call cublas in cuda kernel and use static link

Accelerated Computing GPU-Accelerated Libraries
1

In my project i need to call cublas in cuda kernel, and use static link with cublas. I write a test code to verify my ideas, but compile failed

#include <cuda_runtime.h>
#include <cublas_v2.h>

extern "C" {
__global__ void testcublas(float *d_B, float *d_A, float* d_refC) {
    cublasHandle_t cb_handle = NULL;
    cublasStatus_t status = cublasCreate(&cb_handle);
    int M = 512, N = 512, K = 512;
    float alpha = 1.0, beta = 1.0;
    cublasSgemm(
                cb_handle, CUBLAS_OP_N, CUBLAS_OP_N,
                N, M, K,
                &alpha, d_B, N, d_A, K,
                &beta,  d_refC, N
               );
}

}

compile:

nvcc -arch=sm_70 -lcublas_static -lculibos --relocatable-device-code=true cublas_call.cu -o cublas.cubin

but got errors:
nvlink error : Undefined reference to ‘cublasCreate_v2’ in ‘/tmp/tmpxft_0001d286_00000000-10_cublas_call.o’
nvlink error : Undefined reference to ‘cublasSgemm_v2’ in ‘/tmp/tmpxft_0001d286_00000000-10_cublas_call.o’

Any suggestions about the program would be very helpful!
thanks in advance.

2

You need to be using CUDA 9.2 or before. This capability is deleted and no longer available in CUDA 10.0 and beyond. On Windows I recommend CUDA 9.1 or before.

You need to link against cublas device library.
You need to link against the device runtime library.

-lcublas_device -lcudadevrt

If you have CUDA 9.2 or before, study the Makefile for simpleDevLibCUBLAS sample code.

3

Thanks Robert!

By the way, would you please talk about the reason of delete that capability in CUDA10.0, Is CUDA10.0 has a better solution for that case?