void diff2(Param<T> out, CParam<T> in, int const dim)
{
af::dim4 dims = out.dims();
// Bool for dimension
bool is_dim0 = dim == 0;
bool is_dim1 = dim == 1;
bool is_dim2 = dim == 2;
bool is_dim3 = dim == 3;
T const * const inPtr = in.get();
T * outPtr = out.get();
// TODO: Improve this
for(dim_t l = 0; l < dims[3]; l++) {
for(dim_t k = 0; k < dims[2]; k++) {
for(dim_t j = 0; j < dims[1]; j++) {
for(dim_t i = 0; i < dims[0]; i++) {
// Operation: out[index] = in[index + 1 * dim_size] - in[index]
int idx = getIdx(in.strides(), i, j, k, l);
int jdx = getIdx(in.strides(),
i + is_dim0, j + is_dim1,
k + is_dim2, l + is_dim3);
int kdx = getIdx(in.strides(),
i + 2 * is_dim0, j + 2 * is_dim1,
k + 2 * is_dim2, l + 2 * is_dim3);
int odx = getIdx(out.strides(), i, j, k, l);
outPtr[odx] = inPtr[kdx] + inPtr[idx] - inPtr[jdx] - inPtr[jdx];
}
}
}
}
}
void transpose(Param<T> output, CParam<T> input) {
const dim4 odims = output.dims();
const dim4 ostrides = output.strides();
const dim4 istrides = input.strides();
T *out = output.get();
T const *const in = input.get();
for (dim_t l = 0; l < odims[3]; ++l) {
for (dim_t k = 0; k < odims[2]; ++k) {
// Outermost loop handles batch mode
// if input has no data along third dimension
// this loop runs only once
for (dim_t j = 0; j < odims[1]; ++j) {
for (dim_t i = 0; i < odims[0]; ++i) {
// calculate array indices based on offsets and strides
// the helper getIdx takes care of indices
const dim_t inIdx = getIdx(istrides, j, i, k, l);
const dim_t outIdx = getIdx(ostrides, i, j, k, l);
if (conjugate)
out[outIdx] = getConjugate(in[inIdx]);
else
out[outIdx] = in[inIdx];
}
}
// outData and inData pointers doesn't need to be
// offset as the getIdx function is taking care
// of the batch parameter
}
}
}
void transpose_inplace(Param<T> input) {
const dim4 idims = input.dims();
const dim4 istrides = input.strides();
T *in = input.get();
for (dim_t l = 0; l < idims[3]; ++l) {
for (dim_t k = 0; k < idims[2]; ++k) {
// Outermost loop handles batch mode
// if input has no data along third dimension
// this loop runs only once
//
// Run only bottom triangle. std::swap swaps with upper triangle
for (dim_t j = 0; j < idims[1]; ++j) {
for (dim_t i = j + 1; i < idims[0]; ++i) {
// calculate array indices based on offsets and strides
// the helper getIdx takes care of indices
const dim_t iIdx = getIdx(istrides, j, i, k, l);
const dim_t oIdx = getIdx(istrides, i, j, k, l);
if (conjugate) {
in[iIdx] = getConjugate(in[iIdx]);
in[oIdx] = getConjugate(in[oIdx]);
std::swap(in[iIdx], in[oIdx]);
} else {
std::swap(in[iIdx], in[oIdx]);
}
}
}
}
}
}
void wrap_dim(Param<T> out, CParam<T> in, const dim_t wx, const dim_t wy,
const dim_t sx, const dim_t sy, const dim_t px, const dim_t py) {
const T* inPtr = in.get();
T* outPtr = out.get();
af::dim4 idims = in.dims();
af::dim4 odims = out.dims();
af::dim4 istrides = in.strides();
af::dim4 ostrides = out.strides();
dim_t nx = (odims[0] + 2 * px - wx) / sx + 1;
for (dim_t w = 0; w < idims[3]; w++) {
for (dim_t z = 0; z < idims[2]; z++) {
dim_t cIn = w * istrides[3] + z * istrides[2];
dim_t cOut = w * ostrides[3] + z * ostrides[2];
const T* iptr_ = inPtr + cIn;
T* optr = outPtr + cOut;
for (dim_t col = 0; col < idims[d]; col++) {
// Offset output ptr
const T* iptr = iptr_ + col * istrides[d];
// Calculate input window index
dim_t winy = (col / nx);
dim_t winx = (col % nx);
dim_t startx = winx * sx;
dim_t starty = winy * sy;
dim_t spx = startx - px;
dim_t spy = starty - py;
// Short cut condition ensuring all values within input
// dimensions
bool cond = (spx >= 0 && spx + wx < odims[0] && spy >= 0 &&
spy + wy < odims[1]);
for (dim_t y = 0; y < wy; y++) {
for (dim_t x = 0; x < wx; x++) {
dim_t xpad = spx + x;
dim_t ypad = spy + y;
dim_t iloc = (y * wx + x);
if (d == 0) iloc *= istrides[1];
if (cond || (xpad >= 0 && xpad < odims[0] &&
ypad >= 0 && ypad < odims[1])) {
dim_t oloc =
(ypad * ostrides[1] + xpad * ostrides[0]);
// FIXME: When using threads, atomize this
optr[oloc] += iptr[iloc];
}
}
}
}
}
}
}
void bilateral(Param<OutT> out, CParam<InT> in, float const s_sigma, float const c_sigma)
{
af::dim4 const dims = in.dims();
af::dim4 const istrides = in.strides();
af::dim4 const ostrides = out.strides();
// clamp spatical and chromatic sigma's
float space_ = std::min(11.5f, std::max(s_sigma, 0.f));
float color_ = std::max(c_sigma, 0.f);
dim_t const radius = std::max((dim_t)(space_ * 1.5f), (dim_t)1);
float const svar = space_*space_;
float const cvar = color_*color_;
for(dim_t b3=0; b3<dims[3]; ++b3) {
OutT *outData = out.get() + b3 * ostrides[3];
InT const * inData = in.get() + b3 * istrides[3];
// b3 for loop handles following batch configurations
// - gfor
// - input based batch
// - when input is 4d array for color images
for(dim_t b2=0; b2<dims[2]; ++b2) {
// b2 for loop handles following batch configurations
// - channels
// - input based batch
// - when input is 3d array for grayscale images
for(dim_t j=0; j<dims[1]; ++j) {
// j steps along 2nd dimension
for(dim_t i=0; i<dims[0]; ++i) {
// i steps along 1st dimension
OutT norm = 0.0;
OutT res = 0.0;
OutT const center = (OutT)inData[getIdx(istrides, i, j)];
for(dim_t wj=-radius; wj<=radius; ++wj) {
// clamps offsets
dim_t tj = clamp(j+wj, dim_t(0), dims[1]-1);
for(dim_t wi=-radius; wi<=radius; ++wi) {
// clamps offsets
dim_t ti = clamp(i+wi, dim_t(0), dims[0]-1);
// proceed
OutT const val= (OutT)inData[getIdx(istrides, ti, tj)];
OutT const gauss_space = (wi*wi+wj*wj)/(-2.0*svar);
OutT const gauss_range = ((center-val)*(center-val))/(-2.0*cvar);
OutT const weight = std::exp(gauss_space+gauss_range);
norm += weight;
res += val*weight;
}
} // filter loop ends here
outData[getIdx(ostrides, i, j)] = res/norm;
} //1st dimension loop ends here
} //2nd dimension loop ends here
outData += ostrides[2];
inData += istrides[2];
}
}
}
void select(Param<T> out, CParam<char> cond, CParam<T> a, CParam<T> b) {
af::dim4 adims = a.dims();
af::dim4 astrides = a.strides();
af::dim4 bdims = b.dims();
af::dim4 bstrides = b.strides();
af::dim4 cdims = cond.dims();
af::dim4 cstrides = cond.strides();
af::dim4 odims = out.dims();
af::dim4 ostrides = out.strides();
bool is_a_same[] = {adims[0] == odims[0], adims[1] == odims[1],
adims[2] == odims[2], adims[3] == odims[3]};
bool is_b_same[] = {bdims[0] == odims[0], bdims[1] == odims[1],
bdims[2] == odims[2], bdims[3] == odims[3]};
bool is_c_same[] = {cdims[0] == odims[0], cdims[1] == odims[1],
cdims[2] == odims[2], cdims[3] == odims[3]};
const T *aptr = a.get();
const T *bptr = b.get();
T *optr = out.get();
const char *cptr = cond.get();
for (int l = 0; l < odims[3]; l++) {
int o_off3 = ostrides[3] * l;
int a_off3 = astrides[3] * is_a_same[3] * l;
int b_off3 = bstrides[3] * is_b_same[3] * l;
int c_off3 = cstrides[3] * is_c_same[3] * l;
for (int k = 0; k < odims[2]; k++) {
int o_off2 = ostrides[2] * k + o_off3;
int a_off2 = astrides[2] * is_a_same[2] * k + a_off3;
int b_off2 = bstrides[2] * is_b_same[2] * k + b_off3;
int c_off2 = cstrides[2] * is_c_same[2] * k + c_off3;
for (int j = 0; j < odims[1]; j++) {
int o_off1 = ostrides[1] * j + o_off2;
int a_off1 = astrides[1] * is_a_same[1] * j + a_off2;
int b_off1 = bstrides[1] * is_b_same[1] * j + b_off2;
int c_off1 = cstrides[1] * is_c_same[1] * j + c_off2;
for (int i = 0; i < odims[0]; i++) {
bool cval = is_c_same[0] ? cptr[c_off1 + i] : cptr[c_off1];
T aval = is_a_same[0] ? aptr[a_off1 + i] : aptr[a_off1];
T bval = is_b_same[0] ? bptr[b_off1 + i] : bptr[b_off1];
T oval = cval ? aval : bval;
optr[o_off1 + i] = oval;
}
}
}
}
}
void diagExtract(Param<T> out, CParam<T> in, int const num)
{
af::dim4 const odims = out.dims();
af::dim4 const idims = in.dims();
int const i_off = (num > 0) ? (num * in.strides(1)) : (-num);
for (int l = 0; l < (int)odims[3]; l++) {
for (int k = 0; k < (int)odims[2]; k++) {
const T *iptr = in.get() + l * in.strides(3) + k * in.strides(2) + i_off;
T *optr = out.get() + l * out.strides(3) + k * out.strides(2);
for (int i = 0; i < (int)odims[0]; i++) {
T val = scalar<T>(0);
if (i < idims[0] && i < idims[1]) val = iptr[i * in.strides(1) + i];
optr[i] = val;
}
}
}
}
void approx2(Param<InT> output, CParam<InT> input,
CParam<LocT> xposition, CParam<LocT> yposition,
float const offGrid, af_interp_type method)
{
InT * out = output.get();
const LocT *xpos = xposition.get();
const LocT *ypos = yposition.get();
af::dim4 const odims = output.dims();
af::dim4 const idims = input.dims();
af::dim4 const xdims = xposition.dims();
af::dim4 const ostrides = output.strides();
af::dim4 const istrides = input.strides();
af::dim4 const xstrides = xposition.strides();
af::dim4 const ystrides = yposition.strides();
Interp2<InT, LocT, order> interp;
bool batch = !(xdims[2] == 1 && xdims[3] == 1);
for(dim_t idw = 0; idw < odims[3]; idw++) {
for(dim_t idz = 0; idz < odims[2]; idz++) {
dim_t xoffzw = idw * xstrides[3] + idz * xstrides[2];
dim_t yoffzw = idw * ystrides[3] + idz * ystrides[2];
dim_t ooffzw = idw * ostrides[3] + idz * ostrides[2];
dim_t ioffzw = idw * istrides[3] + idz * istrides[2];
for(dim_t idy = 0; idy < odims[1]; idy++) {
dim_t xoff = xoffzw * batch + idy * xstrides[1];
dim_t yoff = yoffzw * batch + idy * ystrides[1];
dim_t ooff = ooffzw + idy * ostrides[1];
for(dim_t idx = 0; idx < odims[0]; idx++) {
const LocT x = xpos[xoff + idx];
const LocT y = ypos[yoff + idx];
// FIXME: Only cubic interpolation is doing clamping
// We need to make it consistent across all methods
// Not changing the behavior because tests will fail
bool clamp = order == 3;
if (x < 0 || idims[0] < x + 1 ||
y < 0 || idims[1] < y + 1 ) {
out[ooff + idx] = scalar<InT>(offGrid);
} else {
interp(output, ooff + idx, input, ioffzw, x, y, method, 1, clamp);
}
}
}
}
}
}
void diagCreate(Param<T> out, CParam<T> in, int const num)
{
int batch = in.dims(1);
int size = out.dims(0);
T const * iptr = in.get();
T * optr = out.get();
for (int k = 0; k < batch; k++) {
for (int j = 0; j < size; j++) {
for (int i = 0; i < size; i++) {
T val = scalar<T>(0);
if (i == j - num) {
val = (num > 0) ? iptr[i] : iptr[j];
}
optr[i + j * out.strides(1)] = val;
}
}
optr += out.strides(2);
iptr += in.strides(1);
}
}
void copyElemwise(Param<OutT> dst, CParam<InT> src, OutT default_value, double factor)
{
af::dim4 src_dims = src.dims();
af::dim4 dst_dims = dst.dims();
af::dim4 src_strides = src.strides();
af::dim4 dst_strides = dst.strides();
InT const * const src_ptr = src.get();
OutT * dst_ptr = dst.get();
dim_t trgt_l = std::min(dst_dims[3], src_dims[3]);
dim_t trgt_k = std::min(dst_dims[2], src_dims[2]);
dim_t trgt_j = std::min(dst_dims[1], src_dims[1]);
dim_t trgt_i = std::min(dst_dims[0], src_dims[0]);
for(dim_t l=0; l<dst_dims[3]; ++l) {
dim_t src_loff = l*src_strides[3];
dim_t dst_loff = l*dst_strides[3];
bool isLvalid = l<trgt_l;
for(dim_t k=0; k<dst_dims[2]; ++k) {
dim_t src_koff = k*src_strides[2];
dim_t dst_koff = k*dst_strides[2];
bool isKvalid = k<trgt_k;
for(dim_t j=0; j<dst_dims[1]; ++j) {
dim_t src_joff = j*src_strides[1];
dim_t dst_joff = j*dst_strides[1];
bool isJvalid = j<trgt_j;
for(dim_t i=0; i<dst_dims[0]; ++i) {
OutT temp = default_value;
if (isLvalid && isKvalid && isJvalid && i<trgt_i) {
dim_t src_idx = i*src_strides[0] + src_joff + src_koff + src_loff;
temp = OutT(src_ptr[src_idx])*OutT(factor);
}
dim_t dst_idx = i*dst_strides[0] + dst_joff + dst_koff + dst_loff;
dst_ptr[dst_idx] = temp;
}
}
}
}
}
void join(const int dim, Param<T> out, const std::vector<CParam<T>> inputs)
{
af::dim4 zero(0,0,0,0);
af::dim4 d = zero;
switch(dim) {
case 0:
join_append<T, T, 0>(out.get(), inputs[0].get(), zero,
out.dims(), inputs[0].dims(), out.strides(), inputs[0].strides());
for(int i = 1; i < n_arrays; i++) {
d += inputs[i - 1].dims();
join_append<T, T, 0>(out.get(), inputs[i].get(), calcOffset<0>(d),
out.dims(), inputs[i].dims(), out.strides(), inputs[i].strides());
}
break;
case 1:
join_append<T, T, 1>(out.get(), inputs[0].get(), zero,
out.dims(), inputs[0].dims(), out.strides(), inputs[0].strides());
for(int i = 1; i < n_arrays; i++) {
d += inputs[i - 1].dims();
join_append<T, T, 1>(out.get(), inputs[i].get(), calcOffset<1>(d),
out.dims(), inputs[i].dims(), out.strides(), inputs[i].strides());
}
break;
case 2:
join_append<T, T, 2>(out.get(), inputs[0].get(), zero,
out.dims(), inputs[0].dims(), out.strides(), inputs[0].strides());
for(int i = 1; i < n_arrays; i++) {
d += inputs[i - 1].dims();
join_append<T, T, 2>(out.get(), inputs[i].get(), calcOffset<2>(d),
out.dims(), inputs[i].dims(), out.strides(), inputs[i].strides());
}
break;
case 3:
join_append<T, T, 3>(out.get(), inputs[0].get(), zero,
out.dims(), inputs[0].dims(), out.strides(), inputs[0].strides());
for(int i = 1; i < n_arrays; i++) {
d += inputs[i - 1].dims();
join_append<T, T, 3>(out.get(), inputs[i].get(), calcOffset<3>(d),
out.dims(), inputs[i].dims(), out.strides(), inputs[i].strides());
}
break;
}
}
void select_scalar(Param<T> out, CParam<char> cond, CParam<T> a,
const double b) {
af::dim4 astrides = a.strides();
af::dim4 adims = a.dims();
af::dim4 cstrides = cond.strides();
af::dim4 cdims = cond.dims();
af::dim4 odims = out.dims();
af::dim4 ostrides = out.strides();
const T *aptr = a.get();
T *optr = out.get();
const char *cptr = cond.get();
bool is_a_same[] = {adims[0] == odims[0], adims[1] == odims[1],
adims[2] == odims[2], adims[3] == odims[3]};
bool is_c_same[] = {cdims[0] == odims[0], cdims[1] == odims[1],
cdims[2] == odims[2], cdims[3] == odims[3]};
for (int l = 0; l < odims[3]; l++) {
int o_off3 = ostrides[3] * l;
int a_off3 = astrides[3] * is_a_same[3] * l;
int c_off3 = cstrides[3] * is_c_same[3] * l;
for (int k = 0; k < odims[2]; k++) {
int o_off2 = ostrides[2] * k + o_off3;
int a_off2 = astrides[2] * is_a_same[2] * k + a_off3;
int c_off2 = cstrides[2] * is_c_same[2] * k + c_off3;
for (int j = 0; j < odims[1]; j++) {
int o_off1 = ostrides[1] * j + o_off2;
int a_off1 = astrides[1] * is_a_same[1] * j + a_off2;
int c_off1 = cstrides[1] * is_c_same[1] * j + c_off2;
for (int i = 0; i < odims[0]; i++) {
bool cval = is_c_same[0] ? cptr[c_off1 + i] : cptr[c_off1];
T aval = is_a_same[0] ? aptr[a_off1 + i] : aptr[a_off1];
optr[o_off1 + i] = (flip ^ cval) ? aval : b;
}
}
}
}
}
void lookup(Param<InT> out, CParam<InT> input, CParam<IndexT> indices,
unsigned const dim) {
const af::dim4 iDims = input.dims();
const af::dim4 oDims = out.dims();
const af::dim4 iStrides = input.strides();
const af::dim4 oStrides = out.strides();
const InT *inPtr = input.get();
const IndexT *idxPtr = indices.get();
InT *outPtr = out.get();
for (dim_t l = 0; l < oDims[3]; ++l) {
dim_t iLOff = iStrides[3] *
(dim == 3 ? trimIndex((dim_t)idxPtr[l], iDims[3]) : l);
dim_t oLOff = l * oStrides[3];
for (dim_t k = 0; k < oDims[2]; ++k) {
dim_t iKOff =
iStrides[2] *
(dim == 2 ? trimIndex((dim_t)idxPtr[k], iDims[2]) : k);
dim_t oKOff = k * oStrides[2];
for (dim_t j = 0; j < oDims[1]; ++j) {
dim_t iJOff =
iStrides[1] *
(dim == 1 ? trimIndex((dim_t)idxPtr[j], iDims[1]) : j);
dim_t oJOff = j * oStrides[1];
for (dim_t i = 0; i < oDims[0]; ++i) {
dim_t iIOff =
iStrides[0] *
(dim == 0 ? trimIndex((dim_t)idxPtr[i], iDims[0]) : i);
dim_t oIOff = i * oStrides[0];
outPtr[oLOff + oKOff + oJOff + oIOff] =
inPtr[iLOff + iKOff + iJOff + iIOff];
}
}
}
}
}
void join(Param<Tx> out, const int dim, CParam<Tx> first, CParam<Ty> second)
{
Tx* outPtr = out.get();
const Tx* fptr = first.get();
const Ty* sptr = second.get();
af::dim4 zero(0,0,0,0);
const af::dim4 odims = out.dims();
const af::dim4 fdims = first.dims();
const af::dim4 sdims = second.dims();
switch(dim) {
case 0:
join_append<Tx, Tx, 0>(outPtr, fptr, zero,
odims, fdims, out.strides(), first.strides());
join_append<Tx, Ty, 0>(outPtr, sptr, calcOffset<0>(fdims),
odims, sdims, out.strides(), second.strides());
break;
case 1:
join_append<Tx, Tx, 1>(outPtr, fptr, zero,
odims, fdims, out.strides(), first.strides());
join_append<Tx, Ty, 1>(outPtr, sptr, calcOffset<1>(fdims),
odims, sdims, out.strides(), second.strides());
break;
case 2:
join_append<Tx, Tx, 2>(outPtr, fptr, zero,
odims, fdims, out.strides(), first.strides());
join_append<Tx, Ty, 2>(outPtr, sptr, calcOffset<2>(fdims),
odims, sdims, out.strides(), second.strides());
break;
case 3:
join_append<Tx, Tx, 3>(outPtr, fptr, zero,
odims, fdims, out.strides(), first.strides());
join_append<Tx, Ty, 3>(outPtr, sptr, calcOffset<3>(fdims),
odims, sdims, out.strides(), second.strides());
break;
}
}
void fft_c2r(Param<Tr> out, const af::dim4 oDataDims,
CParam<Tc> in, const af::dim4 iDataDims,
const af::dim4 odims)
{
int t_dims[rank];
int in_embed[rank];
int out_embed[rank];
computeDims<rank>(t_dims , odims);
computeDims<rank>(in_embed , iDataDims);
computeDims<rank>(out_embed , oDataDims);
const af::dim4 istrides = in.strides();
const af::dim4 ostrides = out.strides();
typedef typename fftw_real_transform<Tr, Tc>::ctype_t ctype_t;
typename fftw_real_transform<Tr, Tc>::plan_t plan;
fftw_real_transform<Tr, Tc> transform;
int batch = 1;
for (int i = rank; i < 4; i++) {
batch *= odims[i];
}
plan = transform.create(rank,
t_dims,
(int)batch,
(ctype_t *)in.get(),
in_embed, (int)istrides[0],
(int)istrides[rank],
(Tr *)out.get(),
out_embed, (int)ostrides[0],
(int)ostrides[rank],
FFTW_ESTIMATE);
transform.execute(plan);
transform.destroy(plan);
}
static void copy(Param<T> dst, CParam<T> src)
{
af::dim4 src_dims = src.dims();
af::dim4 dst_dims = dst.dims();
af::dim4 src_strides = src.strides();
af::dim4 dst_strides = dst.strides();
T const * src_ptr = src.get();
T * dst_ptr = dst.get();
// find the major-most dimension, which is linear in both arrays
int linear_end = 0;
dim_t count = 1;
while (linear_end < 4
&& count == src_strides[linear_end]
&& count == dst_strides[linear_end]) {
count *= src_dims[linear_end];
++linear_end;
}
// traverse through the array using strides only until neccessary
copy_go(dst_ptr, dst_strides, dst_dims, src_ptr, src_strides, src_dims, 3, linear_end);
}
void fft_inplace(Param<T> in, const af::dim4 iDataDims)
{
int t_dims[rank];
int in_embed[rank];
const af::dim4 idims = in.dims();
computeDims<rank>(t_dims , idims);
computeDims<rank>(in_embed , iDataDims);
const af::dim4 istrides = in.strides();
typedef typename fftw_transform<T>::ctype_t ctype_t;
typename fftw_transform<T>::plan_t plan;
fftw_transform<T> transform;
int batch = 1;
for (int i = rank; i < 4; i++) {
batch *= idims[i];
}
plan = transform.create(rank,
t_dims,
(int)batch,
(ctype_t *)in.get(),
in_embed, (int)istrides[0],
(int)istrides[rank],
(ctype_t *)in.get(),
in_embed, (int)istrides[0],
(int)istrides[rank],
direction ? FFTW_FORWARD : FFTW_BACKWARD,
FFTW_ESTIMATE);
transform.execute(plan);
transform.destroy(plan);
}
void padBorders(Param<T> out, CParam<T> in, const dim4 lBoundPadSize,
const dim4 uBoundPadSize, const af::borderType btype) {
const dim4& oDims = out.dims();
const dim4& oStrs = out.strides();
const dim4& iDims = in.dims();
const dim4& iStrs = in.strides();
T const* const src = in.get();
T* dst = out.get();
const dim4 validRegEnds(
oDims[0] - uBoundPadSize[0], oDims[1] - uBoundPadSize[1],
oDims[2] - uBoundPadSize[2], oDims[3] - uBoundPadSize[3]);
const bool isInputLinear = iStrs[0] == 1;
/*
* VALID REGION COPYING DOES
* NOT NEED ANY BOUND CHECKS
* */
for (dim_t l = lBoundPadSize[3]; l < validRegEnds[3]; ++l) {
dim_t oLOff = oStrs[3] * l;
dim_t iLOff = iStrs[3] * (l - lBoundPadSize[3]);
for (dim_t k = lBoundPadSize[2]; k < validRegEnds[2]; ++k) {
dim_t oKOff = oStrs[2] * k;
dim_t iKOff = iStrs[2] * (k - lBoundPadSize[2]);
for (dim_t j = lBoundPadSize[1]; j < validRegEnds[1]; ++j) {
dim_t oJOff = oStrs[1] * j;
dim_t iJOff = iStrs[1] * (j - lBoundPadSize[1]);
if (isInputLinear) {
T const* const sptr = src + iLOff + iKOff + iJOff;
T* dptr = dst + oLOff + oKOff + oJOff + lBoundPadSize[0];
std::copy(sptr, sptr + iDims[0], dptr);
} else {
for (dim_t i = lBoundPadSize[0]; i < validRegEnds[0]; ++i) {
dim_t oIOff = oStrs[0] * i;
dim_t iIOff = iStrs[0] * (i - lBoundPadSize[0]);
dst[oLOff + oKOff + oJOff + oIOff] =
src[iLOff + iKOff + iJOff + iIOff];
}
}
} // second dimension loop
} // third dimension loop
} // fourth dimension loop
// If we have to do zero padding,
// just return as the output is filled with
// zeros during allocation
if (btype == AF_PAD_ZERO) return;
/*
* PADDED REGIONS NEED BOUND
* CHECKS; FOLLOWING NESTED
* LOOPS SHALL ONLY PROCESS
* PADDED REGIONS AND SKIP REST
* */
for (dim_t l = 0; l < oDims[3]; ++l) {
bool skipL = (l >= lBoundPadSize[3] && l < validRegEnds[3]);
dim_t oLOff = oStrs[3] * l;
dim_t iLOff =
iStrs[3] * idxByndEdge(l, lBoundPadSize[3], iDims[3], btype);
for (dim_t k = 0; k < oDims[2]; ++k) {
bool skipK = (k >= lBoundPadSize[2] && k < validRegEnds[2]);
dim_t oKOff = oStrs[2] * k;
dim_t iKOff =
iStrs[2] * idxByndEdge(k, lBoundPadSize[2], iDims[2], btype);
for (dim_t j = 0; j < oDims[1]; ++j) {
bool skipJ = (j >= lBoundPadSize[1] && j < validRegEnds[1]);
dim_t oJOff = oStrs[1] * j;
dim_t iJOff = iStrs[1] *
idxByndEdge(j, lBoundPadSize[1], iDims[1], btype);
for (dim_t i = 0; i < oDims[0]; ++i) {
bool skipI = (i >= lBoundPadSize[0] && i < validRegEnds[0]);
if (skipI && skipJ && skipK && skipL) continue;
dim_t oIOff = oStrs[0] * i;
dim_t iIOff = iStrs[0] * idxByndEdge(i, lBoundPadSize[0],
iDims[0], btype);
dst[oLOff + oKOff + oJOff + oIOff] =
src[iLOff + iKOff + iJOff + iIOff];
} // first dimension loop
} // second dimension loop
} // third dimension loop
} // fourth dimension loop
}
void transform(Param<T> output, CParam<T> input, CParam<float> transform,
const bool inverse, const bool perspective,
af_interp_type method) {
typedef typename af::dtype_traits<T>::base_type BT;
typedef wtype_t<BT> WT;
const af::dim4 idims = input.dims();
const af::dim4 odims = output.dims();
const af::dim4 tdims = transform.dims();
const af::dim4 tstrides = transform.strides();
const af::dim4 istrides = input.strides();
const af::dim4 ostrides = output.strides();
T *out = output.get();
const float *tf = transform.get();
int batch_size = 1;
if (idims[2] != tdims[2]) batch_size = idims[2];
Interp2<T, WT, order> interp;
for (int idw = 0; idw < (int)odims[3]; idw++) {
dim_t out_offw = idw * ostrides[3];
dim_t in_offw = (idims[3] > 1) * idw * istrides[3];
dim_t tf_offw = (tdims[3] > 1) * idw * tstrides[3];
for (int idz = 0; idz < (int)odims[2]; idz += batch_size) {
dim_t out_offzw = out_offw + idz * ostrides[2];
dim_t in_offzw = in_offw + (idims[2] > 1) * idz * istrides[2];
dim_t tf_offzw = tf_offw + (tdims[2] > 1) * idz * tstrides[2];
const float *tptr = tf + tf_offzw;
float tmat[9];
calc_transform_inverse(tmat, tptr, inverse, perspective,
perspective ? 9 : 6);
for (int idy = 0; idy < (int)odims[1]; idy++) {
for (int idx = 0; idx < (int)odims[0]; idx++) {
WT xidi = idx * tmat[0] + idy * tmat[1] + tmat[2];
WT yidi = idx * tmat[3] + idy * tmat[4] + tmat[5];
if (perspective) {
WT W = idx * tmat[6] + idy * tmat[7] + tmat[8];
xidi /= W;
yidi /= W;
}
// FIXME: Nearest and lower do not do clamping, but other
// methods do Make it consistent
bool clamp = order != 1;
bool condX = xidi >= -0.0001 && xidi < idims[0];
bool condY = yidi >= -0.0001 && yidi < idims[1];
int ooff = out_offzw + idy * ostrides[1] + idx;
if (condX && condY) {
interp(output, ooff, input, in_offzw, xidi, yidi,
method, batch_size, clamp);
} else {
for (int n = 0; n < batch_size; n++) {
out[ooff + n * ostrides[2]] = scalar<T>(0);
}
}
}
}
}
}
}