From d4e3e45e9a8822b85f61f16885876da0ec10fc47 Mon Sep 17 00:00:00 2001 From: Ewan Crawford Date: Wed, 15 Jan 2025 18:09:31 +0000 Subject: [PATCH] [SYCL][Graph][E2E] Reduce work in intensive update tests running (#16632) On certain setups these two 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. --- .../FreeFunctionKernels/update_with_indices_ordering.cpp | 8 +++++--- .../Graph/Update/update_with_indices_ordering.cpp | 2 +- 2 files changed, 6 insertions(+), 4 deletions(-) diff --git a/sycl/test-e2e/Graph/Update/FreeFunctionKernels/update_with_indices_ordering.cpp b/sycl/test-e2e/Graph/Update/FreeFunctionKernels/update_with_indices_ordering.cpp index 4d6aa6445cd0..194d098aec8c 100644 --- a/sycl/test-e2e/Graph/Update/FreeFunctionKernels/update_with_indices_ordering.cpp +++ b/sycl/test-e2e/Graph/Update/FreeFunctionKernels/update_with_indices_ordering.cpp @@ -20,9 +20,11 @@ int main() { // Use a large N to try and make the kernel slow 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; + + // Reduce amount of work compared to version of test without free functions + // due to CMPLRLLVM-64841 + const size_t NumKernelLoops = 1; + const size_t NumSubmitLoops = 1; exp_ext::command_graph Graph{Ctxt, Queue.get_device()}; diff --git a/sycl/test-e2e/Graph/Update/update_with_indices_ordering.cpp b/sycl/test-e2e/Graph/Update/update_with_indices_ordering.cpp index fdd5ffa52fe9..5a9de103053e 100644 --- a/sycl/test-e2e/Graph/Update/update_with_indices_ordering.cpp +++ b/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()};