fcvt.s.d gives wrong answer

[b1f6c1c4]

I tried to run fcvt.s.d to convert a small double (2.386004908190509275e-40 which is 0x37b4c8f800000000) into single, but it gives inf (0x7f800000) instead of 2.38600491e-40 (0x0002991f).

[stmach]

Hello,

Have you checked what is effectively applied to the FPU?
Setting up the operation with

op_i       = fpnew_pkg::F2F
rnd_mode_i = fpnew_pkg::RNE
src_fmt_i  = fpnew_pkg::FP64
dst_fmt_i  = fpnew_pkg::FP32

and applying your given input yields 0xffffffff0002991f as a result, which is what you'd expect.
The issue is likely in the integration of the FPU into your system (either configuration or the application of control signals) then. What system are you using the FPU with?

Cheers!

[b1f6c1c4]

I'm trying with Ariane. How do we debug into it?

[stmach]

Can you please provide a minimal working code snippet that you execute on Ariane which triggers this issue, so that we may reproduce it?

[b1f6c1c4]

I saw this error when I was trying to run riscv-tests (https://github.com/riscv/riscv-tests/blob/master/isa/rv64ud/recoding.S) with OpenPiton.

[zarubaf]

Can you check the signals of the FPU, please?
@b1f6c1c4 and @stmach can we check that you tested the stimuli on the same fpnew source?

[stmach]

I've used the latest tagged release v0.6.1 of the FPU but there haven't been changes to the conversion units in a long time, so it should work with older versions as well.

[zarubaf]

We have rv64ud-p-recoding in the CI, so I don't think it fails for the repo here anyway.

[b1f6c1c4]

Here's what I got. BTW, I'm using fpnew v0.6.1 (4a241d4).

[zarubaf]

@stmach Can this be a VCS/Modelsim difference? When compiling the FPU I am getting two warnings with the FPU in VCS so that might be related (but I don't think so):

Warning-[ENUMASSIGN] Illegal assignment to enum variable
/scratch/zarubaf/snitch/src/fpnew/src/fpnew_opgroup_block.sv, 126
fpnew_opgroup_block, "fpnew_pkg::get_first_enabled_multi(10'b1010101010, FpFmtMask)"
  Only expressions of the enum type can be assigned to an enum variable.
  The type bit [9:0] is incompatible with the enum 'unit_type_t'
  Expression: 10'b1010101010
  Use the static cast operator to convert the expression to enum type.


Warning-[ENUMASSIGN] Illegal assignment to enum variable
/scratch/zarubaf/snitch/src/fpnew/src/fpnew_opgroup_block.sv, 126
fpnew_opgroup_block, "fpnew_pkg::get_first_enabled_multi(10'b1010101010, FpFmtMask)"
  Only expressions of the enum type can be assigned to an enum variable.
  The type bit [9:0] is incompatible with the enum 'unit_type_t'
  Expression: 10'b1010101010
  Use the static cast operator to convert the expression to enum type.

Bender supports generating VCS scripts now and its relatively easy to bring up the simulation in VCS as well.

[zarubaf]

@b1f6c1c4 What's your VCS version? Just to be sure we are not hunting bugs in the FPU although they are in the CPU (wrapper for example). Can you provide us with the same screenshot but signals from the fpnew_top instance (https://github.com/pulp-platform/ariane/blob/9cebe8a238aa8e1450c085f051fefc1e6d55cc41/src/fpu_wrap.sv#L513-L538)?

[stmach]

I tried to run the operation in VCS and indeed i receive Inf (0x7f800000).
@zarubaf I added a static cast to int to solve your warning so it shouldn't be the reason for this mismatch. Feel free to move this issue back to fpnew.
I'll try to find out what triggers VCS to produce the wrong result here... It might take a while though as VCS is not our main simulation target.

[b1f6c1c4]

FYI

vcs script version : O-2018.09
machine name = della2.princeton.edu
machine type = linux64
machine os = Linux 3.10.0-1062.12.1.el7.x86_64
The FLEXlm host ID of this machine is "80000208fe80 00219ba4ae77 00219ba4ae79 00219ba4ae7b 00219ba4ae7d"
Compiler version = VCS-MX O-2018.09-SP2_Full64
VCS Build Date = Feb 28 2019 22:34:30

[stmach]

A post-synthesis netlist obtained from Synopsys DC yields the correct result (simulated with Modelsim since I don't have VCS set up), so the issue seems contained to VCS for now.

[b1f6c1c4]

Any update on this?

[b1f6c1c4]

@stmach @zarubaf Well I'm wondering how can I help for this issue. Will it be helpful if we could agree on a certain version (commit SHA-1) of Ariane and we run VCS and ModelSim separately, and then compare their results? Due to some limitation (broken ModelSim license) I cannot do this on my end. So if you think it will help, you can give me a SHA-1 and I'll do VCS on my end. Otherwise you can debug yourself.