Github_Repos/cvw
1
0
Fork 1
mirror of https://github.com/openhwgroup/cvw synced 2025-02-02 17:55:19 +00:00
Code Issues Packages Projects Releases Wiki Activity

Combine floating point test documentation

Browse Source
This commit is contained in:
Jordan Carlin 2024-10-01 12:44:47 -07:00
parent 6d21e272d0
commit 9363379b55
No known key found for this signature in database
2 changed files with 48 additions and 73 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

56
sim/FPbuild.txt
View File

@ -1,56 +0,0 @@
Procedure for Runnning SoftFloat/TestFloat with Wally
1.) First, compile SoftFloat and TestFloat by going to the addins
directory and finding the specific build directory (e.g.,
Linux_x86_64-GCC. Currently, we are using v3e of
SoftFloat/TestFloat. I am not sure of the order, but I always compile
SoftFloat first as I believe TestFloat uses the static library
SoftFloat creates.
2.) Once compiled both, go to the tests/fp directory and run the
create_vectors.sh Linux script. In the past, we have automated this,
but I believe this has fallen into more of a manual state lately.
3.) Then, run remove_spaces.sh which will remove spaces from the
output and put underscores between vectors (this helps differentiate
the vectors that are generated). Again, this can be combined with
Step 2.
4.) TestFloat is run from wally/cvw/sim and sim-testfloat-batch with
its respective test. The format is ./sim-testfloat-add add. All of
the tests are listed below. This can be augmented or added to for
other FP tests given by the great SoftFloat/TestFloat output.
cvtint - test integer conversion unit (fcvtint)
cvtfp - test floating-point conversion unit (fcvtfp)
cmp - test comparison unit's LT, LE, EQ opperations (fcmp)
add - test addition
fma - test fma
mul - test mult with fma
sub - test subtraction
div - test division
sqrt - test square root
all - test everything
4a.) Each test will test all its vectors - if you want to test a
subset of the vectors (e.g., only binary16), you should modify the
testfloat.do in the sim directory. Change the TEST_SIZE="all" to the
specific test you want to run. For example, if you want to run only
binary16, you should set this variable to TEST_SIZE="HP".
4b.) If you want to turn off the generation of wlf files while running
sim-testfloat-batch, you can modify testfloat.do in the sim
directory. Inside this DO file, modify the WAV file to 0 --> i.e.,
set "quietly set WAV 0;"

65
tests/fp/README.md
View File

@ -1,27 +1,58 @@
james.stine@okstate.edu 14 Jan 2022
jcarlin@hmc.edu Sept 2024
These are the testvectors (TV) to test the floating-point unit using
Berkeley TestFloat written originally by John Hauser. TestFloat
requires both TestFloat and SoftFloat.
## TestFloat for CVW
The locations of these tools at time of this README is found here:
TestFloat-3e: http://www.jhauser.us/arithmetic/TestFloat.html
SoftFloat-3e: http://www.jhauser.us/arithmetic/SoftFloat.html
The CVW floating point unit is tested using testvectors from the Berkeley TestFloat suite, written originally by John Hauser.
These tools have been compiled on a x86_64 environment by going into
their respective build/Linux-x86_64-GCC directories and running make.
TestFloat and SoftFloat can be found as submodules in the addins directory, and are linked here:
- TestFloat: https://github.com/ucb-bar/berkeley-testfloat-3
- SoftFloat: https://github.com/ucb-bar/berkeley-softfloat-3
The makefile in the vectors subdirectory of this directory will generate TV
for each rounding mode and operation. It also puts an underscore between each
vector instead of a space to allow SystemVerilog readmemh to read correctly.
### Compiling SoftFloat/TestFloat and Generating Testvectors
The makefile at the top level of this directory will compile SoftFloat and
TestFloat and then generate all of the TVs. It also generates TVs for the
combined integer floating-point divider.
The entire testvector generation process can be performed by running make in this directory.
Although not needed, a case.sh script is included to change the case
of the hex output. This is for those that do not like to see
hexadecimal capitalized :P.
```bash
make --jobs
```
This compiles SoftFloat for an x86_64 environment in its build/Linux-x86_64-GCC directory using the `SPECIALIZE_TYPE=RISCV` flag to get RISC-V behavior. TestFloat is then compiled in its build/Linux-x86_64-GCC directory using this SoftFloat library.
The Makefile in the vectors subdirectory of this directory is then called to generate testvectors for each rounding mode and operation. It also puts an underscore between each vector instead of a space to allow SystemVerilog `$readmemh` to read correctly.
Testvectors for the combined integer floating-point divider are also generated.
Although not needed, a `case.sh` script is included to change the case of the hex output. This is for those that do not like to see hexadecimal capitalized :P.
### Running TestFloat Vectors on Wally
TestFloat is run using the standard Wally simulation commands.
To run all TestFloat tests on many different derived configurations of Wally, use
```bash
regression-wally --testfloat
```
To run a single test, use
```bash
wsim <config> <test> --tb testbench_fp
```
The choices for `<test>` are as follows:
>cvtint - test integer conversion unit (fcvtint)
cvtfp - test floating-point conversion unit (fcvtfp)
cmp - test comparison unit's LT, LE, EQ opperations (fcmp)
add - test addition
fma - test fma
mul - test mult with fma
sub - test subtraction
div - test division
sqrt - test square root
Any config that includes floating point support can be used. Each test will test all its vectors for all precisions supported by the given config.
### Testvector Count
46464 185856 836352 f16_add_rd.tv
46464 185856 836352 f16_add_rne.tv