ffbidx indexer from PSI

[fleon-psi]

Fast feedback indexer using Nvidia GPUs for indexing.

Short operating instruction:

1. Compile FFBIDX:

export BASE_DIR=
git clone --recurse-submodules https://github.com/paulscherrerinstitute/fast-feedback-indexer/
cd fast-feedback-indexer
mkdir build
cd build
cmake .. -DCMAKE_CUDA_ARCHITECTURES="70;75;80" -DCMAKE_INSTALL_PREFIX=${BASE_DIR}/ffbidx
make
sudo make install

This will place FFBIDX files in $BASE_DIR/ffbidx (if it is not included, /usr/local is the target directory).
The CUDA architecture settings are optimized for V100/A100/T4 GPU devices.
You might want to change CUDA architectures of you have older/newer devices, see Nvidia documentation for details.

2. Compile CrystFEL with the following options:

export BASE_DIR= # the same as above
export PKG_CONFIG_PATH=$PKG_CONFIG_PATH:${BASE_DIR}/ffbidx/lib/pkgconfig
export LD_LIBRARY_PATH=$LD_LIBRARY_PATH:${BASE_DIR}/ffbidx/lib
git clone https://github.com/fleon-psi/crystfel
cd crystfel
git checkout fast_indexer.0.11.0
meson -Dprefix=$BASE_DIR/crystfel/ build
cd build
ninja
ninja install

Requires forked version of CrystFEL at the moment.
BASE_DIR has to be the same as used for FFBIDX.

Then you can add "--indexing=ffbidx" to indexamajig options. FFBIDX lib directory has to be in the LD_LIBRARY_PATH for the execution.

To print possible configuration options, you can run:
indexamajig --help-ffbidx

[pmiddend]

I tried compiling CrystFEL with this out of curiosity and got this error:

include/ffbidx/c_api.h:86:23: error: expected ‘:’, ‘,’, ‘;’, ‘}’ or ‘__attribute__’ before ‘=’ token

which occurs in this piece of code:

struct config_runtime {             // pinned memory
    float_type length_threshold;    // threshold for determining equal vector length (|va| - threshold < |vb| < |va| + threshold)
    float_type triml;               // lower trim value for distance to nearest integer objective value - 0 < triml < trimh
    float_type trimh;               // higher trim value for distance to nearest integer objective value - triml < trimh < 0.5
    float_type delta;               // log2 curve position: score = log2(trim(dist(x)) + delta)
    float_type dist1;               // maximum distance to int for single coordinate
    float_type dist3;               // maximum distance to int for tripple coordinates
    // ============= HERE
    unsigned min_spots=6;           // vector refinement: minimum number of spots to fit against (0: no refinement)
    unsigned num_halfsphere_points; // number of sample points on half sphere for finding vector candidates
    unsigned num_angle_points;      // number of sample points in rotation space for finding cell candidates (0: auto)
};

I've never seen this default value syntax before and gcc (11) doesn't seem to like it. What could be my mistake here?

[fleon-psi]

I tried compiling CrystFEL with this out of curiosity and got this error:
...
I've never seen this default value syntax before and gcc (11) doesn't seem to like it. What could be my mistake here?

Thank you @pmiddend for reporting the issue! Yes, one line from C++ was accidently copied to C code (default initialization of struct members is OK in C++, but not in C) in the fast feedback indexer repository. This should be now fixed.

[pmiddend]

Thanks for the quick answer @fleon-psi! I was not only able to compile it, but also run it on our Maxwell cluster, clocking in at ~50fps. Though I'm not sure if it's really using the GPU. Is there a CPU fallback?

[fleon-psi]

Good to hear compilation works - thanks @pmiddend .
There is no CPU fallback - if the code executed without errors, it run on a GPU.