| 37 | |
| 38 | template<typename T> |
| 39 | Array<T> matmul(const common::SparseArray<T>& lhs, const Array<T>& rhsIn, |
| 40 | af_mat_prop optLhs, af_mat_prop optRhs) { |
| 41 | #if defined(WITH_LINEAR_ALGEBRA) |
| 42 | if (OpenCLCPUOffload( |
| 43 | false)) { // Do not force offload gemm on OSX Intel devices |
| 44 | return cpu::matmul(lhs, rhsIn, optLhs, optRhs); |
| 45 | } |
| 46 | #endif |
| 47 | |
| 48 | int lRowDim = (optLhs == AF_MAT_NONE) ? 0 : 1; |
| 49 | // int lColDim = (optLhs == AF_MAT_NONE) ? 1 : 0; |
| 50 | static const int rColDim = |
| 51 | 1; // Unsupported : (optRhs == AF_MAT_NONE) ? 1 : 0; |
| 52 | |
| 53 | dim4 lDims = lhs.dims(); |
| 54 | dim4 rDims = rhsIn.dims(); |
| 55 | int M = lDims[lRowDim]; |
| 56 | int N = rDims[rColDim]; |
| 57 | // int K = lDims[lColDim]; |
| 58 | |
| 59 | const Array<T> rhs = |
| 60 | (N != 1 && optLhs == AF_MAT_NONE) ? transpose(rhsIn, false) : rhsIn; |
| 61 | Array<T> out = createEmptyArray<T>(af::dim4(M, N, 1, 1)); |
| 62 | |
| 63 | static const T alpha = scalar<T>(1.0); |
| 64 | static const T beta = scalar<T>(0.0); |
| 65 | |
| 66 | const Array<T>& values = lhs.getValues(); |
| 67 | const Array<int>& rowIdx = lhs.getRowIdx(); |
| 68 | const Array<int>& colIdx = lhs.getColIdx(); |
| 69 | |
| 70 | if (optLhs == AF_MAT_NONE) { |
| 71 | if (N == 1) { |
| 72 | kernel::csrmv(out, values, rowIdx, colIdx, rhs, alpha, beta); |
| 73 | } else { |
| 74 | kernel::csrmm_nt(out, values, rowIdx, colIdx, rhs, alpha, beta); |
| 75 | } |
| 76 | } else { |
| 77 | // CSR transpose is a CSC matrix |
| 78 | if (N == 1) { |
| 79 | kernel::cscmv(out, values, rowIdx, colIdx, rhs, alpha, beta, |
| 80 | optLhs == AF_MAT_CTRANS); |
| 81 | } else { |
| 82 | kernel::cscmm_nn(out, values, rowIdx, colIdx, rhs, alpha, beta, |
| 83 | optLhs == AF_MAT_CTRANS); |
| 84 | } |
| 85 | } |
| 86 | return out; |
| 87 | } |
| 88 | |
| 89 | #define INSTANTIATE_SPARSE(T) \ |
| 90 | template Array<T> matmul<T>(const common::SparseArray<T>& lhs, \ |