fix: fix infershape profile #3240
Conversation
test=develop
MyPandaShaoxiang
force-pushed
the
smart_infer_shape
branch
from
d9704c7 to
46b7f03
Compare
| @@ -65,6 +65,7 @@ class OpLite : public Registry { | |||
| virtual bool CheckShape() const { return true; } | |||
| // Inference the outputs' shape. | |||
| virtual bool InferShape() const { return true; } | |||
| virtual bool SmartInferShape() { return this->InferShape(); } | |||
There was a problem hiding this comment.
不需要额外增加一个 SmartInferShape，按需修改单个 Op 特定的 InferShape 实现就可以了
| @@ -150,6 +151,10 @@ class OpLite : public Registry { | |||
| std::vector<Place> valid_places_; | |||
| Place kernel_place_{TARGET(kHost), PRECISION(kFloat)}; | |||
| std::unique_ptr<OpInfo> op_info_; | |||
| std::vector<DDimLite> last_input_shapes; | |||
There was a problem hiding this comment.
按需修改瓶颈Op 的 InferShape 即可，不要在积累加 member，这样所有创建的 Op 都带上这个 member 了。
在实际需要修改的 Op 上加这个 member，其他非瓶颈 op，就不需要改了。
如果这个升级对所有 op 都适用，再考虑加到基类上
There was a problem hiding this comment.
size_t io_shape_lod_hash_{};
存储上次执行前的 input shape, output shape, lod 相关的 hash combined.
for (int ele in elements) io_shape_lod_hash = hash_combine(io_shape_lod_hash , ele);
if (new_hash != io_shape_lod_hash_) {
InferShape();
}
|
|
||
| this->InferShape(); | ||
|
|
||
| if (!last_input_shapes.empty()) { |
There was a problem hiding this comment.
这里跟基类里面已经有的逻辑有啥区别吗？ 为啥要再实现一次？
| @@ -65,6 +65,7 @@ class OpLite : public Registry { | |||
| virtual bool CheckShape() const { return true; } | |||
| // Inference the outputs' shape. | |||
| virtual bool InferShape() const { return true; } | |||
There was a problem hiding this comment.
后续可以考虑 InferShape -> InferShapeImpl
SmartInferShape -> InferShape
| @@ -150,6 +151,10 @@ class OpLite : public Registry { | |||
| std::vector<Place> valid_places_; | |||
| Place kernel_place_{TARGET(kHost), PRECISION(kFloat)}; | |||
| std::unique_ptr<OpInfo> op_info_; | |||
| std::vector<DDimLite> last_input_shapes; | |||
There was a problem hiding this comment.
size_t io_shape_lod_hash_{};
存储上次执行前的 input shape, output shape, lod 相关的 hash combined.
for (int ele in elements) io_shape_lod_hash = hash_combine(io_shape_lod_hash , ele);
if (new_hash != io_shape_lod_hash_) {
InferShape();
}
| last_output_shapes.clear(); | ||
| last_output_lods.clear(); | ||
| } | ||
| last_output_shapes.push_back(param_.output->dims()); |
No description provided.