loongarch使用PP-HumanSeg模型数据输出错误

[sam1234561]

系统架构：loongarch64
版本：Paddle-Lite V2.12，结合#10624patch修改编译
模型：human_pp_humansegv2_mobile_192x192_inference_model_with_softmax
测试demo：mobilenetv1_full

loogarch64运行结果：
run_idx:1 / 10: 38.455 ms
run_idx:2 / 10: 38.395 ms
run_idx:3 / 10: 40.375 ms
run_idx:4 / 10: 38.889 ms
run_idx:5 / 10: 38.832 ms
run_idx:6 / 10: 39.333 ms
run_idx:7 / 10: 38.78 ms
run_idx:8 / 10: 54.512 ms
run_idx:9 / 10: 38.784 ms
run_idx:10 / 10: 38.931 ms

======= benchmark summary =======
input_shape(s) (NCHW):1,3,224,224, :
model_dir:./model/human_pp_humansegv2_mobile_192x192_inference_model_with_softmax
warmup:10
repeats:10
*** time info(ms) ***
1st_duration:42.76
max_duration:54.512
min_duration:38.395
avg_duration:40.5286

====== output summary ======
output tensor num:1

--- output tensor 0 ---
output shape(NCHW):1 2 224 224
output tensor 0 elem num:100352
output tensor 0 standard deviation:1.47064e-05
output tensor 0 mean value:9.96485e-06

x86_64运行结果：
run_idx:1 / 10: 62.989 ms
run_idx:2 / 10: 83.378 ms
run_idx:3 / 10: 89.691 ms
run_idx:4 / 10: 72.524 ms
run_idx:5 / 10: 71.135 ms
run_idx:6 / 10: 82.588 ms
run_idx:7 / 10: 70.5 ms
run_idx:8 / 10: 67.89 ms
run_idx:9 / 10: 72.705 ms
run_idx:10 / 10: 73.236 ms

======= benchmark summary =======
input_shape(s) (NCHW):1,3,224,224, :
model_dir:human_pp_humansegv2_lite_192x192_inference_model_with_softmax/
warmup:10
repeats:10
*** time info(ms) ***
1st_duration:725.606
max_duration:89.691
min_duration:62.989
avg_duration:74.6636

====== output summary ======
output tensor num:1

--- output tensor 0 ---
output shape(NCHW):1 2 224 224
output tensor 0 elem num:100352
output tensor 0 standard deviation:0.250927
output tensor 0 mean value:0.5

对比结果：standard deviation、mean value数据应该是异常的

[ddchenhao66]

精度有问题，可以通过二分法切图来排查是哪个算子引入的问题。使用 Netron 可视化模型，配合 split_model.py 脚本实现切图。