| 889 | |
| 890 | template<typename typ> |
| 891 | static inline void * |
| 892 | delinearize_matrix(typ *dst, |
| 893 | typ *src, |
| 894 | const LINEARIZE_DATA_t* data) |
| 895 | { |
| 896 | using ftyp = fortran_type_t<typ>; |
| 897 | |
| 898 | if (src) { |
| 899 | int i; |
| 900 | typ *rv = src; |
| 901 | fortran_int columns = (fortran_int)data->columns; |
| 902 | fortran_int column_strides = |
| 903 | (fortran_int)(data->column_strides/sizeof(typ)); |
| 904 | fortran_int one = 1; |
| 905 | for (i = 0; i < data->rows; i++) { |
| 906 | if (column_strides > 0) { |
| 907 | copy(&columns, |
| 908 | (ftyp*)src, &one, |
| 909 | (ftyp*)dst, &column_strides); |
| 910 | } |
| 911 | else if (column_strides < 0) { |
| 912 | copy(&columns, |
| 913 | (ftyp*)src, &one, |
| 914 | ((ftyp*)dst + (columns-1)*column_strides), |
| 915 | &column_strides); |
| 916 | } |
| 917 | else { |
| 918 | /* |
| 919 | * Zero stride has undefined behavior in some BLAS |
| 920 | * implementations (e.g. OSX Accelerate), so do it |
| 921 | * manually |
| 922 | */ |
| 923 | if (columns > 0) { |
| 924 | memcpy(dst, |
| 925 | src + (columns-1), |
| 926 | sizeof(typ)); |
| 927 | } |
| 928 | } |
| 929 | src += data->output_lead_dim; |
| 930 | dst += data->row_strides/sizeof(typ); |
| 931 | } |
| 932 | |
| 933 | return rv; |
| 934 | } else { |
| 935 | return src; |
| 936 | } |
| 937 | } |
| 938 | |
| 939 | template<typename typ> |
| 940 | static inline void |
no test coverage detected