How to use negative leading dimension in cuBLASLt matmul interface?

I’m using CUDALibrarySamples provided on Github. CUDALibrarySamples/cuBLASLt/LtSgemm/main.cpp at master · NVIDIA/CUDALibrarySamples · GitHub

I changed the code to the following, which is to change ld_a to -m, and CUDA returned the cuBLAS API failed with status 7

LtSgemm(props.ltHandle,
                CUBLAS_OP_N,
                CUBLAS_OP_N,
                props.m,
                props.n,
                props.k,
                &props.alpha,
                props.Adev,
                -props.m,
                props.Bdev,
                props.k,
                &props.beta,
                props.Cdev,
                props.m,
                props.workspace,
                props.workspaceSize);

For now, I can identify that in CUDALibrarySamples/cuBLASLt/LtSgemm/sample_cublasLt_LtSgemm.cu at master · NVIDIA/CUDALibrarySamples · GitHub

checkCublasStatus(cublasLtMatmulAlgoGetHeuristic(ltHandle, operationDesc, Adesc, Bdesc, Cdesc, Cdesc, preference, 1, &heuristicResult, &returnedResults)); returned an error.

Also using LtSgemmCustomFind, I added following code to select LD NEGATIVE support algo

cublasLtMatmulAlgoCapGetAttribute(&algo, CUBLASLT_ALGO_CAP_LD_NEGATIVE, &ld_negative, sizeof(ld_negative), &sizeWritten);
if (ld_negative) {
    printf("Algo %d support negative ld\n", idx);
} else {
    printf("Algo %d not support negative ld\n", idx);
    continue;
}

and force any check to be true, then I got also status 7 from cublasLtMatmul interface.

I want to know what negative lead dimension means in cuBLASLt? How will the matmul interface read a matrix with a negative ld? Is there any restriction to use negative ld in this interface?

Thanks for your kind help!