riscv-non-isa/riscv-arch-test

How to generate tests for RV64D and RV64F using riscv-ctg?

Opened this issue · 5 comments

riscvzbs commented

When I use 'riscv_ctg -v debug -d ./tests/ -r -cf ./sample_cgfs/dataset.cgf -cf ./sample_cgfs/rv64i.cgf -bi rv64i -p2,' I can generate tests for RV64I. However, I couldn't find the provided rv64d.cgf and rv64f.cgf. Now I want to generate tests for D and F extensions. How can I do that? Looking forward to your guidance. Thank you very much.

davidharrishmc commented

I would like to know this too, and especially whether there is an easy way to modify rv64f.cfg to produce half and quad precision tests.

davidharrishmc commented

@pawks can you comment on whether these configs can be made available?

pawks commented

All the cgfs are already present under the sample_cgfs folder in the repo(link). The cgfs are placed in the directory corresponding to the minimum ISA requirement for the instruction to be included i.e for a RV64FD configuration instructions from all 4 directories will have to be used. This follows the instruction organization in the ISA manual.

davidharrishmc commented

@pawks Thank you for the speedy reply and sorry I missed the subdirectory. A few more questions if you have time to support this.

  1. It looks like riscv_ctg has changed since you generated fadd_b1-01.S June 17, 2022. For example, now it wants a --base-isa and seems to want --flen rather than --xlen. I'm having trouble coming up with a usage string that works with RV32F/fadd.s.cgf. Can you either point me to a string that works, or recommend an older version of riscv_ctg that is compatible with the usage string recorded in fadd_b1-01.S?

  2. RV32F/fadd.s.cgf uses some abstract_comb: 'ibm_b... statements. I'm not finding where these are defined in riscv_ctg. Will riscv_ctg find these combinations somehow, or is there additional information I need outside the riscv_ctg repo?

  3. Related to (2), the abstract_comb ibm_b... takes a parameter of 32. I presume I'll need to change this to 16 and 128 to generate Zfh and Q tests. If this doesn't just work out of the box, is there any documentation about finding and modifying this abstract_comb? I'm comfortable with floating-point and willing to do work if I can find what needs adapting.

allenjbaum commented