* test
* rm in llvmir
* rm in ptx and nir
* hmmmm
* rm in decompositions
* skip tests
* add test
* just this
* rm comment
---------
Co-authored-by: chenyu <[email protected]>
* simple failing jit test case with Tensor.empty
* this used to exist in ops.py...
* Revert "removed if self.buffer.is_allocated() in realized (#14836)"
This reverts commit 72cf603805.
* ptx: use setp.neu for float CMPNE
* test ptx float CMPNE renders setp.neu
* check NaN behavior, not grep ptx strings...
* skip WEBGPU for test_cmpne_nan (Vulkan NaN behavior)
---------
Co-authored-by: Nicolas Pinto <[email protected]>
Co-authored-by: chenyu <[email protected]>
* preallocate all realized buffers
* contiguous
* work
* comment that out
* move to schedule
* better
* correct fix
* just buffer
* disk bufs
* fixes disk tensor stuff
* fix symbolic stuff
* fix multi
* 162 failures
* bugfixes
* don't check that anymore
* fix schedule tests
* mnist should be contiguious
* type and buffer
* fix tests
* shrink axis correction
* mypy fixes
* tests skips
* same 37 failures
* dedup
* no shrink in the graph
* 29 failures
* skips
* fix custom kernel
* fix training
* those optimizations aren't supported currently
* simpler
* more correct
* tests
* 14 failures
* works
* fix that test
* broken
* 11 failures
* only kernel counts left
* fixes
* all tests pass
* remove tensor_map
* op test
* 200 -> 230
* test fixes
* fixes
* revert test_tiny thing
* guard
* revert that
* test tiny passes
* no contigs there
* base realize back
* Revert "no contigs there"
This reverts commit c45bb9fcfd.
* revert that
* chop many assigns
* 12 failures
* fix tests
* tests
* apply after
* pre-commit
* remove old code
* delete that
* fix types
* remove extra contig
* fix dataloader
* torch fix
* disk fix
* update kernel fusion numbres
* runs on amd
* restore kernel count
* add that rule back
* that
* disable that
* wrong
* add the correct rule for that folding
* more tests
* guard c1.arg
* no newlines
* realize those
* split into a different file
* remove detach/contig back
* skip 2
* update that
* fixes and test relaxations from prealloc_bufs
* fix error type and guard _mop
* revert that
* contiguous makes extra/torch_backend/test_kernel_fusion.py fail
* decompositions: fuse (x << n) + c to MULACC
MUL→SHL converts x*(2^n) to x<<n before MULACC can fuse (x*c)+y.
Add pattern to also fuse (x<<n)+c → MULACC(x, 2^n, c) for backends
that support both MULACC and SHL.
* test: add test_mulacc_shl for SHL->MULACC fusion
* test: relax test_mulacc_unrolled to >= 4
SHL->MULACC fusion now also catches power-of-2 address calculations,
increasing MULACC count from 4 to 6 on PTX. the test's intent is that
each unrolled multiply is individually fused (not grouped), so >= 4
is the correct assertion.
---------
Co-authored-by: Prithvish <[email protected]>
Co-authored-by: Nicolas Pinto <[email protected]>
Co-authored-by: Nicolas Pinto <[email protected]>
undo the view and make it a mask, this fuses the setitem with any pending compute too.
one behavior change is that for target not backed by a buffer (const and arange), rangeify makes output contiguous under the hood.
this is stricter better than raise and ask user to call contiguous, as that would no longer be fuse-able.
sirhcmandGitHub