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
162
163
164
165
166
167
168
169
170
171
172
173
174
175
176
177
178
179
180
181
182
183
184
185
186
187
188
189
190
|
"""
Bilateral histogram.
"""
import halide as hl
@hl.alias(
bilateral_grid_Adams2019={"autoscheduler": "Adams2019"},
bilateral_grid_Mullapudi2016={"autoscheduler": "Mullapudi2016"},
bilateral_grid_Li2018={"autoscheduler": "Li2018"},
)
@hl.generator()
class bilateral_grid:
s_sigma = hl.GeneratorParam(8)
input_buf = hl.InputBuffer(hl.Float(32), 2)
r_sigma = hl.InputScalar(hl.Float(32))
bilateral_grid = hl.OutputBuffer(hl.Float(32), 2)
def generate(self):
g = self
x, y, z, c = hl.vars("x y z c")
# Add a boundary condition
clamped = hl.BoundaryConditions.repeat_edge(g.input_buf)
# Construct the bilateral grid
r = hl.RDom([(0, g.s_sigma), (0, g.s_sigma)])
val = clamped[
x * g.s_sigma + r.x - g.s_sigma // 2,
y * g.s_sigma + r.y - g.s_sigma // 2,
]
val = hl.clamp(val, 0.0, 1.0)
zi = hl.i32(val / g.r_sigma + 0.5)
histogram = hl.Func("histogram")
histogram[x, y, z, c] = 0.0
histogram[x, y, zi, c] += hl.mux(c, [val, 1.0])
# Blur the histogram using a five-tap filter
blurx, blury, blurz = hl.funcs("blurx blury blurz")
blurz[x, y, z, c] = (
histogram[x, y, z - 2, c]
+ histogram[x, y, z - 1, c] * 4
+ histogram[x, y, z, c] * 6
+ histogram[x, y, z + 1, c] * 4
+ histogram[x, y, z + 2, c]
)
blurx[x, y, z, c] = (
blurz[x - 2, y, z, c]
+ blurz[x - 1, y, z, c] * 4
+ blurz[x, y, z, c] * 6
+ blurz[x + 1, y, z, c] * 4
+ blurz[x + 2, y, z, c]
)
blury[x, y, z, c] = (
blurx[x, y - 2, z, c]
+ blurx[x, y - 1, z, c] * 4
+ blurx[x, y, z, c] * 6
+ blurx[x, y + 1, z, c] * 4
+ blurx[x, y + 2, z, c]
)
# Take trilinear samples to compute the output
val = hl.clamp(clamped[x, y], 0.0, 1.0)
zv = val / g.r_sigma
zi = hl.i32(zv)
zf = zv - zi
xf = hl.f32(x % g.s_sigma) / g.s_sigma
yf = hl.f32(y % g.s_sigma) / g.s_sigma
xi = x / g.s_sigma
yi = y / g.s_sigma
interpolated = hl.Func("interpolated")
interpolated[x, y, c] = hl.lerp(
hl.lerp(
hl.lerp(blury[xi, yi, zi, c], blury[xi + 1, yi, zi, c], xf),
hl.lerp(blury[xi, yi + 1, zi, c], blury[xi + 1, yi + 1, zi, c], xf),
yf,
),
hl.lerp(
hl.lerp(blury[xi, yi, zi + 1, c], blury[xi + 1, yi, zi + 1, c], xf),
hl.lerp(
blury[xi, yi + 1, zi + 1, c], blury[xi + 1, yi + 1, zi + 1, c], xf
),
yf,
),
zf,
)
# Normalize
g.bilateral_grid[x, y] = interpolated[x, y, 0] / interpolated[x, y, 1]
# ESTIMATES
# (This can be useful in conjunction with RunGen and benchmarks as well
# as auto-schedule, so we do it in all cases.)
# Provide estimates on the input image
g.input_buf.set_estimates([(0, 1536), (0, 2560)])
# Provide estimates on the parameters
g.r_sigma.set_estimate(0.1)
# TODO: Compute estimates from the parameter values
histogram.set_estimate(z, -2, 16)
blurz.set_estimate(z, 0, 12)
blurx.set_estimate(z, 0, 12)
blury.set_estimate(z, 0, 12)
g.bilateral_grid.set_estimates([(0, 1536), (0, 2560)])
if g.using_autoscheduler():
# nothing
pass
elif g.target().has_gpu_feature():
# 0.50ms on an RTX 2060
xi, yi, zi = hl.vars("xi yi zi")
# Schedule blurz in 8x8 tiles. This is a tile in
# grid-space, which means it represents something like
# 64x64 pixels in the input (if s_sigma is 8).
blurz.compute_root().reorder(c, z, x, y).gpu_tile(x, y, xi, yi, 8, 8)
# Schedule histogram to happen per-tile of blurz, with
# intermediate results in shared memory. This means histogram
# and blurz makes a three-stage kernel:
# 1) Zero out the 8x8 set of histograms
# 2) Compute those histogram by iterating over lots of the input image
# 3) Blur the set of histograms in z
histogram.reorder(c, z, x, y).compute_at(blurz, x).gpu_threads(x, y)
histogram.update().reorder(c, r.x, r.y, x, y).gpu_threads(x, y).unroll(c)
# Schedule the remaining blurs and the sampling at the end
# similarly.
(
blurx.compute_root()
.reorder(c, x, y, z)
.reorder_storage(c, x, y, z)
.vectorize(c)
.unroll(y, 2, hl.TailStrategy.RoundUp)
.gpu_tile(x, y, z, xi, yi, zi, 32, 8, 1, hl.TailStrategy.RoundUp)
)
(
blury.compute_root()
.reorder(c, x, y, z)
.reorder_storage(c, x, y, z)
.vectorize(c)
.unroll(y, 2, hl.TailStrategy.RoundUp)
.gpu_tile(x, y, z, xi, yi, zi, 32, 8, 1, hl.TailStrategy.RoundUp)
)
g.bilateral_grid.compute_root().gpu_tile(x, y, xi, yi, 32, 8)
interpolated.compute_at(g.bilateral_grid, xi).vectorize(c)
else:
# CPU schedule.
# 3.98ms on an Intel i9-9960X using 32 threads at 3.7 GHz
# using target x86-64-avx2. This is a little less
# SIMD-friendly than some of the other apps, so we
# benefit from hyperthreading, and don't benefit from
# AVX-512, which on my machine reduces the clock to 3.0
# GHz.
(
blurz.compute_root()
.reorder(c, z, x, y)
.parallel(y)
.vectorize(x, 8)
.unroll(c)
)
histogram.compute_at(blurz, y)
histogram.update().reorder(c, r.x, r.y, x, y).unroll(c)
(
blurx.compute_root() #
.reorder(c, x, y, z) #
.parallel(z) #
.vectorize(x, 8) #
.unroll(c)
)
(
blury.compute_root()
.reorder(c, x, y, z)
.parallel(z)
.vectorize(x, 8)
.unroll(c)
)
g.bilateral_grid.compute_root().parallel(y).vectorize(x, 8)
if __name__ == "__main__":
hl.main()
|
