* Add dtypes (with reasonable defaults) to sum, prod, cumsum, cumprod.
This adds optional dtypes to torch.sum, torch.prod, torch.cumsum, torch.cumprod.
By default, the dtype is torch.float64 for integral types, and the dtype of the input for floating point types.
* Don't use optional<ScalarType>, because the jit can't handle it yet.
Instead, we manually build the overloads. This is fairly painful because of default arguments, but should be easy to pull out once the jit can handle optional<ScalarType>.
* Fix keepdim with out parameters.
* Fix _cudnn_rnn_flatten_weight.
* If dtype is provided to an out function, make sure it matches the dtype of the result.
* Fix typo.
* Split set_default_tensor_type(dtype) into set_default_dtype(dtype).
* Fix flake8.
The difference between this one and set_default_tensor_type is that it only sets scalar type what determines the type + device of a tensor returned from a factory function with defaults is the default tensor type + the current device (if the default tensor type is cuda). This just changes the scalar type of the default tensor type.
We do eventually want to deprecate set_default_tensor_type; it is not clear how to do that in a sensible and backwards compatible way.
* More precise digamma
Fixes#6190.
This is a rebase of #3955 with some tweaks for better performance around
poles. The code is ported over from cephes with permission.
By itself, the cephes code returns inf for the poles.
For better performance around the poles with float32, one intermediate
step is always computed with double precision, regardless of dtype.
This step does `PI / tan(PI * input)`. This is necessary because small (1e-6)
rounding errors for the inputs to tan have strong effects on the output
(ie, the derivative of tan is very large at some points).
* Replace usages of finite-differences digamma with newly implemented digamma
* Better behavior near and at poles
* ScalarConvert -> scalar_cast for readability
* Separate cuda-ness from dtype.
There are no longer torch.cuda.int64, etc; only torch.int64 that correspond to at::ScalarType.
At the python arg parser level, the corresponding ATen type is selected from the combination of (ScalarType, Layout, Device).
There is also currently unused code in here for support ScalarType in native_functions; this will be used for specifying aggregate types
on reduction functions.
* Fix test_autograd.
* Add defaults to randint_like.
* Track is_cuda in py tensor types.
* Fix test_sparse.
* Fix multiprocessing.
* Fix rnn.
* Fix test_nn.
* Fix flake8.
* change irfft signal_sizes arg to be the last
* add docs for fft, ifft, rfft, irfft; update doc for stft
* fix typo in window function docs
* improve gradcheck error message
* implement backward of fft, ifft, rfft, irfft
* add grad tests for fft, ifft, rfft, irfft
* fix nits and typos from #6118
* address comments
* fix fft when any of the input dimensions is not like complex type; add test for ifft+fft
* clarify the comments
* Address comments: add note; add helper function
* use at::nullopt
* add notes on conjugate symmetry; fix complex-to-real cloning condition (should be advanced data layout rather than base_istride)
* add at::sum_intlist and at::prod_intlist
* revert optional<vector> helper due to windows compiler error
* Add string-style devices to all tensors.
Previously, tensors only had a 'get_device' method which would throw an exception on a CPU tensor. This made it necessary to if/else code that
was meant to be device agnostic.
This PR implements the following:
1) Adds a 'device' property to all tensors that returns a string representation of the device for all tensors.
For cpu tensors this is 'cpu'. For cuda tensors this is 'cuda:X', where X is the cuda device ordinal.
2) Adds a DeviceSpec class. This is just a helper class for separating device_type and device_index specification and to allow partial specification.
For example, you can call DeviceSpec('cuda'), DeviceSpec('cuda:0'), DeviceSpec('cuda', 1).
Also has backwards compatibility support for specifying integers, which are treated as cuda devices.
DeviceSpecs have the following properties:
a) device_type: string representation of the device type (i.e. 'cpu' or 'cuda')
b) device_index: integer for the device index (None if not specified)
c) cuda_device_index: for backwards compatibility; behaves roughly like `get_device` did previously. I.e. if a function previously took integers for cuda devices,
it can now take DeviceSpecs (or strings), and can maintain the old functionality by calling `old_index = DeviceSpec(old).cuda_device_index`.
3) tensor methods and torch. functions that took integer devices can now take integers, strings, or DeviceSpecs. For example:
torch.randn((2,3), dtype=torch.cuda.float32, device='cuda:1')
TODO in future PRs:
A) Split out cuda from dtype so you don't need to overspecify cuda-ness
B) We currently only support strings/DeviceSpecs in tensor methods and torch. functions. We should have equivalents torch.cuda.device(...), torch.cuda.device_of, etc.
at the torch. level that work on strings/DeviceSpecs
* Add deviceInt64 to python arg parser.
* device_str.
* Remove device_str.
* remove device prefix from attributes.
* Use const char * instead of string.
* Move autogpu index out of Device.
* comment on is_default.
* Rename torch.DeviceSpec to torch.device.
* comment.
* Fix tests.
* Fix flake8.
* Fix sparse_coo_tensor parameter name.
* Improve error message.
* Remove device_ prefix from C++ device object.
* Allocate static strings.
* Return not implemented from rich compare.
* Move torch::Device to THPDevice.
* Remove cuda index.
* Py_RETURN_NOTIMPLEMENTED doesn't exist in python2.
* Implemented log2 and log10
* Re-add incorrectly removed files
* Fix minor bugs
* Fix log1p docs
* Add a try-except for python2 math module in log2 test
* Revert changes made to aten/doc/*
* Fix docstring errors
* Fix windows build
* Add max_values and argmax convenience functions to ATen
* Add documentation for torch.argmax/argmin and skip max_values
* Add tests for argmax/argmin
* Dont default the dim argument
* Use dim=0 in test_torch.py for argmax tests
* Implement argmin() and argmax() without dim
* Call .contiguous() before .view(-1)
This changes type(tensor) to return `torch.Tensor` instead of
`torch.autograd.Variable`.
This requires a few implementation changes:
- torch.Tensor is now a regular Python class instead of a
pseudo-factory like torch.FloatTensor/torch.DoubleTensor
- torch.autograd.Variable is just a shell with a __new__ function.
Since no instanes are constructed it doesn't have any methods.
- Adds torch.get_default_dtype() since torch.Tensor.dtype returns
<attribute 'dtype' of 'torch._C._TensorBase' objects>
This compares the torch function against the reference math funciton
against a relative small set of inputs, including integers, extremes
of some common functions, zero, a few numbers from randn and a few
numbers near 1e6.
The idea here is not to be completely exhaustive, but rather quickly
expose the most common bugs. For exhaustive checks, we should evaluate
torch functions against all ~4e9 possible float32 value.
We compare the torch function evaluated against contiguous
and non-contiguous inputs and large vs. small tensors.
Also:
- Make torch.allclose work with nan and +/-inf
- Add torch.isclose (like numpy.isclose)
- Add torch.testing.assert_allclose (like
numpy.testing.assert_allclose)
* Introduce torch.layout and split layout from dtypes.
Tensors (and tensor types) now have a 'layout' attribute that returns either 'torch.strided' or 'torch.sparse_coo'.
Previously, dtypes were 1-to-1 with ATen types/PyTensorTypes; the impetus behind this decision was to make things easy in the common case
(i.e. specifying a type in a factory function). But this doesn't really follow for sparity, which isn't a common case.
It also doesn't properly represent the concept or a dtype, which in numpy are proper scalar types (i.e. roughly the type returned from indexing the
last dimension of an n-d array). But this should be the same whether or not the tensor is represented via strides, sparsity, etc.
This is accomplished by:
1) having the dtype of tensor return the (device-type, scalar-type) combination, i.e. torch.cuda.float32, so both
torch.cuda.FloatTensor and torch.cuda.sparse.FloatTensor have the same dtype
2) Adding a layout parameter to python functions, where the combination of (dtype, layout) maps to an ATen type that is used for dispatch.
* Formatting, make init throw python_error.
* Fix cuda not enabled error message.
* Fix test.
Perf numbers:
https://gist.github.com/colesbury/9e28dd7b0f27b0b019f68adbd4bd4b88
I've changed the dispatch stub so that it doesn't require every kernel
to be compiled for every instruction set. Kernel implementations are
stored in the stub's table with the REGISTER_DISPATCH macro.
I've also moved vec256 to it's own folder and split up the
specializations before they get too unwieldy.
Change UnaryOpsKernel to use new DisaptchStub
- Prefer signed integers. Mixing signed and unsigned integers is a
pain and ATen mostly uses signed integers (int64_t).
- Use inline lambda instead of struct for UnaryOps
- Rename partial load overload "load_partial"
This is in preparation for splitting out sparsity (layout) from dtypes; it's complex to maintain these
and tensor.new(...) is a legacy API in any case.
* Add numpy.array-like type inference to torch.tensor.
* Temporary fix for int/double types.
* Treat python floats as the default (scalar) dtype.
* Also make 0-length sequences the default scalar type and add more tests.
* Add type inference to sparse_coo_tensor.
* Fix sparse test.
* Remove allow_variables.
* Check numpy platform bits.
* Address review comments.
* Make suggested changes to constraints.
* More checking windows builds.
* Fix test for windows.
* Support legacy empty tensor behavior in cat
Continuing from #5837:
Fixes#5332.
Currently, the following behavior happens with torch.cat:
```
import torch
x = torch.randn(4, 3, 32, 32)
empty = torch.Tensor([])
res1 = torch.cat([x, empty], dim=1)
res2 = torch.cat([empty, x], dim=1)
```
However, at some point in the past, res1 and res2 were equal. This PR
supports the legacy behavior of ignoring empty tensors when
concatenating a list of tensors, until we have empty tensors that can
have arbitrary shape, at which point we'll stop supporting this
behavior.
* Address comments
* Fix integer overflow in remainder
* Fix remainder operator in CUDA
* Add tests for remainder integer overflow
* Add has_different_sign static function
* 1. Add logdet and slogdet in ATen side
2. Previously, det can return result with incorrect sign upon seeing symmetric
matrices. This is caused by the wrong assumption I had on SVD (when input is
symmetric U=V^T). This fixes it.
3. Moreover, after fixing 2 now QR is always needed for det forward. So I moved
SVD to backward call. Since this is a specific variant of SVD, it is named as
_svd_with_positive_UV_det, with derivative.yaml entry being svd_backward.
4. Updated/added backward functions for det, logdet and slogdet, which uses
_svd_with_positive_UV_det and svd_backward inside.
5. Optimized svd_backward:
a. Avoid unnecessary kernels when only sigma has gradient (this is the usual
case, and also true with *det backward functions).
b. Fix SVD double backward by avoiding a nan.
* 1. Add/update grad checks for det, logdet, and slogdet.
2. Fix an incorrect check for dim_args_idx in test_autograd.py
3. Add option to only test a subset of output values, specified by
test_output_indices, for cases like slogdet where only the
second output is differentiable.
4. Add better doc for the test generating list.
* Add/improve output tests for det, logdet and slogdet
Add a scaling to random matrices so closeness checks are more robust
* Remove unnecessaery Variable wrappers in some test files
* Add logdet slogdet docs
* Improve an err msg in THTensorLapack.c
* add inverse-based backward for invertible matrices
use svd only for non-invertible case, so don't need the special variant anymore
* use LU rather than QR
#5481 was reverted due to a strange test bug. This PR attempts to fix that.
This diff adds vectorization to ATen. It uses intel intrinsics to build a general vec256 class, that represents types of 256bit width. These can then be treated like regular variables. Using those it implements torch.sum() for the contiguous case. It uses Intel TBB for multithreading, which allows workstealing and chunks the reduction operations based on a experimentally chosen value (_THRESHOLD). It uses cpuinfo to pick the right code depending on the host's capabilities.
The kernels are implemented under native/cpu. Each .cpp file is compiled with -avx, -avx2 and no additional flags. A macro is used to append AVX, AVX2 or NONE to the function name. The header then needs to define the functions three times, one for each capability. This could be improved by either changing the cmake file a bit or possibly generating source code using a Python script etc.
For the non-contiguous case this defaults to the current implementation within TH. For CUDA is entirely defaults to the implementation within THC.
There probably needs to be a bit of a debate around the design decisions here, the additional dependencies, parallelization strategy, clarity, etc. The numerical results also diverge from numpy with larger tensors, which is expected since we're summing, for example, 8 numbers and then adding the result to the running sum, instead of each number one by one. But there might be something to be said about accumulating into a double for floats or the degree of divergence, the behavior with respect to CUDA, etc.
I wrote a [small Python script]( https://github.com/cpuhrsch/benchmark/blob/sumall/benchmarks/sum_bench.py) to compare the results with numpy numerically as well as on timing. I ran this script to create timings both on master and this branch.
Here is the command for 1 core
`OMP_NUM_THREAD=1 taskset -c 0 python sum_bench.py --enable_numpy 200`
Here is the command for all cores
`python sum_bench.py --enable_numpy 200`
Here are the results of each:
[Master, 1 core](https://paste.fedoraproject.org/paste/Nho9JzHpPVK9av8a6mByjQ)
[This branch, 1 core](https://paste.fedoraproject.org/paste/6xLHkYvcVJx9z~5MoHxN4w)
[Master, all cores](https://paste.fedoraproject.org/paste/5l3V1d5zGqvJcMXIUteMRw)
[This branch, all cores](https://paste.fedoraproject.org/paste/J4RuDU-0Drz0aZwtphQwEA)
To test the command is
`python sum_bench.py --test 200`
[This branch, test results](https://paste.fedoraproject.org/paste/kTEoUC~oWgXA6XWMAfNfNw)
For this test we look at the average absolute value of the differences. This does not take into account the relative magnitude of the numbers. The numbers are sampled from a standard normal distribution.
In terms of performance this diff should bring PyTorch on par with Numpy and usually exceed it by 1.5 to 2x.
* Revert "ATen ReduceOps (#5481)"
This reverts commit 310c3735b9.
* Revert "Check that new cpuinfo and tbb submodules exist (#5714)"
This reverts commit 1a23c9901d.
* Implement torch.reshape and Tensor.reshape
This implements reshape which has similar semantics to numpy.reshape. It
will return a view of the source tensor if possible. Otherwise, it
returns a copy.
* Remove in-place reshape_ that was an alias for resize_
* Update documentation
This diff adds vectorization to ATen. It uses intel intrinsics to build a general vec256 class, that represents types of 256bit width. These can then be treated like regular variables. Using those it implements torch.sum() for the contiguous case. It uses Intel TBB for multithreading, which allows workstealing and chunks the reduction operations based on a experimentally chosen value (_THRESHOLD). It uses cpuinfo to pick the right code depending on the host's capabilities.
The kernels are implemented under native/cpu. Each .cpp file is compiled with -avx, -avx2 and no additional flags. A macro is used to append AVX, AVX2 or NONE to the function name. The header then needs to define the functions three times, one for each capability. This could be improved by either changing the cmake file a bit or possibly generating source code using a Python script etc.
For the non-contiguous case this defaults to the current implementation within TH. For CUDA is entirely defaults to the implementation within THC.
There probably needs to be a bit of a debate around the design decisions here, the additional dependencies, parallelization strategy, clarity, etc. The numerical results also diverge from numpy with larger tensors, which is expected since we're summing, for example, 8 numbers and then adding the result to the running sum, instead of each number one by one. But there might be something to be said about accumulating into a double for floats or the degree of divergence, the behavior with respect to CUDA, etc.
I wrote a [small Python script]( https://github.com/cpuhrsch/benchmark/blob/sumall/benchmarks/sum_bench.py) to compare the results with numpy numerically as well as on timing. I ran this script to create timings both on master and this branch.
Here is the command for 1 core
`OMP_NUM_THREAD=1 taskset -c 0 python sum_bench.py --enable_numpy 200`
Here is the command for all cores
`python sum_bench.py --enable_numpy 200`
Here are the results of each:
[Master, 1 core](https://paste.fedoraproject.org/paste/Nho9JzHpPVK9av8a6mByjQ)
[This branch, 1 core](https://paste.fedoraproject.org/paste/6xLHkYvcVJx9z~5MoHxN4w)
[Master, all cores](https://paste.fedoraproject.org/paste/5l3V1d5zGqvJcMXIUteMRw)
[This branch, all cores](https://paste.fedoraproject.org/paste/J4RuDU-0Drz0aZwtphQwEA)
To test the command is
`python sum_bench.py --test 200`
[This branch, test results](https://paste.fedoraproject.org/paste/kTEoUC~oWgXA6XWMAfNfNw)
For this test we look at the average absolute value of the differences. This does not take into account the relative magnitude of the numbers. The numbers are sampled from a standard normal distribution.
In terms of performance this diff should bring PyTorch on par with Numpy and usually exceed it by 1.5 to 2x.
* Fix arange floating point error
* fix test
* add type cast when calculating arange size
* fix nit
* update test
* use doubles instead of floats to calculate size
* requested changes
* Add torch.empty, torch.full and new_ size Tensor factory methods.
This adds torch.full, torch.empty equivalents of np.full, np.empty.
In addition, this adds size-based Tensor factory methods new_empty, new_ones, new_full, new_zeros,
which is meant to complete the separation of the legacy "new" method into data-based and size-based
functions.
This also fixes an issue in sparse zeros_like when the dtype didn't match the argument dtype.
* Get rid of unnecessary zero in sparse tensor zeros_like.
* Fix test if only 1 cuda device.
* CPU int-types pow()
* CUDA int-type pow()
* Cleanup + fix deleted line
* Tests for integer-types pow
* Fix build
* Fix windows tests
* Make _test_int_pow static
Questions/possible future works:
How to template-ize to extend support beyond LongTensor?
How to check if autograd works (and if not, how to add explicit gradient)?
CUDA support?
Testing command:
DEBUG=1 NO_CUDA=1 MACOSX_DEPLOYMENT_TARGET=10.9 CC=clang CXX=clang++ python setup.py build && DEBUG=1 NO_CUDA=1 MACOSX_DEPLOYMENT_TARGET=10.9 CC=clang CXX=clang++ python setup.py develop && python3 test/test_torch.py
Partially fixes#2031
* Initial commit for unique op
* Working unique with test
* Make inverse indices shape conform to input
* flake8 whitespace removal
* address review comment nits
* Expose fn and add docs. Explicitly declare no gradients
* Trial generic dispatch implementation
* Add tests for generics
* flake8 whitespace
* Add basic CUDA error throwing and templateize set
* Explicit contiguous and AT_DISPATCH_ALL_TYPES return
* Remove extraneous numpy conversion
* Refactor out .data calls
* Refactored to variable return length API with wrapper fn as opposed to returning a 0-length tensor, per off-line reviewer comments
* Remove A
* Don't use hidden torch._unique() in test
* Fix documentations
