[VP] fix pipecontrol issue

fix pipecontrol issue.
intel · Aug 30, 2023 · 1c13afa · 1c13afa
1 parent 5444c83
commit 1c13afa
Show file tree

Hide file tree

Showing 12 changed files with 22 additions and 10 deletions.
diff --git a/media_common/agnostic/common/hw/mhw_mi.h b/media_common/agnostic/common/hw/mhw_mi.h
@@ -179,6 +179,7 @@ typedef struct _MHW_PIPE_CONTROL_PARAMS
     uint32_t                bInvalidateTextureCache        : 1;
     uint32_t                bGenericMediaStateClear        : 1;
     uint32_t                bIndirectStatePointersDisable  : 1;
+    uint32_t                bUnTypedDataPortCacheFlush     : 1;
     uint32_t                bHdcPipelineFlush              : 1;
     uint32_t                bKernelFenceEnabled            : 1;
     uint32_t                bPPCFlush                      : 1;

diff --git a/media_softlet/agnostic/Xe_M_plus/Xe_LPM_plus_base/hw/mhw_mi_xe_lpm_plus_base_next_impl.h b/media_softlet/agnostic/Xe_M_plus/Xe_LPM_plus_base/hw/mhw_mi_xe_lpm_plus_base_next_impl.h
@@ -662,7 +662,8 @@ class Impl : public mi::Impl<Cmd>
                 cmd.DW1.StateCacheInvalidationEnable    = params.bInvalidateStateCache;
                 cmd.DW1.ConstantCacheInvalidationEnable = params.bInvalidateConstantCache;
             }
-
+            cmd.DW0.UnTypedDataPortCacheFlush        = params.bUnTypedDataPortCacheFlush;
+            cmd.DW0.HdcPipelineFlush                 = params.bHdcPipelineFlush;
             cmd.DW1.RenderTargetCacheFlushEnable     = params.bFlushRenderTargetCache;
             cmd.DW1.DcFlushEnable                    = params.bFlushRenderTargetCache;  // same as above
             cmd.DW1.VfCacheInvalidationEnable        = params.bInvalidateVFECache;

diff --git a/media_softlet/agnostic/common/hw/mhw_mi_cmdpar.h b/media_softlet/agnostic/common/hw/mhw_mi_cmdpar.h
@@ -188,6 +188,7 @@ namespace mi
         bool                    bInvalidateTextureCache       = false;
         bool                    bGenericMediaStateClear       = false;
         bool                    bIndirectStatePointersDisable = false;
+        bool                    bUnTypedDataPortCacheFlush    = false;
         bool                    bHdcPipelineFlush             = false;
         bool                    bKernelFenceEnabled           = false;
         bool                    bPPCFlush                     = false;

diff --git a/media_softlet/agnostic/common/renderhal/renderhal_platform_interface_next.cpp b/media_softlet/agnostic/common/renderhal/renderhal_platform_interface_next.cpp
@@ -504,6 +504,7 @@ MOS_STATUS XRenderHal_Platform_Interface_Next::AddMiPipeControl(
     par.bInvalidateTextureCache = params->bInvalidateTextureCache;
     par.bGenericMediaStateClear = params->bGenericMediaStateClear;
     par.bIndirectStatePointersDisable = params->bIndirectStatePointersDisable;
+    par.bUnTypedDataPortCacheFlush = params->bUnTypedDataPortCacheFlush;
     par.bHdcPipelineFlush = params->bHdcPipelineFlush;
     par.bKernelFenceEnabled = params->bKernelFenceEnabled;
     MHW_CHK_STATUS_RETURN(m_miItf->MHW_ADDCMD_F(PIPE_CONTROL)(pCmdBuffer));

diff --git a/media_softlet/agnostic/common/shared/packet/media_render_cmd_packet.h b/media_softlet/agnostic/common/shared/packet/media_render_cmd_packet.h
@@ -88,6 +88,7 @@ typedef struct _KERNEL_WALKER_PARAMS
     RECT                                alignedRect;
     bool                                isVerticalPattern;
     bool                                bSyncFlag;
+    bool                                bFlushL1;
     bool                                isGroupStartInvolvedInGroupSize;    // true if group start need be involved in the group size.
     bool                                calculateBlockXYByAlignedRect;      // true if iBlocksX/iBlocksY is calculated by alignedRect in RenderCmdPacket instead of kernel object.
     bool                                forcePreferredSLMZero;              // true if preferredSLM need force to 0.

diff --git a/media_softlet/agnostic/common/vp/hal/packet/vp_render_cmd_packet.cpp b/media_softlet/agnostic/common/vp/hal/packet/vp_render_cmd_packet.cpp
@@ -1806,7 +1806,7 @@ MOS_STATUS VpRenderCmdPacket::SendMediaStates(
     MHW_MI_LOAD_REGISTER_IMM_PARAMS loadRegisterImmParams = {};
     PMHW_MI_MMIOREGISTERS           pMmioRegisters        = nullptr;
     MOS_OCA_BUFFER_HANDLE           hOcaBuf               = 0;
-
+    bool                            flushL1               = false;
     //---------------------------------------
     MHW_RENDERHAL_CHK_NULL(pRenderHal);
     MHW_RENDERHAL_CHK_NULL(pRenderHal->pStateHeap);
@@ -1901,6 +1901,11 @@ MOS_STATUS VpRenderCmdPacket::SendMediaStates(
             pipeCtlParams.dwFlushMode             = MHW_FLUSH_CUSTOM;
             pipeCtlParams.bInvalidateTextureCache = true;
             pipeCtlParams.bFlushRenderTargetCache = true;
+            if (flushL1)
+            {   //Flush L1 cache after consumer walker when there is a producer-consumer relationship walker.
+                pipeCtlParams.bUnTypedDataPortCacheFlush = true;
+                pipeCtlParams.bHdcPipelineFlush          = true;
+            }
             MHW_RENDERHAL_CHK_STATUS(pRenderHal->pRenderHalPltInterface->AddMiPipeControl(pRenderHal,
                 pCmdBuffer,
                 &pipeCtlParams));
@@ -1941,7 +1946,8 @@ MOS_STATUS VpRenderCmdPacket::SendMediaStates(
                 pCmdBuffer,
                 &m_gpgpuWalkerParams));
 
-             PrintWalkerParas(m_mediaWalkerParams);
+            flushL1 = it->second.walkerParam.bFlushL1;
+            PrintWalkerParas(m_mediaWalkerParams);
         }
         else
         {

diff --git a/media_softlet/agnostic/common/vp/hal/packet/vp_render_kernel_obj.cpp b/media_softlet/agnostic/common/vp/hal/packet/vp_render_kernel_obj.cpp
@@ -107,7 +107,7 @@ void VpRenderKernelObj::OcaDumpKernelInfo(MOS_COMMAND_BUFFER &cmdBuffer, MOS_CON
 }
 
 // Only for Adv kernels.
-MOS_STATUS VpRenderKernelObj::SetWalkerSetting(KERNEL_THREAD_SPACE& threadSpace, bool bSyncFlag)
+MOS_STATUS VpRenderKernelObj::SetWalkerSetting(KERNEL_THREAD_SPACE &threadSpace, bool bSyncFlag, bool flushL1)
 {
     VP_FUNC_CALL();
     VP_RENDER_CHK_STATUS_RETURN(MOS_STATUS_UNIMPLEMENTED);
@@ -139,7 +139,7 @@ MOS_STATUS VpRenderKernelObj::SetKernelConfigs(
 
     VP_RENDER_CHK_STATUS_RETURN(SetSamplerStates(samplerStateGroup));
 
-    VP_RENDER_CHK_STATUS_RETURN(SetWalkerSetting(kernelParams.kernelThreadSpace, kernelParams.syncFlag));
+    VP_RENDER_CHK_STATUS_RETURN(SetWalkerSetting(kernelParams.kernelThreadSpace, kernelParams.syncFlag,kernelParams.flushL1));
 
     return MOS_STATUS_SUCCESS;
 }

diff --git a/media_softlet/agnostic/common/vp/hal/packet/vp_render_kernel_obj.h b/media_softlet/agnostic/common/vp/hal/packet/vp_render_kernel_obj.h
@@ -86,6 +86,7 @@ typedef struct _KERNEL_PARAMS
     KERNEL_ARGS          kernelArgs;
     KERNEL_THREAD_SPACE  kernelThreadSpace;
     bool                 syncFlag;
+    bool                 flushL1;
 } KERNEL_PARAMS;
 
 struct MEDIA_OBJECT_KA2_INLINE_DATA
@@ -502,7 +503,7 @@ class VpRenderKernelObj
 
 protected:
 
-    virtual MOS_STATUS SetWalkerSetting(KERNEL_THREAD_SPACE& threadSpace, bool bSyncFlag);
+    virtual MOS_STATUS SetWalkerSetting(KERNEL_THREAD_SPACE &threadSpace, bool bSyncFlag, bool flushL1 = false);
 
     virtual MOS_STATUS SetKernelArgs(KERNEL_ARGS &kernelArgs, VP_PACKET_SHARED_CONTEXT *sharedContext);
 

diff --git a/media_softlet/agnostic/common/vp/hal/packet/vp_render_vebox_hdr_3dlut_kernel.cpp b/media_softlet/agnostic/common/vp/hal/packet/vp_render_vebox_hdr_3dlut_kernel.cpp
@@ -461,7 +461,7 @@ MOS_STATUS VpRenderHdr3DLutKernelCM::GetWalkerSetting(KERNEL_WALKER_PARAMS &walk
 }
 
 // Only for Adv kernels.
-MOS_STATUS VpRenderHdr3DLutKernelCM::SetWalkerSetting(KERNEL_THREAD_SPACE &threadSpace, bool bSyncFlag)
+MOS_STATUS VpRenderHdr3DLutKernelCM::SetWalkerSetting(KERNEL_THREAD_SPACE &threadSpace, bool bSyncFlag, bool flushL1)
 {
     VP_FUNC_CALL();
     MOS_ZeroMemory(&m_walkerParam, sizeof(KERNEL_WALKER_PARAMS));

diff --git a/media_softlet/agnostic/common/vp/hal/packet/vp_render_vebox_hdr_3dlut_kernel.h b/media_softlet/agnostic/common/vp/hal/packet/vp_render_vebox_hdr_3dlut_kernel.h
@@ -291,7 +291,7 @@ class VpRenderHdr3DLutKernelCM : public VpRenderHdr3DLutKernel
 protected:
     virtual MOS_STATUS SetupSurfaceState() override;
     virtual MOS_STATUS CpPrepareResources() override;
-    virtual MOS_STATUS SetWalkerSetting(KERNEL_THREAD_SPACE &threadSpace, bool bSyncFlag) override;
+    virtual MOS_STATUS SetWalkerSetting(KERNEL_THREAD_SPACE &threadSpace, bool bSyncFlag, bool flushL1 = false) override;
     virtual MOS_STATUS SetKernelArgs(KERNEL_ARGS &kernelArgs, VP_PACKET_SHARED_CONTEXT *sharedContext) override;
     virtual MOS_STATUS SetKernelConfigs(KERNEL_CONFIGS &kernelConfigs) override;
 

diff --git a/media_softlet/agnostic/common/vp/hal/packet/vp_render_vebox_hvs_kernel.cpp b/media_softlet/agnostic/common/vp/hal/packet/vp_render_vebox_hvs_kernel.cpp
@@ -71,7 +71,7 @@ MOS_STATUS VpRenderHVSKernel::GetWalkerSetting(KERNEL_WALKER_PARAMS &walkerParam
 }
 
 // Only for Adv kernels.
-MOS_STATUS VpRenderHVSKernel::SetWalkerSetting(KERNEL_THREAD_SPACE &threadSpace, bool bSyncFlag)
+MOS_STATUS VpRenderHVSKernel::SetWalkerSetting(KERNEL_THREAD_SPACE &threadSpace, bool bSyncFlag, bool flushL1)
 {
     VP_FUNC_CALL();
     MOS_ZeroMemory(&m_walkerParam, sizeof(KERNEL_WALKER_PARAMS));

diff --git a/media_softlet/agnostic/common/vp/hal/packet/vp_render_vebox_hvs_kernel.h b/media_softlet/agnostic/common/vp/hal/packet/vp_render_vebox_hvs_kernel.h
@@ -227,7 +227,7 @@ class VpRenderHVSKernel : public VpRenderKernelObj
 
 protected:
     virtual MOS_STATUS SetupSurfaceState() override;
-    virtual MOS_STATUS SetWalkerSetting(KERNEL_THREAD_SPACE &threadSpace, bool bSyncFlag) override;
+    virtual MOS_STATUS SetWalkerSetting(KERNEL_THREAD_SPACE &threadSpace, bool bSyncFlag, bool flushL1 = false) override;
     virtual MOS_STATUS SetKernelArgs(KERNEL_ARGS &kernelArgs, VP_PACKET_SHARED_CONTEXT *sharedContext) override;
     virtual MOS_STATUS SetKernelConfigs(KERNEL_CONFIGS &kernelConfigs) override;