Reduced indvar sizes - #1017
Draft
ramonwirsch wants to merge 4 commits into
Draft
Conversation
…r types, test this.
+ StructuredLoop now validates that its indvar container exists and is not unsigned, because our symbolic expressions do not handle unsigned entirely correct * fixed a lot of sdfg tests to not violate new constraint ! lots of opt tests still failing
Daisytuner Report - python_npbench (zinnia)@@ Benchmarks @@
===========================================================================
Benchmark Speedup (Time) ±% Energy Regions
===========================================================================
# adi
# numpy 116.40 ms -2.5% 11.21 J —
# docc-sequential 2.37x (49.16 ms) +4.7% 8.39 J 2
# docc-omp 2.06x (56.37 ms) -1.0% 9.12 J 2
# docc-cuda 2.40x (48.43 ms) -0.2% 4.62 J 6
# atax
+ numpy 53.21 ms -14.4% 9.38 J —
# docc-sequential 3.06x (17.37 ms) +0.4% 3.59 J 3
# docc-omp 2.43x (21.88 ms) +1.1% 4.71 J 3
# docc-cuda 2.87x (18.54 ms) -0.0% 1.84 J 7
# gemm
+ numpy 86.98 ms -15.9% 18.12 J —
# docc-sequential 3.47x (25.09 ms) +0.1% 7.19 J 1
# docc-omp 3.45x (25.21 ms) +0.5% 7.27 J 1
# docc-cuda 0.12x (751.91 ms) -0.0% 72.02 J 1
# gesummv
# numpy 128.51 ms -8.2% 22.04 J —
# docc-sequential 0.73x (176.61 ms) +0.5% 33.02 J 5
# docc-omp 1.76x (73.11 ms) +3.1% 15.43 J 5
# docc-cuda 10.52x (12.22 ms) -0.1% 1.26 J 11
# gemver
+ numpy 91.01 ms -12.1% 15.89 J —
# docc-sequential 11.26x (8.08 ms) -4.6% 1.71 J 7
# docc-omp 6.91x (13.18 ms) +2.0% 2.80 J 7
# docc-cuda 7.19x (12.65 ms) -0.0% 1.26 J 11
# k2mm
+ numpy 89.29 ms -13.0% 18.55 J —
# docc-sequential 0.66x (135.39 ms) +0.6% 24.36 J 3
# docc-omp 0.66x (135.95 ms) +0.1% 27.30 J 3
# docc-cuda 0.09x (940.75 ms) -0.0% 90.14 J 7
# k3mm
# numpy 76.32 ms -5.4% 19.06 J —
# docc-sequential 0.25x (309.73 ms) -0.1% 50.23 J 4
# docc-omp 0.23x (332.13 ms) -0.1% 56.93 J 4
# docc-cuda 0.05x (1.54 s) +0.1% 147.01 J 8
# mvt
+ numpy 45.68 ms -26.5% 8.41 J —
# docc-sequential 2.52x (18.11 ms) -0.2% 3.81 J 2
# docc-omp 2.53x (18.04 ms) +0.0% 3.71 J 2
# docc-cuda 0.13x (361.30 ms) +0.0% 35.45 J 2
# symm
# numpy 63.97 ms -0.1% 6.16 J —
# docc-sequential 4.34x (14.72 ms) +0.1% 2.92 J 3
# docc-omp 3.61x (17.70 ms) -4.3% 3.84 J 3
# syr2k
# numpy 75.25 ms +0.6% 7.30 J —
# docc-sequential 2.07x (36.37 ms) +2.5% 7.14 J 1
# docc-omp 2.11x (35.58 ms) -1.9% 7.04 J 1
# docc-cuda 2.54x (29.62 ms) -3.6% 2.88 J 1
# syrk
# numpy 63.21 ms +0.8% 6.07 J —
# docc-sequential 2.46x (25.71 ms) -0.4% 5.18 J 1
# docc-omp 2.38x (26.54 ms) +2.1% 5.26 J 1
# docc-cuda 2.78x (22.78 ms) +1.8% 2.26 J 1
# trmm
# numpy 69.78 ms +0.1% 6.75 J —
# docc-sequential 4.93x (14.14 ms) -0.3% 2.94 J 3
# docc-omp 4.27x (16.35 ms) -4.1% 3.42 J 3 |
Daisytuner Report - pytorch_models (chamomile)@@ Benchmarks @@
===========================================================================
Benchmark Speedup (Time) ±% Energy Regions
===========================================================================
# resnet18
# torch 74.98 ms -5.3% 19.80 J —
# torch-cuda 19.41 ms +4.7% 3.91 J —
# docc-omp 0.07x (1.08 s) -0.8% 331.94 J 45
# docc-cuda 0.13x (154.63 ms) +0.5% 22.34 J 93
# segformer
# torch-cuda 36.72 ms -1.5% 7.21 J —
# docc-cuda 0.30x (124.16 ms) -2.0% 17.67 J 266 |
…ructuredLoops ~ updated concat_expander, in_local_storage, out_local_storage to also use dynamic indvar types
Daisytuner Report - mlir_torch_models (chamomile)@@ Benchmarks @@
===========================================================================
Benchmark Speedup (Time) ±% Energy Regions
===========================================================================
# resnet18
# torch 78.50 ms +0.8% 20.63 J —
# torch-cuda 19.19 ms +6.0% 3.87 J —
# docc-omp 0.08x (999.06 ms) -1.1% 311.49 J 53
# docc-cuda 0.12x (158.58 ms) +2.3% 31.14 J 101
# segformer
# torch-cuda 36.65 ms -4.0% 7.42 J —
# docc-cuda 0.21x (175.16 ms) -0.6% 34.25 J 402 |
This file contains hidden or 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
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
1 participant
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.
More broadly replace hardcoded Int64 and UInt64 indvars with a function inferring the type from the expression (although for now, it only uses Int32 if it can prove that the immediate result fits into it, otherwise Int64).
Noticed that many places use UInt64, which is not technically valid and potentially problematic. Now enforcing no unsigned indvars on StructuredLoop validator and fixing test cases