CUDA Maximum Reduction Algorithm Not Working - algorithm

A previous question asked how to find to find the maximum value of an array in CUDA efficiently: Finding max value in CUDA, the top response provided a link to a NVIDIA presentation on optimizing reduction kernels.
If you are using Visual Studio, simply remove the header reference, and everything between CPU EXECUTION.
I setup a variant which found the max, but it doesn't match what the CPU is finding:
// Returns the maximum value of
// an array of size n
float GetMax(float *maxes, int n)
{
int i = 0;
float max = -100000;
for(i = 0; i < n; i++)
{
if(maxes[i] > max)
max = maxes[i];
}
return max;
}
// Too obvious...
__device__ float MaxOf2(float a, float b)
{
if(a > b) return a;
else return b;
}
__global__ void MaxReduction(int n, float *g_idata, float *g_odata)
{
extern __shared__ float sdata[];
unsigned int tid = threadIdx.x;
unsigned int i = blockIdx.x*(BLOCKSIZE*2) + tid;
unsigned int gridSize = BLOCKSIZE*2*gridDim.x;
sdata[tid] = 0;
//MMX(index,i)
//MMX(index,i+blockSize)
// Final Optimized Kernel
while (i < n) {
sdata[tid] = MaxOf2(g_idata[i], g_idata[i+BLOCKSIZE]);
i += gridSize;
}
__syncthreads();
if (BLOCKSIZE >= 512) { if (tid < 256) { sdata[tid] = MaxOf2(sdata[tid], sdata[tid + 256]); } __syncthreads(); }
if (BLOCKSIZE >= 256) { if (tid < 128) { sdata[tid] = MaxOf2(sdata[tid], sdata[tid + 128]); } __syncthreads(); }
if (BLOCKSIZE >= 128) { if (tid < 64) { sdata[tid] = MaxOf2(sdata[tid], sdata[tid + 64]); } __syncthreads(); }
if (tid < 32) {
if (BLOCKSIZE >= 64) sdata[tid] = MaxOf2(sdata[tid], sdata[tid + 32]);
if (BLOCKSIZE >= 32) sdata[tid] = MaxOf2(sdata[tid], sdata[tid + 16]);
if (BLOCKSIZE >= 16 ) sdata[tid] = MaxOf2(sdata[tid], sdata[tid + 8]);
if (BLOCKSIZE >= 8) sdata[tid] = MaxOf2(sdata[tid], sdata[tid + 4]);
if (BLOCKSIZE >= 4) sdata[tid] = MaxOf2(sdata[tid], sdata[tid + 2]);
if (BLOCKSIZE >= 2) sdata[tid] = MaxOf2(sdata[tid], sdata[tid + 1]);
}
if (tid == 0) g_odata[blockIdx.x] = sdata[0];
}
I have a giant setup to test this algorithm:
#include <cstdio>
#include <cstdlib>
#include <ctime>
#include <iostream>
#include <sys/time.h>
#include <cuda.h>
#include <cuda_runtime.h>
#include <device_launch_parameters.h>
#include "book.h"
#define ARRAYSIZE 16384
#define GRIDSIZE 60
#define BLOCKSIZE 32
#define SIZEFLOAT 4
using namespace std;
// Function definitions
float GetMax(float *maxes, int n);
__device__ float MaxOf2(float a, float b);
__global__ void MaxReduction(int n, float *g_idata, float *g_odata);
// Returns random floating point number
float RandomReal(float low, float high)
{
float d;
d = (float) rand() / ((float) RAND_MAX + 1);
return (low + d * (high - low));
}
int main()
{
/*****************VARIABLE SETUP*****************/
// Pointer to CPU numbers
float *numbers;
// Pointer to GPU numbers
float *dev_numbers;
// Counter
int i = 0;
// Randomize
srand(time(0));
// Timers
// Kernel timers
cudaEvent_t start_kernel, stop_kernel;
float elapsedTime_kernel;
HANDLE_ERROR(cudaEventCreate(&start_kernel));
HANDLE_ERROR(cudaEventCreate(&stop_kernel));
// cudaMalloc timers
cudaEvent_t start_malloc, stop_malloc;
float elapsedTime_malloc;
HANDLE_ERROR(cudaEventCreate(&start_malloc));
HANDLE_ERROR(cudaEventCreate(&stop_malloc));
// CPU timers
struct timeval start, stop;
float elapsedTime = 0;
/*****************VARIABLE SETUP*****************/
/*****************CPU ARRAY SETUP*****************/
// Setup CPU array
HANDLE_ERROR(cudaHostAlloc((void**)&numbers, ARRAYSIZE * sizeof(float), cudaHostAllocDefault));
for(i = 0; i < ARRAYSIZE; i++)
numbers[i] = RandomReal(0, 50000.0);
/*****************CPU ARRAY SETUP*****************/
/*****************GPU ARRAY SETUP*****************/
// Start recording cuda malloc time
HANDLE_ERROR(cudaEventRecord(start_malloc,0));
// Allocate memory to GPU
HANDLE_ERROR(cudaMalloc((void**)&dev_numbers, ARRAYSIZE * sizeof(float)));
// Transfer CPU array to GPU
HANDLE_ERROR(cudaMemcpy(dev_numbers, numbers, ARRAYSIZE*sizeof(float), cudaMemcpyHostToDevice));
// An array to temporarily store maximum values on the GPU
float *dev_max;
HANDLE_ERROR(cudaMalloc((void**)&dev_max, GRIDSIZE * sizeof(float)));
// An array to hold grab the GPU max
float maxes[GRIDSIZE];
/*****************GPU ARRAY SETUP*****************/
/*****************KERNEL EXECUTION*****************/
// Start recording kernel execution time
HANDLE_ERROR(cudaEventRecord(start_kernel,0));
// Run kernel
MaxReduction<<<GRIDSIZE, BLOCKSIZE, SIZEFLOAT*BLOCKSIZE>>> (ARRAYSIZE, dev_numbers, dev_max);
// Transfer maxes over
HANDLE_ERROR(cudaMemcpy(maxes, dev_max, GRIDSIZE * sizeof(float), cudaMemcpyDeviceToHost));
// Print out the max
cout << GetMax(maxes, GRIDSIZE) << endl;
// Stop recording kernel execution time
HANDLE_ERROR(cudaEventRecord(stop_kernel,0));
HANDLE_ERROR(cudaEventSynchronize(stop_kernel));
// Retrieve recording data
HANDLE_ERROR(cudaEventElapsedTime(&elapsedTime_kernel, start_kernel, stop_kernel));
// Stop recording cuda malloc time
HANDLE_ERROR(cudaEventRecord(stop_malloc,0));
HANDLE_ERROR(cudaEventSynchronize(stop_malloc));
// Retrieve recording data
HANDLE_ERROR(cudaEventElapsedTime(&elapsedTime_malloc, start_malloc, stop_malloc));
// Print results
printf("%5.3f\t%5.3f\n", elapsedTime_kernel, elapsedTime_malloc);
/*****************KERNEL EXECUTION*****************/
/*****************CPU EXECUTION*****************/
// Capture the start time
gettimeofday(&start, NULL);
// Call generic P7Viterbi function
cout << GetMax(numbers, ARRAYSIZE) << endl;
// Capture the stop time
gettimeofday(&stop, NULL);
// Retrieve time elapsed in milliseconds
long int elapsed_sec = stop.tv_sec - start.tv_sec;
long int elapsed_usec = stop.tv_usec - start.tv_usec;
elapsedTime = (float)(1000.0f * elapsed_sec) + (float)(0.001f * elapsed_usec);
// Print results
printf("%5.3f\n", elapsedTime);
/*****************CPU EXECUTION*****************/
// Free memory
cudaFreeHost(numbers);
cudaFree(dev_numbers);
cudaFree(dev_max);
cudaEventDestroy(start_kernel);
cudaEventDestroy(stop_kernel);
cudaEventDestroy(start_malloc);
cudaEventDestroy(stop_malloc);
// Exit program
return 0;
}
I ran cuda-memcheck on this test program, with -g & -G switches on, and it reports 0 problems. Can anyone spot the issue?
NOTE: Be sure to have book.h from the CUDA by Example book in your current directory when you compile the program. Source link here: http://developer.nvidia.com/cuda-example-introduction-general-purpose-gpu-programming
Download the source code, and book.h will be under the common directory/folder.

Your kernel looks broken to me. The thread local search (before the shared memory reduction), should look something like this:
sdata[tid] = g_idata[i];
i += gridSize;
while (i < n) {
sdata[tid] = MaxOf2(sdata[tid], g_idata[i]);
i += gridSize;
}
shouldn't it?
Also note that if you run this on Fermi, the shared memory buffer should be declared volatile, and you will get a noticeable improvement in performance if the thread local search is done with a register variable, rather than in shared memory. There is about an 8 times difference in effective bandwidth between the two.
EDIT: Here is a simplified, working version of your reduction kernel. You should note a number of differences compared to your original:
__global__ void MaxReduction(int n, float *g_idata, float *g_odata)
{
extern __shared__ volatile float sdata[];
unsigned int tid = threadIdx.x;
unsigned int i = blockIdx.x*(BLOCKSIZE) + tid;
unsigned int gridSize = BLOCKSIZE*gridDim.x;
float val = g_idata[i];
i += gridSize;
while (i < n) {
val = MaxOf2(g_idata[i],val);
i += gridSize;
}
sdata[tid] = val;
__syncthreads();
// This versions uses a single warp for the shared memory
// reduction
# pragma unroll
for(int i=(tid+32); ((tid<32)&&(i<BLOCKSIZE)); i+=32)
sdata[tid] = MaxOf2(sdata[tid], sdata[i]);
if (tid < 16) sdata[tid] = MaxOf2(sdata[tid], sdata[tid + 16]);
if (tid < 8) sdata[tid] = MaxOf2(sdata[tid], sdata[tid + 8]);
if (tid < 4) sdata[tid] = MaxOf2(sdata[tid], sdata[tid + 4]);
if (tid < 2) sdata[tid] = MaxOf2(sdata[tid], sdata[tid + 2]);
if (tid == 0) g_odata[blockIdx.x] = MaxOf2(sdata[tid], sdata[tid + 1]);
}
This code should also be safe on Fermi. You should also familiarise yourself with the CUDA math library, because there is a fmax(x,y) intrinsic which you should use in place of your MaxOf2 function.

Related

CUDA - Parallel Reduction Sum of Even and Odd Number Separately

I am trying to implement a parallel reduction sum of even and odd number Separately in CUDA.
I'm new in CUDA programming and I'm trying so hard but I can't find a solution.
I have for example the array : [5, 8, 0, -6, 2]. And the result need to be [4, 5] (Even : 8+0-6+2=4, Odd : 5=5).
But the result of my following code is [8, 5].
I think that my problem is in the notion of "shared" but I do not understand why.
__global__ void sumEvenOdd(int *a, int *b, int N){
int column = blockIdx.x * blockIdx.x + threadIdx.x;
__shared__ int s_data[2];
if (column < N){
if (a[column] % 2 == 0){
s_data[0] += a[column];
}
else{
s_data[1] += a[column];
}
__syncthreads();
b[0] = s_data[0];
b[1] = s_data[1];
}
}
void initArray(int *a, int N){
for (unsigned int i = 0; i < N; i++){
a[i] = rand() % 100;
}
}
void verify_result(int *a, int *b, int N){
int *verify_b;
verify_b = (int*)malloc(2 * sizeof(int));
verify_b[0] = 0;
verify_b[1] = 0;
for (unsigned int i = 0; i < N; i++){
if (a[i] % 2 == 0){
verify_b[0] += a[i];
}
else{
verify_b[1] += a[i];
}
}
for (unsigned int i = 0; i < 2; i++){
assert(verify_b[i] == b[i]);
}
}
void printResult(int *a, int *b, int N){
printf("\n");
for (unsigned int i = 0; i < N; i++){
printf("%d, ", a[i]);
}
printf("\n");
for (unsigned int i = 0; i < 2; i++){
printf("%d, ", b[i]);
}
}
int main(){
//Array sizes;
int N = 5;
//Size (in bytes) of matrix
size_t bytes = N * sizeof(int);
//Host pointers
int *a, *b;
// Allocate host memory
a = (int*)malloc(bytes);
b = (int*)malloc(2 * sizeof(int));
// Initialize array
initArray(a, N);
// Device pointers
int *d_a, *d_b;
// Allocated device memory
cudaMalloc(&d_a, bytes);
cudaMalloc(&d_b, 2 * sizeof(int));
// Copy data to the device
cudaMemcpy(d_a, a, bytes, cudaMemcpyHostToDevice);
//Number of threads
int THREADS = 128;
//Number of blocks
int BLOCKS = (N + THREADS - 1) / THREADS;
// Launch kernel
sumEvenOdd<<<BLOCKS, THREADS>>>(d_a, d_b, N);
cudaDeviceSynchronize();
// Copy back to the host
cudaMemcpy(b, d_b, 2 * sizeof(int), cudaMemcpyDeviceToHost);
// Check result
verify_result(a, b, N);
printResult(a, b, N);
return 0;
}
you cannot just use
s_data[1] += a[column];
remember all units are going to execute this line at the same time, and store in the same position, so all threads are storing into s_data at the same time.
instead you should use atomic add
atomicAdd(&s_data[1], a[column]);
and you should also be initializing s_data to zeros.

Some CUDA computations fail with larger block dimension (< 1024)

I am learning CUDA with a GTX 960 4GB. I wrote a program which performs an element-wise matrix multiplication. When I increase the block dimensions for x and y to lets say (32 x 32) in combination with a large matrix (lets say 15000 x 15000 elements), some but not all multiplication results are wrong (value 0 instead of 6).
When I then decrease the block dimensions to e.g (8 x 8), all results are right again. When I decrease the Matrix size, the results are right again, too.
So in case of this example, there seems to be combinations of total threads and threads per block, which does not work.
I am surprised I can't find any threads regarding this topic. All I can find is about increasing performance and occupancy, but not about when some but not all calculations are aborted.
The grid dimensions are calculated as follows:
dim3 blocks(ceil<int>(COLS / threads.x), ceil<int>(ROWS / threads.y));
Why do some multiplications fail while others are successful?
Some Examples
Block dim : (8, 8)
Matrix shape : (15000, 15000)
Verification : 0 elements have failed, total length 225000000, shape: (15000, 15000)
Block dim : (16, 16)
Matrix shape : (15000, 15000)
Verification : 239936 elements have failed, total length 225000000, shape: (15000, 15000)
Block dim : (32, 32)
Matrix shape : (15000, 15000)
Verification : 719424 elements have failed, total length 225000000, shape: (15000, 15000).
Block dim : (32, 32)
Matrix shape : (10000, 10000)
Verification : 0 elements have failed, total length 100000000, shape: (10000, 10000).
Driver Version
$ cat /proc/driver/nvidia/version
NVRM version: NVIDIA UNIX x86_64 Kernel Module 470.82.00 Thu Oct 14 10:24:40 UTC 2021
Complete Code
#include <stdio.h>
#include <stdlib.h>
#include <math.h>
#include <assert.h>
#include <cuda.h>
#include <cuda_runtime.h>
#define ROWS 10000
#define COLS 10000
#define MAX_ERR 1e-6
typedef struct {
int width;
int height;
float* elements;
} Matrix;
size_t ij(int i, int j){
return j * ROWS + i;
}
__global__ void matrix_multi_elemwise(const Matrix OUT, const Matrix A, const Matrix B) {
int col = blockIdx.x * blockDim.x + threadIdx.x;
int row = blockIdx.y * blockDim.y + threadIdx.y;
if (col < A.width && row < A.height) {
int index = row * A.height + col; // linearisation of index
OUT.elements[index] = A.elements[index] * B.elements[index];
}
}
int main(){
Matrix A, B, OUT;
Matrix dev_A, dev_B, dev_OUT;
size_t SIZE = ROWS * COLS * sizeof(float);
// Allocate host memory
A.elements = (float*) malloc(SIZE);
B.elements = (float*) malloc(SIZE);
OUT.elements = (float*) malloc(SIZE);
// Initialize host matrices
A.height = ROWS; A.width = COLS;
B.height = ROWS; B.width = COLS;
OUT.height = ROWS; OUT.width = COLS;
for (int j = 0; j < ROWS; j++) {
for(int i = 0; i < COLS; i++){
A.elements[ij(i, j)] = 2.0f;
B.elements[ij(i, j)] = 3.0f;
}
}
// Allocate device memory
cudaMalloc((void**) &dev_A.elements, SIZE);
cudaMalloc((void**) &dev_B.elements, SIZE);
cudaMalloc((void**) &dev_OUT.elements, SIZE);
dev_A.height = A.height; dev_A.width = A.width;
dev_B.height = A.height; dev_B.width = B.width;
dev_OUT.height = A.height; dev_OUT.width = OUT.width;
// Transfer data from host to device memory
cudaMemcpy(dev_A.elements, A.elements, SIZE, cudaMemcpyHostToDevice);
cudaMemcpy(dev_B.elements, B.elements, SIZE, cudaMemcpyHostToDevice);
// Executing kernel
dim3 threads(16, 16);
dim3 blocks(ceil<int>(COLS / threads.x), ceil<int>(ROWS / threads.y));
matrix_multi_elemwise<<<blocks, threads>>>(dev_OUT, dev_A, dev_B);
cudaError_t err = cudaGetLastError();
if(err != cudaSuccess) {
printf("CUDA Runtime API Error reported : %s in file %s on line.\n", cudaGetErrorString(err), __FILE__);
}
// Wait for GPU to finish before accessing on host
cudaDeviceSynchronize();
// Transfer data back to host memory
cudaMemcpy(OUT.elements, dev_OUT.elements, SIZE, cudaMemcpyDeviceToHost);
// Verification
int count = 0, length = 0, i = 0, j = 0;
for (j = 0; j < ROWS; j++) {
for(i = 0; i < COLS; i++){
//assert(fabs(OUT.elements[ij(i, j)] / A.elements[ij(i, j)] - B.elements[ij(i, j)]) < MAX_ERR);
if (fabs(OUT.elements[ij(i, j)] / A.elements[ij(i, j)] - B.elements[ij(i, j)]) > MAX_ERR) {
count++;
}
length++;
}
}
printf("Verification: %i elements have failed, total length %i, shape: (%i, %i).\n", count, length, i, j);
// Deallocate device memory
cudaFree(dev_A.elements);
cudaFree(dev_B.elements);
cudaFree(dev_OUT.elements);
// Deallocate host memory
free(A.elements);
free(B.elements);
free(OUT.elements);
}
The number of blocks is wrong. Indeed, COLS and threads.x are both integers. Thus, the result is a truncated integer. ceil<int> cannot ceil the result as it has already been truncated. This cause some blocks not to be computed: 15000 is divisible by 8 but not by 16. You need to either cast COLS to a floating-point number or compute the ceil result manually (safer). Here is an example:
dim3 blocks((COLS + threads.x - 1) / threads.x, (ROWS + threads.y - 1) / threads.y);
As pointed out in the comment, note that row * A.height + col is wrong: it should be row * A.width + col instead. This causes issues for non square matrices.

square Matrix transpose with CUDA

I'm trying to write the matrix transpose algorithm. I test this program with matrix size equal to 1024, the result shows that not all elements are in the right places.
Why isn't my array transposing correctly? Does anyone can help me or give me any hint? I will appreciate it. Thanks a lot!
there is the whole cpu code:
__global__ void transpose_naive (float *out, float *in, int w, int h )
{
unsigned int xIdx = blockDim.x * blockIdx.x + threadIdx.x;
unsigned int yIdx = blockDim.y * blockIdx.y + threadIdx.y;
if ( xIdx <=w && yIdx <=h ) {
unsigned int idx_in = xIdx + w * yIdx;
unsigned int idx_out = yIdx + h * xIdx;
out[idx_out] = in[idx_in];
}
}
int main()
{
int nx=1024;
int mem_size = nx*nx*sizeof(float);
int t=32;
dim3 dimGrid(((nx-1)/t) +1, ((nx-1)/t) +1);
dim3 dimBlock(t,t);
float *h_idata = (float*)malloc(mem_size);
float *h_cdata = (float*)malloc(mem_size);
float *d_idata, *d_cdata;
checkCuda(cudaMalloc(&d_idata, mem_size) );
checkCuda(cudaMalloc(&d_cdata, mem_size) );
// host
for (int j = 0; j < nx; j++)
for (int i = 0; i < nx; i++)
h_idata[j*nx + i] = j*nx + i;
// device
checkCuda(cudaMemcpy(d_idata, h_idata, mem_size, cudaMemcpyHostToDevice) );
// events for timing
cudaEvent_t startEvent, stopEvent;
checkCuda(cudaEventCreate(&startEvent) );
checkCuda(cudaEventCreate(&stopEvent) );
float ms;
checkCuda( cudaEventRecord(startEvent, 0) );
transpose_naive<<<dimGrid, dimBlock>>>(d_cdata, d_idata,nx,nx);
checkCuda(cudaEventRecord(stopEvent, 0) );
checkCuda(cudaEventSynchronize(stopEvent) );
checkCuda(cudaEventElapsedTime(&ms, startEvent, stopEvent) );
checkCuda( cudaMemcpy(h_cdata, d_cdata, mem_size, cudaMemcpyDeviceToHost) );
printf("the time %5f ", ms);
printf("\n");
savetofile(h_idata,"i.txt",nx,nx);
savetofile(h_cdata,"t.txt",nx,nx);
error_exit:
// cleanup
checkCuda(cudaEventDestroy(startEvent) );
checkCuda(cudaEventDestroy(stopEvent) );
checkCuda( cudaFree(d_cdata) );
checkCuda( cudaFree(d_idata) );
free(h_idata);
free(h_cdata);
system("pause");
}
I think there is something wrong with file output "i.txt" and "t.txt" otherwise the program looks to be correct. I have made some minor changes in your code by adding error checking and printing on the standard output stream. I am printing the last (1020 - 1024) 3 x 3 matrix to cross check the transpose. Run it on your system and verify whether the matrix transpose is correct or not?
#include "cuda_runtime.h"
#include <stdio.h>
#include <stdlib.h>
#include "device_launch_parameters.h"
#define gpuErrchk(ans) { gpuAssert((ans), __FILE__, __LINE__); }
inline void gpuAssert(cudaError_t code, const char *file, int line, bool abort = true)
{
if (code != cudaSuccess)
{
fprintf(stderr, "GPUassert: %s %s %d\n", cudaGetErrorString(code),file, line);
if (abort) exit(code);
}
}
__global__ void transpose_naive(float *out, float *in, int w, int h)
{
unsigned int xIdx = blockDim.x * blockIdx.x + threadIdx.x;
unsigned int yIdx = blockDim.y * blockIdx.y + threadIdx.y;
if (xIdx <= w && yIdx <= h) {
unsigned int idx_in = xIdx + w * yIdx;
unsigned int idx_out = yIdx + h * xIdx;
out[idx_out] = in[idx_in];
}
}
int main()
{
int nx = 1024;
int mem_size = nx*nx*sizeof(float);
int t = 32;
dim3 dimGrid(((nx - 1) / t) + 1, (((nx - 1) / t) + 1));
dim3 dimBlock(t, t);
float *h_idata = (float*)malloc(mem_size);
float *h_cdata = (float*)malloc(mem_size);
float *d_idata, *d_cdata;
gpuErrchk(cudaMalloc(&d_idata, mem_size));
gpuErrchk(cudaMalloc(&d_cdata, mem_size));
// host
for (int j = 0; j < nx; j++)
for (int i = 0; i < nx; i++)
h_idata[j*nx + i] = j*nx + i;
// device
gpuErrchk(cudaMemcpy(d_idata,h_idata,mem_size,cudaMemcpyHostToDevice));
// events for timing
cudaEvent_t startEvent, stopEvent;
gpuErrchk(cudaEventCreate(&startEvent));
gpuErrchk(cudaEventCreate(&stopEvent));
float ms;
gpuErrchk(cudaEventRecord(startEvent, 0));
transpose_naive << <dimGrid, dimBlock >> >(d_cdata, d_idata, nx, nx);
gpuErrchk(cudaEventRecord(stopEvent, 0));
gpuErrchk(cudaEventSynchronize(stopEvent));
gpuErrchk(cudaEventElapsedTime(&ms, startEvent, stopEvent));
gpuErrchk(cudaMemcpy(h_cdata,d_cdata,mem_size,cudaMemcpyDeviceToHost));
printf("the time %5f ", ms);
printf("\n");
for (int i = 1020; i < 1024; i++) {
for (int j = 1020; j < 1024; j++) {
printf("%.2f ", h_idata[i*nx + j]);
}
printf("\n");
}
printf("\n");
for (int i = 1020; i < 1024; i++) {
for (int j = 1020; j < 1024; j++) {
printf("%.2f ", h_cdata[i*nx + j]);
}
printf("\n");
}
//savetofile(h_idata, "i.txt", nx, nx);
//savetofile(h_cdata, "t.txt", nx, nx);
//error_exit:
// cleanup
gpuErrchk(cudaEventDestroy(startEvent));
gpuErrchk(cudaEventDestroy(stopEvent));
gpuErrchk(cudaFree(d_cdata));
gpuErrchk(cudaFree(d_idata));
free(h_idata);
free(h_cdata);
//system("pause");
}
The only flaw in the code is the incorrect bound checks in the following line of the kernel.
if ( xIdx <=w && yIdx <=h ) {
As the indices are from 0 to w-1 and 0 to h-1 for x and y dimensions respectively, the if condition should be as follows:
if ( xIdx <w && yIdx <h ) {

FIR filter in CUDA (as a 1D convolution)

I'm trying to implement a FIR (Finite Impulse Response) filter in CUDA. My approach is quite simple and looks somewhat like this:
#include <cuda.h>
__global__ void filterData(const float *d_data,
const float *d_numerator,
float *d_filteredData,
const int numeratorLength,
const int filteredDataLength)
{
int i = blockDim.x * blockIdx.x + threadIdx.x;
float sum = 0.0f;
if (i < filteredDataLength)
{
for (int j = 0; j < numeratorLength; j++)
{
// The first (numeratorLength-1) elements contain the filter state
sum += d_numerator[j] * d_data[i + numeratorLength - j - 1];
}
}
d_filteredData[i] = sum;
}
int main(void)
{
// (Skipping error checks to make code more readable)
int dataLength = 18042;
int filteredDataLength = 16384;
int numeratorLength= 1659;
// Pointers to data, filtered data and filter coefficients
// (Skipping how these are read into the arrays)
float *h_data = new float[dataLength];
float *h_filteredData = new float[filteredDataLength];
float *h_filter = new float[numeratorLength];
// Create device pointers
float *d_data = nullptr;
cudaMalloc((void **)&d_data, dataLength * sizeof(float));
float *d_numerator = nullptr;
cudaMalloc((void **)&d_numerator, numeratorLength * sizeof(float));
float *d_filteredData = nullptr;
cudaMalloc((void **)&d_filteredData, filteredDataLength * sizeof(float));
// Copy data to device
cudaMemcpy(d_data, h_data, dataLength * sizeof(float), cudaMemcpyHostToDevice);
cudaMemcpy(d_numerator, h_numerator, numeratorLength * sizeof(float), cudaMemcpyHostToDevice);
// Launch the kernel
int threadsPerBlock = 256;
int blocksPerGrid = (filteredDataLength + threadsPerBlock - 1) / threadsPerBlock;
filterData<<<blocksPerGrid,threadsPerBlock>>>(d_data, d_numerator, d_filteredData, numeratorLength, filteredDataLength);
// Copy results to host
cudaMemcpy(h_filteredData, d_filteredData, filteredDataLength * sizeof(float), cudaMemcpyDeviceToHost);
// Clean up
cudaFree(d_data);
cudaFree(d_numerator);
cudaFree(d_filteredData);
// Do stuff with h_filteredData...
// Clean up some more
delete [] h_data;
delete [] h_filteredData;
delete [] h_filter;
}
The filter works, but as I'm new to CUDA programming and I'm not sure how to optimize it.
A slight problem that I see is that dataLength, filteredDataLength, and numeratorLength are not known before hand in the application I intend to use the filter in. Also, even though dataLength is a multiple of 32 in the above code, it is not guaranteed to be that in the final application.
When I compare my code above to ArrayFire, my code takes about three times longer to execute.
Does anyone have any ideas on how to speed things up?
EDIT: Have changed all filterLength to numeratorLength.
I can suggest the following to speed up your code:
Use the shared memory: it is a tiny cache-like memory but extremely
faster than the global card memory. You can find more about it by
looking for __shared__ keyword in CUDA documentation. For
example, you can pre-fetch the filter numerators and big chunks
of data in shared memory, this will significantly enhance your
performance. You need to pay extra attention to the data
alignment in this case as it really matters and it can slow down
your code.
Think about unrolling the for-loop of the numerator
sum. You can check the reduce-vector example in CUDA
documentation.
You can also think about parallelizing the
numerator loop itself by itself. This can be done by adding an extra dimension (say 'y') to your thread-block. You will need to make sum a shared vector as well that has the dimension of numeratorLength. You can also check the reduce vector example on how
to quickly take the sum of this vector at the end.
You are attempting at calculating the filter output by directly evaluating the 1D convolution through a CUDA kernel.
In the case when the filter impulse response duration is long, one thing you can do to evaluate the filtered input is performing the calculations directly in the conjugate domain using FFTs. Below I'm reporting a sample code using CUDA Thrust and the cuFFT library. It is a direct translation of the Matlab-based example reported at
Low-Pass Filtering by FFT Convolution
Let me disclaim that some optimizations are possible with this code, but I preferred to leave it as it is so that it could be more easily compared to its Matlab's counterpart.
#include <stdio.h>
#include <math.h>
#include <cufft.h>
#include <thrust\device_vector.h>
#include <thrust\sequence.h>
#define pi_f 3.14159265358979f // Greek pi in single precision
/****************/
/* SIN OPERATOR */
/****************/
class sin_op {
float fk_, Fs_;
public:
sin_op(float fk, float Fs) { fk_ = fk; Fs_ = Fs; }
__host__ __device__ float operator()(float x) const { return sin(2.f*pi_f*x*fk_/Fs_); }
};
/*****************/
/* SINC OPERATOR */
/*****************/
class sinc_op {
float fc_, Fs_;
public:
sinc_op(float fc, float Fs) { fc_ = fc; Fs_ = Fs; }
__host__ __device__ float operator()(float x) const
{
if (x==0) return (2.f*fc_/Fs_);
else return (2.f*fc_/Fs_)*sin(2.f*pi_f*fc_*x/Fs_)/(2.f*pi_f*fc_*x/Fs_);
}
};
/********************/
/* HAMMING OPERATOR */
/********************/
class hamming_op {
int L_;
public:
hamming_op(int L) { L_ = L; }
__host__ __device__ float operator()(int x) const
{
return 0.54-0.46*cos(2.f*pi_f*x/(L_-1));
}
};
/*********************************/
/* MULTIPLY CUFFTCOMPLEX NUMBERS */
/*********************************/
struct multiply_cufftComplex {
__device__ cufftComplex operator()(const cufftComplex& a, const cufftComplex& b) const {
cufftComplex r;
r.x = a.x * b.x - a.y * b.y;
r.y = a.x * b.y + a.y * b.x;
return r;
}
};
/********/
/* MAIN */
/********/
void main(){
// Signal parameters:
int M = 256; // signal length
const int N = 4;
float f[N] = { 440, 880, 1000, 2000 }; // frequencies
float Fs = 5000.; // sampling rate
// Generate a signal by adding up sinusoids:
thrust::device_vector<float> d_x(M,0.f); // pre-allocate 'accumulator'
thrust::device_vector<float> d_n(M); // discrete-time grid
thrust::sequence(d_n.begin(), d_n.end(), 0, 1);
thrust::device_vector<float> d_temp(M);
for (int i=0; i<N; i++) {
float fk = f[i];
thrust::transform(d_n.begin(), d_n.end(), d_temp.begin(), sin_op(fk,Fs));
thrust::transform(d_temp.begin(), d_temp.end(), d_x.begin(), d_x.begin(), thrust::plus<float>());
}
// Filter parameters:
int L = 257; // filter length
float fc = 600.f; // cutoff frequency
// Design the filter using the window method:
thrust::device_vector<float> d_hsupp(L);
thrust::sequence(d_hsupp.begin(), d_hsupp.end(), -(L-1)/2, 1);
thrust::device_vector<float> d_hideal(L);
thrust::transform(d_hsupp.begin(), d_hsupp.end(), d_hideal.begin(), sinc_op(fc,Fs));
thrust::device_vector<float> d_l(L);
thrust::sequence(d_l.begin(), d_l.end(), 0, 1);
thrust::device_vector<float> d_h(L);
thrust::transform(d_l.begin(), d_l.end(), d_h.begin(), hamming_op(L));
// h is our filter
thrust::transform(d_hideal.begin(), d_hideal.end(), d_h.begin(), d_h.begin(), thrust::multiplies<float>());
// --- Choose the next power of 2 greater than L+M-1
int Nfft = pow(2,(ceil(log2((float)(L+M-1))))); // or 2^nextpow2(L+M-1)
// Zero pad the signal and impulse response:
thrust::device_vector<float> d_xzp(Nfft,0.f);
thrust::device_vector<float> d_hzp(Nfft,0.f);
thrust::copy(d_x.begin(), d_x.end(), d_xzp.begin());
thrust::copy(d_h.begin(), d_h.end(), d_hzp.begin());
// Transform the signal and the filter:
cufftHandle plan;
cufftPlan1d(&plan, Nfft, CUFFT_R2C, 1);
thrust::device_vector<cufftComplex> d_X(Nfft/2+1);
thrust::device_vector<cufftComplex> d_H(Nfft/2+1);
cufftExecR2C(plan, (cufftReal*)thrust::raw_pointer_cast(d_xzp.data()), (cufftComplex*)thrust::raw_pointer_cast(d_X.data()));
cufftExecR2C(plan, (cufftReal*)thrust::raw_pointer_cast(d_hzp.data()), (cufftComplex*)thrust::raw_pointer_cast(d_H.data()));
thrust::device_vector<cufftComplex> d_Y(Nfft/2+1);
thrust::transform(d_X.begin(), d_X.end(), d_H.begin(), d_Y.begin(), multiply_cufftComplex());
cufftPlan1d(&plan, Nfft, CUFFT_C2R, 1);
thrust::device_vector<float> d_y(Nfft);
cufftExecC2R(plan, (cufftComplex*)thrust::raw_pointer_cast(d_Y.data()), (cufftReal*)thrust::raw_pointer_cast(d_y.data()));
getchar();
}
Besides my other answer which I expect will be more convenient for convolution kernels with long duration, below I'm reporting a different implementation, which is more compliant with the OP's initial attempt and I expect will be more convenient for convolution kernels with short duration. Such an implementation is based on a hand-written kernel exploiting caching in shared memory. More details can be found in the book by D.B. Kirk and W.-m. W. Hwu
Programming Massively Parallel Processors, Second Edition: A Hands-on Approach
#include <stdio.h>
#include <stdlib.h>
#include "TimingGPU.cuh"
#include "Utilities.cuh"
#define RG 10
#define BLOCKSIZE 8
/****************/
/* CPU FUNCTION */
/****************/
void h_convolution_1D(const float * __restrict__ h_Signal, const float * __restrict__ h_ConvKernel, float * __restrict__ h_Result_CPU,
const int N, const int K) {
for (int i = 0; i < N; i++) {
float temp = 0.f;
int N_start_point = i - (K / 2);
for (int j = 0; j < K; j++) if (N_start_point + j >= 0 && N_start_point + j < N) {
temp += h_Signal[N_start_point+ j] * h_ConvKernel[j];
}
h_Result_CPU[i] = temp;
}
}
/********************/
/* BASIC GPU KERNEL */
/********************/
__global__ void d_convolution_1D_basic(const float * __restrict__ d_Signal, const float * __restrict__ d_ConvKernel, float * __restrict__ d_Result_GPU,
const int N, const int K) {
int i = blockIdx.x * blockDim.x + threadIdx.x;
float temp = 0.f;
int N_start_point = i - (K / 2);
for (int j = 0; j < K; j++) if (N_start_point + j >= 0 && N_start_point + j < N) {
temp += d_Signal[N_start_point+ j] * d_ConvKernel[j];
}
d_Result_GPU[i] = temp;
}
/***************************/
/* GPU KERNEL WITH CACHING */
/***************************/
__global__ void d_convolution_1D_caching(const float * __restrict__ d_Signal, const float * __restrict__ d_ConvKernel, float * __restrict__ d_Result_GPU,
const int N, const int K) {
int i = blockIdx.x * blockDim.x + threadIdx.x;
__shared__ float d_Tile[BLOCKSIZE];
d_Tile[threadIdx.x] = d_Signal[i];
__syncthreads();
float temp = 0.f;
int N_start_point = i - (K / 2);
for (int j = 0; j < K; j++) if (N_start_point + j >= 0 && N_start_point + j < N) {
if ((N_start_point + j >= blockIdx.x * blockDim.x) && (N_start_point + j < (blockIdx.x + 1) * blockDim.x))
// --- The signal element is in the tile loaded in the shared memory
temp += d_Tile[threadIdx.x + j - (K / 2)] * d_ConvKernel[j];
else
// --- The signal element is not in the tile loaded in the shared memory
temp += d_Signal[N_start_point + j] * d_ConvKernel[j];
}
d_Result_GPU[i] = temp;
}
/********/
/* MAIN */
/********/
int main(){
const int N = 15; // --- Signal length
const int K = 5; // --- Convolution kernel length
float *h_Signal = (float *)malloc(N * sizeof(float));
float *h_Result_CPU = (float *)malloc(N * sizeof(float));
float *h_Result_GPU = (float *)malloc(N * sizeof(float));
float *h_ConvKernel = (float *)malloc(K * sizeof(float));
float *d_Signal; gpuErrchk(cudaMalloc(&d_Signal, N * sizeof(float)));
float *d_Result_GPU; gpuErrchk(cudaMalloc(&d_Result_GPU, N * sizeof(float)));
float *d_ConvKernel; gpuErrchk(cudaMalloc(&d_ConvKernel, K * sizeof(float)));
for (int i=0; i < N; i++) { h_Signal[i] = (float)(rand() % RG); }
for (int i=0; i < K; i++) { h_ConvKernel[i] = (float)(rand() % RG); }
gpuErrchk(cudaMemcpy(d_Signal, h_Signal, N * sizeof(float), cudaMemcpyHostToDevice));
gpuErrchk(cudaMemcpy(d_ConvKernel, h_ConvKernel, K * sizeof(float), cudaMemcpyHostToDevice));
h_convolution_1D(h_Signal, h_ConvKernel, h_Result_CPU, N, K);
d_convolution_1D_basic<<<iDivUp(N, BLOCKSIZE), BLOCKSIZE>>>(d_Signal, d_ConvKernel, d_Result_GPU, N, K);
gpuErrchk(cudaPeekAtLastError());
gpuErrchk(cudaDeviceSynchronize());
gpuErrchk(cudaMemcpy(h_Result_GPU, d_Result_GPU, N * sizeof(float), cudaMemcpyDeviceToHost));
for (int i = 0; i < N; i++) if (h_Result_CPU[i] != h_Result_GPU[i]) {printf("mismatch2 at %d, cpu: %d, gpu %d\n", i, h_Result_CPU[i], h_Result_GPU[i]); return 1;}
printf("Test basic passed\n");
d_convolution_1D_caching<<<iDivUp(N, BLOCKSIZE), BLOCKSIZE>>>(d_Signal, d_ConvKernel, d_Result_GPU, N, K);
gpuErrchk(cudaPeekAtLastError());
gpuErrchk(cudaDeviceSynchronize());
gpuErrchk(cudaMemcpy(h_Result_GPU, d_Result_GPU, N * sizeof(float), cudaMemcpyDeviceToHost));
for (int i = 0; i < N; i++) if (h_Result_CPU[i] != h_Result_GPU[i]) {printf("mismatch2 at %d, cpu: %d, gpu %d\n", i, h_Result_CPU[i], h_Result_GPU[i]); return 1;}
printf("Test caching passed\n");
return 0;
}

Unordinary performance gap between OpenCL and CUDA

I have coded a simple tiled matrix multiplication in CUDA. It's like this:
__global__ void matrixMultiplyShared(float * A, float * B, float * C,
int numARows, int numAColumns,
int numBRows, int numBColumns,
int numCRows, int numCColumns) {
__shared__ float ds_A[TILE_WIDTH][TILE_WIDTH];
__shared__ float ds_B[TILE_WIDTH][TILE_WIDTH];
int bx = blockIdx.x; int by = blockIdx.y;
int tx = threadIdx.x; int ty = threadIdx.y;
int row = by * TILE_WIDTH + ty;
int col = bx * TILE_WIDTH + tx;
float Cvalue = 0.0;
// Loop over the M and N tiles required to compute the Pd element
for (int m = 0; m < (numAColumns-1)/TILE_WIDTH+1; ++m) {
if(row<numARows && m*TILE_WIDTH+tx < numAColumns){
ds_A[ty][tx] = A[row*numAColumns + m*TILE_WIDTH+tx];
} else {
ds_A[ty][tx] = 0;
}
if(m*TILE_WIDTH+ty < numBRows && col < numBColumns){
ds_B[ty][tx] = B[(m*TILE_WIDTH+ty)*numBColumns+col];
} else {
ds_B[ty][tx] = 0;
}
__syncthreads();
if(row < numCRows && col < numCColumns){
for (int k = 0; k < TILE_WIDTH; ++k)
Cvalue += ds_A[ty][k] * ds_B[k][tx];
}
__syncthreads();
}
if(row < numCRows && col < numCColumns)
C[row*numCColumns+col] = Cvalue;
}
After that, I used the same above kernel (with some minor changes) in the OpenCL version to compare the performance of CUDA and OpenCL together. But the result was to so far beyond my expectations. OpenCL was 6-7 times faster than CUDA. Is it valid?
The output of Nisght is as follows:
CUDA:
OpenCL:
You can see a large gap between starting the app and executing the kernel. why is that happened?
My GPU is: GTX 580 |
The Kernel Ex time (CUDA): 3.78s |
The Kernel Ex time (OpenCL): 0.53s |
CUDA Code: http://pastebin.com/VQMp3Hba
OpenCL Host Code: http://pastebin.com/cjGYSLQf
OpenCL Kernel Code: http://pastebin.com/KKw3Ayz7
You can try and insert explicit timers in the code instead of trusting the output from the tool. May be the case that the tool is wrong.

Resources