amdgpu/addrlib: Pad pitch to multiples of 256 for DCC surface on Fiji

The change also modifies function CiLib::HwlPadDimensions to report
adjusted pitch alignment.

8 files changed, 131 insertions, 49 deletions

diff --git a/src/amd/addrlib/core/addrlib1.cpp b/src/amd/addrlib/core/addrlib1.cpp
index d48aa7c92bf..040891c431c 100644
--- a/src/amd/addrlib/core/addrlib1.cpp
+++ b/src/amd/addrlib/core/addrlib1.cpp
@@ -3228,13 +3228,14 @@ VOID Lib::PadDimensions(
     UINT_32             padDims,     ///< [in] Dimensions to pad valid value 1,2,3
     UINT_32             mipLevel,    ///< [in] MipLevel
     UINT_32*            pPitch,      ///< [in,out] pitch in pixels
-    UINT_32             pitchAlign,  ///< [in] pitch alignment
+    UINT_32*            pPitchAlign, ///< [in,out] pitch align could be changed in HwlPadDimensions
     UINT_32*            pHeight,     ///< [in,out] height in pixels
     UINT_32             heightAlign, ///< [in] height alignment
     UINT_32*            pSlices,     ///< [in,out] number of slices
     UINT_32             sliceAlign   ///< [in] number of slice alignment
     ) const
 {
+    UINT_32 pitchAlign = *pPitchAlign;
     UINT_32 thickness = Thickness(tileMode);
 
     ADDR_ASSERT(padDims <= 3);
@@ -3302,14 +3303,11 @@ VOID Lib::PadDimensions(
                      flags,
                      numSamples,
                      pTileInfo,
-                     padDims,
                      mipLevel,
                      pPitch,
-                     pitchAlign,
-                     pHeight,
-                     heightAlign,
-                     pSlices,
-                     sliceAlign);
+                     pPitchAlign,
+                     *pHeight,
+                     heightAlign);
 }
 
 
diff --git a/src/amd/addrlib/core/addrlib1.h b/src/amd/addrlib/core/addrlib1.h
index 9c66f53e737..c1fc693b64f 100644
--- a/src/amd/addrlib/core/addrlib1.h
+++ b/src/amd/addrlib/core/addrlib1.h
@@ -71,6 +71,9 @@ struct TileModeFlags
     UINT_32 isBankSwapped   : 1;
 };
 
+static const UINT_32 Block64K = 0x10000;
+static const UINT_32 PrtTileSize = Block64K;
+
 /**
 ****************************************************************************************************
 * @brief This class contains asic independent address lib functionalities
@@ -365,14 +368,13 @@ protected:
     VOID PadDimensions(
         AddrTileMode tileMode, UINT_32 bpp, ADDR_SURFACE_FLAGS flags,
         UINT_32 numSamples, ADDR_TILEINFO* pTileInfo, UINT_32 padDims, UINT_32 mipLevel,
-        UINT_32* pPitch, UINT_32 pitchAlign, UINT_32* pHeight, UINT_32 heightAlign,
+        UINT_32* pPitch, UINT_32* pPitchAlign, UINT_32* pHeight, UINT_32 heightAlign,
         UINT_32* pSlices, UINT_32 sliceAlign) const;
 
     virtual VOID HwlPadDimensions(
         AddrTileMode tileMode, UINT_32 bpp, ADDR_SURFACE_FLAGS flags,
-        UINT_32 numSamples, ADDR_TILEINFO* pTileInfo, UINT_32 padDims, UINT_32 mipLevel,
-        UINT_32* pPitch, UINT_32 pitchAlign, UINT_32* pHeight, UINT_32 heightAlign,
-        UINT_32* pSlices, UINT_32 sliceAlign) const
+        UINT_32 numSamples, ADDR_TILEINFO* pTileInfo, UINT_32 mipLevel,
+        UINT_32* pPitch, UINT_32* pPitchAlign, UINT_32 height, UINT_32 heightAlign) const
     {
     }
 
diff --git a/src/amd/addrlib/r800/ciaddrlib.cpp b/src/amd/addrlib/r800/ciaddrlib.cpp
index 3986c3b2998..3509024f76f 100644
--- a/src/amd/addrlib/r800/ciaddrlib.cpp
+++ b/src/amd/addrlib/r800/ciaddrlib.cpp
@@ -1941,6 +1941,41 @@ UINT_64 CiLib::HwlComputeMetadataNibbleAddress(
 
 /**
 ****************************************************************************************************
+*   CiLib::HwlComputeSurfaceAlignmentsMacroTiled
+*
+*   @brief
+*       Hardware layer function to compute alignment request for macro tile mode
+*
+*   @return
+*       N/A
+*
+****************************************************************************************************
+*/
+VOID CiLib::HwlComputeSurfaceAlignmentsMacroTiled(
+    AddrTileMode        tileMode,           ///< [in] tile mode
+    UINT_32             bpp,                ///< [in] bits per pixel
+    ADDR_SURFACE_FLAGS  flags,              ///< [in] surface flags
+    UINT_32             mipLevel,           ///< [in] mip level
+    UINT_32             numSamples,         ///< [in] number of samples
+    ADDR_TILEINFO*      pTileInfo,          ///< [in,out] bank structure.
+    UINT_32*            pBaseAlign,         ///< [out] base address alignment in bytes
+    UINT_32*            pPitchAlign,        ///< [out] pitch alignment in pixels
+    UINT_32*            pHeightAlign,       ///< [out] height alignment in pixels
+    UINT_32*            pMacroTileWidth,    ///< [out] macro tile width in pixels
+    UINT_32*            pMacroTileHeight    ///< [out] macro tile height in pixels
+    ) const
+{
+    if ((m_settings.isFiji == TRUE) &&
+        (flags.dccCompatible == TRUE) &&
+        (mipLevel == 0) &&
+        (tileMode == ADDR_TM_PRT_TILED_THIN1))
+    {
+        *pPitchAlign = PowTwoAlign(*pPitchAlign, 256);
+    }
+}
+
+/**
+****************************************************************************************************
 *   CiLib::HwlPadDimensions
 *
 *   @brief
@@ -1956,22 +1991,19 @@ VOID CiLib::HwlPadDimensions(
     UINT_32             bpp,         ///< [in] bits per pixel
     ADDR_SURFACE_FLAGS  flags,       ///< [in] surface flags
     UINT_32             numSamples,  ///< [in] number of samples
-    ADDR_TILEINFO*      pTileInfo,   ///< [in,out] bank structure.
-    UINT_32             padDims,     ///< [in] Dimensions to pad valid value 1,2,3
-    UINT_32             mipLevel,    ///< [in] MipLevel
+    ADDR_TILEINFO*      pTileInfo,   ///< [in] tile info
+    UINT_32             mipLevel,    ///< [in] mip level
     UINT_32*            pPitch,      ///< [in,out] pitch in pixels
-    UINT_32             pitchAlign,  ///< [in] pitch alignment
-    UINT_32*            pHeight,     ///< [in,out] height in pixels
-    UINT_32             heightAlign, ///< [in] height alignment
-    UINT_32*            pSlices,     ///< [in,out] number of slices
-    UINT_32             sliceAlign   ///< [in] number of slice alignment
+    UINT_32*            pPitchAlign, ///< [in,out] pitch alignment
+    UINT_32             height,      ///< [in] height in pixels
+    UINT_32             heightAlign  ///< [in] height alignment
     ) const
 {
-    if (m_settings.isVolcanicIslands &&
-        flags.dccCompatible &&
+    if ((m_settings.isVolcanicIslands == TRUE) &&
+        (flags.dccCompatible == TRUE) &&
         (numSamples > 1) &&
         (mipLevel == 0) &&
-        IsMacroTiled(tileMode))
+        (IsMacroTiled(tileMode) == TRUE))
     {
         UINT_32 tileSizePerSample = BITS_TO_BYTES(bpp * MicroTileWidth * MicroTileHeight);
         UINT_32 samplesPerSplit  = pTileInfo->tileSplitBytes / tileSizePerSample;
@@ -1979,7 +2011,7 @@ VOID CiLib::HwlPadDimensions(
         if (samplesPerSplit < numSamples)
         {
             UINT_32 dccFastClearByteAlign = HwlGetPipes(pTileInfo) * m_pipeInterleaveBytes * 256;
-            UINT_32 bytesPerSplit = BITS_TO_BYTES((*pPitch) * (*pHeight) * bpp * samplesPerSplit);
+            UINT_32 bytesPerSplit = BITS_TO_BYTES((*pPitch) * height * bpp * samplesPerSplit);
 
             ADDR_ASSERT(IsPow2(dccFastClearByteAlign));
 
@@ -1988,15 +2020,14 @@ VOID CiLib::HwlPadDimensions(
                 UINT_32 dccFastClearPixelAlign = dccFastClearByteAlign /
                                                 BITS_TO_BYTES(bpp) /
                                                 samplesPerSplit;
-                UINT_32 macroTilePixelAlign = pitchAlign * heightAlign;
+                UINT_32 macroTilePixelAlign = (*pPitchAlign) * heightAlign;
 
                 if ((dccFastClearPixelAlign >= macroTilePixelAlign) &&
                     ((dccFastClearPixelAlign % macroTilePixelAlign) == 0))
                 {
                     UINT_32 dccFastClearPitchAlignInMacroTile =
                         dccFastClearPixelAlign / macroTilePixelAlign;
-                    UINT_32 heightInMacroTile = *pHeight / heightAlign;
-                    UINT_32 dccFastClearPitchAlignInPixels;
+                    UINT_32 heightInMacroTile = height / heightAlign;
 
                     while ((heightInMacroTile > 1) &&
                            ((heightInMacroTile % 2) == 0) &&
@@ -2007,7 +2038,8 @@ VOID CiLib::HwlPadDimensions(
                         dccFastClearPitchAlignInMacroTile >>= 1;
                     }
 
-                    dccFastClearPitchAlignInPixels = pitchAlign * dccFastClearPitchAlignInMacroTile;
+                    UINT_32 dccFastClearPitchAlignInPixels =
+                        (*pPitchAlign) * dccFastClearPitchAlignInMacroTile;
 
                     if (IsPow2(dccFastClearPitchAlignInPixels))
                     {
@@ -2019,6 +2051,8 @@ VOID CiLib::HwlPadDimensions(
                         *pPitch /= dccFastClearPitchAlignInPixels;
                         *pPitch *= dccFastClearPitchAlignInPixels;
                     }
+
+                    *pPitchAlign = dccFastClearPitchAlignInPixels;
                 }
             }
         }
diff --git a/src/amd/addrlib/r800/ciaddrlib.h b/src/amd/addrlib/r800/ciaddrlib.h
index a9b1607ff1d..c59a0b127a1 100644
--- a/src/amd/addrlib/r800/ciaddrlib.h
+++ b/src/amd/addrlib/r800/ciaddrlib.h
@@ -166,9 +166,14 @@ protected:
 
     virtual VOID HwlPadDimensions(
         AddrTileMode tileMode, UINT_32 bpp, ADDR_SURFACE_FLAGS flags,
-        UINT_32 numSamples, ADDR_TILEINFO* pTileInfo, UINT_32 padDims, UINT_32 mipLevel,
-        UINT_32* pPitch, UINT_32 pitchAlign, UINT_32* pHeight, UINT_32 heightAlign,
-        UINT_32* pSlices, UINT_32 sliceAlign) const;
+        UINT_32 numSamples, ADDR_TILEINFO* pTileInfo, UINT_32 mipLevel,
+        UINT_32* pPitch, UINT_32 *PitchAlign, UINT_32 height, UINT_32 heightAlign) const;
+
+    virtual VOID HwlComputeSurfaceAlignmentsMacroTiled(
+        AddrTileMode tileMode, UINT_32 bpp, ADDR_SURFACE_FLAGS flags,
+        UINT_32 mipLevel, UINT_32 numSamples, ADDR_TILEINFO* pTileInfo,
+        UINT_32* pBaseAlign, UINT_32* pPitchAlign, UINT_32* pHeightAlign,
+        UINT_32* pMacroTileWidth, UINT_32* pMacroTileHeight) const;
 
 private:
     VOID ReadGbTileMode(
diff --git a/src/amd/addrlib/r800/egbaddrlib.cpp b/src/amd/addrlib/r800/egbaddrlib.cpp
index b15168263f3..9655c47f7a8 100644
--- a/src/amd/addrlib/r800/egbaddrlib.cpp
+++ b/src/amd/addrlib/r800/egbaddrlib.cpp
@@ -264,7 +264,7 @@ BOOL_32 EgBasedLib::ComputeSurfaceInfoLinear(
                   pOut->pTileInfo,
                   padDims,
                   pIn->mipLevel,
-                  &expPitch, pOut->pitchAlign,
+                  &expPitch, &pOut->pitchAlign,
                   &expHeight, pOut->heightAlign,
                   &expNumSlices, microTileThickness);
 
@@ -378,7 +378,7 @@ BOOL_32 EgBasedLib::ComputeSurfaceInfoMicroTiled(
                   pOut->pTileInfo,
                   padDims,
                   pIn->mipLevel,
-                  &expPitch, pOut->pitchAlign,
+                  &expPitch, &pOut->pitchAlign,
                   &expHeight, pOut->heightAlign,
                   &expNumSlices, microTileThickness);
 
@@ -527,7 +527,7 @@ BOOL_32 EgBasedLib::ComputeSurfaceInfoMacroTiled(
                       pOut->pTileInfo,
                       padDims,
                       pIn->mipLevel,
-                      &paddedPitch, pOut->pitchAlign,
+                      &paddedPitch, &pOut->pitchAlign,
                       &paddedHeight, pOut->heightAlign,
                       &expNumSlices, microTileThickness);
 
@@ -932,22 +932,9 @@ BOOL_32 EgBasedLib::ComputeSurfaceAlignmentsMacroTiled(
         *pBaseAlign = pipes *
             pTileInfo->bankWidth * pTileInfo->banks * pTileInfo->bankHeight * tileSize;
 
-        if ((mipLevel == 0) && (flags.prt) && (m_chipFamily == ADDR_CHIP_FAMILY_SI))
-        {
-            static const UINT_32 PrtTileSize = 0x10000;
-
-            UINT_32 macroTileSize = macroTileWidth * macroTileHeight * numSamples * bpp / 8;
-
-            if (macroTileSize < PrtTileSize)
-            {
-                UINT_32 numMacroTiles = PrtTileSize / macroTileSize;
-
-                ADDR_ASSERT((PrtTileSize % macroTileSize) == 0);
-
-                *pPitchAlign *= numMacroTiles;
-                *pBaseAlign  *= numMacroTiles;
-            }
-        }
+        HwlComputeSurfaceAlignmentsMacroTiled(tileMode, bpp, flags, mipLevel, numSamples,
+                                              pTileInfo, pBaseAlign, pPitchAlign, pHeightAlign,
+                                              pMacroTileWidth, pMacroTileHeight);
     }
 
     return valid;
diff --git a/src/amd/addrlib/r800/egbaddrlib.h b/src/amd/addrlib/r800/egbaddrlib.h
index c397cfc4f81..1a560033681 100644
--- a/src/amd/addrlib/r800/egbaddrlib.h
+++ b/src/amd/addrlib/r800/egbaddrlib.h
@@ -300,6 +300,14 @@ protected:
     static UINT_32 ComputeFmaskNumPlanesFromNumSamples(UINT_32 numSamples);
     static UINT_32 ComputeFmaskResolvedBppFromNumSamples(UINT_32 numSamples);
 
+    virtual VOID HwlComputeSurfaceAlignmentsMacroTiled(
+        AddrTileMode tileMode, UINT_32 bpp, ADDR_SURFACE_FLAGS flags,
+        UINT_32 mipLevel, UINT_32 numSamples, ADDR_TILEINFO* pTileInfo,
+        UINT_32* pBaseAlign, UINT_32* pPitchAlign, UINT_32* pHeightAlign,
+        UINT_32* pMacroTileWidth, UINT_32* pMacroTileHeight) const
+    {
+    }
+
 private:
 
     BOOL_32 ComputeSurfaceInfoLinear(
diff --git a/src/amd/addrlib/r800/siaddrlib.cpp b/src/amd/addrlib/r800/siaddrlib.cpp
index 512fb273243..f8f937950ed 100644
--- a/src/amd/addrlib/r800/siaddrlib.cpp
+++ b/src/amd/addrlib/r800/siaddrlib.cpp
@@ -3327,6 +3327,48 @@ ADDR_E_RETURNCODE SiLib::HwlGetMaxAlignments(
 
 /**
 ****************************************************************************************************
+*   SiLib::HwlComputeSurfaceAlignmentsMacroTiled
+*
+*   @brief
+*       Hardware layer function to compute alignment request for macro tile mode
+*
+*   @return
+*       N/A
+*
+****************************************************************************************************
+*/
+VOID SiLib::HwlComputeSurfaceAlignmentsMacroTiled(
+    AddrTileMode        tileMode,           ///< [in] tile mode
+    UINT_32             bpp,                ///< [in] bits per pixel
+    ADDR_SURFACE_FLAGS  flags,              ///< [in] surface flags
+    UINT_32             mipLevel,           ///< [in] mip level
+    UINT_32             numSamples,         ///< [in] number of samples
+    ADDR_TILEINFO*      pTileInfo,          ///< [in,out] bank structure.
+    UINT_32*            pBaseAlign,         ///< [out] base address alignment in bytes
+    UINT_32*            pPitchAlign,        ///< [out] pitch alignment in pixels
+    UINT_32*            pHeightAlign,       ///< [out] height alignment in pixels
+    UINT_32*            pMacroTileWidth,    ///< [out] macro tile width in pixels
+    UINT_32*            pMacroTileHeight    ///< [out] macro tile height in pixels
+    ) const
+{
+    if ((mipLevel == 0) && (flags.prt))
+    {
+        UINT_32 macroTileSize = (*pMacroTileWidth) * (*pMacroTileHeight) * numSamples * bpp / 8;
+
+        if (macroTileSize < PrtTileSize)
+        {
+            UINT_32 numMacroTiles = PrtTileSize / macroTileSize;
+
+            ADDR_ASSERT((PrtTileSize % macroTileSize) == 0);
+
+            *pPitchAlign *= numMacroTiles;
+            *pBaseAlign  *= numMacroTiles;
+        }
+    }
+}
+
+/**
+****************************************************************************************************
 *   SiLib::InitEquationTable
 *
 *   @brief
diff --git a/src/amd/addrlib/r800/siaddrlib.h b/src/amd/addrlib/r800/siaddrlib.h
index b63d7e878dc..c40b5460b4f 100644
--- a/src/amd/addrlib/r800/siaddrlib.h
+++ b/src/amd/addrlib/r800/siaddrlib.h
@@ -241,6 +241,12 @@ protected:
 
     virtual ADDR_E_RETURNCODE HwlGetMaxAlignments(ADDR_GET_MAX_ALINGMENTS_OUTPUT* pOut) const;
 
+    virtual VOID HwlComputeSurfaceAlignmentsMacroTiled(
+        AddrTileMode tileMode, UINT_32 bpp, ADDR_SURFACE_FLAGS flags,
+        UINT_32 mipLevel, UINT_32 numSamples, ADDR_TILEINFO* pTileInfo,
+        UINT_32* pBaseAlign, UINT_32* pPitchAlign, UINT_32* pHeightAlign,
+        UINT_32* pMacroTileWidth, UINT_32* pMacroTileHeight) const;
+
     // Get equation table pointer and number of equations
     virtual UINT_32 HwlGetEquationTableInfo(const ADDR_EQUATION** ppEquationTable) const
     {