Function dispatch_graph

ppq/api/interface.py:644–699

这个函数执行图切分与调度，你的计算图将被切分成一系列子图，并被调度到不同设备上。 调度的逻辑分为自动控制的部分以及手动覆盖的部分，你可以使用 QuantizationSetting 来向这个函数传递手动调度表 从而覆盖 PPQ 的调度逻辑。 注意：这个函数依据调度器和 TargetPlatform 平台的不同而产生行为差异，生成不同的调度计划。 This function will cut your graph into a series of subgraph and send them to different device. PPQ provides

(
    graph: BaseGraph, platform: TargetPlatform, 
    dispatcher: Union[str, GraphDispatcher] = 'conservative', 
    dispatching_table: DispatchingTable = None)

def dispatch_graph(
    graph: BaseGraph, platform: TargetPlatform, 
    dispatcher: Union[str, GraphDispatcher] = 'conservative', 
    dispatching_table: DispatchingTable = None) -> BaseGraph:
    """这个函数执行图切分与调度，你的计算图将被切分成一系列子图，并被调度到不同设备上。 
    调度的逻辑分为自动控制的部分以及手动覆盖的部分，你可以使用 QuantizationSetting 来向这个函数传递手动调度表 从而覆盖 PPQ 的调度逻辑。

    注意：这个函数依据调度器和 TargetPlatform 平台的不同而产生行为差异，生成不同的调度计划。

    This function will cut your graph into a series of subgraph and send them to different device.
    PPQ provides an automatic dispatcher which, will generate different dispatching scheme on your TargetPlatform.
    A dispatching table can be passed via QuantizationSetting to override
        the default dispatching logic of ppq dispatcher manually.
    """
    dispatching_override = dispatching_table
    quantizer = PFL.Quantizer(platform=platform, graph=graph) # 初始化一个 quantizer 没有很大代价...

    if isinstance(dispatcher, str):
        dispatcher = dispatcher.lower()
        if dispatcher not in DISPATCHER_TABLE:
            raise ValueError(f'Can not found dispatcher type "{dispatcher}", check your input again.')
        dispatcher = DISPATCHER_TABLE[dispatcher](graph)
    else:
        if not isinstance(dispatcher, GraphDispatcher):
            raise TypeError('Parameter "dispachter" of function ppq.api.dispatch_graph must be String or GraphDispatcher, '
                            f'however {type(dispatcher)} was given.')
        dispatcher = dispatcher

    assert isinstance(dispatcher, GraphDispatcher)
    assert isinstance(quantizer, BaseQuantizer)
    quant_types = quantizer.quant_operation_types
    dispatching_table = dispatcher.dispatch(
        graph=graph, quant_types=quant_types,
        quant_platform=TargetPlatform.UNSPECIFIED, # MUST BE UNSPECIFIED, 这里的意思是交由 Quantizer 决定是否量化这个算子
        fp32_platform=TargetPlatform.FP32,
        SOI_platform=TargetPlatform.SOI)

    # override dispatching result
    if dispatching_override is not None:
        if not isinstance(dispatching_override, DispatchingTable):
            raise TypeError('Parameter "dispatching_table" of function ppq.api.dispatch_graph must be DispatchingTable, '
                            f'however {type(dispatching_override)} was given.')

        for opname, platform in dispatching_override.dispatchings.items():
            if opname not in graph.operations: continue
            assert isinstance(platform, int), (
                f'Your dispatching table contains a invalid setting of operation {opname}, '
                'All platform setting given in dispatching table is expected given as int, '
                f'however {type(platform)} was given.')
            dispatching_table[opname] = TargetPlatform(platform)

    for operation in graph.operations.values():
        assert operation.name in dispatching_table, (
            f'Internal Error, Can not find operation {operation.name} in dispatching table.')
        operation.platform = dispatching_table[operation.name]
    return graph