Cache device major version value to avoid multiple calls of getCudaDeviceProperties (facebookresearch#3950)

Summary:
This diff enables to cache the device major version value so getCudaDeviceProperties() doesn't need to be called multiple times. Currently, the profiler of the code looks as so:
{F1933796291}

Pull Request resolved: facebookresearch#3950

Test Plan:
N5114369 -- provides a toy example (2) which exhibits the following timings:
Average timings before change: 3.35s
Average tmings after change:  1.99s

Reviewed By: algoriddle

Differential Revision: D64047778

Pulled By: mlomeli1

fbshipit-source-id: 2f09373944237e80b96d40f35c6714c06f5741a9

Author: mlomeli1

faiss/gpu/GpuDistance.cu

@@ -51,12 +51,18 @@ using namespace raft::distance;
 using namespace raft::neighbors;
 #endif
 
+/// Caches device major version
+int device_major_version = -1;
+
 bool should_use_raft(GpuDistanceParams args) {
-    cudaDeviceProp prop;
-    int dev = args.device >= 0 ? args.device : getCurrentDevice();
-    cudaGetDeviceProperties(&prop, dev);
+    if (device_major_version < 0) {
+        cudaDeviceProp prop;
+        int dev = args.device >= 0 ? args.device : getCurrentDevice();
+        cudaGetDeviceProperties(&prop, dev);
+        device_major_version = prop.major;
+    }
 
-    if (prop.major < 7)
+    if (device_major_version < 7)
         return false;
 
     return args.use_raft;

faiss/gpu/GpuIndex.cu

@@ -42,11 +42,17 @@ constexpr idx_t kAddVecSize = (idx_t)512 * 1024;
 // FIXME: parameterize based on algorithm need
 constexpr idx_t kSearchVecSize = (idx_t)32 * 1024;
 
+/// Caches device major version
+extern int device_major_version;
+
 bool should_use_raft(GpuIndexConfig config_) {
-    cudaDeviceProp prop;
-    cudaGetDeviceProperties(&prop, config_.device);
+    if (device_major_version < 0) {
+        cudaDeviceProp prop;
+        cudaGetDeviceProperties(&prop, config_.device);
+        device_major_version = prop.major;
+    }
 
-    if (prop.major < 7)
+    if (device_major_version < 7)
         return false;
 
     return config_.use_raft;