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
|
#include <torch/csrc/jit/tensorexpr/operators/reduction.h>
using namespace torch::jit::tensorexpr;
// Remove all indices from axes positions.
static std::vector<VarHandle> squeezeIndices(
const ParameterList& indices,
const std::vector<size_t>& axes) {
std::vector<VarHandle> indices_squeezed;
for (size_t dim = 0; dim < indices.size(); ++dim) {
if (!std::count(axes.begin(), axes.end(), dim)) {
indices_squeezed.push_back(indices[dim]);
}
}
return indices_squeezed;
}
namespace torch::jit::tensorexpr {
Tensor computeSum(
const std::vector<ArgValue>& inputs,
const std::vector<ExprHandle>& outputShape,
const std::vector<ExprHandle>& outputStrides,
const std::optional<ScalarType>& outputType,
at::Device device) {
std::vector<size_t> axes;
bool keepdim = false;
// aten::sum takes the input tensor named self.
auto sizes = valueShape(inputs[0]);
size_t rank = sizes.size();
if (inputs.size() > 2) {
if (auto emptyAxes = std::get_if<BufList>(&inputs[1])) {
// If dim-array is an empty list, it will appear as BufList instead of
// IntList, and hence we need a special handling for it.
// In that case, we need to sum over all axes.
TORCH_INTERNAL_ASSERT(emptyAxes->empty());
axes.resize(rank);
std::iota(axes.begin(), axes.end(), 0);
} else if (rank > 0) {
auto const& nodeAxes = std::get<IntList>(inputs[1]);
// Canonicalize axes: wrap around, sort and make unique.
for (auto axis : nodeAxes) {
axes.push_back(at::maybe_wrap_dim(axis, static_cast<int64_t>(rank)));
}
std::sort(axes.begin(), axes.end());
axes.erase(std::unique(axes.begin(), axes.end()), axes.end());
}
keepdim = std::get<bool>(inputs[2]);
} else {
axes.resize(rank);
std::iota(axes.begin(), axes.end(), 0);
}
// Axes go into reduction dimensions.
std::vector<ExprHandle> reductionDims;
reductionDims.reserve(rank);
for (size_t axis : axes) {
reductionDims.emplace_back(sizes[axis]);
}
std::vector<ExprHandle> outputDims;
// Output dimensions are the complement of axes. When keepdim is set, a
// one-sized dimension is inserted for each axis.
for (size_t dim = 0; dim < rank; ++dim) {
if (!std::count(axes.begin(), axes.end(), dim)) {
outputDims.emplace_back(sizes[dim]);
} else if (keepdim) {
outputDims.emplace_back(1);
}
}
return Reduce(
"sum",
outputDims,
outputStrides,
Sum(),
[&](ParameterList& indices) {
// "Squeeze" out indices inserted when keepdim is set.
auto indices_squeezed =
keepdim ? squeezeIndices(indices, axes) : indices;
TORCH_INTERNAL_ASSERT(axes.size() <= indices_squeezed.size());
// Move innermost indices into axes positions:
// 1. Fill the outermost indices first.
// 2. Insert the innermost indices into the correct axis position,
// displacing the outermost indices as needed.
std::vector<ExprHandle> indices_exprs;
size_t i = 0;
for (; i < indices_squeezed.size() - axes.size(); ++i) {
indices_exprs.push_back(indices_squeezed[i]);
}
for (auto axis : axes) {
indices_exprs.insert(
indices_exprs.begin() + static_cast<std::ptrdiff_t>(axis),
indices_squeezed[i]);
++i;
}
auto indexed = tensorOrConstant(inputs[0], indices_exprs);
if (outputType) {
return Cast::make(ToDtype(*outputType), indexed);
} else {
return indexed;
}
},
reductionDims);
}
Tensor computeMean(
const std::vector<ArgValue>& inputs,
const std::vector<ExprHandle>& outputShape,
const std::vector<ExprHandle>& outputStrides,
const std::optional<ScalarType>& outputType,
at::Device device) {
Dtype dtype = kFloat;
if (outputType) {
dtype = Dtype(*outputType);
}
bool keepdim = false;
BufHandle ResultBuf("mean", outputShape, dtype);
auto const& InputBuf = std::get<BufHandle>(inputs[0]);
std::vector<ExprHandle> extra_args;
if (inputs.size() > 2) {
keepdim = std::get<bool>(inputs[2]);
}
if (auto mean_dims = std::get_if<IntList>(&inputs[1])) {
extra_args = c10::fmap<ExprHandle>(*mean_dims);
} else {
// When dims argument is not specified, reduce over all dimensions
for (int64_t idx = 0; idx < static_cast<int64_t>(InputBuf.ndim()); ++idx) {
extra_args.emplace_back(idx);
}
}
extra_args.push_back(LongImm::make(static_cast<int64_t>(keepdim)));
return Tensor(
ResultBuf.node(),
ExternalCall::make(ResultBuf, "nnc_aten_mean", {InputBuf}, extra_args));
}
Tensor computeMax(
const std::vector<ArgValue>& inputs,
const std::vector<ExprHandle>& outputShape,
const std::vector<ExprHandle>& outputStrides,
const std::optional<ScalarType>& outputType,
at::Device device) {
Dtype dtype = kFloat;
if (outputType) {
dtype = Dtype(*outputType);
}
BufHandle ResultBuf("max", outputShape, dtype);
auto const& InputBuf = std::get<BufHandle>(inputs[0]);
auto max_dim = std::get<int64_t>(inputs[1]);
auto keep_dim = std::get<bool>(inputs[2]);
return Tensor(
ResultBuf.node(),
ExternalCall::make(
ResultBuf,
"nnc_aten_max_red",
{InputBuf},
{max_dim, (int64_t)keep_dim}));
}
Tensor computeAdaptiveAvgPool2d(
const std::vector<ArgValue>& inputs,
const std::vector<ExprHandle>& outputShape,
const std::vector<ExprHandle>& outputStrides,
const std::optional<ScalarType>& outputType,
at::Device device) {
Dtype dtype = kFloat;
if (outputType) {
dtype = Dtype(*outputType);
}
BufHandle ResultBuf("adaptive_avgpool2d", outputShape, dtype);
auto const& out_size_param = std::get<IntList>(inputs[1]);
return Tensor(
ResultBuf.node(),
ExternalCall::make(
ResultBuf,
"nnc_aten_adaptive_avg_pool2d",
{std::get<BufHandle>(inputs[0])},
c10::fmap<ExprHandle>(out_size_param)));
}
} // namespace torch::jit::tensorexpr
|
