NLCUDA multi-GPU matrices: Need to keep more RAM on main GPU

BrunoLevy · Dec 1, 2024 · b34f1c9 · b34f1c9
1 parent 08e1310
commit b34f1c9
Showing 1 changed file with 5 additions and 1 deletion.
diff --git a/src/lib/geogram/NL/nl_cuda.c b/src/lib/geogram/NL/nl_cuda.c
@@ -1685,8 +1685,12 @@ static int nlCUDAFindDeviceForMatrix(NLCRSMatrix* M) {
 
     /** Try main device first */
     nlCUDACheck(CUDA()->cudaMemGetInfo(&free_RAM, &total_RAM));
+
     /** On main device, we need to keep a bit of RAM for the rest */
-    free_RAM = (size_t)((double)free_RAM * 0.9);
+    if(CUDA()->nb_devices > 1) {
+	free_RAM = (size_t)((double)free_RAM * 0.8);
+    }
+
     if(free_RAM >= required_RAM) {
 	return dev_id;
     }