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
|
from . import benchmark
import itertools
import numpy as np
import torch
class BroadcastMulBench(benchmark.Benchmark):
def __init__(self, mode, device, dtype, case, M, N, K):
super().__init__(mode, device, dtype)
self.case = case
self.M = M
self.N = N
self.K = K
if case == "row":
self.d1 = self.rand(
[M, N, 1], device=device, dtype=dtype, requires_grad=self.requires_grad
)
self.d2 = self.rand(
[M, 1, K], device=device, dtype=dtype, requires_grad=self.requires_grad
)
elif case == "mid":
self.d1 = self.rand(
[M, N, 1], device=device, dtype=dtype, requires_grad=self.requires_grad
)
self.d2 = self.rand(
[1, N, K], device=device, dtype=dtype, requires_grad=self.requires_grad
)
elif case == "col":
self.d1 = self.rand(
[M, 1, K], device=device, dtype=dtype, requires_grad=self.requires_grad
)
self.d2 = self.rand(
[1, N, K], device=device, dtype=dtype, requires_grad=self.requires_grad
)
else:
raise ValueError("invalid case: %s" % (case))
self.inputs = [self.d1, self.d2]
def forward(self, d1, d2):
y = d1 + d2
return y
def reference(self):
return self.numpy(self.d1) + self.numpy(self.d2)
def config(self):
return [self.M, self.N, self.K]
@staticmethod
def default_configs():
return [[128, 256, 128]]
def memory_workload(self):
if self.mode == "fwd":
sol_count = 1
algorithmic_count = 1
else:
sol_count = (1) + (1)
algorithmic_count = 1 + (1 + 1)
buffer_size = self.M * self.N * self.K
return {
"sol": buffer_size * sol_count,
"algorithmic": buffer_size * algorithmic_count,
}
class BroadcastRowBench(BroadcastMulBench):
def __init__(self, mode, device, dtype, M, N, K):
super(BroadcastRowBench, self).__init__(mode, device, dtype, "row", M, N, K)
@staticmethod
def module():
return "broadcast_row"
class BroadcastMidBench(BroadcastMulBench):
def __init__(self, mode, device, dtype, M, N, K):
super(BroadcastMidBench, self).__init__(mode, device, dtype, "mid", M, N, K)
@staticmethod
def module():
return "broadcast_mid"
class BroadcastColBench(BroadcastMulBench):
def __init__(self, mode, device, dtype, M, N, K):
super(BroadcastColBench, self).__init__(mode, device, dtype, "col", M, N, K)
@staticmethod
def module():
return "broadcast_col"
class BroadcastThreeArgs(benchmark.Benchmark):
def __init__(self, mode, device, dtype, M, N, K, L):
super().__init__(mode, device, dtype)
self.M = M
self.N = N
self.K = K
self.L = L
self.d1 = self.rand([M, N], device=device, dtype=dtype, requires_grad=self.requires_grad)
self.d2 = self.rand([K, M, 1], device=device, dtype=dtype, requires_grad=self.requires_grad)
self.d3 = self.rand(
[L, K, 1, 1], device=device, dtype=dtype, requires_grad=self.requires_grad
)
self.inputs = [self.d1, self.d2, self.d3]
def forward(self, d1, d2, d3):
y = d1 + d2 + d3
return y
def reference(self):
return self.numpy(self.d1) + self.numpy(self.d2) + self.numpy(self.d3)
def config(self):
return [self.M, self.N, self.K, self.L]
@staticmethod
def default_configs():
return [[32, 16, 64, 128]]
def memory_workload(self):
if self.mode == "fwd":
sol_count = 1
algorithmic_count = 1
else:
sol_count = (1) + (1)
algorithmic_count = 1 + (1 + 1 + 1)
buffer_size = self.M * self.N * self.K * self.L * 4
return {
"sol": buffer_size * sol_count,
"algorithmic": buffer_size * algorithmic_count,
}
@staticmethod
def module():
return "broadcast_3args"
# benchmark.register_benchmark_class(BroadcastRowBench)
# benchmark.register_benchmark_class(BroadcastMidBench)
# benchmark.register_benchmark_class(BroadcastColBench)
# benchmark.register_benchmark_class(BroadcastThreeArgs)
# TODO: merge this with elementwise bench
# A template class for elementwise operations.
# A derived class will override the class instance to customize its behavior.
class BroadcastBench(benchmark.Benchmark):
# List of customization class variables.
op_str = None
binary_op_pt_func = None
binary_op_np_func = None
unary_op_pt_func = None
unary_op_np_func = None
split_input = True
def __init__(self, mode, device, dtype, M, N, K):
super().__init__(mode, device, dtype)
self.M = M
self.N = N
self.K = K
self.d1 = self.rand([M, N], device=device, dtype=dtype, requires_grad=self.requires_grad)
self.d2 = self.rand([K, 1, N], device=device, dtype=dtype, requires_grad=self.requires_grad)
self.d3 = self.rand([M, N], device=device, dtype=dtype, requires_grad=self.requires_grad)
self.d4 = self.rand([K, M, 1], device=device, dtype=dtype, requires_grad=self.requires_grad)
self.inputs = [self.d1, self.d2, self.d3, self.d4]
def _eval(self, d1, d2, d3, d4, binary_op, unary_op):
if not binary_op:
def binary_op(x, y):
return x + y
if not unary_op:
def unary_op(x):
return x
if self.split_input:
d1 = unary_op(d1)
d2 = unary_op(d2)
d3 = unary_op(d3)
d4 = unary_op(d4)
else:
d1, d2, d3, d4 = (
unary_op(d1),
unary_op(d2),
unary_op(d1 + 0.001),
unary_op(d4),
)
a = binary_op(d1, d2)
b = binary_op(d3, d4)
c = a + b
return c
def forward(self, d1, d2, d3, d4):
binary_op = self.__class__.binary_op_pt_func
unary_op = self.__class__.unary_op_pt_func
return self._eval(d1, d2, d3, d4, binary_op, unary_op)
def reference(self):
binary_op = self.__class__.binary_op_np_func
unary_op = self.__class__.unary_op_np_func
[d1, d2, d3, d4] = [self.numpy(d) for d in [self.d1, self.d2, self.d3, self.d4]]
return self._eval(d1, d2, d3, d4, binary_op, unary_op)
def config(self):
return [self.M, self.N, self.K]
@classmethod
def module(cls):
return "broadcast_" + cls.op_str
def memory_workload(self):
input_count = len(self.inputs)
if self.mode == "fwd":
if self.split_input:
sol_count = 1
algorithmic_count = 1
else:
sol_count = 1
algorithmic_count = 1
else:
if self.split_input:
sol_count = 1
algorithmic_count = input_count
else:
sol_count = 1
algorithmic_count = input_count
buffer_size = self.M * self.N * self.K * 4
return {
"sol": buffer_size * sol_count,
"algorithmic": buffer_size * algorithmic_count,
}
@staticmethod
def default_configs():
return [[1 << 8, 1 << 7, 1 << 9]]
def register_broadcast_ops():
binary_op_list = [
["mul", lambda a, b: a * b],
["add", lambda a, b: a + b],
["sub", lambda a, b: a - b],
["div", lambda a, b: a / (b + 1e-4)],
[
"pow",
lambda a, b: torch.pow(a, b),
lambda a, b: np.power(a, b),
], # no fuson triggered
["max", lambda a, b: torch.max(a, b), lambda a, b: np.maximum(a, b)],
["min", lambda a, b: torch.min(a, b), lambda a, b: np.minimum(a, b)],
]
unary_op_list = [
["erf", lambda x: torch.erf(x), lambda x: np.erf(x)],
["exp", lambda x: torch.exp(x), lambda x: np.exp(x)],
["sin", lambda x: torch.sin(x), lambda x: np.sin(x)],
["cos", lambda x: torch.cos(x), lambda x: np.cos(x)],
]
for split_input, binary_op in itertools.product([True, False], binary_op_list):
# Make a copy of BroadcastBench
if len(binary_op) == 2:
[op_str, op_pt_func] = binary_op
op_np_func = op_pt_func
elif len(binary_op) == 3:
[op_str, op_pt_func, op_np_func] = binary_op
split_str = "split" if split_input else "shared"
op_str = split_str + "_" + op_str
bm_cls = type("BroadcastBench_" + op_str, (BroadcastBench,), {})
bm_cls.op_str = op_str
bm_cls.binary_op_pt_func = op_pt_func
bm_cls.binary_op_np_func = op_np_func
bm_cls.split_input = split_input
benchmark.register_benchmark_class(bm_cls)
for split_input, unary_op in itertools.product([True, False], unary_op_list):
# Make a copy of BroadcastBench
if len(unary_op) == 2:
[op_str, op_pt_func] = unary_op
op_np_func = op_pt_func
elif len(unary_op) == 3:
[op_str, op_pt_func, op_np_func] = unary_op
split_str = "split" if split_input else "shared"
op_str = split_str + "_" + op_str
bm_cls = type("BroadcastBench_" + op_str, (BroadcastBench,), {})
bm_cls.op_str = op_str
bm_cls.unary_op_pt_func = op_pt_func
bm_cls.unary_op_np_func = op_np_func
bm_cls.split_input = split_input
benchmark.register_benchmark_class(bm_cls)
register_broadcast_ops()
|
