Hendrik/xgboost
1
0
Fork 0
Code Issues Pull Requests Packages Projects Releases Wiki Activity

Pairwise ranking objective implementation on gpu (#4873)

Browse Source 
* - pairwise ranking objective implementation on gpu
   - there are couple of more algorithms (ndcg and map) for which support will be added
     as follow-up pr's
   - with no label groups defined, get gradient is 90x faster on gpu (120m instance
     mortgage dataset)
   - it can perform by an order of magnitude faster with ~ 10 groups (and adequate cores
     for the cpu implementation)

* Add JSON config to rank obj.
This commit is contained in:
sriramch
2019-10-22 20:40:07 -07:00
committed by Jiaming Yuan
parent 5620322a48
commit 310fe60b35
8 changed files with 776 additions and 412 deletions
Download Patch File Download Diff File Expand all files Collapse all files

33
src/tree/updater_gpu_common.cuh
View File

@@ -12,42 +12,9 @@
#include "../common/random.h"
#include "param.h"
#if !defined(__CUDA_ARCH__) || __CUDA_ARCH__ >= 600
#else // In device code and CUDA < 600
XGBOOST_DEVICE __forceinline__ double atomicAdd(double* address, double val) {
unsigned long long int* address_as_ull =
(unsigned long long int*)address; // NOLINT
unsigned long long int old = *address_as_ull, assumed; // NOLINT
do {
assumed = old;
old = atomicCAS(address_as_ull, assumed,
__double_as_longlong(val + __longlong_as_double(assumed)));
// Note: uses integer comparison to avoid hang in case of NaN (since NaN !=
// NaN)
} while (assumed != old);
return __longlong_as_double(old);
}
#endif
namespace xgboost {
namespace tree {
// Atomic add function for gradients
template <typename OutputGradientT, typename InputGradientT>
DEV_INLINE void AtomicAddGpair(OutputGradientT* dest,
const InputGradientT& gpair) {
auto dst_ptr = reinterpret_cast<typename OutputGradientT::ValueT*>(dest);
atomicAdd(dst_ptr,
static_cast<typename OutputGradientT::ValueT>(gpair.GetGrad()));
atomicAdd(dst_ptr + 1,
static_cast<typename OutputGradientT::ValueT>(gpair.GetHess()));
}
struct GPUTrainingParam {
// minimum amount of hessian(weight) allowed in a child
float min_child_weight;

4
src/tree/updater_gpu_hist.cu
View File

@@ -421,7 +421,7 @@ __global__ void SharedMemHistKernel(xgboost::ELLPackMatrix matrix,
// global memory
GradientSumT* atomic_add_ptr =
use_shared_memory_histograms ? smem_arr : d_node_hist;
AtomicAddGpair(atomic_add_ptr + gidx, d_gpair[ridx]);
dh::AtomicAddGpair(atomic_add_ptr + gidx, d_gpair[ridx]);
}
}
@@ -430,7 +430,7 @@ __global__ void SharedMemHistKernel(xgboost::ELLPackMatrix matrix,
__syncthreads();
for (auto i :
dh::BlockStrideRange(static_cast<size_t>(0), matrix.BinCount())) {
AtomicAddGpair(d_node_hist + i, smem_arr[i]);
dh::AtomicAddGpair(d_node_hist + i, smem_arr[i]);
}
}
}