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
258
259
260
261
262
263
264
265
266
267
268
269
270
271
272
273
274
275
276
277
278
279
280
281
282
283
284
285
286
287
288
289
290
291
292
293
294
295
296
297
298
299
300
301
302
303
304
305
306
307
308
309
310
311
312
313
314
315
316
317
318
319
320
321
322
323
324
325
326
327
328
329
330
331
332
333
334
335
336
337
338
339
340
341
342
343
344
345
346
347
348
349
350
351
352
353
354
355
356
357
358
359
360
361
362
363
364
365
366
367
368
369
370
|
#include "caffe2/operators/resize_op.h"
#include "caffe2/utils/cpu_neon.h"
#include "caffe2/utils/math.h"
#ifdef USE_MKLDNN
#include "caffe2/ideep/operators/operator_fallback_ideep.h"
#include "caffe2/ideep/utils/ideep_operator.h"
#endif
namespace caffe2 {
void resizeNearestNCHW2x(
int batch_size,
int num_channels,
int input_height,
int input_width,
const float* input,
float* output) {
const int output_height = input_height * 2;
const int output_width = input_width * 2;
for (int n = 0; n < batch_size; ++n) {
for (int c = 0; c < num_channels; ++c) {
for (int y = 0; y < output_height; ++y) {
const int in_y = y / 2;
#if defined(__ARM_NEON__) || defined(__ARM_NEON)
int vecW = (input_width / 4) * 4; // round down
int x = 0;
for (; x < vecW; x += 4) {
// load 0 1 2 3
float32x4_t v = vld1q_f32(input + in_y * input_width + x);
const int oidx = output_width * y + x * 2;
float32x4x2_t v2 = {{v, v}};
// store 00 11 22 33
vst2q_f32(output + oidx + 0, v2);
}
// handle remainder
for (; x < input_width; ++x) {
const float v = input[in_y * input_width + x];
const int oidx = output_width * y + x * 2;
output[oidx + 0] = v;
output[oidx + 1] = v;
}
#else
for (int x = 0; x < input_width; ++x) {
const float v = input[in_y * input_width + x];
const int oidx = output_width * y + x * 2;
output[oidx + 0] = v;
output[oidx + 1] = v;
}
#endif
}
input += input_height * input_width;
output += output_height * output_width;
}
}
}
template <>
bool ResizeNearestOp<float, CPUContext>::RunOnDeviceWithOrderNCHW() {
const auto& X = Input(0);
const int batch_size = X.dim32(0), num_channels = X.dim32(1),
input_height = X.dim32(2), input_width = X.dim32(3);
if (InputSize() == 2) {
const auto& scales = Input(1);
CAFFE_ENFORCE_EQ(scales.dim(), 1);
CAFFE_ENFORCE_EQ(scales.numel(), 2);
const float* scales_data = scales.data<float>();
height_scale_ = scales_data[0];
width_scale_ = scales_data[1];
}
// NOLINTNEXTLINE(bugprone-narrowing-conversions,cppcoreguidelines-narrowing-conversions)
int output_width = input_width * width_scale_;
// NOLINTNEXTLINE(bugprone-narrowing-conversions,cppcoreguidelines-narrowing-conversions)
int output_height = input_height * height_scale_;
auto* Y = Output(
0,
{batch_size, num_channels, output_height, output_width},
at::dtype<float>());
const float* Xdata = X.data<float>();
float* Ydata = Y->template mutable_data<float>();
// Specialized implementation for fast 2x upsampling
if (width_scale_ == 2.0 && height_scale_ == 2.0) {
resizeNearestNCHW2x(
batch_size, num_channels, input_height, input_width, Xdata, Ydata);
return true;
}
for (int n = 0; n < batch_size; ++n) {
for (int c = 0; c < num_channels; ++c) {
for (int y = 0; y < output_height; ++y) {
// NOLINTNEXTLINE(bugprone-narrowing-conversions,cppcoreguidelines-narrowing-conversions)
const int in_y = std::min((int)(y / height_scale_), (input_height - 1));
for (int x = 0; x < output_width; ++x) {
// NOLINTNEXTLINE(bugprone-narrowing-conversions,cppcoreguidelines-narrowing-conversions)
const int in_x = std::min((int)(x / width_scale_), (input_width - 1));
Ydata[output_width * y + x] = Xdata[input_width * in_y + in_x];
}
}
Xdata += input_height * input_width;
Ydata += output_width * output_height;
}
}
return true;
}
template <>
bool ResizeNearestOp<float, CPUContext>::RunOnDeviceWithOrderNHWC() {
const auto& X = Input(0);
const int batch_size = X.dim32(0), input_height = X.dim32(1),
input_width = X.dim32(2), num_channels = X.dim32(3);
if (InputSize() == 2) {
const auto& scales = Input(1);
CAFFE_ENFORCE_EQ(scales.dim(), 1);
CAFFE_ENFORCE_EQ(scales.numel(), 2);
const float* scales_data = scales.data<float>();
height_scale_ = scales_data[0];
width_scale_ = scales_data[1];
}
// NOLINTNEXTLINE(bugprone-narrowing-conversions,cppcoreguidelines-narrowing-conversions)
int output_width = input_width * width_scale_;
// NOLINTNEXTLINE(bugprone-narrowing-conversions,cppcoreguidelines-narrowing-conversions)
int output_height = input_height * height_scale_;
const int output_width_stride = output_width * num_channels;
const int input_width_stride = input_width * num_channels;
auto* Y = Output(
0,
{batch_size, output_height, output_width, num_channels},
at::dtype<float>());
const float* Xdata = X.data<float>();
float* Ydata = Y->template mutable_data<float>();
for (int n = 0; n < batch_size; ++n) {
for (int y = 0; y < output_height; ++y) {
// NOLINTNEXTLINE(bugprone-narrowing-conversions,cppcoreguidelines-narrowing-conversions)
const int in_y = std::min((int)(y / height_scale_), (input_height - 1));
for (int x = 0; x < output_width; ++x) {
// NOLINTNEXTLINE(bugprone-narrowing-conversions,cppcoreguidelines-narrowing-conversions)
const int in_x = std::min((int)(x / width_scale_), (input_width - 1));
std::memcpy(
&Ydata[output_width_stride * y + num_channels * x],
&Xdata[input_width_stride * in_y + num_channels * in_x],
num_channels * sizeof(float));
}
}
Xdata += input_height * input_width_stride;
Ydata += output_height * output_width_stride;
}
return true;
}
template <>
bool ResizeNearestOp<float, CPUContext>::RunOnDevice() {
switch (order_) {
case StorageOrder::NHWC:
return RunOnDeviceWithOrderNHWC();
case StorageOrder::NCHW:
return RunOnDeviceWithOrderNCHW();
default:
CAFFE_THROW("Unknown Storage order: ", order_);
}
}
template <>
bool ResizeNearestGradientOp<float, CPUContext>::RunOnDeviceWithOrderNCHW() {
const auto& dY = Input(0);
const auto& X = Input(1);
const auto inputDims = dY.sizes();
CAFFE_ENFORCE_EQ(4, inputDims.size());
const int batch_size = dY.dim32(0), num_channels = dY.dim32(1),
input_height = dY.dim32(2), input_width = dY.dim32(3);
const int output_height = X.dim32(2);
const int output_width = X.dim32(3);
if (InputSize() == 3) {
const auto& scales = Input(2);
CAFFE_ENFORCE_EQ(scales.dim(), 1);
CAFFE_ENFORCE_EQ(scales.numel(), 2);
const float* scales_data = scales.data<float>();
height_scale_ = scales_data[0];
width_scale_ = scales_data[1];
}
auto* dX = Output(
0,
{batch_size, num_channels, output_height, output_width},
at::dtype<float>());
math::Set<float, CPUContext>(
dX->numel(), 0.0f, dX->template mutable_data<float>(), &context_);
const float* dYdata = dY.data<float>();
float* dXdata = dX->template mutable_data<float>();
for (int n = 0; n < batch_size; ++n) {
for (int c = 0; c < num_channels; ++c) {
for (int y = 0; y < input_height; ++y) {
const int out_y =
// NOLINTNEXTLINE(bugprone-narrowing-conversions,cppcoreguidelines-narrowing-conversions)
std::min((int)(y / height_scale_), (output_height - 1));
for (int x = 0; x < input_width; ++x) {
const int out_x =
// NOLINTNEXTLINE(bugprone-narrowing-conversions,cppcoreguidelines-narrowing-conversions)
std::min((int)(x / width_scale_), (output_width - 1));
dXdata[output_width * out_y + out_x] += dYdata[input_width * y + x];
}
}
dYdata += input_height * input_width;
dXdata += output_height * output_width;
}
}
return true;
}
template <>
bool ResizeNearestGradientOp<float, CPUContext>::RunOnDeviceWithOrderNHWC() {
const auto& dY = Input(0);
const auto& X = Input(1);
const auto inputDims = dY.sizes();
CAFFE_ENFORCE_EQ(4, inputDims.size());
const int batch_size = dY.dim32(0), input_height = dY.dim32(1),
input_width = dY.dim32(2), num_channels = dY.dim32(3);
const int output_height = X.dim32(1);
const int output_width = X.dim32(2);
if (InputSize() == 3) {
const auto& scales = Input(2);
CAFFE_ENFORCE_EQ(scales.dim(), 1);
CAFFE_ENFORCE_EQ(scales.numel(), 2);
const float* scales_data = scales.data<float>();
height_scale_ = scales_data[0];
width_scale_ = scales_data[1];
}
auto* dX = Output(
0,
{batch_size, output_height, output_width, num_channels},
at::dtype<float>());
math::Set<float, CPUContext>(
dX->numel(), 0.0f, dX->template mutable_data<float>(), &context_);
const int output_width_stride = output_width * num_channels;
const int input_width_stride = input_width * num_channels;
const float* dYdata = dY.data<float>();
float* dXdata = dX->template mutable_data<float>();
for (int n = 0; n < batch_size; ++n) {
for (int y = 0; y < input_height; ++y) {
// NOLINTNEXTLINE(bugprone-narrowing-conversions,cppcoreguidelines-narrowing-conversions)
const int out_y = std::min((int)(y / height_scale_), (output_height - 1));
for (int x = 0; x < input_width; ++x) {
// NOLINTNEXTLINE(bugprone-narrowing-conversions,cppcoreguidelines-narrowing-conversions)
const int out_x = std::min((int)(x / width_scale_), (output_width - 1));
float* dXdata_c0 =
dXdata + output_width_stride * out_y + num_channels * out_x;
const float* dYdata_c0 =
dYdata + input_width_stride * y + num_channels * x;
for (int c = 0; c < num_channels; ++c) {
dXdata_c0[c] += dYdata_c0[c];
}
}
}
dYdata += input_height * input_width_stride;
dXdata += output_height * output_width_stride;
}
return true;
}
template <>
bool ResizeNearestGradientOp<float, CPUContext>::RunOnDevice() {
switch (order_) {
case StorageOrder::NHWC:
return RunOnDeviceWithOrderNHWC();
case StorageOrder::NCHW:
return RunOnDeviceWithOrderNCHW();
default:
CAFFE_THROW("Unknown Storage order: ", order_);
}
}
REGISTER_CPU_OPERATOR(ResizeNearest, ResizeNearestOp<float, CPUContext>);
REGISTER_CPU_GRADIENT_OPERATOR(
ResizeNearestGradient,
ResizeNearestGradientOp<float, CPUContext>);
#ifdef USE_MKLDNN
REGISTER_IDEEP_OPERATOR(
ResizeNearest,
IDEEPFallbackOp<ResizeNearestOp<float, CPUContext>>);
#endif
// Input: X, output: Y
OPERATOR_SCHEMA(ResizeNearest)
.NumInputs(1, 2)
.NumOutputs(1)
.Arg("width_scale", "Scale along width dimension")
.Arg("height_scale", "Scale along height dimension")
.SetDoc(R"DOC(
Resizes the spatial dimensions of the input using nearest neighbor
interpolation. The `width_scale` and `height_scale` arguments
control the size of the output, which is given by:
output_width = floor(input_width * width_scale)
output_height = floor(output_height * height_scale)
)DOC")
.Input(0, "X", "Input tensor")
.Input(
1,
"scales", // the hack to support onnx spec
"1D, 2-element, Scales tensor, [height_scale, width_scale]")
.Output(0, "Y", "Output tensor")
.InheritOnnxSchema("Upsample");
// Input: dY, output: dX
GRADIENT_OPERATOR_SCHEMA(ResizeNearestGradient)
.NumInputs(2, 3)
.NumOutputs(1)
.Arg("width_scale", "Scale along width dimension")
.Arg("height_scale", "Scale along height dimension");
class GetResizeNearestGradient : public GradientMakerBase {
using GradientMakerBase::GradientMakerBase;
vector<OperatorDef> GetGradientDefs() override {
if (def_.input().size() == 2) {
// this is a hack to support the second input as dynamic
// width_scale and height_scale to align with onnx change
return SingleGradientDef(
"ResizeNearestGradient",
"",
vector<string>{GO(0), I(0), I(1)},
vector<string>{GI(0)});
}
return SingleGradientDef(
"ResizeNearestGradient",
"",
vector<string>{GO(0), I(0)},
vector<string>{GI(0)});
}
};
REGISTER_GRADIENT(ResizeNearest, GetResizeNearestGradient);
} // namespace caffe2
using ResizeNearestOpFloatCPU =
caffe2::ResizeNearestOp<float, caffe2::CPUContext>;
// clang-format off
C10_EXPORT_CAFFE2_OP_TO_C10_CPU(
ResizeNearest,
"_caffe2::ResizeNearest("
"Tensor X, "
"str order, "
"float width_scale, "
"float height_scale"
") -> (Tensor Y)",
ResizeNearestOpFloatCPU);
// clang-format on
|
