[webgpu]: optimize pool operators

[xhcao]

The patch optimizes pool operators when output size is small and kernel size is big

Description

Motivation and Context

[xhcao]

The issue is from an user's model.
If the input data shape is [1, 64, 128, 128] (NCHW), the original code needs one work group for 64 output elements. The kernel size is 128128, so there is a 128128 loop.
If the output size is small, one work group is responsible for one output element, there is only 128 loop.
@jchen10 PTAL

[guschmue]

/azp run Linux QNN CI Pipeline,Win_TRT_Minimal_CUDA_Test_CI,Windows ARM64 QNN CI Pipeline,Windows GPU Doc Gen CI Pipeline,Windows x64 QNN CI Pipeline

[azure-pipelines]

Azure Pipelines successfully started running 5 pipeline(s).

[xhcao]

#23614

[guschmue]

this ci error is unrelated to this PR - fix incoming

[vadimkantorov]

does this PR also introduce a shortcut when the input size is equal to kernel size - same code path as simply taking a mean?

[xhcao]

does this PR also introduce a shortcut when the input size is equal to kernel size - same code path as simply taking a mean?

Hi, @vadimkantorov The path is valid when

onnxruntime/onnxruntime/core/providers/webgpu/nn/pool.cc

Line 287 in 7d71975

bool are_small_output_big_kernel = output_size <= 128 && kernel_size >= 128;

, when the kernel size is large and the output size is small.
same code path as simply taking a mean?, Do you mean AveragePool and GlobalAveragePool? The path is also valid for these operators.

[vadimkantorov]

Yes, I mean global pooling, where output_size is strictly 1x1, while input_size (matching kernel_size as in #23614) can be 100 or anything else.

I guess output_size = 1, output_size = kernel_size = 100 would not pass this check, right?

[xhcao]

ut_size = kernel_size = 100 w

In #23614, the input shape is [1, 56, 80, 128], the kernel shape is [80, 128], so the output shape size is 56 (< 128), the kernel size is 80 * 128 (> 128), which will pass this check.