From 198762baa8966f835b616dfd3b4c79ed5edcd824 Mon Sep 17 00:00:00 2001
From: Lukasz Jobczyk <lukasz.jobczyk@intel.com>
Date: Thu, 26 Jun 2025 12:10:18 +0000
Subject: [PATCH] performance: Store pattern allocations on reset and sync

Related-To: NEO-9729

Signed-off-by: Lukasz Jobczyk <lukasz.jobczyk@intel.com>
---
 level_zero/core/source/cmdlist/cmdlist_hw.inl |  5 +++
 .../source/cmdlist/cmdlist_hw_immediate.h     |  1 +
 .../source/cmdlist/cmdlist_hw_immediate.inl   |  5 +++
 .../cmdlist/test_in_order_cmdlist_2.cpp       | 40 +++++++++++++++++++
 4 files changed, 51 insertions(+)

diff --git a/level_zero/core/source/cmdlist/cmdlist_hw.inl b/level_zero/core/source/cmdlist/cmdlist_hw.inl
index adc89f93c9..8158bbd6ec 100644
--- a/level_zero/core/source/cmdlist/cmdlist_hw.inl
+++ b/level_zero/core/source/cmdlist/cmdlist_hw.inl
@@ -112,6 +112,11 @@ void CommandListCoreFamily<gfxCoreFamily>::postInitComputeSetup() {
 
 template <GFXCORE_FAMILY gfxCoreFamily>
 ze_result_t CommandListCoreFamily<gfxCoreFamily>::reset() {
+    for (auto &patternAlloc : this->patternAllocations) {
+        device->storeReusableAllocation(*patternAlloc);
+    }
+    this->patternAllocations.clear();
+
     removeDeallocationContainerData();
     removeHostPtrAllocations();
     removeMemoryPrefetchAllocations();
diff --git a/level_zero/core/source/cmdlist/cmdlist_hw_immediate.h b/level_zero/core/source/cmdlist/cmdlist_hw_immediate.h
index 6604e3075c..09f0f88813 100644
--- a/level_zero/core/source/cmdlist/cmdlist_hw_immediate.h
+++ b/level_zero/core/source/cmdlist/cmdlist_hw_immediate.h
@@ -57,6 +57,7 @@ struct CommandListCoreFamilyImmediate : public CommandListCoreFamily<gfxCoreFami
     using BaseClass::isInOrderExecutionEnabled;
     using BaseClass::isSkippingInOrderBarrierAllowed;
     using BaseClass::isTbxMode;
+    using BaseClass::patternAllocations;
 
     using ComputeFlushMethodType = NEO::CompletionStamp (CommandListCoreFamilyImmediate<gfxCoreFamily>::*)(NEO::LinearStream &, size_t, bool, bool, NEO::AppendOperations, bool);
 
diff --git a/level_zero/core/source/cmdlist/cmdlist_hw_immediate.inl b/level_zero/core/source/cmdlist/cmdlist_hw_immediate.inl
index da7e8ef923..a9c1cd1f73 100644
--- a/level_zero/core/source/cmdlist/cmdlist_hw_immediate.inl
+++ b/level_zero/core/source/cmdlist/cmdlist_hw_immediate.inl
@@ -1197,6 +1197,11 @@ ze_result_t CommandListCoreFamilyImmediate<gfxCoreFamily>::hostSynchronize(uint6
                 if (inOrderExecInfo) {
                     inOrderExecInfo->releaseNotUsedTempTimestampNodes(false);
                 }
+
+                for (auto &patternAlloc : this->patternAllocations) {
+                    this->device->storeReusableAllocation(*patternAlloc);
+                }
+                this->patternAllocations.clear();
             }
 
             bool hangDetected = status == ZE_RESULT_ERROR_DEVICE_LOST;
diff --git a/level_zero/core/test/unit_tests/sources/cmdlist/test_in_order_cmdlist_2.cpp b/level_zero/core/test/unit_tests/sources/cmdlist/test_in_order_cmdlist_2.cpp
index aec886c73d..4e6d74c9bb 100644
--- a/level_zero/core/test/unit_tests/sources/cmdlist/test_in_order_cmdlist_2.cpp
+++ b/level_zero/core/test/unit_tests/sources/cmdlist/test_in_order_cmdlist_2.cpp
@@ -1864,6 +1864,46 @@ HWTEST2_F(InOrderRegularCmdListTests, givenInOrderModeWhenDispatchingRegularCmdL
     EXPECT_FALSE(regularCmdList->latestOperationRequiredNonWalkerInOrderCmdsChaining);
 }
 
+HWTEST2_F(InOrderRegularCmdListTests, givenAppendMemoryFillWhenHostSynchronizeThenStoreFillAllocationsInReusableContainer, IsAtLeastXeCore) {
+    auto immCmdList = createImmCmdList<FamilyType::gfxCoreFamily>();
+    EXPECT_EQ(immCmdList->patternAllocations.size(), 0u);
+    EXPECT_TRUE(static_cast<DeviceImp *>(immCmdList->device)->allocationsForReuse->peekIsEmpty());
+
+    constexpr size_t size = 128 * sizeof(uint32_t);
+    auto data = allocDeviceMem(size);
+    uint64_t pattern = 0u;
+
+    immCmdList->appendMemoryFill(data, &pattern, sizeof(pattern), size, nullptr, 0, nullptr, copyParams);
+    EXPECT_EQ(immCmdList->patternAllocations.size(), 1u);
+    EXPECT_TRUE(static_cast<DeviceImp *>(immCmdList->device)->allocationsForReuse->peekIsEmpty());
+
+    immCmdList->hostSynchronize(std::numeric_limits<uint64_t>::max());
+    EXPECT_EQ(immCmdList->patternAllocations.size(), 0u);
+    EXPECT_FALSE(static_cast<DeviceImp *>(immCmdList->device)->allocationsForReuse->peekIsEmpty());
+
+    context->freeMem(data);
+}
+
+HWTEST2_F(InOrderRegularCmdListTests, givenAppendMemoryFillWhenResetThenStoreFillAllocationsInReusableContainer, IsAtLeastXeCore) {
+    auto regularCmdList = createRegularCmdList<FamilyType::gfxCoreFamily>(false);
+    EXPECT_EQ(regularCmdList->patternAllocations.size(), 0u);
+    EXPECT_TRUE(static_cast<DeviceImp *>(regularCmdList->device)->allocationsForReuse->peekIsEmpty());
+
+    constexpr size_t size = 128 * sizeof(uint32_t);
+    auto data = allocDeviceMem(size);
+    uint64_t pattern = 0u;
+
+    regularCmdList->appendMemoryFill(data, &pattern, sizeof(pattern), size, nullptr, 0, nullptr, copyParams);
+    EXPECT_EQ(regularCmdList->patternAllocations.size(), 1u);
+    EXPECT_TRUE(static_cast<DeviceImp *>(regularCmdList->device)->allocationsForReuse->peekIsEmpty());
+
+    regularCmdList->reset();
+    EXPECT_EQ(regularCmdList->patternAllocations.size(), 0u);
+    EXPECT_FALSE(static_cast<DeviceImp *>(regularCmdList->device)->allocationsForReuse->peekIsEmpty());
+
+    context->freeMem(data);
+}
+
 HWTEST2_F(InOrderRegularCmdListTests, givenInOrderModeWhenDispatchingRegularCmdListThenUpdateCounterAllocation, IsAtLeastXeCore) {
     using MI_STORE_DATA_IMM = typename FamilyType::MI_STORE_DATA_IMM;