How to call cuda functions that are not implemented in CUDA.jl?

Specific Domains GPU
1

If there is a function in CUDA, but have not been implemented in CUDA.jl, is it possible to call that function using ccall?

I am trying it with a call to dgemm ( gemm is available through CUDA.jl, but I want to try a function that’s been implemented to validate the result), using the following script, the return value is CUBLAS_STATUS_SUCCESS, however, I don’t have a way to get the result, and the output indicates that the function didn’t actually succeed.

using LLVM
n = 10000

x = rand(n,n)
y = zeros(n,n)
alpha = rand()

handle = CUBLAS.cublasCreate_v2()

# CUBLAS.handle()
LLVM.@runtime_ccall((:cublasDgemm, CUDA.libcublas()), CUBLAS.cublasStatus_t, (CUBLAS.cublasHandle_t, Char, Char,
                                            Cint, Cint, Cint, 
                                            Cdouble, Ptr{Cdouble}, Ptr{Cdouble}, Cdouble, Ptr{Cdouble}, Cint),
                                            handle,  'N','N', n,n,n, alpha, x,x, alpha, y, n)

Here is the output:

# ** On entry to DGEMM  parameter number 1 had an illegal value
# CUBLAS_STATUS_SUCCESS::cublasStatus_t = 0x00000000

The code probably is not correct. Any help are appreciated.

Thanks.

2

Functions that aren’t wrapped in CUDA.jl are still available through their autogenerated wrappers. So you can just call CUBLAS.cublasDgemm. However, the actual name is cublasDgemm_v2. And that function is being wrapped already, as CUBLAS.gemm! (see wrappers.jl in CUDA.jl/lib/cublas).

EDIT: didn’t read your question properly.

If the returned status is 0, everything was OK; where do you get the other error from?
The ‘result’ here is the fact that this call modifies one of its inputs.

3 

julia> LLVM.@runtime_ccall((:cublasDgemm, CUDA.libcublas()), CUBLAS.cublasStatus_t, (CUBLAS.cublasHandle_t, Char, Char,
                                                   Cint, Cint, Cint, 
                                                   Cdouble, Ptr{Cdouble}, Ptr{Cdouble}, Cdouble, Ptr{Cdouble}, Cint),
                                                   handle,  'N','N', n,n,n, alpha, x,x, alpha, y, n)
 ** On entry to DGEMM  parameter number 1 had an illegal value
CUBLAS_STATUS_SUCCESS::cublasStatus_t = 0x00000000

The output just prints on screen after calling the function. I think it’s coming from CUDA, rather than Julia.

4

I called this function:

julia> @time CUDA.@sync res = CUBLAS.cublasDgemm_v2(handle, 'T', 'N', n,n,n, alpha, x, n, x, n, alpha, y, n)
ERROR: MethodError: Cannot `convert` an object of type Char to an object of type CUDA.CUBLAS.cublasOperation_t
Closest candidates are:
  convert(::Type{T}, ::T) where T at essentials.jl:171
  CUDA.CUBLAS.cublasOperation_t(::Any) where T<:CEnum.Cenum at /home/hu/.julia/packages/CEnum/daZUb/src/operators.jl:19

So how to convert Char to CUDA.CUBLAS.cublasOperation_t ? Thank you much for your reply!

5

Look at how the existing wrappers do that. On CUDA master that convert method is available, for the version you’re using it’ll be explicit still.

1 Like
6

The util section shows that convert function is overloade, But it shows the same error:

julia> CUDA.convert(CUBLAS.cublasOperation_t, 'T')
ERROR: MethodError: Cannot `convert` an object of type Char to an object of type CUDA.CUBLAS.cublasOperation_t
Closest candidates are:
  convert(::Type{T}, ::T) where T at essentials.jl:171
  CUDA.CUBLAS.cublasOperation_t(::Any) where T<:CEnum.Cenum at /home/xiaoqihu/.julia/packages/CEnum/daZUb/src/operators.jl:19
Stacktrace:
 [1] top-level scope at REPL[7]:100:
7

And that works as expected:

julia> convert(CUBLAS.cublasOperation_t, 'T')
CUBLAS_OP_T::cublasOperation_t = 0x00000001

As I said, you need to check for the version of CUDA.jl you’re using. It used to be with a custom function, cublasop, and has been converted to more “Julian” interfaces nowadays. Just look at how CUBLAS.gemm! is implemented.

8

e.g. https://github.com/JuliaGPU/CUDA.jl/blob/1c399bf2ed10624f3961d073f07d55b1a30184ef/lib/cublas/wrappers.jl#L818 on 1.1

9

Thank you very much @maleadt! yes, the gemm! shows how to do it.