code/KwaveUpdateDomainParams.py

#!/usr/bin/env python
# coding: utf-8

import numpy as np
import h5py
import argparse
import random
import shutil
import os
import scipy.interpolate as si

from scipy.signal import gaussian
from itertools import product

from typing import Tuple

def generate_sound_speed(shape = [1152, 512],
        c_ref: float = 1500., c_min: float = 1300., c_max: float = 1800.,
        atn_min: float = 0.05, atn_max: float = 0.15,
        max_ell: int = 6, max_lines: int = 4) -> Tuple[np.ndarray, np.ndarray]:
    """[summary]
    Generate a random speed of sound and attenuation map with up to max_ell ellipses and max_lines straight reflectors

    Args:
        shape (list, optional): [description]. Defaults to [1152, 512].
        c_ref (float, optional): [description]. Defaults to 1500..
        c_min (float, optional): [description]. Defaults to 1300..
        c_max (float, optional): [description]. Defaults to 1800..
        atn_min (float, optional): [description]. Defaults to 0.05.
        atn_max (float, optional): [description]. Defaults to 0.15.
        max_ell (int, optional): [description]. Defaults to 6.
        max_lines (int, optional): [description]. Defaults to 4.

    Returns:
        [type]: [description]
    """

    while True:
        n_ell = np.random.randint(max_ell + 1)
        n_lines = np.random.randint(max_lines + 1)

        if n_ell + n_lines > 0:
            break

    ell_type = np.array([0,] * n_ell + [1,] * n_lines)[np.random.permutation(n_ell + n_lines)]
    c = np.random.rand(n_ell + n_lines + 1) * (c_max - c_min) + c_min
    atn = np.random.rand(n_ell + n_lines + 1) * (atn_max - atn_min) + atn_min

    ell_params = np.random.rand(5, n_ell); # x0, y0, a, b, \theta in columns

    ell_params[0, :] = ell_params[0, :] * shape[-1]
    ell_params[1, :] = ell_params[1, :] * shape[-2]
    ell_params[2, :] = ell_params[2, :] * shape[-1];
    ell_params[3, :] = ell_params[3, :] * shape[-2];
    ell_params[4, :] = (ell_params[4, :] - 0.5) * np.pi / 4;

    line_params = np.random.rand(2, n_lines); # Angle and offset

    line_params[0, :] = (line_params[0, :] - 0.5) * np.pi / 6;
    line_params[1, :] = line_params[1, :] * (shape[-2] - 68) + 68

    sound_speed = np.empty(shape[-2:], dtype=np.single)
    sound_speed[...] = c[-1]
    alpha_coeff = np.empty(shape[-2:], dtype=np.single)
    alpha_coeff[...] = atn[-1]

    i_ellp = 0
    i_lines = 0

    X, Y = np.meshgrid(range(shape[-1]), range(shape[-2]))

    for i,t in enumerate(ell_type):
        if t == 0:
            # elipse
            x0 = np.array([ell_params[0, i_ellp], ell_params[1, i_ellp]])
            a  = np.array([ell_params[2, i_ellp], ell_params[3, i_ellp]])

            th = ell_params[4, i_ellp]
            ath = np.abs(th)

            R = np.array([[np.cos(th), -np.sin(th)], [np.sin(th), np.cos(th)]])
            aR = np.array([[np.cos(ath), -np.sin(ath)], [np.sin(ath), np.cos(ath)]])

            toff = aR.dot(a)[1]

            if x0[1] - toff < 68:
                x0[1] = 68 + toff

            x = X - x0[0]
            y = Y - x0[1]

            u = np.cos(th) * x - np.sin(th) * y
            v = np.sin(th) * x + np.cos(th) * y

            sound_speed[(u / ell_params[2, i_ellp]) ** 2 + (v / min(ell_params[3, i_ellp], ell_params[1, i_ellp] - 68)) ** 2 < 1] = c[i]
            alpha_coeff[(u / ell_params[2, i_ellp]) ** 2 + (v / min(ell_params[3, i_ellp], ell_params[1, i_ellp] - 68)) ** 2 < 1] = atn[i]

            i_ellp += 1
        else:
            # plane
            th =  line_params[0, i_lines]
            off = line_params[1, i_lines]

            if off - shape[-1] * np.tan(np.abs(th)) < 68:
                off = shape[-1] * np.tan(np.abs(th)) + 68

            sound_speed[X * np.sin(th) + (Y - off) * np.cos(th) > 0] = c[i]
            alpha_coeff[X * np.sin(th) + (Y - off) * np.cos(th) > 0] = atn[i]

            i_lines += 1

    return sound_speed, alpha_coeff


def generate_density(shape: tuple = [1152, 1152],
    d_ref: float = 900.,
    n_spec: float = -1.,
    spec_amp: float = 0.1) -> Tuple[np.ndarray, np.ndarray, np.ndarray]:
    """
    Generate a speckle map in the density domain, including the x and y derivatives


    Args:
        shape (_type_, optional): shape of output map. Defaults to [1152, 1152]
        d_ref (float, optional): reference density (kg / m^3). Defaults to 900
        n_spec (float, optional): speckle density. Defaults to -1
        spec_amp (float, optional): speckle amplitude. Defaults to 0.1

    Returns:
        _type_: _description_
    """

    if n_spec < 0:
        n_spec = shape[-1] * shape[-2] * (2. / 64.)

    t_n_spec = int(np.round(n_spec * (10 - np.random.rand() * 8) * 0.1))

    speckle = np.zeros((shape[-2] - 48 - 68, shape[-1] - 48*2), np.single)
    speckle.reshape(-1,1)[np.random.permutation(speckle.size)[:t_n_spec]] = d_ref * (np.random.rand(t_n_spec, 1) - 1/3) * spec_amp

    rho0 = np.empty(shape[-2:], dtype=np.single)
    rho0[...] = d_ref

    rho0[68:-48, 48:-48] += speckle

    rho0_sgx = np.empty_like(rho0)
    rho0_sgx[:,:-1] = (rho0[:,:-1] + rho0[:,1:]) * 0.5
    rho0_sgx[:,-1] = rho0[:,-1]

    rho0_sgy = np.empty_like(rho0)
    rho0_sgy[:-1,:] = (rho0[:-1,:] + rho0[1:,:]) * 0.5
    rho0_sgy[-1,:] = rho0[-1,:]

    return rho0, rho0_sgx, rho0_sgy


def generate_pulse(center_freq=5e6, n_periods=4.4, sampling_freq=40e6):
        samples_per_cycle = sampling_freq / center_freq
        n_samples = np.ceil(samples_per_cycle * n_periods + 1)

        signal = np.sin(np.arange(n_samples, dtype=np.float32) / samples_per_cycle * 2 * np.pi) * gaussian(n_samples, (n_samples - 1) / 6).astype(np.single)

        return signal


if __name__ == "__main__":
    parser = argparse.ArgumentParser()
    parser.add_argument('input', help='reference file')
    parser.add_argument('--device',      '-d', type=int, default=0, help='cuda device to use')
    parser.add_argument('--frequencies', '-f', type=float, nargs='+', default=[5e6, 2.5e6],
        help='at which frequencies to run the simulation')
    parser.add_argument('--offsets',     '-o', type=int, nargs='+', default=[-32, -24, -16, -8, 0, 8, 16, 24, 32],
        help='list of (signed) offsets to use with respect to center, in elements. Angles will be calculated from offsets')
    parser.add_argument('--n_periods', '-p', type=float, default=5.0,
        help='how many periods to use for the pulse')
    parser.add_argument('--ref_offset',  '-r', type=int, default=59 * 1152 + (1152 - (128 + 127) * 4) // 2 + 32 * 2 * 4,
        help='reference offset in entries (where to start offset 0) - default: line 60, element 33, remember the kerf')
    parser.add_argument('--ref_speed',  '-s', type=float, default=1540,
        help='reference speed under which to computer angle')

    args = parser.parse_args()

    hash = f'{random.getrandbits(128):x}'

    tmp_input = f'tmp_input_{hash}.h5'
    tmp_output = f'tmp_output_{hash}.h5'
    sim_result = f'sim_result_{hash}.h5'

    # costs an additional copy but solves problem of opening file multiple times in parallel
    shutil.copyfile(args.input, tmp_input)

    try:
        with h5py.File(tmp_input, 'r+') as f_i:
            shape = np.squeeze(f_i['c0'][()]).shape
            dt = f_i['dt'][0,0,0]
            dx = f_i['dx'][0,0,0]

            c0, alpha_coeff = generate_sound_speed(shape)
            rho0, rho0_sgx, rho0_sgy = generate_density(shape)

            f_i['c0'][...] = c0
            f_i['alpha_coeff'][...] = alpha_coeff

            f_i['rho0'][...] = rho0
            f_i['rho0_sgx'][...] = rho0_sgx
            f_i['rho0_sgy'][...] = rho0_sgy

            # # Also record receive traces
            # smi = np.tile(np.array(f_i['sensor_mask_index']), (1,1,2))
            # smi[...,smi.shape[-1]//2:] += 1152 * (1152 - 60)
            # del f_i['sensor_mask_index']
            # f_i["sensor_mask_index"] = smi
            # f_i["sensor_mask_index"].attrs["data_type"] = np.string_(b"long")
            # f_i["sensor_mask_index"].attrs["domain_type"] = np.string_(b"real")

        with h5py.File(sim_result, 'w') as f_o:
            f_o.create_dataset(f'c0', data=c0, compression="gzip", compression_opts=9)
            f_o.create_dataset(f'alpha_coeff', data=alpha_coeff, compression="gzip", compression_opts=9)
            f_o.create_dataset(f'f', data=np.array(args.frequencies))
            f_o.create_dataset(f'offsets', data=np.array(args.offsets))

        for frequency, offset in product(args.frequencies, args.offsets):
            # TODO: deal with offsets / angles
            w = 127 * 4
            o = offset * 8
            dn = w * o / np.sqrt(64 * o ** 2 + 4 * w ** 2) * (dx / dt) / args.ref_speed

            pulse = generate_pulse(center_freq=frequency, n_periods=args.n_periods, sampling_freq=1.0/dt) * 1e-7

            signal = np.zeros((1, int(np.ceil(np.abs(dn))) + len(pulse), 127 * 4), dtype=np.single)
            for i in range(127 * 4):
                signal[0, int(np.round(-dn * i / 507 if o < 0 else dn - dn * i / 507)) + np.arange(len(pulse)), i] = pulse

            with h5py.File(tmp_input, 'r+') as f_i:
                del f_i["u_source_index"]
                f_i["u_source_index"] = (np.arange((64 + 63) * 4, dtype=np.uint64) + args.ref_offset + offset * 8).reshape(1, 1, -1).astype(np.uint64)
                f_i["u_source_index"].attrs["data_type"] = np.string_(b"long")
                f_i["u_source_index"].attrs["domain_type"] = np.string_(b"real")

                del f_i["uy_source_input"]
                f_i["uy_source_input"] = signal
                f_i["uy_source_input"].attrs["data_type"] = np.string_(b"float")
                f_i["uy_source_input"].attrs["domain_type"] = np.string_(b"real")

                f_i["uy_source_flag"][:] = signal.shape[1]

            os.system(f'../../scripts/kspaceFirstOrder-CUDA -r 20 -t 8 -g {args.device} -i {tmp_input} -o {tmp_output}')

            with h5py.File(tmp_output, 'r+') as f_t:
                p = np.squeeze(f_t['p'][()]).T

            p = 0.35 * (p[0::4,:] + p[3::4,:]) + 0.95 * (p[1::4,:] + p[2::4,:])
            T = (p.shape[1] - 1) * dt

            x_orig = np.linspace(0,T,p.shape[1])
            x_out = np.arange(0,T,1/40e6)

            p_out = si.interp1d(x_orig, p, kind='linear', copy=False, bounds_error=False, fill_value=0, assume_sorted=True)(x_out).astype(np.float32)

            with h5py.File(sim_result, 'r+') as f_o:
                f_o.create_dataset(f'p_f{frequency/1e6}_o{offset}', data=p_out, compression="gzip", compression_opts=9)
    except Exception as e:
        print(f"Error : {e.args[0]}")
        os.system(f'rm {sim_result}')

    try:
        os.system(f'rm {tmp_input}')
    except Exception as e:
        print(f"Error removing {tmp_input} : {e.args[0]}")

    try:
        os.system(f'rm {tmp_output}')
    except Exception as e:
        print(f"Error removing {tmp_output} : {e.args[0]}")