[AMD GPU] Performance optimization

[chhzh123]

A thread to record the weekly progress on the performance optimization project

[taejoonest]

11/27

M = 4096, N = 4096, K = 4096:
sgemm_naive: 686ms
sgemm_coalesced: 449ms
sgemm_shared_memory:82ms
sgemm_1dtile:52ms

sgemm_naive -> sgemm_coalesced: Converting 2D threads to 1D threads
sgemm_coalesced -> sgemm_shared_memory: Dividing computations into smaller tiles of shared memory (equal to block size), and repeating.
sgemm_shared_memory -> sgemm_1dtile:

Note on Dimension: cuBLAS uses each block as 32 x 32. In this example, we use 64 x 16.