1 /* Copyright 2019 The TensorFlow Authors. All Rights Reserved.
2
3 Licensed under the Apache License, Version 2.0 (the "License");
4 you may not use this file except in compliance with the License.
5 You may obtain a copy of the License at
6
7 http://www.apache.org/licenses/LICENSE-2.0
8
9 Unless required by applicable law or agreed to in writing, software
10 distributed under the License is distributed on an "AS IS" BASIS,
11 WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
12 See the License for the specific language governing permissions and
13 limitations under the License.
14 ==============================================================================*/
15
16 #ifndef TENSORFLOW_CORE_KERNELS_STATEFUL_RANDOM_OPS_CPU_GPU_H_
17 #define TENSORFLOW_CORE_KERNELS_STATEFUL_RANDOM_OPS_CPU_GPU_H_
18
19 #include "tensorflow/core/framework/resource_var.h"
20 #include "tensorflow/core/kernels/stateful_random_ops.h"
21
22 namespace tensorflow {
23
24 // The following 5 functions are made templates to avoid duplicate symbols when
25 // linking.
26
27 // The following 2 functions use the contract "lower 32 bits for the first
28 // uint32, higher 32 bits for the second". Note that this is endian-neutral,
29 // unlike a direct memory copy `memcpy(output, &input, 8)`.
Int64ToUint32s(int64 input,uint32 * output1,uint32 * output2)30 PHILOX_DEVICE_INLINE void Int64ToUint32s(int64 input, uint32* output1,
31 uint32* output2) {
32 auto u64 = static_cast<uint64>(input);
33 *output1 = static_cast<uint32>(u64);
34 *output2 = static_cast<uint32>(u64 >> 32);
35 }
36
Uint32sToInt64(uint32 input1,uint32 input2)37 PHILOX_DEVICE_INLINE int64 Uint32sToInt64(uint32 input1, uint32 input2) {
38 auto u64_1 = static_cast<uint64>(input1);
39 auto u64_2 = static_cast<uint64>(input2);
40 return static_cast<int64>(u64_1 | (u64_2 << 32));
41 }
42
43 PHILOX_DEVICE_INLINE PhiloxRandom
GetPhiloxRandomFromMem(StateElementType const * ptr)44 GetPhiloxRandomFromMem(StateElementType const* ptr) {
45 PhiloxRandom::ResultType counter;
46 PhiloxRandom::Key key;
47 Int64ToUint32s(ptr[0], &counter[0], &counter[1]);
48 Int64ToUint32s(ptr[1], &counter[2], &counter[3]);
49 Int64ToUint32s(ptr[2], &key[0], &key[1]);
50 return PhiloxRandom(counter, key);
51 }
52
WritePhiloxRandomToMem(PhiloxRandom const & philox,StateElementType * ptr)53 PHILOX_DEVICE_INLINE void WritePhiloxRandomToMem(PhiloxRandom const& philox,
54 StateElementType* ptr) {
55 PhiloxRandom::ResultType const& counter = philox.counter();
56 PhiloxRandom::Key const& key = philox.key();
57 ptr[0] = Uint32sToInt64(counter[0], counter[1]);
58 ptr[1] = Uint32sToInt64(counter[2], counter[3]);
59 ptr[2] = Uint32sToInt64(key[0], key[1]);
60 }
61
UpdateMemWithPhiloxRandom(PhiloxRandom const & philox,int64 output_size,StateElementType * ptr)62 PHILOX_DEVICE_INLINE void UpdateMemWithPhiloxRandom(PhiloxRandom const& philox,
63 int64 output_size,
64 StateElementType* ptr) {
65 auto new_philox = philox;
66 // Multiplier 256 is the same as in `FillPhiloxRandomTask`; do not change
67 // it just here.
68 auto delta = output_size * 256;
69 new_philox.Skip(delta); // do the actual increasing
70 WritePhiloxRandomToMem(new_philox, ptr);
71 }
72
73 // A per-device helper function that does the actual work for
74 // `UpdateVariableAndFill`.
75 // Reason to use functor: C++ doesn't allow function-template partial
76 // specialization.
77 template <typename Device, typename Distribution>
78 struct UpdateVariableAndFill_Philox;
79
80 template <typename Device>
81 struct RngSkip_Philox;
82
83 using CPUDevice = Eigen::ThreadPoolDevice;
84
85 struct UpdateVariableAndFill_Philox_Arg {
86 int64 output_size;
87 int64 alg_tag_skip;
88 ScopedUnlockUnrefVar* not_used;
89 Tensor* state_tensor;
90 };
91
92 #if GOOGLE_CUDA || TENSORFLOW_USE_ROCM
93
94 using GPUDevice = Eigen::GpuDevice;
95
96 // Declares the partially GPU-specialized functor structs.
97 // must be kept at <=6 arguments because of a gcc/clang ABI incompatibility bug
98 template <typename Distribution>
99 struct UpdateVariableAndFill_Philox<GPUDevice, Distribution> {
100 void operator()(OpKernelContext* ctx, const GPUDevice& device,
101 Distribution dist, UpdateVariableAndFill_Philox_Arg* arg,
102 typename Distribution::ResultElementType* output_data);
103 };
104
105 template <>
106 struct RngSkip_Philox<GPUDevice> {
107 void operator()(const GPUDevice& device, int64 delta, Tensor* state_tensor);
108 };
109
110 #endif // GOOGLE_CUDA || TENSORFLOW_USE_ROCM
111
112 } // end namespace tensorflow
113
114 #endif // TENSORFLOW_CORE_KERNELS_STATEFUL_RANDOM_OPS_CPU_GPU_H_
115