TensorComprehensions

TensorComprehensions copied to clipboard

https://ci.pytorch.org/jenkins/job/tensorcomp-builds/job/linux-xenial-gcc5-cuda9-cudnn7-py3-conda-build-test/682/console reports an illegal memory access, but an attempt to reproduce it in https://github.com/facebookresearch/TensorComprehensions/pull/458 doesn't seem to work.

skimo-openhub

can we build TC without ATen

1

thanks

hutuxianren

Nested Function Calls in TC

8

There is some interest in being able to do nested function calls i.e. define one TC and call it from inside another one. ``` lang = """ def max() {...

prigoyal

Codegen::makeLoopIterators generates names for loop iterators, but they may clash with size parameter or tensor names.

skimo-openhub

Fix the SpatialBatchNorm C++ test to work with latest ATen

1

PR https://github.com/facebookresearch/TensorComprehensions/pull/183 disables SpatialBatchNorm test. It has various issues: 1. dimension mismatching for weight 2. output doesn't match in the specific reference check this is related to changes in ATen...

prigoyal

In the next release, we want to improve the pybindings besides other things: - [ ] PyTorch now provides the cpp extensions that we can use to extract aten tensor...

prigoyal

Bitwise Operation Support

7

Tensor Comprehensions seem like a great way to implement binary approximation of layers as in [BinaryNet](https://arxiv.org/abs/1602.02830). However, tc does not currently support bitwise operations such as xnor (^), or (|),...

jwfromm

(CPU support missing) Check failed: v.is_zero()

5

Here's a bit of C++ code which I'm running, which concatenates two arrays together: ``` #include #include #include #include #include #include #include #include "tc/aten/aten_compiler.h" #include "tc/autotuner/genetic_autotuner_aten.h" #include "tc/core/cuda/cuda_tc_executor.h" #include "tc/core/flags.h"...

shankstm

Atm we only pass TC, input tensors and options.

nicolasvasilache

The current [implementation](https://github.com/facebookresearch/TensorComprehensions/blob/master/tc/autotuner/parameters.cc#L333) is not portable, let's use device queried information to filter. Also make it backends-specific and split ```parameters.cc```, not in #307 though. Related to #354

nicolasvasilache