[Mesa-dev] [PATCH 9/9] swr: [rasterizer core] SIMD16 Frontend WIP

Tim Rowley timothy.o.rowley at intel.com
Tue Apr 4 18:16:24 UTC 2017


Implement widened binner for SIMD16
---
 .../drivers/swr/rasterizer/common/simd16intrin.h   |   44 +-
 src/gallium/drivers/swr/rasterizer/core/binner.cpp | 1642 +++++++++++++++++---
 src/gallium/drivers/swr/rasterizer/core/frontend.h |   98 ++
 src/gallium/drivers/swr/rasterizer/core/utils.h    |   10 +
 4 files changed, 1509 insertions(+), 285 deletions(-)

diff --git a/src/gallium/drivers/swr/rasterizer/common/simd16intrin.h b/src/gallium/drivers/swr/rasterizer/common/simd16intrin.h
index 3b43d51..e5c34c2 100644
--- a/src/gallium/drivers/swr/rasterizer/common/simd16intrin.h
+++ b/src/gallium/drivers/swr/rasterizer/common/simd16intrin.h
@@ -46,10 +46,6 @@ struct simd16scalari
 };
 typedef uint16_t simd16mask;
 
-#define _simd16_masklo(mask) ((mask) & 0xFF)
-#define _simd16_maskhi(mask) (((mask) >> 8))
-#define _simd16_setmask(hi, lo) (((hi) << 8) | (lo))
-
 #else
 typedef __m512 simd16scalar;
 typedef __m512d simd16scalard;
@@ -60,6 +56,10 @@ typedef __mmask16 simd16mask;
 #error Unsupported vector width
 #endif//KNOB_SIMD16_WIDTH == 16
 
+#define _simd16_masklo(mask) ((mask) & 0xFF)
+#define _simd16_maskhi(mask) (((mask) >> 8) & 0xFF)
+#define _simd16_setmask(hi, lo) (((hi) << 8) | (lo))
+
 OSALIGN(union, KNOB_SIMD16_BYTES) simd16vector
 {
     simd16scalar  v[4];
@@ -383,32 +383,26 @@ SIMD16_EMU_AVX512_2(simd16scalar, _simd16_max_ps, _mm256_max_ps)
 
 INLINE simd16mask _simd16_movemask_ps(simd16scalar a)
 {
-    simd16mask mask;
+    simdmask mask_lo = _mm256_movemask_ps(a.lo);
+    simdmask mask_hi = _mm256_movemask_ps(a.hi);
 
-    reinterpret_cast<uint8_t *>(&mask)[0] = _mm256_movemask_ps(a.lo);
-    reinterpret_cast<uint8_t *>(&mask)[1] = _mm256_movemask_ps(a.hi);
-
-    return mask;
+    return static_cast<simd16mask>(mask_lo) | (static_cast<simd16mask>(mask_hi) << 8);
 }
 
 INLINE simd16mask _simd16_movemask_pd(simd16scalard a)
 {
-    simd16mask mask;
+    simdmask mask_lo = _mm256_movemask_pd(a.lo);
+    simdmask mask_hi = _mm256_movemask_pd(a.hi);
 
-    reinterpret_cast<uint8_t *>(&mask)[0] = _mm256_movemask_pd(a.lo);
-    reinterpret_cast<uint8_t *>(&mask)[1] = _mm256_movemask_pd(a.hi);
-
-    return mask;
+    return static_cast<simd16mask>(mask_lo) | (static_cast<simd16mask>(mask_hi) << 4);
 }
 
-INLINE simd16mask _simd16_movemask_epi8(simd16scalari a)
+INLINE uint64_t _simd16_movemask_epi8(simd16scalari a)
 {
-    simd16mask mask;
+    uint32_t mask_lo = _mm256_movemask_epi8(a.lo);
+    uint32_t mask_hi = _mm256_movemask_epi8(a.hi);
 
-    reinterpret_cast<uint8_t *>(&mask)[0] = _mm256_movemask_epi8(a.lo);
-    reinterpret_cast<uint8_t *>(&mask)[1] = _mm256_movemask_epi8(a.hi);
-
-    return mask;
+    return static_cast<uint64_t>(mask_lo) | (static_cast<uint64_t>(mask_hi) << 32);
 }
 
 INLINE simd16scalari _simd16_cvtps_epi32(simd16scalar a)
@@ -809,12 +803,10 @@ INLINE simd16mask _simd16_scalari2mask(simd16scalari mask)
     return _mm512_cmpneq_epu32_mask(mask, _mm512_setzero_epi32());
 }
 
-#if 0
 INLINE simd16mask _simd16_scalard2mask(simd16scalard mask)
 {
-    return _mm512_cmpneq_epu64_mask(mask, _mm512_setzero_epi64());
+    return _mm512_cmpneq_epu64_mask(_mm512_castpd_si512(mask), _mm512_setzero_si512());
 }
-#endif
 
 #define _simd16_setzero_ps      _mm512_setzero_ps
 #define _simd16_setzero_si      _mm512_setzero_si512
@@ -889,6 +881,7 @@ INLINE simd16scalari _simd16_blendv_epi32(simd16scalari a, simd16scalari b, cons
 }
 
 #define _simd16_mul_ps          _mm512_mul_ps
+#define _simd16_div_ps          _mm512_div_ps
 #define _simd16_add_ps          _mm512_add_ps
 #define _simd16_sub_ps          _mm512_sub_ps
 #define _simd16_rsqrt_ps        _mm512_rsqrt14_ps
@@ -900,12 +893,10 @@ INLINE simd16mask _simd16_movemask_ps(simd16scalar a)
     return  _simd16_scalari2mask(_mm512_castps_si512(a));
 }
 
-#if 0
 INLINE simd16mask _simd16_movemask_pd(simd16scalard a)
 {
-    return  _simd16_scalard2mask(_mm512i_castpd_si512(a));
+    return  _simd16_scalard2mask(a);
 }
-#endif
 
 #if 0
 INLINE int _simd16_movemask_epi8(simd16scalari a)
@@ -1040,7 +1031,6 @@ INLINE simd16scalar _simd16_mask_i32gather_ps_temp(simd16scalar a, const float *
 #define _simd16_mask_i32gather_ps(a, m, index, mask, scale) _simd16_mask_i32gather_ps_temp<scale>(a, m, index, mask)
 
 #define _simd16_abs_epi32         _mm512_abs_epi32
-#define _simd16_cmpeq_epi64       _mm512_abs_epi32
 
 INLINE simd16scalari _simd16_cmpeq_epi64(simd16scalari a, simd16scalari b)
 {
diff --git a/src/gallium/drivers/swr/rasterizer/core/binner.cpp b/src/gallium/drivers/swr/rasterizer/core/binner.cpp
index f00701f..9ec5bea 100644
--- a/src/gallium/drivers/swr/rasterizer/core/binner.cpp
+++ b/src/gallium/drivers/swr/rasterizer/core/binner.cpp
@@ -37,15 +37,27 @@
 // Function Prototype
 void BinPostSetupLines(DRAW_CONTEXT *pDC, PA_STATE& pa, uint32_t workerId, simdvector prims[3], simdscalar vRecipW[2], uint32_t primMask, simdscalari primID, simdscalari viewportIdx);
 
+#if USE_SIMD16_FRONTEND
+void BinPostSetupLines_simd16(DRAW_CONTEXT *pDC, PA_STATE& pa, uint32_t workerId, simd16vector prims[3], simd16scalar vRecipW[2], uint32_t primMask, simd16scalari primID, simd16scalari viewportIdx);
+#endif
+
 //////////////////////////////////////////////////////////////////////////
 /// @brief Offsets added to post-viewport vertex positions based on
 /// raster state.
 static const simdscalar g_pixelOffsets[SWR_PIXEL_LOCATION_UL + 1] =
 {
-    _simd_set1_ps(0.0f), // SWR_PIXEL_LOCATION_CENTER
-    _simd_set1_ps(0.5f), // SWR_PIXEL_LOCATION_UL
+    _simd_set1_ps(0.0f),    // SWR_PIXEL_LOCATION_CENTER
+    _simd_set1_ps(0.5f),    // SWR_PIXEL_LOCATION_UL
+};
+
+#if USE_SIMD16_FRONTEND
+static const simd16scalar g_pixelOffsets_simd16[SWR_PIXEL_LOCATION_UL + 1] =
+{
+    _simd16_set1_ps(0.0f),  // SWR_PIXEL_LOCATION_CENTER
+    _simd16_set1_ps(0.5f),  // SWR_PIXEL_LOCATION_UL
 };
 
+#endif
 //////////////////////////////////////////////////////////////////////////
 /// @brief Convert the X,Y coords of a triangle to the requested Fixed 
 /// Point precision from FP32.
@@ -56,6 +68,15 @@ INLINE simdscalari fpToFixedPointVertical(const simdscalar vIn)
     return _simd_cvtps_epi32(vFixed);
 }
 
+#if USE_SIMD16_FRONTEND
+template <typename PT = FixedPointTraits<Fixed_16_8>>
+INLINE simd16scalari fpToFixedPointVertical(const simd16scalar vIn)
+{
+    simd16scalar vFixed = _simd16_mul_ps(vIn, _simd16_set1_ps(PT::ScaleT::value));
+    return _simd16_cvtps_epi32(vFixed);
+}
+
+#endif
 //////////////////////////////////////////////////////////////////////////
 /// @brief Helper function to set the X,Y coords of a triangle to the 
 /// requested Fixed Point precision from FP32.
@@ -72,6 +93,18 @@ INLINE static void FPToFixedPoint(const simdvector * const tri, simdscalari(&vXi
     vYi[2] = fpToFixedPointVertical(tri[2].y);
 }
 
+#if USE_SIMD16_FRONTEND
+INLINE static void FPToFixedPoint(const simd16vector * const tri, simd16scalari(&vXi)[3], simd16scalari(&vYi)[3])
+{
+    vXi[0] = fpToFixedPointVertical(tri[0].x);
+    vYi[0] = fpToFixedPointVertical(tri[0].y);
+    vXi[1] = fpToFixedPointVertical(tri[1].x);
+    vYi[1] = fpToFixedPointVertical(tri[1].y);
+    vXi[2] = fpToFixedPointVertical(tri[2].x);
+    vYi[2] = fpToFixedPointVertical(tri[2].y);
+}
+
+#endif
 //////////////////////////////////////////////////////////////////////////
 /// @brief Calculate bounding box for current triangle
 /// @tparam CT: ConservativeRastFETraits type
@@ -105,6 +138,37 @@ INLINE void calcBoundingBoxIntVertical(const simdvector * const tri, simdscalari
     bbox.ymax = vMaxY;
 }
 
+#if USE_SIMD16_FRONTEND
+template <typename CT>
+INLINE void calcBoundingBoxIntVertical(const simd16vector * const tri, simd16scalari(&vX)[3], simd16scalari(&vY)[3], simd16BBox &bbox)
+{
+    simd16scalari vMinX = vX[0];
+
+    vMinX = _simd16_min_epi32(vMinX, vX[1]);
+    vMinX = _simd16_min_epi32(vMinX, vX[2]);
+
+    simd16scalari vMaxX = vX[0];
+
+    vMaxX = _simd16_max_epi32(vMaxX, vX[1]);
+    vMaxX = _simd16_max_epi32(vMaxX, vX[2]);
+
+    simd16scalari vMinY = vY[0];
+
+    vMinY = _simd16_min_epi32(vMinY, vY[1]);
+    vMinY = _simd16_min_epi32(vMinY, vY[2]);
+
+    simd16scalari vMaxY = vY[0];
+
+    vMaxY = _simd16_max_epi32(vMaxY, vY[1]);
+    vMaxY = _simd16_max_epi32(vMaxY, vY[2]);
+
+    bbox.xmin = vMinX;
+    bbox.xmax = vMaxX;
+    bbox.ymin = vMinY;
+    bbox.ymax = vMaxY;
+}
+
+#endif
 //////////////////////////////////////////////////////////////////////////
 /// @brief FEConservativeRastT specialization of calcBoundingBoxIntVertical
 /// Offsets BBox for conservative rast
@@ -356,6 +420,96 @@ struct GatherScissors<8>
     }
 };
 
+#if USE_SIMD16_FRONTEND
+template<size_t SimdWidth>
+struct GatherScissors_simd16
+{
+    static void Gather(const SWR_RECT* pScissorsInFixedPoint, const uint32_t* pViewportIndex,
+        simd16scalari &scisXmin, simd16scalari &scisYmin,
+        simd16scalari &scisXmax, simd16scalari &scisYmax)
+    {
+        SWR_INVALID("Unhandled Simd Width in Scissor Rect Gather");
+    }
+};
+
+template<>
+struct GatherScissors_simd16<16>
+{
+    static void Gather(const SWR_RECT* pScissorsInFixedPoint, const uint32_t* pViewportIndex,
+        simd16scalari &scisXmin, simd16scalari &scisYmin,
+        simd16scalari &scisXmax, simd16scalari &scisYmax)
+    {
+        scisXmin = _simd16_set_epi32(pScissorsInFixedPoint[pViewportIndex[0]].xmin,
+            pScissorsInFixedPoint[pViewportIndex[1]].xmin,
+            pScissorsInFixedPoint[pViewportIndex[2]].xmin,
+            pScissorsInFixedPoint[pViewportIndex[3]].xmin,
+            pScissorsInFixedPoint[pViewportIndex[4]].xmin,
+            pScissorsInFixedPoint[pViewportIndex[5]].xmin,
+            pScissorsInFixedPoint[pViewportIndex[6]].xmin,
+            pScissorsInFixedPoint[pViewportIndex[7]].xmin,
+            pScissorsInFixedPoint[pViewportIndex[8]].xmin,
+            pScissorsInFixedPoint[pViewportIndex[9]].xmin,
+            pScissorsInFixedPoint[pViewportIndex[10]].xmin,
+            pScissorsInFixedPoint[pViewportIndex[11]].xmin,
+            pScissorsInFixedPoint[pViewportIndex[12]].xmin,
+            pScissorsInFixedPoint[pViewportIndex[13]].xmin,
+            pScissorsInFixedPoint[pViewportIndex[14]].xmin,
+            pScissorsInFixedPoint[pViewportIndex[15]].xmin);
+
+        scisYmin = _simd16_set_epi32(pScissorsInFixedPoint[pViewportIndex[0]].ymin,
+            pScissorsInFixedPoint[pViewportIndex[1]].ymin,
+            pScissorsInFixedPoint[pViewportIndex[2]].ymin,
+            pScissorsInFixedPoint[pViewportIndex[3]].ymin,
+            pScissorsInFixedPoint[pViewportIndex[4]].ymin,
+            pScissorsInFixedPoint[pViewportIndex[5]].ymin,
+            pScissorsInFixedPoint[pViewportIndex[6]].ymin,
+            pScissorsInFixedPoint[pViewportIndex[7]].ymin,
+            pScissorsInFixedPoint[pViewportIndex[8]].ymin,
+            pScissorsInFixedPoint[pViewportIndex[9]].ymin,
+            pScissorsInFixedPoint[pViewportIndex[10]].ymin,
+            pScissorsInFixedPoint[pViewportIndex[11]].ymin,
+            pScissorsInFixedPoint[pViewportIndex[12]].ymin,
+            pScissorsInFixedPoint[pViewportIndex[13]].ymin,
+            pScissorsInFixedPoint[pViewportIndex[14]].ymin,
+            pScissorsInFixedPoint[pViewportIndex[15]].ymin);
+
+        scisXmax = _simd16_set_epi32(pScissorsInFixedPoint[pViewportIndex[0]].xmax,
+            pScissorsInFixedPoint[pViewportIndex[1]].xmax,
+            pScissorsInFixedPoint[pViewportIndex[2]].xmax,
+            pScissorsInFixedPoint[pViewportIndex[3]].xmax,
+            pScissorsInFixedPoint[pViewportIndex[4]].xmax,
+            pScissorsInFixedPoint[pViewportIndex[5]].xmax,
+            pScissorsInFixedPoint[pViewportIndex[6]].xmax,
+            pScissorsInFixedPoint[pViewportIndex[7]].xmax,
+            pScissorsInFixedPoint[pViewportIndex[8]].xmax,
+            pScissorsInFixedPoint[pViewportIndex[9]].xmax,
+            pScissorsInFixedPoint[pViewportIndex[10]].xmax,
+            pScissorsInFixedPoint[pViewportIndex[11]].xmax,
+            pScissorsInFixedPoint[pViewportIndex[12]].xmax,
+            pScissorsInFixedPoint[pViewportIndex[13]].xmax,
+            pScissorsInFixedPoint[pViewportIndex[14]].xmax,
+            pScissorsInFixedPoint[pViewportIndex[15]].xmax);
+
+        scisYmax = _simd16_set_epi32(pScissorsInFixedPoint[pViewportIndex[0]].ymax,
+            pScissorsInFixedPoint[pViewportIndex[1]].ymax,
+            pScissorsInFixedPoint[pViewportIndex[2]].ymax,
+            pScissorsInFixedPoint[pViewportIndex[3]].ymax,
+            pScissorsInFixedPoint[pViewportIndex[4]].ymax,
+            pScissorsInFixedPoint[pViewportIndex[5]].ymax,
+            pScissorsInFixedPoint[pViewportIndex[6]].ymax,
+            pScissorsInFixedPoint[pViewportIndex[7]].ymax,
+            pScissorsInFixedPoint[pViewportIndex[8]].ymax,
+            pScissorsInFixedPoint[pViewportIndex[9]].ymax,
+            pScissorsInFixedPoint[pViewportIndex[10]].ymax,
+            pScissorsInFixedPoint[pViewportIndex[11]].ymax,
+            pScissorsInFixedPoint[pViewportIndex[12]].ymax,
+            pScissorsInFixedPoint[pViewportIndex[13]].ymax,
+            pScissorsInFixedPoint[pViewportIndex[14]].ymax,
+            pScissorsInFixedPoint[pViewportIndex[15]].ymax);
+    }
+};
+
+#endif
 typedef void(*PFN_PROCESS_ATTRIBUTES)(DRAW_CONTEXT*, PA_STATE&, uint32_t, uint32_t, float*);
 
 struct ProcessAttributesChooser
@@ -873,184 +1027,735 @@ void BinTriangles_simd16(
     simd16scalari primID,
     simd16scalari viewportIdx)
 {
-    enum { VERTS_PER_PRIM = 3 };
-
-    simdvector verts[VERTS_PER_PRIM];
-
-    for (uint32_t i = 0; i < VERTS_PER_PRIM; i += 1)
-    {
-        for (uint32_t j = 0; j < 4; j += 1)
-        {
-            verts[i][j] = _simd16_extract_ps(tri[i][j], 0);
-        }
-    }
+    SWR_CONTEXT *pContext = pDC->pContext;
 
-    pa.useAlternateOffset = false;
-    BinTriangles<CT>(pDC, pa, workerId, verts, GetPrimMaskLo(triMask), _simd16_extract_si(primID, 0), _simd16_extract_si(viewportIdx, 0));
+    AR_BEGIN(FEBinTriangles, pDC->drawId);
 
-    if (GetPrimMaskHi(triMask))
-    {
-        for (uint32_t i = 0; i < VERTS_PER_PRIM; i += 1)
-        {
-            for (uint32_t j = 0; j < 4; j += 1)
-            {
-                verts[i][j] = _simd16_extract_ps(tri[i][j], 1);
-            }
-        }
+    const API_STATE& state = GetApiState(pDC);
+    const SWR_RASTSTATE& rastState = state.rastState;
+    const SWR_FRONTEND_STATE& feState = state.frontendState;
+    const SWR_GS_STATE& gsState = state.gsState;
 
-        pa.useAlternateOffset = true;
-        BinTriangles<CT>(pDC, pa, workerId, verts, GetPrimMaskHi(triMask), _simd16_extract_si(primID, 1), _simd16_extract_si(viewportIdx, 1));
-    }
-}
+    MacroTileMgr *pTileMgr = pDC->pTileMgr;
 
-#endif
-struct FEBinTrianglesChooser
-{
-    typedef PFN_PROCESS_PRIMS FuncType;
+    simd16scalar vRecipW0 = _simd16_set1_ps(1.0f);
+    simd16scalar vRecipW1 = _simd16_set1_ps(1.0f);
+    simd16scalar vRecipW2 = _simd16_set1_ps(1.0f);
 
-    template <typename... ArgsB>
-    static FuncType GetFunc()
+    if (feState.vpTransformDisable)
     {
-        return BinTriangles<ConservativeRastFETraits<ArgsB...>>;
+        // RHW is passed in directly when VP transform is disabled
+        vRecipW0 = tri[0].v[3];
+        vRecipW1 = tri[1].v[3];
+        vRecipW2 = tri[2].v[3];
     }
-};
-
-// Selector for correct templated BinTrinagles function
-PFN_PROCESS_PRIMS GetBinTrianglesFunc(bool IsConservative)
-{
-    return TemplateArgUnroller<FEBinTrianglesChooser>::GetFunc(IsConservative);
-}
-
-#if USE_SIMD16_FRONTEND
-struct FEBinTrianglesChooser_simd16
-{
-    typedef PFN_PROCESS_PRIMS_SIMD16 FuncType;
-
-    template <typename... ArgsB>
-    static FuncType GetFunc()
+    else
     {
-        return BinTriangles_simd16<ConservativeRastFETraits<ArgsB...>>;
-    }
-};
-
-// Selector for correct templated BinTrinagles function
-PFN_PROCESS_PRIMS_SIMD16 GetBinTrianglesFunc_simd16(bool IsConservative)
-{
-    return TemplateArgUnroller<FEBinTrianglesChooser_simd16>::GetFunc(IsConservative);
-}
-
-#endif
-
-//////////////////////////////////////////////////////////////////////////
-/// @brief Bin SIMD points to the backend.  Only supports point size of 1
-/// @param pDC - pointer to draw context.
-/// @param pa - The primitive assembly object.
-/// @param workerId - thread's worker id. Even thread has a unique id.
-/// @param tri - Contains point position data for SIMDs worth of points.
-/// @param primID - Primitive ID for each point.
-void BinPoints(
-    DRAW_CONTEXT *pDC,
-    PA_STATE& pa,
-    uint32_t workerId,
-    simdvector prim[3],
-    uint32_t primMask,
-    simdscalari primID,
-    simdscalari viewportIdx)
-{
-    SWR_CONTEXT *pContext = pDC->pContext;
-
-    AR_BEGIN(FEBinPoints, pDC->drawId);
-
-    simdvector& primVerts = prim[0];
+        // Perspective divide
+        vRecipW0 = _simd16_div_ps(_simd16_set1_ps(1.0f), tri[0].w);
+        vRecipW1 = _simd16_div_ps(_simd16_set1_ps(1.0f), tri[1].w);
+        vRecipW2 = _simd16_div_ps(_simd16_set1_ps(1.0f), tri[2].w);
 
-    const API_STATE& state = GetApiState(pDC);
-    const SWR_FRONTEND_STATE& feState = state.frontendState;
-    const SWR_GS_STATE& gsState = state.gsState;
-    const SWR_RASTSTATE& rastState = state.rastState;
-    const uint32_t *pViewportIndex = (uint32_t *)&viewportIdx;
+        tri[0].v[0] = _simd16_mul_ps(tri[0].v[0], vRecipW0);
+        tri[1].v[0] = _simd16_mul_ps(tri[1].v[0], vRecipW1);
+        tri[2].v[0] = _simd16_mul_ps(tri[2].v[0], vRecipW2);
 
-    // Select attribute processor
-    PFN_PROCESS_ATTRIBUTES pfnProcessAttribs = GetProcessAttributesFunc(1,
-        state.backendState.swizzleEnable, state.backendState.constantInterpolationMask);
+        tri[0].v[1] = _simd16_mul_ps(tri[0].v[1], vRecipW0);
+        tri[1].v[1] = _simd16_mul_ps(tri[1].v[1], vRecipW1);
+        tri[2].v[1] = _simd16_mul_ps(tri[2].v[1], vRecipW2);
 
-    if (!feState.vpTransformDisable)
-    {
-        // perspective divide
-        simdscalar vRecipW0 = _simd_div_ps(_simd_set1_ps(1.0f), primVerts.w);
-        primVerts.x = _simd_mul_ps(primVerts.x, vRecipW0);
-        primVerts.y = _simd_mul_ps(primVerts.y, vRecipW0);
-        primVerts.z = _simd_mul_ps(primVerts.z, vRecipW0);
+        tri[0].v[2] = _simd16_mul_ps(tri[0].v[2], vRecipW0);
+        tri[1].v[2] = _simd16_mul_ps(tri[1].v[2], vRecipW1);
+        tri[2].v[2] = _simd16_mul_ps(tri[2].v[2], vRecipW2);
 
-        // viewport transform to screen coords
+        // Viewport transform to screen space coords
         if (state.gsState.emitsViewportArrayIndex)
         {
-            viewportTransform<1>(&primVerts, state.vpMatrices, viewportIdx);
+            viewportTransform<3>(tri, state.vpMatrices, viewportIdx);
         }
         else
         {
-            viewportTransform<1>(&primVerts, state.vpMatrices);
+            viewportTransform<3>(tri, state.vpMatrices);
         }
     }
 
-    // adjust for pixel center location
-    simdscalar offset = g_pixelOffsets[rastState.pixelLocation];
-    primVerts.x = _simd_add_ps(primVerts.x, offset);
-    primVerts.y = _simd_add_ps(primVerts.y, offset);
+    // Adjust for pixel center location
+    const simd16scalar offset = g_pixelOffsets_simd16[rastState.pixelLocation];
 
-    // convert to fixed point
-    simdscalari vXi, vYi;
-    vXi = fpToFixedPointVertical(primVerts.x);
-    vYi = fpToFixedPointVertical(primVerts.y);
+    tri[0].x = _simd16_add_ps(tri[0].x, offset);
+    tri[0].y = _simd16_add_ps(tri[0].y, offset);
 
-    if (CanUseSimplePoints(pDC))
-    {
-        // adjust for ymin-xmin rule
-        vXi = _simd_sub_epi32(vXi, _simd_set1_epi32(1));
-        vYi = _simd_sub_epi32(vYi, _simd_set1_epi32(1));
+    tri[1].x = _simd16_add_ps(tri[1].x, offset);
+    tri[1].y = _simd16_add_ps(tri[1].y, offset);
 
-        // cull points off the ymin-xmin edge of the viewport
-        primMask &= ~_simd_movemask_ps(_simd_castsi_ps(vXi));
-        primMask &= ~_simd_movemask_ps(_simd_castsi_ps(vYi));
+    tri[2].x = _simd16_add_ps(tri[2].x, offset);
+    tri[2].y = _simd16_add_ps(tri[2].y, offset);
 
-        // compute macro tile coordinates 
-        simdscalari macroX = _simd_srai_epi32(vXi, KNOB_MACROTILE_X_DIM_FIXED_SHIFT);
-        simdscalari macroY = _simd_srai_epi32(vYi, KNOB_MACROTILE_Y_DIM_FIXED_SHIFT);
+    simd16scalari vXi[3], vYi[3];
 
-        OSALIGNSIMD(uint32_t) aMacroX[KNOB_SIMD_WIDTH], aMacroY[KNOB_SIMD_WIDTH];
-        _simd_store_si((simdscalari*)aMacroX, macroX);
-        _simd_store_si((simdscalari*)aMacroY, macroY);
+    // Set vXi, vYi to required fixed point precision
+    FPToFixedPoint(tri, vXi, vYi);
 
-        // compute raster tile coordinates
-        simdscalari rasterX = _simd_srai_epi32(vXi, KNOB_TILE_X_DIM_SHIFT + FIXED_POINT_SHIFT);
-        simdscalari rasterY = _simd_srai_epi32(vYi, KNOB_TILE_Y_DIM_SHIFT + FIXED_POINT_SHIFT);
+    // triangle setup
+    simd16scalari vAi[3], vBi[3];
+    triangleSetupABIntVertical(vXi, vYi, vAi, vBi);
 
-        // compute raster tile relative x,y for coverage mask
-        simdscalari tileAlignedX = _simd_slli_epi32(rasterX, KNOB_TILE_X_DIM_SHIFT);
-        simdscalari tileAlignedY = _simd_slli_epi32(rasterY, KNOB_TILE_Y_DIM_SHIFT);
+    // determinant
+    simd16scalari vDet[2];
+    calcDeterminantIntVertical(vAi, vBi, vDet);
 
-        simdscalari tileRelativeX = _simd_sub_epi32(_simd_srai_epi32(vXi, FIXED_POINT_SHIFT), tileAlignedX);
-        simdscalari tileRelativeY = _simd_sub_epi32(_simd_srai_epi32(vYi, FIXED_POINT_SHIFT), tileAlignedY);
+    // cull zero area
+    simd16mask maskLo = _simd16_movemask_pd(_simd16_castsi_pd(_simd16_cmpeq_epi64(vDet[0], _simd16_setzero_si())));
+    simd16mask maskHi = _simd16_movemask_pd(_simd16_castsi_pd(_simd16_cmpeq_epi64(vDet[1], _simd16_setzero_si())));
 
-        OSALIGNSIMD(uint32_t) aTileRelativeX[KNOB_SIMD_WIDTH];
-        OSALIGNSIMD(uint32_t) aTileRelativeY[KNOB_SIMD_WIDTH];
-        _simd_store_si((simdscalari*)aTileRelativeX, tileRelativeX);
-        _simd_store_si((simdscalari*)aTileRelativeY, tileRelativeY);
+    uint32_t cullZeroAreaMask = maskLo | (maskHi << (KNOB_SIMD16_WIDTH / 2));
 
-        OSALIGNSIMD(uint32_t) aTileAlignedX[KNOB_SIMD_WIDTH];
-        OSALIGNSIMD(uint32_t) aTileAlignedY[KNOB_SIMD_WIDTH];
-        _simd_store_si((simdscalari*)aTileAlignedX, tileAlignedX);
-        _simd_store_si((simdscalari*)aTileAlignedY, tileAlignedY);
+    // don't cull degenerate triangles if we're conservatively rasterizing
+    uint32_t origTriMask = triMask;
+    if (rastState.fillMode == SWR_FILLMODE_SOLID && !CT::IsConservativeT::value)
+    {
+        triMask &= ~cullZeroAreaMask;
+    }
 
-        OSALIGNSIMD(float) aZ[KNOB_SIMD_WIDTH];
-        _simd_store_ps((float*)aZ, primVerts.z);
+    // determine front winding tris
+    // CW  +det
+    // CCW det < 0;
+    // 0 area triangles are marked as backfacing regardless of winding order,
+    // which is required behavior for conservative rast and wireframe rendering
+    uint32_t frontWindingTris;
+    if (rastState.frontWinding == SWR_FRONTWINDING_CW)
+    {
+        maskLo = _simd16_movemask_pd(_simd16_castsi_pd(_simd16_cmpgt_epi64(vDet[0], _simd16_setzero_si())));
+        maskHi = _simd16_movemask_pd(_simd16_castsi_pd(_simd16_cmpgt_epi64(vDet[1], _simd16_setzero_si())));
+    }
+    else
+    {
+        maskLo = _simd16_movemask_pd(_simd16_castsi_pd(_simd16_cmpgt_epi64(_simd16_setzero_si(), vDet[0])));
+        maskHi = _simd16_movemask_pd(_simd16_castsi_pd(_simd16_cmpgt_epi64(_simd16_setzero_si(), vDet[1])));
+    }
 
-        // store render target array index
-        OSALIGNSIMD(uint32_t) aRTAI[KNOB_SIMD_WIDTH];
-        if (gsState.gsEnable && gsState.emitsRenderTargetArrayIndex)
+    frontWindingTris = maskLo | (maskHi << (KNOB_SIMD16_WIDTH / 2));
+
+    // cull
+    uint32_t cullTris;
+    switch ((SWR_CULLMODE)rastState.cullMode)
+    {
+    case SWR_CULLMODE_BOTH:  cullTris = 0xffffffff; break;
+    case SWR_CULLMODE_NONE:  cullTris = 0x0; break;
+    case SWR_CULLMODE_FRONT: cullTris = frontWindingTris; break;
+        // 0 area triangles are marked as backfacing, which is required behavior for conservative rast
+    case SWR_CULLMODE_BACK:  cullTris = ~frontWindingTris; break;
+    default: SWR_INVALID("Invalid cull mode: %d", rastState.cullMode); cullTris = 0x0; break;
+    }
+
+    triMask &= ~cullTris;
+
+    if (origTriMask ^ triMask)
+    {
+        RDTSC_EVENT(FECullZeroAreaAndBackface, _mm_popcnt_u32(origTriMask ^ triMask), 0);
+    }
+
+    // Simple non-conformant wireframe mode, useful for debugging
+    if (rastState.fillMode == SWR_FILLMODE_WIREFRAME)
+    {
+        // construct 3 SIMD lines out of the triangle and call the line binner for each SIMD
+        simd16vector line[2];
+        simd16scalar recipW[2];
+        line[0] = tri[0];
+        line[1] = tri[1];
+        recipW[0] = vRecipW0;
+        recipW[1] = vRecipW1;
+        BinPostSetupLines_simd16(pDC, pa, workerId, line, recipW, triMask, primID, viewportIdx);
+
+        line[0] = tri[1];
+        line[1] = tri[2];
+        recipW[0] = vRecipW1;
+        recipW[1] = vRecipW2;
+        BinPostSetupLines_simd16(pDC, pa, workerId, line, recipW, triMask, primID, viewportIdx);
+
+        line[0] = tri[2];
+        line[1] = tri[0];
+        recipW[0] = vRecipW2;
+        recipW[1] = vRecipW0;
+        BinPostSetupLines_simd16(pDC, pa, workerId, line, recipW, triMask, primID, viewportIdx);
+
+        AR_END(FEBinTriangles, 1);
+        return;
+    }
+
+    /// Note: these variable initializations must stay above any 'goto endBenTriangles'
+    // compute per tri backface
+    uint32_t frontFaceMask = frontWindingTris;
+    uint32_t *pPrimID = (uint32_t *)&primID;
+    const uint32_t *pViewportIndex = (uint32_t *)&viewportIdx;
+    DWORD triIndex = 0;
+
+    uint32_t edgeEnable;
+    PFN_WORK_FUNC pfnWork;
+    if (CT::IsConservativeT::value)
+    {
+        // determine which edges of the degenerate tri, if any, are valid to rasterize.
+        // used to call the appropriate templated rasterizer function
+        if (cullZeroAreaMask > 0)
+        {
+            // e0 = v1-v0
+            const simd16scalari x0x1Mask = _simd16_cmpeq_epi32(vXi[0], vXi[1]);
+            const simd16scalari y0y1Mask = _simd16_cmpeq_epi32(vYi[0], vYi[1]);
+
+            uint32_t e0Mask = _simd16_movemask_ps(_simd16_castsi_ps(_simd16_and_si(x0x1Mask, y0y1Mask)));
+
+            // e1 = v2-v1
+            const simd16scalari x1x2Mask = _simd16_cmpeq_epi32(vXi[1], vXi[2]);
+            const simd16scalari y1y2Mask = _simd16_cmpeq_epi32(vYi[1], vYi[2]);
+
+            uint32_t e1Mask = _simd16_movemask_ps(_simd16_castsi_ps(_simd16_and_si(x1x2Mask, y1y2Mask)));
+
+            // e2 = v0-v2
+            // if v0 == v1 & v1 == v2, v0 == v2
+            uint32_t e2Mask = e0Mask & e1Mask;
+            SWR_ASSERT(KNOB_SIMD_WIDTH == 8, "Need to update degenerate mask code for avx512");
+
+            // edge order: e0 = v0v1, e1 = v1v2, e2 = v0v2
+            // 32 bit binary: 0000 0000 0010 0100 1001 0010 0100 1001
+            e0Mask = pdep_u32(e0Mask, 0x00249249);
+
+            // 32 bit binary: 0000 0000 0100 1001 0010 0100 1001 0010
+            e1Mask = pdep_u32(e1Mask, 0x00492492);
+
+            // 32 bit binary: 0000 0000 1001 0010 0100 1001 0010 0100
+            e2Mask = pdep_u32(e2Mask, 0x00924924);
+
+            edgeEnable = (0x00FFFFFF & (~(e0Mask | e1Mask | e2Mask)));
+        }
+        else
+        {
+            edgeEnable = 0x00FFFFFF;
+        }
+    }
+    else
+    {
+        // degenerate triangles won't be sent to rasterizer; just enable all edges
+        pfnWork = GetRasterizerFunc(rastState.sampleCount, rastState.bIsCenterPattern, (rastState.conservativeRast > 0),
+            (SWR_INPUT_COVERAGE)pDC->pState->state.psState.inputCoverage, ALL_EDGES_VALID, (state.scissorsTileAligned == false));
+    }
+
+    if (!triMask)
+    {
+        goto endBinTriangles;
+    }
+
+    // Calc bounding box of triangles
+    simd16BBox bbox;
+    calcBoundingBoxIntVertical<CT>(tri, vXi, vYi, bbox);
+
+    // determine if triangle falls between pixel centers and discard
+    // only discard for non-MSAA case and when conservative rast is disabled
+    // (xmin + 127) & ~255
+    // (xmax + 128) & ~255
+    if ((rastState.sampleCount == SWR_MULTISAMPLE_1X || rastState.bIsCenterPattern) &&
+        (!CT::IsConservativeT::value))
+    {
+        origTriMask = triMask;
+
+        int cullCenterMask;
+
+        {
+            simd16scalari xmin = _simd16_add_epi32(bbox.xmin, _simd16_set1_epi32(127));
+            xmin = _simd16_and_si(xmin, _simd16_set1_epi32(~255));
+            simd16scalari xmax = _simd16_add_epi32(bbox.xmax, _simd16_set1_epi32(128));
+            xmax = _simd16_and_si(xmax, _simd16_set1_epi32(~255));
+
+            simd16scalari vMaskH = _simd16_cmpeq_epi32(xmin, xmax);
+
+            simd16scalari ymin = _simd16_add_epi32(bbox.ymin, _simd16_set1_epi32(127));
+            ymin = _simd16_and_si(ymin, _simd16_set1_epi32(~255));
+            simd16scalari ymax = _simd16_add_epi32(bbox.ymax, _simd16_set1_epi32(128));
+            ymax = _simd16_and_si(ymax, _simd16_set1_epi32(~255));
+
+            simd16scalari vMaskV = _simd16_cmpeq_epi32(ymin, ymax);
+
+            vMaskV = _simd16_or_si(vMaskH, vMaskV);
+            cullCenterMask = _simd16_movemask_ps(_simd16_castsi_ps(vMaskV));
+        }
+
+        triMask &= ~cullCenterMask;
+
+        if (origTriMask ^ triMask)
+        {
+            RDTSC_EVENT(FECullBetweenCenters, _mm_popcnt_u32(origTriMask ^ triMask), 0);
+        }
+    }
+
+    // Intersect with scissor/viewport. Subtract 1 ULP in x.8 fixed point since xmax/ymax edge is exclusive.
+    // Gather the AOS effective scissor rects based on the per-prim VP index.
+    /// @todo:  Look at speeding this up -- weigh against corresponding costs in rasterizer.
+    simd16scalari scisXmin, scisYmin, scisXmax, scisYmax;
+
+    if (state.gsState.emitsViewportArrayIndex)
+    {
+        GatherScissors_simd16<KNOB_SIMD16_WIDTH>::Gather(&state.scissorsInFixedPoint[0], pViewportIndex,
+            scisXmin, scisYmin, scisXmax, scisYmax);
+    }
+    else // broadcast fast path for non-VPAI case.
+    {
+        scisXmin = _simd16_set1_epi32(state.scissorsInFixedPoint[0].xmin);
+        scisYmin = _simd16_set1_epi32(state.scissorsInFixedPoint[0].ymin);
+        scisXmax = _simd16_set1_epi32(state.scissorsInFixedPoint[0].xmax);
+        scisYmax = _simd16_set1_epi32(state.scissorsInFixedPoint[0].ymax);
+    }
+
+    bbox.xmin = _simd16_max_epi32(bbox.xmin, scisXmin);
+    bbox.ymin = _simd16_max_epi32(bbox.ymin, scisYmin);
+    bbox.xmax = _simd16_min_epi32(_simd16_sub_epi32(bbox.xmax, _simd16_set1_epi32(1)), scisXmax);
+    bbox.ymax = _simd16_min_epi32(_simd16_sub_epi32(bbox.ymax, _simd16_set1_epi32(1)), scisYmax);
+
+    if (CT::IsConservativeT::value)
+    {
+        // in the case where a degenerate triangle is on a scissor edge, we need to make sure the primitive bbox has
+        // some area. Bump the xmax/ymax edges out 
+        simd16scalari topEqualsBottom = _simd16_cmpeq_epi32(bbox.ymin, bbox.ymax);
+        bbox.ymax = _simd16_blendv_epi32(bbox.ymax, _simd16_add_epi32(bbox.ymax, _simd16_set1_epi32(1)), topEqualsBottom);
+        simd16scalari leftEqualsRight = _simd16_cmpeq_epi32(bbox.xmin, bbox.xmax);
+        bbox.xmax = _simd16_blendv_epi32(bbox.xmax, _simd16_add_epi32(bbox.xmax, _simd16_set1_epi32(1)), leftEqualsRight);
+    }
+
+    // Cull tris completely outside scissor
+    {
+        simd16scalari maskOutsideScissorX = _simd16_cmpgt_epi32(bbox.xmin, bbox.xmax);
+        simd16scalari maskOutsideScissorY = _simd16_cmpgt_epi32(bbox.ymin, bbox.ymax);
+        simd16scalari maskOutsideScissorXY = _simd16_or_si(maskOutsideScissorX, maskOutsideScissorY);
+        uint32_t maskOutsideScissor = _simd16_movemask_ps(_simd16_castsi_ps(maskOutsideScissorXY));
+        triMask = triMask & ~maskOutsideScissor;
+    }
+
+    if (!triMask)
+    {
+        goto endBinTriangles;
+    }
+
+    // Convert triangle bbox to macrotile units.
+    bbox.xmin = _simd16_srai_epi32(bbox.xmin, KNOB_MACROTILE_X_DIM_FIXED_SHIFT);
+    bbox.ymin = _simd16_srai_epi32(bbox.ymin, KNOB_MACROTILE_Y_DIM_FIXED_SHIFT);
+    bbox.xmax = _simd16_srai_epi32(bbox.xmax, KNOB_MACROTILE_X_DIM_FIXED_SHIFT);
+    bbox.ymax = _simd16_srai_epi32(bbox.ymax, KNOB_MACROTILE_Y_DIM_FIXED_SHIFT);
+
+    OSALIGNSIMD16(uint32_t) aMTLeft[KNOB_SIMD16_WIDTH], aMTRight[KNOB_SIMD16_WIDTH], aMTTop[KNOB_SIMD16_WIDTH], aMTBottom[KNOB_SIMD16_WIDTH];
+
+    _simd16_store_si(reinterpret_cast<simd16scalari *>(aMTLeft),    bbox.xmin);
+    _simd16_store_si(reinterpret_cast<simd16scalari *>(aMTRight),   bbox.xmax);
+    _simd16_store_si(reinterpret_cast<simd16scalari *>(aMTTop),     bbox.ymin);
+    _simd16_store_si(reinterpret_cast<simd16scalari *>(aMTBottom),  bbox.ymax);
+
+    // transpose verts needed for backend
+    /// @todo modify BE to take non-transformed verts
+    __m128 vHorizX[2][KNOB_SIMD_WIDTH]; // KNOB_SIMD16_WIDTH
+    __m128 vHorizY[2][KNOB_SIMD_WIDTH]; // KNOB_SIMD16_WIDTH
+    __m128 vHorizZ[2][KNOB_SIMD_WIDTH]; // KNOB_SIMD16_WIDTH
+    __m128 vHorizW[2][KNOB_SIMD_WIDTH]; // KNOB_SIMD16_WIDTH
+
+    vTranspose3x8(vHorizX[0], _simd16_extract_ps(tri[0].x, 0), _simd16_extract_ps(tri[1].x, 0), _simd16_extract_ps(tri[2].x, 0));
+    vTranspose3x8(vHorizY[0], _simd16_extract_ps(tri[0].y, 0), _simd16_extract_ps(tri[1].y, 0), _simd16_extract_ps(tri[2].y, 0));
+    vTranspose3x8(vHorizZ[0], _simd16_extract_ps(tri[0].z, 0), _simd16_extract_ps(tri[1].z, 0), _simd16_extract_ps(tri[2].z, 0));
+    vTranspose3x8(vHorizW[0], _simd16_extract_ps(vRecipW0, 0), _simd16_extract_ps(vRecipW1, 0), _simd16_extract_ps(vRecipW2, 0));
+
+    vTranspose3x8(vHorizX[1], _simd16_extract_ps(tri[0].x, 1), _simd16_extract_ps(tri[1].x, 1), _simd16_extract_ps(tri[2].x, 1));
+    vTranspose3x8(vHorizY[1], _simd16_extract_ps(tri[0].y, 1), _simd16_extract_ps(tri[1].y, 1), _simd16_extract_ps(tri[2].y, 1));
+    vTranspose3x8(vHorizZ[1], _simd16_extract_ps(tri[0].z, 1), _simd16_extract_ps(tri[1].z, 1), _simd16_extract_ps(tri[2].z, 1));
+    vTranspose3x8(vHorizW[1], _simd16_extract_ps(vRecipW0, 1), _simd16_extract_ps(vRecipW1, 1), _simd16_extract_ps(vRecipW2, 1));
+
+    // store render target array index
+    OSALIGNSIMD16(uint32_t) aRTAI[KNOB_SIMD16_WIDTH];
+    if (gsState.gsEnable && gsState.emitsRenderTargetArrayIndex)
+    {
+        simd16vector vRtai[3];
+        pa.Assemble_simd16(VERTEX_RTAI_SLOT, vRtai);
+        simd16scalari vRtaii;
+        vRtaii = _simd16_castps_si(vRtai[0].x);
+        _simd16_store_si(reinterpret_cast<simd16scalari *>(aRTAI), vRtaii);
+    }
+    else
+    {
+        _simd16_store_si(reinterpret_cast<simd16scalari *>(aRTAI), _simd16_setzero_si());
+    }
+
+endBinTriangles:
+
+
+    // scan remaining valid triangles and bin each separately
+    while (_BitScanForward(&triIndex, triMask))
+    {
+        uint32_t linkageCount = state.backendState.numAttributes;
+        uint32_t numScalarAttribs = linkageCount * 4;
+
+        BE_WORK work;
+        work.type = DRAW;
+
+        bool isDegenerate;
+        if (CT::IsConservativeT::value)
+        {
+            // only rasterize valid edges if we have a degenerate primitive
+            int32_t triEdgeEnable = (edgeEnable >> (triIndex * 3)) & ALL_EDGES_VALID;
+            work.pfnWork = GetRasterizerFunc(rastState.sampleCount, rastState.bIsCenterPattern, (rastState.conservativeRast > 0),
+                (SWR_INPUT_COVERAGE)pDC->pState->state.psState.inputCoverage, triEdgeEnable, (state.scissorsTileAligned == false));
+
+            // Degenerate triangles are required to be constant interpolated
+            isDegenerate = (triEdgeEnable != ALL_EDGES_VALID) ? true : false;
+        }
+        else
+        {
+            isDegenerate = false;
+            work.pfnWork = pfnWork;
+        }
+
+        // Select attribute processor
+        PFN_PROCESS_ATTRIBUTES pfnProcessAttribs = GetProcessAttributesFunc(3,
+            state.backendState.swizzleEnable, state.backendState.constantInterpolationMask, isDegenerate);
+
+        TRIANGLE_WORK_DESC &desc = work.desc.tri;
+
+        desc.triFlags.frontFacing = state.forceFront ? 1 : ((frontFaceMask >> triIndex) & 1);
+        desc.triFlags.primID = pPrimID[triIndex];
+        desc.triFlags.renderTargetArrayIndex = aRTAI[triIndex];
+        desc.triFlags.viewportIndex = pViewportIndex[triIndex];
+
+        auto pArena = pDC->pArena;
+        SWR_ASSERT(pArena != nullptr);
+
+        // store active attribs
+        float *pAttribs = (float*)pArena->AllocAligned(numScalarAttribs * 3 * sizeof(float), 16);
+        desc.pAttribs = pAttribs;
+        desc.numAttribs = linkageCount;
+        pfnProcessAttribs(pDC, pa, triIndex, pPrimID[triIndex], desc.pAttribs);
+
+        // store triangle vertex data
+        desc.pTriBuffer = (float*)pArena->AllocAligned(4 * 4 * sizeof(float), 16);
+
+        {
+            const uint32_t i = triIndex >> 3;   // triIndex / KNOB_SIMD_WIDTH
+            const uint32_t j = triIndex & 7;    // triIndex % KNOB_SIMD_WIDTH
+
+            _mm_store_ps(&desc.pTriBuffer[ 0], vHorizX[i][j]);
+            _mm_store_ps(&desc.pTriBuffer[ 4], vHorizY[i][j]);
+            _mm_store_ps(&desc.pTriBuffer[ 8], vHorizZ[i][j]);
+            _mm_store_ps(&desc.pTriBuffer[12], vHorizW[i][j]);
+        }
+
+        // store user clip distances
+        if (rastState.clipDistanceMask)
+        {
+            uint32_t numClipDist = _mm_popcnt_u32(rastState.clipDistanceMask);
+            desc.pUserClipBuffer = (float*)pArena->Alloc(numClipDist * 3 * sizeof(float));
+            ProcessUserClipDist<3>(pa, triIndex, rastState.clipDistanceMask, &desc.pTriBuffer[12], desc.pUserClipBuffer);
+        }
+
+        for (uint32_t y = aMTTop[triIndex]; y <= aMTBottom[triIndex]; ++y)
+        {
+            for (uint32_t x = aMTLeft[triIndex]; x <= aMTRight[triIndex]; ++x)
+            {
+#if KNOB_ENABLE_TOSS_POINTS
+                if (!KNOB_TOSS_SETUP_TRIS)
+#endif
+                {
+                    pTileMgr->enqueue(x, y, &work);
+                }
+            }
+        }
+
+        triMask &= ~(1 << triIndex);
+    }
+
+    AR_END(FEBinTriangles, 1);
+}
+
+#endif
+struct FEBinTrianglesChooser
+{
+    typedef PFN_PROCESS_PRIMS FuncType;
+
+    template <typename... ArgsB>
+    static FuncType GetFunc()
+    {
+        return BinTriangles<ConservativeRastFETraits<ArgsB...>>;
+    }
+};
+
+// Selector for correct templated BinTrinagles function
+PFN_PROCESS_PRIMS GetBinTrianglesFunc(bool IsConservative)
+{
+    return TemplateArgUnroller<FEBinTrianglesChooser>::GetFunc(IsConservative);
+}
+
+#if USE_SIMD16_FRONTEND
+struct FEBinTrianglesChooser_simd16
+{
+    typedef PFN_PROCESS_PRIMS_SIMD16 FuncType;
+
+    template <typename... ArgsB>
+    static FuncType GetFunc()
+    {
+        return BinTriangles_simd16<ConservativeRastFETraits<ArgsB...>>;
+    }
+};
+
+// Selector for correct templated BinTrinagles function
+PFN_PROCESS_PRIMS_SIMD16 GetBinTrianglesFunc_simd16(bool IsConservative)
+{
+    return TemplateArgUnroller<FEBinTrianglesChooser_simd16>::GetFunc(IsConservative);
+}
+
+#endif
+
+//////////////////////////////////////////////////////////////////////////
+/// @brief Bin SIMD points to the backend.  Only supports point size of 1
+/// @param pDC - pointer to draw context.
+/// @param pa - The primitive assembly object.
+/// @param workerId - thread's worker id. Even thread has a unique id.
+/// @param tri - Contains point position data for SIMDs worth of points.
+/// @param primID - Primitive ID for each point.
+void BinPoints(
+    DRAW_CONTEXT *pDC,
+    PA_STATE& pa,
+    uint32_t workerId,
+    simdvector prim[3],
+    uint32_t primMask,
+    simdscalari primID,
+    simdscalari viewportIdx)
+{
+    SWR_CONTEXT *pContext = pDC->pContext;
+
+    AR_BEGIN(FEBinPoints, pDC->drawId);
+
+    simdvector& primVerts = prim[0];
+
+    const API_STATE& state = GetApiState(pDC);
+    const SWR_FRONTEND_STATE& feState = state.frontendState;
+    const SWR_GS_STATE& gsState = state.gsState;
+    const SWR_RASTSTATE& rastState = state.rastState;
+    const uint32_t *pViewportIndex = (uint32_t *)&viewportIdx;
+
+    // Select attribute processor
+    PFN_PROCESS_ATTRIBUTES pfnProcessAttribs = GetProcessAttributesFunc(1,
+        state.backendState.swizzleEnable, state.backendState.constantInterpolationMask);
+
+    if (!feState.vpTransformDisable)
+    {
+        // perspective divide
+        simdscalar vRecipW0 = _simd_div_ps(_simd_set1_ps(1.0f), primVerts.w);
+        primVerts.x = _simd_mul_ps(primVerts.x, vRecipW0);
+        primVerts.y = _simd_mul_ps(primVerts.y, vRecipW0);
+        primVerts.z = _simd_mul_ps(primVerts.z, vRecipW0);
+
+        // viewport transform to screen coords
+        if (state.gsState.emitsViewportArrayIndex)
+        {
+            viewportTransform<1>(&primVerts, state.vpMatrices, viewportIdx);
+        }
+        else
+        {
+            viewportTransform<1>(&primVerts, state.vpMatrices);
+        }
+    }
+
+    // adjust for pixel center location
+    simdscalar offset = g_pixelOffsets[rastState.pixelLocation];
+    primVerts.x = _simd_add_ps(primVerts.x, offset);
+    primVerts.y = _simd_add_ps(primVerts.y, offset);
+
+    // convert to fixed point
+    simdscalari vXi, vYi;
+    vXi = fpToFixedPointVertical(primVerts.x);
+    vYi = fpToFixedPointVertical(primVerts.y);
+
+    if (CanUseSimplePoints(pDC))
+    {
+        // adjust for ymin-xmin rule
+        vXi = _simd_sub_epi32(vXi, _simd_set1_epi32(1));
+        vYi = _simd_sub_epi32(vYi, _simd_set1_epi32(1));
+
+        // cull points off the ymin-xmin edge of the viewport
+        primMask &= ~_simd_movemask_ps(_simd_castsi_ps(vXi));
+        primMask &= ~_simd_movemask_ps(_simd_castsi_ps(vYi));
+
+        // compute macro tile coordinates 
+        simdscalari macroX = _simd_srai_epi32(vXi, KNOB_MACROTILE_X_DIM_FIXED_SHIFT);
+        simdscalari macroY = _simd_srai_epi32(vYi, KNOB_MACROTILE_Y_DIM_FIXED_SHIFT);
+
+        OSALIGNSIMD(uint32_t) aMacroX[KNOB_SIMD_WIDTH], aMacroY[KNOB_SIMD_WIDTH];
+        _simd_store_si((simdscalari*)aMacroX, macroX);
+        _simd_store_si((simdscalari*)aMacroY, macroY);
+
+        // compute raster tile coordinates
+        simdscalari rasterX = _simd_srai_epi32(vXi, KNOB_TILE_X_DIM_SHIFT + FIXED_POINT_SHIFT);
+        simdscalari rasterY = _simd_srai_epi32(vYi, KNOB_TILE_Y_DIM_SHIFT + FIXED_POINT_SHIFT);
+
+        // compute raster tile relative x,y for coverage mask
+        simdscalari tileAlignedX = _simd_slli_epi32(rasterX, KNOB_TILE_X_DIM_SHIFT);
+        simdscalari tileAlignedY = _simd_slli_epi32(rasterY, KNOB_TILE_Y_DIM_SHIFT);
+
+        simdscalari tileRelativeX = _simd_sub_epi32(_simd_srai_epi32(vXi, FIXED_POINT_SHIFT), tileAlignedX);
+        simdscalari tileRelativeY = _simd_sub_epi32(_simd_srai_epi32(vYi, FIXED_POINT_SHIFT), tileAlignedY);
+
+        OSALIGNSIMD(uint32_t) aTileRelativeX[KNOB_SIMD_WIDTH];
+        OSALIGNSIMD(uint32_t) aTileRelativeY[KNOB_SIMD_WIDTH];
+        _simd_store_si((simdscalari*)aTileRelativeX, tileRelativeX);
+        _simd_store_si((simdscalari*)aTileRelativeY, tileRelativeY);
+
+        OSALIGNSIMD(uint32_t) aTileAlignedX[KNOB_SIMD_WIDTH];
+        OSALIGNSIMD(uint32_t) aTileAlignedY[KNOB_SIMD_WIDTH];
+        _simd_store_si((simdscalari*)aTileAlignedX, tileAlignedX);
+        _simd_store_si((simdscalari*)aTileAlignedY, tileAlignedY);
+
+        OSALIGNSIMD(float) aZ[KNOB_SIMD_WIDTH];
+        _simd_store_ps((float*)aZ, primVerts.z);
+
+        // store render target array index
+        OSALIGNSIMD(uint32_t) aRTAI[KNOB_SIMD_WIDTH];
+        if (gsState.gsEnable && gsState.emitsRenderTargetArrayIndex)
+        {
+            simdvector vRtai;
+            pa.Assemble(VERTEX_RTAI_SLOT, &vRtai);
+            simdscalari vRtaii = _simd_castps_si(vRtai.x);
+            _simd_store_si((simdscalari*)aRTAI, vRtaii);
+        }
+        else
+        {
+            _simd_store_si((simdscalari*)aRTAI, _simd_setzero_si());
+        }
+
+        uint32_t *pPrimID = (uint32_t *)&primID;
+        DWORD primIndex = 0;
+
+        const SWR_BACKEND_STATE& backendState = pDC->pState->state.backendState;
+
+        // scan remaining valid triangles and bin each separately
+        while (_BitScanForward(&primIndex, primMask))
+        {
+            uint32_t linkageCount = backendState.numAttributes;
+            uint32_t numScalarAttribs = linkageCount * 4;
+
+            BE_WORK work;
+            work.type = DRAW;
+
+            TRIANGLE_WORK_DESC &desc = work.desc.tri;
+
+            // points are always front facing
+            desc.triFlags.frontFacing = 1;
+            desc.triFlags.primID = pPrimID[primIndex];
+            desc.triFlags.renderTargetArrayIndex = aRTAI[primIndex];
+            desc.triFlags.viewportIndex = pViewportIndex[primIndex];
+
+            work.pfnWork = RasterizeSimplePoint;
+
+            auto pArena = pDC->pArena;
+            SWR_ASSERT(pArena != nullptr);
+
+            // store attributes
+            float *pAttribs = (float*)pArena->AllocAligned(3 * numScalarAttribs * sizeof(float), 16);
+            desc.pAttribs = pAttribs;
+            desc.numAttribs = linkageCount;
+
+            pfnProcessAttribs(pDC, pa, primIndex, pPrimID[primIndex], pAttribs);
+
+            // store raster tile aligned x, y, perspective correct z
+            float *pTriBuffer = (float*)pArena->AllocAligned(4 * sizeof(float), 16);
+            desc.pTriBuffer = pTriBuffer;
+            *(uint32_t*)pTriBuffer++ = aTileAlignedX[primIndex];
+            *(uint32_t*)pTriBuffer++ = aTileAlignedY[primIndex];
+            *pTriBuffer = aZ[primIndex];
+
+            uint32_t tX = aTileRelativeX[primIndex];
+            uint32_t tY = aTileRelativeY[primIndex];
+
+            // pack the relative x,y into the coverageMask, the rasterizer will
+            // generate the true coverage mask from it
+            work.desc.tri.triFlags.coverageMask = tX | (tY << 4);
+
+            // bin it
+            MacroTileMgr *pTileMgr = pDC->pTileMgr;
+#if KNOB_ENABLE_TOSS_POINTS
+            if (!KNOB_TOSS_SETUP_TRIS)
+#endif
+            {
+                pTileMgr->enqueue(aMacroX[primIndex], aMacroY[primIndex], &work);
+            }
+            primMask &= ~(1 << primIndex);
+        }
+    }
+    else
+    {
+        // non simple points need to be potentially binned to multiple macro tiles
+        simdscalar vPointSize;
+        if (rastState.pointParam)
         {
-            simdvector vRtai;
-            pa.Assemble(VERTEX_RTAI_SLOT, &vRtai);
-            simdscalari vRtaii = _simd_castps_si(vRtai.x);
+            simdvector size[3];
+            pa.Assemble(VERTEX_POINT_SIZE_SLOT, size);
+            vPointSize = size[0].x;
+        }
+        else
+        {
+            vPointSize = _simd_set1_ps(rastState.pointSize);
+        }
+
+        // bloat point to bbox
+        simdBBox bbox;
+        bbox.xmin = bbox.xmax = vXi;
+        bbox.ymin = bbox.ymax = vYi;
+
+        simdscalar vHalfWidth = _simd_mul_ps(vPointSize, _simd_set1_ps(0.5f));
+        simdscalari vHalfWidthi = fpToFixedPointVertical(vHalfWidth);
+        bbox.xmin = _simd_sub_epi32(bbox.xmin, vHalfWidthi);
+        bbox.xmax = _simd_add_epi32(bbox.xmax, vHalfWidthi);
+        bbox.ymin = _simd_sub_epi32(bbox.ymin, vHalfWidthi);
+        bbox.ymax = _simd_add_epi32(bbox.ymax, vHalfWidthi);
+
+        // Intersect with scissor/viewport. Subtract 1 ULP in x.8 fixed point since xmax/ymax edge is exclusive.
+        // Gather the AOS effective scissor rects based on the per-prim VP index.
+        /// @todo:  Look at speeding this up -- weigh against corresponding costs in rasterizer.
+        simdscalari scisXmin, scisYmin, scisXmax, scisYmax;
+        if (state.gsState.emitsViewportArrayIndex)
+        {
+            GatherScissors<KNOB_SIMD_WIDTH>::Gather(&state.scissorsInFixedPoint[0], pViewportIndex,
+                scisXmin, scisYmin, scisXmax, scisYmax);
+        }
+        else // broadcast fast path for non-VPAI case.
+        {
+            scisXmin = _simd_set1_epi32(state.scissorsInFixedPoint[0].xmin);
+            scisYmin = _simd_set1_epi32(state.scissorsInFixedPoint[0].ymin);
+            scisXmax = _simd_set1_epi32(state.scissorsInFixedPoint[0].xmax);
+            scisYmax = _simd_set1_epi32(state.scissorsInFixedPoint[0].ymax);
+        }
+
+        bbox.xmin = _simd_max_epi32(bbox.xmin, scisXmin);
+        bbox.ymin = _simd_max_epi32(bbox.ymin, scisYmin);
+        bbox.xmax = _simd_min_epi32(_simd_sub_epi32(bbox.xmax, _simd_set1_epi32(1)), scisXmax);
+        bbox.ymax = _simd_min_epi32(_simd_sub_epi32(bbox.ymax, _simd_set1_epi32(1)), scisYmax);
+
+        // Cull bloated points completely outside scissor
+        simdscalari maskOutsideScissorX = _simd_cmpgt_epi32(bbox.xmin, bbox.xmax);
+        simdscalari maskOutsideScissorY = _simd_cmpgt_epi32(bbox.ymin, bbox.ymax);
+        simdscalari maskOutsideScissorXY = _simd_or_si(maskOutsideScissorX, maskOutsideScissorY);
+        uint32_t maskOutsideScissor = _simd_movemask_ps(_simd_castsi_ps(maskOutsideScissorXY));
+        primMask = primMask & ~maskOutsideScissor;
+
+        // Convert bbox to macrotile units.
+        bbox.xmin = _simd_srai_epi32(bbox.xmin, KNOB_MACROTILE_X_DIM_FIXED_SHIFT);
+        bbox.ymin = _simd_srai_epi32(bbox.ymin, KNOB_MACROTILE_Y_DIM_FIXED_SHIFT);
+        bbox.xmax = _simd_srai_epi32(bbox.xmax, KNOB_MACROTILE_X_DIM_FIXED_SHIFT);
+        bbox.ymax = _simd_srai_epi32(bbox.ymax, KNOB_MACROTILE_Y_DIM_FIXED_SHIFT);
+
+        OSALIGNSIMD(uint32_t) aMTLeft[KNOB_SIMD_WIDTH], aMTRight[KNOB_SIMD_WIDTH], aMTTop[KNOB_SIMD_WIDTH], aMTBottom[KNOB_SIMD_WIDTH];
+        _simd_store_si((simdscalari*)aMTLeft, bbox.xmin);
+        _simd_store_si((simdscalari*)aMTRight, bbox.xmax);
+        _simd_store_si((simdscalari*)aMTTop, bbox.ymin);
+        _simd_store_si((simdscalari*)aMTBottom, bbox.ymax);
+
+        // store render target array index
+        OSALIGNSIMD(uint32_t) aRTAI[KNOB_SIMD_WIDTH];
+        if (gsState.gsEnable && gsState.emitsRenderTargetArrayIndex)
+        {
+            simdvector vRtai[2];
+            pa.Assemble(VERTEX_RTAI_SLOT, vRtai);
+            simdscalari vRtaii = _simd_castps_si(vRtai[0].x);
             _simd_store_si((simdscalari*)aRTAI, vRtaii);
         }
         else
@@ -1058,6 +1763,207 @@ void BinPoints(
             _simd_store_si((simdscalari*)aRTAI, _simd_setzero_si());
         }
 
+        OSALIGNSIMD(float) aPointSize[KNOB_SIMD_WIDTH];
+        _simd_store_ps((float*)aPointSize, vPointSize);
+
+        uint32_t *pPrimID = (uint32_t *)&primID;
+
+        OSALIGNSIMD(float) aPrimVertsX[KNOB_SIMD_WIDTH];
+        OSALIGNSIMD(float) aPrimVertsY[KNOB_SIMD_WIDTH];
+        OSALIGNSIMD(float) aPrimVertsZ[KNOB_SIMD_WIDTH];
+
+        _simd_store_ps((float*)aPrimVertsX, primVerts.x);
+        _simd_store_ps((float*)aPrimVertsY, primVerts.y);
+        _simd_store_ps((float*)aPrimVertsZ, primVerts.z);
+
+        // scan remaining valid prims and bin each separately
+        const SWR_BACKEND_STATE& backendState = state.backendState;
+        DWORD primIndex;
+        while (_BitScanForward(&primIndex, primMask))
+        {
+            uint32_t linkageCount = backendState.numAttributes;
+            uint32_t numScalarAttribs = linkageCount * 4;
+
+            BE_WORK work;
+            work.type = DRAW;
+
+            TRIANGLE_WORK_DESC &desc = work.desc.tri;
+
+            desc.triFlags.frontFacing = 1;
+            desc.triFlags.primID = pPrimID[primIndex];
+            desc.triFlags.pointSize = aPointSize[primIndex];
+            desc.triFlags.renderTargetArrayIndex = aRTAI[primIndex];
+            desc.triFlags.viewportIndex = pViewportIndex[primIndex];
+
+            work.pfnWork = RasterizeTriPoint;
+
+            auto pArena = pDC->pArena;
+            SWR_ASSERT(pArena != nullptr);
+
+            // store active attribs
+            desc.pAttribs = (float*)pArena->AllocAligned(numScalarAttribs * 3 * sizeof(float), 16);
+            desc.numAttribs = linkageCount;
+            pfnProcessAttribs(pDC, pa, primIndex, pPrimID[primIndex], desc.pAttribs);
+
+            // store point vertex data
+            float *pTriBuffer = (float*)pArena->AllocAligned(4 * sizeof(float), 16);
+            desc.pTriBuffer = pTriBuffer;
+            *pTriBuffer++ = aPrimVertsX[primIndex];
+            *pTriBuffer++ = aPrimVertsY[primIndex];
+            *pTriBuffer = aPrimVertsZ[primIndex];
+
+            // store user clip distances
+            if (rastState.clipDistanceMask)
+            {
+                uint32_t numClipDist = _mm_popcnt_u32(rastState.clipDistanceMask);
+                desc.pUserClipBuffer = (float*)pArena->Alloc(numClipDist * 3 * sizeof(float));
+                float dists[8];
+                float one = 1.0f;
+                ProcessUserClipDist<1>(pa, primIndex, rastState.clipDistanceMask, &one, dists);
+                for (uint32_t i = 0; i < numClipDist; i++) {
+                    desc.pUserClipBuffer[3*i + 0] = 0.0f;
+                    desc.pUserClipBuffer[3*i + 1] = 0.0f;
+                    desc.pUserClipBuffer[3*i + 2] = dists[i];
+                }
+            }
+
+            MacroTileMgr *pTileMgr = pDC->pTileMgr;
+            for (uint32_t y = aMTTop[primIndex]; y <= aMTBottom[primIndex]; ++y)
+            {
+                for (uint32_t x = aMTLeft[primIndex]; x <= aMTRight[primIndex]; ++x)
+                {
+#if KNOB_ENABLE_TOSS_POINTS
+                    if (!KNOB_TOSS_SETUP_TRIS)
+#endif
+                    {
+                        pTileMgr->enqueue(x, y, &work);
+                    }
+                }
+            }
+
+            primMask &= ~(1 << primIndex);
+        }
+    }
+
+    AR_END(FEBinPoints, 1);
+}
+
+#if USE_SIMD16_FRONTEND
+void BinPoints_simd16(
+    DRAW_CONTEXT *pDC,
+    PA_STATE& pa,
+    uint32_t workerId,
+    simd16vector prim[3],
+    uint32_t primMask,
+    simd16scalari primID,
+    simd16scalari viewportIdx)
+{
+    SWR_CONTEXT *pContext = pDC->pContext;
+
+    AR_BEGIN(FEBinPoints, pDC->drawId);
+
+    simd16vector& primVerts = prim[0];
+
+    const API_STATE& state = GetApiState(pDC);
+    const SWR_FRONTEND_STATE& feState = state.frontendState;
+    const SWR_GS_STATE& gsState = state.gsState;
+    const SWR_RASTSTATE& rastState = state.rastState;
+    const uint32_t *pViewportIndex = (uint32_t *)&viewportIdx;
+
+    // Select attribute processor
+    PFN_PROCESS_ATTRIBUTES pfnProcessAttribs = GetProcessAttributesFunc(1,
+        state.backendState.swizzleEnable, state.backendState.constantInterpolationMask);
+
+    if (!feState.vpTransformDisable)
+    {
+        // perspective divide
+        simd16scalar vRecipW0 = _simd16_div_ps(_simd16_set1_ps(1.0f), primVerts.w);
+
+        primVerts.x = _simd16_mul_ps(primVerts.x, vRecipW0);
+        primVerts.y = _simd16_mul_ps(primVerts.y, vRecipW0);
+        primVerts.z = _simd16_mul_ps(primVerts.z, vRecipW0);
+
+        // viewport transform to screen coords
+        if (state.gsState.emitsViewportArrayIndex)
+        {
+            viewportTransform<1>(&primVerts, state.vpMatrices, viewportIdx);
+        }
+        else
+        {
+            viewportTransform<1>(&primVerts, state.vpMatrices);
+        }
+    }
+
+    const simd16scalar offset = g_pixelOffsets_simd16[rastState.pixelLocation];
+
+    primVerts.x = _simd16_add_ps(primVerts.x, offset);
+    primVerts.y = _simd16_add_ps(primVerts.y, offset);
+
+    // convert to fixed point
+    simd16scalari vXi, vYi;
+
+    vXi = fpToFixedPointVertical(primVerts.x);
+    vYi = fpToFixedPointVertical(primVerts.y);
+
+    if (CanUseSimplePoints(pDC))
+    {
+        // adjust for ymin-xmin rule
+        vXi = _simd16_sub_epi32(vXi, _simd16_set1_epi32(1));
+        vYi = _simd16_sub_epi32(vYi, _simd16_set1_epi32(1));
+
+        // cull points off the ymin-xmin edge of the viewport
+        primMask &= ~_simd16_movemask_ps(_simd16_castsi_ps(vXi));
+        primMask &= ~_simd16_movemask_ps(_simd16_castsi_ps(vYi));
+
+        // compute macro tile coordinates 
+        simd16scalari macroX = _simd16_srai_epi32(vXi, KNOB_MACROTILE_X_DIM_FIXED_SHIFT);
+        simd16scalari macroY = _simd16_srai_epi32(vYi, KNOB_MACROTILE_Y_DIM_FIXED_SHIFT);
+
+        OSALIGNSIMD16(uint32_t) aMacroX[KNOB_SIMD16_WIDTH], aMacroY[KNOB_SIMD16_WIDTH];
+
+        _simd16_store_si(reinterpret_cast<simd16scalari *>(aMacroX), macroX);
+        _simd16_store_si(reinterpret_cast<simd16scalari *>(aMacroY), macroY);
+
+        // compute raster tile coordinates
+        simd16scalari rasterX = _simd16_srai_epi32(vXi, KNOB_TILE_X_DIM_SHIFT + FIXED_POINT_SHIFT);
+        simd16scalari rasterY = _simd16_srai_epi32(vYi, KNOB_TILE_Y_DIM_SHIFT + FIXED_POINT_SHIFT);
+
+        // compute raster tile relative x,y for coverage mask
+        simd16scalari tileAlignedX = _simd16_slli_epi32(rasterX, KNOB_TILE_X_DIM_SHIFT);
+        simd16scalari tileAlignedY = _simd16_slli_epi32(rasterY, KNOB_TILE_Y_DIM_SHIFT);
+
+        simd16scalari tileRelativeX = _simd16_sub_epi32(_simd16_srai_epi32(vXi, FIXED_POINT_SHIFT), tileAlignedX);
+        simd16scalari tileRelativeY = _simd16_sub_epi32(_simd16_srai_epi32(vYi, FIXED_POINT_SHIFT), tileAlignedY);
+
+        OSALIGNSIMD16(uint32_t) aTileRelativeX[KNOB_SIMD16_WIDTH];
+        OSALIGNSIMD16(uint32_t) aTileRelativeY[KNOB_SIMD16_WIDTH];
+
+        _simd16_store_si(reinterpret_cast<simd16scalari *>(aTileRelativeX), tileRelativeX);
+        _simd16_store_si(reinterpret_cast<simd16scalari *>(aTileRelativeY), tileRelativeY);
+
+        OSALIGNSIMD16(uint32_t) aTileAlignedX[KNOB_SIMD16_WIDTH];
+        OSALIGNSIMD16(uint32_t) aTileAlignedY[KNOB_SIMD16_WIDTH];
+
+        _simd16_store_si(reinterpret_cast<simd16scalari *>(aTileAlignedX), tileAlignedX);
+        _simd16_store_si(reinterpret_cast<simd16scalari *>(aTileAlignedY), tileAlignedY);
+
+        OSALIGNSIMD16(float) aZ[KNOB_SIMD16_WIDTH];
+        _simd16_store_ps(reinterpret_cast<float *>(aZ), primVerts.z);
+
+        // store render target array index
+        OSALIGNSIMD16(uint32_t) aRTAI[KNOB_SIMD16_WIDTH];
+        if (gsState.gsEnable && gsState.emitsRenderTargetArrayIndex)
+        {
+            simd16vector vRtai;
+            pa.Assemble_simd16(VERTEX_RTAI_SLOT, &vRtai);
+            simd16scalari vRtaii = _simd16_castps_si(vRtai.x);
+            _simd16_store_si(reinterpret_cast<simd16scalari *>(aRTAI), vRtaii);
+        }
+        else
+        {
+            _simd16_store_si(reinterpret_cast<simd16scalari *>(aRTAI), _simd16_setzero_si());
+        }
+
         uint32_t *pPrimID = (uint32_t *)&primID;
         DWORD primIndex = 0;
 
@@ -1114,103 +2020,108 @@ void BinPoints(
             {
                 pTileMgr->enqueue(aMacroX[primIndex], aMacroY[primIndex], &work);
             }
+
             primMask &= ~(1 << primIndex);
         }
     }
     else
     {
         // non simple points need to be potentially binned to multiple macro tiles
-        simdscalar vPointSize;
+        simd16scalar vPointSize;
+
         if (rastState.pointParam)
         {
-            simdvector size[3];
-            pa.Assemble(VERTEX_POINT_SIZE_SLOT, size);
+            simd16vector size[3];
+            pa.Assemble_simd16(VERTEX_POINT_SIZE_SLOT, size);
             vPointSize = size[0].x;
         }
         else
         {
-            vPointSize = _simd_set1_ps(rastState.pointSize);
+            vPointSize = _simd16_set1_ps(rastState.pointSize);
         }
 
         // bloat point to bbox
-        simdBBox bbox;
+        simd16BBox bbox;
+
         bbox.xmin = bbox.xmax = vXi;
         bbox.ymin = bbox.ymax = vYi;
 
-        simdscalar vHalfWidth = _simd_mul_ps(vPointSize, _simd_set1_ps(0.5f));
-        simdscalari vHalfWidthi = fpToFixedPointVertical(vHalfWidth);
-        bbox.xmin = _simd_sub_epi32(bbox.xmin, vHalfWidthi);
-        bbox.xmax = _simd_add_epi32(bbox.xmax, vHalfWidthi);
-        bbox.ymin = _simd_sub_epi32(bbox.ymin, vHalfWidthi);
-        bbox.ymax = _simd_add_epi32(bbox.ymax, vHalfWidthi);
+        simd16scalar vHalfWidth = _simd16_mul_ps(vPointSize, _simd16_set1_ps(0.5f));
+        simd16scalari vHalfWidthi = fpToFixedPointVertical(vHalfWidth);
+
+        bbox.xmin = _simd16_sub_epi32(bbox.xmin, vHalfWidthi);
+        bbox.xmax = _simd16_add_epi32(bbox.xmax, vHalfWidthi);
+        bbox.ymin = _simd16_sub_epi32(bbox.ymin, vHalfWidthi);
+        bbox.ymax = _simd16_add_epi32(bbox.ymax, vHalfWidthi);
 
         // Intersect with scissor/viewport. Subtract 1 ULP in x.8 fixed point since xmax/ymax edge is exclusive.
         // Gather the AOS effective scissor rects based on the per-prim VP index.
         /// @todo:  Look at speeding this up -- weigh against corresponding costs in rasterizer.
-        simdscalari scisXmin, scisYmin, scisXmax, scisYmax;
+        simd16scalari scisXmin, scisYmin, scisXmax, scisYmax;
         if (state.gsState.emitsViewportArrayIndex)
         {
-            GatherScissors<KNOB_SIMD_WIDTH>::Gather(&state.scissorsInFixedPoint[0], pViewportIndex,
+            GatherScissors_simd16<KNOB_SIMD16_WIDTH>::Gather(&state.scissorsInFixedPoint[0], pViewportIndex,
                 scisXmin, scisYmin, scisXmax, scisYmax);
         }
         else // broadcast fast path for non-VPAI case.
         {
-            scisXmin = _simd_set1_epi32(state.scissorsInFixedPoint[0].xmin);
-            scisYmin = _simd_set1_epi32(state.scissorsInFixedPoint[0].ymin);
-            scisXmax = _simd_set1_epi32(state.scissorsInFixedPoint[0].xmax);
-            scisYmax = _simd_set1_epi32(state.scissorsInFixedPoint[0].ymax);
+            scisXmin = _simd16_set1_epi32(state.scissorsInFixedPoint[0].xmin);
+            scisYmin = _simd16_set1_epi32(state.scissorsInFixedPoint[0].ymin);
+            scisXmax = _simd16_set1_epi32(state.scissorsInFixedPoint[0].xmax);
+            scisYmax = _simd16_set1_epi32(state.scissorsInFixedPoint[0].ymax);
         }
 
-        bbox.xmin = _simd_max_epi32(bbox.xmin, scisXmin);
-        bbox.ymin = _simd_max_epi32(bbox.ymin, scisYmin);
-        bbox.xmax = _simd_min_epi32(_simd_sub_epi32(bbox.xmax, _simd_set1_epi32(1)), scisXmax);
-        bbox.ymax = _simd_min_epi32(_simd_sub_epi32(bbox.ymax, _simd_set1_epi32(1)), scisYmax);
+        bbox.xmin = _simd16_max_epi32(bbox.xmin, scisXmin);
+        bbox.ymin = _simd16_max_epi32(bbox.ymin, scisYmin);
+        bbox.xmax = _simd16_min_epi32(_simd16_sub_epi32(bbox.xmax, _simd16_set1_epi32(1)), scisXmax);
+        bbox.ymax = _simd16_min_epi32(_simd16_sub_epi32(bbox.ymax, _simd16_set1_epi32(1)), scisYmax);
 
         // Cull bloated points completely outside scissor
-        simdscalari maskOutsideScissorX = _simd_cmpgt_epi32(bbox.xmin, bbox.xmax);
-        simdscalari maskOutsideScissorY = _simd_cmpgt_epi32(bbox.ymin, bbox.ymax);
-        simdscalari maskOutsideScissorXY = _simd_or_si(maskOutsideScissorX, maskOutsideScissorY);
-        uint32_t maskOutsideScissor = _simd_movemask_ps(_simd_castsi_ps(maskOutsideScissorXY));
+        simd16scalari maskOutsideScissorX = _simd16_cmpgt_epi32(bbox.xmin, bbox.xmax);
+        simd16scalari maskOutsideScissorY = _simd16_cmpgt_epi32(bbox.ymin, bbox.ymax);
+        simd16scalari maskOutsideScissorXY = _simd16_or_si(maskOutsideScissorX, maskOutsideScissorY);
+        uint32_t maskOutsideScissor = _simd16_movemask_ps(_simd16_castsi_ps(maskOutsideScissorXY));
         primMask = primMask & ~maskOutsideScissor;
 
         // Convert bbox to macrotile units.
-        bbox.xmin = _simd_srai_epi32(bbox.xmin, KNOB_MACROTILE_X_DIM_FIXED_SHIFT);
-        bbox.ymin = _simd_srai_epi32(bbox.ymin, KNOB_MACROTILE_Y_DIM_FIXED_SHIFT);
-        bbox.xmax = _simd_srai_epi32(bbox.xmax, KNOB_MACROTILE_X_DIM_FIXED_SHIFT);
-        bbox.ymax = _simd_srai_epi32(bbox.ymax, KNOB_MACROTILE_Y_DIM_FIXED_SHIFT);
+        bbox.xmin = _simd16_srai_epi32(bbox.xmin, KNOB_MACROTILE_X_DIM_FIXED_SHIFT);
+        bbox.ymin = _simd16_srai_epi32(bbox.ymin, KNOB_MACROTILE_Y_DIM_FIXED_SHIFT);
+        bbox.xmax = _simd16_srai_epi32(bbox.xmax, KNOB_MACROTILE_X_DIM_FIXED_SHIFT);
+        bbox.ymax = _simd16_srai_epi32(bbox.ymax, KNOB_MACROTILE_Y_DIM_FIXED_SHIFT);
 
-        OSALIGNSIMD(uint32_t) aMTLeft[KNOB_SIMD_WIDTH], aMTRight[KNOB_SIMD_WIDTH], aMTTop[KNOB_SIMD_WIDTH], aMTBottom[KNOB_SIMD_WIDTH];
-        _simd_store_si((simdscalari*)aMTLeft, bbox.xmin);
-        _simd_store_si((simdscalari*)aMTRight, bbox.xmax);
-        _simd_store_si((simdscalari*)aMTTop, bbox.ymin);
-        _simd_store_si((simdscalari*)aMTBottom, bbox.ymax);
+        OSALIGNSIMD16(uint32_t) aMTLeft[KNOB_SIMD16_WIDTH], aMTRight[KNOB_SIMD16_WIDTH], aMTTop[KNOB_SIMD16_WIDTH], aMTBottom[KNOB_SIMD16_WIDTH];
+
+        _simd16_store_si(reinterpret_cast<simd16scalari *>(aMTLeft),      bbox.xmin);
+        _simd16_store_si(reinterpret_cast<simd16scalari *>(aMTRight),     bbox.xmax);
+        _simd16_store_si(reinterpret_cast<simd16scalari *>(aMTTop),       bbox.ymin);
+        _simd16_store_si(reinterpret_cast<simd16scalari *>(aMTBottom),    bbox.ymax);
 
         // store render target array index
-        OSALIGNSIMD(uint32_t) aRTAI[KNOB_SIMD_WIDTH];
+        OSALIGNSIMD16(uint32_t) aRTAI[KNOB_SIMD16_WIDTH];
         if (gsState.gsEnable && gsState.emitsRenderTargetArrayIndex)
         {
-            simdvector vRtai[2];
-            pa.Assemble(VERTEX_RTAI_SLOT, vRtai);
-            simdscalari vRtaii = _simd_castps_si(vRtai[0].x);
-            _simd_store_si((simdscalari*)aRTAI, vRtaii);
+            simd16vector vRtai[2];
+            pa.Assemble_simd16(VERTEX_RTAI_SLOT, vRtai);
+            simd16scalari vRtaii = _simd16_castps_si(vRtai[0].x);
+            _simd16_store_si(reinterpret_cast<simd16scalari *>(aRTAI), vRtaii);
         }
         else
         {
-            _simd_store_si((simdscalari*)aRTAI, _simd_setzero_si());
+            _simd16_store_si(reinterpret_cast<simd16scalari *>(aRTAI), _simd16_setzero_si());
         }
 
-        OSALIGNSIMD(float) aPointSize[KNOB_SIMD_WIDTH];
-        _simd_store_ps((float*)aPointSize, vPointSize);
+        OSALIGNSIMD16(float) aPointSize[KNOB_SIMD16_WIDTH];
+        _simd16_store_ps(reinterpret_cast<float *>(aPointSize), vPointSize);
 
         uint32_t *pPrimID = (uint32_t *)&primID;
 
-        OSALIGNSIMD(float) aPrimVertsX[KNOB_SIMD_WIDTH];
-        OSALIGNSIMD(float) aPrimVertsY[KNOB_SIMD_WIDTH];
-        OSALIGNSIMD(float) aPrimVertsZ[KNOB_SIMD_WIDTH];
+        OSALIGNSIMD16(float) aPrimVertsX[KNOB_SIMD16_WIDTH];
+        OSALIGNSIMD16(float) aPrimVertsY[KNOB_SIMD16_WIDTH];
+        OSALIGNSIMD16(float) aPrimVertsZ[KNOB_SIMD16_WIDTH];
 
-        _simd_store_ps((float*)aPrimVertsX, primVerts.x);
-        _simd_store_ps((float*)aPrimVertsY, primVerts.y);
-        _simd_store_ps((float*)aPrimVertsZ, primVerts.z);
+        _simd16_store_ps(reinterpret_cast<float *>(aPrimVertsX), primVerts.x);
+        _simd16_store_ps(reinterpret_cast<float *>(aPrimVertsY), primVerts.y);
+        _simd16_store_ps(reinterpret_cast<float *>(aPrimVertsZ), primVerts.z);
 
         // scan remaining valid prims and bin each separately
         const SWR_BACKEND_STATE& backendState = state.backendState;
@@ -1257,9 +2168,9 @@ void BinPoints(
                 float one = 1.0f;
                 ProcessUserClipDist<1>(pa, primIndex, rastState.clipDistanceMask, &one, dists);
                 for (uint32_t i = 0; i < numClipDist; i++) {
-                    desc.pUserClipBuffer[3*i + 0] = 0.0f;
-                    desc.pUserClipBuffer[3*i + 1] = 0.0f;
-                    desc.pUserClipBuffer[3*i + 2] = dists[i];
+                    desc.pUserClipBuffer[3 * i + 0] = 0.0f;
+                    desc.pUserClipBuffer[3 * i + 1] = 0.0f;
+                    desc.pUserClipBuffer[3 * i + 2] = dists[i];
                 }
             }
 
@@ -1284,46 +2195,6 @@ void BinPoints(
     AR_END(FEBinPoints, 1);
 }
 
-#if USE_SIMD16_FRONTEND
-void BinPoints_simd16(
-    DRAW_CONTEXT *pDC,
-    PA_STATE& pa,
-    uint32_t workerId,
-    simd16vector prim[3],
-    uint32_t primMask,
-    simd16scalari primID,
-    simd16scalari viewportIdx)
-{
-    enum { VERTS_PER_PRIM = 1 };
-
-    simdvector verts[VERTS_PER_PRIM];
-
-    for (uint32_t i = 0; i < VERTS_PER_PRIM; i += 1)
-    {
-        for (uint32_t j = 0; j < 4; j += 1)
-        {
-            verts[i][j] = _simd16_extract_ps(prim[i][j], 0);
-        }
-    }
-
-    pa.useAlternateOffset = false;
-    BinPoints(pDC, pa, workerId, verts, GetPrimMaskLo(primMask), _simd16_extract_si(primID, 0), _simd16_extract_si(viewportIdx, 0));
-
-    if (GetPrimMaskHi(primMask))
-    {
-        for (uint32_t i = 0; i < VERTS_PER_PRIM; i += 1)
-        {
-            for (uint32_t j = 0; j < 4; j += 1)
-            {
-                verts[i][j] = _simd16_extract_ps(prim[i][j], 1);
-            }
-        }
-
-        pa.useAlternateOffset = true;
-        BinPoints(pDC, pa, workerId, verts, GetPrimMaskHi(primMask), _simd16_extract_si(primID, 1), _simd16_extract_si(viewportIdx, 1));
-    }
-}
-
 #endif
 //////////////////////////////////////////////////////////////////////////
 /// @brief Bin SIMD lines to the backend.
@@ -1537,6 +2408,232 @@ endBinLines:
     AR_END(FEBinLines, 1);
 }
 
+#if USE_SIMD16_FRONTEND
+void BinPostSetupLines_simd16(
+    DRAW_CONTEXT *pDC,
+    PA_STATE& pa,
+    uint32_t workerId,
+    simd16vector prim[],
+    simd16scalar recipW[],
+    uint32_t primMask,
+    simd16scalari primID,
+    simd16scalari viewportIdx)
+{
+    SWR_CONTEXT *pContext = pDC->pContext;
+
+    AR_BEGIN(FEBinLines, pDC->drawId);
+
+    const API_STATE& state = GetApiState(pDC);
+    const SWR_RASTSTATE& rastState = state.rastState;
+    const SWR_FRONTEND_STATE& feState = state.frontendState;
+    const SWR_GS_STATE& gsState = state.gsState;
+
+    // Select attribute processor
+    PFN_PROCESS_ATTRIBUTES pfnProcessAttribs = GetProcessAttributesFunc(2,
+        state.backendState.swizzleEnable, state.backendState.constantInterpolationMask);
+
+    simd16scalar& vRecipW0 = recipW[0];
+    simd16scalar& vRecipW1 = recipW[1];
+
+    // convert to fixed point
+    simd16scalari vXi[2], vYi[2];
+
+    vXi[0] = fpToFixedPointVertical(prim[0].x);
+    vYi[0] = fpToFixedPointVertical(prim[0].y);
+    vXi[1] = fpToFixedPointVertical(prim[1].x);
+    vYi[1] = fpToFixedPointVertical(prim[1].y);
+
+    // compute x-major vs y-major mask
+    simd16scalari xLength = _simd16_abs_epi32(_simd16_sub_epi32(vXi[0], vXi[1]));
+    simd16scalari yLength = _simd16_abs_epi32(_simd16_sub_epi32(vYi[0], vYi[1]));
+    simd16scalar vYmajorMask = _simd16_castsi_ps(_simd16_cmpgt_epi32(yLength, xLength));
+    uint32_t yMajorMask = _simd16_movemask_ps(vYmajorMask);
+
+    // cull zero-length lines
+    simd16scalari vZeroLengthMask = _simd16_cmpeq_epi32(xLength, _simd16_setzero_si());
+    vZeroLengthMask = _simd16_and_si(vZeroLengthMask, _simd16_cmpeq_epi32(yLength, _simd16_setzero_si()));
+
+    primMask &= ~_simd16_movemask_ps(_simd16_castsi_ps(vZeroLengthMask));
+
+    uint32_t *pPrimID = (uint32_t *)&primID;
+    const uint32_t *pViewportIndex = (uint32_t *)&viewportIdx;
+
+    // Calc bounding box of lines
+    simd16BBox bbox;
+    bbox.xmin = _simd16_min_epi32(vXi[0], vXi[1]);
+    bbox.xmax = _simd16_max_epi32(vXi[0], vXi[1]);
+    bbox.ymin = _simd16_min_epi32(vYi[0], vYi[1]);
+    bbox.ymax = _simd16_max_epi32(vYi[0], vYi[1]);
+
+    // bloat bbox by line width along minor axis
+    simd16scalar vHalfWidth = _simd16_set1_ps(rastState.lineWidth / 2.0f);
+    simd16scalari vHalfWidthi = fpToFixedPointVertical(vHalfWidth);
+
+    simd16BBox bloatBox;
+
+    bloatBox.xmin = _simd16_sub_epi32(bbox.xmin, vHalfWidthi);
+    bloatBox.xmax = _simd16_add_epi32(bbox.xmax, vHalfWidthi);
+    bloatBox.ymin = _simd16_sub_epi32(bbox.ymin, vHalfWidthi);
+    bloatBox.ymax = _simd16_add_epi32(bbox.ymax, vHalfWidthi);
+
+    bbox.xmin = _simd16_blendv_epi32(bbox.xmin, bloatBox.xmin, vYmajorMask);
+    bbox.xmax = _simd16_blendv_epi32(bbox.xmax, bloatBox.xmax, vYmajorMask);
+    bbox.ymin = _simd16_blendv_epi32(bloatBox.ymin, bbox.ymin, vYmajorMask);
+    bbox.ymax = _simd16_blendv_epi32(bloatBox.ymax, bbox.ymax, vYmajorMask);
+
+    // Intersect with scissor/viewport. Subtract 1 ULP in x.8 fixed point since xmax/ymax edge is exclusive.
+    simd16scalari scisXmin, scisYmin, scisXmax, scisYmax;
+
+    if (state.gsState.emitsViewportArrayIndex)
+    {
+        GatherScissors_simd16<KNOB_SIMD16_WIDTH>::Gather(&state.scissorsInFixedPoint[0], pViewportIndex,
+            scisXmin, scisYmin, scisXmax, scisYmax);
+    }
+    else // broadcast fast path for non-VPAI case.
+    {
+        scisXmin = _simd16_set1_epi32(state.scissorsInFixedPoint[0].xmin);
+        scisYmin = _simd16_set1_epi32(state.scissorsInFixedPoint[0].ymin);
+        scisXmax = _simd16_set1_epi32(state.scissorsInFixedPoint[0].xmax);
+        scisYmax = _simd16_set1_epi32(state.scissorsInFixedPoint[0].ymax);
+    }
+
+    bbox.xmin = _simd16_max_epi32(bbox.xmin, scisXmin);
+    bbox.ymin = _simd16_max_epi32(bbox.ymin, scisYmin);
+    bbox.xmax = _simd16_min_epi32(_simd16_sub_epi32(bbox.xmax, _simd16_set1_epi32(1)), scisXmax);
+    bbox.ymax = _simd16_min_epi32(_simd16_sub_epi32(bbox.ymax, _simd16_set1_epi32(1)), scisYmax);
+
+    // Cull prims completely outside scissor
+    {
+        simd16scalari maskOutsideScissorX = _simd16_cmpgt_epi32(bbox.xmin, bbox.xmax);
+        simd16scalari maskOutsideScissorY = _simd16_cmpgt_epi32(bbox.ymin, bbox.ymax);
+        simd16scalari maskOutsideScissorXY = _simd16_or_si(maskOutsideScissorX, maskOutsideScissorY);
+        uint32_t maskOutsideScissor = _simd16_movemask_ps(_simd16_castsi_ps(maskOutsideScissorXY));
+        primMask = primMask & ~maskOutsideScissor;
+    }
+
+    if (!primMask)
+    {
+        goto endBinLines;
+    }
+
+    // Convert triangle bbox to macrotile units.
+    bbox.xmin = _simd16_srai_epi32(bbox.xmin, KNOB_MACROTILE_X_DIM_FIXED_SHIFT);
+    bbox.ymin = _simd16_srai_epi32(bbox.ymin, KNOB_MACROTILE_Y_DIM_FIXED_SHIFT);
+    bbox.xmax = _simd16_srai_epi32(bbox.xmax, KNOB_MACROTILE_X_DIM_FIXED_SHIFT);
+    bbox.ymax = _simd16_srai_epi32(bbox.ymax, KNOB_MACROTILE_Y_DIM_FIXED_SHIFT);
+
+    OSALIGNSIMD16(uint32_t) aMTLeft[KNOB_SIMD16_WIDTH], aMTRight[KNOB_SIMD16_WIDTH], aMTTop[KNOB_SIMD16_WIDTH], aMTBottom[KNOB_SIMD16_WIDTH];
+
+    _simd16_store_si(reinterpret_cast<simd16scalari *>(aMTLeft),    bbox.xmin);
+    _simd16_store_si(reinterpret_cast<simd16scalari *>(aMTRight),   bbox.xmax);
+    _simd16_store_si(reinterpret_cast<simd16scalari *>(aMTTop),     bbox.ymin);
+    _simd16_store_si(reinterpret_cast<simd16scalari *>(aMTBottom),  bbox.ymax);
+
+    // transpose verts needed for backend
+    /// @todo modify BE to take non-transformed verts
+    __m128 vHorizX[2][KNOB_SIMD_WIDTH]; // KNOB_SIMD16_WIDTH
+    __m128 vHorizY[2][KNOB_SIMD_WIDTH]; // KNOB_SIMD16_WIDTH
+    __m128 vHorizZ[2][KNOB_SIMD_WIDTH]; // KNOB_SIMD16_WIDTH
+    __m128 vHorizW[2][KNOB_SIMD_WIDTH]; // KNOB_SIMD16_WIDTH
+
+    const simdscalar unused = _simd_setzero_ps();
+
+    vTranspose3x8(vHorizX[0], _simd16_extract_ps(prim[0].x, 0), _simd16_extract_ps(prim[1].x, 0), unused);
+    vTranspose3x8(vHorizY[0], _simd16_extract_ps(prim[0].y, 0), _simd16_extract_ps(prim[1].y, 0), unused);
+    vTranspose3x8(vHorizZ[0], _simd16_extract_ps(prim[0].z, 0), _simd16_extract_ps(prim[1].z, 0), unused);
+    vTranspose3x8(vHorizW[0], _simd16_extract_ps(vRecipW0, 0),  _simd16_extract_ps(vRecipW1, 0),  unused);
+
+    vTranspose3x8(vHorizX[1], _simd16_extract_ps(prim[0].x, 1), _simd16_extract_ps(prim[1].x, 1), unused);
+    vTranspose3x8(vHorizY[1], _simd16_extract_ps(prim[0].y, 1), _simd16_extract_ps(prim[1].y, 1), unused);
+    vTranspose3x8(vHorizZ[1], _simd16_extract_ps(prim[0].z, 1), _simd16_extract_ps(prim[1].z, 1), unused);
+    vTranspose3x8(vHorizW[1], _simd16_extract_ps(vRecipW0, 1),  _simd16_extract_ps(vRecipW1, 1),  unused);
+
+    // store render target array index
+    OSALIGNSIMD16(uint32_t) aRTAI[KNOB_SIMD16_WIDTH];
+    if (gsState.gsEnable && gsState.emitsRenderTargetArrayIndex)
+    {
+        simd16vector vRtai[2];
+        pa.Assemble_simd16(VERTEX_RTAI_SLOT, vRtai);
+        simd16scalari vRtaii = _simd16_castps_si(vRtai[0].x);
+        _simd16_store_si(reinterpret_cast<simd16scalari *>(aRTAI), vRtaii);
+    }
+    else
+    {
+        _simd16_store_si(reinterpret_cast<simd16scalari *>(aRTAI), _simd16_setzero_si());
+    }
+
+    // scan remaining valid prims and bin each separately
+    DWORD primIndex;
+    while (_BitScanForward(&primIndex, primMask))
+    {
+        uint32_t linkageCount = state.backendState.numAttributes;
+        uint32_t numScalarAttribs = linkageCount * 4;
+
+        BE_WORK work;
+        work.type = DRAW;
+
+        TRIANGLE_WORK_DESC &desc = work.desc.tri;
+
+        desc.triFlags.frontFacing = 1;
+        desc.triFlags.primID = pPrimID[primIndex];
+        desc.triFlags.yMajor = (yMajorMask >> primIndex) & 1;
+        desc.triFlags.renderTargetArrayIndex = aRTAI[primIndex];
+        desc.triFlags.viewportIndex = pViewportIndex[primIndex];
+
+        work.pfnWork = RasterizeLine;
+
+        auto pArena = pDC->pArena;
+        SWR_ASSERT(pArena != nullptr);
+
+        // store active attribs
+        desc.pAttribs = (float*)pArena->AllocAligned(numScalarAttribs * 3 * sizeof(float), 16);
+        desc.numAttribs = linkageCount;
+        pfnProcessAttribs(pDC, pa, primIndex, pPrimID[primIndex], desc.pAttribs);
+
+        // store line vertex data
+        desc.pTriBuffer = (float*)pArena->AllocAligned(4 * 4 * sizeof(float), 16);
+
+        {
+            const uint32_t i = primIndex >> 3;  // triIndex / KNOB_SIMD_WIDTH
+            const uint32_t j = primIndex & 7;   // triIndex % KNOB_SIMD_WIDTH
+
+            _mm_store_ps(&desc.pTriBuffer[ 0], vHorizX[i][j]);
+            _mm_store_ps(&desc.pTriBuffer[ 4], vHorizY[i][j]);
+            _mm_store_ps(&desc.pTriBuffer[ 8], vHorizZ[i][j]);
+            _mm_store_ps(&desc.pTriBuffer[12], vHorizW[i][j]);
+        }
+
+        // store user clip distances
+        if (rastState.clipDistanceMask)
+        {
+            uint32_t numClipDist = _mm_popcnt_u32(rastState.clipDistanceMask);
+            desc.pUserClipBuffer = (float*)pArena->Alloc(numClipDist * 2 * sizeof(float));
+            ProcessUserClipDist<2>(pa, primIndex, rastState.clipDistanceMask, &desc.pTriBuffer[12], desc.pUserClipBuffer);
+        }
+
+        MacroTileMgr *pTileMgr = pDC->pTileMgr;
+        for (uint32_t y = aMTTop[primIndex]; y <= aMTBottom[primIndex]; ++y)
+        {
+            for (uint32_t x = aMTLeft[primIndex]; x <= aMTRight[primIndex]; ++x)
+            {
+#if KNOB_ENABLE_TOSS_POINTS
+                if (!KNOB_TOSS_SETUP_TRIS)
+#endif
+                {
+                    pTileMgr->enqueue(x, y, &work);
+                }
+            }
+        }
+
+        primMask &= ~(1 << primIndex);
+    }
+
+endBinLines:
+
+    AR_END(FEBinLines, 1);
+}
+
+#endif
 //////////////////////////////////////////////////////////////////////////
 /// @brief Bin SIMD lines to the backend.
 /// @param pDC - pointer to draw context.
@@ -1622,34 +2719,63 @@ void BinLines_simd16(
     simd16scalari primID,
     simd16scalari viewportIdx)
 {
-    enum { VERTS_PER_PRIM = 2 };
+    SWR_CONTEXT *pContext = pDC->pContext;
 
-    simdvector verts[VERTS_PER_PRIM];
+    const API_STATE& state = GetApiState(pDC);
+    const SWR_RASTSTATE& rastState = state.rastState;
+    const SWR_FRONTEND_STATE& feState = state.frontendState;
+    const SWR_GS_STATE& gsState = state.gsState;
 
-    for (uint32_t i = 0; i < VERTS_PER_PRIM; i += 1)
-    {
-        for (uint32_t j = 0; j < 4; j += 1)
-        {
-            verts[i][j] = _simd16_extract_ps(prim[i][j], 0);
-        }
-    }
+    // Select attribute processor
+    PFN_PROCESS_ATTRIBUTES pfnProcessAttribs = GetProcessAttributesFunc(2,
+        state.backendState.swizzleEnable, state.backendState.constantInterpolationMask);
 
-    pa.useAlternateOffset = false;
-    BinLines(pDC, pa, workerId, verts, GetPrimMaskLo(primMask), _simd16_extract_si(primID, 0), _simd16_extract_si(viewportIdx, 0));
+    simd16scalar vRecipW[2] = { _simd16_set1_ps(1.0f), _simd16_set1_ps(1.0f) };
 
-    if (GetPrimMaskHi(primMask))
+    if (!feState.vpTransformDisable)
     {
-        for (uint32_t i = 0; i < VERTS_PER_PRIM; i += 1)
+        // perspective divide
+        vRecipW[0] = _simd16_div_ps(_simd16_set1_ps(1.0f), prim[0].w);
+        vRecipW[1] = _simd16_div_ps(_simd16_set1_ps(1.0f), prim[1].w);
+
+        prim[0].v[0] = _simd16_mul_ps(prim[0].v[0], vRecipW[0]);
+        prim[1].v[0] = _simd16_mul_ps(prim[1].v[0], vRecipW[1]);
+
+        prim[0].v[1] = _simd16_mul_ps(prim[0].v[1], vRecipW[0]);
+        prim[1].v[1] = _simd16_mul_ps(prim[1].v[1], vRecipW[1]);
+
+        prim[0].v[2] = _simd16_mul_ps(prim[0].v[2], vRecipW[0]);
+        prim[1].v[2] = _simd16_mul_ps(prim[1].v[2], vRecipW[1]);
+
+        // viewport transform to screen coords
+        if (state.gsState.emitsViewportArrayIndex)
         {
-            for (uint32_t j = 0; j < 4; j += 1)
-            {
-                verts[i][j] = _simd16_extract_ps(prim[i][j], 1);
-            }
+            viewportTransform<2>(prim, state.vpMatrices, viewportIdx);
+        }
+        else
+        {
+            viewportTransform<2>(prim, state.vpMatrices);
         }
+}
 
-        pa.useAlternateOffset = true;
-        BinLines(pDC, pa, workerId, verts, GetPrimMaskHi(primMask), _simd16_extract_si(primID, 1), _simd16_extract_si(viewportIdx, 1));
-    }
+    // adjust for pixel center location
+    simd16scalar offset = g_pixelOffsets_simd16[rastState.pixelLocation];
+
+    prim[0].x = _simd16_add_ps(prim[0].x, offset);
+    prim[0].y = _simd16_add_ps(prim[0].y, offset);
+
+    prim[1].x = _simd16_add_ps(prim[1].x, offset);
+    prim[1].y = _simd16_add_ps(prim[1].y, offset);
+
+    BinPostSetupLines_simd16(
+        pDC,
+        pa,
+        workerId,
+        prim,
+        vRecipW,
+        primMask,
+        primID,
+        viewportIdx);
 }
 
 #endif
diff --git a/src/gallium/drivers/swr/rasterizer/core/frontend.h b/src/gallium/drivers/swr/rasterizer/core/frontend.h
index 37b7215..e880ead 100644
--- a/src/gallium/drivers/swr/rasterizer/core/frontend.h
+++ b/src/gallium/drivers/swr/rasterizer/core/frontend.h
@@ -112,6 +112,23 @@ void triangleSetupABIntVertical(const simdscalari vX[3], const simdscalari vY[3]
     vB[1] = _simd_sub_epi32(vX[2], vX[1]);
     vB[2] = _simd_sub_epi32(vX[0], vX[2]);
 }
+
+#if ENABLE_AVX512_SIMD16
+INLINE
+void triangleSetupABIntVertical(const simd16scalari vX[3], const simd16scalari vY[3], simd16scalari(&vA)[3], simd16scalari(&vB)[3])
+{
+    // A = y0 - y1
+    // B = x1 - x0
+    vA[0] = _simd16_sub_epi32(vY[0], vY[1]);
+    vA[1] = _simd16_sub_epi32(vY[1], vY[2]);
+    vA[2] = _simd16_sub_epi32(vY[2], vY[0]);
+
+    vB[0] = _simd16_sub_epi32(vX[1], vX[0]);
+    vB[1] = _simd16_sub_epi32(vX[2], vX[1]);
+    vB[2] = _simd16_sub_epi32(vX[0], vX[2]);
+}
+
+#endif
 // Calculate the determinant of the triangle
 // 2 vectors between the 3 points: P, Q
 // Px = x0-x2, Py = y0-y2
@@ -185,6 +202,44 @@ void calcDeterminantIntVertical(const simdscalari vA[3], const simdscalari vB[3]
     pvDet[1] = vResultHi;
 }
 
+#if ENABLE_AVX512_SIMD16
+INLINE
+void calcDeterminantIntVertical(const simd16scalari vA[3], const simd16scalari vB[3], simd16scalari *pvDet)
+{
+    // refer to calcDeterminantInt comment for calculation explanation
+    // A1*B2
+    simd16scalari vA1Lo = _simd16_unpacklo_epi32(vA[1], vA[1]); // 0 0 1 1 4 4 5 5
+    simd16scalari vA1Hi = _simd16_unpackhi_epi32(vA[1], vA[1]); // 2 2 3 3 6 6 7 7
+
+    simd16scalari vB2Lo = _simd16_unpacklo_epi32(vB[2], vB[2]);
+    simd16scalari vB2Hi = _simd16_unpackhi_epi32(vB[2], vB[2]);
+
+    simd16scalari vA1B2Lo = _simd16_mul_epi32(vA1Lo, vB2Lo);    // 0 1 4 5
+    simd16scalari vA1B2Hi = _simd16_mul_epi32(vA1Hi, vB2Hi);    // 2 3 6 7
+
+    // B1*A2
+    simd16scalari vA2Lo = _simd16_unpacklo_epi32(vA[2], vA[2]);
+    simd16scalari vA2Hi = _simd16_unpackhi_epi32(vA[2], vA[2]);
+
+    simd16scalari vB1Lo = _simd16_unpacklo_epi32(vB[1], vB[1]);
+    simd16scalari vB1Hi = _simd16_unpackhi_epi32(vB[1], vB[1]);
+
+    simd16scalari vA2B1Lo = _simd16_mul_epi32(vA2Lo, vB1Lo);
+    simd16scalari vA2B1Hi = _simd16_mul_epi32(vA2Hi, vB1Hi);
+
+    // A1*B2 - A2*B1
+    simd16scalari detLo = _simd16_sub_epi64(vA1B2Lo, vA2B1Lo);
+    simd16scalari detHi = _simd16_sub_epi64(vA1B2Hi, vA2B1Hi);
+
+    // shuffle 0 1 4 5 -> 0 1 2 3
+    simd16scalari vResultLo = _simd16_permute2f128_si(detLo, detHi, 0x20);
+    simd16scalari vResultHi = _simd16_permute2f128_si(detLo, detHi, 0x31);
+
+    pvDet[0] = vResultLo;
+    pvDet[1] = vResultHi;
+}
+
+#endif
 INLINE
 void triangleSetupC(const __m128 vX, const __m128 vY, const __m128 vA, const __m128 &vB, __m128 &vC)
 {
@@ -227,6 +282,27 @@ void viewportTransform(simdvector *v, const SWR_VIEWPORT_MATRICES & vpMatrices)
     }
 }
 
+#if USE_SIMD16_FRONTEND
+template<uint32_t NumVerts>
+INLINE
+void viewportTransform(simd16vector *v, const SWR_VIEWPORT_MATRICES & vpMatrices)
+{
+    const simd16scalar m00 = _simd16_broadcast_ss(&vpMatrices.m00[0]);
+    const simd16scalar m30 = _simd16_broadcast_ss(&vpMatrices.m30[0]);
+    const simd16scalar m11 = _simd16_broadcast_ss(&vpMatrices.m11[0]);
+    const simd16scalar m31 = _simd16_broadcast_ss(&vpMatrices.m31[0]);
+    const simd16scalar m22 = _simd16_broadcast_ss(&vpMatrices.m22[0]);
+    const simd16scalar m32 = _simd16_broadcast_ss(&vpMatrices.m32[0]);
+
+    for (uint32_t i = 0; i < NumVerts; ++i)
+    {
+        v[i].x = _simd16_fmadd_ps(v[i].x, m00, m30);
+        v[i].y = _simd16_fmadd_ps(v[i].y, m11, m31);
+        v[i].z = _simd16_fmadd_ps(v[i].z, m22, m32);
+    }
+}
+
+#endif
 template<uint32_t NumVerts>
 INLINE
 void viewportTransform(simdvector *v, const SWR_VIEWPORT_MATRICES & vpMatrices, simdscalari vViewportIdx)
@@ -247,6 +323,28 @@ void viewportTransform(simdvector *v, const SWR_VIEWPORT_MATRICES & vpMatrices,
     }
 }
 
+#if USE_SIMD16_FRONTEND
+template<uint32_t NumVerts>
+INLINE
+void viewportTransform(simd16vector *v, const SWR_VIEWPORT_MATRICES & vpMatrices, simd16scalari vViewportIdx)
+{
+    // perform a gather of each matrix element based on the viewport array indexes
+    const simd16scalar m00 = _simd16_i32gather_ps(&vpMatrices.m00[0], vViewportIdx, 4);
+    const simd16scalar m30 = _simd16_i32gather_ps(&vpMatrices.m30[0], vViewportIdx, 4);
+    const simd16scalar m11 = _simd16_i32gather_ps(&vpMatrices.m11[0], vViewportIdx, 4);
+    const simd16scalar m31 = _simd16_i32gather_ps(&vpMatrices.m31[0], vViewportIdx, 4);
+    const simd16scalar m22 = _simd16_i32gather_ps(&vpMatrices.m22[0], vViewportIdx, 4);
+    const simd16scalar m32 = _simd16_i32gather_ps(&vpMatrices.m32[0], vViewportIdx, 4);
+
+    for (uint32_t i = 0; i < NumVerts; ++i)
+    {
+        v[i].x = _simd16_fmadd_ps(v[i].x, m00, m30);
+        v[i].y = _simd16_fmadd_ps(v[i].y, m11, m31);
+        v[i].z = _simd16_fmadd_ps(v[i].z, m22, m32);
+    }
+}
+
+#endif
 INLINE
 void calcBoundingBoxInt(const __m128i &vX, const __m128i &vY, SWR_RECT &bbox)
 {
diff --git a/src/gallium/drivers/swr/rasterizer/core/utils.h b/src/gallium/drivers/swr/rasterizer/core/utils.h
index 9dfa16a..660a63f 100644
--- a/src/gallium/drivers/swr/rasterizer/core/utils.h
+++ b/src/gallium/drivers/swr/rasterizer/core/utils.h
@@ -83,6 +83,16 @@ struct simdBBox
     simdscalari xmax;
 };
 
+#if ENABLE_AVX512_SIMD16
+struct simd16BBox
+{
+    simd16scalari ymin;
+    simd16scalari ymax;
+    simd16scalari xmin;
+    simd16scalari xmax;
+};
+
+#endif
 INLINE
 void vTranspose(__m128 &row0, __m128 &row1, __m128 &row2, __m128 &row3)
 {
-- 
2.7.4



More information about the mesa-dev mailing list