Update the heuristic for AArch64 bmm/baddbmm #149122

michalowski-arm · 2025-03-13T13:25:15Z

Updates heuristic for bmm/baddbmm and consolidates all heuristic logic in a single location

The goal of the consolidation is to improve maintainability and readability of the heuristic logic. Instead of different parts scattered across two files, this patch centralizes everything inside Matmul.cpp, where there already exists heuristic-based selection for mkldnn.
The logic of the check itself doesn't change (existing code is reused where possible) but a separate heuristic threshold for bmm/baddbmm is introduced based on newer, benchmarking data. Use the script below to see the performance improvement for bmm from the new heuristic:

import torch
import time

# Set below to True to use cases selected by only one of the hueristics.
USE_ONLY_DIVERGENT_TEST_CASES = True   
BATCH_SIZES  = [ 1, 8, 32, 64, 128, 256 ]
M_DIMS       = [ 4, 8, 16, 32, 64, 256, 512 ]
N_DIMS       = [ 4, 8, 16, 32, 64, 256, 512 ]
K_DIMS       = [ 4, 8, 16, 32, 64, 256, 512 ]
ITERS = 50

def old_heuristic(m, n, k):
    is_above_min_dims = m > 8 and n > 8 and k > 8
    is_above_min_size = m*n*k > 8_192
    return is_above_min_dims and is_above_min_size

def new_heuristic(b, m, n, k):
    return b*b*m*n*k >= 4_194_304

def generate_test_cases():
   test_cases = []
   for b in BATCH_SIZES:
       for m in M_DIMS:
           for n in N_DIMS:
                   for k in K_DIMS:
                       if USE_ONLY_DIVERGENT_TEST_CASES:
                           if old_heuristic(m, n, k) != new_heuristic(b, m, n, k):
                               test_cases.append([b, m, n, k])
                       else:
                           test_cases.append([b, m, n, k])
   return test_cases

def test(x, y):
   for _ in range(5):
       torch.bmm(x, y)
   perf = 0.0
   for _ in range(ITERS):
       start = time.time()
       torch.bmm(x, y)
       end = time.time()
       perf += (end - start) / ITERS
   return perf

def main():
   print(f"{'b':<10}{'m':<10}{'n':<10}{'k':<10}{'time (s)':10}")
   cumulative_mean_time = 0.0
   for b, m, n, k in generate_test_cases():
       mean_time = test(torch.rand(b, m, n), torch.rand(b, n, k))
       cumulative_mean_time += mean_time
       print(f"{b:<10}{m:<10}{n:<10}{k:<10}{mean_time:10.3e}")
   print(f"Cumulative mean time = {cumulative_mean_time:.4f} s")


if __name__ == "__main__":
   main()

From the script we see that cumulative mean time from all test cases (at 16 threads) is:

1.6195 s for the old heuristic
0.7012 s for the new heuristic

cc @jgong5 @mingfeima @XiaobingSuper @sanchitintel @ashokei @jingxu10 @jerryzh168 @malfet @snadampal @milpuz01 @aditew01 @nikhil-arm @fadara01

pytorch-bot · 2025-03-13T13:25:20Z

🔗 Helpful Links

🧪 See artifacts and rendered test results at hud.pytorch.org/pr/149122

📄 Preview Python docs built from this PR
📄 Preview C++ docs built from this PR
❓ Need help or want to give feedback on the CI? Visit the bot commands wiki or our office hours

Note: Links to docs will display an error until the docs builds have been completed.

✅ No Failures

As of commit 53a6da1 with merge base a53d14d ():
💚 Looks good so far! There are no failures yet. 💚

This comment was automatically generated by Dr. CI and updates every 15 minutes.

malfet

Can you provide a bit more description on what this consolidation is trying to do?
Or is it just a better engineering change?
If it changes the semantic of the check, would be nice to have some sort of unit test added, if it's a performance only change would be nice to have script in PR description that one can run to observe those perf changes

michalowski-arm · 2025-03-14T12:05:40Z

Can you provide a bit more description on what this consolidation is trying to do? Or is it just a better engineering change? If it changes the semantic of the check, would be nice to have some sort of unit test added, if it's a performance only change would be nice to have script in PR description that one can run to observe those perf changes

Added some extra information in the PR description, let me know if anything more is needed.

aditew01 · 2025-03-19T09:55:10Z

aten/src/ATen/native/mkldnn/Matmul.cpp

+#if defined(__aarch64__)
+    const int64_t mkldnn_acl_bmm_baddbmm_threshold = get_mkldnn_acl_bmm_baddbmm_threshold();
+    // BATCH_SIZE^2 * M * N * K >= THRESHOLD
+    return mat1.size(0) * mat1.size(0) * mat1.size(1) * mat1.size(2) * mat2.size(2) >= mkldnn_acl_bmm_baddbmm_threshold;


For Neoverse cores, mkldnn_acl_bmm_baddbmm_threshold == 1L<<22
Does this imply we don't want to dispatch to MKLDNN for aten::baddbmm or aten::bmm cases?

The script attached only tests for float32 cases. Did we test the performance for float16 / bfloat16?

For Neoverse cores, mkldnn_acl_bmm_baddbmm_threshold == 1L<<22
Does this imply we don't want to dispatch to MKLDNN for aten::baddbmm or aten::bmm cases?

1L<<22 is 4,194,304 which is not as large as it seems.
for K, N = 1024 (which is not large by today's standards), we'll go to oneDNN/ACL with M=4 and batch size = 1

malfet · 2025-03-19T23:40:31Z

aten/src/ATen/native/mkldnn/Matmul.cpp

@@ -322,6 +322,42 @@ void mkldnn_matmul(

 }

+#if defined(__aarch64__)


This should be guarded behind USE_MKLDNN_ACL or something like that

fadara01 · 2025-04-07T07:07:54Z

@pytorchbot label "ciflow/linux-aarch64"

aten/src/ATen/native/mkldnn/Matmul.cpp

michalowski-arm · 2025-04-08T10:47:33Z

Some more performance data

Running the script from the description with dtype=torch.bfloat16 we get 11.1x speedup (no change for torch.float16)
Running t5-small, t5-base and t5-large models with dtype=torch.float32 at 16 threads:

=== Performance Comparison: Old vs New Heuristic (t5-small) ===
Batch Size | Model Speedup (%) | BMM Speedup (%) |   Old BMM CPU % |   New BMM CPU %
-------------------------------------------------------------------------------------
         1 |              0.70 |            0.00 |            3.21 |            3.23
         2 |             -1.47 |            1.05 |            5.03 |            4.91
         4 |              0.92 |            5.97 |            7.84 |            7.44
         8 |              4.28 |            9.75 |           11.63 |           10.97
        16 |             36.45 |           85.63 |           10.92 |            2.47
        32 |             48.33 |           90.33 |           14.85 |            2.78
        64 |             58.10 |           94.50 |           19.29 |            2.53
       128 |             69.77 |           96.93 |           21.09 |            2.14
       256 |             70.06 |           97.67 |           22.05 |            1.71
-------------------------------------------------------------------------------------

=== Performance Comparison: Old vs New Heuristic (t5-base) ===
Batch Size | Model Speedup (%) | BMM Speedup (%) |   Old BMM CPU % |   New BMM CPU %
-------------------------------------------------------------------------------------
         1 |             -3.39 |           -9.53 |            4.38 |            4.64
         2 |             -3.03 |           -5.84 |            6.00 |            6.17
         4 |             -0.41 |            3.05 |            9.41 |            9.09
         8 |             41.05 |           85.02 |           12.97 |            3.29
        16 |             44.94 |           88.77 |           13.24 |            2.70
        32 |             53.44 |           93.33 |           17.73 |            2.54
        64 |             65.31 |           96.36 |           19.65 |            2.06
       128 |             71.74 |           97.70 |           20.75 |            1.69
       256 |             71.65 |           98.22 |           21.30 |            1.34
-------------------------------------------------------------------------------------

=== Performance Comparison: Old vs New Heuristic (t5-large) ===
Batch Size | Model Speedup (%) | BMM Speedup (%) |   Old BMM CPU % |   New BMM CPU %
-------------------------------------------------------------------------------------
         1 |              0.51 |           -3.64 |            5.20 |            5.42
         2 |              2.66 |            4.22 |            6.74 |            6.63
         4 |              5.15 |            8.84 |            9.82 |            9.43
         8 |             44.21 |           85.09 |           13.58 |            3.63
        16 |             47.11 |           90.62 |           15.40 |            2.73
        32 |             57.76 |           94.74 |           17.58 |            2.19
        64 |             66.74 |           96.99 |           19.28 |            1.74
       128 |             70.75 |           97.94 |           20.13 |            1.42
       256 |             68.83 |           98.20 |           20.61 |            1.19
-------------------------------------------------------------------------------------

Running t5-small and t5-base models with ONEDNN_DEFAULT_FPMATH_MODE=BF16 at 16 threads:

=== Performance Comparison: Old vs New Heuristic (t5-small) ===
Batch Size | Model Speedup (%) | BMM Speedup (%) |   Old BMM CPU % |   New BMM CPU %
-------------------------------------------------------------------------------------
         1 |             -2.34 |           -3.67 |            3.69 |            3.74
         2 |             -1.30 |           -4.07 |            5.45 |            5.60
         4 |              0.82 |            9.57 |            9.01 |            8.22
         8 |              9.69 |           24.32 |           13.43 |           11.25
        16 |             56.74 |           91.30 |           17.83 |            3.58
        32 |             70.97 |           94.43 |           21.48 |            4.12
        64 |             80.46 |           96.38 |           23.81 |            4.41
       128 |             87.78 |           97.57 |           25.62 |            5.10
-------------------------------------------------------------------------------------

=== Performance Comparison: Old vs New Heuristic (t5-base) ===
Batch Size | Model Speedup (%) | BMM Speedup (%) |   Old BMM CPU % |   New BMM CPU %
-------------------------------------------------------------------------------------
         1 |              4.55 |           -2.32 |            3.82 |            4.10
         2 |              3.20 |            7.15 |            5.64 |            5.41
         4 |              3.72 |            7.05 |            9.20 |            8.88
         8 |             47.54 |           86.87 |           14.34 |            3.59
        16 |             63.61 |           93.27 |           18.58 |            3.44
        32 |             75.78 |           95.62 |           21.81 |            3.95
        64 |             84.57 |           97.02 |           24.09 |            4.66
       128 |             89.44 |           97.69 |           25.36 |            5.54
-------------------------------------------------------------------------------------

fadara01 · 2025-04-29T06:28:55Z

@pytorchbot label "ciflow/linux-aarch64"

fadara01 · 2025-04-29T06:33:24Z

@malfet @aditew01 can you guys give this another look?

aditew01

LGTM!

malfet

LGTM

malfet · 2025-05-13T23:28:22Z

@pytorchbot merge

pytorchmergebot · 2025-05-13T23:30:42Z

Merge started

Your change will be merged once all checks pass (ETA 0-4 Hours).

Learn more about merging in the wiki.

Questions? Feedback? Please reach out to the PyTorch DevX Team

Advanced Debugging

Check the merge workflow status
here

malfet · 2025-05-23T13:32:59Z

FYI this change is being reverted right now as causing 50% perf regression...

jeanschmidt · 2025-05-23T14:52:55Z

@pytorchbot revert -m "breaking internal models, @malfet may you help merge this?" -c ghfirst

pytorchmergebot · 2025-05-23T14:54:45Z

@pytorchbot successfully started a revert job. Check the current status here.
Questions? Feedback? Please reach out to the PyTorch DevX Team

@malfet

This reverts commit d759a51. Reverted #149122 on behalf of https://github.com/jeanschmidt due to breaking internal models, @malfet may you help merge this? ([comment](#149122 (comment)))

pytorchmergebot · 2025-05-23T14:54:58Z

@michalowski-arm your PR has been successfully reverted.

This PR was reopened (likely due to being reverted), so your approval was removed. Please request another review.

aditew01 · 2025-05-27T11:47:41Z

@malfet @jeanschmidt where can I look for the model break / regression suite to fix the heuristics ?

michalowski-arm · 2025-05-29T13:28:08Z

@malfet @jeanschmidt just pinging this to ask where we can see the perf results again? In my tests there was a general speedup with this heuristic for all except the smallest shapes, so to fix it I'd need to see the shapes being ran.

michalowski-arm · 2025-06-09T11:22:34Z

@malfet @jeanschmidt In this CI run we don't see any regressions: TorchInductor Performance DashBoard. This was for bf16 only, I don't think we can run f32. Are there other tests we are supposed to be running? Also, presumably the regression you saw was for AArch64, or did this change somehow impact x86 performance?

nikhil-arm · 2025-07-17T13:28:39Z

Waiting for Meta team to share details on internal regressions

malfet · 2025-08-06T22:40:28Z

@pytorchbot rebase

pytorchmergebot · 2025-08-06T22:42:09Z

@pytorchbot started a rebase job onto refs/remotes/origin/viable/strict. Check the current status here

pytorchmergebot · 2025-08-06T22:42:11Z

Rebase failed due to Command git -C /home/runner/work/pytorch/pytorch rebase refs/remotes/origin/viable/strict pull/149122/head returned non-zero exit code 1

Rebasing (1/2)
Auto-merging aten/src/ATen/native/LinearAlgebra.cpp
CONFLICT (content): Merge conflict in aten/src/ATen/native/LinearAlgebra.cpp
Auto-merging aten/src/ATen/native/mkldnn/Matmul.cpp
CONFLICT (content): Merge conflict in aten/src/ATen/native/mkldnn/Matmul.cpp
error: could not apply b382d29a47a... Update the heuristic for AArch64 bmm/baddbmm
hint: Resolve all conflicts manually, mark them as resolved with
hint: "git add/rm <conflicted_files>", then run "git rebase --continue".
hint: You can instead skip this commit: run "git rebase --skip".
hint: To abort and get back to the state before "git rebase", run "git rebase --abort".
hint: Disable this message with "git config set advice.mergeConflict false"
Could not apply b382d29a47a... # Update the heuristic for AArch64 bmm/baddbmm

Raised by https://github.com/pytorch/pytorch/actions/runs/16790309166

Replaced the defined(__aarch64__) with AT_MKLDNN_ACL_ENABLED() in the parts of Matmul.cpp relevant to the bmm heuristic

michalowski-arm requested review from lezcano, nikitaved and IvanYashchuk as code owners March 13, 2025 13:25

pytorch-bot bot added module: cpu CPU specific problem (e.g., perf, algorithm) release notes: linalg_frontend release notes category labels Mar 13, 2025

pytorchbot added the open source label Mar 13, 2025

malfet added triaged This issue has been looked at a team member, and triaged and prioritized into an appropriate module module: arm Related to ARM architectures builds of PyTorch. Includes Apple M1 labels Mar 13, 2025

malfet reviewed Mar 13, 2025

View reviewed changes

aditew01 requested a review from malfet March 19, 2025 09:31

aditew01 reviewed Mar 19, 2025

View reviewed changes

malfet reviewed Mar 19, 2025

View reviewed changes

pytorch-bot bot added the ciflow/linux-aarch64 linux aarch64 CI workflow label Apr 7, 2025

fadara01 reviewed Apr 7, 2025

View reviewed changes

aten/src/ATen/native/mkldnn/Matmul.cpp Show resolved Hide resolved

pytorch-bot bot removed the ciflow/linux-aarch64 linux aarch64 CI workflow label Apr 8, 2025

michalowski-arm requested review from malfet and aditew01 April 9, 2025 10:44

aditew01 added the ciflow/linux-aarch64 linux aarch64 CI workflow label Apr 10, 2025

pytorch-bot bot removed the ciflow/linux-aarch64 linux aarch64 CI workflow label Apr 22, 2025

pytorch-bot bot added the ciflow/linux-aarch64 linux aarch64 CI workflow label Apr 29, 2025

fadara01 previously approved these changes Apr 29, 2025

View reviewed changes

aditew01 previously approved these changes May 6, 2025

View reviewed changes

pytorch-bot bot removed the ciflow/linux-aarch64 linux aarch64 CI workflow label May 7, 2025

malfet previously approved these changes May 13, 2025

View reviewed changes

malfet added the ciflow/trunk Trigger trunk jobs on your pull request label May 13, 2025

pytorchmergebot added the merging label May 13, 2025

pytorchmergebot added the Merged label May 14, 2025

pytorchmergebot closed this in d759a51 May 14, 2025

pytorchmergebot removed the merging label May 14, 2025

pytorchmergebot added Reverted ci-no-td Do not run TD on this PR labels May 23, 2025

pytorchmergebot reopened this May 23, 2025

michalowski-arm added 2 commits August 7, 2025 13:54

Update the heuristic for AArch64 bmm/baddbmm

dc2c95e

Replace defined(__aarch64__) guard in Matmul.cpp

53a6da1

Replaced the defined(__aarch64__) with AT_MKLDNN_ACL_ENABLED() in the parts of Matmul.cpp relevant to the bmm heuristic

michalowski-arm force-pushed the bmm_heuristic branch from 4ac203f to 53a6da1 Compare August 8, 2025 09:59

pytorch-bot bot removed the ciflow/trunk Trigger trunk jobs on your pull request label Aug 8, 2025

		@@ -322,6 +322,42 @@ void mkldnn_matmul(

		}

		#if defined(__aarch64__)

Update the heuristic for AArch64 bmm/baddbmm #149122

Are you sure you want to change the base?

Update the heuristic for AArch64 bmm/baddbmm #149122

Uh oh!

Conversation

michalowski-arm commented Mar 13, 2025 • edited by pytorch-bot bot Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

Uh oh!

pytorch-bot bot commented Mar 13, 2025 • edited Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

🔗 Helpful Links

🧪 See artifacts and rendered test results at hud.pytorch.org/pr/149122

✅ No Failures

Uh oh!

malfet left a comment

Choose a reason for hiding this comment

Uh oh!

michalowski-arm commented Mar 14, 2025

Uh oh!

aditew01 Mar 19, 2025

Choose a reason for hiding this comment

Uh oh!

fadara01 Apr 7, 2025 • 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!

malfet Mar 19, 2025

Choose a reason for hiding this comment

Uh oh!

michalowski-arm Apr 8, 2025

Choose a reason for hiding this comment

Uh oh!

fadara01 commented Apr 7, 2025

Uh oh!

Uh oh!

michalowski-arm commented Apr 8, 2025

Uh oh!

fadara01 commented Apr 29, 2025

Uh oh!

fadara01 commented Apr 29, 2025

Uh oh!

aditew01 left a comment

Choose a reason for hiding this comment

Uh oh!

malfet left a comment

Choose a reason for hiding this comment

Uh oh!

malfet commented May 13, 2025

Uh oh!

pytorchmergebot commented May 13, 2025

Merge started

Uh oh!

malfet commented May 23, 2025

Uh oh!

jeanschmidt commented May 23, 2025

Uh oh!

pytorchmergebot commented May 23, 2025

Uh oh!

pytorchmergebot commented May 23, 2025

Uh oh!

aditew01 commented May 27, 2025

Uh oh!

michalowski-arm commented May 29, 2025

Uh oh!

michalowski-arm commented Jun 9, 2025

Uh oh!

nikhil-arm commented Jul 17, 2025

Uh oh!

malfet commented Aug 6, 2025

Uh oh!

pytorchmergebot commented Aug 6, 2025

Uh oh!

pytorchmergebot commented Aug 6, 2025

Uh oh!

Uh oh!

michalowski-arm commented Mar 13, 2025 •

edited by pytorch-bot bot

Loading

pytorch-bot bot commented Mar 13, 2025 •

edited

Loading

fadara01 Apr 7, 2025 •

edited

Loading