Why PyCUDA is faster than C CUDA in this example - cuda

I am exploring to move from OpenCL to CUDA, and did a few tests to benchmark the speed of CUDA in various implementations. To my surprise, in the examples below, the PyCUDA implementation is about 20% faster than the C CUDA example.
I read many posts talking about "release build" of C CUDA code. I did try having -Xptxas -O3 in the makefile and that really did not make a difference. I also tried to adjust the block size, with which the kernel was executed. Unfortunately, it did not help improve the speed, either.
My questions here are:
What could be the reasons leading to the speed difference between C CUDA and PYCUDA?
If the "advanced" (lack of a better word) compiling in PYCUDA is one of reasons, how can I optimize the compiling of my C CUDA code?
Are there any other ways to improve the speed of C CUDA in this case?
While I appreciate general comments, I am looking for actionable suggestions that I can validate on my machine. Thanks!
import pycuda.autoinit
import pycuda.driver as drv
import numpy as np
from pycuda.compiler import SourceModule
import time
mod = SourceModule(
"""
__global__ void saxpy(int n, const float a, float *x, float *y)
{
int i = blockIdx.x * blockDim.x + threadIdx.x;
if (i < n){
y[i] = a * x[i] + y[i];
}
}
"""
)
saxpy = mod.get_function("saxpy")
N = 1 << 25
time_elapse = 0.0
for i in range(100):
# print(i)
# print(N)
x = np.ones(N).astype(np.float32)
y = 2 * np.ones(N).astype(np.float32)
start = time.time()
saxpy(
np.int32(N),
np.float32(2.0),
drv.In(x),
drv.InOut(y),
block=(512, 1, 1),
grid=(int(N / 512) + 1, 1),
)
time_elapse += (time.time() - start)
print(time_elapse )
print(y[-100:-1])
print(y.sum())
print(N * 4.0)
#include <stdio.h>
#include <time.h>
#define DIM 512
__global__ void saxpy(int n, float a, float *x, float *y)
{
int i = blockIdx.x * blockDim.x + threadIdx.x;
if (i < n)
y[i] = a * x[i] + y[i];
}
int main(int num_iterations)
{
double start;
double cputime;
int N = 1 << 25;
float *x, *y, *d_x, *d_y;
int i, j;
for (j = 0; j < num_iterations; j++)
{
x = (float *)malloc(N * sizeof(float));
y = (float *)malloc(N * sizeof(float));
cudaMalloc(&d_x, N * sizeof(float));
cudaMalloc(&d_y, N * sizeof(float));
for (i = 0; i < N; i++)
{
x[i] = 1.0f;
y[i] = 2.0f;
}
cudaMemcpy(d_x, x, N * sizeof(float), cudaMemcpyHostToDevice);
cudaMemcpy(d_y, y, N * sizeof(float), cudaMemcpyHostToDevice);
// Perform SAXPY on 1M elements
start = clock();
saxpy<<<(N + DIM) / DIM, DIM>>>(N, 2.0f, d_x, d_y);
cputime += ((double)(clock() - start) / CLOCKS_PER_SEC);
cudaMemcpy(y, d_y, N * sizeof(float), cudaMemcpyDeviceToHost);
// float maxError = 0.0f;
// for (int i = 0; i < N; i++){
// maxError = max(maxError, abs(y[i] - 4.0f));
// //printf("y[%d]: %f\n", i,y[i]);
// }
// printf("Max error: %f\n", maxError);
cudaFree(d_x);
cudaFree(d_y);
free(x);
free(y);
}
printf("cpu time is %f\n", cputime);
return 0;
}
I saved the above file as cuda_example.cu and compile it with the following commands in a makefile:
nvcc -arch=sm_61 -Xptxas -O3,-v -o main cuda_example.cu

If I execute your CUDA-C code as is, and set num_iterations to 300 like this:
int num_iterations =300;
then the execution of your program takes about 60s on a Geforce GTX 1650. Your code is extremely inefficient, as you copy data back and forth between GPU and device at every iteration.
So, lets restrict the loop to just the kernel execution:
#include <stdio.h>
#include <time.h>
#define DIM 512
__global__ void saxpy(int n, float a, float *x, float *y)
{
int i = blockIdx.x * blockDim.x + threadIdx.x;
if (i < n)
y[i] = a * x[i] + y[i];
}
int main()
{
double start = clock();
int N = 1 << 25;
float *x, *y, *d_x, *d_y;
int i, j;
int num_iterations = 300;
x = (float *)malloc(N * sizeof(float));
y = (float *)malloc(N * sizeof(float));
cudaMalloc(&d_x, N * sizeof(float));
cudaMalloc(&d_y, N * sizeof(float));
for (i = 0; i < N; i++)
{
x[i] = 1.0f;
y[i] = 2.0f;
}
cudaMemcpy(d_x, x, N * sizeof(float), cudaMemcpyHostToDevice);
cudaMemcpy(d_y, y, N * sizeof(float), cudaMemcpyHostToDevice);
for (j = 0; j < num_iterations; j++){
saxpy<<<(N + DIM) / DIM, DIM>>>(N, 2.0f, d_x, d_y);
cudaDeviceSynchronize();
}
cudaMemcpy(y, d_y, N * sizeof(float), cudaMemcpyDeviceToHost);
cudaFree(d_x);
cudaFree(d_y);
free(x);
free(y);
double cputime = ((double)(clock() - start) / CLOCKS_PER_SEC);
printf("cpu time is %f\n", cputime);
return 0;
}
If I do that, then the execution time becomes 1.36 seconds. Doing sth similar to the PyCUDA code I got about 19s of execution time.

Related

Problem with response times in CUDA why cpu is faster than gpu?

Hi i have a problem with the time response I am getting a longer response time on GPU than CPU
the algorithm used is a matrix multiplication algorithm
using the next functions:
// Start timers
cudaEvent_t timer1, timer2;
cudaEventCreate(&timer1);
cudaEventCreate(&timer2);
cudaEventRecord(timer1, 0);
cudaEventSynchronize(timer1);
// Stop timers
cudaEventRecord(timer2, 0);
cudaEventSynchronize(timer1);
cudaEventSynchronize(timer2);
float elapsed;
cudaEventElapsedTime(&elapsed, timer1, timer2);
cudaDeviceReset();
return elapsed;
here is my code on GPU:
float Mult_gpu(float* hostPtr, float* hostPtr2, float* hostPtr3, int size, int Ncols, int Nrows, int n) {
size_t pitch;
check("Creating timers");
cudaEvent_t timer1, timer2;
cudaEventCreate(&timer1);
cudaEventCreate(&timer2);
cudaEventRecord(timer1, 0);
cudaEventSynchronize(timer1);
/******************************************/
/***Configuracion de las matrices en gpu***/
/******************************************/
float* devPtr;
cudaMallocPitch(&devPtr, &pitch, n * sizeof(float), Nrows);
cudaMemcpy2D(devPtr, pitch, hostPtr, n * sizeof(float), n * sizeof(float), Nrows, cudaMemcpyHostToDevice);
float* devPtr2;
cudaMallocPitch(&devPtr2, &pitch, Ncols * sizeof(float), n);
cudaMemcpy2D(devPtr2, pitch, hostPtr2, Ncols * sizeof(float), Ncols * sizeof(float), n, cudaMemcpyHostToDevice);
float* devPtr3;
cudaMallocPitch(&devPtr3, &pitch, Ncols * sizeof(float), Nrows);
//dim3 gridSize(iDivUp(Ncols3, BLOCKSIZE_x), iDivUp(Nrows3, BLOCKSIZE_y));
//dim3 blockSize(BLOCKSIZE_y, BLOCKSIZE_x);
dim3 block(32, 32); //hilos por bloque
dim3 grid((size / block.x) + 1, (size / block.y) + 1); //numero de bloques
/**************************/
/**Lanzamiento del kernel**/
/**************************/
Mult << <grid, block >> > (devPtr, devPtr2, devPtr3, pitch, Ncols, Nrows, n);
cudaDeviceSynchronize();
/*********************************/
/***Copiado de devPtr a hosPtr2***/
/*********************************/
cudaMemcpy2D(hostPtr3, Ncols * sizeof(float), devPtr3, pitch, Ncols * sizeof(float), Nrows, cudaMemcpyDeviceToHost);
//cudaMemcpy(hostPtr3, devPtr3, size * sizeof(float), cudaMemcpyDeviceToHost);
cudaFree(devPtr);
cudaFree(devPtr2);
cudaFree(devPtr3);
// Stop timers
cudaEventRecord(timer2, 0);
cudaEventSynchronize(timer1);
cudaEventSynchronize(timer2);
float elapsed;
cudaEventElapsedTime(&elapsed, timer1, timer2);
cudaDeviceReset();
return elapsed;
}
and here is my code on CPU
float Mult_cpu(float* hostPtrA, float* HostPtrB, float* hostPtrC, int Ncols, int Nrows, int n)
{
cudaEvent_t timer1, timer2;
cudaEventCreate(&timer1);
cudaEventCreate(&timer2);
cudaEventRecord(timer1, 0);
cudaEventSynchronize(timer1);
for (int i = 0; i < Nrows; ++i) {
for (int j = 0; j < Ncols; ++j) {
float suma = 0;
for (int k = 0; k < n; ++k) {
suma += hostPtrA[i * n + k] * HostPtrB[k * Ncols + j];
}
hostPtrC[i * Ncols + j] = suma;
}
}
// Stop timers
cudaEventRecord(timer2, 0);
cudaEventSynchronize(timer1);
cudaEventSynchronize(timer2);
float elapsed;
cudaEventElapsedTime(&elapsed, timer1, timer2);
return elapsed;
}
when i use a matrix 500x500 or any matrix the CPU is faster than GPU and i don't understand why i don't know if the problem is my kernel program or the CUDA functions that im using
my kernel code
__global__ void Mult(float* devPtrA, float* devPtrB, float* devPtrC, size_t pitch, int Ncols, int Nrows, int n)
{
float temp;
int r = blockDim.y * blockIdx.y + threadIdx.y; //for (int f = 0; f <= fil - 1; f += 1) equivalencia en for
int c = blockDim.x * blockIdx.x + threadIdx.x; //for (int c = 0; c <= col - 1; c += 1)
if ((r < Ncols) && (c < Nrows)) {
for (int c2 = 0; c2 < n; c2++) {
float* vertical = (float*)((char*)devPtrA + r * pitch);
float element1 = vertical[c2];
float* horizontal = (float*)((char*)devPtrB + c2 * pitch);
float element2 = horizontal[c];
temp += element1 * element2;
}
//printf("\nla fila es: %d la columna es: %d el valor es: %8.4f\n\n", r, c, temp);
float* vertical2 = (float*)((char*)devPtrC + r * pitch);
vertical2[c] = temp;
}
}
You should read on the concept of SIMT architecture, CUDA execution model and branch divergence. Analyze your CUDA kernel performance with a profiler. I suspect that the condition if ((r < Ncols) && (c < Nrows)) in your kernel causes threads in each warp to diverge and hence the reduced performance. Also pitch affects the global memory access pattern in your code which is another factor in the performance of CUDA kernels. Some other excellent optimization tips can be found here.
CudaMalloc is really slow. If you know the size of your matrices beforehand, do the initialization at the beginning of your program.

Dot product in Cuda by example does not work for me

I'm starting to read "Cuda By Example" Book and I've been a problem with the dot example using "shared memory". I copy-paste the example from the book and I set: N = x * 1024; threadsPerBlock = 32; blocksPerGrid = 8. Where I test the "x" values with 2, 3, 4, 5.
If I set x = 3, the result is bad, but when I used x = 2,4,5 all is ok. I don't understand where is the problem. The code is:
#include "cuda_runtime.h"
#include "device_launch_parameters.h"
#include <stdio.h>
#define imin(a, b) (a<b?a:b)
#define sum_squares(x) (x*(x+1)*(2*x+1)/6)
const int x = 3;
const int N = 3 * 1024;
const int threadsPerBlock = 32;
const int blocksPerGrid = 8;
__global__ void dot(float *a, float *b, float *c)
{
__shared__ float cache[threadsPerBlock];
int tid = threadIdx.x + blockIdx.x * blockDim.x;
int cacheIndex = threadIdx.x;
float temp = 0;
while (tid < N)
{
temp += a[tid] * b[tid];
tid += blockDim.x * gridDim.x;
}
cache[cacheIndex] = temp;
__syncthreads();
int i = blockDim.x / 2;
while (i != 0)
{
if (cacheIndex < i)
cache[cacheIndex] += cache[cacheIndex + i];
__syncthreads();
i /= 2;
}
if (cacheIndex == 0)
c[blockIdx.x] = cache[0];
}
int main()
{
float *a, *b, *partial_c, result;
float *d_a, *d_b, *d_partial_c;
a = (float *)malloc(N * sizeof(float));
b = (float *)malloc(N * sizeof(float));
partial_c = (float *)malloc(blocksPerGrid * sizeof(float));
cudaMalloc((void **)&d_a, N * sizeof(float));
cudaMalloc((void **)&d_b, N * sizeof(float));
cudaMalloc((void **)&d_partial_c, blocksPerGrid * sizeof(float));
for (int i = 0; i < N; i++)
{
a[i] = i;
b[i] = 2 * i;
}
cudaMemcpy(d_a, a, N * sizeof(float), cudaMemcpyHostToDevice);
cudaMemcpy(d_b, b, N * sizeof(float), cudaMemcpyHostToDevice);
dot << <blocksPerGrid, threadsPerBlock >> >(d_a, d_b, d_partial_c);
cudaMemcpy(partial_c, d_partial_c, blocksPerGrid * sizeof(float), cudaMemcpyDeviceToHost);
result = 0;
for (int i = 0; i < blocksPerGrid; i++)
result += partial_c[i];
if (2 * sum_squares((float)(N - 1)) == result)
printf(":)\n");
else
printf(":(\n");
cudaFree(d_a);
cudaFree(d_b);
cudaFree(d_partial_c);
free(a);
free(b);
free(partial_c);
getchar();
return 0;
}
Because float does not have enough precision, which is ~7 decimal digits only. But for x=3; your expected result is
19317916672
containing 11 digits.
for x=4,5, the results are bad on my machine too.

cuda calc distance of two points

Here I want to calculate the distance of each two points, and decide if they are neighbours. here is my simple code in cuda.
__global__ void calcNeighbors(const DataPoint* points,
const float doubleRadius, bool* neighbors) {
int tid = threadIdx.x + blockIdx.x * blockDim.x;
float dis = 0.0f;
while (tid < N) {
DataPoint p1 = points[tid];
for (int i=0; i<N; i++) {
DataPoint p2 = points[i];
dis = 0;
dis += (p1.pfDimens[0]-p2.pfDimens[0]) * (p1.pfDimens[0]-p2.pfDimens[0]) +
(p1.pfDimens[1]-p2.pfDimens[1]) * (p1.pfDimens[1]-p2.pfDimens[1]) +
(p1.pfDimens[2]-p2.pfDimens[2]) * (p1.pfDimens[2]-p2.pfDimens[2]);
if (dis <= doubleRadius) {
neighbors[tid*N+i] = true;
} else {
neighbors[tid*N+i] = false;
}
}
tid += blockDim.x * gridDim.x;
}
}
The DataPoint is a struct is
typedef struct DataPoint {
float pfDimens[3];
} DataPoint;
so here i want to reduce the time, How can i do? I have tried to use memory coalesing and share memory, but i didn't get a good speed up?
===============use share memory==============
__global__ void calcNeighbors2(const DataPoint* points,
const float doubleRadius, bool* neighbors) {
__shared__ DataPoint sharedpoints[threadsPerBlock];
int start = blockIdx.x * blockDim.x;
int len = start+threadIdx.x;
if (len < N) {
sharedpoints[threadIdx.x] = points[len];
}
len = imin(N, blockDim.x + start);
__syncthreads();
int tid = threadIdx.x;
float dis;
while (tid < N) {
DataPoint p1 = points[tid];
for (int i=start; i<len; i++) {
dis = 0;
dis += (p1.pfDimens[0]-sharedpoints[i-start].pfDimens[0]) * (p1.pfDimens[0]-sharedpoints[i-start].pfDimens[0]) +
(p1.pfDimens[1]-sharedpoints[i-start].pfDimens[1]) * (p1.pfDimens[1]-sharedpoints[i-start].pfDimens[1]) +
(p1.pfDimens[2]-sharedpoints[i-start].pfDimens[2]) * (p1.pfDimens[2]-sharedpoints[i-start].pfDimens[2]);
if (dis <= doubleRadius) {
neighbors[i*N+tid] = true;
} else {
neighbors[i*N+tid] = false;
}
}
tid += blockDim.x;
}
}
Here i changed the neighbors[tid*N+i] to neighbors[i*N+tid], it give me amlost 8x speed up on Tesla K10.G2.8GB. But when i use share memory to store some points, it is no use?
There are at least 4 ideas, some of which have already been stated in the comments:
Transform your point distance storage from AoS format:
struct DataPoint {
float pfDimens[3];
};
to SoA format:
struct DataPoint {
float pfDimens_x[NPTS];
float pfDimens_y[NPTS];
float pfDimens_z[NPTS];
};
this will enable full coalescing on loading of the data. In fact, to help with point 4 below, I would just switch to using 3 bare arrays, rather than a structure.
reduce the computation to (slightly less than) half:
for (int i=N-1; i>tid; i--) {
then, either in the thread code itself, or in the host, you can populate the other "half" of the output matrix by copying data.
Transpose the storage in your output matrix, so that you can write a storage operation like this:
neighbors[i*N+tid] = true;
which will nicely coalesce, as opposed to this:
neighbors[tid*N+i] = true;
which will not.
Since your input point data is read only, mark the kernel parameter appropriately:
const float * __restrict__ points_x, const float * __restrict__ points_y, const float * __restrict__ points_z
in some cases, and on some GPUs, this will often lead to a speed-up due to use of the read-only cache. If you really want to get aggressive with caching, and your data array is small enough (4K or less float points), you could put a copy of the point data in global memory as well as a copy in __constant__ memory, and load the "uniform" load you are doing here through constant memory:
DataPoint p2 = c_points[i];
thus you could perform the coalesced load through the read-only cache, the uniform load through the constant cache, and the coalesced store going to ordinary global memory.
On a K40c, on linux/CUDA 7, for N = 4096, the net effect of these changes appears to be about a 3.5x speedup, at the kernel level:
$ cat t749.cu
#include <stdio.h>
#define N 4096
// if N is 16K/3 or less, we can use constant
#define USE_CONSTANT
#define THRESH 0.2f
#define nTPB 256
#define nBLK (N/nTPB+1)
#define cudaCheckErrors(msg) \
do { \
cudaError_t __err = cudaGetLastError(); \
if (__err != cudaSuccess) { \
fprintf(stderr, "Fatal error: %s (%s at %s:%d)\n", \
msg, cudaGetErrorString(__err), \
__FILE__, __LINE__); \
fprintf(stderr, "*** FAILED - ABORTING\n"); \
exit(1); \
} \
} while (0)
#include <time.h>
#include <sys/time.h>
#define USECPSEC 1000000ULL
unsigned long long dtime_usec(unsigned long long start){
timeval tv;
gettimeofday(&tv, 0);
return ((tv.tv_sec*USECPSEC)+tv.tv_usec)-start;
}
struct DataPoint {
float pfDimens[3];
};
__global__ void calcNeighbors(const DataPoint* points,
const float doubleRadius, bool* neighbors) {
int tid = threadIdx.x + blockIdx.x * blockDim.x;
float dis = 0.0f;
while (tid < N) {
DataPoint p1 = points[tid];
for (int i=0; i<N; i++) {
DataPoint p2 = points[i];
dis = 0;
dis += (p1.pfDimens[0]-p2.pfDimens[0]) * (p1.pfDimens[0]-p2.pfDimens[0]) +
(p1.pfDimens[1]-p2.pfDimens[1]) * (p1.pfDimens[1]-p2.pfDimens[1]) +
(p1.pfDimens[2]-p2.pfDimens[2]) * (p1.pfDimens[2]-p2.pfDimens[2]);
if (dis <= doubleRadius) {
neighbors[tid*N+i] = true;
} else {
neighbors[tid*N+i] = false;
}
}
tid += blockDim.x * gridDim.x;
}
}
#ifdef USE_CONSTANT
__constant__ float cpx[N];
__constant__ float cpy[N];
__constant__ float cpz[N];
#endif
__global__ void calcNeighbors2(const float * __restrict__ pts_x, const float * __restrict__ pts_y, const float * __restrict__ pts_z, const float doubleRadius, bool * __restrict__ neighbors) {
int tid = threadIdx.x+blockDim.x*blockIdx.x;
while (tid < N) {
float p1x = pts_x[tid];
float p1y = pts_y[tid];
float p1z = pts_z[tid];
for (int i = N-1; i > tid; i--){
float p2x, p2y, p2z;
#ifdef USE_CONSTANT
p2x = cpx[i];
p2y = cpy[i];
p2z = cpz[i];
#else
p2x = pts_x[i];
p2y = pts_y[i];
p2z = pts_z[i];
#endif
float dis = ((p1x-p2x)*(p1x-p2x)) + ((p1y-p2y)*(p1y-p2y)) + ((p1z-p2z)*(p1z-p2z));
neighbors[i*N+tid] = (dis <= doubleRadius);
}
tid += blockDim.x * gridDim.x;
}
}
int main(){
float *dx, *dy, *dz, *hx, *hy, *hz;
DataPoint *dp, *hp;
bool *dn, *hn1, *hn2;
hx =(float *)malloc(N*sizeof(float));
hy =(float *)malloc(N*sizeof(float));
hz =(float *)malloc(N*sizeof(float));
hp =(DataPoint *)malloc(N*sizeof(DataPoint));
hn1=(bool *)malloc(N*N*sizeof(bool));
hn2=(bool *)malloc(N*N*sizeof(bool));
cudaMalloc(&dx, N*sizeof(float));
cudaMalloc(&dy, N*sizeof(float));
cudaMalloc(&dz, N*sizeof(float));
cudaMalloc(&dp, N*sizeof(DataPoint));
cudaMalloc(&dn, N*N*sizeof(bool));
for (int i =0; i < N; i++){
hx[i] = rand()/(float)RAND_MAX;
hy[i] = rand()/(float)RAND_MAX;
hz[i] = rand()/(float)RAND_MAX;
hp[i].pfDimens[0] = hx[i];
hp[i].pfDimens[1] = hy[i];
hp[i].pfDimens[2] = hz[i];}
cudaMemcpy(dx, hx, N*sizeof(float), cudaMemcpyHostToDevice);
cudaMemcpy(dy, hy, N*sizeof(float), cudaMemcpyHostToDevice);
cudaMemcpy(dz, hz, N*sizeof(float), cudaMemcpyHostToDevice);
cudaMemcpy(dp, hp, N*sizeof(DataPoint), cudaMemcpyHostToDevice);
// warm-up
calcNeighbors<<<nBLK, nTPB>>>(dp, THRESH, dn);
cudaDeviceSynchronize();
cudaMemset(dn, 0, N*N*sizeof(bool));
unsigned long long t1 = dtime_usec(0);
calcNeighbors<<<nBLK, nTPB>>>(dp, THRESH, dn);
cudaDeviceSynchronize();
cudaCheckErrors("kernel 1 error");
t1 = dtime_usec(t1);
cudaMemcpy(hn1, dn, N*N*sizeof(bool), cudaMemcpyDeviceToHost);
// warm-up
calcNeighbors2<<<nBLK, nTPB>>>(dx, dy, dz, THRESH, dn);
cudaDeviceSynchronize();
cudaMemset(dn, 0, N*N*sizeof(bool));
unsigned long long t2 = dtime_usec(0);
calcNeighbors2<<<nBLK, nTPB>>>(dx, dy, dz, THRESH, dn);
cudaDeviceSynchronize();
cudaCheckErrors("kernel 2 error");
t2 = dtime_usec(t2);
cudaMemcpy(hn2, dn, N*N*sizeof(bool), cudaMemcpyDeviceToHost);
cudaCheckErrors("some error");
printf("t1: %fs, t2: %fs\n", t1/(float)USECPSEC, t2/(float)USECPSEC);
// results validation
for (int i = 0; i < N; i++)
for (int j = i+1; j < N; j++)
if (hn1[i*N+j] != hn2[j*N+i]) {printf("mismatch at %d, %d, was: %d, should be: %d\n", i, j, hn2[j*N+i], hn1[i*N+j]); return 1;}
return 0;
}
$ nvcc -arch=sm_35 -o t749 t749.cu
$ ./t749
t1: 0.004903s, t2: 0.001395s
$
In the case of K40c, the limited number of blocks being launched above (16) is a significant impediment to performance, due to latency. If we comment out the USE_CONSTANT define, and change N to 16384, we observe an even higher speedup with the improved kernel:
$ ./t749
t1: 0.267107s, t2: 0.008209s
$
the resultant ~48 blocks being enough to approximately "fill" the K40c which has 15 SMs.
EDIT: now that you've posted a shared memory kernel, I added it to my test case as calcNeighbors3 and compared it's timing performance (as t3). It is almost as fast as my kernel, and it seems to provide the correct result (matches your original kernel) so I'm not sure what your concerns are.
Here's the updated code and test case:
$ cat t749.cu
#include <stdio.h>
#include <math.h>
#define imin(X,Y) ((X)<(Y))?(X):(Y)
#define N 32768
// if N is 16K/3 or less, we can use constant
// #define USE_CONSTANT
#define THRESH 0.2f
#define nTPB 256
#define nBLK (N/nTPB+1)
#define cudaCheckErrors(msg) \
do { \
cudaError_t __err = cudaGetLastError(); \
if (__err != cudaSuccess) { \
fprintf(stderr, "Fatal error: %s (%s at %s:%d)\n", \
msg, cudaGetErrorString(__err), \
__FILE__, __LINE__); \
fprintf(stderr, "*** FAILED - ABORTING\n"); \
exit(1); \
} \
} while (0)
#include <time.h>
#include <sys/time.h>
#define USECPSEC 1000000ULL
unsigned long long dtime_usec(unsigned long long start){
timeval tv;
gettimeofday(&tv, 0);
return ((tv.tv_sec*USECPSEC)+tv.tv_usec)-start;
}
struct DataPoint {
float pfDimens[3];
};
__global__ void calcNeighbors(const DataPoint* points,
const float doubleRadius, bool* neighbors) {
int tid = threadIdx.x + blockIdx.x * blockDim.x;
float dis = 0.0f;
while (tid < N) {
DataPoint p1 = points[tid];
for (int i=0; i<N; i++) {
DataPoint p2 = points[i];
dis = 0;
dis += (p1.pfDimens[0]-p2.pfDimens[0]) * (p1.pfDimens[0]-p2.pfDimens[0]) +
(p1.pfDimens[1]-p2.pfDimens[1]) * (p1.pfDimens[1]-p2.pfDimens[1]) +
(p1.pfDimens[2]-p2.pfDimens[2]) * (p1.pfDimens[2]-p2.pfDimens[2]);
if (dis <= doubleRadius) {
neighbors[tid*N+i] = true;
} else {
neighbors[tid*N+i] = false;
}
}
tid += blockDim.x * gridDim.x;
}
}
#ifdef USE_CONSTANT
__constant__ float cpx[N];
__constant__ float cpy[N];
__constant__ float cpz[N];
#endif
__global__ void calcNeighbors2(const float * __restrict__ pts_x, const float * __restrict__ pts_y, const float * __restrict__ pts_z, const float doubleRadius, bool * __restrict__ neighbors) {
int tid = threadIdx.x+blockDim.x*blockIdx.x;
while (tid < N) {
float p1x = pts_x[tid];
float p1y = pts_y[tid];
float p1z = pts_z[tid];
for (int i = N-1; i > tid; i--){
float p2x, p2y, p2z;
#ifdef USE_CONSTANT
p2x = cpx[i];
p2y = cpy[i];
p2z = cpz[i];
#else
p2x = pts_x[i];
p2y = pts_y[i];
p2z = pts_z[i];
#endif
float dis = ((p1x-p2x)*(p1x-p2x)) + ((p1y-p2y)*(p1y-p2y)) + ((p1z-p2z)*(p1z-p2z));
neighbors[i*N+tid] = (dis <= doubleRadius);
}
tid += blockDim.x * gridDim.x;
}
}
__global__ void calcNeighbors3(const DataPoint* points,
const float doubleRadius, bool* neighbors) {
__shared__ DataPoint sharedpoints[nTPB];
int start = blockIdx.x * blockDim.x;
int len = start+threadIdx.x;
if (len < N) {
sharedpoints[threadIdx.x] = points[len];
}
len = imin(N, blockDim.x + start);
__syncthreads();
int tid = threadIdx.x;
float dis;
while (tid < N) {
DataPoint p1 = points[tid];
for (int i=start; i<len; i++) {
dis = 0;
dis += (p1.pfDimens[0]-sharedpoints[i-start].pfDimens[0]) * (p1.pfDimens[0]-sharedpoints[i-start].pfDimens[0]) +
(p1.pfDimens[1]-sharedpoints[i-start].pfDimens[1]) * (p1.pfDimens[1]-sharedpoints[i-start].pfDimens[1]) +
(p1.pfDimens[2]-sharedpoints[i-start].pfDimens[2]) * (p1.pfDimens[2]-sharedpoints[i-start].pfDimens[2]);
if (dis <= doubleRadius) {
neighbors[i*N+tid] = true;
} else {
neighbors[i*N+tid] = false;
}
}
tid += blockDim.x;
}
}
int main(){
float *dx, *dy, *dz, *hx, *hy, *hz;
DataPoint *dp, *hp;
bool *dn, *hn1, *hn2, *hn3;
hx =(float *)malloc(N*sizeof(float));
hy =(float *)malloc(N*sizeof(float));
hz =(float *)malloc(N*sizeof(float));
hp =(DataPoint *)malloc(N*sizeof(DataPoint));
hn1=(bool *)malloc(N*N*sizeof(bool));
hn2=(bool *)malloc(N*N*sizeof(bool));
hn3=(bool *)malloc(N*N*sizeof(bool));
cudaMalloc(&dx, N*sizeof(float));
cudaMalloc(&dy, N*sizeof(float));
cudaMalloc(&dz, N*sizeof(float));
cudaMalloc(&dp, N*sizeof(DataPoint));
cudaMalloc(&dn, N*N*sizeof(bool));
for (int i =0; i < N; i++){
hx[i] = rand()/(float)RAND_MAX;
hy[i] = rand()/(float)RAND_MAX;
hz[i] = rand()/(float)RAND_MAX;
hp[i].pfDimens[0] = hx[i];
hp[i].pfDimens[1] = hy[i];
hp[i].pfDimens[2] = hz[i];}
cudaMemcpy(dx, hx, N*sizeof(float), cudaMemcpyHostToDevice);
cudaMemcpy(dy, hy, N*sizeof(float), cudaMemcpyHostToDevice);
cudaMemcpy(dz, hz, N*sizeof(float), cudaMemcpyHostToDevice);
cudaMemcpy(dp, hp, N*sizeof(DataPoint), cudaMemcpyHostToDevice);
#ifdef USE_CONSTANT
cudaMemcpyToSymbol(cpx, hx, N*sizeof(float));
cudaMemcpyToSymbol(cpy, hy, N*sizeof(float));
cudaMemcpyToSymbol(cpz, hz, N*sizeof(float));
#endif
// warm-up
calcNeighbors<<<nBLK, nTPB>>>(dp, THRESH, dn);
cudaDeviceSynchronize();
cudaMemset(dn, 0, N*N*sizeof(bool));
unsigned long long t1 = dtime_usec(0);
calcNeighbors<<<nBLK, nTPB>>>(dp, THRESH, dn);
cudaDeviceSynchronize();
cudaCheckErrors("kernel 1 error");
t1 = dtime_usec(t1);
cudaMemcpy(hn1, dn, N*N*sizeof(bool), cudaMemcpyDeviceToHost);
// warm-up
calcNeighbors2<<<nBLK, nTPB>>>(dx, dy, dz, THRESH, dn);
cudaDeviceSynchronize();
cudaMemset(dn, 0, N*N*sizeof(bool));
unsigned long long t2 = dtime_usec(0);
calcNeighbors2<<<nBLK, nTPB>>>(dx, dy, dz, THRESH, dn);
cudaDeviceSynchronize();
cudaCheckErrors("kernel 2 error");
t2 = dtime_usec(t2);
cudaMemcpy(hn2, dn, N*N*sizeof(bool), cudaMemcpyDeviceToHost);
// warm-up
calcNeighbors3<<<nBLK, nTPB>>>(dp, THRESH, dn);
cudaDeviceSynchronize();
cudaMemset(dn, 0, N*N*sizeof(bool));
unsigned long long t3 = dtime_usec(0);
calcNeighbors3<<<nBLK, nTPB>>>(dp, THRESH, dn);
cudaDeviceSynchronize();
cudaCheckErrors("kernel 3 error");
t3 = dtime_usec(t3);
cudaMemcpy(hn3, dn, N*N*sizeof(bool), cudaMemcpyDeviceToHost);
cudaCheckErrors("some error");
printf("t1: %fs, t2: %fs, t3: %fs\n", t1/(float)USECPSEC, t2/(float)USECPSEC, t3/(float)USECPSEC);
// results validation
for (int i = 0; i < N; i++)
for (int j = i+1; j < N; j++)
if (hn1[i*N+j] != hn2[j*N+i]) {printf("1:2 mismatch at %d, %d, was: %d, should be: %d\n", i, j, hn2[j*N+i], hn1[i*N+j]); return 1;}
for (int i = 0; i < N*N; i++)
if (hn1[i] != hn3[i]) {printf("1:3 mismatch at %d, was: %d, should be: %d\n", i, hn1[i], hn3[i]); return 1;}
return 0;
}
$ nvcc -arch=sm_35 -o t749 t749.cu
$ ./t749
t1: 1.260010s, t2: 0.022661s, t3: 0.029632s
$
For this test, I have changed the data set size to 32768 since that is closer to the range you care about. Your shared memory kernel shows about a 42x speedup over your original kernel, and my kernel shows about a 55x speedup, on my K40c.

Matrix Multiplication giving wrong output [duplicate]

This question already has an answer here:
Unable to execute device kernel in CUDA
(1 answer)
Closed 7 years ago.
What I am attempting to do is Multiply Matrix A & Matrix B and then from the product matrix I get the index of the maximum value per column. But unfortunately, only the first 128*128 values of the matrix multiplication are correct while others are just garbage. I do not quite understand how this works. I request you to kindly guide me with this ..
#include<stdio.h>
#include "cuda.h"
#include<stdlib.h>
#define blockD 32
const int wA = 128;
const int hA = 4096;
const int wB = 4096;
const int hB = wA;
main(void){
void MatrixMultiplication(float *, float *, float *, float *);
int size_A = wA * hA * sizeof(float);
int size_B = wB * hB * sizeof(float);
int size_C = wB * hA * sizeof(float);
int size_max = 2 * wB * sizeof(float);
float *M, *N, *P, *C;
// allocate memory on the CPU
M = (float*)malloc(size_A);
N = (float*)malloc(size_B);
P = (float*)malloc(size_max);
C = (float*)malloc(size_C);
// initialize the matrices
for (int y=0; y < hA; y++) {
for (int x=0; x < wA; x++){
M[y*wA + x] = 32; //x + y*wA;
}
}
for (int y=0; y<hB; y++) {
for (int x=0; x<wB; x++){
N[y*wB + x] = 21; //x + y*wB;
}
}
MatrixMultiplication(M, N, P, C);
//Write
FILE *f1;
int i,j;
f1 = fopen("C.txt","w");
for(i = hA - 2 ; i < hA; i ++){
for(j = 0; j < wB; j++){
fprintf(f1,"%d\t",int(C[i*wB + j]));
}
fprintf(f1,"\n");
}
fclose(f1);
// free the memory allocated on the CPU
free( M );
free( N );
free( P );
free( C );
cudaDeviceReset();
return 0;
}
__device__ void MaxFunction(float* Pd, float* max)
{
int x = (threadIdx.x + blockIdx.x * blockDim.x);
int y = (threadIdx.y + blockIdx.y * blockDim.y);
int k = 0;
int temp = 0; int temp_idx = 0;
for (k = 0; k < wB; ++k) {
if(Pd[x*wB + k] > temp){
temp = Pd[x*wB + k];
temp_idx = x*wB + k;
}
}
max[y*2 + 0] = temp;
max[y*2 + 1] = temp_idx;
}
__global__ void MatrixMulKernel(float* Md, float* Nd, float* Pd, float* max)
{
// declare cache in the shared memory
__shared__ float Mds[blockD][blockD];
__shared__ float Nds[blockD][blockD];
float Pvalue = 0;
// Loop over the Md and Nd block dimension required to compute the Pd element
for (int m = (wA * blockD * blockIdx.y), n = (blockD * blockIdx.x);
m < ((wA * blockD * blockIdx.y)+wA-1);
m += blockD, n += (blockD*hB)){
// collaboratively loading of Md and Nd blocks into shared memory
Mds[threadIdx.y][threadIdx.x] = Md[m + wA * threadIdx.y + threadIdx.x];
Nds[threadIdx.y][threadIdx.x] = Nd[n + wA * threadIdx.y + threadIdx.x];
__syncthreads();
// keep track of the running sum
for (int k = 0; k < blockD; k++)
Pvalue += Mds[threadIdx.y][k] * Nds[k][threadIdx.x];
__syncthreads();
}
// write back to the global memory
int p = hB * blockD * blockIdx.y + blockD * blockIdx.x;
Pd[p + hB * threadIdx.y + threadIdx.x] = Pvalue;
__syncthreads();
MaxFunction(Pd, max);
}
void MatrixMultiplication(float *M, float *N, float *P, float *C) {
int size_A = wA * hA * sizeof(float);
int size_B = wB * hB * sizeof(float);
int size_C = wB * hA * sizeof(float);
int size_max = 2 * wB * sizeof(float);
float *Md, *Nd, *Pd, *max;
// allocate memory on the GPU
cudaMalloc((void**)&Md, size_A);
cudaMalloc((void**)&Nd, size_B);
cudaMalloc((void**)&Pd, size_C);
cudaMalloc((void**)&max, size_max);
// transfer M and N to device memory
cudaMemcpy(Md, M, size_A, cudaMemcpyHostToDevice);
cudaMemcpy(Nd, N, size_B, cudaMemcpyHostToDevice);
// kernel invocation code
dim3 dimBlock(blockD, blockD);
dim3 dimGrid(wA/blockD, hB/blockD);
//Execute Kernel
MatrixMulKernel<<<dimGrid, dimBlock>>>( Md, Nd, Pd, max);
// transfer P from device
cudaMemcpy(P, max, size_max, cudaMemcpyDeviceToHost);
cudaMemcpy(C, Pd, size_C, cudaMemcpyDeviceToHost);
// free the memory allocated on the GPU
cudaFree(Md);
cudaFree(Nd);
cudaFree(Pd);
cudaFree(max);
}
In your code you seem to have more than one problem. One of the problems is, in place of this:
dim3 dimGrid(wA/blockD, hB/blockD);
You should have this:
dim3 dimGrid(wB/blockD, hA/blockD);
Ultimately you need one thread in your grid for each output point. Your formulation was giving you a grid of 4 blocks by 4 blocks, whereas you need a grid of 128 blocks by 128 blocks.
The other problem I found with your code was in these lines in the kernel:
int p = hB * blockD * blockIdx.y + blockD * blockIdx.x;
Pd[p + hB * threadIdx.y + threadIdx.x] = Pvalue;
They are not indexing properly through the output array. Rather than try to sort it out using your scheme, I used this instead:
Pd[(threadIdx.x + (blockIdx.x * blockDim.x)) + ((threadIdx.y + (blockIdx.y * blockDim.y))*(gridDim.x*blockDim.x))] = Pvalue;
When I made the above two changes to your code, I got what I believe are correct results throughout the array. And it took about 32 seconds on my machine to run it. (Note that I haven't tried fixing your original max-finding code -- see below for a better approach.)
Based on your previous question, you seemed to be concerned about speed. If you want to do fast matrix multiply, you should use cublas. The following code shows how to use cublas to multiply two ordinary C-style matrices (they don't have to be square). I've also included a column-max finding kernel that will be fast when the number of columns is large (say, over 500 or so. You have 4096 columns in your example). For small numbers of columns, there may be quicker ways to perform this function, but small numbers of columns also suggests that the overall problem size may be small and so speed (of this piece of code) will not really be an issue.
Here's the code:
#include <stdio.h>
#include <cublas_v2.h>
#define VERBOSE 1
#define nTPB 64
#define ROW_A 4
#define COL_A 4
#define ROW_B COL_A
#define COL_B 4
#define ROW_C ROW_A
#define COL_C COL_B
#define SIZ_A (ROW_A*COL_A)
#define SIZ_B (ROW_B*COL_B)
#define SIZ_C (ROW_C*COL_C)
// error check macros
#define cudaCheckErrors(msg) \
do { \
cudaError_t __err = cudaGetLastError(); \
if (__err != cudaSuccess) { \
fprintf(stderr, "Fatal error: %s (%s at %s:%d)\n", \
msg, cudaGetErrorString(__err), \
__FILE__, __LINE__); \
fprintf(stderr, "*** FAILED - ABORTING\n"); \
exit(1); \
} \
} while (0)
// for CUBLAS V2 API
#define cublasCheckErrors(fn) \
do { \
cublasStatus_t __err = fn; \
if (__err != CUBLAS_STATUS_SUCCESS) { \
fprintf(stderr, "Fatal cublas error: %d (at %s:%d)\n", \
(int)(__err), \
__FILE__, __LINE__); \
fprintf(stderr, "*** FAILED - ABORTING\n"); \
exit(1); \
} \
} while (0)
__global__ void col_max(float *mat, float *max, unsigned int *midx, unsigned int rows, unsigned int cols){
int idx = threadIdx.x + blockDim.x*blockIdx.x;
if (idx < cols){
float tempmax = mat[idx];
unsigned int tempmidx = 0;
for (int i = 1; i< rows; i++)
if (mat[idx + (i*cols)] > tempmax){
tempmax = mat[idx + (i*cols)];
tempmidx = i;}
max[idx] = tempmax;
midx[idx] = tempmidx;
}
}
int main(){
float *h_A, *h_B, *h_C, *d_A, *d_B, *d_C, *h_max, *d_max;
unsigned int *h_idx, *d_idx;
h_A = (float *)malloc(SIZ_A*sizeof(float));
if (h_A==0) {printf("malloc fail\n"); return -1;}
h_B = (float *)malloc(SIZ_B*sizeof(float));
if (h_B==0) {printf("malloc fail\n"); return -1;}
h_C = (float *)malloc(SIZ_C*sizeof(float));
if (h_C==0) {printf("malloc fail\n"); return -1;}
h_max = (float *)malloc(COL_C*sizeof(float));
if (h_max==0) {printf("malloc fail\n"); return -1;}
h_idx = (unsigned int*)malloc(COL_C*sizeof(unsigned int));
if (h_idx==0) {printf("malloc fail\n"); return -1;}
cudaMalloc((void **)&d_A, SIZ_A*sizeof(float));
cudaMalloc((void **)&d_B, SIZ_B*sizeof(float));
cudaMalloc((void **)&d_C, SIZ_C*sizeof(float));
cudaMalloc((void **)&d_max, COL_C*sizeof(float));
cudaMalloc((void **)&d_idx, COL_C*sizeof(unsigned int));
cudaCheckErrors("cuda malloc fail");
// initialize data
for (int i=0; i< SIZ_A; i++) h_A[i] = (float)(i+1);
for (int i=0; i< SIZ_B; i++) h_B[i] = (float)(i+2);
cudaMemcpy(d_A, h_A, SIZ_A*sizeof(float), cudaMemcpyHostToDevice);
cudaMemcpy(d_B, h_B, SIZ_B*sizeof(float), cudaMemcpyHostToDevice);
cudaCheckErrors("cuda memcpy 1 fail");
const float alpha = 1.0f;
const float beta = 0.0f;
cublasHandle_t handle;
cublasCheckErrors(cublasCreate(&handle));
// C = A*B
// due to cublas expecting column-major storage, parameters
// are scrambled
cublasCheckErrors(cublasSgemm(handle, CUBLAS_OP_N, CUBLAS_OP_N, COL_B, ROW_A, COL_A, &alpha, d_B, COL_B, d_A, COL_A, &beta, d_C, COL_C));
cudaMemcpy(h_C, d_C, SIZ_C*sizeof(float), cudaMemcpyDeviceToHost);
cudaCheckErrors("cuda memcpy 2 fail");
col_max<<<(COL_C + nTPB - 1)/nTPB, nTPB>>>(d_C, d_max, d_idx, ROW_C, COL_C);
cudaCheckErrors("kernel launch fail");
cudaMemcpy(h_max, d_max, COL_C*sizeof(float), cudaMemcpyDeviceToHost);
cudaMemcpy(h_idx, d_idx, COL_C*sizeof(unsigned int), cudaMemcpyDeviceToHost);
cudaCheckErrors("cuda memcpy 3 fail/kernel fail");
if (VERBOSE){
printf("A: \n");
for (int i=0; i< ROW_A; i++){
for (int j=0; j< COL_A; j++)
printf("%7.5G", h_A[j+(i*COL_A)]);
printf("\n");}
printf("B: \n");
for (int i=0; i< ROW_B; i++){
for (int j=0; j< COL_B; j++)
printf("%7.5G", h_B[j+(i*COL_B)]);
printf("\n");}
printf("C = A*B: \n");
for (int i=0; i< ROW_C; i++){
for (int j=0; j< COL_C; j++)
printf("%7.5G", h_C[j+(i*COL_C)]);
printf("\n");}
printf("COLUMN MAX:\n");
for (int i=0; i< COL_C; i++)
printf("%7.5G", h_max[i]);
printf("\nCOLUMN MAX IDX:\n");
for (int i=0; i< COL_C; i++)
printf("%7d", h_idx[i]);
}
printf("\n finished!\n");
return 0;
}
Here's what I used to compile:
$ nvcc -arch=sm_20 -O3 -o t221 t221.cu -lcublas
And here's the sample output:
$ cuda-memcheck ./t221
========= CUDA-MEMCHECK
A:
1 2 3 4
5 6 7 8
9 10 11 12
13 14 15 16
B:
2 3 4 5
6 7 8 9
10 11 12 13
14 15 16 17
C = A*B:
100 110 120 130
228 254 280 306
356 398 440 482
484 542 600 658
COLUMN MAX:
484 542 600 658
COLUMN MAX IDX:
3 3 3 3
finished!
========= ERROR SUMMARY: 0 errors
$
When I extended my code to handle the same sizes you indicated, (A = 4096x128, B=128x4096) it took about 1 second on my machine. So it's much faster than your code. However, when I take your code and comment out your call to MaxFunction in the kernel, it also only takes about 1 second to compute the matrix multiply result. So if you wanted to keep your matrix multiply code (i.e. not use cublas) you could break the code into 2 kernels, and use your multiply routine in the first kernel with my max-finding routine (col_max) in the second kernel, and also probably get a pretty fast result.
As #talonmies indicated, if you are running on a windows machine, be sure you are aware of the ramifications of windows TDR. (search that in the upper right corner search box if needed)

Unable to execute device kernel in CUDA

I am trying to call a device kernel within a global kernel. My global kernel is a Matrix Multiplication and my device kernel is finding the maximum value and the index in each column of the product matrix. Following is the code :
__device__ void MaxFunction(float* Pd, float* max)
{
int x = (threadIdx.x + blockIdx.x * blockDim.x);
int y = (threadIdx.y + blockIdx.y * blockDim.y);
int k = 0;
int temp = 0; int temp_idx = 0;
for (k = 0; k < wB; ++k) {
if(Pd[x*wB + y] > temp){
temp = Pd[x*wB + y];
temp_idx = x*wB + y;
}
max[y*2 + 0] = temp;
max[y*2 + 1] = temp_idx;
}
}
__global__ void MatrixMulKernel(float* Md, float* Nd, float* Pd, float* max)
{
// declare cache in the shared memory
__shared__ float Mds[blockD][blockD];
__shared__ float Nds[blockD][blockD];
float Pvalue = 0;
// Loop over the Md and Nd block dimension required to compute the Pd element
for (int m = (wA * blockD * blockIdx.y), n = (blockD * blockIdx.x);
m < ((wA * blockD * blockIdx.y)+wA-1);
m += blockD, n += (blockD*hB)){
// collaboratively loading of Md and Nd blocks into shared memory
Mds[threadIdx.y][threadIdx.x] = Md[m + wA * threadIdx.y + threadIdx.x];
Nds[threadIdx.y][threadIdx.x] = Nd[n + wA * threadIdx.y + threadIdx.x];
__syncthreads();
// keep track of the running sum
for (int k = 0; k < blockD; k++)
Pvalue += Mds[threadIdx.y][k] * Nds[k][threadIdx.x];
__syncthreads();
}
// write back to the global memory
int p = hB * blockD * blockIdx.y + blockD * blockIdx.x;
Pd[p + hB * threadIdx.y + threadIdx.x] = Pvalue;
__syncthreads();
MaxFunction(Pd, max);
}
The Main code :
#include<stdio.h>
#include "cuda.h"
#include<stdlib.h>
#define blockD 32
const int wA = 128;
const int hA = 1024;
const int wB = 128;
const int hB = wA;
main(void){
void MatrixMultiplication(float *, float *, float *, float *);
int size_A = wA * hA * sizeof(float);
int size_B = wB * hB * sizeof(float);
int size_C = wB * hA * sizeof(float);
int size_max = 2 * wB * sizeof(float);
float *M, *N, *P, *C;
// allocate memory on the CPU
M = (float*)malloc(size_A);
N = (float*)malloc(size_B);
P = (float*)malloc(size_max);
C = (float*)malloc(size_C);
// initialize the matrices
for (int y=0; y < hA; y++) {
for (int x=0; x < wA; x++){
M[y*wA + x] = x;
}
}
for (int y=0; y<hB; y++) {
for (int x=0; x<wB; x++){
N[y*wB + x] = x;
}
}
MatrixMultiplication(M, N, P, C);
//Write
FILE *f1;
int i, j;
f1 = fopen("max_val.txt","w");
for(i=0; i < (wB * 2); i+=2){
fprintf(f1,"%d\t%d\n",int(P[i]),int(P[i+1]));
}
fclose(f1);
f1 = fopen("Prod_mat.txt","w");
for(i=0; i < 2; i++){
for(j=0; j < wB; j++){
fprintf(f1,"%d\t",int(C[i*wB + j]));
}
fprintf(f1,"\n");
}
fclose(f1);
free( M );
free( N );
free( P );
free( C );
cudaDeviceReset();
return 0;
}
void MatrixMultiplication(float *M, float *N, float *P, float *C) {
int size_A = wA * hA * sizeof(float);
int size_B = wB * hB * sizeof(float);
int size_C = wB * hA * sizeof(float);
int size_max = 2 * wB * sizeof(float);
float *Md, *Nd, *Pd, *max;
// allocate memory on the GPU
cudaMalloc((void**)&Md, size_A);
cudaMalloc((void**)&Nd, size_B);
cudaMalloc((void**)&Pd, size_C);
cudaMalloc((void**)&max, size_max);
// transfer M and N to device memory
cudaMemcpy(Md, M, size_A, cudaMemcpyHostToDevice);
cudaMemcpy(Nd, N, size_B, cudaMemcpyHostToDevice);
// kernel invocation code
dim3 dimBlock(blockD, blockD);
dim3 dimGrid(wA/blockD, hB/blockD);
//Execute Kernel
MatrixMulKernel<<<dimGrid, dimBlock>>>( Md, Nd, Pd, max);
// transfer P from device
cudaMemcpy(P, max, size_max, cudaMemcpyDeviceToHost);
cudaMemcpy(C, Pd, size_C, cudaMemcpyDeviceToHost);
cudaFree(Md);
cudaFree(Nd);
cudaFree(Pd);
cudaFree(max);
}
The Matrix Multiplication result is fine (Verified using Matlab), but I am not able to get the max values and their corresponding index. I would appreciate if anyone can kindly point out at what I am doing wrong. The max variable has only garbage when I run the above code.
Apparently you are attempting to find the maximum value in each column, as well as the offset to that value.
But all of your threads in y are hammering on the same location for max value (max[x*2 + 0]). This isn't recommended, as there is no way to sort out a race condition. You should use atomic operations, or other methods (e.g. reduction) to handle multiple threads updating a single max value this way.
Since you have a need to update two values atomically (the max value and it's location), it's not a simple matter of replacing your plain access with a standard atomic function. However, since you are dealing with two 32-bit adjacent quantities, you may be interested in my answer here.
By the way I think matlab's native matrix multiply on gpuArray should be faster than any matrix multiply code you write. But it would require the Parallel Compute Toolbox.