• Home
  • Line#
  • Scopes#
  • Navigate#
  • Raw
  • Download
1 /* Copyright 2017 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 // This header declares functions which may be called by the generated code on
17 // the CPU. Calls to these functions must be resolved explicitly in the JIT in
18 // xla::cpu::SimpleResolver.  It also defines a per-CpuExecutable context
19 // which is used to cache expensive state and resources utilized by the
20 // aforementioned functions.
21 //
22 // Other functions are declared in individual libraries as well, such as
23 // runtime_conv2d and runtime_matmul. As individual libraries, callers for
24 // ahead-of-time compilation can link only the required subset.
25 
26 #ifndef TENSORFLOW_COMPILER_XLA_SERVICE_CPU_CPU_RUNTIME_H_
27 #define TENSORFLOW_COMPILER_XLA_SERVICE_CPU_CPU_RUNTIME_H_
28 
29 #include "tensorflow/compiler/xla/executable_run_options.h"
30 #include "tensorflow/compiler/xla/service/cpu/xfeed_manager.h"
31 #include "tensorflow/compiler/xla/service/hlo_instructions.h"
32 #include "tensorflow/compiler/xla/types.h"
33 
34 namespace xla {
35 namespace cpu {
36 namespace runtime {
37 
38 // Names of runtime functions. These get resolved from the generated code to the
39 // right symbol at link time in one of two ways:
40 // 1. When using the JIT, the symbol resolver (SimpleResolver in
41 //    third_party/tensorflow/compiler/xla/service/cpu/simple_orc_jit.cc) maps
42 //    this symbol name to
43 //    the actual symbol.
44 // 2. When using ahead-of-time compilation, the linker can resolve the name
45 //    because it is a symbol in the cpu_runtime library.
46 extern const char* const kEigenMatMulF16SymbolName;
47 extern const char* const kEigenMatMulF32SymbolName;
48 extern const char* const kEigenMatMulF64SymbolName;
49 extern const char* const kEigenMatMulC64SymbolName;
50 extern const char* const kEigenMatMulC128SymbolName;
51 extern const char* const kEigenMatMulS32SymbolName;
52 extern const char* const kMKLConvF32SymbolName;
53 extern const char* const kMKLMatMulF32SymbolName;
54 extern const char* const kMKLMatMulF64SymbolName;
55 extern const char* const kMKLSingleThreadedMatMulF32SymbolName;
56 extern const char* const kMKLSingleThreadedMatMulF64SymbolName;
57 extern const char* const kEigenConvF16SymbolName;
58 extern const char* const kEigenConvF32SymbolName;
59 extern const char* const kEigenFftSymbolName;
60 extern const char* const kEigenSingleThreadedFftSymbolName;
61 extern const char* const kEigenSingleThreadedMatMulF16SymbolName;
62 extern const char* const kEigenSingleThreadedMatMulF32SymbolName;
63 extern const char* const kEigenSingleThreadedMatMulF64SymbolName;
64 extern const char* const kEigenSingleThreadedMatMulC64SymbolName;
65 extern const char* const kEigenSingleThreadedMatMulC128SymbolName;
66 extern const char* const kEigenSingleThreadedMatMulS32SymbolName;
67 extern const char* const kEigenSingleThreadedConvF16SymbolName;
68 extern const char* const kEigenSingleThreadedConvF32SymbolName;
69 extern const char* const kAcquireInfeedBufferForDequeueSymbolName;
70 extern const char* const kReleaseInfeedBufferAfterDequeueSymbolName;
71 extern const char* const kAcquireOutfeedBufferForPopulationSymbolName;
72 extern const char* const kReleaseOutfeedBufferAfterPopulationSymbolName;
73 extern const char* const kParallelForkJoinSymbolName;
74 extern const char* const kKeyValueSortSymbolName;
75 extern const char* const kTopKF32SymbolName;
76 extern const char* const kAllReduceSymbolName;
77 extern const char* const kCollectivePermuteSymbolName;
78 extern const char* const kReplicaIdSymbolName;
79 extern const char* const kTracingStartSymbolName;
80 extern const char* const kTracingEndSymbolName;
81 extern const char* const kAllToAllSymbolName;
82 
83 // All symbol names for XLA CPU runtime functions need to start with this
84 // prefix.
85 extern const char* const kXlaCpuRuntimeSymbolNamePrefix;
86 
87 // Returns the infeed manager used by the CPU runtime for the CPU device
88 // `device_ordinal`.  Note the device ordinal does not name a CPU
89 XfeedManager* GetXfeedManager(int device_ordinal);
90 
91 }  // namespace runtime
92 }  // namespace cpu
93 }  // namespace xla
94 
95 extern "C" {
96 
97 extern xla::int64 __xla_cpu_runtime_TracingStart(
98     const void* /* xla::ExecutableRunOptions* */ run_options_ptr,
99     const char* name);
100 extern void __xla_cpu_runtime_TracingEnd(
101     const void* /* xla::ExecutableRunOptions* */ run_options_ptr,
102     xla::int64 id);
103 
104 // Some things common to all of the runtime entry points below:
105 //
106 //  * The shape pointer and shape_length reflect values that can be deserialized
107 //    via llvm_ir::DecodeSelfDescribingShapeConstant. This is the way we pass
108 //    reified type information from the generated program to the runtime, which
109 //    helps check the type safety and contract for the emitted-code/runtime
110 //    communication.
111 //
112 //  * run_options is used to look up the device ordinal for the stream executor
113 //    we're executing under.  If it is null the device ordinal is assumed to be
114 //    0 (this behavior helps in writing tests).
115 
116 // Note: in the runtime entry points below, the shape pointer and shape_length
117 // reflect values that can be deserialized via
118 // llvm_ir::DecodeSelfDescribingShapeConstant. This is the way we pass reified
119 // type information from the generated program to the runtime, which helps check
120 // the type safety and contract for the emitted-code/runtime communication.
121 
122 // Blocks until the next infeed buffer is ready to be dequeued, then
123 // returns it. Fails catastrophically if the next enqueued buffer is
124 // not of the correct length in bytes. Checking the shape rather than
125 // the length would be more exact, but the length check is chosen as a
126 // tradeoff between error checking and speed/simplicity.
127 extern void* __xla_cpu_runtime_AcquireInfeedBufferForDequeue(
128     const xla::ExecutableRunOptions* run_options, xla::int32 buffer_length,
129     const void* shape, xla::int32 shape_length);
130 
131 // Relinquishes the next infeed buffer that was returned by
132 // __xla_cpu_runtime_AcquireInfeedBufferForDequeue. Once this call
133 // completes the data at buffer_ptr may no longer be
134 // accessed. buffer_length must match the length passed to the call to
135 // __xla_cpu_runtime_AcquireInfeedBufferForDequeue that returned
136 // buffer_ptr. This function must be called before the next buffer is
137 // acquired, i.e., there may only be one outstanding infeed buffer in
138 // use by the runtime.  TODO(b/31340454) investigate whether or not it
139 // is worth supporting zero-copy infeed where the buffer is retained
140 // by the compiled code until it has been used. If zero-copy infeed is
141 // implemented we will add support for multiple outstanding buffers
142 // that can be returned out of order.
143 extern void __xla_cpu_runtime_ReleaseInfeedBufferAfterDequeue(
144     const xla::ExecutableRunOptions* run_options, xla::int32 buffer_length,
145     void* buffer_ptr, const void* shape_ptr, xla::int32 shape_length);
146 
147 // Blocks until the next outfeed buffer is available to be populated, then
148 // returns it.
149 extern void* __xla_cpu_runtime_AcquireOutfeedBufferForPopulation(
150     const xla::ExecutableRunOptions* run_options, xla::int32 buffer_length,
151     const void* shape_ptr, xla::int32 shape_length);
152 
153 // Relinquishes the outfeed buffer after it has been populated.
154 // buffer_ptr must have been previously returned by
155 // __xla_cpu_runtime_AcquireOutfeedBufferForPopulation.
156 // Once this call completes, buffer_ptr may no longer be accessed.
157 // buffer_length must match the length passed to the call to
158 // __xla_cpu_runtime_AcquireInfeedBufferForDequeue that returned
159 // buffer_ptr. This function must be called before the next buffer is
160 // acquired, i.e., there may only be one outstanding outfeed buffer in
161 // use by the runtime.
162 extern void __xla_cpu_runtime_ReleaseOutfeedBufferAfterPopulation(
163     const xla::ExecutableRunOptions* run_options, xla::int32 buffer_length,
164     void* buffer_ptr, const void* shape_ptr, xla::int32 shape_length);
165 
166 // Perform all reduce on a CPU.
167 //
168 // participating_replicas: array of replica IDs participating in the reduction,
169 // cf. GetParticipatingReplicas.
170 // channel_id_present, op_id: whether op_id is a channel ID or a module ID.
171 // reduction_kind: operator used for a reduction, cf. ReductionKind.
172 // shape_ptr: shape of all input/output buffers.
173 extern void __xla_cpu_runtime_AllReduce(
174     const xla::ExecutableRunOptions* run_options,
175     const void* replica_groups_str, xla::int32 replica_groups_str_size,
176     xla::int32 channel_id_present, xla::int64 op_id, xla::int32 reduction_kind,
177     const void* shape_ptr, xla::int32 shape_length, xla::int32 num_buffers,
178     void** input_buffers, void** output_buffers);
179 
180 extern void __xla_cpu_runtime_CollectivePermute(
181     const xla::ExecutableRunOptions* run_options, xla::int32 channel_id_present,
182     xla::int64 op_id, xla::int32 byte_size, void* input_buffer,
183     void* output_buffer, const void* source_target_pairs,
184     xla::int32 source_target_pairs_size);
185 
186 extern void __xla_cpu_runtime_AllToAll(
187     const xla::ExecutableRunOptions* run_options, xla::int32 channel_id_present,
188     xla::int64 op_id, const void* replica_groups_str,
189     xla::int32 replica_groups_str_size, xla::int32 num_buffers,
190     xla::int64 buffer_size, void** source_buffers, void** destination_buffers);
191 
192 // Write the replica ID into the output buffer.
193 extern void __xla_cpu_runtime_ReplicaId(
194     const xla::ExecutableRunOptions* run_options, void* output_buffer);
195 
196 }  // extern "C"
197 
198 #endif  // TENSORFLOW_COMPILER_XLA_SERVICE_CPU_CPU_RUNTIME_H_
199