[PIR] convert pd_op.sum to cinn_op.reduce_sum #58207
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. |
… replace_sum_to_cinn_reduce_sum
… replace_sum_to_cinn_reduce_sum
… replace_sum_to_cinn_reduce_sum
… replace_sum_to_cinn_reduce_sum
… replace_sum_to_cinn_reduce_sum
| // See the License for the specific language governing permissions and | ||
| // limitations under the License. | ||
|
|
||
| #include "paddle/cinn/hlir/dialect/operator/transforms/pd_op_to_cinn_op_convert_pass.h" |
There was a problem hiding this comment.
文件名是不是可精简为：pd_to_cinn_pass.h，因为目前我们没有Kernel层的转换需求，所以默认是Op层面的。或者pd_to_cinn_op_pass.h、pd_op_to_cinn_pass.h
| namespace dialect { | ||
| namespace ir { | ||
|
|
||
| class PDSum2CINNReduceSumPattern |
There was a problem hiding this comment.
| class PDSum2CINNReduceSumPattern | |
| class SumOpPattern |
是不是可以约定下这个命名范式，因为源Dialect是pd_op，所以我们只需要以其为命名即可？即使后续出现一对多的场景，这里的class name也会比较简洁。具体的映射到哪个算子，看code也是很清楚的。
| public: | ||
| void operator()(pir::drr::DrrPatternContext *ctx) const override { | ||
| // Source Pattern | ||
| pir::drr::SourcePattern pat = ctx->SourcePattern(); |
There was a problem hiding this comment.
| pir::drr::SourcePattern pat = ctx->SourcePattern(); | |
| pir::drr::SourcePattern patttern = ctx->SourcePattern(); |
pat 作为 pattern 缩写似乎比较少见？不像idx这些
| void operator()(pir::drr::DrrPatternContext *ctx) const override { | ||
| // Source Pattern | ||
| pir::drr::SourcePattern pat = ctx->SourcePattern(); | ||
| const auto &full_int_array = pat.Op("pd_op.full_int_array", |
There was a problem hiding this comment.
| const auto &full_int_array = pat.Op("pd_op.full_int_array", | |
| const auto &full_int_array = pat.Op(paddle::dialect::FullIntArrayOp::name(), |
我们应该尽量避免直接使用 op_name_str
| {"place", pat.Attr("place_2")}}); | ||
|
|
||
| const auto &sum = pat.Op( | ||
| "pd_op.sum", |
| } | ||
|
|
||
| bool PdOpToCinnOpPass::CanApplyOn(pir::Operation *op) const { | ||
| return op->name() == "builtin.module" && op->num_regions() > 0; |
There was a problem hiding this comment.
| return op->name() == "builtin.module" && op->num_regions() > 0; | |
| return op->isa<ModuleOp>() && op->num_regions() > 0; |
… replace_sum_to_cinn_reduce_sum
| } | ||
| }; | ||
|
|
||
| PdOpToCinnOpPass::PdOpToCinnOpPass() : pir::Pass("PdOpToCinnOpPass", 1) {} |
There was a problem hiding this comment.
关于pass name的命名规范，之前的想法是统一使用小写+下划线的形式，我已提了统一修改之前pass的pr，见#58205 辛苦这里改一下？
paddle/phi/infermeta/nullary.cc
Outdated
| @@ -41,13 +41,11 @@ void CreateInferMeta(const IntArray& shape, DataType dtype, MetaTensor* out) { | |||
| CreateInferMetaBase(shape.GetData(), dtype, DataLayout::NCHW, out); | |||
| } | |||
|
|
|||
| void CreateIntArrayInferMeta(const IntArray& data, | |||
| void CreateIntArrayInferMeta(const std::vector<int64_t>& shape, | |||
There was a problem hiding this comment.
CreateIntArrayInferMeta的函数名是不也需要调整下？
| const auto &new_perm_attr = | ||
| res.Attr([](const pir::drr::MatchContext &match_ctx) -> phi::IntArray { | ||
| auto shape = | ||
| match_ctx.Attr<std::vector<int64_t>>("expand_shape_value"); | ||
|
|
||
| return phi::IntArray(shape); | ||
| }); | ||
| const auto &full2 = res.Op("pd_op.full", | ||
| {{"shape", pat.Attr("expand_shape_value")}, | ||
| {{"shape", new_perm_attr}, |
There was a problem hiding this comment.
full的shape参数还是vector<int64_t>类型的，这里换成int_array是用来测试吗？
There was a problem hiding this comment.
这个类似我理解和参数表的里面的不是一样的，主要看底层的attribute map里面存的是啥， 这个full 里面底层存的是int array
… replace_sum_to_cinn_reduce_sum
… replace_sum_to_cinn_reduce_sum
… replace_sum_to_cinn_reduce_sum
… replace_sum_to_cinn_reduce_sum
* update * update * update * update * fix bug * add test flag * fix bug * update * fix cmake bug * remove cinn_op header * fix full int array bug * fix vjp gene bug * fix bug * fix bug * polish code * update * update * update * fix bug * fix bug
PR types
Others
PR changes
Others
Description
由于编译器中间部分算子定义和主框架不一样，主框架的reduce 系列，支持axis为可变attribute，但是编译器中当前不支持axis 为可变attribute，因此做了一个变换映射
对于full int array内部存储的phi::IntArray 改成vector,方便drr的适配，这个算子仅在新IR中使用，不存在兼容性问题
Pcard-67164