[Rocm] fix test of random_crop_op & logsumexp #32824
Conversation
|
Thanks for your contribution! |
Ray2020BD
changed the title
| @@ -69,6 +70,7 @@ HOSTDEVICE inline void StridedMemcpy(const T* x, const size_t* x_dims, T* out, | |||
| "expected to be 1, but got %ld. Please check input " | |||
| "value.", | |||
| i, rank, out_stride); | |||
| #endif | |||
There was a problem hiding this comment.
看了下你的分析文档，但当前这个修复方式不能满足OP要，我们OP修复的原则是
- 不可以因为这两段代码会挂就不检查x_stride和out_stride，而是需要找到能工作的代码替换这两段代码，或者修复PADDLE_ENFORCE的逻辑使其能够工作
我们修复OP的目的不是为了单测可以通过，而是保证功能的正确性，当前的diff只是简单跳过了x_stride和out_stride的检查，不符合我们OP修复的目的。
There was a problem hiding this comment.
这个直接printf不符合paddle msg的标准，请参考这个文档写关于paddle的error msg
https://github.com/PaddlePaddle/Paddle/wiki/Paddle-Error-Message-Writing-Specification
Ray2020BD
changed the title
| @@ -69,6 +70,7 @@ HOSTDEVICE inline void StridedMemcpy(const T* x, const size_t* x_dims, T* out, | |||
| "expected to be 1, but got %ld. Please check input " | |||
| "value.", | |||
| i, rank, out_stride); | |||
| #endif | |||
There was a problem hiding this comment.
这个直接printf不符合paddle msg的标准，请参考这个文档写关于paddle的error msg
https://github.com/PaddlePaddle/Paddle/wiki/Paddle-Error-Message-Writing-Specification
| dy = np.ones(1, dtype=self.dtype) | ||
| x = self.inputs['X'] | ||
| y = self.outputs['Out'] | ||
| return dy * np.exp(x - y) |
There was a problem hiding this comment.
这里能解释一下为什么get_numeric_gradient得到的结果在ROCm上无法对齐，需要定义user_defined_grads呢？CUDA上为什么可以直接使用get_numeric_gradient？
There was a problem hiding this comment.
logsumexp这个op本身运算出来的analytic_grads是没问题的，有问题的是numeric_grads（cuda下该计算没问题），所以干脆自己计算numeric_grads
| ) else np.float64 | ||
| self.sequence_length = None if fluid.core.is_compiled_with_rocm( | ||
| ) else np.array( | ||
| [12, 11, 10, 9, 8], dtype=np.int32) |
There was a problem hiding this comment.
MIOPEN不支持sequence_length的话，SimpleRNN的API也同样需要修改，对于sequence_length非空的输入需要抛出正确的error msg.
There was a problem hiding this comment.
simpleRNN的修改以及该单测的修改从本pr中删除，将和其他rnn单测修改统一放到其他的pr中。
| if self.is_bidirec: | ||
| for i in range(0, len(flat_w), 4): | ||
| flat_w[i + 1], flat_w[i + 2] = flat_w[i + 2], flat_w[i + 1] | ||
|
|
There was a problem hiding this comment.
这里能解释一下为什么需要对flat_w的数据做交换吗？是miopenSetRNNDescriptor_V2 里面的 miopenRNNbidirection 和 cudnnSetRNNDescriptor_v6 的 CUDNN_BIDIRECTIONAL 的行为不一致？
There was a problem hiding this comment.
miopen相关api的实现中，layout并没有和cuda相关pai对齐，需要手动调节。
Ray2020BD
changed the title
PR types
Bug fixes
PR changes
OPs
Describe
[Rocm] fix test_random_crop_op ( under release-compile )
[Rocm] fix test_logsumexp