[SYCL][Graph] Reduce work in update tests

On certain setups these tests take a long time to execute,
impairing the ability to run the whole suite of graph E2E
tests in lit.

Reduce the number of graph execution iterations used in tests to
alleviate this situation. There is an issue with free-function
kernels taking longer that the non-free function equivalent, documented in
CMPLRLLVM-64841, so only use a single execution iteration for now.

Author: EwanC

sycl/test-e2e/Graph/Update/FreeFunctionKernels/update_with_indices_ordering.cpp

@@ -20,9 +20,13 @@ int main() {
 
   // Use a large N to try and make the kernel slow
   const size_t N = 1 << 16;
+
+  // Reduce amount of work compared to version of test without free functions
+  // due to CMPLRLLVM-64841
+  //
   // Loop inside kernel to make even slower (too large N runs out of memory)
-  const size_t NumKernelLoops = 4;
-  const size_t NumSubmitLoops = 8;
+  const size_t NumKernelLoops = 2;
+  const size_t NumSubmitLoops = 1;
 
   exp_ext::command_graph Graph{Ctxt, Queue.get_device()};
 

sycl/test-e2e/Graph/Update/update_with_indices_ordering.cpp

@@ -18,7 +18,7 @@ int main() {
   const size_t N = 1 << 16;
   // Loop inside kernel to make even slower (too large N runs out of memory)
   const size_t NumKernelLoops = 4;
-  const size_t NumSubmitLoops = 8;
+  const size_t NumSubmitLoops = 2;
 
   exp_ext::command_graph Graph{Queue.get_context(), Queue.get_device()};