Method Run

catboost/cuda/targets/kernel.h:190–226

        }

        void Run(const TCudaStream& stream, TKernelContext& context) {
            if (FunctionValue.Size()) {
                NKernel::FillBuffer(FunctionValue.Get(), 0.0f, 1, stream.GetStream());
            }
            if (Predictions.Size() == 0) {
                return;
            }

            if (Der.Size()) {
                CB_ENSURE(Der.Size() == Predictions.Size());
            }
            CB_ENSURE(QuerySizes.Size() == QueryOffsets.Size());
            if (Indices.Size()) {
                CB_ENSURE(Indices.Size() == Predictions.Size());
                NKernel::Gather(context.MseDer.Get(), Predictions.Get(), Indices.Get(), Indices.Size(), stream.GetStream());
            } else {
                CopyMemoryAsync(Predictions.Get(), context.MseDer.Get(), Predictions.Size(), stream);
            }

            NKernel::MultiplyVector(context.MseDer.Get(), -1.0f, Predictions.Size(), stream.GetStream());
            NKernel::AddVector(context.MseDer.Get(), Relevs.Get(), Relevs.Size(), stream.GetStream());
            NKernel::ComputeGroupMeans(context.MseDer.Get(), Weights.Get(), QueryOffsets.Get(), QueryOffsetsBias,
                                       QuerySizes.Get(), QueryOffsets.Size(), context.QueryMeans,
                                       stream.GetStream());
            NKernel::ComputeGroupIds(QuerySizes.Get(), QueryOffsets.Get(), QueryOffsetsBias, QueryOffsets.Size(),
                                     context.Qids, stream.GetStream());
            NKernel::ApproximateQueryRmse(context.MseDer.Get(),
                                          Weights.Get(),
                                          context.Qids,
                                          static_cast<ui32>(Predictions.Size()),
                                          context.QueryMeans,
                                          Indices.Get(),
                                          FunctionValue.Get(),
                                          Der.Get(),
                                          Der2.Get(),
                                          stream.GetStream());
        }
    };

    class TQuerySoftMaxKernel: public TKernelBase<NKernel::TQuerySoftMaxContext, false> {