`AMDComputeQueue.__del__` frees `hw_page` which is safe because
`AMDAllocator._free` does `self.dev.synchronize()` which is supposed
to wait for execution of IB to finish, however that doesn't happen if
AMDComputeQueue is dropped right after submit before timeline signal is
incremented, which it is in most places leading to a race if .bind() is
also used (required for multi-xcc because bug in mec fw treats all
PACKET3_PRED_EXECs outside IBs as if they had EXEC_COUNT of zero).
* why does max_unpool2d feel slower than out.gradient ...
* slightly cleaner
* what happened to ruff
* need to think about this some more
* slightly faster now?
* clean up, 1 more failing edge case
* ok good
* working TINY_BACKEND
* nit doc wording
* retry CI
* wrong but uses less shared
* for size 8 tc1 with devectorize in 0 loads into local before wmma and works
* improvements over tc1 devectorize
* fix tc=3
* works for handcoded tc opts
* clean bugfix tc=3
* fix
* revert changes
* add validation of the bounds of Ops.INDEX
* do mask properly
* more validation
* correct
* fix gated
* add CAST support to vmin/vmax
* fix ptx and image
* ptx no diff
* upat.index also stays
---------
Co-authored-by: qazal <qazal.software@gmail.com>
* wow argmax is so good
* 1 less line
* clean up and better variable names
* is this torch thing right...?
* add more tests
* slap a TODO on it
* clean ups
* prettier looking code and fix ceil mode test
* add return types and some docs
* ok that was a bad example since indices == value, just no example
* extra/gemm/max_matmul: start of custom kernels for GEMM
* add an unoptimized FP16/FP16 MMA example
* add slow 3-stage fp16 acc example
* add correct 3-stage pipeline with unswizzled/flat smem input (slow)
* add acc fp16 example with 3 stages and swizzle (no bank conflicts)
* add max version of NV fp16_fp16_fp16
* fix up comments and removed unused code in max variations
* add start of no_xor example
* fix to account for UOps to Ops
* train_shakespeare_char.py works
* move aten.where.self_out to tiny_backend_out
* fix memory leak
* corealize in the backward_hook
* Update backend.py
---------
Co-authored-by: George Hotz <72895+geohot@users.noreply.github.com>
* jit the forward
* might timeout, idk just send it
* this is dumb
* naive bitonic lol
* idk if this is correct, but that squeeze before is definitly not
* vectorized bitonic sort, but still slow
* yay 1 layer is correct
* alright its pretty good
* good enough
* rerun CI
* nit improve comment
* add f16/f32 mfma support for MI300
- add 16x16 mfma shape support for f16 with f32 acc
- add ops_python mfma emulation
- add arch to AMDRenderer
* minor cleanup
* minor cleanup
* add mfma emulation task to ci
* add back todo
* hotfix: comment
* add tc=3 job to ci
