1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
|
--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("mriq.c", "ComputeQCPU", 0)
dofile("cudaize.lua") --defines custom tile_by_index, copy_to_registers,
--copy_to_shared methods
N=32768
M=3072
TI=128
TJ=128
permute(0,{"j","i"})
--tile_by_index({"j","i"}, {TI,TJ}, {l1_control="jj", l2_control="ii"}, {"jj","ii", "j", "i"})
tile_by_index({"i"}, {TJ}, {l1_control="ii",l1_tile="i"}, {"ii", "j","i"})
tile_by_index({"j"}, {TI}, {l1_control="jj"}, {"ii","jj", "j", "i"})
--tile_by_index({"i"}, {TI}, {l1_control="ii"}, {"ii", "i", "j"})
--print_code()
normalize_index("j")
normalize_index("i")
--print_code()
--tile_by_index({"i"}, {TI}, {l1_control="iii",l1_tile="i"}, {"ii","jj", "iii","j","i"})
--print_code()
cudaize("Kernel_GPU", {x=N,y=N,z=N,Qr=N,Qi=N,kVals=M},{block={"jj"}, thread={"j"}})
copy_to_shared("tx","kVals",1)
--copy_to_shared("tx","x",1)
--copy_to_shared("tx","y",1)
--copy_to_shared("tx","z",1)
--copy_to_texture("kVals")
--datacopy(0, 3, "kVals", {"tt","t"},false,0,1,-16,true)
--print_code()
--datacopy_privatized(0,"tx","kVals",{"tx"})
--copy_to_registers("tx","kVals")
copy_to_registers("ii","x")
copy_to_registers("ii","y")
copy_to_registers("ii","z")
copy_to_registers("ii","Qi")
copy_to_registers("ii","Qr")
--[[datacopy_privatized(0,"tx","x",{"tx"})
datacopy_privatized(0,"tx","y",{"tx"})
datacopy_privatized(0,"tx","z",{"tx"})
datacopy_privatized(0,"tx","Qi",{"tx"})
datacopy_privatized(0,"tx","Qr",{"tx"})
]]--
--unroll(0,5,64)
print_code()
--unroll_to_depth(1) --won't unroll past thread/loop mapping, unrolls up to two loop levels
|
