Use TensorIndexer for the view tests

[naoyam]

Enabled TensorIndexer for the reshape tests.

I temporarily added a codegen diff result to this PR. This one is more concise as I disabled index hoisting. As far as I can see, there's no concerning change. I haven't verified everything, but I believe most of them are because TensorIndexer can detect more divisible splits, which helps generate simplified indices through more aggressive contig indexing.

Once approved, I'll remove the html file.

Context

Part of #4175.

I'm planning to enable the new indexer globally by default once we are sufficiently confident with it. I'm going to enable it for some of the C++ tests for now. Just manually checking the diff results seems to be the only way to gain some confidence.

All the tests are passing in my local branch, but just having green test results don't necessarily mean everything is properly ported to the new indexer. I'll also check perf changes with the benchmarks, but they may not give clear signals as indexing is just one piece of performance bottlenecks.

[naoyam]

!test --diff

[github-actions]

Review updated until commit 26ad85b

Description

• Enabled TensorIndexer for view and reshape tests

• Added setup methods to enable TensorIndexer options

• Temporarily included codegen diff results for review

---

Changes walkthrough 📝

	Relevant files
Enhancement	test_gpu_view.cpp Enable TensorIndexer in view and reshape tests                      tests/cpp/test_gpu_view.cpp Modified GpuViewTest to inherit from NVFuserTest and added SetUp method to enable TensorIndexer Modified ReshapeReduction to inherit from NVFuserFixtureParamTest and added SetUp method to enable TensorIndexer +15/-2

---

PR Reviewer Guide 🔍

Here are some key observations to aid the review process:

🧪 PR contains tests

⚡ Recommended focus areas for review Code Duplication The SetUp method is duplicated in both GpuViewTest and ReshapeReduction classes. Consider creating a base class to avoid code duplication. class GpuViewTest : public NVFuserTest { protected: void SetUp() override { NVFuserTest::SetUp(); EnableOptionsGuard::getCurOptions().set(EnableOption::IdModel, {"all"}); } }; Test Coverage Ensure that the new tests cover all edge cases and scenarios, especially since the new indexer is being enabled. Verify that the existing tests are still relevant and effective. TEST_F(GpuViewTest, FusionViewDtypeSameSizeOutput) { Performance Impact Evaluate the performance impact of enabling TensorIndexer globally. Conduct thorough benchmarking to ensure that the benefits outweigh any potential regressions. class GpuViewTest : public NVFuserTest { protected: void SetUp() override { NVFuserTest::SetUp(); EnableOptionsGuard::getCurOptions().set(EnableOption::IdModel, {"all"}); } };

[naoyam]

!test --diff

[jjsjann123]

github cannot show the diff 😢

Not totally sure how I should interpret the diff.

Searching by ^-
generated cuda indexing does look simpler. I'm a bit surprised to see the first code diff like these (line 11925 in the diff):

- if (threadIdx.x + (128 * blockIdx.x)) < 120)
+ if (threadIdx < 120)

I'm not holding anything against this PR, since it's only turning it on in the test. But let's remove the temporary file first before stamping it. I don't want to accidentally add that in our history.

[naoyam]

github cannot show the diff 😢

Not totally sure how I should interpret the diff.

Searching by ^- generated cuda indexing does look simpler. I'm a bit surprised to see the first code diff like these (line 11925 in the diff):

- if (threadIdx.x + (128 * blockIdx.x)) < 120)
+ if (threadIdx < 120)

I'm not holding anything against this PR, since it's only turning it on in the test. But let's remove the temporary file first before stamping it. I don't want to accidentally add that in our history.

IIRC, that's because the loop ID parallelized by BIDx is actually just a broadcast ID and that the new indexer is able to simplify the index.

Do you have any questions with other changes? Any concern?

I'm planning to enable the new indexer globally by default once we are sufficiently confident with it. I'm going to enable it for some of the C++ tests for now. Just manually checking the diff results seems to be the only way to gain some confidence.

All the tests are passing in my local branch, but just having green test results don't necessarily mean everything is properly ported to the new indexer. I'll also check perf changes with the benchmarks, but they may not give clear signals as indexing is just one piece of performance bottlenecks.

[naoyam]

github cannot show the diff 😢

No, it doesn't. Please download it and open it locally.

[jjsjann123]

IIRC, that's because the loop ID parallelized by BIDx is actually just a broadcast ID and that the new indexer is able to simplify the index.
Do you have any questions with other changes? Any concern?

Thanks. my earlier quick scanning does seem to see indexing code getting at least shorter. So that's a positive thing.

I'll also check perf changes with the benchmarks, but they may not give clear signals as indexing is just one piece of performance bottlenecks.

Are we seeing mixed performance impact? I don't think we necessarily have to answer all those questions, but is there any significant regression that's worth investigation?
Since this PR switches the indexing mode on only on cpp tests, not benchmark. Does this mean perf impact would still be later conducted on python benchmark, when we switch to use TensorIndexer by default?

I don't have much concern on this PR other than that question above.

But most importantly, let's remove the diff code so I can stamp it.

[naoyam]

!test

[jjsjann123]

thanks for getting me to double check.
Looks like tmp files are cleaned up.

🚢