Use torch.matmul in nn.Linear #1935

colesbury · 2017-06-28T23:08:57Z

This takes advantage of the broadcasting behavior of torch.matmul to
support inputs with more than two dimensions. The extra dimensions are
treated like part of the batch dimension, much like nn.Bottle in Lua
Torch.

There are a few related small performance changes:

Addmm computes the gradient in column-major for inputs in
column-major format
Variable.mm calls Addmm in-place with the desired output buffer

colesbury · 2017-06-28T23:43:15Z

Some micro-benchmark performance here:
https://gist.github.com/colesbury/fcfed34f6ddfde391adee71c67ec24a8

There's not a significant difference in perf with this PR for the sizes I tried.

One weird thing is that passing in a 3D input can sometimes be faster than the equivalent 2D input (and likely slower for some other cases). The difference is in the bias.grad computation. It's because:

maybe_unexpand is implemented using repeated calls to sum(0) for each dimension
Strangely, this is actually sometimes faster, probably because sum(dim) is not implemented very efficiently. (Too much work in each thread, not enough parallelism)

ngimel · 2017-06-29T00:07:41Z

Yeah, abandoned @csarofeen's PR torch/cutorch#751 tried to address that, I don't remember what happened to it. Most likely reduction kernels are starting a single threadblock.

This takes advantage of the broadcasting behavior of torch.matmul to support inputs with more than two dimensions. The extra dimensions are treated like part of the batch dimension, much like nn.Bottle in Lua Torch. There are a few related small performance changes: * Addmm computes the gradient in column-major for inputs in column-major format * Variable.mm calls Addmm in-place with the desired output buffer

ngimel · 2017-06-30T22:38:50Z

Microbenchmark 3 (64 x 2408) is much slower for me after this PR (.54 s vs .40 s before), because 3 memcopies are now done, one in each addmm (as opposed to 2 before). Even those 2 should not be needed, and at some point were not needed, I'll have to see how and why they appeared.

soumith · 2017-07-01T06:25:33Z

@ngimel if you can investigate and issue a fix, that'd be awesome. Over the next 1/2 weeks we lost Trevor and Paszke over time off, so things will be slower.

ngimel · 2017-07-01T16:52:54Z

I'm on vacation too:-)

soumith · 2017-07-01T16:53:18Z

haha okay, then i'll look into this. thanks :D

ngimel · 2017-07-01T17:32:23Z

What I've learned yesterday is that all 3 matrix multiplies in linear go through https://github.com/pytorch/pytorch/blob/master/torch/autograd/_functions/blas.py#L19-L26, and even if output and add_matrix data pointers there are the same, after passing through wrappers t and r_ in THC are not the same, and memcopy here https://github.com/pytorch/pytorch/blob/master/torch/lib/THC/generic/THCTensorMathBlas.cu#L245-L249 is triggered. Before PR forward matrix multiply did not trigger the memcopy, both backward did. After PR, all three MMs do. That benchmark looks like it's close to being CPU bound, so nvprof is pretty much useless for timings, can be used only for the order in which kernels are invoked and kernel durations. (I'm getting .8 s instead of .4-.5 when running under nvprof). I've seen the difference only with the latest (8.0.88) cublas, that improved pre-PR time to .4. With 8.0.61 both pre-PR and post-PR is around .5 s (times are on P100).

colesbury · 2017-07-03T18:29:37Z

I'll send a PR to remove the extra copies in the backwards.

The forward didn't previously trigger a memcopy because the bias addition was separate.

Previously it was:

gemm
add bias

Now it is:

copy bias to output
gemm

So there's an extra copy, but one fewer addition. I'm surprised this is slower. We can get old behavior by removing these three lines in nn/functional.py:

ngimel · 2017-07-03T18:51:41Z

I was also surprised this was slower. Extra Memcopy in THC addmm can't explain it, because it is tiny, must be something about the cpu overhead not being hidden with the new order, but hard to say for sure.

Syncing nvfuser devel branch to upstream master. https://github.com/csarofeen/pytorch/ Codegen changes include: - codegen improvement: i. improved view support on pointwise and transpose scheduler ii. grouped grid welford added for better outer-norm grid persistence in normalization - misc: i. new composite ops added: variance_mean , arange, ii. fixes misaligned address for transpose scheduler iii. refactor on separation of compilation API from execution API to prepare us for async compilation iv. double type support on expression evaluator v. PYTORCH_NVFUSER_DUMP refactor to save PTX and CUBIN Commits that's in this PR from the devel branch: ``` 89330aa Tensor factories must set the output shape as its input (#1939) b2fd01e arange support (#1933) 56c00fd Double support on all expression evaluators (#1937) 371f282 Improve trivial reduction merge support (#1931) 1d0c267 Test `rand` in a fusion with zero tensor input (#1932) 0dab160 Fix softmax bwd sizes. (#1890) ef98f36 Fix a bug (#1936) 63132a0 Propagate permissive mapping information into indexing pass (#1929) b4ac2c8 Map IterationDomains through view operations. (#1919) c0a187a do not use deprecated functions (#1935) 88de85e Upstream cherry pick fixes 0811 (#1934) b247dcf Separate kernel compilation API from kernel execution API (#1914) b34e3b9 Fix `ir_utils::hasBlockSync` + misc fixes in transpose scheduler (#1924) 14a53e6 Nullary RNGOp (#1892) 3c3c89e Misc fixes/tuning for transpose scheduler (#1912) 20cf109 Grouped grid welford (#1921) 6cf7eb0 Transpose scheduler small dim sizes better support (#1910) 9341ea9 Disabled ViewPersistentShmoo sizes that results in NAN (#1922) 057237f Fix CUDA driver error: misaligned address for transpose scheduler (#1918) 3fb3d80 Add variance_mean function using Welford (#1907) 98febf6 Remove DisableOption::UnrollWithRng (#1913) ee8ef33 Minor fix for the debug interface of using PTX directly (#1917) 6e8f953 Add PYTORCH_NVFUSER_DUMP options to save PTX and CUBIN (#1916) 5eefa9a dopt is only available since nvrtc 11.7 (#1915) 2ec8fc7 Kill computeAtBetween (#1911) d0d106a Improve view support on pointwise and transpose scheduler (#1906) e71e1ec Fix name clash of RNG with shared memory (#1904) 3381793 Fix mutator and sameAs for expanded IterDomain (#1902) ``` RUN_TORCHBENCH: nvfuser Differential Revision: [D39324552](https://our.internmc.facebook.com/intern/diff/D39324552) [ghstack-poisoned]

Syncing nvfuser devel branch to upstream master. https://github.com/csarofeen/pytorch/ Codegen changes include: - codegen improvement: i. improved view support on pointwise and transpose scheduler ii. grouped grid welford added for better outer-norm grid persistence in normalization - misc: i. new composite ops added: variance_mean , arange, ii. fixes misaligned address for transpose scheduler iii. refactor on separation of compilation API from execution API to prepare us for async compilation iv. double type support on expression evaluator v. PYTORCH_NVFUSER_DUMP refactor to save PTX and CUBIN Commits that's in this PR from the devel branch: ``` 89330aa Tensor factories must set the output shape as its input (#1939) b2fd01e arange support (#1933) 56c00fd Double support on all expression evaluators (#1937) 371f282 Improve trivial reduction merge support (#1931) 1d0c267 Test `rand` in a fusion with zero tensor input (#1932) 0dab160 Fix softmax bwd sizes. (#1890) ef98f36 Fix a bug (#1936) 63132a0 Propagate permissive mapping information into indexing pass (#1929) b4ac2c8 Map IterationDomains through view operations. (#1919) c0a187a do not use deprecated functions (#1935) 88de85e Upstream cherry pick fixes 0811 (#1934) b247dcf Separate kernel compilation API from kernel execution API (#1914) b34e3b9 Fix `ir_utils::hasBlockSync` + misc fixes in transpose scheduler (#1924) 14a53e6 Nullary RNGOp (#1892) 3c3c89e Misc fixes/tuning for transpose scheduler (#1912) 20cf109 Grouped grid welford (#1921) 6cf7eb0 Transpose scheduler small dim sizes better support (#1910) 9341ea9 Disabled ViewPersistentShmoo sizes that results in NAN (#1922) 057237f Fix CUDA driver error: misaligned address for transpose scheduler (#1918) 3fb3d80 Add variance_mean function using Welford (#1907) 98febf6 Remove DisableOption::UnrollWithRng (#1913) ee8ef33 Minor fix for the debug interface of using PTX directly (#1917) 6e8f953 Add PYTORCH_NVFUSER_DUMP options to save PTX and CUBIN (#1916) 5eefa9a dopt is only available since nvrtc 11.7 (#1915) 2ec8fc7 Kill computeAtBetween (#1911) d0d106a Improve view support on pointwise and transpose scheduler (#1906) e71e1ec Fix name clash of RNG with shared memory (#1904) 3381793 Fix mutator and sameAs for expanded IterDomain (#1902) ``` RUN_TORCHBENCH: nvfuser ghstack-source-id: d0d88cf Pull Request resolved: #84626

Syncing nvfuser devel branch to upstream master. https://github.com/csarofeen/pytorch/ Codegen changes include: - codegen improvement: i. improved view support on pointwise and transpose scheduler ii. grouped grid welford added for better outer-norm grid persistence in normalization - misc: i. new composite ops added: variance_mean , arange, ii. fixes misaligned address for transpose scheduler iii. refactor on separation of compilation API from execution API to prepare us for async compilation iv. double type support on expression evaluator v. PYTORCH_NVFUSER_DUMP refactor to save PTX and CUBIN Commits that's in this PR from the devel branch: ``` 89330aa Tensor factories must set the output shape as its input (#1939) b2fd01e arange support (#1933) 56c00fd Double support on all expression evaluators (#1937) 371f282 Improve trivial reduction merge support (#1931) 1d0c267 Test `rand` in a fusion with zero tensor input (#1932) 0dab160 Fix softmax bwd sizes. (#1890) ef98f36 Fix a bug (#1936) 63132a0 Propagate permissive mapping information into indexing pass (#1929) b4ac2c8 Map IterationDomains through view operations. (#1919) c0a187a do not use deprecated functions (#1935) 88de85e Upstream cherry pick fixes 0811 (#1934) b247dcf Separate kernel compilation API from kernel execution API (#1914) b34e3b9 Fix `ir_utils::hasBlockSync` + misc fixes in transpose scheduler (#1924) 14a53e6 Nullary RNGOp (#1892) 3c3c89e Misc fixes/tuning for transpose scheduler (#1912) 20cf109 Grouped grid welford (#1921) 6cf7eb0 Transpose scheduler small dim sizes better support (#1910) 9341ea9 Disabled ViewPersistentShmoo sizes that results in NAN (#1922) 057237f Fix CUDA driver error: misaligned address for transpose scheduler (#1918) 3fb3d80 Add variance_mean function using Welford (#1907) 98febf6 Remove DisableOption::UnrollWithRng (#1913) ee8ef33 Minor fix for the debug interface of using PTX directly (#1917) 6e8f953 Add PYTORCH_NVFUSER_DUMP options to save PTX and CUBIN (#1916) 5eefa9a dopt is only available since nvrtc 11.7 (#1915) 2ec8fc7 Kill computeAtBetween (#1911) d0d106a Improve view support on pointwise and transpose scheduler (#1906) e71e1ec Fix name clash of RNG with shared memory (#1904) 3381793 Fix mutator and sameAs for expanded IterDomain (#1902) ``` RUN_TORCHBENCH: nvfuser Differential Revision: [D39324552](https://our.internmc.facebook.com/intern/diff/D39324552) [ghstack-poisoned]

Syncing nvfuser devel branch to upstream master. https://github.com/csarofeen/pytorch/ Codegen changes include: - codegen improvement: i. improved view support on pointwise and transpose scheduler ii. grouped grid welford added for better outer-norm grid persistence in normalization - misc: i. new composite ops added: variance_mean , arange, ii. fixes misaligned address for transpose scheduler iii. refactor on separation of compilation API from execution API to prepare us for async compilation iv. double type support on expression evaluator v. PYTORCH_NVFUSER_DUMP refactor to save PTX and CUBIN Commits that's in this PR from the devel branch: ``` 89330aa Tensor factories must set the output shape as its input (#1939) b2fd01e arange support (#1933) 56c00fd Double support on all expression evaluators (#1937) 371f282 Improve trivial reduction merge support (#1931) 1d0c267 Test `rand` in a fusion with zero tensor input (#1932) 0dab160 Fix softmax bwd sizes. (#1890) ef98f36 Fix a bug (#1936) 63132a0 Propagate permissive mapping information into indexing pass (#1929) b4ac2c8 Map IterationDomains through view operations. (#1919) c0a187a do not use deprecated functions (#1935) 88de85e Upstream cherry pick fixes 0811 (#1934) b247dcf Separate kernel compilation API from kernel execution API (#1914) b34e3b9 Fix `ir_utils::hasBlockSync` + misc fixes in transpose scheduler (#1924) 14a53e6 Nullary RNGOp (#1892) 3c3c89e Misc fixes/tuning for transpose scheduler (#1912) 20cf109 Grouped grid welford (#1921) 6cf7eb0 Transpose scheduler small dim sizes better support (#1910) 9341ea9 Disabled ViewPersistentShmoo sizes that results in NAN (#1922) 057237f Fix CUDA driver error: misaligned address for transpose scheduler (#1918) 3fb3d80 Add variance_mean function using Welford (#1907) 98febf6 Remove DisableOption::UnrollWithRng (#1913) ee8ef33 Minor fix for the debug interface of using PTX directly (#1917) 6e8f953 Add PYTORCH_NVFUSER_DUMP options to save PTX and CUBIN (#1916) 5eefa9a dopt is only available since nvrtc 11.7 (#1915) 2ec8fc7 Kill computeAtBetween (#1911) d0d106a Improve view support on pointwise and transpose scheduler (#1906) e71e1ec Fix name clash of RNG with shared memory (#1904) 3381793 Fix mutator and sameAs for expanded IterDomain (#1902) ``` RUN_TORCHBENCH: nvfuser ghstack-source-id: 0443a2a Pull Request resolved: #84626

Syncing nvfuser devel branch to upstream master. https://github.com/csarofeen/pytorch/ Codegen changes include: - codegen improvement: i. improved view support on pointwise and transpose scheduler ii. grouped grid welford added for better outer-norm grid persistence in normalization - misc: i. new composite ops added: variance_mean , arange, ii. fixes misaligned address for transpose scheduler iii. refactor on separation of compilation API from execution API to prepare us for async compilation iv. double type support on expression evaluator v. PYTORCH_NVFUSER_DUMP refactor to save PTX and CUBIN Commits that's in this PR from the devel branch: ``` 89330aa Tensor factories must set the output shape as its input (#1939) b2fd01e arange support (#1933) 56c00fd Double support on all expression evaluators (#1937) 371f282 Improve trivial reduction merge support (#1931) 1d0c267 Test `rand` in a fusion with zero tensor input (#1932) 0dab160 Fix softmax bwd sizes. (#1890) ef98f36 Fix a bug (#1936) 63132a0 Propagate permissive mapping information into indexing pass (#1929) b4ac2c8 Map IterationDomains through view operations. (#1919) c0a187a do not use deprecated functions (#1935) 88de85e Upstream cherry pick fixes 0811 (#1934) b247dcf Separate kernel compilation API from kernel execution API (#1914) b34e3b9 Fix `ir_utils::hasBlockSync` + misc fixes in transpose scheduler (#1924) 14a53e6 Nullary RNGOp (#1892) 3c3c89e Misc fixes/tuning for transpose scheduler (#1912) 20cf109 Grouped grid welford (#1921) 6cf7eb0 Transpose scheduler small dim sizes better support (#1910) 9341ea9 Disabled ViewPersistentShmoo sizes that results in NAN (#1922) 057237f Fix CUDA driver error: misaligned address for transpose scheduler (#1918) 3fb3d80 Add variance_mean function using Welford (#1907) 98febf6 Remove DisableOption::UnrollWithRng (#1913) ee8ef33 Minor fix for the debug interface of using PTX directly (#1917) 6e8f953 Add PYTORCH_NVFUSER_DUMP options to save PTX and CUBIN (#1916) 5eefa9a dopt is only available since nvrtc 11.7 (#1915) 2ec8fc7 Kill computeAtBetween (#1911) d0d106a Improve view support on pointwise and transpose scheduler (#1906) e71e1ec Fix name clash of RNG with shared memory (#1904) 3381793 Fix mutator and sameAs for expanded IterDomain (#1902) ``` RUN_TORCHBENCH: nvfuser Differential Revision: [D39324552](https://our.internmc.facebook.com/intern/diff/D39324552) [ghstack-poisoned]

Syncing nvfuser devel branch to upstream master. https://github.com/csarofeen/pytorch/ Codegen changes include: - codegen improvement: i. improved view support on pointwise and transpose scheduler ii. grouped grid welford added for better outer-norm grid persistence in normalization - misc: i. new composite ops added: variance_mean , arange, ii. fixes misaligned address for transpose scheduler iii. refactor on separation of compilation API from execution API to prepare us for async compilation iv. double type support on expression evaluator v. PYTORCH_NVFUSER_DUMP refactor to save PTX and CUBIN Commits that's in this PR from the devel branch: ``` 89330aa Tensor factories must set the output shape as its input (#1939) b2fd01e arange support (#1933) 56c00fd Double support on all expression evaluators (#1937) 371f282 Improve trivial reduction merge support (#1931) 1d0c267 Test `rand` in a fusion with zero tensor input (#1932) 0dab160 Fix softmax bwd sizes. (#1890) ef98f36 Fix a bug (#1936) 63132a0 Propagate permissive mapping information into indexing pass (#1929) b4ac2c8 Map IterationDomains through view operations. (#1919) c0a187a do not use deprecated functions (#1935) 88de85e Upstream cherry pick fixes 0811 (#1934) b247dcf Separate kernel compilation API from kernel execution API (#1914) b34e3b9 Fix `ir_utils::hasBlockSync` + misc fixes in transpose scheduler (#1924) 14a53e6 Nullary RNGOp (#1892) 3c3c89e Misc fixes/tuning for transpose scheduler (#1912) 20cf109 Grouped grid welford (#1921) 6cf7eb0 Transpose scheduler small dim sizes better support (#1910) 9341ea9 Disabled ViewPersistentShmoo sizes that results in NAN (#1922) 057237f Fix CUDA driver error: misaligned address for transpose scheduler (#1918) 3fb3d80 Add variance_mean function using Welford (#1907) 98febf6 Remove DisableOption::UnrollWithRng (#1913) ee8ef33 Minor fix for the debug interface of using PTX directly (#1917) 6e8f953 Add PYTORCH_NVFUSER_DUMP options to save PTX and CUBIN (#1916) 5eefa9a dopt is only available since nvrtc 11.7 (#1915) 2ec8fc7 Kill computeAtBetween (#1911) d0d106a Improve view support on pointwise and transpose scheduler (#1906) e71e1ec Fix name clash of RNG with shared memory (#1904) 3381793 Fix mutator and sameAs for expanded IterDomain (#1902) ``` RUN_TORCHBENCH: nvfuser ghstack-source-id: 34c0b92 Pull Request resolved: #84626

Syncing nvfuser devel branch to upstream master. https://github.com/csarofeen/pytorch/ Codegen changes include: - codegen improvement: i. improved view support on pointwise and transpose scheduler ii. grouped grid welford added for better outer-norm grid persistence in normalization - misc: i. new composite ops added: variance_mean , arange, ii. fixes misaligned address for transpose scheduler iii. refactor on separation of compilation API from execution API to prepare us for async compilation iv. double type support on expression evaluator v. PYTORCH_NVFUSER_DUMP refactor to save PTX and CUBIN Commits that's in this PR from the devel branch: ``` 89330aa Tensor factories must set the output shape as its input (#1939) b2fd01e arange support (#1933) 56c00fd Double support on all expression evaluators (#1937) 371f282 Improve trivial reduction merge support (#1931) 1d0c267 Test `rand` in a fusion with zero tensor input (#1932) 0dab160 Fix softmax bwd sizes. (#1890) ef98f36 Fix a bug (#1936) 63132a0 Propagate permissive mapping information into indexing pass (#1929) b4ac2c8 Map IterationDomains through view operations. (#1919) c0a187a do not use deprecated functions (#1935) 88de85e Upstream cherry pick fixes 0811 (#1934) b247dcf Separate kernel compilation API from kernel execution API (#1914) b34e3b9 Fix `ir_utils::hasBlockSync` + misc fixes in transpose scheduler (#1924) 14a53e6 Nullary RNGOp (#1892) 3c3c89e Misc fixes/tuning for transpose scheduler (#1912) 20cf109 Grouped grid welford (#1921) 6cf7eb0 Transpose scheduler small dim sizes better support (#1910) 9341ea9 Disabled ViewPersistentShmoo sizes that results in NAN (#1922) 057237f Fix CUDA driver error: misaligned address for transpose scheduler (#1918) 3fb3d80 Add variance_mean function using Welford (#1907) 98febf6 Remove DisableOption::UnrollWithRng (#1913) ee8ef33 Minor fix for the debug interface of using PTX directly (#1917) 6e8f953 Add PYTORCH_NVFUSER_DUMP options to save PTX and CUBIN (#1916) 5eefa9a dopt is only available since nvrtc 11.7 (#1915) 2ec8fc7 Kill computeAtBetween (#1911) d0d106a Improve view support on pointwise and transpose scheduler (#1906) e71e1ec Fix name clash of RNG with shared memory (#1904) 3381793 Fix mutator and sameAs for expanded IterDomain (#1902) ``` RUN_TORCHBENCH: nvfuser Differential Revision: [D39324552](https://our.internmc.facebook.com/intern/diff/D39324552) Pull Request resolved: #84626 Approved by: https://github.com/malfet

soumith added the in progress label Jun 28, 2017

jekbradbury mentioned this pull request Jun 28, 2017

[feature request] time-distributed layers for application of normal layers to sequence data #1927

Open

colesbury force-pushed the linear branch from e54cca0 to e018821 Compare June 28, 2017 23:19

colesbury force-pushed the linear branch from e018821 to d9db6fa Compare June 29, 2017 16:08

soumith approved these changes Jun 30, 2017

View reviewed changes

soumith merged commit da0fad8 into pytorch:master Jun 30, 2017

soumith removed the in progress label Jun 30, 2017

colesbury deleted the linear branch July 3, 2017 17:28

bunelr mentioned this pull request Aug 14, 2017

Update documentation to reflect Linear with 2+D inputs #2410

Merged

ssnl mentioned this pull request Jan 19, 2018

Heuristic-based autograd execution order #4746

Merged

jjsjann123 pushed a commit to jjsjann123/pytorch that referenced this pull request Aug 30, 2022

do not use deprecated functions (pytorch#1935)

c0a187a

Provide feedback

Saved searches

Use saved searches to filter your results more quickly

Uh oh!

Use torch.matmul in nn.Linear #1935

Use torch.matmul in nn.Linear #1935

Uh oh!

colesbury commented Jun 28, 2017

Uh oh!

colesbury commented Jun 28, 2017

Uh oh!

ngimel commented Jun 29, 2017

Uh oh!

ngimel commented Jun 30, 2017

Uh oh!

soumith commented Jul 1, 2017

Uh oh!

ngimel commented Jul 1, 2017

Uh oh!

soumith commented Jul 1, 2017

Uh oh!

ngimel commented Jul 1, 2017

Uh oh!

colesbury commented Jul 3, 2017

Uh oh!

ngimel commented Jul 3, 2017

Uh oh!

Reviewers

Assignees

Labels

Projects

Milestone

Development

Uh oh!

3 participants

Use torch.matmul in nn.Linear #1935

Use torch.matmul in nn.Linear #1935

Uh oh!

Conversation

colesbury commented Jun 28, 2017

Uh oh!

colesbury commented Jun 28, 2017

Uh oh!

ngimel commented Jun 29, 2017

Uh oh!

ngimel commented Jun 30, 2017

Uh oh!

soumith commented Jul 1, 2017

Uh oh!

ngimel commented Jul 1, 2017

Uh oh!

soumith commented Jul 1, 2017

Uh oh!

ngimel commented Jul 1, 2017

Uh oh!

colesbury commented Jul 3, 2017

Uh oh!

ngimel commented Jul 3, 2017

Uh oh!

Reviewers

Assignees

Labels

Projects

Milestone

Development

Uh oh!

3 participants