George Hotz
George Hotz
The code in llvm and OpenCL isn't optimized and isn't well tested (probably shouldn't be touched), but I will merge the changes that remove the extra array construction from tuple,...
Yea, only in the core tinygrad repo, accel is all WIP.
You haven't, there's still changes in the accel/ folder.
Also in extra/, the issue is this stuff isn't well tested. The 1000 lines is only for the files under tinygrad/
Hmm perhaps these other changes are harmless enough. lmk when you are ready to merge.
Err, I wouldn't do examples either. If you want to change code, you need a test for it.
I don't think we're communicating well. If you want to refactor code, you need coverage in a test. The code in tinygrad/ is (mostly) covered, including tinygrad/llops, but not the...
Doesn't work for me. Run the test like this, max should be similar speed to sum when vectorized. ``` kafka@tubby:~/fun/tinygrad$ OPT=2 LLVM=1 python3 test/test_speed_v_torch.py TestSpeed.test_sum TestSpeed.test_max sum 4096x4096 1.26 ms...
The main issue with using a BLAS library is operations after the SUM, say like a RELU. The BLAS library would have to write the matrix out to RAM, the...
I don't understand this, and this fix is pretty ugly. Anything cleaner?