【PaddlePaddle Hackathon 4】No.56 : add fp16 test and bf16 for searchsorted #53205
Conversation
|
你的PR提交成功，感谢你对开源项目的贡献! |
|
✅ This PR's description meets the template requirements! |
|
Sorry to inform you that c07c5f5's CIs have passed for more than 7 days. To prevent PR conflicts, you need to re-run all CIs manually. |
paddle/phi/kernels/funcs/algorithm.h
Outdated
| @@ -48,7 +48,7 @@ HOSTDEVICE inline size_t LowerBound(const T1 *x, size_t num, const T2 &val) { | |||
| while (count > 0) { | |||
| int64_t step = (count >> 1); | |||
| auto *it = first + step; | |||
| if (*it < val) { | |||
| if (static_cast<float>(*it) < static_cast<float>(val)) { | |||
There was a problem hiding this comment.
建议不要直接转换为float，最好使用文档里提供的MyTypeTrait来判断需要转换的类型
paddle/phi/kernels/funcs/algorithm.h
Outdated
| @@ -71,7 +71,7 @@ HOSTDEVICE inline size_t UpperBound(const T1 *x, size_t num, const T2 &val) { | |||
| while (count > 0) { | |||
| auto step = (count >> 1); | |||
| auto *it = first + step; | |||
| if (val < *it) { | |||
| if (static_cast<float>(val) < static_cast<float>(*it)) { | |||
There was a problem hiding this comment.
建议不要直接转换为float，最好使用文档里提供的MyTypeTrait来判断需要转换的类型
| @@ -215,5 +219,46 @@ def test_sortedsequence_values_type_error(): | |||
| self.assertRaises(TypeError, test_sortedsequence_values_type_error) | |||
|
|
|||
|
|
|||
| class TestSearchSortedFP16OP(TestSearchSorted): | |||
|
Sorry to inform you that 5134014's CIs have passed for more than 7 days. To prevent PR conflicts, you need to re-run all CIs manually. |
|
Sorry to inform you that 33ab075's CIs have passed for more than 7 days. To prevent PR conflicts, you need to re-run all CIs manually. |
| } else if (type == phi::DataType::FLOAT16) { | ||
| visitor.template apply<int16_t>(); | ||
| } else if (type == phi::DataType::BFLOAT16) { | ||
| visitor.template apply<uint16_t>(); |
There was a problem hiding this comment.
这里如果是phi::DataType::FLOAT16，apply应该是需要写成phi::dtype::float16
如果是phi::DataType::BFLOAT16，应该是phi::dtype::bfloat16
|
@zhangting2020 辛苦帮忙看看这个问题 一开始报错如下：
经分析是IsInf和IsNan函数被重载了多次,编译器无法区分调用哪个重载版本，于是针对float16和bfloat16类型做出如下修改：
但还是无法进行解决
|
PR types
Others
PR changes
APIs
Description
相关链接：
#51281
#54871