This is the release note of v9.0.0a2. See here for the complete list of solved issues and merged PRs.
Update (2020-12-02): Unfortunately, the Windows build of this release is not working. We have taken down Windows wheels from PyPI, but if you need one for reference purposes you can still download them from the Assets section below. We are working hard to resolve this issue towards the next v9.0.0b1 release.
Highlights
cupy.vectorize & Initial CUDA JIT support
With this release, we are including a very early version of a Python to CUDA transpiler that will allow users to write their own CUDA kernels in Python, similarly to what Numba does. However, while Numba works on the bytecode and directly outputs the PTX code using LLVM, our approach uses the Python AST to directly translate the source code to C-CUDA and compile it using the NVIDIA toolchain, aiming to achieve a higher performance in the long run.
import cupy
def f(x, y):
# This code will be compiled to a CUDA kernel by our JIT
return x * x + y * y
x = cupy.linspace(0, 10, 6)
y = cupy.linspace(0, 20, 6)
func = cupy.vectorize(f)
out = func(x, y)
# out is [ 0. 20. 80. 180. 320. 500.]The initial version provides the limited support of primitive operators but we will be going forward in the upcoming releases. Check out #4290 if you are interested.
Jitify for raw kernels and modules
Thanks to @leofang now it is possible to use headers and libraries that were not possible before in RawKernel or RawModule due to the NVRTC reliance. With the new jitify=True option, Jitify is applied to your code so that you can use libraries such as the cuRAND device API, or CUB device routines in your raw kernels.
cupyx.lapack now as a public interface to cuBLAS
Until now, cuBLAS & cuSOLVER bindings were not publicly exposed in the API. However, with the introduction of cupyx.lapack by @anaruse, now it is possible to use LAPACK compatible routines backed by cuBLAS & cuSOLVER with a much simpler interface.
Deprecations in upcoming releases
We are going to drop support for Python 3.5 and obsolete libraries such as CUDA 9.0 and NumPy 1.16. Leave a comment in #4300 if you have any concerns in your use-case.
Changes
New Features
- Add
cupy.cusolver.gesvthat usescusolverDn<t1><t2>gesv(#3917) - Add
cupy.cusolver.gelsthat usescusolverDn<t1><t2>gels(#4073) - Add
cupy.vectorize(#4135) - Support cuFFT callbacks (#4141)
- Add
spline_filter1dandspline_filtertocupyx.scipy.ndimage.interpolation(#4145) - Add
cupyx.scipy.sparse.linalg.svds(#4155) - Improve coverage of cuBLAS L1 functions (#4205)
- Change the wrapper arguments for cuBLAS L2 functions (#4221)
- Add
cupyx.scipy.sparse.linalg.cg(#4222) - Support Jitify (#4228)
- Add
cupyx.lapack(#4235)
Enhancements
- Remove
cupy.testing.NumpyError(#4225) - Fix an issue in
cupyx.scipy.sparse.linalg.eigshwith CUDA 9.2 (#4231) - Add parallel build feature (#4240)
- Support compile-time constants in CuPy JIT (#4241)
- Detect the CC of the device when building (#4242)
- Reduce the file size of cuFFT callback modules (#4267)
- Bump cuDNN to v8.0.5 (#4303)
- Detect Jitify version (#4306)
Performance Improvements
- Improve
cupy.random.randint(#4160) - Improve performance of
cupyx.scipy.sparse.linalg.eigsh(#4214) - Improve
convolve/correlate(#4248) - Improve Jitify performance (#4277)
Bug Fixes
- Respect user-supplied output array in all binary morphology functions (#4157)
- Refactor
AssertFunctionIsCalled(#4233) - Fix possible redefinition of "-ccbin" in
cupy.fft._callback(#4276) - Fix undefined symbols in
cupy.fft._callback(#4283) - Fix issues when coo sparse matrix is created from dense matrix (#4295)
- Fix
cupy.random.bytesnot working (#4318)
Code Fixes
- Use
.imag = 0at hipFFT workaround (#4234) - Use
assertstatement instead ofself.assert*methods (#4292)
Documentation
- Add a note in
free_all_blocksreference (#4196) - Add cupy-cuda111 to README (#4210)
- Add missing functions to the API reference (#4215)
- Add description of env vars for parallel build and auto cc detection (#4250)
- Add
spline_filterfunctions to ndimage docs (#4265) cupy-cuda111package now on PyPI (#4333)
Installation
- Reset
extra_compile_argsfor each module (#4336) - Tentatively hide
CUPY_NUM_BUILD_JOBSoption (#4339)
HIP/ROCm
Tests
- Remove unused features in
testing.parameterized(#4178) - Add pytest backend implementation of
testing.parameterize(#4192) - Use Python 3.6 in Travis CI (#4206)
- Add repr of parameterized test and stop adding error message (#4211)
- Fix
numpy_cupy_equalfor case that both numpy cupy raise errors (#4244) - Fix tests of
__bytes__(#4252) - Use GitHub Actions (#4261)
- Skip some failing tests for fp16 + CUDA 9.0 (#4299)
- Assert the results are bool (#4310)
- Add import test for ROCm (#4326)
Others
- Bump version to v9.0.0a2 (#4331)
- Add
equal_nantoggle forNaNvalues inarray_equal(#4203) - Fix output dtype of
linalg.norm(#4227) - Warn non-tuple sequence for multidimensional indexing (#4245)
DeprecationWaringon truth value on empty array (#4308)
Contributors
The CuPy Team would like to thank all those who contributed to this release!