Tech Eva

CUDA

Compute Capabilities: 3.0 (Andreas Desktop) VS 5.0 (Johannes)

• 32bit integer atomics (shared + global)
• 64bit atomics (shared + global)
• Voting (+ballot)
• atomic add float
• memory write fence
• warp shuffle fkts
• 3d grid
• 48kb shared mem
• 1D Tex 65536 .. oder 2^27
• 2D Tex 65536 * 65536
• 3D Tex 65536 * 32768 * 2048

GLSL:

• 32bit integer atomics (shared + global): YES ** float atomics
• 64bit atomics (shared + global): Extension puzzle.. no idea...
• Voting (+ballot): Possible with ARB_shader_group_vote
• atomic add float: NV_shader_atomic_float
• memory write fence: YES
• warp shuffle fkts: NV_shader_thread_shuffle
• 3d grid: YES
• 32KB shared mem
• 1D Tex 16k
• 2D Tex 16k * 16k
• 3D Tex 2048 * 2048 * 2048

Pro/Con

PRO OpenGL:

• Rasterizer more intuitive
• no large vendor specific library

CONTRA:

• driver mud

PRO CUDA:

• 64 bit atomics
• NSight
• more memory

CONTRA CUDA:

• fat lib