zhenjing
zhenjing
将队列换成无锁队列 https://github.com/cameron314/concurrentqueue 做过测试。数据如下： 线程池： 1、采用多个子线程池，每个线程池4个并发线程，任务队列采用无锁阻塞队列 2、每个算法句柄绑定特定线程池 +-------------+-----------+----------+----------+----------+--------+--------+---------------+---------------+------------+ | HandleCount | ThreadNum | AVG (ms) | Min (ms) | Max (ms) | MaxCPU | AvgCPU | MaxMemory(MB) | userTimeRatio | throughput...
./benchmark.out models/ 50 0 0 1 0 0 1 1 MNN benchmark Forward type: CPU thread=1 precision=0 sparsity=0 sparseBlockOC=1 testQuantizedModel=1 --------> Benchmarking... loop = 50, warmup = 0 [-INFO-]: precision!=2,...
版本：2.8.4
测试用哪个模型？有参考的开源模型吗？
When rapidjson write String value，it scan all char to deal with Unicode escaping. So string is larger, it costs more time. Refer : rapidjson -> write.h -> bool WriteString(const Ch*...
fixbug: np.int -> int, np.float -> float refer: https://github.com/zhenjing/robotic-grasping
bugfix: evaluate rot, zoom convert to number from tensor refer to: https://github.com/zhenjing/robotic-grasping
> 大概率是 session api 的调用代码问题。fp16 的输入输出必须用 copyFromHost / copyToHost ，不能直接访问 tensor 的 host 指针。建议都用 Module API. Module API比session api慢30%。session api将来不支持fp16吗？
是bug吗？