cudaStream strange performance - cuda

I try to develop an example of sobel with cudaStream. Here is the program:
void SobelStream(void)
{
cv::Mat imageGrayL2 = cv::imread("/home/xavier/Bureau/Image1.png",0);
u_int8_t *u8_PtImageHost;
u_int8_t *u8_PtImageDevice;
u_int8_t *u8_ptDataOutHost;
u_int8_t *u8_ptDataOutDevice;
u_int8_t u8_Used[NB_STREAM];
u8_ptDataOutHost = (u_int8_t *)malloc(WIDTH*HEIGHT*sizeof(u_int8_t));
checkCudaErrors(cudaMalloc((void**)&u8_ptDataOutDevice,WIDTH*HEIGHT*sizeof(u_int8_t)));
u8_PtImageHost = (u_int8_t *)malloc(WIDTH*HEIGHT*sizeof(u_int8_t));
checkCudaErrors(cudaMalloc((void**)&u8_PtImageDevice,WIDTH*HEIGHT*sizeof(u_int8_t)));
cudaChannelFormatDesc channelDesc = cudaCreateChannelDesc<unsigned char>();
checkCudaErrors(cudaMallocArray(&Array_PatchsMaxDevice, &channelDesc,WIDTH,HEIGHT ));
checkCudaErrors(cudaBindTextureToArray(Image,Array_PatchsMaxDevice));
dim3 threads(BLOC_X,BLOC_Y);
dim3 blocks(ceil((float)WIDTH/BLOC_X),ceil((float)HEIGHT/BLOC_Y));
ClearKernel<<<blocks,threads>>>(u8_ptDataOutDevice,WIDTH,HEIGHT);
int blockh = HEIGHT/NB_STREAM;
Stream = (cudaStream_t *) malloc(NB_STREAM * sizeof(cudaStream_t));
for (int i = 0; i < NB_STREAM; i++)
{
checkCudaErrors(cudaStreamCreate(&(Stream[i])));
}
// for(int i=0;i<NB_STREAM;i++)
// {
// cudaSetDevice(0);
// cudaStreamCreate(&Stream[i]);
// }
cudaEvent_t Start;
cudaEvent_t Stop;
cudaEventCreate(&Start);
cudaEventCreate(&Stop);
cudaEventRecord(Start, 0);
//////////////////////////////////////////////////////////
for(int i=0;i<NB_STREAM;i++)
{
if(i == 0)
{
int localHeight = blockh;
checkCudaErrors(cudaMemcpy2DToArrayAsync( Array_PatchsMaxDevice,
0,
0,
imageGrayL2.data,//u8_PtImageDevice,
WIDTH,
WIDTH,
blockh,
cudaMemcpyHostToDevice ,
Stream[i]));
dim3 threads(BLOC_X,BLOC_Y);
dim3 blocks(ceil((float)WIDTH/BLOC_X),ceil((float)localHeight/BLOC_Y));
SobelKernel<<<blocks,threads,0,Stream[i]>>>(u8_ptDataOutDevice,0,WIDTH,localHeight-1);
checkCudaErrors(cudaGetLastError());
u8_Used[i] = 1;
}else{
int ioffsetImage = WIDTH*(HEIGHT/NB_STREAM );
int hoffset = HEIGHT/NB_STREAM *i;
int hoffsetkernel = HEIGHT/NB_STREAM -1 + HEIGHT/NB_STREAM* (i-1);
int localHeight = min(HEIGHT - (blockh*i),blockh);
//printf("hoffset: %d hoffsetkernel %d localHeight %d rest %d ioffsetImage %d \n",hoffset,hoffsetkernel,localHeight,HEIGHT - (blockh +1 +blockh*(i-1)),ioffsetImage*i/WIDTH);
checkCudaErrors(cudaMemcpy2DToArrayAsync( Array_PatchsMaxDevice,
0,
hoffset,
&imageGrayL2.data[ioffsetImage*i],//&u8_PtImageDevice[ioffset*i],
WIDTH,
WIDTH,
localHeight,
cudaMemcpyHostToDevice ,
Stream[i]));
u8_Used[i] = 1;
if(HEIGHT - (blockh +1 +blockh*(i-1))<=0)
{
break;
}
}
}
///////////////////////////////////////////
for(int i=0;i<NB_STREAM;i++)
{
if(i == 0)
{
int localHeight = blockh;
dim3 threads(BLOC_X,BLOC_Y);
dim3 blocks(1,1);
SobelKernel<<<blocks,threads,0,Stream[i]>>>(u8_ptDataOutDevice,0,WIDTH,localHeight-1);
checkCudaErrors(cudaGetLastError());
u8_Used[i] = 1;
}else{
int ioffsetImage = WIDTH*(HEIGHT/NB_STREAM );
int hoffset = HEIGHT/NB_STREAM *i;
int hoffsetkernel = HEIGHT/NB_STREAM -1 + HEIGHT/NB_STREAM* (i-1);
int localHeight = min(HEIGHT - (blockh*i),blockh);
dim3 threads(BLOC_X,BLOC_Y);
dim3 blocks(1,1);
SobelKernel<<<blocks,threads,0,Stream[i]>>>(u8_ptDataOutDevice,hoffsetkernel,WIDTH,localHeight);
checkCudaErrors(cudaGetLastError());
u8_Used[i] = 1;
if(HEIGHT - (blockh +1 +blockh*(i-1))<=0)
{
break;
}
}
}
///////////////////////////////////////////////////////
for(int i=0;i<NB_STREAM;i++)
{
if(i == 0)
{
int localHeight = blockh;
checkCudaErrors(cudaMemcpyAsync(u8_ptDataOutHost,u8_ptDataOutDevice,WIDTH*(localHeight-1)*sizeof(u_int8_t),cudaMemcpyDeviceToHost,Stream[i]));
u8_Used[i] = 1;
}else{
int ioffsetImage = WIDTH*(HEIGHT/NB_STREAM );
int hoffset = HEIGHT/NB_STREAM *i;
int hoffsetkernel = HEIGHT/NB_STREAM -1 + HEIGHT/NB_STREAM* (i-1);
int localHeight = min(HEIGHT - (blockh*i),blockh);
checkCudaErrors(cudaMemcpyAsync(&u8_ptDataOutHost[hoffsetkernel*WIDTH],&u8_ptDataOutDevice[hoffsetkernel*WIDTH],WIDTH*localHeight*sizeof(u_int8_t),cudaMemcpyDeviceToHost,Stream[i]));
u8_Used[i] = 1;
if(HEIGHT - (blockh +1 +blockh*(i-1))<=0)
{
break;
}
}
}
for(int i=0;i<NB_STREAM;i++)
{
cudaStreamSynchronize(Stream[i]);
}
cudaEventRecord(Stop, 0);
cudaEventSynchronize(Start);
cudaEventSynchronize(Stop);
float dt_ms;
cudaEventElapsedTime(&dt_ms, Start, Stop);
printf("dt_ms %f \n",dt_ms);
}
I had a really strange performance on th execution of my program. I decided to profile my example and I get that:
I don't understand it seems that each stream are waiting each other.
Can someone help me about that?

First of all, in the future, please provide a complete code. I'm also working off of your cross-posting here to fill in some details such as kernel sizes.
You have two issues to address:
First, any time you wish to use cudaMemcpyAsync, you will most likely want to be working with pinned host allocations. If you use allocations created e.g. with malloc, you will not get the expected behavior from cudaMemcpyAsync as far as asynchronous concurrent execution is concerned. This necessity is covered in the programming guide:
If host memory is involved in the copy, it must be page-locked.
So the first change to make to your code is to convert this:
u8_PtImageHost = (u_int8_t *)malloc(WIDTH*HEIGHT*sizeof(u_int8_t));
u8_ptDataOutHost = (u_int8_t *)malloc(WIDTH*HEIGHT*sizeof(u_int8_t));
to this:
checkCudaErrors(cudaHostAlloc(&u8_PtImageHost, WIDTH*HEIGHT*sizeof(u_int8_t), cudaHostAllocDefault));
checkCudaErrors(cudaHostAlloc(&u8_ptDataOutHost, WIDTH*HEIGHT*sizeof(u_int8_t), cudaHostAllocDefault));
with that change alone, your execution duration drops from about 21ms to 7ms according to my testing. The reason for this is that without the change, we get no overlap whatsoever:
With the change, the copy activity can overlap with each other (H->D and D->H) and with kernel execution:
The second issue you face to get to concurrent kernel execution is that your kernels are just too large (too many blocks/threads):
#define WIDTH 6400
#define HEIGHT 4800
#define NB_STREAM 10
#define BLOC_X 32
#define BLOC_Y 32
dim3 threads(BLOC_X,BLOC_Y);
dim3 blocks(ceil((float)WIDTH/BLOC_X),ceil((float)HEIGHT/BLOC_Y));
I would suggest that if these are the sizes of kernels you need to run, there's probably not much benefit to try and strive for kernel overlap - each kernel is launching enough blocks to "fill" the GPU, so you have already exposed enough parallelism to keep the GPU busy. But if you are desperate to witness kernel concurrency, you could make your kernels use a smaller number of blocks while causing each kernel to spend more time executing. We could do this by launching 1 block, and have just the the threads in each block perform the image filtering.

Related

cudaMallocManaged and cudaDeviceSynchronize()

I have the following two mostly identical example codes. code1.cu use cudaMalloc and cudaMemcpy to handling device/host variable value exchange.
The code2.cu use cudaMallocManaged and thus cudaMemcpy is not needed. When cudaMallocManaged is used, I have to include cudaDeviceSynchronize() to get the correct results, while for the one with cudaMalloc, this is not needed. I would appreciate some hint on why this is happening
code2.cu
#include <iostream>
#include <math.h>
#include <vector>
//
using namespace std;
// Kernel function to do nested loops
__global__
void add(int max_x, int max_y, float *tot, float *x, float *y)
{
int i = blockIdx.x*blockDim.x + threadIdx.x;
int j = blockIdx.y*blockDim.y + threadIdx.y;
if(i < max_x && j<max_y) {
atomicAdd(tot, x[i] + y[j]);
}
}
int main(void)
{
int Nx = 1<<15;
int Ny = 1<<15;
float *d_x = NULL, *d_y = NULL;
float *d_tot = NULL;
cudaMalloc((void **)&d_x, sizeof(float)*Nx);
cudaMalloc((void **)&d_y, sizeof(float)*Ny);
cudaMallocManaged((void **)&d_tot, sizeof(float));
// Allocate Unified Memory – accessible from CPU or GPU
vector<float> vx;
vector<float> vy;
// initialize x and y arrays on the host
for (int i = 0; i < Nx; i++)
vx.push_back(i);
for (int i = 0; i < Ny; i++)
vy.push_back(i*10);
//
float tot = 0;
for(int i = 0; i<vx.size(); i++)
for(int j = 0; j<vy.size(); j++)
tot += vx[i] + vy[j];
cout<<"CPU: tot: "<<tot<<endl;
//
cudaMemcpy(d_x, vx.data(), vx.size()*sizeof(float), cudaMemcpyHostToDevice);
cudaMemcpy(d_y, vy.data(), vy.size()*sizeof(float), cudaMemcpyHostToDevice);
//
int blockSize; // The launch configurator returned block size
int minGridSize; // The minimum grid size needed to achieve the
cudaOccupancyMaxPotentialBlockSize( &minGridSize, &blockSize, add, 0, Nx+Ny);
//.. bx*by can not go beyond the blockSize, or hardware limit, which is 1024;
//.. bx*bx = blockSize && bx/by=Nx/Ny, solve the equation
int bx = sqrt(blockSize*Nx/(float)Ny);
int by = bx*Ny/(float)Nx;
dim3 blockSize_3D(bx, by);
dim3 gridSize_3D((Nx+bx-1)/bx, (Ny+by+1)/by);
cout<<"blockSize: "<<blockSize<<endl;
cout<<"bx: "<<bx<<" by: "<<by<<" gx: "<<gridSize_3D.x<<" gy: "<<gridSize_3D.y<<endl;
// calculate theoretical occupancy
int maxActiveBlocks;
cudaOccupancyMaxActiveBlocksPerMultiprocessor( &maxActiveBlocks, add, blockSize, 0);
int device;
cudaDeviceProp props;
cudaGetDevice(&device);
cudaGetDeviceProperties(&props, device);
float occupancy = (maxActiveBlocks * blockSize / props.warpSize) /
(float)(props.maxThreadsPerMultiProcessor /
props.warpSize);
printf("Launched blocks of size %d. Theoretical occupancy: %f\n",
blockSize, occupancy);
// Run kernel on 1M elements on the GPU
tot = 0;
add<<<gridSize_3D, blockSize_3D>>>(Nx, Ny, d_tot, d_x, d_y);
// Wait for GPU to finish before accessing on host
//cudaDeviceSynchronize();
tot =*d_tot;
//
//
cout<<" GPU: tot: "<<tot<<endl;
// Free memory
cudaFree(d_x);
cudaFree(d_y);
cudaFree(d_tot);
return 0;
}
code1.cu
#include <iostream>
#include <math.h>
#include <vector>
//
using namespace std;
// Kernel function to do nested loops
__global__
void add(int max_x, int max_y, float *tot, float *x, float *y)
{
int i = blockIdx.x*blockDim.x + threadIdx.x;
int j = blockIdx.y*blockDim.y + threadIdx.y;
if(i < max_x && j<max_y) {
atomicAdd(tot, x[i] + y[j]);
}
}
int main(void)
{
int Nx = 1<<15;
int Ny = 1<<15;
float *d_x = NULL, *d_y = NULL;
float *d_tot = NULL;
cudaMalloc((void **)&d_x, sizeof(float)*Nx);
cudaMalloc((void **)&d_y, sizeof(float)*Ny);
cudaMalloc((void **)&d_tot, sizeof(float));
// Allocate Unified Memory – accessible from CPU or GPU
vector<float> vx;
vector<float> vy;
// initialize x and y arrays on the host
for (int i = 0; i < Nx; i++)
vx.push_back(i);
for (int i = 0; i < Ny; i++)
vy.push_back(i*10);
//
float tot = 0;
for(int i = 0; i<vx.size(); i++)
for(int j = 0; j<vy.size(); j++)
tot += vx[i] + vy[j];
cout<<"CPU: tot: "<<tot<<endl;
//
cudaMemcpy(d_x, vx.data(), vx.size()*sizeof(float), cudaMemcpyHostToDevice);
cudaMemcpy(d_y, vy.data(), vy.size()*sizeof(float), cudaMemcpyHostToDevice);
//
int blockSize; // The launch configurator returned block size
int minGridSize; // The minimum grid size needed to achieve the
cudaOccupancyMaxPotentialBlockSize( &minGridSize, &blockSize, add, 0, Nx+Ny);
//.. bx*by can not go beyond the blockSize, or hardware limit, which is 1024;
//.. bx*bx = blockSize && bx/by=Nx/Ny, solve the equation
int bx = sqrt(blockSize*Nx/(float)Ny);
int by = bx*Ny/(float)Nx;
dim3 blockSize_3D(bx, by);
dim3 gridSize_3D((Nx+bx-1)/bx, (Ny+by+1)/by);
cout<<"blockSize: "<<blockSize<<endl;
cout<<"bx: "<<bx<<" by: "<<by<<" gx: "<<gridSize_3D.x<<" gy: "<<gridSize_3D.y<<endl;
// calculate theoretical occupancy
int maxActiveBlocks;
cudaOccupancyMaxActiveBlocksPerMultiprocessor( &maxActiveBlocks, add, blockSize, 0);
int device;
cudaDeviceProp props;
cudaGetDevice(&device);
cudaGetDeviceProperties(&props, device);
float occupancy = (maxActiveBlocks * blockSize / props.warpSize) /
(float)(props.maxThreadsPerMultiProcessor /
props.warpSize);
printf("Launched blocks of size %d. Theoretical occupancy: %f\n",
blockSize, occupancy);
// Run kernel on 1M elements on the GPU
tot = 0;
add<<<gridSize_3D, blockSize_3D>>>(Nx, Ny, d_tot, d_x, d_y);
// Wait for GPU to finish before accessing on host
//cudaDeviceSynchronize();
//
cudaMemcpy(&tot, d_tot, sizeof(float), cudaMemcpyDeviceToHost);
//
cout<<" GPU: tot: "<<tot<<endl;
// Free memory
cudaFree(d_x);
cudaFree(d_y);
cudaFree(d_tot);
return 0;
}
//Code2.cu has the following output:
//
//CPU: tot: 8.79609e+12
//blockSize: 1024
//bx: 32 by: 32 gx: 1024 gy: 1025
//Launched blocks of size 1024. Theoretical occupancy: 1.000000
//GPU: tot: 0
After remove the comment on cudaDeviceSynchronize(),
GPU: tot: 8.79609e+12
CUDA kernel launches are asynchronous. That means that they execute independently of the CPU thread that launched them.
Because of this asynchronous launch, the CUDA kernel is not guaranteed to be finished (or even started) by the time your CPU thread code begins testing the result.
Therefore it is necessary to wait until the GPU kernel is complete, and cudaDeviceSynchronize() does exactly that. cudaMemcpy also has a synchronizing effect, so when you remove the cudaMemcpy operations, you lose that synchronization, but cudaDeviceSynchronize() restores it.

Has cudaMalloc changed to be asynchronous?

I've read in other places that cudaMalloc will synchronize across kernels.
(e.g. will cudaMalloc synchronize host and device?)
However, I just tested this code out and based on what I'm seeing in the visual profiler, it seems like cudaMalloc is not synchronizing. if you add cudaFree into the loop, that does synchronize. I'm using CUDA 7.5. Does anyone know if cudaMalloc changed its behavior? Or am I missing some subtlety? Thanks very much!
__global__ void slowKernel()
{
float input = 5;
for( int i = 0; i < 1000000; i++ ){
input = input * .9999999;
}
}
__global__ void fastKernel()
{
float input = 5;
for( int i = 0; i < 100000; i++ ){
input = input * .9999999;
}
}
void mallocSynchronize(){
cudaStream_t stream1, stream2;
cudaStreamCreate( &stream1 );
cudaStreamCreate( &stream2 );
slowKernel <<<1, 1, 0, stream1 >>>();
int *dev_a = 0;
for( int i = 0; i < 10; i++ ){
cudaMalloc( &dev_a, 4 * 1024 * 1024 );
fastKernel <<<1, 1, 0, stream2 >>>();
// cudaFree( dev_a ); // If you uncomment this, the second fastKernel launch will wait until slowKernel completes
}
}
Your methodology is flawed, but you conclusion looks correct to me (if you look at your profile data you should see that both long and short kernels are taking the same amount of time and run very quickly, because aggressive compiler optimisation is eliminating all the code in both cases).
I turned your example into something more reasonable
#include <time.h>
__global__ void slowKernel(float *output, bool write=false)
{
float input = 5;
#pragma unroll
for( int i = 0; i < 10000000; i++ ){
input = input * .9999999;
}
if (write) *output -= input;
}
__global__ void fastKernel(float *output, bool write=false)
{
float input = 5;
#pragma unroll
for( int i = 0; i < 100000; i++ ){
input = input * .9999999;
}
if (write) *output -= input;
}
void burntime(long val) {
struct timespec tv[] = {{0, val}};
nanosleep(tv, 0);
}
void mallocSynchronize(){
cudaStream_t stream1, stream2;
cudaStreamCreate( &stream1 );
cudaStreamCreate( &stream2 );
const size_t sz = 1 << 21;
slowKernel <<<1, 1, 0, stream1 >>>((float *)(0));
burntime(500000000L); // 500ms wait - slowKernel around 1300ms
int *dev_a = 0;
for( int i = 0; i < 10; i++ ){
cudaMalloc( &dev_a, sz );
fastKernel <<<1, 1, 0, stream2 >>>((float *)(0));
burntime(1000000L); // 1ms wait - fastKernel around 15ms
}
}
int main()
{
mallocSynchronize();
cudaDeviceSynchronize();
cudaDeviceReset();
return 0;
}
[note requires POSIX time functions so this won't run on Windows]
On a fairly fast Maxwell device (GTX970), I see that cudaMalloc calls in the loop overlap with the still executing slowKernel call in the profile trace, and then with running fastKernel calls in the other stream. I was willing to accept the initial conclusion that minor timing variations could be cause the effect you saw in your broken example. However, in this code, 0.5 seconds time shift in synchronisation between the host and device traces seems very improbable. You might need to vary the duration of the burntime calls to get the same effect, depending on how fast your GPU is.
So this is a very long way of saying, yes it looks like it is a non-synchronising call on Linux with CUDA 7.5 and a Maxwell device. I don't believe that has always been the case, but then again the documentation has never, as best as I can tell, said whether is should block/synchronize or not. I don't have access to older CUDA versions and supported hardware to see what this example would do with an older driver and a Fermi or Kepler device.

cuda display driver stppoed [duplicate]

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.

The result of an experiment different from CUDA Occupancy Calculator

I study CUDA architecture.
I made some of parallel processing code in environment like below.
GPU : GTX580 (CC is 2.0)
Threads Per Block : 16x16 = 256
Registers Per Thread : 16
Shared Memory Per Block : 48 bytes
I know the number of Registers and Shared Memory size by the compile option: --ptxas-options=-v
In addition, grid size is 32x32 = 1024 and there is not extra shared memory.
So, I tried to use CUDA_Occupancy_Calculator by NVIDIA.
then, It said,
3.) GPU Occupancy Data is displayed here and in the graphs:
Active Threads per Multiprocessor 1536
Active Warps per Multiprocessor 48
Active Thread Blocks per Multiprocessor 6
Occupancy of each Multiprocessor 100%
So, I run the application.
But, the result showed that the block size is 8x8 faster than 16x16.
8x8 means the block size, and the gird size is 64x64.
16x16 means the block size, and the grid size is 32x32.
So, the total amount of threads is same. It's unchanged.
I don't know the why. Please help me.
Following code is a part of my Program.
void LOAD_VERTEX(){
MEM[0] = 60; //y0
MEM[1] = 50; //x0
MEM[2] = 128; //r0
MEM[3] = 0; //g0
MEM[4] = 70; //b0
MEM[5] = 260;
MEM[6] = 50;
MEM[7] = 135;
MEM[8] = 70;
MEM[9] = 0;
MEM[10] = 260;
MEM[11] = 250;
MEM[12] = 0;
MEM[13] = 200;
MEM[14] = 55;
MEM[15] = 60;
MEM[16] = 250;
MEM[17] = 55;
MEM[18] = 182;
MEM[19] = 100;
MEM[20] = 30;
MEM[21] = 330;
MEM[22] = 72;
MEM[23] = 12;
MEM[24] = 25;
MEM[25] = 30;
MEM[26] = 130;
MEM[27] = 80;
MEM[28] = 255;
MEM[29] = 15;
MEM[30] = 230;
MEM[31] = 330;
MEM[32] = 56;
MEM[33] = 186;
MEM[34] = 201;
}
__global__ void PRINT_POLYGON( unsigned char *IMAGEin, int *MEMin, int dev_ID, int a, int b, int c)
{
int i = blockIdx.x*TILE_WIDTH + threadIdx.x;
int j = blockIdx.y*TILE_HEIGHT + threadIdx.y;
float result_a, result_b;
int temp[15];
int k;
for(k = 0; k < 5; k++){
temp[k] = a*5+k;
temp[k+5] = b*5+k;
temp[k+10] = c*5+k;
}
int result_a_up = ((MEMin[temp[11]]-MEMin[temp[1]])*(i-MEMin[temp[0]]))-((MEMin[temp[10]]-MEMin[temp[0]])*(j-MEMin[temp[1]]));
int result_a_down = ((MEMin[temp[11]]-MEMin[temp[1]])*(MEMin[temp[5]]-MEMin[temp[0]]))-((MEMin[temp[6]]-MEMin[temp[1]])*(MEMin[temp[10]]-MEMin[temp[0]]));
int result_b_up = ((MEMin[temp[6]] -MEMin[temp[1]])*(MEMin[temp[0]]-i))-((MEMin[temp[5]] -MEMin[temp[0]])*(MEMin[temp[1]]-j));
int result_b_down = ((MEMin[temp[11]]-MEMin[temp[1]])*(MEMin[temp[5]]-MEMin[temp[0]]))-((MEMin[temp[6]]-MEMin[temp[1]])*(MEMin[temp[10]]-MEMin[temp[0]]));
result_a = float(result_a_up) / float(result_a_down);
result_b = float(result_b_up) / float(result_b_down);
int isIn = (0 <= result_a && result_a <=1) && ((0 <= result_b && result_b <= 1)) && ((0 <= (result_a+result_b) && (result_a+result_b) <= 1));
IMAGEin[(i*HEIGHTs+j)*CHANNELS] += (int)(float(MEMin[temp[2]]) + (float(MEMin[temp[7]])-float(MEMin[temp[2]]))*result_a + (float(MEMin[temp[12]])-float(MEMin[temp[2]]))*result_b) * isIn; //Red Channel
IMAGEin[(i*HEIGHTs+j)*CHANNELS+1] += (int)(float(MEMin[temp[3]]) + (float(MEMin[temp[8]])-float(MEMin[temp[3]]))*result_a + (float(MEMin[temp[13]])-float(MEMin[temp[3]]))*result_b) * isIn; //Green Channel
IMAGEin[(i*HEIGHTs+j)*CHANNELS+2] += (int)(float(MEMin[temp[4]]) + (float(MEMin[temp[9]])-float(MEMin[temp[4]]))*result_a + (float(MEMin[temp[14]])-float(MEMin[temp[4]]))*result_b) * isIn; //Blue Channel
}
//The information each device
struct DataStruct {
int deviceID;
unsigned char IMAGE_SEG[WIDTH*HEIGHTs*CHANNELS];
};
void* routine( void *pvoidData ) {
DataStruct *data = (DataStruct*)pvoidData;
unsigned char *dev_IMAGE;
int *dev_MEM;
unsigned char *IMAGE_SEG = data->IMAGE_SEG;
HANDLE_ERROR(cudaSetDevice(data->deviceID));
//initialize array
memset(IMAGE_SEG, 0, WIDTH*HEIGHTs*CHANNELS);
printf("Device %d Starting..\n", data->deviceID);
//Evaluate Time
cudaEvent_t start, stop;
cudaEventCreate( &start );
cudaEventCreate( &stop );
HANDLE_ERROR( cudaMalloc( (void **)&dev_MEM, sizeof(int)*35) ); //Creating int array each Block
HANDLE_ERROR( cudaMalloc( (void **)&dev_IMAGE, sizeof(unsigned char)*WIDTH*HEIGHTs*CHANNELS) ); //output array
cudaMemcpy(dev_MEM, MEM, sizeof(int)*256, cudaMemcpyHostToDevice);
cudaMemset(dev_IMAGE, 0, sizeof(unsigned char)*WIDTH*HEIGHTs*CHANNELS);
dim3 grid(WIDTH/TILE_WIDTH, HEIGHTs/TILE_HEIGHT); //blocks in a grid
dim3 block(TILE_WIDTH, TILE_HEIGHT); //threads in a block
cudaEventRecord(start, 0);
PRINT_POLYGON<<<grid,block>>>( dev_IMAGE, dev_MEM, data->deviceID, 0, 1, 2); //Start the Kernel
PRINT_POLYGON<<<grid,block>>>( dev_IMAGE, dev_MEM, data->deviceID, 0, 2, 3); //Start the Kernel
PRINT_POLYGON<<<grid,block>>>( dev_IMAGE, dev_MEM, data->deviceID, 0, 3, 4); //Start the Kernel
PRINT_POLYGON<<<grid,block>>>( dev_IMAGE, dev_MEM, data->deviceID, 0, 4, 5); //Start the Kernel
PRINT_POLYGON<<<grid,block>>>( dev_IMAGE, dev_MEM, data->deviceID, 3, 2, 4); //Start the Kernel
PRINT_POLYGON<<<grid,block>>>( dev_IMAGE, dev_MEM, data->deviceID, 2, 6, 4); //Start the Kernel
cudaEventRecord(stop, 0);
cudaEventSynchronize(stop);
HANDLE_ERROR( cudaMemcpy( IMAGE_SEG, dev_IMAGE, sizeof(unsigned char)*WIDTH*HEIGHTs*CHANNELS, cudaMemcpyDeviceToHost ) );
HANDLE_ERROR( cudaFree( dev_MEM ) );
HANDLE_ERROR( cudaFree( dev_IMAGE ) );
cudaEventElapsedTime( &elapsed_time_ms[data->deviceID], start, stop ); //Calculate elapsed time
cudaEventDestroy(start);
cudaEventDestroy(stop);
printf("Algorithm Elapsed Time : %f ms(Device %d)\n", elapsed_time_ms[data->deviceID], data->deviceID);
printf("Device %d Complete!\n", data->deviceID);
return 0;
}
int main( void )
{
int i;
CUTThread thread[7];
printf("Program Start.\n");
LOAD_VERTEX();
DataStruct data[DEVICENUM]; //define device info
for(i = 0; i < DEVICENUM; i++){
data[i].deviceID = i;
thread[i] = start_thread(routine, &(data[i]));
}
for(i = 0; i < DEVICENUM; i++){
end_thread(thread[i]);
}
cudaFreeHost(MEM);
return 0;
}
Since you copied over your question from the Nvidia forum, I'll copy my answer as well:
For your kernel your finding of reduced performance with higher occupancy is easily explained by the cache overflowing for higher occupancy.
The local array temp[] at full occupancy requires 1536×15×4=92160 bytes of cache, while at 33% occupancy (for the smaller 8×8 block size) only 512×15×4=30720 bytes are required per SM. With the larger 48kB cache/SM setting the latter could be fully cached eliminating off-chip memory accesses for temp[] almost completely, but even in the default 16kB cache/SM setting the cache hit probability is substantially higher.
As the temp[] array is not needed anyway, the fastest option (at either occupancy) would be to completely eliminate it. The compiler might already be able to achieve this if you just insert a #pragma unroll before the initialization loop. Otherwise replace all uses of temp[] with a little macro or inline function, or even just substitute the result into the code (which in this case I would even find more readable).

Problem with CUDA operation overlapping example

all
I referred to simpleMultiCopy.cu in CUDA SDK 4.0 and wrote one, see code below.
simpleMultiCopy.cu is an example of operation overlapping in a loop. And mine is similar, it will send a slice of data to GPU to compute each iteration in a loop where I perform the overlapping operation.
This is just a test/demo, don't care the logic of the kernel(increment_kernel), it was used just to delay some time. The main logic lies in processWithStreams function.
But this program works incorrectly with this out put:
i: 0, current_stream: 0, next_stream: 1
i: 1, current_stream: 1, next_stream: 0
Cuda error in file 'ttt.cu' in line 132 : unspecified launch failure.
line 132 is:
CUDA_SAFE_CALL( cudaMemcpyAsync(
d_data_in[next_stream],
h_data_in[next_stream],
memsize,
cudaMemcpyHostToDevice,
stream[next_stream]) ); //this is line 132
I don't have much ideas about how CUDA works, so please help.
Any help will be appreciate.
Code:
#include <stdio.h>
#include <cutil_inline.h>
float processWithStreams(int streams_used);
#define STREAM_COUNT 2
int N = 1 << 24;
int *h_data_source;
int *h_data_sink;
int *h_data_in[STREAM_COUNT];
int *d_data_in[STREAM_COUNT];
int *h_data_out[STREAM_COUNT];
int *d_data_out[STREAM_COUNT];
cudaEvent_t cycleDone[STREAM_COUNT];
cudaStream_t stream[STREAM_COUNT];
cudaEvent_t start, stop;
dim3 block(512);
dim3 grid;
int memsize;
__global__ void increment_kernel(int *g_data, int inc_value)
{
int idx = blockIdx.x * blockDim.x + threadIdx.x;
//g_data[idx] = g_data[idx] + inc_value;
int i = blockDim.x * gridDim.x;
for(; i > 0; i /= 2)
{
if(idx > i)
g_data[idx]++;
}
}
int main(int argc, char *argv[])
{
if( cutCheckCmdLineFlag(argc, (const char**)argv, "device") )
cutilDeviceInit(argc, argv);
else
cudaSetDevice( cutGetMaxGflopsDeviceId());
h_data_source = (int *)malloc(sizeof(int) * N);
memset(h_data_source, 0, sizeof(int) * N);
int i;
memsize = 1024 * 1024 * sizeof(int);
for(i = 0; i < STREAM_COUNT; i++)
{
CUDA_SAFE_CALL( cudaHostAlloc(&h_data_in[i], memsize, cudaHostAllocDefault) );
CUDA_SAFE_CALL( cudaMalloc(&d_data_in[i], memsize) );
CUDA_SAFE_CALL( cudaHostAlloc(&h_data_out[i], memsize, cudaHostAllocDefault) );
CUDA_SAFE_CALL( cudaMalloc(&d_data_out[i], memsize) );
CUDA_SAFE_CALL( cudaStreamCreate(&stream[i]) );
CUDA_SAFE_CALL( cudaEventCreate(&cycleDone[i]) );
cudaEventRecord(cycleDone[i], stream[i]);
}
CUDA_SAFE_CALL( cudaEventCreate(&start) );
CUDA_SAFE_CALL( cudaEventCreate(&stop) );
grid.x = N / block.x;
grid.y = 1;
float time1 = processWithStreams(STREAM_COUNT);
printf("time: %f\n", time1);
free( h_data_source );
free( h_data_sink );
for( i = 0; i < STREAM_COUNT; ++i ) {
cudaFreeHost(h_data_in[i]);
cudaFree(d_data_in[i]);
cudaStreamDestroy(stream[i]);
cudaEventDestroy(cycleDone[i]);
}
cudaEventDestroy(start);
cudaEventDestroy(stop);
cudaThreadExit();
cutilExit(argc, argv);
return 0;
}
float processWithStreams(int streams_used) {
int current_stream = 0;
float time;
cudaEventRecord(start, 0);
for( int i=0; i < N / 1024 / 1024; ++i ) {
int next_stream = (current_stream + 1 ) % streams_used;
printf("i: %d, current_stream: %d, next_stream: %d\n", i, current_stream, next_stream);
// Ensure that processing and copying of the last cycle has finished
cudaEventSynchronize(cycleDone[next_stream]);
// Process current frame
increment_kernel<<<grid, block, 0, stream[current_stream]>>>(
d_data_in[current_stream], 1);
// Upload next frame
CUDA_SAFE_CALL( cudaMemcpyAsync(
d_data_in[next_stream],
h_data_in[next_stream],
memsize,
cudaMemcpyHostToDevice,
stream[next_stream]) );
CUDA_SAFE_CALL( cudaEventRecord(
cycleDone[next_stream],
stream[next_stream]) );
// Download current frame
CUDA_SAFE_CALL( cudaMemcpyAsync(
h_data_out[current_stream],
d_data_out[current_stream],
memsize,
cudaMemcpyDeviceToHost,
stream[current_stream]) );
CUDA_SAFE_CALL( cudaEventRecord(
cycleDone[current_stream],
stream[current_stream]) );
current_stream = next_stream;
}
cudaEventRecord(stop, 0);
cudaEventElapsedTime(&time, start, stop);
return time;
}
The problem is in your kernel. One thing that happens when checking errors in CUDA is that errors that occurred previously and were not checked will be reported next time you check for an error. That line is the first time you check for errors after the kernel launch which returned the error your are seeing.
The error unspecified launch failure is usually associated with out of bounds accesses to memory if I recall correctly.
You are launching your kernel with 32768 blocks and 512 threads per block. Calculating the idx value for the last thread of the last block we have 32767 * 512 + 511 = 16777215. In the first iteration idx < i and in the following ones you are trying to read and write to position 16777215 of g_data when you only allocated space for 1024 * 1024 integers.
edit: just noticed, why the tag operator overloading?