Optimize Maxpooling forward/backward: Use a more appropriate workgroup size and use registers to optimize performance

[chunhuanMeng]

Use a more appropriate workgroup size and use registers to optimize performance

[chunhuanMeng]

Maxpooling3d

dtype	Shape (N, C, H, W, D, oH, oW, oD, channels_last)	Before	With PR
bf16	16, 32, 64, 64, 64, 32, 32, 32, False	801.072us	818.840us
bf16	1, 4, 144, 144, 144, 72, 72, 72, False	91.756us	69.788us
bf16	512, 512, 12, 12, 12, 6, 6, 6, False	16.053ms	2.542ms
bf16	16, 32, 64, 64, 64, 32, 32, 32, True	2.342ms	1.093ms
bf16	1, 4, 144, 144, 144, 72, 72, 72, True	112.492us	81.068us
bf16	512, 512, 12, 12, 12, 6, 6, 6, True	33.218ms	3.645ms
fp32	16, 32, 64, 64, 64, 32, 32, 32, False	818.772us	809.988us
fp32	1, 4, 144, 144, 144, 72, 72, 72, False	90.272us	66.144us
fp32	512, 512, 12, 12, 12, 6, 6, 6, False	15.779ms	2.482ms
fp32	16, 32, 64, 64, 64, 32, 32, 32, True	2.433ms	1.088ms
fp32	1, 4, 144, 144, 144, 72, 72, 72, True	118.156us	77.212us
fp32	512, 512, 12, 12, 12, 6, 6, 6, True	32.778ms	3.644ms

Maxpooling3d_backward

dtype	Shape (N, C, H, W, D, oH, oW, oD, channels_last)	Before	With PR
bf16	16, 32, 64, 64, 64, 32, 32, 32, False	770.588us	744.892us
bf16	1, 4, 144, 144, 144, 72, 72, 72, False	66.820us	42.584us
bf16	512, 512, 12, 12, 12, 6, 6, 6, False	13.586ms	2.366ms
bf16	16, 32, 64, 64, 64, 32, 32, 32, True	1.339ms	825.320us
bf16	1, 4, 144, 144, 144, 72, 72, 72, True	94.536us	51.188us
bf16	512, 512, 12, 12, 12, 6, 6, 6, True	23.022ms	2.722ms
fp32	16, 32, 64, 64, 64, 32, 32, 32, False	1.241ms	1.252ms
fp32	1, 4, 144, 144, 144, 72, 72, 72, False	37.060us	20.936us
fp32	512, 512, 12, 12, 12, 6, 6, 6, False	12.595ms	5.450ms
fp32	16, 32, 64, 64, 64, 32, 32, 32, True	2.860ms	1.902ms
fp32	1, 4, 144, 144, 144, 72, 72, 72, True	57.360us	20.672us
fp32	512, 512, 12, 12, 12, 6, 6, 6, True	21.412ms	6.434ms