any way to avoid this read-back-from-GPU?


    Aug 27 2006 | 8:35 am
    hello, thanks to the patch that joshua posted in respond to keiths question, i can now do what i was trying to do with jit.gl.mesh and indeed it is faster than handing geometries to jit.gl.render by hand.
    i have one question, however; i'd like to make the displacements of my geometry's vertices "smooth". i can accomplish that by blurring the matrix that defines the z-displacement. to get a nice cheap blur, i'd like to use the GPU bassed gaussian blur from the examples; however, in order to provide the blurred matrix to jit.gl.mesh, i need to copy it back to RAM from VRAM with jit.matrix which is an expensive step. is there any clever way to avoid that step?
    the patch is below.
    ----------------------

    • Aug 27 2006 | 10:00 pm
      On Aug 27, 2006, at 1:35 AM, Ali Momeni wrote:
      > i have one question, however; i'd like to make the displacements of > my geometry's vertices "smooth". i can accomplish that by blurring > the matrix that defines the z-displacement. to get a nice cheap > blur, i'd like to use the GPU bassed gaussian blur from the > examples; however, in order to provide the blurred matrix to > jit.gl.mesh, i need to copy it back to RAM from VRAM with > jit.matrix which is an expensive step. is there any clever way to > avoid that step?
      Not currently. Technically there is a way to use Frame Buffer Objects (FBOs) as Vertex Buffer Objects (VBOs), but this is not exposed in Jitter at this time. You will have to copy back to RAM, use fewer vertices, or figure out a different approach.
      -Joshua