Array<T> createSubArray(const Array<T>& parent,
const std::vector<af_seq> &index,
bool copy)
{
parent.eval();
dim4 dDims = parent.getDataDims();
dim4 pDims = parent.dims();
dim4 dims = toDims (index, pDims);
dim4 offset = toOffset(index, dDims);
dim4 stride = toStride (index, dDims);
Array<T> out = Array<T>(parent, dims, offset, stride);
if (!copy) return out;
if (stride[0] != 1 ||
stride[1] < 0 ||
stride[2] < 0 ||
stride[3] < 0) {
out = copyArray(out);
}
return out;
}
Array<T> createSubArray(const Array<T> &parent, const vector<af_seq> &index,
bool copy) {
parent.eval();
dim4 dDims = parent.getDataDims();
dim4 dStrides = calcStrides(dDims);
dim4 parent_strides = parent.strides();
if (dStrides != parent_strides) {
const Array<T> parentCopy = copyArray(parent);
return createSubArray(parentCopy, index, copy);
}
dim4 pDims = parent.dims();
dim4 dims = toDims(index, pDims);
dim4 strides = toStride(index, dDims);
// Find total offsets after indexing
dim4 offsets = toOffset(index, pDims);
dim_t offset = parent.getOffset();
for (int i = 0; i < 4; i++) offset += offsets[i] * parent_strides[i];
Array<T> out = Array<T>(parent, dims, offset, strides);
if (!copy) return out;
if (strides[0] != 1 || strides[1] < 0 || strides[2] < 0 || strides[3] < 0) {
out = copyArray(out);
}
return out;
}
void assign(Param<T> out, af::dim4 dDims,
CParam<T> rhs, std::vector<bool> const isSeq,
std::vector<af_seq> const seqs, std::vector< CParam<uint> > idxArrs)
{
af::dim4 pDims = out.dims();
// retrieve dimensions & strides for array to which rhs is being copied to
af::dim4 dst_offsets = toOffset(seqs, dDims);
af::dim4 dst_strides = toStride(seqs, dDims);
// retrieve rhs array dimenesions & strides
af::dim4 src_dims = rhs.dims();
af::dim4 src_strides = rhs.strides();
// declare pointers to af_array index data
uint const * const ptr0 = idxArrs[0].get();
uint const * const ptr1 = idxArrs[1].get();
uint const * const ptr2 = idxArrs[2].get();
uint const * const ptr3 = idxArrs[3].get();
const T * src= rhs.get();
T * dst = out.get();
for(dim_t l=0; l<src_dims[3]; ++l) {
dim_t src_loff = l*src_strides[3];
dim_t dst_lIdx = trimIndex(isSeq[3] ? l+dst_offsets[3] : ptr3[l], pDims[3]);
dim_t dst_loff = dst_lIdx * dst_strides[3];
for(dim_t k=0; k<src_dims[2]; ++k) {
dim_t src_koff = k*src_strides[2];
dim_t dst_kIdx = trimIndex(isSeq[2] ? k+dst_offsets[2] : ptr2[k], pDims[2]);
dim_t dst_koff = dst_kIdx * dst_strides[2];
for(dim_t j=0; j<src_dims[1]; ++j) {
dim_t src_joff = j*src_strides[1];
dim_t dst_jIdx = trimIndex(isSeq[1] ? j+dst_offsets[1] : ptr1[j], pDims[1]);
dim_t dst_joff = dst_jIdx * dst_strides[1];
for(dim_t i=0; i<src_dims[0]; ++i) {
dim_t src_ioff = i*src_strides[0];
dim_t src_idx = src_ioff + src_joff + src_koff + src_loff;
dim_t dst_iIdx = trimIndex(isSeq[0] ? i+dst_offsets[0] : ptr0[i], pDims[0]);
dim_t dst_ioff = dst_iIdx * dst_strides[0];
dim_t dst_idx = dst_ioff + dst_joff + dst_koff + dst_loff;
dst[dst_idx] = src[src_idx];
}
}
}
}
}
Array<T> index(const Array<T>& in, const af_index_t idxrs[])
{
kernel::IndexKernelParam_t p;
std::vector<af_seq> seqs(4, af_span);
// create seq vector to retrieve output
// dimensions, offsets & offsets
for (dim_t x=0; x<4; ++x) {
if (idxrs[x].isSeq) {
seqs[x] = idxrs[x].idx.seq;
}
}
// retrieve dimensions, strides and offsets
dim4 iDims = in.dims();
dim4 dDims = in.getDataDims();
dim4 oDims = toDims (seqs, iDims);
dim4 iOffs = toOffset(seqs, dDims);
dim4 iStrds= toStride(seqs, dDims);
for (dim_t i=0; i<4; ++i) {
p.isSeq[i] = idxrs[i].isSeq;
p.offs[i] = iOffs[i];
p.strds[i] = iStrds[i];
}
Buffer* bPtrs[4];
std::vector< Array<uint> > idxArrs(4, createEmptyArray<uint>(dim4()));
// look through indexs to read af_array indexs
for (dim_t x=0; x<4; ++x) {
// set index pointers were applicable
if (!p.isSeq[x]) {
idxArrs[x] = castArray<uint>(idxrs[x].idx.arr);
bPtrs[x] = idxArrs[x].get();
// set output array ith dimension value
oDims[x] = idxArrs[x].elements();
}
else {
// alloc an 1-element buffer to avoid OpenCL from failing
bPtrs[x] = bufferAlloc(sizeof(uint));
}
}
Array<T> out = createEmptyArray<T>(oDims);
if(oDims.elements() == 0) { return out; }
kernel::index<T>(out, in, p, bPtrs);
for (dim_t x=0; x<4; ++x) {
if (p.isSeq[x]) bufferFree(bPtrs[x]);
}
return out;
}
void index(Array<T> out, Array<T> const in,
std::vector<bool> const isSeq, std::vector<af_seq> const seqs,
std::vector< Array<uint> > const idxArrs)
{
const af::dim4 iDims = in.dims();
const af::dim4 dDims = in.getDataDims();
const af::dim4 iOffs = toOffset(seqs, dDims);
const af::dim4 iStrds = toStride(seqs, dDims);
const af::dim4 oDims = out.dims();
const af::dim4 oStrides = out.strides();
const T *src = in.get();
T *dst = out.get();
const uint* ptr0 = idxArrs[0].get();
const uint* ptr1 = idxArrs[1].get();
const uint* ptr2 = idxArrs[2].get();
const uint* ptr3 = idxArrs[3].get();
for (dim_t l=0; l<oDims[3]; ++l) {
dim_t lOff = l*oStrides[3];
dim_t inIdx3 = trimIndex(isSeq[3] ? l+iOffs[3] : ptr3[l], iDims[3]);
dim_t inOff3 = inIdx3*iStrds[3];
for (dim_t k=0; k<oDims[2]; ++k) {
dim_t kOff = k*oStrides[2];
dim_t inIdx2 = trimIndex(isSeq[2] ? k+iOffs[2] : ptr2[k], iDims[2]);
dim_t inOff2 = inIdx2*iStrds[2];
for (dim_t j=0; j<oDims[1]; ++j) {
dim_t jOff = j*oStrides[1];
dim_t inIdx1 = trimIndex(isSeq[1] ? j+iOffs[1] : ptr1[j], iDims[1]);
dim_t inOff1 = inIdx1*iStrds[1];
for (dim_t i=0; i<oDims[0]; ++i) {
dim_t iOff = i*oStrides[0];
dim_t inIdx0 = trimIndex(isSeq[0] ? i+iOffs[0] : ptr0[i], iDims[0]);
dim_t inOff0 = inIdx0*iStrds[0];
dst[lOff+kOff+jOff+iOff] = src[inOff3+inOff2+inOff1+inOff0];
}
}
}
}
}
