at::Tensor nms_cuda(const at::Tensor input,
float thresh)
{
AT_CHECK(input.ndimension() == 3,
"First argument should be a 3D Tensor, (batch_sz x n_boxes x 4)");
// AT_CHECK(scores.ndimens/ion() == 2,
// "Second argument should be a 2D Tensor, (batch_sz x n_boxes)");
// AT_CHECK(input.size(0) == scores.size(0),
// "First and second arguments must have equal-sized first dimensions");
// AT_CHECK(input.size(1) == scores.size(1),
// "First and second arguments must have equal-sized second dimensions");
AT_CHECK(input.size(2) == 4,
"First argument dimension 2 must have size 4, and should be of the form [x, y, w, h]");
AT_CHECK(input.is_contiguous(), "First argument must be a contiguous Tensor");
// AT_CHECK(scores.is_contiguous(), "Second argument must be a contiguous Tensor");
AT_CHECK(input.type().scalarType() == at::kFloat || input.type().scalarType() == at::kDouble,
"First argument must be Float or Double Tensor");
// AT_CHECK(scores.type().scalarType() == at::kFloat || scores.type().scalarType() == at::kDouble,
// "Second argument must be Float or Double Tensor");
AT_CHECK(input.is_contiguous(), "First argument must be a contiguous Tensor");
// AT_CHECK(scores.is_contiguous(), "Second argument must be a contiguous Tensor");
return non_max_suppression_cuda(input, thresh);
}
at::Tensor sigmoid_add(at::Tensor x, at::Tensor y) {
AT_CHECK(x.type().is_cuda(), "x must be a CUDA tensor");
AT_CHECK(y.type().is_cuda(), "y must be a CUDA tensor");
auto output = at::zeros_like(x);
sigmoid_add_cuda(
x.data<float>(), y.data<float>(), output.data<float>(), output.numel());
return output;
}
size_t ReplaceAll(std::string& s, const char* from, const char* to) {
AT_CHECK(from && *from, "");
AT_CHECK(to, "");
size_t numReplaced = 0;
std::string::size_type lenFrom = std::strlen(from);
std::string::size_type lenTo = std::strlen(to);
for (auto pos = s.find(from); pos != std::string::npos;
pos = s.find(from, pos + lenTo)) {
s.replace(pos, lenFrom, to);
numReplaced++;
}
return numReplaced;
}
Tensor PerTensorAffineQuantizer::quantize(Tensor tensor) {
IntArrayRef sizes = tensor.sizes();
// Here we need a std::intrusive_ptr<Quantizer>.. but actually "this" is the
// quantizer that can be reused, so I'm using intrusive_from_this here
AT_CHECK(
tensor.options().device() == kCPU,
"quantize only works for CPU backend right now.");
Tensor qv = new_qtensor_cpu(
sizes,
tensor.options().dtype(at::kQInt8),
intrusive_from_this());
tensor = tensor.contiguous();
const float* svd = tensor.data<float>();
#ifdef USE_FBGEMM
auto qvd = reinterpret_cast<uint8_t*>(qv.data<qint8>());
fbgemm::TensorQuantizationParams qparams;
qparams.scale = scale_;
qparams.zero_point = zero_point_;
qparams.precision = 8;
fbgemm::Quantize<uint8_t>(/*src=*/svd,
/*dst=*/qvd,
/*len=*/tensor.numel(),
/*qparams=*/qparams);
#else
auto qvd = qv.data<qint8>();
for (int i = 0; i < tensor.numel(); ++i) {
qvd[i] = quantize_uint8(scale_, zero_point_, svd[i]);
}
#endif
return qv;
}
Tensor Type::copy(const Tensor & src, bool non_blocking) const {
AT_CHECK(src.defined(), "attempt to copy an undefined tensor");
if (is_sparse()) {
auto indices = src._indices();
auto values = src._values();
auto & this_dense = toBackend(is_cuda() ? Backend::CUDA : Backend::CPU);
auto & this_dense_idx = this_dense.toScalarType(ScalarType::Long);
auto indices_copy = this_dense_idx.copy(indices, non_blocking);
auto values_copy = this_dense.copy(values, non_blocking);
return _sparse_coo_tensor_unsafe(indices_copy, values_copy, src.sizes());
} else {
Tensor r = this->tensor(src.sizes());
r.copy_(src, non_blocking);
return r;
}
}
inline Tensor new_qtensor_cpu(
IntArrayRef sizes,
const TensorOptions& options,
QuantizerPtr quantizer) {
AT_ASSERT(options.device().is_cpu());
native::check_size_nonnegative(sizes);
auto* allocator = at::getCPUAllocator();
int64_t nelements = at::prod_intlist(sizes);
auto dtype = options.dtype();
AT_CHECK(isQIntType(typeMetaToScalarType(dtype)),
"ScalarType is not supported in new_qtensor_cpu.");
auto storage = c10::make_intrusive<StorageImpl>(
dtype,
nelements,
allocator->allocate(nelements * dtype.itemsize()),
allocator,
/*resizable=*/true);
auto tensor = detail::make_tensor<QTensorImpl>(
storage, at::QuantizedCPUTensorId(), quantizer);
get_qtensorimpl(tensor)->set_sizes_contiguous(sizes);
return tensor;
}
void THNN_(SpatialMaxUnpooling_updateOutput)(
THNNState *state,
THTensor *input,
THTensor *output,
THIndexTensor *indices,
int owidth, int oheight)
{
int dimw = 2;
int dimh = 1;
int nbatch = 1;
int nslices;
int iheight;
int iwidth;
scalar_t *input_data;
scalar_t *output_data;
THIndex_t *indices_data;
AT_CHECK(!input->is_empty() && (input->dim() == 3 || input->dim() == 4),
"non-empty 3D or 4D (batch mode) tensor expected for input, but got sizes: ", input->sizes());
THNN_CHECK_SHAPE_INDICES(input, indices);
if (input->dim() == 4)
{
nbatch = input->size(0);
dimw++;
dimh++;
}
/* sizes */
nslices = input->size(dimh-1);
iheight = input->size(dimh);
iwidth = input->size(dimw);
/* get contiguous input and indices */
input = THTensor_(newContiguous)(input);
indices = THIndexTensor_(newContiguous)(indices);
/* resize output */
if (input->dim() == 3)
{
THTensor_(resize3d)(output, nslices, oheight, owidth);
THTensor_(zero)(output);
input_data = input->data<scalar_t>();
output_data = output->data<scalar_t>();
indices_data = THIndexTensor_(data)(indices);
THNN_(SpatialMaxUnpooling_updateOutput_frame)(input_data, output_data,
indices_data,
nslices,
iwidth, iheight,
owidth, oheight);
}
else
{
int p;
THTensor_(resize4d)(output, nbatch, nslices, oheight, owidth);
THTensor_(zero)(output);
input_data = input->data<scalar_t>();
output_data = output->data<scalar_t>();
indices_data = THIndexTensor_(data)(indices);
for (p = 0; p < nbatch; p++)
{
THNN_(SpatialMaxUnpooling_updateOutput_frame)(
input_data+p*nslices*iwidth*iheight,
output_data+p*nslices*owidth*oheight,
indices_data+p*nslices*iwidth*iheight,
nslices,
iwidth, iheight,
owidth, oheight);
}
}
/* cleanup */
c10::raw::intrusive_ptr::decref(input);
THIndexTensor_(free)(indices);
}