Is there anyway to restrict the precision of float-point precision to a higher level with dpcpp?

[nahso]

Hello,

I am using dpcpp to compute the sparse matrix-matrix multiplication(spmm) in sequential minimal optimization(SMO), but the result is not precise enough. I specified -fp-model=double, -fp=model=restrict, -fimf-precision=high, -fimf-max-error=0.1, -fimf-accuracy-bits=52 one by one, and used sycl::gpu_selector and sycl::cpu_selector, but the final result of SMO was not right.

Here is the spmm code with dpcpp:

q.submit([&](sycl::handler& h) {
h.parallel_for(sycl::nd_range<2>(global_range, local_range), [=](sycl::nd_item<2> it)
{
auto i = it.get_global_id(0);
auto jj = it.get_global_id(1);
float sum = 0;
for (size_t j = row_ptr_usm[i]; j < row_ptr_usm[i + 1]; ++j) {
sum += val_usm[j] * dense_trans[col_ptr_usm[j] * n + jj];
}
res[i * n + jj] = sum;
});
}).wait();

[NoorjahanSk_Intel]

Hi,

Thanks for reaching out to us.

Could you please try with below compiler option which enables precision?

"-fp-model = precise"

If we use this option  (-fp-model=precise) , it implies -fimf-precision=high.

Please refer to below link for more details:

https://www.intel.com/content/www/us/en/develop/documentation/oneapi-dpcpp-cpp-compiler-dev-guide-and-reference/top/compiler-reference/compiler-options/floating-point-options/fp-model-fp.html

If this doesn't resolve your issue then please provide us with below details

• Complete reproducer code and steps to reproduce the issue.
• Expected and observed results. 
• Hardware details(device you are using to run your code) and compiler version being used.

Thanks & Regards,

Noorjahan

[nahso]

No, I tried -fp-model = precise with O0, O2, O3, Ofast, but none of these flags make the final result precious enough.

[NoorjahanSk_Intel]

Hi,

Thanks for the update.

Could you please provide us with the following details so that we can try it from our end?

• Complete reproducer code and steps to reproduce the issue.
• Expected and observed results. 
• Hardware details(the device you are using to run your code) and compiler version being used.

Thanks & Regards,

Noorjahan.

[NoorjahanSk_Intel]

Hi,

We haven't heard back from you. Could you please provide an update on your issue?

Thanks & Regards,

Noorjahan.

[nahso · Accepted solution]

Hi,

I'm sorry for no reply, we reported this issue to the competition team, and they modified the precision requirement just before. So current precision is OK. Thank you!

[NoorjahanSk_Intel]

Hi,

Glad to know that your issue is resolved. If you need any additional information, please post a new question as this thread will no longer be monitored by Intel.

Thanks & Regards,

Noorjahan.