Problem on tutorial get_started.py

[vweilv]

When I run the get_started.py tutorial, the following errors occur:

/tools/tvm/dmlc-core/include/dmlc/logging.h:308: [10:48:17] src/codegen/codegen.cc:27: Check failed: bf != nullptr Target cuda is not enabled

It seems that cuda is NOT enabled when tvm is compiled. But I indeed enabled cuda when tvm is compiled. The following is the config.mk file of Tvm:

USE_CUDA = 1
CUDA_PATH = /usr/local/cuda
USE_CUDNN = 1

One more thing which is sure is that cuda is correctly installed at /usr/local/cuda. Other softwares such as torch\tensorflow can run correctly with cuda.

So, what is the problem? Thank you very much.

Other:
OS: ubuntu 14.04
Cuda: 8.0
CuDnn: 5.0

[tqchen]

try to do a make clean, and build again. The error message indicates that the CUDA is not enabled, it means this file never get compiled https://github.com/dmlc/tvm/blob/master/src/codegen/build_cuda.cc#L11

[vweilv]

Yeah, the problem is CUDA is not enabled. I have re-built it several times before I post this issue. I also traced the runtime registry process in which llvm, OpenCL, verilog are all registered except CUDA.

Maybe the last thing I could do is to re-clone it from GitHub.

[vweilv]

I solved the problem by directly revising the Makefile.

I now still do not know why the following code in Makefile will NEVER succeed: ifeq ($(USE_CUDA), 1). When I print the value of USE_CUDA, the output is real 1, but it still will NOT execute the ifeq branch and so the -DTVM_CUDA_RUNTIME=0 will always be set.

Really amazing... Anyone can tell me why?

Anyway, it is a good news. Thank you Tianqi very much.

[tqchen]

hmm, i do not know what is happening on your makefile, please let us know if you find a good fix to this issue

[tqchen]

for now i am going to close this since the original issue has been resolved

[vweilv]

I know why ifeq will always fail: there is a blank after USE_CUDA = 1! I'm used to delete some characters in vi by x, after which the cursor will forward one character. At this time, if I continue to insert words with out any additional operation, a blank will exist.

So naive mistake!

[binbinmeng]

tvm/src/codegen/codegen.cc:27: Check failed: bf != nullptr Target llvm -device=tracing is not enabled
But the log of make is :
-- Build with RPC support...
-- Build with Graph runtime support...
-- Build VTA runtime with target: sim
-- Found CUDA_TOOLKIT_ROOT_DIR=/usr/local/cuda-9.0
-- Found CUDA_CUDA_LIBRARY=/usr/local/cuda-9.0/targets/x86_64-linux/lib/stubs/libcuda.so
-- Found CUDA_CUDART_LIBRARY=/usr/local/cuda-9.0/lib64/libcudart.so
-- Found CUDA_NVRTC_LIBRARY=/usr/local/cuda-9.0/lib64/libnvrtc.so
-- Found CUDA_CUDNN_LIBRARY=/usr/local/cuda-9.0/lib64/libcudnn.so
-- Found CUDA_CUBLAS_LIBRARY=/usr/local/cuda-9.0/lib64/libcublas.so
-- Build with CUDA support
-- Build with cuDNN support
-- Build with cuBLAS support
-- Use llvm-config=/usr/bin/llvm-config-6.0
-- /usr/lib/llvm-6.0/include
-- Found LLVM_INCLUDE_DIRS=/usr/lib/llvm-6.0/include
-- Found LLVM_DEFINITIONS= -DNDEBUG -D_GNU_SOURCE -D__STDC_CONSTANT_MACROS -D__STDC_FORMAT_MACROS -D__STDC_LIMIT_MACROS
-- Found TVM_LLVM_VERSION=60
-- Build with LLVM
-- Set TVM_LLVM_VERSION=60
-- Configuring done
-- Generating done
So, it is very strage. Do you have any idea ?