[PIR][oneDNN] Add matmul_activation_fuse_pass #62901
Conversation
|
你的PR提交成功，感谢你对开源项目的贡献! |
LLee233
force-pushed
the
matmul_act_pass
branch
4 times, most recently
from
e26289f to
ff633b9
Compare
|
Hi @onecatcn，这个PR的CI已经过了，可以麻烦帮忙请相应的owner来review一下吗？Thx~ |
| return true; | ||
| }); | ||
| pat.RequireNativeCall([&](const paddle::drr::MatchContext &match_ctx) { | ||
| auto result_gelu = match_ctx.Attr<bool>("approximate"); |
There was a problem hiding this comment.
approximate为false就不能匹配吗？
There was a problem hiding this comment.
我把两种Gelu的情况拆开来写了，这个是新加的一个MatmulGeluTanhPattern，用来单独表示approximate为true时可以fuse的pattern，approximate 为false的时候是跟MatmulActivationFusePattern一起的，会进matmul+gelu_erf fusion。
LLee233
force-pushed
the
matmul_act_pass
branch
from
5776f37 to
e4cd1f1
Compare
| pir::RewritePatternSet ps(context); | ||
| // std::vector<bool> bool_set = {false, true}; | ||
| int benefit_idx = 1; | ||
| for (auto act_op : act_ops) { |
There was a problem hiding this comment.
| for (auto act_op : act_ops) { | |
| for (const auto& act_op : act_ops) { |
另外关于benefit_idx的设置可能有些问题，从代码逻辑上看，从上到下依次添加的这些pattern，他们的benefit_idx是依次变大的，benefit_idx值越大的pattern，我们期望是越优先被应用，以MatmulClipFusePattern和FusedMatmulClipFusePattern来说，我们期望MatmulClipFusePattern比FusedMatmulClipFusePattern先应用，因此MatmulClipFusePattern的benefit_idx应该较大才对
There was a problem hiding this comment.
从我的理解来看，因为这个pass目前matmul相关pass里不属于第一个pass，也就是说，在之后遇到fused_matmul的可能性较大，所以这种情况下优先match fused_matmul会不会更好一点？
There was a problem hiding this comment.
不过如果是以单个pattern为准在整个graph上取多个符合要求的pattern，那确实应该是原生matmul+act被match到的个数更多的可能性较大
There was a problem hiding this comment.
好的，那我就先不改了哈。如果之后pass order有变化或者说还是需要以非fused_matmul的pattern为主我再随时更改~
| paddle::dialect::MatmulOp::name(), | ||
| paddle::onednn::dialect::FusedMatmulOp::name(), | ||
| benefit_idx++)); | ||
| for (auto act_op : act_ops) { |
There was a problem hiding this comment.
| for (auto act_op : act_ops) { | |
| for (const auto& act_op : act_ops) { |
|
Hi @wanghuancoder , @XiaoguangHu01 可以麻烦帮忙review/approve一下吗？Thx~ |
5 participants
PR types
New features
PR changes
Others
Description
Based on new pass mechanism, here we add pass "matmul_activation_fuse_pass" for PIR.
The new pass is same as "matmul_activation_mkldnn_fuse_pass" in "/paddle/fluid/framework/ir/mkldnn/matmul_activation_mkldnn_fuse_pass.cc"