一段程序代码，需要改成cuda格式之行，可是效果非常差，希望大神看看

[小牛牛K]

[code=c]__global__ void LU_SGS(double *temperature) {
int tx = threadIdx.x;
double Residual = NUM;
//int stableTimes = 0;
while (Residual > EPS)
{
Residual = 0.0;
for (int i = NUM - 2; i > 0; i--)
{
double preResidual = temperature[i];
temperature[i] = (temperature[i + 1] + temperature[i - 1]) / 2.0;
Residual += temperature[i] - preResidual;
}
for (int i = 1; i < NUM-1; i++)
{
double preResidual = temperature[i];
temperature[i] = (temperature[i + 1] + temperature[i - 1]) / 2.0;
Residual += temperature[i] - preResidual;
}
//stableTimes++;
}
}
/code]

[tengwl]

开多个blocks啊，你程序中只用了一个block

[stellarman]

如果num大的话，当然慢了，这个没有并行，而且double变量的话，用Geforce卡的话只有float的1/8. Tesla 是1/2. for (int i = 1; i < NUM-1; i++) { double preResidual = temperature[i]; temperature[i] = (temperature[i + 1] + temperature[i - 1]) / 2.0; Residual += temperature[i] - preResidual; }