forked from JuliaGPU/CUDA.jl
/
codegen.jl
161 lines (118 loc) · 3.77 KB
/
codegen.jl
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
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
@testset "LLVM IR" begin
@testset "JuliaLang/julia#21121" begin
function foobar()
weight_matrix = CuStaticSharedArray(Float32, (16, 16))
sync_threads()
weight_matrix[1, 16] *= 2
sync_threads()
end
ir = sprint(io->CUDA.code_llvm(io, foobar, Tuple{}))
@test !occursin("inttoptr", ir)
end
@testset "CUDA.jl#553" begin
function kernel(ptr)
unsafe_store!(ptr, CUDA.fma(unsafe_load(ptr), unsafe_load(ptr,2), unsafe_load(ptr,3)))
return
end
ir = sprint(io->CUDA.code_llvm(io, kernel, Tuple{Ptr{Float32}}))
@test !occursin("@__nv_fmaf", ir)
end
@testset "assume" begin
foo(i) = cld(42, i)
ir = sprint(io->CUDA.code_llvm(io, foo, Tuple{Int}))
@test occursin("@gpu_report_exception", ir)
bar(i) = (CUDA.assume(i > 0); cld(42, i))
ir = sprint(io->CUDA.code_llvm(io, bar, Tuple{Int}))
@test !occursin("gpu_report_exception", ir)
end
@testset "stripping invariant.load" begin
function kernel(ptr, x)
i = CUDA.threadIdx_x()
@inbounds ptr[] = x[i]
return
end
arr = CuArray(zeros(Float64))
@cuda kernel(arr, (1., 2., ))
@test Array(arr)[] == 1.
end
@testset "stripping const TBAA" begin
# this one is particularly nasty because it occurs in a nested function
_a = rand(Int, 2, 1)
b = ((1,9999),(1,9999))
out = CuArray(zeros(Int, 2,1))
a = Tuple(_a)
function kernel(out, a, b)
i = threadIdx().x
blockIdx().x
@inbounds out[i,1] = a[i] + b[i][1]
return
end
@cuda threads=2 kernel(out, a, b)
@test Array(out) == (_a .+ 1)
end
@testset "ptxas-compatible control flow" begin
@noinline function throw_some()
throw(42)
return
end
@inbounds function kernel(input, output, n)
i = threadIdx().x
temp = CuStaticSharedArray(Int, 1)
if i == 1
1 <= n || throw_some()
temp[1] = input
end
sync_threads()
1 <= n || throw_some()
unsafe_store!(output, temp[1], i)
return
end
function gpu(input)
output = CuArray(zeros(eltype(input), 2))
ptr = pointer(output)
ptr = reinterpret(Ptr{eltype(input)}, ptr)
@cuda threads=2 kernel(input, ptr, 99)
return Array(output)
end
function cpu(input)
output = zeros(eltype(input), 2)
for j in 1:2
@inbounds output[j] = input
end
return output
end
input = rand(1:100)
@test cpu(input) == gpu(input)
end
end
############################################################################################
@testset "PTX" begin
@testset "local memory stores due to byval" begin
# JuliaGPU/GPUCompiler.jl#92
function kernel(y1, y2)
y = threadIdx().x == 1 ? y1 : y2
@inbounds y[] = 0
return
end
asm = sprint(io->CUDA.code_ptx(io, kernel, NTuple{2,CuDeviceArray{Float32,1,AS.Global}}))
@test !occursin(".local", asm)
end
end
############################################################################################
@testset "SASS" begin
@testset "basic reflection" begin
valid_kernel() = return
invalid_kernel() = 1
@not_if_sanitize @test CUDA.code_sass(devnull, valid_kernel, Tuple{}) == nothing
@not_if_sanitize @test_throws CUDA.KernelError CUDA.code_sass(devnull, invalid_kernel, Tuple{})
end
@testset "function name mangling" begin
@eval @noinline $(Symbol("dummy_^"))(x) = x
@eval kernel_341(ptr) = (@inbounds unsafe_store!(ptr, $(Symbol("dummy_^"))(unsafe_load(ptr))); nothing)
@not_if_sanitize CUDA.code_sass(devnull, kernel_341, Tuple{Ptr{Int}})
end
@testset "device runtime" begin
kernel() = (CUDA.cudaGetLastError(); return)
@not_if_sanitize CUDA.code_sass(devnull, kernel, Tuple{})
end
end