| 7575f334ee0879825ceed23928f5e99d0f71b5f8 |
|
14-Feb-2018 |
Justin Lebar <jlebar@google.com> |
[XLA:GPU] Don't crash when the root instruction of a computation is a multi-output fusion node, and avoid some pointer chasing with tuples.
Previously, the kernels we generated would have one argument per
*top-level* buffer of the input/output. This was fine for inputs. But
it doesn't work for outputs: Imagine you're a node that returns a tuple
-- e.g. multi-output fusion -- if all you get is a pointer to the
top-level buffer of your output (which should contain pointers to the
lower-level buffers at some point, but at the moment is just empty), how
are you supposed to figure out where to write your output?
(This usually worked because most of the time your output would live
inside of the big XLA temp buffer, and kernels always get a pointer to
that.)
Now we pass all the buffers, top-level and otherwise, to our kernel. In
addition, we're now willing to dereference statically tuples that live
entirely in XLA's temp buffer. Pointers in input tuples must still be
dereferenced dynamically, because the caller has the option of giving us
these values or not when invoking XLA.
This change makes some parts of BufferAssignment/BufferAllocations more
truthful. Previously, if you passed a tuple-shaped input to XLA, we'd
say in BufferAllocations that the pointer for some subshape of the param
was the *top-level tuple pointer*. XLA then knew that this was a lie
and would dereference it accordingly. Now we have an explicit notion of
a BufferAllocation pointing to a subshape of an input parameter.
PiperOrigin-RevId: 185614060
/external/tensorflow/tensorflow/compiler/xla/tests/tuple_test.cc
|
| 7022e6b62908f688cae446f1abbc667c5273db04 |
|
13-Jan-2018 |
A. Unique TensorFlower <gardener@tensorflow.org> |
[XLA] Add some scalar tests.
PiperOrigin-RevId: 181819491
/external/tensorflow/tensorflow/compiler/xla/tests/tuple_test.cc
|
| 0683cdbd8701e4e6a582db1e71d58fcad628e070 |
|
11-Dec-2017 |
A. Unique TensorFlower <gardener@tensorflow.org> |
Enriching some C64 test coverage.
PiperOrigin-RevId: 178624364
/external/tensorflow/tensorflow/compiler/xla/tests/tuple_test.cc
|
| a235f23d5babcffa05b6d190c3e1a8909afb5273 |
|
22-Nov-2017 |
Mark Heffernan <meheff@google.com> |
Roll forward new copy insertion pass.
Original cl: cl/174423881, rollback cl: cl/174505237.
This roll forward includes the following changes from the original to address various issues uncovered with the rollback:
(1) A fix for a problem with fusion instruction serialization was broken out and submitted separately (cl/176035108).
(2) A dataflow analysis fix was broken out and submitted separately (cl/176035108)
(3) Adding RunBenchmarks to our unit test main was broken out. Fix for uncovered segv in while_test_cpu benchmark in pending cl/176068232.
(4) Moved a cpu-specific copy-insertion pass into it's own file, and added tests.
(5) Renamed gpu/copy_insertion.* to gpu/gpu_copy_insertion.* to match cpu side.
PiperOrigin-RevId: 176658339
/external/tensorflow/tensorflow/compiler/xla/tests/tuple_test.cc
|
| 456929281592f14d50443cfbdaa2f6b36167a134 |
|
03-Nov-2017 |
Mark Heffernan <meheff@google.com> |
Rollback copy insertion change because it results in a DCHECK with an internal model.
END_PUBLIC
BEGIN_PUBLIC
Automated g4 rollback of changelist 174423881
PiperOrigin-RevId: 174505237
/external/tensorflow/tensorflow/compiler/xla/tests/tuple_test.cc
|
| 7bb2d57b0b051d1cf8dd74d3276bf5a452774172 |
|
03-Nov-2017 |
Mark Heffernan <meheff@google.com> |
Rewrite CopyInsertion to use module-scoped HloAliasAnalysis. The net effect (number of copies inserted) is roughly similar to the existing implementation, but the new implementation is much more general. The new implementation can handle entry argument buffer reuse with minimal modification, for example.
Some unnecessary copies are still added due to deficiencies in buffer assignment (b/62548313), but these can be removed when buffer assignment also uses HloAliasAnalysis.
Also address a few issues uncovered with this cl:
(1) For inplace dynamic slice in llvm backends, truncate do not wrap the slice. This matches the behavior of the non-inplace variant.
(2) Disable SelectBetweenPredTuples test on GPU. The test introduces top-level buffer ambiguity which is not tolerated by the gpu backend.
(3) When deserializing HLO form a proto, do not uniquify instruction names in fused computations.
(4) In dataflow analysis, don't deallocate deleted HloValues during propagation.
(5) In dataflow analysis, fix issue with live_out_of_computation property.
PiperOrigin-RevId: 174423881
/external/tensorflow/tensorflow/compiler/xla/tests/tuple_test.cc
|
| b29b839215fa9bf5a00ca97e19673cfa5f780314 |
|
26-Sep-2017 |
A. Unique TensorFlower <gardener@tensorflow.org> |
[XLA] Map API change to enable mapping over an arbitrary set of dimensions.
PiperOrigin-RevId: 170090055
/external/tensorflow/tensorflow/compiler/xla/tests/tuple_test.cc
|
| dc1eda8a6d06cff541be768a0c8e2b22b376651c |
|
13-Sep-2017 |
Peter Hawkins <phawkins@google.com> |
[XLA] Fix CHECK-failure crash if a non-tuple was passed to GetTupleElement.
PiperOrigin-RevId: 168550703
/external/tensorflow/tensorflow/compiler/xla/tests/tuple_test.cc
|
| ddd8e21b7c1d23bf80ddf0141b44e168c17647f3 |
|
27-Jul-2017 |
Eli Bendersky <eliben@google.com> |
[XLA] Consolidate all similar main()s in tests into a single target.
PiperOrigin-RevId: 163354724
/external/tensorflow/tensorflow/compiler/xla/tests/tuple_test.cc
|
| c0f475b03eeca19b217527f7946da837e8438e27 |
|
06-Jul-2017 |
Peter Hawkins <phawkins@google.com> |
[XLA] Add support for tuple constants via a rewrite that replaces constant tuples with tuples of non-tuple constants.
Fix crash in HloInstruction::ToString() for tuple constants.
PiperOrigin-RevId: 161085252
/external/tensorflow/tensorflow/compiler/xla/tests/tuple_test.cc
|
| a4a46983233690f17ca794c2907b2b11c8cd2060 |
|
21-Jun-2017 |
A. Unique TensorFlower <gardener@tensorflow.org> |
[XLA] Add tests for select ops and while loops that produce tuples that contain predicates.
PiperOrigin-RevId: 159645900
/external/tensorflow/tensorflow/compiler/xla/tests/tuple_test.cc
|
| 46737e4e81314f7482bfd6a710f126a27f5d7975 |
|
19-Jun-2017 |
A. Unique TensorFlower <gardener@tensorflow.org> |
Remove class xla::LiteralUtil. NFC (mind-numbingly so).
This patch removes class xla::LiteralUtil and rewrites every call to use class
xla::Literal instead.
PiperOrigin-RevId: 159446373
/external/tensorflow/tensorflow/compiler/xla/tests/tuple_test.cc
|
| c77399d44fc2ed6912e7f301839ad3e404739b80 |
|
14-Jun-2017 |
Eli Bendersky <eliben@google.com> |
[XLA] Remove remaining flags from cpu_compiler_flags
And move them to debug_options_flags; these two flags (embed_ir_in_executable,
dump_debug_json_to) are also unified with similarly named GPU compiler flags.
This lets us completely remove the cpu_compiler_flags module.
PiperOrigin-RevId: 158989621
/external/tensorflow/tensorflow/compiler/xla/tests/tuple_test.cc
|
| 8cedce4b806639c351d45a00324fcc269704f42b |
|
12-Jun-2017 |
Eli Bendersky <eliben@google.com> |
[XLA] Replace some XLA CPU compiler specific options by generic "debug options".
LLVM optimization level, extra LLVM flags and "cpu parallel" all turn into
debug options on the xla proto. "cpu parallel" is combined with "backend extra
options" as a map.
PiperOrigin-RevId: 158751784
/external/tensorflow/tensorflow/compiler/xla/tests/tuple_test.cc
|
| 1e67c90e2caceeff82d09793d1ef5fa0300d219b |
|
09-Jan-2017 |
Peter Hawkins <phawkins@google.com> |
Initial open-source release of XLA: Accelerated Linear Algebra.
XLA is a compiler-based linear algebra execution engine that targets CPUs, GPUs and custom accelerators.
XLA is still experimental; we are releasing it early to get the community involved.
Change: 143990941
/external/tensorflow/tensorflow/compiler/xla/tests/tuple_test.cc
|