cuBLAS works with 11.2, but not with 11.3 on RTX 3080 Mobile. On A100 both work

Accelerated Computing GPU-Accelerated Libraries
1

Hello developers-community,

I use the following cubLASS call:

    c = cublasGemmEx(cublas,
        cublasOperation_t::CUBLAS_OP_N,
        cublasOperation_t::CUBLAS_OP_N,
        src_h,
        src_w,
        src_w,
        &alpha,
        dev_x,
        cublasDataType_t::CUDA_R_16F,
        src_w,
        dev_y,
        cublasDataType_t::CUDA_R_16F,
        src_w,
        &beta,
        dev_z,
        cublasDataType_t::CUDA_R_16F,
        src_w,
        cublasComputeType_t::CUBLAS_COMPUTE_16F,
        cublasGemmAlgo_t::CUBLAS_GEMM_DEFAULT);

on RTX 3080 Mobile with cuda 11.2 it works perfectly, but on 11.3 it returns following error.

On entry to GEMM_EX parameter number 9 had an illegal value

All input matrices are half and CUDA_R_16F is valid.

The same code runs perfectly on A100 with 11.3

Why RTX 3080 doesn’t work with 11.3 and cublass?

Thank you!

2

Do you mind providing an complete example we can run?

3

Upgrading directly to cuda 11.3.1 fixed the issue from cuda 11.3.0.

Thank you!