static
void assign(af_array &out, const unsigned &ndims, const af_seq *index, const af_array &in)
{
ArrayInfo iInfo = getInfo(in);
ArrayInfo oInfo = getInfo(out);
af_dtype iType = iInfo.getType();
dim4 const outDs = oInfo.dims();
dim4 const iDims = iInfo.dims();
ARG_ASSERT(0, (outDs.ndims()>=iDims.ndims()));
ARG_ASSERT(1, (outDs.ndims()>=(int)ndims));
AF_CHECK(af_eval(out));
vector<af_seq> index_(index, index+ndims);
dim4 const oStrides = af::toStride(index_, outDs);
dim4 oDims = af::toDims(index_, outDs);
dim4 oOffsets = af::toOffset(index_, outDs);
Array<T> *dst = createRefArray<T>(getArray<T>(out), oDims, oOffsets, oStrides);
for (int i = 0; i < 4; i++) {
if (oDims[i] != iDims[i])
AF_ERROR("Size mismatch between input and output", AF_ERR_SIZE);
}
bool noCaseExecuted = true;
if (isComplex) {
noCaseExecuted = false;
switch(iType) {
case c64: copy<cdouble, T>(*dst, getArray<cdouble>(in), scalar<T>(0), 1.0); break;
case c32: copy<cfloat , T>(*dst, getArray<cfloat >(in), scalar<T>(0), 1.0); break;
default : noCaseExecuted = true; break;
}
}
static const T ZERO = scalar<T>(0);
if(noCaseExecuted) {
noCaseExecuted = false;
switch(iType) {
case f64: copy<double , T>(*dst, getArray<double>(in), ZERO, 1.0); break;
case f32: copy<float , T>(*dst, getArray<float >(in), ZERO, 1.0); break;
case s32: copy<int , T>(*dst, getArray<int >(in), ZERO, 1.0); break;
case u32: copy<uint , T>(*dst, getArray<uint >(in), ZERO, 1.0); break;
case u8 : copy<uchar , T>(*dst, getArray<uchar >(in), ZERO, 1.0); break;
case b8 : copy<char , T>(*dst, getArray<char >(in), ZERO, 1.0); break;
default : noCaseExecuted = true; break;
}
}
if (noCaseExecuted)
TYPE_ERROR(1, iType);
delete dst;
}
static
void assign(Array<Tout> &out, const unsigned &ndims, const af_seq *index, const Array<Tin> &in_)
{
dim4 const outDs = out.dims();
dim4 const iDims = in_.dims();
DIM_ASSERT(0, (outDs.ndims()>=iDims.ndims()));
DIM_ASSERT(0, (outDs.ndims()>=(dim_t)ndims));
out.eval();
vector<af_seq> index_(index, index+ndims);
dim4 oDims = toDims(index_, outDs);
bool is_vector = true;
for (int i = 0; is_vector && i < (int)oDims.ndims() - 1; i++) {
is_vector &= oDims[i] == 1;
}
is_vector &= in_.isVector() || in_.isScalar();
for (dim_t i = ndims; i < (int)in_.ndims(); i++) {
oDims[i] = 1;
}
if (is_vector) {
if (oDims.elements() != (dim_t)in_.elements() &&
in_.elements() != 1) {
AF_ERROR("Size mismatch between input and output", AF_ERR_SIZE);
}
// If both out and in are vectors of equal elements, reshape in to out dims
Array<Tin> in = in_.elements() == 1 ? tile(in_, oDims) : modDims(in_, oDims);
Array<Tout> dst = createSubArray<Tout>(out, index_, false);
copyArray<Tin , Tout>(dst, in);
} else {
for (int i = 0; i < 4; i++) {
if (oDims[i] != iDims[i]) {
AF_ERROR("Size mismatch between input and output", AF_ERR_SIZE);
}
}
Array<Tout> dst = createSubArray<Tout>(out, index_, false);
copyArray<Tin , Tout>(dst, in_);
}
}
array constant(cfloat val, const dim4 &dims)
{
af_array res;
AF_THROW(af_constant_complex(&res, real(val), imag(val),
dims.ndims(), dims.get(), c32));
return array(res);
}
array
constant(T val, const dim4 &dims, const af::dtype type)
{
af_array res;
if (type != s64 && type != u64) {
AF_THROW(af_constant(&res, (double)val,
dims.ndims(), dims.get(), type));
}
else if (type == s64) {
AF_THROW(af_constant_long (&res, ( intl)val,
dims.ndims(),
dims.get()));
} else {
AF_THROW(af_constant_ulong(&res, (uintl)val,
dims.ndims(),
dims.get()));
}
return array(res);
}
AFAPI array constant(cdouble val, const dim4 &dims, const af::dtype type)
{
if (type != c32 && type != c64) {
return constant(real(val), dims, type);
}
af_array res;
AF_THROW(af_constant_complex(&res,
real(val),
imag(val),
dims.ndims(),
dims.get(), type));
return array(res);
}
AF_BATCH_KIND identifyBatchKind(const dim4 &sDims, const dim4 &fDims) {
dim_t sn = sDims.ndims();
dim_t fn = fDims.ndims();
if (sn == baseDim && fn == baseDim)
return AF_BATCH_NONE;
else if (sn == baseDim && (fn > baseDim && fn <= 4))
return AF_BATCH_RHS;
else if ((sn > baseDim && sn <= 4) && fn == baseDim)
return AF_BATCH_LHS;
else if ((sn > baseDim && sn <= 4) && (fn > baseDim && fn <= 4)) {
bool doesDimensionsMatch = true;
bool isInterleaved = true;
for (dim_t i = baseDim; i < 4; i++) {
doesDimensionsMatch &= (sDims[i] == fDims[i]);
isInterleaved &=
(sDims[i] == 1 || fDims[i] == 1 || sDims[i] == fDims[i]);
}
if (doesDimensionsMatch) return AF_BATCH_SAME;
return (isInterleaved ? AF_BATCH_DIFF : AF_BATCH_UNSUPPORTED);
} else
return AF_BATCH_UNSUPPORTED;
}
ConvolveBatchKind identifyBatchKind(const dim4 &sDims, const dim4 &fDims)
{
dim_t sn = sDims.ndims();
dim_t fn = fDims.ndims();
if (sn==baseDim && fn==baseDim)
return ONE2ONE;
else if (sn==baseDim && (fn>baseDim && fn<=4))
return ONE2MANY;
else if ((sn>baseDim && sn<=4) && fn==baseDim)
return MANY2ONE;
else if ((sn>baseDim && sn<=4) && (fn>baseDim && fn<=4)) {
bool doesDimensionsMatch = true;
for (dim_t i=baseDim; i<4; i++) {
if (sDims[i]!=fDims[i]) {
doesDimensionsMatch = false;
break;
}
}
return (doesDimensionsMatch ? MANY2MANY : CONVOLVE_UNSUPPORTED_BATCH_MODE);
}
else
return CONVOLVE_UNSUPPORTED_BATCH_MODE;
}
Array<in_t> lookup(const Array<in_t> &input,
const Array<idx_t> &indices, const unsigned dim)
{
const dim4 iDims = input.dims();
dim4 oDims(1);
for (int d=0; d<4; ++d)
oDims[d] = (d==int(dim) ? indices.elements() : iDims[d]);
Array<in_t> out = createEmptyArray<in_t>(oDims);
dim_t nDims = iDims.ndims();
switch(dim) {
case 0: kernel::lookup<in_t, idx_t, 0>(out, input, indices, nDims); break;
case 1: kernel::lookup<in_t, idx_t, 1>(out, input, indices, nDims); break;
case 2: kernel::lookup<in_t, idx_t, 2>(out, input, indices, nDims); break;
case 3: kernel::lookup<in_t, idx_t, 3>(out, input, indices, nDims); break;
}
return out;
}
array iota(const dim4 &dims, const dim4 &tile_dims, const af::dtype ty)
{
af_array out;
AF_THROW(af_iota(&out, dims.ndims(), dims.get(), tile_dims.ndims(), tile_dims.get(), ty));
return array(out);
}
// Assign values to an array
array::array_proxy&
af::array::array_proxy::operator=(const array &other)
{
unsigned nd = numDims(impl->parent_->get());
const dim4 this_dims = getDims(impl->parent_->get());
const dim4 other_dims = other.dims();
int dim = gforDim(impl->indices_);
af_array other_arr = other.get();
bool batch_assign = false;
bool is_reordered = false;
if (dim >= 0) {
//FIXME: Figure out a faster, cleaner way to do this
dim4 out_dims = seqToDims(impl->indices_, this_dims, false);
batch_assign = true;
for (int i = 0; i < AF_MAX_DIMS; i++) {
if (this->impl->indices_[i].isBatch) batch_assign &= (other_dims[i] == 1);
else batch_assign &= (other_dims[i] == out_dims[i]);
}
if (batch_assign) {
af_array out;
AF_THROW(af_tile(&out, other_arr,
out_dims[0] / other_dims[0],
out_dims[1] / other_dims[1],
out_dims[2] / other_dims[2],
out_dims[3] / other_dims[3]));
other_arr = out;
} else if (out_dims != other_dims) {
// HACK: This is a quick check to see if other has been reordered inside gfor
// TODO: Figure out if this breaks and implement a cleaner method
other_arr = gforReorder(other_arr, dim);
is_reordered = true;
}
}
af_array par_arr = 0;
if (impl->is_linear_) {
AF_THROW(af_flat(&par_arr, impl->parent_->get()));
nd = 1;
} else {
par_arr = impl->parent_->get();
}
af_array tmp = 0;
AF_THROW(af_assign_gen(&tmp, par_arr, nd, impl->indices_, other_arr));
af_array res = 0;
if (impl->is_linear_) {
AF_THROW(af_moddims(&res, tmp, this_dims.ndims(), this_dims.get()));
AF_THROW(af_release_array(par_arr));
AF_THROW(af_release_array(tmp));
} else {
res = tmp;
}
impl->parent_->set(res);
if (dim >= 0 && (is_reordered || batch_assign)) {
if (other_arr) AF_THROW(af_release_array(other_arr));
}
return *this;
}
array constant(double val, const dim4 &dims, af_dtype type)
{
af_array res;
AF_THROW(af_constant(&res, val, dims.ndims(), dims.get(), type));
return array(res);
}
array iota(const dim4 &dims, const unsigned rep, af_dtype ty)
{
af_array out;
AF_THROW(af_iota(&out, dims.ndims(), dims.get(), rep, ty));
return array(out);
}
af_err af_approx1_uniform(af_array *yo, const af_array yi,
const af_array xo, const int xdim,
const double xi_beg, const double xi_step,
const af_interp_type method, const float offGrid)
{
try {
const ArrayInfo& yi_info = getInfo(yi);
const ArrayInfo& xo_info = getInfo(xo);
const dim4 yi_dims = yi_info.dims();
const dim4 xo_dims = xo_info.dims();
ARG_ASSERT(1, yi_info.isFloating()); // Only floating and complex types
ARG_ASSERT(2, xo_info.isRealFloating()) ; // Only floating types
ARG_ASSERT(1, yi_info.isSingle() == xo_info.isSingle()); // Must have same precision
ARG_ASSERT(1, yi_info.isDouble() == xo_info.isDouble()); // Must have same precision
ARG_ASSERT(3, xdim >= 0 && xdim < 4);
// POS should either be (x, 1, 1, 1) or (1, yi_dims[1], yi_dims[2], yi_dims[3])
if (xo_dims[xdim] != xo_dims.elements()) {
for (int i = 0; i < 4; i++) {
if (xdim != i) DIM_ASSERT(2, xo_dims[i] == yi_dims[i]);
}
}
ARG_ASSERT(5, xi_step != 0);
ARG_ASSERT(6, (method == AF_INTERP_CUBIC ||
method == AF_INTERP_CUBIC_SPLINE ||
method == AF_INTERP_LINEAR ||
method == AF_INTERP_LINEAR_COSINE ||
method == AF_INTERP_LOWER ||
method == AF_INTERP_NEAREST));
if (yi_dims.ndims() == 0 || xo_dims.ndims() == 0) {
*yo = createHandle(dim4(0,0,0,0), yi_info.getType());
return AF_SUCCESS;
}
dim4 yo_dims = yi_dims;
yo_dims[xdim] = xo_dims[xdim];
if (*yo == 0) {
*yo = createHandle(yo_dims, yi_info.getType());
}
DIM_ASSERT(1, getInfo(*yo).dims() == yo_dims);
switch(yi_info.getType()) {
case f32: approx1<float , float >(yo, yi, xo, xdim,
xi_beg, xi_step,
method, offGrid); break;
case f64: approx1<double , double>(yo, yi, xo, xdim,
xi_beg, xi_step,
method, offGrid); break;
case c32: approx1<cfloat , float >(yo, yi, xo, xdim,
xi_beg, xi_step,
method, offGrid); break;
case c64: approx1<cdouble, double>(yo, yi, xo, xdim,
xi_beg, xi_step,
method, offGrid); break;
default: TYPE_ERROR(1, yi_info.getType());
}
}
CATCHALL;
return AF_SUCCESS;
}
array randn(const dim4 &dims, const af::dtype type)
{
af_array res;
AF_THROW(af_randn(&res, dims.ndims(), dims.get(), type));
return array(res);
}
array moddims(const array& in, const dim4& dims)
{
return af::moddims(in, dims.ndims(), dims.get());
}
array identity(const dim4 &dims, const af::dtype type)
{
af_array res;
AF_THROW(af_identity(&res, dims.ndims(), dims.get(), type));
return array(res);
}
array randn(const dim4 &dims, const dtype ty, randomEngine &r)
{
af_array out;
AF_THROW(af_random_normal(&out, dims.ndims(), dims.get(), ty, r.get()));
return array(out);
}
array range(const dim4 &dims, const int seq_dim, const af::dtype ty)
{
af_array out;
AF_THROW(af_range(&out, dims.ndims(), dims.get(), seq_dim, ty));
return array(out);
}