Modify converted_call to accept a separate argument when the callable is an attribute, representing the attribute's owner.

This gives converted_call a chance to override the attribute lookup, and in turn allows working around the limitation `super` that prevents it from resolving dynamic attributes.

PiperOrigin-RevId: 216727687

Also fix the DotLayout test, it would pass even when commenting out the dot specific logic in GpuLayoutAssignment.

PiperOrigin-RevId: 216726796

PiperOrigin-RevId: 216725062

patch command can exceed the 10s timeout limit on some slow machines.

Fixes #22761

PiperOrigin-RevId: 216721485

Use the functional operator pattern for logical expressions. The current implementation relied on the TF import, which is now gone.

PiperOrigin-RevId: 216719675

PiperOrigin-RevId: 216718180

PiperOrigin-RevId: 216714487

Also fix the initial_state and constants param in the eager context.

PiperOrigin-RevId: 216712447

PiperOrigin-RevId: 216706771

PiperOrigin-RevId: 216706627

PiperOrigin-RevId: 216705389

PiperOrigin-RevId: 216705354

PiperOrigin-RevId: 216697820

PiperOrigin-RevId: 216697425

PiperOrigin-RevId: 216691593

PiperOrigin-RevId: 216688700

PiperOrigin-RevId: 216686849

PiperOrigin-RevId: 216686136

We are working on disallowing xla::Add on PRED types because they can be
confusing. This is a prerequiste for that to not fail over in the new
shape check.

PiperOrigin-RevId: 216685015

Wonder how that could have ever let this CL to go in in the first place.

Automated rollback of commit 905deeaa

PiperOrigin-RevId: 216683052

PiperOrigin-RevId: 216663386

PiperOrigin-RevId: 216639755

PiperOrigin-RevId: 216638239

PiperOrigin-RevId: 216633097

PiperOrigin-RevId: 216632480

We have a lot of such warnings and don't make an attempt to clean them up
internally.

PiperOrigin-RevId: 216631010

No functional change.

PiperOrigin-RevId: 216629980

PiperOrigin-RevId: 216628574

PiperOrigin-RevId: 216627219

[XLA::GPU] Explicitly use {lhs,rhs}_contracting_dimension from DotDimensionNumbers as the reduction dimension for Dot.

PiperOrigin-RevId: 216627100

 - Array fields minmax and quantization_params are pointers.
      (so the current check could have false negatives as identical objects
       have different addresses)
 - also compare narrow_range.
      (so the current check could have false positives --- my bad, I added
       narrow_range later and forgot to update this code).

PiperOrigin-RevId: 216626868

We fuse everything into the scatter now, and emit two kernels. The first kernel
fills the output buffer with the computation fused into the scatter operand.
The second kernel is a regular scatter, which also contains the fused
operations from the updates and scatter_indices inputs.

PiperOrigin-RevId: 216624225

PiperOrigin-RevId: 216624182

 - Make the channel-dims-equal-to-3 heuristic more specific.  The relevant
   cudnn optimization only occurs with input-channels equal to 3 and
   output-channels equal to 32 or 64.

 - Simplify the logic that decided which shapes' dimensions to pad by using a
   clever (if I do say so myself) trick. Really all we need is a new layer of
   indirection which allows us to say "modify the input shape's feature dim"
   and have that be reflected in the lhs shape's feature dim.

 - Rename it to CudnnConvPadForTensorCores.  I initially thought that the
   3-input-channels optimization applied to Pascal, but experimentally this
   seems to be incorrect.

PiperOrigin-RevId: 216623854

PiperOrigin-RevId: 216623201

PiperOrigin-RevId: 216622520

INFINITY define requires the cmath header.

PiperOrigin-RevId: 216621867

PiperOrigin-RevId: 216619617

Make them shorter and more consistent.

- CudnnConvolutionFoo -> CudnnConvFoo
- PadInsertion -> CudnnConvPaddingLegalization
- PadForTensorCores -> CudnnConvPadForSpeed (padding channel dimensions from
  3 -> 4 is not a tensor-cores-related optimization and ideally should be run
  on P100s as well).

PiperOrigin-RevId: 216618934

PiperOrigin-RevId: 216618378