File: test_layernorm_nnpi_fp16.py

package info (click to toggle)
pytorch 1.13.1%2Bdfsg-4
  • links: PTS, VCS
  • area: main
  • in suites: bookworm
  • size: 139,252 kB
  • sloc: cpp: 1,100,274; python: 706,454; ansic: 83,052; asm: 7,618; java: 3,273; sh: 2,841; javascript: 612; makefile: 323; xml: 269; ruby: 185; yacc: 144; objc: 68; lex: 44
file content (240 lines) | stat: -rw-r--r-- 8,760 bytes parent folder | download
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
import numpy as np
import caffe2.python.fakelowp.init_shared_libs  # noqa
from caffe2.proto import caffe2_pb2
from caffe2.python import core
from caffe2.python import workspace
from caffe2.python.onnx.onnxifi import onnxifi_caffe2_net
from caffe2.python.fakelowp.test_utils import print_test_debug_info
from hypothesis import given, settings
from hypothesis import strategies as st
import caffe2.python.serialized_test.serialized_test_util as serial
import datetime

core.GlobalInit(["caffe2",
                 "--glow_global_fp16=1",
                 "--glow_global_fused_scale_offset_fp16=1",
                 "--glow_global_force_sls_fp16_accum=1"])

GLOW_LOWERED_BATCHNORM = False


# Test the lowered LayerNorm op
class LayerNorm(serial.SerializedTestCase):

    @given(seed=st.integers(0, 65535),
           batch_size=st.integers(min_value=1, max_value=50),
           size=st.integers(min_value=2, max_value=128),
           epsilon=st.floats(min_value=1e-4, max_value=1e-3),
           elementwise_affine=st.booleans())
    @settings(deadline=datetime.timedelta(seconds=10))
    def test_layernorm(self, seed, batch_size, size, epsilon, elementwise_affine):
        np.random.seed(seed)
        # Reset the workspace
        workspace.ResetWorkspace()
        axis = 1

        dims = np.array(([batch_size, size]))
        X = np.random.uniform(size=dims).astype(np.float32) - 0.5
        gamma = np.random.randn(*X.shape[axis:]).astype(np.float32)
        beta = np.random.randn(*X.shape[axis:]).astype(np.float32)

        pred_net = caffe2_pb2.NetDef()
        pred_net.name = "pred"
        pred_net.external_input.extend(["X", "gamma", "beta"])
        pred_net.external_output.extend(["Y", "mean", "rstd"])
        pred_net.op.add().CopyFrom(
            core.CreateOperator(
                "LayerNorm",
                ["X", "gamma", "beta"] if elementwise_affine else ["X"],
                ["Y", "mean", "rstd"],
                axis=axis,
                epsilon=epsilon,
                elementwise_affine=elementwise_affine
            )
        )

        pred_net_ref = caffe2_pb2.NetDef()
        pred_net_ref.name = "pred_ref"
        pred_net_ref.external_input.extend(["X", "gamma", "beta"])
        pred_net_ref.external_output.extend(["Y", "mean", "rstd"])
        pred_net_ref.op.add().CopyFrom(
            core.CreateOperator(
                "LayerNormFakeFP16NNPI",
                ["X", "gamma", "beta"] if elementwise_affine else ["X"],
                ["Y", "mean", "rstd"],
                axis=axis,
                epsilon=epsilon,
                elementwise_affine=elementwise_affine
            )
        )

        shape_hits = {"X": X.shape, "gamma": gamma.shape, "beta": beta.shape}
        pred_net_onnxified = onnxifi_caffe2_net(
            pred_net,
            shape_hits,
            debug=True,
            adjust_batch=True,
            use_onnx=False
        )
        num_onnxified_ops = sum(
            1 if o.type == "Onnxifi" else 0 for o in pred_net_onnxified.op)
        np.testing.assert_equal(num_onnxified_ops, 1)

        workspace.FeedBlob("X", X)
        workspace.FeedBlob("gamma", gamma)
        workspace.FeedBlob("beta", beta)

        workspace.CreateNet(pred_net_ref)
        workspace.CreateNet(pred_net_onnxified)

        workspace.RunNet(pred_net_ref.name)
        Y_c2 = workspace.FetchBlob("Y")

        dims1 = np.array(([1, *dims]))
        X_glow = X.reshape(dims1)
        workspace.FeedBlob("X", X_glow)

        workspace.RunNet(pred_net_onnxified.name)
        Y_glow = workspace.FetchBlob("Y")

        if not np.allclose(Y_glow, Y_c2):
            diff_Y = np.abs(Y_glow - Y_c2)
            print_test_debug_info(
                "layernorm",
                {
                    "seed": seed,
                    "size": size,
                    "batch_size": batch_size,
                    "epsilon": epsilon,
                    "gamma": gamma,
                    "beta": beta,
                    "elementwise_affine": elementwise_affine,
                    "X": X,
                    "Y_glow": Y_glow,
                    "Y_c2": Y_c2,
                    "diff_Y": diff_Y,
                }
            )
            assert(0)

    def _get_scale_zp(self, tensor):
        tensor_max = np.max(tensor)
        tensor_min = min(0, np.min(tensor))
        scale = np.float32(np.float16((tensor_max - tensor_min) / 255.0))
        if scale < 1e-6:
            scale = np.float32(1e-6)
        zero_point = 0 - tensor_min / scale
        zero_point = int(round(np.clip(zero_point, 0, 255.0)))
        return (scale, zero_point)

    def _layernorm_transform(self, X):
        mean = np.mean(X, axis=1)
        mean_exp = np.outer(mean, np.ones(X.shape[1]))
        std = np.std(X, axis=1)
        std_exp = np.outer(std, np.ones(X.shape[1]))
        Y = (X - mean_exp) / std_exp
        return Y

    @given(seed=st.integers(0, 65535),
           batch_size=st.integers(min_value=1, max_value=50),
           size=st.integers(min_value=2, max_value=128),
           epsilon=st.floats(min_value=1e-4, max_value=1e-3),
           elementwise_affine=st.booleans())
    @settings(deadline=datetime.timedelta(seconds=10))
    # re-enable when T74553975 gets fixed
    def test_fused_ln_quantize(self, seed, batch_size, size, epsilon, elementwise_affine):
        np.random.seed(seed)

        # Reset the workspace
        workspace.ResetWorkspace()
        axis = 1

        dims = np.array(([batch_size, size]))
        X = np.random.uniform(size=dims).astype(np.float32) - 0.5
        gamma = np.random.randn(*X.shape[axis:]).astype(np.float32)
        beta = np.random.randn(*X.shape[axis:]).astype(np.float32)

        Y = self._layernorm_transform(X)
        scale, zp = self._get_scale_zp(Y)

        pred_net = caffe2_pb2.NetDef()
        pred_net.name = "pred"
        pred_net.external_input.extend(["X", "gamma", "beta"])
        pred_net.external_output.extend(["Y_q"])
        pred_net.op.add().CopyFrom(
            core.CreateOperator(
                "LayerNorm",
                ["X", "gamma", "beta"] if elementwise_affine else ["X"],
                ["Y", "mean", "rstd"],
                axis=axis,
                epsilon=epsilon,
                elementwise_affine=elementwise_affine
            )
        )
        pred_net.op.add().CopyFrom(
            core.CreateOperator(
                "Int8Quantize", ["Y"], ["Y_q"], Y_scale=scale, Y_zero_point=zp
            )
        )

        print(pred_net)
        pred_net_ref = caffe2_pb2.NetDef()
        pred_net_ref.name = "pred_ref"
        pred_net_ref.external_input.extend(["X", "gamma", "beta"])
        pred_net_ref.external_output.extend(["Y_q"])
        pred_net_ref.op.add().CopyFrom(
            core.CreateOperator(
                "LayerNormInt8QuantizeFakeNNPI",
                ["X", "gamma", "beta"] if elementwise_affine else ["X"],
                ["Y_q", "mean", "rstd"],
                axis=axis,
                epsilon=epsilon,
                elementwise_affine=elementwise_affine,
                Y_scale=scale, Y_zero_point=zp
            )
        )
        shape_hits = {"X": X.shape, "gamma": gamma.shape, "beta": beta.shape}
        pred_net_onnxified = onnxifi_caffe2_net(
            pred_net,
            shape_hits,
            debug=True,
            adjust_batch=True,
            use_onnx=False
        )
        num_onnxified_ops = sum(
            1 if o.type == "Onnxifi" else 0 for o in pred_net_onnxified.op)
        np.testing.assert_equal(num_onnxified_ops, 1)

        workspace.FeedBlob("X", X)
        workspace.FeedBlob("gamma", gamma)
        workspace.FeedBlob("beta", beta)

        workspace.CreateNet(pred_net_ref)
        workspace.CreateNet(pred_net_onnxified)

        workspace.RunNet(pred_net_ref.name)
        Y_c2 = workspace.FetchInt8Blob("Y_q")

        workspace.RunNet(pred_net_onnxified.name)
        Y_glow = workspace.FetchInt8Blob("Y_q")

        if not np.allclose(Y_glow.data, Y_c2.data) or \
           Y_glow.scale != Y_c2.scale or Y_glow.zero_point != Y_c2.zero_point:
            diff_Y = np.abs(Y_glow.data.astype(np.float32) - Y_c2.data.astype(np.float32))
            print_test_debug_info(
                "layernorm",
                {
                    "seed": seed,
                    "size": size,
                    "batch_size": batch_size,
                    "epsilon": epsilon,
                    "gamma": gamma,
                    "beta": beta,
                    "elementwise_affine": elementwise_affine,
                    "X": X,
                    "Y_glow": Y_glow,
                    "Y_c2": Y_c2,
                    "diff_Y": diff_Y,
                }
            )
            assert(0)