Launch GPU kernels on CUDA streams

src/TrixiCUDA.jl

@@ -62,7 +62,7 @@ export semidiscretizeGPU
 # Get called every time the package is loaded
 function __init__()
 
-    # Initialize the device properties
+    # Initialize the device
     init_device()
 end
 

src/auxiliary/auxiliary.jl

@@ -1,29 +1,60 @@
 include("configurators.jl")
 include("stable.jl")
 
-# Initialize the device properties
+# Initialize the device 
 function init_device()
     try
-        # Consider single GPU for now
         # TODO: Consider multiple GPUs later
         device = CUDA.device()
 
-        # Get the device properties
-        global MULTIPROCESSOR_COUNT = CUDA.attribute(device, CUDA.CU_DEVICE_ATTRIBUTE_MULTIPROCESSOR_COUNT)
-        global MAX_THREADS_PER_BLOCK = CUDA.attribute(device, CUDA.CU_DEVICE_ATTRIBUTE_MAX_THREADS_PER_BLOCK)
-        global MAX_SHARED_MEMORY_PER_BLOCK = CUDA.attribute(device, CUDA.CU_DEVICE_ATTRIBUTE_MAX_SHARED_MEMORY_PER_BLOCK)
+        try
+            # Get properties 
+            global MULTIPROCESSOR_COUNT = CUDA.attribute(device, CUDA.CU_DEVICE_ATTRIBUTE_MULTIPROCESSOR_COUNT)
+            global MAX_THREADS_PER_BLOCK = CUDA.attribute(device, CUDA.CU_DEVICE_ATTRIBUTE_MAX_THREADS_PER_BLOCK)
+            global MAX_SHARED_MEMORY_PER_BLOCK = CUDA.attribute(device, CUDA.CU_DEVICE_ATTRIBUTE_MAX_SHARED_MEMORY_PER_BLOCK)
+        catch e
+            # Handle CUDA errors
+            println("Error getting device properties: ", e.msg)
+            println("Ensure that your GPU device properties can be retrieved successfully.")
+
+            # Fall back to set zeros
+            global MULTIPROCESSOR_COUNT = 0
+            global MAX_THREADS_PER_BLOCK = 0
+            global MAX_SHARED_MEMORY_PER_BLOCK = 0
+            println("Device properties have been set to zero. CUDA operations 
+            will fail because no valid properties are configured.")
+        end
+
+        try
+            # Create CUDA streams
+            global STREAM1 = CUDA.CuStream()
+            global STREAM2 = CUDA.CuStream()
+        catch e
+            # Handle CUDA errors
+            println("Error initializing CUDA streams: ", e.msg)
+            println("Ensure there are enough GPU resources available for stream creation.")
+
+            # Fall back to set default streams
+            global STREAM1 = CUDA.stream()
+            global STREAM2 = CUDA.stream()
+            println("Streams have been set to the default CUDA stream. 
+            This will impact performance due to no concurrency.")
+        end
+
     catch e
-        # Handle the errors
+        # Handle CUDA errors or not
         if e isa CUDA.CuError
-            println("Error initializing device: ", e.msg)
-            println("Ensure a CUDA-enabled GPU is available and properly configured.")
+            println("Error detecting device: ", e.msg)
+            println("Ensure a CUDA compatible GPU is available and properly configured.")
         else
             println("An unexpected error occurred: ", e)
         end
 
-        # Fall back to set default values
-        global MULTIPROCESSOR_COUNT = 0
-        global MAX_THREADS_PER_BLOCK = 0
-        global MAX_SHARED_MEMORY_PER_BLOCK = 0
+        # Fall back to set nothing
+        global MULTIPROCESSOR_COUNT = nothing
+        global MAX_THREADS_PER_BLOCK = nothing
+        global MAX_SHARED_MEMORY_PER_BLOCK = nothing
+        global STREAM1 = nothing
+        global STREAM2 = nothing
     end
 end