#include <ATen/native/Pool.h>
#include "op_plugin/AclOpsInterface.h"
#include "op_plugin/utils/OpAdapter.h"
#include "torch_npu/csrc/framework/utils/UtilForOpAdapter.h"
namespace acl_op {
static const int64_t INDICES_TYPE_CONVERT = 2;
static const int64_t BLOCKSIZE = 16;
using npu_preparation = at_npu::native::OpPreparation;
using npu_utils = at_npu::native::NpuUtils;
namespace {
at::Tensor &max_pool2d_with_indices_backward_out_nocheck(at::Tensor &grad_input, const at::Tensor &grad_output,
const at::Tensor &self, at::IntArrayRef kernel_size,
at::IntArrayRef stride, at::IntArrayRef padding,
at::IntArrayRef dilation, bool ceil_mode,
const at::Tensor &indices)
{
at::Tensor self_cp = self;
at::Tensor grad_output_cp = grad_output;
at::Tensor indices_cp = indices;
if (self.dim() == 3) {
self_cp = self.unsqueeze(0);
grad_output_cp = grad_output.unsqueeze(0);
indices_cp = indices.unsqueeze(0);
grad_input.unsqueeze_(0);
}
int64_t stride_H = 1;
int64_t stride_W = 1;
if (stride.empty()) {
stride_H = kernel_size[0];
stride_W = kernel_size[1];
} else {
stride_H = stride[0];
stride_W = stride[1];
}
c10::SmallVector<int64_t, N> kernel_size_new = {1, kernel_size[0], kernel_size[1], 1};
c10::SmallVector<int64_t, N> strides_size = {1, stride_H, stride_W, 1};
c10::SmallVector<int64_t, N> paddings = {1, padding[0], padding[1], 1};
c10::SmallVector<int64_t, N> dilations = {1, dilation[0], dilation[1], 1};
at_npu::native::OpCommand cmd;
if (indices.dtype() == at::kChar) {
** The indices of op_api foward, has continuous storage memory, and dtype is int8, format is NCHW, shape is 4D.
** To avoid FE insert transdata node, and damage the storage memory, modifing indices's format to NC1HWC0.
** And modify the dtype to short to align with argmax of op parameter.
*/
at::SmallVector<int64_t, N> shape;
for (int i = 0; i < indices.dim(); i++) {
shape.emplace_back((i == 3) ? indices.size(i) / INDICES_TYPE_CONVERT / BLOCKSIZE : indices.size(i));
}
shape.emplace_back(BLOCKSIZE);
at::Tensor indices_para = indices.view(at::kShort);
torch_npu::NPUStorageDesc &desc = torch_npu::NPUBridge::GetNpuStorageImpl(indices_para)->npu_desc_;
desc.npu_format_ = ACL_FORMAT_NC1HWC0;
desc.storage_sizes_ = shape;
desc.data_type_ = at::ScalarType::Short;
desc.origin_format_ = ACL_FORMAT_NCHW;
desc.base_sizes_ = indices_para.sizes();
desc.base_strides_ = indices_para.strides();
cmd.Name("MaxPoolGradWithArgmaxV1")
.Input(self, "x")
.Input(grad_output, "grad")
.Input(indices_para, "argmax", c10::nullopt, "uint16")
.Output(grad_input, "y")
.Attr("ksize", kernel_size_new)
.Attr("strides", strides_size)
.Attr("pads", paddings)
.Attr("dilations", dilations)
.Attr("ceil_mode", ceil_mode)
.Run();
} else {
cmd.Name("MaxPoolGradWithArgmaxV1")
.Input(self, "x")
.Input(grad_output, "grad")
.Input(indices, "argmax", c10::nullopt, "uint16")
.Output(grad_input, "y")
.Attr("ksize", kernel_size_new)
.Attr("strides", strides_size)
.Attr("pads", paddings)
.Attr("dilations", dilations)
.Attr("ceil_mode", ceil_mode)
.Run();
}
if (self.dim() == 3) {
grad_input.squeeze_(0);
}
return grad_input;
}
void max_pool2d_with_indices_backward_parameter_check(const at::Tensor &self, at::IntArrayRef kernel_size,
at::IntArrayRef stride, at::IntArrayRef padding,
at::IntArrayRef dilation)
{
TORCH_CHECK(kernel_size.size() == 1 || kernel_size.size() == 2,
"max_pool2d: kernel_size must either be a single int, or a tuple of two ints"
+ OPS_ERROR(ErrCode::PARAM));
TORCH_CHECK(stride.size() == 0 || stride.size() == 1 || stride.size() == 2,
"max_pool2d: stride must either be omitted, a single int, or a tuple of two ints"
+ OPS_ERROR(ErrCode::PARAM));
TORCH_CHECK(padding.size() == 1 || padding.size() == 2,
"max_pool2d: padding must be either be a single int, or a tuple of two ints"
+ OPS_ERROR(ErrCode::PARAM));
TORCH_CHECK(dilation.size() == 1 || dilation.size() == 2,
"max_pool2d: dilation must be either a single int, or a tuple of two ints"
+ OPS_ERROR(ErrCode::PARAM));
TORCH_CHECK((self.ndimension() == 3 || self.ndimension() == 4),
"non-empty 3D or 4D (batch mode) tensor expected for input"
+ OPS_ERROR(ErrCode::PARAM));
}
}
at::Tensor &max_pool2d_with_indices_backward_out(const at::Tensor &grad_output, const at::Tensor &self,
at::IntArrayRef kernel_size, at::IntArrayRef stride,
at::IntArrayRef padding, at::IntArrayRef dilation, bool ceil_mode,
const at::Tensor &indices, at::Tensor &grad_input)
{
max_pool2d_with_indices_backward_parameter_check(self, kernel_size, stride, padding, dilation);
npu_preparation::CheckOut({self, grad_output, indices}, grad_input, self);
if (!npu_utils::check_match(&grad_input)) {
at::Tensor contig_grad_input = npu_utils::format_contiguous(grad_input);
max_pool2d_with_indices_backward_out_nocheck(contig_grad_input, grad_output, self, kernel_size, stride, padding,
dilation, ceil_mode, indices);
npu_utils::format_fresh_view(grad_input, contig_grad_input);
} else {
max_pool2d_with_indices_backward_out_nocheck(grad_input, grad_output, self, kernel_size, stride, padding,
dilation, ceil_mode, indices);
}
return grad_input;
}
at::Tensor max_pool2d_with_indices_backward(const at::Tensor &grad_output_var, const at::Tensor &self,
at::IntArrayRef kernel_size, at::IntArrayRef stride,
at::IntArrayRef padding, at::IntArrayRef dilation, bool ceil_mode,
const at::Tensor &indices)
{
max_pool2d_with_indices_backward_parameter_check(self, kernel_size, stride, padding, dilation);
const int k_H = at::native::safe_downcast<int, int64_t>(kernel_size[0]);
const int k_W = kernel_size.size() == 1 ? k_H : at::native::safe_downcast<int, int64_t>(kernel_size[1]);
c10::SmallVector<int64_t, SIZE> kernel_sizes = {k_H, k_W};
at::IntArrayRef kernel_sizess = at::IntArrayRef(kernel_sizes);
const int d_H = stride.empty() ? k_H : at::native::safe_downcast<int, int64_t>(stride[0]);
const int d_W = stride.empty() ? k_W :
stride.size() == 1 ? d_H :
at::native::safe_downcast<int, int64_t>(stride[1]);
c10::SmallVector<int64_t, SIZE> strides = {d_H, d_W};
at::IntArrayRef stridess = at::IntArrayRef(strides);
const int pad_H = at::native::safe_downcast<int, int64_t>(padding[0]);
const int pad_W = padding.size() == 1 ? pad_H : at::native::safe_downcast<int, int64_t>(padding[1]);
c10::SmallVector<int64_t, SIZE> paddings = {pad_H, pad_W};
at::IntArrayRef padss = at::IntArrayRef(paddings);
const int dilation_H = at::native::safe_downcast<int, int64_t>(dilation[0]);
const int dilation_W = dilation.size() == 1 ? dilation_H : at::native::safe_downcast<int, int64_t>(dilation[1]);
c10::SmallVector<int64_t, SIZE> dilations = {dilation_H, dilation_W};
at::IntArrayRef dilationss = at::IntArrayRef(dilations);
at::Tensor grad_input = npu_preparation::apply_tensor(self);
max_pool2d_with_indices_backward_out_nocheck(grad_input, grad_output_var, self, kernel_sizess, stridess, padss,
dilationss, ceil_mode, indices);
return grad_input;
}
}