Create and use getMaxSharedMemory utility

src/fvdb/detail/ops/gsplat/GaussianRasterizeBackward.cu

@@ -1149,11 +1149,7 @@ callRasterizeBackwardWithCorrectSharedChannels(
     if (numSharedSharedChannelsOverride > 0) {
         return callWithSharedChannels(numSharedSharedChannelsOverride);
     } else {
-        cudaDeviceProp deviceProperties;
-        if (cudaGetDeviceProperties(&deviceProperties, stream.device_index()) != cudaSuccess) {
-            AT_ERROR("Failed to query device properties");
-        }
-        const size_t maxSharedMemory = deviceProperties.sharedMemPerBlockOptin;
+        const size_t maxSharedMemory = fvdb::detail::getMaxSharedMemory(stream.device_index());
 
         const size_t sharedMemChannelOptions[4] = {NUM_CHANNELS, 64, 32, 16};
         for (size_t i = 0; i < 4; ++i) {

src/fvdb/detail/utils/cuda/Utils.cuh

@@ -72,6 +72,43 @@ mergeStreams() {
     C10_CUDA_CHECK(cudaEventDestroy(mergeEvent));
 }
 
+/// @brief Get the maximum shared memory per block optin for a given device
+///
+/// This function memoizes the result for each device to minimize calls to
+/// cudaDeviceGetAttribute.
+///
+/// This function is thread-safe.
+///
+/// @param device The device to get the maximum shared memory per block optin for
+/// @return The maximum shared memory per block optin in bytes
+inline int
+getMaxSharedMemory(int device) {
+    TORCH_CHECK(device >= 0, "Invalid device index: ", device);
+
+    static const int deviceCount = c10::cuda::device_count();
+    TORCH_CHECK(device < deviceCount, "Device index out of range: ", device);
+
+    static std::vector<std::atomic<int>> cache(deviceCount);
+    static std::vector<std::once_flag> initFlags(deviceCount);
+
+    if (int cached = cache[device].load(std::memory_order_relaxed); cached > 0) {
+        return cached;
+    }
+
+    std::call_once(initFlags[device], [device]() {
+        int maxSharedMemory = 0;
+        TORCH_CHECK(cudaDeviceGetAttribute(&maxSharedMemory,
+                                           cudaDevAttrMaxSharedMemoryPerBlockOptin,
+                                           device) == cudaSuccess,
+                    "Failed to get max shared memory per block optin");
+        cache[device].store(maxSharedMemory, std::memory_order_relaxed);
+    });
+
+    int cached = cache[device].load(std::memory_order_relaxed);
+    TORCH_CHECK(cached > 0, "Failed to initialize max shared memory for device ", device);
+    return cached;
+};
+
 } // namespace detail
 
 } // namespace fvdb