Integrate llvm-project at 063c42e919c0

[hanhanW]

• Reset third_party/llvm-project: 063c42e919c01d7e64c1af5a10898fc84b06dfe8 (2023-08-16 10:45:54 -0700): [clang-format] Handle NamespaceMacro string arg for FixNamespaceComments

Cherry-pick commits:

• llvm/llvm-project@73ddc44
• llvm/llvm-project@2cc5f5d
• llvm/llvm-project@0f8bab8
• llvm/llvm-project@199442ea
• llvm/llvm-project@22a28d8
• llvm/llvm-project@4ffc63a

Revert commits:

• llvm/llvm-project@dad9de0

[stellaraccident]

We've got a lot of commits landing today. Let's see how things look but probably better to leave this tomorrow.

[hanhanW]

We've got a lot of commits landing today. Let's see how things look but probably better to leave this tomorrow.

okay, SG. I will rebase it tomorrow.

[hanhanW]

There are integration tests failure about tosa::ReduceMax should take axis as i32 type. I've fixed some C++ file to use i32 type attribute, but it does not help. Is it because we use pre-built binary? How do I fix it to unblock the integrate?

@pzread @rsuderman could you help on it if you have contexts? Thanks

[hanhanW]

There are some failures about TOSA attribute type should be i32 (not i64). @jpienaar do you know who can help on this?

Failed Tests (11):
  TENSORFLOW_TESTS :: iree_tfl_tests/east_text_detector.run
  TENSORFLOW_TESTS :: iree_tfl_tests/gpt2.run
  TENSORFLOW_TESTS :: iree_tfl_tests/llvmcpu_mobilenet_v1.run
  TENSORFLOW_TESTS :: iree_tfl_tests/llvmcpu_mobilenet_v3-large_uint8.run
  TENSORFLOW_TESTS :: iree_tfl_tests/llvmcpu_resnet_50_int8.run
  TENSORFLOW_TESTS :: iree_tfl_tests/mnasnet.run
  TENSORFLOW_TESTS :: iree_tfl_tests/mobilenet_v3.run
  TENSORFLOW_TESTS :: iree_tfl_tests/person_detect.run
  TENSORFLOW_TESTS :: iree_tfl_tests/vmvx_mobilebert_tf2_quant.run
  TENSORFLOW_TESTS :: iree_tfl_tests/vmvx_person_detect.run
  TENSORFLOW_TESTS :: iree_tfl_tests/vulkan_mobilenet_v1.run

[jpienaar]

There are some failures about TOSA attribute type should be i32 (not i64). @jpienaar do you know who can help on this?

Failed Tests (11):
  TENSORFLOW_TESTS :: iree_tfl_tests/east_text_detector.run
  TENSORFLOW_TESTS :: iree_tfl_tests/gpt2.run
  TENSORFLOW_TESTS :: iree_tfl_tests/llvmcpu_mobilenet_v1.run
  TENSORFLOW_TESTS :: iree_tfl_tests/llvmcpu_mobilenet_v3-large_uint8.run
  TENSORFLOW_TESTS :: iree_tfl_tests/llvmcpu_resnet_50_int8.run
  TENSORFLOW_TESTS :: iree_tfl_tests/mnasnet.run
  TENSORFLOW_TESTS :: iree_tfl_tests/mobilenet_v3.run
  TENSORFLOW_TESTS :: iree_tfl_tests/person_detect.run
  TENSORFLOW_TESTS :: iree_tfl_tests/vmvx_mobilebert_tf2_quant.run
  TENSORFLOW_TESTS :: iree_tfl_tests/vmvx_person_detect.run
  TENSORFLOW_TESTS :: iree_tfl_tests/vulkan_mobilenet_v1.run

For these we'll need to bump the TF nightly version. @rsuderman was looking at doing it, I'm not sure how far along he is and if we can do this along with the integrate or do in 2 steps.

[MaheshRavishankar]

There are some failures about TOSA attribute type should be i32 (not i64). @jpienaar do you know who can help on this?

Failed Tests (11):
  TENSORFLOW_TESTS :: iree_tfl_tests/east_text_detector.run
  TENSORFLOW_TESTS :: iree_tfl_tests/gpt2.run
  TENSORFLOW_TESTS :: iree_tfl_tests/llvmcpu_mobilenet_v1.run
  TENSORFLOW_TESTS :: iree_tfl_tests/llvmcpu_mobilenet_v3-large_uint8.run
  TENSORFLOW_TESTS :: iree_tfl_tests/llvmcpu_resnet_50_int8.run
  TENSORFLOW_TESTS :: iree_tfl_tests/mnasnet.run
  TENSORFLOW_TESTS :: iree_tfl_tests/mobilenet_v3.run
  TENSORFLOW_TESTS :: iree_tfl_tests/person_detect.run
  TENSORFLOW_TESTS :: iree_tfl_tests/vmvx_mobilebert_tf2_quant.run
  TENSORFLOW_TESTS :: iree_tfl_tests/vmvx_person_detect.run
  TENSORFLOW_TESTS :: iree_tfl_tests/vulkan_mobilenet_v1.run

For these we'll need to bump the TF nightly version. @rsuderman was looking at doing it, I'm not sure how far along he is and if we can do this along with the integrate or do in 2 steps.

If this is just a TOSA issue, maybe just mark these as XFAIL, file a bug with this list. When the TFNightly gets bumped we can enable those again.

[hanhanW]

There are other similar failures in e2e model compilation.

[github-actions]

Abbreviated Benchmark Summary

@ commit 646fa07a5e43f6892f7cd2d91def16f9175f4e34 (no previous benchmark results to compare)

Raw Latencies

Benchmark Name	Average Latency (ms)	Median Latency (ms)	Latency Standard Deviation (ms)
BertForMaskedLMTF(stablehlo) [cuda-sm\_80-linux\_gnu-cuda][default-flags] cuda(none)[full-inference,default-flags] with zeros @ a2-highgpu-1g[gpu]	6.329	6.293	0.113
BertLargeTF(stablehlo) [cuda-sm\_80-linux\_gnu-cuda][default-flags] cuda(none)[full-inference,default-flags] with zeros @ a2-highgpu-1g[gpu]	10.523	10.511	0.036
BertLargefp16PTBatch1(linalg) [cuda-sm\_80-linux\_gnu-cuda][default-flags] cuda(none)[full-inference,default-flags] with zeros @ a2-highgpu-1g[gpu]	12.457	12.456	0.003

[Top 3 out of 40 results showed]

No improved or regressed compilation metrics 🏖️

For more information:

• Full benchmark result tables
• Build

Source Workflow Run

[MaheshRavishankar]

Thanks a lot @hanhanW . This was quite a lift.

@stellaraccident maybe you should make a go/no-go on this.. basically TOSA models need to be regenerated after this integrate. Not sure the disruption it will cause downstream.

It looks good to me.

[hcindyl]

Thanks a lot @hanhanW . This was quite a lift.

@stellaraccident maybe you should make a go/no-go on this.. basically TOSA models need to be regenerated after this integrate. Not sure the disruption it will cause downstream.

It looks good to me.

The breakage in tosa also means iree-samples tflitehub regression will be broken after this, as well as any other downstream projects that uses tflite as input.

[MaheshRavishankar]

Thanks a lot @hanhanW . This was quite a lift.
@stellaraccident maybe you should make a go/no-go on this.. basically TOSA models need to be regenerated after this integrate. Not sure the disruption it will cause downstream.
It looks good to me.

The breakage in tosa also means iree-samples tflitehub regression will be broken after this, as well as any other downstream projects that uses tflite as input.

I have no idea how else to do this apart from landing this and having downstream projects also adapt here...

[benvanik]

SGTM - such breakages happen - at least there's no cycles :)

[hanhanW]

Others seem to be depended on this, we should land this and re-enable the rest as a second step?

[hanhanW]

I think this is ready to land if someone can stamp it.