Executing CUDA kernels requires registering the kernel with the CUDA runtime beforehand. This is achieved by keeping track of procedures marked with CUDA_Global when analyzing packages and then registering each of the procedures with a fat binary created by the CUDA toolchain.
Tested on x86_64-pc-linux-gnu, committed on trunk
gcc/ada/
* debug.adb: Document -gnatd_c flag as being used for CUDA.
* gnat_cuda.ads: New file.
* gnat_cuda.adb: New file.
* rtsfind.ads: Add Interfaces_C_Strings package and
RE_Fatbin_Wrapper, RE_Register_Fat_Binary,
RE_Register_Fat_Binary_End, RE_Register_Function, RE_Chars_Ptr,
RE_New_Char_Array entities.
* rtsfind.adb: Create new Interfaces_C_Descendant subtype,
handle it.
* sem_ch7.adb (Analyze_Package_Body_Helper): Call CUDA init
procedure.
* sem_prag.adb (Analyze_Pragma): Call Add_Cuda_Kernel procedure.
* gcc-interface/Make-lang.in (GNAT_ADA_OBJS): Add gnat_cuda.o.
patch.diff.gz
Description: application/gzip
