Calling cuBLAS from device?

Accelerated Computing GPU-Accelerated Libraries
1

Hi All,

Is there a way to call cuBLAS or its lite version for gemm inside the GPU kernel?

Thanks

2

Not today, but we will be releasing cuBLASDx (Device eXtensions) that will allow this in the future. Check out cuFFTDx to get an idea about the API and implementation

3

Interesting, is there a public timeline for that?
Out of curiosity- would the library call get inlined to the cuda kernel?

4

Interesting, is there a public timeline for that?

No details at the moment

Out of curiosity- would the library call get inlined to the cuda kernel?

Yes

See the later part of the Math Libs GTC presentation

5

If you need in-kernel GEMMs, CUTLASS might be up your alley.