-
Notifications
You must be signed in to change notification settings - Fork 13
/
cl_abstraction.py
174 lines (147 loc) · 6.14 KB
/
cl_abstraction.py
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
from . import pycl as cl
import numpy as np
from ctypes import c_void_p as void_p
def min_ptwo(val, pt):
"Gives the minimum divisionally aligned value for input value"
assert val > 0
assert pt > 0
return ((val - 1) // pt + 1) * pt
class CLType:
def __init__(self, cldev, w, h, simple=False):
self.cldev = cldev
self.simple = simple
self.original_width = w
self.original_height = h
self.width = min_ptwo(w, 8)
self.height = min_ptwo(h, 8)
self.shape = (self.width, self.height)
class CLImage(CLType):
def __init__(self, cldev, w, h):
super().__init__(cldev, w, h, simple=False)
assert self.width % 8 == 0, "Image width must be divisible by 8"
assert self.height % 8 == 0, "Image height must be divisible by 8"
# Default memflags are CL_MEM_READ_WRITE
self.data = cl.clCreateImage2D(
cldev.ctx, self.width, self.height, imgformat=cldev.image_format
)
def from_numpy(self, source):
h, w = source.shape[0], source.shape[1]
if h % 8 != 0 or w % 8 != 0:
padding = np.zeros(
(min_ptwo(h, 8), min_ptwo(w, 8), source.shape[2]), dtype=source.dtype
)
padding[:h, :w] = source[:, :]
source = padding
ary = np.ascontiguousarray(source)
if ary.__array_interface__["strides"]:
raise ValueError("I don't know how to handle strided arrays yet.")
ptr = void_p(ary.__array_interface__["data"][0])
# print(self.data._height, self.data._width, ary.shape)
evt = cl.clEnqueueWriteImage(
self.cldev.queue, self.data, ptr, (0, 0, 0), (ary.shape[1], ary.shape[0], 1), 0, 0
)
evt.wait()
def to_numpy(self):
"See pycl.py for buffer_to_ndarray"
out = np.empty((self.height, self.width, 4), dtype=np.float32)
assert out.flags.contiguous, "Don't know how to write non-contiguous yet."
evt = cl.clEnqueueReadImage(
self.cldev.queue,
self.data,
void_p(out.__array_interface__["data"][0]),
(0, 0, 0),
(self.width, self.height, 1),
self.width * 4 * 4,
0,
)
evt.wait()
return out[: self.original_height, : self.original_width]
class CLFloat2D(CLType):
def __init__(self, cldev, a):
super().__init__(cldev, a.shape[1], a.shape[0], simple=False)
assert a.dtype == np.float32
res, evt = cl.buffer_from_ndarray(self.cldev.queue, a)
evt.wait()
self.data = res
def to_numpy(self):
res, evt = cl.buffer_to_ndarray(
self.cldev.queue, self.data, dtype=np.float32, shape=self.shape
)
evt.wait()
return res
class CLDev:
""" OpenCL device class for 2D-3D image array processing """
def __init__(self, dev_id):
self.ctx = cl.clCreateContext()
for dev in self.ctx.devices:
print(dev.name)
dvids = cl.clGetDeviceIDs()
# print(len(dvids))
d = dvids[dev_id]
cl.clGetDeviceInfo(d, cl.cl_device_info.CL_DEVICE_NAME)
cl.clGetDeviceInfo(d, cl.cl_device_info.CL_DEVICE_TYPE)
print(f"Device {dev_id} available:", d.available)
print("Max work item sizes:", d.max_work_item_sizes)
print("Supported image formats for RGBA:")
self.supported_rgba = [
i.image_channel_data_type
for i in cl.clGetSupportedImageFormats(self.ctx)
if i.image_channel_order == cl.cl_channel_order.CL_RGBA
]
print(self.supported_rgba)
# Ensure we have RGBA float32
assert (
cl.cl_channel_type.CL_FLOAT in self.supported_rgba
), "Your device doesn't support CL_FLOAT for RGBA"
self.queue = cl.clCreateCommandQueue(self.ctx)
self.kernels = {}
self.mem_flags = cl.cl_mem_flags
self.image_format = cl.cl_image_format(
cl.cl_channel_order.CL_RGBA, cl.cl_channel_type.CL_FLOAT
)
def build(self, name, source, argtypes=None):
"Build CL kernel. Load from cache if exists. Returns CL kernel."
if name in self.kernels:
# print("Cache:", name)
return self.kernels[name]
print("Build:", name)
try:
prg = cl.clCreateProgramWithSource(self.ctx, source)
b = prg.build()
kernel = b[name]
except KeyError as e:
print(e)
prg_info = cl.cl_program_build_info
print(cl.clGetProgramBuildInfo(prg, prg_info.CL_PROGRAM_BUILD_STATUS, 0))
print(cl.clGetProgramBuildInfo(prg, prg_info.CL_PROGRAM_BUILD_OPTIONS, 0))
print(cl.clGetProgramBuildInfo(prg, prg_info.CL_PROGRAM_BUILD_LOG, 0))
raise
# kernel.argtypes = (cl.cl_int, cl.cl_int, cl.cl_int, cl.cl_mem, cl.cl_mem, cl.cl_mem)
kernel.argtypes = argtypes
self.kernels[name] = kernel
return kernel
def new_image(self, width, height):
return CLImage(self, width, height)
def new_image_from_ndarray(self, arr):
"shape[0]=height, shape[1]=width"
i = CLImage(self, arr.shape[1], arr.shape[0])
i.from_numpy(arr)
return i
def to_buffer(self, narray):
gc_c, gc_e = cl.buffer_from_ndarray(self.queue, narray)
gc_e.wait()
return gc_c
def run(self, kernel, params, inputs, outputs, shape=None):
"Run CL kernel on params. Multiple in, single out"
assert len(outputs) > 0, "No outputs given"
assert shape is not None, "Invalid kernel shape"
assert type(shape[1]) == int, "Invalid kernel shape"
assert type(shape[0]) == int, "Invalid kernel shape"
assert shape in [(a.width, a.height) for a in outputs], "Kernel shape not in outputs"
assert shape[1] % 8 == 0, "Input image height must be divisible by 8"
assert shape[0] % 8 == 0, "Input image width must be divisible by 8"
# width, height, params, inputs, outputs
run_evt = kernel(shape[0], shape[1], *params, *inputs, *outputs).on(
self.queue, offset=(0, 0), gsize=shape, lsize=(8, 8)
)
run_evt.wait()