Develop upstream sync 241210 #2783
Merged
Conversation
This file contains bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
Adds a new boolean `xla_dump_hlo_unoptimized_snapshots` to the `DebugOptions` protobuf. When enabled, we'll dump an `HloUnoptimizedSnapshot` for each execution of an HLO module. This option only affects GPU targets for now. PiperOrigin-RevId: 703009410
PiperOrigin-RevId: 703011765
PiperOrigin-RevId: 703016890
PiperOrigin-RevId: 703016903
PiperOrigin-RevId: 703019838
PiperOrigin-RevId: 703025710
PiperOrigin-RevId: 703026703
PiperOrigin-RevId: 703028747
…en HostOffloadLegalize moves copies out of host-memory-only offloading. PiperOrigin-RevId: 703033475
PiperOrigin-RevId: 703033719
PiperOrigin-RevId: 703035110
… log Imported from GitHub PR openxla/xla#19913 Error started occurring from this commit for exp openxla/xla@6e9eefe (originally introduced here openxla/xla@9b19353#diff-61ab646c9c3b8b0fc5ed1e9a62f535e9df5843adddd071250343f3bec48eacb6) and from this one openxla/xla@53d5338 for log. Trying to compile following MLIR code: ``` HloModule module ENTRY main { p0 = bf16[4] parameter(0) ROOT exp = bf16[4] exp(p0) } ``` would result in: ``` UNKNOWN: <unknown>:0: error: loc(callsite("wrapped_exponential" at "wrapped_exponential")): failed to legalize operation 'math.exp' <unknown>:0: note: loc("wrapped_exponential"): called from <unknown>:0: note: loc(callsite("wrapped_exponential" at "wrapped_exponential")): see current operation: %7 = "math.exp"(%6) <{fastmath = #arith.fastmath<afn>}> : (bf16) -> bf16 ``` Copybara import of the project: -- 616c10b5308cb827c593a89455fea4b772d6e870 by Milica Makevic <[email protected]>: Do not use fast approximation for exp and log for ROCm -- 3fa4914f90458a0285deb8801c5689421f945fe4 by Milica Makevic <[email protected]>: Add unit test for log and exp lowering on ROCm Merging this change closes tensorflow#19913 PiperOrigin-RevId: 703035402
…utions. This extends the custom algorithm to cover 2D cases. Benchmarks show about 50 times better performance than the generic algorithm, detailed results: name old cpu/op new cpu/op delta BM_Conv2DStrided/process_time 35.2ms ± 9% 34.3ms ± 6% ~ (p=0.690 n=5+5) BM_Conv2DTransposedStrided/process_time 8.25s ± 8% 0.03s ± 3% -99.62% (p=0.008 n=5+5) name old time/op new time/op delta BM_Conv2DStrided/process_time 3.06ms ±19% 2.88ms ± 6% ~ (p=0.421 n=5+5) BM_Conv2DTransposedStrided/process_time 415ms ±12% 9ms ± 4% -97.93% (p=0.008 n=5+5) Planned improvements of this algorithm: - support feature_group_size > 1 (grouped convolution), - parallel packing of the patches (second algorithm step), - support the case with multiple input channels and output channels at the same time, - explore input kernel rotation possibilities & perf impact, PiperOrigin-RevId: 703036601
Updates LLVM usage to match [71ac1eb50955](llvm/llvm-project@71ac1eb50955) PiperOrigin-RevId: 703048823
PiperOrigin-RevId: 703050199
…n in GEMM Rewriter Imported from GitHub PR openxla/xla#20153 Removes collectives from the set of ops that can be exchanged with dequantization in the GEMM rewriter. Copybara import of the project: -- e2efa84143fe30c5c6b25132831a62707c2a8f75 by Philipp Hack <[email protected]>: Removes collectives from the set of ops exchanged with dequantization in the GEMM rewriter. Merging this change closes tensorflow#20153 PiperOrigin-RevId: 703051850
PiperOrigin-RevId: 703052110
…peration with oneDNN primitives Imported from GitHub PR openxla/xla#18616 This PR refactors the code that fuses add operation to matmul / convolution primitives. It removes usage of macros and separate templatized handlers for matmul and convolution cases. Copybara import of the project: -- 68bcdf81a47fb0f753d837c034931094c5cd8017 by Akhil Goel <[email protected]>: Refactor Add Handler -- 462890bb75f2fcea3fdc5966bfa7a2b8f94b255a by Akhil Goel <[email protected]>: Address review comments Merging this change closes tensorflow#18616 PiperOrigin-RevId: 703054496
PiperOrigin-RevId: 703063087
Add the dtypes for which a CUB kernel is unavailable to the log output. PiperOrigin-RevId: 703067645
PiperOrigin-RevId: 703074301
PiperOrigin-RevId: 703076882
…dent ops. PiperOrigin-RevId: 703082625
`xla::Compiler` manages instances of `Compiler` that are registered statically. `StreamExecutorGpuClient` used to get its `Compiler` instance during static initialization which might fail if the `Compiler` instance gets registered later. As a fix we will get the needed `Compiler` instance during every compilation call. PiperOrigin-RevId: 703087406
PiperOrigin-RevId: 703091758
PiperOrigin-RevId: 703092539
Updates LLVM usage to match [dd7a3d4d798e](llvm/llvm-project@dd7a3d4d798e) PiperOrigin-RevId: 703100529
|
is this weekly-sync no problem now? |
i-chaochen
requested review from
mmakevic-amd,
draganmladjenovic,
i-chaochen and
hsharsha
i-chaochen
reviewed
Dec 16, 2024
There was a problem hiding this comment.
have you followed the insturctions https://github.com/ROCmSoftwarePlatform/tensorflow-upstream/blob/develop-upstream/rocm_docs/SYNC_UPSTREAM.md
to do this check?
○ When all merge conflict resolved, do a grep -rn "<<<<<<" to make sure no diff symbols exist in the source
|
it seems there are still 9 failed tests on gpu-pycpp, could you able to reproduce it? |
|
i-chaochen
reviewed
Dec 19, 2024
|
Retest cpu-pycpp please |
|
Rerun cpu-pycpp please |
|
Rebuild cpu-pycpp please |
mmakevic-amd
force-pushed
the
develop-upstream-sync-241210
branch
from
2277278 to
b1e817d
Compare
…flow-upstream into develop-upstream-sync-241210
|
//tensorflow/compiler/mlir/quantization/tensorflow/python:quantize_model_test FAILED in 49 out of 50 in 302.4s locally: |
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
Add this suggestion to a batch that can be applied as a single commit.
This suggestion is invalid because no changes were made to the code.
Suggestions cannot be applied while the pull request is closed.
Suggestions cannot be applied while viewing a subset of changes.
Only one suggestion per line can be applied in a batch.
Add this suggestion to a batch that can be applied as a single commit.
Applying suggestions on deleted lines is not supported.
You must change the existing code in this line in order to create a valid suggestion.
Outdated suggestions cannot be applied.
This suggestion has been applied or marked resolved.
Suggestions cannot be applied from pending reviews.
Suggestions cannot be applied on multi-line comments.
Suggestions cannot be applied while the pull request is queued to merge.
Suggestion cannot be applied right now. Please check back later.
Weekly sync from 12.10 with the upstream.