From ee8c95de1c8b3d80692f8e13d898bd87a80f592d Mon Sep 17 00:00:00 2001
From: Filip Leonarski <filip.leonarski@psi.ch>
Date: Sun, 11 Oct 2026 14:25:56 +0200
Subject: [PATCH] CUDAMemHelpers test: the FFT indexer wait case trims the
 allocation pool after its pools

The pools' indexers return their memory to the allocation pool when destroyed; without a
trim the next GPU test in the same process (AdaptiveSpotFinderGPU_RunToRunReproducible)
found the card taken and failed with out of memory.

Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com>
---
 tests/CUDAMemHelpersTest.cpp | 61 +++++++++++++++++++-----------------
 1 file changed, 33 insertions(+), 28 deletions(-)

diff --git a/tests/CUDAMemHelpersTest.cpp b/tests/CUDAMemHelpersTest.cpp
index 8c08d74f0..e095333a7 100644
--- a/tests/CUDAMemHelpersTest.cpp
+++ b/tests/CUDAMemHelpersTest.cpp
@@ -154,37 +154,42 @@ TEST_CASE("IndexerThreadPool_FFTIndexerWaitsForGPUWorkBeside", "[CUDAMemHelpers]
                 }
             }
     };
-    IndexerThreadPool probe(settings, IndexerConstruction::OnFirstUse, /*beside=*/true);
-    IndexerThreadPool pool(settings, IndexerConstruction::OnFirstUse);
+    // The pools' indexers go back to the allocation pool when they are destroyed; trim it after them, so
+    // the next test finds the card free.
+    {
+        IndexerThreadPool probe(settings, IndexerConstruction::OnFirstUse, /*beside=*/true);
+        IndexerThreadPool pool(settings, IndexerConstruction::OnFirstUse);
 
-    fill();
-    {
-        GPUWorkBeside work;
-        CHECK(probe.Run(x, recip).error != nullptr);
-    }
-    {
-        std::optional<GPUImagePass> image_pass(std::in_place);
-        std::thread ends([&] {
-            std::this_thread::sleep_for(std::chrono::milliseconds(500));
-            held.reset();
-            image_pass.reset();
-        });
-        GPUWorkBeside work;
-        CHECK(probe.Run(x, recip).error == nullptr);
-        ends.join();
-    }
+        fill();
+        {
+            GPUWorkBeside work;
+            CHECK(probe.Run(x, recip).error != nullptr);
+        }
+        {
+            std::optional<GPUImagePass> image_pass(std::in_place);
+            std::thread ends([&] {
+                std::this_thread::sleep_for(std::chrono::milliseconds(500));
+                held.reset();
+                image_pass.reset();
+            });
+            GPUWorkBeside work;
+            CHECK(probe.Run(x, recip).error == nullptr);
+            ends.join();
+        }
 
-    fill();
-    {
-        std::optional<GPUWorkBeside> work(std::in_place);
-        std::thread ends([&] {
-            std::this_thread::sleep_for(std::chrono::milliseconds(500));
-            held.reset();
-            work.reset();
-        });
-        CHECK(pool.Run(x, recip).error == nullptr);
-        ends.join();
+        fill();
+        {
+            std::optional<GPUWorkBeside> work(std::in_place);
+            std::thread ends([&] {
+                std::this_thread::sleep_for(std::chrono::milliseconds(500));
+                held.reset();
+                work.reset();
+            });
+            CHECK(pool.Run(x, recip).error == nullptr);
+            ends.join();
+        }
     }
+    cuda_trim_allocation_pool();
 }
 
 #endif