AMD SI cards - weird results

[skn123]

I have raised an issue in the vexcl repo
ddemidov/vexcl#254
and a solution was provided that solved the issue. Can a similar workaround be provided for ViennaCL also given that a lot of examples are used by vexcl also?

[skn123]

Here is an example of the pronblem

naths@naths-HP-Laptop-15-bs1xx:~/build/viennacl/examples/tutorial$ ./eigen-with-viennacl

Single precision

---

VCL sparsematrix dimensions: 6, 5
Difference for dense matrix-vector product: 42.0595
Difference for dense matrix-vector product (Eigen->ViennaCL->Eigen): 42.0595
Difference for sparse matrix-vector product: 23.6854
Difference for sparse matrix-vector product (Eigen->ViennaCL->Eigen): 23.6854

Double precision

---

VCL sparsematrix dimensions: 6, 5
Difference for dense matrix-vector product: inf
Difference for dense matrix-vector product (Eigen->ViennaCL->Eigen): inf
Difference for sparse matrix-vector product: inf
Difference for sparse matrix-vector product (Eigen->ViennaCL->Eigen): inf

!!!! TUTORIAL COMPLETED SUCCESSFULLY !!!!

naths@naths-HP-Laptop-15-bs1xx:~/build/viennacl/examples/tutorial$ ./eigen-with-viennacl

Single precision

---

VCL sparsematrix dimensions: 6, 5
Difference for dense matrix-vector product: 42.0595
Difference for dense matrix-vector product (Eigen->ViennaCL->Eigen): 42.0595
Difference for sparse matrix-vector product: 23.6854
Difference for sparse matrix-vector product (Eigen->ViennaCL->Eigen): 23.6854

Double precision

---

VCL sparsematrix dimensions: 6, 5
Difference for dense matrix-vector product: 42.0595
Difference for dense matrix-vector product (Eigen->ViennaCL->Eigen): 42.0595
Difference for sparse matrix-vector product: 23.6854
Difference for sparse matrix-vector product (Eigen->ViennaCL->Eigen): 23.6854

!!!! TUTORIAL COMPLETED SUCCESSFULLY !!!!

[karlrupp]

Thanks, @skn123 . This is a terrible bug on AMD's end. Launching an empty kernel after regular kernels is bogus.

I'll see what I can do in terms of a proper workaround.

[skn123]

@karlrupp It will indeed be helpful. However, a temporary fix would be helpful as I need ViennaCL to run with caffe opencl. If a proper workaround is found, then I will share it with other sites / libraries also. The sad part is that the standard answer AMD gives is that "this card is not supported" (when in the README it says otherwise)

[skn123]

@karlrupp any update on this? I am working to fix the OpenCL backend of Caffe also and need this fix from ViennaCL also

[karlrupp]

Not yet - I'll need another week or so.

[skn123]

@karlrupp .. bump

[skn123]

I can confirm that the fix I am using works here also