Regression on GEMM allocation

[mratsim]

Regression on GEMM allocation

[mratsim]

Commenting out extras bench like OpenBLAS and compiling with nim cpp -r -d:useSysAssert -d:useGcAssert --stacktrace:on -d:release -o:build/gemm_f32_omp benchmarks/gemm/gemm_bench_float32.nim

we get [SYSASSERT] dealloc: object header corrupted (e898f02)
or [SYSASSERT] result wrong pointer! (current master c4ac93d) depending on the commit we are on.

[mratsim]

It can be reproduced with full debug:

A matrix shape: (M: 1920, N: 1920)
B matrix shape: (M: 1920, N: 1920)
Output shape: (M: 1920, N: 1920)
Required number of operations: 14155.776 millions
Required bytes:                   29.491 MB
Arithmetic intensity:            480.000 FLOP/byte
Theoretical peak single-core:    224.000 GFLOP/s
Theoretical peak multi:         4032.000 GFLOP/s
Make sure to not bench Apple Accelerate or the default Linux BLAS.
15052863
3932223
15052863
3932223
15052863
3932223
15052863
3932223
Traceback (most recent call last)
gemm_bench_float32.nim(346) gemm_bench_float32
gemm_bench_float32.nim(181) benchLaserGEMM
gemm.nim(215)            gemm_strided
gemm_tiling.nim(341)     newTiles
alloc.nim(929)           alloc
alloc.nim(795)           rawAlloc
alloc.nim(596)           getBigChunk
alloc.nim(211)           removeChunkFromMatrix2
SIGSEGV: Illegal storage access. (Attempt to read from nil?

The crash is in the second allocShare at the start of the 6th benchmark iteration

[mratsim]

Fixed in #28, the newTiles proc wasn't getting mc, nc and kc from partitionMNK in the correct order