how could RTT be so slow on my intel card?

Question

I simply draw something into a render target, and use it as an normal texture. It's always working greate to me with my nvidia video card. But today I found my program ran terribly slow (less than 5 fps) on a intel card.

After running an analyse I found glGenerateMipmapEXT is the trouble maker, it costs most of the cpu time. Here's how I bind a RT texture:

  glBindTexture(GL_TEXTURE_2D, m_textureID);<br>
  glGenerateMipmapEXT(GL_TEXTURE_2D);

without glGenerateMipmapEXT the texture is nothing but pure white pic. Something wrong with my RTT?

Answer 1

Please remove the MipMapping flag for your texture... You need glGenerateMipmap because you use MipMapping and OpenGL need a smaller texture. On bad cards(INTEL), GenerateMipmap is a costly operation. This operation must be called one times.

Add this instruction when you load your texture after glBindTexture :

glTexParameteri(GL_TEXTURE_2D, GL_TEXTURE_MIN_FILTER, GL_LINEAR);

Answer 2

First reason - Intel's OpenGL drivers are universally acknowledged to be rubbish. Sudden and mysterious crashes, artefacts, performance differences, etc are the order of the day here. This is normal, nothing to see here, move along.

Second reason - RTT is always going to be slower than just rendering direct to the backbuffer. Not RTT in and of itself, but the overhead of rendering to a texture, then rendering that same texture to your backbuffer - that's some extra fillrate overhead you're incurring. Even on a very very good driver/hardware combo with fillrate being the only major bottleneck and everything else being equal, it will always be slower.

Third reason - mipmapping a rendertarget texture is not a great idea. It may not even be supported in hardware, which means that the driver must pull the texture back to system memory, update the miplevels in software, then put it back to GPU memory.