From 4cb69e817c0cec430ff956b979409bcfbfb4e3ca Mon Sep 17 00:00:00 2001 From: Tim Rowley Date: Tue, 7 Mar 2017 16:23:18 -0800 Subject: [PATCH] swr: [rasterizer core] SIMD16 Frontend WIP - fix tesselation crashes Reviewed-by: Bruce Cherniak --- .../drivers/swr/rasterizer/core/frontend.cpp | 19 ++++----- src/gallium/drivers/swr/rasterizer/core/pa.h | 7 ++-- .../drivers/swr/rasterizer/core/pa_avx.cpp | 40 ++++++++++--------- 3 files changed, 35 insertions(+), 31 deletions(-) diff --git a/src/gallium/drivers/swr/rasterizer/core/frontend.cpp b/src/gallium/drivers/swr/rasterizer/core/frontend.cpp index 99d047da420..acbd779c328 100644 --- a/src/gallium/drivers/swr/rasterizer/core/frontend.cpp +++ b/src/gallium/drivers/swr/rasterizer/core/frontend.cpp @@ -610,6 +610,8 @@ INLINE static T RoundDownEven(T value) /// /// attribCount will limit the vector copies to those attribs specified /// +/// note: the stride between vertexes is determinded by KNOB_NUM_ATTRIBUTES +/// void PackPairsOfSimdVertexIntoSimd16VertexInPlace(simdvertex *vertex, uint32_t vertexCount, uint32_t attribCount) { SWR_ASSERT(vertex); @@ -1244,7 +1246,7 @@ static void TessellationStages( uint32_t requiredDSVectorInvocations = AlignUp(tsData.NumDomainPoints, KNOB_SIMD_WIDTH) / KNOB_SIMD_WIDTH; size_t requiredDSOutputVectors = requiredDSVectorInvocations * tsState.numDsOutputAttribs; #if USE_SIMD16_FRONTEND - size_t requiredAllocSize = sizeof(simdvector) * RoundUpEven(requiredDSOutputVectors); // simd8 -> simd16, padding + size_t requiredAllocSize = sizeof(simdvector) * RoundUpEven(requiredDSVectorInvocations) * tsState.numDsOutputAttribs; // simd8 -> simd16, padding #else size_t requiredAllocSize = sizeof(simdvector) * requiredDSOutputVectors; #endif @@ -1253,7 +1255,7 @@ static void TessellationStages( AlignedFree(gt_pTessellationThreadData->pDSOutput); gt_pTessellationThreadData->pDSOutput = (simdscalar*)AlignedMalloc(requiredAllocSize, 64); #if USE_SIMD16_FRONTEND - gt_pTessellationThreadData->numDSOutputVectors = RoundUpEven(requiredDSOutputVectors); // simd8 -> simd16, padding + gt_pTessellationThreadData->numDSOutputVectors = RoundUpEven(requiredDSVectorInvocations) * tsState.numDsOutputAttribs; // simd8 -> simd16, padding #else gt_pTessellationThreadData->numDSOutputVectors = requiredDSOutputVectors; #endif @@ -1272,7 +1274,11 @@ static void TessellationStages( dsContext.pDomainU = (simdscalar*)tsData.pDomainPointsU; dsContext.pDomainV = (simdscalar*)tsData.pDomainPointsV; dsContext.pOutputData = gt_pTessellationThreadData->pDSOutput; +#if USE_SIMD16_FRONTEND + dsContext.vectorStride = RoundUpEven(requiredDSVectorInvocations); // simd8 -> simd16 +#else dsContext.vectorStride = requiredDSVectorInvocations; +#endif uint32_t dsInvocations = 0; @@ -1289,19 +1295,14 @@ static void TessellationStages( UPDATE_STAT_FE(DsInvocations, tsData.NumDomainPoints); #if USE_SIMD16_FRONTEND - // TEMPORARY: DS outputs simdvertex, PA inputs simd16vertex, so convert simdvertex to simd16vertex, in-place - - PackPairsOfSimdVertexIntoSimd16VertexInPlace( - reinterpret_cast(dsContext.pOutputData), - RoundUpEven(dsContext.vectorStride), // simd8 -> simd16 - tsState.numDsOutputAttribs); + SWR_ASSERT(IsEven(dsContext.vectorStride)); // simd8 -> simd16 #endif PA_TESS tessPa( pDC, #if USE_SIMD16_FRONTEND reinterpret_cast(dsContext.pOutputData), // simd8 -> simd16 - RoundUpEven(dsContext.vectorStride) / 2, // simd8 -> simd16 + dsContext.vectorStride / 2, // simd8 -> simd16 #else dsContext.pOutputData, dsContext.vectorStride, diff --git a/src/gallium/drivers/swr/rasterizer/core/pa.h b/src/gallium/drivers/swr/rasterizer/core/pa.h index afeed657333..6160996cad3 100644 --- a/src/gallium/drivers/swr/rasterizer/core/pa.h +++ b/src/gallium/drivers/swr/rasterizer/core/pa.h @@ -233,8 +233,6 @@ struct PA_STATE_OPT : public PA_STATE this->reset = false; } - this->pfnPaFunc = this->pfnPaNextFunc; - if (!HasWork()) { morePrims = false; // no more to do @@ -290,12 +288,14 @@ struct PA_STATE_OPT : public PA_STATE #if ENABLE_AVX512_SIMD16 void SetNextState_simd16(PA_STATE_OPT::PFN_PA_FUNC_SIMD16 pfnPaNextFunc_simd16, + PA_STATE_OPT::PFN_PA_FUNC pfnPaNextFunc, PA_STATE_OPT::PFN_PA_SINGLE_FUNC pfnPaNextSingleFunc, uint32_t numSimdPrims = 0, uint32_t numPrimsIncrement = 0, bool reset = false) { this->pfnPaNextFunc_simd16 = pfnPaNextFunc_simd16; + this->pfnPaNextFunc = pfnPaNextFunc; this->nextNumSimdPrims = numSimdPrims; this->nextNumPrimsIncrement = numPrimsIncrement; this->nextReset = reset; @@ -344,12 +344,13 @@ INLINE void SetNextPaState(PA_STATE_OPT& pa, PA_STATE_OPT::PFN_PA_FUNC pfnPaNext #if ENABLE_AVX512_SIMD16 INLINE void SetNextPaState_simd16(PA_STATE_OPT& pa, PA_STATE_OPT::PFN_PA_FUNC_SIMD16 pfnPaNextFunc_simd16, + PA_STATE_OPT::PFN_PA_FUNC pfnPaNextFunc, PA_STATE_OPT::PFN_PA_SINGLE_FUNC pfnPaNextSingleFunc, uint32_t numSimdPrims = 0, uint32_t numPrimsIncrement = 0, bool reset = false) { - return pa.SetNextState_simd16(pfnPaNextFunc_simd16, pfnPaNextSingleFunc, numSimdPrims, numPrimsIncrement, reset); + return pa.SetNextState_simd16(pfnPaNextFunc_simd16, pfnPaNextFunc, pfnPaNextSingleFunc, numSimdPrims, numPrimsIncrement, reset); } #endif diff --git a/src/gallium/drivers/swr/rasterizer/core/pa_avx.cpp b/src/gallium/drivers/swr/rasterizer/core/pa_avx.cpp index 23bf1e0fff4..2a4104aacd3 100644 --- a/src/gallium/drivers/swr/rasterizer/core/pa_avx.cpp +++ b/src/gallium/drivers/swr/rasterizer/core/pa_avx.cpp @@ -469,6 +469,7 @@ static bool PaPatchList_simd16(PA_STATE_OPT& pa, uint32_t slot, simd16vector ver SetNextPaState_simd16( pa, PaPatchList_simd16, + PaPatchList, PaPatchListSingle); return false; @@ -505,6 +506,7 @@ static bool PaPatchListTerm_simd16(PA_STATE_OPT& pa, uint32_t slot, simd16vector SetNextPaState_simd16( pa, PaPatchList_simd16, + PaPatchList, PaPatchListSingle, 0, KNOB_SIMD16_WIDTH, @@ -741,13 +743,13 @@ bool PaTriList2(PA_STATE_OPT& pa, uint32_t slot, simdvector verts[]) #if ENABLE_AVX512_SIMD16 bool PaTriList0_simd16(PA_STATE_OPT& pa, uint32_t slot, simd16vector verts[]) { - SetNextPaState_simd16(pa, PaTriList1_simd16, PaTriListSingle0); + SetNextPaState_simd16(pa, PaTriList1_simd16, PaTriList1, PaTriListSingle0); return false; // Not enough vertices to assemble 16 triangles } bool PaTriList1_simd16(PA_STATE_OPT& pa, uint32_t slot, simd16vector verts[]) { - SetNextPaState_simd16(pa, PaTriList2_simd16, PaTriListSingle0); + SetNextPaState_simd16(pa, PaTriList2_simd16, PaTriList2, PaTriListSingle0); return false; // Not enough vertices to assemble 16 triangles } @@ -781,7 +783,7 @@ bool PaTriList2_simd16(PA_STATE_OPT& pa, uint32_t slot, simd16vector verts[]) v2[i] = _simd16_permute_ps(temp2, perm2); } - SetNextPaState_simd16(pa, PaTriList0_simd16, PaTriListSingle0, 0, KNOB_SIMD16_WIDTH, true); + SetNextPaState_simd16(pa, PaTriList0_simd16, PaTriList0, PaTriListSingle0, 0, KNOB_SIMD16_WIDTH, true); return true; } @@ -1019,7 +1021,7 @@ bool PaTriStrip1(PA_STATE_OPT& pa, uint32_t slot, simdvector verts[]) #if ENABLE_AVX512_SIMD16 bool PaTriStrip0_simd16(PA_STATE_OPT& pa, uint32_t slot, simd16vector verts[]) { - SetNextPaState_simd16(pa, PaTriStrip1_simd16, PaTriStripSingle0); + SetNextPaState_simd16(pa, PaTriStrip1_simd16, PaTriStrip1, PaTriStripSingle0); return false; // Not enough vertices to assemble 16 triangles. } @@ -1050,7 +1052,7 @@ bool PaTriStrip1_simd16(PA_STATE_OPT& pa, uint32_t slot, simd16vector verts[]) v2[i] = _simd16_shuffle_ps(a[i], shuff, _MM_SHUFFLE(2, 2, 2, 2)); // a2 a2 a4 a4 a6 a6 a8 a8 aA aA aC aC aE aE b0 b0 } - SetNextPaState_simd16(pa, PaTriStrip1_simd16, PaTriStripSingle0, 0, KNOB_SIMD16_WIDTH); + SetNextPaState_simd16(pa, PaTriStrip1_simd16, PaTriStrip1, PaTriStripSingle0, 0, KNOB_SIMD16_WIDTH); return true; } @@ -1285,7 +1287,7 @@ bool PaTriFan1(PA_STATE_OPT& pa, uint32_t slot, simdvector verts[]) #if ENABLE_AVX512_SIMD16 bool PaTriFan0_simd16(PA_STATE_OPT& pa, uint32_t slot, simd16vector verts[]) { - SetNextPaState_simd16(pa, PaTriFan1_simd16, PaTriFanSingle0); + SetNextPaState_simd16(pa, PaTriFan1_simd16, PaTriFan1, PaTriFanSingle0); return false; // Not enough vertices to assemble 16 triangles. } @@ -1319,7 +1321,7 @@ bool PaTriFan1_simd16(PA_STATE_OPT& pa, uint32_t slot, simd16vector verts[]) v1[i] = _simd16_shuffle_ps(b[i], v2[i], _MM_SHUFFLE(2, 1, 2, 1)); // b1 b2 b3 b4 b5 b6 b7 b8 b9 bA bB bC bD bE bF c0 } - SetNextPaState_simd16(pa, PaTriFan1_simd16, PaTriFanSingle0, 0, KNOB_SIMD16_WIDTH); + SetNextPaState_simd16(pa, PaTriFan1_simd16, PaTriFan1, PaTriFanSingle0, 0, KNOB_SIMD16_WIDTH); return true; } @@ -1457,7 +1459,7 @@ bool PaQuadList1(PA_STATE_OPT& pa, uint32_t slot, simdvector verts[]) #if ENABLE_AVX512_SIMD16 bool PaQuadList0_simd16(PA_STATE_OPT& pa, uint32_t slot, simd16vector verts[]) { - SetNextPaState_simd16(pa, PaQuadList1_simd16, PaQuadListSingle0); + SetNextPaState_simd16(pa, PaQuadList1_simd16, PaQuadList1, PaQuadListSingle0); return false; // Not enough vertices to assemble 16 triangles. } @@ -1485,7 +1487,7 @@ bool PaQuadList1_simd16(PA_STATE_OPT& pa, uint32_t slot, simd16vector verts[]) v2[i] = _simd16_shuffle_ps(temp0, temp1, _MM_SHUFFLE(3, 2, 3, 2)); // a2 a3 a6 a7 aA aB aE aF b2 b3 b6 b7 bA bB bE bF } - SetNextPaState_simd16(pa, PaQuadList0_simd16, PaQuadListSingle0, 0, KNOB_SIMD16_WIDTH, true); + SetNextPaState_simd16(pa, PaQuadList0_simd16, PaQuadList0, PaQuadListSingle0, 0, KNOB_SIMD16_WIDTH, true); return true; } @@ -1712,7 +1714,7 @@ bool PaLineLoop1(PA_STATE_OPT& pa, uint32_t slot, simdvector verts[]) #if ENABLE_AVX512_SIMD16 bool PaLineLoop0_simd16(PA_STATE_OPT& pa, uint32_t slot, simd16vector verts[]) { - SetNextPaState_simd16(pa, PaLineLoop1_simd16, PaLineLoopSingle0); + SetNextPaState_simd16(pa, PaLineLoop1_simd16, PaLineLoop1, PaLineLoopSingle0); return false; } @@ -1735,7 +1737,7 @@ bool PaLineLoop1_simd16(PA_STATE_OPT& pa, uint32_t slot, simd16vector verts[]) } } - SetNextPaState_simd16(pa, PaLineLoop1_simd16, PaLineLoopSingle0, 0, KNOB_SIMD16_WIDTH); + SetNextPaState_simd16(pa, PaLineLoop1_simd16, PaLineLoop1, PaLineLoopSingle0, 0, KNOB_SIMD16_WIDTH); return true; } @@ -1824,7 +1826,7 @@ bool PaLineList1(PA_STATE_OPT& pa, uint32_t slot, simdvector verts[]) #if ENABLE_AVX512_SIMD16 bool PaLineList0_simd16(PA_STATE_OPT& pa, uint32_t slot, simd16vector verts[]) { - SetNextPaState_simd16(pa, PaLineList1_simd16, PaLineListSingle0); + SetNextPaState_simd16(pa, PaLineList1_simd16, PaLineList1, PaLineListSingle0); return false; // Not enough vertices to assemble 16 lines } @@ -1849,7 +1851,7 @@ bool PaLineList1_simd16(PA_STATE_OPT& pa, uint32_t slot, simd16vector verts[]) v1[i] = _simd16_shuffle_ps(temp0, temp1, _MM_SHUFFLE(3, 1, 3, 1)); // a1 a3 a5 a7 a9 aB aD aF b1 b3 b5 b7 b9 bB bD bF } - SetNextPaState_simd16(pa, PaLineList0_simd16, PaLineListSingle0, 0, KNOB_SIMD16_WIDTH, true); + SetNextPaState_simd16(pa, PaLineList0_simd16, PaLineList0, PaLineListSingle0, 0, KNOB_SIMD16_WIDTH, true); return true; } @@ -2042,7 +2044,7 @@ bool PaLineStrip1(PA_STATE_OPT& pa, uint32_t slot, simdvector verts[]) #if ENABLE_AVX512_SIMD16 bool PaLineStrip0_simd16(PA_STATE_OPT& pa, uint32_t slot, simd16vector verts[]) { - SetNextPaState_simd16(pa, PaLineStrip1_simd16, PaLineStripSingle0); + SetNextPaState_simd16(pa, PaLineStrip1_simd16, PaLineStrip1, PaLineStripSingle0); return false; // Not enough vertices to assemble 16 lines } @@ -2069,7 +2071,7 @@ bool PaLineStrip1_simd16(PA_STATE_OPT& pa, uint32_t slot, simd16vector verts[]) v1[i] = _simd16_permute_ps(temp, perm); // a1 a2 a3 a4 a5 a6 a7 a8 a9 aA aB aC aD aE aF b0 } - SetNextPaState_simd16(pa, PaLineStrip1_simd16, PaLineStripSingle0, 0, KNOB_SIMD16_WIDTH); + SetNextPaState_simd16(pa, PaLineStrip1_simd16, PaLineStrip1, PaLineStripSingle0, 0, KNOB_SIMD16_WIDTH); return true; } @@ -2234,7 +2236,7 @@ bool PaPoints0_simd16(PA_STATE_OPT& pa, uint32_t slot, simd16vector verts[]) verts[0] = a; // points only have 1 vertex. - SetNextPaState_simd16(pa, PaPoints0_simd16, PaPointsSingle0, 0, KNOB_SIMD16_WIDTH, true); + SetNextPaState_simd16(pa, PaPoints0_simd16, PaPoints0, PaPointsSingle0, 0, KNOB_SIMD16_WIDTH, true); return true; } @@ -2390,7 +2392,7 @@ bool PaRectList2( /// There is not enough to assemble 8 triangles. bool PaRectList0_simd16(PA_STATE_OPT& pa, uint32_t slot, simd16vector verts[]) { - SetNextPaState_simd16(pa, PaRectList1_simd16, PaRectListSingle0); + SetNextPaState_simd16(pa, PaRectList1_simd16, PaRectList1, PaRectListSingle0); return false; } @@ -2494,7 +2496,7 @@ bool PaRectList1_simd16( v2[i] = _simd16_insert_ps(_simd16_setzero_ps(), v2_lo, 0); } - SetNextPaState_simd16(pa, PaRectList1_simd16, PaRectListSingle0, 0, KNOB_SIMD16_WIDTH, true); + SetNextPaState_simd16(pa, PaRectList1_simd16, PaRectList1, PaRectListSingle0, 0, KNOB_SIMD16_WIDTH, true); return true; } @@ -2510,7 +2512,7 @@ bool PaRectList2_simd16( simd16vector verts[]) { SWR_INVALID("Is rect list used for anything other then clears?"); - SetNextPaState_simd16(pa, PaRectList0_simd16, PaRectListSingle0, 0, KNOB_SIMD16_WIDTH, true); + SetNextPaState_simd16(pa, PaRectList0_simd16, PaRectList0, PaRectListSingle0, 0, KNOB_SIMD16_WIDTH, true); return true; } -- 2.30.2