Hello everyone! I run the benchmark on an Android device with RPC(adreno_gpu_bench_texture. py). Compared to [Benchmark]（ https://github.com/apache/tvm/wiki/Benchmark ）The test runing time cost almost 20 times than benchmark. My device is also 8Gen1, and the build target="opencl "(Native OpenCL). According to the document: mobilenetv1, it takes 3.82ms (fp32) and 2.72ms (fp16), while on my device it is 66.93ms (fp32) and 56.59ms (fp16). When target=“opencl device=adneno” (OpenCLML), it raise Error: CL_OUT_OF_HOST_MEMORY. I would like to know what is causing this? How to solve it? Device: xiaomi 12p(8Gen1)/Ubuntu20.04
我使用RPC方式在安卓设备上运行TVM的benchmark(adreno_gpu_bench_texture.py)。相比于Benchmark提供的数据，我测试运行的时间要多20倍左右的时间。我的设备同样是8Gen1,我构建的target=‘opencl’ （Native OpenCL），根据文档mobilenetv1花费3.82ms(fp32) 2.72ms(fp16)，我设备上是66.93ms(fp32) 56.59ms(fp16)。当target=“opencl -device=adreno” (OpenCLML)会出现Error: CL_OUT_OF_HOST_MEMORY无法运行。我想知道这是什么原因导致的？该怎么解决呢？ 设备：小米12p（8Gen1）/Ubuntu20.04