock
mb_preimg_dofs = plan.aligned_preimage_dofs_per_microblock
preimg_dofs_over_dofs = (mb_preimg_dofs+mb_img_dofs-1) // mb_img_dofs
load_code = []
store_code = []
var_num = 0
for load_block in range(preimg_dofs_over_dofs):
for inl in range(par.inline):
# load and store are split for better pipelining
# compiler can't figure that out because of branch
var = "tmp%d" % var_num
var_num += 1
load_code.append(POD(float_type, var))
block_addr = "%d * ALIGNED_IMAGE_DOFS_PER_MB + IMAGE_MB_DOF" % load_block
load_instr = Assign(var,
"field[GLOBAL_MB_PREIMAGE_DOF_BASE"
" + %d*ALIGNED_PREIMAGE_DOFS_PER_MB"
" + %s]" % (inl, block_addr))
store_instr = Assign(
"smem_field[PAR_MB_NR][%d][%s]" % (inl, block_addr),
var
)
if (load_block+1)*mb_img_dofs >= mb_preimg_dofs:
cond = "%s < ALIGNED_PREIMAGE_DOFS_PER_MB" % block_addr
load_instr = If(cond, load_instr)
store_instr = If(cond, store_instr)
load_code.append(load_instr)
store_code.append(store_instr)
return Block(load_code + [Line()] + store_code)
def get_scalar_diff_code():
code = []
for inl in range(par.inline):
for axis in dims:
code.append(
Initializer(POD(float_type, "d%drst%d" % (inl, axis)), 0))
code.append(Line())
tex_channels = ["x", "y", "z", "w"]
store_code = Block()
for inl in range(par.inline):
for rst_axis in dims:
store_code.append(Assign(
"drst%d_global[GLOBAL_MB_IMAGE_DOF_BASE + "
"%d*ALIGNED_IMAGE_DOFS_PER_MB + IMAGE_MB_DOF]"
% (rst_axis, inl),
"d%drst%d" % (inl, rst_axis)
))
from hedge.backends.cuda.tools import unroll
code.extend([
Comment("everybody needs to be done with the old data"),
S("__syncthreads()"),
Line(),
get_load_code(),
Line(),
Comment("all the new data must be loaded"),
S("__syncthreads()"),
Line(),
])
if float_type == numpy.float32:
code.append(Value("float%d" % rst_channels, "dmat_entries"))
code.extend([
POD(float_type, "field_value%d" % inl)
for inl in range(par.inline)
]+[Line()])
def unroll_body(j):
result = [
Assign("field_value%d" % inl,
"smem_field[PAR_MB_NR][%d][mb_el*PREIMAGE_DOFS_PER_EL+%s]" % (inl, j))
for inl in range(par.inline)
]
if float_type == numpy.float32:
result.append(Assign("dmat_entries",
"tex1Dfetch(diff_rst_mat_tex, IMAGE_EL_DOF + %s*IMAGE_DOFS_PER_EL)" % j))
result.extend(
S("d%drst%d += dmat_entries.%s * field_value%d"
% (inl, axis, tex_channels[axis], inl))
for inl in range(par.inline)
for axis in dims)
elif float_type == numpy.float64:
result.extend(
S("d%(inl)drst%(axis)d += "
"fp_tex1Dfetch(diff_rst_mat_tex, %(axis)d "
"+ DIMENSIONS*(IMAGE_EL_DOF + %(j)d*IMAGE_DOFS_PER_EL))"
"* field_value%(inl)d" % {
"inl": inl,
"axis": axis,
"j": j
})
for inl in range(par.inline)
for axis in dims)
else:
assert False
return result
code.append(If("IMAGE_MB_DOF < IMAGE_DOFS_PER_MB", Block(unroll(unroll_body,
total_number=plan.preimage_dofs_per_el)
+[store_code])))
return code
f_body.extend([
For("unsigned short seq_mb_number = 0",
"seq_mb_number < SEQ_MB_COUNT",
"++seq_mb_number",
Block(get_scalar_diff_code())
)
])
# finish off ----------------------------------------------------------
cmod.append(FunctionBody(f_decl, f_body))
if not for_benchmark and "cuda_dump_kernels" in discr.debug:
from hedge.tools import open_unique_debug_file
open_unique_debug_file("diff", ".cu").write(str(cmod))
mod = SourceModule(cmod,
keep="cuda_keep_kernels" in discr.debug,
#options=["--maxrregcount=16"]
)
func = mod.get_function("apply_diff_mat_smem")
if "cuda_diff" in discr.debug:
print "diff: lmem=%d smem=%d regs=%d" % (
func.local_size_bytes,
func.shared_size_bytes,
func.registers)
diff_rst_mat_texref = mod.get_texref("diff_rst_mat_tex")
gpu_diffmats = self.gpu_diffmats(diff_op, elgroup)
if given.float_type == numpy.float32:
gpu_diffmats.bind_to_texref_ext(diff_rst_mat_texref, rst_channels)
elif given.float_type == numpy.float64:
gpu_diffmats.bind_to_texref_ext(diff_rst_mat_texref,
allow_double_hack=True)
else:
assert False
assert given.microblock.aligned_floats % chunk_size == 0
block = (
chunk_size,
plan.parallelism.parallel,
given.microblock.aligned_floats//chunk_size)
func.prepare(
["PP"] + discr.dimensions*["P"],
texrefs=[diff_rst_mat_texref])
return block, func
def get_kernel(self, fdata, ilist_data, for_benchmark):
from cgen.cuda import CudaShared, CudaGlobal
from pycuda.tools import dtype_to_ctype
discr = self.discr
given = self.plan.given
fplan = self.plan
d = discr.dimensions
dims = range(d)
elgroup, = discr.element_groups
float_type = given.float_type
f_decl = CudaGlobal(
FunctionDeclaration(Value("void", "apply_flux"), [
Pointer(POD(float_type, "debugbuf")),
Pointer(POD(numpy.uint8, "gmem_facedata")),
] + [
Pointer(POD(float_type, "gmem_fluxes_on_faces%d" % flux_nr))
for flux_nr in range(len(self.fluxes))
]))
cmod = Module()
cmod.append(Include("pycuda-helpers.hpp"))
for dep_expr in self.all_deps:
cmod.extend([
Value(
"texture<%s, 1, cudaReadModeElementType>" %
dtype_to_ctype(float_type, with_fp_tex_hack=True),
"field%d_tex" % self.dep_to_index[dep_expr])
])
if fplan.flux_count != len(self.fluxes):
from warnings import warn
warn(
"Flux count in flux execution plan different from actual flux count.\n"
"You may want to specify the tune_for= kwarg in the Discretization\n"
"constructor.")
cmod.extend([
Line(),
Typedef(POD(float_type, "value_type")),
Line(),
flux_header_struct(float_type, discr.dimensions),
Line(),
face_pair_struct(float_type, discr.dimensions),
Line(),
Define("DIMENSIONS", discr.dimensions),
Define("DOFS_PER_FACE", fplan.dofs_per_face),
Define("THREADS_PER_FACE", fplan.threads_per_face()),
Line(),
Define("CONCURRENT_FACES", fplan.parallel_faces),
Define("BLOCK_MB_COUNT", fplan.mbs_per_block),
Line(),
Define("FACEDOF_NR", "threadIdx.x"),
Define("BLOCK_FACE", "threadIdx.y"),
Line(),
Define("FLUX_COUNT", len(self.fluxes)),
Line(),
Define("THREAD_NUM", "(FACEDOF_NR + BLOCK_FACE*THREADS_PER_FACE)"),
Define("THREAD_COUNT", "(THREADS_PER_FACE*CONCURRENT_FACES)"),
Define(
"COALESCING_THREAD_COUNT",
"(THREAD_COUNT < 0x10 ? THREAD_COUNT : THREAD_COUNT & ~0xf)"),
Line(),
Define("DATA_BLOCK_SIZE", fdata.block_bytes),
Define("ALIGNED_FACE_DOFS_PER_MB",
fplan.aligned_face_dofs_per_microblock()),
Define("ALIGNED_FACE_DOFS_PER_BLOCK",
"(ALIGNED_FACE_DOFS_PER_MB*BLOCK_MB_COUNT)"),
Line(),
Define("FOF_BLOCK_BASE",
"(blockIdx.x*ALIGNED_FACE_DOFS_PER_BLOCK)"),
Line(),
] + ilist_data.code + [
Line(),
Value("texture<index_list_entry_t, 1, cudaReadModeElementType>",
"tex_index_lists"),
Line(),
fdata.struct,
Line(),
CudaShared(Value("flux_data", "data")),
])
if not fplan.direct_store:
cmod.extend([
CudaShared(
ArrayOf(
ArrayOf(POD(float_type, "smem_fluxes_on_faces"),
"FLUX_COUNT"),
"ALIGNED_FACE_DOFS_PER_MB*BLOCK_MB_COUNT")),
Line(),
])
S = Statement
f_body = Block()
from hedge.backends.cuda.tools import get_load_code
f_body.extend(
get_load_code(dest="&data",
base="gmem_facedata + blockIdx.x*DATA_BLOCK_SIZE",
bytes="sizeof(flux_data)",
descr="load face_pair data") +
[S("__syncthreads()"), Line()])
def get_flux_code(flux_writer):
flux_code = Block([])
flux_code.extend([
Initializer(Pointer(Value("face_pair", "fpair")),
"data.facepairs+fpair_nr"),
Initializer(
MaybeUnused(POD(numpy.uint32, "a_index")),
"fpair->a_base + tex1Dfetch(tex_index_lists, "
"fpair->a_ilist_index + FACEDOF_NR)"),
Initializer(
MaybeUnused(POD(numpy.uint32, "b_index")),
"fpair->b_base + tex1Dfetch(tex_index_lists, "
"fpair->b_ilist_index + FACEDOF_NR)"),
Line(),
flux_writer(),
Line(),
S("fpair_nr += CONCURRENT_FACES")
])
return flux_code
flux_computation = Block([
Comment("fluxes for dual-sided (intra-block) interior face pairs"),
While("fpair_nr < data.header.same_facepairs_end",
get_flux_code(lambda: self.write_interior_flux_code(True))),
Line(),
Comment("work around nvcc assertion failure"),
S("fpair_nr+=1"),
S("fpair_nr-=1"),
Line(),
Comment(
"fluxes for single-sided (inter-block) interior face pairs"),
While("fpair_nr < data.header.diff_facepairs_end",
get_flux_code(lambda: self.write_interior_flux_code(False))),
Line(),
Comment("fluxes for single-sided boundary face pairs"),
While(
"fpair_nr < data.header.bdry_facepairs_end",
get_flux_code(
lambda: self.write_boundary_flux_code(for_benchmark))),
])
f_body.extend_log_block("compute the fluxes", [
Initializer(POD(numpy.uint32, "fpair_nr"), "BLOCK_FACE"),
If("FACEDOF_NR < DOFS_PER_FACE", flux_computation)
])
if not fplan.direct_store:
f_body.extend([Line(), S("__syncthreads()"), Line()])
f_body.extend_log_block(
"store fluxes",
[
#Assign("debugbuf[blockIdx.x]", "FOF_BLOCK_BASE"),
#Assign("debugbuf[0]", "FOF_BLOCK_BASE"),
#Assign("debugbuf[0]", "sizeof(face_pair)"),
For(
"unsigned word_nr = THREAD_NUM",
"word_nr < ALIGNED_FACE_DOFS_PER_MB*BLOCK_MB_COUNT",
"word_nr += COALESCING_THREAD_COUNT",
Block([
Assign(
"gmem_fluxes_on_faces%d[FOF_BLOCK_BASE+word_nr]"
% flux_nr,
"smem_fluxes_on_faces[%d][word_nr]" % flux_nr)
for flux_nr in range(len(self.fluxes))
]
#+[If("isnan(smem_fluxes_on_faces[%d][word_nr])" % flux_nr,
#Block([
#Assign("debugbuf[blockIdx.x]", "word_nr"),
#])
#)
#for flux_nr in range(len(self.fluxes))]
))
])
if False:
f_body.extend([
Assign("debugbuf[blockIdx.x*96+32+BLOCK_FACE*32+threadIdx.x]",
"fpair_nr"),
Assign("debugbuf[blockIdx.x*96+16]",
"data.header.same_facepairs_end"),
Assign("debugbuf[blockIdx.x*96+17]",
"data.header.diff_facepairs_end"),
Assign("debugbuf[blockIdx.x*96+18]",
"data.header.bdry_facepairs_end"),
])
# finish off ----------------------------------------------------------
cmod.append(FunctionBody(f_decl, f_body))
if not for_benchmark and "cuda_dump_kernels" in discr.debug:
from hedge.tools import open_unique_debug_file
open_unique_debug_file("flux_gather", ".cu").write(str(cmod))
#from pycuda.tools import allow_user_edit
mod = SourceModule(
#allow_user_edit(cmod, "kernel.cu", "the flux kernel"),
cmod,
keep="cuda_keep_kernels" in discr.debug)
expr_to_texture_map = dict(
(dep_expr,
mod.get_texref("field%d_tex" % self.dep_to_index[dep_expr]))
for dep_expr in self.all_deps)
index_list_texref = mod.get_texref("tex_index_lists")
index_list_texref.set_address(ilist_data.device_memory,
ilist_data.bytes)
index_list_texref.set_format(
cuda.dtype_to_array_format(ilist_data.type), 1)
index_list_texref.set_flags(cuda.TRSF_READ_AS_INTEGER)
func = mod.get_function("apply_flux")
block = (fplan.threads_per_face(), fplan.parallel_faces, 1)
func.prepare(
(2 + len(self.fluxes)) * "P",
texrefs=expr_to_texture_map.values() + [index_list_texref])
if "cuda_flux" in discr.debug:
print "flux: lmem=%d smem=%d regs=%d" % (
func.local_size_bytes, func.shared_size_bytes, func.num_regs)
return block, func, expr_to_texture_map
def get_boundary_flux_mod(fluxes, fvi, discr, dtype):
from cgen import \
FunctionDeclaration, FunctionBody, Typedef, Struct, \
Const, Reference, Value, POD, MaybeUnused, \
Statement, Include, Line, Block, Initializer, Assign, \
CustomLoop, For
from pytools import to_uncomplex_dtype, flatten
from codepy.bpl import BoostPythonModule
mod = BoostPythonModule()
mod.add_to_preamble([
Include("cstdlib"),
Include("algorithm"),
Line(),
Include("boost/foreach.hpp"),
Line(),
Include("hedge/face_operators.hpp"),
])
S = Statement
mod.add_to_module([
S("using namespace hedge"),
S("using namespace pyublas"),
Line(),
Typedef(POD(dtype, "value_type")),
Typedef(POD(to_uncomplex_dtype(dtype), "uncomplex_type")),
])
arg_struct = Struct("arg_struct", [
Value("numpy_array<value_type>", "flux%d_on_faces" % i)
for i in range(len(fluxes))
]+[
Value("numpy_array<value_type>", arg_name)
for arg_name in fvi.arg_names
])
mod.add_struct(arg_struct, "ArgStruct")
mod.add_to_module([Line()])
fdecl = FunctionDeclaration(
Value("void", "gather_flux"),
[
Const(Reference(Value("face_group<face_pair<straight_face> >" , "fg"))),
Reference(Value("arg_struct", "args"))
])
from pymbolic.mapper.stringifier import PREC_PRODUCT
def gen_flux_code():
f2cm = FluxToCodeMapper()
result = [
Assign("fof%d_it[loc_fof_base+i]" % flux_idx,
"uncomplex_type(fp.int_side.face_jacobian) * " +
flux_to_code(f2cm, False, flux_idx, fvi, flux.op.flux, PREC_PRODUCT))
for flux_idx, flux in enumerate(fluxes)
]
return [
Initializer(Value("value_type", cse_name), cse_str)
for cse_name, cse_str in f2cm.cse_name_list] + result
fbody = Block([
Initializer(
Const(Value("numpy_array<value_type>::iterator", "fof%d_it" % i)),
"args.flux%d_on_faces.begin()" % i)
for i in range(len(fluxes))
]+[
Initializer(
Const(Value("numpy_array<value_type>::const_iterator",
"%s_it" % arg_name)),
"args.%s.begin()" % arg_name)
for arg_name in fvi.arg_names
]+[
Line(),
CustomLoop("BOOST_FOREACH(const face_pair<straight_face> &fp, fg.face_pairs)", Block(
list(flatten([
Initializer(Value("node_number_t", "%s_ebi" % where),
"fp.%s.el_base_index" % where),
Initializer(Value("index_lists_t::const_iterator", "%s_idx_list" % where),
"fg.index_list(fp.%s.face_index_list_number)" % where),
Line(),
]
for where in ["int_side", "ext_side"]
))+[
Line(),
Initializer(Value("node_number_t", "loc_fof_base"),
"fg.face_length()*(fp.%(where)s.local_el_number*fg.face_count"
" + fp.%(where)s.face_id)" % {"where": "int_side"}),
Line(),
For(
"unsigned i = 0",
"i < fg.face_length()",
"++i",
Block(
[
Initializer(MaybeUnused(
Value("node_number_t", "%s_idx" % where)),
"%(where)s_ebi + %(where)s_idx_list[i]"
% {"where": where})
for where in ["int_side", "ext_side"]
]+gen_flux_code()
)
)
]))
])
mod.add_function(FunctionBody(fdecl, fbody))
#print "----------------------------------------------------------------"
#print mod.generate()
#raw_input("[Enter]")
return mod.compile(get_flux_toolchain(discr, fluxes))
def get_kernel(self, diff_op_cls, elgroup, for_benchmark=False):
from cgen import \
Pointer, POD, Value, ArrayOf, \
Module, FunctionDeclaration, FunctionBody, Block, \
Line, Define, Include, \
Initializer, If, For, Statement, Assign
from cgen import dtype_to_ctype
from cgen.cuda import CudaShared, CudaGlobal
discr = self.discr
d = discr.dimensions
dims = range(d)
given = self.plan.given
par = self.plan.parallelism
diffmat_data = self.gpu_diffmats(diff_op_cls, elgroup)
elgroup, = discr.element_groups
float_type = given.float_type
f_decl = CudaGlobal(
FunctionDeclaration(
Value("void", "apply_diff_mat"),
[
Pointer(POD(numpy.uint8, "gmem_diff_rst_mat")),
#Pointer(POD(float_type, "debugbuf")),
] +
[Pointer(POD(float_type, "drst%d_global" % i)) for i in dims]))
rst_channels = given.devdata.make_valid_tex_channel_count(d)
cmod = Module([
Include("pycuda-helpers.hpp"),
Line(),
Value(
"texture<fp_tex_%s, 1, cudaReadModeElementType>" %
dtype_to_ctype(float_type), "field_tex"),
Line(),
Define("DIMENSIONS", discr.dimensions),
Define("DOFS_PER_EL", given.dofs_per_el()),
Line(),
Define("SEGMENT_DOF", "threadIdx.x"),
Define("PAR_MB_NR", "threadIdx.y"),
Line(),
Define("MB_SEGMENT", "blockIdx.x"),
Define("MACROBLOCK_NR", "blockIdx.y"),
Line(),
Define("DOFS_PER_SEGMENT", self.plan.segment_size),
Define("SEGMENTS_PER_MB", self.plan.segments_per_microblock()),
Define("ALIGNED_DOFS_PER_MB", given.microblock.aligned_floats),
Define("ELS_PER_MB", given.microblock.elements),
Line(),
Define("PAR_MB_COUNT", par.parallel),
Define("INLINE_MB_COUNT", par.inline),
Define("SEQ_MB_COUNT", par.serial),
Line(),
Define("THREAD_NUM", "(SEGMENT_DOF+PAR_MB_NR*DOFS_PER_SEGMENT)"),
Define("COALESCING_THREAD_COUNT",
"(PAR_MB_COUNT*DOFS_PER_SEGMENT)"),
Line(),
Define("MB_DOF_BASE", "(MB_SEGMENT*DOFS_PER_SEGMENT)"),
Define("MB_DOF", "(MB_DOF_BASE+SEGMENT_DOF)"),
Define(
"GLOBAL_MB_NR_BASE",
"(MACROBLOCK_NR*PAR_MB_COUNT*INLINE_MB_COUNT*SEQ_MB_COUNT)"),
Define(
"GLOBAL_MB_NR", "(GLOBAL_MB_NR_BASE"
"+ (seq_mb_number*PAR_MB_COUNT + PAR_MB_NR)*INLINE_MB_COUNT)"),
Define("GLOBAL_MB_DOF_BASE", "(GLOBAL_MB_NR*ALIGNED_DOFS_PER_MB)"),
Line(),
Define("DIFFMAT_SEGMENT_FLOATS", diffmat_data.block_floats),
Define("DIFFMAT_SEGMENT_BYTES",
"(DIFFMAT_SEGMENT_FLOATS*%d)" % given.float_size()),
Define("DIFFMAT_COLUMNS", diffmat_data.matrix_columns),
Line(),
CudaShared(
ArrayOf(POD(float_type, "smem_diff_rst_mat"),
"DIFFMAT_COLUMNS*DOFS_PER_SEGMENT")),
Line(),
])
S = Statement
f_body = Block()
f_body.extend_log_block("calculate responsibility data", [
Initializer(POD(numpy.uint16, "mb_el"), "MB_DOF/DOFS_PER_EL"),
])
from hedge.backends.cuda.tools import get_load_code
f_body.extend(
get_load_code(
dest="smem_diff_rst_mat",
base="gmem_diff_rst_mat + MB_SEGMENT*DIFFMAT_SEGMENT_BYTES",
bytes="DIFFMAT_SEGMENT_BYTES",
descr="load diff mat segment") +
[S("__syncthreads()"), Line()])
# ---------------------------------------------------------------------
def get_scalar_diff_code():
code = []
for inl in range(par.inline):
for axis in dims:
code.append(
Initializer(POD(float_type, "d%drst%d" % (inl, axis)),
0))
code.append(Line())
def get_mat_entry(row, col, axis):
return ("smem_diff_rst_mat["
"%(row)s*DIFFMAT_COLUMNS + %(axis)s*DOFS_PER_EL"
" + %(col)s"
"]" % {
"row": row,
"col": col,
"axis": axis
})
tex_channels = ["x", "y", "z", "w"]
from hedge.backends.cuda.tools import unroll
code.extend([
POD(float_type, "field_value%d" % inl)
for inl in range(par.inline)
] + [Line()] + unroll(
lambda j: [
Assign(
"field_value%d" % inl,
"fp_tex1Dfetch(field_tex, GLOBAL_MB_DOF_BASE + %d*ALIGNED_DOFS_PER_MB "
"+ mb_el*DOFS_PER_EL + %s)" % (inl, j))
for inl in range(par.inline)
] + [Line()] + [
S("d%drst%d += %s * field_value%d" %
(inl, axis, get_mat_entry("SEGMENT_DOF", j, axis), inl))
for axis in dims for inl in range(par.inline)
] + [Line()], given.dofs_per_el(), self.plan.max_unroll))
store_code = Block()
for inl in range(par.inline):
for rst_axis in dims:
store_code.append(
Assign(
"drst%d_global[GLOBAL_MB_DOF_BASE"
" + %d*ALIGNED_DOFS_PER_MB + MB_DOF]" %
(rst_axis, inl),
"d%drst%d" % (inl, rst_axis),
))
code.append(If("MB_DOF < DOFS_PER_EL*ELS_PER_MB", store_code))
return code
f_body.extend([
For("unsigned short seq_mb_number = 0",
"seq_mb_number < SEQ_MB_COUNT", "++seq_mb_number",
Block(get_scalar_diff_code()))
])
# finish off ----------------------------------------------------------
cmod.append(FunctionBody(f_decl, f_body))
if not for_benchmark and "cuda_dump_kernels" in discr.debug:
from hedge.tools import open_unique_debug_file
open_unique_debug_file("diff", ".cu").write(str(cmod))
mod = SourceModule(
cmod,
keep="cuda_keep_kernels" in discr.debug,
#options=["--maxrregcount=10"]
)
field_texref = mod.get_texref("field_tex")
func = mod.get_function("apply_diff_mat")
func.prepare(discr.dimensions * [float_type] + ["P"],
block=(self.plan.segment_size, par.parallel, 1),
texrefs=[field_texref])
if "cuda_diff" in discr.debug:
print "diff: lmem=%d smem=%d regs=%d" % (
func.local_size_bytes, func.shared_size_bytes, func.num_regs)
return func, field_texref
def get_kernel(self, with_scaling, for_benchmark=False):
from cgen import \
Pointer, POD, Value, ArrayOf, \
Module, FunctionDeclaration, FunctionBody, Block, \
Line, Define, Include, \
Initializer, If, For, Statement, Assign, \
ArrayInitializer
from cgen import dtype_to_ctype
from cgen.cuda import CudaShared, CudaConstant, CudaGlobal
discr = self.discr
d = discr.dimensions
dims = range(d)
given = self.plan.given
float_type = given.float_type
f_decl = CudaGlobal(
FunctionDeclaration(Value("void", "apply_el_local_mat_smem_mat"), [
Pointer(POD(float_type, "out_vector")),
Pointer(POD(numpy.uint8, "gmem_matrix")),
Pointer(POD(float_type, "debugbuf")),
POD(numpy.uint32, "microblock_count"),
]))
cmod = Module([
Include("pycuda-helpers.hpp"),
Line(),
Value(
"texture<fp_tex_%s, 1, cudaReadModeElementType>" %
dtype_to_ctype(float_type), "in_vector_tex"),
])
if with_scaling:
cmod.append(
Value(
"texture<fp_tex_%s, 1, cudaReadModeElementType>" %
dtype_to_ctype(float_type), "scaling_tex"), )
par = self.plan.parallelism
cmod.extend([
Line(),
Define("DIMENSIONS", discr.dimensions),
Define("DOFS_PER_EL", given.dofs_per_el()),
Define("PREIMAGE_DOFS_PER_EL", self.plan.preimage_dofs_per_el),
Line(),
Define("SEGMENT_DOF", "threadIdx.x"),
Define("PAR_MB_NR", "threadIdx.y"),
Line(),
Define("MB_SEGMENT", "blockIdx.x"),
Define("MACROBLOCK_NR", "blockIdx.y"),
Line(),
Define("DOFS_PER_SEGMENT", self.plan.segment_size),
Define("SEGMENTS_PER_MB", self.plan.segments_per_microblock()),
Define("ALIGNED_DOFS_PER_MB", given.microblock.aligned_floats),
Define("ALIGNED_PREIMAGE_DOFS_PER_MB",
self.plan.aligned_preimage_dofs_per_microblock),
Define("MB_EL_COUNT", given.microblock.elements),
Line(),
Define("PAR_MB_COUNT", par.parallel),
Define("INLINE_MB_COUNT", par.inline),
Define("SEQ_MB_COUNT", par.serial),
Line(),
Define("THREAD_NUM", "(SEGMENT_DOF+PAR_MB_NR*DOFS_PER_SEGMENT)"),
Define("COALESCING_THREAD_COUNT",
"(PAR_MB_COUNT*DOFS_PER_SEGMENT)"),
Line(),
Define("MB_DOF_BASE", "(MB_SEGMENT*DOFS_PER_SEGMENT)"),
Define("MB_DOF", "(MB_DOF_BASE+SEGMENT_DOF)"),
Define(
"GLOBAL_MB_NR_BASE",
"(MACROBLOCK_NR*PAR_MB_COUNT*INLINE_MB_COUNT*SEQ_MB_COUNT)"),
Define(
"GLOBAL_MB_NR", "(GLOBAL_MB_NR_BASE"
"+ (seq_mb_number*PAR_MB_COUNT + PAR_MB_NR)*INLINE_MB_COUNT)"),
Define("GLOBAL_MB_DOF_BASE", "(GLOBAL_MB_NR*ALIGNED_DOFS_PER_MB)"),
Define("GLOBAL_MB_PREIMG_DOF_BASE",
"(GLOBAL_MB_NR*ALIGNED_PREIMAGE_DOFS_PER_MB)"),
Line(),
Define("MATRIX_COLUMNS", self.plan.gpu_matrix_columns()),
Define("MATRIX_SEGMENT_FLOATS",
self.plan.gpu_matrix_block_floats()),
Define("MATRIX_SEGMENT_BYTES",
"(MATRIX_SEGMENT_FLOATS*%d)" % given.float_size()),
Line(),
CudaShared(
ArrayOf(POD(float_type, "smem_matrix"),
"MATRIX_SEGMENT_FLOATS")),
CudaShared(
ArrayOf(
ArrayOf(
ArrayOf(POD(float_type, "dof_buffer"), "PAR_MB_COUNT"),
"INLINE_MB_COUNT"), "DOFS_PER_SEGMENT"), ),
CudaShared(POD(numpy.uint16, "segment_start_el")),
CudaShared(POD(numpy.uint16, "segment_stop_el")),
CudaShared(POD(numpy.uint16, "segment_el_count")),
Line(),
ArrayInitializer(
CudaConstant(
ArrayOf(POD(numpy.uint32, "segment_start_el_lookup"),
"SEGMENTS_PER_MB")),
[(chk * self.plan.segment_size) // given.dofs_per_el()
for chk in range(self.plan.segments_per_microblock())]),
ArrayInitializer(
CudaConstant(
ArrayOf(POD(numpy.uint32, "segment_stop_el_lookup"),
"SEGMENTS_PER_MB")),
[
min(given.microblock.elements,
(chk * self.plan.segment_size +
self.plan.segment_size - 1) // given.dofs_per_el() +
1)
for chk in range(self.plan.segments_per_microblock())
]),
])
S = Statement
f_body = Block()
f_body.extend_log_block(
"calculate this dof's element",
[Initializer(POD(numpy.uint8, "mb_el"), "MB_DOF/DOFS_PER_EL")])
if self.plan.use_prefetch_branch:
f_body.extend_log_block("calculate segment responsibility data", [
If(
"THREAD_NUM==0",
Block([
Assign("segment_start_el",
"segment_start_el_lookup[MB_SEGMENT]"),
Assign("segment_stop_el",
"segment_stop_el_lookup[MB_SEGMENT]"),
Assign("segment_el_count",
"segment_stop_el-segment_start_el"),
])),
S("__syncthreads()")
])
from hedge.backends.cuda.tools import get_load_code
f_body.extend(
get_load_code(dest="smem_matrix",
base=(
"gmem_matrix + MB_SEGMENT*MATRIX_SEGMENT_BYTES"),
bytes="MATRIX_SEGMENT_BYTES",
descr="load matrix segment") +
[S("__syncthreads()")])
# ---------------------------------------------------------------------
def get_batched_fetch_mat_mul_code(el_fetch_count):
result = []
dofs = range(self.plan.preimage_dofs_per_el)
for load_segment_start in range(0, self.plan.preimage_dofs_per_el,
self.plan.segment_size):
result.extend([S("__syncthreads()")] + [
Assign(
"dof_buffer[PAR_MB_NR][%d][SEGMENT_DOF]" %
inl, "fp_tex1Dfetch(in_vector_tex, "
"GLOBAL_MB_PREIMG_DOF_BASE"
" + %d*ALIGNED_PREIMAGE_DOFS_PER_MB"
" + (segment_start_el)*PREIMAGE_DOFS_PER_EL + %d + SEGMENT_DOF)"
% (inl, load_segment_start))
for inl in range(par.inline)
] + [
S("__syncthreads()"),
Line(),
])
for dof in dofs[load_segment_start:load_segment_start +
self.plan.segment_size]:
for inl in range(par.inline):
result.append(
S("result%d += "
"smem_matrix[SEGMENT_DOF*MATRIX_COLUMNS + %d]"
"*"
"dof_buffer[PAR_MB_NR][%d][%d]" %
(inl, dof, inl, dof - load_segment_start)))
result.append(Line())
return result
from hedge.backends.cuda.tools import unroll
def get_direct_tex_mat_mul_code():
return (
[POD(float_type, "fof%d" % inl) for inl in range(par.inline)] +
[POD(float_type, "lm"), Line()] + unroll(
lambda j: [
Assign(
"fof%d" % inl,
"fp_tex1Dfetch(in_vector_tex, "
"GLOBAL_MB_PREIMG_DOF_BASE"
" + %(inl)d * ALIGNED_PREIMAGE_DOFS_PER_MB"
" + mb_el*PREIMAGE_DOFS_PER_EL+%(j)s)" % {
"j": j,
"inl": inl,
"row": "SEGMENT_DOF"
},
) for inl in range(par.inline)
] + [
Assign(
"lm",
"smem_matrix["
"%(row)s*MATRIX_COLUMNS + %(j)s]" % {
"j": j,
"row": "SEGMENT_DOF"
},
)
] + [
S("result%(inl)d += fof%(inl)d*lm" % {"inl": inl})
for inl in range(par.inline)
],
total_number=self.plan.preimage_dofs_per_el,
max_unroll=self.plan.max_unroll) + [Line()])
def get_mat_mul_code(el_fetch_count):
if el_fetch_count == 1:
return get_batched_fetch_mat_mul_code(el_fetch_count)
else:
return get_direct_tex_mat_mul_code()
def mat_mul_outer_loop(fetch_count):
if with_scaling:
inv_jac_multiplier = (
"fp_tex1Dfetch(scaling_tex,"
"(GLOBAL_MB_NR + %(inl)d)*MB_EL_COUNT + mb_el)")
else:
inv_jac_multiplier = "1"
write_condition = "MB_DOF < DOFS_PER_EL*MB_EL_COUNT"
if self.with_index_check:
write_condition += " && GLOBAL_MB_NR < microblock_count"
return For(
"unsigned short seq_mb_number = 0",
"seq_mb_number < SEQ_MB_COUNT", "++seq_mb_number",
Block([
Initializer(POD(float_type, "result%d" % inl), 0)
for inl in range(par.inline)
] + [Line()] + get_mat_mul_code(fetch_count) + [
If(
write_condition,
Block([
Assign(
"out_vector[GLOBAL_MB_DOF_BASE"
" + %d*ALIGNED_DOFS_PER_MB"
" + MB_DOF]" % inl, "result%d * %s" %
(inl, (inv_jac_multiplier % {
"inl": inl
}))) for inl in range(par.inline)
]))
]))
if self.plan.use_prefetch_branch:
from cgen import make_multiple_ifs
f_body.append(
make_multiple_ifs([
("segment_el_count == %d" % fetch_count,
mat_mul_outer_loop(fetch_count)) for fetch_count in range(
1,
self.plan.max_elements_touched_by_segment() + 1)
]))
else:
f_body.append(mat_mul_outer_loop(0))
# finish off ----------------------------------------------------------
cmod.append(FunctionBody(f_decl, f_body))
if not for_benchmark and "cuda_dump_kernels" in discr.debug:
from hedge.tools import open_unique_debug_file
open_unique_debug_file(self.plan.debug_name,
".cu").write(str(cmod))
mod = SourceModule(
cmod,
keep="cuda_keep_kernels" in discr.debug,
#options=["--maxrregcount=12"]
)
func = mod.get_function("apply_el_local_mat_smem_mat")
if self.plan.debug_name in discr.debug:
print "%s: lmem=%d smem=%d regs=%d" % (
self.plan.debug_name, func.local_size_bytes,
func.shared_size_bytes, func.num_regs)
in_vector_texref = mod.get_texref("in_vector_tex")
texrefs = [in_vector_texref]
if with_scaling:
scaling_texref = mod.get_texref("scaling_tex")
texrefs.append(scaling_texref)
else:
scaling_texref = None
func.prepare("PPPI",
block=(self.plan.segment_size,
self.plan.parallelism.parallel, 1),
texrefs=texrefs)
return func, in_vector_texref, scaling_texref
def make_cuda_kernel(self, discr, dtype, eg):
given = discr.given
ldis = eg.local_discretization
microblocks_per_block = 1
from cgen.cuda import CudaGlobal
from cgen import (Module, Value, Include,
Typedef, FunctionBody, FunctionDeclaration, Const,
Line, POD, LiteralBlock,
Define, Pointer)
cmod = Module([
Include("pycuda-helpers.hpp"),
Line(),
Typedef(POD(dtype, "value_type")),
Line(),
Define("DOFS_PER_EL", given.dofs_per_el()),
Define("ALIGNED_DOFS_PER_MB", given.microblock.aligned_floats),
Define("VERTICES_PER_EL", ldis.vertex_count()),
Define("ELS_PER_MB", given.microblock.elements),
Define("MBS_PER_BLOCK", microblocks_per_block),
Line(),
Define("DOF_IN_MB_IDX", "threadIdx.x"),
Define("DOF_IN_EL_IDX", "(DOF_IN_MB_IDX-el_idx_in_mb*DOFS_PER_EL)"),
Define("MB_IN_BLOCK_IDX", "threadIdx.y"),
Define("BLOCK_IDX", "blockIdx.x"),
Define("MB_NUMBER", "(BLOCK_IDX * MBS_PER_BLOCK + MB_IN_BLOCK_IDX)"),
Define("BLOCK_DATA", "whole_block[MB_IN_BLOCK_IDX]")]
+ self.get_cuda_extra_preamble(discr, dtype, eg)
+ [FunctionBody(
CudaGlobal(FunctionDeclaration(
Value("void", "elwise_kernel"), [
Pointer(Const(POD(dtype, "field"))),
Pointer(POD(dtype, "result")),
POD(numpy.uint32, "mb_count"),
])),
LiteralBlock("""
int el_idx_in_mb = DOF_IN_MB_IDX / DOFS_PER_EL;
if (MB_NUMBER >= mb_count)
return;
int idx = MB_NUMBER * ALIGNED_DOFS_PER_MB + DOF_IN_MB_IDX;
int element_base_idx = ALIGNED_DOFS_PER_MB * MB_IN_BLOCK_IDX +
(DOF_IN_MB_IDX / DOFS_PER_EL) * DOFS_PER_EL;
int dof_in_element = DOF_IN_MB_IDX-el_idx_in_mb*DOFS_PER_EL;
__shared__ value_type whole_block[MBS_PER_BLOCK][ALIGNED_DOFS_PER_MB+1];
int idx_in_block = ALIGNED_DOFS_PER_MB * MB_IN_BLOCK_IDX + DOF_IN_MB_IDX;
BLOCK_DATA[idx_in_block] = field[idx];
__syncthreads();
%s
result[idx] = node_result;
""" % self.get_cuda_code(discr, dtype, eg)))
])
if False:
for i, l in enumerate(str(cmod).split("\n")):
print i+1, l
raw_input()
from pycuda.compiler import SourceModule
mod = SourceModule(
cmod,
keep="cuda_keep_kernels" in discr.debug,
)
func = mod.get_function("elwise_kernel")
func.prepare(
"PPI", block=(
given.microblock.aligned_floats,
microblocks_per_block, 1))
mb_count = len(discr.blocks) * discr.given.microblocks_per_block
grid_dim = (mb_count + microblocks_per_block - 1) \
// microblocks_per_block
from pytools import Record
class KernelInfo(Record):
pass
return KernelInfo(
func=func,
grid_dim=grid_dim,
mb_count=mb_count)
Pointer, Module, Block, Initializer, Assign, Const
from cgen.opencl import CLKernel, CLGlobal, \
CLRequiredWorkGroupSize
mod = Module([
FunctionBody(
CLKernel(
CLRequiredWorkGroupSize(
(local_size, ),
FunctionDeclaration(Value("void", "add"),
arg_decls=[
CLGlobal(
Pointer(Const(POD(dtype, name))))
for name in ["tgt", "op1", "op2"]
]))),
Block([
Initializer(
POD(numpy.int32,
"idx"), "get_local_id(0) + %d * get_group_id(0)" %
(local_size * thread_strides))
] + [
Assign(
"tgt[idx+%d]" % (o * local_size), "op1[idx+%d] + op2[idx+%d]" %
(o * local_size, o * local_size))
for o in range(thread_strides)
]))
])
knl = cl.Program(ctx, str(mod)).build().add
knl(queue, (local_size * macroblock_count, ), (local_size, ), c_buf, a_buf,
def make_diff(self, elgroup, dtype, shape):
"""
:param shape: If non-square, the resulting code takes two element_ranges
arguments and supports non-square matrices.
"""
from hedge._internal import UniformElementRanges
assert isinstance(elgroup.ranges, UniformElementRanges)
ldis = elgroup.local_discretization
discr = self.discr
from cgen import (
FunctionDeclaration, FunctionBody, Typedef,
Const, Reference, Value, POD,
Statement, Include, Line, Block, Initializer, Assign,
For, If,
Define)
from pytools import to_uncomplex_dtype
from codepy.bpl import BoostPythonModule
mod = BoostPythonModule()
# {{{ preamble
S = Statement
mod.add_to_preamble([
Include("hedge/volume_operators.hpp"),
Include("boost/foreach.hpp"),
])
mod.add_to_module([
S("namespace ublas = boost::numeric::ublas"),
S("using namespace hedge"),
S("using namespace pyublas"),
Line(),
Define("ROW_COUNT", shape[0]),
Define("COL_COUNT", shape[1]),
Define("DIMENSIONS", discr.dimensions),
Line(),
Typedef(POD(dtype, "value_type")),
Typedef(POD(to_uncomplex_dtype(dtype), "uncomplex_type")),
])
fdecl = FunctionDeclaration(
Value("void", "diff"),
[
Const(Reference(Value("uniform_element_ranges", "from_ers"))),
Const(Reference(Value("uniform_element_ranges", "to_ers"))),
Value("numpy_array<value_type>", "field")
]+[
Value("ublas::matrix<uncomplex_type>", "diffmat_rst%d" % rst)
for rst in range(discr.dimensions)
]+[
Value("numpy_array<value_type>", "result%d" % i)
for i in range(discr.dimensions)
]
)
# }}}
# {{{ set-up
def make_it(name, is_const=True, tpname="value_type"):
if is_const:
const = "const_"
else:
const = ""
return Initializer(
Value("numpy_array<%s>::%siterator" % (tpname, const), name+"_it"),
"%s.begin()" % name)
fbody = Block([
If("ROW_COUNT != diffmat_rst%d.size1()" % i,
S('throw(std::runtime_error("unexpected matrix size"))'))
for i in range(discr.dimensions)
] + [
If("COL_COUNT != diffmat_rst%d.size2()" % i,
S('throw(std::runtime_error("unexpected matrix size"))'))
for i in range(discr.dimensions)
]+[
If("ROW_COUNT != to_ers.el_size()",
S('throw(std::runtime_error("unsupported image element size"))')),
If("COL_COUNT != from_ers.el_size()",
S('throw(std::runtime_error("unsupported preimage element size"))')),
If("from_ers.size() != to_ers.size()",
S('throw(std::runtime_error("image and preimage element groups '
'do nothave the same element count"))')),
Line(),
make_it("field"),
]+[
make_it("result%d" % i, is_const=False)
for i in range(discr.dimensions)
]+[
Line(),
# }}}
# {{{ computation
For("element_number_t eg_el_nr = 0",
"eg_el_nr < to_ers.size()",
"++eg_el_nr",
Block([
Initializer(
Value("node_number_t", "from_el_base"),
"from_ers.start() + eg_el_nr*COL_COUNT"),
Initializer(
Value("node_number_t", "to_el_base"),
"to_ers.start() + eg_el_nr*ROW_COUNT"),
Line(),
For("unsigned i = 0",
"i < ROW_COUNT",
"++i",
Block([
Initializer(Value("value_type", "drst_%d" % rst), 0)
for rst in range(discr.dimensions)
]+[
Line(),
]+[
For("unsigned j = 0",
"j < COL_COUNT",
"++j",
Block([
S("drst_%(rst)d += "
"diffmat_rst%(rst)d(i, j)*field_it[from_el_base+j]"
% {"rst":rst})
for rst in range(discr.dimensions)
])
),
Line(),
]+[
Assign("result%d_it[to_el_base+i]" % rst,
"drst_%d" % rst)
for rst in range(discr.dimensions)
])
)
])
)
])
# }}}
# {{{ compilation
mod.add_function(FunctionBody(fdecl, fbody))
#print "----------------------------------------------------------------"
#print mod.generate()
#raw_input()
compiled_func = mod.compile(self.discr.toolchain).diff
if self.discr.instrumented:
from hedge.tools import time_count_flop
compiled_func = time_count_flop(compiled_func,
discr.diff_timer, discr.diff_counter,
discr.diff_flop_counter,
flops=discr.dimensions*(
2 # mul+add
* ldis.node_count() * len(elgroup.members)
* ldis.node_count()
+
2 * discr.dimensions
* len(elgroup.members) * ldis.node_count()),
increment=discr.dimensions)
return compiled_func