tomflinda comments

Results 62 comments of


                                            tomflinda

include additional kernel function information

@zjin-lcf Currently, no further action is needed in SYCLomatic; I close this issue.

DPCT1013 rounding modes could not be specified

@zjin-lcf in SYCL 2020 spec, there is no different rounding mode version for one math API.

The bh program hangs

@zjin-lcf have you tried it on cpu device with small configuration, like "./main 10 10"? on my machine, on cpu device "./main 10 10" run passed.

The bh program hangs

@zjin-lcf I have reported this issue to SYCL compiler team to locate the root cause and wait for their feedback.

The bh program hangs

@zjin-lcf these it no further action needed in SYCLomatic side, so I close this issue.

error: assigning to 'void ()(const int , int , int)' from incompatible type 'void (const int , int , int, const sycl::nd_item<3> &, int )': different number of parameters (3 vs 5)

@tomflinda currently not.

performance increase with the always_inline attribute

@zjin-lcf, not sure we can approve that “__attribute__((always_inline))” always gets good performance in SYCL side (inline can cause extra pressure on Registers). We will consider to implement them if that...

performance increase with the always_inline attribute

We have implemented a PR https://github.com/oneapi-src/SYCLomatic/pull/745 to introduce a new option to add inline for kernel function

deviceProp.maxThreadsPerMultiProcessor != deviceProp.get_max_work_items_per_compute_unit() ?

@zjin-lcf we plan to sync with compiler team what the root cause is.

deviceProp.maxThreadsPerMultiProcessor != deviceProp.get_max_work_items_per_compute_unit() ?

@zjin-lcf We have reported this issue to the compiler team and https://github.com/intel/llvm/issues/7997 also track its status, so there is no further action needed in SYCLomatic, I close this issue.