[XPU][PHI Kernels] bind reduce_max_int64 set_value_bool sin_grad_fp32 cos_grad_fp32 for XPU #55375
Conversation
|
你的PR提交成功，感谢你对开源项目的贡献! |
|
❌ The PR is not created using PR's template. You can refer to this Demo. |
test/xpu/test_set_value_op_xpu.py
Outdated
| @@ -171,6 +173,8 @@ class XPUTestSetValueItemSliceInWhile(XPUTestSetValueApi): | |||
| def set_dtype(self): | |||
| if self.in_type == np.float16: | |||
| self.dtype = "float32" | |||
| elif self.in_type == np.bool_: | |||
| self.dtyper = "bool" | |||
| @@ -57,6 +57,8 @@ def set_value(self): | |||
|
|
|||
| def set_dtype(self): | |||
| self.dtype = self.in_type | |||
| if self.in_type == np.bool_: | |||
There was a problem hiding this comment.
这里对XPUTestSetValueItemBool重载了set_dtype，这几个测试跑只跑fp32，其他的单测还是测所有类型
| @@ -166,7 +169,7 @@ void SetValueImpl(const Context& dev_ctx, | |||
| } | |||
| } | |||
|
|
|||
| // Because strided_slice does not support the case of stride < 0 | |||
| // Because strided_slice_view_update does not support the case of stride < 0 | |||
| value_dims_vec.end(), | ||
| ext_value_dims_vec.begin() + (max_dims - value_dims.size())); | ||
|
|
||
| r = xpu::broadcast<XPUType>(dev_ctx.x_context(), |
There was a problem hiding this comment.
这里的broadcast以及上面slice_dims_for_assign的处理最好封装在XPUElementwise函数里面，然后以functor的形式传进去，增加代码的复用程度
| self.dtype = "float32" | ||
| else: | ||
| self.dtype = self.in_type | ||
| self.dtype = "float32" |
There was a problem hiding this comment.
如前面PR的description所述，添加了Bool类型后XPU单测挂了，后面我试过CPU也会报同样的错误且那边的单测只测float32
CPU那边的单测，TestItemBool继承了这里的set_dtype。
CPU的报错信息：
核心yr原因还是目前框架会在图里红圈的地方调用elementwise_sub，但是这个操作CPU上也不支持bool类型。换成FP32后结果就正常了。
|
你的PR已合入Paddle库，请关注后续测试结果。 |
… cos_grad_fp32 for XPU (PaddlePaddle#55375) * bind kernels for xpu * format code * format code * 0d support for set value * refine set_value
… cos_grad_fp32 for XPU (PaddlePaddle#55375) * bind kernels for xpu * format code * format code * 0d support for set value * refine set_value
PR types
New features
PR changes
OPs
Description