1 /* Copyright 2015 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 #include "tensorflow/stream_executor/device_description.h"
17 
18 #include <algorithm>
19 
20 #include "absl/strings/str_cat.h"
21 #include "tensorflow/stream_executor/lib/human_readable.h"
22 #include "tensorflow/stream_executor/lib/mathutil.h"
23 
24 namespace stream_executor {
25 
26 static const uint64 kUninitializedUint64 = -1ULL;
27 /* static */ const char *DeviceDescription::kUndefinedString = "<undefined>";
28 
DeviceDescription()29 DeviceDescription::DeviceDescription()
30     : device_vendor_(kUndefinedString),
31       platform_version_(kUndefinedString),
32       driver_version_(kUndefinedString),
33       runtime_version_(kUndefinedString),
34       pci_bus_id_(kUndefinedString),
35       name_(kUndefinedString),
36       thread_dim_limit_(kUninitializedUint64, kUninitializedUint64,
37                         kUninitializedUint64),
38       block_dim_limit_(kUninitializedUint64, kUninitializedUint64,
39                        kUninitializedUint64),
40       threads_per_core_limit_(kUninitializedUint64),
41       threads_per_block_limit_(kUninitializedUint64),
42       threads_per_warp_(kUninitializedUint64),
43       registers_per_core_limit_(kUninitializedUint64),
44       registers_per_block_limit_(kUninitializedUint64),
45       device_address_bits_(kUninitializedUint64),
46       device_memory_size_(kUninitializedUint64),
47       memory_bandwidth_(kUninitializedUint64),
48       shared_memory_per_core_(kUninitializedUint64),
49       shared_memory_per_block_(kUninitializedUint64),
50       clock_rate_ghz_(-1.0),
51       cuda_compute_capability_major_(-1),
52       cuda_compute_capability_minor_(-1),
53       rocm_amdgpu_isa_version_(-1),
54       rocm_amdgpu_gcn_arch_name_(kUndefinedString),
55       numa_node_(-1),
56       core_count_(-1),
57       ecc_enabled_(false) {}
58 
ToMap() const59 std::unique_ptr<std::map<std::string, std::string>> DeviceDescription::ToMap()
60     const {
61   std::unique_ptr<std::map<std::string, std::string>> owned_result{
62       new std::map<std::string, std::string>};
63   std::map<std::string, std::string> &result = *owned_result;
64   result["Device Vendor"] = device_vendor();
65   result["Platform Version"] = platform_version();
66   result["Driver Version"] = driver_version();
67   result["Runtime Version"] = runtime_version();
68   result["PCI bus ID"] = pci_bus_id_;
69   result["Device Name"] = name_;
70 
71   const ThreadDim &thread_dim = thread_dim_limit();
72   result["ThreadDim Limit"] =
73       absl::StrCat(thread_dim.x, ",", thread_dim.y, ",", thread_dim.z);
74   const BlockDim &block_dim = block_dim_limit();
75   result["BlockDim Limit"] =
76       absl::StrCat(block_dim.x, ",", block_dim.y, ",", block_dim.z);
77 
78   result["Threads Per Core Limit"] = absl::StrCat(threads_per_core_limit());
79   result["Threads Per Block Limit"] = absl::StrCat(threads_per_block_limit());
80   result["Registers Per Block Limit"] =
81       absl::StrCat(registers_per_block_limit());
82 
83   result["Device Address Bits"] = absl::StrCat(device_address_bits());
84   result["Device Memory Size"] =
85       port::HumanReadableNumBytes::ToString(device_memory_size());
86   result["Memory Bandwidth"] = absl::StrCat(
87       port::HumanReadableNumBytes::ToString(memory_bandwidth_), "/s");
88 
89   result["Shared Memory Per Core"] =
90       port::HumanReadableNumBytes::ToString(shared_memory_per_core_);
91   result["Shared Memory Per Block"] =
92       port::HumanReadableNumBytes::ToString(shared_memory_per_block_);
93 
94   result["Clock Rate GHz"] = absl::StrCat(clock_rate_ghz());
95 
96   result["CUDA Compute Capability"] = absl::StrCat(
97       cuda_compute_capability_major_, ".", cuda_compute_capability_minor_);
98 
99   result["AMDGPU GCN Arch Name"] = rocm_amdgpu_gcn_arch_name_;
100 
101   result["NUMA Node"] = absl::StrCat(numa_node());
102   result["Core Count"] = absl::StrCat(core_count());
103   result["ECC Enabled"] = absl::StrCat(ecc_enabled());
104   return owned_result;
105 }
106 
107 namespace internal {
108 
DeviceDescriptionBuilder()109 DeviceDescriptionBuilder::DeviceDescriptionBuilder()
110     : device_description_(new DeviceDescription) {}
111 
112 }  // namespace internal
113 
cuda_compute_capability(int * major,int * minor) const114 bool DeviceDescription::cuda_compute_capability(int *major, int *minor) const {
115   *major = cuda_compute_capability_major_;
116   *minor = cuda_compute_capability_minor_;
117   return cuda_compute_capability_major_ != 0;
118 }
119 
rocm_amdgpu_isa_version(int * version) const120 bool DeviceDescription::rocm_amdgpu_isa_version(int *version) const {
121   bool status = false;
122   if (rocm_amdgpu_isa_version_ > 0) {
123     *version = rocm_amdgpu_isa_version_;
124     status = true;
125   }
126   return status;
127 }
128 
ThreadDimOk(const DeviceDescription & device_description,const ThreadDim & thread_dim)129 bool ThreadDimOk(const DeviceDescription &device_description,
130                  const ThreadDim &thread_dim) {
131   const int64 total_threads = thread_dim.x * thread_dim.y * thread_dim.z;
132   const int64 threads_per_block_limit =
133       device_description.threads_per_block_limit();
134   if (total_threads > threads_per_block_limit) {
135     VLOG(2) << "exceeded total-thread-per-block limit: " << total_threads
136             << " vs limit " << threads_per_block_limit;
137     return false;
138   }
139 
140   const auto &limit = device_description.thread_dim_limit();
141   bool ok = thread_dim.x <= limit.x && thread_dim.y <= limit.y &&
142             thread_dim.z <= limit.z;
143   if (!ok) {
144     VLOG(2) << "thread dim " << thread_dim.ToString()
145             << " exceeds limit constraints of " << limit.ToString();
146   }
147   return ok;
148 }
149 
DivideCeil(uint64 x,uint64 y)150 uint64 DivideCeil(uint64 x, uint64 y) {
151   return port::MathUtil::CeilOfRatio(x, y);
152 }
153 
CalculateDimensionality(const DeviceDescription & device_description,int64 element_count,int64 * threads_per_block,int64 * block_count)154 void CalculateDimensionality(const DeviceDescription &device_description,
155                              int64 element_count, int64 *threads_per_block,
156                              int64 *block_count) {
157   *threads_per_block = device_description.threads_per_block_limit();
158   *block_count = port::MathUtil::CeilOfRatio(element_count, *threads_per_block);
159   if (*block_count == 1) {
160     CHECK_LE(element_count, *threads_per_block);
161     *threads_per_block = element_count;
162   }
163 }
164 
165 }  // namespace stream_executor
166