Related
Nvprof reported that there are about 200 milion shared_ld_bank_conflict and some shared_st_bank_conflict in my sgemm kernel. I tried the padding trick __shared__ float smem[SIZE + OFFSET];, it reduced store bank conflicts to 0, but load bank conflicts are still there. I don't know how to further improve it.
__global__ void sgemm(
const float* __restrict__ A,
const float* __restrict__ B,
float* __restrict__ C,
int M, int N, int K
){
int tid = threadIdx.x;
int gStartx = blockIdx.x * 128;
int gStarty = blockIdx.y * 128;
int dx = tid % 8;
int dy = tid / 8;
int vx = tid % 16;
int vy = tid / 16;
__shared__ volatile float aSM[8][128+4];
__shared__ volatile float bSM[8][128+4];
float aBuffer1[4];
float bBuffer1[4];
float aBuffer2[4];
float bBuffer2[4];
float cCache[8][8];
#pragma unroll
for (int i=0; i<8; i++)
#pragma unroll
for (int j=0; j<8; j++)
cCache[i][j] = 0.f;
//load first two tiles
#pragma unroll
for (int i=0; i<4; i++){
aBuffer1[i] = A[(gStarty + dy + i*32)*K + (dx)];
bBuffer1[i] = B[(gStartx + dy + i*32)*K + (dx)];
}
int nIt = (K + 8 - 1) / 8;
#pragma unroll
for (int itr=0; itr<nIt; itr++){
int gStartk = itr * 8;
int is_odd = itr & 1;
if (is_odd == 0){
#pragma unroll
for (int i=0; i<4; i++){
if (itr != (nIt - 1)){
// prefetch next tiles
aBuffer2[i] = A[(gStarty + i*32 + dy)*K + (gStartk + 8 + dx)];
bBuffer2[i] = B[(gStartx + i*32 + dy)*K + (gStartk + 8 + dx)];
}
//move current tiles to SMEM
aSM[dx][dy+i*32] = aBuffer1[i];
bSM[dx][dy+i*32] = bBuffer1[i];
}
} else {
#pragma unroll
for (int i=0; i<4; i++){
if (itr != (nIt - 1)){
//prefetch next tiles to another buffer
aBuffer1[i] = A[(gStarty + i*32 + dy)*K + (gStartk + 8 + dx)];
bBuffer1[i] = B[(gStartx + i*32 + dy)*K + (gStartk + 8 + dx)];
}
aSM[dx][dy+i*32] = aBuffer2[i];
bSM[dx][dy+i*32] = bBuffer2[i];
}
}
__syncthreads();
float aCache[8][4];
#pragma unroll
for (int p=0; p<2; p++){
#pragma unroll
for (int ki=0; ki<8; ki++){
#pragma unroll
for (int mi=0; mi<4; mi++){
aCache[ki][mi] = aSM[ki][8*vy + 4*p +mi];
}
}
#pragma unroll
for (int ki=0; ki<8; ki++){
#pragma unroll
for (int ni=0; ni<8; ni++){
float b = bSM[ki][8*vx + ni];
#pragma unroll
for (int mi=0; mi<4; mi++){
float a = aCache[ki][mi];
cCache[mi + 4*p][ni] = fma(a, b, cCache[mi + 4*p][ni] );
}
}
}
}
__syncthreads();
}
#pragma unroll
for (int i=0; i<8; i++){
for (int j=0; j<8; j++){
C[(gStarty + vy*8 + i)*N + (gStartx + vx*8 + j)] = cCache[i][j];
}
}
}
A (2048x2048) matrix is row major, B (2048x2048) is column major, each block has 256 threads, each block calculates 128x128 portion of C, and each thread calculates 8x8x8. the gpu is Tesla P100.
Ok I found a solution: when storing to bSM, insert one padding word between every 32 words in the second dimention
//bSM[dx][dy+i*32] = bBuffer1[i];
bSM[dx][dy+i*33] = bBuffer1[i]; //we're skipping column 32, 65, 98, 131
when reading bSM[i][j], read it like this: bSM[i][j/32 + j]
//float b = bSM[ki][8*vx + ni];
float b = bSM[ki][(8*vx) / 32 + 8*vx + ni];
// (8*vx+ni)/32 is the same as (8*vx)/32, since vi is always less than 8
now it's giving me 55% performance of cublas gemm on tesla p4
I was trying to parallelize the gaussian blur function using OpenMP,
but I am new at OpenMP, and when I tried to parallelize the two for loops (I don't think there are any variables that need to be private for each thread), it ended up
running even slower than before, and the output was different. So did I do anything wrong? What should I do to make it run faster?
void gaussian_blur(float **src, float **dst, int w, int h, float sigma)
{
int x, y, i;
int ksize = (int)(sigma * 2.f * 4.f + 1) | 1;
int halfk = ksize / 2;
float scale = -0.5f/(sigma*sigma);
float sum = 0.f;
float *kernel, *ringbuf;
int xmax = w - halfk;
int ymax = h - halfk;
// if sigma too small, just copy src to dst
if (ksize <= 1)
{
for (y = 0; y < h; y++)
for (x = 0; x < w; x++)
dst[y][x] = src[y][x];
return;
}
// create Gaussian kernel
kernel = malloc(ksize * sizeof(float));
ringbuf = malloc(ksize * sizeof(float));
#pragma omp parallel for reduction(+ : sum)
for (i = 0; i < ksize; i++)
{
float x = (float)(i - halfk);
float t = expf(scale * x * x);
kernel[i] = t;
sum += t;
}
scale = 1.f / sum;
#pragma omp parallel for
for (i = 0; i < ksize; i++)
kernel[i] *= scale;
// blur each row
#pragma omp parallel for // this is the for loop I parallelized but ended up with wrong output and running slower
for (y = 0; y < h; y++)
{
int x1;
int bufi0 = ksize-1;
float tmp = src[y][0];
for (x1 = 0; x1 < halfk ; x1++) ringbuf[x1] = tmp;
for (; x1 < ksize-1; x1++) ringbuf[x1] = src[y][x1-halfk];
for (x1 = 0; x1 < w; x1++)
{
if(x1 < xmax)
ringbuf[bufi0++] = src[y][x1+halfk];
else
ringbuf[bufi0++] = src[y][w-1];
if (bufi0 == ksize) bufi0 = 0;
dst[y][x1] = convolve(kernel, ringbuf, ksize, bufi0);
}
}
// blur each column
#pragma omp parallel for // this is the for loop I parallelized but ended up with wrong output and running slower
for (x = 0; x < w; x++)
{
int y1;
int bufi0 = ksize-1;
float tmp = dst[0][x];
for (y1 = 0; y1 < halfk ; y1++) ringbuf[y1] = tmp;
for ( ; y1 < ksize-1; y1++) ringbuf[y1] = dst[y1-halfk][x];
for (y1 = 0; y1 < h; y1++)
{
if(y1 < ymax)
ringbuf[bufi0++] = dst[y1+halfk][x];
else
ringbuf[bufi0++] = dst[h-1][x];
if (bufi0 == ksize) bufi0 = 0;
dst[y1][x] = convolve(kernel, ringbuf, ksize, bufi0);
}
}
// clean up
free(kernel);
free(ringbuf);
}
Besides the need to properly identify private and shared data, there are several things that you could do in order to speed up your program.
As a first step you should remove any unnecessary concurrency. For example, how big ksize happens to be on average? If it is less than several hundred elements, it makes absolutely no sense to employ OpenMP for such simple operations as computing the kernel and then normalising it:
#pragma omp parallel for reduction(+ : sum)
for (i = 0; i < ksize; i++)
{
float x = (float)(i - halfk);
float t = expf(scale * x * x);
kernel[i] = t;
sum += t;
}
scale = 1.f / sum;
#pragma omp parallel for
for (i = 0; i < ksize; i++)
kernel[i] *= scale;
On a typical modern CPU it would take more cycles to bootstrap the parallel regions than to compute this on a single core. Also on modern CPUs these loops can be unrolled and vectorised and you can get up to 8x boost on a single core. If the kernel is too small, then besides OpenMP overhead you will also get slowdown from excessive false sharing. You have to make sure that each thread gets an exact multiple of 16 elements (64 bytes of cache line size / sizeof(float)) to work on in order to prevent false sharing.
You also have to make sure that threads do not share cache lines in the column blur section.
// blur each column
#pragma omp parallel for
for (x = 0; x < w; x++)
{
...
for (y1 = 0; y1 < h; y1++)
{
...
dst[y1][x] = convolve(kernel, ringbuf, ksize, bufi0);
}
}
Because of the access pattern here, you have to make sure that each thread gets a chunk of columns that is a multiple of 16 or else there will be a border overlap area of 16*y1 pixels shared by every two consecutive threads where excessive false sharing will occur. If you cannot guarantee that w is divisible by 16, then you can give each thread a starting offset in the y direction, e.g. the innermost loop becomes:
int tid = omp_get_thread_num();
for (y1 = 2*tid; y1 < h; y1++)
{
...
}
for (y1 = 0; y1 < 2*tid; y1++)
{
...
}
The multiplier 2 is arbitrary. The idea is to give the next thread several rows of advance in comparison to the current one so that both threads will not be processing the same line at once at any moment in time. You could also use addition and modulo arithmetic to compute y1, i.e.
for (y2 = 0; y2 < h; y2++)
{
y1 = (y2 + 2*tid) % h;
...
}
but this is generally slower than just separating the loop in two parts.
Also mind your data size. The last level cache (LLC) has very high but still limited bandwidth. If data cannot fit in the private cache of each core then compiler optimisations such as loop vectorisations can put very high pressure on the LLC. Things get more ugly if data doesn't fit in the LLC and therefore the main memory has to be accessed.
If you don't know what false sharing is, there is an article in Dr.Dobb's that kind of explains it here.
I may have fixed your code. You did not post your convolve function so it's difficult to say for sure but I'm not sure it matters. There are at least two bugs. There is a race condition in the ringbuf array. To fix this I extend the array times the number of threads.
ringbuf = (float*)malloc(nthreads*ksize * sizeof(float));
To access the array do something like this
int ithread = omp_get_thread_num();
ringbuf[ksize*ithread + x1]
Edit: I added some code which defines ringbuf inside the parallel block. That way you don't have to access ringbuf based on the thread number.
The second errors is the ibufi0 variable. I defined a new one like this
const int ibufi0_fix = (x1+ksize-1)%ksize;
Below is the code I used to check it. Replace with your convolve function. Note, this may still be quite inefficient. There are probably cache issues such as cache misses and false sharing (particularly when you convolve vertically). Hopefully, though, the image will be correct now.
Edit: here is a paper by Intel that shows how to do this best with AVX. It's optimized to minimize the cache misses. I'm not sure it's optimized for threading though.
http://software.intel.com/en-us/articles/iir-gaussian-blur-filter-implementation-using-intel-advanced-vector-extensions
I'm writing my own function on this (it's actually the reason I started learning OpenMP) which uses SSE/AVX as well. There are a lot of similarities with matrix multiplication and image filtering so I learned how to optimized matrix multiplication first and will do Gaussian Blur shortly...
#include "math.h"
#include "omp.h"
#include "stdio.h"
#include <nmmintrin.h>
float convolve(const float *kernel, const float *ringbuf, const int ksize, const int bufi0) {
float sum = 0.0f;
for(int i=0; i<ksize; i++) {
sum += kernel[i]*ringbuf[i];
}
return sum;
}
void gaussian_blur(float *src, float *dst, int w, int h, float sigma, int nthreads)
{
int x, y, i;
int ksize = (int)(sigma * 2.f * 4.f + 1) | 1;
int halfk = ksize / 2;
printf("ksize %d\n", ksize);
float scale = -0.5f/(sigma*sigma);
float sum = 0.f;
float *kernel, *ringbuf;
int xmax = w - halfk;
int ymax = h - halfk;
// if sigma too small, just copy src to dst
if (ksize <= 1)
{
for (y = 0; y < h; y++)
for (x = 0; x < w; x++)
dst[y*w + x] = src[y*w + x];
return;
}
// create Gaussian kernel
//kernel = malloc(ksize * sizeof(float));
kernel = (float*)_mm_malloc(ksize * sizeof(float),16);
//ringbuf = malloc(ksize * sizeof(float));
ringbuf = (float*)_mm_malloc(nthreads*ksize * sizeof(float),16);
#pragma omp parallel for reduction(+ : sum) if(nthreads>1)
for (i = 0; i < ksize; i++)
{
float x = (float)(i - halfk);
float t = expf(scale * x * x);
kernel[i] = t;
sum += t;
}
scale = 1.f / sum;
#pragma omp parallel for if(nthreads>1)
for (i = 0; i < ksize; i++)
kernel[i] *= scale;
// blur each row
#pragma omp parallel for if(nthreads>1)// this is the for loop I parallelized but ended up with wrong output and running slower
for (y = 0; y < h; y++)
{
int ithread = omp_get_thread_num();
//printf("nthread %d\n", nthread);
int x1;
int bufi0 = ksize-1;
float tmp = src[y*w + 0];
for (x1 = 0; x1 < halfk ; x1++) ringbuf[ksize*ithread + x1] = tmp;
for (; x1 < ksize-1; x1++) ringbuf[ksize*ithread + x1] = src[y*w + x1-halfk];
for (x1 = 0; x1 < w; x1++)
{
const int ibufi0_fix = (x1+ksize-1)%ksize;
if(x1 < xmax)
ringbuf[ksize*ithread + ibufi0_fix] = src[y*w + x1+halfk];
else
ringbuf[ksize*ithread + ibufi0_fix] = src[y*w + w-1];
if (bufi0 == ksize) bufi0 = 0;
dst[y*w + x1] = convolve(kernel, &ringbuf[ksize*ithread], ksize, bufi0);
}
}
// blur each column
#pragma omp parallel for if(nthreads>1)// this is the for loop I parallelized but ended up with wrong output and running slower
for (x = 0; x < w; x++)
{
int ithread = omp_get_thread_num();
int y1;
int bufi0 = ksize-1;
float tmp = dst[0*w + x];
for (y1 = 0; y1 < halfk ; y1++) ringbuf[ksize*ithread + y1] = tmp;
for ( ; y1 < ksize-1; y1++) ringbuf[ksize*ithread + y1] = dst[(y1-halfk)*w + x];
for (y1 = 0; y1 < h; y1++)
{
const int ibufi0_fix = (y1+ksize-1)%ksize;
if(y1 < ymax)
ringbuf[ibufi0_fix] = dst[(y1+halfk)*w + x];
else
ringbuf[ibufi0_fix] = dst[(h-1)*w + x];
if (bufi0 == ksize) bufi0 = 0;
dst[y1*w + x] = convolve(kernel, &ringbuf[ksize*ithread], ksize, bufi0);
}
}
// clean up
_mm_free(kernel);
_mm_free(ringbuf);
}
int compare(float *dst1, float *dst2, const int n) {
int error = 0;
for(int i=0; i<n; i++) {
if(*dst1 != *dst2) error++;
}
return error;
}
int main() {
const int w = 20;
const int h = 20;
float *src = (float*)_mm_malloc(w*h*sizeof(float),16);
float *dst1 = (float*)_mm_malloc(w*h*sizeof(float),16);
float *dst2 = (float*)_mm_malloc(w*h*sizeof(float),16);
for(int i=0; i<w*h; i++) {
src[i] = i;
}
gaussian_blur(src, dst1, w, h, 1.0f, 1);
gaussian_blur(src, dst2, w, h, 1.0f, 4);
int error = compare(dst1, dst2, w*h);
printf("error %d\n", error);
_mm_free(src);
_mm_free(dst1);
_mm_free(dst2);
}
Edit: here is code which defines ringbuf inside the parallel block based on the comment by Hristo. It should be equivalent.
#include "math.h"
#include "omp.h"
#include "stdio.h"
#include <nmmintrin.h>
float convolve(const float *kernel, const float *ringbuf, const int ksize, const int bufi0) {
float sum = 0.0f;
for(int i=0; i<ksize; i++) {
sum += kernel[i]*ringbuf[i];
}
return sum;
}
void gaussian_blur(float *src, float *dst, int w, int h, float sigma, int nthreads)
{
int x, y, i;
int ksize = (int)(sigma * 2.f * 4.f + 1) | 1;
int halfk = ksize / 2;
printf("ksize %d\n", ksize);
float scale = -0.5f/(sigma*sigma);
float sum = 0.f;
float *kernel;
int xmax = w - halfk;
int ymax = h - halfk;
// if sigma too small, just copy src to dst
if (ksize <= 1)
{
for (y = 0; y < h; y++)
for (x = 0; x < w; x++)
dst[y*w + x] = src[y*w + x];
return;
}
// create Gaussian kernel
//kernel = malloc(ksize * sizeof(float));
kernel = (float*)_mm_malloc(ksize * sizeof(float),16);
#pragma omp parallel for reduction(+ : sum) if(nthreads>1)
for (i = 0; i < ksize; i++)
{
float x = (float)(i - halfk);
float t = expf(scale * x * x);
kernel[i] = t;
sum += t;
}
scale = 1.f / sum;
#pragma omp parallel for if(nthreads>1)
for (i = 0; i < ksize; i++)
kernel[i] *= scale;
// blur each row
//#pragma omp parallel for if(nthreads>1)// this is the for loop I parallelized but ended up with wrong output and running slower
#pragma omp parallel if(nthreads>1)
{
float *ringbuf = (float*)_mm_malloc(ksize * sizeof(float),16);
#pragma omp for// this is the for loop I parallelized but ended up with wrong output and running slower
for (y = 0; y < h; y++)
{
//printf("nthread %d\n", nthread);
int x1;
int bufi0 = ksize-1;
float tmp = src[y*w + 0];
for (x1 = 0; x1 < halfk ; x1++) ringbuf[x1] = tmp;
for (; x1 < ksize-1; x1++) ringbuf[x1] = src[y*w + x1-halfk];
for (x1 = 0; x1 < w; x1++)
{
const int ibufi0_fix = (x1+ksize-1)%ksize;
if(x1 < xmax)
ringbuf[ibufi0_fix] = src[y*w + x1+halfk];
else
ringbuf[ibufi0_fix] = src[y*w + w-1];
if (bufi0 == ksize) bufi0 = 0;
dst[y*w + x1] = convolve(kernel, ringbuf, ksize, bufi0);
}
}
_mm_free(ringbuf);
}
// blur each column
#pragma omp parralel if(ntheads>1)
{
float *ringbuf = (float*)_mm_malloc(ksize * sizeof(float),16);
#pragma omp for// this is the for loop I parallelized but ended up with wrong output and running slower
for (x = 0; x < w; x++)
{
int y1;
int bufi0 = ksize-1;
float tmp = dst[0*w + x];
for (y1 = 0; y1 < halfk ; y1++) ringbuf[y1] = tmp;
for ( ; y1 < ksize-1; y1++) ringbuf[y1] = dst[(y1-halfk)*w + x];
for (y1 = 0; y1 < h; y1++)
{
const int ibufi0_fix = (y1+ksize-1)%ksize;
if(y1 < ymax)
ringbuf[ibufi0_fix] = dst[(y1+halfk)*w + x];
else
ringbuf[ibufi0_fix] = dst[(h-1)*w + x];
if (bufi0 == ksize) bufi0 = 0;
dst[y1*w + x] = convolve(kernel, ringbuf, ksize, bufi0);
}
}
_mm_free(ringbuf);
}
// clean up
_mm_free(kernel);
}
int compare(float *dst1, float *dst2, const int n) {
int error = 0;
for(int i=0; i<n; i++) {
if(*dst1 != *dst2) error++;
}
return error;
}
int main() {
const int w = 20;
const int h = 20;
float *src = (float*)_mm_malloc(w*h*sizeof(float),16);
float *dst1 = (float*)_mm_malloc(w*h*sizeof(float),16);
float *dst2 = (float*)_mm_malloc(w*h*sizeof(float),16);
for(int i=0; i<w*h; i++) {
src[i] = i;
}
gaussian_blur(src, dst1, w, h, 1.0f, 1);
gaussian_blur(src, dst2, w, h, 1.0f, 4);
int error = compare(dst1, dst2, w*h);
printf("error %d\n", error);
_mm_free(src);
_mm_free(dst1);
_mm_free(dst2);
}
I am learning CUDA and still at the very beginner level. I am trying a simple assignment but my code crashes when I run it and I am not sure why. Any help would be appreciated.
EDIT: Crashes on cudaMemcpy and in Image structure, the pixelVal is of type int**. Is that the cause?
Original C++ code:
void Image::reflectImage(bool flag, Image& oldImage)
/*Reflects the Image based on users input*/
{
int rows = oldImage.N;
int cols = oldImage.M;
Image tempImage(oldImage);
for(int i = 0; i < rows; i++)
{
for(int j = 0; j < cols; j++)
tempImage.pixelVal[rows - (i + 1)][j] = oldImage.pixelVal[i][j];
}
oldImage = tempImage;
}
My CUDA kernel & code:
#define NTPB 512
__global__ void fliph(int* a, int* b, int r, int c)
{
int i = blockIdx.x * blockDim.x + threadIdx.x;
int j = blockIdx.y * blockDim.y + threadIdx.y;
if (i >= r || j >= c)
return;
a[(r - i * c) + j] = b[i * c + j];
}
void Image::reflectImage(bool flag, Image& oldImage)
/*Reflects the Image based on users input*/
{
int rows = oldImage.N;
int cols = oldImage.M;
Image tempImage(oldImage);
if(flag == true) //horizontal reflection
{
//Allocate device memory
int* dpixels;
int* oldPixels;
int n = rows * cols;
cudaMalloc((void**)&dpixels, n * sizeof(int));
cudaMalloc((void**)&oldPixels, n * sizeof(int));
cudaMemcpy(dpixels, tempImage.pixelVal, n * sizeof(int), cudaMemcpyHostToDevice);
cudaMemcpy(oldPixels, oldImage.pixelVal, n * sizeof(int), cudaMemcpyHostToDevice);
int nblks = (n + NTPB - 1) / NTPB;
fliph<<<nblks, NTPB>>>(dpixels, oldPixels, rows, cols);
cudaMemcpy(tempImage.pixelVal, dpixels, n * sizeof(int), cudaMemcpyDeviceToHost);
cudaFree(dpixels);
cudaFree(oldPixels);
}
oldImage = tempImage;
}
You have to create a 2D Grid in order to process the image using 2D indices i and j. In the current case, the kernel is processing only the first row of the image.
To create a 2D indexing mechanism, create a 2D block and 2D grid like this:
const int BLOCK_DIM = 16;
dim3 Block(BLOCK_DIM,BLOCK_DIM);
dim3 Grid;
Grid.x = (cols + Block.x - 1)/Block.x;
Grid.y = (rows + Block.y - 1)/Block.y;
fliph<<<Grid, Block>>>(dpixels, oldPixels, rows, cols);
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;
}
I want to write a CUDA kernel that will multiply 2 matrices NxN size. I did manage to do it, but without the thread cooperation... Now I want to do it with thread cooperation, and I followed the code provided in the SDK. But for some reason kernel returns different result. So here is the .cu file:
#include<stdio.h>
#include<cuda.h>
#include<cuda_runtime.h>
#include<cuda_runtime_api.h>
#include<device_functions.h>
static void HandleError(cudaError_t err, const char *file, int line)
{
if(err!=cudaSuccess){
printf("%s in %s file at line %s\n", cudaGetErrorString(err), file, line);
exit(EXIT_FAILURE);
}
}
#define HANDLE_ERROR(err) (HandleError(err, __FILE__, __LINE__))
#ifndef _MATRIXMUL_KERNEL_H_
#define _MATRIXMUL_KERNEL_H_
#define ORDER 4
__global__ void matrixMul( int* A, int* B, int* C, int wA, int wB)
{
int bx = blockIdx.x;
int by = blockIdx.y;
int tx = threadIdx.x;
int ty = threadIdx.y;
int aBegin = wA * ORDER * by;
int aEnd = aBegin + wA - 1;
int aStep = ORDER;
int bBegin = ORDER * bx;
int bStep = ORDER * wB;
int Csub=0;
for (int a = aBegin, b = bBegin; a <= aEnd; a += aStep, b += bStep)
{
__shared__ int As[ORDER][ORDER];
__shared__ int Bs[ORDER][ORDER];
As[ty][tx] = A[a + wA * ty + tx];
Bs[ty][tx] = B[b + wB * ty + tx];
__syncthreads();
#pragma unroll
for (int k = 0; k < ORDER; ++k)
Csub += As[ty][k] * Bs[k][tx];
__syncthreads();
}
int c = wB * ORDER * by + ORDER * bx;
C[c + wB * ty + tx] = Csub;
}
#endif
int main()
{
int *a=(int*)malloc(ORDER*ORDER*sizeof(int));
int *b=(int*)malloc(ORDER*ORDER*sizeof(int));
int *c=(int*)malloc(ORDER*ORDER*sizeof(int));
int *dev_a, *dev_b, *dev_c;
HANDLE_ERROR(cudaMalloc((void**)&dev_a, ORDER*ORDER*sizeof(int*)));
HANDLE_ERROR(cudaMalloc((void**)&dev_b, ORDER*ORDER*sizeof(int*)));
HANDLE_ERROR(cudaMalloc((void**)&dev_c, ORDER*ORDER*sizeof(int*)));
for(int i=0; i<ORDER*ORDER; i++)
{
a[i]=1;
b[i]=2;
}
HANDLE_ERROR(cudaMemcpy(dev_a, a, ORDER*ORDER*sizeof(int), cudaMemcpyHostToDevice));
HANDLE_ERROR(cudaMemcpy(dev_b, b, ORDER*ORDER*sizeof(int), cudaMemcpyHostToDevice));
matrixMul<<<ORDER, ORDER>>>(dev_a, dev_b, dev_c, ORDER, ORDER);
HANDLE_ERROR(cudaMemcpy(c, dev_c, ORDER*ORDER*sizeof(int), cudaMemcpyDeviceToHost));
for(int i=0; i<ORDER*ORDER; i++)
{
if((i%ORDER)==0)
printf("\n\n");
printf("%d\t", a[i]);
}
for(int i=0; i<ORDER*ORDER; i++)
{
if((i%ORDER)==0)
printf("\n\n");
printf("%d\t", b[i]);
}
for(int i=0; i<ORDER*ORDER; i++)
{
if((i%ORDER)==0)
printf("\n\n");
printf("%d\t", c[i]);
}
cudaFree(dev_a);
cudaFree(dev_b);
cudaFree(dev_c);
return 0;
}
Yes, I know that there is no "real" question... But if anyone could point me to wright direction I would be grateful. Thank you!
If you need more code example, let me know and I'll edit the question.
EDIT #1: I forgot to mention... I haven't been able to implement nvcc in Visual Studi 2010 so I'm unable to use debugger. Any suggestion about that?
EDIT #2: Updated question so it shows both CUDA kernel and main.
Your kernel seems right, if your thread-geometry is BLOCKSIZE x BLOCKSIZE. Is that the case?
If that isn't your problem:
Since you said you got it working without thread synchronization, you probably got the memory allocation correct.
Try testing with a thread-geometry of 4x4 and the following two matrices:
1 1 1 1 1 0 0 0
2 2 2 2 0 1 0 0
3 3 3 3 0 0 1 0
5 5 5 5 0 0 0 1
The output should give you a hint as to what might be going wrong.