[Encode] Wait HW semaphore before CSC

In remote gaming usage, when input is RGB surface, insert MI_SEMAPHORE_WAIT to wait for external copy to be done before CSC. It can help to remove the latency gap between external copy and CSC. This change is for AVC VDEnc only, and it is off by default for now. Change-Id: I2070816e5408d75983a2638784ceca416c10f601
intel · May 16, 2019 · 7c0258c · 7c0258c
1 parent 676951b
commit 7c0258c
Show file tree

Hide file tree

Showing 6 changed files with 53 additions and 1 deletion.
diff --git a/media_driver/agnostic/common/codec/hal/codechal_encode_csc_ds.cpp b/media_driver/agnostic/common/codec/hal/codechal_encode_csc_ds.cpp
@@ -1455,12 +1455,42 @@ MOS_STATUS CodechalEncodeCscDs::CscKernel(
         walkerParams.GroupIdLoopSelect = m_groupId;
     }
 
+    // In remote gaming scenario, when input is RGB surface, insert HW semaphore to wait for 
+    // external copy completion before CSC. Once the marker is overwritten by external copy,
+    // HW semaphore will be signalled and CSC will start. 
+    bool isRgbInput = (m_surfaceParamsCsc.psInputSurface->Format == Format_A8R8G8B8) ||
+                      (m_surfaceParamsCsc.psInputSurface->Format == Format_A8B8G8R8) ||
+                      (m_surfaceParamsCsc.psInputSurface->Format == Format_X8R8G8B8) ||
+                      (m_surfaceParamsCsc.psInputSurface->Format == Format_X8B8G8R8);
+    uint32_t offset = m_surfaceParamsCsc.psInputSurface->dwWidth * (m_surfaceParamsCsc.psInputSurface->dwHeight - 16) * 4 - 4;
+    if (m_externalCopySync && isRgbInput)
+    {
+        MHW_MI_SEMAPHORE_WAIT_PARAMS miSemaphoreWaitParams;
+        MOS_ZeroMemory((&miSemaphoreWaitParams), sizeof(miSemaphoreWaitParams));
+        miSemaphoreWaitParams.presSemaphoreMem = &m_surfaceParamsCsc.psInputSurface->OsResource;
+        miSemaphoreWaitParams.dwResourceOffset = offset;
+        miSemaphoreWaitParams.bPollingWaitMode = true;
+        miSemaphoreWaitParams.dwSemaphoreData  = 0x01234501; // the marker to check in source surface
+        miSemaphoreWaitParams.CompareOperation = MHW_MI_SAD_NOT_EQUAL_SDD;
+        CODECHAL_ENCODE_CHK_STATUS_RETURN(m_miInterface->AddMiSemaphoreWaitCmd(&cmdBuffer, &miSemaphoreWaitParams));
+    }
+
     CODECHAL_ENCODE_CHK_STATUS_RETURN(m_renderInterface->AddMediaObjectWalkerCmd(&cmdBuffer, &walkerParams));
 
     CODECHAL_ENCODE_CHK_STATUS_RETURN(m_encoder->EndStatusReport(&cmdBuffer, encFunctionType));
 
     CODECHAL_ENCODE_CHK_STATUS_RETURN(m_stateHeapInterface->SubmitBlocks(m_cscKernelState));
 
+    // write the marker to source surface for next MI_SEMAPHORE_WAIT to check.
+    if (m_externalCopySync && isRgbInput)
+    {
+        MHW_MI_STORE_DATA_PARAMS storeDataParams;
+        storeDataParams.pOsResource      = &m_surfaceParamsCsc.psInputSurface->OsResource;
+        storeDataParams.dwResourceOffset = offset;
+        storeDataParams.dwValue          = 0x01234501; // write the marker to source surface 
+        CODECHAL_ENCODE_CHK_STATUS_RETURN(m_miInterface->AddMiStoreDataImmCmd(&cmdBuffer, &storeDataParams));
+    }
+
     if (!m_singleTaskPhaseSupported || m_lastTaskInPhase)
     {
         CODECHAL_ENCODE_CHK_STATUS_RETURN(m_stateHeapInterface->UpdateGlobalCmdBufId());
@@ -1798,7 +1828,8 @@ CodechalEncodeCscDs::CodechalEncodeCscDs(CodechalEncoderState *encoder)
       m_currRefList(encoder->m_currRefList),
       m_resMbStatsBuffer(encoder->m_resMbStatsBuffer),
       m_rawSurfaceToEnc(encoder->m_rawSurfaceToEnc),
-      m_rawSurfaceToPak(encoder->m_rawSurfaceToPak)
+      m_rawSurfaceToPak(encoder->m_rawSurfaceToPak),
+      m_externalCopySync(encoder->m_externalCopySync)
 {
     // Initilize interface pointers
     m_encoder = encoder;

diff --git a/media_driver/agnostic/common/codec/hal/codechal_encode_csc_ds.h b/media_driver/agnostic/common/codec/hal/codechal_encode_csc_ds.h
@@ -652,6 +652,7 @@ class CodechalEncodeCscDs
     bool&                       m_singleTaskPhaseSupported;
     bool&                       m_firstTaskInPhase;
     bool&                       m_lastTaskInPhase;
+    bool&                       m_externalCopySync;
     uint8_t&                    m_groupId;
     uint8_t&                    m_outputChromaFormat;
     uint32_t&                   m_standard;

diff --git a/media_driver/agnostic/common/codec/hal/codechal_encoder_base.h b/media_driver/agnostic/common/codec/hal/codechal_encoder_base.h
@@ -1378,6 +1378,7 @@ class CodechalEncoderState : public Codechal
     bool                            m_2xScalingEnabled = false;   //!< 2x Scaling kernel only used by HEVC now
     bool                            m_useRawForRef = false;       //!< Flag to indicate if using raw surface for reference
     bool                            m_disableReconMMCD = false;   //!< disable Recon surface's MMC
+    bool                            m_externalCopySync  = false;  //!< Flag to indicate if GPU polling based sync for raw surface copy is enabled
     uint8_t                         m_prevReconFrameIdx = 0;      //!< Previous reconstruct frame index
     uint8_t                         m_currReconFrameIdx = 0;      //!< Current reconstruct frame index
 

diff --git a/media_driver/agnostic/common/codec/hal/codechal_vdenc_avc.cpp b/media_driver/agnostic/common/codec/hal/codechal_vdenc_avc.cpp
@@ -1186,6 +1186,13 @@ MOS_STATUS CodechalVdencAvcState::Initialize(CodechalSetting * settings)
             &userFeatureData);
         m_tlbPrefetchEnable = userFeatureData.bData == 1;
 
+        MOS_ZeroMemory(&userFeatureData, sizeof(userFeatureData));
+        MOS_UserFeature_ReadValue_ID(
+            nullptr,
+            __MEDIA_USER_FEATURE_VALUE_EXTERNAL_COPY_SYNC_ID,
+            &userFeatureData);
+        m_externalCopySync = userFeatureData.bData == 1;
+
         MOS_ZeroMemory(&userFeatureData, sizeof(userFeatureData));
         MOS_UserFeature_ReadValue_ID(
             nullptr,
@@ -3155,6 +3162,8 @@ MOS_STATUS CodechalVdencAvcState::ExecuteKernelFunctions()
 
         m_firstTaskInPhase = true;
 
+    m_externalCopySync &= (m_avcSeqParam->ScenarioInfo == ESCENARIO_REMOTEGAMING);
+
     if (m_cscDsState->UseSfc() && m_cscDsState->RequireCsc())
     {
         CODECHAL_ENCODE_CHK_STATUS_RETURN(m_cscDsState->CscUsingSfc(m_avcSeqParam->InputColorSpace));

diff --git a/media_driver/agnostic/common/os/mos_utilities.c b/media_driver/agnostic/common/os/mos_utilities.c
@@ -3438,6 +3438,15 @@ static MOS_USER_FEATURE_VALUE MOSUserFeatureDescFields[__MOS_USER_FEATURE_KEY_MA
        MOS_USER_FEATURE_VALUE_TYPE_UINT32,
        "0",
        "Enable MOS to simualte random memory allocate fail. "),
+        MOS_DECLARE_UF_KEY_DBGONLY(__MEDIA_USER_FEATURE_VALUE_EXTERNAL_COPY_SYNC_ID,
+       "External Copy Sync",
+       __MEDIA_USER_FEATURE_SUBKEY_INTERNAL,
+       __MEDIA_USER_FEATURE_SUBKEY_INTERNAL,
+       "Encode",
+       MOS_USER_FEATURE_TYPE_USER,
+       MOS_USER_FEATURE_VALUE_TYPE_UINT32,
+       "0",
+       "Enable GPU polling based sync for external raw surface copy."),
 };
 
 #define MOS_NUM_USER_FEATURE_VALUES     (sizeof(MOSUserFeatureDescFields) / sizeof(MOSUserFeatureDescFields[0]))

diff --git a/media_driver/agnostic/common/os/mos_utilities.h b/media_driver/agnostic/common/os/mos_utilities.h
@@ -552,6 +552,7 @@ typedef enum _MOS_USER_FEATURE_VALUE_ID
     __MEDIA_USER_FEATURE_VALUE_CMD_PARS_FILES_DIRECORY_ID,
     __MEDIA_USER_FEATURE_VALUE_APOGEIOS_ENABLE_ID,
     __MEDIA_USER_FEATURE_VALUE_SIMULATE_RANDOM_ALLOC_MEMORY_FAIL_ID,
+    __MEDIA_USER_FEATURE_VALUE_EXTERNAL_COPY_SYNC_ID,
     __MOS_USER_FEATURE_KEY_MAX_ID,
 } MOS_USER_FEATURE_VALUE_ID;