Floating point exception (core dumped) ？

[brilliant-yan]

What are the problems?(screenshots or detailed error messages)

floating point exception (core dumped)
( Debug: cta_num_limit_by_smem == 0 in ppl.nn/deps/ppl.kernel.cuda/src/nn/conv/conv_jit.cc on 3090？)

What are the types of GPU/CPU you are using?

3090

What's the operating system ppl.nn runs on?

ubuntu20.04

What's the compiler and its version?

nvcc 11.2

Which version(commit id or tag) of ppl.nn is used?

current master

What are the commands used to build ppl.nn?

./build.sh -DPPLNN_USE_X86_64=ON -DPPLNN_USE_CUDA=ON -DPPLNN_ENABLE_PYTHON_API=ON

What are the execution commands?

pplnn-build/tools/pplnn --onnx-model ./tests/testdata/mnasnet0_5.onnx --use-cuda --warmup-iterations 500 --enable-profiling

minimal code snippets for reproducing these problems(if necessary)

models and inputs for reproducing these problems (send them to openppl.ai@hotmail.com if necessary)

[brilliant-yan]

如果关掉这个选项 -DPPLNN_ENABLE_CUDA_JIT=OFF，就不会有上面的错误。但是不打开这个选项，性能是不是也会有所下降，我测试了benchmark，各个模型的性能都不如tensorrt好

[qiumeng6]

我也遇到类似的问题，还不知道如何解决。

[jianfei-wangg]

我也遇到类似的问题，还不知道如何解决。
Hello, I have no RTX3090 to be tested. And I tested mnasnet0_5.onnx on T4 and A100 with cuda=11.2 and JIT=ON, but these all works fine. I suggest you to substitute conv_jit.cc:Line599~600 with:
int cta_num_limit_by_regs = (regs_per_cta == 0) ? cta_num_limit_by_thds : max_regs_per_sm / regs_per_cta; int cta_num_limit_by_smem = (smem_per_cta == 0) ? cta_num_limit_by_thds : max_smem_per_sm / smem_per_cta;
and then test again. Please report the result on RTX3090, thx.

[brilliant-yan]

我也遇到类似的问题，还不知道如何解决。
Hello, I have no RTX3090 to be tested. And I tested mnasnet0_5.onnx on T4 and A100 with cuda=11.2 and JIT=ON, but these all works fine. I suggest you to substitute conv_jit.cc:Line599~600 with:
int cta_num_limit_by_regs = (regs_per_cta == 0) ? cta_num_limit_by_thds : max_regs_per_sm / regs_per_cta; int cta_num_limit_by_smem = (smem_per_cta == 0) ? cta_num_limit_by_thds : max_smem_per_sm / smem_per_cta;
and then test again. Please report the result on RTX3090, thx.

Hi,
I tried your suggestion and modified the code, but the problem still persists on 3090. In addition, there is no problem in my multiple tests on the T4 and Jetson series.