[NewIR]support new ir load combine #56101
Conversation
|
你的PR提交成功，感谢你对开源项目的贡献! |
|
❌ The PR is not created using PR's template. You can refer to this Demo. |
phlrain
changed the title
| op_yaml_info_parser, | ||
| &(op_func_node.kernel_context_)); | ||
| } | ||
| ::ir::BuildPhiContext<phi::KernelContext, |
There was a problem hiding this comment.
随着执行器接入的完善，BuildOpFuncList函数后续应该没有什么用处了。TODO：后续随执行器代码清理 pr 一并清理
| @@ -1547,7 +1547,8 @@ void NewIRInterpreter::BuildInstruction() { | |||
| } | |||
|
|
|||
| if (op_name == "pd.fused_softmax_mask_upper_triangle" || | |||
| op_name == "pd.fused_softmax_mask_upper_triangle_grad") { | |||
| op_name == "pd.fused_softmax_mask_upper_triangle_grad" || | |||
| op_name == "pd.load_combine") { | |||
There was a problem hiding this comment.
看起来 Legacy 的策略是用「白名单」来识别的？感觉这里的判断连同「白名单」一起封装成一个函数 IsLegacyKernelOp(op_name) 比较好一些？
There was a problem hiding this comment.
已经优化； 理想情况，这款应该用dialect来区分，但是暂时暂时缺少对应的dialect，所以暂时用了白名单
| std::vector<paddle::framework::Variable*> vec_tmp = {var}; | ||
|
|
||
| runtime_ctx->outputs[legacy_arg_name] = vec_tmp; |
There was a problem hiding this comment.
| std::vector<paddle::framework::Variable*> vec_tmp = {var}; | |
| runtime_ctx->outputs[legacy_arg_name] = vec_tmp; | |
| runtime_ctx->outputs[legacy_arg_name] = {var}; |
|
|
||
| auto& val = attr.second; | ||
|
|
||
| if (val.isa<ir::StrAttribute>()) { |
There was a problem hiding this comment.
这里的Attribute解析数据的逻辑在paddle/fluid/ir/dialect/utils.h 已经有GetAttributeData 函数实现了，可以看下是否能够复用？
There was a problem hiding this comment.
我确认了下 GetAttributeData 返回的是 phi的attribute， 这里需要的是framework::attribute, 定义还是有些差别，不能直接用同一个
| @@ -387,6 +388,9 @@ std::unique_ptr<ir::Program> PdOpLowerToKernelPass(ir::Program* prog, | |||
| GetKernelKey(op_item, place, map_value_pair, op_info_parser.get()); | |||
| VLOG(6) << "kernel type " << kernel_key; | |||
|
|
|||
| if (op_item->name() == "pd.load_combine") { | |||
| kernel_key.set_dtype(phi::DataType::FLOAT32); | |||
There was a problem hiding this comment.
这里为什么直接设置 kernel_key 为FP32了？我看load_combile 是支持很多其他数据类型的？是先验证了FP32，还是后面逻辑会自适应的修改？
若是前者，这里是否需要加一下 TODO 标记下？
There was a problem hiding this comment.
load_combine 是一个没有输入的op，无法从输入的参数表里面推导输出的类型，但是kernel key不是有一个类型
… support_new_ir_load_combine
… support_new_ir_load_combine
… support_new_ir_load_combine
PR types
New features
PR changes
Others
Description
新IR 支持load combine op
Others
Pcard-67164