From c12e35065af693fcad866d2089adf277a6109683 Mon Sep 17 00:00:00 2001
From: Carlos Xiong <clever.xiong@amd.com>
Date: Wed, 2 Jul 2014 01:46:06 -0400
Subject: [PATCH] amdgpu/addrlib: Add a flag "tcCompatible" to surface info
 output structure.

Even if surface info input flag "tcComaptible" is enabled, tc
compatible may be not supported if tile split happens for depth
surfaces. Add a new flag in output structure to notify client to
disable tc compatible in this case.
---
 src/amd/addrlib/addrinterface.h    | 13 ++++++--
 src/amd/addrlib/core/addrlib.cpp   |  1 +
 src/amd/addrlib/r800/ciaddrlib.cpp | 51 +++++++++++++++++++++++-------
 3 files changed, 50 insertions(+), 15 deletions(-)

diff --git a/src/amd/addrlib/addrinterface.h b/src/amd/addrlib/addrinterface.h
index d9a2b82c043..d05c6ef08db 100644
--- a/src/amd/addrlib/addrinterface.h
+++ b/src/amd/addrlib/addrinterface.h
@@ -534,9 +534,16 @@ typedef struct _ADDR_COMPUTE_SURFACE_INFO_OUTPUT
     INT_32          tileIndex;      ///< Tile index, MAY be "downgraded"
 
     INT_32          macroModeIndex; ///< Index in macro tile mode table if there is one (CI)
-    /// Special information to work around SI mipmap swizzle bug UBTS #317508
-    BOOL_32         last2DLevel;    ///< TRUE if this is the last 2D(3D) tiled
-                                    ///< Only meaningful when create flag checkLast2DLevel is set
+    /// Output flags
+    struct
+    {
+        /// Special information to work around SI mipmap swizzle bug UBTS #317508
+        UINT_32     last2DLevel  : 1;  ///< TRUE if this is the last 2D(3D) tiled
+                                       ///< Only meaningful when create flag checkLast2DLevel is set
+        UINT_32     tcCompatible : 1;  ///< If the surface can be shader compatible
+        UINT_32     reserved     :30; ///< Reserved bits
+    };
+
     /// Stereo info
     ADDR_QBSTEREOINFO*  pStereoInfo;///< Stereo information, needed when .qbStereo flag is TRUE
 } ADDR_COMPUTE_SURFACE_INFO_OUTPUT;
diff --git a/src/amd/addrlib/core/addrlib.cpp b/src/amd/addrlib/core/addrlib.cpp
index b829eec474b..8cf4a245229 100644
--- a/src/amd/addrlib/core/addrlib.cpp
+++ b/src/amd/addrlib/core/addrlib.cpp
@@ -449,6 +449,7 @@ ADDR_E_RETURNCODE AddrLib::ComputeSurfaceInfo(
         pOut->pixelBits = localIn.bpp;
         pOut->numSamples = localIn.numSamples;
         pOut->last2DLevel = FALSE;
+        pOut->tcCompatible = FALSE;
 
 #if !ALT_TEST
         if (localIn.numSamples > 1)
diff --git a/src/amd/addrlib/r800/ciaddrlib.cpp b/src/amd/addrlib/r800/ciaddrlib.cpp
index 5f8a1fee527..1024ff24549 100644
--- a/src/amd/addrlib/r800/ciaddrlib.cpp
+++ b/src/amd/addrlib/r800/ciaddrlib.cpp
@@ -483,7 +483,8 @@ INT_32 CiAddrLib::HwlPostCheckTileIndex(
                         // tileSplitBytes stored in m_tileTable is only valid for depth entries
                         if (type == ADDR_DEPTH_SAMPLE_ORDER)
                         {
-                            if (pInfo->tileSplitBytes == m_tileTable[index].info.tileSplitBytes)
+                            if (Min(m_tileTable[index].info.tileSplitBytes,
+                                    m_rowSize) == pInfo->tileSplitBytes)
                             {
                                 break;
                             }
@@ -621,6 +622,9 @@ ADDR_E_RETURNCODE CiAddrLib::HwlComputeSurfaceInfo(
         pOut->macroModeIndex = TileIndexInvalid;
     }
 
+    // Pass tcCompatible flag from input to output; and turn off it if tile split occurs
+    pOut->tcCompatible = pIn->flags.tcCompatible;
+
     ADDR_E_RETURNCODE retCode = SiAddrLib::HwlComputeSurfaceInfo(pIn,pOut);
 
     if (pOut->macroModeIndex == TileIndexNoMacroIndex)
@@ -1010,11 +1014,18 @@ VOID CiAddrLib::HwlSetupTileInfo(
         // See table entries 0-4
         if (flags.depth || flags.stencil)
         {
-            if (flags.depth && (flags.nonSplit || flags.tcCompatible))
+            // tileSize = thickness * bpp * numSamples * 8 * 8 / 8
+            UINT_32 tileSize = thickness * bpp * numSamples * 8;
+
+            // Turn off tc compatible if row_size is smaller than tile size (tile split occurs).
+            if (m_rowSize < tileSize)
             {
-                // tileSize = bpp * numSamples * 8 * 8 / 8
-                UINT_32 tileSize = bpp * numSamples * 8;
+                flags.tcCompatible = FALSE;
+                pOut->tcCompatible = FALSE;
+            }
 
+            if (flags.depth && (flags.nonSplit || flags.tcCompatible))
+            {
                 // Texure readable depth surface should not be split
                 switch (tileSize)
                 {
@@ -1215,6 +1226,29 @@ VOID CiAddrLib::HwlSetupTileInfo(
         pOut->tileIndex = 8;
         *pTileInfo = m_tileTable[8].info;
     }
+
+    // Turn off tcCompatible for color surface if tileSplit happens. Depth/stencil is
+    // handled at tileIndex selecting time.
+    if (pOut->tcCompatible && (inTileType != ADDR_DEPTH_SAMPLE_ORDER))
+    {
+        if (IsMacroTiled(tileMode))
+        {
+            // Non-depth entries store a split factor
+            UINT_32 sampleSplit = m_tileTable[pOut->tileIndex].info.tileSplitBytes;
+            UINT_32 tileBytes1x = BITS_TO_BYTES(bpp * MicroTilePixels * thickness);
+            UINT_32 colorTileSplit = Max(256u, sampleSplit * tileBytes1x);
+
+            if (m_rowSize < colorTileSplit)
+            {
+                pOut->tcCompatible = FALSE;
+            }
+        }
+        else
+        {
+            // Client should not enable tc compatible for linear and 1D tile modes.
+            pOut->tcCompatible = FALSE;
+        }
+    }
 }
 
 /**
@@ -1517,14 +1551,7 @@ INT_32 CiAddrLib::HwlComputeMacroModeIndex(
 
         pTileInfo->pipeConfig = m_tileTable[tileIndex].info.pipeConfig;
 
-        if (m_tileTable[tileIndex].type != ADDR_DEPTH_SAMPLE_ORDER)
-        {
-            pTileInfo->tileSplitBytes = tileSplitC;
-        }
-        else
-        {
-            pTileInfo->tileSplitBytes = m_tileTable[tileIndex].info.tileSplitBytes;
-        }
+        pTileInfo->tileSplitBytes = tileSplitC;
     }
 
     if (NULL != pTileMode)
-- 
2.30.2