-
Notifications
You must be signed in to change notification settings - Fork 203
/
classifyfloatstructs.t
137 lines (119 loc) · 4.38 KB
/
classifyfloatstructs.t
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
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
-- reading and returning by reference produces far more compact assembly
-- in terra than reading and returning by value, even when a by-value function
-- is embedded in a by-reference wrapper. Also, optimization behavior varies
-- wildly. Here's a demonstration:
local float4 = tuple(float,float,float,float)
-- first, comparing assembly code size of stand-alone functions which all do
-- the same operation, in different styles. This is after aggressive
-- optimization by the LLVM backend
--------------------------------------------------------------------------------
-- read by value, return by value
-- 26 instructions, 120 bytes
local terra add4_val_val(a : float4, b : float4) : float4
return {
a._0+b._0,
a._1+b._1,
a._2+b._2,
a._3+b._3
}
end
-- read by value, return by reference
-- 24 instructions, 103 bytes
local terra add4_val_ref(c : &float4, a : float4, b: float4) : {}
c._0 = a._0 + b._0
c._1 = a._1 + b._1
c._2 = a._2 + b._2
c._3 = a._3 + b._3
end
-- read by reference, return by value
-- 14 instructions, 74 bytes
local terra add4_ref_val(a : &float4, b : &float4) : float4
return {
a._0+b._0,
a._1+b._1,
a._2+b._2,
a._3+b._3
}
end
-- read by reference, return by reference
-- 12 instructions, 57 bytes
local terra add4_ref_ref(c : &float4, a : &float4, b: &float4) : {}
c._0 = a._0 + b._0
c._1 = a._1 + b._1
c._2 = a._2 + b._2
c._3 = a._3 + b._3
end
-- read by reference, return by reference, BUT use temporary variables
-- 4 instructions, 12 bytes (!!!)
-- what happens here is that the tempvars are understood and treated
-- as a single SIMD register, and so only one addition is executed.
-- this is already reflected in the bytecode passed to LLVM, so I suppose
-- terra does this optimization.
local terra add4_ref_ref_tempvar(c : &float4, a : &float4, b: &float4) : {}
var x = a._0 + b._0
var y = a._1 + b._1
var z = a._2 + b._2
var w = a._3 + b._3
c._0 = x
c._1 = y
c._2 = z
c._3 = w
end
-- turn on always-inline for later
add4_val_val:setinlined(true)
add4_val_ref:setinlined(true)
add4_ref_val:setinlined(true)
add4_ref_ref:setinlined(true)
-- uncomment to look at individual LLVM bytecode & disassembly
add4_val_val:disas()
-- add4_val_ref:disas()
-- add4_ref_val:disas()
-- add4_ref_ref:disas()
-- add4_ref_ref_tempvar:disas()
if terralib.lookupsymbol and require("ffi").os ~= "Windows" then
terra sizecheck()
var si : terralib.SymbolInfo
terralib.lookupsymbol(add4_val_val,&si)
return si.size
end
assert(sizecheck() < 16)
end
--------------------------------------------------------------------------------
-- up to this point, one could argue that functions are always inlined into
-- other functions, and that the instructions dissolve in the greater scheme
-- of things.
-- if that is true, let's attempt to convert one style into another and see
-- if the optimizations catch up.
-- read by value, return by value -> read by reference, return by reference
-- the clunky solution to the slim interface
-- 38 instructions (!), 193 bytes (!!), so it gets *worse*
local terra add4_val_val_to_ref_ref(c : &float4, a : &float4, b: &float4) : {}
@c = add4_val_val(@a, @b)
end
-- read by reference, return by reference -> read by value, return by value
-- the slim solution to the clunky interface
-- 13 instructions, 61 bytes with tempvar -- wow, that's better than
-- the original function with the actual code in it!
local terra add4_ref_ref_to_val_val(a : float4, b : float4) : float4
var c : float4
add4_ref_ref_tempvar(&c, &a, &b)
return c
end
-- so what happens if we do a conversion back to the by-reference interface?
-- 41 instructions, 194 bytes
local terra add4_ref_ref_to_val_val_to_ref_ref(c : &float4, a : &float4, b: &float4) : {}
@c = add4_ref_ref_to_val_val(@a, @b)
end
-- and nest it once more, back to the by-value interface
-- 47 instructions, 208 bytes
-- so once we pass structs by-value, we'll never recover.
local terra add4_ref_ref_to_val_val_to_ref_ref_to_val_val(a : float4, b : float4) : float4
var c : float4
add4_ref_ref_to_val_val_to_ref_ref(&c, &a, &b)
return c
end
-- uncomment to look at individual disassembly
-- add4_val_val_to_ref_ref:disas()
-- add4_ref_ref_to_val_val:disas()
-- add4_ref_ref_to_val_val_to_ref_ref:disas()
-- add4_ref_ref_to_val_val_to_ref_ref_to_val_val:disas()