feature: initial support for patching region params

Related-To: NEO-8070

Signed-off-by: Dunajski, Bartosz <bartosz.dunajski@intel.com>

level_zero/core/source/cmdlist/CMakeLists.txt

@@ -16,6 +16,7 @@ target_sources(${L0_STATIC_LIB_NAME}
                ${CMAKE_CURRENT_SOURCE_DIR}/cmdlist_imp.h
                ${CMAKE_CURRENT_SOURCE_DIR}/cmdlist_hw_immediate.h
                ${CMAKE_CURRENT_SOURCE_DIR}/cmdlist_hw_immediate.inl
+               ${CMAKE_CURRENT_SOURCE_DIR}/cmdlist_launch_params.h
                ${CMAKE_CURRENT_SOURCE_DIR}/cmdlist_extended${BRANCH_DIR_SUFFIX}cmdlist_extended.inl
 )
 

level_zero/core/source/cmdlist/cmdlist.h

@@ -18,6 +18,7 @@
 #include "shared/source/unified_memory/unified_memory.h"
 #include "shared/source/utilities/stackvec.h"
 
+#include "level_zero/core/source/cmdlist/cmdlist_launch_params.h"
 #include <level_zero/ze_api.h>
 #include <level_zero/zet_api.h>
 
@@ -34,23 +35,6 @@ struct Event;
 struct Kernel;
 struct CommandQueue;
 
-struct CmdListKernelLaunchParams {
-    NEO::RequiredPartitionDim requiredPartitionDim = NEO::RequiredPartitionDim::none;
-    NEO::RequiredDispatchWalkOrder requiredDispatchWalkOrder = NEO::RequiredDispatchWalkOrder::none;
-    uint32_t additionalSizeParam = NEO::additionalKernelLaunchSizeParamNotSet;
-    uint32_t numKernelsInSplitLaunch = 0;
-    uint32_t numKernelsExecutedInSplitLaunch = 0;
-    bool isIndirect = false;
-    bool isPredicate = false;
-    bool isCooperative = false;
-    bool isKernelSplitOperation = false;
-    bool isBuiltInKernel = false;
-    bool isDestinationAllocationInSystemMemory = false;
-    bool isHostSignalScopeEvent = false;
-    bool skipInOrderNonWalkerSignaling = false;
-    bool pipeControlSignalling = false;
-};
-
 struct CmdListReturnPoint {
     NEO::StreamProperties configSnapshot;
     uint64_t gpuAddress = 0;

level_zero/core/source/cmdlist/cmdlist_hw_xehp_and_later.inl

@@ -163,6 +163,7 @@ ze_result_t CommandListCoreFamily<gfxCoreFamily>::appendLaunchKernelWithParams(K
     auto kernelPreemptionMode = obtainKernelPreemptionMode(kernel);
 
     kernel->patchGlobalOffset();
+    kernel->patchRegionParams(launchParams);
     this->allocateOrReuseKernelPrivateMemoryIfNeeded(kernel, kernelDescriptor.kernelAttributes.perHwThreadPrivateMemorySize);
 
     if (launchParams.isIndirect) {

level_zero/core/source/cmdlist/cmdlist_launch_params.h

@@ -0,0 +1,31 @@
+/*
+ * Copyright (C) 2023 Intel Corporation
+ *
+ * SPDX-License-Identifier: MIT
+ *
+ */
+
+#pragma once
+
+#include "shared/source/helpers/definitions/command_encoder_args.h"
+
+#include <cstdint>
+
+namespace L0 {
+struct CmdListKernelLaunchParams {
+    NEO::RequiredPartitionDim requiredPartitionDim = NEO::RequiredPartitionDim::none;
+    NEO::RequiredDispatchWalkOrder requiredDispatchWalkOrder = NEO::RequiredDispatchWalkOrder::none;
+    uint32_t additionalSizeParam = NEO::additionalKernelLaunchSizeParamNotSet;
+    uint32_t numKernelsInSplitLaunch = 0;
+    uint32_t numKernelsExecutedInSplitLaunch = 0;
+    bool isIndirect = false;
+    bool isPredicate = false;
+    bool isCooperative = false;
+    bool isKernelSplitOperation = false;
+    bool isBuiltInKernel = false;
+    bool isDestinationAllocationInSystemMemory = false;
+    bool isHostSignalScopeEvent = false;
+    bool skipInOrderNonWalkerSignaling = false;
+    bool pipeControlSignalling = false;
+};
+} // namespace L0

level_zero/core/source/kernel/kernel.h

@@ -30,6 +30,7 @@ class MemoryManager;
 namespace L0 {
 struct Device;
 struct Module;
+struct CmdListKernelLaunchParams;
 
 struct KernelImmutableData {
     KernelImmutableData(L0::Device *l0device = nullptr);
@@ -137,6 +138,7 @@ struct Kernel : _ze_kernel_handle_t, virtual NEO::DispatchKernelEncoderI {
     virtual uint32_t *getGlobalOffsets() = 0;
     virtual ze_result_t setGlobalOffsetExp(uint32_t offsetX, uint32_t offsetY, uint32_t offsetZ) = 0;
     virtual void patchGlobalOffset() = 0;
+    virtual void patchRegionParams(const CmdListKernelLaunchParams &launchParams) = 0;
 
     virtual ze_result_t suggestMaxCooperativeGroupCount(uint32_t *totalGroupCount, NEO::EngineGroupType engineGroupType,
                                                         bool isEngineInstanced) = 0;

level_zero/core/source/kernel/kernel_imp.h

@@ -58,6 +58,7 @@ struct KernelImp : Kernel {
     ze_result_t setArgumentValue(uint32_t argIndex, size_t argSize, const void *pArgValue) override;
 
     void setGroupCount(uint32_t groupCountX, uint32_t groupCountY, uint32_t groupCountZ) override;
+    void patchRegionParams(const CmdListKernelLaunchParams &launchParams) override;
 
     ze_result_t setGroupSize(uint32_t groupSizeX, uint32_t groupSizeY,
                              uint32_t groupSizeZ) override;

level_zero/core/source/kernel/kernel_imp_helper.cpp

@@ -5,6 +5,7 @@
  *
  */
 
+#include "level_zero/core/source/cmdlist/cmdlist_launch_params.h"
 #include "level_zero/core/source/kernel/kernel_imp.h"
 
 namespace L0 {
@@ -13,4 +14,6 @@ KernelExt *KernelImp::getExtension(uint32_t extensionType) { return nullptr; }
 
 void KernelImp::getExtendedKernelProperties(ze_base_desc_t *pExtendedProperties) {}
 
+void KernelImp::patchRegionParams(const CmdListKernelLaunchParams &launchParams) {}
+
 } // namespace L0

level_zero/core/test/unit_tests/fixtures/module_fixture.h

@@ -88,6 +88,7 @@ struct ModuleImmutableDataFixture : public DeviceFixture {
         using KernelImp::kernelRequiresUncachedMocsCount;
         using KernelImp::midThreadPreemptionDisallowedForRayTracingKernels;
         using KernelImp::patchBindlessOffsetsInCrossThreadData;
+        using KernelImp::pImplicitArgs;
         using KernelImp::printfBuffer;
         using KernelImp::privateMemoryGraphicsAllocation;
         using KernelImp::requiredWorkgroupOrder;