update:
another d3d11 advantage is that it has cross vendor binary shaders avoiding need to ship source code also cl has binary support but device dependent and perhaps not future proof among same devices/platforms..
so it's of no use you must ship code similar to glsl..
note that tegra2 sdk has extension (opengl es 2.0 has)
for compiling the gl shader sources to binaries in GLES 2.0.. and using it altough it has some issues with fixed state setup as alpha blending or others and you must follow rules but perhaps this would indicate Nvidia will add binary shaders to desktop gl 3.x drivers?..
A d3d11 disadvantage is multi GPU explicit management seems more intrincate.. and what about events, command buffers, etc.. also no profiling info?..
also no concept of platform and devices in d3d11 is all graphic devices and due to that multidevice management seems more complex need to test a compute shader program load balacing ati and nvidia gpu's: note this is achieved now in ATI+Nvidia GPUs in DirectCompute benchamrk in OCL mode and also smallpt 2.0 beta OCL uses CPU+GPU (multidevice) load balancer..
so good..
also found today that at least since cuda 2.3 we have:
CURTEXPORT cudaError_t CUDARTAPI cudaSetValidDevices(int *device_arr, int len);
seems similar to compute excluse mode..
also cudamemgetinfo in cuda 3.0 is new..
can't get that by other means? must see cudevicequery..
remember also ati ska 1.53 with d3d11 support which says info of kernels as alu limited,tex limited, etc..
no similar info for opencl in profiler in vs..
also ska not so info for glsl shaders.. only hlsl.
Thursday, 14 January 2010
Subscribe to:
Post Comments (Atom)
0 comments:
Post a Comment