#include <vector>
#include "op_plugin/OpApiInterface.h"
#include "op_plugin/utils/op_api_common.h"
#include "torch_npu/csrc/framework/utils/InternalFormatOpAdapter.h"
namespace op_api {
constexpr int X1_DIM_NUM = 3;
constexpr int X1_SCALE_DIM = 2;
constexpr int X_LAST_DIM_INDEX = 2;
constexpr int Y_DIM_NUM = 2;
using npu_preparation = at_npu::native::OpPreparation;
static bool is_nz_format(const at::Tensor &w)
{
const torch_npu::NPUStorageDesc &tensor_desc = torch_npu::NPUBridge::GetNpuStorageImpl(w)->npu_desc_;
return tensor_desc.npu_format_ == ACL_FORMAT_FRACTAL_NZ;
}
at::Tensor npu_quant_matmul_reduce_sum(
const at::Tensor &x1,
const at::Tensor &x2,
const c10::optional<at::Tensor> &x1_scale_optional,
const c10::optional<at::Tensor> &x2_scale_optional)
{
bool is_x2_nz = is_nz_format(x2);
TORCH_CHECK(is_x2_nz, "only support x2's format is nz.", OPS_ERROR(ErrCode::PARAM));
TORCH_CHECK(x1.dim() == X1_DIM_NUM, "x1 dim should be ", X1_DIM_NUM, OPS_ERROR(ErrCode::PARAM));
TORCH_CHECK(x2.dim() == X1_DIM_NUM, "x2 dim should be ", X1_DIM_NUM, OPS_ERROR(ErrCode::PARAM));
const at::Tensor &x1_scale = x1_scale_optional.value_or(at::Tensor());
TORCH_CHECK(x1_scale.dim() == X1_SCALE_DIM, "x1_scale dim should be ", X1_SCALE_DIM, OPS_ERROR(ErrCode::PARAM));
const at::Tensor &x2_scale = x2_scale_optional.value_or(at::Tensor());
TORCH_CHECK(x2_scale.dim() == 1, "x2_scale dim should be ", 1, OPS_ERROR(ErrCode::PARAM));
auto b_dim = x1.size(0);
auto m_dim = x1.size(1);
auto n_dim = x2.size(X_LAST_DIM_INDEX);
TORCH_CHECK(b_dim == x2.size(0), "the first dim of x1 and x2 must be same", OPS_ERROR(ErrCode::PARAM));
TORCH_CHECK(
x1.size(X_LAST_DIM_INDEX) == x2.size(1), "the K dim of x1 and x2 must be same", OPS_ERROR(ErrCode::PARAM));
TORCH_CHECK(x2_scale.size(0) == n_dim, "the shape of x2_scale must be (N,)", OPS_ERROR(ErrCode::PARAM));
TORCH_CHECK(x1_scale.size(1) == m_dim && x1_scale.size(0) == b_dim,
"the shape of x1_scale must be (B, M)",
OPS_ERROR(ErrCode::PARAM));
c10::SmallVector<int64_t, Y_DIM_NUM> output_size;
output_size.push_back(m_dim);
output_size.push_back(n_dim);
c10::TensorOptions options = x1.options().dtype(at::ScalarType::BFloat16);
at::Tensor result = npu_preparation::apply_tensor_without_format(output_size, options);
at::Tensor y_scale;
at::Tensor x1_offset;
at::Tensor x2_offset;
at::Tensor y_offset;
at::Tensor bias;
bool transpose_x1 = false;
bool transpose_x2 = false;
int64_t group_size = -1;
c10::IntArrayRef dims = {0};
bool keep_dims = false;
EXEC_NPU_CMD(aclnnQuantMatmulReduceSumWeightNz,
x1, x2, x1_scale, x2_scale, y_scale, x1_offset, x2_offset, y_offset, bias,
transpose_x1, transpose_x2, group_size, dims, keep_dims, result);
return result;
}
}