Hecmay opened a new issue #6634:
URL: https://github.com/apache/incubator-tvm/issues/6634
## Environment
* Ubuntu 18.04.5 LTS
* CUDA Version: 11.0.
* GPU: RTX 2080 Ti. Compute Compatibility : 7.5
## Description
I created a simple MatMul tensor expression in C++, and want to use Anosr to
search the optimal schedule for it. Here is the skeleton of the program that I
am using. The `args` is an array of tensor inputs and output of the MatMul op.
```c++
// Create DAG and search task
const auto& dag = tvm::auto_scheduler::ComputeDAG(args);
auto task = SearchTask(dag, "test", Target("cuda"), Target("llvm"),
Optional<HardwareParams>());
// Create tuning options and search policy
auto options = TuningOptions(num_measure_trials, early_stopping,
num_measures_per_round, verbose, builder,
runner,
Optional<Array<MeasureCallback>>());
auto policy = SketchPolicy(task, cost_model, params, seed, verbose,
callbacks);
// Launch Ansor
std::pair<te::Schedule, Array<te::Tensor>> res = AutoSchedule(policy,
options);
```
When running the program, `measure.py` complains that the GPU architecture
cannot be found. The error message is as followed. I tried some solutions in
this link:
https://discuss.tvm.apache.org/t/solved-compile-error-related-to-autotvm/804/11.
But none of them worked in my case.
```shell
------------------------------------------------------------
------------------------- [ Search ]
------------------------------------------------------------
Generate Sketches #s: 1
Sample Initial Population #s: 744 fail_ct: 3352 Time elapsed: 0.39
GA Iter: 0 Max score: 0.0000 Min score: 0.0000 #Pop: 744
#M+: 0 #M-: 0
GA Iter: 5 Max score: 0.0000 Min score: 0.0000 #Pop: 2048
#M+: 1446 #M-: 89
GA Iter: 10 Max score: 0.0000 Min score: 0.0000 #Pop: 2048
#M+: 1573 #M-: 94
EvolutionarySearch #s: 128 Time elapsed: 3.50
------------------------------------------------------------
------------------------- [ Measure ]
------------------------------------------------------------
Get 10 programs for measure. (This may take a while)
.E.E.E.E.E.E.E.E.E.E
==================================================
No: 1 GFLOPS: 0.00 / 0.00 results:
MeasureResult(error_type:CompileHostError, error_msg:Traceback (most recent
call last):
File "/home/sx/dlcb/build/tvm/python/tvm/auto_scheduler/measure.py", line
516, in timed_func
sch, args, target=task.target, target_host=task.target_host
File "/home/sx/dlcb/build/tvm/python/tvm/driver/bu
...
et_arch)
File "/home/sx/dlcb/build/tvm/python/tvm/contrib/nvcc.py", line 71, in
compile_cuda
raise ValueError("arch(sm_xy) is not passed, and we cannot detect it
from env")
ValueError: arch(sm_xy) is not passed, and we cannot detect it from env
, all_cost:0.02, Tstamp:1601955085.48)
==================================================
```
I was able to run the Ansor tutorial, so I guess my env settings should be
correct. I also tried to specify the architecture string (i.e. `sm_75` in my
case). Ansor will throw the following error when I do so:
```shell
==================================================
No: 4 GFLOPS: 0.00 / 0.00 results:
MeasureResult(error_type:RuntimeDeviceError, error_msg:Traceback (most recent
call last):
File "/home/sx/dlcb/build/tvm/python/tvm/auto_scheduler/measure.py", line
672, in timed_func
ndarray.empty(get_const_tuple(x.shape), x.dtype, ctx) for x in
build_res.args
File "/home/sx/dlcb/build/tvm/py
...
] (0) /home/sx/dlcb/build/tvm/build/libtvm.so(+0x1791844) [0x7fd1ecfb3844]
File "/home/sx/dlcb/build/tvm/src/runtime/cuda/cuda_device_api.cc", line
115
CUDA: Check failed: e == cudaSuccess || e == cudaErrorCudartUnloading:
initialization error
, all_cost:0.46, Tstamp:1601956508.32)
==================================================
```
----------------------------------------------------------------
This is an automated message from the Apache Git Service.
To respond to the message, please log on to GitHub and use the
URL above to go to the specific comment.
For queries about this service, please contact Infrastructure at:
[email protected]
