[PHI Kernels] add pad op for xpu #53684
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. |
| #include "paddle/phi/kernels/pad_grad_kernel.h" | ||
|
|
||
| #include "paddle/phi/backends/xpu/enforce_xpu.h" | ||
| #include "paddle/phi/backends/xpu/xpu_context.h" |
| pad_left, | ||
| pad_right, | ||
| value); | ||
| PADDLE_ENFORCE_XDNN_SUCCESS(r, "pad"); |
There was a problem hiding this comment.
请问这里不应该是和实际调用的XDNN API保持一致吗，pad_grad_kernel里面实际调用的也是xpu::pad
| class TestPadOpError(unittest.TestCase): | ||
| def test_errors(self): | ||
| with paddle.fluid.framework._static_guard(): | ||
| with program_guard(Program(), Program()): | ||
| input_data = np.random.random((2, 2)).astype("float32") | ||
|
|
||
| def test_Variable(): | ||
| paddle.nn.functional.pad(x=input_data, pad=[1, 1, 1, 1]) | ||
|
|
||
| self.assertRaises(TypeError, test_Variable) | ||
|
|
||
| data = paddle.static.data( | ||
| name='data', shape=[4], dtype='float16' | ||
| ) | ||
| paddle.nn.functional.pad(x=data, pad=[0, 1]) |
There was a problem hiding this comment.
对于xpu算子允许范围以外的数据类型，我不确定单测里需不需要测，需要看看@ykkk2333怎么说
There was a problem hiding this comment.
XPU实际是支持FP16的，已经加上了
| fc = paddle.nn.Linear(4, 10) | ||
| x = paddle.randn([2, 4]) | ||
| x.stop_gradient = False | ||
| feat = fc(x) # [2,3,10] | ||
|
|
||
| out = self.call_func(feat) | ||
|
|
||
| sgd = paddle.optimizer.SGD() | ||
| sgd.minimize(paddle.mean(out)) | ||
| self.assertTrue(self.var_prefix() in str(main_prog)) | ||
| exe = paddle.static.Executor(paddle.XPUPlace(0)) | ||
| exe.run(starup_prog) | ||
| res = exe.run(fetch_list=[feat, out]) |
There was a problem hiding this comment.
这里为什么要实际跑一次训练？上面TestCase1-3测的和这里有什么不同？
There was a problem hiding this comment.
这个是我看CPU和GPU的单测里面有就加上了，看内容似乎是和框架里面的其他部分一起做的一个小的集成测试？
PR types
Others
PR changes
OPs
Description
为XPU添加pad_kernel及对应的反向算子
nn.functional.pad在
mode="constant"和len(pad) == dim * 2的情况下会调用pad op。该Op在CPU和GPU中复用了paddle/phi/kernels/func/padding.h中基于EigenTensor的实现，EigenTensor不支持XPU，需要调用xdnn的pad api实现。
单测部分的代码参考了https://github.com/PaddlePaddle/Paddle/blob/develop/python/paddle/fluid/tests/unittests/test_pad_op.py