partial texture transfers using PBOs

Hello,

I’m using PBOs to transfer texture data from system memory to the GPU.
I’m basically following the tutorial here:
http://www.mathematik.uni-dortmund.de/~goeddeke/gpgpu/tutorial3.html#pbodownload

Except, I’m modifying the code to only transfer a region of the texture:


struct Rect
{
int x;          // bottom-left corner of the rectangle (assuming lower-left origin)
int y; 

int nx;         // height and width of the rectangle
int ny;
};

Rect dirtyRect; // region of the texture to update

glBindBuffer(GL_PIXEL_UNPACK_BUFFER_ARB, ioBuf[array]);
glBufferData(GL_PIXEL_UNPACK_BUFFER_ARB, dirtyRect.nx*dirtyRect.ny*sizeof(float), 
                        NULL, GL_STREAM_DRAW);
void* ioMem = glMapBuffer(GL_PIXEL_UNPACK_BUFFER_ARB, GL_WRITE_ONLY);
assert(ioMem); 

// modified to copy only the portion of data specified by dirtyRect...
copy(chunks[chunk][array], ioMem);


glUnmapBuffer(GL_PIXEL_UNPACK_BUFFER_ARB);
glTexSubImage2D(GL_TEXTURE_RECTANGLE_ARB, 0, dirtyRect.x, dirtyRect.y, dirtyRect.nx, dirtyRect.ny, 
                GL_LUMINANCE, GL_FLOAT, BUFFER_OFFSET(0));
glBindBuffer(GL_PIXEL_UNPACK_BUFFER_ARB, 0);

This seems to work when the format of the texture is GL_RGBA and the type of data is GL_UNSIGNED_BYTE, however it does not always work when the format is GL_LUMINANCE_ALPHA. Sometimes it succeeds, and sometimes GL_INVALID_OPERATION is returned from the glTexSubImage call. (This is on a Quadro 4800) The success of the operation seems to be dependent on the region of the texture to be uploaded.

The docs on glBufferData aren’t clear to me as to what should be specified for the size argument. Should it be the size of the dirty region passed into glTexSubImage? Or should it be the whole texture?

It could be a byte alignment issue, where previously you ‘got away with it’ via a 4-byte RGBA format.
Now that you are using LUMINANCE_ALPHA (2 byte) you may need to specify a different pack alignment:

glPixelStorei(GL_UNPACK_ALIGNMENT, 1);
or
glPixelStorei(GL_UNPACK_ALIGNMENT, 2);
should get the job done.

Ah, yeah, that was the problem. Setting the unpack alignment fixes the problem. Thanks so much!

It probably also “sometimes” succeeds if the width is even.

The reason is alignment depends not only on the number of components, the size in bytes of the type of the component but also depends on the width of a row.

I usually uses this kind of code to determine the optimal value for alignment (it goes from 1 (slowest) to 8 (fastest)):


int components; // 1,2,3,4
int width;

// row size in bytes
int rowsize=components*sizeof(component_type)*width;

int alignment;
if(rowsize%8==0) {
  alignment=8;
} else {
  if(rowsize%4==0) {
    alignment=4;
  } else {
    if(rowsize%2==0) {
      alignment=2;
    } else {
      alignment=1;
    }
  }
}