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
|
#include "caffe2/core/common.h"
#include "caffe2/core/test_utils.h"
#include "caffe2/opt/converter.h"
#include "caffe2/opt/custom/concat_elim.h"
#include "caffe2/predictor/emulator/data_filler.h"
#include "caffe2/utils/proto_utils.h"
#include <gtest/gtest.h>
using namespace caffe2::testing;
using namespace caffe2::emulator;
TEST(gatherFuse8BitRowwiseQuantFloatMulLengthsSumElim, Basic) {
using namespace caffe2;
caffe2::NetDef net;
NetMutator(&net)
.newOp("Gather", {"Data0", "Idx"}, {"Gout"})
.newOp("Fused8BitRowwiseQuantizedToFloat", {"Gout"}, {"Fout"})
.newOp("Mul", {"Fout", "Min"}, {"Mout"})
.addArgument("axis", 0)
.addArgument("broadcast", 1)
.newOp("LengthsSum", {"Mout", "Len"}, {"Out"});
auto nn = caffe2::convertToNNModule(net);
caffe2::opt::gatherFuse8BitRowwiseQuantFloatMulLengthsSumElim(&nn);
auto optimized_net = caffe2::convertToCaffe2Proto(nn, net);
EXPECT_EQ(optimized_net.op().size(), 1);
// Fill in a float typed dummy tensor first
Workspace workspace;
std::vector<int64_t> lut_dims{20, 20};
TensorFiller lut_filler(lut_dims);
const auto dummy = "dummy";
Blob* blob = workspace.CreateBlob(dummy);
::fill_with_type(lut_filler, "float", BlobGetMutableTensor(blob, CPU));
CAFFE_ENFORCE(workspace.GetBlob(dummy)->GetRaw());
// Run the FloatToFused8BitRowwiseQuantized operator to do proper
// quantization
OperatorDef op_def = CreateOperatorDef(
"FloatToFused8BitRowwiseQuantized", "", {dummy}, {"Data0"}, {});
workspace.RunOperatorOnce(op_def);
auto* blob2 = workspace.GetBlob("Data0");
CAFFE_ENFORCE(blob2->GetRaw());
// Fill in the rest of the inputs
blob = workspace.CreateBlob("Idx");
auto* t = BlobGetMutableTensor(blob, CPU);
ReinitializeTensor(t, {5}, at::dtype<int32_t>().device(CPU));
int32_t* data = t->mutable_data<int32_t>();
for (int i = 0; i < 5; ++i) {
data[i] = i;
}
blob = workspace.CreateBlob("Min");
t = BlobGetMutableTensor(blob, CPU);
ReinitializeTensor(t, {5}, at::dtype<float>().device(CPU));
float* fdata = t->mutable_data<float>();
for (int i = 0; i < 5; ++i) {
fdata[i] = 0.9;
}
blob = workspace.CreateBlob("Len");
t = BlobGetMutableTensor(blob, CPU);
ReinitializeTensor(t, {2}, at::dtype<int32_t>().device(CPU));
data = t->mutable_data<int32_t>();
data[0] = 2;
data[1] = 3;
workspace.RunNetOnce(net);
auto outBefore = getTensor(workspace, "Out").Clone();
workspace.RemoveBlob("out");
workspace.RunNetOnce(optimized_net);
auto outAfter = getTensor(workspace, "Out").Clone();
assertTensorEquals(outBefore, outAfter);
}
TEST(gatherFuse8BitRowwiseQuantFloatMulLengthsSumElim, NoFuse) {
using namespace caffe2;
caffe2::NetDef net;
NetMutator(&net)
.newOp("Gather", {"Data0", "Idx"}, {"Gout"})
.newOp("Fused8BitRowwiseQuantizedToFloat", {"Gout"}, {"Fout"})
.newOp("Mul", {"Fout", "Min"}, {"Mout"})
.addArgument("axis", 0)
.addArgument("broadcast", 1)
.newOp("LengthsSum", {"Mout", "Len"}, {"Out"})
.newOp("Copy", {"Fout"}, {"Fout2"});
auto nn = caffe2::convertToNNModule(net);
caffe2::opt::gatherFuse8BitRowwiseQuantFloatMulLengthsSumElim(&nn);
auto optimized_net = caffe2::convertToCaffe2Proto(nn, net);
EXPECT_EQ(optimized_net.op().size(), 5);
}
TEST(ConcatElim, BasicNet) {
caffe2::NetDef net;
NetMutator(&net)
.newOp("Concat", {"X0", "X1", "X2"}, {"concat_out", "split_info"})
.addArgument("axis", 1)
.addArgument("add_axis", 1)
.newOp("BatchMatMul", {"concat_out", "concat_out"}, {"matmul_out"})
.addArgument("trans_a", 0)
.addArgument("trans_b", 1)
.addArgument("broadcast", 0)
.newOp("Flatten", {"matmul_out"}, {"flatten_out"})
.newOp("BatchGather", {"flatten_out", "indices"}, {"out"});
auto nn = caffe2::convertToNNModule(net);
caffe2::opt::concatElim(&nn);
auto optimized_net = caffe2::convertToCaffe2Proto(nn, net);
EXPECT_EQ(optimized_net.op().size(), 1);
std::vector<int64_t> input_dim = {30, 20};
std::vector<std::vector<std::vector<int64_t>>> input_dims = {
{/* X0 */ input_dim, /* X1 */ input_dim, /* X2 */ input_dim},
{/* indices */ {3}}};
std::vector<std::vector<std::string>> input_types = {
{"float", "float", "float"}, {"int"}};
auto filler = TestDataRandomFiller(net, input_dims, input_types);
caffe2::Workspace workspace;
filler.fillInputToWorkspace(&workspace);
workspace.RunNetOnce(net);
auto outBefore = getTensor(workspace, "out").Clone();
workspace.RunNetOnce(optimized_net);
auto outAfter = getTensor(workspace, "out_cc_bmm_bg").Clone();
assertTensorEquals(outBefore, outAfter);
}
TEST(ConcatElim, ProdNet) {
// Test concatElim on a realistic prod model.
caffe2::NetDef net;
ReadProtoFromFile("caffe2/caffe2/opt/custom/concat_elim_test_net.pb", &net);
EXPECT_EQ(net.op().size(), 176);
auto nn = caffe2::convertToNNModule(net);
caffe2::opt::concatElim(&nn);
auto optimized_net = caffe2::convertToCaffe2Proto(nn, net);
EXPECT_EQ(optimized_net.op().size(), 173);
}
TEST(ConcatAddMulNaNClipElim, BasicNet) {
caffe2::NetDef net;
NetMutator(&net)
.newOp("Concat", {"X0", "X1", "X2"}, {"concat_out", "split_info"})
.addArgument("axis", 1)
.newOp("Add", {"concat_out", "add_in"}, {"add_out"})
.addArgument("broadcast", 1)
.newOp("Mul", {"add_out", "mul_in"}, {"mul_out"})
.addArgument("broadcast", 1)
.newOp("ReplaceNaN", {"mul_out"}, {"replace_out"})
.addArgument("value", 0.0001f)
.newOp("Clip", {"replace_out"}, {"out"});
auto nn = caffe2::convertToNNModule(net);
caffe2::opt::concatAddMulNaNClipElim(&nn);
auto optimized_net = caffe2::convertToCaffe2Proto(nn, net);
EXPECT_EQ(optimized_net.op().size(), 1);
std::vector<int64_t> input_dim = {30, 20};
std::vector<std::vector<std::vector<int64_t>>> input_dims = {
{/* X0 */ input_dim, /* X1 */ input_dim, /* X2 */ input_dim},
{/* add_in */ {60}},
{/* mul_in */ {60}}};
std::vector<std::vector<std::string>> input_types = {
{"float", "float", "float"}, {"float"}, {"float"}};
auto filler = TestDataRandomFiller(net, input_dims, input_types);
caffe2::Workspace workspace;
filler.fillInputToWorkspace(&workspace);
workspace.RunNetOnce(net);
auto outBefore = getTensor(workspace, "out").Clone();
workspace.RemoveBlob("out");
workspace.RunNetOnce(optimized_net);
auto outAfter = getTensor(workspace, "out").Clone();
assertTensorEquals(outBefore, outAfter);
}
TEST(ConcatAddMulNaNClipElim, ProdNet) {
// Test ConcatAddMulNaNClipElim on a realistic prod model.
caffe2::NetDef net;
ReadProtoFromFile("caffe2/caffe2/opt/custom/test_cc_amcr_net.pb", &net);
auto size = net.op().size();
auto nn = caffe2::convertToNNModule(net);
caffe2::opt::concatAddMulNaNClipElim(&nn);
auto optimized_net = caffe2::convertToCaffe2Proto(nn, net);
// Ensure that optimization happens: number of ops is smaller than before.
EXPECT_LT(optimized_net.op().size(), size);
}
|
