Argthtjtr

Continuing on my CUDA beginner's adventure, I've been introduced to Thrust, which seems a convenient lib that saves me the hassle of explicit memory (de-)allocation.

I've already tried combining it with a few cuBLAS routines, e.g. gemv, by generating a raw pointer to the underlying storage with thrust::raw_pointer_cast(array.data()) and then feeding this to the routines, and it works just fine.

The current task is to get the inverse of a matrix, and for that I'm using getrfBatched and getriBatched. From the documentation:

cublasStatus_t cublasDgetrfBatched(cublasHandle_t handle,

                                   int n, 

                                   double *Aarray,

                                   int lda, 

                                   int *PivotArray,

                                   int *infoArray,

                                   int batchSize);

where

Aarray - device - array of pointers to <type> array

Naturally I thought I could use another layer of Thrust vector to express this array of pointers and again feed its raw pointer to cuBLAS, so here's what I did:

void test()

{

    thrust::device_vector<double> in(4);

    in[0] = 1;

    in[1] = 3;

    in[2] = 2;

    in[3] = 4;

    cublasStatus_t stat;

    cublasHandle_t handle;

    stat = cublasCreate(&handle);

    thrust::device_vector<double> out(4, 0);

    thrust::device_vector<int> pivot(2, 0);

    int info = 0;

    thrust::device_vector<double*> in_array(1);

    in_array[0] = thrust::raw_pointer_cast(in.data());

    thrust::device_vector<double*> out_array(1);

    out_array[0] = thrust::raw_pointer_cast(out.data());

    stat = cublasDgetrfBatched(handle, 2,

        (double**)thrust::raw_pointer_cast(in_array.data()), 2,

        thrust::raw_pointer_cast(pivot.data()), &info, 1);

    stat = cublasDgetriBatched(handle, 2,

        (const double**)thrust::raw_pointer_cast(in_array.data()), 2,

        thrust::raw_pointer_cast(pivot.data()),

        (double**)thrust::raw_pointer_cast(out_array.data()), 2, &info, 1);

}

When executed, stat says CUBLAS_STATUS_SUCCESS (0) and info says 0 (execution successful), yet if I try to access the elements of in, pivot or out with standard bracket notation, I hit a thrust::system::system_error. Seems to me that the corresponding memory got corrupted somehow.

Anything obvious that I'm missing here?

The documentation for cublas<t>getrfBatched indicates that the infoArray parameter is expected to be a pointer to device memory.

Instead you have passed a pointer to host memory:

int info = 0;

...

stat = cublasDgetrfBatched(handle, 2,

    (double**)thrust::raw_pointer_cast(in_array.data()), 2,

    thrust::raw_pointer_cast(pivot.data()), &info, 1);

                                            ^^^^^

If you run your code with cuda-memcheck (always a good practice, in my opinion, any time you are having trouble with a CUDA code, before asking others for help) you will receive an error of "invalid global write of size 4". This is due to the fact that a kernel launched by cublasDgetrfBatched() is attempting to write the info data to device memory using an ordinary host pointer that you provided, which is always illegal in CUDA.

CUBLAS itself does not trap errors like this for performance reasons. However the thrust API uses more rigorous synchronization and error checking, in some cases. Therefore, the use of thrust code after this error reports the error, even though the error had nothing to do with thrust (it was an asynchronously reported error from a previous kernel launch).

The solution is straightforward; provide device storage for info:

$ cat t329.cu

#include <thrust/device_vector.h>

#include <cublas_v2.h>

#include <iostream>



void test()

{

    thrust::device_vector<double> in(4);

    in[0] = 1;

    in[1] = 3;

    in[2] = 2;

    in[3] = 4;

    cublasStatus_t stat;

    cublasHandle_t handle;

    stat = cublasCreate(&handle);

    thrust::device_vector<double> out(4, 0);

    thrust::device_vector<int> pivot(2, 0);

    thrust::device_vector<int> info(1, 0);

    thrust::device_vector<double*> in_array(1);

    in_array[0] = thrust::raw_pointer_cast(in.data());

    thrust::device_vector<double*> out_array(1);

    out_array[0] = thrust::raw_pointer_cast(out.data());

    stat = cublasDgetrfBatched(handle, 2,

        (double**)thrust::raw_pointer_cast(in_array.data()), 2,

        thrust::raw_pointer_cast(pivot.data()), thrust::raw_pointer_cast(info.data()), 1);

    stat = cublasDgetriBatched(handle, 2,

        (const double**)thrust::raw_pointer_cast(in_array.data()), 2,

        thrust::raw_pointer_cast(pivot.data()),

        (double**)thrust::raw_pointer_cast(out_array.data()), 2, thrust::raw_pointer_cast(info.data()), 1);

    for (int i = 0; i < 4; i++) {

      double test = in[i];

      std::cout << test << std::endl;

      }

}





int main(){



  test();

}

$ nvcc -o t329 t329.cu -lcublas

t329.cu(12): warning: variable "stat" was set but never used



$ cuda-memcheck ./t329

========= CUDA-MEMCHECK

3

0.333333

4

0.666667

========= ERROR SUMMARY: 0 errors

$

You'll note this change in the above code is applied to usage for both cublas calls, as the infoArray parameter has the same expectations for both.