[PIR] pir onednn support mixed instruction #60754
Conversation
|
你的PR提交成功，感谢你对开源项目的贡献! |
… pir_onednn_mixedinstr
… pir_onednn_mixedinstr
… pir_onednn_mixedinstr
| gpu_pm.Run(pir_program_.get()); | ||
| } else { | ||
| ::pir::PassManager cpu_pm(::pir::IrContext::Instance(), 2); | ||
| cpu_pm.AddPass(::pir::CreateReplaceFetchWithShadowOutputPass()); |
| data_format_tensors : x, out, mid_out, out_grad | ||
|
|
||
| - op : pad3d | ||
| extra_args : |
There was a problem hiding this comment.
没有extra_args仍需要指定这个key吗？
There was a problem hiding this comment.
这个底层能力上是可以不指定的。但是我写这里的时候，想了一下，还是想放个空的。大多数op都有extra_args，写一个空的extra_args更能让看代码的人一眼就知道他的extra_args是空的。
test/ir/inference/auto_scan_test.py
Outdated
| dic["use_gpu"] = enable_gpu | ||
| return str(dic) | ||
|
|
||
| def pir_run_test_config( |
There was a problem hiding this comment.
给MkldnnAutoScanTest派生出PirMkldnnAutoScanTest是不是更好点？而不是给MkldnnAutoScanTest增加新成员函数。之后的单测可继承PirMkldnnAutoScanTest
There was a problem hiding this comment.
能复用的代码，还是建议抽出来尽量复用下，减少后期维护成本
| #ifdef PADDLE_WITH_DNNL | ||
| auto new_in_type = new_in.type(); | ||
| if (new_in_type.isa<AllocatedDenseTensorType>()) { | ||
| if (new_in_type.dyn_cast<AllocatedDenseTensorType>().data_layout() == | ||
| phi::DataLayout::ONEDNN) { | ||
| new_in = AddOneDNN2PaddleLayoutTransferOp( | ||
| new_in, phi::DataLayout::ANY, block); | ||
| } | ||
| } | ||
| #endif |
There was a problem hiding this comment.
这个不太合理吧，大家编译的时候带上WITH_MKLDNN，但是不跑mkldnn模式，这里会影响其他模式逻辑？
There was a problem hiding this comment.
看错了，有这个判断不会的if (new_in_type.dyn_cast().data_layout() == phi::DataLayout::ONEDNN)
There was a problem hiding this comment.
在不跑mkldnn模式的时候，new_in_type.dyn_cast().data_layout() == phi::DataLayout::ONEDNN 为 false。所以不会影响其它模式。但位置在当前pd_op_to_kernel_pass.cc的代码框架下，只能放在这里。
| #ifdef PADDLE_WITH_DNNL | ||
| for (auto& ins : vec_instruction_base_) { | ||
| if (static_cast<OneDNNPhiKernelInstruction*>(ins.get()) != nullptr || | ||
| static_cast<OneDNNLegacyKernelInstruction*>(ins.get()) != nullptr || |
There was a problem hiding this comment.
C++11 有 std::is_base_of<A, B>::Value ，另外只需要判断是否是OnDNN的基类是否就可以了？
… pir_onednn_mixedinstr
5 participants
PR types
New features
PR changes
Others
Description
Pcard-67164
PIR适配OneDNN 支持OneDNNMixedPhiKernelInstruction，并支持pad3d。
此外还做了如下处理：