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
|
/**
* Copyright (c) 2016-present, Facebook, Inc.
*
* Licensed under the Apache License, Version 2.0 (the "License");
* you may not use this file except in compliance with the License.
* You may obtain a copy of the License at
*
* http://www.apache.org/licenses/LICENSE-2.0
*
* Unless required by applicable law or agreed to in writing, software
* distributed under the License is distributed on an "AS IS" BASIS,
* WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
* See the License for the specific language governing permissions and
* limitations under the License.
*/
#pragma once
#include <fbgemm/FbgemmFP16.h>
#include "caffe2/core/context.h"
#include "caffe2/core/operator.h"
#include "caffe2/operators/fully_connected_op.h"
#include "caffe2/utils/conversions.h"
#include "caffe2/utils/math.h"
namespace caffe2 {
/**
* C2 wrapper for fp16 gemm
*
* Suppose your predict_net has an FC operator in fp32 as follows:
* op {
* input: "x"
* input: "w"
* input: "b"
* output: "y"
* type: "FC"
* }
* ...
* external_input: "w"
*
* To use FbFCPacked operator with fp16 fbgemm, in init_net
* ... # an operator that generates w
* op {
* input: "w"
* output: "w_packed"
* type: "FbGemmPack"
* }
* ...
* external_output: "w_packed"
*
* in predict_net:
* op {
* input: "x"
* input: "w_packed"
* input: "b"
* output: "y"
* type: "FbFCPacked"
* }
* ...
* external_input: "w_packed"
*/
template <
class Context,
class Engine = DefaultEngine,
typename T_W = fbgemm::float16>
class FbFCPackedOperator final : public Operator<Context> {
public:
USE_OPERATOR_CONTEXT_FUNCTIONS;
FbFCPackedOperator(const OperatorDef& operator_def, Workspace* ws)
: Operator<Context>(operator_def, ws),
axis_(this->template GetSingleArgument<int32_t>("axis", 1)),
axis_w_(this->template GetSingleArgument<int32_t>("axis_w", 1)) {}
~FbFCPackedOperator() {}
// template on X, B, and Y.
template <typename T_X, typename T_B, typename T_Y>
bool DoRunWithType() {
const auto& X = Input(0);
const auto& b = Input(2);
CAFFE_ENFORCE(b.dim() == 1, b.dim());
// batch size
const auto canonical_axis = X.canonical_axis_index(axis_);
const int M = X.size_to_dim(canonical_axis);
const int N = b.numel();
// Load the packed matrix
auto* W =
OperatorBase::Input<caffe2::unique_ptr<fbgemm::PackedGemmMatrixFP16>>(1)
.get();
const int K = W->numRows();
if (!W->packed()) {
if (!packed_w_) {
std::vector<float> src_mat(W->matSize());
for (int i = 0; i < W->matSize(); ++i) {
src_mat[i] =
fbgemm::cpu_half2float(W->pmat()[i]);
}
packed_w_ = std::make_unique<fbgemm::PackedGemmMatrixFP16>(
fbgemm::matrix_op_t::Transpose,
W->numRows(), W->numCols(),
1.0,
src_mat.data());
}
W = packed_w_.get();
}
auto dimErrorString = [&]() {
return c10::str(
"Dimension mismatch: ",
"X: ",
X.sizes(),
", W: ",
std::vector<int>({K, W->numCols()}),
", b: ",
b.sizes(),
", axis: ",
axis_,
", M: ",
M,
", N: ",
N,
", K: ",
K);
};
// Error checking
CAFFE_ENFORCE(M == X.numel() / K, dimErrorString());
CAFFE_ENFORCE(K == X.size_from_dim(canonical_axis), dimErrorString());
CAFFE_ENFORCE(N == W->numCols(), dimErrorString());
Y_shape_cache_ = X.sizes().vec();
// This is an invariant of canonical_axis, so we can DCHECK.
TORCH_DCHECK_LE(canonical_axis + 1, Y_shape_cache_.size());
Y_shape_cache_.resize(canonical_axis + 1);
Y_shape_cache_[canonical_axis] = N;
auto* Y = Output(0, Y_shape_cache_, at::dtype<T_Y>());
if (X.numel() == 0) {
// skip the rest of the computation if X is empty
Y->template mutable_data<T_Y>();
return true;
}
// Call the fp16 gemm interface
fbgemm::cblas_gemm_compute(
fbgemm::matrix_op_t::NoTranspose,
M,
X.template data<T_X>(),
*W,
0.f,
Y->template mutable_data<T_Y>());
// Add bias term, accumulation is still in fp32.
TensorProto::DataType math_type = TensorProto_DataType_FLOAT;
if (bias_multiplier_.numel() != M) {
// If the helper bias multiplier is not M, reshape and fill it with one.
bias_multiplier_.Resize(M);
math::Set<T_B, Context>(
M,
convert::To<float, T_B>(1),
bias_multiplier_.template mutable_data<T_B>(),
&context_);
}
math::Gemm<T_B, Context, Engine>(
CblasNoTrans,
CblasNoTrans,
M,
N,
1,
1,
bias_multiplier_.template data<T_B>(),
b.template data<T_B>(),
1,
Y->template mutable_data<T_Y>(),
&context_,
math_type);
return true;
}
bool RunOnDevice() override {
return DoRunWithType<
float, // X
float, // B
float>(); // Y
}
protected:
size_t axis_{1};
size_t axis_w_{1};
// A local vector to cache the output shape so we don't need to recreate
// a vector object every time we run Run().
vector<int64_t> Y_shape_cache_;
Tensor bias_multiplier_{Context::GetDeviceType()};
caffe2::unique_ptr<fbgemm::PackedGemmMatrixFP16> packed_w_{nullptr};
};
class PackedGemmMatrixFP16ShapeFunctions : public ExternalTensorFunctionsBase {
public:
explicit PackedGemmMatrixFP16ShapeFunctions()
: ExternalTensorFunctionsBase() {}
~PackedGemmMatrixFP16ShapeFunctions() override {}
bool isQuantized() const override {
return false;
}
bool IsSameMetaType(TypeIdentifier id) override;
void SetupExternalTensorDescriptor(
const Blob* blob,
std::vector<std::vector<uint64_t>>* shapes,
std::vector<std::vector<float>>* all_scales,
std::vector<std::vector<int32_t>>* all_offsets,
ExternalTensorDescriptor* desc) override;
void LoadInfoOfBlob(
const Blob* /* unused */,
std::vector<float>* /* unused */,
std::vector<float>* /* unused */,
uint32_t* /* unused */) override {}
TypeIdentifier GetTypeMetaId() override;
TypeMeta GetExternalTensorType(const void* c) override;
vector<int64_t> GetExternalTensorInfo(
const void* c,
size_t* capacity,
DeviceOption* device) override;
};
} // namespace caffe2
|
