🦎 Triturus 🦎

The following table describes the implemented kernels.

Kernel ID	Description	Operation	Source
vadd	Vector addition	$a_i+b_i$	add
vamax	Vector maximum	$\max_i a_i$	max
vmax	Vector maximum with indices	$(\max_i a_i, \arg\max_i a_i)$	max
matmax	Matrix maximum along one axis	$\max_i a_{ij}$ or $\max_j a_{ij}$	max
mm	Matrix multiplication	$\sum_j a_{ij}b_{jk}$	mm
lm2exp	Batch log-matmul, one matrix in log-space	$\log(\sum_j a_{rij} \exp b_{rjk})$	lm2exp
lt2exp	Batch log-Tucker2, two matrices in log-space	$\log(\sum_{i,j} w_{rsij} \exp a_{rik} \exp b_{rjk})$	lt2exp

Benchmarks Gallery

Kernel ID	Benchmark Description	Baselines	Results
vmax	Vector maximum with and without indices	torch	here
matmax	Matrix maximum along rows and columns	torch	here
mm	Matrix multiplication with square matrices	torch	here
lm2exp	Batch log-matmul, square and rectangular batch matrices	torch + jit	here
lt2exp	Batch log-Tucker2, square and rectangular batch matrices	torch + jit	here

---

Benchmark of vmax

Benchmark of matmax

Benchmark of mm

Benchmark of lm2exp

Benchmark of lt2exp