diff options
| author | Derick Huth <derickhuth@gmail.com> | 2014-10-06 12:42:34 -0600 |
|---|---|---|
| committer | Derick Huth <derickhuth@gmail.com> | 2014-10-06 12:42:34 -0600 |
| commit | 8d73c8fcc75556c1df71dd39dd99783f8f86fc3e (patch) | |
| tree | 157d627863d76a4c256a27cae27ce2e8566c7ea0 /test-chill/test-cases/examples/cuda-chill/nbody.lua | |
| parent | e87b55ad69f0ac6211daae741b32c8ee9dcbe470 (diff) | |
| parent | 8c646f24570079eac53e58fcf42d0d4fbc437ee3 (diff) | |
| download | chill-8d73c8fcc75556c1df71dd39dd99783f8f86fc3e.tar.gz chill-8d73c8fcc75556c1df71dd39dd99783f8f86fc3e.tar.bz2 chill-8d73c8fcc75556c1df71dd39dd99783f8f86fc3e.zip | |
Merge pull request #2 from dhuth/master
Moved omega into chill.
Diffstat (limited to 'test-chill/test-cases/examples/cuda-chill/nbody.lua')
| -rw-r--r-- | test-chill/test-cases/examples/cuda-chill/nbody.lua | 53 |
1 files changed, 53 insertions, 0 deletions
diff --git a/test-chill/test-cases/examples/cuda-chill/nbody.lua b/test-chill/test-cases/examples/cuda-chill/nbody.lua new file mode 100644 index 0000000..08f88a9 --- /dev/null +++ b/test-chill/test-cases/examples/cuda-chill/nbody.lua @@ -0,0 +1,53 @@ +--CUBLAS 2 MM Multiply + +--This function form intializes "CUDAIZE v2" versus "CUDAIZE v1" if you +--call init() and use global variables to specify procedure and loop + +--Second parameter is procedure # and third is loop # +init("nbody.c", "nbody_cpu" , 0) + +dofile("cudaize.lua") --defines custom tile_by_index, copy_to_registers, + --copy_to_shared methods +NBODIES=16384 + + +--Tj=128 CHANGE FOR BEST..... BEST IS 64BLOCKS 128THREADS +--Ti=256 +Tj=64 +Ti=32 +Tjjj=1 +Tiii=1 +Tn=0.1 +--normalize_index("j") +-- +--print_code() +--normalize_index("n") +-- TILE COMMANDS ZEROOOOOOOOOOO:3 +--tile_by_index({"i","j"},{Ti,Tj},{l1_control="ii",l2_control="jj"},{"ii","jj","i","j"})--CU=-1 +tile_by_index({"i"},{Ti},{l1_control="ii"},{"ii","i","j"})--CU=-1 +--normalize_index("i") +--tile_by_index({"n"},{Tn},{l1_control="nn"},{"jj","ii","nn","j","i","n"})--CU=-1 + +--tile_by_index({"j","i"},{Tjjj,Tiii},{l1_control="jjj",l2_control="iii"},{"jj","ii","nn","jjj","j","iii","i","n"})--CU=3 +--tile_by_index({"j"}, {Tn}, {l1_control="j",l1_tile="jjj"}, {"ii", "jj", "nn","jjj","j","i","n"}) +--tile_by_index({"i"}, {Ti/2}, {l1_control="iii"}, {"ii","iii", "jj","i","j"}) +--print_code() +cudaize("kernel_GPU",{oldpos=4*NBODIES,oldpos1=4*NBODIES,oldvel=4*NBODIES,force=4*NBODIES,newpos=4*NBODIES,newvel=4*NBODIES},{block={"ii"}, thread={"i"}})--CU=3 +print_code() +--tile(0,6,6) +--copy_to_shared("tx","oldpos",-16) +--copy_to_registers("j","oldpos") +--copy_to_registers("j","oldpos1") +--copy_to_registers("j","force") + +--copy_to_texture("oldpos") +--tile(1,3,3) +--tile(2,3,3) + +print_code() +--unroll_to_depth(1) +-- +--tile(2,3,3) +--unroll(2,3,0) +--unroll(0,5,0) +--print_code() |