[libcusmm] libcusmm_unittest_transpose test failure

[dev-zero]

[libcusmm] libcusmm_unittest_transpose test failure

[alazzaro]

Are you using CUBLAS flag during the compilation? Such a big kernel should not be there, unless we apply densification and that case we should not transpose...

[dev-zero]

no, I used cmake -DUSE_MPI=OFF -DUSE_CUDA=ON -DWITH_GPU=K20X to build, which leaves CUBLAS turned off by default

[alazzaro]

Ah OK, I see the problem now. The file

https://github.com/cp2k/dbcsr/blob/develop/src/acc/libsmm_acc/libcusmm/parameters_K20X.json

reports "huge" kernels (I think for testing purpose):

{"m": 81, "n": 9, "k": 9, "tile_m": 3, "tile_n": 3, "w": 4, "v": 6, "threads": 128, "grouping": 16, "minblocks": 8, "algorithm": "largeDB1", "perf": 189.642},
{"m": 96, "n": 96, "k": 96, "tile_m": 6, "tile_n": 3, "w": 14, "v": 48, "threads": 512, "grouping": 16, "minblocks": 1, "algorithm": "largeDB1", "perf": 614.588},
{"m": 100, "n": 10, "k": 10, "tile_m": 2, "tile_n": 2, "threads": 256, "grouping": 16, "minblocks": 4, "algorithm": "medium", "perf": 226.917},
{"m": 121, "n": 11, "k": 11, "tile_m": 5, "tile_n": 3, "threads": 128, "grouping": 16, "minblocks": 1, "algorithm": "medium", "perf": 233.211},
{"m": 144, "n": 12, "k": 12, "tile_m": 2, "tile_n": 4, "w": 6, "v": 8, "threads": 288, "grouping": 16, "minblocks": 4, "algorithm": "largeDB1", "perf": 268.209},
{"m": 169, "n": 13, "k": 13, "tile_m": 3, "tile_n": 4, "w": 6, "v": 10, "threads": 256, "grouping": 16, "minblocks": 1, "algorithm": "largeDB1", "perf": 221.427},
{"m": 196, "n": 14, "k": 14, "tile_m": 6, "tile_n": 2, "threads": 256, "grouping": 16, "minblocks": 1, "algorithm": "medium", "perf": 243.838},
{"m": 225, "n": 15, "k": 15, "tile_m": 3, "tile_n": 3, "w": 4, "v": 12, "threads": 384, "grouping": 16, "minblocks": 1, "algorithm": "largeDB1", "perf": 248.307},
{"m": 256, "n": 16, "k": 16, "tile_m": 2, "tile_n": 6, "w": 6, "v": 10, "threads": 384, "grouping": 16, "minblocks": 1, "algorithm": "largeDB1", "perf": 309.19}

@shoshijak
Those kernels are too big to run on the GPU, I wonder how they made it...
I see two solutions:

1. remove those kernels from the file
2. replace the assert with the product of dimensions (808080 as maximum).

Personally, I will go for the first solution...

[shoshijak]

The libcusmm_unittest_transpose test just tests all transposition operations that could arise given the kernels defined in parameters_GPU.json. I see that in parameters_K20X.json (https://github.com/cp2k/dbcsr/blob/develop/src/acc/libsmm_acc/libcusmm/parameters_K20X.json), there are a number of kernels with m, n, or k > 80. I'm guessing they were introduced to the parameter file before the Cannot transpose matrices with dimensions above 80-limitation.

If really libcusmm should not be transposing kernels with a dimension above 80, then these kernels should be removed from parameters_K20X.json.