[VP] Enable 64b Effinicient

Enable 64b Effinicient: (1) Add compute walker 2 and remove cfe and compute walker (2) Enable bindless mode
intel · Jul 8, 2024 · 32a5aa6 · 32a5aa6
1 parent 34429a3
commit 32a5aa6
Show file tree

Hide file tree

Showing 21 changed files with 643 additions and 229 deletions.
diff --git a/media_common/agnostic/common/hw/mhw_render.h b/media_common/agnostic/common/hw/mhw_render.h
@@ -77,6 +77,7 @@ typedef struct _MHW_RENDER_ENGINE_L3_CACHE_CONFIG
 
 typedef enum _MHW_RENDER_ENGINE_ADDRESS_SHIFT
 {
+    MHW_RENDER_ENGINE_KERNEL_POINTER_SHIFT = 6,
     MHW_RENDER_ENGINE_STATE_BASE_ADDRESS_SHIFT  = 12
 } MHW_RENDER_ENGINE_ADDRESS_SHIFT;
 
@@ -146,6 +147,7 @@ typedef struct _MHW_STATE_BASE_ADDR_PARAMS
     uint32_t                mocs4IndirectObjectBuffer;
     uint32_t                mocs4StatelessDataport;
     uint32_t                l1CacheConfig;
+    bool                    addressDis;
 } MHW_STATE_BASE_ADDR_PARAMS, *PMHW_STATE_BASE_ADDR_PARAMS;
 
 typedef struct _MHW_VFE_SCOREBOARD_DELTA
@@ -309,7 +311,8 @@ typedef struct _MHW_GPGPU_WALKER_PARAMS
     MHW_EMIT_LOCAL_MODE        emitLocal;
 
     bool                       hasBarrier;
-
+    PMHW_INLINE_DATA_PARAMS    inlineDataParamBase;
+
 } MHW_GPGPU_WALKER_PARAMS, *PMHW_GPGPU_WALKER_PARAMS;
 
 typedef struct _MHW_MEDIA_OBJECT_PARAMS

diff --git a/media_common/agnostic/common/hw/mhw_state_heap.h b/media_common/agnostic/common/hw/mhw_state_heap.h
@@ -83,6 +83,8 @@ typedef struct MHW_BLOCK_MANAGER *PMHW_BLOCK_MANAGER;
 
 #define MHW_INVALID_SYNC_TAG            0xFFFFFFFF
 
+#define MAX_INLINE_DATA_PARAMS          32
+
 enum MW_RENDER_ENGINE_ADDRESS_SHIFT
 {
     MHW_STATE_HEAP_SURFACE_STATE_SHIFT = 0
@@ -699,6 +701,14 @@ typedef enum _MHW_CHROMAKEY_MODE
     MHW_CHROMAKEY_MODE_REPLACE_BLACK = 1
 } MHW_CHROMAKEY_MODE;
 
+typedef struct _MHW_INLINE_DATA_PARAMS
+{
+    uint32_t          dwOffset;
+    uint32_t          dwSize;
+    PMOS_RESOURCE     resource;
+    bool              isPtrType;
+} MHW_INLINE_DATA_PARAMS, *PMHW_INLINE_DATA_PARAMS;
+
 typedef struct _MHW_SAMPLER_STATE_UNORM_PARAM
 {
     MHW_SAMPLER_FILTER_MODE      SamplerFilterMode;
@@ -845,6 +855,7 @@ struct MHW_STATE_HEAP_SETTINGS
     HeapManager::Behavior m_dshBehavior = HeapManager::Behavior::wait;    //!< DSH behavior
 
     uint32_t        dwNumSyncTags = 0; //!< to be removed with old interfaces
+    MOS_HW_RESOURCE_DEF m_heapUsageType = MOS_CODEC_RESOURCE_USAGE_BEGIN_CODEC;
 };
 
 typedef struct _MHW_STATE_HEAP_DYNAMIC_ALLOC_PARAMS

diff --git a/media_common/agnostic/common/os/mos_os_hw.h b/media_common/agnostic/common/os/mos_os_hw.h
@@ -293,6 +293,7 @@ typedef enum MOS_HW_COMMAND_ENUM
     MOS_SURFACE_STATE,
     MOS_SURFACE_STATE_ADV,
     MOS_STATE_BASE_ADDR,
+    MOS_COMPUTE_WALKER,
     MOS_VEBOX_STATE,
     MOS_VEBOX_DI_IECP,
     MOS_VEBOX_TILING_CONVERT,

diff --git a/media_common/agnostic/common/renderhal/renderhal.h b/media_common/agnostic/common/renderhal/renderhal.h
@@ -841,6 +841,7 @@ typedef struct _RENDERHAL_STATE_HEAP_SETTINGS
     int32_t             iSurfaceStates;                                         // Number of Surfaces per SSH
     int32_t             iSurfacesPerBT;                                         // Number of Surfaces per BT
     int32_t             iBTAlignment;                                           // BT Alignment size
+    MOS_HW_RESOURCE_DEF heapUsageType;
 } RENDERHAL_STATE_HEAP_SETTINGS, *PRENDERHAL_STATE_HEAP_SETTINGS;
 
 typedef struct _RENDERHAL_STATE_HEAP
@@ -1197,6 +1198,7 @@ typedef struct _RENDERHAL_INTERFACE
     bool                        bEnableGpgpuMidBatchPreEmption;                 // Middle Batch Buffer Preemption
     bool                        bEnableGpgpuMidThreadPreEmption;                // Middle Thread Preemption
     bool                        bComputeContextInUse;                           // Compute Context use for media
+    bool                        isBindlessHeapInUse;                            // Bindless Heap Mode use
 
     uint32_t                    dwMaskCrsThdConDataRdLn;                        // Unifies pfnSetupInterfaceDescriptor for g75,g8,...
     uint32_t                    dwMinNumberThreadsInGroup;                      // Unifies pfnSetupInterfaceDescriptor for g75,g8,...
@@ -1460,6 +1462,13 @@ typedef struct _RENDERHAL_INTERFACE
                 PMHW_SAMPLER_STATE_PARAM    pSamplerParams,
                 int32_t                     iSamplers);
 
+    MOS_STATUS (*pfnSetAndGetSamplerStates) (
+                PRENDERHAL_INTERFACE     pRenderHal,
+                int32_t                  iMediaID,
+                PMHW_SAMPLER_STATE_PARAM pSamplerParams,
+                int32_t                  iSamplers,
+                std::map<uint32_t, uint32_t> &samplerMap);
+
     int32_t (* pfnAllocateMediaID) (
                 PRENDERHAL_INTERFACE        pRenderHal,
                 int32_t                     iKernelAllocationID,

diff --git a/media_common/agnostic/common/renderhal/renderhal_platform_interface.h b/media_common/agnostic/common/renderhal/renderhal_platform_interface.h
@@ -565,6 +565,12 @@ class XRenderHal_Platform_Interface
         return MOS_STATUS_SUCCESS;
     };
 
+    virtual bool IsBindlessHeapInUse(
+        PRENDERHAL_INTERFACE pRenderHal)
+    {
+        return false;
+    }
+
     //! \brief    Allocates scratch space buffer.
     //! \details  On some new pltforms, a single scratch space buffer may be allocated and used for
     //!           all threads.

diff --git a/media_driver/media_softlet/agnostic/Xe_R/Xe_HPG/renderhal/renderhal_xe_hpg.cpp b/media_driver/media_softlet/agnostic/Xe_R/Xe_HPG/renderhal/renderhal_xe_hpg.cpp
@@ -67,7 +67,8 @@ extern const RENDERHAL_STATE_HEAP_SETTINGS g_cRenderHal_State_Heap_Settings_xe_h
     RENDERHAL_SSH_BINDING_TABLES,              //!< iBindingTables
     RENDERHAL_SSH_SURFACE_STATES,              //!< iSurfaceStates
     RENDERHAL_SSH_SURFACES_PER_BT,             //!< iSurfacesPerBT
-    RENDERHAL_SSH_BINDING_TABLE_ALIGN          //!< iBTAlignment
+    RENDERHAL_SSH_BINDING_TABLE_ALIGN,         //!< iBTAlignment
+    MOS_CODEC_RESOURCE_USAGE_BEGIN_CODEC       //!< heapUsageType
 };
 
 extern const RENDERHAL_ENLARGE_PARAMS g_cRenderHal_Enlarge_State_Heap_Settings_Adv_xe_hpg =

diff --git a/media_driver/media_softlet/agnostic/g12/g12_base/renderhal/renderhal_g12_base.cpp b/media_driver/media_softlet/agnostic/g12/g12_base/renderhal/renderhal_g12_base.cpp
@@ -71,7 +71,8 @@ extern const RENDERHAL_STATE_HEAP_SETTINGS g_cRenderHal_State_Heap_Settings_g12
     RENDERHAL_SSH_BINDING_TABLES,              //!< iBindingTables
     RENDERHAL_SSH_SURFACE_STATES,              //!< iSurfaceStates
     RENDERHAL_SSH_SURFACES_PER_BT,             //!< iSurfacesPerBT
-    RENDERHAL_SSH_BINDING_TABLE_ALIGN          //!< iBTAlignment
+    RENDERHAL_SSH_BINDING_TABLE_ALIGN,         //!< iBTAlignment
+    MOS_CODEC_RESOURCE_USAGE_BEGIN_CODEC       //!< heapUsageType
 };
 
 const uint32_t g_cLookup_RotationMode_g12[8] = 

diff --git a/media_softlet/agnostic/Xe_R/Xe2_HPG/renderhal/renderhal_xe2_hpg_next.cpp b/media_softlet/agnostic/Xe_R/Xe2_HPG/renderhal/renderhal_xe2_hpg_next.cpp
@@ -67,7 +67,8 @@ extern const RENDERHAL_STATE_HEAP_SETTINGS g_cRenderHal_State_Heap_Settings_xe2_
     RENDERHAL_SSH_BINDING_TABLES,      //!< iBindingTables
     RENDERHAL_SSH_SURFACE_STATES,      //!< iSurfaceStates
     RENDERHAL_SSH_SURFACES_PER_BT,     //!< iSurfacesPerBT
-    RENDERHAL_SSH_BINDING_TABLE_ALIGN  //!< iBTAlignment
+    RENDERHAL_SSH_BINDING_TABLE_ALIGN,  //!< iBTAlignment
+    MOS_CODEC_RESOURCE_USAGE_BEGIN_CODEC //!< heapUsageType
 };
 
 extern const RENDERHAL_ENLARGE_PARAMS g_cRenderHal_Enlarge_State_Heap_Settings_Adv_xe2_hpg =

diff --git a/media_softlet/agnostic/Xe_R/Xe_HPG_Base/renderhal/renderhal_xe_hpg_base.cpp b/media_softlet/agnostic/Xe_R/Xe_HPG_Base/renderhal/renderhal_xe_hpg_base.cpp
@@ -61,7 +61,8 @@ extern const RENDERHAL_STATE_HEAP_SETTINGS g_cRenderHal_State_Heap_Settings_hpg_
     RENDERHAL_SSH_BINDING_TABLES,              //!< iBindingTables
     RENDERHAL_SSH_SURFACE_STATES,              //!< iSurfaceStates
     RENDERHAL_SSH_SURFACES_PER_BT,             //!< iSurfacesPerBT
-    RENDERHAL_SSH_BINDING_TABLE_ALIGN          //!< iBTAlignment
+    RENDERHAL_SSH_BINDING_TABLE_ALIGN,         //!< iBTAlignment
+    MOS_CODEC_RESOURCE_USAGE_BEGIN_CODEC       //!< heapUsageType
 };
 
 extern const RENDERHAL_ENLARGE_PARAMS g_cRenderHal_Enlarge_State_Heap_Settings_Adv_hpg_base =

diff --git a/media_softlet/agnostic/common/hw/mhw_render_cmdpar.h b/media_softlet/agnostic/common/hw/mhw_render_cmdpar.h
@@ -61,6 +61,12 @@ struct MHW_VFE_SCOREBOARD
     MHW_VFE_SCOREBOARD_DELTA ScoreboardDelta[MHW_MAX_DEPENDENCY_COUNT] = {};
 };
 
+struct MHW_HEAPS_RESOURCE
+{
+    PMOS_RESOURCE          presInstructionBuffer = nullptr;
+    PMHW_INLINE_DATA_PARAMS inlineDataParamsBase   = nullptr;
+};
+
 enum MHW_VFE_SLICE_DISABLE
 {
     MHW_VFE_SLICE_ALL = 0,
@@ -219,6 +225,7 @@ struct _MHW_PAR_T(STATE_BASE_ADDRESS)
     uint32_t                mocs4IndirectObjectBuffer  = 0;
     uint32_t                mocs4StatelessDataport     = 0;
     uint32_t                l1CacheConfig              = 0;
+    bool                    addressDis                 = false;
 };
 
 struct _MHW_PAR_T(MEDIA_VFE_STATE)
@@ -391,7 +398,9 @@ struct _MHW_PAR_T(COMPUTE_WALKER)
     bool                      isGenerateLocalId             = false;
     MHW_EMIT_LOCAL_MODE       emitLocal                     = MHW_EMIT_LOCAL_NONE;
     uint32_t                  preferredSlmAllocationSize    = 0;
-
+    _MHW_PAR_T(CFE_STATE)     cfeState                      = {};
+    MHW_HEAPS_RESOURCE        heapsResource                 = {};
+
 };
 
 struct _MHW_PAR_T(STATE_COMPUTE_MODE)