[IR] Ir fill constant #56520
[IR] Ir fill constant #56520
Conversation
… ir_fill_constant
|
你的PR提交成功，感谢你对开源项目的贡献! |
| ir::Program* program, | ||
| const std::vector<std::string>& args, | ||
| int axis = 0) { | ||
| ir::OpInfo op_info = ctx->GetRegisteredOpInfo("pd.stack"); |
There was a problem hiding this comment.
| ir::OpInfo op_info = ctx->GetRegisteredOpInfo("pd.stack"); | |
| ir::OpInfo op_info = ctx->GetRegisteredOpInfo(dialect::StackOp::name()); |
| @@ -1175,6 +1198,149 @@ struct TrilAndTriuOpTranscriber : public OpTranscriber { | |||
| } | |||
| }; | |||
|
|
|||
| struct FillConstantTranscriber : public OpTranscriber { | |||
| ir::Operation* operator()(ir::IrContext* ctx, | |||
There was a problem hiding this comment.
不建议直接重载 operator，可以看看OpTranscriber::operator()的实现，里面做的事情其实是比较多的，自己实现一个容易遗漏。
如果只需要修改处理输入的方式，可以只重载OpTranscriber::GenerateOperationInput；
如果需要修改处理属性的方式，可以重载OpTranscriber::TranslateOpAttribute,
如果需要处理特定输入/属性，可以考虑重载GetSpecialInputHandlers/GetSpecialAttributeHandlers
| has_mutable_attribute |= op_desc.HasInput("ValueTensor", true) && | ||
| op_desc.Input("ValueTensor", true).size() > 0; | ||
|
|
||
| if (!has_mutable_attribute) { |
There was a problem hiding this comment.
建议写两个不同的OpTranscriber，通过Op名指定分发方式，如
struct FillConstantTranscriber : public OpTranscriber {
ir::Operation* operator()(ir::IrContext* ctx,
TranslationContext* param_map,
const OpDesc& op_desc,
ir::Program* program) override {
bool has_mutable_attribute = op_desc.HasInput("ShapeTensor", true) && op_desc.Input("ShapeTensor", true).size() > 0;
has_mutable_attribute |= op_desc.HasInput("ShapeTensorList", true) && op_desc.Input("ShapeTensorList", true).size() > 0;
has_mutable_attribute |= op_desc.HasInput("ValueTensor", true) && op_desc.Input("ValueTensor", true).size() > 0;
if (has_mutable_attribute) {
return FullOpTranscriber()(ctx, param_map, op_desc, program);
} else {
return FullWithInputOpTranscriber()(ctx, param_map, op_desc, program);
}
}
};然后FullOpTranscriber和FullWithInputOpTranscriber可以通过继承的方式做一下代码复用
| auto op_info = ctx->GetRegisteredOpInfo("pd.fill_with_tensor"); | ||
| if (!op_info) { | ||
| IR_THROW( | ||
| "Op tril_triu should have corresponding OpInfo " |
There was a problem hiding this comment.
| "Op tril_triu should have corresponding OpInfo " | |
| "Op fill_constant with mutable attribute should have corresponding OpInfo " |
| op_desc.Input("ValueTensor", true).size() > 0; | ||
|
|
||
| if (!has_mutable_attribute) { | ||
| auto op_info = ctx->GetRegisteredOpInfo("pd.full"); |
There was a problem hiding this comment.
| auto op_info = ctx->GetRegisteredOpInfo("pd.full"); | |
| auto op_info = ctx->GetRegisteredOpInfo(dialect::FullOp::name()); |
| ir::Operation* operation = ir::Operation::Create( | ||
| {}, attribute_map, {defining_info.value.type()}, op_info); | ||
| program->block()->push_back(operation); | ||
| return operation; |
There was a problem hiding this comment.
现在的实现里缺少对OpTranscriber::RecordOpResultMapping的调用，这个函数是用来记录VarDesc和ir::Value之间的映射的，缺失会导致bug。
| param : [shape, dtype] | ||
| kernel : | ||
| func : full_with_tensor | ||
| data_type : dtype |
There was a problem hiding this comment.
需要支持 inplace 版本么？参考 full_
* support ir fill constant
PR types
Bug fixes
PR changes
Others
Description
Pcard-67164
解决IR不支持fill_constant传递ShapeTensor、ShapeTensorList、ValueTensor问题。
关于fill_with_tensor的实现，由于Tensor shape可能在运行时才有值，且InforMata传入的是shape的MetaTensor。导致InforMata的时候无法推到shape。所以InforMata将shape设置成[-1]，在FullWithTensorKernel中，再跟进shape做一次Resize。