Batch potrf #11796

t-vi · 2018-09-18T13:34:03Z

Native batched potrf native. One item of #7500
Also batch tril/triu in native.

This doesn't speparate out the single matrix versions, but I don't
think it is too inefficient.

This builds on the batch linear algebra systematic of #9949 by @vishwakftw

I didn't know how getri worked after using getrf, which led to some issues

…tch-inverse

@vishwakftw

Native batched potrf native. One item of pytorch#7500 Also batch tril/triu in native. This doesn't speparate out the single matrix versions, but I don't think it is too inefficient. This builds on the batch linear algebra systematic of pytorch#9949 by @vishwakftw .

t-vi · 2018-09-18T13:35:57Z

With apologies to @ethanluoyc for not being aware of #9623 before just the backward and tests were missing.

t-vi · 2018-09-18T14:30:18Z

Hmm. This looks like something in #9949 (@vishwakftw ).
Edit: The signature of random_fullrank_matrix_distinct_singular_value changed.

14:00:34   File "test_autograd.py", line 3083, in <module>
14:00:34     ('inverse', random_fullrank_matrix_distinct_singular_value(S, [2, 3]),
14:00:34   File "/var/lib/jenkins/workspace/test/common.py", line 660, in random_fullrank_matrix_distinct_singular_value
14:00:34     return torch.stack(all_matrices).reshape(*(batches + (l, l)))
14:00:34 RuntimeError: shape '[2, 3]' is invalid for input of size 150
14:00:34 Traceback (most recent call last):
14:00:34   File "test/run_test.py", line 392, in <module>
14:00:34     main()
14:00:34   File "test/run_test.py", line 384, in main
14:00:34     raise RuntimeError(message)
14:00:34 RuntimeError: test_autograd failed!
14:00:35 + cleanup

vishwakftw · 2018-09-18T14:43:30Z

#9949 requires a rebase, yes.

aten/src/ATen/native/BatchLinearAlgebra.cpp

+  auto m = self.size(-1);
+  auto self_batched_ = self.view({-1, n, m});
+  auto self_batched = self_batched_.accessor<scalar_t, 3>();
+  auto result_batched_ = result.view({-1, n, m});


ssnl · 2018-09-25T16:32:29Z

Thanks @t-vi, @zou3519 says that he will be taking a look. Could you rebase on top of master please?

t-vi · 2018-09-25T18:48:25Z

Sure thing. Where should I put the code? @vishwakftw moved the Gesv.* to BatchLinearAlgebra.* in his batch inverse PR (#9949) and I put my stuff in there. Should I keep the renaming?

Balandat · 2018-10-14T14:02:08Z

Is the plan to add a batched version of potrs as well? Doesn't seem that's part of this PR.

t-vi · 2018-10-14T14:26:01Z

**Edit**: I mixed that up with pstrf, which isn't on GPU as far as I know, sorry. Thanks vishwakftw for correcting me! My other priority is "pure GPU potrs".

vishwakftw · 2018-10-14T15:15:13Z

potrs is there for GPU, and the batched version too.

vishwakftw · 2018-11-10T01:22:10Z

@t-vi could you rebase?

t-vi · 2018-11-11T10:09:48Z

I looked into this, rebasing seems more effort than starting over.

vishwakftw · 2018-11-24T04:09:14Z

aten/src/ATen/native/cuda/BatchLinearAlgebra.cu

+      n = n % stride_batch;
+    } else if (stride_batch > stride_min) {
+      n = n - n % stride_max + n % stride_batch; // eliminate batch part
+    } // if stride_batch < stride min, the divisions below will eliminate batch


@t-vi Could you please give a small explanation of the logic here? It doesn't seem very obvious to me, unfortunately. Thank you.

Here, you have stride_max > stride_batch > stride_min you now want to eliminate the batch contribution in the offset. Subtracting n % stride_max will eliminate both the batch and the "lower" contribution, so adding n % stride_batch adds the second part back. As a result exactly the batch offset is removed.

I got it. Thank you very much.

vishwakftw added 30 commits July 27, 2018 15:34

Add batched inverse

0096374

Remove pin_memory from Gesv.cu

157280b

Missed #endif

bd43e52

Fix computation of inverses

6ed2f09

I didn't know how getri worked after using getrf, which led to some issues

Silly typographical error; my bad

0d5a133

Modify docs, derivative, and add test in test_autograd

47ffc82

Add tests in test_cuda, test_torch

69db465

Remove _batch_inverse from MVN

bf165f1

Fix include error

24b5d97

Rename files for consistency

1dfabaf

Clean up error messages, pass infos by ref

9a22157

Modify test case for full rank matrices

5fa97b3

Some changes

cc6b9e4

Merge branch 'master' into batch-inverse

50ef95d

exclude gesv and inverse functions from test_jit

5c78d86

fix nit

53fb608

Remove unnecessary Python bindings, fix inverse for empty tensors

752763f

Merge branch 'master' into batch-inverse

fba711b

fix CUDA build failure

b4ff55f

Merge branch 'master' into batch-inverse

5f94114

Update MiscUtils.h

2c86c31

Fix lint

a5b9fd8

Fix CUDA build failure

639f938

Merge branch 'master' into batch-inverse

dcf99d0

Fix some nits

241ae6b

Merge branch 'batch-inverse' of github.com:vishwakftw/pytorch into ba…

61802a4

…tch-inverse

Merge branch 'master' into batch-inverse

5866244

Merge branch 'master' into batch-inverse

c293f40

Remove check after batched getrf

d383d9f

self.type()._inverse_helper(self) --> at::_inverse_helper(self)

4c8cb4a

t-vi added 2 commits September 18, 2018 15:19

Batch potrf

f42c25a

Native batched potrf native. One item of pytorch#7500 Also batch tril/triu in native. This doesn't speparate out the single matrix versions, but I don't think it is too inefficient. This builds on the batch linear algebra systematic of pytorch#9949 by @vishwakftw .

Merge branch 'master' into batch_potrf

6a58987

t-vi requested review from apaszke, colesbury, ezyang, gchanan, soumith and zdevito as code owners September 18, 2018 13:34

update test after merge

6979ebd

fmassa reviewed Sep 18, 2018

View reviewed changes

add half for tril/triu, fix more tests

3e2edf7

t-vi mentioned this pull request Sep 20, 2018

MultivariateNormal and potrf is slow on gpu and seems to have some memory leak #11333

Closed

ssnl assigned zou3519 Sep 25, 2018

jacobrgardner mentioned this pull request Oct 8, 2018

Preconditioning for SV-DKL is slow cornellius-gp/gpytorch#320

Closed

Balandat mentioned this pull request Oct 14, 2018

[WIP] LazyTensors can handle multiple batches cornellius-gp/gpytorch#322

Merged

vishwakftw mentioned this pull request Oct 18, 2018

Rename potrf to cholesky #12699

Closed

t-vi closed this Nov 11, 2018

jjbouza mentioned this pull request Nov 19, 2018

Batched SVD using cuSolver #14175

Closed

vishwakftw reviewed Nov 24, 2018

View reviewed changes

ezyang added the open source label Jun 24, 2019

Batch potrf #11796

Batch potrf #11796

Uh oh!

Conversation

t-vi commented Sep 18, 2018

Uh oh!

t-vi commented Sep 18, 2018

Uh oh!

t-vi commented Sep 18, 2018 • edited Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

Uh oh!

vishwakftw commented Sep 18, 2018

Uh oh!

This comment was marked as off-topic.

Uh oh!

This comment was marked as off-topic.

Uh oh!

This comment was marked as off-topic.

Uh oh!

This comment was marked as off-topic.

Uh oh!

ssnl commented Sep 25, 2018

Uh oh!

t-vi commented Sep 25, 2018

Uh oh!

Balandat commented Oct 14, 2018

Uh oh!

t-vi commented Oct 14, 2018 via email • edited Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

Uh oh!

vishwakftw commented Oct 14, 2018

Uh oh!

vishwakftw commented Nov 10, 2018

Uh oh!

t-vi commented Nov 11, 2018

Uh oh!

vishwakftw Nov 24, 2018

Choose a reason for hiding this comment

Uh oh!

t-vi Nov 24, 2018

Choose a reason for hiding this comment

Uh oh!

vishwakftw Nov 24, 2018

Choose a reason for hiding this comment

Uh oh!

Reviewers

Assignees

Labels

Projects

Milestone

Development

Uh oh!

8 participants

t-vi commented Sep 18, 2018 •

edited

Loading

t-vi commented Oct 14, 2018 via email •

edited

Loading