I’m trying to run TVM on Snapdragon board locally via Termux, since using RPC does not work for me.
To enable LLVM, I pkg-installed llvm, clang, llvm-static, etc… To enable OpenCL, I copied libOpenCL.so from /system/vendor/lib to termux local directory (since termux does not have root permission) and OpenCL libraries offered from Qualcomm developer website.
As a result, TVM local build on the board succeeded while enabling LLVM and OpenCL.
However, if I try to run tutorials, cpu target works:
target=tvm.target.Target('llvm -mtriple=aarch64-unknown-linux-android')
while opencl target 'opencl' or 'opencl -device=adreno' does not work:
Cannot find config for target=opencl -keys=opencl,gpu -max_num_threads=256, workload=('dense_small_batch.cuda', ('TENSOR', (1, 512), 'float32'), ('TENSOR', (1000, 512), 'float32'), None, 'float32'). A fallback configuration is used, which may bring great performance regression.
[09:09:33] /data/data/com.termux/files/home/tvm/src/runtime/opencl/opencl_device_api.cc:240: No OpenCL platform matched given existing options ...
/data/data/com.termux/files/home/tvm/tutorials/get_started/relay_quick_start.py:102: DeprecationWarning: legacy graph runtime behavior of producing json / lib / params will be removed in the next release. Please see documents of tvm.contrib.graph_runtime.GraphModule for the new recommended usage.
g,m,p = relay.build(mod, target, target_host, params=params)
[09:09:34] /data/data/com.termux/files/home/tvm/src/runtime/opencl/opencl_device_api.cc:240: No OpenCL platform matched given existing options ...
Traceback (most recent call last):
File "/data/data/com.termux/files/home/tvm/tutorials/get_started/relay_quick_start.py", line 114, in <module>
module = graph_runtime.create(g, m, ctx)
File "/data/data/com.termux/files/home/tvm/python/tvm/contrib/graph_runtime.py", line 66, in create
return GraphModule(fcreate(graph_json_str, libmod, *device_type_id))
File "/data/data/com.termux/files/home/tvm/python/tvm/_ffi/_ctypes/packed_func.py", line 237, in __call__
raise get_last_ffi_error()
tvm._ffi.base.TVMError: Traceback (most recent call last):
File "/data/data/com.termux/files/home/tvm/src/runtime/opencl/opencl_device_api.cc", line 122
TVMError:
---------------------------------------------------------------
An internal invariant was violated during the execution of TVM.
Please read TVM's error reporting guidelines.
More details can be found here: https://discuss.tvm.ai/t/error-reporting/7793.
---------------------------------------------------------------
Check failed: context != nullptr == false: No OpenCL device
It seems that TVM can build relay module relay.build but the error occurs at the line calling opencl context that I defined as ctx=tvm.cl().
I first thought that termux might not detect OpenCL but running clinfo on termux shows as below:
~/clinfo $ ./clinfo
Number of platforms 1
Platform Name QUALCOMM Snapdragon(TM)
Platform Vendor QUALCOMM
Platform Version OpenCL 2.0 QUALCOMM build: commit #ffa098c changeid #I3cb6626e76 Date: 01/16/20 Thu Local Branch: Remote Branch:
Platform Profile FULL_PROFILE
Platform Extensions
Platform Name QUALCOMM Snapdragon(TM)
Number of devices 1
Device Name QUALCOMM Adreno(TM)
Device Vendor QUALCOMM
Device Vendor ID 0xbf4d3c4b
Device Version OpenCL 2.0 Adreno(TM) 650
Driver Version OpenCL 2.0 QUALCOMM build: commit #ffa098c changeid #I3cb6626e76 Date: 01/16/20 Thu Local Branch: Remote Branch: Compiler E031.37.07.00
Device OpenCL C Version OpenCL C 2.0 Adreno(TM) 650
Device Type GPU
Does anyone have an idea about this issue?