[CUDA] Add quantile regression objective for new CUDA version

[shiyu1994]

Add quantile regression objective for new CUDA version.

[shiyu1994]

It is weird that

nvlink error   : Entry function '_ZN8LightGBM33CUDAConstructHistogramDenseKernelIhfLm11120EEEvPKNS_20CUDALeafSplitsStructEPKfS5_PKT_PKjSA_PKii' uses too much shared data (0xcbfc bytes, 0xc000 max) (target: sm_60)
nvlink error   : Entry function '_ZN8LightGBM34CUDAConstructHistogramSparseKernelItmfLm11120EEEvPKNS_20CUDALeafSplitsStructEPKfS5_PKT_PKT0_SB_PKji' uses too much shared data (0xcbfc bytes, 0xc000 max) (target: sm_60)

should arise when no related code of CUDAConstructHistogramSparseKernel or CUDAConstructHistogramDenseKernel is changed in this PR for cuda 10.0.

Note that for cuda 10.0, in master branch we've already used a smaller shared memory size than the claimed maximum 0c000 bytes (with which we should be able to use DP_SHARED_HIST_SIZE = 6144) to avoid the error above.

LightGBM/include/LightGBM/cuda/cuda_row_data.hpp

Lines 22 to 27 in 8811063

	#if CUDART_VERSION == 10000
	#define DP_SHARED_HIST_SIZE (5560)
	#else
	#define DP_SHARED_HIST_SIZE (6144)
	#endif
	#define SP_SHARED_HIST_SIZE (DP_SHARED_HIST_SIZE * 2)

It is unclear to me why adding code in this PR will have to make the DP_SHARED_HIST_SIZE = 5560 for cuda 10.0 smaller.

@guolinke Do you have any idea?

[guolinke]

how many bytes were exceeded? It looks quite large. are there any functions that possibly inherit or call other functions with shared memory?

[shiyu1994]

how many bytes were exceeded?

0xcbfc - 0xc000 = 0x0bfc = 3068 bytes

are there any functions that possibly inherit or call other functions with shared memory?

No. The code compiles with cuda 11.0 but not with cuda 10.0.

[shiyu1994]

@guolinke Let me adjust the #define DP_SHARED_HIST_SIZE (5560) to a lower value so that the compilation can succeed. And let's drop the support for CUDA 10.0 in another PR.

[shiyu1994]

It is unclear to me why adding code in this PR will have to make the DP_SHARED_HIST_SIZE = 5560 for cuda 10.0 smaller.

Just use binary search to find out that the maximum allowed value is 5176.

[shiyu1994]

@guolinke This is ready for review.

[github-actions]

This pull request has been automatically locked since there has not been any recent activity since it was closed.
To start a new related discussion, open a new issue at https://github.com/microsoft/LightGBM/issues
including a reference to this.