I am just beginning CUDA and C and I am trying to do simple addition.
When I try to print the result, I am getting the following as output:
" 3 + 4 is 1"
To compile the code, I am running the command "nvcc test.cu" which generates a.out
Thanks for your help.
Here is test.cu:
#include <stdio.h>
__global__ void add(int a, int b, int *c){
*c = a + b;
}
int main(){
int a,b,c;
int *dev_c;
a=3;
b=4;
cudaMalloc((void**)&dev_c, sizeof(int));
add<<<1,1>>>(a,b,dev_c);
cudaMemcpy(&c, dev_c, sizeof(int), cudaMemcpyDeviceToHost);
printf("%d + %d is %d\n", a, b, c);
cudaFree(dev_c);
return 0;
}
for debugging purposes, you should use printf inside the kernel. But I think your problem is that dev_c is not a raw pointer so cudaMemcpy didn't work well
cudaMemcpy(&c, dev_c, sizeof(int), cudaMemcpyDeviceToHost);
Related
Why am I getting a wrong result for adding two numbers in cuda?
I am getting 1 as an answer instead of 9. Can anybody tell me why? Does this have something to do with the pointers? I have used the following code:
#include <iostream>
#include <cuda_runtime.h>
#include <cuda.h>
using namespace std;
__global__ void add(int *a, int *b, int *c)
{
*c = *a + *b;
}
int main(void) {
int a, b, c; // host copies of a, b, c
int *d_a, *d_b, *d_c; // device copies of a, b, c
int size = sizeof(int);
// Allocate space for device copies of a, b, c
cudaMalloc((void **)&d_a, size);
cudaMalloc((void **)&d_b, size);
cudaMalloc((void **)&d_c, size);
// Setup input values
a = 2;
b = 7;
cudaMemcpy(d_a, &a, size, cudaMemcpyHostToDevice);
cudaMemcpy(d_b, &b, size, cudaMemcpyHostToDevice);
// Launch add() kernel on GPU
add<<<1,1>>>(d_a, d_b, d_c);
// Copy result back to host
cudaMemcpy(&c, d_c, size, cudaMemcpyDeviceToHost);
cout << "answer is " << c <<endl;
// Cleanup
cudaFree(d_a); cudaFree(d_b); cudaFree(d_c);
//return 0;
}
Problem was with the persistence mode. The persistence mode was disabled. Doing this worked:
$ sudo -i
# nvidia-smi -pm 1
This question already has answers here:
Trouble compiling helloworld.cu
(2 answers)
Closed 3 years ago.
I am a beginner for cuda. I wrote a test code for testing GPU device. my gpu model is k80.
There are 8 gpu cards in one node.
#include <iostream>
#include <cuda_runtime.h>
#include <device_launch_parameters.h>
#define N 10000
__global__ void add(int *a, int *b, int *c)
{
int tid = blockIdx.x;
if (tid < N)
c[tid] = a[tid] + b[tid];
}
int main()
{
int a[N], b[N], c[N];
int *dev_a, *dev_b, *dev_c;
cudaMalloc((void**)&dev_a, N * sizeof(int));
cudaMalloc((void**)&dev_b, N * sizeof(int));
cudaMalloc((void**)&dev_c, N * sizeof(int));
for (int i = 0;i < N;i++)
{
a[i] = -i;
b[i] = i*i;
}
cudaMemcpy(dev_a, a, N * sizeof(int), cudaMemcpyHostToDevice);
cudaMemcpy(dev_b, b, N * sizeof(int), cudaMemcpyHostToDevice);
add << <N, 1 >> > (dev_a, dev_b, dev_c);
cudaMemcpy(c, dev_c, N * sizeof(int), cudaMemcpyDeviceToHost);
for (int i = 0;i < N;i++)
{
printf("%d + %d = %d\\n", a[i], b[i], c[i]);
}
cudaFree(dev_a);
cudaFree(dev_b);
cudaFree(dev_c);
return 0;
}
When i compile the code:
nvcc gputest.cu -o gputest
I got errors :
gputest.cu(38): error: identifier "printf" is undefined
1 error detected in the compilation of "/tmp/tmpxft_000059a6_00000000-4_gputest.cpp4.ii".
I think printf is a function in iostream file, but i have already included the iostream. I don't know why?
Add:
#include <stdio.h>
and it will compile is OK.
printf is a function defined in the C standard library cstdio, so inclusion of stdio.h makes sense here. Different compilers may have different behavior here, but in the case of nvcc this is generally the right way to do it.
(It's not valid to assume in all cases that inclusion of iostream will satisfy the reference here.)
I am trying to call cudaMemsetAsync from kernel (so called "dynamic parallelism"). But no matter what value I use, it always set memory to 0.
Here is my test code:
#include "cuda_runtime.h"
#include "device_launch_parameters.h"
#include "cuda_device_runtime_api.h"
#include <stdio.h>
const int size = 5;
__global__ void kernel(int *c)
{
cudaMemsetAsync(c, 0x7FFFFFFF, size * 4, NULL);
}
int main()
{
cudaError_t cudaStatus;
int c[size] = { 12, 12, 12, 12, 12 };
int *dev_c = 0;
cudaStatus = cudaSetDevice(0);
cudaStatus = cudaMalloc((void**)&dev_c, size * sizeof(int));
cudaStatus = cudaMemcpy(dev_c, c, size * sizeof(int), cudaMemcpyHostToDevice);
kernel <<< 1, 1 >>>(dev_c);
cudaStatus = cudaMemcpy(c, dev_c, size * sizeof(int), cudaMemcpyDeviceToHost);
cudaFree(dev_c);
cudaStatus = cudaDeviceReset();
printf("%d\n", cudaStatus);
printf("{%d,%d,%d,%d,%d}\n", c[0], c[1], c[2], c[3], c[4]);
return 0;
}
And if I run it, I got output like this:
>nvcc -run kernel.cu -gencode=arch=compute_35,code=\"sm_35,compute_35\" -rdc=true -lcudadevrt
kernel.cu
Creating library a.lib and object a.exp
0
{0,0,0,0,0}
When I call memory set, I use value 0x7FFFFFFF. I'm expecting non-zero numbers, but it always shows zero.
Is this a bug? or I did something wrong? I'm using CUDA 8.0
I can confirm this appears not to work in CUDA 8 on the systems I tested it with.
If you want a single thread to perform the operation, you can use memset directly in device code (it, like memcpy, has been supported forever). The kernel will emit a byte sized loop inline within your kernel and the operation will be handled by each running thread.
If you want a dynamic parallelism style memset operation, then the easiest thing is to make your own. A trivial (and very, very lightly tested) implementation in the code you posted might look like this:
#include <cstring>
#include <cstdio>
const int size = 5;
__global__ void myMemset_kernel(void* p, unsigned char val, size_t sz)
{
size_t tid = threadIdx.x + blockDim.x * blockIdx.x;
unsigned char* _p = (unsigned char*)p;
for(; tid < sz; tid += blockDim.x * gridDim.x) {
_p[tid] = val;
}
}
__device__ void myMemset(void* p, unsigned int val, size_t sz, cudaStream_t s=NULL)
{
const dim3 blocksz(256,1,1);
size_t nblocks = (sz + blocksz.x -1) / blocksz.x;
unsigned charval = val & 0xff;
myMemset_kernel<<< dim3(nblocks,1,1), blocksz, 0, s >>>(p, charval, sz);
}
__global__ void kernel(int *c)
{
cudaStream_t s;
cudaStreamCreateWithFlags(&s, cudaStreamNonBlocking);
myMemset(c, 0x7FFFFFFF, size * 4, s);
cudaDeviceSynchronize();
}
int main()
{
int c[size];
int *dev_c;
memset(&c[0], 0xffffff0c, size * sizeof(int));
printf("{%08x,%08x,%08x,%08x,%08x}\n", c[0], c[1], c[2], c[3], c[4]);
cudaMalloc((void**)&dev_c, size * sizeof(int));
cudaMemcpy(dev_c, c, size * sizeof(int), cudaMemcpyHostToDevice);
kernel <<< 1, 1 >>>(dev_c);
cudaMemcpy(c, dev_c, size * sizeof(int), cudaMemcpyDeviceToHost);
cudaFree(dev_c);
printf("{%08x,%08x,%08x,%08x,%08x}\n", c[0], c[1], c[2], c[3], c[4]);
return 0;
}
which compiles and does this:
$ nvcc -rdc=true -arch=sm_52 -o memset memset.cu -lcudadevrt
$ ./memset
{0c0c0c0c,0c0c0c0c,0c0c0c0c,0c0c0c0c,0c0c0c0c}
{ffffffff,ffffffff,ffffffff,ffffffff,ffffffff}
A final point -- note the values above and read this question and answer. In your code, it is not possible to use cudaMemset to apply a value of 0x7FFFFFFF. Although the value argument is an unsigned integer, cudaMemset and its relatives work like regular memset and set byte values. Only the least significant byte of the 32 bit argument is used to set values. If your objective is to set 32 bit values, then you will need to make your own version of memset for that purpose anyway.
I'm trying to learn CUDA by myself, and I'm now into the issue of branch divergence. As far as I understand, this is the name given to the problem that arises when several threads in a block are said to take a branch (due to if or switch statements, for example), but others in that block don't have to take it.
In order to investigate a little bit further this phenomena and its consequences, I've written a little file with a couple of CUDA functions. One of them is supposed to take lots of time, since the threads are stopped for much more time (9999... iterations) than in the other one (in which they're only stopped for an assignation).
However, when I run the code, I'm getting very similar times. Furthermore, even measuring the time that running both of them takes I get a time similar to running only one. Did I code anything wrong, or is there a logical explanation for this?
Code:
#include <stdio.h>
#include <stdlib.h>
#include <cutil.h>
#define ITERATIONS 9999999999999999999
#define BLOCK_SIZE 16
unsigned int hTimer;
void checkCUDAError (const char *msg)
{
cudaError_t err = cudaGetLastError();
if (cudaSuccess != err)
{
fprintf(stderr, "Cuda error: %s: %s.\n", msg,cudaGetErrorString( err) );
getchar();
exit(EXIT_FAILURE);
}
}
__global__ void divergence(float *A, float *B){
float result = 0;
if(threadIdx.x % 2 == 0)
{
for(int i=0;i<ITERATIONS;i++){
result+=A[threadIdx.x]*A[threadIdx.x];
}
} else
for(int i=0;i<ITERATIONS;i++){
result+=A[threadIdx.x]*B[threadIdx.x];
}
}
__global__ void betterDivergence(float *A, float *B){
float result = 0;
float *aux;
//This structure should not affect performance that much
if(threadIdx.x % 2 == 0)
aux = A;
else
aux = B;
for(int i=0;i<ITERATIONS;i++){
result+=A[threadIdx.x]*aux[threadIdx.x];
}
}
// ------------------------
// MAIN function
// ------------------------
int main(int argc, char ** argv){
float* d_a;
float* d_b;
float* d_result;
float *elementsA;
float *elementsB;
elementsA = (float *)malloc(BLOCK_SIZE*sizeof(float));
elementsB = (float *)malloc(BLOCK_SIZE*sizeof(float));
//"Randomly" filling the arrays
for(int x=0;x<BLOCK_SIZE;x++){
elementsA[x] = (x%2==0)?2:1;
elementsB[x] = (x%2==0)?1:3;
}
cudaMalloc((void**) &d_a, BLOCK_SIZE*sizeof(float));
cudaMalloc((void**) &d_b, BLOCK_SIZE*sizeof(float));
cudaMalloc((void**) &d_result, sizeof(float));
cudaMemcpy(d_a, elementsA, BLOCK_SIZE*sizeof(float), cudaMemcpyHostToDevice);
cudaMemcpy(d_b, elementsB, BLOCK_SIZE*sizeof(float), cudaMemcpyHostToDevice);
CUT_SAFE_CALL(cutCreateTimer(&hTimer));
CUT_CHECK_ERROR("cudaCreateTimer\n");
CUT_SAFE_CALL( cutResetTimer(hTimer) );
CUT_CHECK_ERROR("reset timer\n");
CUT_SAFE_CALL( cutStartTimer(hTimer) );
CUT_CHECK_ERROR("start timer\n");
float timerValue;
dim3 dimBlock(BLOCK_SIZE,BLOCK_SIZE);
dim3 dimGrid(32/dimBlock.x, 32/dimBlock.y);
divergence<<<dimBlock, dimGrid>>>(d_a, d_b);
betterDivergence<<<dimBlock, dimGrid>>>(d_a, d_b);
checkCUDAError("kernel invocation");
cudaThreadSynchronize();
CUT_SAFE_CALL(cutStopTimer(hTimer));
CUT_CHECK_ERROR("stop timer\n");
timerValue = cutGetTimerValue(hTimer);
printf("kernel execution time (secs): %f s\n", timerValue);
return 0;
}
1) You have no memory writes in your __global__ code except the local variable(result). I'm not sure that cuda compiler does that, but all your code can be safely removed with no side effect(and maybe the compiler had done that).
2) All your reads from device memory in __global__ functions are from one place on each iteration. Cuda will store the value in register memory and the longest operation(memory access) will be done very fast here.
3) May be the compiler had replaced your cycles with single multiplication like `result=ITERATIONS*A[threadIdx.x]*B[threadIdx.x]
4) If all the code in your functions will be executed as you wrote it, your betterDivergence is going to be approximately 2 times faster than your another function because you have the loops in if branches in slower one and no loops in branches in faster one. But there won't be any idle time in threads among the threads that execute same loop because all threads are going to execute the body of the loop each iteration.
I suggest you to write another example where you will store the result in some device memory and then copy that memory back to host and make some more unpredictable calculations to prevent possible optimizations.
Below is shown the final, tested, right example of a code that allows to compare the performance between CUDA code with and without branch divergence:
#include <stdio.h>
#include <stdlib.h>
#include <cutil.h>
//#define ITERATIONS 9999999999999999999
#define ITERATIONS 999999
#define BLOCK_SIZE 16
#define WARP_SIZE 32
unsigned int hTimer;
void checkCUDAError (const char *msg)
{
cudaError_t err = cudaGetLastError();
if (cudaSuccess != err)
{
fprintf(stderr, "Cuda error: %s: %s.\n", msg,cudaGetErrorString( err) );
getchar();
exit(EXIT_FAILURE);
}
}
__global__ void divergence(float *A, float *B){
int a = blockIdx.x*blockDim.x + threadIdx.x;
if (a >= ITERATIONS) return;
if(threadIdx.x > 2)
{
for(int i=0;i<ITERATIONS;i++){
B[a]=A[a]+1;
}
} else
for(int i=0;i<ITERATIONS;i++){
B[a]=A[a]-1;
}
}
__global__ void noDivergence(float *A, float *B){
int a = blockIdx.x*blockDim.x + threadIdx.x;
if (a >= ITERATIONS) return;
if(threadIdx.x > WARP_SIZE)
{
for(int i=0;i<ITERATIONS;i++){
B[a]=A[a]+1;
}
} else
for(int i=0;i<ITERATIONS;i++){
B[a]=A[a]-1;
}
}
// ------------------------
// MAIN function
// ------------------------
int main(int argc, char ** argv){
float* d_a;
float* d_b;
float* d_result;
float *elementsA;
float *elementsB;
elementsA = (float *)malloc(BLOCK_SIZE*sizeof(float));
elementsB = (float *)malloc(BLOCK_SIZE*sizeof(float));
//"Randomly" filling the arrays
for(int x=0;x<BLOCK_SIZE;x++){
elementsA[x] = (x%2==0)?2:1;
}
cudaMalloc((void**) &d_a, BLOCK_SIZE*sizeof(float));
cudaMalloc((void**) &d_b, BLOCK_SIZE*sizeof(float));
cudaMalloc((void**) &d_result, sizeof(float));
cudaMemcpy(d_a, elementsA, BLOCK_SIZE*sizeof(float), cudaMemcpyHostToDevice);
cudaMemcpy(d_b, elementsB, BLOCK_SIZE*sizeof(float), cudaMemcpyHostToDevice);
CUT_SAFE_CALL(cutCreateTimer(&hTimer));
CUT_CHECK_ERROR("cudaCreateTimer\n");
CUT_SAFE_CALL( cutResetTimer(hTimer) );
CUT_CHECK_ERROR("reset timer\n");
CUT_SAFE_CALL( cutStartTimer(hTimer) );
CUT_CHECK_ERROR("start timer\n");
float timerValue;
dim3 dimBlock(BLOCK_SIZE,BLOCK_SIZE);
dim3 dimGrid(128/dimBlock.x, 128/dimBlock.y);
//divergence<<<dimGrid, dimBlock>>>(d_a, d_b);
noDivergence<<<dimGrid, dimBlock>>>(d_a, d_b);
checkCUDAError("kernel invocation");
cudaThreadSynchronize();
CUT_SAFE_CALL(cutStopTimer(hTimer));
CUT_CHECK_ERROR("stop timer\n");
timerValue = cutGetTimerValue(hTimer)/1000;
printf("kernel execution time (secs): %f s\n", timerValue);
cudaMemcpy(elementsB, d_b, BLOCK_SIZE*sizeof(float), cudaMemcpyDeviceToHost);
return 0;
}
I don't know why my kernel function doesn't work. Theoretically my program should display a = 14 but it displays a = 5.
#include <iostream>
#include <cuda.h>
#include <cuda_runtime.h>
using namespace std;
__global__ void AddIntCUDA(int* a, int* b)
{
a[0] += b[0];
}
int main()
{
int a = 5;
int b = 9;
int *d_a ;
int *d_b ;
cudaMalloc(&d_a, sizeof(int));
cudaMalloc(&d_b, sizeof(int));
cudaMemcpy(d_a, &a, sizeof(int), cudaMemcpyHostToDevice);
cudaMemcpy(d_b, &b, sizeof(int), cudaMemcpyHostToDevice);
AddIntCUDA<<<1, 1>>>(d_a, d_b);
cudaMemcpy(&a, d_a, sizeof(int), cudaMemcpyDeviceToHost);
cout<<"The answer is a = "<<a<<endl;
cudaFree(d_a);
cudaFree(d_b);
return 0;
}
Also I don't understand why if I have:
cudaMemcpy(d_b, &b, sizeof(int), cudaMemcpyHostToDevice); //d_b = 9 on device
cudaMemcpy(&a, d_b, sizeof(int), cudaMemcpyDeviceToHost); //a = 9 on host
a is still 5?
Whenever you are having trouble with a CUDA program, the first step should be to use proper cuda error checking on all cuda API calls and kernel calls. With error checking, this error (driver issue) would have been immediately obvious.
Additional suggestions can be found on the cuda tag info tab.
Maybe you need to put cudaDeviceSynchronize(); after AddIntCUDA<<<1, 1>>>(d_a, d_b);
When you executed AddIntCUDA<<<1, 1>>>(d_a, d_b); The host doesn't wait to the CUDA kernel if you don't put cudaDeviceSynchronize();