From 964fe7e672690be5f1f380673126a1f82f3253e2 Mon Sep 17 00:00:00 2001 From: "Tian Zheng (Engrg-Hardware 1)" Date: Tue, 12 Dec 2023 15:40:07 +0000 Subject: [PATCH] Add comments --- paddle/phi/api/yaml/fused_ops.yaml | 3 +++ 1 file changed, 3 insertions(+) diff --git a/paddle/phi/api/yaml/fused_ops.yaml b/paddle/phi/api/yaml/fused_ops.yaml index 328c9ee83d29a7..d346ae4d36b7cc 100644 --- a/paddle/phi/api/yaml/fused_ops.yaml +++ b/paddle/phi/api/yaml/fused_ops.yaml @@ -383,6 +383,9 @@ func : layer_norm_act_xpu data_type : x +# This op is implemented using CUDNN Frontend API, which serves as a supplement to +# legacy max pooling implementation. It shows better performance with NHWC layout and +# half precision. - op : max_pool2d_v2 args : (Tensor x, int[] kernel_size, int[] strides= {1, 1}, int[] paddings = {0, 0}, str data_format = "NCHW", bool global_pooling = false, bool adaptive = false) output : Tensor(out), Tensor(saved_idx)