- The current implementation computes `size + total_threads`, which can
overflow and cause CUDA_ERROR_ILLEGAL_ADDRESS when size is close to
the maximum representable value.
- The num_blocks calculation can also overflow due to the implementation
of divup().
- This patch prevents these overflows and allows the kernel to work
correctly for the full representable range of tensor sizes.
- Also adds relevant tests.
The sum accuracy test currently uses the default test precision for
the given scalar type. However, scalars are generated via a normal
distribution, and given a large enough count and strong enough random
generator, the expected sum is zero. This causes the test to
periodically fail.
Here we estimate an upper-bound for the error as `sqrt(N) * prec` for
summing N values, with each having an approximate epsilon of `prec`.
Also fixed a few warnings generated by MSVC when compiling the
reduction test.
reducer0.reducePacket(accum1, accum0);
reducer0.reducePacket(accum2, accum0);
reducer0.reducePacket(accum3, accum0);
For the mean reducer this will increment the count as well as adding together the accumulators and result in the wrong count being divided into the sum at the end.
- The current implementation computes `size + total_threads`, which can
overflow and cause CUDA_ERROR_ILLEGAL_ADDRESS when size is close to
the maximum representable value.
- The num_blocks calculation can also overflow due to the implementation
of divup().
- This patch prevents these overflows and allows the kernel to work
correctly for the full representable range of tensor sizes.
- Also adds relevant tests.
For vectorized 1-dimensional inputs that do not take the special
blocking path (e.g. `std::complex<...>`), there was an
index-out-of-bounds error causing the broadcast size to be
computed incorrectly. Here we fix this, and make other minor
cleanup changes.
Fixes#2351.
& and | short-circuit, && and || don't. When both arguments to those
are boolean, the short-circuiting version is usually the desired one, so
clang warns on this.
Here, it is inconsequential, so switch to && and || to suppress the warning.
For moderately sized inputs, running the Tree reduction quickly
fills/overflows the GPU thread stack space, leading to memory errors.
This was happening in the `cxx11_tensor_complex_gpu` test, for example.
Disabling tree reduction on GPU fixes this.
The `Complex.h` file applies equally to HIP/CUDA, so placing under the
generic `GPU` folder.
The `TensorReductionCuda.h` has already been deprecated, now removing
for the next Eigen version.
This is to make way for a new `Tuple` class that mimics `std::tuple`,
but can be reliably used on device and with aligned Eigen types.
The existing Tuple has very few references, and is actually an
analogue of `std::pair`.
All cuda `__half` functions are device-only in CUDA 9, including
conversions. Host-side conversions were added in CUDA 10.
The existing code doesn't build prior to 10.0.
All arithmetic functions are always device-only, so there's
therefore no reason to use vectorization on the host at all.
Modified the code to disable vectorization for `__half` on host,
which required also updating the `TensorReductionGpu` implementation
which previously made assumptions about available packets.
The original `fill` implementation introduced a 5x regression on my
nvidia Quadro K1200. @rohitsan reported up to 100x regression for
HIP. This restores performance.
For custom scalars, zero is not necessarily represented by
a zeroed-out memory block (e.g. gnu MPFR). We therefore
cannot rely on `memset` if we want to fill a matrix or tensor
with zeroes. Instead, we should rely on `fill`, which for trivial
types does end up getting converted to a `memset` under-the-hood
(at least with gcc/clang).
Requires adding a `fill(begin, end, v)` to `TensorDevice`.
Replaced all potentially bad instances of memset with fill.
Fixes#2245.
The extra [TOC] tag is generating a huge floating duplicated
table-of-contents, which obscures the majority of the page
(see bottom of https://eigen.tuxfamily.org/dox/unsupported/eigen_tensors.html).
Remove it.
Also, headers do not support markup (see
[doxygen bug](https://github.com/doxygen/doxygen/issues/7467)), so
backticks like
```
```
end up generating titles that looks like
```
Constructor <tt>Tensor<double,2></tt>
```
Removing backticks for now. To generate proper formatted headers, we
must directly use html instead of markdown, i.e.
```
<h2>Constructor <code>Tensor<double,2></code></h2>
```
which is ugly.
Fixes#2254.
- Move constructors can only be defaulted as NOEXCEPT if all members
have NOEXCEPT move constructors.
- gcc 4.8 has some funny parsing bug in `a < b->c`, thinking `b-` is a template parameter.
As written, depending on multithreading/gpu, the returned index from
`argmin`/`argmax` is not currently stable. Here we modify the functors
to always keep the first occurence (i.e. if the value is equal to the
current min/max, then keep the one with the smallest index).
This is otherwise causing unpredictable results in some TF tests.
