def compare_impl(full_raw_stack,
full_actuals,
chunk_raw_stack,
chunk_actuals,
fancy_func=False,
external_trim=None):
# cross check both stacks -- make sure they have the same shape
cross_check(full_raw_stack, chunk_raw_stack)
# build a full tree -- just to get the aten op name
root = stack_parser.exec_time_tree(full_raw_stack, fancy_func=fancy_func)
# this must be true -- the root has 2 children -- the aten op and other
op_name = root.children[0].func
# process CPU_log and HB_log
# CPU_log should be given by full input data
# HB_log should be given by chunk input data
cpu_log = process_CPU_stack.parse(full_raw_stack, fancy_func=fancy_func)
hb_log = process_HB_stack.parse(chunk_raw_stack,
fancy_func=fancy_func,
trimming=True)
# re-apply trim of 0 if using external tirmming is enabled
# since we can have more than one @TRIM@ node, we cannot just adjust these nodes
if external_trim is not None:
def reset_trim(root):
if root.func == "@TRIM@":
root.time = float(0)
hb_log = traversal(hb_log, reset_trim)
stack_parser.exec_time_apply_trim(hb_log)
# get total time on device
# so we accumulate all simualted time
total_time_on_HB = 0
if external_trim is None:
def acc_trim(root):
nonlocal total_time_on_HB
if root.func == "@TRIM@":
total_time_on_HB += root.time
traversal(hb_log, acc_trim)
else:
total_time_on_HB = external_trim
# debug
print(stack_parser.exec_time_print_tree(cpu_log))
print(stack_parser.exec_time_print_tree(hb_log))
print("total time on HB = " + str(total_time_on_HB))
# process input tensors
actuals = actual_parser.parse(full_actuals, chunk_actuals)
return ATen_OP(op_name, cpu_log, hb_log, total_time_on_HB, actuals)
def test_compare_aten_op_1():
aten_op = compare(current_path + "/demo/full.std",
current_path + "/demo/chunk.std",
current_path + "/demo/manycore_stats.log")
# pytest assertion
assert aten_op.hb_device_time == 1204.603
assert aten_op.hb_host_time == 170.0
assert aten_op.xeon_time == 34.451
cpu_graph = stack_parser.exec_time_print_tree(aten_op.cpu_log)
assert cpu_graph == "|- Node(@CPU_LOG@ : 34.451)\n |- Node(at::Tensor at::CPUType::{anonymous}::empty(c10::IntArrayRef, const c10::TensorOptions&, c10::optional<c10::MemoryFormat>) : 0.043)\n |- Node(at::native::add_stub::add_stub() : 34.316)"
hb_graph = stack_parser.exec_time_print_tree(aten_op.hb_log)
assert hb_graph == "|- Node(@HB_LOG@ : 170.0)\n |- Node(at::Tensor at::HammerBladeType::{anonymous}::empty(c10::IntArrayRef, const c10::TensorOptions&, c10::optional<c10::MemoryFormat>) : 0.004)\n |- Node(at::native::add_stub::add_stub() : 170.0)\n |- Node(@OFFLOAD_KERNEL@__tensorlib_add : 0.0)\n |- Node(@TRIM@ : 0.0)"
def test_hb_stack_1():
cpu_log = process_HB_stack.parse(raw_stack)
graph = stack_parser.exec_time_print_tree(cpu_log)
assert graph == """|- Node(@HB_LOG@ : 0.234)\n |- Node(at::Tensor at::HammerBladeType::{anonymous}::empty(c10::IntArrayRef, const c10::TensorOptions&, c10::optional<c10::MemoryFormat>) : 0.01)\n |- Node(at::native::add_stub::add_stub() : 0.187)\n |- Node(@OFFLOAD_KERNEL@__tensorlib_add : 0.145)\n |- Node(@TRIM@ : 0.0)"""
def test_cpu_stack_1():
cpu_log = process_CPU_stack.parse(raw_stack)
graph = stack_parser.exec_time_print_tree(cpu_log)
assert graph == """|- Node(@CPU_LOG@ : 0.067)\n |- Node(at::Tensor at::CPUType::{anonymous}::empty(c10::IntArrayRef, const c10::TensorOptions&, c10::optional<c10::MemoryFormat>) : 0.015)\n |- Node(at::native::add_stub::add_stub() : 0.009)"""
def test_stack_parsing_1():
root = stack_parser.exec_time_tree(raw_stack)
graph = stack_parser.exec_time_print_tree(root)
assert graph == """|- Node(time_in_roi : 0.399)\n |- Node(at::Tensor at::CPUType::{anonymous}::add(const at::Tensor&, const at::Tensor&, c10::Scalar) : 0.399)\n |- Node(@CPU_LOG@ : 0.067)\n |- Node(at::Tensor at::CPUType::{anonymous}::empty(c10::IntArrayRef, const c10::TensorOptions&, c10::optional<c10::MemoryFormat>) : 0.015)\n |- Node(at::native::add_stub::add_stub() : 0.009)\n |- Node(@HB_LOG@ : 0.234)\n |- Node(at::Tensor at::HammerBladeType::{anonymous}::empty(c10::IntArrayRef, const c10::TensorOptions&, c10::optional<c10::MemoryFormat>) : 0.01)\n |- Node(at::native::add_stub::add_stub() : 0.187)\n |- Node(@OFFLOAD_KERNEL@__tensorlib_add : 0.145)\n |- Node(@TRIM@ : 0.0)\n |- Node(at::Tensor at::CPUType::{anonymous}::llcopy(const at::Tensor&) : 0.025)\n |- Node(other : 0.0)"""