cudaCheckError() failed : invalid device function
hugobordigoni opened this issue · 10 comments
I spent so much time debugging this issue that I give the answer here:
When running the demo.py as stated in README, I was getting an error cudaCheckError() failed : invalid device function with no traceback. It happen when this line was executed : https://github.com/smallcorgi/Faster-RCNN_TF/blob/master/lib/fast_rcnn/test.py#L169
I have never seen this error in any of my other tensorflow project.
This issue was similar to this one in Faster-RCNN for python : rbgirshick/py-faster-rcnn#2
And i solved it by updating the arch code in https://github.com/smallcorgi/Faster-RCNN_TF/blob/master/lib/make.sh#L9 and https://github.com/smallcorgi/Faster-RCNN_TF/blob/master/lib/setup.py#L137
I don't know how to find the arch code of any GPU, but for Tesla K80, sm_37 seems to work.
I don't know if we can change something so that it works for any GPU or maybe we can add an information in the README?
Hope it can help people having the same issue.
Just adding to this since it was useful to me.
I hit this same problem when testing on AWS EC2 instances with GPU. I had to use sm_20 in two places as mentioned above:
lib/make.sh
lib/setup.py
and force the rebuild of the python modules:
cd lib
python setup.py build_ext --inplace
When I ran $ python ./tools/demo.py --model ./VGGnet_fast_rcnn_iter_70000.ckpt,
I had exactly the same error cudaCheckError() failed : invalid device function
I tried to follow having the sm_37 in lib/make.sh and lib/setup.py. I think the setting is almost there. What can I do? I am using AWS EC2 g2.2xlarge. Below messages say I am using NVIDA GRID K520.
I tensorflow/stream_executor/dso_loader.cc:128] successfully opened CUDA library libcublas.so locally
I tensorflow/stream_executor/dso_loader.cc:128] successfully opened CUDA library libcudnn.so locally
I tensorflow/stream_executor/dso_loader.cc:128] successfully opened CUDA library libcufft.so locally
I tensorflow/stream_executor/dso_loader.cc:128] successfully opened CUDA library libcuda.so.1 locally
I tensorflow/stream_executor/dso_loader.cc:128] successfully opened CUDA library libcurand.so locally
I tensorflow/stream_executor/cuda/cuda_gpu_executor.cc:937] successful NUMA node read from SysFS had negative value (-1), but there must be at least one NUMA node, so returning NUMA node zero
I tensorflow/core/common_runtime/gpu/gpu_device.cc:885] Found device 0 with properties:
name: GRID K520
major: 3 minor: 0 memoryClockRate (GHz) 0.797
pciBusID 0000:00:03.0
Total memory: 3.94GiB
Free memory: 3.91GiB
I tensorflow/core/common_runtime/gpu/gpu_device.cc:906] DMA: 0
I tensorflow/core/common_runtime/gpu/gpu_device.cc:916] 0: Y
I tensorflow/core/common_runtime/gpu/gpu_device.cc:975] Creating TensorFlow device (/gpu:0) -> (device: 0, name: GRID K520, pci bus id: 0000:00:03.0)
Tensor("Placeholder:0", shape=(?, ?, ?, 3), dtype=float32)
Tensor("conv5_3/conv5_3:0", shape=(?, ?, ?, 512), dtype=float32)
Tensor("rpn_conv/3x3/rpn_conv/3x3:0", shape=(?, ?, ?, 512), dtype=float32)
Tensor("rpn_cls_score/rpn_cls_score:0", shape=(?, ?, ?, 18), dtype=float32)
Tensor("rpn_cls_prob:0", shape=(?, ?, ?, ?), dtype=float32)
Tensor("rpn_cls_prob_reshape:0", shape=(?, ?, ?, 18), dtype=float32)
Tensor("rpn_bbox_pred/rpn_bbox_pred:0", shape=(?, ?, ?, 36), dtype=float32)
Tensor("Placeholder_1:0", shape=(?, 3), dtype=float32)
Tensor("conv5_3/conv5_3:0", shape=(?, ?, ?, 512), dtype=float32)
Tensor("rois:0", shape=(?, 5), dtype=float32)
[<tf.Tensor 'conv5_3/conv5_3:0' shape=(?, ?, ?, 512) dtype=float32>, <tf.Tensor 'rois:0' shape=(?, 5) dtype=float32>]
Tensor("fc7/fc7:0", shape=(?, 4096), dtype=float32)
Loaded network ./VGGnet_fast_rcnn_iter_70000.ckpt
cudaCheckError() failed : invalid device function
I got the same error (i.e., cudaCheckError() failed : invalid device function) with my Tesla K40. When I changed the -arch parameter in lib/make.sh to sm_35, and rerun make.sh, it worked.
@ahmedammar you could even use sm_30 for AWS
@eakbas Thanks!
For other GPUs:
# Which CUDA capabilities do we want to pre-build for?
# https://developer.nvidia.com/cuda-gpus
# Compute/shader model Cards
# 6.1 P4, P40, Titan X so CUDA_MODEL = 61
# 6.0 P100 so CUDA_MODEL = 60
# 5.2 M40
# 3.7 K80
# 3.5 K40, K20
# 3.0 K10, Grid K520 (AWS G2)
# Other Nvidia shader models should work, but they will require extra startup
# time as the code is pre-optimized for them.
CUDA_MODELS=30 35 37 52 60 61
Credit to https://github.com/mldbai/mldb/blob/master/ext/tensorflow.mk
@VitaliKaiser Hello, I am using AWS EC2 GPU to run demo.py, getting 'cudaCheckError() failed : invalid device function'.
Since GPU used is GRID K520 on my instance, I follow your post to change -arch parameter in steup.py and make.sh to sm_30 and rerun make.sh, but this error is still there when I run './tools/demo.py --model ./VGG.....ckpt'. Could you please give me some help?
@fangyan93 It´s quite a while since I last looked into it, but I had lost a lot of time to figure out things were not rebuild!
You have delete every binary which is build with the make.sh script. And then build it one more time.
@VitaliKaiser Thanks for reply. Yes, I remove the previous build files and rebuild from very beginning, it works!
So for those who are still lost. Here are a few clean steps to resolve the issue (you need to recompile your CUDA):
- Got to the following page https://developer.nvidia.com/cuda-gpus and find your GPU
- Find the number (the "Compute Capability") next to your GPU name, e.g. for 680 it is 3.0
- Remove the dot from it so it becomes 30
- In the make_cuda.sh file required for compiling, change the number after "arch" flag in the nvcc command to the one you found above. Example:
nvcc -c -o corr_cuda_kernel.cu.o corr_cuda_kernel.cu -x cu -Xcompiler -fPIC -arch=sm_52
nvcc -c -o corr_cuda_kernel.cu.o corr_cuda_kernel.cu -x cu -Xcompiler -fPIC -arch=sm_30 - Delete the folder of built files if you have already compiled it before
- Run the make_cuda.sh and continue as usual