#include "op_plugin/AclOpsInterface.h"
#include "op_plugin/utils/OpAdapter.h"
namespace acl_op {
using npu_preparation = at_npu::native::OpPreparation;
namespace {
bool is_special_conv1d(const at::Tensor &input, const at::Tensor &weight, at::IntArrayRef stride,
at::IntArrayRef padding, at::IntArrayRef dilation, int64_t groups)
{
if (stride[1] > 63 && stride[1] == weight.size(3) && padding[1] == 0 && dilation[1] == 1 && groups == 1 &&
input.size(1) == 1) {
return true;
} else {
return false;
}
}
}
at::Tensor &npu_conv2d_out(const at::Tensor &input, const at::Tensor &weight, const c10::optional<at::Tensor> &bias_opt,
at::IntArrayRef stride, at::IntArrayRef padding, at::IntArrayRef dilation, int64_t groups,
at::Tensor &result)
{
TORCH_CHECK(stride.size() >= 2, "stride has to contain more than 2 elements, but got ", stride.size(),
OPS_ERROR(ErrCode::PARAM));
TORCH_CHECK(padding.size() >= 2, "padding has to contain more than 2 elements, but got ", padding.size(),
OPS_ERROR(ErrCode::PARAM));
TORCH_CHECK(dilation.size() >= 2, "dilation has to contain more than 2 elements, but got ", dilation.size(),
OPS_ERROR(ErrCode::PARAM));
const at::Tensor &bias = c10::value_or_else(bias_opt, [] { return at::Tensor(); });
c10::SmallVector<int64_t, N> strides_size = {1, 1, stride[0], stride[1]};
c10::SmallVector<int64_t, N> paddings = {padding[0], padding[0], padding[1], padding[1]};
c10::SmallVector<int64_t, N> dilations = {1, 1, dilation[0], dilation[1]};
at_npu::native::OpCommand cmd;
cmd.Name("Conv2D").Input(input, "x").Input(weight, "filter");
if (bias.defined()) {
cmd.Input(bias);
}
cmd.Output(result, "y")
.Attr("strides", strides_size)
.Attr("pads", paddings)
.Attr("dilations", dilations)
.Attr("groups", groups)
.Attr("data_format", (string) "NCHW")
.Run();
return result;
}
at::Tensor npu_conv2d(const at::Tensor &input, const at::Tensor &weight, const c10::optional<at::Tensor> &bias,
at::IntArrayRef stride, at::IntArrayRef padding, at::IntArrayRef dilation, int64_t groups)
{
TORCH_CHECK(input.dim() >= 4, "input has to be more than 4D, but got Tensor of dimension ", input.dim(),
OPS_ERROR(ErrCode::PARAM));
TORCH_CHECK(weight.dim() >= 4, "weight has to more than 4D, but got Tensor of dimension ", weight.dim(),
OPS_ERROR(ErrCode::PARAM));
TORCH_CHECK(stride.size() >= 2, "stride has to contain more than 2 elements, but got ", stride.size(),
OPS_ERROR(ErrCode::PARAM));
TORCH_CHECK(padding.size() >= 2, "padding has to contain more than 2 elements, but got ", padding.size(),
OPS_ERROR(ErrCode::PARAM));
TORCH_CHECK(dilation.size() >= 2, "dilation has to contain more than 2 elements, but got ", dilation.size(),
OPS_ERROR(ErrCode::PARAM));
TORCH_CHECK(weight.size(3) != 0, "4th dim of weight cannot be 0" + OPS_ERROR(ErrCode::PARAM));
TORCH_CHECK(stride[0] * stride[1] != 0, "Stride cannot contain 0" + OPS_ERROR(ErrCode::PARAM));
if (is_special_conv1d(input, weight, stride, padding, dilation, groups)) {
at::Tensor mm_input = input.view({input.size(0), input.size(3) / weight.size(3), weight.size(3)});
at::Tensor mm_other = weight.view({weight.size(0), weight.size(3)}).permute({1, 0});
at::Tensor mm_result = at::matmul(mm_input, mm_other);
at::Tensor result = mm_result.permute({0, 2, 1}).unsqueeze(2);
return result;
}
int64_t N = input.size(0);
int64_t H = input.size(2);
int64_t W = input.size(3);
int64_t Co = weight.size(0);
auto kernel_size = weight.sizes().slice(2);
int64_t Ho = (H + 2 * padding[0] - dilation[0] * (kernel_size[0] - 1) - 1) / stride[0] + 1;
int64_t Wo = (W + 2 * padding[1] - dilation[1] * (kernel_size[1] - 1) - 1) / stride[1] + 1;
TORCH_CHECK(Ho > 0, "Ho has to be positive, but got ", Ho,
OPS_ERROR(ErrCode::PARAM));
TORCH_CHECK(Wo > 0, "Wo has to be positive, but got ", Wo,
OPS_ERROR(ErrCode::PARAM));
c10::SmallVector<int64_t, SIZE> output_size = {N, Co, Ho, Wo};
int64_t result_format = input.dtype() == at::kHalf ? ACL_FORMAT_NC1HWC0 : ACL_FORMAT_ND;
at::Tensor result = npu_preparation::apply_tensor_with_format(input, output_size, result_format);
acl_op::npu_conv2d_out(input, weight, bias, stride, padding, dilation, groups, result);
return result;
}
}