def get_func_and_var_info():
needed = []
for n in loopnode.dfg.get_nodes_deep():
if isinstance(n, DataFlowGraph.ExtCallNode):
needed.append(NeededInfo(loopnode.dfg.stmt_sequence, n.line.stmt_idx, n.ast_node.func.as_string(), n))
elif isinstance(n.ast_node, astroid.Name):
needed.append(NeededInfo(loopnode.dfg.stmt_sequence, n.line.stmt_idx, n.ast_node.name, n,
transform=lambda x: (type(x), x.dtype, x.shape) if isinstance(x, np.ndarray) else (type(x), None, None)))
return watcher.get_runtime_info(func, needed)
def get_dot_shapes(funccall_info):
#print "In get dot shapes with funccall_info", funccall_info
need_arg_shapes = []
#print "got funccall_info"
#pprint(funccall_info)
for (ni, f) in funccall_info.iteritems():
#print "ni, f, ", ni, f
if f == np.dot:
need_arg_shapes.append(
NeededInfo(
ni.stmt_sequence, ni.stmt_idx, '[' + (','.join([
a.as_string() + '.shape'
for a in ni.dfg_node.ast_node.args
])) + ']', ni.dfg_node))
#print "get_dot_shapes needs ", need_arg_shapes
return watcher.get_runtime_info(func, need_arg_shapes)
def get_dot_shapes(funccall_info):
# print "In get dot shapes with funccall_info", funccall_info
need_arg_shapes = []
# print "got funccall_info"
# pprint(funccall_info)
for (ni, f) in funccall_info.iteritems():
# print "ni, f, ", ni, f
if f == np.dot:
need_arg_shapes.append(
NeededInfo(
ni.stmt_sequence,
ni.stmt_idx,
"[" + (",".join([a.as_string() + ".shape" for a in ni.dfg_node.ast_node.args])) + "]",
ni.dfg_node,
)
)
# print "get_dot_shapes needs ", need_arg_shapes
return watcher.get_runtime_info(func, need_arg_shapes)
def get_func_and_var_info():
needed = []
for n in loopnode.dfg.get_nodes_deep():
if isinstance(n, DataFlowGraph.ExtCallNode):
needed.append(
NeededInfo(loopnode.dfg.stmt_sequence, n.line.stmt_idx,
n.ast_node.func.as_string(), n))
elif isinstance(n.ast_node, astroid.Name):
needed.append(
NeededInfo(loopnode.dfg.stmt_sequence,
n.line.stmt_idx,
n.ast_node.name,
n,
transform=lambda x: (type(x), x.dtype, x.shape)
if isinstance(x, np.ndarray) else
(type(x), None, None)))
return watcher.get_runtime_info(func, needed)
def optimize_matrix_chain(self, func, dfg):
"""
Look for (effectively) nested calls to dot.
The meat of this function is
1) get information about all external calls - to filter those which are to np.dot
2) get information on the shapes of the arguments to each np.dot (get_dot_shapes)
3) find chained dot calls
4) for each chain, calculate the optimal order, generate a new expression, and
call replace_subgraph_and_code to put it in.
"""
print "optimize_matrix_chain called with", func, dfg
assumptions = {}
done_chains = {}
subchains = set(
) # Chains that are reachable recursively; these are not full chains
def chain_for(dfg_node, shape, dot_shapes):
""" For a given call to dot (represented by dfg_node), move up the data flow graph
and find other dot calls that are chained together and whose intermediate results
are not used elsewhere (so we can freely change the order of multiplications).
When this function completes the full chain will be stored in done_chains."""
cprint("chain_for called with " + str(dfg_node), 'red')
cprint("dot_shapes is " + str(dot_shapes), 'red')
res = []
a_inp = dfg.get_callarg_value_node(dfg_node, 'a', 0)
b_inp = dfg.get_callarg_value_node(dfg_node, 'b', 1)
assert a_inp is not None and b_inp is not None
for i, arg in enumerate((a_inp, b_inp)):
# Ensure that we don't need the intermediate result elsewhere
arg_source = dfg.last_transform(arg)
print "arg source for ", arg, "is", arg_source
if (arg_source in dot_shapes
and len(dot_shapes[arg_source]) == 2
and not (dfg.has_nonlocal_scope(arg)
or dfg.fanout(arg_source) > 1)):
if arg_source in done_chains:
res += done_chains[arg_source]
else:
res += chain_for(arg_source, dot_shapes[arg_source],
dot_shapes)
subchains.add(arg_source)
else:
res += [(arg_source, shape[i])]
done_chains[dfg_node] = res
print "chain_for returning", res
return res
def mult_order_to_expr(inputs, order, func):
if isinstance(order, Number):
return inputs[order] #.ast_node.as_string()
else:
return (func + "(" +
mult_order_to_expr(inputs, order[0], func) + ", " +
mult_order_to_expr(inputs, order[1], func) + ")")
def optimize_chain_inner(dot_shapes):
print "In optimize_chain_inner! Looking to optimize these dot calls/shapes ", dot_shapes
dot_call_dfg_nodes = {
dotcall_ni.dfg_node: shape
for (dotcall_ni, shape) in dot_shapes.items()
}
for (dotcall_ni, shape) in dot_shapes.iteritems():
if len(shape) == 2:
chain_for(dotcall_ni.dfg_node, shape, dot_call_dfg_nodes)
# Remove incomplete chains
for d in subchains:
done_chains.pop(d, None)
print "And here are the chains: "
pprint(done_chains)
for (end, inputs) in done_chains.iteritems():
for (source, shape) in inputs:
assumptions[source] = "{1}.shape == %s" % str(shape)
chain_inputs = [inp for (inp, shp) in inputs]
chain_shapes = [shp for (inp, shp) in inputs]
optimal_order = matrix_chain.matrix_chain_tree(chain_shapes)
print "Optimal order", optimal_order
placeholders = ['{%i}' % i for i in range(len(chain_inputs))]
new_chain_expr = mult_order_to_expr(
placeholders, optimal_order, end.ast_node.func.as_string())
nodes_to_replace, edges_to_replace = dfg.subgraph_between(
chain_inputs, end)
cprint(
"Going to replace a subgraph, saved as 'last_subgraph', with %s"
% new_chain_expr, 'red')
mainvar('last_subgraph', nodes_to_replace)
mainvar('input_nodes', chain_inputs)
mainvar('new_chain_expr', new_chain_expr)
mainvar('assumptions', assumptions)
if not DRYRUN:
modcode.replace_subgraph_and_code(dfg, nodes_to_replace,
chain_inputs,
new_chain_expr,
assumptions)
#print "func is", func
def get_dot_shapes(funccall_info):
#print "In get dot shapes with funccall_info", funccall_info
need_arg_shapes = []
#print "got funccall_info"
#pprint(funccall_info)
for (ni, f) in funccall_info.iteritems():
#print "ni, f, ", ni, f
if f == np.dot:
need_arg_shapes.append(
NeededInfo(
ni.stmt_sequence, ni.stmt_idx, '[' + (','.join([
a.as_string() + '.shape'
for a in ni.dfg_node.ast_node.args
])) + ']', ni.dfg_node))
#print "get_dot_shapes needs ", need_arg_shapes
return watcher.get_runtime_info(func, need_arg_shapes)
# Get (references to) all the function calls in the graph
print "Nodes", dfg.nodes
func_calls = [
NeededInfo(dfg.stmt_sequence, n.line.stmt_idx,
n.ast_node.func.as_string(), n) for n in dfg.nodes
if isinstance(n, DataFlowGraph.ExtCallNode)
]
print "func_calls looking for info for "
pprint(func_calls)
p = watcher.get_runtime_info(
func,
func_calls).then(get_dot_shapes).then(optimize_chain_inner).done(
None, onError)
def optimize_matrix_chain(self, func, dfg):
"""
Look for (effectively) nested calls to dot.
The meat of this function is
1) get information about all external calls - to filter those which are to np.dot
2) get information on the shapes of the arguments to each np.dot (get_dot_shapes)
3) find chained dot calls
4) for each chain, calculate the optimal order, generate a new expression, and
call replace_subgraph_and_code to put it in.
"""
print "optimize_matrix_chain called with", func, dfg
assumptions = {}
done_chains = {}
subchains = set() # Chains that are reachable recursively; these are not full chains
def chain_for(dfg_node, shape, dot_shapes):
""" For a given call to dot (represented by dfg_node), move up the data flow graph
and find other dot calls that are chained together and whose intermediate results
are not used elsewhere (so we can freely change the order of multiplications).
When this function completes the full chain will be stored in done_chains."""
cprint("chain_for called with " + str(dfg_node), "red")
cprint("dot_shapes is " + str(dot_shapes), "red")
res = []
a_inp = dfg.get_callarg_value_node(dfg_node, "a", 0)
b_inp = dfg.get_callarg_value_node(dfg_node, "b", 1)
assert a_inp is not None and b_inp is not None
for i, arg in enumerate((a_inp, b_inp)):
# Ensure that we don't need the intermediate result elsewhere
arg_source = dfg.last_transform(arg)
print "arg source for ", arg, "is", arg_source
if (
arg_source in dot_shapes
and len(dot_shapes[arg_source]) == 2
and not (dfg.has_nonlocal_scope(arg) or dfg.fanout(arg_source) > 1)
):
if arg_source in done_chains:
res += done_chains[arg_source]
else:
res += chain_for(arg_source, dot_shapes[arg_source], dot_shapes)
subchains.add(arg_source)
else:
res += [(arg_source, shape[i])]
done_chains[dfg_node] = res
print "chain_for returning", res
return res
def mult_order_to_expr(inputs, order, func):
if isinstance(order, Number):
return inputs[order] # .ast_node.as_string()
else:
return (
func
+ "("
+ mult_order_to_expr(inputs, order[0], func)
+ ", "
+ mult_order_to_expr(inputs, order[1], func)
+ ")"
)
def optimize_chain_inner(dot_shapes):
print "In optimize_chain_inner! Looking to optimize these dot calls/shapes ", dot_shapes
dot_call_dfg_nodes = {dotcall_ni.dfg_node: shape for (dotcall_ni, shape) in dot_shapes.items()}
for (dotcall_ni, shape) in dot_shapes.iteritems():
if len(shape) == 2:
chain_for(dotcall_ni.dfg_node, shape, dot_call_dfg_nodes)
# Remove incomplete chains
for d in subchains:
done_chains.pop(d, None)
print "And here are the chains: "
pprint(done_chains)
for (end, inputs) in done_chains.iteritems():
for (source, shape) in inputs:
assumptions[source] = "{1}.shape == %s" % str(shape)
chain_inputs = [inp for (inp, shp) in inputs]
chain_shapes = [shp for (inp, shp) in inputs]
optimal_order = matrix_chain.matrix_chain_tree(chain_shapes)
print "Optimal order", optimal_order
placeholders = ["{%i}" % i for i in range(len(chain_inputs))]
new_chain_expr = mult_order_to_expr(placeholders, optimal_order, end.ast_node.func.as_string())
nodes_to_replace, edges_to_replace = dfg.subgraph_between(chain_inputs, end)
cprint("Going to replace a subgraph, saved as 'last_subgraph', with %s" % new_chain_expr, "red")
mainvar("last_subgraph", nodes_to_replace)
mainvar("input_nodes", chain_inputs)
mainvar("new_chain_expr", new_chain_expr)
mainvar("assumptions", assumptions)
if not DRYRUN:
modcode.replace_subgraph_and_code(dfg, nodes_to_replace, chain_inputs, new_chain_expr, assumptions)
# print "func is", func
def get_dot_shapes(funccall_info):
# print "In get dot shapes with funccall_info", funccall_info
need_arg_shapes = []
# print "got funccall_info"
# pprint(funccall_info)
for (ni, f) in funccall_info.iteritems():
# print "ni, f, ", ni, f
if f == np.dot:
need_arg_shapes.append(
NeededInfo(
ni.stmt_sequence,
ni.stmt_idx,
"[" + (",".join([a.as_string() + ".shape" for a in ni.dfg_node.ast_node.args])) + "]",
ni.dfg_node,
)
)
# print "get_dot_shapes needs ", need_arg_shapes
return watcher.get_runtime_info(func, need_arg_shapes)
# Get (references to) all the function calls in the graph
print "Nodes", dfg.nodes
func_calls = [
NeededInfo(dfg.stmt_sequence, n.line.stmt_idx, n.ast_node.func.as_string(), n)
for n in dfg.nodes
if isinstance(n, DataFlowGraph.ExtCallNode)
]
print "func_calls looking for info for "
pprint(func_calls)
p = (
watcher.get_runtime_info(func, func_calls)
.then(get_dot_shapes)
.then(optimize_chain_inner)
.done(None, onError)
)
