Python CPU and GPU accelerated TDEs, over 100 million TDEs per second!

Note from Dec 2022: The code here work beautifully and I plan to continue making minor bug fixes maintaining the current functionality. But I no longer will be making any improvements to this project. I would be happy to chat or email to help anyone who wants to dive in. For the biggest potential improvement, see this issue: https://github.com/tbenthompson/cutde/issues/23

cutde: CUDA, OpenCL and C++ enabled fullspace and halfspace triangle dislocation elements (TDEs), benchmarked at 130 million TDEs per second. cutde is a translation and optimization of the original MATLAB code from Nikhoo and Walter 2015.. In addition to the basic pair-wise TDE operations for displacement and strain, cutde also has:

• all pairs matrix construction functions.
• matrix free functions for low memory usage settings.
• block-wise functions that are especially helpful in an FMM or hierarchical matrix setting.
• an adaptive cross approximation implementation for building hierarchical matrices.

See below for basic usage and installation instructions. For more realistic usage examples, please check out the TDE examples in these BIE tutorials. You'll find examples of using all the above variants.

• Python CPU and GPU accelerated TDEs, over 100 million TDEs per second!
• Usage documentation
  • Simple pair-wise TDEs
  • I want stress.
  • All pairs interactions matrix
  • Matrix-free all pairs interactions
  • Block-wise interaction matrices
  • Adaptive cross approximation (ACA)
• Installation
  • GPU installation
    • PyCUDA
    • Mac OS X
    • Ubuntu + PyOpenCL/PoCL
    • Ubuntu + PyOpenCL with system drivers
    • Windows
    • Something else
    • Why can't I use Apple CPU OpenCL?
• Development
  • Architecture

import matplotlib.pyplot as plt
import numpy as np

import cutde.fullspace as FS

xs = np.linspace(-2, 2, 200)
ys = np.linspace(-2, 2, 200)
obsx, obsy = np.meshgrid(xs, ys)
pts = np.array([obsx, obsy, 0 * obsy]).reshape((3, -1)).T.copy()

fault_pts = np.array([[-1, 0, 0], [1, 0, 0], [1, 0, -1], [-1, 0, -1]])
fault_tris = np.array([[0, 1, 2], [0, 2, 3]], dtype=np.int64)

disp_mat = FS.disp_matrix(obs_pts=pts, tris=fault_pts[fault_tris], nu=0.25)

slip = np.array([[1, 0, 0], [1, 0, 0]])
disp = disp_mat.reshape((-1, 6)).dot(slip.flatten())

disp_grid = disp.reshape((*obsx.shape, 3))

plt.figure(figsize=(5, 5), dpi=300)
cntf = plt.contourf(obsx, obsy, disp_grid[:, :, 0], levels=21)
plt.contour(
    obsx,
    obsy,
    disp_grid[:, :, 0],
    colors="k",
    linestyles="-",
    linewidths=0.5,
    levels=21,
)
plt.colorbar(cntf)
plt.title("$u_x$")
plt.tight_layout()
plt.savefig("docs/example.png", bbox_inches="tight")

Usage documentation

Simple pair-wise TDEs

Computing TDEs for observation point/source element pairs is really simple:

from cutde.fullspace import disp, strain

disp = disp(obs_pts, src_tris, slips, nu)
strain = strain(obs_pts, src_tris, slips, nu)

Replace the cutde.fullspace with cutde.halfspace to use halfspace TDEs.

The parameters:

• obs_pts is a np.array with shape (N, 3)
• src_tris is a np.array with shape (N, 3, 3) where the second dimension corresponds to each vertex and the third dimension corresponds to the cooordinates of those vertices.
• slips is a np.array with shape (N, 3) where slips[:,0] is the strike slip component, while component 1 is the dip slip and component 2 is the tensile/opening component.
• the last parameter, nu, is the Poisson ratio.

IMPORTANT: N should be the same for all these arrays. There is exactly one triangle and slip value used for each observation point.

• The output disp is a (N, 3) array with displacement components in the x, y, z directions.
• The output strain is a (N, 6) array representing a symmetric tensor. strain[:,0] is the xx component of strain, 1 is yy, 2 is zz, 3 is xy, 4 is xz, and 5 is yz.

I want stress.

Use:

stress = cutde.fullspace.strain_to_stress(strain, sm, nu)

to convert from stress to strain assuming isotropic linear elasticity. sm is the shear modulus and nu is the Poisson ratio. The strain array here is expected to have a shape (N, 6). In case you have a matrix or other shaped array, this snippet might help:

strain_to_stress(strain_matrix.reshape((-1, 6)), mu, nu).reshape(strain_matrix.shape)

All pairs interactions matrix

If, instead, you want to create a matrix representing the interaction between every observation point and every source triangle, there is a different interface:

from cutde.fullspace import disp_matrix, strain_matrix

disp_mat = disp_matrix(obs_pts, src_tris, nu)
strain_mat = strain_matrix(obs_pts, src_tris, nu)

• obs_pts is a np.array with shape (N_OBS_PTS, 3)
• src_tris is a np.array with shape (N_SRC_TRIS, 3, 3) where the second dimension corresponds to each vertex and the third dimension corresponds to the cooordinates of those vertices.
• the last parameter, nu, is the Poisson ratio.
• The output disp_mat is a (N_OBS_PTS, 3, N_SRC_TRIS, 3) array. The second dimension corresponds to the components of the observed displacement while the fourth dimension corresponds to the component of the source slip vector. The slip vector components are ordered the same way as in disp(...) and strain(...).
• The output strain_mat is a (N_OBS_PTS, 6, N_SRC_TRIS, 3) array. Like above, the dimension corresponds to the components of the observation strain with the ordering identical to strain(...).

Note that to use the strain_to_stress function with a matrix output like this, you'll need to re-order the axes of the strain array so that the 6-d strain axis is the last axis. You can do this with np.transpose(...).

Matrix-free all pairs interactions

A common use of the matrices produced above by disp_matrix(...) would be to perform matrix-vector products with a input vector with (N_SRC_TRIS * 3) entries and an output vector with (N_OBS_PTS * 6) entries. But, building the entire matrix can require a very large amount of memory. In some situations, it's useful to compute matrix-vector products without ever computing the matrix itself, a so-called "matrix-free" operation. In order to do this, the matrix entries are recomputed whenever they are needed. As a result, performing a matrix-vector product is much slower -- on my machine, about 20x slower. But, the trade-off may be worthwhile if you are memory-constrained.

from cutde.fullspace import disp_free, strain_free
disp = disp_free(obs_pts, src_tris, slips, nu)
strain = strain_free(obs_pts, src_tris, slips, nu)

The parameters are the same as for disp_matrix(...) with the addition of slips. The slips array is a (N_SRC_TRIS, 3) array containing the source slip vectors.

Block-wise interaction matrices

In some settings, it is useful to compute many sub-blocks of a matrix without computing the full matrix. For example, this is useful for the nearfield component of a hierarchical matrix or fast multipole approximation.

from cutde.fullspace import disp_block, strain_block
disp_matrices, block_idxs = disp_block(
    obs_pts, src_tris, obs_start, obs_end, src_start, src_end, nu
)
strain_matrices, strain_block_idxs = strain_block(
    obs_pts, src_tris, obs_start, obs_end, src_start, src_end, nu
)

• obs_pts, src_tris and nu are the same as for disp_matrix.
• obs_starts and obs_end are arrays with N_BLOCKS elements representing the first and last observation point indices in each block.
• src_starts and src_end are arrays with N_BLOCKS elements representing the first and last source triangle indices in each block.

The output disp_matrices and strain_matrices will be a densely packed representation with each block's boundaries demarcated by block_idxs. As an example of extracting a single block:

disp_matrices, block_idxs = disp_block(obs_pts, src_tris, [0, 5], [5, 10], [0, 2], [2, 4], nu)
block1 = disp_matrices[block_idxs[0]:block_idxs[1]].reshape((5, 3, 2, 3))

Adaptive cross approximation (ACA)

Sometimes the matrix blocks we want to compute represent far-field interactions where the observation points are all sufficiently far away and separated as a group from the source triangles. In this situation, the matrix blocks are approximately low rank. An approximate matrix will require much less storage space and allow for more efficient matrix-vector products. Adaptive cross approximation is an algorithm for computing such a low rank representation. See Grasedyck 2005 for an accessible and general introduction to ACA. Or, see the ACA section here for an introduction that builds up to using the cutde.fullspace.disp_aca(...) implementation.

disp_appxs = cutde.fullspace.disp_aca(
    obs_pts, tris, obs_start, obs_end, src_start, src_end, nu, tol, max_iter
)

Like all the other functions, this function is provided by both cutde.fullspace and cutde.halfspace.

The parameters are the same as disp_block(...) with the addition of tol and max_iter. The tolerance, tol, is specified as an array of length N_BLOCKS in terms of the Frobenius norm of the error matrix between the true matrix and the approximation. The algorithm is not guaranteed to reach the specified tolerance but should come very close. The maximum number of iterations (equal to the maximum rank of the approximation) is also specified as an array of length N_BLOCKS.

The output disp_appxs will be a list of (U, V) pairs representing the left and right vectors of the low rank approximation. To approximate a matrix vector product:

U, V = disp_appxs[0]
y = U.dot(V.dot(x))

Installation

Installing from conda-forge is preferable because there should be fewer issues involving compilers. To install cutde from conda-forge:

conda install -c conda-forge cutde

or to install from pypi with pip:

pip install cutde

Installing via pip will build the C++ extensions from source which will require access to a non-ancient version of either GCC or Clang.

That should be sufficient to use the C++/CPU backend. If you want to use the GPU backend via PyCUDA or PyOpenCL, follow along below.

GPU installation

Install either PyCUDA or PyOpenCL following the directions below.

PyCUDA

If you have an NVIDIA GPU, install PyCUDA with:

conda config --prepend channels conda-forge
conda install -c conda-forge pycuda

Mac OS X

Install PyOpenCL and the PoCL OpenCL driver with:

conda config --prepend channels conda-forge
conda install pocl pyopencl

Ubuntu + PyOpenCL/PoCL

Just like on a Mac:

conda config --prepend channels conda-forge
conda install pocl pyopencl

Ubuntu + PyOpenCL with system drivers

conda install pyopencl ocl-icd ocl-icd-system

You will need to install the system OpenCL drivers yourself depending on the hardware you have. See the "Something else" section below.

Windows

See the PyCUDA instructions if you have an NVIDIA GPU.

I'm not aware of anyone testing cutde on OpenCL on Windows yet. It should not be difficult to install. I would expect that you install pyopencl via conda and then install the OpenCL libraries and drivers that are provided by your hardware vendor. See the "Something else" section below.

Something else

I'd suggest starting by trying the instructions for the system most similar to yours above. If that doesn't work, never fear! OpenCL should be installable on almost all recent hardware and typical operating systems. These directions can be helpful.. I am happy to try to help if you have OpenCL installation issues, but I can't promise to be useful.

Why can't I use Apple CPU OpenCL?

You might have gotten the message: cutde does not support the Apple CPU OpenCL implementation and no other platform or device was found. Please consult the cutde README.

The Apple OpenCL implementation for Intel CPUs has very poor support for the OpenCL standard and causes lots of difficult-to-resolve errors. Instead, please use the PoCL implementation. You can install it with conda install -c conda-forge pocl.

Development

For developing cutde, clone the repo and set up your conda environment based on the environment.yml with:

conda env create

Then, pip install -e ..

Next, for developing on a GPU, please install either pycuda or pyopencl as instructed in the Installation section above.

Then, you should re-generate the baseline test data derived from the MATLAB code from Mehdi Nikhoo. To do this, first install octave. On Ubuntu, this is just:

sudo apt-get install octave

And run

./tests/setup_test_env

which will run the tests/matlab/gen_test_data.m script.

Finally, to check that cutde is working properly, run pytest!

Architecture

A summary of the modules.

• halfspace.py and fullspace.py - the main entrypoints. These are very thin wrapper layers that provide the user-facing API.
• coordinators.py - the driver functions that call the CUDA kernels. I would suggest starting here!
• geometry.py - geometry helper functions
• common.cu - a semi-direct translation of the main computation kernels in the MATLAB. These are called by the other CUDA kernels below.
• pairs.cu - the CUDA kernels for the pair-wise TDE calculations.
• matrix.cu - the CUDA kernels for the all pairs TDE calculation that constructs a matrix.
• blocks.cu - the CUDA kernels for the block-wise matrix calculation.
• free.cu - the CUDA kernels for the matrix-free matrix-vector product calculation. This can be used if the matrix you'd like to construct is too large to hold in memory.
• aca.cu - the CUDA kernels for the adaptive cross approximation implementation.
• backend.py - a layer that abstracts between the CUDA, OpenCL and C++.
• gpu_backend.py - some helper functions for the CUDA and OpenCL backends
• cuda.py - the PyCUDA backend.
• opencl.py - the PyOpenCL backend.
• cpp.py and cutde.cpp_backend - combined, these two files provide a portability layer so that the CUDA code can actually be compiled as C++ and run, albeit a bit slowly, on the CPU.

The tests/tde_profile.py script is useful for assessing performance.

Some tests are marked as slow. To run these, run pytest --runslow.

If you several backends available and installed cutde will prefer CUDA, then OpenCL and finally fall back to the C++ backend. If you would prefer to specify which backend to use, you can set the environment variable CUTDE_USE_BACKEND to either cuda, opencl or cpp.

The README.md is auto-generated from a template in docs/. To run this process, run docs/build_readme.