Started as a group project for Imperial College London's year 2 end of year project with Ryan Voecks, Petr Olšan, Orlan Forshaw, Zian Lin, and Gabriel Tetrault.
I was responsible for part of the GPU design in System Verilog and mainly the whole design of the Compiler. Originally, the compiler just generated GPU machine code to plot math expressions like these on the screen:
2.3sin(x/y) + 2.3cos(x/y)
Then I came up with a programming language designed for linear algebra operations and adapted the compiler to output GPU machine code to execute matrix operations heavily parallelised.
Example code for dot product:
$z = |2,1|[1., 2.]
$m = |1,2|[1., 2.] dot $z
$mc = |1,1|[5.]
$first = |3,1|[1.,2.,3.]
$second = |3,1|[1., 2., 3.]
$final = $first + $second
$finalCheck = |3,1|[1.,4.,6.]
$beforeRelu = |2,2|[-3.0, -5.5, 3., 8.]
$afterRelu = relu(-$beforeRelu)
$reluCheck = |2,2|[3.0, 5.5, 0., 0.]
.plot $m 0.0 10.
.plot $mc 0.0 10.0
.plot $final 0.0 10.0
.plot $finalCheck 0.0 10.0
.plot $beforeRelu 0.0 10.0
.plot $afterRelu 0.0 10.0
.plot $reluCheck 0.0 10.0
where .plot $someVarName min max will plot the array as a rectangular heatmap
(they look something like the logo above) with min the coldest value and max
the hottest value of the heatmap.
To store matrices multi-dimensional arrays in memory the compiler outputs code to allocate and populate the according memory. It also uses an automatic garbage collection mechanism to reuse memory of variables that are not used anymore.
The Go program testgen.go was used to generate C++ testbenches for Verilator as well as scripts to compile the RTL with Verilator along with the testbench. To do this testgen.go used template files for the script and the testbench and then used the text/template package to fill them with values from the YAML file. Moreover, the YAML file can specify assembly files to be used which testgen.go will assemble (after first recompiling the assembler), and then place correctly inside the testbench to send them to the GPU interface in different cycles among the correct control instructions. It is even possible to just specify math equations in a file in the YAML’s directory which will be automatically compiled before being assembled. In the top-level directory of the repository a Bash script automatically runs all the tests in the directory test where each test is one directory containing at least a YAML file, and, optionally, math expression files or assembly. It generates the corresponding testbenches in a different directory along with all error messages and runs them.
See example tests in the test directory.
Build the docker container: (need to be inside repo top level directory)
docker build -t epic_accelerator .
To run all tests in directory test run this in the container:
./test.sh
Or to just run the test sampletest in the test directory run:
./test.sh test/sampletes
(Can specify any number of tests as arguments.)