I have a GPU card GeForce GTX 295 and visual studio 2012 and cuda with version 6.5. I run a simple code like
#include "stdafx.h"
#include <stdio.h>
#include <cuda.h>
// Kernel that executes on the CUDA device
__global__ void square_array(float *a, int N)
{
int idx = blockIdx.x * blockDim.x + threadIdx.x;
if (idx<N) a[idx] = a[idx] * a[idx]; }
// main routine that executes on the host
int main(void)
{ float *a_h, *a_d; // Pointer to host & device arrays
const int N = 10; // Number of elements in arrays
size_t size = N * sizeof(float);
a_h = (float *)malloc(size); // Allocate array on host
cudaMalloc((void **) &a_d, size); // Allocate array on device // Initialize host array and copy it to CUDA device
for (int i=0; i<N; i++) a_h[i] = (float)i;
cudaMemcpy(a_d, a_h, size, cudaMemcpyHostToDevice); // Do calculation on device:
int block_size = 4;
int n_blocks = N/block_size + (N%block_size == 0 ? 0:1);
square_array <<< n_blocks, block_size >>> (a_d, N);
// Retrieve result from device and store it in host array
cudaMemcpy(a_h, a_d, sizeof(float)*N, cudaMemcpyDeviceToHost);
// Print results
for (int i=0; i<N; i++)
printf("%d %f\n", i, a_h[i]);
// Cleanup
free(a_h);
cudaFree(a_d); }
In this code ,when I use command cudaGetLastError (void) after calling the kernel, at console window an error display "Invalid device function" .How can I get rid of it?
Sample codes of cuda kit 6.5 are being run successfully with visual studio 2012.enter code here
GTX 295 has compute capability 1.3 I believe. It may be worth checking your solution compiler settings to see whether you are not compiling the solution using something like compute_20,sm_20. If so, try to change these values to e.g. compute_10,sm_10, rebuild and see whether it helps. See here for details on setting these values.
EDIT:
According to njuffa and also CUDA documentation support for cc1.0 devices was removed in CUDA 6.5 so you'll have to use compute_13,sm_13.
Related
My monte carlo pi calculation CUDA program is causing my nvidia driver to crash when I exceed around 500 trials and 256 full blocks. It seems to be happening in the monteCarlo kernel function.Any help is appreciated.
#include <stdio.h>
#include <stdlib.h>
#include <cuda.h>
#include <curand.h>
#include <curand_kernel.h>
#define NUM_THREAD 256
#define NUM_BLOCK 256
///////////////////////////////////////////////////////////////////////////////////////////
///////////////////////////////////////////////////////////////////////////////////////////
// Function to sum an array
__global__ void reduce0(float *g_odata) {
extern __shared__ int sdata[];
// each thread loads one element from global to shared mem
unsigned int tid = threadIdx.x;
unsigned int i = blockIdx.x*blockDim.x + threadIdx.x;
sdata[tid] = g_odata[i];
__syncthreads();
// do reduction in shared mem
for (unsigned int s=1; s < blockDim.x; s *= 2) { // step = s x 2
if (tid % (2*s) == 0) { // only threadIDs divisible by the step participate
sdata[tid] += sdata[tid + s];
}
__syncthreads();
}
// write result for this block to global mem
if (tid == 0) g_odata[blockIdx.x] = sdata[0];
}
///////////////////////////////////////////////////////////////////////////////////////////
///////////////////////////////////////////////////////////////////////////////////////////
__global__ void monteCarlo(float *g_odata, int trials, curandState *states){
// unsigned int tid = threadIdx.x;
unsigned int i = blockIdx.x*blockDim.x + threadIdx.x;
unsigned int incircle, k;
float x, y, z;
incircle = 0;
curand_init(1234, i, 0, &states[i]);
for(k = 0; k < trials; k++){
x = curand_uniform(&states[i]);
y = curand_uniform(&states[i]);
z =(x*x + y*y);
if (z <= 1.0f) incircle++;
}
__syncthreads();
g_odata[i] = incircle;
}
///////////////////////////////////////////////////////////////////////////////////////////
///////////////////////////////////////////////////////////////////////////////////////////
int main() {
float* solution = (float*)calloc(100, sizeof(float));
float *sumDev, *sumHost, total;
const char *error;
int trials;
curandState *devStates;
trials = 500;
total = trials*NUM_THREAD*NUM_BLOCK;
dim3 dimGrid(NUM_BLOCK,1,1); // Grid dimensions
dim3 dimBlock(NUM_THREAD,1,1); // Block dimensions
size_t size = NUM_BLOCK*NUM_THREAD*sizeof(float); //Array memory size
sumHost = (float*)calloc(NUM_BLOCK*NUM_THREAD, sizeof(float));
cudaMalloc((void **) &sumDev, size); // Allocate array on device
error = cudaGetErrorString(cudaGetLastError());
printf("%s\n", error);
cudaMalloc((void **) &devStates, (NUM_THREAD*NUM_BLOCK)*sizeof(curandState));
error = cudaGetErrorString(cudaGetLastError());
printf("%s\n", error);
// Do calculation on device by calling CUDA kernel
monteCarlo <<<dimGrid, dimBlock>>> (sumDev, trials, devStates);
error = cudaGetErrorString(cudaGetLastError());
printf("%s\n", error);
// call reduction function to sum
reduce0 <<<dimGrid, dimBlock, (NUM_THREAD*sizeof(float))>>> (sumDev);
error = cudaGetErrorString(cudaGetLastError());
printf("%s\n", error);
dim3 dimGrid1(1,1,1);
dim3 dimBlock1(256,1,1);
reduce0 <<<dimGrid1, dimBlock1, (NUM_THREAD*sizeof(float))>>> (sumDev);
error = cudaGetErrorString(cudaGetLastError());
printf("%s\n", error);
// Retrieve result from device and store it in host array
cudaMemcpy(sumHost, sumDev, sizeof(float), cudaMemcpyDeviceToHost);
error = cudaGetErrorString(cudaGetLastError());
printf("%s\n", error);
*solution = 4*(sumHost[0]/total);
printf("%.*f\n", 1000, *solution);
free (solution);
free(sumHost);
cudaFree(sumDev);
cudaFree(devStates);
//*solution = NULL;
return 0;
}
If smaller numbers of trials work correctly, and if you are running on MS Windows without the NVIDIA Tesla Compute Cluster (TCC) driver and/or the GPU you are using is attached to a display, then you are probably exceeding the operating system's "watchdog" timeout. If the kernel occupies the display device (or any GPU on Windows without TCC) for too long, the OS will kill the kernel so that the system does not become non-interactive.
The solution is to run on a non-display-attached GPU and if you are on Windows, use the TCC driver. Otherwise, you will need to reduce the number of trials in your kernel and run the kernel multiple times to compute the number of trials you need.
EDIT: According to the CUDA 4.0 curand docs(page 15, "Performance Notes"), you can improve performance by copying the state for a generator to local storage inside your kernel, then storing the state back (if you need it again) when you are finished:
curandState state = states[i];
for(k = 0; k < trials; k++){
x = curand_uniform(&state);
y = curand_uniform(&state);
z =(x*x + y*y);
if (z <= 1.0f) incircle++;
}
Next, it mentions that setup is expensive, and suggests that you move curand_init into a separate kernel. This may help keep the cost of your MC kernel down so you don't run up against the watchdog.
I recommend reading that section of the docs, there are several useful guidelines.
For those of you having a geforce GPU which does not support TCC driver there is another solution based on:
http://msdn.microsoft.com/en-us/library/windows/hardware/ff569918(v=vs.85).aspx
start regedit,
navigate to HKEY_LOCAL_MACHINE\System\CurrentControlSet\Control\GraphicsDrivers
create new DWORD key called TdrLevel, set value to 0,
restart PC.
Now your long-running kernels should not be terminated. This answer is based on:
Modifying registry to increase GPU timeout, windows 7
I just thought it might be useful to provide the solution here as well.
My monte carlo pi calculation CUDA program is causing my nvidia driver to crash when I exceed around 500 trials and 256 full blocks. It seems to be happening in the monteCarlo kernel function.Any help is appreciated.
#include <stdio.h>
#include <stdlib.h>
#include <cuda.h>
#include <curand.h>
#include <curand_kernel.h>
#define NUM_THREAD 256
#define NUM_BLOCK 256
///////////////////////////////////////////////////////////////////////////////////////////
///////////////////////////////////////////////////////////////////////////////////////////
// Function to sum an array
__global__ void reduce0(float *g_odata) {
extern __shared__ int sdata[];
// each thread loads one element from global to shared mem
unsigned int tid = threadIdx.x;
unsigned int i = blockIdx.x*blockDim.x + threadIdx.x;
sdata[tid] = g_odata[i];
__syncthreads();
// do reduction in shared mem
for (unsigned int s=1; s < blockDim.x; s *= 2) { // step = s x 2
if (tid % (2*s) == 0) { // only threadIDs divisible by the step participate
sdata[tid] += sdata[tid + s];
}
__syncthreads();
}
// write result for this block to global mem
if (tid == 0) g_odata[blockIdx.x] = sdata[0];
}
///////////////////////////////////////////////////////////////////////////////////////////
///////////////////////////////////////////////////////////////////////////////////////////
__global__ void monteCarlo(float *g_odata, int trials, curandState *states){
// unsigned int tid = threadIdx.x;
unsigned int i = blockIdx.x*blockDim.x + threadIdx.x;
unsigned int incircle, k;
float x, y, z;
incircle = 0;
curand_init(1234, i, 0, &states[i]);
for(k = 0; k < trials; k++){
x = curand_uniform(&states[i]);
y = curand_uniform(&states[i]);
z =(x*x + y*y);
if (z <= 1.0f) incircle++;
}
__syncthreads();
g_odata[i] = incircle;
}
///////////////////////////////////////////////////////////////////////////////////////////
///////////////////////////////////////////////////////////////////////////////////////////
int main() {
float* solution = (float*)calloc(100, sizeof(float));
float *sumDev, *sumHost, total;
const char *error;
int trials;
curandState *devStates;
trials = 500;
total = trials*NUM_THREAD*NUM_BLOCK;
dim3 dimGrid(NUM_BLOCK,1,1); // Grid dimensions
dim3 dimBlock(NUM_THREAD,1,1); // Block dimensions
size_t size = NUM_BLOCK*NUM_THREAD*sizeof(float); //Array memory size
sumHost = (float*)calloc(NUM_BLOCK*NUM_THREAD, sizeof(float));
cudaMalloc((void **) &sumDev, size); // Allocate array on device
error = cudaGetErrorString(cudaGetLastError());
printf("%s\n", error);
cudaMalloc((void **) &devStates, (NUM_THREAD*NUM_BLOCK)*sizeof(curandState));
error = cudaGetErrorString(cudaGetLastError());
printf("%s\n", error);
// Do calculation on device by calling CUDA kernel
monteCarlo <<<dimGrid, dimBlock>>> (sumDev, trials, devStates);
error = cudaGetErrorString(cudaGetLastError());
printf("%s\n", error);
// call reduction function to sum
reduce0 <<<dimGrid, dimBlock, (NUM_THREAD*sizeof(float))>>> (sumDev);
error = cudaGetErrorString(cudaGetLastError());
printf("%s\n", error);
dim3 dimGrid1(1,1,1);
dim3 dimBlock1(256,1,1);
reduce0 <<<dimGrid1, dimBlock1, (NUM_THREAD*sizeof(float))>>> (sumDev);
error = cudaGetErrorString(cudaGetLastError());
printf("%s\n", error);
// Retrieve result from device and store it in host array
cudaMemcpy(sumHost, sumDev, sizeof(float), cudaMemcpyDeviceToHost);
error = cudaGetErrorString(cudaGetLastError());
printf("%s\n", error);
*solution = 4*(sumHost[0]/total);
printf("%.*f\n", 1000, *solution);
free (solution);
free(sumHost);
cudaFree(sumDev);
cudaFree(devStates);
//*solution = NULL;
return 0;
}
If smaller numbers of trials work correctly, and if you are running on MS Windows without the NVIDIA Tesla Compute Cluster (TCC) driver and/or the GPU you are using is attached to a display, then you are probably exceeding the operating system's "watchdog" timeout. If the kernel occupies the display device (or any GPU on Windows without TCC) for too long, the OS will kill the kernel so that the system does not become non-interactive.
The solution is to run on a non-display-attached GPU and if you are on Windows, use the TCC driver. Otherwise, you will need to reduce the number of trials in your kernel and run the kernel multiple times to compute the number of trials you need.
EDIT: According to the CUDA 4.0 curand docs(page 15, "Performance Notes"), you can improve performance by copying the state for a generator to local storage inside your kernel, then storing the state back (if you need it again) when you are finished:
curandState state = states[i];
for(k = 0; k < trials; k++){
x = curand_uniform(&state);
y = curand_uniform(&state);
z =(x*x + y*y);
if (z <= 1.0f) incircle++;
}
Next, it mentions that setup is expensive, and suggests that you move curand_init into a separate kernel. This may help keep the cost of your MC kernel down so you don't run up against the watchdog.
I recommend reading that section of the docs, there are several useful guidelines.
For those of you having a geforce GPU which does not support TCC driver there is another solution based on:
http://msdn.microsoft.com/en-us/library/windows/hardware/ff569918(v=vs.85).aspx
start regedit,
navigate to HKEY_LOCAL_MACHINE\System\CurrentControlSet\Control\GraphicsDrivers
create new DWORD key called TdrLevel, set value to 0,
restart PC.
Now your long-running kernels should not be terminated. This answer is based on:
Modifying registry to increase GPU timeout, windows 7
I just thought it might be useful to provide the solution here as well.
I have a simple program to calculate square root, loop unrolling was done as
loop unrolling
#include <stdio.h>
#include <cuda.h>
__global__ void square(float *a, int N,int idx);
// Kernel that executes on the CUDA device
__global__ void first(float *arr, int N)
{
int idx = 2*(blockIdx.x * blockDim.x + threadIdx.x);
int n=N;
//printf("%d\n",n);
for(int q=0;q<2;q++)
{
if(N<2000)
{
arr[idx+q] = arr[idx+q] * arr[idx+q];
}
}
}
// main routine that executes on the host
int main(void)
{
clock_t start = clock(),diff;
float *a_h, *a_d; // Pointer to host & device arrays
const int N = 1000; // Number of elements in arrays
size_t size = N * sizeof(float);
a_h = (float *)malloc(size); // Allocate array on host
cudaMalloc((void **) &a_d, size); // Allocate array on device
// Initialize host array and copy it to CUDA device
for (int i=0; i<N; i++) a_h[i] = (float)i;
cudaMemcpy(a_d, a_h, size, cudaMemcpyHostToDevice);
// Do calculation on device:
int block_size = 4;
//int n_blocks = N/block_size + (N%block_size == 0 ? 0:1);
first <<< 4, 128 >>> (a_d, N);
//cudaThreadSynchronize();
// Retrieve result from device and store it in host array
cudaMemcpy(a_h, a_d, sizeof(float)*N, cudaMemcpyDeviceToHost);
// Print results
for (int i=0; i<N; i++) printf("%d %f\n", i, a_h[i]);
// Cleanup
free(a_h); cudaFree(a_d);
diff = clock() - start;
int msec = diff * 1000 / CLOCKS_PER_SEC;
printf("Time taken %d seconds %d milliseconds\n", msec/1000, msec%1000);
}
then realizing that the loop calculation can be minimized with dynamic parallelism .
unrolling with dynamic parallelism was implemented as
unrolling with dynamic parallelism
#include <stdio.h>
#include <cuda.h>
__global__ void square(float *a, int N,int idx);
// Kernel that executes on the CUDA device
__global__ void first(float *arr, int N)
{
int idx = 2*(blockIdx.x * blockDim.x + threadIdx.x);
int n=N;
square <<< 1,2 >>> (arr, n,idx);
}
__global__ void square(float *a, int N,int idx)
{
int tdx = blockIdx.x * blockDim.x + threadIdx.x;
printf("%d\n",N);
if(N<2000)
{
a[tdx+idx] = a[tdx+idx] * a[tdx+idx];
}
}
// main routine that executes on the host
int main(void)
{
clock_t start = clock(),diff;
float *a_h, *a_d; // Pointer to host & device arrays
const int N = 1000; // Number of elements in arrays
size_t size = N * sizeof(float);
a_h = (float *)malloc(size); // Allocate array on host
cudaMalloc((void **) &a_d, size); // Allocate array on device
// Initialize host array and copy it to CUDA device
for (int i=0; i<N; i++) a_h[i] = (float)i;
cudaMemcpy(a_d, a_h, size, cudaMemcpyHostToDevice);
// Do calculation on device:
int block_size = 4;
//int n_blocks = N/block_size + (N%block_size == 0 ? 0:1);
first <<< 4, 128 >>> (a_d, N);
//cudaThreadSynchronize();
// Retrieve result from device and store it in host array
cudaMemcpy(a_h, a_d, sizeof(float)*N, cudaMemcpyDeviceToHost);
// Print results
for (int i=0; i<N; i++) printf("%d %f\n", i, a_h[i]);
// Cleanup
free(a_h); cudaFree(a_d);
diff = clock() - start;
int msec = diff * 1000 / CLOCKS_PER_SEC;
printf("Time taken %d seconds %d milliseconds\n", msec/1000, msec%1000);
}
the implementation of dynamic parallelism with unrolling takes more time for executio than only unrolling. Aren,t we suppose to improve execution time with dynamic parallelism in such case?
Dynamic parallelism is mainly useful in cases where you have parallelism that is dynamic. That is: cases where you don't know how much parallelism you're going to need until you've done some calculation. Rather than transfer data back to the host which is then instantly fed into parameterising another launch, you launch from within the kernel. In this pattern, with memcpys between kernel launches avoided, you'll see speedup.
In your example above this is not the case. You could have just launched twice as many threads from the host. There's nothing dynamic required as there's no parallelism available there that you didn't know about at the time of the first kernel launch.
Furthermore, performance requirements for kernels launched using dynamic parallelism are similar to that of those launched from the host. You have to launch a reasonable amount of work or the launch latency will dominate your computation time.
I have a cuda thrust program as
#include <stdio.h>
#include<iostream>
#include <cuda.h>
#include <thrust/sort.h>
// main routine that executes on the host
int main(void)
{
int *a_h, *a_d; // Pointer to host & device arrays
const int N = 10; // Number of elements in arrays
size_t size = N * sizeof(int);
a_h = (int *)malloc(size); // Allocate array on host
cudaMalloc((void **) &a_d, size);// Allocate array on device
std::cout<<"enter the 10 numbers";
// Initialize host array and copy it to CUDA device
for (int i=0; i<N; i++)
{
std::cin>>a_h[i];
}
for (int i=0; i<N; i++) printf("%d %d\n", i, a_h[i]);
cudaMemcpy(a_d, a_h, size, cudaMemcpyHostToDevice);
thrust::sort(a_d, a_d + N);
// Do calculation on device:
cudaMemcpy(a_h, a_d, sizeof(int)*N, cudaMemcpyDeviceToHost);
// Print results
for (int i=0; i<N; i++) printf("%d %d\n", i, a_h[i]);
// Cleanup
free(a_h); cudaFree(a_d);
}
but it is not running to give the desired output.
Are we supposed to use the host vector and device vector for sorting in thrust????
For device operations you should use either a device pointer, or a device_vector iterator, not raw pointers. Raw pointers (that point to host memory) can be used for operations on the host.
So if you modify your code as follows:
#include <thrust/device_ptr.h>
...
thrust::device_ptr<int> t_a(a_d); // add this line before the sort line
thrust::sort(t_a, t_a + N); // modify your sort line
I believe it will work for you.
You may wish to read the thrust quick start guide. In particular note this section:
You may wonder what happens when a "raw" pointer is used as an argument to a Thrust function. Like the STL, Thrust permits this usage and it will dispatch the host path of the algorithm. If the pointer in question is in fact a pointer to device memory then you'll need to wrap it with thrust::device_ptr before calling the function
I am a newbie to Thrust. I see that all Thrust presentations and examples only show host code.
I would like to know if I can pass a device_vector to my own kernel? How?
If yes, what are the operations permitted on it inside kernel/device code?
As it was originally written, Thrust is purely a host side abstraction. It cannot be used inside kernels. You can pass the device memory encapsulated inside a thrust::device_vector to your own kernel like this:
thrust::device_vector< Foo > fooVector;
// Do something thrust-y with fooVector
Foo* fooArray = thrust::raw_pointer_cast( fooVector.data() );
// Pass raw array and its size to kernel
someKernelCall<<< x, y >>>( fooArray, fooVector.size() );
and you can also use device memory not allocated by thrust within thrust algorithms by instantiating a thrust::device_ptr with the bare cuda device memory pointer.
Edited four and half years later to add that as per #JackOLantern's answer, thrust 1.8 adds a sequential execution policy which means you can run single threaded versions of thrust's alogrithms on the device. Note that it still isn't possible to directly pass a thrust device vector to a kernel and device vectors can't be directly used in device code.
Note that it is also possible to use the thrust::device execution policy in some cases to have parallel thrust execution launched by a kernel as a child grid. This requires separate compilation/device linkage and hardware which supports dynamic parallelism. I am not certain whether this is actually supported in all thrust algorithms or not, but certainly works with some.
This is an update to my previous answer.
Starting from Thrust 1.8.1, CUDA Thrust primitives can be combined with the thrust::device execution policy to run in parallel within a single CUDA thread exploiting CUDA dynamic parallelism. Below, an example is reported.
#include <stdio.h>
#include <thrust/reduce.h>
#include <thrust/execution_policy.h>
#include "TimingGPU.cuh"
#include "Utilities.cuh"
#define BLOCKSIZE_1D 256
#define BLOCKSIZE_2D_X 32
#define BLOCKSIZE_2D_Y 32
/*************************/
/* TEST KERNEL FUNCTIONS */
/*************************/
__global__ void test1(const float * __restrict__ d_data, float * __restrict__ d_results, const int Nrows, const int Ncols) {
const unsigned int tid = threadIdx.x + blockDim.x * blockIdx.x;
if (tid < Nrows) d_results[tid] = thrust::reduce(thrust::seq, d_data + tid * Ncols, d_data + (tid + 1) * Ncols);
}
__global__ void test2(const float * __restrict__ d_data, float * __restrict__ d_results, const int Nrows, const int Ncols) {
const unsigned int tid = threadIdx.x + blockDim.x * blockIdx.x;
if (tid < Nrows) d_results[tid] = thrust::reduce(thrust::device, d_data + tid * Ncols, d_data + (tid + 1) * Ncols);
}
/********/
/* MAIN */
/********/
int main() {
const int Nrows = 64;
const int Ncols = 2048;
gpuErrchk(cudaFree(0));
// size_t DevQueue;
// gpuErrchk(cudaDeviceGetLimit(&DevQueue, cudaLimitDevRuntimePendingLaunchCount));
// DevQueue *= 128;
// gpuErrchk(cudaDeviceSetLimit(cudaLimitDevRuntimePendingLaunchCount, DevQueue));
float *h_data = (float *)malloc(Nrows * Ncols * sizeof(float));
float *h_results = (float *)malloc(Nrows * sizeof(float));
float *h_results1 = (float *)malloc(Nrows * sizeof(float));
float *h_results2 = (float *)malloc(Nrows * sizeof(float));
float sum = 0.f;
for (int i=0; i<Nrows; i++) {
h_results[i] = 0.f;
for (int j=0; j<Ncols; j++) {
h_data[i*Ncols+j] = i;
h_results[i] = h_results[i] + h_data[i*Ncols+j];
}
}
TimingGPU timerGPU;
float *d_data; gpuErrchk(cudaMalloc((void**)&d_data, Nrows * Ncols * sizeof(float)));
float *d_results1; gpuErrchk(cudaMalloc((void**)&d_results1, Nrows * sizeof(float)));
float *d_results2; gpuErrchk(cudaMalloc((void**)&d_results2, Nrows * sizeof(float)));
gpuErrchk(cudaMemcpy(d_data, h_data, Nrows * Ncols * sizeof(float), cudaMemcpyHostToDevice));
timerGPU.StartCounter();
test1<<<iDivUp(Nrows, BLOCKSIZE_1D), BLOCKSIZE_1D>>>(d_data, d_results1, Nrows, Ncols);
gpuErrchk(cudaPeekAtLastError());
gpuErrchk(cudaDeviceSynchronize());
printf("Timing approach nr. 1 = %f\n", timerGPU.GetCounter());
gpuErrchk(cudaMemcpy(h_results1, d_results1, Nrows * sizeof(float), cudaMemcpyDeviceToHost));
for (int i=0; i<Nrows; i++) {
if (h_results1[i] != h_results[i]) {
printf("Approach nr. 1; Error at i = %i; h_results1 = %f; h_results = %f", i, h_results1[i], h_results[i]);
return 0;
}
}
timerGPU.StartCounter();
test2<<<iDivUp(Nrows, BLOCKSIZE_1D), BLOCKSIZE_1D>>>(d_data, d_results1, Nrows, Ncols);
gpuErrchk(cudaPeekAtLastError());
gpuErrchk(cudaDeviceSynchronize());
printf("Timing approach nr. 2 = %f\n", timerGPU.GetCounter());
gpuErrchk(cudaMemcpy(h_results1, d_results1, Nrows * sizeof(float), cudaMemcpyDeviceToHost));
for (int i=0; i<Nrows; i++) {
if (h_results1[i] != h_results[i]) {
printf("Approach nr. 2; Error at i = %i; h_results1 = %f; h_results = %f", i, h_results1[i], h_results[i]);
return 0;
}
}
printf("Test passed!\n");
}
The above example performs reductions of the rows of a matrix in the same sense as Reduce matrix rows with CUDA, but it is done differently from the above post, namely, by calling CUDA Thrust primitives directly from user written kernels. Also, the above example serves to compare the performance of the same operations when done with two execution policies, namely, thrust::seq and thrust::device. Below, some graphs showing the difference in performance.
The performance has been evaluated on a Kepler K20c and on a Maxwell GeForce GTX 850M.
I would like to provide an updated answer to this question.
Starting from Thrust 1.8, CUDA Thrust primitives can be combined with the thrust::seq execution policy to run sequentially within a single CUDA thread (or sequentially within a single CPU thread). Below, an example is reported.
If you want parallel execution within a thread, then you may consider using CUB which provides reduction routines that can be called from within a threadblock, provided that your card enables dynamic parallelism.
Here is the example with Thrust
#include <stdio.h>
#include <thrust/reduce.h>
#include <thrust/execution_policy.h>
/********************/
/* CUDA ERROR CHECK */
/********************/
#define gpuErrchk(ans) { gpuAssert((ans), __FILE__, __LINE__); }
inline void gpuAssert(cudaError_t code, char *file, int line, bool abort=true)
{
if (code != cudaSuccess)
{
fprintf(stderr,"GPUassert: %s %s %d\n", cudaGetErrorString(code), file, line);
if (abort) exit(code);
}
}
__global__ void test(float *d_A, int N) {
float sum = thrust::reduce(thrust::seq, d_A, d_A + N);
printf("Device side result = %f\n", sum);
}
int main() {
const int N = 16;
float *h_A = (float*)malloc(N * sizeof(float));
float sum = 0.f;
for (int i=0; i<N; i++) {
h_A[i] = i;
sum = sum + h_A[i];
}
printf("Host side result = %f\n", sum);
float *d_A; gpuErrchk(cudaMalloc((void**)&d_A, N * sizeof(float)));
gpuErrchk(cudaMemcpy(d_A, h_A, N * sizeof(float), cudaMemcpyHostToDevice));
test<<<1,1>>>(d_A, N);
}
If you mean to use the data allocated / processed by thrust yes you can, just get the raw pointer of the allocated data.
int * raw_ptr = thrust::raw_pointer_cast(dev_ptr);
if you want to allocate thrust vectors in the kernel I never tried but I don't think will work
and also if it works I don't think it will provide any benefit.