I am trying to solve a problem in which i am supposed to change a colour image to a greyscale image. For this purpose i am using CUDA parallel approach. The kerne code i am invoking on the GPU is as follows.
__global__
void rgba_to_greyscale(const uchar4* const rgbaImage,
unsigned char* const greyImage,
int numRows, int numCols)
{
int absolute_image_position_x = blockIdx.x;
int absolute_image_position_y = blockIdx.y;
if ( absolute_image_position_x >= numCols ||
absolute_image_position_y >= numRows )
{
return;
}
uchar4 rgba = rgbaImage[absolute_image_position_x + absolute_image_position_y];
float channelSum = .299f * rgba.x + .587f * rgba.y + .114f * rgba.z;
greyImage[absolute_image_position_x + absolute_image_position_y] = channelSum;
}
void your_rgba_to_greyscale(const uchar4 * const h_rgbaImage,
uchar4 * const d_rgbaImage,
unsigned char* const d_greyImage,
size_t numRows,
size_t numCols)
{
//You must fill in the correct sizes for the blockSize and gridSize
//currently only one block with one thread is being launched
const dim3 blockSize(numCols/32, numCols/32 , 1); //TODO
const dim3 gridSize(numRows/12, numRows/12 , 1); //TODO
rgba_to_greyscale<<<gridSize, blockSize>>>(d_rgbaImage,
d_greyImage,
numRows,
numCols);
cudaDeviceSynchronize(); checkCudaErrors(cudaGetLastError());
}
i see a line of dots in the first pixel line.
error i am getting is
libdc1394 error: Failed to initialize libdc1394
Difference at pos 51 exceeds tolerance of 5
Reference: 255
GPU : 0
my input/output images
Can anyone help me with this??? thanks in advance.
I recently joined this course and tried your solution but it don't work so, i tried my own. You are almost correct. The correct solution is this:
__global__`
void rgba_to_greyscale(const uchar4* const rgbaImage,
unsigned char* const greyImage,
int numRows, int numCols)
{`
int pos_x = (blockIdx.x * blockDim.x) + threadIdx.x;
int pos_y = (blockIdx.y * blockDim.y) + threadIdx.y;
if(pos_x >= numCols || pos_y >= numRows)
return;
uchar4 rgba = rgbaImage[pos_x + pos_y * numCols];
greyImage[pos_x + pos_y * numCols] = (.299f * rgba.x + .587f * rgba.y + .114f * rgba.z);
}
The rest is same as your code.
Now, since I posted this question I have been continuously working on this problem there are a couple of improvements that should be done in order to get this problem correct now I realize my initial solution was wrong . Changes to be done:-
1. absolute_position_x =(blockIdx.x * blockDim.x) + threadIdx.x;
2. absolute_position_y = (blockIdx.y * blockDim.y) + threadIdx.y;
Secondly,
1. const dim3 blockSize(24, 24, 1);
2. const dim3 gridSize((numCols/16), (numRows/16) , 1);
In the solution we are using a grid of numCols/16 * numCols/16
and blocksize of 24 * 24
code executed in 0.040576 ms
#datenwolf : thanks for answering above!!!
Since you are not aware of the image size. It is best to choose any reasonable dimension of the two-dimensional block of threads and then check for two conditions. The first one is that the pos_x and pos_y indexes in the kernel do not exceed numRows and numCols. Secondly the grid size should be just above the total number of threads in all the blocks.
const dim3 blockSize(16, 16, 1);
const dim3 gridSize((numCols%16) ? numCols/16+1 : numCols/16,
(numRows%16) ? numRows/16+1 : numRows/16, 1);
libdc1394 error: Failed to initialize libdc1394
I don't think that this is a CUDA problem. libdc1394 is a library used to access IEEE1394 aka FireWire aka iLink video devices (DV camcorders, Apple iSight camera). That library doesn'r properly initialize, hence you're not getting usefull results. Basically it's NINO: Nonsens In Nonsens Out.
the calculation of absolute x & y image positions is perfect.
but when u need to access that particular pixel in the coloured image , shouldn't you u use the following code??
uchar4 rgba = rgbaImage[absolute_image_position_x + (absolute_image_position_y * numCols)];
I thought so, when comparing it to a code you'd write to execute the same problem in serial code.
Please let me know :)
You still should have a problem with run time - the conversion will not give a proper result.
The lines:
uchar4 rgba = rgbaImage[absolute_image_position_x + absolute_image_position_y];
greyImage[absolute_image_position_x + absolute_image_position_y] = channelSum;
should be changed to:
uchar4 rgba = rgbaImage[absolute_image_position_x + absolute_image_position_y*numCols];
greyImage[absolute_image_position_x + absolute_image_position_y*numCols] = channelSum;
__global__
void rgba_to_greyscale(const uchar4* const rgbaImage,
unsigned char* const greyImage,
int numRows, int numCols)
{
int rgba_x = blockIdx.x * blockDim.x + threadIdx.x;
int rgba_y = blockIdx.y * blockDim.y + threadIdx.y;
int pixel_pos = rgba_x+rgba_y*numCols;
uchar4 rgba = rgbaImage[pixel_pos];
unsigned char gray = (unsigned char)(0.299f * rgba.x + 0.587f * rgba.y + 0.114f * rgba.z);
greyImage[pixel_pos] = gray;
}
void your_rgba_to_greyscale(const uchar4 * const h_rgbaImage, uchar4 * const d_rgbaImage,
unsigned char* const d_greyImage, size_t numRows, size_t numCols)
{
//You must fill in the correct sizes for the blockSize and gridSize
//currently only one block with one thread is being launched
const dim3 blockSize(24, 24, 1); //TODO
const dim3 gridSize( numCols/24+1, numRows/24+1, 1); //TODO
rgba_to_greyscale<<<gridSize, blockSize>>>(d_rgbaImage, d_greyImage, numRows, numCols);
cudaDeviceSynchronize(); checkCudaErrors(cudaGetLastError());
}
The libdc1394 error is not related to firewire etc in this case - it is the library that udacity is using to compare the image your program creates to the reference image. And what is is saying is that the difference between your image and the reference image has been been exceeded by a specific threshold, for that position ie. pixel.
You are running following number of block and grids:
const dim3 blockSize(numCols/32, numCols/32 , 1); //TODO
const dim3 gridSize(numRows/12, numRows/12 , 1); //TODO
yet you are not using any threads in your kernel code!
int absolute_image_position_x = blockIdx.x;
int absolute_image_position_y = blockIdx.y;
think this way, the width of an image can be divide into absolute_image_position_x parts of column and the height of an image can be divide into absolute_image_position_y parts of row. Now the box each of the cross section it creates you need to change/redraw all the pixels in terms of greyImage, parallely. Enough spoiler for an assignment :)
same code with with ability to handle non-standard input size images
int idx=blockDim.x*blockIdx.x+threadIdx.x;
int idy=blockDim.y*blockIdx.y+threadIdx.y;
uchar4 rgbcell=rgbaImage[idx*numCols+idy];
greyImage[idx*numCols+idy]=0.299*rgbcell.x+0.587*rgbcell.y+0.114*rgbcell.z;
}
void your_rgba_to_greyscale(const uchar4 * const h_rgbaImage, uchar4 * const d_rgbaImage,
unsigned char* const d_greyImage, size_t numRows, size_t numCols)
{
//You must fill in the correct sizes for the blockSize and gridSize
//currently only one block with one thread is being launched
int totalpixels=numRows*numCols;
int factors[]={2,4,8,16,24,32};
vector<int> numbers(factors,factors+sizeof(factors)/sizeof(int));
int factor=1;
while(!numbers.empty())
{
if(totalpixels%numbers.back()==0)
{
factor=numbers.back();
break;
}
else
{
numbers.pop_back();
}
}
const dim3 blockSize(factor, factor, 1); //TODO
const dim3 gridSize(numRows/factor+1, numCols/factor+1,1); //TODO
rgba_to_greyscale<<<gridSize, blockSize>>>(d_rgbaImage, d_greyImage, numRows, numCols);
1- int x =(blockIdx.x * blockDim.x) + threadIdx.x;
2- int y = (blockIdx.y * blockDim.y) + threadIdx.y;
And in grid and block size
1- const dim3 blockSize(32, 32, 1);
2- const dim3 gridSize((numCols/32+1), (numRows/32+1) , 1);
Code executed in 0.036992 ms.
const dim3 blockSize(16, 16, 1); //TODO
const dim3 gridSize( (numRows+15)/16, (numCols+15)/16, 1); //TODO
int x = blockIdx.x * blockDim.x + threadIdx.x;
int y = blockIdx.y * blockDim.y + threadIdx.y;
uchar4 rgba = rgbaImage[y*numRows + x];
float channelSum = .299f * rgba.x + .587f * rgba.y + .114f * rgba.z;
greyImage[y*numRows + x] = channelSum;
Related
I am implementing a nearest neighborhood kernel function to resize the input image. But the result is wrong and I have no idea.
Here is the input image
the result is wrong.
I use opencv to read the input image.
cv::Mat image = cv::imread("/home/tumh/test.jpg");
unsigned char* data = image.data;
int outH, outW;
float *out_data_host = test(data, image.rows, image.cols, outH, outW);
cv::Mat out_image(outH, outW, CV_32FC3);
memcpy(out_image.data, out_data_host, outH * outW * 3 * sizeof(float));
float* test(unsigned char* in_data_host, const int &inH, const int &inW, int &outH, int &outW) {
// get the output size
int im_size_min = std::min(inW, inH);
int im_size_max = std::max(inW, inH);
float scale_factor = static_cast<float>(640) / im_size_min;
float im_scale_x = std::floor(inW * scale_factor / 64) * 64 / inW;
float im_scale_y = std::floor(inH * scale_factor / 64) * 64 / inH;
outW = inW * im_scale_x;
outH = inH * im_scale_y;
int channel = 3;
unsigned char* in_data_dev;
CUDA_CHECK(cudaMalloc(&in_data_dev, sizeof(unsigned char) * channel * inH * inW));
CUDA_CHECK(cudaMemcpy(in_data_dev, in_data_host, 1 * sizeof(unsigned char) * channel * inH * inW, cudaMemcpyHostToDevice));
// image pre process
const float2 scale = make_float2( im_scale_x, im_scale_y);
float * out_buffer = NULL;
CUDA_CHECK(cudaMalloc(&out_buffer, sizeof(float) * channel * outH * outW));
float *out_data_host = new float[sizeof(float) * channel * outH * outW];
const dim3 threads(32, 32);
const dim3 block(iDivUp(outW, threads.x), iDivUp(outW, threads.y));
gpuPreImageNet<<<block, threads>>>(scale, in_data_dev, inW, out_buffer, outW, outH);
CUDA_CHECK(cudaFree(in_data_dev));
CUDA_CHECK(cudaMemcpy(out_data_host, out_buffer, sizeof(float) * channel * outH * outW, cudaMemcpyDeviceToHost));
CUDA_CHECK(cudaFree(out_buffer));
return out_data_host;
}
Here is the resize kernel function
__global__ void gpuPreImageNet( float2 scale, unsigned char* input, int iWidth, float* output, int oWidth, int oHeight )
{
const int x = blockIdx.x * blockDim.x + threadIdx.x;
const int y = blockIdx.y * blockDim.y + threadIdx.y;
const int n = oWidth * oHeight;
int channel = 3;
if( x >= oWidth || y >= oHeight )
return;
const int dx = ((float)x * scale.x);
const int dy = ((float)y * scale.y);
const unsigned char* px = input + dy * iWidth * channel + dx * channel ;
const float3 bgr = make_float3(*(px + 0), *(px + 1), *(px + 2));
output[channel * y * oWidth + channel * x + 0] = bgr.x;
output[channel * y * oWidth + channel * x + 1] = bgr.y;
output[channel * y * oWidth + channel * x + 2] = bgr.z;
}
Most of the implementation is from https://github.com/soulsheng/ResizeNN/blob/master/resizeCUDA/resizeNN.cu
Any idea?
Maybe you are observing an uninitialized memory problem.
As i understand your code, out_data_host allocation is too big
new float[sizeof(float) * channel * outH * outW];
should be
new float[channel * outH * outW]
Then out_buffer is uninitialized, add a cudaMemset after the cudaMalloc line.
To clarify your code, since you already use OpenCV to load images, why don't you use opencv to resize your images ?
cv::resize // Host side method is probably better since you'll have less data copied through PCI-Express
// or
cv::cuda::resize
It took me around two days to figure out a solution for this problem. Basically, I was building a GPU based image preprocessing pipeline for my project. Here's the custom Cuda Kernel.
For Gray scale Image Resizing, change channel from 3 -> 1 and it should work.
__global__ void resize_kernel( real* pIn, real* pOut, int widthIn, int heightIn, int widthOut, int heightOut)
{
int i = blockDim.y * blockIdx.y + threadIdx.y;
int j = blockDim.x * blockIdx.x + threadIdx.x;
int channel = 3;
if( i < heightOut && j < widthOut )
{
int iIn = i * heightIn / heightOut;
int jIn = j * widthIn / widthOut;
for(int c = 0; c < channel; c++)
pOut[ (i*widthOut + j)*channel + c ] = pIn[ (iIn*widthIn + jIn)*channel + c ];
}
}
I need to do a RGB2GRAY image processing algorithm. I just need some help in completing the global function or how I can access the * d_src pointer. This is my code, your help will be greatly appreciated.
#include "cuda_runtime.h"
#include "device_launch_parameters.h"
#include "CImg.h"
#include <iostream>
using namespace std;
using namespace cimg_library;
__global__ void rgb2gray(unsigned char * d_src, unsigned char * d_dst, int width, int height){
int pos_x = blockIdx.x * blockDim.x + threadIdx.x;
int pos_y = blockIdx.y * blockDim.y + threadIdx.y;
if (pos_x >= width || pos_y >= height)
return;
}
int main(){
//Load image
CImg<unsigned char> src("lena.jpg");
int width = src.width();
int height = src.height();
unsigned long sizee = src.size();
int sze = width * height;
cout << sze << endl;
//create pointer to image
unsigned char *h_src = src.data();
CImg<unsigned char> dst(width, height, 1, 1);
unsigned char *h_dst = dst.data();
unsigned char *d_src;
unsigned char *d_dst;
cout << sizee << endl;
cudaMalloc((void**)&d_src, sizee);
cudaMalloc((void**)&d_dst, width*height*sizeof(int));
cudaMemcpy(d_src, h_src, sizee, cudaMemcpyHostToDevice);
//launch the kernel
rgb2gray << <(width/16,height/16,1), (16, 16, 1) >> >(d_src, d_dst, width, height);
//force the printf()s to flush
cudaDeviceSynchronize();
// copy back the result array to the CPU
cudaMemcpy(h_dst, d_dst, width*height, cudaMemcpyDeviceToHost);
cudaFree(d_src);
cudaFree(d_dst);
CImgDisplay main_disp(dst, "After Processing");
while (!main_disp.is_closed())
main_disp.wait();
return 0;
}
Firstly, since your dst object consists of unsigned char, allocate d_dst as follows;
cudaMalloc((void**)&d_dst, width*height*sizeof(unsigned char));
Next, grid must cover every pixels, considering cases when width or height are not a multiple of 16. Launch kernel with following kernel configuration.
dim3 blkDim (16, 16, 1);
dim3 grdDim ((width + 15)/16, (height + 15)/16, 1);
rgb2gray<<<grdDim, blkDim>>>(d_src, d_dst, width, height);
Lastly, your kernel should look like this. Note that RGB channels are split in d_src.
int pos_x = blockIdx.x * blockDim.x + threadIdx.x;
int pos_y = blockIdx.y * blockDim.y + threadIdx.y;
if (pos_x >= width || pos_y >= height)
return;
unsigned char r = d_src[pos_y * width + pos_x];
unsigned char g = d_src[(height + pos_y) * width + pos_x];
unsigned char b = d_src[(height * 2 + pos_y) * width + pos_x];
unsigned int _gray = (unsigned int)((float)(r + g + b) / 3.0f + 0.5);
unsigned char gray = _gray > 255 ? 255 : _gray;
d_dst[pos_y * width + pos_x] = gray;
You can see the full code here.
I'd like to know how to access pixel information when using OpenCV GPU. I'm currently downloading gpu::mat information to a mat variable but it is too slow.
Does anyone know how to do it?
You could access to the data inside a kernel.
For (row,col) the row and column number,
the value of a pixel with channel number ch < 3 will be:
uint8_t val = gpumat.data[ (row*gpumat.step) + col*gpumat.channels() + ch];
So let say you have an BGR input image stored on a GpuMat called src and you want to
assign each pixel value to a destination image called dst (also GpuMat).
You could call the kernel
kernel_assign_pixel<<<gridDim, blockDim>>>
(src.data, dst.data, src.rows, src.cols, src.step, dst.step, src.channels());
Defined in the following way
__global__ void kernel_assign_pixel
(uint8_t* src, uint8_t* dst, int MaxRows, int MaxCols, int iStep, int oStep, int MaxC)
{
unsigned int row = blockIdx.x * blockDim.x + threadIdx.x; //Row number
unsigned int col = blockIdx.y * blockDim.y + threadIdx.y; //Column number
unsigned int ch = blockIdx.z * blockDim.z + threadIdx.z; //Channel number
if (row<MaxRows && col<MaxCols && ch < MaxC)
{
int tidIn = row * iStep + col * MaxC + ch;
int tidOut = row * oStep + col * MaxC + ch;
dst[tidOut]=src[tidIn];
}
}
I'm doing linear filtering on images using CUDA. I use 2D thread blocks and 2D grid to make the problem natural. Here's how I index: (height and width are image dimensions)
dim3 BlockDim(16,16);
dim3 GridDim;
GridDim.x = (width + 15) / 16;
GridDim.y = (height + 15) / 16;
In kernel I access the locations as follows:
unsigned int xIndex = blockIdx.x*16+ threadIdx.x;
unsigned int yIndex = blockIdx.y*16+ threadIdx.y;
unsigned int tid = yIndex * width + xIndex;
And I want to return four boundaries (i'll cater them later on). I do this as:
if(yIndex>=height-N || xIndex>=width-N || yIndex<N || xIndex<N)
return;
Where N is the number of pixels at each boundary I don't want to calculate.
Problem:
The code runs fine on all standard images sizes. But for some random image sizes it shows diagonal line(s). For example in my case 500x333 image (even when no dimension is multiple of 16) is showing correct output whereas 450x365 is showing diagonal lines in the output. The problem remains even if I just return the extra threads of grid and nothing else like this:
if(yIndex>=height || xIndex>=width)
return;
The code remains the same, some inputs run fine while others don't. Can anybody spot the bug? I have attached the input and output samples here: IMAGES Thanks!
Update:
Kernel Code (Simplified to return input image, but gives the same problem)
__global__ void filter_8u_c1_kernel(unsigned char* in, unsigned char* out, int width, int height, float* filter, int fSize)
{
unsigned int xIndex = blockIdx.x*BLOCK_SIZE + threadIdx.x;
unsigned int yIndex = blockIdx.y*BLOCK_SIZE + threadIdx.y;
unsigned int tid = yIndex * width + xIndex;
unsigned int N = filterSize/2;
if(yIndex>=height-N || xIndex>=width-N || yIndex<N || xIndex<N)
return;
/*Filter code removed, still gives the same problem*/
out[tid] = in[tid];
}
Update 2:
I have also removed the return statement by reversing the if condition. But the problem persists.
if(yIndex<=height-N && xIndex<=width-N && yIndex>N && xIndex>N){
/*Kernel Code*/
}
There are quite a few things you still haven't described very well, but based on the information you have posted, I built what I am guessing is a reasonable repro case with parameters which match a case you say it failing (450 x 364 with filterSize=5):
#include <stdio.h>
#include <assert.h>
template<int filterSize>
__global__ void filter_8u_c1_kernel(unsigned char* in, unsigned char* out, int width, int height, float* filter, int fSize)
{
unsigned int xIndex = blockIdx.x*blockDim.x + threadIdx.x;
unsigned int yIndex = blockIdx.y*blockDim.y + threadIdx.y;
unsigned int tid = yIndex * width + xIndex;
unsigned int N = filterSize/2;
if(yIndex>=height-N || xIndex>=width-N || yIndex<N || xIndex<N)
return;
out[tid] = in[tid];
}
int main(void)
{
const int width = 450, height = 365, filterSize=5;
const size_t isize = sizeof(unsigned char) * size_t(width * height);
unsigned char * _in, * _out, * out;
assert( cudaMalloc((void **)&_in, isize) == cudaSuccess );
assert( cudaMalloc((void **)&_out, isize) == cudaSuccess );
assert( cudaMemset(_in, 'Z', isize) == cudaSuccess );
assert( cudaMemset(_out, 'A', isize) == cudaSuccess );
const dim3 BlockDim(16,16);
dim3 GridDim;
GridDim.x = (width + BlockDim.x - 1) / BlockDim.x;
GridDim.y = (height + BlockDim.y - 1) / BlockDim.y;
filter_8u_c1_kernel<filterSize><<<GridDim,BlockDim>>>(_in,_out,width,height,0,0);
assert( cudaPeekAtLastError() == cudaSuccess );
out = (unsigned char *)malloc(isize);
assert( cudaMemcpy(out, _out, isize, cudaMemcpyDeviceToHost) == cudaSuccess);
for(int i=0; i<width; i++) {
fprintf(stdout, "%d: ", i);
for(int j=0; j<height; j++) {
unsigned int idx = i + j*width;
fprintf(stdout, "%c", out[idx]);
}
fprintf(stdout, "\n");
}
return cudaThreadExit();
}
When run it does exactly what I would expect, overwriting the output memory with the input everywhere except for the first and last two lines and the first and last two entries in all the lines in between. This is running with CUDA 3.2 on OS X 10.6.5 with a compute 1.2 GPU. So whatever is happening in you code, it isn't happening in my repro case, which either means I have misinterpreted what you have written, or there is something else you haven't described which is causing the problem.
i began to implement some simple image processing using cuda but i have an error in my code
the error happens when i copy pixels from device to host
this is my try
#include "cuda_runtime.h"
#include "device_launch_parameters.h"
#include <opencv2\core\core.hpp>
#include <opencv2\highgui\highgui.hpp>
#include <stdio.h>
using namespace cv;
unsigned char *h_pixels;
unsigned char *d_pixels;
int bufferSize;
int width,height;
const int BLOCK_SIZE = 32;
Mat image;
void get_pixels(const char* fileName)
{
image = imread(fileName);
bufferSize = image.size().width * image.size().height * 3 * sizeof(unsigned char);
width = image.size().width;
height = image.size().height;
h_pixels = new unsigned char[bufferSize];
memcpy(h_pixels,image.data,bufferSize);
}
__global__ void invert_image(unsigned char* pixels,int width,int height)
{
int row = blockIdx.y * BLOCK_SIZE + threadIdx.y;
int col = blockIdx.x * BLOCK_SIZE + threadIdx.x;
int cidx = (row * width + col) * 3;
pixels[cidx] = 255 - pixels[cidx];
pixels[cidx + 1] = 255 - pixels[cidx + 1];
pixels[cidx + 2] = 255 - pixels[cidx + 2];
}
int main()
{
get_pixels("D:\\photos\\z.jpg");
cudaError_t err = cudaMalloc((void**)&d_pixels,bufferSize);
err = cudaMemcpy(d_pixels,h_pixels,bufferSize,cudaMemcpyHostToDevice);
dim3 dimBlock(BLOCK_SIZE,BLOCK_SIZE);
dim3 dimGrid(width/dimBlock.x,height/dimBlock.y);
invert_image<<<dimBlock,dimGrid>>>(d_pixels,width,height);
unsigned char *pixels = new unsigned char[bufferSize];
err= cudaMemcpy(pixels,d_pixels,bufferSize,cudaMemcpyDeviceToHost);// unknown error
const char * errStr = cudaGetErrorString(err);
cudaFree(d_pixels);
image.data = pixels;
namedWindow("display image");
imshow("display image",image);
waitKey();
return 0;
}
also how can i find out error that occurs in cuda device
thanks for your help
OpenCV images are not continuous. Each row is 4 byte or 8 byte aligned. You should also pass the step field of the Mat to the CUDA kernel, so that you can calculate the cidx correctly. The generic formula to calculate the output index is:
cidx = row * (step/elementSize) + (NumberOfChannels * col);
in your case, it will be:
cidx = row * step + (3 * col);
Referring to the alignment of images, you buffer size is equal to image.step * image.size().height.
Next thing is the one pointed out by #phoad in the third point. You should create enough number of thread blocks to cover the whole image.
Here is a generic formula for Grid which will create enough number of blocks for any image size.
dim3 block(BLOCK_SIZE,BLOCK_SIZE);
dim3 grid((width + block.x - 1)/block.x,(height + block.y - 1)/block.y);
First of all be sure that the image file is read correctly.
Check if the device memory is allocated with CUDA_SAFE_CALL(cudaMalloc(..))
Check the dimensions of the image. If the dimension of the image is not multiples of BLOCKSIZE than you might be missing some indices and the image is not fully inverted.
Call cudaDeviceSynchronize after the kernel call and check its return value.
Do you get any error when you run the code without calling the kernel anyway?
You are not freeing the h_pixels and might have a memory leak.
Instead of using BLOCKSIZE in the kernel you might use "blockDim.x". So calculating indices like "blockIdx.x * blockDim.x + threadIdx.x"
Try to do not touch the memory area in the kernel code, namely comment out the memory updates at the kernel (the lines where you access the pixels array) and check if the program continues to fail. If it does not continue to fail you might be accessing out of the bounds.
Use this command immediately after the kernel invocation to print the kernel errors:
printf("error code: %s\n",cudaGetErrorString(cudaGetLastError()))