Performance: Evaluate MaxPool2d forward backward performance gap between IPEX.

intel / torch-xpu-ops

Apache License 2.0

30 stars 21 forks source link

Performance: Evaluate MaxPool2d forward backward performance gap between IPEX. #938

Open fengyuan14 opened 1 month ago

fengyuan14 commented 1 month ago

🚀 The feature, motivation and pitch

Backward: Torch-xpu-ops align with CUDA implementation, using deterministic algorithm.
Forward: To investigate the gap, after PTI is fixed.

Alternatives

No response

Additional context

No response

huaiyuzh commented 2 weeks ago

pointnet_bf16_ipex25_1108 is using IPEX max_pool2d, pointnet_bf16_ipex25_1108_stock_pytorch is using stock pytorch max_pool2d. pointnet_bf16_ipex25_1108.zip