add reduce_all and reduce_any #49765
Conversation
|
你的PR提交成功，感谢你对开源项目的贡献! |
| "keep_dim": keep_dim, | ||
| "dim": dim, | ||
| "reduce_all": reduce_all, |
There was a problem hiding this comment.
这里不能去掉，会报错
ci里面有一个验证(op_teller)是 getAttr 说要没有这些的时候，我想过这个ci，不知道那个是怎么回事。
There was a problem hiding this comment.
现在的覆盖率也是很差，不懂为什么 reduce_any 整个不会进去
| cast_layer->setOutputType(0, nvinfer1::DataType::kFLOAT); | ||
| cast_layer->getOutput(0)->setType(nvinfer1::DataType::kFLOAT); |
There was a problem hiding this comment.
这里用int32是不是更合适，0对应false，1对应true？
There was a problem hiding this comment.
确实，我以为只支持float,刚看了下kFLOAT and kHALF kINT32 kINT8 都行
| outputs=["reduce_output_data"], | ||
| ) |
There was a problem hiding this comment.
输出为bool类型时，需添加outputs_dtype设定输出类型。可参考test_trt_convert_compare_and_logical.py
| "reduce_prod", | ||
| ]: | ||
| dics1 = [ | ||
| for out_dtype in [-1, 0, 2, 5]: |
There was a problem hiding this comment.
generate_input1 也需要添加bool支持
|
使用TRT 8.0+，该PR TRT部分逻辑正确，GPU算子存在问题导致单测无法正常通过。 GPU算子在单独PR修复后该PR可合入。 |
|
Paddle-TRT 下的reduce_all 和 reduce_any 算子已经添加了。 |
PR types
Others
PR changes
Others
Describe
add reduce_all and reduce_any op
#48292