"""Add 算子测试规范(TTK TestSpec)。
kernel/GEIR = add(numpy 入/出):真值经 numpy_to_torch_tensor 转 torch 计算
torch.add(fp16/bf16 抬 fp32 中间精度以匹配 PrecisionReduceFlag;complex32
拆实虚部抬 complex64 相加后拼回),再经 torch_to_numpy_tensor 回转 numpy;
浮点档走 cross_check 三方交叉校验(对比 torch 标杆);
aclnn = aclnnAdd / aclnnAdds / aclnnInplaceAdd / aclnnInplaceAdds /
aclnnAddV3 / aclnnInplaceAddV3(torch 入/出):
真值 out = self + alpha * other(Adds 的 other 为标量;AddV3 的 self 为标量、
other 为张量;InplaceAdd/Adds 写回 selfRef,InplaceAddV3 写回 other);
e2e = torch.add(torch 入/出):真值 torch.add(input, other, alpha)。
"""
import torch
__spec__ = {
"add": "AddTestSpec",
"aclnnAdd": "AclnnAddTestSpec",
"aclnnAdds": "AclnnAddTestSpec",
"aclnnInplaceAdd": "AclnnInplaceAddTestSpec",
"aclnnInplaceAdds": "AclnnInplaceAddTestSpec",
"aclnnAddV3": "AclnnAddV3TestSpec",
"aclnnInplaceAddV3": "AclnnInplaceAddV3TestSpec",
"torch.add": "TorchAddTestSpec",
}
def _torch_add(t1, t2, alpha=1.0):
"""torch.add 包装:整数输入时 alpha 必须为整数(torch 拒绝浮点 alpha)。
t1/t2 可为张量或标量(AddV3 的 self 为标量、other 为张量),整型判定仅对
张量生效,标量不触发 torch.is_floating_point 的 Tensor 类型要求。
"""
if isinstance(alpha, float) and alpha.is_integer():
is_integral = (torch.is_tensor(t1) and not torch.is_floating_point(t1)) or (
torch.is_tensor(t2) and not torch.is_floating_point(t2)
)
if is_integral:
alpha = int(alpha)
return torch.add(t1, t2, alpha=alpha)
def _cast_to_out(result, out):
"""按 out 张量的 dtype 回写结果(对齐 NPU 输出 dtype)。"""
if out is not None and result.dtype != out.dtype:
return result.to(out.dtype)
return result
def _third_party_inputs(
args,
kwargs,
first_names=("self", "selfT", "x1", "input"),
second_names=("other", "x2"),
):
"""从三方标杆调用参数中提取两个输入张量(位置优先,kwargs 兜底)。
位置不足时按候选名从 kwargs 回退(兼容关键字喂参如 self=/other=/x1=/x2=);
仍缺失时抛含 args/kwargs 的明确异常,避免裸 IndexError。
"""
def _pick(idx, names):
if len(args) > idx:
return args[idx]
for name in names:
if name in kwargs:
return kwargs[name]
return None
first = _pick(0, first_names)
second = _pick(1, second_names)
if first is None or second is None:
raise TypeError(
"third_party: two input tensors not provided (args=%r, kwargs=%r)"
% (args, sorted(kwargs))
)
return first, second
class ThirdPartyImpl:
"""torch 三方标杆(各模式共用):alpha 属性绑定,两个输入按位置接收。
Add 的入参名含 Python 保留名 ``self``,故 ``__call__`` 按位置取
args[0]/args[1] 而非按名绑定。
"""
def __init__(self, *, alpha=1.0, **kwargs):
self.alpha = alpha
def __call__(self, *args, **kwargs):
a, b = _third_party_inputs(args, kwargs)
return [_torch_add(a, b, self.alpha)]
class AddTestSpec:
"""kernel/GEIR 流程:numpy 入/出。"""
@staticmethod
def golden(x1, x2, alpha=1.0, **kwargs):
"""kernel/GEIR golden:numpy → torch 计算 → numpy 回转。"""
from ttk.utilities import numpy_to_torch_tensor, torch_to_numpy_tensor
out_dtypes = tuple(kwargs.get("output_dtypes") or ())
out_dtype = str(out_dtypes[0]).lower() if out_dtypes else None
if out_dtype == "complex32":
import numpy
a = x1[..., 0].astype(numpy.float32) + 1j * x1[..., 1].astype(numpy.float32)
b = x2[..., 0].astype(numpy.float32) + 1j * x2[..., 1].astype(numpy.float32)
res = a + b
return [
numpy.stack(
(res.real.astype(numpy.float16), res.imag.astype(numpy.float16)),
axis=-1,
)
]
x1_torch = numpy_to_torch_tensor(x1)
x2_torch = numpy_to_torch_tensor(x2)
return [torch_to_numpy_tensor(_torch_add(x1_torch, x2_torch, alpha))]
third_party = {"torch": ThirdPartyImpl}
class AclnnAddTestSpec:
"""aclnnAdd / aclnnAdds 流程:out = self + alpha * other(other 张量或标量)。"""
def golden(self, other, alpha, out, **kwargs):
"""参数跟随 aclnnAdd/aclnnAdds:self 张量、other 张量/标量、alpha 标量、out 张量。"""
return [_cast_to_out(_torch_add(self, other, alpha), out)]
third_party = {"torch": ThirdPartyImpl}
class AclnnInplaceAddTestSpec:
"""aclnnInplaceAdd / aclnnInplaceAdds 流程:selfRef += alpha * other(other 张量或标量)。"""
def golden(selfRef, other, alpha, **kwargs):
"""参数跟随 aclnnInplaceAdd/aclnnInplaceAdds:selfRef 张量、other 张量/标量、alpha 标量。"""
return [_torch_add(selfRef, other, alpha)]
third_party = {"torch": ThirdPartyImpl}
class AclnnAddV3TestSpec:
"""aclnnAddV3 流程:out = self + alpha * other(self 标量、other 张量)。"""
def golden(self, other, alpha, out, **kwargs):
"""参数跟随 aclnnAddV3:self 标量、other 张量、alpha 标量、out 张量。"""
return [_cast_to_out(_torch_add(self, other, alpha), out)]
third_party = {"torch": ThirdPartyImpl}
class AclnnInplaceAddV3TestSpec:
"""aclnnInplaceAddV3 流程:other = selfRef + alpha * other(selfRef 标量、other 张量输入输出)。"""
def golden(selfRef, other, alpha, **kwargs):
"""参数跟随 aclnnInplaceAddV3:selfRef 标量、other 张量(输入输出)、alpha 标量。"""
return [_torch_add(selfRef, other, alpha)]
third_party = {"torch": ThirdPartyImpl}
class TorchAddTestSpec:
"""e2e 流程:torch 入/出。"""
def golden(input, other, alpha=1.0, **kwargs):
"""参数跟随 torch.add:input/other 位置张量,alpha 标量属性。"""
return [_torch_add(input, other, alpha)]
third_party = {"torch": ThirdPartyImpl}