我建议如下所示操作上述差分方程,然后使用 CUDA Thrust 原语。
微分方程操作 - 微分方程的显式形式
通过简单的代数,您可以找到以下内容:
y[1] = beta * y[0] + alpha * x[1]
y[2] = beta^2 * y[0] + alpha * beta * x[1] + alpha * x[2]
y[3] = beta^3 * y[0] + alpha * beta^2 * x[1] + alpha * beta * x[2] + alpha * x[3]
相应地,显式形式如下:
y[n] / beta^n = y[0] + alpha * x[1] / beta + alpha * x[2] / beta^2 + ...
CUDA 推力实现
您可以通过以下步骤实现上述显式形式:
- 将输入序列
d_input初始化为alpha,d_input[0] = 1.除外;
- 定义一个向量
d_1_over_beta_to_the_n等于1, 1/beta, 1/beta^2, 1/beta^3, ...;
- 将
d_input 乘以d_1_over_beta_to_the_n;
- 执行
inclusive_scan获取y[n] / beta^n的序列;
- 将上述序列除以
1, 1/beta, 1/beta^2, 1/beta^3, ...。
编辑
上述方法可以推荐用于线性时变 (LTV) 系统。对于线性时不变 (LTI) 系统,可以推荐 Paul 提到的 FFT 方法。我在对FIR filter in CUDA 的回答中使用 CUDA Thrust 和 cuFFT 提供了该方法的示例。
完整代码
#include <thrust/sequence.h>
#include <thrust/device_vector.h>
#include <thrust/host_vector.h>
int main(void)
{
int N = 20;
// --- Filter parameters
double alpha = 2.7;
double beta = -0.3;
// --- Defining and initializing the input vector on the device
thrust::device_vector<double> d_input(N,alpha * 1.);
d_input[0] = d_input[0]/alpha;
// --- Defining the output vector on the device
thrust::device_vector<double> d_output(d_input);
// --- Defining the {1/beta^n} sequence
thrust::device_vector<double> d_1_over_beta(N,1./beta);
thrust::device_vector<double> d_1_over_beta_to_the_n(N,1./beta);
thrust::device_vector<double> d_n(N);
thrust::sequence(d_n.begin(), d_n.end());
thrust::inclusive_scan(d_1_over_beta.begin(), d_1_over_beta.end(), d_1_over_beta_to_the_n.begin(), thrust::multiplies<double>());
thrust::transform(d_1_over_beta_to_the_n.begin(), d_1_over_beta_to_the_n.end(), d_input.begin(), d_input.begin(), thrust::multiplies<double>());
thrust::inclusive_scan(d_input.begin(), d_input.end(), d_output.begin(), thrust::plus<double>());
thrust::transform(d_output.begin(), d_output.end(), d_1_over_beta_to_the_n.begin(), d_output.begin(), thrust::divides<double>());
for (int i=0; i<N; i++) {
double val = d_output[i];
printf("Device vector element number %i equal to %f\n",i,val);
}
// --- Defining and initializing the input vector on the host
thrust::host_vector<double> h_input(N,1.);
// --- Defining the output vector on the host
thrust::host_vector<double> h_output(h_input);
h_output[0] = h_input[0];
for(int i=1; i<N; i++)
{
h_output[i] = h_input[i] * alpha + beta * h_output[i-1];
}
for (int i=0; i<N; i++) {
double val = h_output[i];
printf("Host vector element number %i equal to %f\n",i,val);
}
for (int i=0; i<N; i++) {
double val = h_output[i] - d_output[i];
printf("Difference between host and device vector element number %i equal to %f\n",i,val);
}
getchar();
}