From a42e88d55b343a69c24f771fdeb01c260fd9c1f6 Mon Sep 17 00:00:00 2001
From: "Dunajski, Bartosz" <bartosz.dunajski@intel.com>
Date: Fri, 11 Aug 2023 14:39:33 +0000
Subject: [PATCH] fix: stall RelaxedOrdering scheduler when programming OCL
 dependencies

Signed-off-by: Dunajski, Bartosz <bartosz.dunajski@intel.com>
---
 opencl/source/command_queue/enqueue_common.h  |  3 +-
 .../command_queue/enqueue_kernel_2_tests.cpp  | 49 +++++++++++++++++++
 2 files changed, 51 insertions(+), 1 deletion(-)

diff --git a/opencl/source/command_queue/enqueue_common.h b/opencl/source/command_queue/enqueue_common.h
index c768e869ca..12cd8c90b4 100644
--- a/opencl/source/command_queue/enqueue_common.h
+++ b/opencl/source/command_queue/enqueue_common.h
@@ -865,6 +865,7 @@ CompletionStamp CommandQueueHw<GfxFamily>::enqueueNonBlocked(
     }
 
     auto memoryCompressionState = getGpgpuCommandStreamReceiver().getMemoryCompressionState(auxTranslationRequired);
+    bool hasStallingCmds = !relaxedOrderingEnabled && (eventsRequest.numEventsInWaitList > 0 || timestampPacketDependencies.previousEnqueueNodes.peekNodes().size() > 0);
 
     DispatchFlags dispatchFlags(
         {},                                                                                                     // csrDependencies
@@ -895,7 +896,7 @@ CompletionStamp CommandQueueHw<GfxFamily>::enqueueNonBlocked(
         kernel->areMultipleSubDevicesInContext(),                                                               // areMultipleSubDevicesInContext
         kernel->requiresMemoryMigration(),                                                                      // memoryMigrationRequired
         isTextureCacheFlushNeeded(commandType),                                                                 // textureCacheFlush
-        false,                                                                                                  // hasStallingCmds
+        hasStallingCmds,                                                                                        // hasStallingCmds
         relaxedOrderingEnabled,                                                                                 // hasRelaxedOrderingDependencies
         false,                                                                                                  // stateCacheInvalidation
         isStallingCommandsOnNextFlushRequired(),                                                                // isStallingCommandsOnNextFlushRequired
diff --git a/opencl/test/unit_test/command_queue/enqueue_kernel_2_tests.cpp b/opencl/test/unit_test/command_queue/enqueue_kernel_2_tests.cpp
index 3d1e6abdf7..15772750a4 100644
--- a/opencl/test/unit_test/command_queue/enqueue_kernel_2_tests.cpp
+++ b/opencl/test/unit_test/command_queue/enqueue_kernel_2_tests.cpp
@@ -1208,6 +1208,55 @@ HWTEST2_F(RelaxedOrderingEnqueueKernelTests, givenEnqueueKernelWhenProgrammingDe
     clReleaseEvent(outEvent);
 }
 
+HWTEST2_F(RelaxedOrderingEnqueueKernelTests, givenRelaxedOrderingDisabledWhenDispatchingWithDependencyThenMarkAsStallingCmd, IsAtLeastXeHpcCore) {
+    auto &ultCsr = pDevice->getUltCommandStreamReceiver<FamilyType>();
+
+    MockKernelWithInternals mockKernel(*pClDevice);
+
+    {
+        MockCommandQueueHw<FamilyType> ioq{context, pClDevice, nullptr};
+
+        ioq.enqueueKernel(mockKernel.mockKernel, 1, nullptr, gws, nullptr, 0, nullptr, nullptr);
+
+        // IOQ without dependency
+        EXPECT_FALSE(ultCsr.recordedDispatchFlags.hasStallingCmds);
+        EXPECT_FALSE(ultCsr.recordedDispatchFlags.hasRelaxedOrderingDependencies);
+
+        ioq.enqueueKernel(mockKernel.mockKernel, 1, nullptr, gws, nullptr, 0, nullptr, nullptr);
+
+        // IOQ with implicit dependency
+        EXPECT_TRUE(ultCsr.recordedDispatchFlags.hasStallingCmds);
+        EXPECT_FALSE(ultCsr.recordedDispatchFlags.hasRelaxedOrderingDependencies);
+    }
+
+    {
+        MockCommandQueueHw<FamilyType> ooq{context, pClDevice, nullptr};
+        ooq.setOoqEnabled();
+
+        cl_event event;
+
+        ooq.enqueueKernel(mockKernel.mockKernel, 1, nullptr, gws, nullptr, 0, nullptr, nullptr);
+
+        // OOQ without dependency
+        EXPECT_FALSE(ultCsr.recordedDispatchFlags.hasStallingCmds);
+        EXPECT_FALSE(ultCsr.recordedDispatchFlags.hasRelaxedOrderingDependencies);
+
+        ooq.enqueueKernel(mockKernel.mockKernel, 1, nullptr, gws, nullptr, 0, nullptr, &event);
+
+        // OOQ without implicit dependency
+        EXPECT_FALSE(ultCsr.recordedDispatchFlags.hasStallingCmds);
+        EXPECT_FALSE(ultCsr.recordedDispatchFlags.hasRelaxedOrderingDependencies);
+
+        ooq.enqueueKernel(mockKernel.mockKernel, 1, nullptr, gws, nullptr, 1, &event, nullptr);
+
+        // OOQ with explicit dependency
+        EXPECT_TRUE(ultCsr.recordedDispatchFlags.hasStallingCmds);
+        EXPECT_FALSE(ultCsr.recordedDispatchFlags.hasRelaxedOrderingDependencies);
+
+        clReleaseEvent(event);
+    }
+}
+
 HWTEST2_F(RelaxedOrderingEnqueueKernelTests, givenBarrierWithDependenciesWhenFlushingThenAllowForRelaxedOrdering, IsAtLeastXeHpcCore) {
     using MI_LOAD_REGISTER_REG = typename FamilyType::MI_LOAD_REGISTER_REG;
     using MI_STORE_DATA_IMM = typename FamilyType::MI_STORE_DATA_IMM;