wip(standalone): GPU step 28 -- skirted near tiles on GPU (correct, but SLOWER); 16 shape workers

Luna: near skirts were CPU-only by implementation gap, not correctness. Ported with near-resolution refinement and overflow fallback; near-probe preparation split out so the skirt kernels no longer inline the near field (first attempt: two fxc /O3 compiles ran 10.5 h and were stopped). Shape compile proven identical at W1/W8/W16. -SkipCpuDetailBand retired (154/640 skirt audit failures, +174 ms planning). Audits: hashes W1/W8/W16, walk 5,642/5,642, whole-near 960 tiles + 640 skirt faces 0 sign/topology diffs (one 0.001 cm vertex-key delta), far/thin 0 diffs, both survivals pass with 2,074 candidates at 7.00 / 7.02 GiB (was ~6.2), ten captures pass. My runs outside the sandbox (Step28.exe). NOTE the machine was ~25% slower than last night for every build (step 27 CPU full machine 0.97-1.02 s at night, 1.22-1.29 s now): 0xFF00: CPU 1.337 / 1.291 / 1.295 s; GPU miss 1.337 s; GPU cached 1.663 / 1.424 / 1.617 s (GPU tiles 1,427 / 2,072 / 1,222) 0xFFFF: CPU 1.221 / 1.231 / 1.236 s (step 27 same session 1.285 / 1.222 s) GPU cached 1.582 / 1.550 / 1.416, 1.635 / 1.441 / 1.470 / 1.445 s (GPU tiles vary 593-2,074; timestamps up to ~470 ms) => GPU route now slower than the CPU route measured alongside it. Step 27 GPU control was not clean (both exes load the same precompiled .cso files). Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_01MmSoF64BbRFmk9C3p1VsBJ
2026-10-08 17:07:40 +02:00
1 parent 7d1808fdf2
commit 5aa285c3ef
@@ -66,7 +66,6 @@ public:
        uint64 LocalBandSkips=0;
        uint64 GpuVisualCandidates=0,CpuVisualCandidates=0,GpuOnlyCandidates=0;
        uint64 GpuVisualDetailEvaluations=0,GpuOnlyDetailEvaluations=0;
-        uint64 GpuOnlySkippedCandidateMasks=0;
        uint64 DistanceSamples=0,DistanceSampleNanoseconds=0;
        uint64 DetailSamples=0,DetailSampleNanoseconds=0;
        uint64 ReliefSampleNanoseconds=0,LayerLookupSampleNanoseconds=0;
@@ -282,18 +281,14 @@ public:
               bool bInRoundTunnels=false,bool bInProfileDetailBand=false,
               bool bValidatePassageRoutes=true,
               const std::vector<VoxelForgeCore::DetailBandTileRegion>& InGpuVisualDetailTiles={},
-               const std::vector<VoxelForgeCore::DetailBandTileRegion>& InCpuVisualDetailTiles={},
-               bool bInSkipCpuDetailBandForGpuVisuals=false)
+               const std::vector<VoxelForgeCore::DetailBandTileRegion>& InCpuVisualDetailTiles={})
    {
        using Clock=std::chrono::steady_clock;
        const auto TotalStart=Clock::now();
-        Owner=&World;WorkerCount=std::clamp(Workers,1,8);DetailVariant=Variant;
+        Owner=&World;WorkerCount=std::clamp(Workers,1,16);DetailVariant=Variant;
        Scheduler=InScheduler;RoundTunnelsEnabled=bInRoundTunnels;
        bAnalyticOnly=false;
        bProfileDetailBand=bInProfileDetailBand;
-        bSkipCpuDetailBandForGpuVisuals=bInSkipCpuDetailBandForGpuVisuals;
-        if(bSkipCpuDetailBandForGpuVisuals&&InGpuVisualDetailTiles.empty())
-        {Error="GPU visual detail-band skipping requires at least one GPU near tile region";return false;}
        for(int32 A=0;A<3;++A)
        {
            Min[A]=(&MinVoxels.X)[A]-2;
@@ -333,9 +328,6 @@ public:
            Stats.CandidateMergeSeconds);
        if(!BuildGpuVisualDetailNodeSets(InGpuVisualDetailTiles,InCpuVisualDetailTiles,Error))
            return false;
-        GpuVisualProxyValidWords.clear();
-        if(bSkipCpuDetailBandForGpuVisuals)
-            GpuVisualProxyValidWords.assign(CandidateWords.size(),0u);
        Stats.IntervalSeconds=std::chrono::duration<double>(Clock::now()-Stage).count();

        Stage=Clock::now();
@@ -777,14 +769,6 @@ public:
                const uint64 Mask=uint64{1}<<(Zi&63u);
                if((CandidateWords[Word]&Mask)==0)
                    Out.nearPackedProxy[SampleIndex]=0u;
-                else if(IsSkippedGpuVisualDetailNode(WX,WY,WZ)
-                    &&(GpuVisualProxyValidWords[Word]&Mask)==0)
-                {
-                    // UINT_MAX asks ApplyNearFieldDetailCS to use its own exact
-                    // GPU roughness result instead of a CPU packed proxy.
-                    Out.nearPackedProxy[SampleIndex]=std::numeric_limits<uint32>::max();
-                    ++Out.nearPackedSkippedSampleCount;
-                }
                else
                {
                    ++Out.nearPackedCandidateSampleCount;
@@ -933,8 +917,6 @@ public:
    {
        int32 IX=0,IY=0,IZ=0;
        const bool bInteger=NearInteger(X,IX)&&NearInteger(Y,IY)&&NearInteger(Z,IZ);
-        if(bInteger&&IsSkippedGpuVisualDetailNode(IX,IY,IZ))
-            return GpuSkippedDetailDensityAt(IX,IY,IZ,X,Y,Z);
        const bool bPacked=!bAnalyticOnly&&bInteger&&Contains(IX,IY,IZ);
        if(bPacked)
            return PackedDensityAt(IX,IY,IZ,IsRock(IX,IY,IZ),X,Y,Z);
@@ -943,9 +925,6 @@ public:

    float DensityAtLattice(int32 X,int32 Y,int32 Z) const
    {
-        if(IsSkippedGpuVisualDetailNode(X,Y,Z))
-            return GpuSkippedDetailDensityAt(X,Y,Z,static_cast<float>(X),
-                static_cast<float>(Y),static_cast<float>(Z));
        const bool bPacked=!bAnalyticOnly&&Contains(X,Y,Z);
        if(bPacked)
            return PackedDensityAt(X,Y,Z,IsRock(X,Y,Z),static_cast<float>(X),
@@ -959,39 +938,6 @@ public:
        return AnalyticDensityAt(X,Y,Z,false);
    }

-    float GpuSkippedDetailDensityAt(int32 X,int32 Y,int32 Z,
-                                    float FX,float FY,float FZ) const
-    {
-        const size_t Column=static_cast<size_t>(Y-Min[1])*Size[0]
-            +static_cast<size_t>(X-Min[0]);
-        const size_t Zi=static_cast<size_t>(Z-Min[2]);
-        const size_t Word=Column*WordsPerZ+(Zi>>6);
-        const uint64 Mask=uint64{1}<<(Zi&63u);
-        if(!GpuVisualProxyValidWords.empty()&&(GpuVisualProxyValidWords[Word]&Mask))
-            return PackedDensityAt(X,Y,Z,IsRock(X,Y,Z),FX,FY,FZ);
-
-        if(CandidateWords[Word]&Mask)
-        {
-            // The GPU-only compile omitted this packed cache entry. Reconstruct the
-            // CPU reference's un-clamped detail density and the same signed Q8.8
-            // proxy when an exceptional CPU mesh or field query needs the sample.
-            const float ExactDensity=AnalyticDensityAt(FX,FY,FZ,false,
-                nullptr,nullptr,nullptr,false);
-            int32 Q=static_cast<int32>(std::lround(ExactDensity*256.0f));
-            Q=std::clamp(Q,-32767,32767);
-            if(Q==0)Q=ExactDensity<0.0f?-1:1;
-            float Density=static_cast<float>(Q)*(1.0f/256.0f);
-            const bool bRock=ExactDensity<=0.0f;
-            if((Density<0.0f)!=bRock)
-                Density=(bRock?-1.0f:1.0f)*std::max(std::abs(Density),1.0f/256.0f);
-            return ApplyEditBrushes(Density,FX,FY,FZ);
-        }
-
-        const bool bRock=((RockWords[Word]&Mask)!=0);
-        constexpr float ProxyBandEdge=DetailSupremum+0.125f;
-        return ApplyEditBrushes(bRock?-ProxyBandEdge:ProxyBandEdge,FX,FY,FZ);
-    }
-
    VoxelForgeCore::DetailFieldSample DetailSampleAt(float X,float Y,float Z) const
    {
        VoxelForgeCore::DetailFieldSample Sample;
@@ -1018,18 +964,6 @@ public:
                OutDensity[Lane]=DensityAtLattice(XStart+Lane*XStep,Y,Z);
            return;
        }
-        if(bSkipCpuDetailBandForGpuVisuals)
-        {
-            bool bContainsSkippedNode=false;
-            for(int32 Lane=0;Lane<Count;++Lane)
-                bContainsSkippedNode|=IsSkippedGpuVisualDetailNode(XStart+Lane*XStep,Y,Z);
-            if(bContainsSkippedNode)
-            {
-                for(int32 Lane=0;Lane<Count;++Lane)
-                    OutDensity[Lane]=DensityAtLattice(XStart+Lane*XStep,Y,Z);
-                return;
-            }
-        }
        bool bFast=EditBrushes.empty()&&Y>=Min[1]&&Y<=Max[1]
            &&Z>=Min[2]&&Z<=Max[2]
            &&XStart>=Min[0]
@@ -1185,9 +1119,6 @@ public:

    bool IsRock(int32 X,int32 Y,int32 Z) const
    {
-        if(IsSkippedGpuVisualDetailNode(X,Y,Z))
-            return AnalyticDensityAt(static_cast<float>(X),static_cast<float>(Y),
-                static_cast<float>(Z))<=0.0f;
        if(!Contains(X,Y,Z))return EvaluateFast(X,Y,Z);
        const size_t Line=static_cast<size_t>(Y-Min[1])*Size[0]
            +static_cast<size_t>(X-Min[0]);
@@ -3342,17 +3273,14 @@ private:
        if(GpuTiles.empty()&&CpuTiles.empty())
        {
            GpuVisualDetailWords.clear();CpuVisualDetailWords.clear();
-            CpuVisualDetailGuardWords.clear();
            if(bProfileDetailBand)
            {
                GpuVisualDetailWords.assign(CandidateWords.size(),0u);
                CpuVisualDetailWords.assign(CandidateWords.size(),0u);
-                CpuVisualDetailGuardWords.assign(CandidateWords.size(),0u);
            }
            GpuVisualDetailTileCount=CpuVisualDetailTileCount=0;
            GpuVisualDetailGridSamples=CpuVisualDetailGridSamples=0;
            GpuVisualDetailNodeCount=CpuVisualDetailNodeCount=0;
-            CpuVisualDetailGuardNodeCount=0;
            return true;
        }
        const auto BuildSet=[&](const std::vector<VoxelForgeCore::DetailBandTileRegion>& Tiles,
@@ -3403,55 +3331,16 @@ private:
        GpuVisualDetailNodeCount=0;CpuVisualDetailNodeCount=0;
        for(const uint64 Bits:GpuVisualDetailWords)GpuVisualDetailNodeCount+=std::popcount(Bits);
        for(const uint64 Bits:CpuVisualDetailWords)CpuVisualDetailNodeCount+=std::popcount(Bits);
-        CpuVisualDetailGuardWords.assign(CandidateWords.size(),0u);
-        if(bSkipCpuDetailBandForGpuVisuals&&!CpuTiles.empty())
-        {
-            constexpr int32 AxisOffsets[6][3]={{-1,0,0},{1,0,0},{0,-1,0},
-                {0,1,0},{0,0,-1},{0,0,1}};
-            for(size_t Word=0;Word<CpuVisualDetailWords.size();++Word)
-            {
-                uint64 Bits=CpuVisualDetailWords[Word];
-                const size_t Column=Word/WordsPerZ;
-                const int32 X=Min[0]+static_cast<int32>(Column%Size[0]);
-                const int32 Y=Min[1]+static_cast<int32>(Column/Size[0]);
-                const size_t ZiBase=(Word%WordsPerZ)*64u;
-                while(Bits)
-                {
-                    const uint32 Bit=std::countr_zero(Bits);Bits&=Bits-1;
-                    const size_t Zi=ZiBase+Bit;
-                    if(Zi>=static_cast<size_t>(Size[2]))continue;
-                    const int32 Z=Min[2]+static_cast<int32>(Zi);
-                    for(const auto& Offset:AxisOffsets)
-                    {
-                        const int32 GX=X+Offset[0],GY=Y+Offset[1],GZ=Z+Offset[2];
-                        if(GX<Min[0]||GX>Max[0]||GY<Min[1]||GY>Max[1]
-                            ||GZ<Min[2]||GZ>Max[2])continue;
-                        const size_t GuardColumn=static_cast<size_t>(GY-Min[1])*Size[0]
-                            +static_cast<size_t>(GX-Min[0]);
-                        const size_t GuardZi=static_cast<size_t>(GZ-Min[2]);
-                        const size_t GuardWord=GuardColumn*WordsPerZ+(GuardZi>>6);
-                        const uint64 GuardMask=uint64{1}<<(GuardZi&63u);
-                        if(CandidateWords[GuardWord]&GuardMask)
-                            CpuVisualDetailGuardWords[GuardWord]|=GuardMask;
-                    }
-                }
-            }
-        }
-        CpuVisualDetailGuardNodeCount=0;
-        for(const uint64 Bits:CpuVisualDetailGuardWords)
-            CpuVisualDetailGuardNodeCount+=std::popcount(Bits);
        GpuVisualDetailTileCount=GpuTiles.size();CpuVisualDetailTileCount=CpuTiles.size();
        if(!GpuTiles.empty()||!CpuTiles.empty())
        {
-            std::printf("[GpuVisualDetailBandPlan] gpu_tiles=%llu gpu_sample_nodes=%llu gpu_unique_candidate_nodes=%llu cpu_near_tiles=%llu cpu_sample_nodes=%llu cpu_unique_candidate_nodes=%llu cpu_boundary_guard_candidate_nodes=%llu skip_enabled=%d\n",
+            std::printf("[GpuVisualDetailBandPlan] gpu_tiles=%llu gpu_sample_nodes=%llu gpu_unique_candidate_nodes=%llu cpu_near_tiles=%llu cpu_sample_nodes=%llu cpu_unique_candidate_nodes=%llu\n",
                static_cast<unsigned long long>(GpuVisualDetailTileCount),
                static_cast<unsigned long long>(GpuVisualDetailGridSamples),
                static_cast<unsigned long long>(GpuVisualDetailNodeCount),
                static_cast<unsigned long long>(CpuVisualDetailTileCount),
                static_cast<unsigned long long>(CpuVisualDetailGridSamples),
-                static_cast<unsigned long long>(CpuVisualDetailNodeCount),
-                static_cast<unsigned long long>(CpuVisualDetailGuardNodeCount),
-                bSkipCpuDetailBandForGpuVisuals?1:0);
+                static_cast<unsigned long long>(CpuVisualDetailNodeCount));
        }
        return true;
    }
@@ -3476,25 +3365,6 @@ private:
            &(uint64{1}<<(Zi&63u)))!=0;
    }

-    bool IsSkippedGpuVisualDetailWord(size_t Word,uint64 Mask) const
-    {
-        return bSkipCpuDetailBandForGpuVisuals
-            &&(GpuVisualDetailWords[Word]&Mask)!=0
-            &&(CpuVisualDetailWords[Word]&Mask)==0
-            &&(CpuVisualDetailGuardWords[Word]&Mask)==0;
-    }
-
-    bool IsSkippedGpuVisualDetailNode(int32 X,int32 Y,int32 Z) const
-    {
-        if(!bSkipCpuDetailBandForGpuVisuals||!Contains(X,Y,Z)
-            ||GpuVisualDetailWords.empty())return false;
-        const size_t Column=static_cast<size_t>(Y-Min[1])*Size[0]
-            +static_cast<size_t>(X-Min[0]);
-        const size_t Zi=static_cast<size_t>(Z-Min[2]);
-        return IsSkippedGpuVisualDetailWord(Column*WordsPerZ+(Zi>>6),
-            uint64{1}<<(Zi&63u));
-    }
-
    void ApplyDetailBand(FParallelTiming& Timing)
    {
        using Clock=std::chrono::steady_clock;
@@ -3566,25 +3436,16 @@ private:
                        if(std::abs(SDF)>DetailSupremum+0.125f)continue;
                        const size_t VisualWordIndex=WordBase+(Zi>>6);
                        const uint64 VisualMask=uint64{1}<<(Zi&63u);
-                        const bool bGpuVisualNode=(bProfile||bSkipCpuDetailBandForGpuVisuals)
+                        const bool bGpuVisualNode=bProfile
                            &&(GpuVisualDetailWords[VisualWordIndex]&VisualMask)!=0;
-                        const bool bCpuVisualNode=(bProfile||bSkipCpuDetailBandForGpuVisuals)
+                        const bool bCpuVisualNode=bProfile
                            &&(CpuVisualDetailWords[VisualWordIndex]&VisualMask)!=0;
-                        const bool bSkipDetail=bSkipCpuDetailBandForGpuVisuals&&bGpuVisualNode
-                            &&!bCpuVisualNode
-                            &&(CpuVisualDetailGuardWords[VisualWordIndex]&VisualMask)==0;
                        if(bProfile)
                        {
                            if(bGpuVisualNode)++Profile.GpuVisualCandidates;
                            if(bCpuVisualNode)++Profile.CpuVisualCandidates;
                            if(bGpuVisualNode&&!bCpuVisualNode)++Profile.GpuOnlyCandidates;
                        }
-                        if(bSkipDetail)
-                        {
-                            ValidBits|=uint64{1}<<(Zi&63u);
-                            if(bProfile)++Profile.GpuOnlySkippedCandidateMasks;
-                            continue;
-                        }
                        const float SurfaceSDF=Sample.SurfaceSDF;
                        const float CaveInfluence=Sample.CaveInfluence;
                        const float FloorInfluence=Sample.FloorInfluence;
@@ -3681,7 +3542,7 @@ private:
            uint64 LayerLookupSampleNs=0,OwnerSampleNs=0,OwnerFieldCalls=0,OwnerActiveCalls=0;
            uint64 OwnerCacheHits=0,OwnerCacheMisses=0;
            uint64 GpuVisualCandidates=0,CpuVisualCandidates=0,GpuOnlyCandidates=0;
-            uint64 GpuVisualDetailCalls=0,GpuOnlyDetailCalls=0,GpuOnlySkippedMasks=0;
+            uint64 GpuVisualDetailCalls=0,GpuOnlyDetailCalls=0;
            for(int32 Worker=0;Worker<WorkerCount;++Worker)
            {
                const FDetailWorkerProfile& P=WorkerProfile[static_cast<size_t>(Worker)];
@@ -3701,7 +3562,6 @@ private:
                GpuOnlyCandidates+=P.GpuOnlyCandidates;
                GpuVisualDetailCalls+=P.GpuVisualDetailEvaluations;
                GpuOnlyDetailCalls+=P.GpuOnlyDetailEvaluations;
-                GpuOnlySkippedMasks+=P.GpuOnlySkippedCandidateMasks;
            }
            const double DistanceAverageNs=DistanceSamples?static_cast<double>(DistanceSampleNs)/DistanceSamples:0.0;
            const double DetailAverageNs=DetailSamples?static_cast<double>(DetailSampleNs)/DetailSamples:0.0;
@@ -3726,14 +3586,12 @@ private:
                static_cast<unsigned long long>(OwnerCacheHits),
                static_cast<unsigned long long>(OwnerCacheMisses),
                static_cast<unsigned long long>(LocalBandSkips));
-            std::printf("[GpuDetailConsumerProfile] gpu_near_proxy_nodes=%llu cpu_near_mesh_nodes=%llu gpu_only_nodes=%llu gpu_near_detail_noise_evaluations=%llu gpu_only_detail_noise_evaluations=%llu gpu_only_skipped_candidate_masks=%llu skip_enabled=%d\n",
+            std::printf("[GpuDetailConsumerProfile] gpu_near_proxy_nodes=%llu cpu_near_mesh_nodes=%llu gpu_only_nodes=%llu gpu_near_detail_noise_evaluations=%llu gpu_only_detail_noise_evaluations=%llu\n",
                static_cast<unsigned long long>(GpuVisualCandidates),
                static_cast<unsigned long long>(CpuVisualCandidates),
                static_cast<unsigned long long>(GpuOnlyCandidates),
                static_cast<unsigned long long>(GpuVisualDetailCalls),
-                static_cast<unsigned long long>(GpuOnlyDetailCalls),
-                static_cast<unsigned long long>(GpuOnlySkippedMasks),
-                bSkipCpuDetailBandForGpuVisuals?1:0);
+                static_cast<unsigned long long>(GpuOnlyDetailCalls));
            for(int32 Worker=0;Worker<WorkerCount;++Worker)
            {
                const FDetailWorkerProfile& P=WorkerProfile[static_cast<size_t>(Worker)];
@@ -3769,8 +3627,6 @@ private:
                    +static_cast<size_t>(X-Min[0]);
                StoreSurfaceDistance(Node,Density);
                CandidateWords[Word]|=Mask;
-                if(IsSkippedGpuVisualDetailNode(X,Y,Z))
-                    GpuVisualProxyValidWords[Word]|=Mask;
                ++AddedThisPass;

                const bool bExactRock=Density<0.0f;
@@ -4821,14 +4677,10 @@ private:
    std::unique_ptr<int16[]> SurfaceDistanceQ;
    std::array<std::vector<uint64>,3> CandidateAxisWords;
    std::vector<uint64> CandidateWords;
-    bool bSkipCpuDetailBandForGpuVisuals=false;
    std::vector<uint64> GpuVisualDetailWords,CpuVisualDetailWords;
-    std::vector<uint64> CpuVisualDetailGuardWords;
    uint64 GpuVisualDetailTileCount=0,CpuVisualDetailTileCount=0;
    uint64 GpuVisualDetailGridSamples=0,CpuVisualDetailGridSamples=0;
    uint64 GpuVisualDetailNodeCount=0,CpuVisualDetailNodeCount=0;
-    uint64 CpuVisualDetailGuardNodeCount=0;
-    std::vector<uint64> GpuVisualProxyValidWords;
    FBuildStats Stats;
    bool bAnalyticOnly=false;
 };
@@ -307,7 +307,7 @@ namespace VoxelForgeCore
            static_cast<int32>(std::clamp<std::uint32_t>(request.workerLimit,1,8)),ToNative(request.detail),
            request.includeFarWorld,error,&scheduler,request.roundTunnels,request.profileDetailBand,
            request.validatePassageRoutes,request.gpuVisualDetailTiles,
-            request.cpuVisualDetailTiles,request.skipCpuDetailBandForGpuVisuals);
+            request.cpuVisualDetailTiles);
        if(ok)
        {
            impl_->shape=std::move(shape);