[pytorch upstream] pointwise is slow compare to cuda

[jianyizh]

Describe the issue

in torchbench model pyhpc_equation_of_state inference, the whole model is fused into a large pointwise kernel, and it's slow compare to a100. This model is faster than a100 in eager mode. FP64 maybe a reason, but I manually change it to fp32, and performance is still low.
pointwise_test.zip
| | 1550 | A100 |
|fp64| 0.77ms | 0.43ms|
|fp32| 0.30ms | 0.08ms|

Environment details

triton 3.3.0+git0bcc8265
pytorch: 3ed5f1fb77669c8ac5d02e7acc0218e31b71c0b6