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
|
import torch
import torch._C._te as te
import time
import numpy as np
import pandas as pd
import matplotlib.pyplot as plt
import seaborn as sns
import argparse
class kernel_arena_scope(object):
def __enter__(self):
self.scope = te.KernelScope()
def __exit__(self, typ, val, traceback):
self.scope = None
unary_ops = [
("sin", torch.sin),
("cos", torch.cos),
("tan", torch.tan),
("asin", torch.asin),
("acos", torch.acos),
("atan", torch.atan),
("sinh", torch.sinh),
("cosh", torch.cosh),
("tanh", torch.tanh),
("sigmoid", torch.sigmoid),
("exp", torch.exp),
("expm1", torch.expm1),
("expm1", torch.expm1),
("abs", torch.abs),
("log", torch.log),
("fast_log", torch.log),
("log2", torch.log2),
("log10", torch.log10),
("log1p", torch.log1p),
("erf", torch.erf),
("erfc", torch.erfc),
("sqrt", torch.sqrt),
("rsqrt", torch.rsqrt),
("ceil", torch.ceil),
("floor", torch.floor),
("round", torch.round),
("trunc", torch.trunc),
("lgamma", torch.lgamma),
# ("frac", torch.frac), # seems unimplemented
# ("isnan", torch.isnan), # no out variant
]
def gen_unary_nnc_fun(nnc_name):
def nnc_fun(A, B):
def compute(i, j):
return getattr(A.load([i, j]), nnc_name)()
return compute
return nnc_fun
def gen_unary_torch_fun(torch_op):
def torch_fun(a, b, out):
def fun():
return torch_op(a, out=out)
return fun
return torch_fun
def gen_binary_nnc_fun(fn):
def nnc_fun(A, B):
def compute(i, j):
return fn(A.load([i, j]), B.load([i, j]))
return compute
return nnc_fun
def gen_binary_torch_fun(fn):
def pt_fun(a, b, out):
def fun():
return fn(a, b, out=out)
return fun
return pt_fun
def gen_int_comparison_tensors(N, M):
return (torch.randint(0, 3, (N, M)), torch.randint(0, 3, (N, M)), torch.empty((N, M), dtype=torch.bool))
def gen_float_comparison_tensors(N, M):
return (torch.rand(N, M), torch.rand(N, M), torch.empty((N, M), dtype=torch.bool))
te_bool = te.Dtype.Bool
binary_ops = [
('add', (lambda a, b: a + b), torch.add),
('mul', (lambda a, b: a * b), torch.mul),
('sub', (lambda a, b: a - b), torch.sub),
('div', (lambda a, b: a / b), torch.div),
('eq', (lambda a, b: te.Cast.make(te_bool, a == b)), torch.eq, gen_int_comparison_tensors),
('gt', (lambda a, b: te.Cast.make(te_bool, a > b)), torch.gt, gen_float_comparison_tensors),
('lt', (lambda a, b: te.Cast.make(te_bool, a < b)), torch.lt, gen_float_comparison_tensors),
('gte', (lambda a, b: te.Cast.make(te_bool, a >= b)), torch.greater_equal, gen_float_comparison_tensors),
('lte', (lambda a, b: te.Cast.make(te_bool, a <= b)), torch.less_equal, gen_float_comparison_tensors),
# ('neq', (lambda a, b: a != b), None)), # no one-op equivalent
# ('&', (lambda a, b: a & b), torch.bitwise_and), # requires more work to test
]
def nnc_relu(A, B):
def f(i, j):
return torch._C._te.ifThenElse(A.load([i, j]) < torch._C._te.ExprHandle.float(0),
torch._C._te.ExprHandle.float(0), A.load([i, j]))
return f
def pt_relu(a, b, c):
return torch.relu(a)
custom_ops = [
('relu', nnc_relu, pt_relu),
# ('nnc_mul_relu', nnc_mul_relu, pt_mul_relu)
# ('manual_sigmoid', nnc_manual_sigmoid, lambda a, b, c: torch.sigmoid(a, out=c))
]
def gen_custom_torch_fun(fn):
def pt_fun(a, b, out):
def fun():
return fn(a, b, out)
return fun
return pt_fun
def normalize_benchmarks(ops):
return [i + (None,) if len(i) == 3 else i for i in ops]
names = []
nnc_fns = []
pt_fns = []
shape_fns = []
for nnc_name, pt_op in unary_ops:
names.append(nnc_name)
nnc_fns.append(gen_unary_nnc_fun(nnc_name))
pt_fns.append(gen_unary_torch_fun(pt_op))
shape_fns.append(None)
for name, lmbda, pt_fn, shape_fn in normalize_benchmarks(binary_ops):
names.append(name)
nnc_fns.append(gen_binary_nnc_fun(lmbda))
pt_fns.append(gen_binary_torch_fun(pt_fn))
shape_fns.append(shape_fn)
for name, lmbda, pt_fn, shape_fn in normalize_benchmarks(custom_ops):
names.append(name)
nnc_fns.append(lmbda)
pt_fns.append(gen_custom_torch_fun(pt_fn))
shape_fns.append(shape_fn)
benchmarks = list(zip(names, nnc_fns, pt_fns, shape_fns))
def run_benchmarks(benchmarks, sizes):
df = pd.DataFrame(columns=['name', 'N', 'M', 'nnc_time', 'torch_time', 'ratio'])
with torch.no_grad():
for name, nnc_fun, torch_fun, shape_fn in benchmarks:
for N, M in sizes:
iters = int(1e6 / (N + M))
with kernel_arena_scope():
if shape_fn is None:
tA = torch.rand(M, N).clamp(0.01, 0.99)
tB = torch.rand(M, N).clamp(0.01, 0.99)
tX = torch.empty(M, N)
tR = torch.empty(M, N)
else:
tA, tB, tX = shape_fn(M, N)
tR = tX.clone()
def get_nnc_type(dtype):
if dtype == torch.float:
return torch._C._te.Dtype.Float
elif dtype == torch.long:
return torch._C._te.Dtype.Long
dtype = get_nnc_type(tA.dtype)
dM = torch._C._te.ExprHandle.int(M)
dN = torch._C._te.ExprHandle.int(N)
A = torch._C._te.Placeholder('A', dtype, [dM, dN])
B = torch._C._te.Placeholder('B', dtype, [dM, dN])
dim_args = [torch._C._te.DimArg(*args) for args in [(dM, 'm'), (dN, 'n')]]
compute = nnc_fun(A, B)
X = torch._C._te.Compute('X', dim_args, compute)
loopnest = torch._C._te.LoopNest([X])
loopnest.prepare_for_codegen()
stmt = torch._C._te.simplify(loopnest.root_stmt())
cg = torch._C._te.construct_codegen('llvm', stmt, [torch._C._te.BufferArg(x) for x in [A, B, X]])
# warmup
for _ in range(10):
cg.call([tA, tB, tX])
start = time.time()
for it in range(iters):
cg.call([tA, tB, tX])
time1 = time.time() - start
fn = torch_fun(tA, tB, tR)
# warmup
for _ in range(10):
tR = fn()
start = time.time()
for it in range(iters):
tR = fn()
time2 = time.time() - start
df = df.append({'name': name, 'N': N, 'M': M, 'nnc_time': time1,
'torch_time': time2, 'ratio': time2 / time1}, ignore_index=True)
print(name, N, M)
print(time2 / time1, time1, time2)
print()
def check_correctness(a, b):
if not np.allclose(a, b):
print(name)
assert(np.allclose(a, b))
check_correctness(tX, tR)
return df
def dump_plot(df, sizes):
keys = []
vals = []
indexed = df[df['N'] == df['M']]
for index, row in indexed.iterrows():
keys.append(row['name'])
vals.append(row['ratio'])
keys = keys[::len(sizes)]
sns.set(rc={'figure.figsize' : (5.0, len(keys) * 0.5)})
cmap = sns.diverging_palette(10, 120, n=9, as_cmap=True)
np_vals = np.array([vals]).reshape(-1, len(sizes))
g = sns.heatmap(np_vals, annot=True, cmap=cmap, center=1.0, yticklabels=True)
plt.yticks(rotation=0)
plt.title('PyTorch performance divided by NNC performance (single core)')
plt.xlabel('Size of NxN matrix')
plt.ylabel('Operation')
g.set_yticklabels(keys)
g.set_xticklabels(sizes)
plt.savefig('nnc.png')
if __name__ == "__main__":
parser = argparse.ArgumentParser(description='Runs NNC microbenchmarks')
parser.add_argument('--multi_threaded', action='store_true', help='Run with more than one thread')
args = parser.parse_args()
if not args.multi_threaded:
torch.set_num_threads(1)
sizes = [1, 4, 16, 64, 256, 1024]
df = run_benchmarks(benchmarks, [(i, i) for i in sizes])
dump_plot(df, sizes)
|
