Reading and storing large matrix file for GPU - c++

Question

Goal: Storing a large matrix in memory (Radon matrix), and transferring it into GPU memory for massively parallel operations.
Problem: Horrible reading time, and potentially sub-optimal use of space (but non-limiting for the program's usage)
I have the possibility of doing this in either C or C++.
The files which I'm receiving are parsed as follows:
0.70316,0.71267,0.72221,0.73177,0.74135,0.75094,0.76053,0.77011,0.77967,0.7892,0.79868,0.80811,0.81747
and this goes on for at least 50MB.
My naïve implementation:
float ** Radon;
Radon = (float **)malloc(HeightxNproj * sizeof(float *));
for (int i = 0; i < Hauteur * Nproj; i++)
Radon[i] = (float *)malloc(WidthSquared * sizeof(float));
FILE *radonFile;
radonFile = fopen("radon.txt", "r");
if (radonFile == NULL)
{
printf("Radon file opening failed.");
return -1;
}
for (int i = 0; i < HeightxNproj; i++)
{
for (int j = 0; j < WidthSquared; j++)
{
fscanf(radonFile, "%f,", &Radon[i][j]);
}
}
fclose(radonFile);
printf("Radon loaded.");
I'm programming for windows. I've read a bit about File Memory Mapping, but I don't know if this method, which is not actually storing the matrix in memory, is compatible with GPGPU programming. I'm using CUDA, and I'll have to pass this matrix onto GPU memory for parallel operations.
This file-reading method performs terribly, it's taking roughly a minute to read and parse the 50MB file. Is there a way to shorten reading and parsing time? The matrix is also a sparse matrix, are there common ways to deal with such matrix?

Answer 1

The more separate accessing of a file the more performance you lose. The first step you should take is to estimate number of information you need to read from the file and read it in one take. It will increase your performance by huge amount. You can use memory mapped files.
and this goes on for at least 50MB.
This is not that much.
The files i'm receiving are parsed as follows:
0.70316,0.71267,0.72221,0.73177,0.74135,0.75094,0.76053,0.77011,0.77967,0.7892,0.79868,0.80811,0.81747
Save it in binary to save about half of the memory (maybe even more). This will also increase reading speed.
Read the whole file at one time.
An example will make you realize how naive and slow is your approach:
Once I was implementing algorithm that was reading .obj 3d model. The model was like 10 MB and it took around 1-2 minute to load. This was very strange, because Blender could load it immediately - maybe 1 or 2 seconds. Mapping whole file to memory and pre-allocating buffers allowed me to load the file in less than 5 secs.
Note:
I can do this in either C or C++, both are ok.
Don't ever mix C with C++ when it comes to memory management, unless you are sure what you are doing. C++ exceptions can cause huge memory leaks if you don't protect C dynamically allocated memory using RAII.