[英]CUDA: embedded for loop kernel
我有一些代码,我想进入一个cuda内核。 看吧:
for (r = Y; r < Y + H; r+=2)
{
ch1RowSum = ch2RowSum = ch3RowSum = 0;
for (c = X; c < X + W; c+=2)
{
chan1Value = //some calc'd value
chan3Value = //some calc'd value
chan2Value = //some calc'd value
ch2RowSum += chan2Value;
ch3RowSum += chan3Value;
ch1RowSum += chan1Value;
}
ch1Mean += ch1RowSum / W;
ch2Mean += ch2RowSum / W;
ch3Mean += ch3RowSum / W;
}
这应该分成两个内核,一个用于计算RowSums,另一个用于计算均值,我应该如何处理我的循环索引从0开始并在N结束的事实?
假设你有一个计算三个值的内核。 配置中的每个线程将计算每个(r,c)对的三个值。
__global__ value_kernel(Y, H, X, W)
{
r = blockIdx.x + Y;
c = threadIdx.x + W;
chan1value = ...
chan2value = ...
chan3value = ...
}
我不相信你可以在上面的内核中计算总和(至少完全平行)。 你将不能像上面那样使用+ =。 你可以把它全部放在一个内核中,如果你在每个块(行)中只有一个线程做总和和意思,就像这样......
__global__ both_kernel(Y, H, X, W)
{
r = blockIdx.x + Y;
c = threadIdx.x + W;
chan1value = ...
chan2value = ...
chan3value = ...
if(threadIdx.x == 0)
{
ch1RowSum = 0;
ch2RowSum = 0;
ch3RowSum = 0;
for(i=0; i<blockDim.x; i++)
{
ch1RowSum += chan1value;
ch2RowSum += chan2value;
ch3RowSum += chan3value;
}
ch1Mean = ch1RowSum / blockDim.x;
ch2Mean = ch2RowSum / blockDim.x;
ch3Mean = ch3RowSum / blockDim.x;
}
}
但是最好使用第一个值内核,然后使用第二个内核来获得总和,这意味着...可以进一步并行化下面的内核,如果它是独立的,你可以在准备好时专注于它。
__global__ sum_kernel(Y,W)
{
r = blockIdx.x + Y;
ch1RowSum = 0;
ch2RowSum = 0;
ch3RowSum = 0;
for(i=0; i<W; i++)
{
ch1RowSum += chan1value;
ch2RowSum += chan2value;
ch3RowSum += chan3value;
}
ch1Mean = ch1RowSum / W;
ch2Mean = ch2RowSum / W;
ch3Mean = ch3RowSum / W;
}
声明:本站的技术帖子网页,遵循CC BY-SA 4.0协议,如果您需要转载,请注明本站网址或者原文地址。任何问题请咨询:yoyou2525@163.com.