#include "op_plugin/AclOpsInterface.h"
#include "op_plugin/utils/OpAdapter.h"
#include "torch_npu/csrc/framework/utils/UtilForOpAdapter.h"
namespace acl_op {
using npu_preparation = at_npu::native::OpPreparation;
using npu_utils = at_npu::native::NpuUtils;
namespace {
inline void alpha_check_npu(const at::ScalarType dtype, at::Scalar alpha)
{
TORCH_CHECK(!alpha.isBoolean() || dtype == at::kBool, "Boolean alpha only supported for Boolean results."
+ OPS_ERROR(ErrCode::TYPE));
TORCH_CHECK(isFloatingType(dtype) || alpha.isIntegral(true),
"For integral input tensors, argument alpha must not be a floating point number."
+ OPS_ERROR(ErrCode::TYPE));
}
at::Tensor add_dest_output(const at::Tensor &self, const at::Tensor &other)
{
bool is_self_wrapped = npu_preparation::is_scalar_wrapped_to_tensor(self);
return is_self_wrapped ? other : self;
}
at::Tensor &adds_out_npu_nocheck(at::Tensor &result, const at::Tensor &self, const at::Scalar other,
const at::Scalar alpha)
{
alpha_check_npu(result.scalar_type(), alpha);
float other_value = op_plugin::utils::get_scalar_float_value(other);
float alpha_value = op_plugin::utils::get_scalar_float_value(alpha);
float value = other_value * alpha_value;
at_npu::native::OpCommand cmd;
std::string real_type = "";
if (self.scalar_type() == at::kBool) {
auto unified_result = npu_preparation::binary_op_check(result, self, other, true);
if (unified_result.common_type == at::kBool) {
unified_result.common_type = at::kByte;
unified_result.result_type_defined = true;
real_type = "uint8";
}
cmd.Expect(unified_result);
}
cmd.Name("Add")
.Input(self)
.Input(at::Scalar(value), self.scalar_type())
.Output(result, "", c10::nullopt, real_type)
.Run();
return result;
}
at::Tensor &add_out_npu_nocheck(at::Tensor &result, const at::Tensor &self, const at::Tensor &other, at::Scalar alpha)
{
auto unified_result = npu_preparation::binary_op_check(result, self, other, true);
if (npu_preparation::IsCPUScalar(other)) {
adds_out_npu_nocheck(result, self, other.item(), alpha);
} else if (npu_preparation::IsCPUScalar(self)) {
adds_out_npu_nocheck(result, other, self.item(), alpha);
} else {
alpha_check_npu(result.scalar_type(), alpha);
at_npu::native::OpCommand cmd;
cmd.Expect(unified_result);
if (op_plugin::utils::is_scalar_one(alpha)) {
if (self.scalar_type() == at::kLong) {
TORCH_NPU_WARN_ONCE(
"The oprator of add is executed, Currently High Accuracy but Low Performance OP with 64-bit has "
"been used, Please Do Some Cast at Python Functions with 32-bit for Better Performance!");
}
std::string real_type = "";
if (self.scalar_type() == at::kBool && other.scalar_type() == at::kBool) {
unified_result.common_type = at::kByte;
unified_result.result_type_defined = true;
cmd.Expect(unified_result);
real_type = "uint8";
}
cmd.Name("Add").Input(self).Input(other).Output(result, "", c10::nullopt, real_type).Run();
} else {
cmd.Name("AxpyV2").Input(self).Input(other).Input(alpha, self.scalar_type()).Output(result).Run();
}
}
return result;
}
bool check_size(const at::Tensor &self, const at::Tensor &other)
{
if (self.dim() != other.dim()) {
return false;
}
for (size_t i = 0; i < static_cast<uint64_t>(self.dim()); i++) {
if (self.size(i) != other.size(i)) {
return false;
}
}
return true;
}
at::Tensor stride_add_tensor_get(const at::Tensor &src)
{
if (src.is_contiguous()) {
return src;
} else {
auto src_desc = torch_npu::NPUBridge::GetNpuStorageImpl(src)->npu_desc_;
at::Tensor src_new =
npu_preparation::apply_tensor_with_format(src_desc.base_sizes_, src.options(), ACL_FORMAT_NC1HWC0);
src_new.set_(src.storage(), src_new.storage_offset(), src_new.sizes(), src_new.strides());
return src_new;
}
}
}
at::Tensor add(const at::Tensor &self, const at::Tensor &other, const at::Scalar &alpha)
{
alpha_check_npu(at::native::result_type(self, other), alpha);
if ((!(self.is_contiguous() && other.is_contiguous())) &&
(npu_utils::check_5d_5d_match(self) || npu_utils::check_5d_5d_match(other)) && check_size(self, other)) {
int64_t c0_len = 16;
at::Tensor self_use = stride_add_tensor_get(self);
TORCH_CHECK(self.numel() > 0, "self must be non-empty tensor" + OPS_ERROR(ErrCode::PARAM));
TORCH_CHECK(other.numel() > 0, "other must be non-empty tensor" + OPS_ERROR(ErrCode::PARAM));
at::Scalar self_c1_offset(self.storage_offset() / (self.size(2) * self.size(3) * c0_len));
at::Tensor other_use = stride_add_tensor_get(other);
at::Scalar other_c1_offset(other.storage_offset() / (other.size(2) * other.size(3) * c0_len));
at::Scalar stride_len(self.size(1) / c0_len);
at::Tensor result = acl_op::npu_stride_add(self_use, other_use, self_c1_offset, other_c1_offset, stride_len);
return result;
}
at::Tensor output_tensor = add_dest_output(self, other);
auto output_size = op_infer::broadcast_ops_npu_output_size(self, other);
at::ScalarType result_type = at::native::result_type(self, other);
at::Tensor self_cp = (self.scalar_type() != result_type && !npu_preparation::is_scalar_wrapped_to_tensor(self)) ?
at_npu::native::custom_ops::_npu_dtype_cast(self, result_type) :
self;
at::Tensor other_cp = (other.scalar_type() != result_type && !npu_preparation::is_scalar_wrapped_to_tensor(other)) ?
at_npu::native::custom_ops::_npu_dtype_cast(other, result_type) :
other;
at::Tensor result = npu_preparation::apply_tensor_with_format(
output_size, output_tensor.options().dtype(result_type), npu_preparation::get_tensor_npu_format(output_tensor));
add_out_npu_nocheck(result, self_cp, other_cp, alpha);
return result;
}
at::Tensor add(const at::Tensor &self, const at::Scalar &other, const at::Scalar &alpha)
{
alpha_check_npu(at::native::result_type(self, other), alpha);
at::Tensor result = npu_preparation::apply_tensor(self);
adds_out_npu_nocheck(result, self, other, alpha);
return result;
}
at::Tensor &add_(at::Tensor &self, const at::Tensor &other, const at::Scalar &alpha)
{
at::ScalarType result_type = at::native::result_type(self, other);
at::ScalarType self_type = self.scalar_type();
TORCH_CHECK(canCast(result_type, self_type), "result type ", result_type,
" can't be cast to the desired output type ", self_type, OPS_ERROR(ErrCode::TYPE));
at::Tensor self_cp = (self_type != result_type && !npu_preparation::is_scalar_wrapped_to_tensor(self)) ?
at_npu::native::custom_ops::_npu_dtype_cast(self, result_type) :
self;
at::Tensor other_cp = (other.scalar_type() != result_type && !npu_preparation::is_scalar_wrapped_to_tensor(other)) ?
at_npu::native::custom_ops::_npu_dtype_cast(other, result_type) :
other;
npu_preparation::CheckMemory({self_cp, other_cp}, {self_cp});
if (!npu_utils::check_match(&self_cp)) {
at::Tensor contiguous_self;
if (c10_npu::GetSocVersion() >= c10_npu::SocVersion::Ascend910B1) {
contiguous_self = npu_utils::format_contiguous_add_copy_optimize(self_cp);
} else {
contiguous_self = npu_utils::format_contiguous(self_cp);
}
add_out_npu_nocheck(contiguous_self, contiguous_self, other_cp, alpha);
npu_utils::format_fresh_view(self_cp, contiguous_self);
} else {
add_out_npu_nocheck(self_cp, self_cp, other_cp, alpha);
}
if (self_type == result_type) {
self = self_cp;
} else {
self.copy_(self_cp);
}
return self;
}
at::Tensor &add_(at::Tensor &self, const at::Scalar &other, const at::Scalar &alpha)
{
if (!npu_utils::check_match(&self)) {
at::Tensor contiguous_self;
if (c10_npu::GetSocVersion() >= c10_npu::SocVersion::Ascend910B1) {
contiguous_self = npu_utils::format_contiguous_add_copy_optimize(self);
} else {
contiguous_self = npu_utils::format_contiguous(self);
}
adds_out_npu_nocheck(contiguous_self, contiguous_self, other, alpha);
npu_utils::format_fresh_view(self, contiguous_self);
} else {
adds_out_npu_nocheck(self, self, other, alpha);
}
return self;
}
at::Tensor &add_out(const at::Tensor &self, const at::Tensor &other, const at::Scalar &alpha, at::Tensor &result)
{
at::Tensor output_tensor = add_dest_output(self, other);
auto output_size = op_infer::broadcast_ops_npu_output_size(self, other);
at::ScalarType result_type = at::native::result_type(self, other);
at::Tensor self_cp = (self.scalar_type() != result_type && !npu_preparation::is_scalar_wrapped_to_tensor(self)) ?
at_npu::native::custom_ops::_npu_dtype_cast(self, result_type) :
self;
at::Tensor other_cp = (other.scalar_type() != result_type && !npu_preparation::is_scalar_wrapped_to_tensor(other)) ?
at_npu::native::custom_ops::_npu_dtype_cast(other, result_type) :
other;
npu_preparation::CheckOut({self_cp, other_cp}, result, npu_preparation::get_tensor_npu_format(result), result_type,
output_size);
if (!npu_utils::check_match(&result)) {
at::Tensor contiguous_result;
if (c10_npu::GetSocVersion() >= c10_npu::SocVersion::Ascend910B1) {
contiguous_result = npu_utils::format_contiguous_add_copy_optimize(result);
} else {
contiguous_result = npu_utils::format_contiguous(result);
}
add_out_npu_nocheck(contiguous_result, self_cp, other_cp, alpha);
npu_utils::format_fresh_view(result, contiguous_result);
} else {
add_out_npu_nocheck(result, self_cp, other_cp, alpha);
}
return result;
}
}