Message boards :
Number crunching :
Does the CPU app & GPU app use the same floating-point format?
Message board moderation
Author | Message |
---|---|
Send message Joined: 11 Nov 07 Posts: 232 Credit: 178,229,009 RAC: 0 |
Just wondering if the CPU application use the same floating-point format as the GPU application? Or is the CPU calculation done by using Extended_precision and the GPU calculation is done by using Double_precision_floating-point_format? |
Send message Joined: 8 May 10 Posts: 576 Credit: 15,979,383 RAC: 0 |
Just wondering if the CPU application use the same floating-point format as the GPU application?No. The extended precision is a bad feature of the x87 FPU. I've been avoiding it like the plague. Not only is the x87 FPU slower than using SSE2, it adds inconsistency. The final results depend on whether intermediate results are kept in a register or not, so you get slightly different results between different compilers and between debug and optimized builds, which is just more difficult to work with and test. When it does have to be used (on 32 bit x86 systems, except for the critical part which uses SSE2+ if available), it currently explicitly turns off the extended precision. |
©2024 Astroinformatics Group