experiments/STT-tensorflow
1
0
Fork 0
Code Issues Pull Requests Projects Releases Wiki Activity
STT-tensorflow/tensorflow/compiler/xla/service/llvm_ir
History
Adrian Kuegel 87c1cf490f Use a pair of (building block, index) as key for the cache. 
Previously, we only used index as key, and then checked whether the building
block matches. If it didn't match, we overwrote the cache entry. In
pathological cases, this means we *never* used a cached entry, because even in
cases we had already computed a value that could be used, it was already
replaced by a value which could not be used. This lead to exponential compile
time and *huge* generated LLVM IR.

PiperOrigin-RevId: 338237297
Change-Id: I258202c6238056bfb4466a95c30b6cd48700fe21
2020-10-21 04:20:37 -07:00
..
alias_analysis_test.cc
Change XLA's default to disable cpu_fast_math options with the exception of min_max behavior.
2020-05-15 13:11:47 -07:00
alias_analysis.cc
…
alias_analysis.h
…
buffer_assignment_util.cc
[XLA/GPU] Remove uses of BufferAssignment in nested computation lowering. NFC
2020-07-28 14:31:02 -07:00
buffer_assignment_util.h
[XLA/GPU] Remove uses of BufferAssignment in nested computation lowering. NFC
2020-07-28 14:31:02 -07:00
BUILD
Refactor the code to avoid duplication (NFC).
2020-10-01 05:16:36 -07:00
dynamic_update_slice_util.cc
Some refactoring groundwork towards an LHLO backend for XLA GPU; NFC
2020-06-25 20:22:18 -07:00
dynamic_update_slice_util.h
Some refactoring groundwork towards an LHLO backend for XLA GPU; NFC
2020-06-25 20:22:18 -07:00
fused_ir_emitter.cc
Use a pair of (building block, index) as key for the cache.
2020-10-21 04:20:37 -07:00
fused_ir_emitter.h
Use a pair of (building block, index) as key for the cache.
2020-10-21 04:20:37 -07:00
ir_array.cc
Implement the llvm lowering for the customcall sliceToDynamic and padToStatic for on XLA:GPU.
2020-06-10 15:56:58 -07:00
ir_array.h
Implement the llvm lowering for the customcall sliceToDynamic and padToStatic for on XLA:GPU.
2020-06-10 15:56:58 -07:00
ir_builder_mixin.h
[XLA] Remove deprecated function. It's unused.
2020-04-14 04:15:47 -07:00
kernel_support_library.cc
…
kernel_support_library.h
…
llvm_loop.cc
…
llvm_loop.h
…
llvm_util.cc
Roll forward XLA GPU LHLO sort op migration
2020-08-13 12:00:17 -07:00
llvm_util.h
Roll forward XLA GPU LHLO sort op migration
2020-08-13 12:00:17 -07:00
loop_emitter.cc
Merge pull request #42683 from nouiz:upstream_master_grid_size
2020-10-06 04:04:24 -07:00
loop_emitter.h
Merge pull request #42683 from nouiz:upstream_master_grid_size
2020-10-06 04:04:24 -07:00
math_ops.cc
Enable XLA:CPU fast math for min/max by default to be similar to TF's behavior.
2020-06-17 01:27:08 -07:00
math_ops.h
…
README.md
…
sort_util.cc
Some refactoring groundwork towards an LHLO backend for XLA GPU; NFC
2020-06-25 20:22:18 -07:00
sort_util.h
Some refactoring groundwork towards an LHLO backend for XLA GPU; NFC
2020-06-25 20:22:18 -07:00
tuple_ops.cc
[XLA/GPU] Add a gold CHECK file for sorting. NFC.
2020-08-05 15:39:42 -07:00
tuple_ops.h
…

README.md

Common utilities and abstractions for handling and emitting LLVM IR for XLA backends.