#include "op_plugin/AclOpsInterface.h"
#include "op_plugin/utils/OpAdapter.h"
namespace acl_op {
using npu_preparation = at_npu::native::OpPreparation;
using npu_utils = at_npu::native::NpuUtils;
namespace {
c10::SmallVector<int64_t, SIZE> upsample_nearest1d_backward_infer_size(at::IntArrayRef input_size)
{
TORCH_CHECK(
input_size.size() == 3,
"It is expected input_size equals to 3, but got size ",
input_size.size(), OPS_ERROR(ErrCode::PARAM));
c10::SmallVector<int64_t, SIZE> output_size;
int64_t N = input_size[0];
int64_t C = input_size[1];
int64_t W = input_size[2];
output_size = {N, C, 1, W};
return output_size;
}
at::Tensor& upsample_nearest1d_backward_out_nocheck(
at::Tensor& grad_input,
const at::Tensor& grad_output,
at::IntArrayRef output_size,
at::IntArrayRef input_size,
c10::optional<double> scales)
{
at::Tensor grad_cp = grad_output.unsqueeze(2);
at_npu::native::OpCommand cmd;
if (grad_output.scalar_type() == at::kFloat || grad_output.scalar_type() == at::kHalf) {
c10::SmallVector<int64_t, SIZE> result_size = {1, input_size[2]};
cmd.Name("ResizeNearestNeighborV2Grad")
.Input(grad_cp)
.Input(result_size, at::kInt)
.Output(grad_input)
.Attr("align_corners", false)
.Attr("half_pixel_centers", false)
.Run();
} else {
TORCH_CHECK(output_size[0] != 0, "output_size must not equals to 0, but got ", output_size[0],
OPS_ERROR(ErrCode::PARAM));
c10::SmallVector<int64_t, SIZE> origin_size = upsample_nearest1d_backward_infer_size(input_size);
at::Scalar scales_cp = scales.has_value() ? scales.value() : -1;
cmd.Name("ResizeGrad")
.Input(grad_cp)
.Input(scales_cp, at::kFloat)
.Input(scales_cp, at::kFloat)
.Input(origin_size, at::kLong)
.Output(grad_input)
.Attr("coordinate_transformation_mode", (string)"pytorch_half_pixel")
.Attr("cubic_coeff_a", (float)-0.75)
.Attr("exclude_outside", (int64_t)0)
.Attr("extrapolation_value", (float)0.0)
.Attr("mode", (string)"nearest")
.Attr("nearest_mode", (string)"floor")
.Run();
}
grad_input = grad_input.squeeze(2);
return grad_input;
}
}
at::Tensor& upsample_nearest1d_backward_out(
const at::Tensor& grad_output,
at::IntArrayRef output_size,
at::IntArrayRef input_size,
c10::optional<double> scales,
at::Tensor& grad_input)
{
c10::SmallVector<int64_t, SIZE> op_infer_output_size = upsample_nearest1d_backward_infer_size(input_size);
npu_preparation::CheckOut(
{grad_output},
grad_input,
grad_output,
op_infer_output_size);
if (!npu_utils::check_match(&grad_input)) {
at::Tensor contiguous_result = npu_utils::format_contiguous(grad_input);
upsample_nearest1d_backward_out_nocheck(contiguous_result, grad_output, output_size, input_size, scales);
npu_utils::format_fresh_view(grad_input, contiguous_result);
} else {
upsample_nearest1d_backward_out_nocheck(grad_input, grad_output, output_size, input_size, scales);
}
return grad_input;
}
at::Tensor upsample_nearest1d_backward(
const at::Tensor& grad_output,
at::IntArrayRef output_size,
at::IntArrayRef input_size,
c10::optional<double> scales)
{
c10::SmallVector<int64_t, SIZE> op_infer_output_size = upsample_nearest1d_backward_infer_size(input_size);
at::Tensor grad_input = npu_preparation::apply_tensor(grad_output, op_infer_output_size);
upsample_nearest1d_backward_out_nocheck(
grad_input, grad_output, output_size, input_size, scales);
return grad_input;
}
}