[Paddle TensorRT No.9-10] Add pd_op.(argmin,argsort) converter
#69261
Conversation
… add_pirtensort_converter_argmin_argsort
|
你的PR提交成功，感谢你对开源项目的贡献! |
ooooo-create
changed the title
pd_op.(argmin,argsort) converter
ooooo-create
changed the title
pd_op.(argmin,argsort) converterpd_op.(argmin,argsort) converter
| return false; | ||
| } | ||
| auto x = op.x(); | ||
| auto x_tensor_type = x.type().dyn_cast<paddle::dialect::DenseTensorType>(); |
There was a problem hiding this comment.
获取dtype使用pir::GetDataTypeFromValue(x)，可参考ScaleOpPattern
There was a problem hiding this comment.
有如流的账号，但是不知道怎么加，手机vx是 18268023940
| data_type == phi::DataType::FLOAT64)) { | ||
| return false; | ||
| } | ||
| int axis = static_cast<int>(op.axis() |
There was a problem hiding this comment.
这里首先判断下，pir::GetDefiningOpForInput(op,1)->isapaddle:::dialect::FullOp，然后再去做下面的限制
There was a problem hiding this comment.
前面已经判断过了，如果不是，就返回 false
| phi::DataType dtype = | ||
| op.attribute<paddle::dialect::DataTypeAttribute>("dtype").data(); | ||
| if (axis == 0 || flatten || | ||
| (dtype != phi::DataType::INT32 && dtype != phi::DataType::INT64)) |
There was a problem hiding this comment.
这里面加一个VLOG(3)的打印，pd_op.argmin因为什么条件不能进入trt
| op->set_attribute(kCanRunTrtAttr, rewriter.bool_attr(true)); | ||
| return true; | ||
| } | ||
| }; |
| const std::vector<std::string> required_attrs = {"axis", "descending"}; | ||
| for (const auto &attr : required_attrs) { | ||
| if (!op->HasAttribute(attr)) { | ||
| VLOG(3) << "Argsort " << attr << " attribute does not exist"; |
| axis += x_shape.size(); | ||
| } | ||
| if (x_shape[axis] > 3840 || x_shape[axis] < 0) { | ||
| VLOG(3) << "The axis dim of input should be less than 3840 and greater " |
| x = inputs[0] | ||
| input_dims = x.shape | ||
| rank = len(input_dims) | ||
| axis = int( |
There was a problem hiding this comment.
这里还需要支持axis为pir::value的输入，也需要进入trt，同理可以把pd_op.argmax补充一下
There was a problem hiding this comment.
Marker Pass 除了 full op 产生的 Value 情况会不进入 Tensorrt
| descending = paddle_op.attrs().get("descending", False) | ||
| if axis < 0: | ||
| axis += len(input_shape) | ||
| topk_op = trt.TopKOperation.MAX if descending else trt.TopKOperation.MIN |
There was a problem hiding this comment.
这里converter可以参考argsort_op.cc，这里应该是少了很多情况
| @@ -51,6 +63,64 @@ def test_trt_result(self): | |||
| self.check_trt_result() | |||
|
|
|||
|
|
|||
| class TestArgminCase2TRTPattern(TensorRTBaseTest): | |||
There was a problem hiding this comment.
这里再补充一下axis为pir::value的场景，使用np.array([1])，feed_list中加入axis，但是min_shape,和max_shape不需要写
There was a problem hiding this comment.
Marker Pass 除了 full op 产生的 Value 情况会不进入 Tensorrt
| def setUp(self): | ||
| self.python_api = paddle.argsort | ||
| self.api_args = { | ||
| "x": np.random.randn(2, 3).astype(np.float32), |
| self.check_trt_result() | ||
|
|
||
|
|
||
| class TestArgsortCase4TRTPattern(TensorRTBaseTest): |
There was a problem hiding this comment.
这个其实不用再加一个吧，能不能把self.check_marker(expected_result=False)放在前面某个单测里面检查下
There was a problem hiding this comment.
前面的 chekc_marker 应该都是 True 的吧，False 可以放在前面吗
PR Category
Inference
PR Types
New features
Description
新增了 pd_op.argmix Marker 和 Converter
新增了 pd_op.argsort Marker 和 Converter
links: 【开源任务】PIR-TensorRT converter推全升级 #69178