Support LightGBM on macOS with real (possibly external) GPU device

[kshv190]

Description

Getting error of invalid kernel on MacBook Pro i9 with AMD Radeon Pro 5500M GPU.
Here is the log

$ ../../lightgbm config=train.conf data=regression.train

[LightGBM] [Warning] data is set=regression.train, data=regression.train will be ignored. Current value: data=regression.train
[LightGBM] [Info] Finished loading parameters
[LightGBM] [Info] Loading initial scores...
[LightGBM] [Info] Construct bin mappers from text data time 0.01 seconds
[LightGBM] [Info] Loading initial scores...
[LightGBM] [Info] Finished loading data in 0.050829 seconds
[LightGBM] [Info] This is the GPU trainer!!
[LightGBM] [Info] Total Bins 6132
[LightGBM] [Info] Number of data points in the train set: 7000, number of used features: 28
[LightGBM] [Info] Using requested OpenCL platform 0 device 2
[LightGBM] [Info] Using GPU Device: AMD Radeon Pro 5500M Compute Engine, Vendor: AMD
[LightGBM] [Info] Compiling OpenCL Kernel with 256 bins...
[LightGBM] [Info] GPU programs have been built
[LightGBM] [Info] Size of histogram bin entry: 8
[LightGBM] [Info] 27 dense feature groups (0.19 MB) transferred to GPU in 0.181396 secs. 1 sparse feature groups
Met Exceptions:
Invalid Kernel

Reproducible example

I am using the regression example in the example folders.
I modified the train.conf file with the following lines:

device = gpu
gpu_platform_id = 0
gpu_device_id = 2

Environment info

Mac book Pro 16 inch 2019.
GPU: AMD Radeon Pro 5500M

LightGBM version or commit hash: 3.2.1.99

Command(s) you used to install LightGBM

git clone --recursive https://github.com/microsoft/LightGBM
cd LightGBM
mkdir build
cd build
export CXX=g++-11 CC=gcc-11
cmake -DUSE_GPU=1 ..
make -j4

Additional Comments

[StrikerRUS]

LightGBM GPU version is not supported on macOS for now: #1523 (comment). But theoretically it's supposed to be working with a real GPU device (AMD Radeon Pro 5500M in this case): #1523 (comment).

[kshv190]

I built the R package with:
Rscript build_r.R --use-gpu

and it seems to work, please look at the log.

model <- lgb.cv(params = list(objective = "regression", metric = "l2", device='gpu', gpu_platform_id=0, gpu_device_id = 2), data = dtrain)
[LightGBM] [Info] This is the GPU trainer!!
[LightGBM] [Info] Total Bins 412
[LightGBM] [Info] Number of data points in the train set: 18834, number of used features: 9
[LightGBM] [Info] Using requested OpenCL platform 0 device 2
[LightGBM] [Info] Using GPU Device: AMD Radeon Pro 5500M Compute Engine, Vendor: AMD
[LightGBM] [Info] Compiling OpenCL Kernel with 64 bins...
[LightGBM] [Info] GPU programs have been built
[LightGBM] [Info] Size of histogram bin entry: 8
[LightGBM] [Info] 9 dense feature groups (0.22 MB) transferred to GPU in 0.006214 secs. 0 sparse feature groups
[LightGBM] [Info] This is the GPU trainer!!
[LightGBM] [Info] Total Bins 412
[LightGBM] [Info] Number of data points in the train set: 18833, number of used features: 9
[LightGBM] [Info] Using requested OpenCL platform 0 device 2
[LightGBM] [Info] Using GPU Device: AMD Radeon Pro 5500M Compute Engine, Vendor: AMD
[LightGBM] [Info] Compiling OpenCL Kernel with 64 bins...
[LightGBM] [Info] GPU programs have been built
[LightGBM] [Info] Size of histogram bin entry: 8
[LightGBM] [Info] 9 dense feature groups (0.22 MB) transferred to GPU in 0.003041 secs. 0 sparse feature groups
[LightGBM] [Info] This is the GPU trainer!!
[LightGBM] [Info] Total Bins 412
[LightGBM] [Info] Number of data points in the train set: 18833, number of used features: 9
[LightGBM] [Info] Using requested OpenCL platform 0 device 2
[LightGBM] [Info] Using GPU Device: AMD Radeon Pro 5500M Compute Engine, Vendor: AMD
[LightGBM] [Info] Compiling OpenCL Kernel with 64 bins...
[LightGBM] [Info] GPU programs have been built
[LightGBM] [Info] Size of histogram bin entry: 8
[LightGBM] [Info] 9 dense feature groups (0.22 MB) transferred to GPU in 0.003050 secs. 0 sparse feature groups
[LightGBM] [Info] Start training from score -0.000390
[LightGBM] [Info] Start training from score -0.000361
[LightGBM] [Info] Start training from score -0.000363
[1] "[1]: valid's l2:0.000198784+1.9672e-05"

[StrikerRUS]

Closed in favor of being in #2302. We decided to keep all feature requests in one place.

Welcome to contribute this feature! Please re-open this issue (or post a comment if you are not a topic starter) if you are actively working on implementing this feature.