【发布时间】:2020-02-28 09:47:51
【问题描述】:
我正在尝试为 arrayfire 版本 3.6.4 调整匹配过滤器算法(给定 here)。 这是我最终得到的结果:
#include <arrayfire.h>
using namespace af;
struct SAR_data { //! SAR data structure format
double deltaF; //! step size of frequency data (Hz)
double maxWr; //! maximum scene size in range direction (m)
double maxWx; //! maximum scene size in cross-range direction (m)
array minF; //! (Np x 1): Vector containing the start frequency of
//! each pulse (Hz)
array x_mat; //! (Sx x Sy): the x-positions of each pixel (m)
array y_mat; //! (Sx x Sy): the y-positions of each pixel (m)
array z_mat; //! (Sx x Sy): the z-positions of each pixel (m)
array AntX; //! (Np x 1): the x-position of the sensor at each pulse (m)
array AntY; //! (Np x 1): the y-position of the sensor at each pulse (m)
array AntZ; //! (Np x 1): the z-position of the sensor at each pulse (m)
array R0; //! (Np x 1): the range to scene center (m)
array phdata; //! (K x Np): phase history data (frequency domain),
//! fast time in rows, slow time in columns
array im_final; //! (Sx x Sy): the complex image value at each pixel
};
// matched filter realization using ArrayFire
void matched_filter_ArrayFire(SAR_data& data) {
double c = 299792458.0; // speed of light
// Determine the size of the phase history data
int K = data.phdata.dims(0); // # of frequency bins per pulse
int Np = data.phdata.dims(1); // # of pulses
// Initialize the image with all zero values (complex)
data.im_final = constant<cdouble>(0, data.x_mat.dims(), c64);
array im_slices = constant<cdouble>(0, K, data.x_mat.dims(0),
data.x_mat.dims(1), c64);
array fspan = array(seq(0.0, K-1)) * data.deltaF;
cdouble unit = {0, 1};
for (int ii = 0; ii < Np; ii++) {
// compute differential range for each image pixel (m)
array dx = data.AntX(ii) - data.x_mat;
array dy = data.AntY(ii) - data.y_mat;
array dz = data.AntZ(ii) - data.z_mat;
array dR = sqrt(dx*dx + dy*dy + dz*dz) - data.R0(ii);
// calculate the frequency of each sample in the pulse (Hz)
array freq = data.minF(ii) + fspan;
array tt = data.phdata(span,ii);
// perform the Matched Filter operation
gfor (array jj, K) {
im_slices(jj,span,span) =
tt(jj)*exp(unit*((double)(4.0*Pi/c)*freq(jj)*dR));
}
tt = sum(im_slices,0);
data.im_final = data.im_final + moddims(tt, data.x_mat.dims());
}
}
代码编译得很好,但是当我尝试运行它时,它会抛出异常:
What() is:ArrayFire Exception (Invalid input size:203):
In function class af::dim4 __cdecl getOutDims(const class af::dim4 &,const class af::dim4 &,bool)
In file src\backend\common\ArrayInfo.cpp:130
Invalid dimension for argument 1
Expected: ldims == rdims
In function class af::array __cdecl af::operator -(const class af::array &,const class af::array &)
In file src\api\cpp\array.cpp:838
我发现,它抱怨以下几行:
array dx = data.AntX(ii) - data.x_mat;
array dy = data.AntY(ii) - data.y_mat;
array dz = data.AntZ(ii) - data.z_mat;
因为 data.AntX(ii)(即 1x1)和 data.x_mat(即 Sx x Sy)的维度不匹配.. 不过,在旧版本的 ArrayFire 中,它曾经可以正常工作。 显然,我需要一些方法来告诉 ArrayFire 为 data.x_mat 的整个矩阵维度“扩展”data.AntX(ii) 的值。但是怎么做呢?
【问题讨论】:
-
您可以将数组 data.AntX(ii) 平铺到另一个矩阵的维度上。
array tx = tile(data.AntX(ii), data.x_mat.dims()). -
另外,供您参考,该博客文章在 ArrayFire 2 时代已经相当老了。* 当 ArrayFire 是封闭源代码并且 API 在少数地方有点不同但应该与 ArrayFire 3 相同.*(开源)在许多地方。您也可以通过join.slack.com/t/arrayfire-org/shared_invite/… 与我们联系