Merge pull request #37 from mrphys/nufft-1d-gpu

Implement 1D NUFFT for GPU

.devcontainer/Dockerfile

@@ -1,4 +1,4 @@
-FROM ghcr.io/mrphys/tensorflow-manylinux:1.14.0
+FROM ghcr.io/mrphys/tensorflow-manylinux:1.15.0
 
 # To enable plotting.
 RUN apt-get update && \

.github/workflows/build-package.yml

@@ -18,7 +18,7 @@ jobs:
     runs-on: ubuntu-latest
 
     container:
-      image: ghcr.io/mrphys/tensorflow-manylinux:1.14.0
+      image: ghcr.io/mrphys/tensorflow-manylinux:1.15.0
 
     strategy:
       matrix:

README.md

@@ -43,6 +43,7 @@ versions of TensorFlow and TensorFlow NUFFT according to the table below.
 
 | TensorFlow NUFFT Version | TensorFlow Compatibility | Release Date |
 | ------------------------ | ------------------------ | ------------ |
+| v0.12.0                  | v2.11.x                  | Nov 27, 2022 |
 | v0.11.0                  | v2.10.x                  | Oct 12, 2022 |
 | v0.10.1                  | v2.10.x                  | Sep 26, 2022 |
 | v0.10.0                  | v2.10.x                  | Sep 7, 2022  |

RELEASE.md

@@ -1,19 +1,7 @@
-# Release 0.11.0
+# Release 0.12.0
 
-## Major Features and Improvements
-
-- Added new option `points_range` to control the range supported by the
-  algorithm. This option provides a trade-off between flexibility and
-  performance.
-- Added new option `debugging.check_points_range` to assert that the input
-  points lie within the supported range.
+This release is compiled against TensorFlow 2.11.
 
-## Bug Fixes and Other Changes
+## Major Features and Improvements
 
-- `nufft` type-1 will now raise an error when `source` and `points` have
-  incompatible samples dimensions. Previously the computation would have
-  proceeded, ignoring any additional samples in `source`.
-- Improved error reporting for invalid `grid_shape` arguments. `nufft` will
-  now raise an informative error when `grid_shape` has an invalid length or
-  when the user fails to provide it for type-1 transforms. Previously, `nufft`
-  would have behaved erratically or crashed.
+- Added support for 1D transforms to the GPU kernel.

requirements.txt

@@ -1,3 +1,3 @@
 packaging
 pydantic
-tensorflow>=2.10.0,<2.11.0
+tensorflow>=2.11.0,<2.12.0

tensorflow_nufft/__about__.py

@@ -28,7 +28,7 @@
 __summary__ = "A fast, native non-uniform FFT op for TensorFlow."
 __uri__ = "https://github.com/mrphys/tensorflow-nufft"
 
-__version__ = "0.11.0"
+__version__ = "0.12.0"
 
 __author__ = "Javier Montalt Tordera"
 __email__ = "[email protected]"

tensorflow_nufft/cc/kernels/nufft_kernels.cc

@@ -388,15 +388,15 @@ class NUFFTBaseOp : public OpKernel {
                  int64_t batch_rank,
                  int64_t* source_batch_dims,
                  int64_t* points_batch_dims,
-                 int64_t* num_modes,
+                 int64_t* grid_dims,
                  int64_t num_points,
                  FloatType* points,
                  Complex<Device, FloatType>* source,
                  Complex<Device, FloatType>* target) {
     // Number of coefficients.
     int num_coeffs = 1;
     for (int d = 0; d < rank; d++) {
-      num_coeffs *= num_modes[d];
+      num_coeffs *= grid_dims[d];
     }
 
     // Number of calls to FINUFFT execute.
@@ -465,10 +465,10 @@ class NUFFTBaseOp : public OpKernel {
     options.num_threads = worker_threads.num_threads;
 
     // Make inlined vector from pointer to number of modes. TODO: use inlined
-    // vector for all of num_modes.
+    // vector for all of grid_dims.
     int num_modes_int[3] = {1, 1, 1};
     for (int i = 0; i < rank; ++i) {
-      num_modes_int[i] = static_cast<int>(num_modes[i]);
+      num_modes_int[i] = static_cast<int>(grid_dims[i]);
     }
 
     // Make the NUFFT plan.