Since NVC0 chipset doesn't allocate it, this causes segv.
Since no initialization, max value is undefined.
If we don't install gdev yet, cmake build instruction fails since gdev_api.h is not found. We added common to include directory to search header, gdev_api.h.
We define gdev_io_memcpy as utility function and use it for memcpy on MMIO memory.
Since some versions of GCC requires libraries are appended to the end of build command line. For instance, GCC in Ubuntu 12.04 (LTS) requires it.
libdrm & libdrm_nouveau is needed.
Install libucuda to gdev/lib and cudump to gdev/bin.
To use strdup function in kernel space, we implement STRDUP and use it in gdev_cuda.c.
Now we can share C files in util directory
Since -std=c++11 option is supported by gcc 4.7 or later, we should check it and switch c++0x if gcc version is less than 4.7. For instance, Ubuntu 12.04 (LTS) gcc version is still 4.6.
Some functions such as MALLOC, FREE etc. are used, but definitions are scattered. We introduce util directory to share utility functions.
This reverts commit dc5b1ea.
Since CUDA 5.0~ compilers don't guarantee that .nv.XXX appears orderly. So we should lookup actual function by using XXX name.
Since memcpy may use SSE code, it's not appropriate for IO memory
After that, we're planning to use util directory for utility functions.
nvrm: force memcpy not to be extracted by SSE