GroupedGEMM more bigger tiles.

[aosewski]

I've added some more bigger tiles' configurations.

@zjing14 Please note that I've removed the smallest tiles (16x16) as they exhibit the worst performance in tested scenarios.

[zjing14]

@aosewski Any performance improve with larger tiles?

[aosewski]

For row 12 (irregular mk-nk-mn)

# enlarging tiles first part
Best Perf: 1.03203 ms, 56.1824 TFlops, 268.232 GB/s, DeviceGroupedGemm_Xdl<256, 128, 64, 32, 8, 8, 32, 32, 2, 1>
# enlarging tiles second part
Best Perf: 0.752408 ms, 77.062 TFlops, 367.918 GB/s, DeviceGroupedGemm_Xdl<256, 128, 128, 32, 8, 8, 32, 32, 2, 2>

overall speedup (wrt to table): 11.469ms -> 0.752ms =~ 15.25x

For row 13 (irregular mk-kn-mn)

# enlarging tiles first part
Best Perf: 0.334912 ms, 86.5632 TFlops, 469.635 GB/s, DeviceGroupedGemm_Xdl<256, 64, 128, 32, 8, 2, 32, 32, 1, 2>

speedup (wrt to table) 1.113ms -> 0.3349 =~ 3.32x

[zjing14]

For row 12 (irregular mk-nk-mn)

# enlarging tiles first part
Best Perf: 1.03203 ms, 56.1824 TFlops, 268.232 GB/s, DeviceGroupedGemm_Xdl<256, 128, 64, 32, 8, 8, 32, 32, 2, 1>
# enlarging tiles second part
Best Perf: 0.752408 ms, 77.062 TFlops, 367.918 GB/s, DeviceGroupedGemm_Xdl<256, 128, 128, 32, 8, 8, 32, 32, 2, 2>

overall speedup (wrt to table): 11.469ms -> 0.752ms =~ 15.25x

For row 13 (irregular mk-kn-mn)

# enlarging tiles first part
Best Perf: 0.334912 ms, 86.5632 TFlops, 469.635 GB/s, DeviceGroupedGemm_Xdl<256, 64, 128, 32, 8, 2, 32, 32, 1, 2>

speedup (wrt to table) 1.113ms -> 0.3349 =~ 3.32x

That is awesome. Will merge the PR once it passes CI.