• notice
  • Congratulations on the launch of the Sought Tech site

Image processing with CUDA and libpng

I've been trying to use CUDA and the libpng library to process imagery (i.e.apply a black and white filter).However, I'm not sure how to pass the image data to the device.

From what I understand, the libpng library reads png and uses this function to save the information line by line in the name In png_bytep structure for row_pointers.

 void read_png_file(char* file_name){
        char header[8]; //8 is the maximum size that can be checked

        FILE *fp=fopen(file_name, "rb");
        if (!fp)
                abort_("[read_png_file] File %s could not be opened for reading" file_name);
        fread(header, 1, 8, fp);
        //if (png_sig_cmp(header, 0, 8))
        //abort_("[read_png_file] File %s is not recognized as a PNG file", file_name);

        //Inicializa variables necesarias para libpng
        png_ptr=png_create_read_struct (PNG_LIBPNG_VER_STRING, NULL, NULL, NULL);

        if (!png_ptr)
                abort_("[read_png_file] png_create_read_struct failed");

        if (!info_ptr)
                abort_("[read_png_file] png_create_info_struct failed");

        if (setjmp(png_jmpbuf(png_ptr)))
                abort_("[read_png_file] Error during init_io");

        //Inicializa el input/output para el archivo PNG
        png_init_io(png_ptr, fp);
        png_set_sig_bytes(png_ptr, 8);

        //Lee la información anterior a los datos de los píxeles como tal
        png_read_info(png_ptr, info_ptr);

        //Almacena información del archivo PNG 
        width=png_get_image_width(png_ptr, info_ptr);
        height=png_get_image_height(png_ptr, info_ptr);
        color_type=png_get_color_type(png_ptr, info_ptr);
        bit_depth=png_get_bit_depth(png_ptr, info_ptr);

        png_read_update_info(png_ptr, info_ptr);

        //Lectura del archivo PNG
        if (setjmp(png_jmpbuf(png_ptr)))
                abort_("[read_png_file] Error during read_image");

        //Reserva el espacio necesario para almacenar los datos del archivo PNG por filas
        row_pointers=(png_bytep*) malloc(sizeof(png_bytep) * height);
        for (y=0; y<height; y)
                row_pointers[y]=(png_byte*) malloc(png_get_rowbytes(png_ptr,info_ptr));

        //Y para la copia para el device
        d_row_pointers=(png_bytep*) malloc(sizeof(png_bytep) * height);
        for (y=0; y<height; y)
                d_row_pointers[y]=(png_byte*) malloc(png_get_rowbytes(png_ptr,info_ptr));

        png_read_image(png_ptr, row_pointers);

I've been trying to pass the information saved in row_pointers, create a copy called d_row_pointers, and Use the cudaMalloc and cudaMemcpy functions as follows:

png_bytep * row_pointers;
png_bytep * d_row_pointers;

int main(int argc, char **argv)
        //Verifica los parámetros para ejecutar el programa
        if (argc !=3)
                abort_("Uso:./Nombre_del_Programa <file_in> <file_out>");


        int size=sizeof(png_bytep);
        int int_size=sizeof(int);
        cudaMalloc((void **)&d_row_pointers, size);for (y=0; y<height; y)
                 cudaMalloc((void **)&d_row_pointers[y],png_get_rowbytes(png_ptr,info_ptr));

        cudaMalloc((void **)&d_width, int_size);
        cudaMalloc((void **)&d_height, int_size);

        cudaMemcpy(d_row_pointers, &row_pointers, size, cudaMemcpyHostToDevice);
        for (y=0; y<height; y)
                 cudaMemcpy(d_row_pointers[y], &row_pointers[y], png_get_rowbytes(png_ptr,info_ptr), cudaMemcpyHostToDevice);
        cudaMemcpy(d_width, &width, int_size, cudaMemcpyHostToDevice);
        cudaMemcpy(d_height, &height, int_size, cudaMemcpyHostToDevice);

        //Lanzar el kernel
        process_file<<<1,1>>>(d_row_pointers, d_width, d_height);

        //Copiar los resultados de vuelta al host
        cudaMemcpy(&row_pointers, d_row_pointers, size, cudaMemcpyDeviceToHost);

        for (y=0; y<height; y)

        //Escritura de la imagen con los resultados

        return 0;

This is the process_file function where the filter should be applied

__global__ void process_file(png_bytep * d_row_pointers, int * d_width, int * d_height)
        //Se realizan los cambios deseados en la imagen
        //Verificar los datos recibidos
        printf("Width=%d , Height=%d ", *d_width, *d_height);

        int rgb_total=0;
        float rgb_average=0.0;
        int x=0;
        int y=0;
        png_byte *row;
        png_byte *ptr;

        for (y=0; y<*d_height-1; y) {

                for (x=0; x<*d_width; x) {

                        printf("Pixel %d-%d, Rgb values: %d-%d-%d \n", x, y, ptr[0], ptr[1], ptr[2]); 
                        rgb_total=ptr[0] ptr[1] ptr[2];
                        //Calculando el promedios RGB
                        //printf("Average: %d \n", (int)rgb_average);
                        //printf("Changed to %d-%d-%d \n",ptr[0], ptr[1], ptr[2]); 
                        //printf("Pixel %d-%d done\n",x,y); 

        //printf("Para la imagen de resolución: %d x %d-", *width, *height);

but I can't get it to work, even though the kernel seems to be receiving the structure, when I access the data it's full Zero when it is considered an RGB value.I would appreciate any help on the correct way to pass this data to the kernel.Thanks!

PD: The complete code can be found in herefound

uj5u.com enthusiastic netizens replied:

I'm pretty sure this line is causing the problem:

cudaMemcpy(d_row_pointers[y], &row_pointers[y], 
    png_get_rowbytes(png_ptr,info_ptr), cudaMemcpyHostToDevice);

You cannotd_row_pointers[y] from cpu codeaccess data font>.cudaMalloc expects a pointer to a pointer on the cpu and will handle arguments as such.

You need to be careful with multidimensional arrays in CUDA.Most people choose to use flat arrays, even for images.

Here's what I suggest: copy the material to a flat (1D) buffer, then copy it Copy to device.This makes it harder to make mistakes.

Ah, one more thing: you don't need to explicitly reference the width and height of the cudaMalloc and cudaMemcpy kernels number.You can pass them by value as if it were a normal function.You only need those functions for arrays.


Technical otaku

Sought technology together

Related Topic


Leave a Reply