Improve zeCommandListAppendMemoryFill Performance (2)

Add missing kernel for remainder kernel when pattern size is 1.

Signed-off-by: Jaime Arteaga <[email protected]>

Author: Jaime Arteaga

level_zero/core/source/cmdlist/cmdlist_hw.inl

@@ -1182,12 +1182,25 @@ ze_result_t CommandListCoreFamily<gfxCoreFamily>::appendMemoryFill(void *ptr,
 
         uint32_t groups = static_cast<uint32_t>(size) / groupSizeX;
         ze_group_count_t dispatchFuncArgs{groups, 1u, 1u};
-        res = CommandListCoreFamily<gfxCoreFamily>::appendLaunchKernel(builtinFunction->toHandle(),
-                                                                       &dispatchFuncArgs, nullptr,
-                                                                       0, nullptr);
+        res = appendLaunchKernelSplit(builtinFunction->toHandle(), &dispatchFuncArgs, hSignalEvent);
         if (res) {
             return res;
         }
+
+        uint32_t groupRemainderSizeX = static_cast<uint32_t>(size) % groupSizeX;
+        if (groupRemainderSizeX) {
+            builtinFunction->setGroupSize(groupRemainderSizeX, 1u, 1u);
+            ze_group_count_t dispatchFuncRemainderArgs{1u, 1u, 1u};
+
+            size_t dstOffset = dstAllocation.offset + (size - groupRemainderSizeX);
+            builtinFunction->setArgBufferWithAlloc(0, dstAllocation.alignedAllocationPtr, dstAllocation.alloc);
+            builtinFunction->setArgumentValue(1, sizeof(dstOffset), &dstOffset);
+
+            res = appendLaunchKernelSplit(builtinFunction->toHandle(), &dispatchFuncRemainderArgs, hSignalEvent);
+            if (res) {
+                return res;
+            }
+        }
     } else {
         auto builtinFunction = device->getBuiltinFunctionsLib()->getFunction(Builtin::FillBufferMiddle);
 

level_zero/core/test/unit_tests/sources/cmdlist/test_cmdlist_fill.cpp

@@ -68,6 +68,7 @@ class AppendFillFixture : public DeviceFixture, public ::testing::Test {
 
     void SetUp() override {
         dstPtr = new uint8_t[allocSize];
+        immediateDstPtr = new uint8_t[allocSize];
 
         neoDevice = NEO::MockDevice::createWithNewExecutionEnvironment<NEO::MockDevice>(NEO::defaultHwInfo.get());
         auto mockBuiltIns = new MockBuiltins();
@@ -80,20 +81,38 @@ class AppendFillFixture : public DeviceFixture, public ::testing::Test {
     }
 
     void TearDown() override {
+        delete[] immediateDstPtr;
         delete[] dstPtr;
     }
 
     std::unique_ptr<Mock<MockDriverFillHandle>> driverHandle;
     NEO::MockDevice *neoDevice = nullptr;
     L0::Device *device = nullptr;
-    static constexpr size_t allocSize = 512;
-    static constexpr size_t patternSize = 4;
+    static constexpr size_t allocSize = 70;
+    static constexpr size_t patternSize = 8;
     uint8_t *dstPtr = nullptr;
     uint8_t pattern[patternSize] = {1, 2, 3, 4};
+
+    static constexpr size_t immediateAllocSize = 106;
+    uint8_t immediatePattern = 4;
+    uint8_t *immediateDstPtr = nullptr;
 };
 
 using Platforms = IsAtLeastProduct<IGFX_SKYLAKE>;
 
+HWTEST2_F(AppendFillFixture,
+          givenCallToAppendMemoryFillWithImmediateValueThenSuccessIsReturned, Platforms) {
+    using GfxFamily = typename NEO::GfxFamilyMapper<gfxCoreFamily>::GfxFamily;
+
+    auto commandList = std::make_unique<WhiteBox<MockCommandList<gfxCoreFamily>>>();
+    commandList->initialize(device, NEO::EngineGroupType::RenderCompute);
+
+    auto result = commandList->appendMemoryFill(immediateDstPtr, &immediatePattern,
+                                                sizeof(immediatePattern),
+                                                immediateAllocSize, nullptr, 0, nullptr);
+    EXPECT_EQ(ZE_RESULT_SUCCESS, result);
+}
+
 HWTEST2_F(AppendFillFixture,
           givenCallToAppendMemoryFillThenSuccessIsReturned, Platforms) {
     using GfxFamily = typename NEO::GfxFamilyMapper<gfxCoreFamily>::GfxFamily;