From f41959ccb2d9d4c722fe8fc3351401d53bcf4900 Mon Sep 17 00:00:00 2001
From: Manjunath Kudlur <keveman@gmail.com>
Date: Fri, 6 Nov 2015 16:27:58 -0800
Subject: TensorFlow: Initial commit of TensorFlow library. TensorFlow is an
 open source software library for numerical computation using data flow
 graphs.

Base CL: 107276108
---
 tensorflow/stream_executor/gcuda.cc | 87 +++++++++++++++++++++++++++++++++++++
 1 file changed, 87 insertions(+)
 create mode 100644 tensorflow/stream_executor/gcuda.cc

(limited to 'tensorflow/stream_executor/gcuda.cc')

diff --git a/tensorflow/stream_executor/gcuda.cc b/tensorflow/stream_executor/gcuda.cc
new file mode 100644
index 0000000000..505534c08f
--- /dev/null
+++ b/tensorflow/stream_executor/gcuda.cc
@@ -0,0 +1,87 @@
+#include "tensorflow/stream_executor/gcuda.h"
+
+namespace perftools {
+namespace gputools {
+
+// Returns the mapping of gcudacc kernel stub to preferred cache
+// configuration. C++ static singleton pattern.
+std::map<void *, KernelCacheConfig> &GetGcudaccStubToCacheConfigMap() {
+  static std::map<void *, KernelCacheConfig> cache_config_by_stub;
+  return cache_config_by_stub;
+}
+
+shared_mem_config::SharedMemConfig DeviceGetSharedMemConfig(
+    StreamExecutor *stream_exec) {
+  SharedMemoryConfig config = stream_exec->GetDeviceSharedMemoryConfig();
+
+  switch (config) {
+    case SharedMemoryConfig::kDefault:
+      return shared_mem_config::kDefaultBankSize;
+    case SharedMemoryConfig::kFourByte:
+      return shared_mem_config::kFourByteBankSize;
+    case SharedMemoryConfig::kEightByte:
+      return shared_mem_config::kEightByteBankSize;
+    default:
+      LOG(FATAL) << "Impossible shared memory config returned: "
+                 << static_cast<int>(config);
+  }
+}
+
+void DeviceSetSharedMemConfig(StreamExecutor *stream_exec,
+                              shared_mem_config::SharedMemConfig config) {
+  SharedMemoryConfig executor_config;
+  switch (config) {
+    case shared_mem_config::kDefaultBankSize:
+      executor_config = SharedMemoryConfig::kDefault;
+      break;
+    case shared_mem_config::kFourByteBankSize:
+      executor_config = SharedMemoryConfig::kFourByte;
+      break;
+    case shared_mem_config::kEightByteBankSize:
+      executor_config = SharedMemoryConfig::kEightByte;
+      break;
+    default:
+      LOG(FATAL) << "Impossible shared memory config specified: "
+                 << static_cast<int>(config);
+  }
+
+  if (!stream_exec->SetDeviceSharedMemoryConfig(executor_config).ok()) {
+    // The message is logged at a higher level.
+    LOG(INFO) << "Unable to set cache configuration; proceeding.";
+  }
+}
+
+template <>
+void FuncSetCacheConfig<void *>(Stream *stream, void *fptr,
+                                cache_config::CacheConfig cache_config) {
+  // Map from the legacy to the C++11 type.
+  KernelCacheConfig kernel_cache_config;
+  switch (cache_config) {
+    case cache_config::kPreferShared:
+      kernel_cache_config = KernelCacheConfig::kPreferShared;
+      break;
+    case cache_config::kPreferL1:
+      kernel_cache_config = KernelCacheConfig::kPreferL1;
+      break;
+    case cache_config::kPreferEqual:
+      kernel_cache_config = KernelCacheConfig::kPreferEqual;
+      break;
+    default:
+      kernel_cache_config = KernelCacheConfig::kNoPreference;
+  }
+  auto cache_config_map = GetGcudaccStubToCacheConfigMap();
+  cache_config_map[fptr] = kernel_cache_config;
+}
+
+template <>
+KernelCacheConfig FuncGetCacheConfig<void *>(void *fptr) {
+  auto cache_config_map = GetGcudaccStubToCacheConfigMap();
+  auto iter = cache_config_map.find(fptr);
+  if (iter == cache_config_map.end()) {
+    return KernelCacheConfig::kNoPreference;
+  }
+  return cache_config_map[fptr];
+}
+
+}  // namespace gputools
+}  // namespace perftools
-- 
cgit v1.2.3