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 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 blocks_per_core_limit_(kUninitializedUint64), 41 threads_per_core_limit_(kUninitializedUint64), 42 threads_per_block_limit_(kUninitializedUint64), 43 threads_per_warp_(kUninitializedUint64), 44 registers_per_core_limit_(kUninitializedUint64), 45 registers_per_block_limit_(kUninitializedUint64), 46 device_address_bits_(kUninitializedUint64), 47 device_memory_size_(kUninitializedUint64), 48 memory_bandwidth_(kUninitializedUint64), 49 shared_memory_per_core_(kUninitializedUint64), 50 shared_memory_per_block_(kUninitializedUint64), 51 clock_rate_ghz_(-1.0), 52 cuda_compute_capability_major_(-1), 53 cuda_compute_capability_minor_(-1), 54 rocm_amdgpu_isa_version_(-1), 55 numa_node_(-1), 56 core_count_(-1), 57 ecc_enabled_(false) {} 58 59 std::unique_ptr<std::map<string, string>> DeviceDescription::ToMap() const { 60 std::unique_ptr<std::map<string, string>> owned_result{ 61 new std::map<string, string>}; 62 std::map<string, string> &result = *owned_result; 63 result["Device Vendor"] = device_vendor(); 64 result["Platform Version"] = platform_version(); 65 result["Driver Version"] = driver_version(); 66 result["Runtime Version"] = runtime_version(); 67 result["PCI bus ID"] = pci_bus_id_; 68 result["Device Name"] = name_; 69 70 const ThreadDim &thread_dim = thread_dim_limit(); 71 result["ThreadDim Limit"] = 72 absl::StrCat(thread_dim.x, ",", thread_dim.y, ",", thread_dim.z); 73 const BlockDim &block_dim = block_dim_limit(); 74 result["BlockDim Limit"] = 75 absl::StrCat(block_dim.x, ",", block_dim.y, ",", block_dim.z); 76 77 result["Threads Per Core Limit"] = absl::StrCat(threads_per_core_limit()); 78 result["Threads Per Block Limit"] = absl::StrCat(threads_per_block_limit()); 79 result["Registers Per Block Limit"] = 80 absl::StrCat(registers_per_block_limit()); 81 82 result["Device Address Bits"] = absl::StrCat(device_address_bits()); 83 result["Device Memory Size"] = 84 port::HumanReadableNumBytes::ToString(device_memory_size()); 85 result["Memory Bandwidth"] = absl::StrCat( 86 port::HumanReadableNumBytes::ToString(memory_bandwidth_), "/s"); 87 88 result["Shared Memory Per Core"] = 89 port::HumanReadableNumBytes::ToString(shared_memory_per_core_); 90 result["Shared Memory Per Block"] = 91 port::HumanReadableNumBytes::ToString(shared_memory_per_block_); 92 93 result["Clock Rate GHz"] = absl::StrCat(clock_rate_ghz()); 94 95 result["CUDA Compute Capability"] = absl::StrCat( 96 cuda_compute_capability_major_, ".", cuda_compute_capability_minor_); 97 98 result["NUMA Node"] = absl::StrCat(numa_node()); 99 result["Core Count"] = absl::StrCat(core_count()); 100 result["ECC Enabled"] = absl::StrCat(ecc_enabled()); 101 return owned_result; 102 } 103 104 namespace internal { 105 106 DeviceDescriptionBuilder::DeviceDescriptionBuilder() 107 : device_description_(new DeviceDescription) {} 108 109 } // namespace internal 110 111 bool DeviceDescription::cuda_compute_capability(int *major, int *minor) const { 112 *major = cuda_compute_capability_major_; 113 *minor = cuda_compute_capability_minor_; 114 return cuda_compute_capability_major_ != 0; 115 } 116 117 bool DeviceDescription::rocm_amdgpu_isa_version(int *version) const { 118 bool status = false; 119 if (rocm_amdgpu_isa_version_ > 0) { 120 *version = rocm_amdgpu_isa_version_; 121 status = true; 122 } 123 return status; 124 } 125 126 bool ThreadDimOk(const DeviceDescription &device_description, 127 const ThreadDim &thread_dim) { 128 auto total_threads = thread_dim.x * thread_dim.y * thread_dim.z; 129 auto threads_per_block_limit = device_description.threads_per_block_limit(); 130 if (total_threads > threads_per_block_limit) { 131 VLOG(2) << "exceeded total-thread-per-block limit: " << total_threads 132 << " vs limit " << threads_per_block_limit; 133 return false; 134 } 135 136 const auto &limit = device_description.thread_dim_limit(); 137 bool ok = thread_dim.x <= limit.x && thread_dim.y <= limit.y && 138 thread_dim.z <= limit.z; 139 if (!ok) { 140 VLOG(2) << "thread dim " << thread_dim.ToString() 141 << " exceeds limit contraints of " << limit.ToString(); 142 } 143 return ok; 144 } 145 146 uint64 DivideCeil(uint64 x, uint64 y) { 147 return port::MathUtil::CeilOfRatio(x, y); 148 } 149 150 void CalculateDimensionality(const DeviceDescription &device_description, 151 int64 element_count, int64 *threads_per_block, 152 int64 *block_count) { 153 *threads_per_block = device_description.threads_per_block_limit(); 154 *block_count = port::MathUtil::CeilOfRatio(element_count, *threads_per_block); 155 if (*block_count == 1) { 156 CHECK_LE(element_count, *threads_per_block); 157 *threads_per_block = element_count; 158 } 159 } 160 161 } // namespace stream_executor 162
