Method MakeStripeGrouping

catboost/cuda/gpu_data/samples_grouping_gpu.cpp:102–151

    }

    TGpuSamplesGrouping<NCudaLib::TStripeMapping> TGpuSamplesGroupingHelper<NCudaLib::TMirrorMapping>::MakeStripeGrouping(const TGpuSamplesGrouping<NCudaLib::TMirrorMapping>& mirrorMapping,
                                                                                                                          const TCudaBuffer<const ui32, NCudaLib::TStripeMapping>& indices) {
        CB_ENSURE(indices.GetObjectsSlice() == mirrorMapping.CurrentDocsSlice);
        TSlice docsSlice = indices.GetObjectsSlice();
        docsSlice.Left += mirrorMapping.CurrentDocsSlice.Left;
        docsSlice.Right += mirrorMapping.CurrentDocsSlice.Left;
        const IQueriesGrouping& grouping = *mirrorMapping.Grouping;
        const ui32 queryIdOffset = grouping.GetQueryId(docsSlice.Left);

        const NCudaLib::TDistributedObject<ui32>& baseBias = mirrorMapping.GetOffsetsBias();
        NCudaLib::TDistributedObject<ui32> deviceOffsetsBiases = NCudaLib::GetCudaManager().CreateDistributedObject<ui32>(0);

        TVector<TSlice> groupMetaSlices(deviceOffsetsBiases.DeviceCount());

        ui32 offset = 0;
        for (ui32 dev = 0; dev < deviceOffsetsBiases.DeviceCount(); ++dev) {
            TSlice deviceSlice = indices.GetMapping().DeviceSlice(dev);
            const ui32 localBias = baseBias.At(dev) + offset;
            deviceOffsetsBiases.Set(dev, localBias);
            ui32 firstGroupId = grouping.GetQueryId(docsSlice.Left + offset) - queryIdOffset;
            ui32 lastGroupId = grouping.GetQueryId(docsSlice.Left + offset + deviceSlice.Size()) - queryIdOffset;
            offset += deviceSlice.Size();
            groupMetaSlices[dev] = TSlice(firstGroupId, lastGroupId);
        }

        NCudaLib::TStripeMapping groupMetaMapping(std::move(groupMetaSlices));
        CB_ENSURE(groupMetaMapping.GetObjectsSlice() == TSlice(0, grouping.GetQueryId(docsSlice.Right) - queryIdOffset));
        NCudaLib::TCudaBuffer<const ui32, NCudaLib::TStripeMapping> offsets = NCudaLib::StripeView(mirrorMapping.Offsets, groupMetaMapping);
        NCudaLib::TCudaBuffer<const ui32, NCudaLib::TStripeMapping> sizes = NCudaLib::StripeView(mirrorMapping.Sizes, groupMetaMapping);

        TGpuSamplesGrouping<NCudaLib::TStripeMapping> samplesGrouping(&grouping,
                                                                      mirrorMapping.CurrentDocsSlice,
                                                                      std::move(offsets),
                                                                      std::move(sizes),
                                                                      std::move(deviceOffsetsBiases));

        {
            const TQueriesGrouping* pointwiseSamplesGrouping = dynamic_cast<const TQueriesGrouping*>(&grouping);
            if (pointwiseSamplesGrouping && pointwiseSamplesGrouping->GetFlatQueryPairs().size()) {
                NCudaLib::TStripeMapping pairsMapping = groupMetaMapping.Transform([&](const TSlice& groupsSlice) -> ui64 {
                    ui32 firstGroupId = groupsSlice.Left + queryIdOffset;
                    ui32 lastGroupId = groupsSlice.Right + queryIdOffset;
                    return pointwiseSamplesGrouping->GetQueryPairOffset(lastGroupId) - pointwiseSamplesGrouping->GetQueryPairOffset(firstGroupId);
                });
                samplesGrouping.Pairs = NCudaLib::StripeView(mirrorMapping.Pairs, pairsMapping);
                samplesGrouping.PairsWeights = NCudaLib::StripeView(mirrorMapping.PairsWeights, pairsMapping);
            }
        }
        return samplesGrouping;
    }

    TGpuSamplesGrouping<NCudaLib::TStripeMapping>
    TGpuSamplesGroupingHelper<NCudaLib::TStripeMapping>::CreateGpuGrouping(const TDocParallelDataSet& dataSet) {