Merge pull request #13747 from panyx0718/doc

clarify the fraction_of_gpu_memory flag

paddle/fluid/framework/rw_lock.h

@@ -46,6 +46,7 @@ struct RWLock {
  private:
   pthread_rwlock_t lock_;
 };
+// TODO(paddle-dev): Support RWLock for WIN32 for correctness.
 #else
 // https://stackoverflow.com/questions/7125250/making-pthread-rwlock-wrlock-recursive
 // In windows, rw_lock seems like a hack. Use empty object and do nothing.

paddle/fluid/platform/gpu_info.cc

@@ -20,8 +20,11 @@ limitations under the License. */
 #include "paddle/fluid/platform/enforce.h"
 
 DEFINE_double(fraction_of_gpu_memory_to_use, 0.92,
-              "Default use 92% of GPU memory for PaddlePaddle,"
+              "Allocate a trunk of gpu memory that is this fraction of the "
-              "reserve the rest for page tables, etc");
+              "total gpu memory size. Future memory usage will be allocated "
+              "from the trunk. If the trunk doesn't have enough gpu memory, "
+              "additional trunks of the same size will be requested from gpu "
+              "until the gpu has no memory left for another trunk.");
 
 namespace paddle {
 namespace platform {