| 28 | |
| 29 | private: |
| 30 | void Compute(KernelComputeContext* ctx) const override { |
| 31 | const Tensor* input_tensor = ctx->Tensor4ArgNameAndIndex("input", 0); |
| 32 | const Tensor* index_tensor = ctx->Tensor4ArgNameAndIndex("index", 0); |
| 33 | Tensor* out_tensor = ctx->Tensor4ArgNameAndIndex("output", 0); |
| 34 | const Tensor* src_tensor = ctx->Tensor4ArgNameAndIndex("src", 0); |
| 35 | const int32_t dim = ctx->Attr<int32_t>("dim"); |
| 36 | |
| 37 | const IDX_T* index = index_tensor->dptr<IDX_T>(); |
| 38 | IN_T* output = out_tensor->mut_dptr<IN_T>(); |
| 39 | size_t out_bytes_size = |
| 40 | out_tensor->shape_view().elem_cnt() * GetSizeOfDataType(out_tensor->data_type()); |
| 41 | |
| 42 | Tensor* like_tensor = ctx->Tensor4ArgNameAndIndex("like", 0); |
| 43 | const IN_T* src = src_tensor->dptr<IN_T>(); |
| 44 | |
| 45 | if (input_tensor) { |
| 46 | Memcpy<device_type>(ctx->stream(), output, input_tensor->dptr<IN_T>(), out_bytes_size); |
| 47 | } else if (like_tensor) { |
| 48 | Memset<device_type>(ctx->stream(), output, 0, out_bytes_size); |
| 49 | } else { |
| 50 | UNIMPLEMENTED() << "Input tensor and like tensor cannot be empty simultaneously."; |
| 51 | } |
| 52 | |
| 53 | const Shape src_shape = ExpandDimIf0D(src_tensor->shape_view()); |
| 54 | const Shape index_shape = ExpandDimIf0D(index_tensor->shape_view()); |
| 55 | const int ndim = src_shape.NumAxes(); |
| 56 | DimOpIndexNdHelper<IDX_T> src_nd_helper(src_shape.data(), ndim); |
| 57 | DimOpIndexNdHelper<IDX_T> idx_nd_helper(index_shape.data(), ndim); |
| 58 | DimOpIndexNdHelper<IDX_T> output_nd_helper(out_tensor->shape_view().data(), ndim); |
| 59 | |
| 60 | const int64_t upper_bound = [&]() { |
| 61 | if (input_tensor) { |
| 62 | const Shape input_shape = ExpandDimIf0D(input_tensor->shape_view()); |
| 63 | return input_shape.At(dim); |
| 64 | } else { |
| 65 | const Shape like_shape = ExpandDimIf0D(like_tensor->shape_view()); |
| 66 | return like_shape.At(dim); |
| 67 | } |
| 68 | }(); |
| 69 | |
| 70 | DimScatterFunctor<device_type, IN_T, IDX_T, Opt>()( |
| 71 | ctx->stream(), src_nd_helper, idx_nd_helper, output_nd_helper, ndim, index_shape.elem_cnt(), |
| 72 | dim, upper_bound, index, src, output); |
| 73 | } |
| 74 | bool AlwaysComputeWhenAllOutputsEmpty() const override { return false; } |
| 75 | }; |
| 76 |
nothing calls this directly
no test coverage detected