Apple GPU
This is a project working on reverse engineering the Apple G13 GPU architecture (as used by the M1), and hacking together documentation, a disassembler, an emulator and an assembler along the way. This is a somewhat messy work in progress, but it reflects the best of my understanding.
Documentation can be viewed at https://dougallj.github.io/applegpu/docs.html
Disassembler
This is probably the most useful bit, still an early work in progress:
$ python3 disassemble.py code.bin
0: 0501040d00c43200 device_load 2, 0, i32, pair, 4, r0_r1, u2_u3, 0, lsl 1
8: 3800 wait 0
a: be890a042c00 convert u32_to_f, $r2, r0.discard, 1
10: be810a242c00 convert u32_to_f, $r0, r1.discard, 1
16: 9a85c4020200 fmul $r1, r2.discard, 0.5
1c: 0a05c282 rcp r1, r1.discard
20: 9a81c0020200 fmul $r0, r0.discard, 0.5
26: 0a01c082 rcp r0, r0.discard
2a: c508803d00803000 uniform_store 2, i16, pair, 0, r1l_r1h, 8
32: c500a03d00803000 uniform_store 2, i16, pair, 0, r0l_r0h, 10
Assembler
Useful for generating tests. Very hacky, and missing a lot of error checking, so check the disassembly output to see if it gave you what you asked for.
$ python3 assemble.py 'fmul $r1, r2.discard, 0.5 ; rcp r1, r1.discard'
9a85c4020200 fmul $r1, r2.discard, 0.5
0a05c282 rcp r1, r1.discard
Emulator
Emulator is not a standalone tool, nor a final API, and so far has only been used for testing the logic of instructions against hardware. There's no flow control yet, but if you want to use it, it looks something like:
cs = applegpu.CoreState()
# use cs.set_reg32/cs.set_reg16 to initialise core state
remaining = instructions
while remaining:
n = applegpu.opcode_to_number(remaining)
desc = applegpu.get_instruction_descriptor(n)
desc.exec(n, cs)
size = desc.decode_size(n)
remaining = remaining[size:]
# use cs.get_reg32/cs.get_reg16 to dump core state
However, it is used in the hardware tests.
Hardware tests
Hardware tests run instructions on the actual GPU (so can only run on Apple Silicon), and emulate them, and compare state afterwards. This is achieved using a modified version of Alyssa Rosenzweig's wrap.dylib from https://github.com/AsahiLinux/gpu
To build the testbed: cd hwtestbed && make
To run the tests: python3 hwtest.py
This will take ages, but builds a cache in hwtestbed/cache allowing future runs to be a bit faster. I generally comment out what I'm not working on. The disk is used extensively to communicate between python and the injected dylib - this could be fixed. It's probably not the worst culprit, but I wouldn't run the tests if you're trying to avoid hammering your disk (https://twitter.com/marcan42/status/1361160838854316032)
With a bit of work it should be possible to test much better and much more quickly (i.e. explicitly test cases that actually matter, rather than iterating through every combination we think might do something that might matter).
Hardware testing is great because it's hard to mess up, quick to add new tests, and we have tests. Most other things about it are currently not great. The tests are a bit scattershot - some tests are thorough, but others only test what had to be work to implement other tests. It can be hard to see what's covered, so I've found it useful to experimentally break things to see if they're being covered by the tests or not.
Also, there's also no cache invalidation, there are hardcode offsets, there's tons of dead code, and I'm generally afraid to modify the code in hwtestbed because any changes to how it configures the GPU might break all the tests, and it'd be painful to rebuild my cache to make sure it's all still working. main.mm and compute.metal were written with this in mind, and should be flexible enough to allow uniform register tests and device memory tests to be added.
Documentation
Documentation is generated, partially based on the instruction descriptions in applegpu.py using:
$ python3 genhtml.py > docs.html
Most of the text is in genhtml.py.