Lay the source tree out in directories instead of one flat root

[galekseev]

Summary

The repository root held the C++ sources, the OpenCL kernels, the Dockerfile, the build artifacts and the documentation side by side. Everything now has a directory:

src/       C++ sources of the host program
kernels/   keccak.cl and profanity.cl
bin/       what make produces: the binary and a copy of the kernels
build/     object files
docker/    the shipping image and its entrypoint
bench/     revision comparison tooling, with its recorded runs in bench/logs/
tests/     correctness tests and the mp-math microbenchmark
docs/      build guides and the vast.ai guide, with the screenshot in docs/img/

Two things needed real changes rather than a move:

Finding the kernels. profanity.cpp read them from the working directory. It now looks next to the executable first, then in kernels/ beside it, and falls back to the working directory, which is where every earlier release kept them. The compiled-kernel cache moves next to the binary for the same reason. So ./bin/profanity2.x64 works from the repository root, from inside bin/, and by absolute path from an unrelated directory.

Comparing across the move. bench/ builds arbitrary revisions out of history, and has to go on comparing a revision from before this change against one from after it. bench/Dockerfile and bench/prepare-native.sh now take the build output from whichever layout a revision uses. The second commit fixes a false positive this introduced: a git diff of SpeedSample.cpp reported the move itself as a change, so every comparison spanning the reorganization warned that the two revisions time a round differently. The timer is now compared by blob rather than by path.

Test plan

• make from clean builds bin/profanity2.x64 with no warnings, kernels copied next to it
• runs from the repository root, from bin/, and by absolute path from /tmp; the cache lands in bin/ in all three cases and nothing is left in the working directory
• make -C tests builds, and test_correctness_pr49.x64 passes run both from tests/ and from the root
• bench/prepare-native.sh 90c61e3 868be80 builds both layouts and lays them out identically; timer.state is same, as it should be for a file that only moved
• no performance change: Dispatcher.o, Mode.o, precomp.o and SpeedSample.o are byte-identical between the two builds, only profanity.o differs. A speed comparison on the Mac this was written on was thermally saturated and could not tell the two apart from noise, which is why the object files are the evidence here
• CI builds the image from docker/Dockerfile (no local Docker daemon to check it with)

Made with Cursor