[VP] Add L0 Mode for KernelBinToSource

Enable level zero mode for KernelBinToSourceNativeForAdvKrnl
intel · Oct 30, 2023 · 25422ea · 25422ea
1 parent 0147a3a
commit 25422ea
Show file tree

Hide file tree

Showing 4 changed files with 125 additions and 1 deletion.
diff --git a/media_softlet/agnostic/common/vp/hal/features/vp_filter.h b/media_softlet/agnostic/common/vp/hal/features/vp_filter.h
@@ -329,6 +329,7 @@ using VEBOX_CSC_PARAMS      = _VEBOX_CSC_PARAMS;
 using PVEBOX_CSC_PARAMS     = VEBOX_CSC_PARAMS *;
 
 using KERNEL_ARGS = std::vector<KRN_ARG>;
+using KERNEL_BTIS = std::map<uint32_t,uint32_t>;
 
 struct _VEBOX_UPDATE_PARAMS
 {

diff --git a/media_softlet/agnostic/common/vp/hal/packet/vp_render_common.h b/media_softlet/agnostic/common/vp/hal/packet/vp_render_common.h
@@ -76,6 +76,28 @@ struct KRN_ARG
     KRN_ARG_KIND           eArgKind;
 };
 
+//for L0 use only
+struct KRN_BTI
+{
+    uint32_t               uIndex;
+    uint32_t               uBTI;
+};
+
+//for L0 use only
+struct KRN_EXECUTE_ENV
+{
+    uint32_t uBarrierCount;
+    bool     bDisableMidThreadPreemption;
+    uint32_t uGrfCount;
+    bool     bHasGlobalAtomics;
+    bool     bHasNoStatelessWrite;
+    uint32_t uInlineDataPayloadSize;
+    uint32_t uOffsetToSkipPerThreadDataLoad;
+    uint32_t uSimdSize;
+    uint32_t uSubgroupIndependentForwardProgress;
+    uint32_t uEuThreadCount;
+};
+
 using SurfaceIndex = uint32_t;
 using SamplerIndex = uint32_t;
 using KernelIndex  = uint32_t;              // index of current kernel in KERNEL_PARAMS_LIST

diff --git a/media_softlet/agnostic/common/vp/hal/platform_interface/vp_platform_interface.cpp b/media_softlet/agnostic/common/vp/hal/platform_interface/vp_platform_interface.cpp
@@ -674,3 +674,64 @@ void VpPlatformInterface::DisableRender()
     VP_PUBLIC_NORMALMESSAGE("Disable Render.");
     m_isRenderDisabled = true;
 }
+
+//for L0 use only
+MOS_STATUS VpRenderKernel::SetKernelExeEnv(KRN_EXECUTE_ENV &exeEnv)
+{
+    VP_FUNC_CALL();
+    m_kernelExeEnv = exeEnv;
+    return MOS_STATUS_SUCCESS;
+}
+
+//for L0 use only
+MOS_STATUS VpRenderKernel::AddKernelBti(KRN_BTI &bti)
+{
+    VP_FUNC_CALL();
+    m_kernelBtis.emplace(bti.uIndex, bti.uBTI);
+    return MOS_STATUS_SUCCESS;
+}
+
+//for L0 use only
+MOS_STATUS VpRenderKernel::SetKernelCurbeSize(uint32_t size)
+{
+    VP_FUNC_CALL();
+
+    m_curbeSize = size;
+    return MOS_STATUS_SUCCESS;
+}
+
+//for L0 use only
+void VpPlatformInterface::InitVpDelayedNativeAdvKernel(
+    const uint32_t  *kernelBin,
+    uint32_t         kernelBinSize,
+    KRN_ARG         *kernelArgs,
+    uint32_t         kernelArgSize,
+    uint32_t         kernelCurbeSize,
+    KRN_EXECUTE_ENV &kernelExeEnv,
+    KRN_BTI         *kernelBtis,
+    uint32_t         kernelBtiSize,
+    std::string      kernelName)
+{
+    VP_FUNC_CALL();
+
+    VpRenderKernel vpKernel;
+
+    vpKernel.SetKernelBinPointer((void *)kernelBin);
+    vpKernel.SetKernelName(kernelName);
+    vpKernel.SetKernelBinOffset(0x0);
+    vpKernel.SetKernelBinSize(kernelBinSize);
+    vpKernel.SetKernelExeEnv(kernelExeEnv);
+    vpKernel.SetKernelCurbeSize(kernelCurbeSize);
+
+    for (uint32_t i = 0; i < kernelArgSize; ++i)
+    {
+        vpKernel.AddKernelArg(kernelArgs[i]);
+    }
+
+    for (uint32_t i = 0; i < kernelBtiSize; ++i)
+    {
+        vpKernel.AddKernelBti(kernelBtis[i]);
+    }
+
+    m_kernelPool.insert(std::make_pair(vpKernel.GetKernelName(), vpKernel));
+}
diff --git a/media_softlet/agnostic/common/vp/hal/platform_interface/vp_platform_interface.h b/media_softlet/agnostic/common/vp/hal/platform_interface/vp_platform_interface.h
@@ -113,6 +113,29 @@ class VpRenderKernel
         return m_kernelArgs;
     }
 
+
+    //for L0 use only
+    uint32_t &GetCurbeSize()
+    {
+        return m_curbeSize;
+    }
+
+    MOS_STATUS SetKernelExeEnv(KRN_EXECUTE_ENV &exeEnv);
+
+    MOS_STATUS SetKernelCurbeSize(uint32_t size);
+
+    MOS_STATUS AddKernelBti(KRN_BTI &bti);
+
+    KERNEL_BTIS GetKernelBtis()
+    {
+        return m_kernelBtis;
+    }
+
+    KRN_EXECUTE_ENV &GetKernelExeEnv()
+    {
+        return m_kernelExeEnv;
+    }
+
 protected:
     // Compositing Kernel DLL/Search state
     const Kdll_RuleEntry        *m_kernelDllRules = nullptr;
@@ -122,13 +145,18 @@ class VpRenderKernel
     uint32_t                    m_kernelBinSize = 0;
     // CM Kernel Execution Code Offset
     uint32_t                    m_kernelBinOffset = 0;
-    // CM Kernel Arguments
+    // CM Kernel Arguments or L0 Kernel Arguments
     KERNEL_ARGS                 m_kernelArgs;
     std::string                 m_kernelName = {};
     // CM Compositing Kernel patch file buffer and size
     const void                  *m_fcPatchBin = nullptr;
     uint32_t                    m_fcPatchBinSize = 0;
 
+    //for L0 use only
+    uint32_t m_curbeSize = 0;
+    KERNEL_BTIS     m_kernelBtis;
+    KRN_EXECUTE_ENV m_kernelExeEnv = {};
+
 public:
     const static std::string          s_kernelNameNonAdvKernels;
 
@@ -320,6 +348,18 @@ class VpPlatformInterface
         uint32_t        kernelBinSize,
         std::string     kernelName);
 
+    //for L0 kernel use only
+    virtual void InitVpDelayedNativeAdvKernel(
+        const uint32_t  *kernelBin,
+        uint32_t         kernelBinSize,
+        KRN_ARG         *kernelArgs,
+        uint32_t         kernelArgSize,
+        uint32_t         kernelCurbeSize,
+        KRN_EXECUTE_ENV& kernelExeEnv,
+        KRN_BTI         *kernelBtis,
+        uint32_t         kernelBtiSize,
+        std::string      kernelName);
+
     virtual void AddNativeAdvKernelToDelayedList(
         DelayLoadedKernelType kernelType,
         DelayLoadedFunc       func);