[NPU] Support npu op reciprocal and reciprocal grad #34531
Conversation
|
Thanks for your contribution! |
| import paddle | ||
| import paddle.fluid as fluid | ||
|
|
||
| paddle.enable_static() |
There was a problem hiding this comment.
- 如果不是用API组网，静态图执行的方式写单测，这个enable_static应该是没有必要的
- 上面的import，看看没有必要的也可以删掉了
There was a problem hiding this comment.
- 此处参考了npu pow op的单测写法，使用的是静态图的写法.
- 无用的import已remove.
| return | ||
| self.check_grad_with_place( | ||
| self.place, ['X'], 'Out', max_relative_error=0.01) | ||
|
|
There was a problem hiding this comment.
FP16类型不用check_grad吗？这个NPU算子开发要求里有没有明确提到
There was a problem hiding this comment.
有明确提到：参见NPU开发算子文档：check_grad_with_place检查反向梯度时，不支持对float16数据类型的检查，不用写相关单测。
| self.dtype = np.float32 | ||
|
|
||
| def init_kernel_type(self): | ||
| pass |
There was a problem hiding this comment.
init_kernel_type这个方法在哪里用到了吗？
| ctx.template device_context<paddle::platform::NPUDeviceContext>() | ||
| .stream(); | ||
|
|
||
| const auto& runner = NpuOpRunner("Reciprocal", {*x}, {*out}, {}); |
| def set_npu(self): | ||
| self.__class__.use_npu = True | ||
| self.place = paddle.NPUPlace(0) | ||
| self.__class__.no_need_check_grad = True |
There was a problem hiding this comment.
如果fp16类型确定不需要梯度检查，需要使用skip_check_grad_ci装饰器，并描述原因（可以在源码里搜索下参考下别的单测）
| ops::ReciprocalGradNPUKernel<paddle::platform::NPUDeviceContext, int>, | ||
| ops::ReciprocalGradNPUKernel<paddle::platform::NPUDeviceContext, int64_t>, | ||
| ops::ReciprocalGradNPUKernel<paddle::platform::NPUDeviceContext, | ||
| paddle::platform::float16>); |
There was a problem hiding this comment.
看PR描述里提到 complex64, complex12是可以支持的，但这里没有注册，需要确认下开发要求
There was a problem hiding this comment.
已确定开发需求（@彭宇琪）：npu上支持的数据类型与cpu上保持一致即可。故最终调整支持的数据类型为float, double和float16.
limin2021
force-pushed
the
add_npu_reciprocal
branch
from
873f47b to
758f37c
Compare
limin2021
force-pushed
the
add_npu_reciprocal
branch
from
758f37c to
541df3d
Compare
| def init_dtype(self): | ||
| self.dtype = np.float16 | ||
|
|
||
|
|
There was a problem hiding this comment.
建议增加对double数据类型的单测，C++算子支持的数据类型要尽量在单测中覆盖
There was a problem hiding this comment.
Done.
测试fp64类型时，报了如下错误：
ExternalError: ACL error, the error code is : 500002. (at /workspace/limin-workspace/npu-paddle/Paddle/paddle/fluid/operators/npu_op_runner.cc:380) 1226: [operator < mean > error]。
原因：reciprocal的单测（反向）会自动添加mean op，但是mean的前向ReduceMeanD不支持double和int，注册mean npu kernel时却支持了fp64和int类型。
解决：将mean_op_npu.cc中对double和int类型的支持去掉。
limin2021
force-pushed
the
add_npu_reciprocal
branch
from
541df3d to
17912cf
Compare
PR types
New features
PR changes
OPs
Describe
[NPU] Support npu op reciprocal and reciprocal grad
说明：
NPU端支持的数据类型与CPU端的保持一致：float, float16, double。
测试fp64类型时，报了如下错误：
ExternalError: ACL error, the error code is : 500002. (at /workspace/limin-workspace/npu-paddle/Paddle/paddle/fluid/operators/npu_op_runner.cc:380) 1226: [operator < mean > error]。原因：reciprocal的单测（反向）会自动添加mean op，但是mean的前向ReduceMeanD不支持double和int，注册mean kernel时却支持了fp64和int类型。
解决：将mean_op_npu.cc中对double和int类型的支持去掉。
运行结果
单测运行结果
调用 reciprocal npu kernel 和 reciprocal grad npu kernel
fp32：
fp64：
fp16：
