[Encode] Adaptive TU Enable

* [Encode] adaptive TU for AVC/HEVC/AV1 adaptive tu
intel · Oct 14, 2024 · 1290bba · 1290bba
1 parent 4fe64bd
commit 1290bba
Show file tree

Hide file tree

Showing 14 changed files with 169 additions and 24 deletions.
diff --git a/media_common/agnostic/common/codec/shared/codec_def_encode_av1.h b/media_common/agnostic/common/codec/shared/codec_def_encode_av1.h
@@ -366,7 +366,7 @@ typedef struct _CODEC_AV1_ENCODE_PICTURE_PARAMS
     uint8_t         ref_frame_idx[7];       // [0..6]
     uint8_t         HierarchLevelPlus1;
     uint8_t         primary_ref_frame;      // [0..7]
-    uint8_t         Reserved8b3;
+    uint8_t         AdaptiveTUEnabled;
     uint8_t         Reserved8b4;
     uint8_t         order_hint;
 

diff --git a/media_common/agnostic/common/codec/shared/codec_def_encode_avc.h b/media_common/agnostic/common/codec/shared/codec_def_encode_avc.h
@@ -1050,6 +1050,8 @@ typedef struct _CODEC_AVC_ENCODE_PIC_PARAMS
     */
     uint8_t         QpModulationStrength;
 
+    uint8_t         AdaptiveTUEnabled;
+
     /*! \brief StatusReportEnable
     *
     *  Request features to be enabled at status report.

diff --git a/media_common/agnostic/common/codec/shared/codec_def_encode_hevc.h b/media_common/agnostic/common/codec/shared/codec_def_encode_hevc.h
@@ -978,6 +978,8 @@ typedef struct _CODEC_HEVC_ENCODE_PICTURE_PARAMS
         } fields;
         uint8_t value;
     } QualityInfoSupportFlags;
+
+    uint8_t AdaptiveTUEnabled;
 } CODEC_HEVC_ENCODE_PICTURE_PARAMS, *PCODEC_HEVC_ENCODE_PICTURE_PARAMS;
 
 /*! \brief Slice-level parameters of a compressed picture for HEVC encoding.

diff --git a/media_softlet/agnostic/common/codec/hal/enc/av1/features/encode_av1_brc.cpp b/media_softlet/agnostic/common/codec/hal/enc/av1/features/encode_av1_brc.cpp
@@ -85,6 +85,15 @@ namespace encode
             "Encode RateControl Method",
             m_rcMode,
             MediaUserSetting::Group::Sequence);
+
+        ENCODE_CHK_NULL_RETURN(m_basicFeature->m_av1PicParams);
+        MediaUserSetting::Value outValue;
+        ReadUserSetting(
+            m_userSettingPtr,
+            outValue,
+            "Adaptive TU Enable",
+            MediaUserSetting::Group::Sequence);
+        m_basicFeature->m_av1PicParams->AdaptiveTUEnabled |= outValue.Get<uint8_t>(); 
 #endif
         return MOS_STATUS_SUCCESS;
     }
@@ -211,6 +220,7 @@ namespace encode
         dmem->UPD_CurHeight = (uint16_t)m_basicFeature->m_oriFrameHeight;
         dmem->UPD_Asyn = 0;
         dmem->UPD_EnableAdaptiveRounding = (m_basicFeature->m_roundingMethod == RoundingMethod::adaptiveRounding);
+        dmem->UPD_AdaptiveTUEnabled = picParams->AdaptiveTUEnabled;
 
         if (seqParams->GopRefDist == 16 && m_rcMode == RATECONTROL_CQL)
             dmem->UPD_MaxBRCLevel = 4;

diff --git a/media_softlet/agnostic/common/codec/hal/enc/av1/packet/encode_av1_brc_update_packet.cpp b/media_softlet/agnostic/common/codec/hal/enc/av1/packet/encode_av1_brc_update_packet.cpp
@@ -115,11 +115,19 @@ namespace encode
             {
                 // VDEnc read batch buffer (input for HuC FW)
                 allocParamsForBufferLinear.dwBytes = MOS_ALIGN_CEIL(m_hwInterface->m_vdencReadBatchBufferSize, CODECHAL_PAGE_SIZE);
-                allocParamsForBufferLinear.pBufName = "VDENC Read Batch Buffer";
+                allocParamsForBufferLinear.pBufName = "VDENC Read Origin Batch Buffer";
                 allocParamsForBufferLinear.ResUsageType = MOS_HW_RESOURCE_USAGE_ENCODE_INTERNAL_WRITE;
                 allocatedbuffer                     = m_allocator->AllocateResource(allocParamsForBufferLinear, true);
                 ENCODE_CHK_NULL_RETURN(allocatedbuffer);
-                m_vdencReadBatchBuffer[k][i] = *allocatedbuffer;
+                m_vdencReadBatchBufferOrigin[k][i] = *allocatedbuffer;
+
+                // VDEnc read batch buffer (input for HuC FW)
+                allocParamsForBufferLinear.dwBytes      = MOS_ALIGN_CEIL(m_hwInterface->m_vdencReadBatchBufferSize, CODECHAL_PAGE_SIZE);
+                allocParamsForBufferLinear.pBufName     = "VDENC Read TU7 Batch Buffer";
+                allocParamsForBufferLinear.ResUsageType = MOS_HW_RESOURCE_USAGE_ENCODE_INTERNAL_WRITE;
+                allocatedbuffer                         = m_allocator->AllocateResource(allocParamsForBufferLinear, true);
+                ENCODE_CHK_NULL_RETURN(allocatedbuffer);
+                m_vdencReadBatchBufferTU7[k][i] = *allocatedbuffer;
 
                 // BRC update DMEM
                 allocParamsForBufferLinear.dwBytes = MOS_ALIGN_CEIL(m_vdencBrcUpdateDmemBufferSize, CODECHAL_CACHELINE_SIZE);
@@ -156,8 +164,19 @@ namespace encode
 
         ENCODE_CHK_NULL_RETURN(m_basicFeature);
         ENCODE_CHK_NULL_RETURN(m_basicFeature->m_recycleBuf);
+        ENCODE_CHK_NULL_RETURN(m_basicFeature->m_av1PicParams);
+        ENCODE_CHK_NULL_RETURN(m_basicFeature->m_av1SeqParams);
+
+        ENCODE_CHK_STATUS_RETURN(ConstructBatchBufferHuCBRC(&m_vdencReadBatchBufferOrigin[m_pipeline->m_currRecycledBufIdx][m_pipeline->GetCurrentPass()]));
+
+        if (m_basicFeature->m_av1PicParams->AdaptiveTUEnabled != 0)
+        {
+            auto original_TU              = m_basicFeature->m_targetUsage;
+            m_basicFeature->m_targetUsage = m_basicFeature->m_av1SeqParams->TargetUsage = 7;
+            ENCODE_CHK_STATUS_RETURN(ConstructBatchBufferHuCBRC(&m_vdencReadBatchBufferTU7[m_pipeline->m_currRecycledBufIdx][m_pipeline->GetCurrentPass()]));
+            m_basicFeature->m_targetUsage = m_basicFeature->m_av1SeqParams->TargetUsage = original_TU;
+        }        
 
-        ENCODE_CHK_STATUS_RETURN(ConstructBatchBufferHuCBRC(&m_vdencReadBatchBuffer[m_pipeline->m_currRecycledBufIdx][m_pipeline->GetCurrentPass()]));
         ENCODE_CHK_STATUS_RETURN(ConstructPakInsertHucBRC(&m_vdencPakInsertBatchBuffer[m_pipeline->m_currRecycledBufIdx]));
 
         bool firstTaskInPhase = packetPhase & firstPacket;
@@ -502,11 +521,15 @@ namespace encode
 
         ENCODE_CHK_STATUS_RETURN(DumpRegion(0, "_BrcHistory", true, hucRegionDumpUpdate, 6080));
         ENCODE_CHK_STATUS_RETURN(DumpRegion(1, "_VdencStats", true, hucRegionDumpUpdate, 48*4));
-        ENCODE_CHK_STATUS_RETURN(DumpRegion(3, "_InputSLBB", true, hucRegionDumpUpdate, 600*4));
+        ENCODE_CHK_STATUS_RETURN(DumpRegion(3, "_InputSLBB_Origin", true, hucRegionDumpUpdate, 600*4));
         ENCODE_CHK_STATUS_RETURN(DumpRegion(5, "_ConstData", true, hucRegionDumpUpdate, MOS_ALIGN_CEIL(m_vdencBrcConstDataBufferSize, CODECHAL_PAGE_SIZE)));
         ENCODE_CHK_STATUS_RETURN(DumpRegion(7, "_PakMmio", true, hucRegionDumpUpdate, 16*4));
         ENCODE_CHK_STATUS_RETURN(DumpRegion(8, "_InputPakInsert", true, hucRegionDumpUpdate, 100));
         ENCODE_CHK_STATUS_RETURN(DumpRegion(10, "_InputCdfTable", true, hucRegionDumpUpdate, 4 * MOS_ALIGN_CEIL(m_basicFeature->m_cdfMaxNumBytes, CODECHAL_CACHELINE_SIZE)));
+        if (m_basicFeature->m_av1PicParams->AdaptiveTUEnabled != 0)
+        {
+            ENCODE_CHK_STATUS_RETURN(DumpRegion(12, "_InputSLBB_TU7", true, hucRegionDumpUpdate, 600 * 4));
+        }
 
         return MOS_STATUS_SUCCESS;
     }
@@ -598,8 +621,8 @@ namespace encode
         // Region 1 - VDenc Stats Buffer (Input)
         params.regionParams[1].presRegion = resTileBasedStatisticsBuffer;
         params.regionParams[1].dwOffset   = offset;
-        // Region 3 - Input SLB Buffer (Input)
-        params.regionParams[3].presRegion = const_cast<PMOS_RESOURCE>(&m_vdencReadBatchBuffer[bufIdx][currentPass]);
+        // Region 3 - Input SLB Buffer (Input Origin)
+        params.regionParams[3].presRegion = const_cast<PMOS_RESOURCE>(&m_vdencReadBatchBufferOrigin[bufIdx][currentPass]);
         // Region 4 - BRC Data for next frame's width/height - (Output)
         params.regionParams[4].presRegion = resBrcDataBuffer;
         params.regionParams[4].isWritable = true;
@@ -620,6 +643,11 @@ namespace encode
         // Region 11 - CDF (output)
         params.regionParams[11].presRegion = m_basicFeature->m_defaultCdfBufferInUse;
         params.regionParams[11].isWritable = true;
+        // Region 12 - Input SLB Buffer (Input TU7)
+        if (m_basicFeature->m_av1PicParams->AdaptiveTUEnabled != 0)
+        {
+            params.regionParams[12].presRegion = const_cast<PMOS_RESOURCE>(&m_vdencReadBatchBufferTU7[bufIdx][currentPass]);
+        }
 
         return MOS_STATUS_SUCCESS;
     }

diff --git a/media_softlet/agnostic/common/codec/hal/enc/av1/packet/encode_av1_brc_update_packet.h b/media_softlet/agnostic/common/codec/hal/enc/av1/packet/encode_av1_brc_update_packet.h
@@ -143,8 +143,9 @@ namespace encode
         uint8_t     UPD_CQMEnabled;
         uint8_t     UPD_TempCurrentlayer;
         uint8_t     UPD_TempScalable;
+        uint8_t     UPD_AdaptiveTUEnabled;
 
-        uint8_t     RSVD8[62];
+        uint8_t     RSVD8[61];
     };
 
     typedef struct _HUC_MODE_COST
@@ -235,7 +236,8 @@ namespace encode
         static constexpr uint32_t               m_vdboxHucAv1BrcUpdateKernelDescriptor = 19;//!< Huc AV1 Brc init kernel descriptor
 
         // Batch Buffer for VDEnc
-        MOS_RESOURCE                            m_vdencReadBatchBuffer[CODECHAL_ENCODE_RECYCLED_BUFFER_NUM][VDENC_BRC_NUM_OF_PASSES] = {};  //!< VDEnc read batch buffer
+        MOS_RESOURCE                            m_vdencReadBatchBufferOrigin[CODECHAL_ENCODE_RECYCLED_BUFFER_NUM][VDENC_BRC_NUM_OF_PASSES] = {};  //!< VDEnc read batch buffer
+        MOS_RESOURCE                            m_vdencReadBatchBufferTU7[CODECHAL_ENCODE_RECYCLED_BUFFER_NUM][VDENC_BRC_NUM_OF_PASSES]    = {};  //!< VDEnc read batch buffer
         MOS_RESOURCE                            m_vdencPakInsertBatchBuffer[CODECHAL_ENCODE_RECYCLED_BUFFER_NUM] = {};                      //!< VDEnc read batch buffer
         MOS_RESOURCE                            m_vdencBrcConstDataBuffer[CODECHAL_ENCODE_RECYCLED_BUFFER_NUM] = {};                        //!< VDEnc brc constant data buffer
 

diff --git a/media_softlet/agnostic/common/codec/hal/enc/avc/features/encode_avc_brc.cpp b/media_softlet/agnostic/common/codec/hal/enc/avc/features/encode_avc_brc.cpp
@@ -190,6 +190,14 @@ MOS_STATUS AvcEncodeBRC::Update(void *params)
         "Encode RateControl Method",
         m_rcMode,
         MediaUserSetting::Group::Sequence);
+
+    MediaUserSetting::Value outValue;
+    ReadUserSetting(
+        m_userSettingPtr,
+        outValue,
+        "Adaptive TU Enable",
+        MediaUserSetting::Group::Sequence);
+    m_basicFeature->m_picParam->AdaptiveTUEnabled |= outValue.Get<uint8_t>();
 #endif
     return MOS_STATUS_SUCCESS;
 }
@@ -640,6 +648,8 @@ MOS_STATUS AvcEncodeBRC::SetDmemForUpdate(void *params, uint16_t currPass, bool
     hucVdencBrcUpdateDmem->UPD_PAKPassNum_U8 = (uint8_t)currPass;
     hucVdencBrcUpdateDmem->UPD_MaxNumPass_U8 = m_featureManager->GetNumPass();
 
+    hucVdencBrcUpdateDmem->UPD_AdaptiveTUEnabled = avcPicParams->AdaptiveTUEnabled;
+
     uint32_t numP = 0;
     if (avcSeqParams->GopRefDist && (avcSeqParams->GopPicSize > 0))
     {

diff --git a/media_softlet/agnostic/common/codec/hal/enc/avc/packet/encode_avc_huc_brc_update_packet.cpp b/media_softlet/agnostic/common/codec/hal/enc/avc/packet/encode_avc_huc_brc_update_packet.cpp
@@ -141,7 +141,15 @@ MOS_STATUS AvcHucBrcUpdatePkt::AllocateResources()
         allocParamsForBufferLinear.ResUsageType = MOS_HW_RESOURCE_USAGE_ENCODE_INTERNAL_READ_WRITE_CACHE;
         allocatedbuffer = m_allocator->AllocateResource(allocParamsForBufferLinear, true);
         ENCODE_CHK_NULL_RETURN(allocatedbuffer);
-        m_vdencBrcImageStatesReadBuffer[k] = allocatedbuffer;
+        m_vdencBrcImageStatesReadBufferOrigin[k] = allocatedbuffer;
+
+        // VDENC IMG STATE read buffer
+        allocParamsForBufferLinear.dwBytes      = m_brcFeature->GetVdencBRCImgStateBufferSize();
+        allocParamsForBufferLinear.pBufName     = "VDENC BRC IMG State Read Buffer";
+        allocParamsForBufferLinear.ResUsageType = MOS_HW_RESOURCE_USAGE_ENCODE_INTERNAL_READ_WRITE_CACHE;
+        allocatedbuffer                         = m_allocator->AllocateResource(allocParamsForBufferLinear, true);
+        ENCODE_CHK_NULL_RETURN(allocatedbuffer);
+        m_vdencBrcImageStatesReadBufferTU7[k] = allocatedbuffer;
 
         for (auto i = 0; i < VDENC_BRC_NUM_OF_PASSES; i++)
         {
@@ -334,8 +342,18 @@ MOS_STATUS AvcHucBrcUpdatePkt::Execute(PMOS_COMMAND_BUFFER cmdBuffer, bool store
 MOS_STATUS AvcHucBrcUpdatePkt::Submit(MOS_COMMAND_BUFFER *commandBuffer, uint8_t packetPhase)
 {
     ENCODE_FUNC_CALL();
+    ENCODE_CHK_NULL_RETURN(m_basicFeature->m_picParam);
+    ENCODE_CHK_NULL_RETURN(m_basicFeature->m_seqParam);
+
+    ENCODE_CHK_STATUS_RETURN(ConstructImageStateReadBuffer(m_vdencBrcImageStatesReadBufferOrigin[m_pipeline->m_currRecycledBufIdx]));
 
-    ENCODE_CHK_STATUS_RETURN(ConstructImageStateReadBuffer(m_vdencBrcImageStatesReadBuffer[m_pipeline->m_currRecycledBufIdx]));
+    if (m_basicFeature->m_picParam->AdaptiveTUEnabled != 0)
+    {
+        auto original_TU              = m_basicFeature->m_targetUsage;
+        m_basicFeature->m_targetUsage = m_basicFeature->m_seqParam->TargetUsage = 7;
+        ENCODE_CHK_STATUS_RETURN(ConstructImageStateReadBuffer(m_vdencBrcImageStatesReadBufferTU7[m_pipeline->m_currRecycledBufIdx]));
+        m_basicFeature->m_targetUsage = m_basicFeature->m_seqParam->TargetUsage = original_TU;
+    }
 
     bool firstTaskInPhase = packetPhase & firstPacket;
     bool requestProlog = false;
@@ -525,7 +543,7 @@ MHW_SETPAR_DECL_SRC(HUC_VIRTUAL_ADDR_STATE, AvcHucBrcUpdatePkt)
     // Input regions
     params.regionParams[1].presRegion = m_basicFeature->m_recycleBuf->GetBuffer(VdencStatsBuffer, 0);
     params.regionParams[2].presRegion = m_basicFeature->m_recycleBuf->GetBuffer(BrcPakStatisticBuffer, 0);
-    params.regionParams[3].presRegion = m_vdencBrcImageStatesReadBuffer[m_pipeline->m_currRecycledBufIdx];
+    params.regionParams[3].presRegion = m_vdencBrcImageStatesReadBufferOrigin[m_pipeline->m_currRecycledBufIdx];
     params.regionParams[5].presRegion = m_vdencBrcConstDataBuffer[GetCurrConstDataBufIdx()];
     params.regionParams[7].presRegion = m_basicFeature->m_recycleBuf->GetBuffer(PakSliceSizeStreamOutBuffer, m_pipeline->GetCurrentPass() ?
         m_basicFeature->m_frameNum : m_basicFeature->m_frameNum ? m_basicFeature->m_frameNum-1 : 0);   // use stats from previous frame for pass 0
@@ -542,6 +560,11 @@ MHW_SETPAR_DECL_SRC(HUC_VIRTUAL_ADDR_STATE, AvcHucBrcUpdatePkt)
         params.regionParams[8].presRegion = m_basicFeature->m_recycleBuf->GetBuffer(RecycleResId::HucRoiMapBuffer, m_basicFeature->m_frameNum);
     }
 
+    if (m_basicFeature->m_picParam->AdaptiveTUEnabled != 0)
+    {
+        params.regionParams[12].presRegion = m_vdencBrcImageStatesReadBufferTU7[m_pipeline->m_currRecycledBufIdx];
+    }
+
     return MOS_STATUS_SUCCESS;
 }
 

diff --git a/media_softlet/agnostic/common/codec/hal/enc/avc/packet/encode_avc_huc_brc_update_packet.h b/media_softlet/agnostic/common/codec/hal/enc/avc/packet/encode_avc_huc_brc_update_packet.h
@@ -123,7 +123,8 @@ struct VdencAvcHucBrcUpdateDmem
     uint16_t     UPD_NumSlicesForRounding;
     uint32_t     UPD_UserMaxFramePB;        // In Bytes
     uint8_t      UPD_ExtCurrFrameType;      // correctly calculated FrameType for all cases (including hierarchy golden BGops)
-    uint8_t      RSVD2[3];
+    uint8_t      UPD_AdaptiveTUEnabled;
+    uint8_t      RSVD2[2];
 };
 
 struct VdencAvcHucBrcConstantData
@@ -248,7 +249,8 @@ class AvcHucBrcUpdatePkt : public EncodeHucPkt
     uint32_t m_vdencBrcUpdateDmemBufferSize = sizeof(VdencAvcHucBrcUpdateDmem);         //!< Offset of BRC update DMEM buffer
     uint32_t m_vdencBrcConstDataBufferSize  = sizeof(VdencAvcHucBrcConstantData);       //!< Offset of BRC const data buffer
 
-    PMOS_RESOURCE m_vdencBrcImageStatesReadBuffer[CODECHAL_ENCODE_RECYCLED_BUFFER_NUM]                     = {};  //!< Read-only VDENC+PAK IMG STATE buffer.
+    PMOS_RESOURCE m_vdencBrcImageStatesReadBufferOrigin[CODECHAL_ENCODE_RECYCLED_BUFFER_NUM]               = {};  //!< Read-only VDENC+PAK IMG STATE buffer.
+    PMOS_RESOURCE m_vdencBrcImageStatesReadBufferTU7[CODECHAL_ENCODE_RECYCLED_BUFFER_NUM]                  = {};  //!< Read-only VDENC+PAK IMG STATE buffer.
     PMOS_RESOURCE m_vdencBrcUpdateDmemBuffer[CODECHAL_ENCODE_RECYCLED_BUFFER_NUM][VDENC_BRC_NUM_OF_PASSES] = {};  //!< Brc Update DMEM Buffer Array.
     PMOS_RESOURCE m_vdencBrcConstDataBuffer[CODECHAL_ENCODE_VDENC_BRC_CONST_BUFFER_NUM]                    = {};  //!< BRC Const Data Buffer for each frame type.
 

diff --git a/media_softlet/agnostic/common/codec/hal/enc/hevc/features/encode_hevc_brc.cpp b/media_softlet/agnostic/common/codec/hal/enc/hevc/features/encode_hevc_brc.cpp
@@ -104,6 +104,15 @@ namespace encode
             "HEVC VDEnc ACQP Enable",
             m_hevcVDEncAcqpEnabled,
             MediaUserSetting::Group::Sequence);
+
+        ENCODE_CHK_NULL_RETURN(m_basicFeature->m_hevcPicParams);
+        MediaUserSetting::Value outValue;
+        ReadUserSetting(
+            m_userSettingPtr,
+            outValue,
+            "Adaptive TU Enable",
+            MediaUserSetting::Group::Sequence);
+        m_basicFeature->m_hevcPicParams->AdaptiveTUEnabled |= outValue.Get<uint8_t>();
 #endif
         return MOS_STATUS_SUCCESS;
     }
@@ -286,6 +295,8 @@ namespace encode
 
         hucVdencBrcUpdateDmem->UPD_Randomaccess = m_basicFeature->m_hevcSeqParams->LowDelayMode == 1 ? 0 : 1;
 
+        hucVdencBrcUpdateDmem->UPD_AdaptiveTUEnabled = m_basicFeature->m_hevcPicParams->AdaptiveTUEnabled;
+
         return MOS_STATUS_SUCCESS;
     }
 

diff --git a/media_softlet/agnostic/common/codec/hal/enc/hevc/packet/encode_hevc_vdenc_packet.cpp b/media_softlet/agnostic/common/codec/hal/enc/hevc/packet/encode_hevc_vdenc_packet.cpp
@@ -1397,6 +1397,10 @@ namespace encode
                 secondLevelBatchBufferUsed->dwOffset,
                 false,
                 m_basicFeature->m_vdencBatchBufferPerSlicePart2Size[currSlcIdx]);
+            if (m_basicFeature->m_hevcPicParams->AdaptiveTUEnabled == 0)
+            {
+                SETPAR_AND_ADDCMD(VDENC_HEVC_VP9_TILE_SLICE_STATE, m_vdencItf, &cmdBuffer);
+            }
         }
         else
         {
@@ -1412,8 +1416,10 @@ namespace encode
             ENCODE_CHK_STATUS_RETURN(AddAllCmds_HCP_PAK_INSERT_OBJECT(&cmdBuffer));
 
             SETPAR_AND_ADDCMD(VDENC_WEIGHTSOFFSETS_STATE, m_vdencItf, &cmdBuffer);
+
+            SETPAR_AND_ADDCMD(VDENC_HEVC_VP9_TILE_SLICE_STATE, m_vdencItf, &cmdBuffer);
         }
-        SETPAR_AND_ADDCMD(VDENC_HEVC_VP9_TILE_SLICE_STATE, m_vdencItf, &cmdBuffer);
+
         SETPAR_AND_ADDCMD(VDENC_WALKER_STATE, m_vdencItf, &cmdBuffer);
         return eStatus;
     }