Summary:
Pull Request resolved: https://github.com/pytorch/pytorch/pull/27561
Adds memory_format keyword argument (positional for cpp).
'Preserve' behavior now follows next rules:
1) If tensor is non-overlapping and dense - output tensor will have the same strides as input tensor.
2) If not (1) and tensor is stored in the channels last format, output tensor going to have channels last format.
3) Output tensor is going to be contiguous in all other cases.
---
Dense tensor is the tensor that store values in a contiguous block of memory.
Non-overlapping tensor is the tensor in which elements occupy individual non-repetitive memory.
Test Plan: Imported from OSS
Differential Revision: D17980316
Pulled By: VitalyFedyunin
fbshipit-source-id: 2a1d47571268673de0c6f5ae1b6d4f9110962ab0
Summary:
This PR excises the last of SymbolicVariable. There should be no change in functionality. One new test for addmm fusion was added. A case where the peephole optimizer might convert a scalar argument remains untested.
Pull Request resolved: https://github.com/pytorch/pytorch/pull/25077
Test Plan: Refactors existing code so mostly covered by current tests. One test for addmm fusion was added.
Differential Revision: D17145334
Pulled By: mruberry
fbshipit-source-id: 6b68faf764f9ee8398b55c43110228ed9faf81eb
Summary:
Pull Request resolved: https://github.com/pytorch/pytorch/pull/24284
This PR finishes the unification of all Tensor types into a single object.
ProfiledTensorType is renamed to TensorType and the old TensorType is
deleted.
Notes:
* Fixes bug in merge for VaryingShape by changing its representation to an
optional list of optional ints.
* Removes ProfiledTensorType::create(type) invocations that can now
simply be expect calls on tensor type.
Test Plan: Imported from OSS
Differential Revision: D16794034
Pulled By: zdevito
fbshipit-source-id: 10362398d0bb166d0d385d74801e95d9b87d9dfc
Summary:
Fixes: https://github.com/pytorch/pytorch/issues/22833
grad_sum_to_size does not commute with AutogradAdd after all because it turns the broadcasting AutogradAdd into a broadcasting add.
Chillee did actually do most of the tracking down to the fusion of grad_sum_to_size and pinging me when he had found the cause. Thank you!
About the choice of removing the fusion completely instead of being more precise:
- We do have grad_sum_to_size elimination which works for cases where broadcasting does not actually happen in the forward, so the cases where the fusing of grad_sum_to_size is actually beneficial is much smaller than when initially proposed.
- There will be less fusion, in terms of the tests, IOU stops being fully fused. I vaguely think that it is a case we could handle with refined logic.
- Keeping it would add complexity in checking when to merge fusion groups to the complexities that this PR removes.
- The future of fusion probably lies more in more complete solutions including reductions (TVM or KeOps or our own or ...).
Pull Request resolved: https://github.com/pytorch/pytorch/pull/23372
Differential Revision: D16489930
Pulled By: soumith
fbshipit-source-id: bc0431b0d3eda264c401b634675872c4ce46f0f4
Summary:
This PR is a eliminates unneeded grad_sum_to_size and in particular speeds up the LSTM backward by allowing better fusion.
It consists of two parts:
- In AutoDiff, record broadcasting sizes only if the broadcast output size is different from the input size, otherwise record None.
- The specialization of Optional arguments (#18407) allows us to then eliminate ` _grad_sum_to_size(t, None)` in the peephole optimization step.
Thus, in the LSTM case, no SumToSize remain in the crucial fusion group. The trick here is that we can specialize on the runtime information from the forward.
I'm testing that different broadcasting situations lead to different graphs.
I didn't move all symbolic_script _grad_sum_to_size to the new logic, but it might be better to do this incrementally, anyway.
Pull Request resolved: https://github.com/pytorch/pytorch/pull/18697
Differential Revision: D15482076
Pulled By: wanchaol
fbshipit-source-id: 7f89367e35b8729910077c95c02bccefc8678afb
Summary:
move the insert_guard all the way up to the beginning of the decomposation, this will fix the case that we lose insert_point context after decomposeCommonNormalization and we still need to modify the graph.
fixes#19502
Pull Request resolved: https://github.com/pytorch/pytorch/pull/19646
Differential Revision: D15058040
Pulled By: wanchaol
fbshipit-source-id: ebdbf8623ebfe4556c461e1b650e94b905791adb
Summary:
I believe the existing check in FuseGraph was only `false` if PyTorch was built with NO_CUDA=1. Otherwise, we would create fusion groups even if we're on a CPU-only machine running CPU code. This is confusing. Instead I've made it so that the decision to fuse or not is dependent on if the producer Value is a known CPU tensor. If it is, we skip fusion.
Pull Request resolved: https://github.com/pytorch/pytorch/pull/19342
Differential Revision: D15038351
Pulled By: jamesr66a
fbshipit-source-id: fce9d83929309a7bf14346833f84b996f3e7f6db
Summary:
Fixes: #19253
Fixing pow(Tensor, float) is straightforward.
The breakage for pow(float, Tensor) is a bit more subtle to trigger, and fixing needs `torch.log` (`math.log` didn't work) from the newly merged #19115 (Thanks ngimel for pointing out this has landed.)
Pull Request resolved: https://github.com/pytorch/pytorch/pull/19324
Differential Revision: D15003531
Pulled By: ailzhang
fbshipit-source-id: 8b22138fa27a43806b82886fb3a7b557bbb5a865
Summary:
Partially fuse layer_norm by decomposing layer_norm into the batchnorm kernel that computes the stats, and then fusing the affine operations after the reduce operations, this is similar to the batchnorm fusion that apaszke did, it also only works in inference mode now.
Pull Request resolved: https://github.com/pytorch/pytorch/pull/18266
Differential Revision: D14879877
Pulled By: wanchaol
fbshipit-source-id: 0197d8f2a17ec438d3e53f4c411d759c1ae81efe
Summary:
This PR propagates where we use first-class modules objects into the compiler. This creates a transitionary state where:
* compiler.cpp creates Graphs where `self` is a Module class and attributes/parameters/buffers/submodules are looked up with `prim::GetAttr`
* GraphExecutor still runs "lowered graphs" where the self object has been removed by a compiler pass `lower_first_class_method`.
* Tracing still creates "lowered graphs", and a pass "lift_lowered_method" creates a first-class method graph for things.
* This PR separates out Method and Function. A script::Function is a pure Graph with no `self` bound. Similar to Python, a script::Method is just a bound `self` and its underlying `script::Function`.
* This PR also separates CompilationUnit from Module. A CompilationUnit is just a list of named script::Functions. Class's have a CompilationUnit holding the class methods, and Modules also have a CompilationUnit holding their Methods. This avoids the weird circular case Module --has a-> Class -> has a -> Module ...
Details:
* In this transitionary state, we maintain two copies of a Graph, first-class module and lowered. Th first-class one has a self argument that is the module's class type. The lowered one is the lowered graph that uses the initial_ivalues inputs.
* When defining lowered methods using `_defined_lowered` we immediately create the first-class equivalent. The reverse is done lazily, creating lowered_methods on demand from the class.
* The two way conversions will be deleted in a future PR when the executor itself runs first-class objects. However this requires more changes to (1) the traces, (2) the python bindings, and (3) the onnx export pass and would make this PR way to large.
Pull Request resolved: https://github.com/pytorch/pytorch/pull/19167
Differential Revision: D14891966
Pulled By: zdevito
fbshipit-source-id: 0b5f03118aa65448a15c7a7818e64089ec93d7ea
Summary:
Bug fix for https://github.com/pytorch/pytorch/issues/15043, where a large fusion in JIT with a large number of kernel arguments, which exceeds the limit allowed by nvrtc on a cuda device.
The fix is to check the number of arguments before a cuda kernel is generated. If the number exceeds the limit, take the runFallBack() path.
Add a reduced test from the original issue to keep the test time low. The test would fail without this fix.
Pull Request resolved: https://github.com/pytorch/pytorch/pull/18063
Differential Revision: D14691401
Pulled By: soumith
fbshipit-source-id: b98829bc89ed7724e91eda82ae3a5a1151af721a
Summary:
This PR did two things:
1. Enable scalar->float specialization in symbolic script, so AD formula that contains scalar in the schema, should write `float` instead.
2. add addcmul, lerp to AD and fuser.
Pull Request resolved: https://github.com/pytorch/pytorch/pull/18081
Differential Revision: D14490493
Pulled By: wanchaol
fbshipit-source-id: b3b86d960d5f051b30733bc908b19786111cdaa4
Summary:
so that functions like `def fn(x, p:float)` can be fused. Fixes#9940 and #11186. Fuses only float (not integer) arguments to simplify assembling arguments for fusion launch.
CPU fusion is disabled in CI and this won't be tested, but I tested it locally.
cc t-vi, apaszke
Pull Request resolved: https://github.com/pytorch/pytorch/pull/18087
Differential Revision: D14581206
Pulled By: wanchaol
fbshipit-source-id: ccb0cf79b1751706f9b2cdf1715115eae5a39fb6
Summary:
Pull Request resolved: https://github.com/pytorch/pytorch/pull/17585
Create a sugared value that represents a class during initialization. This is
so that assignments to attributes correctly define attributes in __init__ but
raise an error elsewhere.
Reviewed By: shannonzhu
Differential Revision: D14263403
fbshipit-source-id: 09b2feeb272302f00a79c2a0302fbdf5483aed6a
Summary:
Trying to land again, make prim::None into a case of prim::Constant. Reverted the previous landing because it broke an important onnx export test.
https://github.com/pytorch/pytorch/pull/16160
Pull Request resolved: https://github.com/pytorch/pytorch/pull/17186
Differential Revision: D14115304
Pulled By: eellison
fbshipit-source-id: 161435fc30460b4e116cdd62c7b2e5b94581dcb7
Summary:
Reenables rand_like fusion if no tensor is broadcasted in the fusion group. This is a sufficient but not necessary condition for fused rand_like to produce correct results, and it has an unpleasant side effect of falling back to non-fused path if rand_like was optimistically included in the fusion group, but there is a broadcast in the fusion group not necessarily related to rand_like. E.g. before this PR, if the network had (biasAdd -> relu -> dropout), fuser could fuse biasAdd and relu, now it will try fusing the whole thing (if dropout is expressed via rand_like) and fall back every time.
Pull Request resolved: https://github.com/pytorch/pytorch/pull/16087
Differential Revision: D13720232
Pulled By: zou3519
fbshipit-source-id: 1e19203bec4a59257bfc7078b054a19f00fab4ad
Summary:
This change simplifies analysis done on constants since prim::None does not need to be handled separately now. To check if a constant node is None, use node->isNone().
Next step will be to remove prim::Undefined.
Pull Request resolved: https://github.com/pytorch/pytorch/pull/16160
Differential Revision: D14109636
Pulled By: eellison
fbshipit-source-id: d26fd383976163a2ddd4c24984bd672a541cc876
Summary:
- Moved a few functions from `autograd` namespace to `aten` namespace to be visible from JIT nativeResolver.
- Added a hack to loop up keyword only argument. Will add proper support for kw only later
- Simulate function overload in aten using `_<number>` as function name suffix.
- Even `forward` returns multiple outputs like in `kthvalue`, there's at most one requires grad that we currently support.
- Removed the `TensorList` related ops here since partial `TensorList` support is prone to bugs. Our symbolic diff for `cat` was never tested with autodiff, and it seems broken. Need to find another proper way to support these ops(either by properly supporting `TensorList` or sth like `prim::ConstantChunk` and leave them for next PR.
Ops supported in this PR:
```
erf
expand_as
index
kthvalue
mean
permute
pow
rsub
select
sqrt
squeeze
t
to
topk
transpose
view
var
embedding
logsumexp
// grad is None
_dim_arange
contiguous
nonzero
ones_like
```
Pull Request resolved: https://github.com/pytorch/pytorch/pull/16689
Differential Revision: D14020806
Pulled By: ailzhang
fbshipit-source-id: a5e2c144a7be5a0d39d7ac5f93cb402ec12503a5
Summary:
Here is a fresh attempt at getting some fusion back in autodiff-generated graphs in the presence of SumToSize.
- The sum to size operator is now `aten::_grad_sum_to_size` to allow symbolic script differentiation (and that in turn would need to use this in place of sum_to_size to signal that it strictly operates on gradients). This is also used in the autodiff code, replacing `prim::SumToSize`.
- `_grad_sum_to_size` is now fusable, `cat`s - which are fused afterwards thanks to Adam's simplification of the code - are only fused if there is no `_grad_sum_to_size` in the fusion group.
- I push the `_grad_sum_to_size` out of the the fusion group when compiling and record the desired summations in the KernelSpec. The reasoning is the following:
- As the autodiff is a repeated applicaiton of the chain rule, we always have the pattern `grad_in = mm(A, grad_out)`, with A often diagonal for cases interesting to the fuser, whence it is `grad_in = a * grad_out` (a pointwise multiplication). We know that only `grad_out` may have AutodiffGradSumToSize applied, so we can commute AutodiffGradSumToSize with the `mul` (and `div` and `neg` are of similar origin).
- For `type_as` the gradient might be giving the type, so just skip SumToSize,
- `add` (which was inserted as `prim::AutogradAdd`) adding gradients when the forward used the same value in several places. This is non-broadcasting, so we know that the two arguments would have the same sizes as inputs - which is good so we don't have to do bookkeeping of the two parts.
Details:
- During fusion, the Tensor arguments are always kept as the first parameters of the fusion group to accomodate indexing assumptions in the fuser.
- The rewriting of the fusion group to record the necessary output transformation and eliminate `_grad_sum_to_size` from the fusion group is now in the fuser compile step.
- In the execution step, the arguments are split into Tensor / Non-Tensor and the non-tensor args are mostly forgotten about except for doing `sum_to_size` at the end. This would want to be improved if/when we fuse nonconstant scalar arguments.
- In a number of places in the fuser, the non-Tensor arguments to the fusion group needed to be ignored.
Thank you, apaszke for the insightful discussion. All bad ideas and errors are my own.
Pull Request resolved: https://github.com/pytorch/pytorch/pull/14957
Differential Revision: D13888173
Pulled By: zou3519
fbshipit-source-id: 071992c876e8b845f2b3e6329ae03a835d39a0ea
Summary:
This PR changes the way we store aliasing information from a "set" approach to a "points-to" analysis. Set-based approaches lose information in ways that make it difficult to do "live" updates to the alias DB as one as mutating the graph.
The tradeoff is that simple queries get more expensive, since they require traversing the points-to graph to answer most questions. In practice, this is unlikely to be that costly since we don't have massive aliasing chains, but we could create an approximation/caching layer if this becomes a problem.
My rough plan is:
1. This PR, switching to a points-to graph
2. Make it "live": analyzing a node should record all the edges the node added, so that we can rollback when the node is destroyed.
3. Reduce wildcard scope: we can make the wildcard a special vertex that points to anything that we're not "sure" about; namely, things that have been put inside lists, or graph inputs.
Pull Request resolved: https://github.com/pytorch/pytorch/pull/16386
Differential Revision: D13855117
Pulled By: suo
fbshipit-source-id: f009f58143173c275501624eb105d07ab60fe5e1
Summary:
This is the first of several PRs to simplify AliasDb usage.
- Hide the concept wildcards from users. They are too hard to think about and too easy to forget about.
- Start moving "mutability-safe" graph mutation methods into AliasDb (right now, the various methods that deal with topological move).
Eventually I want to create a "mutability-aware" handle to the graph. If you only use that handle to transform the graph, you can be sure that all transformations are safe with respect to mutability.
Pull Request resolved: https://github.com/pytorch/pytorch/pull/15656
Differential Revision: D13615492
Pulled By: suo
fbshipit-source-id: 5c39a157b4ea76f1f976315d06a314a89cc4f22f
Summary:
That makes that definition of a "fusable node" much simpler,
as we don't need to keep considering whether something has to be an
"exit node" at every step. The fuser now tries to maximize the
pointwise fusions first, and proceeds to prepending chunks and appending
concats only once a fix point is reached.
This patch not only makes the fuser much simpler to reason about,
making it siginifcantly easier to implement features like SumToSize
fusion, to improve performance of derivative graphs.
cc zou3519 mruberry
Pull Request resolved: https://github.com/pytorch/pytorch/pull/15633
Differential Revision: D13575306
Pulled By: zou3519
fbshipit-source-id: 0c55ea61d65d1f1ed3d75a8e1e83bc85a83f3aff
Summary:
We don't support reductions yet, but simply decomposing batch_norm
into a kernel that computes the stats, and the fusing everything else
with ReLU and following pointwise ops provides nice speedups.
Note that this is only limited to inference mode for now, because we
don't support convolutions and batch norm in AD, so the fuser isn't
applied to those parts.
This commit gives us a 7% end-to-end speedup for ResNet50 with batch size 32. Note that this only applies to inference mode at the moment due to lack of AD support for CNN operations (I'll be adding that soon), and not to the standard `torchvision` models, because they use in-place ops which aren't supported by the fuser (we need a way of proving that de-inplacing them is safe).
cc zou3519 zdevito mruberry ngimel
Pull Request resolved: https://github.com/pytorch/pytorch/pull/15146
Differential Revision: D13548303
Pulled By: zou3519
fbshipit-source-id: a2e2e5abc383f637fae19bd1b423f20c2cbc056a
Summary:
The PR clang-formats everything in `torch/csrc/jit/` and adds it to the pre-commit hook.
Here is a list of non-mechanical changes:
- I went over each file and fixed up whenever I could tell that clang-format was clobbering comment formatting.
- Made the macros in register_prim_ops a little more clang-format friendly by omitting trailing commas
- Refactored autodiff.cpp to use a helper class with explicit state rather than a bunch of capturing lambdas
- Small improvements to the precommit hook clang-format
Pull Request resolved: https://github.com/pytorch/pytorch/pull/15524
Differential Revision: D13547989
Pulled By: suo
fbshipit-source-id: 3ff1541bb06433ccfe6de6e33f29227a2b5bb493
Summary:
This PR adds the final set of clang-tidy checks we should add for our codebase: a last set of performance-related checks. Most fixes here are around changing `auto` to `const auto&` in a few places where unnecessary copies were made, and adding `reserve()` calls before loops doing repeated `push_back()`. Also a few cases of calling `std::string::find` with a single-character string literal instead of a single char, which uses a less efficient string search algorithm meant for searching larger substrings.
ezyang apaszke
Pull Request resolved: https://github.com/pytorch/pytorch/pull/15198
Differential Revision: D13468797
Pulled By: goldsborough
fbshipit-source-id: 2bed1ea1c7c162b7f3e0e1026f17125e88c4d5b2
Summary:
Anywhere we used #include "foo.h", we now say #include <foo.h>
Paths are adjusted to be rooted out of aten/src, torch/lib, or
the root level directory.
I modified CMakeLists.txt by hand to remove TH and THC from
the include paths.
I used the following script to do the canonicalization:
```
import subprocess
import re
import os.path
files = subprocess.check_output(['git', 'ls-files']).decode('utf-8').rstrip().split('\n')
for fn in files:
if not any(fn.endswith(suff) for suff in ['.cu', '.cpp', '.in', '.h', '.hpp', '.cu', '.cuh', '.cc']):
continue
if not any(fn.startswith(pref) for pref in ["aten/", "torch/"]):
continue
with open(fn, 'r') as f:
c = f.read()
def fmt(p):
return "#include <{}>".format(p)
def repl(m):
p = m.group(1)
if p in ["dlfcn.h", "unistd.h", "nvrtc.h", "cuda.h", "cuda_runtime.h", "cstdint", "cudnn.h", "Python.h", "cusparse.h", "cuda_runtime_api.h", "cuda_fp16.h", "cublas_v2.h", "stdint.h", "curand_kernel.h"]:
return fmt(p)
if any(p.startswith(pref) for pref in ["torch/csrc", "c10/", "ATen/", "caffe2/", "TH/", "THC/", "Eigen/", "gtest/", "zdl/", "gloo/", "onnx/", "miopen/"]):
return fmt(p)
for root in ["aten/src", "torch/lib", ""]:
for bad_root in [os.path.dirname(fn), "aten/src/TH", "aten/src/THC", "torch/csrc"]:
new_p = os.path.relpath(os.path.join(bad_root, p), root)
if not new_p.startswith("../") and (os.path.exists(os.path.join(root, new_p)) or os.path.exists(os.path.join(root, new_p + ".in"))):
return fmt(new_p)
print("ERROR: ", fn, p)
return m.group(0)
new_c = re.sub(r'#include "([^"]+)"', repl, c)
if new_c != c:
print(fn)
with open(fn, 'w') as f:
f.write(new_c)
```
Signed-off-by: Edward Z. Yang <ezyang@fb.com>
Pull Request resolved: https://github.com/pytorch/pytorch/pull/14849
Reviewed By: dzhulgakov
Differential Revision: D13363445
Pulled By: ezyang
fbshipit-source-id: 52361f878a672785f9306c9e9ab2513128092b68
Summary:
**Review only the last commit.**
This commit adds a few optimizations to AD, that let us dramatically
reduce the number of sizes we capture from forward.
We now:
- collapse chains of SumToSize
- avoid capturing sizes of tensors that are captured anyway
- more aggressively DCE the reverse code
- run CSE on the primal code to deduplicate `aten::size` calls
cc zou3519 zdevito
Pull Request resolved: https://github.com/pytorch/pytorch/pull/14758
Differential Revision: D13324440
Pulled By: zou3519
fbshipit-source-id: 45ccbc13605adcef2b461840c6089d3200000c72
Summary:
Dealing with so many `aten::size` calls (in particular calls on elements computed inside fusion groups) requires us to do some extra graph processing in the fuser (to compute the sizes by explicit broadcasts, instead of writing the intermediate tensors only to check their size). This restores the forward expects of LSTM and MiLSTM to a single big kernel. Unfortunately the backward is much harder, because as long as we can't prove that the reductions are unnecessary (or if we can't distribute them over the op), we will not be able to fuse them.
Pull Request resolved: https://github.com/pytorch/pytorch/pull/14558
Differential Revision: D13321748
Pulled By: zou3519
fbshipit-source-id: c04fc2f70d106d2bfb56206b5aec517a93b79d1f
Summary:
Fixes#14674. We won't have time for a proper fix before the release, so at least disable fusion of nodes that trigger incorrect behavior.
Pull Request resolved: https://github.com/pytorch/pytorch/pull/14752
Differential Revision: D13320407
Pulled By: zou3519
fbshipit-source-id: 2400f7c2cd332b957c248e755fdb0dadee68da5d
