Too imprecise native_sin/native_cos implementation.

Too imprecise native_sin/native_cos implementation.

Running my OpenCL app on Intel GPUs I encountered precision issue. Results were too imprecise and often failed to pass validation vs CPU-based (or AMD/nVidia GPU based) ones.

Considerable improvement was only when I refused from native_cos/native_sin in oclFFT  (Apple's FFT release for OpenCL) and in other parts of app code.

I have no such issues with same OpenCL code on GPUs from other vendors (both AMD and nVidia).

Please consider some improvements for native trigonometric function presion in future SDK/hardware releases. Too low precision makes them just un-useful no matter how speedy they are. Make precision at least in line with other vendors.

2 帖子 / 0 全新

Native functions should be faster but not necessarily precise. I am not sure how much difference is acceptable. Are you seeing precision issues with native_sin and native_cos only? Thanks for the feedback.