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
|
/* Copyright 2015 The TensorFlow Authors. All Rights Reserved.
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.
==============================================================================*/
// See docs in ../ops/math_ops.cc.
#define EIGEN_USE_THREADS
#include "tensorflow/core/kernels/cast_op.h"
#include "tensorflow/core/common_runtime/device.h"
#include "tensorflow/core/framework/op.h"
#include "tensorflow/core/framework/op_kernel.h"
#include "tensorflow/core/framework/types.h"
#include "tensorflow/core/platform/logging.h"
#include "tensorflow/core/platform/macros.h"
#include "tensorflow/core/platform/types.h"
#include "tensorflow/core/util/work_sharder.h"
#include "tensorflow/core/kernels/cast_op_impl.h"
namespace tensorflow {
typedef Eigen::ThreadPoolDevice CPUDevice;
typedef Eigen::GpuDevice GPUDevice;
#ifdef TENSORFLOW_USE_SYCL
typedef Eigen::SyclDevice SYCLDevice;
#endif // TENSORFLOW_USE_SYCL
#define CURRY_TYPES2(FN, arg0) \
FN(arg0, bool); \
FN(arg0, uint8); \
FN(arg0, int8); \
FN(arg0, uint16); \
FN(arg0, int16); \
FN(arg0, int32); \
FN(arg0, int64); \
FN(arg0, Eigen::half); \
FN(arg0, float); \
FN(arg0, double); \
FN(arg0, std::complex<float>); \
FN(arg0, std::complex<double>)
CastOpBase::CastOpBase(OpKernelConstruction* ctx) : OpKernel(ctx) {
OP_REQUIRES_OK(ctx, ctx->GetAttr("SrcT", &src_dtype_));
OP_REQUIRES_OK(ctx, ctx->GetAttr("DstT", &dst_dtype_));
}
void CastOpBase::Compute(OpKernelContext* ctx) {
const Tensor& inp = ctx->input(0);
if (work_ == nullptr) {
ctx->set_output(0, inp);
} else {
Tensor* out = nullptr;
OP_REQUIRES_OK(ctx, ctx->allocate_output(0, inp.shape(), &out));
work_(ctx, inp, out);
}
}
Status CastOpBase::Unimplemented() {
return errors::Unimplemented("Cast ", DataTypeString(src_dtype_), " to ",
DataTypeString(dst_dtype_), " is not supported");
}
CpuCastOp::CpuCastOp(OpKernelConstruction* ctx) : CastOpBase(ctx) {
OP_REQUIRES_OK(ctx, Prepare());
}
Status CpuCastOp::Prepare() {
if (src_dtype_ == dst_dtype_) {
work_ = nullptr; // Identity
return Status::OK();
}
if (src_dtype_ == DT_BOOL) {
work_ = GetCpuCastFromBool(dst_dtype_);
} else if (src_dtype_ == DT_UINT8) {
work_ = GetCpuCastFromUint8(dst_dtype_);
} else if (src_dtype_ == DT_INT8) {
work_ = GetCpuCastFromInt8(dst_dtype_);
} else if (src_dtype_ == DT_UINT16) {
work_ = GetCpuCastFromUint16(dst_dtype_);
} else if (src_dtype_ == DT_INT16) {
work_ = GetCpuCastFromInt16(dst_dtype_);
} else if (src_dtype_ == DT_INT32) {
work_ = GetCpuCastFromInt32(dst_dtype_);
} else if (src_dtype_ == DT_INT64) {
work_ = GetCpuCastFromInt64(dst_dtype_);
} else if (src_dtype_ == DT_HALF) {
work_ = GetCpuCastFromHalf(dst_dtype_);
} else if (src_dtype_ == DT_FLOAT) {
work_ = GetCpuCastFromFloat(dst_dtype_);
} else if (src_dtype_ == DT_DOUBLE) {
work_ = GetCpuCastFromDouble(dst_dtype_);
} else if (src_dtype_ == DT_COMPLEX64) {
work_ = GetCpuCastFromComplex64(dst_dtype_);
} else if (src_dtype_ == DT_COMPLEX128) {
work_ = GetCpuCastFromComplex128(dst_dtype_);
} else if (src_dtype_ == DT_BFLOAT16) {
work_ = GetCpuCastFromBfloat(dst_dtype_);
}
// TODO(sesse): If CPU casting to or from Eigen::half ever becomes a
// bottleneck, we could probably implement specialized support for
// vectorized versions (not the least based on F16C for Haswell
// or newer).
return work_ == nullptr ? Unimplemented() : Status::OK();
}
#if GOOGLE_CUDA
class GpuCastOp : public CastOpBase {
public:
explicit GpuCastOp(OpKernelConstruction* ctx) : CastOpBase(ctx) {
OP_REQUIRES_OK(ctx, Prepare());
}
private:
Status Prepare() {
if (src_dtype_ == dst_dtype_) {
work_ = nullptr; // Identity
return Status::OK();
}
if (src_dtype_ == DT_BOOL) {
work_ = GetGpuCastFromBool(dst_dtype_);
} else if (src_dtype_ == DT_UINT8) {
work_ = GetGpuCastFromUint8(dst_dtype_);
} else if (src_dtype_ == DT_INT8) {
work_ = GetGpuCastFromInt8(dst_dtype_);
} else if (src_dtype_ == DT_UINT16) {
work_ = GetGpuCastFromUint16(dst_dtype_);
} else if (src_dtype_ == DT_INT16) {
work_ = GetGpuCastFromInt16(dst_dtype_);
} else if (src_dtype_ == DT_INT32) {
work_ = GetGpuCastFromInt32(dst_dtype_);
} else if (src_dtype_ == DT_INT64) {
work_ = GetGpuCastFromInt64(dst_dtype_);
} else if (src_dtype_ == DT_HALF) {
work_ = GetGpuCastFromHalf(dst_dtype_);
} else if (src_dtype_ == DT_FLOAT) {
work_ = GetGpuCastFromFloat(dst_dtype_);
} else if (src_dtype_ == DT_DOUBLE) {
work_ = GetGpuCastFromDouble(dst_dtype_);
} else if (src_dtype_ == DT_COMPLEX64) {
work_ = GetGpuCastFromComplex64(dst_dtype_);
} else if (src_dtype_ == DT_COMPLEX128) {
work_ = GetGpuCastFromComplex128(dst_dtype_);
} else if (src_dtype_ == DT_BFLOAT16) {
work_ = GetGpuCastFromBfloat(dst_dtype_);
}
return work_ == nullptr ? Unimplemented() : Status::OK();
}
};
#endif // GOOGLE_CUDA
#undef CAST_CASE
REGISTER_KERNEL_BUILDER(Name("Cast").Device(DEVICE_CPU), CpuCastOp);
#if GOOGLE_CUDA
#define REGISTER_CAST_GPU(srctype, dsttype) \
REGISTER_KERNEL_BUILDER(Name("Cast") \
.TypeConstraint<srctype>("SrcT") \
.TypeConstraint<dsttype>("DstT") \
.Device(DEVICE_GPU), \
GpuCastOp)
CURRY_TYPES2(REGISTER_CAST_GPU, bool);
CURRY_TYPES2(REGISTER_CAST_GPU, uint8);
CURRY_TYPES2(REGISTER_CAST_GPU, int8);
CURRY_TYPES2(REGISTER_CAST_GPU, uint16);
CURRY_TYPES2(REGISTER_CAST_GPU, int16);
CURRY_TYPES2(REGISTER_CAST_GPU, int32);
CURRY_TYPES2(REGISTER_CAST_GPU, int64);
CURRY_TYPES2(REGISTER_CAST_GPU, Eigen::half);
CURRY_TYPES2(REGISTER_CAST_GPU, float);
CURRY_TYPES2(REGISTER_CAST_GPU, double);
CURRY_TYPES2(REGISTER_CAST_GPU, std::complex<float>);
CURRY_TYPES2(REGISTER_CAST_GPU, std::complex<double>);
REGISTER_CAST_GPU(float, bfloat16);
REGISTER_CAST_GPU(bfloat16, float);
#undef REGISTER_CAST_GPU
#endif // GOOGLE_CUDA
#ifdef TENSORFLOW_USE_SYCL
class SyclCastOp : public CastOpBase {
public:
explicit SyclCastOp(OpKernelConstruction* ctx) : CastOpBase(ctx) {
OP_REQUIRES_OK(ctx, Prepare());
}
private:
Status Prepare() {
if (src_dtype_ == dst_dtype_) {
work_ = nullptr; // Identity
return Status::OK();
}
if (src_dtype_ == DT_BOOL) {
work_ = GetSyclCastFromBool(dst_dtype_);
} else if (src_dtype_ == DT_INT32) {
work_ = GetSyclCastFromInt32(dst_dtype_);
} else if (src_dtype_ == DT_INT64) {
work_ = GetSyclCastFromInt64(dst_dtype_);
} else if (src_dtype_ == DT_FLOAT) {
work_ = GetSyclCastFromFloat(dst_dtype_);
} else if (src_dtype_ == DT_DOUBLE) {
work_ = GetSyclCastFromDouble(dst_dtype_);
}
return work_ == nullptr ? Unimplemented() : Status::OK();
}
};
#define REGISTER_CAST_SYCL(srctype, dsttype) \
REGISTER_KERNEL_BUILDER(Name("Cast") \
.TypeConstraint<srctype>("SrcT") \
.TypeConstraint<dsttype>("DstT") \
.Device(DEVICE_SYCL), \
SyclCastOp)
CURRY_TYPES2(REGISTER_CAST_SYCL, bool);
CURRY_TYPES2(REGISTER_CAST_SYCL, int32);
CURRY_TYPES2(REGISTER_CAST_SYCL, int64);
CURRY_TYPES2(REGISTER_CAST_SYCL, float);
CURRY_TYPES2(REGISTER_CAST_SYCL, double);
#undef REGISTER_CAST_SYCL
#endif // TENSORFLOW_USE_SYCL
#undef CURRY_TYPES2
// HostCast differs from Cast in that its input and output are in host memory.
REGISTER_KERNEL_BUILDER(Name("_HostCast").Device(DEVICE_CPU), CpuCastOp);
REGISTER_KERNEL_BUILDER(
Name("_HostCast").Device(DEVICE_GPU).HostMemory("x").HostMemory("y"),
CpuCastOp);
#ifdef TENSORFLOW_USE_SYCL
REGISTER_KERNEL_BUILDER(
Name("_HostCast").Device(DEVICE_SYCL).HostMemory("x").HostMemory("y"),
CpuCastOp);
#endif // TENSORFLOW_USE_SYCL
} // end namespace tensorflow
|
