swr/rast: reduce simd{16}vertex stack for VS output

Frontend - reduce simdvertex/simd16vertex stack usage for VS output in
ProcessDraw, fixes stack overflow in some of the deeper call stacks under
SIMD16.

1. Move the vertex store out of PA_FACTORY, and off the stack
2. Allocate the vertex store out of the aligned heap (pointer is
   temporarily stored in TLS, but will be migrated to thread pool
   along with other frontend temporary buffers).
3. Grow the vertex store as necessary for the number of verts per
   primitive, in chunks of 8/4 simdvertex/simd16vertex

Reviewed-by: Bruce Cherniak <[email protected]>

2 files changed, 54 insertions, 16 deletions

diff --git a/src/gallium/drivers/swr/rasterizer/core/frontend.cpp b/src/gallium/drivers/swr/rasterizer/core/frontend.cpp
index b09a7985ccc..3886c64ccf6 100644
--- a/src/gallium/drivers/swr/rasterizer/core/frontend.cpp
+++ b/src/gallium/drivers/swr/rasterizer/core/frontend.cpp
@@ -551,6 +551,7 @@ static void StreamOut(
 
                 _mm_store_ps((float*)pPrimDataAttrib, attrib[v]);
             }
+
             soMask &= ~(1 << slot);
         }
 
@@ -1345,8 +1346,6 @@ static void TessellationStages(
             const uint32_t numPrims_lo = std::min<uint32_t>(numPrims, KNOB_SIMD_WIDTH);
             const uint32_t numPrims_hi = std::max<uint32_t>(numPrims, KNOB_SIMD_WIDTH) - KNOB_SIMD_WIDTH;
 
-            const uint32_t primMask = GenMask(numPrims);
-
             const simd16scalari primID = _simd16_set1_epi32(dsContext.PrimitiveID);
             const simdscalari primID_lo = _simd16_extract_si(primID, 0);
             const simdscalari primID_hi = _simd16_extract_si(primID, 1);
@@ -1390,9 +1389,9 @@ static void TessellationStages(
                 if (HasRastT::value)
                 {
 #if USE_SIMD16_FRONTEND
-                    simd16vector    prim_simd16[3];
+                    simd16vector    prim_simd16[3]; // Only deal with triangles, lines, or points
 #else
-                    simdvector      prim[3]; // Only deal with triangles, lines, or points
+                    simdvector      prim[3];        // Only deal with triangles, lines, or points
 #endif
                     AR_BEGIN(FEPAAssemble, pDC->drawId);
                     bool assemble =
@@ -1407,7 +1406,7 @@ static void TessellationStages(
                     SWR_ASSERT(pfnClipFunc);
 #if USE_SIMD16_FRONTEND
                     tessPa.useAlternateOffset = false;
-                    pfnClipFunc(pDC, tessPa, workerId, prim_simd16, primMask, primID, _simd16_set1_epi32(0));
+                    pfnClipFunc(pDC, tessPa, workerId, prim_simd16, GenMask(numPrims), primID, _simd16_set1_epi32(0));
 #else
                     pfnClipFunc(pDC, tessPa, workerId, prim,
                         GenMask(tessPa.NumPrims()), _simd_set1_epi32(dsContext.PrimitiveID), _simd_set1_epi32(0));
@@ -1420,9 +1419,21 @@ static void TessellationStages(
         } // while (tessPa.HasWork())
     } // for (uint32_t p = 0; p < numPrims; ++p)
 
+#if USE_SIMD16_FRONTEND
+    if (gt_pTessellationThreadData->pDSOutput != nullptr)
+    {
+        AlignedFree(gt_pTessellationThreadData->pDSOutput);
+        gt_pTessellationThreadData->pDSOutput = nullptr;
+    }
+    gt_pTessellationThreadData->numDSOutputVectors = 0;
+
+#endif
     TSDestroyCtx(tsCtx);
 }
 
+THREAD PA_STATE::SIMDVERTEX *pVertexStore = nullptr;
+THREAD uint32_t gVertexStoreSize = 0;
+
 //////////////////////////////////////////////////////////////////////////
 /// @brief FE handler for SwrDraw.
 /// @tparam IsIndexedT - Is indexed drawing enabled
@@ -1530,8 +1541,36 @@ void ProcessDraw(
         pSoPrimData = (uint32_t*)pDC->pArena->AllocAligned(4096, 16);
     }
 
+    const uint32_t vertexCount = NumVertsPerPrim(state.topology, state.gsState.gsEnable);
+
+    SWR_ASSERT(vertexCount <= MAX_NUM_VERTS_PER_PRIM);
+
+    // grow the vertex store for the PA as necessary
+    if (gVertexStoreSize < vertexCount)
+    {
+        if (pVertexStore != nullptr)
+        {
+            AlignedFree(pVertexStore);
+        }
+
+        while (gVertexStoreSize < vertexCount)
+        {
+#if USE_SIMD16_FRONTEND
+            gVertexStoreSize += 4;  // grow in chunks of 4 simd16vertex
+#else
+            gVertexStoreSize += 8;  // grow in chunks of 8 simdvertex
+#endif
+        }
+
+        SWR_ASSERT(gVertexStoreSize <= MAX_NUM_VERTS_PER_PRIM);
+
+        pVertexStore = reinterpret_cast<PA_STATE::SIMDVERTEX *>(AlignedMalloc(gVertexStoreSize * sizeof(pVertexStore[0]), 64));
+
+        SWR_ASSERT(pVertexStore != nullptr);
+    }
+
     // choose primitive assembler
-    PA_FACTORY<IsIndexedT, IsCutIndexEnabledT> paFactory(pDC, state.topology, work.numVerts);
+    PA_FACTORY<IsIndexedT, IsCutIndexEnabledT> paFactory(pDC, state.topology, work.numVerts, pVertexStore, gVertexStoreSize);
     PA_STATE& pa = paFactory.GetPA();
 
 #if USE_SIMD16_FRONTEND
@@ -1689,8 +1728,6 @@ void ProcessDraw(
                             const uint32_t numPrims_lo = std::min<uint32_t>(numPrims, KNOB_SIMD_WIDTH);
                             const uint32_t numPrims_hi = std::max<uint32_t>(numPrims, KNOB_SIMD_WIDTH) - KNOB_SIMD_WIDTH;
 
-                            const uint32_t primMask = GenMask(numPrims);
-
                             const simd16scalari primID = pa.GetPrimID(work.startPrimID);
                             const simdscalari primID_lo = _simd16_extract_si(primID, 0);
                             const simdscalari primID_hi = _simd16_extract_si(primID, 1);
@@ -1732,7 +1769,7 @@ void ProcessDraw(
                                         StreamOut(pDC, pa, workerId, pSoPrimData, numPrims_hi, 0);
                                     }
 #else
-                                    pa.useAlternateOffset = false;  // StreamOut() is SIMD16-compatible..
+                                    pa.useAlternateOffset = false;
                                     StreamOut(pDC, pa, workerId, pSoPrimData, 0);
 #endif
                                 }
@@ -1742,7 +1779,7 @@ void ProcessDraw(
                                     SWR_ASSERT(pDC->pState->pfnProcessPrims_simd16);
 
                                     pa.useAlternateOffset = false;
-                                    pDC->pState->pfnProcessPrims_simd16(pDC, pa, workerId, prim_simd16, primMask, primID, _simd16_setzero_si());
+                                    pDC->pState->pfnProcessPrims_simd16(pDC, pa, workerId, prim_simd16, GenMask(numPrims), primID, _simd16_setzero_si());
                                 }
                             }
                         }
diff --git a/src/gallium/drivers/swr/rasterizer/core/pa.h b/src/gallium/drivers/swr/rasterizer/core/pa.h
index 10570f43f0f..403efe057d0 100644
--- a/src/gallium/drivers/swr/rasterizer/core/pa.h
+++ b/src/gallium/drivers/swr/rasterizer/core/pa.h
@@ -136,9 +136,9 @@ struct PA_STATE_OPT : public PA_STATE
     uint32_t primIDIncr{ 0 };            // how much to increment for each vector (typically vector / {1, 2})
     SIMDSCALARI primID;
 
-    typedef bool(*PFN_PA_FUNC)(PA_STATE_OPT& state, uint32_t slot, simdvector verts[]);
+    typedef bool(*PFN_PA_FUNC)(PA_STATE_OPT& pa, uint32_t slot, simdvector verts[]);
 #if ENABLE_AVX512_SIMD16
-    typedef bool(*PFN_PA_FUNC_SIMD16)(PA_STATE_OPT& state, uint32_t slot, simd16vector verts[]);
+    typedef bool(*PFN_PA_FUNC_SIMD16)(PA_STATE_OPT& pa, uint32_t slot, simd16vector verts[]);
 #endif
     typedef void(*PFN_PA_SINGLE_FUNC)(PA_STATE_OPT& pa, uint32_t slot, uint32_t primIndex, __m128 verts[]);
 
@@ -691,6 +691,7 @@ PRAGMA_WARNING_PUSH_DISABLE(4789)
                 pBase += SIMD_WIDTH;
             }
         }
+
         return true;
     }
 PRAGMA_WARNING_POP()
@@ -1392,7 +1393,7 @@ private:
 template <typename IsIndexedT, typename IsCutIndexEnabledT>
 struct PA_FACTORY
 {
-    PA_FACTORY(DRAW_CONTEXT* pDC, PRIMITIVE_TOPOLOGY in_topo, uint32_t numVerts) : topo(in_topo)
+    PA_FACTORY(DRAW_CONTEXT* pDC, PRIMITIVE_TOPOLOGY in_topo, uint32_t numVerts, PA_STATE::SIMDVERTEX *pVertexStore, uint32_t vertexStoreSize) : topo(in_topo)
     {
 #if KNOB_ENABLE_CUT_AWARE_PA == TRUE
         const API_STATE& state = GetApiState(pDC);
@@ -1408,7 +1409,7 @@ struct PA_FACTORY
             memset(&indexStore, 0, sizeof(indexStore));
             uint32_t numAttribs = state.feNumAttributes;
 
-            new (&this->paCut) PA_STATE_CUT(pDC, (uint8_t*)&this->vertexStore[0], MAX_NUM_VERTS_PER_PRIM * PA_STATE::SIMD_WIDTH,
+            new (&this->paCut) PA_STATE_CUT(pDC, reinterpret_cast<uint8_t *>(pVertexStore), vertexStoreSize * PA_STATE::SIMD_WIDTH,
                 &this->indexStore[0], numVerts, numAttribs, state.topology, false);
             cutPA = true;
         }
@@ -1416,7 +1417,7 @@ struct PA_FACTORY
 #endif
         {
             uint32_t numPrims = GetNumPrims(in_topo, numVerts);
-            new (&this->paOpt) PA_STATE_OPT(pDC, numPrims, (uint8_t*)&this->vertexStore[0], MAX_NUM_VERTS_PER_PRIM * PA_STATE::SIMD_WIDTH, false);
+            new (&this->paOpt) PA_STATE_OPT(pDC, numPrims, reinterpret_cast<uint8_t *>(pVertexStore), vertexStoreSize * PA_STATE::SIMD_WIDTH, false);
             cutPA = false;
         }
 
@@ -1438,10 +1439,10 @@ struct PA_FACTORY
 
     PA_STATE_OPT paOpt;
     PA_STATE_CUT paCut;
+
     bool cutPA{ false };
 
     PRIMITIVE_TOPOLOGY topo{ TOP_UNKNOWN };
 
-    PA_STATE::SIMDVERTEX    vertexStore[MAX_NUM_VERTS_PER_PRIM];
     PA_STATE::SIMDMASK      indexStore[MAX_NUM_VERTS_PER_PRIM];
 };