def expansion(node: 'Reduce', state: SDFGState, sdfg: SDFG):
node.validate(sdfg, state)
input_edge: graph.MultiConnectorEdge = state.in_edges(node)[0]
output_edge: graph.MultiConnectorEdge = state.out_edges(node)[0]
input_dims = len(input_edge.data.subset)
input_data = sdfg.arrays[input_edge.data.data]
output_data = sdfg.arrays[output_edge.data.data]
# Setup all locations in which code will be written
cuda_globalcode = CodeIOStream()
localcode = CodeIOStream()
# Try to autodetect reduction type
redtype = detect_reduction_type(node.wcr)
node_id = state.node_id(node)
state_id = sdfg.node_id(state)
idstr = '{sdfg}_{state}_{node}'.format(sdfg=sdfg.name,
state=state_id,
node=node_id)
# Obtain some SDFG-related information
input_memlet = input_edge.data
output_memlet = output_edge.data
if node.out_connectors:
dtype = next(node.out_connectors.values())
else:
dtype = sdfg.arrays[output_memlet.data].dtype
output_type = dtype.ctype
if node.identity is None:
raise ValueError('For device reduce nodes, initial value must be '
'specified')
# Create a functor or use an existing one for reduction
if redtype == dtypes.ReductionType.Custom:
body, [arg1, arg2] = unparse_cr_split(sdfg, node.wcr)
cuda_globalcode.write(
"""
struct __reduce_{id} {{
template <typename T>
DACE_HDFI T operator()(const T &{arg1}, const T &{arg2}) const {{
{contents}
}}
}};""".format(id=idstr, arg1=arg1, arg2=arg2, contents=body), sdfg,
state_id, node_id)
reduce_op = ', __reduce_' + idstr + '(), ' + symstr(node.identity)
elif redtype in ExpandReduceCUDADevice._SPECIAL_RTYPES:
reduce_op = ''
else:
credtype = 'dace::ReductionType::' + str(
redtype)[str(redtype).find('.') + 1:]
reduce_op = ((', dace::_wcr_fixed<%s, %s>()' %
(credtype, output_type)) + ', ' +
symstr(node.identity))
# Try to obtain the number of threads in the block, or use the default
# configuration
block_threads = devicelevel_block_size(sdfg, state, node)
if block_threads is not None:
block_threads = functools.reduce(lambda a, b: a * b, block_threads,
1)
# Checks
if block_threads is None:
raise ValueError('Block-wide GPU reduction must occur within'
' a GPU kernel')
if issymbolic(block_threads, sdfg.constants):
raise ValueError('Block size has to be constant for block-wide '
'reduction (got %s)' % str(block_threads))
if (node.axes is not None and len(node.axes) < input_dims):
raise ValueError(
'Only full reduction is supported for block-wide reduce,'
' please use the pure expansion')
if (input_data.storage != dtypes.StorageType.Register
or output_data.storage != dtypes.StorageType.Register):
raise ValueError(
'Block-wise reduction only supports GPU register inputs '
'and outputs')
if redtype in ExpandReduceCUDABlock._SPECIAL_RTYPES:
raise ValueError('%s block reduction not supported' % redtype)
credtype = 'dace::ReductionType::' + str(
redtype)[str(redtype).find('.') + 1:]
if redtype == dtypes.ReductionType.Custom:
redop = '__reduce_%s()' % idstr
else:
redop = 'dace::_wcr_fixed<%s, %s>()' % (credtype, output_type)
# Allocate shared memory for block reduce
localcode.write("""
typedef cub::BlockReduce<{type}, {numthreads}> BlockReduce_{id};
__shared__ typename BlockReduce_{id}::TempStorage temp_storage_{id};
""".format(id=idstr,
type=output_data.dtype.ctype,
numthreads=block_threads))
input = (input_memlet.data + ' + ' +
cpp_array_expr(sdfg, input_memlet, with_brackets=False))
output = cpp_array_expr(sdfg, output_memlet)
localcode.write("""
{output} = BlockReduce_{id}(temp_storage_{id}).Reduce({input}, {redop});
""".format(id=idstr,
redop=redop,
input=input_memlet.data,
output=output))
# Make tasklet
tnode = dace.nodes.Tasklet('reduce',
{'_in': dace.pointer(input_data.dtype)},
{'_out': dace.pointer(output_data.dtype)},
localcode.getvalue(),
language=dace.Language.CPP)
# Add the rest of the code
sdfg.append_global_code(cuda_globalcode.getvalue(), 'cuda')
# Rename outer connectors and add to node
input_edge._dst_conn = '_in'
output_edge._src_conn = '_out'
node.add_in_connector('_in')
node.add_out_connector('_out')
return tnode
def expansion(node: 'Reduce', state: SDFGState, sdfg: SDFG):
node.validate(sdfg, state)
input_edge: graph.MultiConnectorEdge = state.in_edges(node)[0]
output_edge: graph.MultiConnectorEdge = state.out_edges(node)[0]
input_dims = len(input_edge.data.subset)
output_dims = len(output_edge.data.subset)
input_data = sdfg.arrays[input_edge.data.data]
output_data = sdfg.arrays[output_edge.data.data]
# Setup all locations in which code will be written
cuda_globalcode = CodeIOStream()
cuda_initcode = CodeIOStream()
cuda_exitcode = CodeIOStream()
host_globalcode = CodeIOStream()
host_localcode = CodeIOStream()
output_memlet = output_edge.data
# Try to autodetect reduction type
redtype = detect_reduction_type(node.wcr)
node_id = state.node_id(node)
state_id = sdfg.node_id(state)
idstr = '{sdfg}_{state}_{node}'.format(sdfg=sdfg.name,
state=state_id,
node=node_id)
if node.out_connectors:
dtype = next(node.out_connectors.values())
else:
dtype = sdfg.arrays[output_memlet.data].dtype
output_type = dtype.ctype
if node.identity is None:
raise ValueError('For device reduce nodes, initial value must be '
'specified')
# Create a functor or use an existing one for reduction
if redtype == dtypes.ReductionType.Custom:
body, [arg1, arg2] = unparse_cr_split(sdfg, node.wcr)
cuda_globalcode.write(
"""
struct __reduce_{id} {{
template <typename T>
DACE_HDFI T operator()(const T &{arg1}, const T &{arg2}) const {{
{contents}
}}
}};""".format(id=idstr, arg1=arg1, arg2=arg2, contents=body), sdfg,
state_id, node_id)
reduce_op = ', __reduce_' + idstr + '(), ' + symstr(node.identity)
elif redtype in ExpandReduceCUDADevice._SPECIAL_RTYPES:
reduce_op = ''
else:
credtype = 'dace::ReductionType::' + str(
redtype)[str(redtype).find('.') + 1:]
reduce_op = ((', dace::_wcr_fixed<%s, %s>()' %
(credtype, output_type)) + ', ' +
symstr(node.identity))
# Obtain some SDFG-related information
input_memlet = input_edge.data
reduce_shape = input_memlet.subset.bounding_box_size()
num_items = ' * '.join(symstr(s) for s in reduce_shape)
input = (input_memlet.data + ' + ' +
cpp_array_expr(sdfg, input_memlet, with_brackets=False))
output = (output_memlet.data + ' + ' +
cpp_array_expr(sdfg, output_memlet, with_brackets=False))
input_dims = input_memlet.subset.dims()
output_dims = output_memlet.subset.data_dims()
reduce_all_axes = (node.axes is None or len(node.axes) == input_dims)
if reduce_all_axes:
reduce_last_axes = False
else:
reduce_last_axes = sorted(node.axes) == list(
range(input_dims - len(node.axes), input_dims))
if (not reduce_all_axes) and (not reduce_last_axes):
raise NotImplementedError(
'Multiple axis reductions not supported on GPUs. Please use '
'the pure expansion or make reduce axes the last in the array.'
)
# Verify that data is on the GPU
if input_data.storage not in [
dtypes.StorageType.GPU_Global, dtypes.StorageType.CPU_Pinned
]:
raise ValueError('Input of GPU reduction must either reside '
' in global GPU memory or pinned CPU memory')
if output_data.storage not in [
dtypes.StorageType.GPU_Global, dtypes.StorageType.CPU_Pinned
]:
raise ValueError('Output of GPU reduction must either reside '
' in global GPU memory or pinned CPU memory')
# Determine reduction type
kname = (ExpandReduceCUDADevice._SPECIAL_RTYPES[redtype] if redtype
in ExpandReduceCUDADevice._SPECIAL_RTYPES else 'Reduce')
# Create temp memory for this GPU
cuda_globalcode.write(
"""
void *__cub_storage_{sdfg}_{state}_{node} = NULL;
size_t __cub_ssize_{sdfg}_{state}_{node} = 0;
""".format(sdfg=sdfg.name, state=state_id, node=node_id), sdfg,
state_id, node)
if reduce_all_axes:
reduce_type = 'DeviceReduce'
reduce_range = num_items
reduce_range_def = 'size_t num_items'
reduce_range_use = 'num_items'
reduce_range_call = num_items
elif reduce_last_axes:
num_reduce_axes = len(node.axes)
not_reduce_axes = reduce_shape[:-num_reduce_axes]
reduce_axes = reduce_shape[-num_reduce_axes:]
num_segments = ' * '.join([symstr(s) for s in not_reduce_axes])
segment_size = ' * '.join([symstr(s) for s in reduce_axes])
reduce_type = 'DeviceSegmentedReduce'
iterator = 'dace::stridedIterator({size})'.format(
size=segment_size)
reduce_range = '{num}, {it}, {it} + 1'.format(num=num_segments,
it=iterator)
reduce_range_def = 'size_t num_segments, size_t segment_size'
iterator_use = 'dace::stridedIterator(segment_size)'
reduce_range_use = 'num_segments, {it}, {it} + 1'.format(
it=iterator_use)
reduce_range_call = '%s, %s' % (num_segments, segment_size)
# Call CUB to get the storage size, allocate and free it
cuda_initcode.write(
"""
cub::{reduce_type}::{kname}(nullptr, __cub_ssize_{sdfg}_{state}_{node},
({intype}*)nullptr, ({outtype}*)nullptr, {reduce_range}{redop});
cudaMalloc(&__cub_storage_{sdfg}_{state}_{node}, __cub_ssize_{sdfg}_{state}_{node});
""".format(sdfg=sdfg.name,
state=state_id,
node=node_id,
reduce_type=reduce_type,
reduce_range=reduce_range,
redop=reduce_op,
intype=input_data.dtype.ctype,
outtype=output_data.dtype.ctype,
kname=kname), sdfg, state_id, node)
cuda_exitcode.write(
'cudaFree(__cub_storage_{sdfg}_{state}_{node});'.format(
sdfg=sdfg.name, state=state_id, node=node_id), sdfg, state_id,
node)
# Write reduction function definition
cuda_globalcode.write("""
DACE_EXPORTED void __dace_reduce_{id}({intype} *input, {outtype} *output, {reduce_range_def}, cudaStream_t stream);
void __dace_reduce_{id}({intype} *input, {outtype} *output, {reduce_range_def}, cudaStream_t stream)
{{
cub::{reduce_type}::{kname}(__cub_storage_{id}, __cub_ssize_{id},
input, output, {reduce_range_use}{redop}, stream);
}}
""".format(id=idstr,
intype=input_data.dtype.ctype,
outtype=output_data.dtype.ctype,
reduce_type=reduce_type,
reduce_range_def=reduce_range_def,
reduce_range_use=reduce_range_use,
kname=kname,
redop=reduce_op))
# Write reduction function definition in caller file
host_globalcode.write(
"""
DACE_EXPORTED void __dace_reduce_{id}({intype} *input, {outtype} *output, {reduce_range_def}, cudaStream_t stream);
""".format(id=idstr,
reduce_range_def=reduce_range_def,
intype=input_data.dtype.ctype,
outtype=output_data.dtype.ctype), sdfg, state_id, node)
# Call reduction function where necessary
host_localcode.write(
'__dace_reduce_{id}({input}, {output}, {reduce_range_call}, __dace_current_stream);'
.format(id=idstr,
input=input,
output=output,
reduce_range_call=reduce_range_call))
# Make tasklet
tnode = dace.nodes.Tasklet('reduce',
{'_in': dace.pointer(input_data.dtype)},
{'_out': dace.pointer(output_data.dtype)},
host_localcode.getvalue(),
language=dace.Language.CPP)
# Add the rest of the code
sdfg.append_global_code(host_globalcode.getvalue())
sdfg.append_global_code(cuda_globalcode.getvalue(), 'cuda')
sdfg.append_init_code(cuda_initcode.getvalue(), 'cuda')
sdfg.append_exit_code(cuda_exitcode.getvalue(), 'cuda')
# Rename outer connectors and add to node
input_edge._dst_conn = '_in'
output_edge._src_conn = '_out'
node.add_in_connector('_in')
node.add_out_connector('_out')
return tnode
