执行内核函数后,打印输出始终为0。 经过一些测试,cudaMemcpy仍然是正确的。但内核似乎无法正常工作,无法从d_inputs获取正确的数据。 有人可以帮忙解释一下吗?谢谢!
#include <cuda_runtime.h>
#include <cuda.h>
#include <stdio.h>
#include <sys/time.h>
#include <math.h>
#define N 32
__global__ void Kernel_double(int niters, int* d_inputs,double* d_outputs)
{
int tid = blockIdx.x * blockDim.x + threadIdx.x;
if (tid<N) {
double val =(double) d_inputs[tid];
/*for (int iter=0; iter < niters; iter++){
val = (sqrt(pow(val,2.0)) + 5.0) - 101.0;
val = (val / 3.0) + 102.0;
val = (val + 1.07) - 103.0;
val = (val / 1.037) + 104.0;
val = (val + 3.00) - 105.0;
val = (val / 0.22) + 106.0;
}*/
val = val + 1.0;
//printf("This is %f\n",val);
d_outputs[tid] = val;
}
}
int main(int argc, char **argv)
{
int niters = 10;
printf("Iterate %d times with GPU 0 or CPU 1: %d\n", niters, cpu);
int inputs[N];
for (int i = 0; i<N; i++){
inputs[i] = i+1;
}
int d_inputs[N];
double d_outputs[N];
double outputs[N];
cudaMalloc( (void**)&d_inputs, N*sizeof(int));
cudaMalloc( (void**)&d_outputs, N*sizeof(double));
printf("test %d \n", inputs[3]);
cudaMemcpy(d_inputs, inputs, N*sizeof(int), cudaMemcpyHostToDevice);
printf("test %d \n", d_inputs[1]);
Kernel_double<<<16,2>>>(niters, d_inputs,d_outputs);
//cudaDeviceSynchronize();
cudaMemcpy(outputs, d_outputs, N*sizeof(double), cudaMemcpyDeviceToHost);
for(int j =0;j<10; j++){
printf("Outputs[%d] is: %f and %f\n",j, d_outputs[j], outputs[j]);
}
cudaFree(d_inputs);
cudaFree(d_outputs);
return EXIT_SUCCESS;
}
答案 0 :(得分:4)
cuda-memcheck
,运行代码向其他人寻求帮助。即使您不了解错误输出,对于其他试图帮助您的人也会有所帮助。如果你在这里使用了正确的cuda错误检查,那么你会被告知你的cudaMemcpy操作报告了一个无效的参数,由于下面的第3项。cpu
未在任何地方定义。我们不会为此分配或创建设备指针:
int d_inputs[N];
double d_outputs[N];
那些正在创建堆栈变量(数组),允许编译器将其视为常量指针。相反,你应该这样做:
int *d_inputs;
double *d_outputs;
编译器理解这些是可修改的指针(稍后将使用cudaMalloc
修改)。
在第3项中解决问题后,这将不合法:
printf("test %d \n", d_inputs[1]);
因为这需要在主机代码中取消引用设备指针(d_inputs
),这在CUDA中是非法的,至少在此处已经这样做了。您的代码后面的printf
语句也存在类似问题(使用d_outputs
)。
以下代码在某种程度上解决了上述问题,似乎对我来说正确运行:
$ cat t44.cu
#include <cuda_runtime.h>
#include <cuda.h>
#include <stdio.h>
#include <sys/time.h>
#include <math.h>
#define N 32
__global__ void Kernel_double(int niters, int* d_inputs,double* d_outputs)
{
int tid = blockIdx.x * blockDim.x + threadIdx.x;
if (tid<N) {
double val =(double) d_inputs[tid];
/*for (int iter=0; iter < niters; iter++){
val = (sqrt(pow(val,2.0)) + 5.0) - 101.0;
val = (val / 3.0) + 102.0;
val = (val + 1.07) - 103.0;
val = (val / 1.037) + 104.0;
val = (val + 3.00) - 105.0;
val = (val / 0.22) + 106.0;
}*/
val = val + 1.0;
//printf("This is %f\n",val);
d_outputs[tid] = val;
}
}
int main(int argc, char **argv)
{
int niters = 10;
int cpu = 0;
printf("Iterate %d times with GPU 0 or CPU 1: %d\n", niters, cpu);
int inputs[N];
for (int i = 0; i<N; i++){
inputs[i] = i+1;
}
int *d_inputs;
double *d_outputs;
double outputs[N];
cudaMalloc( (void**)&d_inputs, N*sizeof(int));
cudaMalloc( (void**)&d_outputs, N*sizeof(double));
printf("test %d \n", inputs[3]);
cudaMemcpy(d_inputs, inputs, N*sizeof(int), cudaMemcpyHostToDevice);
// printf("test %d \n", d_inputs[1]);
Kernel_double<<<16,2>>>(niters, d_inputs,d_outputs);
//cudaDeviceSynchronize();
cudaMemcpy(outputs, d_outputs, N*sizeof(double), cudaMemcpyDeviceToHost);
for(int j =0;j<10; j++){
printf("Outputs[%d] is: %f\n",j, outputs[j]);
}
cudaFree(d_inputs);
cudaFree(d_outputs);
return EXIT_SUCCESS;
}
$ nvcc -lineinfo -arch=sm_61 -o t44 t44.cu
$ cuda-memcheck ./t44
========= CUDA-MEMCHECK
Iterate 10 times with GPU 0 or CPU 1: 0
test 4
Outputs[0] is: 2.000000
Outputs[1] is: 3.000000
Outputs[2] is: 4.000000
Outputs[3] is: 5.000000
Outputs[4] is: 6.000000
Outputs[5] is: 7.000000
Outputs[6] is: 8.000000
Outputs[7] is: 9.000000
Outputs[8] is: 10.000000
Outputs[9] is: 11.000000
========= ERROR SUMMARY: 0 errors
$