I'm using cub::DeviceScan functiona and the sample code snippet has a parameter temp_storage_bytes, which it uses to allocate memory (which, incidentally, the code snippet never frees).
The code snippet calls cub::DeviceScan functions with a pointer to NULL memory which triggers it to calculate the required amount of temporary device memory needed for the function, and then returns. The necessary temporary memory is allocated with cudaMalloc, and the function call is repeated pointing to this memory. The temporary memory is then freed with cudaFree (or probably should be).
I'm doing many repetitions of the device scan on different float arrays, but each float array is identical length.
My question is, can I assume that temp_storage_bytes will always be the same value? If so, I can then do a single cudaMalloc and a single cudaFree for many function calls.
The example is unclear on how the required memory is determined and whether it can change for a given array of a given length.
You can assume you need only one call to cub::DeviceScan::InclusiveScan to determine the amount of temporary temp_storage_bytes bytes required if you have repeated calls to cub::DeviceScan::InclusiveScan over different arrays of same length. In the example below, I'm calling several times cub::DeviceScan::InclusiveScan over different arrays of same length and using only one call to cub::DeviceScan::InclusiveScan to determine the amount of temporary size-
// Ensure printing of CUDA runtime errors to console
#define CUB_STDERR
#include <stdio.h>
#include <algorithm> // std::generate
#include <cub/cub.cuh> // or equivalently <cub/device/device_scan.cuh>
#include <thrust\device_vector.h>
#include <thrust\host_vector.h>
void main(void)
{
// Declare, allocate, and initialize device pointers for input and output
int num_items = 7;
thrust::device_vector<int> d_in(num_items);
thrust::device_vector<int> d_out(num_items);
// Determine temporary device storage requirements for inclusive prefix sum
void *d_temp_storage = NULL;
size_t temp_storage_bytes = 0;
cub::DeviceScan::InclusiveSum(d_temp_storage, temp_storage_bytes, d_in.data(), d_out.data(), num_items);
// Allocate temporary storage for inclusive prefix sum
cudaMalloc(&d_temp_storage, temp_storage_bytes);
for (int k=0; k<10; k++) {
thrust::host_vector<int> h_in(num_items);
thrust::host_vector<int> h_out(num_items,0);
std::generate(h_in.begin(), h_in.end(), rand);
d_in = h_in;
// Run inclusive prefix sum
cub::DeviceScan::InclusiveSum(d_temp_storage, temp_storage_bytes, d_in.data(), d_out.data(), num_items);
int difference = 0;
int prev = 0;
for (int i=0; i<num_items; i++) {
h_out[i] = prev + h_in[i];
prev = h_out[i];
int val = d_out[i];
printf("%i %i %i %i\n",i,difference,h_out[i],d_out[i]);
difference = difference + abs(h_out[i] - d_out[i]);
}
if (difference == 0) printf("Test passed!\n");
else printf("A problem occurred!\n");
h_in.shrink_to_fit();
h_out.shrink_to_fit();
}
getchar();
}
Related
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 have a class FPlan that has a number of methods such as permute and packing.
__host__ __device__ void Perturb_action(FPlan *dfp){
dfp->perturb();
dfp->packing();
}
__global__ void Vector_Perturb(FPlan **dfp, int n){
int i=threadIx.x;
if(i<n) Perturb_action(dfp[i]);
}
in main:
FPlan **fp_vec;
fp_vec=(FPlan**)malloc(VEC_SIZE*sizeof(FPlan*));
//initialize the vec
for(int i=0; i<VEC_SIZE;i++)
fp_vec[i]=&fp;
//fp of type FPlan that is initialized
int v_sz=sizeof(fp_vec);
double test=fp_vec[0]->getCost();
printf("the cost before perturb %f\n"test);
FPlan **value;
cudaMalloc(&value,v_sz);
cudaMemcpy(value,&fp_vec,v_sz,cudaMemcpyHostToDevice);
//call kernel
dim3 threadsPerBlock(VEC_SIZE);
dim3 numBlocks(1);
Vector_Perturb<<<numBlocks,threadsPerBlock>>> (value,VEC_SIZE);
cudaMemcpy(fp_vec,value,v_sz,cudaMemcpyDeviceToHost);
test=fp_vec[0]->getCost();
printf("the cost after perturb %f\n"test);
test=fp_vec[1]->getCost();
printf("the cost after perturb %f\n"test);
I am getting before permute for fp_vec[0] printf the cost 0.8.
After permute for fp_vec[0] the value inf and for fp_vec[1] the value 0.8.
The expected output after the permutation should be something like fp_vec[0] = 0.7 and fp_vec[1] = 0.9. I want to apply these permutations to an array of type FPlan.
What am I missing? Is calling an external function supported in CUDA?
This seems to be a common problem these days:
Consider the following code:
#include <stdio.h>
#include <stdlib.h>
int main() {
int* arr = (int*) malloc(100);
printf("sizeof(arr) = %i", sizeof(arr));
return 0;
}
what is the expected ouptut? 100? no its 4 (at least on a 32 bit machine). sizeof() returns the size of the type of a variable not the allocated size of an array.
int v_sz=sizeof(fp_vec);
double test=fp_vec[0]->getCost();
printf("the cost before perturb %f\n"test);
FPlan **value;
cudaMalloc(&value,v_sz);
cudaMemcpy(value,&fp_vec,v_sz,cudaMemcpyHostToDevice);
You are allocating 4 (or 8) bytes on the device and copy 4 (or 8) bytes. The result is undefined (and maybe every time garbage).
Besides that, you shold do proper error checking of your CUDA calls.
Have a look: What is the canonical way to check for errors using the CUDA runtime API?
Here is my kernel code
typedef unsigned char Npp8u;
...
// Kernel Implementation
__device__ unsigned int min_device;
__device__ unsigned int max_device;
__global__ void findMax_Min(Npp8u * data, int numEl){
int index = blockDim.x*blockIdx.x + threadIdx.x;
int shared_index = threadIdx.x;
__shared__ Npp8u data_shared_min[BLOCKDIM];
__shared__ Npp8u data_shared_max[BLOCKDIM];
// check index condition
if(index < numEl){
data_shared_min[shared_index] = data[index]; //pass values from global to shared memory
__syncthreads();
data_shared_max[shared_index] = data[index]; //pass values from global to shared memory
for (unsigned int stride = BLOCKDIM/2; stride > 0; stride >>= 1) {
if(threadIdx.x < stride){
if(data_shared_max[threadIdx.x] < data_shared_max[threadIdx.x+stride]) data_shared_max[shared_index] = data_shared_max[shared_index+stride];
if(data_shared_min[threadIdx.x]> data_shared_min[threadIdx.x+stride]) data_shared_min[shared_index] = data_shared_min[shared_index+stride];
}
__syncthreads();
}
if(threadIdx.x == 0 ){
atomicMin(&(min_device), (unsigned int)data_shared_min[threadIdx.x ]);
//min_device =10;
__syncthreads();
atomicMax(&(max_device), (unsigned int)data_shared_max[threadIdx.x ]);
}
}else{
data_shared_min[shared_index] = 9999;
}
}
I have an image that is 512x512 and I want to find the min and max pixel values. data is the 1-D version of the image. This code works for max but not for min value. As I checked from matlab max value is 202 and min value is 10 but it finds 0 for the min value. Here is my kernel codes and memcpy calls
int main(){
// Host parameter declarations.
Npp8u * imageHost;
int nWidth, nHeight, nMaxGray;
// Load image to the host.
std::cout << "Load PGM file." << std::endl;
imageHost = LoadPGM("lena_before.pgm", nWidth, nHeight, nMaxGray);
// Device parameter declarations.
Npp8u * imageDevice;
unsigned int max, min;
size_t size = sizeof(Npp8u)*nWidth*nHeight;
cudaMalloc((Npp8u**)&imageDevice, size);
cudaMemcpy(imageDevice, imageHost, size, cudaMemcpyHostToDevice);
int numPixels = nWidth*nHeight;
dim3 numThreads(BLOCKDIM);
dim3 numBlocks(numPixels/BLOCKDIM + (numPixels%BLOCKDIM == 0 ? 0 : 1));
findMax_Min<<<numBlocks, numThreads>>>(imageDevice,numPixels);
cudaMemcpyFromSymbol(&max,max_device, sizeof(max_device), 0, cudaMemcpyDeviceToHost);
cudaMemcpyFromSymbol(&min,min_device, sizeof(min_device), 0, cudaMemcpyDeviceToHost);
printf("Min value for image : %i\n", min);
printf("Max value for image : %i\n", max);
...
Another interesting thing is changing the order of cudaMemcpy just after the kernel call also causes malfunctioning and values both are read as zero. I do not see the problem. Is there anyone sees the obstructed part?
You might want to do cuda error checking. You might also want to initialize min_device to a large value and max_device to zero. There are other problems with your reduction method related to stride (what happens in the last block of an odd size image when you add stride to threadIdx.x, it may exceed the defined image range in shared memory) , but I don't think it matters for a 512x512 image. If min_device just happened to start out at zero, all of your atomicMin operations would always leave zero there.
You can try initializing min_device and max_device like this:
__device__ unsigned int min_device = 9999;
__device__ unsigned int max_device = 0;
For the cudamemcpy calls at the end, you are copying 4 bytes (size of max_device) into a one-byte variable (Npp8u max) and likewise for min. So that's a problem. Since you're using pointers, the copy operation is definitely overwriting something that you don't intend. If the compiler stores the variables sequentially the way you have them defined, one copy operation is overwriting the other variable, which I think would explain the behavior you're seeing. If you created min and max as unsigned int quantities, I think this problem would go away.
EDIT: Since you haven't shown your actual block dimensions, it's possible that you still have a problem with your reduction. You might want to change this line:
if(threadIdx.x < stride){
To something like:
if((threadIdx.x < stride) && ((index + stride)< numEl)){
This or something like it should correct the hazard I mention in the first paragraph. I guess you're trying to account for the hazard using this line:
data_shared_min[shared_index] = 9999;
But there's no guarantee that line of code gets executed before the data element that it is setting in shared memory gets read by some other thread. I also don't know what happens when you assign a value of 9999 to a byte quantity, but it's probably not what you expect.
I have an array of doubles stored in GPU global memory and i need to find the maximum value in it. I have read some texts about parallel reduction, so i know that one should divide the array between blocks and make them find their "global maximum", and so on.
But they never seem to address the issue of threads trying to write to the same memory position simultaneously.
Let's say that local_max=0.0 in the beginning of a block execution. Then each thread reads their value from the input vector, decides that is larger than local_max, and then try to write their value to local_max. When all of this happens at the exact same time (atleast when inside the same warp), how can this work and end up with the actual maximum within this block?
I would think either an atomic function or some kind of lock or critical section would be needed, but i haven't seen this addressed in the answers i have found. (ex http://developer.download.nvidia.com/compute/cuda/1_1/Website/projects/reduction/doc/reduction.pdf )
The answer to your questions are contained in the very document you linked to, and the SDK reduction example shows concrete implementations of the reduction concept.
For completeness, here is a concrete example of a reduction kernel:
template <typename T, int BLOCKSIZE>
__global__ reduction(T *inputvals, T *outputvals, int N)
{
__shared__ volatile T data[BLOCKSIZE];
T maxval = inputvals[threadIdx.x];
for(int i=blockDim.x + threadIdx.x; i<N; i+=blockDim.x)
{
maxfunc(maxval, inputvals[i]);
}
data[threadIdx.x] = maxval;
__syncthreads();
// Here maxfunc(a,b) sets a to the minimum of a and b
if (threadIdx.x < 32) {
for(int i=32+threadIdx.x; i < BLOCKSIZE; i+= 32) {
maxfunc(data[threadIdx.x], data[i]);
}
if (threadIdx.x < 16) maxfunc(data[threadIdx.x], data[threadIdx.x+16]);
if (threadIdx.x < 8) maxfunc(data[threadIdx.x], data[threadIdx.x+8]);
if (threadIdx.x < 4) maxfunc(data[threadIdx.x], data[threadIdx.x+4]);
if (threadIdx.x < 2) maxfunc(data[threadIdx.x], data[threadIdx.x+2]);
if (threadIdx.x == 0) {
maxfunc(data[0], data[1]);
outputvals[blockIdx.x] = data[0];
}
}
}
The key point is using the synchronization that is implicit within a warp to perform the reduction in shared memory. The result is a single per-block maximum value. A second reduction pass is required to reduce the set of block maximums to the global maximum (often it is faster to o this on the host). In this example, maxvals is the "compare and set" function which could be as simple as
template<T>
__device__ void maxfunc(T & a, T & b)
{
a = (b > a) ? b : a;
}
Dont' cook your own code, use some thrust (included in version 4.0 of the Cuda sdk) :
#include <thrust/device_vector.h>
#include <thrust/sequence.h>
#include <thrust/copy.h>
#include <iostream>
int main(void)
{
thrust::host_vector<int> h_vec(10000);
thrust::sequence(h_vec.begin(), h_vec.end());
// show hvec
thrust::copy(h_vec.begin(), h_vec.end(),
std::ostream_iterator<int>(std::cout, "\n"));
// transfer to device
thrust::device_vector<int> d_vec = h_vec;
int max_dvec_value = *thrust::max_element(d_vec.begin(), d_vec.end());
std::cout << "max value: " << max_dvec_value << "\n";
return 0;
}
And watch out that thrust::max_element returns a pointer.
Your question is clearly answered in the document you link to. I think you just need to spend some more time reading it and understanding the CUDA concepts used in it. In particular, I would focus on shared memory, the __syncthreads() method, and how to uniquely identify a thread while inside a kernel. Additionally, you should try to understand why the reduction may need to be run in 2 passes to find the global maximum.
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
}
}