fastconv

A simple library for 1D and 2D convolutions in Modern Fortran, compatible with fpm. Focused on speed in image processing. Optimized for convolutions of 1D, 2D and 3D tensors with small kernels. Currently only implemented for real32.

Installation

The simplest way to use fastconv is to add it as dependency in fpm.toml:

[dependencies]
fastconv = { git = "https://github.com/gronki/fastconv.git" }

Examples

Convolution 1D

program fctest

    use conv1d_m, only: conv1d_t
    use iso_fortran_env, only: fp => real_k

    type(conv1d_t) :: conv1d
    real(fp) :: inp(4), out1(2)
    real(fp), allocatable :: out2(:)

    inp(:) = [0, -1, 2, 0]
    call conv1d % set_kernel ([-1.0_fp, 2.0_fp, -1.0_fp])
    conv1d % preserve_shape = .false.

    call conv1d % conv(inp, out1)
    out2 = conv1d % apply(inp)

    print *, "out1 = ", out1
    print *, "shape(out1) = ", shape(out1)
    print *, "out1 == out2 = ", out1 == out2

    conv1d % preserve_shape = .true.
    out2 = conv1d % apply(inp)
    print *, "out2 = ", out2
    print *, "shape(out2) = ", shape(out2)

end program

Output:

 out1 =   -4.00000000       5.00000000    
 shape(out1) =            2
 out1 == out2 =  T T
 out2 =    5.54353672E-42  -4.00000000       5.00000000       0.00000000    
 shape(out2) =            4

Benchmarks

Benchmarks are available that allow to scan through setting space and choose the most optimal convolution parameters. You might also be interested how Fortran performance compares to C++ (see C++ translation of parts of this library).

To run benchmark, you can use the wrapper script (requires fpm):

./run.sh gcc
./run.sh oneapi

To print more detailed statistics:

export EXTRA_OUTPUT=1

or by building and executing Docker image:

docker build -t fastconv --build-arg FFLAGS="-O3 -funsafe-math-optimizations -funroll-loops -ffree-line-length-none -march=native" https://github.com/gronki/fastconv.git
docker run -it fastconv

You should get the output like:

 conv1d_ref_t() --> 
time =    12.62  verif_x =       3142758373.  verif_y =       3117590248.
 conv1d_pad_t(pad_modulo=4) --> 
time =     9.13  verif_x =       3142758373.  verif_y =       3117590248.
 conv1d_pad_t(pad_modulo=8) --> 
time =     9.47  verif_x =       3142758373.  verif_y =       3117590248.
...