parkipy.ewald.PerfModel

class parkipy.ewald.PerfModel(
p2p_time,
p2g_time,
fft_time,
cnv_time,
ifft_time,
g2p_time,
kernel,
N_out,
N_in,
N_grid,
d_out,
d_in,
fft_dim,
ifft_dim,
fft_shape,
cell_size,
window_P,
dtype,
execution_space,
)

Performance model for an Ewald summation run.

Stores per-stage wall-clock times and computes floating-point operation (FLOP) and memory operation (MOP) count estimates for each stage of the Ewald algorithm: P2P, P2G, FFT, CNV, IFFT, and G2P.

An instance is returned as the second element of the kernel output when options.return_walltime == True.

FLOP counts are device-aware: hardware-specific instruction costs are looked up from an internal table for known GPUs (currently the A100 and GH200). For unrecognised devices all instruction costs default to 1 and a UserWarning is raised.

Parameters:

  • p2p_time (dict) – Wall-clock time dict for the P2P (near-field) stage.

  • p2g_time (dict) – Wall-clock time dict for the P2G (particle-to-grid spreading) stage.

  • fft_time (dict) – Wall-clock time dict for the forward FFT stage.

  • cnv_time (dict) – Wall-clock time dict for the convolution (CNV) stage.

  • ifft_time (dict) – Wall-clock time dict for the inverse FFT stage.

  • g2p_time (dict) – Wall-clock time dict for the G2P (grid-to-particle interpolation) stage.

  • kernel (str) – Name of the Ewald kernel (e.g. 'stokes_sl', 'stokes_comb', 'laplace'). Used to select the correct FLOP cost model.

  • N_out (int) – Number of target particles.

  • N_in (int) – Number of source particles.

  • N_grid (int) – Number of far-field grid points.

  • d_out (int) – Output dimension (number of components per target).

  • d_in (int) – Total input dimension, counting source positions, densities, and (where applicable) normal vectors.

  • fft_dim (int) – Number of components transformed in the forward FFT.

  • ifft_dim (int) – Number of components transformed in the inverse FFT.

  • fft_shape (array_like) – Shape of the (upsampled) FFT grid.

  • cell_size (int) – Near-field cell size used by the P2P stage.

  • window_P (int) – Window function support size in grid subintervals.

  • dtype (numpy.dtype) – Floating-point dtype of the input arrays. Used to compute memory operation counts (itemsize bytes per real element, 2 * itemsize per complex element).

  • execution_space (pykokkos.ExecutionSpace) – Kokkos execution space. Used to determine whether to query the GPU device name for hardware-specific FLOP constants.

Notes

Throughput (FLOP/s) and bandwidth (bytes/s) for each stage can be read from the __repr__ output, or computed directly as perf.flop_<stage> / perf.time_<stage>['tot'].