-
Notifications
You must be signed in to change notification settings - Fork 288
Expand file tree
/
Copy pathtest_launch_latency.py
More file actions
executable file
·336 lines (234 loc) · 9.24 KB
/
test_launch_latency.py
File metadata and controls
executable file
·336 lines (234 loc) · 9.24 KB
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
191
192
193
194
195
196
197
198
199
200
201
202
203
204
205
206
207
208
209
210
211
212
213
214
215
216
217
218
219
220
221
222
223
224
225
226
227
228
229
230
231
232
233
234
235
236
237
238
239
240
241
242
243
244
245
246
247
248
249
250
251
252
253
254
255
256
257
258
259
260
261
262
263
264
265
266
267
268
269
270
271
272
273
274
275
276
277
278
279
280
281
282
283
284
285
286
287
288
289
290
291
292
293
294
295
296
297
298
299
300
301
302
303
304
305
306
307
308
309
310
311
312
313
314
315
316
317
318
319
320
321
322
323
324
325
326
327
328
329
330
331
332
333
334
335
336
# SPDX-FileCopyrightText: Copyright (c) 2021-2025 NVIDIA CORPORATION & AFFILIATES. All rights reserved.
# SPDX-License-Identifier: LicenseRef-NVIDIA-SOFTWARE-LICENSE
import ctypes
import pytest
from kernels import kernel_string
from conftest import ASSERT_DRV
from cuda.bindings import driver as cuda
def launch(kernel, stream, args=(), arg_types=()):
cuda.cuLaunchKernel(
kernel,
1,
1,
1, # grid dim
1,
1,
1, # block dim
0,
stream, # shared mem and stream
(args, arg_types),
0,
) # arguments
def launch_packed(kernel, stream, params):
cuda.cuLaunchKernel(
kernel,
1,
1,
1, # grid dim
1,
1,
1, # block dim
0,
stream, # shared mem and stream
params,
0,
) # arguments
# Measure launch latency with no parmaeters
@pytest.mark.benchmark(group="launch-latency")
def test_launch_latency_empty_kernel(benchmark, init_cuda, load_module):
device, ctx, stream = init_cuda
module = load_module(kernel_string, device)
err, func = cuda.cuModuleGetFunction(module, b"empty_kernel")
ASSERT_DRV(err)
benchmark(launch, func, stream)
cuda.cuCtxSynchronize()
# Measure launch latency with a single parameter
@pytest.mark.benchmark(group="launch-latency")
def test_launch_latency_small_kernel(benchmark, init_cuda, load_module):
device, ctx, stream = init_cuda
module = load_module(kernel_string, device)
err, func = cuda.cuModuleGetFunction(module, b"small_kernel")
ASSERT_DRV(err)
err, f = cuda.cuMemAlloc(ctypes.sizeof(ctypes.c_float))
ASSERT_DRV(err)
benchmark(launch, func, stream, args=(f,), arg_types=(None,))
cuda.cuCtxSynchronize()
(err,) = cuda.cuMemFree(f)
ASSERT_DRV(err)
# Measure launch latency with many parameters using builtin parameter packing
@pytest.mark.benchmark(group="launch-latency")
def test_launch_latency_small_kernel_512_args(benchmark, init_cuda, load_module):
device, ctx, stream = init_cuda
module = load_module(kernel_string, device)
err, func = cuda.cuModuleGetFunction(module, b"small_kernel_512_args")
ASSERT_DRV(err)
args = []
arg_types = [None] * 512
for _ in arg_types:
err, p = cuda.cuMemAlloc(ctypes.sizeof(ctypes.c_int))
ASSERT_DRV(err)
args.append(p)
args = tuple(args)
arg_types = tuple(arg_types)
benchmark(launch, func, stream, args=args, arg_types=arg_types)
cuda.cuCtxSynchronize()
for p in args:
(err,) = cuda.cuMemFree(p)
ASSERT_DRV(err)
@pytest.mark.benchmark(group="launch-latency")
def test_launch_latency_small_kernel_512_bools(benchmark, init_cuda, load_module):
device, ctx, stream = init_cuda
module = load_module(kernel_string, device)
err, func = cuda.cuModuleGetFunction(module, b"small_kernel_512_bools")
ASSERT_DRV(err)
args = [True] * 512
arg_types = [ctypes.c_bool] * 512
args = tuple(args)
arg_types = tuple(arg_types)
benchmark(launch, func, stream, args=args, arg_types=arg_types)
cuda.cuCtxSynchronize()
@pytest.mark.benchmark(group="launch-latency")
def test_launch_latency_small_kernel_512_doubles(benchmark, init_cuda, load_module):
device, ctx, stream = init_cuda
module = load_module(kernel_string, device)
err, func = cuda.cuModuleGetFunction(module, b"small_kernel_512_doubles")
ASSERT_DRV(err)
args = [1.2345] * 512
arg_types = [ctypes.c_double] * 512
args = tuple(args)
arg_types = tuple(arg_types)
benchmark(launch, func, stream, args=args, arg_types=arg_types)
cuda.cuCtxSynchronize()
@pytest.mark.benchmark(group="launch-latency")
def test_launch_latency_small_kernel_512_ints(benchmark, init_cuda, load_module):
device, ctx, stream = init_cuda
module = load_module(kernel_string, device)
err, func = cuda.cuModuleGetFunction(module, b"small_kernel_512_ints")
ASSERT_DRV(err)
args = [123] * 512
arg_types = [ctypes.c_int] * 512
args = tuple(args)
arg_types = tuple(arg_types)
benchmark(launch, func, stream, args=args, arg_types=arg_types)
cuda.cuCtxSynchronize()
@pytest.mark.benchmark(group="launch-latency")
def test_launch_latency_small_kernel_512_bytes(benchmark, init_cuda, load_module):
device, ctx, stream = init_cuda
module = load_module(kernel_string, device)
err, func = cuda.cuModuleGetFunction(module, b"small_kernel_512_chars")
ASSERT_DRV(err)
args = [127] * 512
arg_types = [ctypes.c_byte] * 512
args = tuple(args)
arg_types = tuple(arg_types)
benchmark(launch, func, stream, args=args, arg_types=arg_types)
cuda.cuCtxSynchronize()
@pytest.mark.benchmark(group="launch-latency")
def test_launch_latency_small_kernel_512_longlongs(benchmark, init_cuda, load_module):
device, ctx, stream = init_cuda
module = load_module(kernel_string, device)
err, func = cuda.cuModuleGetFunction(module, b"small_kernel_512_longlongs")
ASSERT_DRV(err)
args = [9223372036854775806] * 512
arg_types = [ctypes.c_longlong] * 512
args = tuple(args)
arg_types = tuple(arg_types)
benchmark(launch, func, stream, args=args, arg_types=arg_types)
cuda.cuCtxSynchronize()
# Measure launch latency with many parameters using builtin parameter packing
@pytest.mark.benchmark(group="launch-latency")
def test_launch_latency_small_kernel_256_args(benchmark, init_cuda, load_module):
device, ctx, stream = init_cuda
module = load_module(kernel_string, device)
err, func = cuda.cuModuleGetFunction(module, b"small_kernel_256_args")
ASSERT_DRV(err)
args = []
arg_types = [None] * 256
for _ in arg_types:
err, p = cuda.cuMemAlloc(ctypes.sizeof(ctypes.c_int))
ASSERT_DRV(err)
args.append(p)
args = tuple(args)
arg_types = tuple(arg_types)
benchmark(launch, func, stream, args=args, arg_types=arg_types)
cuda.cuCtxSynchronize()
for p in args:
(err,) = cuda.cuMemFree(p)
ASSERT_DRV(err)
# Measure launch latency with many parameters using builtin parameter packing
@pytest.mark.benchmark(group="launch-latency")
def test_launch_latency_small_kernel_16_args(benchmark, init_cuda, load_module):
device, ctx, stream = init_cuda
module = load_module(kernel_string, device)
err, func = cuda.cuModuleGetFunction(module, b"small_kernel_16_args")
ASSERT_DRV(err)
args = []
arg_types = [None] * 16
for _ in arg_types:
err, p = cuda.cuMemAlloc(ctypes.sizeof(ctypes.c_int))
ASSERT_DRV(err)
args.append(p)
args = tuple(args)
arg_types = tuple(arg_types)
benchmark(launch, func, stream, args=args, arg_types=arg_types)
cuda.cuCtxSynchronize()
for p in args:
(err,) = cuda.cuMemFree(p)
ASSERT_DRV(err)
# Measure launch latency with many parameters, excluding parameter packing
@pytest.mark.benchmark(group="launch-latency")
def test_launch_latency_small_kernel_512_args_ctypes(benchmark, init_cuda, load_module):
device, ctx, stream = init_cuda
module = load_module(kernel_string, device)
err, func = cuda.cuModuleGetFunction(module, b"small_kernel_512_args")
ASSERT_DRV(err)
vals = []
val_ps = []
for i in range(512):
err, p = cuda.cuMemAlloc(ctypes.sizeof(ctypes.c_int))
ASSERT_DRV(err)
vals.append(p)
val_ps.append(ctypes.c_void_p(int(vals[i])))
packagedParams = (ctypes.c_void_p * 512)()
for i in range(512):
packagedParams[i] = ctypes.addressof(val_ps[i])
benchmark(launch_packed, func, stream, packagedParams)
cuda.cuCtxSynchronize()
for p in vals:
(err,) = cuda.cuMemFree(p)
ASSERT_DRV(err)
def pack_and_launch(kernel, stream, params):
packed_params = (ctypes.c_void_p * len(params))()
ptrs = [0] * len(params)
for i in range(len(params)):
ptrs[i] = ctypes.c_void_p(int(params[i]))
packed_params[i] = ctypes.addressof(ptrs[i])
cuda.cuLaunchKernel(kernel, 1, 1, 1, 1, 1, 1, 0, stream, packed_params, 0)
# Measure launch latency plus parameter packing using ctypes
@pytest.mark.benchmark(group="launch-latency")
def test_launch_latency_small_kernel_512_args_ctypes_with_packing(benchmark, init_cuda, load_module):
device, ctx, stream = init_cuda
module = load_module(kernel_string, device)
err, func = cuda.cuModuleGetFunction(module, b"small_kernel_512_args")
ASSERT_DRV(err)
vals = []
for i in range(512):
err, p = cuda.cuMemAlloc(ctypes.sizeof(ctypes.c_int))
ASSERT_DRV(err)
vals.append(p)
benchmark(pack_and_launch, func, stream, vals)
cuda.cuCtxSynchronize()
for p in vals:
(err,) = cuda.cuMemFree(p)
ASSERT_DRV(err)
# Measure launch latency with a single large struct parameter
@pytest.mark.benchmark(group="launch-latency")
def test_launch_latency_small_kernel_2048B(benchmark, init_cuda, load_module):
device, ctx, stream = init_cuda
module = load_module(kernel_string, device)
err, func = cuda.cuModuleGetFunction(module, b"small_kernel_2048B")
ASSERT_DRV(err)
class struct_2048B(ctypes.Structure):
_fields_ = [("values", ctypes.c_uint8 * 2048)]
benchmark(launch, func, stream, args=(struct_2048B(),), arg_types=(None,))
cuda.cuCtxSynchronize()