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

fix auc.cu

Browse Source
This commit is contained in:
amdsc21
2023-03-09 20:29:38 +01:00
parent 6eba0a56ec
commit a56055225a
7 changed files with 205 additions and 20 deletions
Download Patch File Download Diff File Expand all files Collapse all files

43
src/common/algorithm.cuh
View File

@@ -10,14 +10,26 @@
#include <cstddef> // size_t
#include <cstdint> // int32_t
#if defined(XGBOOST_USE_HIP)
#include <hipcub/hipcub.hpp>
#elif defined(XGBOOST_USE_CUDA)
#include <cub/cub.cuh> // DispatchSegmentedRadixSort,NullType,DoubleBuffer
#endif
#include <iterator> // distance
#include <limits> // numeric_limits
#include <type_traits> // conditional_t,remove_const_t
#include "common.h" // safe_cuda
#include "cuda_context.cuh" // CUDAContext
#if defined(XGBOOST_USE_HIP)
#include "device_helpers.hip.h"
#elif defined(XGBOOST_USE_CUDA)
#include "device_helpers.cuh" // TemporaryArray,SegmentId,LaunchN,Iota,device_vector
#endif
#include "xgboost/base.h" // XGBOOST_DEVICE
#include "xgboost/context.h" // Context
#include "xgboost/logging.h" // CHECK
@@ -39,6 +51,7 @@ static void DeviceSegmentedRadixSortKeys(CUDAContext const *ctx, void *d_temp_st
using OffsetT = int;
// Null value type
#if defined(XGBOOST_USE_CUDA)
cub::DoubleBuffer<KeyT> d_keys(const_cast<KeyT *>(d_keys_in), d_keys_out);
cub::DoubleBuffer<cub::NullType> d_values;
@@ -47,6 +60,20 @@ static void DeviceSegmentedRadixSortKeys(CUDAContext const *ctx, void *d_temp_st
OffsetT>::Dispatch(d_temp_storage, temp_storage_bytes, d_keys, d_values, num_items,
num_segments, d_begin_offsets, d_end_offsets, begin_bit,
end_bit, false, ctx->Stream(), debug_synchronous)));
#elif defined(XGBOOST_USE_HIP)
if (IS_DESCENDING) {
rocprim::segmented_radix_sort_pairs_desc<KeyT, hipcub::NullType, BeginOffsetIteratorT>(d_temp_storage,
temp_storage_bytes, d_keys_in, d_keys_out, nullptr, nullptr, num_items,
num_segments, d_begin_offsets, d_end_offsets,
begin_bit, end_bit, ctx->Stream(), debug_synchronous);
}
else {
rocprim::segmented_radix_sort_pairs<KeyT, hipcub::NullType, BeginOffsetIteratorT>(d_temp_storage,
temp_storage_bytes, d_keys_in, d_keys_out, nullptr, nullptr, num_items,
num_segments, d_begin_offsets, d_end_offsets,
begin_bit, end_bit, ctx->Stream(), debug_synchronous);
}
#endif
}
// Wrapper around cub sort for easier `descending` sort.
@@ -60,14 +87,18 @@ void DeviceSegmentedRadixSortPair(void *d_temp_storage,
BeginOffsetIteratorT d_begin_offsets,
EndOffsetIteratorT d_end_offsets, dh::CUDAStreamView stream,
int begin_bit = 0, int end_bit = sizeof(KeyT) * 8) {
#if defined(XGBOOST_USE_CUDA)
cub::DoubleBuffer<KeyT> d_keys(const_cast<KeyT *>(d_keys_in), d_keys_out);
cub::DoubleBuffer<ValueT> d_values(const_cast<ValueT *>(d_values_in), d_values_out);
#endif
// In old version of cub, num_items in dispatch is also int32_t, no way to change.
using OffsetT = std::conditional_t<dh::BuildWithCUDACub() && dh::HasThrustMinorVer<13>(),
std::size_t, std::int32_t>;
CHECK_LE(num_items, std::numeric_limits<OffsetT>::max());
// For Thrust >= 1.12 or CUDA >= 11.4, we require system cub installation
#if defined(XGBOOST_USE_CUDA)
#if THRUST_MAJOR_VERSION >= 2
dh::safe_cuda((cub::DispatchSegmentedRadixSort<
descending, KeyT, ValueT, BeginOffsetIteratorT, EndOffsetIteratorT,
@@ -88,6 +119,18 @@ void DeviceSegmentedRadixSortPair(void *d_temp_storage,
d_begin_offsets, d_end_offsets, begin_bit,
end_bit, false, stream, false)));
#endif
#elif defined(XGBOOST_USE_HIP)
if (descending) {
rocprim::segmented_radix_sort_pairs_desc(d_temp_storage, temp_storage_bytes, d_keys_in, d_keys_out,
d_values_in, d_values_out, num_items, num_segments,
d_begin_offsets, d_end_offsets, begin_bit, end_bit, stream, false);
}
else {
rocprim::segmented_radix_sort_pairs(d_temp_storage, temp_storage_bytes, d_keys_in, d_keys_out,
d_values_in, d_values_out, num_items, num_segments, d_begin_offsets, d_end_offsets,
begin_bit, end_bit, stream, false);
}
#endif
}
} // namespace detail

26
src/common/device_helpers.hip.h
View File

@@ -1208,8 +1208,7 @@ void InclusiveScan(InputIteratorT d_in, OutputIteratorT d_out, ScanOpT scan_op,
#endif
#endif
safe_cuda((rocprim::inclusive_scan<InputIteratorT, OutputIteratorT, ScanOpT>(nullptr,
bytes, d_in, d_out, num_items, scan_op)));
safe_cuda((rocprim::inclusive_scan(nullptr, bytes, d_in, d_out, (size_t) num_items, scan_op)));
TemporaryArray<char> storage(bytes);
@@ -1229,8 +1228,7 @@ void InclusiveScan(InputIteratorT d_in, OutputIteratorT d_out, ScanOpT scan_op,
#endif
#endif
safe_cuda((rocprim::inclusive_scan<InputIteratorT, OutputIteratorT, ScanOpT>(
storage.data().get(), bytes, d_in, d_out, num_items, scan_op)));
safe_cuda((rocprim::inclusive_scan(storage.data().get(), bytes, d_in, d_out, (size_t) num_items, scan_op)));
}
template <typename InIt, typename OutIt, typename Predicate>
@@ -1262,11 +1260,7 @@ void ArgSort(xgboost::common::Span<U> keys, xgboost::common::Span<IdxT> sorted_i
using ValueT = std::remove_const_t<IdxT>;
TemporaryArray<KeyT> out(keys.size());
hipcub::DoubleBuffer<KeyT> d_keys(const_cast<KeyT *>(keys.data()),
out.data().get());
TemporaryArray<IdxT> sorted_idx_out(sorted_idx.size());
hipcub::DoubleBuffer<ValueT> d_values(const_cast<ValueT *>(sorted_idx.data()),
sorted_idx_out.data().get());
// track https://github.com/NVIDIA/cub/pull/340 for 64bit length support
using OffsetT = std::conditional_t<!BuildWithCUDACub(), std::ptrdiff_t, int32_t>;
@@ -1286,8 +1280,8 @@ void ArgSort(xgboost::common::Span<U> keys, xgboost::common::Span<IdxT> sorted_i
#endif
#endif
safe_cuda((rocprim::radix_sort_pairs<KeyT, ValueT, OffsetT>(d_temp_storage,
bytes, d_keys, d_values, sorted_idx.size(), 0,
safe_cuda((rocprim::radix_sort_pairs(d_temp_storage,
bytes, keys.data(), out.data().get(), sorted_idx.data(), sorted_idx_out.data().get(), sorted_idx.size(), 0,
sizeof(KeyT) * 8)));
TemporaryArray<char> storage(bytes);
@@ -1305,8 +1299,8 @@ void ArgSort(xgboost::common::Span<U> keys, xgboost::common::Span<IdxT> sorted_i
#endif
#endif
safe_cuda((rocprim::radix_sort_pairs<KeyT, ValueT, OffsetT>(d_temp_storage,
bytes, d_keys, d_values, sorted_idx.size(), 0,
safe_cuda((rocprim::radix_sort_pairs(d_temp_storage,
bytes, keys.data(), out.data().get(), sorted_idx.data(), sorted_idx_out.data().get(), sorted_idx.size(), 0,
sizeof(KeyT) * 8)));
} else {
void *d_temp_storage = nullptr;
@@ -1323,8 +1317,8 @@ void ArgSort(xgboost::common::Span<U> keys, xgboost::common::Span<IdxT> sorted_i
#endif
#endif
safe_cuda((rocprim::radix_sort_pairs<KeyT, ValueT, OffsetT>(d_temp_storage,
bytes, d_keys, d_values, sorted_idx.size(), 0,
safe_cuda((rocprim::radix_sort_pairs(d_temp_storage,
bytes, keys.data(), out.data().get(), sorted_idx.data(), sorted_idx_out.data().get(), sorted_idx.size(), 0,
sizeof(KeyT) * 8)));
TemporaryArray<char> storage(bytes);
@@ -1341,8 +1335,8 @@ void ArgSort(xgboost::common::Span<U> keys, xgboost::common::Span<IdxT> sorted_i
sizeof(KeyT) * 8, false, nullptr, false)));
#endif
#endif
safe_cuda((rocprim::radix_sort_pairs<KeyT, ValueT, OffsetT>(d_temp_storage,
bytes, d_keys, d_values, sorted_idx.size(), 0,
safe_cuda((rocprim::radix_sort_pairs(d_temp_storage,
bytes, keys.data(), out.data().get(), sorted_idx.data(), sorted_idx_out.data().get(), sorted_idx.size(), 0,
sizeof(KeyT) * 8)));
}

8
src/common/threading_utils.cuh
View File

@@ -9,7 +9,13 @@
#include "./math.h" // Sqr
#include "common.h"
#if defined(XGBOOST_USE_HIP)
#include "device_helpers.hip.h"
#elif defined(XGBOOST_USE_CUDA)
#include "device_helpers.cuh" // LaunchN
#endif
#include "xgboost/base.h" // XGBOOST_DEVICE
#include "xgboost/span.h" // Span
@@ -67,7 +73,7 @@ SegmentedTrapezoidThreads(xgboost::common::Span<U> group_ptr,
dh::safe_cuda(hipMemcpy(
&total, out_group_threads_ptr.data() + out_group_threads_ptr.size() - 1,
sizeof(total), hipMemcpyDeviceToHost));
#else
#elif defined(XGBOOST_USE_CUDA)
dh::safe_cuda(cudaMemcpy(
&total, out_group_threads_ptr.data() + out_group_threads_ptr.size() - 1,
sizeof(total), cudaMemcpyDeviceToHost));