Using OpenCL target in TVM auto-scheduler

Troubleshooting
#1

Hi,

I am trying to use TVM’s new auto-scheduler feature according to Auto-scheduling Matrix Multiplication for CPU — tvm 0.8.dev0 documentation

This works fine, but I would like to use OpenCL as a backend. When changing

target = tvm.target.Target("llvm")

to

target = tvm.target.Target("opencl")

I get the following error:

Traceback (most recent call last):
  File "tune_matmul.py", line 54, in <module>
    task = tvm.auto_scheduler.SearchTask(func=matmul, args=(N, L, M, "float32"), target=target)
  File "/tvm/python/tvm/auto_scheduler/search_task.py", line 241, in __init__
    layout_rewrite_option,
  File "/tvm/python/tvm/_ffi/_ctypes/object.py", line 136, in __init_handle_by_constructor__
    handle = __init_by_constructor__(fconstructor, args)
  File "/tvm/python/tvm/_ffi/_ctypes/packed_func.py", line 260, in __init_handle_by_constructor__
    raise get_last_ffi_error()
tvm._ffi.base.TVMError: Traceback (most recent call last):
  [bt] (4) /tvm/build/libtvm.so(TVMFuncCall+0x48) [0x2ad689d596d8]
  [bt] (3) /tvm/build/libtvm.so(+0x793c89) [0x2ad6891bfc89]
  [bt] (2) /tvm/build/libtvm.so(tvm::auto_scheduler::SearchTask::SearchTask(tvm::auto_scheduler::ComputeDAG, tvm::runtime::String,     tvm::Target, tvm::Target, tvm::runtime::Optional<tvm::auto_scheduler::HardwareParams>, tvm::auto_scheduler::LayoutRewriteOption)+0x36a) [0x2ad6891bf38a]
  [bt] (1) /tvm/build/libtvm.so(tvm::auto_scheduler::HardwareParamsNode::GetDefaultHardwareParams(tvm::Target const&, tvm::Target     const&)+0x6a8) [0x2ad6891bee18]
  [bt] (0) /tvm/build/libtvm.so(+0x7924f0) [0x2ad6891be4f0]
  File "/tvm/src/auto_scheduler/search_task.cc", line 107
TVMError: No default hardware parameters for opencl target device: ""

Can you help me sort this out? Many thanks in advance!

#2

Can you try create

target_host = tvm.target.Target("llvm")

and when search task, add the target_host

task = tvm.auto_scheduler.SearchTask(func=matmul_add, args=(N, L, M, "float32"), target=target, target_host=target_host)

#3

We only implement Mali / ROCM default parameter for now. you could refer the note part Auto-scheduling a Neural Network for mali GPU — tvm 0.8.dev0 documentation how to set your device opencl hardware parameter.

#4

This works, many thanks!

#5

Do you have any ideas about setting hardware parameters for adreno gpus? Thank you

#6

Can I use opencl for amd gpu and not use rocm, now?

#7

Auto-scheduling a Neural Network for mali GPU — tvm 0.15.dev0 documentation Please refer Note part: How to get the hardware parameters from remote device

#8

of course yes, but be careful auto scheduler doesn’t consider matrix core.

#9

Thanks for your reply. But, there are two different kind of amd gpu on the platform and I want to one of them for auto shedule. So, how can I set the target when using auto shedule for tuning.

Thanks.

#10

Adreno GPUs give misleading information here. They report their warp/wave front sizes as 1.