#include "op_plugin/AclOpsInterface.h"
#include "op_plugin/OpApiInterface.h"
#include "op_plugin/utils/op_api_common.h"
#include "op_plugin/utils/OpUtils.h"
namespace {
using npu_preparation = at_npu::native::OpPreparation;
bool is_three_tensor_base_format(const at::Tensor &input0, const at::Tensor &input1, const at::Tensor &input2)
{
return at_npu::native::FormatHelper::IsOpInputBaseFormat(input0) &&
op_plugin::utils::is_two_tensor_base_format(input1, input2);
}
bool is_nd_nd_nz_format(const at::Tensor &self, const at::Tensor &mat1, const at::Tensor &mat2)
{
auto dim_tensor0 = self.dim();
return (dim_tensor0 == 2 || dim_tensor0 == 1) && !op_plugin::utils::is_nz_format(self) &&
op_plugin::utils::is_nd_nz_format(mat1, mat2);
}
}
#define DO_ADDMM_COMPATIBILITY(aclnn_nz_api, aclnn_nd_api, input0, input1, input2, aclop_func_call) \
do { \
if (is_three_tensor_base_format(input0, input1, input2)) { \
DO_COMPATIBILITY(aclnn_nd_api, aclop_func_call); \
} else { \
static bool is_support_soc = (c10_npu::GetSocVersion() >= c10_npu::SocVersion::Ascend910B1 && \
c10_npu::GetSocVersion() < c10_npu::SocVersion::Ascend310B1) || \
(c10_npu::GetSocVersion() > c10_npu::SocVersion::Ascend310B4); \
if (is_nd_nd_nz_format(input0, input1, input2) && is_support_soc) { \
DO_COMPATIBILITY(aclnn_nz_api, aclop_func_call); \
} else { \
if (!c10_npu::IsAclnnOnly()) { \
return aclop_func_call; \
} \
const torch_npu::NPUStorageDesc &tensor_desc0 = \
torch_npu::NPUBridge::GetNpuStorageImpl(input0)->npu_desc_; \
const torch_npu::NPUStorageDesc &tensor_desc1 = \
torch_npu::NPUBridge::GetNpuStorageImpl(input1)->npu_desc_; \
const torch_npu::NPUStorageDesc &tensor_desc2 = \
torch_npu::NPUBridge::GetNpuStorageImpl(input2)->npu_desc_; \
TORCH_CHECK(false, \
"matmul got not support format in current device: ", \
"(", \
tensor_desc0.npu_format_, \
", ", \
tensor_desc1.npu_format_, \
", ", \
tensor_desc2.npu_format_, \
")", \
OPS_ERROR(ErrCode::PARAM)); \
} \
} \
} while (0)
namespace op_api {
using npu_preparation = at_npu::native::OpPreparation;
at::Tensor &addmm_out(
const at::Tensor &self,
const at::Tensor &mat1,
const at::Tensor &mat2,
const at::Scalar &beta,
const at::Scalar &alpha,
at::Tensor &out)
{
DO_ADDMM_COMPATIBILITY(aclnnAddmmWeightNz, aclnnAddmm, self, mat1, mat2,
acl_op::addmm_out(self, mat1, mat2, beta, alpha, out));
int8_t cube_math_type = op_plugin::utils::get_cube_math_type_with_passthrough();
auto output_size = op_infer::addmm_npu_output_size(self, mat1, mat2);
npu_preparation::check_tensor({self, mat1, mat2}, out, out.scalar_type(), output_size);
if (is_nd_nd_nz_format(self, mat1, mat2)) {
EXEC_NPU_CMD(aclnnAddmmWeightNz, self, mat1, mat2, beta, alpha, out, cube_math_type);
} else {
EXEC_NPU_CMD(aclnnAddmm, self, mat1, mat2, beta, alpha, out, cube_math_type);
}
auto names = at::namedinference::propagate_names_for_addmm(mat1, mat2, self);
at::namedinference::propagate_names_if_nonempty(out, names);
return out;
}
at::Tensor addmm(
const at::Tensor &self,
const at::Tensor &mat1,
const at::Tensor &mat2,
const at::Scalar &beta,
const at::Scalar &alpha)
{
DO_ADDMM_COMPATIBILITY(aclnnAddmmWeightNz, aclnnAddmm, self, mat1, mat2,
acl_op::addmm(self, mat1, mat2, beta, alpha));
auto output_size = op_infer::addmm_npu_output_size(self, mat1, mat2);
at::Tensor result = npu_preparation::apply_tensor_without_format(output_size, self.options());
int8_t cube_math_type = op_plugin::utils::get_cube_math_type_with_passthrough();
if (is_nd_nd_nz_format(self, mat1, mat2)) {
EXEC_NPU_CMD(aclnnAddmmWeightNz, self, mat1, mat2, beta, alpha, result, cube_math_type);
} else {
EXEC_NPU_CMD(aclnnAddmm, self, mat1, mat2, beta, alpha, result, cube_math_type);
}
auto names = at::namedinference::propagate_names_for_addmm(mat1, mat2, self);
at::namedinference::propagate_names_if_nonempty(result, names);
FLOP_COUNT(FlopCounter::addmm_flop, mat1, mat2);
return result;
}
at::Tensor &addmm_(
at::Tensor &self,
const at::Tensor &mat1,
const at::Tensor &mat2,
const at::Scalar &beta,
const at::Scalar &alpha)
{
DO_COMPATIBILITY(aclnnInplaceAddmm, acl_op::addmm_(self, mat1, mat2, beta, alpha));
auto output_size = op_infer::addmm_npu_output_size(self, mat1, mat2);
npu_preparation::check_tensor({self, mat1, mat2}, self, self.scalar_type(), output_size);
int8_t cube_math_type = op_plugin::utils::get_cube_math_type_with_passthrough();
EXEC_NPU_CMD(aclnnInplaceAddmm, self, mat1, mat2, beta, alpha, cube_math_type);
auto names = at::namedinference::propagate_names_for_addmm(mat1, mat2, self);
at::namedinference::propagate_names_if_nonempty(self, names);
return self;
}
}