[GpuOclRuntime] Avoid error when iterating over non-gpu devices

[akroviakov]

Consider these devices visible to OpenCL (CPU and GPU):

Platform: Intel(R) OpenCL
  Device: Intel(R) Xeon(R) Gold 6438Y+
Platform: Intel(R) OpenCL Graphics
  Device: Intel(R) Data Center GPU Max 1100

If you try to run the mlp.mlir example in the current main, it would display an error (-1) for CPU, but would still run the test. The error is displayed due to reading all devices using CL_DEVICE_TYPE_GPU. Instead, we can first read all devices using CL_DEVICE_TYPE_ALL and then use CL_DEVICE_TYPE to select GPUs, this way, no error is displayed.

[AndreyPavlenko]

Does it mean, that clGetDeviceIDs(platform, CL_DEVICE_TYPE_ALL, 0, nullptr, &numDevices) returns error? Could it be an environment configuration issue?
Btw, mlp.mlir is running by CI and it does not fail.

[AndreyPavlenko]

Sorry, I closed it accidently.

[kurapov-peter]

What's the use case for running CPU through opencl anyway? The runtime is not supposed to be called during CPU execution.

[akroviakov]

I build like this:

git clone ...
./scripts/compile.sh --dev --imex
cd build
cmake --build . --target gc-check

then I run mlp.mlir like this:

/home/.../graph-compiler/build/bin/gc-gpu-runner --shared-libs=/home/.../graph-compiler/externals/llvm-project/build/lib/libmlir_runner_utils.so /home/.../graph-compiler/test/mlir/test/gc/gpu-runner/mlp.mlir

Assuming I am using gpu-runner in a way I am not supposed to, what would be the correct command to run a single arbitrary .mlir test for GPU then?

[AndreyPavlenko]

Assuming I am using gpu-runner in a way I am not supposed to, what would be the correct command to run a single arbitrary .mlir test for GPU then?

You are using it in the right way and the same command works for me:

$ ./bin/gc-gpu-runner --shared-libs=../../../graph-compiler/externals/llvm-project/build/Release/lib/libmlir_runner_utils.so ../../../graph-compiler/test/mlir/test/gc/gpu-runner/mlp.mlir
Unranked Memref base@ = 0x55c191cd5180 rank = 2 offset = 0 sizes = [1, 10] strides = [10, 1] data =
[[0.1,   0.1,   0.1,   0.1,   0.1,   0.1,   0.1,   0.1,   0.1,   0.1]]

Does this command fails for you? What's the failure?

[akroviakov]

As I have mentioned in the beginning, the test runs:

~/graph-compiler/build$ /home/.../graph-compiler/build/bin/gc-gpu-runner --shared-libs=/home/.../graph-compiler/externals/llvm-project/build/lib/libmlir_runner_utils.so /home/.../graph-compiler/test/mlir/test/gc/gpu-runner/mlp.mlir 
[ERROR] [/home/.../graph-compiler/lib/gc/ExecutionEngine/GPURuntime/ocl/GpuOclRuntime.cpp:357] Failed to get the number of devices on the platform.0x55bc84c46970 Error: -1
Unranked Memref base@ = 0x55bc85e94480 rank = 2 offset = 0 sizes = [1, 10] strides = [10, 1] data = 
[[0.1,   0.1,   0.1,   0.1,   0.1,   0.1,   0.1,   0.1,   0.1,   0.1]]

However, it also displays an error for one of the platforms. This error is displayed because we ask the following platform:

Platform: Intel(R) OpenCL
  Device: Intel(R) Xeon(R) Gold 6438Y+

for a device ID of type CL_DEVICE_TYPE_GPU which triggers:

    if (err != CL_SUCCESS) {
      gcLogE("Failed to get the number of devices on the platform.", platform,
             " Error: ", err);
      continue;
    }

So is it ok for GpuOclRuntime to see platforms that are not OpenCL Graphics (e.g., by simply iterating over them)? If yes, does that mean that I can simply ignore the error?

[AndreyPavlenko]

does that mean that I can simply ignore the error?

Got it. The error message is redundant here. It should be either replaced with a debug message or removed.