refactor: add SetMaxBVHLevels debug flag

Signed-off-by: Kamil Kopryk <kamil.kopryk@intel.com>

level_zero/core/source/device/device_imp.cpp

@@ -902,6 +902,11 @@ ze_result_t DeviceImp::getKernelProperties(ze_device_module_properties_t *pKerne
             if (releaseHelper && releaseHelper->isRayTracingSupported()) {
                 rtProperties->flags = ZE_DEVICE_RAYTRACING_EXT_FLAG_RAYQUERY;
                 rtProperties->maxBVHLevels = NEO::RayTracingHelper::maxBvhLevels;
+
+                if (NEO::debugManager.flags.SetMaxBVHLevels.get() != -1) {
+                    rtProperties->maxBVHLevels = static_cast<uint32_t>(NEO::debugManager.flags.SetMaxBVHLevels.get());
+                }
+
             } else {
                 rtProperties->flags = 0;
                 rtProperties->maxBVHLevels = 0;

level_zero/core/source/kernel/kernel_imp.cpp

@@ -1199,6 +1199,11 @@ ze_result_t KernelImp::initialize(const ze_kernel_desc_t *desc) {
 
     if (this->usesRayTracing()) {
         uint32_t bvhLevels = NEO::RayTracingHelper::maxBvhLevels;
+
+        if (NEO::debugManager.flags.SetMaxBVHLevels.get() != -1) {
+            bvhLevels = static_cast<uint32_t>(NEO::debugManager.flags.SetMaxBVHLevels.get());
+        }
+
         auto arg = this->getImmutableData()->getDescriptor().payloadMappings.implicitArgs.rtDispatchGlobals;
         neoDevice->initializeRayTracing(bvhLevels);
 

level_zero/core/test/unit_tests/sources/device/test_l0_device.cpp

@@ -1205,6 +1205,33 @@ HWTEST2_F(DeviceTest, whenPassingRaytracingExpStructToGetPropertiesThenPropertie
     EXPECT_EQ(expectedMaxBVHLevels, rayTracingProperties.maxBVHLevels);
 }
 
+HWTEST2_F(DeviceTest, givenSetMaxBVHLevelsWhenPassingRaytracingExpStructToGetPropertiesThenPropertiesWithCorrectFlagIsReturned, MatchAny) {
+
+    DebugManagerStateRestore dbgRestorer;
+    debugManager.flags.SetMaxBVHLevels.set(7);
+
+    ze_device_module_properties_t kernelProperties = {};
+    kernelProperties.stype = ZE_STRUCTURE_TYPE_KERNEL_PROPERTIES;
+
+    ze_device_raytracing_ext_properties_t rayTracingProperties = {};
+    rayTracingProperties.stype = ZE_STRUCTURE_TYPE_DEVICE_RAYTRACING_EXT_PROPERTIES;
+    rayTracingProperties.flags = ZE_DEVICE_RAYTRACING_EXT_FLAG_FORCE_UINT32;
+    rayTracingProperties.maxBVHLevels = 37u;
+
+    kernelProperties.pNext = &rayTracingProperties;
+
+    ze_result_t res = device->getKernelProperties(&kernelProperties);
+    EXPECT_EQ(ZE_RESULT_SUCCESS, res);
+    EXPECT_NE(ZE_DEVICE_RAYTRACING_EXT_FLAG_FORCE_UINT32, rayTracingProperties.flags);
+
+    auto releaseHelper = this->neoDevice->getReleaseHelper();
+    if (releaseHelper && releaseHelper->isRayTracingSupported()) {
+        EXPECT_EQ(7u, rayTracingProperties.maxBVHLevels);
+    } else {
+        EXPECT_EQ(0u, rayTracingProperties.maxBVHLevels);
+    }
+}
+
 TEST_F(DeviceTest, givenKernelPropertiesStructureWhenKernelPropertiesCalledThenAllPropertiesAreAssigned) {
     const auto &hardwareInfo = this->neoDevice->getHardwareInfo();
 

shared/source/debug_settings/debug_variables_base.inl

@@ -304,6 +304,7 @@ DECLARE_DEBUG_VARIABLE(int32_t, DeferStateInitSubmissionToFirstRegularUsage, -1,
 DECLARE_DEBUG_VARIABLE(int32_t, ForceNonWalkerSplitMemoryCopy, -1, "-1: default, 0: disabled, 1: enabled. If set, memory copy will be executed as single byte copy Walker without performance optimizations")
 DECLARE_DEBUG_VARIABLE(int32_t, OverrideTimestampWidth, -1, "-1: default from KMD, > 0: Override timestamp width used for profiling. Requires XeKMD kernel.")
 DECLARE_DEBUG_VARIABLE(int32_t, DebugUmdFifoPollInterval, -1, "-1: default , > 0: Fifo will be polled based on input in milliseconds.")
+DECLARE_DEBUG_VARIABLE(int32_t, SetMaxBVHLevels, -1, "-1: default , > 0: Set maxBVHLevel in RTDispatchGlobal.")
 DECLARE_DEBUG_VARIABLE(int32_t, DebugUmdInterruptTimeout, -1, "-1: default , > 0: interruptTimeout based on input in milliseconds. Default is 2000 milliseconds")
 DECLARE_DEBUG_VARIABLE(int32_t, DebugUmdMaxReadWriteRetry, -1, "-1: default , > 0: max pread/pwrite retry attempts in read/writeGpuMemory calls based on input in milliseconds. Default is 3")
 DECLARE_DEBUG_VARIABLE(int32_t, ForceIndirectDetectionForCMKernels, -1, "-1: default , 0 : disable indirect detection for CM kernels, 1 : enable indirect detection for CM kernels")

shared/source/device/device.cpp

@@ -1134,18 +1134,24 @@ void Device::allocateRTDispatchGlobals(uint32_t maxBvhLevels) {
             break;
         }
 
-        struct RTDispatchGlobals dispatchGlobals = {0};
+        RTDispatchGlobals dispatchGlobals = {0};
 
         dispatchGlobals.rtMemBasePtr = rtStackAllocation->getGpuAddress() + rtStackSize;
         dispatchGlobals.callStackHandlerKSP = reinterpret_cast<uint64_t>(nullptr);
         auto releaseHelper = getReleaseHelper();
         dispatchGlobals.stackSizePerRay = releaseHelper ? releaseHelper->getStackSizePerRay() : 0;
-        dispatchGlobals.numDSSRTStacks = RayTracingHelper::getNumRtStacksPerDss(*this);
-        dispatchGlobals.maxBVHLevels = maxBvhLevels;
 
+        auto rtStacksPerDss = RayTracingHelper::getNumRtStacksPerDss(*this);
+        dispatchGlobals.numDSSRTStacks = rtStacksPerDss;
+        dispatchGlobals.maxBVHLevels = maxBvhLevels;
         uint32_t *dispatchGlobalsAsArray = reinterpret_cast<uint32_t *>(&dispatchGlobals);
         dispatchGlobalsAsArray[7] = 1;
 
+        if (releaseHelper) {
+            bool heaplessEnabled = this->getCompilerProductHelper().isHeaplessModeEnabled();
+            releaseHelper->adjustRTDispatchGlobals(static_cast<void *>(&dispatchGlobals), rtStacksPerDss, heaplessEnabled, maxBvhLevels);
+        }
+
         MemoryTransferHelper::transferMemoryToAllocation(productHelper.isBlitCopyRequiredForLocalMemory(this->getRootDeviceEnvironment(), *dispatchGlobalsArrayAllocation),
                                                          *this,
                                                          dispatchGlobalsArrayAllocation,

shared/source/release_helper/release_helper.h

@@ -57,6 +57,7 @@ class ReleaseHelper {
     virtual uint32_t getAdditionalFp16Caps() const = 0;
     virtual uint32_t getAdditionalExtraCaps() const = 0;
     virtual uint32_t getStackSizePerRay() const = 0;
+    virtual void adjustRTDispatchGlobals(void *rtDispatchGlobals, uint32_t rtStacksPerDss, bool heaplessEnabled, uint32_t maxBvhLevels) const = 0;
     virtual bool isLocalOnlyAllowed() const = 0;
     virtual bool isDummyBlitWaRequired() const = 0;
     virtual bool isDirectSubmissionLightSupported() const = 0;
@@ -100,6 +101,7 @@ class ReleaseHelperHw : public ReleaseHelper {
     uint32_t getAdditionalFp16Caps() const override;
     uint32_t getAdditionalExtraCaps() const override;
     uint32_t getStackSizePerRay() const override;
+    void adjustRTDispatchGlobals(void *rtDispatchGlobals, uint32_t rtStacksPerDss, bool heaplessEnabled, uint32_t maxBvhLevels) const override;
     bool isLocalOnlyAllowed() const override;
     bool isDummyBlitWaRequired() const override;
     bool isDirectSubmissionLightSupported() const override;

shared/source/release_helper/release_helper_base.inl

@@ -137,6 +137,10 @@ uint32_t ReleaseHelperHw<releaseType>::getStackSizePerRay() const {
     return 0u;
 }
 
+template <ReleaseType releaseType>
+void ReleaseHelperHw<releaseType>::adjustRTDispatchGlobals(void *rtDispatchGlobals, uint32_t rtStacksPerDss, bool heaplessEnabled, uint32_t maxBvhLevels) const {
+}
+
 template <ReleaseType releaseType>
 bool ReleaseHelperHw<releaseType>::isLocalOnlyAllowed() const {
     return true;

shared/test/common/mocks/mock_release_helper.h

@@ -42,6 +42,7 @@ class MockReleaseHelper : public ReleaseHelper {
     ADDMETHOD_CONST_NOBASE(getFtrXe2Compression, bool, false, ());
     ADDMETHOD_CONST_NOBASE(isDirectSubmissionLightSupported, bool, false, ());
     ADDMETHOD_CONST_NOBASE(computeSlmValues, uint32_t, {}, (uint32_t slmSize, bool isHeapless));
+    ADDMETHOD_CONST_NOBASE_VOIDRETURN(adjustRTDispatchGlobals, (void *rtDispatchGlobals, uint32_t rtStacksPerDss, bool heaplessEnabled, uint32_t maxBvhLevels));
 
     const SizeToPreferredSlmValueArray &getSizeToPreferredSlmValue(bool isHeapless) const override {
         static SizeToPreferredSlmValueArray sizeToPreferredSlmValue = {};

shared/test/common/test_files/igdrcl.config

@@ -663,6 +663,7 @@ PipelinedEuThreadArbitration = -1
 ExperimentalUSMAllocationReuseCleaner = -1
 DummyPageBackingEnabled = 0
 EnableDeferBacking = 0
+SetMaxBVHLevels = -1
 GetSipBinaryFromExternalLib = -1
 LogUsmReuse = 0
 # Please don't edit below this line

shared/test/unit_test/device/neo_device_tests.cpp

@@ -253,7 +253,7 @@ TEST_F(DeviceTest, whenAllocateRTDispatchGlobalsIsCalledThenStackSizePerRayIsSet
     pDevice->initializeRayTracing(5);
     pDevice->allocateRTDispatchGlobals(3);
     EXPECT_NE(nullptr, pDevice->getRTDispatchGlobals(3));
-    struct RTDispatchGlobals dispatchGlobals = *reinterpret_cast<struct RTDispatchGlobals *>(pDevice->getRTDispatchGlobals(3)->rtDispatchGlobalsArray->getUnderlyingBuffer());
+    RTDispatchGlobals dispatchGlobals = *reinterpret_cast<struct RTDispatchGlobals *>(pDevice->getRTDispatchGlobals(3)->rtDispatchGlobalsArray->getUnderlyingBuffer());
 
     auto releaseHelper = getReleaseHelper();
     if (releaseHelper) {