Fix contiguity analysis of predicates to match updated contiguity. #1991

csarofeen · 2022-09-17T22:24:32Z

This PR depends on #1990

Extends contiguity method introduced there to predicate analysis. A few questions marked in comments but should be safe as is so the questions could be pushed to a follow up.

  int w = 15, x = 31;

  auto tv0 = makeContigTensor(4);
  fusion.addInput(tv0);
  auto tv1 = sin(tv0);
  auto tv2 = view(tv1, {w, x}, {x, w});

  fusion.addOutput(tv2);

  tv2->merge(0)->split(0, 4)->split(0, 8, false);

  TransformPropagator propagator(tv2);
  MaxRootDomainInfoSpanningTree spanning_tree(tv2);
  spanning_tree.traverse(&propagator);

Will now generate:

__global__ void CUDAGeneratedKernel(Tensor<float, 2> T0, Tensor<float, 2> T2) {
  NVFUSER_DEFINE_MAGIC_ZERO
  float T1[((8 * (ceilDiv((ceilDiv((31 * (ceilDiv((T0.size[0] * T0.size[1]), 31))), 4)), 8))) * 4)];
  #pragma unroll
  for(nvfuser_index_t i28 = 0; i28 < 8; ++i28) {
    #pragma unroll 1
    for(nvfuser_index_t i29 = 0; i29 < (ceilDiv((ceilDiv((31 * (ceilDiv((T0.size[0] * T0.size[1]), 31))), 4)), 8)); ++i29) {
      #pragma unroll
      for(nvfuser_index_t i30 = 0; i30 < 4; ++i30) {
        int64_t i66;
        i66 = (((i28 * (ceilDiv((ceilDiv((31 * (ceilDiv((T0.size[0] * T0.size[1]), 31))), 4)), 8))) + i29) * 4) + (i30 + nvfuser_zero);
        if ((i66 < (31 * (ceilDiv((T0.size[0] * T0.size[1]), 31))))) {
          T1[((((i28 * (ceilDiv((ceilDiv((31 * (ceilDiv((T0.size[0] * T0.size[1]), 31))), 4)), 8))) + i29) * 4) + i30)]
             = sinf(T0[i66]);
        }
      }
    }
  }
  NVFUSER_UPDATE_MAGIC_ZERO
  #pragma unroll
  for(nvfuser_index_t i31 = 0; i31 < 8; ++i31) {
    #pragma unroll 1
    for(nvfuser_index_t i32 = 0; i32 < (ceilDiv((ceilDiv((31 * (ceilDiv((T0.size[0] * T0.size[1]), 31))), 4)), 8)); ++i32) {
      #pragma unroll
      for(nvfuser_index_t i33 = 0; i33 < 4; ++i33) {
        int64_t i151;
        i151 = (((i31 * (ceilDiv((ceilDiv((31 * (ceilDiv((T0.size[0] * T0.size[1]), 31))), 4)), 8))) + i32) * 4) + (i33 + nvfuser_zero);
        if ((i151 < (31 * (ceilDiv((T0.size[0] * T0.size[1]), 31))))) {
          T2[i151]
             = T1[((((i31 * (ceilDiv((ceilDiv((31 * (ceilDiv((T0.size[0] * T0.size[1]), 31))), 4)), 8))) + i32) * 4) + i33)];
        }
      }
    }
  }
  NVFUSER_UPDATE_MAGIC_ZERO
}

No significant perf change detected with this PR which is expected

…n ca_map.

… lower_utils_cleanup

naoyam · 2022-09-28T06:18:21Z

This PR should close #2001

… contig_preds

csarofeen · 2022-09-29T12:31:25Z

@naoyam this should be ready for review.

naoyam

Comments so far

torch/csrc/jit/codegen/cuda/iter_visitor.h

torch/csrc/jit/codegen/cuda/lower_divisble_split.cpp

naoyam · 2022-09-29T17:43:37Z

torch/csrc/jit/codegen/cuda/contiguity.cpp

+        // stride. If we're computing predicates, then we don't want to do this,
+        // as when we mark something as non-contiguous we don't want it merged
+        // with any other domains.


I don't follow this. First of all, are we always assuming tensors are fully contiguous when generating predicates? That's what I remember, and if it's still the case, this part of the code should never be reachable for predicates.

I just tried to split up the conditional into two stages. The predicate path gets caught below in !ignore_consistent_ordering_ sorry for the confusion. Predicate will have ignore_consistent_ordering_ == true so it won't exit/return non-contiguous based on this check.

Comment is confusing, let me take another attempt at it.

naoyam · 2022-09-29T19:38:43Z

torch/csrc/jit/codegen/cuda/index_compute.cpp

+  for (auto root_i : c10::irange(predicate_contiguity.size())) {
+    auto root_id = consumer_root_domain[root_i];
+    if (root_id->maybePartial()) {
+      // predicate_contiguity[root_i] = false;


Is this safe to mark it as contiguous? Probably no unit test exercises this case, but as it was commented, predicates should be done at root domains in this case.

Fixed and added a unit test (FusionContigPredicateShift). The test should fail without the fix.

naoyam · 2022-09-29T19:40:43Z

torch/csrc/jit/codegen/cuda/index_compute.cpp

+         gather_expr->windowShape().at(root_i) != 1)) {
+      // TODO: The following line commented out didn't have any failures, is it
+      // needed?
+      predicate_contiguity[root_i] = false;


I think this is necessary. It probably just doesn't have tests. Let me add tests.

Fixed and added a unit test (FusionContigPredicateShift). The test should fail without the fix.

naoyam · 2022-09-29T22:01:21Z

torch/csrc/jit/codegen/cuda/contiguity.cpp

+      // If we're computing predicates (ignore_consistent_ordering_==true),
+      // then we don't have this same constraint, we can just ignore
+      // contiguity of the roots all together.


Makes sense. Thanks for clarification.

Actually, this isn't correct. Even for predicates, some root domains are marked as non-contiguous when they need to be predicated at root domains.

Some cleanups would be needed here. I'll work on it.

A problem with the design is root_contiguity is used both as indicating contiguity AND if domains are not necessary to be predicated at root. These are two separate concepts, so I split out the latter as final_ids.

naoyam · 2022-09-29T22:07:08Z

torch/csrc/jit/codegen/cuda/contiguity.cpp

+        // TODO: This didn't error when I removed "!ignore_consistent_ordering_"
+        // is it really needed?


Do you still want to keep this comment? It should actually not be necessary. If this is used for indexing, is_indexing_pass is true, so nothing should change. If used for predicates, !root_contiguity_ should be always true, so is_indexing_pass doesn't matter.

I actually like the current code as it's easier to understand the underlying logic.

naoyam

Stamping for now

naoyam · 2022-09-30T16:50:23Z

@csarofeen I pushed a few commits. Please let me know if they make sense. I already approved the PR so that this can be merged if they look good.

csarofeen · 2022-09-30T17:42:17Z

LGTM Thanks!

Syncing nvfuser devel branch to upstream master. https://github.com/csarofeen/pytorch/ Codegen changes include: * codegen improvement: i. allow non-root trivial reductions, allow empty/no-op fusion ii. fixes vectorization checks and size calculation iii. bank conflict handle improvement iv. enables transpose scheduler * misc: i. CI tests failure fixes ii. cpp tests file clean up iii. trivial forwarding supports added in codegen runtime iv. added factory methods support in codegen Commits that's in this PR from the devel branch: ``` 7117a7e patching nvfuser conv cudnn test numerics mismatch (#2048) 65af1a4 Inserting sync for redundant parallel types is already done at the (#2023) 6ac74d1 Fix sync map (#2047) f5bca33 Bank conflict checker improvements (#2032) d2ca7e3 Minor update on cp.async code generation. (#1901) d36cf61 Test file cleanup (#2040) 0b8e83f Allow non-root trivial reductions (#2037) a2dfe40 Fix vectorize size calculation (#2035) e040676 Use withPredicate to replace setPredicate to maintain Exprs immutable (#2025) 197221b removing ci workflow (#2034) 40e2703 Reduction rand like patch (#2031) bc77266 Add utility for checking bank conflict of shared memory (#2029) ddd1cf7 Add back FusionReductionWithTrivialReduction_CUDA (#2030) fbd97e5 Revert "Cleanup trivial reduction workarounds (#2006)" (#2024) bca20c1 Cleanup trivial reduction workarounds (#2006) e4b6585 Trivial forwarding (#1995) 1a0e355 Fix contiguity analysis of predicates to match updated contiguity. (#1991) a4effa6 Enable output allocation cache (#2010) 35440b7 Patching bn inference (#2016) 0f9f0b4 Add matmul benchmark (#2007) 45045cd Enable tests previously disabled due to an aliasing bug (#2005) 967aa77 Contiguous indexing for View operations (#1990) a43cb20 Make inlining even more modular (#2004) dc45835 Test util cleanup (#2003) 3ca21eb More strict validation (#2000) a7a7d57 Fix build problem (#1999) fc235b0 Just fixes comments (#1998) 482386c cleanup (#1997) 4cbe0db Improve divisible split detection (#1970) 42ccc52 Minor build fix. (#1996) fcf8c09 Cleanup of lower_utils.cpp: Isolate out GpuLower usage (#1989) 15f2f6d Move ConcretizedBroadcastDomains to shared_ptr in GpuLower. (#1988) 8f1c7f5 Minor cleanup lower_unroll.cpp (#1994) 1d9858c Minor cleanup (#1992) f262d9c Add support for uniform RNG (#1986) eb1dad1 Remove non-const functions, remove GpuLower instance on build, pass in ca_map. (#1987) 634820c Add support for some empty fusion (#1981) eabe8d8 Segment self mapping fusions (#1954) e96aacf Enable Transpose operation (#1882) 425dce2 Add a null scheduler that helps segmenting away no-op schedules (#1835) 306d4a6 Fix canScheduleCompileTime check of transpose scheduler (#1969) b1bd32c Minor fix (#1967) bd93578 Enable transpose scheduler (#1927) b7a206e Move scheduler vectorize utilities into their own file (#1959) d9420e4 View scheduling (#1928) c668e13 Upstream push ci fixes (#1965) c40202b Fix dump effective bandwidth (#1962) 93505bc WAR on index mapping when exact and permissive maps differ (#1960) 45e95fd Allow splitting inner-most ID to create virtual innermost ID in transpose scheduler (#1930) a3ecb33 Improve the comments at the beginning of index_compute.h (#1946) f7bc341 Remove unused variables (#1955) df3393a Some cleanup (#1957) 7d1d7c8 TVDomainGuard factory (#1953) 357ba22 Fill allocation with nan on tests (#1956) 8eafc54 Fix detection of unmappable root domains (#1952) 90a51f2 Some indexing cleanups, Add eye support (#1940) ddc01e4 Exclude unsupported data types (#1951) 992e17c test the groups the same order as they are merged (#1949) 208262b Move detection of self mapping IDs to IterDomainGraph from (#1941) ac4de38 Merge pull request #1945 from csarofeen/master_merge_0828 6310948 Add full, full_like, zeros, zeros_like, ones, ones_like (#1943) aab10bc Merge remote-tracking branch 'upstream/viable/strict' into HEAD 4c254c0 Fix arange when step is negative (#1942) 89330aa Tensor factories must set the output shape as its input (#1939) ``` RUN_TORCHBENCH: nvfuser Differential Revision: [D40869846](https://our.internmc.facebook.com/intern/diff/D40869846) Pull Request resolved: pytorch#87779 Approved by: https://github.com/davidberard98

csarofeen added 11 commits September 13, 2022 13:47

Propogate divisible split information.

d932fa6

Remove non-const functions, remove GpuLower instance on build, pass i…

7587976

…n ca_map.

Merge branch 'halo_info_cleanup' into contig_merge_split

e3bdff8

IterVisitor cleanup.

ea8405d

Move ConcretizedBroadcastDomains to shared_ptr in GpuLower.

45d291a

Isolate GpuLower usage into lower_utils.

41d5f9e

Merge branch 'devel' of https://www.github.com/csarofeen/pytorch into…

a5a393d

… lower_utils_cleanup

Merge branch 'concrete_bcast_cleanup' into contig_merge_split

e7b78a7

Merge branch 'lower_utils_cleanup' into contig_merge_split

445ffef

Rework contiguity analysis to work through view splits.

a23a7c3

Fix contiguity analysis of predicates to match updated contiguity.

1212cff

csarofeen requested a review from naoyam September 17, 2022 22:24

clang

3a38f0c

Base automatically changed from contig_merge_split to devel September 28, 2022 22:56

csarofeen added 2 commits September 28, 2022 19:01

Merge branch 'devel' of https://www.github.com/csarofeen/pytorch into…

10336d8

… contig_preds

Merge conflicts.

59b2639

naoyam reviewed Sep 29, 2022

View reviewed changes

csarofeen added 3 commits September 29, 2022 14:15

Please don't come back.

f8be3e0

Bad merge conflict resolution.

a6dc018

Clarity fix.

11b352e

naoyam reviewed Sep 29, 2022

View reviewed changes

naoyam added 3 commits September 29, 2022 17:34

Make sure partial IDs and shift are predicated at root

44da63e

cleanup

d550cf2

comment

54f805d

naoyam approved these changes Sep 30, 2022

View reviewed changes

csarofeen merged commit 1a0e355 into devel Sep 30, 2022

		// TODO: This didn't error when I removed "!ignore_consistent_ordering_"
		// is it really needed?

Fix contiguity analysis of predicates to match updated contiguity. #1991

Fix contiguity analysis of predicates to match updated contiguity. #1991

Uh oh!

Conversation

csarofeen commented Sep 17, 2022 • edited Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

Uh oh!

naoyam commented Sep 28, 2022

Uh oh!

csarofeen commented Sep 29, 2022

Uh oh!

naoyam left a comment

Choose a reason for hiding this comment

Uh oh!

Uh oh!

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

naoyam Sep 30, 2022 • edited Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

naoyam left a comment

Choose a reason for hiding this comment

Uh oh!

naoyam commented Sep 30, 2022

Uh oh!

csarofeen commented Sep 30, 2022

Uh oh!

Uh oh!

csarofeen commented Sep 17, 2022 •

edited

Loading

naoyam Sep 30, 2022 •

edited

Loading