From d2a5f19435b00a1741e82a3d84e785b040a1a7d6 Mon Sep 17 00:00:00 2001 From: Emilien Bauer Date: Sat, 21 Oct 2023 16:12:54 +0100 Subject: [PATCH] Update README.md. --- fast/README.md | 45 ++++++++++++++++++++------------------------- 1 file changed, 20 insertions(+), 25 deletions(-) diff --git a/fast/README.md b/fast/README.md index 7f52741d302..99037ebee96 100644 --- a/fast/README.md +++ b/fast/README.md @@ -4,40 +4,36 @@ This folder is where benchmarking and correctness checks will happen in. ## Structure -Each benchmark has a `name`, and it's devito definition is stored in `name.py`. +Each benchmark has a `name`, and it's devito definition is stored in a python script `.py`. +Those are: -All intermediate files discussed in this section are valid `make` targets, to enable partial compilation and result inspection. +- wave2d_b +- wave3d_b +- diffusion_2D +- diffusion_3D -Running this file with `-xdsl` will generate all the stencil source files and input data. -The initial data will be stored in `.input.data`, the stencil source in `.main.mlir` and the "driver" file containing the main and timers in `.mlir`. +Slurm batch files are provided to reproduce the experiments from the paper on ARCHER2 in slurm-jobs. -The kernel file is then compiled to either `.cpu.o` or `.gpu.o`. This is controllable using the `MODE` variabe: `make 2d5pt.out MODE=gpu` - -The main file is compiled to `main.o`, and the `interop.c` file is compiled to `.interop.o`, to read the `.input.data` file and write to the correct output. - -The `.out` file, the actual executable, is then compiled from these three object files. - -Running devito and saving the output in `.devito.data` is also a valid makefile target. - -Finally, the `.bench` will run both devito and the stencil version and compare the results. +- `diffusion-X.slurm` / `wave-X.slurm` run the corresponding script distributed among X nodes. ## Usage -Passing options to the benchmark files is done with the `BENCH_OPTS="..."` variable for make. - -`make -B 2d5pt.bench BENCH_OPTS="-d 100 100 -nt 100" MODE=CPU` -`make -B 3d_diff.bench BENCH_OPTS="-d 100 100 100 -nt 100" MODE=CPU` +Each python script has: +- a `-d` flag, expecting the size of the grid on which to execute the stencil operators. +- a `-nt` flag, expecting the number of iterations to execute. +- a `-so` flag, expecting a space discretization order. we used 2, 4 and 8 in our experiments. +To run the Devito implementation, run the script with `--devito 1`. -To conclude, running the `2d5pt` example on `gpu` and compare the results, use: +To run the xDSL implementation, run the script with `--xdsl 1`. -`make 2d5pt.bench BENCH_OPTS="-d 1000 1000 -nt 1000" MODE=gpu` +⚠️ Because of runtime environment incompabilities, please always run each implementation seprately to measure performance. -Current modes are: `cpu`(default), `openmp`, `gpu` and `mpi` +`setup_wave2d.py` and `setup_wave3d.py` are provided to set up the necessary data to run the corresponding scripts. Run them first with the same `-d -so ` to generate the expected files. -## ToDos: +TODO: put data comparison in-place when using devito and xdsl? -- Controlling devito omp flags / gpu usage is currently not done in the `Makefile` +To plot the final values, run the script with `--plot 1`. ## Passing environment variables to devito/omp @@ -46,7 +42,6 @@ Prefixing the `make` command with `NAME=val` will make the variable `NAME` avail Example: ```bash -DEVITO_ARCH=gcc DEVITO_LANGUAGE=openmp DEVITO_LOGGING=DEBUG python 3d_diff.py -d 300 300 300 -nt 300 -xdsl -DEVITO_ARCH=gcc DEVITO_LANGUAGE=openmp DEVITO_LOGGING=DEBUG python 3d_diff.py -d 300 300 300 -nt 300 -make 3d_diff.bench BENCH_OPTS="-d 300 300 300 -nt 300" MODE=cpu +DEVITO_ARCH=gcc DEVITO_LANGUAGE=openmp DEVITO_LOGGING=DEBUG python diffusion_3D.py -d 300 300 300 -nt 300 --xdsl 1 +DEVITO_ARCH=gcc DEVITO_LANGUAGE=openmp DEVITO_LOGGING=DEBUG python diffusion_3D.py -d 300 300 300 -nt 300 --devito 1 ```