diff options
Diffstat (limited to 'numpy')
| -rw-r--r-- | numpy/core/src/multiarray/new_iterator.c.src | 347 |
1 files changed, 189 insertions, 158 deletions
diff --git a/numpy/core/src/multiarray/new_iterator.c.src b/numpy/core/src/multiarray/new_iterator.c.src index 9dc3bcf40..06daed3ba 100644 --- a/numpy/core/src/multiarray/new_iterator.c.src +++ b/numpy/core/src/multiarray/new_iterator.c.src @@ -67,9 +67,19 @@ * needs to be changed as well. */ +struct NpyIter_InternalOnly { + /* Initial fixed position data */ + npy_uint32 itflags; + npy_uint16 ndim, niter; + npy_intp itersize; + /* The rest is variable */ + char iter_flexdata; +}; + +typedef struct NpyIter_AD NpyIter_AxisData; +typedef struct NpyIter_BD NpyIter_BufferData; + /* Byte sizes of the iterator members */ -#define NIT_ITERSIZE_SIZEOF() \ - (NPY_SIZEOF_INTP) #define NIT_PERM_SIZEOF(itflags, ndim, niter) \ ((NPY_SIZEOF_INTP)*(ndim)) #define NIT_DTYPES_SIZEOF(itflags, ndim, niter) \ @@ -85,12 +95,9 @@ #define NIT_BUFFERDATA_SIZEOF(itflags, ndim, niter) \ ((itflags&NPY_ITFLAG_BUFFER) ? ((NPY_SIZEOF_INTP)*(3 + 7*niter)) : 0) -/* Byte offsets of the iterator members */ -#define NIT_ITERSIZE_OFFSET() \ - (8) +/* Byte offsets of the iterator members starting from iter->iter_flexdata */ #define NIT_PERM_OFFSET() \ - (NIT_ITERSIZE_OFFSET() + \ - NIT_ITERSIZE_SIZEOF()) + (0) #define NIT_DTYPES_OFFSET(itflags, ndim, niter) \ (NIT_PERM_OFFSET() + \ NIT_PERM_SIZEOF(itflags, ndim, niter)) @@ -115,52 +122,65 @@ /* Internal-only ITERATOR DATA MEMBER ACCESS */ #define NIT_ITFLAGS(iter) \ - (*((npy_uint32*)(iter))) + ((iter)->itflags) #define NIT_NDIM(iter) \ - (*((npy_uint16*)(iter) + 2)) + ((iter)->ndim) #define NIT_NITER(iter) \ - (*((npy_uint16*)(iter) + 3)) -#define NIT_ITERSIZE(iter) (*((npy_intp *)( \ - (char*)(iter) + NIT_ITERSIZE_OFFSET()))) + ((iter)->niter) +#define NIT_ITERSIZE(iter) \ + (iter->itersize) #define NIT_PERM(iter) ((npy_intp*)( \ - (char*)(iter) + NIT_PERM_OFFSET())) + &(iter)->iter_flexdata + NIT_PERM_OFFSET())) #define NIT_DTYPES(iter) ((PyArray_Descr **)( \ - (char*)(iter) + NIT_DTYPES_OFFSET(itflags, ndim, niter))) + &(iter)->iter_flexdata + NIT_DTYPES_OFFSET(itflags, ndim, niter))) #define NIT_RESETDATAPTR(iter) ((char **)( \ - (char*)(iter) + NIT_RESETDATAPTR_OFFSET(itflags, ndim, niter))) + &(iter)->iter_flexdata + NIT_RESETDATAPTR_OFFSET(itflags, ndim, niter))) #define NIT_BASEOFFSETS(iter) ((npy_intp *)( \ - (char*)(iter) + NIT_BASEOFFSETS_OFFSET(itflags, ndim, niter))) + &(iter)->iter_flexdata + NIT_BASEOFFSETS_OFFSET(itflags, ndim, niter))) #define NIT_OBJECTS(iter) ((PyArrayObject **)( \ - (char*)(iter) + NIT_OBJECTS_OFFSET(itflags, ndim, niter))) + &(iter)->iter_flexdata + NIT_OBJECTS_OFFSET(itflags, ndim, niter))) #define NIT_OPITFLAGS(iter) ( \ - (char*)(iter) + NIT_OPITFLAGS_OFFSET(itflags, ndim, niter)) -#define NIT_BUFFERDATA(iter) \ - ((char*)(iter) + NIT_BUFFERDATA_OFFSET(itflags, ndim, niter)) -#define NIT_AXISDATA(iter) \ - ((char*)(iter) + NIT_AXISDATA_OFFSET(itflags, ndim, niter)) + &(iter)->iter_flexdata + NIT_OPITFLAGS_OFFSET(itflags, ndim, niter)) +#define NIT_BUFFERDATA(iter) ((NpyIter_BufferData *)( \ + &(iter)->iter_flexdata + NIT_BUFFERDATA_OFFSET(itflags, ndim, niter))) +#define NIT_AXISDATA(iter) ((NpyIter_AxisData *)( \ + &(iter)->iter_flexdata + NIT_AXISDATA_OFFSET(itflags, ndim, niter))) /* Internal-only BUFFERDATA MEMBER ACCESS */ -#define NBF_BUFFERSIZE(bufferdata) (*((npy_intp *)(bufferdata))) -#define NBF_SIZE(bufferdata) (*((npy_intp *)(bufferdata) + 1)) -#define NBF_POS(bufferdata) (*((npy_intp *)(bufferdata) + 2)) -#define NBF_STRIDES(bufferdata) ((npy_intp *)(bufferdata) + 3) -#define NBF_PTRS(bufferdata) ((char **)(bufferdata) + 3 + 1*(niter)) -#define NBF_READTRANSFERFN(bufferdata) \ - ((PyArray_StridedTransferFn *)(bufferdata) + 3 + 2*(niter)) -#define NBF_READTRANSFERDATA(bufferdata) \ - ((void **)(bufferdata) + 3 + 3*(niter)) -#define NBF_WRITETRANSFERFN(bufferdata) \ - ((PyArray_StridedTransferFn *)(bufferdata) + 3 + 4*(niter)) -#define NBF_WRITETRANSFERDATA(bufferdata) \ - ((void **)(bufferdata) + 3 + 5*(niter)) -#define NBF_BUFFERS(bufferdata) ((char **)(bufferdata) + 3 + 6*(niter)) +struct NpyIter_BD { + npy_intp buffersize, size, pos; + npy_intp bd_flexdata; +}; +#define NBF_BUFFERSIZE(bufferdata) ((bufferdata)->buffersize) +#define NBF_SIZE(bufferdata) ((bufferdata)->size) +#define NBF_POS(bufferdata) ((bufferdata)->pos) +#define NBF_STRIDES(bufferdata) ( \ + &(bufferdata)->bd_flexdata + 0) +#define NBF_PTRS(bufferdata) ((char **) \ + (&(bufferdata)->bd_flexdata + 1*(niter))) +#define NBF_READTRANSFERFN(bufferdata) ((PyArray_StridedTransferFn *) \ + (&(bufferdata)->bd_flexdata + 2*(niter))) +#define NBF_READTRANSFERDATA(bufferdata) ((void **) \ + (&(bufferdata)->bd_flexdata + 3*(niter))) +#define NBF_WRITETRANSFERFN(bufferdata) ((PyArray_StridedTransferFn *) \ + (&(bufferdata)->bd_flexdata + 4*(niter))) +#define NBF_WRITETRANSFERDATA(bufferdata) ((void **) \ + (&(bufferdata)->bd_flexdata + 5*(niter))) +#define NBF_BUFFERS(bufferdata) ((char **) \ + (&(bufferdata)->bd_flexdata + 6*(niter))) /* Internal-only AXISDATA MEMBER ACCESS. */ -#define NAD_SHAPE(axisdata) (*((npy_intp *)(axisdata))) -#define NAD_COORD(axisdata) (*((npy_intp *)(axisdata) + 1)) -#define NAD_STRIDES(axisdata) ((npy_intp *)(axisdata) + 2) -#define NAD_PTRS(axisdata) \ - ((char**)(axisdata) + 2 + NAD_NSTRIDES()) +struct NpyIter_AD { + npy_intp shape, coord; + npy_intp ad_flexdata; +}; +#define NAD_SHAPE(axisdata) ((axisdata)->shape) +#define NAD_COORD(axisdata) ((axisdata)->coord) +#define NAD_STRIDES(axisdata) ( \ + &(axisdata)->ad_flexdata + 0) +#define NAD_PTRS(axisdata) ((char **) \ + &(axisdata)->ad_flexdata + 1*(niter+1)) + #define NAD_NSTRIDES() \ ((niter) + ((itflags&NPY_ITFLAG_HASINDEX) ? 1 : 0)) @@ -170,14 +190,23 @@ 1 + \ /* intp coord */ \ 1 + \ - /* intp stride[niter] AND char* ptr[niter] */ \ - 2*(niter) + \ - /* intp indexstride AND intp index (when index is provided) */ \ - ((itflags&NPY_ITFLAG_HASINDEX) ? 2 : 0) \ + /* intp stride[niter+1] AND char* ptr[niter+1] */ \ + 2*((niter)+1) \ )*NPY_SIZEOF_INTP ) \ +/* + * Macro to advance an AXISDATA pointer by a specified count. + * Requires that sizeof_axisdata be previously initialized + * to NIT_SIZEOF_AXISDATA(itflags, ndim, niter). + */ +#define NIT_ADVANCE_AXISDATA(axisdata, count) \ + (*((char **)(&axisdata))) += (count)*sizeof_axisdata +#define NIT_INDEX_AXISDATA(axisdata, index) ((NpyIter_AxisData *) \ + (((char *)(axisdata)) + (index)*sizeof_axisdata)) + /* Size of the whole iterator */ #define NIT_SIZEOF_ITERATOR(itflags, ndim, niter) ( \ + sizeof(struct NpyIter_InternalOnly) + \ NIT_AXISDATA_OFFSET(itflags, ndim, niter) + \ NIT_SIZEOF_AXISDATA(itflags, ndim, niter)*(ndim)) @@ -268,7 +297,7 @@ NpyIter_MultiNew(npy_intp niter, PyArrayObject **op_in, npy_uint32 flags, char **op_dataptr; npy_intp *perm; - char *bufferdata = NULL; + NpyIter_BufferData *bufferdata = NULL; char axes_dupcheck[NPY_MAXDIMS]; int any_allocate_if_null = 0, any_missing_dtypes = 0, allocate_output_scalars = 0; @@ -601,7 +630,6 @@ NpyIter_MultiNew(npy_intp niter, PyArrayObject **op_in, npy_uint32 flags, Py_DECREF(dtype); } - /* * All of the data types have been settled, so it's time * to check that data type conversions are following the @@ -742,7 +770,7 @@ NpyIter_MultiNew(npy_intp niter, PyArrayObject **op_in, npy_uint32 flags, (op_itflags[iiter]&NPY_OP_ITFLAG_WRITE))) { int is_one_to_one = 1; npy_intp stride, shape, innerstride = 0, innershape; - char *axisdata = NIT_AXISDATA(iter); + NpyIter_AxisData *axisdata = NIT_AXISDATA(iter); npy_intp sizeof_axisdata = NIT_SIZEOF_AXISDATA(itflags, ndim, niter); /* Find stride of the first non-empty shape */ @@ -755,10 +783,10 @@ NpyIter_MultiNew(npy_intp niter, PyArrayObject **op_in, npy_uint32 flags, } break; } - axisdata += sizeof_axisdata; + NIT_ADVANCE_AXISDATA(axisdata, 1); } ++idim; - axisdata += sizeof_axisdata; + NIT_ADVANCE_AXISDATA(axisdata, 1); /* Check that everything could have coalesced together */ for (; idim < ndim; ++idim) { stride = NAD_STRIDES(axisdata)[iiter]; @@ -778,7 +806,7 @@ NpyIter_MultiNew(npy_intp niter, PyArrayObject **op_in, npy_uint32 flags, innershape *= shape; } } - axisdata += sizeof_axisdata; + NIT_ADVANCE_AXISDATA(axisdata, 1); } /* * If we looped all the way to the end, one stride works. @@ -815,7 +843,7 @@ NpyIter_MultiNew(npy_intp niter, PyArrayObject **op_in, npy_uint32 flags, /* Now that the axes are finished, adjust ITERSIZE if necessary */ if ((itflags&NPY_ITFLAG_NOINNER) && !(itflags&NPY_ITFLAG_BUFFER)) { - char *axisdata = NIT_AXISDATA(iter); + NpyIter_AxisData *axisdata = NIT_AXISDATA(iter); NIT_ITERSIZE(iter) /= NAD_SHAPE(axisdata); } @@ -875,7 +903,7 @@ int NpyIter_Deallocate(NpyIter *iter) /* Deallocate any buffers and buffering data */ if (itflags&NPY_ITFLAG_BUFFER) { - char *bufferdata = NIT_BUFFERDATA(iter); + NpyIter_BufferData *bufferdata = NIT_BUFFERDATA(iter); char **buffers; void **transferdata; @@ -957,7 +985,7 @@ int NpyIter_RemoveInnerLoop(NpyIter *iter) } else { - char *axisdata = NIT_AXISDATA(iter); + NpyIter_AxisData *axisdata = NIT_AXISDATA(iter); NIT_ITERSIZE(iter) /= NAD_SHAPE(axisdata); } } @@ -976,7 +1004,7 @@ void NpyIter_Reset(NpyIter *iter) npy_intp niter = NIT_NITER(iter); char **resetdataptr; - char *axisdata; + NpyIter_AxisData *axisdata; npy_intp sizeof_axisdata; npy_intp istrides, nstrides; @@ -990,7 +1018,7 @@ void NpyIter_Reset(NpyIter *iter) npyiter_copy_from_buffers(iter); } - for (idim = 0; idim < ndim; ++idim, axisdata += sizeof_axisdata) { + for (idim = 0; idim < ndim; ++idim, NIT_ADVANCE_AXISDATA(axisdata, 1)) { char **ptrs; NAD_COORD(axisdata) = 0; ptrs = NAD_PTRS(axisdata); @@ -1013,7 +1041,7 @@ void NpyIter_ResetBasePointers(NpyIter *iter, char **baseptrs) npy_intp iiter, niter = NIT_NITER(iter); char **resetdataptr; - char *axisdata; + NpyIter_AxisData *axisdata; npy_intp *baseoffsets; npy_intp sizeof_axisdata; npy_intp istrides, nstrides; @@ -1034,7 +1062,7 @@ void NpyIter_ResetBasePointers(NpyIter *iter, char **baseptrs) resetdataptr[iiter] = baseptrs[iiter] + baseoffsets[iiter]; } - for (idim = 0; idim < ndim; ++idim, axisdata += sizeof_axisdata) { + for (idim = 0; idim < ndim; ++idim, NIT_ADVANCE_AXISDATA(axisdata, 1)) { char **ptrs; NAD_COORD(axisdata) = 0; ptrs = NAD_PTRS(axisdata); @@ -1064,7 +1092,7 @@ int NpyIter_GotoCoords(NpyIter *iter, npy_intp *coords) npy_intp niter = NIT_NITER(iter); char **dataptr; - char *ad, *axisdata; + NpyIter_AxisData *ad, *axisdata; npy_intp sizeof_axisdata; npy_intp istrides, nstrides; npy_intp *perm, perm_coords[NPY_MAXDIMS]; @@ -1090,7 +1118,7 @@ int NpyIter_GotoCoords(NpyIter *iter, npy_intp *coords) /* Permute the input coordinates to match the iterator */ ad = axisdata; - for (idim = 0; idim < ndim; ++idim, ad += sizeof_axisdata) { + for (idim = 0; idim < ndim; ++idim, NIT_ADVANCE_AXISDATA(ad, 1)) { npy_intp p = perm[idim], i, shape; if (p < 0) { /* If the perm entry is negative, reverse the coordinate */ @@ -1121,8 +1149,8 @@ int NpyIter_GotoCoords(NpyIter *iter, npy_intp *coords) * fastest-changing. The successive pointers accumulate * the offsets, starting from the original data pointers. */ - axisdata += (ndim-1)*sizeof_axisdata; - for (idim = ndim-1; idim >= 0; --idim, axisdata -= sizeof_axisdata) { + axisdata = NIT_INDEX_AXISDATA(NIT_AXISDATA(iter), ndim-1); + for (idim = ndim-1; idim >= 0; --idim, NIT_ADVANCE_AXISDATA(axisdata, -1)) { npy_intp i, *strides; char **ptrs; @@ -1151,7 +1179,7 @@ int NpyIter_GotoIndex(NpyIter *iter, npy_intp index) npy_intp niter = NIT_NITER(iter); char **dataptr; - char *axisdata; + NpyIter_AxisData *axisdata; npy_intp sizeof_axisdata; npy_intp istrides, nstrides; @@ -1186,8 +1214,8 @@ int NpyIter_GotoIndex(NpyIter *iter, npy_intp index) * fastest-changing. The successive pointers accumulate * the offsets, starting from the original data pointers. */ - axisdata += (ndim-1)*sizeof_axisdata; - for (idim = 0; idim < ndim; ++idim, axisdata -= sizeof_axisdata) { + axisdata = NIT_INDEX_AXISDATA(NIT_AXISDATA(iter), ndim-1); + for (idim = 0; idim < ndim; ++idim, NIT_ADVANCE_AXISDATA(axisdata, -1)) { npy_intp i, shape, iterstride, *strides; char **ptrs; @@ -1255,13 +1283,13 @@ npyiter_iternext_itflags@tag_itflags@_dims@tag_ndim@_iters@tag_niter@( npy_intp istrides, nstrides, sizeof_axisdata; #if @const_ndim@ > 0 - char* axisdata0; + NpyIter_AxisData *axisdata0; #endif #if @const_ndim@ > 1 - char* axisdata1; + NpyIter_AxisData *axisdata1; #endif #if @const_ndim@ > 2 - char* axisdata2; + NpyIter_AxisData *axisdata2; #endif nstrides = NAD_NSTRIDES(); @@ -1297,7 +1325,7 @@ npyiter_iternext_itflags@tag_itflags@_dims@tag_ndim@_iters@tag_niter@( } # endif - axisdata1 = axisdata0 + sizeof_axisdata; + axisdata1 = NIT_INDEX_AXISDATA(axisdata0, 1); /* Increment coordinate 1 */ NAD_COORD(axisdata1)++; /* Increment pointer 1 */ @@ -1319,7 +1347,7 @@ npyiter_iternext_itflags@tag_itflags@_dims@tag_ndim@_iters@tag_niter@( return 0; # else - axisdata2 = axisdata1 + sizeof_axisdata; + axisdata2 = NIT_INDEX_AXISDATA(axisdata1, 1); /* Increment coordinate 2 */ NAD_COORD(axisdata2)++; /* Increment pointer 2 */ @@ -1340,7 +1368,7 @@ npyiter_iternext_itflags@tag_itflags@_dims@tag_ndim@_iters@tag_niter@( } for (idim = 3; idim < ndim; ++idim) { - axisdata2 += sizeof_axisdata; + NIT_ADVANCE_AXISDATA(axisdata2, 1); /* Increment the coordinate */ NAD_COORD(axisdata2)++; /* Increment the pointer */ @@ -1353,7 +1381,7 @@ npyiter_iternext_itflags@tag_itflags@_dims@tag_ndim@_iters@tag_niter@( /* Reset the coordinates and pointers of all previous axisdatas */ axisdata1 = axisdata2; do { - axisdata1 -= sizeof_axisdata; + NIT_ADVANCE_AXISDATA(axisdata1, -1); /* Reset the coordinate to 0 */ NAD_COORD(axisdata1) = 0; /* Reset the pointer to the updated value */ @@ -1386,7 +1414,7 @@ npyiter_buffered_iternext(NpyIter *iter) npy_intp ndim = NIT_NDIM(iter); npy_intp niter = NIT_NITER(iter); - char *bufferdata = NIT_BUFFERDATA(iter); + NpyIter_BufferData *bufferdata = NIT_BUFFERDATA(iter); /* * If the iterator handles the inner loop, need to increment all @@ -1544,7 +1572,7 @@ npyiter_getcoord_itflags@tag_itflags@(NpyIter *iter, npy_intp *outcoord) npy_intp niter = NIT_NITER(iter); npy_intp idim, sizeof_axisdata; - char* axisdata; + NpyIter_AxisData *axisdata; #if !((@const_itflags@)&NPY_ITFLAG_IDENTPERM) npy_intp* perm = NIT_PERM(iter); #endif @@ -1554,16 +1582,16 @@ npyiter_getcoord_itflags@tag_itflags@(NpyIter *iter, npy_intp *outcoord) #if ((@const_itflags@)&NPY_ITFLAG_IDENTPERM) outcoord += ndim-1; for(idim = 0; idim < ndim; ++idim, --outcoord, - axisdata += sizeof_axisdata) { + NIT_ADVANCE_AXISDATA(axisdata, 1)) { *outcoord = NAD_COORD(axisdata); } #elif !((@const_itflags@)&NPY_ITFLAG_NEGPERM) - for(idim = 0; idim < ndim; ++idim, axisdata += sizeof_axisdata) { + for(idim = 0; idim < ndim; ++idim, NIT_ADVANCE_AXISDATA(axisdata, 1)) { npy_intp p = perm[idim]; outcoord[ndim-p-1] = NAD_COORD(axisdata); } #else - for(idim = 0; idim < ndim; ++idim, axisdata += sizeof_axisdata) { + for(idim = 0; idim < ndim; ++idim, NIT_ADVANCE_AXISDATA(axisdata, 1)) { npy_intp p = perm[idim]; if (p < 0) { /* If the perm entry is negative, reverse the coordinate */ @@ -1668,7 +1696,7 @@ int NpyIter_GetShape(NpyIter *iter, npy_intp *outshape) npy_intp niter = NIT_NITER(iter); npy_intp idim, sizeof_axisdata; - char* axisdata; + NpyIter_AxisData *axisdata; npy_intp* perm; if (!(itflags&NPY_ITFLAG_HASCOORDS)) { @@ -1681,7 +1709,7 @@ int NpyIter_GetShape(NpyIter *iter, npy_intp *outshape) perm = NIT_PERM(iter); axisdata = NIT_AXISDATA(iter); sizeof_axisdata = NIT_SIZEOF_AXISDATA(itflags, ndim, niter); - for(idim = 0; idim < ndim; ++idim, axisdata += sizeof_axisdata) { + for(idim = 0; idim < ndim; ++idim, NIT_ADVANCE_AXISDATA(axisdata, 1)) { npy_intp p = perm[idim]; if (p < 0) { outshape[ndim+p] = NAD_SHAPE(axisdata); @@ -1700,15 +1728,13 @@ char **NpyIter_GetDataPtrArray(NpyIter *iter) npy_intp ndim = NIT_NDIM(iter); npy_intp niter = NIT_NITER(iter); - char* data; - if (itflags&NPY_ITFLAG_BUFFER) { - data = NIT_BUFFERDATA(iter); - return NBF_PTRS(data); + NpyIter_BufferData *bufferdata = NIT_BUFFERDATA(iter); + return NBF_PTRS(bufferdata); } else { - data = NIT_AXISDATA(iter); - return NAD_PTRS(data); + NpyIter_AxisData *axisdata = NIT_AXISDATA(iter); + return NAD_PTRS(axisdata); } } @@ -1739,7 +1765,8 @@ PyArrayObject *NpyIter_GetIterView(NpyIter *iter, npy_intp i) npy_intp shape[NPY_MAXDIMS], strides[NPY_MAXDIMS]; PyArrayObject *obj, *view; PyArray_Descr *dtype; - char *dataptr, *axisdata; + char *dataptr; + NpyIter_AxisData *axisdata; npy_intp sizeof_axisdata; int writeable; @@ -1764,7 +1791,7 @@ PyArrayObject *NpyIter_GetIterView(NpyIter *iter, npy_intp i) sizeof_axisdata = NIT_SIZEOF_AXISDATA(itflags, ndim, niter); /* Retrieve the shape and strides from the axisdata */ - for (idim = 0; idim < ndim; ++idim, axisdata += sizeof_axisdata) { + for (idim = 0; idim < ndim; ++idim, NIT_ADVANCE_AXISDATA(axisdata, 1)) { shape[ndim-idim-1] = NAD_SHAPE(axisdata); strides[ndim-idim-1] = NAD_STRIDES(axisdata)[i]; } @@ -1792,7 +1819,7 @@ npy_intp *NpyIter_GetIndexPtr(NpyIter *iter) npy_intp ndim = NIT_NDIM(iter); npy_intp niter = NIT_NITER(iter); - char* axisdata = NIT_AXISDATA(iter); + NpyIter_AxisData *axisdata = NIT_AXISDATA(iter); if (itflags&NPY_ITFLAG_HASINDEX) { /* The index is just after the data pointers */ @@ -1836,15 +1863,13 @@ npy_intp *NpyIter_GetInnerStrideArray(NpyIter *iter) npy_intp ndim = NIT_NDIM(iter); npy_intp niter = NIT_NITER(iter); - char* data; - if (itflags&NPY_ITFLAG_BUFFER) { - data = NIT_BUFFERDATA(iter); + NpyIter_BufferData *data = NIT_BUFFERDATA(iter); return NBF_STRIDES(data); } else { - data = NIT_AXISDATA(iter); - return NAD_STRIDES(data); + NpyIter_AxisData *axisdata = NIT_AXISDATA(iter); + return NAD_STRIDES(axisdata); } } @@ -1854,15 +1879,13 @@ npy_intp* NpyIter_GetInnerLoopSizePtr(NpyIter *iter) npy_intp ndim = NIT_NDIM(iter); npy_intp niter = NIT_NITER(iter); - char* data; - if (itflags&NPY_ITFLAG_BUFFER) { - data = NIT_BUFFERDATA(iter); + NpyIter_BufferData *data = NIT_BUFFERDATA(iter); return &NBF_SIZE(data); } else { - data = NIT_AXISDATA(iter); - return &NAD_SHAPE(data); + NpyIter_AxisData *axisdata = NIT_AXISDATA(iter); + return &NAD_SHAPE(axisdata); } } @@ -2259,7 +2282,8 @@ npyiter_fill_axisdata(NpyIter *iter, PyArrayObject **op, npy_intp iiter, niter = NIT_NITER(iter); npy_intp ondim; - char *odataptr, *axisdata0, *axisdata; + char *odataptr; + NpyIter_AxisData *axisdata0, *axisdata; npy_intp sizeof_axisdata; axisdata0 = NIT_AXISDATA(iter); @@ -2298,7 +2322,7 @@ npyiter_fill_axisdata(NpyIter *iter, PyArrayObject **op, } NAD_PTRS(axisdata)[0] = odataptr; - axisdata += sizeof_axisdata; + NIT_ADVANCE_AXISDATA(axisdata, 1); } for (idim = ondim; idim < ndim; ++idim) { NAD_SHAPE(axisdata) = 1; @@ -2306,7 +2330,7 @@ npyiter_fill_axisdata(NpyIter *iter, PyArrayObject **op, NAD_STRIDES(axisdata)[0] = 0; NAD_PTRS(axisdata)[0] = odataptr; - axisdata += sizeof_axisdata; + NIT_ADVANCE_AXISDATA(axisdata, 1); } } else { @@ -2352,7 +2376,7 @@ npyiter_fill_axisdata(NpyIter *iter, PyArrayObject **op, return 0; } - axisdata += sizeof_axisdata; + NIT_ADVANCE_AXISDATA(axisdata, 1); } } @@ -2400,13 +2424,13 @@ npyiter_fill_axisdata(NpyIter *iter, PyArrayObject **op, } NAD_PTRS(axisdata)[iiter] = odataptr; - axisdata += sizeof_axisdata; + NIT_ADVANCE_AXISDATA(axisdata, 1); } for (idim = ondim; idim < ndim; ++idim) { NAD_STRIDES(axisdata)[iiter] = 0; NAD_PTRS(axisdata)[iiter] = odataptr; - axisdata += sizeof_axisdata; + NIT_ADVANCE_AXISDATA(axisdata, 1); } } else { @@ -2459,7 +2483,7 @@ npyiter_fill_axisdata(NpyIter *iter, PyArrayObject **op, return 0; } - axisdata += sizeof_axisdata; + NIT_ADVANCE_AXISDATA(axisdata, 1); } } } @@ -2479,7 +2503,7 @@ npyiter_fill_axisdata(NpyIter *iter, PyArrayObject **op, axes = op_axes ? op_axes[iiter] : NULL; axisdata = axisdata0; - for (idim = 0; idim < ndim; ++idim, axisdata += sizeof_axisdata) { + for (idim = 0; idim < ndim; ++idim, NIT_ADVANCE_AXISDATA(axisdata, 1)) { npy_intp i; if (axes) { i = axes[ndim-idim-1]; @@ -2505,9 +2529,8 @@ npyiter_fill_axisdata(NpyIter *iter, PyArrayObject **op, for (idim = 0; idim < ndim; ++idim) { NIT_ITERSIZE(iter) *= NAD_SHAPE(axisdata); - axisdata += sizeof_axisdata; + NIT_ADVANCE_AXISDATA(axisdata, 1); } - return 1; } @@ -2530,7 +2553,7 @@ npyiter_replace_axisdata(NpyIter *iter, npy_intp iiter, npy_intp idim, ndim = NIT_NDIM(iter); npy_intp niter = NIT_NITER(iter); - char *axisdata0, *axisdata; + NpyIter_AxisData *axisdata0, *axisdata; npy_intp sizeof_axisdata; npy_intp *perm; npy_intp baseoffset = 0; @@ -2545,7 +2568,7 @@ npyiter_replace_axisdata(NpyIter *iter, npy_intp iiter, */ axisdata = axisdata0; - for (idim = 0; idim < ndim; ++idim, axisdata += sizeof_axisdata) { + for (idim = 0; idim < ndim; ++idim, NIT_ADVANCE_AXISDATA(axisdata, 1)) { npy_intp i, p, shape; /* Apply the perm to get the original axis */ @@ -2587,7 +2610,7 @@ npyiter_replace_axisdata(NpyIter *iter, npy_intp iiter, NIT_RESETDATAPTR(iter)[iiter] = op_dataptr; NIT_BASEOFFSETS(iter)[iiter] = baseoffset; axisdata = axisdata0; - for (idim = 0; idim < ndim; ++idim, axisdata += sizeof_axisdata) { + for (idim = 0; idim < ndim; ++idim, NIT_ADVANCE_AXISDATA(axisdata, 1)) { NAD_PTRS(axisdata)[iiter] = op_dataptr; } } @@ -2607,7 +2630,7 @@ npyiter_compute_index_strides(NpyIter *iter, npy_uint32 flags) npy_intp niter = NIT_NITER(iter); npy_intp indexstride; - char* axisdata; + NpyIter_AxisData *axisdata; npy_intp sizeof_axisdata; /* @@ -2627,7 +2650,7 @@ npyiter_compute_index_strides(NpyIter *iter, npy_uint32 flags) sizeof_axisdata = NIT_SIZEOF_AXISDATA(itflags, ndim, niter); axisdata = NIT_AXISDATA(iter); indexstride = 1; - for(idim = 0; idim < ndim; ++idim, axisdata += sizeof_axisdata) { + for(idim = 0; idim < ndim; ++idim, NIT_ADVANCE_AXISDATA(axisdata, 1)) { npy_intp shape = NAD_SHAPE(axisdata); if (shape == 1) { @@ -2642,9 +2665,9 @@ npyiter_compute_index_strides(NpyIter *iter, npy_uint32 flags) } else if (flags&NPY_ITER_F_ORDER_INDEX) { sizeof_axisdata = NIT_SIZEOF_AXISDATA(itflags, ndim, niter); - axisdata = NIT_AXISDATA(iter) + (ndim-1)*sizeof_axisdata; + axisdata = NIT_INDEX_AXISDATA(NIT_AXISDATA(iter), ndim-1); indexstride = 1; - for(idim = 0; idim < ndim; ++idim, axisdata -= sizeof_axisdata) { + for(idim = 0; idim < ndim; ++idim, NIT_ADVANCE_AXISDATA(axisdata, -1)) { npy_intp shape = NAD_SHAPE(axisdata); if (shape == 1) { @@ -2723,14 +2746,14 @@ npyiter_flip_negative_strides(NpyIter *iter) npy_intp iiter, niter = NIT_NITER(iter); npy_intp istrides, nstrides = NAD_NSTRIDES(); - char *axisdata, *axisdata0; + NpyIter_AxisData *axisdata, *axisdata0; npy_intp *baseoffsets; npy_intp sizeof_axisdata = NIT_SIZEOF_AXISDATA(itflags, ndim, niter); int any_flipped = 0; axisdata0 = axisdata = NIT_AXISDATA(iter); baseoffsets = NIT_BASEOFFSETS(iter); - for (idim = 0; idim < ndim; ++idim, axisdata += sizeof_axisdata) { + for (idim = 0; idim < ndim; ++idim, NIT_ADVANCE_AXISDATA(axisdata, 1)) { npy_intp *strides = NAD_STRIDES(axisdata); int any_negative = 0; @@ -2779,7 +2802,7 @@ npyiter_flip_negative_strides(NpyIter *iter) resetdataptr[istrides] += baseoffsets[istrides]; } axisdata = axisdata0; - for (idim = 0; idim < ndim; ++idim, axisdata += sizeof_axisdata) { + for (idim = 0; idim < ndim; ++idim, NIT_ADVANCE_AXISDATA(axisdata, 1)) { char **ptrs = NAD_PTRS(axisdata); for (istrides = 0; istrides < nstrides; ++istrides) { ptrs[istrides] = resetdataptr[istrides]; @@ -2848,7 +2871,7 @@ npyiter_find_best_axis_ordering(NpyIter *iter) npy_intp i0, i1, ipos, j0, j1; npy_intp *perm; - char *axisdata = NIT_AXISDATA(iter); + NpyIter_AxisData *axisdata = NIT_AXISDATA(iter); npy_intp sizeof_axisdata = NIT_SIZEOF_AXISDATA(itflags, ndim, niter); int permuted = 0; @@ -2865,13 +2888,15 @@ npyiter_find_best_axis_ordering(NpyIter *iter) /* 'ipos' is where perm[i0] will get inserted */ ipos = i0; j0 = perm[i0]; - strides0 = NAD_STRIDES(axisdata + j0*sizeof_axisdata); + + strides0 = NAD_STRIDES(NIT_INDEX_AXISDATA(axisdata, j0)); for (i1 = i0-1; i1 >= 0; --i1) { int ambig = 1, shouldswap = 0; npy_intp *strides1; j1 = perm[i1]; - strides1 = NAD_STRIDES(axisdata + j1*sizeof_axisdata); + + strides1 = NAD_STRIDES(NIT_INDEX_AXISDATA(axisdata, j1)); for (iiter = 0; iiter < niter; ++iiter) { if (strides0[iiter] != 0 && strides1[iiter] != 0) { @@ -2926,20 +2951,21 @@ npyiter_find_best_axis_ordering(NpyIter *iter) /* Apply the computed permutation to the AXISDATA array */ if (permuted == 1) { npy_intp i, size = sizeof_axisdata/NPY_SIZEOF_INTP; - char *ad_i; + NpyIter_AxisData *ad_i; /* Use the coord as a flag, set each to 1 */ - for (idim = 0; idim < ndim; ++idim) { - NAD_COORD(axisdata + idim*sizeof_axisdata) = 1; + ad_i = axisdata; + for (idim = 0; idim < ndim; ++idim, NIT_ADVANCE_AXISDATA(ad_i, 1)) { + NAD_COORD(ad_i) = 1; } /* Apply the permutation by following the cycles */ for (idim = 0; idim < ndim; ++idim) { - ad_i = axisdata + idim*sizeof_axisdata; + ad_i = NIT_INDEX_AXISDATA(axisdata, idim); /* If this axis hasn't been touched yet, process it */ if (NAD_COORD(ad_i) == 1) { npy_intp pidim = perm[idim], qidim, tmp; - char *ad_p, *ad_q; + NpyIter_AxisData *ad_p, *ad_q; if (pidim != idim) { /* Follow the cycle, copying the data */ @@ -2949,7 +2975,7 @@ npyiter_find_best_axis_ordering(NpyIter *iter) ad_q = ad_i; tmp = *((npy_intp*)ad_q + i); while (pidim != idim) { - ad_p = axisdata + pidim*sizeof_axisdata; + ad_p = NIT_INDEX_AXISDATA(axisdata, pidim); *((npy_intp*)ad_q + i) = *((npy_intp*)ad_p + i); qidim = pidim; @@ -2960,8 +2986,9 @@ npyiter_find_best_axis_ordering(NpyIter *iter) } /* Follow the cycle again, marking it as done */ pidim = perm[idim]; + while (pidim != idim) { - NAD_COORD(axisdata + pidim*sizeof_axisdata) = 0; + NAD_COORD(NIT_INDEX_AXISDATA(axisdata, pidim)) = 0; pidim = perm[pidim]; } } @@ -2981,9 +3008,9 @@ npyiter_coalesce_axes(NpyIter *iter) npy_intp niter = NIT_NITER(iter); npy_intp istrides, nstrides = NAD_NSTRIDES(); - char *axisdata = NIT_AXISDATA(iter); + NpyIter_AxisData *axisdata = NIT_AXISDATA(iter); npy_intp sizeof_axisdata = NIT_SIZEOF_AXISDATA(itflags, ndim, niter); - char *ad_compress; + NpyIter_AxisData *ad_compress; npy_intp new_ndim = 1; /* The HASCOORDS or IDENTPERM flags do not apply after coalescing */ @@ -2995,9 +3022,9 @@ npyiter_coalesce_axes(NpyIter *iter) for (idim = 0; idim < ndim-1; ++idim) { int can_coalesce = 1; npy_intp shape0 = NAD_SHAPE(ad_compress); - npy_intp shape1 = NAD_SHAPE(axisdata + sizeof_axisdata); + npy_intp shape1 = NAD_SHAPE(NIT_INDEX_AXISDATA(axisdata, 1)); npy_intp *strides0 = NAD_STRIDES(ad_compress); - npy_intp *strides1 = NAD_STRIDES(axisdata + sizeof_axisdata); + npy_intp *strides1 = NAD_STRIDES(NIT_INDEX_AXISDATA(axisdata, 1)); /* Check that all the axes can be coalesced */ for (istrides = 0; istrides < nstrides; ++istrides) { @@ -3012,7 +3039,7 @@ npyiter_coalesce_axes(NpyIter *iter) if (can_coalesce) { npy_intp *strides = NAD_STRIDES(ad_compress); - axisdata += sizeof_axisdata; + NIT_ADVANCE_AXISDATA(axisdata, 1); NAD_SHAPE(ad_compress) *= NAD_SHAPE(axisdata); for (istrides = 0; istrides < nstrides; ++istrides) { if (strides[istrides] == 0) { @@ -3021,8 +3048,8 @@ npyiter_coalesce_axes(NpyIter *iter) } } else { - axisdata += sizeof_axisdata; - ad_compress += sizeof_axisdata; + NIT_ADVANCE_AXISDATA(axisdata, 1); + NIT_ADVANCE_AXISDATA(ad_compress, 1); if (ad_compress != axisdata) { memcpy(ad_compress, axisdata, sizeof_axisdata); } @@ -3052,7 +3079,8 @@ npyiter_shrink_ndim(NpyIter *iter, npy_intp new_ndim) npy_intp ndim = NIT_NDIM(iter); npy_intp niter = NIT_NITER(iter); - char *iterdata = (char*)iter; + char *to, *from; + npy_intp size; /* * The only place that will shift is perm[ndim], so only one @@ -3060,10 +3088,10 @@ npyiter_shrink_ndim(NpyIter *iter, npy_intp new_ndim) * extra buffers stored after the main iterator structure shouldn't * be moved. */ - memmove(iterdata + NIT_DTYPES_OFFSET(itflags, new_ndim, niter), - iterdata + NIT_DTYPES_OFFSET(itflags, ndim, niter), - NIT_SIZEOF_ITERATOR(itflags, new_ndim, niter) - - NIT_DTYPES_OFFSET(itflags, new_ndim, niter)); + from = &iter->iter_flexdata + NIT_DTYPES_OFFSET(itflags, ndim, niter); + to = &iter->iter_flexdata + NIT_DTYPES_OFFSET(itflags, new_ndim, niter); + size = ((char *)iter) + NIT_SIZEOF_ITERATOR(itflags, new_ndim, niter) - to; + memmove(to, from, size); NIT_NDIM(iter) = new_ndim; } @@ -3091,7 +3119,7 @@ npyiter_new_temp_array(NpyIter *iter, PyTypeObject *subtype, npy_intp new_shape[NPY_MAXDIMS], strides[NPY_MAXDIMS], stride = op_dtype->elsize; char reversestride[NPY_MAXDIMS], anyreverse = 0; - char *axisdata = NIT_AXISDATA(iter); + NpyIter_AxisData *axisdata = NIT_AXISDATA(iter); npy_intp sizeof_axisdata = NIT_SIZEOF_AXISDATA(itflags, ndim, niter); PyArrayObject *ret; @@ -3102,7 +3130,7 @@ npyiter_new_temp_array(NpyIter *iter, PyTypeObject *subtype, reversestride[idim] = 0; } - for (idim = 0; idim < ndim; ++idim, axisdata += sizeof_axisdata) { + for (idim = 0; idim < ndim; ++idim, NIT_ADVANCE_AXISDATA(axisdata, 1)) { npy_intp i, p; /* Apply the perm to get the original axis */ @@ -3405,7 +3433,8 @@ npyiter_allocate_buffers(NpyIter *iter) npy_intp iiter, niter = NIT_NITER(iter); char *op_itflags = NIT_OPITFLAGS(iter); - char *bufferdata = NIT_BUFFERDATA(iter), *axisdata = NIT_AXISDATA(iter); + NpyIter_BufferData *bufferdata = NIT_BUFFERDATA(iter); + NpyIter_AxisData *axisdata = NIT_AXISDATA(iter); PyArrayObject **op = NIT_OBJECTS(iter); PyArray_Descr **op_dtype = NIT_DTYPES(iter); npy_intp *strides = NAD_STRIDES(axisdata), op_stride; @@ -3492,7 +3521,8 @@ npyiter_jumpforward(NpyIter *iter, npy_intp count) npy_intp idim, ndim = NIT_NDIM(iter); npy_intp niter = NIT_NITER(iter); - char *axisdata, **dataptr; + NpyIter_AxisData *axisdata; + char **dataptr; npy_intp sizeof_axisdata; npy_intp delta, coord, shape; npy_intp istrides, nstrides; @@ -3501,7 +3531,7 @@ npyiter_jumpforward(NpyIter *iter, npy_intp count) axisdata = NIT_AXISDATA(iter); /* Go forward through axisdata, incrementing the coordinates */ - for (idim = 0; idim < ndim; ++idim, axisdata += sizeof_axisdata) { + for (idim = 0; idim < ndim; ++idim, NIT_ADVANCE_AXISDATA(axisdata, 1)) { shape = NAD_SHAPE(axisdata); delta = count % shape; coord = NAD_COORD(axisdata) + delta; @@ -3520,7 +3550,7 @@ npyiter_jumpforward(NpyIter *iter, npy_intp count) /* If it incremented past the end, set to the last coordinate */ if (count > 0) { axisdata = NIT_AXISDATA(iter); - for (idim = 0; idim < ndim; ++idim, axisdata += sizeof_axisdata) { + for (idim = 0; idim < ndim; ++idim, NIT_ADVANCE_AXISDATA(axisdata, 1)) { NAD_COORD(axisdata) = NAD_SHAPE(axisdata)-1; } return 0; @@ -3535,8 +3565,9 @@ npyiter_jumpforward(NpyIter *iter, npy_intp count) * fastest-changing. The successive pointers accumulate * the offsets, starting from the original data pointers. */ - axisdata = NIT_AXISDATA(iter) + (ndim-1)*sizeof_axisdata; - for (idim = 0; idim < ndim; ++idim, axisdata -= sizeof_axisdata) { + axisdata = NIT_AXISDATA(iter); + NIT_ADVANCE_AXISDATA(axisdata, ndim-1); + for (idim = 0; idim < ndim; ++idim, NIT_ADVANCE_AXISDATA(axisdata, -1)) { npy_intp *strides; char **ptrs; @@ -3567,7 +3598,7 @@ npyiter_checkspaceleft(NpyIter *iter, npy_intp count, npy_intp idim, ndim = NIT_NDIM(iter); npy_intp niter = NIT_NITER(iter); - char *axisdata; + NpyIter_AxisData *axisdata; npy_intp sizeof_axisdata; npy_intp coord, shape; npy_intp spaceleft = 1, factor = 1; @@ -3577,7 +3608,7 @@ npyiter_checkspaceleft(NpyIter *iter, npy_intp count, /* Go forward through axisdata, calculating the space left */ for (idim = 0; idim < ndim && spaceleft < count; - ++idim, axisdata += sizeof_axisdata) { + ++idim, NIT_ADVANCE_AXISDATA(axisdata, 1)) { shape = NAD_SHAPE(axisdata); coord = NAD_COORD(axisdata); spaceleft += (shape-coord-1) * factor; @@ -3614,8 +3645,8 @@ npyiter_copy_from_buffers(NpyIter *iter) npy_intp iiter, niter = NIT_NITER(iter); char *op_itflags = NIT_OPITFLAGS(iter); - char *bufferdata = NIT_BUFFERDATA(iter), - *axisdata = NIT_AXISDATA(iter); + NpyIter_BufferData *bufferdata = NIT_BUFFERDATA(iter); + NpyIter_AxisData *axisdata = NIT_AXISDATA(iter); PyArray_Descr **dtypes = NIT_DTYPES(iter); npy_intp transfersize = NBF_SIZE(bufferdata); @@ -3668,8 +3699,8 @@ npyiter_copy_to_buffers(NpyIter *iter) npy_intp iiter, niter = NIT_NITER(iter); char *op_itflags = NIT_OPITFLAGS(iter); - char *bufferdata = NIT_BUFFERDATA(iter), - *axisdata = NIT_AXISDATA(iter); + NpyIter_BufferData *bufferdata = NIT_BUFFERDATA(iter); + NpyIter_AxisData *axisdata = NIT_AXISDATA(iter); PyArray_Descr **dtypes = NIT_DTYPES(iter); npy_intp *strides = NBF_STRIDES(bufferdata), @@ -3771,12 +3802,12 @@ NpyIter_DebugPrint(NpyIter *iter) npy_intp idim, ndim = NIT_NDIM(iter); npy_intp iiter, niter = NIT_NITER(iter); - char *axisdata; + NpyIter_AxisData *axisdata; npy_intp sizeof_axisdata; printf("\n------ BEGIN ITERATOR DUMP ------\n"); printf("Iterator Address: %p\n", iter); - printf("Flags: "); + printf("ItFlags: "); if (itflags&NPY_ITFLAG_IDENTPERM) printf("IDENTPERM "); if (itflags&NPY_ITFLAG_NEGPERM) @@ -3863,7 +3894,7 @@ NpyIter_DebugPrint(NpyIter *iter) printf("\n"); if (itflags&NPY_ITFLAG_BUFFER) { - char *bufferdata = NIT_BUFFERDATA(iter); + NpyIter_BufferData *bufferdata = NIT_BUFFERDATA(iter); printf("BufferData:\n"); printf(" BufferSize: %d\n", (int)NBF_BUFFERSIZE(bufferdata)); printf(" Size: %d\n", (int)NBF_SIZE(bufferdata)); @@ -3901,7 +3932,7 @@ NpyIter_DebugPrint(NpyIter *iter) axisdata = NIT_AXISDATA(iter); sizeof_axisdata = NIT_SIZEOF_AXISDATA(itflags, ndim, niter); - for (idim = 0; idim < ndim; ++idim, axisdata += sizeof_axisdata) { + for (idim = 0; idim < ndim; ++idim, NIT_ADVANCE_AXISDATA(axisdata, 1)) { printf("AxisData[%d]:\n", (int)idim); printf(" Shape: %d\n", (int)NAD_SHAPE(axisdata)); printf(" Coord: %d\n", (int)NAD_COORD(axisdata)); |