使用MNN2QNNModel 报错Segmentation fault (core dumped)

[pipiyaoo]

我当前的一个静态的onnx模型，经过MNNConvert转换之后，再使用MNN2QNNModel转成QNN的产物，报错
param dims: 1x3x32x900
param dims: 1x3x32x900
Total input shape type size:1
[Temp Product]: Qnn temp product generate at /.../.../...
CPU Group: [ 17 18 16 19 ], 800000 - 3600000
CPU Group: [ 0 3 5 7 15 1 12 14 2 4 6 13 ], 800000 - 4800000
CPU Group: [ 9 10 11 8 ], 800000 - 4900000
The device supports: i8sdot:0, fp16:0, i8mm: 0, sve2: 0, sme2: 0
Can't Find type=5 backend, use 0 instead
Load Cache file error.
input 0 shape:1 3 32 900
Segmentation fault (core dumped)

我的MNN版本是3.3.0
QNN版本是2.40

[Qxinyu]

可以先把MNN2QNNModel.cpp的258行修改下：
config.type = MNN_FORWARD_NN;
改为：
config.type = MNN_CONVERT_QNN;
后续我们会修改MNN2QNNModel的转换工具。

[pipiyaoo]

可以先把MNN2QNNModel.cpp的258行修改下： config.type = MNN_FORWARD_NN; 改为： config.type = MNN_CONVERT_QNN; 后续我们会修改MNN2QNNModel的转换工具。

修改了之后，执行MNN2QNNModel工具，还是报错
[Pass]: qnn-model-lib-generator success!
qnn-context-binary-generator pid:79124
Initializing HtpProvider
0.0ms [ ERROR ] Dimension at index 1 cannot be broadcast between in[0] (192) and in[1] (225).

 0.0ms [ ERROR ] Dimension at index 1 cannot be broadcast between in[0] (192) and in[1] (225).

 0.0ms [ ERROR ] Dimension at index 1 cannot be broadcast between in[0] (192) and in[1] (225).

 0.0ms [ ERROR ] Dimension at index 1 cannot be broadcast between in[0] (192) and in[1] (225).

 0.0ms [ ERROR ] Op specific validation failed.

189.6ms [ ERROR ] validateNativeOps master op validator BinaryOp_I_9_10_O_12:qti.aisw:ElementWiseAdd failed 3110

189.6ms [ ERROR ] QnnBackend_validateOpConfig failed 3110

189.6ms [ ERROR ] Failed to validate op BinaryOp_I_9_10_O_12 with error 0xc26

[ ERROR ] QnnModel::addNode() validating node BinaryOp_I_9_10_O_12 failed.
[ ERROR ] all_in_one_rec_quant_900_0.addNode(QNN_OPCONFIG_VERSION_1, "BinaryOp_I_9_10_O_12", "qti.aisw", "ElementWiseAdd", params_BinaryOp_I_9_10_O_12, 0, inputs_BinaryOp_I_9_10_O_12, 2, outputs_BinaryOp_I_9_10_O_12, 1) expected MODEL_NO_ERROR, got MODEL_GRAPH_OP_VALIDATION_ERROR
189.6ms [ ERROR ] Failed in composeGraphs()
189.7ms [ ERROR ] ComposeGraphs Failed with error = 1
Graph Compose failure
Backend 1 free cleanup called during process exit
[Error]: qnn-context-binary-generator error!

qnn-model-lib-generator使用成功了，但是qnn-context-binary-generator 失败

[Qxinyu]

方便提供模型用来测试吗

[pipiyaoo]

方便提供模型用来测试吗

这个确实不方便，还有个问题，如果我使用MNNConvert --saveStaticModel true --inputConfigFile /config.txt 转换出来一个静态的模型，再使用quantized.out进行量化，json文件如下
{
"format":"BGR",
"mean":[0.5, 0.5, 0.5],
"normal":[0.5, 0.5, 0.5],
"width":900,
"height":32,
"path":"/imgs/",
"feature_quantize_method":"KL",
"weight_quantize_method":"ADMM",
"input_type":"image"
}

也是会报段错误
CPU Group: [ 17 18 16 19 ], 800000 - 3600000
CPU Group: [ 0 3 5 7 15 1 12 14 2 4 6 13 ], 800000 - 4800000
CPU Group: [ 9 10 11 8 ], 800000 - 4900000
The device supports: i8sdot:0, fp16:0, i8mm: 0, sve2: 0, sme2: 0
[11:07:06] /home/host/MNN-master/tools/quantization/Helper.cpp:137: used dataset num: 1000
[11:07:07] /home/host/MNN-master/tools/quantization/calibration.cpp:1053: fake quant weights done.
Segmentation fault (core dumped)

请问这又是什么原因呢

[Qxinyu]

能获取到崩溃的堆栈信息吗

[pipiyaoo]

能获取到崩溃的堆栈信息吗

并没有，只有段错误这一行