0824 MLIR和AST相关资料

• 简单直观的理解：【编译原理：构建抽象语法树】
  编译原理：构建抽象语法树_哔哩哔哩_bilibili
• 程序转为AST可视化：https://astexplorer.net/
• MLIR的博客：https://www.cnblogs.com/YuanZiming/p/18956253
• 学习MLIR的仓库：https://github.com/KEKE046/mlir-tutorial
• for循环：https://blog.csdn.net/ReadyShowShow/article/details/137275857
• if判断

DIST-IR中有用的测试用例

test/test_sequential_executor.py中的两个元素加法

test/test_simulator.py中的两个连续的矩阵乘法_test_data_parallel

test/test_pytorch_backend.py中的Send函数测试：test_send_recv

test/test_prettyprint.py中的两个连续的矩阵乘法：IR

tensor乘法simulator.py

完成图的构建，在这里采用了手写中间表示的方式

FunctionMaker中包含四个重要函数

• add_op：参数为：操作码、输入、属性、子函数
• add_input_value：添加输入值
• set_outputs：设置元组输出
• finalize：根据上面的参数构造Function类

IR的格式

大萨

simulator的实现

类型

Function = {# Invariant: ops are in topological orderops: List[Op]inputs: List[Value]outputs: List[Value]
}Device = {device_id: Int# Unique device IDdevice_type: String# Device type (e.g. "gpu")
}Op = {name: String# Do we need this to be unique in a function? Across all functions?op_type: OpType# The type of operatorin_edges: List[Value]# Pointer to a Value object either in Function.inputs or another Op.out_edgesout_edges: List[Value]# To support ops that have more than one outputattributes: Dict[String, Any]# Constant data for ops, e.g. stride of convolution or devices to scattersubfunctions: List[Function]
}OpType = | MatMul | Add | Relu | ...| MatMulGrad | ReluGrad | ...| N11Loss | SGDOptimizer | ...| Pmap | AllReduce | ...Value = {name: String# Again, does it need to be unique in a function?type: Typedevice: DeviceID# Which device this value lives on# TODO pointer to source op that generated this value?
}Type =| Tensor{shape: Optional[Shape], dtype: Type, device: Device}| Float | Int | ...| TupleType{elems: List[Type]}Topology = {devices: List[Device]# The list of all devices in the topology.bandwidths: Dict[Device, Dict[Device, Float]]# The bandwidth between each device.
}

dispatch是如何绑定的

更新字典的函数

• 操作码类型推断：
  

• 混合输入的操作码类型推断：
  

• 实际操作码计算
  

• 通信类型操作码：