| 44 | |
| 45 | template<typename Tk, typename Tv> |
| 46 | void sortByKeyBatched(Param<Tk> pKey, Param<Tv> pVal, const int dim, |
| 47 | bool isAscending) { |
| 48 | af::dim4 inDims; |
| 49 | for (int i = 0; i < 4; i++) inDims[i] = pKey.dims[i]; |
| 50 | |
| 51 | const dim_t elements = inDims.elements(); |
| 52 | |
| 53 | // Sort dimension |
| 54 | // tileDims * seqDims = inDims |
| 55 | af::dim4 tileDims(1); |
| 56 | af::dim4 seqDims = inDims; |
| 57 | tileDims[dim] = inDims[dim]; |
| 58 | seqDims[dim] = 1; |
| 59 | |
| 60 | // Create/call iota |
| 61 | Array<uint> Seq = iota<uint>(seqDims, tileDims); |
| 62 | |
| 63 | Tk *Key = pKey.ptr; |
| 64 | auto cKey = memAlloc<Tk>(elements); |
| 65 | CUDA_CHECK(cudaMemcpyAsync(cKey.get(), Key, elements * sizeof(Tk), |
| 66 | cudaMemcpyDeviceToDevice, getActiveStream())); |
| 67 | |
| 68 | Tv *Val = pVal.ptr; |
| 69 | thrustSortByKey(Key, Val, elements, isAscending); |
| 70 | thrustSortByKey(cKey.get(), Seq.get(), elements, isAscending); |
| 71 | |
| 72 | auto cSeq = memAlloc<uint>(elements); |
| 73 | CUDA_CHECK(cudaMemcpyAsync(cSeq.get(), Seq.get(), elements * sizeof(uint), |
| 74 | cudaMemcpyDeviceToDevice, getActiveStream())); |
| 75 | |
| 76 | // This always needs to be ascending |
| 77 | thrustSortByKey(Seq.get(), Val, elements, true); |
| 78 | thrustSortByKey(cSeq.get(), Key, elements, true); |
| 79 | |
| 80 | // No need of doing moddims here because the original Array<T> |
| 81 | // dimensions have not been changed |
| 82 | // val.modDims(inDims); |
| 83 | } |
| 84 | |
| 85 | template<typename Tk, typename Tv> |
| 86 | void sort0ByKey(Param<Tk> okey, Param<Tv> oval, bool isAscending) { |
nothing calls this directly
no test coverage detected