I want to launch multiple times the following kernel in a FOR LOOP (pseudo):
__global__ void kernel(t_dev is input array in global mem) {
__shared__ PREC tt[BLOCK_DIM];
if (thid < m) {
tt[thid] = t_dev.data[ii]; // MEM READ!
}
... // MODIFY
__syncthreads();
if (thid < m) {
t_dev.data[thid] = tt[thid]; // MEM WRITE!
}
__threadfence(); // or __syncthreads(); //// NECESSARY!! but why?
}
What I do conceptually is I read in values from t_dev . modify them, and write out to global mem again! and then I start the same kernel again!!
Why do I need obviously the _threadfence or __syncthread
otherwise the result get wrong, because, memory writes are not finished when the same kernel starts again. Thats what happens here, my GTX580 has device overlap enabled,
But why are global mem writes not finished when the next kernel starts... is this because of the device overlap or because its always like that? I thought, when we launch kernel after kernel, mem write/reads are finished after one kernel... :-)
Thanks for your answers!
SOME CODE :
for(int kernelAIdx = 0; kernelAIdx < loops; kernelAIdx++){
proxGPU::sorProxContactOrdered_1threads_StepA_kernelWrap<PREC,SorProxSettings1>(
mu_dev,x_new_dev,T_dev,x_old_dev,d_dev,
t_dev,
kernelAIdx,
pConvergedFlag_dev,
m_absTOL,m_relTOL);
proxGPU::sorProx_StepB_kernelWrap<PREC,SorProxSettings1>(
t_dev,
T_dev,
x_new_dev,
kernelAIdx
);
}
These are thw two kernels which are in the loop, the t_dev and x_new_dev, is moved from Step A to Step B,
Kernel A looks as follows:
template<typename PREC, int THREADS_PER_BLOCK, int BLOCK_DIM, int PROX_PACKAGES, typename TConvexSet>
__global__ void sorProxContactOrdered_1threads_StepA_kernel(
utilCuda::Matrix<PREC> mu_dev,
utilCuda::Matrix<PREC> y_dev,
utilCuda::Matrix<PREC> T_dev,
utilCuda::Matrix<PREC> x_old_dev,
utilCuda::Matrix<PREC> d_dev,
utilCuda::Matrix<PREC> t_dev,
int kernelAIdx,
int maxNContacts,
bool * convergedFlag_dev,
PREC _absTOL, PREC _relTOL){
//__threadfence() HERE OR AT THE END; THEN IT WORKS???? WHY
// Assumend 1 Block, with THREADS_PER_BLOCK Threads and Column Major Matrix T_dev
int thid = threadIdx.x;
int m = min(maxNContacts*PROX_PACKAGE_SIZE, BLOCK_DIM); // this is the actual size of the diagonal block!
int i = kernelAIdx * BLOCK_DIM;
int ii = i + thid;
//First copy x_old_dev in shared
__shared__ PREC xx[BLOCK_DIM]; // each thread writes one element, if its in the limit!!
__shared__ PREC tt[BLOCK_DIM];
if(thid < m){
xx[thid] = x_old_dev.data[ii];
tt[thid] = t_dev.data[ii];
}
__syncthreads();
PREC absTOL = _absTOL;
PREC relTOL = _relTOL;
int jj;
//PREC T_iijj;
//Offset the T_dev_ptr to the start of the Block
PREC * T_dev_ptr = PtrElem_ColM(T_dev,i,i);
PREC * mu_dev_ptr = &mu_dev.data[PROX_PACKAGES*kernelAIdx];
__syncthreads();
for(int j_t = 0; j_t < m ; j_t+=PROX_PACKAGE_SIZE){
//Select the number of threads we need!
// Here we process one [m x PROX_PACKAGE_SIZE] Block
// First Normal Direction ==========================================================
jj = i + j_t;
__syncthreads();
if( ii == jj ){ // select thread on the diagonal ...
PREC x_new_n = (d_dev.data[ii] + tt[thid]);
//Prox Normal!
if(x_new_n <= 0.0){
x_new_n = 0.0;
}
/* if( !checkConverged(x_new,xx[thid],absTOL,relTOL)){
*convergedFlag_dev = 0;
}*/
xx[thid] = x_new_n;
tt[thid] = 0.0;
}
// all threads not on the diagonal fall into this sync!
__syncthreads();
// Select only m threads!
if(thid < m){
tt[thid] += T_dev_ptr[thid] * xx[j_t];
}
// ====================================================================================
// wee need to syncronize here because one threads finished lambda_t2 with shared mem tt, which is updated from another thread!
__syncthreads();
// Second Tangential Direction ==========================================================
jj++;
__syncthreads();
if( ii == jj ){ // select thread on diagonal, one thread finishs T1 and T2 directions.
// Prox tangential
PREC lambda_T1 = (d_dev.data[ii] + tt[thid]);
PREC lambda_T2 = (d_dev.data[ii+1] + tt[thid+1]);
PREC radius = (*mu_dev_ptr) * xx[thid-1];
PREC absvalue = sqrt(lambda_T1*lambda_T1 + lambda_T2*lambda_T2);
if(absvalue > radius){
lambda_T1 = (lambda_T1 * radius ) / absvalue;
lambda_T2 = (lambda_T2 * radius ) / absvalue;
}
/*if( !checkConverged(lambda_T1,xx[thid],absTOL,relTOL)){
*convergedFlag_dev = 0;
}
if( !checkConverged(lambda_T2,xx[thid+1],absTOL,relTOL)){
*convergedFlag_dev = 0;
}*/
//Write the two values back!
xx[thid] = lambda_T1;
tt[thid] = 0.0;
xx[thid+1] = lambda_T2;
tt[thid+1] = 0.0;
}
// all threads not on the diagonal fall into this sync!
__syncthreads();
T_dev_ptr = PtrColOffset_ColM(T_dev_ptr,1,T_dev.outerStrideBytes);
__syncthreads();
if(thid < m){
tt[thid] += T_dev_ptr[thid] * xx[j_t+1];
}
__syncthreads();
T_dev_ptr = PtrColOffset_ColM(T_dev_ptr,1,T_dev.outerStrideBytes);
__syncthreads();
if(thid < m){
tt[thid] += T_dev_ptr[thid] * xx[j_t+2];
}
// ====================================================================================
__syncthreads();
// move T_dev_ptr 1 column
T_dev_ptr = PtrColOffset_ColM(T_dev_ptr,1,T_dev.outerStrideBytes);
// move mu_ptr to nex contact
__syncthreads();
mu_dev_ptr = &mu_dev_ptr[1];
__syncthreads();
}
__syncthreads();
// Write back the results, dont need to syncronize because
// do it anyway to be safe for testing first!
if(thid < m){
y_dev.data[ii] = xx[thid]; THIS IS UPDATED IN KERNEL B
t_dev.data[ii] = tt[thid]; THIS IS UPDATED IN KERNEL B
}
//__threadfence(); /// THIS STUPID THREADFENCE MAKES IT WORKING!
I compare the solution at the end with the CPU, and HERE I put everywhere I can a syncthread around only to be safe, for the start! (this code does gauss seidel stuff)
but it does not work at all without the THREAD_FENCE at the END or at the BEGINNIG where it does not make sense...
Sorry for so much code, but probably you can guess where the problem comes, frome because I am bit at my end, with explainig why this happens?
We checked the algorithm several times, there is no memory error (reported from Nsight) or
other stuff, every thing works fine... Kernel A is launched with ONE Block only!
If you launch the successive instances of the kernel into the same stream, each kernel launch is synchronous compared to the kernel instance before and after it. The programming model guarantees it. CUDA only permits simultaneous kernel execution on kernels launched into different streams of the same context, and even then overlapping kernel execution only happens if the scheduler determines that sufficient resources are available to do so.
Neither __threadfence nor __syncthreads will have the effect you seem to be thinking about - __threadfence works only at the scope of all active threads and __syncthreads is an intra-block barrier operation. If you really want kernel to kernel synchronization, you need to use one of the host side synchronization calls, like cudaThreadSynchronize (pre CUDA 4.0) or cudaDeviceSynchronize (cuda 4.0 and later), or the per-stream equivalent if you are using streams.
While I am a bit surprised by what you are experiencing, I believe your explanation may be correct.
Writes to global memory, with an exception of atomic functions, are not guaranteed to be immediately visible by other threads (from the same, or from different blocks). By putting __threadfence() you halt the current thread until the writes are in fact visible. This might be important in particular when you are using global memory with a cache (the Fermi series).
One thing to note: Kernel calls are asynchronous. While your first kernel call is being handled by the GPU, the host may issue another call. The next kernel will not run in parallel with your current one, but will launch as soon as the current one finishes, esentially hiding the latency caused by the CPU->GPU communication.
Using cudaThreadSynchronise halts the host thread until all the CUDA tasks are done. It may help you, but it will also prevent you from hiding the CPU->GPU communication latency. Do note, that using synchronous memory access (e.g. cudaMemcpy, without "Async" suffix) esentially behaves like cudaThreadSynchronise too.
Related
I like to do CUDA synchronization for multiple blocks. It is not for each block where __syncthreads() can easily handle it.
I saw there are exiting discussions on this topic, for example cuda block synchronization, and I like the simple solution brought up by #johan, https://stackoverflow.com/a/67252761/3188690, essentially it uses a 64 bits counter to track the synchronized blocks.
However, I wrote the following code trying to accomplish the similar job but meet a problem. Here I used the term environment so that the wkNumberEnvs of blocks within this environment shall be synchronized. It has a counter. I used atomicAdd() to count how many blocks have already been synchronized themselves, once the number of sync blocks == wkBlocksPerEnv, I know all blocks finished sync and it is free to go. However, it has a strange outcome that I am not sure why.
The problem comes from this while loop. Since the first threads of all blocks are doing the atomicAdd, there is a while loop to check until the condition meets. But I find that some blocks will be stuck into the endless loop, which I am not sure why the condition cannot be met eventually? And if I printf some messages either in *** I can print here 1 or *** I can print here 2, there is no endless loop and everything is perfect. I do not see something obvious.
const int wkBlocksPerEnv = 2;
__device__ int env_sync_block_count[wkNumberEnvs];
__device__ void syncthreads_for_env(){
// sync threads for each block so all threads in this block finished the previous tasks
__syncthreads();
// sync threads for wkBlocksPerEnv blocks for each environment
if(wkBlocksPerEnv > 1){
const int kThisEnvId = get_env_scope_block_id(blockIdx.x);
if (threadIdx.x == 0){
// incrementing env_sync_block_count by 1
atomicAdd(&env_sync_block_count[kThisEnvId], 1);
// *** I can print here 1
while(env_sync_block_count[kThisEnvId] != wkBlocksPerEnv){
// *** I can print here 2
}
// Do the next job ...
}
}
There are two potential issues with your code. Caching and block scheduling.
Caching can prevent you from observing an updated value during the while loop.
Block scheduling can cause a dead-lock if you wait for an update of a block which has not yet been scheduled. Since CUDA does not guarantee a specific order of scheduled blocks, the only way to prevent this dead-lock is to limit the number of blocks in the grid such that all blocks can run simultaneously.
Following code shows how you could synchronize multiple blocks while avoiding above issues. I adapted the code from the multi-grid synchronization given in the CUDA-sample conjugateGradientMultiDeviceCG https://github.com/NVIDIA/cuda-samples/blob/master/Samples/4_CUDA_Libraries/conjugateGradientMultiDeviceCG/conjugateGradientMultiDeviceCG.cu#L186
On pre-Volta devices, it uses volatile memory accesses. Volta and later uses acquire/release semantics.
Grid size is limited by querying device properties.
#include <cassert>
#include <cstdio>
constexpr int wkBlocksPerEnv = 13;
__device__
int getEnv(int blockId){
return blockId / wkBlocksPerEnv;
}
__device__
int getRankInEnv(int blockId){
return blockId % wkBlocksPerEnv;
}
__device__
unsigned char load_arrived(unsigned char *arrived) {
#if __CUDA_ARCH__ < 700
return *(volatile unsigned char *)arrived;
#else
unsigned int result;
asm volatile("ld.acquire.gpu.global.u8 %0, [%1];"
: "=r"(result)
: "l"(arrived)
: "memory");
return result;
#endif
}
__device__
void store_arrived(unsigned char *arrived,
unsigned char val) {
#if __CUDA_ARCH__ < 700
*(volatile unsigned char *)arrived = val;
#else
unsigned int reg_val = val;
asm volatile(
"st.release.gpu.global.u8 [%1], %0;" ::"r"(reg_val) "l"(arrived)
: "memory");
// Avoids compiler warnings from unused variable val.
(void)(reg_val = reg_val);
#endif
}
#if 0
//wrong implementation which does not synchronize. to check that kernel assert does trigger without proper synchronization
__device__
void syncthreads_for_env(unsigned char* temp){
}
#else
//temp must have at least size sizeof(unsigned char) * total_number_of_blocks in grid
__device__
void syncthreads_for_env(unsigned char* temp){
__syncthreads();
const int env = getEnv(blockIdx.x);
const int blockInEnv = getRankInEnv(blockIdx.x);
unsigned char* const mytemp = temp + env * wkBlocksPerEnv;
if(threadIdx.x == 0){
if(blockInEnv == 0){
// Leader block waits for others to join and then releases them.
// Other blocks in env can arrive in any order, so the leader have to wait for
// all others.
for (int i = 0; i < wkBlocksPerEnv - 1; i++) {
while (load_arrived(&mytemp[i]) == 0)
;
}
for (int i = 0; i < wkBlocksPerEnv - 1; i++) {
store_arrived(&mytemp[i], 0);
}
__threadfence();
}else{
// Other blocks in env note their arrival and wait to be released.
store_arrived(&mytemp[blockInEnv - 1], 1);
while (load_arrived(&mytemp[blockInEnv - 1]) == 1)
;
}
}
__syncthreads();
}
#endif
__global__
void kernel(unsigned char* synctemp, int* array){
const int env = getEnv(blockIdx.x);
const int blockInEnv = getRankInEnv(blockIdx.x);
if(threadIdx.x == 0){
array[blockIdx.x] = 1;
}
syncthreads_for_env(synctemp);
if(threadIdx.x == 0){
int sum = 0;
for(int i = 0; i < wkBlocksPerEnv; i++){
sum += array[env * wkBlocksPerEnv + i];
}
assert(sum == wkBlocksPerEnv);
}
}
int main(){
const int smem = 0;
const int blocksize = 128;
int deviceId = 0;
int numSMs = 0;
int maxBlocksPerSM = 0;
cudaGetDevice(&deviceId);
cudaDeviceGetAttribute(&numSMs, cudaDevAttrMultiProcessorCount, deviceId);
cudaOccupancyMaxActiveBlocksPerMultiprocessor(
&maxBlocksPerSM,
kernel,
blocksize,
smem
);
int maxBlocks = maxBlocksPerSM * numSMs;
maxBlocks -= maxBlocks % wkBlocksPerEnv; //round down to nearest multiple of wkBlocksPerEnv
printf("wkBlocksPerEnv %d, maxBlocks: %d\n", wkBlocksPerEnv, maxBlocks);
int* d_array;
unsigned char* d_synctemp;
cudaMalloc(&d_array, sizeof(int) * maxBlocks);
cudaMalloc(&d_synctemp, sizeof(unsigned char) * maxBlocks);
cudaMemset(d_synctemp, 0, sizeof(unsigned char) * maxBlocks);
kernel<<<maxBlocks, blocksize>>>(d_synctemp, d_array);
cudaFree(d_synctemp);
cudaFree(d_array);
return 0;
}
Atomic value is going to global memory but in the while-loop you read it directly and it must be coming from the cache which will not automatically synchronize between threads (cache-coherence only handled by explicit synchronizations like threadfence). Thread gets its own synchronization but other threads may not see it.
Even if you use threadfence, the threads in same warp would be in dead-lock waiting forever if they were the first to check the value before any other thread updates it. But should work with newest GPUs supporting independent thread scheduling.
I like to do CUDA synchronization for multiple blocks.
You should learn to dis-like it. Synchronization is always costly, even when implemented just right, and inter-core synchronization all the more so.
if (threadIdx.x == 0){
// incrementing env_sync_block_count by 1
atomicAdd(&env_sync_block_count[kThisEnvId], 1);
while(env_sync_block_count[kThisEnvId] != wkBlocksPerEnv)
// OH NO!!
{
}
}
This is bad. With this code, the first warp of each block will perform repeated reads of env_sync_block_count[kThisEnvId]. First, and as #AbatorAbetor mentioned, you will face the problem of cache incoherence, causing your blocks to potentially read the wrong value from a local cache well after the global value has long changed.
Also, your blocks will hog up the multiprocessors. Blocks will stay resident and have at least one active warp, indefinitely. Who's to say the will be evicted from their multiprocessor to schedule additional blocks to execute? If I were the GPU, I wouldn't allow more and more active blocks to pile up. Even if you don't deadlock - you'll be wasting a lot of time.
Now, #AbatorAbetor's answer avoids the deadlock by limiting the grid size. And I guess that works. But unless you have a very good reason to write your kernels this way - the real solution is to just break up your algorithm into consecutive kernels (or better yet, figure out how to avoid the need to synchronize altogether).
a mid-way approach is to only have some blocks get past the point of synchronization. You could do that by not waiting except on some condition which holds for a very limited number of blocks (say you had a single workgroup - then only the blocks which got the last K possible counter values, wait).
Suppose I have 8 blocks of 32 threads each running on a GTX 970. Each blcok either writes all 1's or all 0's to an array of length 32 in global memory, where thread 0 in a block writes to position 0 in the array.
Now to write the actual values atomicExch is used, exchanging the current value in the array with the value that the block attempts to write. Because of SIMD, atomic operation and the fact that a warp executes in lockstep I would expect the array to, at any point in time, only contain 1's or 0's. But never a mix of the two.
However, while running code like this there are several cases where at some point in time the array contains of a mix of 0's and 1's. Which appears to point to the fact that atomic operations are not executed per warp, and instead scheduled using some other scheme.
From other sources I have not really found a conclusive write-up detailing the scheduling of atomic operations across different warps (please correct me if I'm wrong), so I was wondering if there is any information on this topic. Since I need to write many small vectors consisting of several 32 bit integers atomically to global memory, and an atomic operation that is guaranteed to write a single vector atomically is obviously very important.
For those wondering, the code I wrote was executed on a GTX 970, compiled on compute capability 5.2, using CUDA 8.0.
The atomic instructions, like all instructions, are scheduled per warp. However there is an unspecified pipeline associated with atomics, and the scheduled instruction flow through the pipeline is not guaranteed to be executed in lockstep, for every thread, for every stage through the pipeline. This gives rise to the possibility for your observations.
I believe a simple thought experiment will demonstrate that this must be true: what if 2 threads in the same warp targeted the same location? Clearly every aspect of the processing could not proceed in lockstep. We could extend this thought experiment to the case where we have multiple issue per clock within an SM and even across SMs, to as additional examples.
If the vector length were short enough (16 bytes or less) then it should be possible to accomplish this ("atomic update") simply by having a thread in a warp write an appropriate vector-type quantity, e.g. int4. As long as all threads (regardless of where they are in the grid) are attempting to update a naturally aligned location, the write should not be corrupted by other writes.
However, after discussion in the comments, it seems that OP's goal is to be able to have a warp or threadblock update a vector of some length, without interference from other warps or threadblocks. It seems to me that really what is desired is access control (so that only one warp or threadblock is updating a particular vector at a time) and OP had some code that wasn't working as desired.
This access control can be enforced using an ordinary atomic operation (atomicCAS in the example below) to permit only one "producer" to update a vector at a time.
What follows is an example producer-consumer code, where there are multiple threadblocks that are updating a range of vectors. Each vector "slot" has a "slot control" variable, which is atomically updated to indicate:
vector is empty
vector is being filled
vector is filled, ready for "consumption"
with this 3-level scheme, we can allow for ordinary access to the vector by both consumer and multiple producer workers, with a single ordinary atomic variable access mechanism. Here is an example code:
#include <assert.h>
#include <iostream>
#include <stdio.h>
const int num_slots = 256;
const int slot_length = 32;
const int max_act = 65536;
const int slot_full = 2;
const int slot_filling = 1;
const int slot_empty = 0;
const int max_sm = 64; // needs to be greater than the maximum number of SMs for any GPU that it will be run on
__device__ int slot_control[num_slots] = {0};
__device__ int slots[num_slots*slot_length];
__device__ int observations[max_sm] = {0}; // reported by consumer
__device__ int actives[max_sm] = {0}; // reported by producers
__device__ int correct = 0;
__device__ int block_id = 0;
__device__ volatile int restricted_sm = -1;
__device__ int num_act = 0;
static __device__ __inline__ int __mysmid(){
int smid;
asm volatile("mov.u32 %0, %%smid;" : "=r"(smid));
return smid;}
// this code won't work on a GPU with a single SM!
__global__ void kernel(){
__shared__ volatile int done, update, next_slot;
int my_block_id = atomicAdd(&block_id, 1);
int my_sm = __mysmid();
if (my_block_id == 0){
if (!threadIdx.x){
restricted_sm = my_sm;
__threadfence();
// I am "block 0" and process the vectors, checking for coherency
// "consumer"
next_slot = 0;
volatile int *vslot_control = slot_control;
volatile int *vslots = slots;
int scount = 0;
while(scount < max_act){
if (vslot_control[next_slot] == slot_full){
scount++;
int slot_val = vslots[next_slot*slot_length];
for (int i = 1; i < slot_length; i++) if (slot_val != vslots[next_slot*slot_length+i]) { assert(0); /* badness - incoherence */}
observations[slot_val]++;
vslot_control[next_slot] = slot_empty;
correct++;
__threadfence();
}
next_slot++;
if (next_slot >= num_slots) next_slot = 0;
}
}}
else {
// "producer"
while (restricted_sm < 0); // wait for signaling
if (my_sm == restricted_sm) return;
next_slot = 0;
done = 0;
__syncthreads();
while (!done) {
if (!threadIdx.x){
while (atomicCAS(slot_control+next_slot, slot_empty, slot_filling) > slot_empty) {
next_slot++;
if (next_slot >= num_slots) next_slot = 0;}
// we grabbed an empty slot, fill it with my_sm
if (atomicAdd(&num_act, 1) < max_act) update = 1;
else {done = 1; update = 0;}
}
__syncthreads();
if (update) slots[next_slot*slot_length+threadIdx.x] = my_sm;
__threadfence(); //enforce ordering
if ((update) && (!threadIdx.x)){
slot_control[next_slot] = 2; // mark slot full
atomicAdd(actives+my_sm, 1);}
__syncthreads();
}
}
}
int main(){
kernel<<<256, slot_length>>>();
cudaDeviceSynchronize();
cudaError_t res= cudaGetLastError();
if (res != cudaSuccess) printf("kernel failure: %d\n", (int)res);
int *h_obs = new int[max_sm];
int *h_act = new int[max_sm];
int h_correct;
cudaMemcpyFromSymbol(h_obs, observations, sizeof(int)*max_sm);
cudaMemcpyFromSymbol(h_act, actives, sizeof(int)*max_sm);
cudaMemcpyFromSymbol(&h_correct, correct, sizeof(int));
int h_total_act = 0;
int h_total_obs = 0;
for (int i = 0; i < max_sm; i++){
std::cout << h_act[i] << "," << h_obs[i] << " ";
h_total_act += h_act[i];
h_total_obs += h_obs[i];}
std::cout << std::endl << h_total_act << "," << h_total_obs << "," << h_correct << std::endl;
}
I don't claim this code to be defect free for any use case. It is advanced to demonstrate the workability of a concept, not as production-ready code. It seems to work for me on linux, on a couple different systems I tested it on. It should not be run on GPUs that have only a single SM, as one SM is reserved for the consumer, and the remaining SMs are used by the producers.
I am currently learning CUDA streams through the computation of a dot product between two vectors. The ingredients are a kernel function that takes in vectors x and y and returns a vector result of size equal to the number of blocks, where each block contributes its own reduced sum.
I also have a host function dot_gpu that calls the kernel and reduces the vector result to the final dot product value.
The synchronous version does just this:
// copy to device
copy_to_device<double>(x_h, x_d, n);
copy_to_device<double>(y_h, y_d, n);
// kernel
double result = dot_gpu(x_d, y_d, n, blockNum, blockSize);
while the async one goes like:
double result[numChunks];
for (int i = 0; i < numChunks; i++) {
int offset = i * chunkSize;
// copy to device
copy_to_device_async<double>(x_h+offset, x_d+offset, chunkSize, stream[i]);
copy_to_device_async<double>(y_h+offset, y_d+offset, chunkSize, stream[i]);
// kernel
result[i] = dot_gpu(x_d+offset, y_d+offset, chunkSize, blockNum, blockSize, stream[i]);
}
for (int i = 0; i < numChunks; i++) {
finalResult += result[i];
cudaStreamDestroy(stream[i]);
}
I am getting worse performance when using streams and was trying to investigate the reasons. I tried to pipeline the downloads, kernel calls and uploads, but with no results.
// accumulate the result of each block into a single value
double dot_gpu(const double *x, const double* y, int n, int blockNum, int blockSize, cudaStream_t stream=NULL)
{
double* result = malloc_device<double>(blockNum);
dot_gpu_kernel<<<blockNum, blockSize, blockSize * sizeof(double), stream>>>(x, y, result, n);
#if ASYNC
double* r = malloc_host_pinned<double>(blockNum);
copy_to_host_async<double>(result, r, blockNum, stream);
CudaEvent copyResult;
copyResult.record(stream);
copyResult.wait();
#else
double* r = malloc_host<double>(blockNum);
copy_to_host<double>(result, r, blockNum);
#endif
double dotProduct = 0.0;
for (int i = 0; i < blockNum; i ++) {
dotProduct += r[i];
}
cudaFree(result);
#if ASYNC
cudaFreeHost(r);
#else
free(r);
#endif
return dotProduct;
}
My guess is that the problem is inside the dot_gpu() functions that doesn't only call the kernel. Tell me if I understand correctly the following stream executions
foreach stream {
cudaMemcpyAsync( device[stream], host[stream], ... stream );
LaunchKernel<<<...stream>>>( ... );
cudaMemcpyAsync( host[stream], device[stream], ... stream );
}
The host executes all the three instructions without being blocked, since cudaMemcpyAsync and kernel return immediately (however on the GPU they will execute sequentially as they are assigned to the same stream). So host goes on to the next stream (even if stream1 who knows what stage it is at, but who cares.. it's doing his job on the GPU, right?) and executes the three instructions again without being blocked.. and so on and so forth. However, my code blocks the host before it can process the next stream, somewhere inside the dot_gpu() function. Is it because I am allocating & freeing stuff, as well as reducing the array returned by the kernel to a single value?
Assuming your objectified CUDA interface does what the function and method names suggest, there are three reasons why work from subsequent calls to dot_gpu() might not overlap:
Your code explicitly blocks by recording an event and waiting for it.
If it weren't blocking for 1. already, your code would block on the pinned host side allocation and deallocation, as you suspected.
If your code weren't blocking for 2. already, work from subsequent calls to dot_gpu() might still not overlap depending on compute capbility. Devices of compute capability 3.0 or lower do not reorder operations even if they are enqueued to different streams.
Even for devices of compute capability 3.5 and higher the number of streams whose operations can be reordered is limited by the CUDA_DEVICE_MAX_CONNECTIONS environment variable, which defaults to 8 and can be set to values as large as 32.
I have a code to calculate primes which I have parallelized using OpenMP:
#pragma omp parallel for private(i,j) reduction(+:pcount) schedule(dynamic)
for (i = sqrt_limit+1; i < limit; i++)
{
check = 1;
for (j = 2; j <= sqrt_limit; j++)
{
if ( !(j&1) && (i&(j-1)) == 0 )
{
check = 0;
break;
}
if ( j&1 && i%j == 0 )
{
check = 0;
break;
}
}
if (check)
pcount++;
}
I am trying to port it to GPU, and I would want to reduce the count as I did for the OpenMP example above. Following is my code, which apart from giving incorrect results is also slower:
__global__ void sieve ( int *flags, int *o_flags, long int sqrootN, long int N)
{
long int gid = blockIdx.x*blockDim.x+threadIdx.x, tid = threadIdx.x, j;
__shared__ int s_flags[NTHREADS];
if (gid > sqrootN && gid < N)
s_flags[tid] = flags[gid];
else
return;
__syncthreads();
s_flags[tid] = 1;
for (j = 2; j <= sqrootN; j++)
{
if ( gid%j == 0 )
{
s_flags[tid] = 0;
break;
}
}
//reduce
for(unsigned int s=1; s < blockDim.x; s*=2)
{
if( tid % (2*s) == 0 )
{
s_flags[tid] += s_flags[tid + s];
}
__syncthreads();
}
//write results of this block to the global memory
if (tid == 0)
o_flags[blockIdx.x] = s_flags[0];
}
First of all, how do I make this kernel fast, I think the bottleneck is the for loop, and I am not sure how to replace it. And next, my counts are not correct. I did change the '%' operator and noticed some benefit.
In the flags array, I have marked the primes from 2 to sqroot(N), in this kernel I am calculating primes from sqroot(N) to N, but I would need to check whether each number in {sqroot(N),N} is divisible by primes in {2,sqroot(N)}. The o_flags array stores the partial sums for each block.
EDIT: Following the suggestion, I modified my code (I understand about the comment on syncthreads now better); I realized that I do not need the flags array and just the global indexes work in my case. What concerns me at this point is the slowness of the code (more than correctness) that could be attributed to the for loop. Also, after a certain data size (100000), the kernel was producing incorrect results for subsequent data sizes. Even for data sizes less than 100000, the GPU reduction results are incorrect (a member in the NVidia forum pointed out that that may be because my data size is not of a power of 2).
So there are still three (may be related) questions -
How could I make this kernel faster? Is it a good idea to use shared memory in my case where I have to loop over each tid?
Why does it produce correct results only for certain data sizes?
How could I modify the reduction?
__global__ void sieve ( int *o_flags, long int sqrootN, long int N )
{
unsigned int gid = blockIdx.x*blockDim.x+threadIdx.x, tid = threadIdx.x;
volatile __shared__ int s_flags[NTHREADS];
s_flags[tid] = 1;
for (unsigned int j=2; j<=sqrootN; j++)
{
if ( gid % j == 0 )
s_flags[tid] = 0;
}
__syncthreads();
//reduce
reduce(s_flags, tid, o_flags);
}
While I profess to know nothing about sieving for primes, there are a host of correctness problems in your GPU version which will stop it from working correctly irrespective of whether the algorithm you are implementing is correct or not:
__syncthreads() calls must be unconditional. It is incorrect to write code where branch divergence could leave some threads within the same warp unable to execute a __syncthreads() call. The underlying PTX is bar.sync and the PTX guide says this:
Barriers are executed on a per-warp basis as if all the threads in a
warp are active. Thus, if any thread in a warp executes a bar
instruction, it is as if all the threads in the warp have executed the
bar instruction. All threads in the warp are stalled until the barrier
completes, and the arrival count for the barrier is incremented by the
warp size (not the number of active threads in the warp). In
conditionally executed code, a bar instruction should only be used if
it is known that all threads evaluate the condition identically (the
warp does not diverge). Since barriers are executed on a per-warp
basis, the optional thread count must be a multiple of the warp size.
Your code unconditionally sets s_flags to one after conditionally loading some values from global memory. Surely that cannot be the intent of the code?
The code lacks a synchronization barrier between the sieving code and the reduction, this can lead to a shared memory race and incorrect results from the reduction.
If you are planning on running this code on a Fermi class card, the shared memory array should be declared volatile to prevent compiler optimization from potentially breaking the shared memory reduction.
If you fix those things, the code might work. Performance is a completely different issue. Certainly on older hardware, the integer modulo operation was very, very slow and not recommended. I can recall reading some material suggesting that Sieve of Atkin was a useful approach to fast prime generation on GPUs.
I had a simple CUDA problem for a class assignment, but the professor added an optional task to implement the same algorithm using shared memory instead. I was unable to finish it before the deadline (as in, the turn-in date was a week ago) but I'm still curious so now I'm going to ask the internet ;).
The basic assignment was to implement a bastardized version of a red-black successive over-relaxation both sequentially and in CUDA, make sure you got the same result in both and then compare the speedup. Like I said, doing it with shared memory was an optional +10% add-on.
I'm going to post my working version and pseudocode what I've attempted to do since I don't have the code in my hands at the moment, but I can update this later with the actual code if someone needs it.
Before anyone says it: Yes, I know using CUtil is lame, but it made the comparison and timers easier.
Working global memory version:
#include <stdlib.h>
#include <stdio.h>
#include <cutil_inline.h>
#define N 1024
__global__ void kernel(int *d_A, int *d_B) {
unsigned int index_x = blockIdx.x * blockDim.x + threadIdx.x;
unsigned int index_y = blockIdx.y * blockDim.y + threadIdx.y;
// map the two 2D indices to a single linear, 1D index
unsigned int grid_width = gridDim.x * blockDim.x;
unsigned int index = index_y * grid_width + index_x;
// check for boundaries and write out the result
if((index_x > 0) && (index_y > 0) && (index_x < N-1) && (index_y < N-1))
d_B[index] = (d_A[index-1]+d_A[index+1]+d_A[index+N]+d_A[index-N])/4;
}
main (int argc, char **argv) {
int A[N][N], B[N][N];
int *d_A, *d_B; // These are the copies of A and B on the GPU
int *h_B; // This is a host copy of the output of B from the GPU
int i, j;
int num_bytes = N * N * sizeof(int);
// Input is randomly generated
for(i=0;i<N;i++) {
for(j=0;j<N;j++) {
A[i][j] = rand()/1795831;
//printf("%d\n",A[i][j]);
}
}
cudaEvent_t start_event0, stop_event0;
float elapsed_time0;
CUDA_SAFE_CALL( cudaEventCreate(&start_event0) );
CUDA_SAFE_CALL( cudaEventCreate(&stop_event0) );
cudaEventRecord(start_event0, 0);
// sequential implementation of main computation
for(i=1;i<N-1;i++) {
for(j=1;j<N-1;j++) {
B[i][j] = (A[i-1][j]+A[i+1][j]+A[i][j-1]+A[i][j+1])/4;
}
}
cudaEventRecord(stop_event0, 0);
cudaEventSynchronize(stop_event0);
CUDA_SAFE_CALL( cudaEventElapsedTime(&elapsed_time0,start_event0, stop_event0) );
h_B = (int *)malloc(num_bytes);
memset(h_B, 0, num_bytes);
//ALLOCATE MEMORY FOR GPU COPIES OF A AND B
cudaMalloc((void**)&d_A, num_bytes);
cudaMalloc((void**)&d_B, num_bytes);
cudaMemset(d_A, 0, num_bytes);
cudaMemset(d_B, 0, num_bytes);
//COPY A TO GPU
cudaMemcpy(d_A, A, num_bytes, cudaMemcpyHostToDevice);
// create CUDA event handles for timing purposes
cudaEvent_t start_event, stop_event;
float elapsed_time;
CUDA_SAFE_CALL( cudaEventCreate(&start_event) );
CUDA_SAFE_CALL( cudaEventCreate(&stop_event) );
cudaEventRecord(start_event, 0);
// TODO: CREATE BLOCKS AND THREADS AND INVOKE GPU KERNEL
dim3 block_size(256,1,1); //values experimentally determined to be fastest
dim3 grid_size;
grid_size.x = N / block_size.x;
grid_size.y = N / block_size.y;
kernel<<<grid_size,block_size>>>(d_A,d_B);
cudaEventRecord(stop_event, 0);
cudaEventSynchronize(stop_event);
CUDA_SAFE_CALL( cudaEventElapsedTime(&elapsed_time,start_event, stop_event) );
//COPY B BACK FROM GPU
cudaMemcpy(h_B, d_B, num_bytes, cudaMemcpyDeviceToHost);
// Verify result is correct
CUTBoolean res = cutComparei( (int *)B, (int *)h_B, N*N);
printf("Test %s\n",(1 == res)?"Passed":"Failed");
printf("Elapsed Time for Sequential: \t%.2f ms\n", elapsed_time0);
printf("Elapsed Time for CUDA:\t%.2f ms\n", elapsed_time);
printf("CUDA Speedup:\t%.2fx\n",(elapsed_time0/elapsed_time));
cudaFree(d_A);
cudaFree(d_B);
free(h_B);
cutilDeviceReset();
}
For the shared memory version, this is what I've tried so far:
#define N 1024
__global__ void kernel(int *d_A, int *d_B, int width) {
//assuming width is 64 because that's the biggest number I can make it
//each MP has 48KB of shared mem, which is 12K ints, 32 threads/warp, so max 375 ints/thread?
__shared__ int A_sh[3][66];
//get x and y index and turn it into linear index
for(i=0; i < width+2; i++) //have to load 2 extra values due to the -1 and +1 in algo
A_sh[index_y%3][i] = d_A[index+i-1]; //so A_sh[index_y%3][0] is actually d_A[index-1]
__syncthreads(); //and hope that previous and next row have been loaded by other threads in the block?
//ignore boundary conditions because it's pseudocode
for(i=0; i < width; i++)
d_B[index+i] = A_sh[index_y%3][i] + A_sh[index_y%3][i+2] + A_sh[index_y%3-1][i+1] + A_sh[index_y%3+1][i+1];
}
main(){
//same init as above until threads/grid init
dim3 threadsperblk(32,16);
dim3 numblks(32,64);
kernel<<<numblks,threadsperblk>>>(d_A,d_B,64);
//rest is the same
}
This shared mem code crashes ("launch failed due to unspecified error") since I haven't caught all the boundary conditions yet, but I'm not worried about that as much as finding the correct way to get things going. I feel that my code is way too complicated to be the correct path (especially compared to the SDK examples), but I also can't see another way to do it since my array doesn't fit into shared mem like all the examples I can find.
And frankly, I'm not sure it would be that much faster on my hardware (GTX 560 Ti - runs the global memory version in 0.121ms), but I need to prove it to myself first :P
Edit 2: For anyone who runs across this in the future, the code in the answer is a good starting point if you want to do some shared memory.
The key to getting the maximum out of these sort of stencil operators in CUDA is data re-usage. I have found that the best approach is usually to have each block "walk" through a dimension of the grid. After the block has loaded an initial tile of data into shared memory, only a single dimension (so row in a row-major order 2D problem ) needs to be read from global memory to have the necessary data in shared memory for the second and subsequent row calculations. The rest of the data can just be reused. To visualise how the shared memory buffer looks through the first four steps of this sort of algorithm:
Three "rows" (a,b,c) of the input grid are loaded to shared memory, and the stencil computed for row (b) and written to global memory
aaaaaaaaaaaaaaaa
bbbbbbbbbbbbbbbb
cccccccccccccccc
Another row (d) is loaded into the shared memory buffer, replacing row (a), and the calculations made for row (c) using a different stencil, reflecting where the row data is in shared memory
dddddddddddddddd
bbbbbbbbbbbbbbbb
cccccccccccccccc
Another row (e) is loaded into the shared memory buffer, replacing row (b), and the calculations made for row (d), using a different stencil from either step 1 or 2.
dddddddddddddddd
eeeeeeeeeeeeeeee
cccccccccccccccc
Another row (f) is loaded into the shared memory buffer, replacing row (c), and the calculations made for row (e). Now the data is back to the same layout as used in step 1, and the same stencil used in step 1 can be used.
dddddddddddddddd
eeeeeeeeeeeeeeee
ffffffffffffffff
The whole cycle repeats until the block has traverse full column length of the input grid. The reason for using different stencils rather than shifting the data in the shared memory buffer is down to performance - shared memory only has about 1000 Gb/s bandwidth on Fermi, and the shifting of data will become a bottleneck in fully optimal code. You should try different buffer sizes, because you might find smaller buffers allows for higher occupancy and improved kernel throughput.
EDIT: To give a concrete example of how that might be implemented:
template<int width>
__device__ void rowfetch(int *in, int *out, int col)
{
*out = *in;
if (col == 1) *(out-1) = *(in-1);
if (col == width) *(out+1) = *(in+1);
}
template<int width>
__global__ operator(int *in, int *out, int nrows, unsigned int lda)
{
// shared buffer holds three rows x (width+2) cols(threads)
__shared__ volatile int buffer [3][2+width];
int colid = threadIdx.x + blockIdx.x * blockDim.x;
int tid = threadIdx.x + 1;
int * rowpos = &in[colid], * outpos = &out[colid];
// load the first three rows (compiler will unroll loop)
for(int i=0; i<3; i++, rowpos+=lda) {
rowfetch<width>(rowpos, &buffer[i][tid], tid);
}
__syncthreads(); // shared memory loaded and all threads ready
int brow = 0; // brow is the next buffer row to load data onto
for(int i=0; i<nrows; i++, rowpos+=lda, outpos+=lda) {
// Do stencil calculations - use the value of brow to determine which
// stencil to use
result = ();
// write result to outpos
*outpos = result;
// Fetch another row
__syncthreads(); // Wait until all threads are done calculating
rowfetch<width>(rowpos, &buffer[brow][tid], tid);
brow = (brow < 2) ? (brow+1) : 0; // Increment or roll brow over
__syncthreads(); // Wait until all threads have updated the buffer
}
}