# Copyright (c) 2022 VisualDL Authors. All Rights Reserve. # # Licensed under the Apache License, Version 2.0 (the "License"); # you may not use this file except in compliance with the License. # You may obtain a copy of the License at # # http://www.apache.org/licenses/LICENSE-2.0 # # Unless required by applicable law or agreed to in writing, software # distributed under the License is distributed on an "AS IS" BASIS, # WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. # See the License for the specific language governing permissions and # limitations under the License. # ======================================================================= import collections class TC_Allowlist(dict): # Refer to https://github.com/NVIDIA/PyProf/blob/fd1b2902e3306119eee40ba6b6e8b2f816920c29/pyprof/prof/tc.py#L19 allowlist = [ 'h884', 's884', 'h1688', 's1688', 'hmma', 'i8816', '16816', 'dgrad_1x1_stride_2x2', 'first_layer_wgrad_kernel', 'conv1x1', 'conv2d_c1_k1', 'direct_group', 'xmma_implicit_gemm', 'xmma_sparse_conv', 'xmma_warp_specialized_implicit_gemm', 'xmma_gemm', 'xmma_sparse_gemm', 'c1688' ] def __init__(self): pass def __contains__(self, item): # If kernel name contains substring equal to any one in allowlist, then it uses tensor core. for pattern in self.allowlist: if pattern in item: return True return False _allow_list = TC_Allowlist() class DeviceItem: def __init__(self, name): self.name = name self.call = 0 self.gpu_time = 0 self.max_gpu_time = 0 self.min_gpu_time = float('inf') self.tensorcore_used = True if name in _allow_list else False self.sum_blocks_per_sm = 0.0 self.sum_occupancy = 0.0 @property def avg_gpu_time(self): return self.gpu_time / self.call def add_gpu_time(self, time): if time > self.max_gpu_time: self.max_gpu_time = time if time < self.min_gpu_time: self.min_gpu_time = time self.gpu_time += time def add_item(self, node): self.call += 1 self.add_gpu_time(node.end_ns - node.start_ns) self.sum_blocks_per_sm += node.blocks_per_sm self.sum_occupancy += node.occupancy class KernelParser: def __init__(self): self.kernel_items = {} # for kernel summary self.kernel_items_with_op_name_attributes = collections.defaultdict( dict) self.gpu_ids = set() self.occupancy = 0.0 self.sm_efficiency = 0.0 self.tensor_core_ratio = 0.0 def parse(self, nodelists): # noqa: C901 total_duration = 0.0 weighted_occupancy = 0.0 weighted_sm_efficiency = 0.0 for threadid, nodes in nodelists.items(): for node in nodes: if node.type == 'Operator': op_name = node.name for children in node.children_node: if children.type == 'OperatorInner': for runtime_node in children.runtime_node: for device_node in runtime_node.device_node: if device_node.type == 'Kernel': op_attribute_name = self._translate_op_name_attributes_to_string( op_name, device_node) if op_attribute_name not in self.kernel_items_with_op_name_attributes[ device_node.name]: self.kernel_items_with_op_name_attributes[ device_node.name][ op_attribute_name] = DeviceItem( device_node.name) self.kernel_items_with_op_name_attributes[ device_node. name][op_attribute_name].add_item( device_node) for runtime_node in node.runtime_node: for device_node in runtime_node.device_node: if device_node.type == 'Kernel': op_attribute_name = self._translate_op_name_attributes_to_string( op_name, device_node) if op_attribute_name not in self.kernel_items_with_op_name_attributes[ device_node.name]: self.kernel_items_with_op_name_attributes[ device_node. name][op_attribute_name] = DeviceItem( device_node.name) self.kernel_items_with_op_name_attributes[ device_node. name][op_attribute_name].add_item( device_node) elif node.type == 'OperatorInner': continue op_name = node.name for runtime_node in node.runtime_node: for device_node in runtime_node.device_node: if device_node.type == 'Kernel': op_attribute_name = self._translate_op_name_attributes_to_string( op_name, device_node) if op_attribute_name not in self.kernel_items_with_op_name_attributes[ device_node.name]: self.kernel_items_with_op_name_attributes[ device_node. name][op_attribute_name] = DeviceItem( device_node.name) self.kernel_items_with_op_name_attributes[ device_node.name][op_attribute_name].add_item( device_node) for threadid, nodes in nodelists.items(): for node in nodes: for runtime_node in node.runtime_node: for device_node in runtime_node.device_node: if device_node.type == 'Kernel': name = device_node.name if name not in self.kernel_items: self.kernel_items[name] = DeviceItem(name) self.kernel_items[name].add_item(device_node) weighted_occupancy += ( device_node.occupancy / 100) * ( device_node.end_ns - device_node.start_ns) if device_node.blocks_per_sm > 1: sm_efficiency = 1 else: sm_efficiency = device_node.blocks_per_sm weighted_sm_efficiency += sm_efficiency * ( device_node.end_ns - device_node.start_ns) total_duration += ( device_node.end_ns - device_node.start_ns) self.gpu_ids.add(device_node.device_id) self.occupancy = weighted_occupancy / total_duration if total_duration != 0 else 0.0 self.sm_efficiency = weighted_sm_efficiency # to divide ProfileStep time in ProfileData total_time = 0 total_tensorcore_time = 0 for name, node in self.kernel_items.items(): if node.tensorcore_used: total_tensorcore_time += node.gpu_time total_time += node.gpu_time self.tensor_core_ratio = total_tensorcore_time / total_time if total_time != 0 else 0.0 def _translate_op_name_attributes_to_string(self, op_name, event): result = '{}-[{},{},{}]-[{},{},{}]-{}-{}'.format( op_name, event.grid_x, event.grid_y, event.grid_z, event.block_x, event.block_y, event.block_z, event.registers_per_thread, event.shared_memory) return result