[Mesa-dev] [PATCH 09/24] swr/rast: SIMD16 FE - interleaved simdvertex output in GS

[Tim Rowley]

Eliminates conversion copies on GS output from simdvertex to simd16vertex.
---
 .../drivers/swr/rasterizer/core/frontend.cpp       | 22 ++++------------
 src/gallium/drivers/swr/swr_shader.cpp             | 29 +++++++++++++++++++---
 2 files changed, 31 insertions(+), 20 deletions(-)

diff --git a/src/gallium/drivers/swr/rasterizer/core/frontend.cpp b/src/gallium/drivers/swr/rasterizer/core/frontend.cpp
index 3886c64..e88246f 100644
--- a/src/gallium/drivers/swr/rasterizer/core/frontend.cpp
+++ b/src/gallium/drivers/swr/rasterizer/core/frontend.cpp
@@ -717,10 +717,6 @@ void ProcessStreamIdBuffer(uint32_t stream, uint8_t* pStreamIdBase, uint32_t num
 
 THREAD SWR_GS_CONTEXT tlsGsContext;
 
-#if USE_SIMD16_FRONTEND
-THREAD simd16vertex tempVertex_simd16[128];
-
-#endif
 template<typename SIMDVERTEX, uint32_t SIMD_WIDTH>
 struct GsBufferInfo
 {
@@ -819,7 +815,11 @@ static void GeometryShaderStage(
         tlsGsContext.vert[i].attrib[VERTEX_POSITION_SLOT] = attrib[i];
     }
 
+#if USE_SIMD16_FRONTEND
+    const GsBufferInfo<simd16vertex, KNOB_SIMD16_WIDTH> bufferInfo(state.gsState);
+#else
     const GsBufferInfo<simdvertex, KNOB_SIMD_WIDTH> bufferInfo(state.gsState);
+#endif
 
     // record valid prims from the frontend to avoid over binning the newly generated
     // prims from the GS
@@ -923,19 +923,7 @@ static void GeometryShaderStage(
                 }
 
 #if USE_SIMD16_FRONTEND
-                // TEMPORARY: GS outputs simdvertex, PA inputs simd16vertex, so convert simdvertex to simd16vertex
-
-                SWR_ASSERT(numEmittedVerts <= 256);
-
-                PackPairsOfSimdVertexIntoSimd16Vertex(
-                    tempVertex_simd16,
-                    reinterpret_cast<const simdvertex *>(pBase),
-                    numEmittedVerts,
-                    SWR_VTX_NUM_SLOTS);
-
-#endif
-#if USE_SIMD16_FRONTEND
-                PA_STATE_CUT gsPa(pDC, reinterpret_cast<uint8_t *>(tempVertex_simd16), numEmittedVerts, reinterpret_cast<simd16mask *>(pCutBuffer), numEmittedVerts, numAttribs, pState->outputTopology, processCutVerts);
+                PA_STATE_CUT gsPa(pDC, pBase, numEmittedVerts, reinterpret_cast<simd16mask *>(pCutBuffer), numEmittedVerts, numAttribs, pState->outputTopology, processCutVerts);
 
 #else
                 PA_STATE_CUT gsPa(pDC, pBase, numEmittedVerts, pCutBuffer, numEmittedVerts, numAttribs, pState->outputTopology, processCutVerts);
diff --git a/src/gallium/drivers/swr/swr_shader.cpp b/src/gallium/drivers/swr/swr_shader.cpp
index d55820e..2f495f5 100644
--- a/src/gallium/drivers/swr/swr_shader.cpp
+++ b/src/gallium/drivers/swr/swr_shader.cpp
@@ -370,8 +370,13 @@ BuilderSWR::swr_gs_llvm_emit_vertex(const struct lp_build_tgsi_gs_iface *gs_base
 
     IRB()->SetInsertPoint(unwrap(LLVMGetInsertBlock(gallivm->builder)));
 
+#if USE_SIMD16_FRONTEND
+    const uint32_t simdVertexStride = sizeof(simdvertex) * 2;
+    const uint32_t numSimdBatches = (pGS->maxNumVerts + (mVWidth * 2) - 1) / (mVWidth * 2);
+#else
     const uint32_t simdVertexStride = sizeof(simdvertex);
-    const uint32_t numSimdBatches = (pGS->maxNumVerts + 7) / 8;
+    const uint32_t numSimdBatches = (pGS->maxNumVerts + mVWidth - 1) / mVWidth;
+#endif
     const uint32_t inputPrimStride = numSimdBatches * simdVertexStride;
 
     Value *pStream = LOAD(iface->pGsCtx, { 0, SWR_GS_CONTEXT_pStream });
@@ -388,8 +393,14 @@ BuilderSWR::swr_gs_llvm_emit_vertex(const struct lp_build_tgsi_gs_iface *gs_base
           inputPrimStride * 6,
           inputPrimStride * 7 } );
 
-    Value *vVertexSlot = ASHR(unwrap(emitted_vertices_vec), 3);
-    Value *vSimdSlot = AND(unwrap(emitted_vertices_vec), 7);
+#if USE_SIMD16_FRONTEND
+    const uint32_t simdShift = log2(mVWidth * 2);
+    Value *vSimdSlot = AND(unwrap(emitted_vertices_vec), (mVWidth * 2) - 1);
+#else
+    const uint32_t simdShift = log2(mVWidth);
+    Value *vSimdSlot = AND(unwrap(emitted_vertices_vec), mVWidth - 1);
+#endif
+    Value *vVertexSlot = ASHR(unwrap(emitted_vertices_vec), simdShift);
 
     for (uint32_t attrib = 0; attrib < iface->num_outputs; ++attrib) {
        uint32_t attribSlot = attrib;
@@ -400,10 +411,17 @@ BuilderSWR::swr_gs_llvm_emit_vertex(const struct lp_build_tgsi_gs_iface *gs_base
        else if (iface->info->output_semantic_name[attrib] == TGSI_SEMANTIC_LAYER)
           attribSlot = VERTEX_RTAI_SLOT;
 
+#if USE_SIMD16_FRONTEND
+       Value *vOffsetsAttrib =
+          ADD(vOffsets, MUL(vVertexSlot, VIMMED1((uint32_t)sizeof(simdvertex) * 2)));
+       vOffsetsAttrib =
+          ADD(vOffsetsAttrib, VIMMED1((uint32_t)(attribSlot*sizeof(simdvector) * 2)));
+#else
        Value *vOffsetsAttrib =
           ADD(vOffsets, MUL(vVertexSlot, VIMMED1((uint32_t)sizeof(simdvertex))));
        vOffsetsAttrib =
           ADD(vOffsetsAttrib, VIMMED1((uint32_t)(attribSlot*sizeof(simdvector))));
+#endif
        vOffsetsAttrib =
           ADD(vOffsetsAttrib, MUL(vSimdSlot, VIMMED1((uint32_t)sizeof(float))));
 
@@ -416,8 +434,13 @@ BuilderSWR::swr_gs_llvm_emit_vertex(const struct lp_build_tgsi_gs_iface *gs_base
 
           MASKED_SCATTER(vData, vPtrs, 32, vMask1);
 
+#if USE_SIMD16_FRONTEND
+          vOffsetsAttrib =
+             ADD(vOffsetsAttrib, VIMMED1((uint32_t)sizeof(simdscalar) * 2));
+#else
           vOffsetsAttrib =
              ADD(vOffsetsAttrib, VIMMED1((uint32_t)sizeof(simdscalar)));
+#endif
        }
     }
 }
-- 
2.7.4