amd/addrlib: expose DCC address equations to drivers

Acked-by: Pierre-Eric Pelloux-Prayer <pierre-eric.pelloux-prayer@amd.com>
Reviewed-by: Bas Nieuwenhuizen <bas@basnieuwenhuizen.nl>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/10003>

diff --git a/src/amd/addrlib/inc/addrinterface.h b/src/amd/addrlib/inc/addrinterface.h
index bbde582..807098f 100644 (file)
--- a/src/amd/addrlib/inc/addrinterface.h
+++ b/src/amd/addrlib/inc/addrinterface.h
@@ -3394,6 +3394,35 @@ typedef struct _ADDR2_COMPUTE_DCCINFO_OUTPUT
     };
 
     ADDR2_META_MIP_INFO* pMipInfo;      ///< DCC mip information
+
+    /* The equation for doing DCC address computations in shaders. */
+    union {
+       /* This is chip-specific, and it varies with:
+        * - resource type
+        * - swizzle_mode
+        * - bpp
+        * - number of fragments
+        * - pipe_aligned
+        * - rb_aligned
+        */
+       struct {
+          UINT_8 num_bits;
+
+          struct {
+             struct {
+                UINT_8 dim; /* 0..4 as index, 5 means invalid */
+                UINT_8 ord; /* 0..31 */
+             } coord[8]; /* 0..num_coords */
+          } bit[32]; /* 0..num_bits */
+          UINT_8 numPipeBits;
+       } gfx9;
+
+       /* This is chip-specific, it requires 64KB_R_X, and it varies with:
+        * - bpp
+        * - pipe_aligned
+        */
+       UINT_16 *gfx10_bits; /* 68 2-byte elements */
+    } equation;
 } ADDR2_COMPUTE_DCCINFO_OUTPUT;
 
 /**

diff --git a/src/amd/addrlib/src/gfx10/gfx10addrlib.cpp b/src/amd/addrlib/src/gfx10/gfx10addrlib.cpp
index 0ffee23..4ded4c3 100644 (file)
--- a/src/amd/addrlib/src/gfx10/gfx10addrlib.cpp
+++ b/src/amd/addrlib/src/gfx10/gfx10addrlib.cpp
@@ -496,6 +496,53 @@ ADDR_E_RETURNCODE Gfx10Lib::HwlComputeDccInfo(
                     pOut->pMipInfo[0].sliceSize = pOut->dccRamSliceSize;
                 }
             }
+
+            // Get the DCC address equation (copied from DccAddrFromCoord)
+            const UINT_32 elemLog2    = Log2(pIn->bpp >> 3);
+            const UINT_32 numPipeLog2 = m_pipesLog2;
+            UINT_32       index       = m_dccBaseIndex + elemLog2;
+            const UINT_8* patIdxTable;
+
+            if (m_settings.supportRbPlus)
+            {
+                patIdxTable = GFX10_DCC_64K_R_X_RBPLUS_PATIDX;
+
+                if (pIn->dccKeyFlags.pipeAligned)
+                {
+                    index += MaxNumOfBpp;
+
+                    if (m_numPkrLog2 < 2)
+                    {
+                        index += m_pipesLog2 * MaxNumOfBpp;
+                    }
+                    else
+                    {
+                        // 4 groups for "m_numPkrLog2 < 2" case
+                        index += 4 * MaxNumOfBpp;
+
+                        const UINT_32 dccPipePerPkr = 3;
+
+                        index += (m_numPkrLog2 - 2) * dccPipePerPkr * MaxNumOfBpp +
+                                 (m_pipesLog2 - m_numPkrLog2) * MaxNumOfBpp;
+                    }
+                }
+            }
+            else
+            {
+                patIdxTable = GFX10_DCC_64K_R_X_PATIDX;
+
+                if (pIn->dccKeyFlags.pipeAligned)
+                {
+                    index += (numPipeLog2 + UnalignedDccType) * MaxNumOfBpp;
+                }
+                else
+                {
+                    index += Min(numPipeLog2, UnalignedDccType - 1) * MaxNumOfBpp;
+                }
+            }
+
+            ADDR_C_ASSERT(sizeof(GFX10_DCC_64K_R_X_SW_PATTERN[patIdxTable[index]]) == 68 * 2);
+            pOut->equation.gfx10_bits = (UINT_16*)GFX10_DCC_64K_R_X_SW_PATTERN[patIdxTable[index]];
         }
     }
 
@@ -2279,7 +2326,7 @@ ADDR_E_RETURNCODE Gfx10Lib::HwlComputeNonBlockCompressedView(
 
             // For any mipmap level, move nonBc view base address by offset
             HwlComputeSubResourceOffsetForSwizzlePattern(&subOffIn, &subOffOut);
-            pOut->offset = subOffOut.offset;\r
+            pOut->offset = subOffOut.offset;
 
             ADDR2_COMPUTE_SLICE_PIPEBANKXOR_INPUT slicePbXorIn = {};
             slicePbXorIn.bpe             = infoIn.bpp;

diff --git a/src/amd/addrlib/src/gfx9/gfx9addrlib.cpp b/src/amd/addrlib/src/gfx9/gfx9addrlib.cpp
index 59a4c86..7461650 100644 (file)
--- a/src/amd/addrlib/src/gfx9/gfx9addrlib.cpp
+++ b/src/amd/addrlib/src/gfx9/gfx9addrlib.cpp
@@ -665,6 +665,55 @@ ADDR_E_RETURNCODE Gfx9Lib::HwlComputeDccInfo(
         pOut->metaBlkNumPerSlice = numMetaBlkX * numMetaBlkY;
         pOut->fastClearSizePerSlice =
             pOut->metaBlkNumPerSlice * numCompressBlkPerMetaBlk * Min(numFrags, m_maxCompFrag);
+
+        // Get the DCC address equation (copied from DccAddrFromCoord)
+        UINT_32 elementBytesLog2  = Log2(pIn->bpp >> 3);
+        UINT_32 numSamplesLog2    = Log2(pIn->numFrags);
+        UINT_32 metaBlkWidthLog2  = Log2(pOut->metaBlkWidth);
+        UINT_32 metaBlkHeightLog2 = Log2(pOut->metaBlkHeight);
+        UINT_32 metaBlkDepthLog2  = Log2(pOut->metaBlkDepth);
+        UINT_32 compBlkWidthLog2  = Log2(pOut->compressBlkWidth);
+        UINT_32 compBlkHeightLog2 = Log2(pOut->compressBlkHeight);
+        UINT_32 compBlkDepthLog2  = Log2(pOut->compressBlkDepth);
+
+        MetaEqParams metaEqParams = {0, elementBytesLog2, numSamplesLog2, pIn->dccKeyFlags,
+                                     Gfx9DataColor, pIn->swizzleMode, pIn->resourceType,
+                                     metaBlkWidthLog2, metaBlkHeightLog2, metaBlkDepthLog2,
+                                     compBlkWidthLog2, compBlkHeightLog2, compBlkDepthLog2};
+
+        CoordEq *eq = (CoordEq *)((Gfx9Lib *)this)->GetMetaEquation(metaEqParams);
+
+        // Generate the DCC address equation.
+        pOut->equation.gfx9.num_bits = Min(32u, eq->getsize());
+        bool checked = false;
+        for (unsigned b = 0; b < pOut->equation.gfx9.num_bits; b++) {
+           CoordTerm &bit = (*eq)[b];
+
+           unsigned c;
+           for (c = 0; c < bit.getsize(); c++) {
+              Coordinate &coord = bit[c];
+              pOut->equation.gfx9.bit[b].coord[c].dim = coord.getdim();
+              pOut->equation.gfx9.bit[b].coord[c].ord = coord.getord();
+           }
+           for (; c < 5; c++)
+              pOut->equation.gfx9.bit[b].coord[c].dim = 5; /* meaning invalid */
+        }
+
+        // Reduce num_bits because DIM_M fills the rest of the bits monotonically.
+        for (int b = pOut->equation.gfx9.num_bits - 1; b >= 1; b--) {
+           CoordTerm &prev = (*eq)[b - 1];
+           CoordTerm &cur = (*eq)[b];
+
+           if (cur.getsize() == 1 && cur[0].getdim() == DIM_M &&
+               prev.getsize() == 1 && prev[0].getdim() == DIM_M &&
+               prev[0].getord() + 1 == cur[0].getord())
+              pOut->equation.gfx9.num_bits = b;
+           else
+              break;
+        }
+
+        pOut->equation.gfx9.numPipeBits = GetPipeLog2ForMetaAddressing(pIn->dccKeyFlags.pipeAligned,
+                                                                       pIn->swizzleMode);
     }
 
     return ADDR_OK;