[CUDA] CUDA Quantized Training (fixes #5606) (#5933)

* add quantized training (first stage)

* add histogram construction functions for integer gradients

* add stochastic rounding

* update docs

* fix compilation errors by adding template instantiations

* update files for compilation

* fix compilation of gpu version

* initialize gradient discretizer before share states

* add a test case for quantized training

* add quantized training for data distributed training

* Delete origin.pred

* Delete ifelse.pred

* Delete LightGBM_model.txt

* remove useless changes

* fix lint error

* remove debug loggings

* fix mismatch of vector and allocator types

* remove changes in main.cpp

* fix bugs with uninitialized gradient discretizer

* initialize ordered gradients in gradient discretizer

* disable quantized training with gpu and cuda

fix msvc compilation errors and warnings

* fix bug in data parallel tree learner

* make quantized training test deterministic

* make quantized training in test case more accurate

* refactor test_quantized_training

* fix leaf splits initialization with quantized training

* check distributed quantized training result

* add cuda gradient discretizer

* add quantized training for CUDA version in tree learner

* remove cuda computability 6.1 and 6.2

* fix parts of gpu quantized training errors and warnings

* fix build-python.sh to install locally built version

* fix memory access bugs

* fix lint errors

* mark cuda quantized training on cuda with categorical features as unsupported

* rename cuda_utils.h to cuda_utils.hu

* enable quantized training with cuda

* fix cuda quantized training with sparse row data

* allow using global memory buffer in histogram construction with cuda quantized training

* recover build-python.sh

enlarge allowed package size to 100M

.ci/check_python_dists.sh

@@ -25,7 +25,7 @@ if [ $PY_MINOR_VER -gt 7 ]; then
         pydistcheck \
             --inspect \
             --ignore 'compiled-objects-have-debug-symbols,distro-too-large-compressed' \
-            --max-allowed-size-uncompressed '70M' \
+            --max-allowed-size-uncompressed '100M' \
             --max-allowed-files 800 \
             ${DIST_DIR}/* || exit -1
     elif { test $(uname -m) = "aarch64"; }; then

include/LightGBM/cuda/cuda_algorithms.hpp

@@ -13,7 +13,7 @@
 #include <stdio.h>
 
 #include <LightGBM/bin.h>
-#include <LightGBM/cuda/cuda_utils.h>
+#include <LightGBM/cuda/cuda_utils.hu>
 #include <LightGBM/utils/log.h>
 
 #include <algorithm>

include/LightGBM/cuda/cuda_column_data.hpp

@@ -9,7 +9,7 @@
 #define LIGHTGBM_CUDA_CUDA_COLUMN_DATA_HPP_
 
 #include <LightGBM/config.h>
-#include <LightGBM/cuda/cuda_utils.h>
+#include <LightGBM/cuda/cuda_utils.hu>
 #include <LightGBM/bin.h>
 #include <LightGBM/utils/openmp_wrapper.h>
 

include/LightGBM/cuda/cuda_metadata.hpp

@@ -8,7 +8,7 @@
 #ifndef LIGHTGBM_CUDA_CUDA_METADATA_HPP_
 #define LIGHTGBM_CUDA_CUDA_METADATA_HPP_
 
-#include <LightGBM/cuda/cuda_utils.h>
+#include <LightGBM/cuda/cuda_utils.hu>
 #include <LightGBM/meta.h>
 
 #include <vector>

include/LightGBM/cuda/cuda_metric.hpp

@@ -9,7 +9,7 @@
 
 #ifdef USE_CUDA
 
-#include <LightGBM/cuda/cuda_utils.h>
+#include <LightGBM/cuda/cuda_utils.hu>
 #include <LightGBM/metric.h>
 
 namespace LightGBM {

include/LightGBM/cuda/cuda_objective_function.hpp

@@ -9,7 +9,7 @@
 
 #ifdef USE_CUDA
 
-#include <LightGBM/cuda/cuda_utils.h>
+#include <LightGBM/cuda/cuda_utils.hu>
 #include <LightGBM/objective_function.h>
 #include <LightGBM/meta.h>
 

include/LightGBM/cuda/cuda_row_data.hpp

@@ -10,7 +10,7 @@
 
 #include <LightGBM/bin.h>
 #include <LightGBM/config.h>
-#include <LightGBM/cuda/cuda_utils.h>
+#include <LightGBM/cuda/cuda_utils.hu>
 #include <LightGBM/dataset.h>
 #include <LightGBM/train_share_states.h>
 #include <LightGBM/utils/openmp_wrapper.h>

include/LightGBM/cuda/cuda_split_info.hpp

@@ -24,12 +24,14 @@ class CUDASplitInfo {
 
   double left_sum_gradients;
   double left_sum_hessians;
+  int64_t left_sum_of_gradients_hessians;
   data_size_t left_count;
   double left_gain;
   double left_value;
 
   double right_sum_gradients;
   double right_sum_hessians;
+  int64_t right_sum_of_gradients_hessians;
   data_size_t right_count;
   double right_gain;
   double right_value;

include/LightGBM/cuda/cuda_utils.hu

@@ -7,15 +7,21 @@
 #define LIGHTGBM_CUDA_CUDA_UTILS_H_
 
 #ifdef USE_CUDA
+
 #include <cuda.h>
 #include <cuda_runtime.h>
 #include <stdio.h>
+
 #include <LightGBM/utils/log.h>
+
+#include <algorithm>
 #include <vector>
 #include <cmath>
 
 namespace LightGBM {
 
+typedef unsigned long long atomic_add_long_t;
+
 #define CUDASUCCESS_OR_FATAL(ans) { gpuAssert((ans), __FILE__, __LINE__); }
 inline void gpuAssert(cudaError_t code, const char *file, int line, bool abort = true) {
   if (code != cudaSuccess) {
@@ -125,13 +131,19 @@ class CUDAVector {
     T* new_data = nullptr;
     AllocateCUDAMemory<T>(&new_data, size, __FILE__, __LINE__);
     if (size_ > 0 && data_ != nullptr) {
-      CopyFromCUDADeviceToCUDADevice<T>(new_data, data_, size, __FILE__, __LINE__);
+      const size_t size_for_old_content = std::min<size_t>(size_, size);
+      CopyFromCUDADeviceToCUDADevice<T>(new_data, data_, size_for_old_content, __FILE__, __LINE__);
     }
     DeallocateCUDAMemory<T>(&data_, __FILE__, __LINE__);
     data_ = new_data;
     size_ = size;
   }
 
+  void InitFromHostVector(const std::vector<T>& host_vector) {
+    Resize(host_vector.size());
+    CopyFromHostToCUDADevice(data_, host_vector.data(), host_vector.size(), __FILE__, __LINE__);
+  }
+
   void Clear() {
     if (size_ > 0 && data_ != nullptr) {
       DeallocateCUDAMemory<T>(&data_, __FILE__, __LINE__);
@@ -171,6 +183,10 @@ class CUDAVector {
     return data_;
   }
 
+  void SetValue(int value) {
+    SetCUDAMemory<T>(data_, value, size_, __FILE__, __LINE__);
+  }
+
   const T* RawDataReadOnly() const {
     return data_;
   }

include/LightGBM/sample_strategy.h

@@ -6,7 +6,7 @@
 #ifndef LIGHTGBM_SAMPLE_STRATEGY_H_
 #define LIGHTGBM_SAMPLE_STRATEGY_H_
 
-#include <LightGBM/cuda/cuda_utils.h>
+#include <LightGBM/cuda/cuda_utils.hu>
 #include <LightGBM/utils/random.h>
 #include <LightGBM/utils/common.h>
 #include <LightGBM/utils/threading.h>

src/boosting/cuda/cuda_score_updater.hpp

@@ -8,7 +8,7 @@
 
 #ifdef USE_CUDA
 
-#include <LightGBM/cuda/cuda_utils.h>
+#include <LightGBM/cuda/cuda_utils.hu>
 
 #include "../score_updater.hpp"
 

src/cuda/cuda_utils.cpp

@@ -5,7 +5,7 @@
 
 #ifdef USE_CUDA
 
-#include <LightGBM/cuda/cuda_utils.h>
+#include <LightGBM/cuda/cuda_utils.hu>
 
 namespace LightGBM {
 

src/io/config.cpp

@@ -389,10 +389,6 @@ void Config::CheckParamConflict() {
     if (deterministic) {
       Log::Warning("Although \"deterministic\" is set, the results ran by GPU may be non-deterministic.");
     }
-    if (use_quantized_grad) {
-      Log::Warning("Quantized training is not supported by CUDA tree learner. Switch to full precision training.");
-      use_quantized_grad = false;
-    }
   }
   // linear tree learner must be serial type and run on CPU device
   if (linear_tree) {

src/metric/cuda/cuda_binary_metric.hpp

@@ -10,7 +10,7 @@
 #ifdef USE_CUDA
 
 #include <LightGBM/cuda/cuda_metric.hpp>
-#include <LightGBM/cuda/cuda_utils.h>
+#include <LightGBM/cuda/cuda_utils.hu>
 
 #include <vector>
 

src/metric/cuda/cuda_pointwise_metric.hpp

@@ -10,7 +10,7 @@
 #ifdef USE_CUDA
 
 #include <LightGBM/cuda/cuda_metric.hpp>
-#include <LightGBM/cuda/cuda_utils.h>
+#include <LightGBM/cuda/cuda_utils.hu>
 
 #include <vector>
 

src/metric/cuda/cuda_regression_metric.hpp

@@ -10,7 +10,7 @@
 #ifdef USE_CUDA
 
 #include <LightGBM/cuda/cuda_metric.hpp>
-#include <LightGBM/cuda/cuda_utils.h>
+#include <LightGBM/cuda/cuda_utils.hu>
 
 #include <vector>
 

src/treelearner/cuda/cuda_best_split_finder.cpp

@@ -40,6 +40,9 @@ CUDABestSplitFinder::CUDABestSplitFinder(
   select_features_by_node_(select_features_by_node),
   cuda_hist_(cuda_hist) {
   InitFeatureMetaInfo(train_data);
+  if (has_categorical_feature_ && config->use_quantized_grad) {
+    Log::Fatal("Quantized training on GPU with categorical features is not supported yet.");
+  }
   cuda_leaf_best_split_info_ = nullptr;
   cuda_best_split_info_ = nullptr;
   cuda_best_split_info_buffer_ = nullptr;
@@ -326,13 +329,23 @@ void CUDABestSplitFinder::FindBestSplitsForLeaf(
   const data_size_t num_data_in_smaller_leaf,
   const data_size_t num_data_in_larger_leaf,
   const double sum_hessians_in_smaller_leaf,
-  const double sum_hessians_in_larger_leaf) {
+  const double sum_hessians_in_larger_leaf,
+  const score_t* grad_scale,
+  const score_t* hess_scale,
+  const uint8_t smaller_num_bits_in_histogram_bins,
+  const uint8_t larger_num_bits_in_histogram_bins) {
   const bool is_smaller_leaf_valid = (num_data_in_smaller_leaf > min_data_in_leaf_ &&
     sum_hessians_in_smaller_leaf > min_sum_hessian_in_leaf_);
   const bool is_larger_leaf_valid = (num_data_in_larger_leaf > min_data_in_leaf_ &&
     sum_hessians_in_larger_leaf > min_sum_hessian_in_leaf_ && larger_leaf_index >= 0);
+  if (grad_scale != nullptr && hess_scale != nullptr) {
+    LaunchFindBestSplitsDiscretizedForLeafKernel(smaller_leaf_splits, larger_leaf_splits,
+      smaller_leaf_index, larger_leaf_index, is_smaller_leaf_valid, is_larger_leaf_valid,
+      grad_scale, hess_scale, smaller_num_bits_in_histogram_bins, larger_num_bits_in_histogram_bins);
+  } else {
     LaunchFindBestSplitsForLeafKernel(smaller_leaf_splits, larger_leaf_splits,
       smaller_leaf_index, larger_leaf_index, is_smaller_leaf_valid, is_larger_leaf_valid);
+  }
   global_timer.Start("CUDABestSplitFinder::LaunchSyncBestSplitForLeafKernel");
   LaunchSyncBestSplitForLeafKernel(smaller_leaf_index, larger_leaf_index, is_smaller_leaf_valid, is_larger_leaf_valid);
   SynchronizeCUDADevice(__FILE__, __LINE__);

src/treelearner/cuda/cuda_best_split_finder.cu

@@ -320,6 +320,175 @@ __device__ void FindBestSplitsForLeafKernelInner(
   }
 }
 
+template <bool USE_RAND, bool USE_L1, bool USE_SMOOTHING, bool REVERSE, typename BIN_HIST_TYPE, typename ACC_HIST_TYPE, bool USE_16BIT_BIN_HIST, bool USE_16BIT_ACC_HIST>
+__device__ void FindBestSplitsDiscretizedForLeafKernelInner(
+  // input feature information
+  const BIN_HIST_TYPE* feature_hist_ptr,
+  // input task information
+  const SplitFindTask* task,
+  CUDARandom* cuda_random,
+  // input config parameter values
+  const double lambda_l1,
+  const double lambda_l2,
+  const double path_smooth,
+  const data_size_t min_data_in_leaf,
+  const double min_sum_hessian_in_leaf,
+  const double min_gain_to_split,
+  // input parent node information
+  const double parent_gain,
+  const int64_t sum_gradients_hessians,
+  const data_size_t num_data,
+  const double parent_output,
+  // gradient scale
+  const double grad_scale,
+  const double hess_scale,
+  // output parameters
+  CUDASplitInfo* cuda_best_split_info) {
+  const double sum_hessians = static_cast<double>(sum_gradients_hessians & 0x00000000ffffffff) * hess_scale;
+  const double cnt_factor = num_data / sum_hessians;
+  const double min_gain_shift = parent_gain + min_gain_to_split;
+
+  cuda_best_split_info->is_valid = false;
+
+  ACC_HIST_TYPE local_grad_hess_hist = 0;
+  double local_gain = 0.0f;
+  bool threshold_found = false;
+  uint32_t threshold_value = 0;
+  __shared__ int rand_threshold;
+  if (USE_RAND && threadIdx.x == 0) {
+    if (task->num_bin - 2 > 0) {
+      rand_threshold = cuda_random->NextInt(0, task->num_bin - 2);
+    }
+  }
+  __shared__ uint32_t best_thread_index;
+  __shared__ double shared_double_buffer[32];
+  __shared__ bool shared_bool_buffer[32];
+  __shared__ uint32_t shared_int_buffer[64];
+  const unsigned int threadIdx_x = threadIdx.x;
+  const bool skip_sum = REVERSE ?
+    (task->skip_default_bin && (task->num_bin - 1 - threadIdx_x) == static_cast<int>(task->default_bin)) :
+    (task->skip_default_bin && (threadIdx_x + task->mfb_offset) == static_cast<int>(task->default_bin));
+  const uint32_t feature_num_bin_minus_offset = task->num_bin - task->mfb_offset;
+  if (!REVERSE) {
+    if (threadIdx_x < feature_num_bin_minus_offset && !skip_sum) {
+      const unsigned int bin_offset = threadIdx_x;
+      if (USE_16BIT_BIN_HIST && !USE_16BIT_ACC_HIST) {
+        const int32_t local_grad_hess_hist_int32 = feature_hist_ptr[bin_offset];
+        local_grad_hess_hist = (static_cast<int64_t>(static_cast<int16_t>(local_grad_hess_hist_int32 >> 16)) << 32) | (static_cast<int64_t>(local_grad_hess_hist_int32 & 0x0000ffff));
+      } else {
+        local_grad_hess_hist = feature_hist_ptr[bin_offset];
+      }
+    }
+  } else {
+    if (threadIdx_x >= static_cast<unsigned int>(task->na_as_missing) &&
+      threadIdx_x < feature_num_bin_minus_offset && !skip_sum) {
+      const unsigned int read_index = feature_num_bin_minus_offset - 1 - threadIdx_x;
+      if (USE_16BIT_BIN_HIST && !USE_16BIT_ACC_HIST) {
+        const int32_t local_grad_hess_hist_int32 = feature_hist_ptr[read_index];
+        local_grad_hess_hist = (static_cast<int64_t>(static_cast<int16_t>(local_grad_hess_hist_int32 >> 16)) << 32) | (static_cast<int64_t>(local_grad_hess_hist_int32 & 0x0000ffff));
+      } else {
+        local_grad_hess_hist = feature_hist_ptr[read_index];
+      }
+    }
+  }
+  __syncthreads();
+  local_gain = kMinScore;
+  local_grad_hess_hist = ShufflePrefixSum<ACC_HIST_TYPE>(local_grad_hess_hist, reinterpret_cast<ACC_HIST_TYPE*>(shared_int_buffer));
+  double sum_left_gradient = 0.0f;
+  double sum_left_hessian = 0.0f;
+  double sum_right_gradient = 0.0f;
+  double sum_right_hessian = 0.0f;
+  data_size_t left_count = 0;
+  data_size_t right_count = 0;
+  int64_t sum_left_gradient_hessian = 0;
+  int64_t sum_right_gradient_hessian = 0;
+  if (REVERSE) {
+    if (threadIdx_x >= static_cast<unsigned int>(task->na_as_missing) && threadIdx_x <= task->num_bin - 2 && !skip_sum) {
+      sum_right_gradient_hessian = USE_16BIT_ACC_HIST ?
+        (static_cast<int64_t>(static_cast<int16_t>(local_grad_hess_hist >> 16)) << 32) | static_cast<int64_t>(local_grad_hess_hist & 0x0000ffff) :
+        local_grad_hess_hist;
+      sum_right_gradient = static_cast<double>(static_cast<int32_t>((sum_right_gradient_hessian & 0xffffffff00000000) >> 32)) * grad_scale;
+      sum_right_hessian = static_cast<double>(static_cast<int32_t>(sum_right_gradient_hessian & 0x00000000ffffffff)) * hess_scale;
+      right_count = static_cast<data_size_t>(__double2int_rn(sum_right_hessian * cnt_factor));
+      sum_left_gradient_hessian = sum_gradients_hessians - sum_right_gradient_hessian;
+      sum_left_gradient = static_cast<double>(static_cast<int32_t>((sum_left_gradient_hessian & 0xffffffff00000000)>> 32)) * grad_scale;
+      sum_left_hessian = static_cast<double>(static_cast<int32_t>(sum_left_gradient_hessian & 0x00000000ffffffff)) * hess_scale;
+      left_count = num_data - right_count;
+      if (sum_left_hessian >= min_sum_hessian_in_leaf && left_count >= min_data_in_leaf &&
+        sum_right_hessian >= min_sum_hessian_in_leaf && right_count >= min_data_in_leaf &&
+        (!USE_RAND || static_cast<int>(task->num_bin - 2 - threadIdx_x) == rand_threshold)) {
+        double current_gain = CUDALeafSplits::GetSplitGains<USE_L1, USE_SMOOTHING>(
+          sum_left_gradient, sum_left_hessian + kEpsilon, sum_right_gradient,
+          sum_right_hessian + kEpsilon, lambda_l1,
+          lambda_l2, path_smooth, left_count, right_count, parent_output);
+        // gain with split is worse than without split
+        if (current_gain > min_gain_shift) {
+          local_gain = current_gain - min_gain_shift;
+          threshold_value = static_cast<uint32_t>(task->num_bin - 2 - threadIdx_x);
+          threshold_found = true;
+        }
+      }
+    }
+  } else {
+    if (threadIdx_x <= feature_num_bin_minus_offset - 2 && !skip_sum) {
+      sum_left_gradient_hessian = USE_16BIT_ACC_HIST ?
+        (static_cast<int64_t>(static_cast<int16_t>(local_grad_hess_hist >> 16)) << 32) | static_cast<int64_t>(local_grad_hess_hist & 0x0000ffff) :
+        local_grad_hess_hist;
+      sum_left_gradient = static_cast<double>(static_cast<int32_t>((sum_left_gradient_hessian & 0xffffffff00000000) >> 32)) * grad_scale;
+      sum_left_hessian = static_cast<double>(static_cast<int32_t>(sum_left_gradient_hessian & 0x00000000ffffffff)) * hess_scale;
+      left_count = static_cast<data_size_t>(__double2int_rn(sum_left_hessian * cnt_factor));
+      sum_right_gradient_hessian = sum_gradients_hessians - sum_left_gradient_hessian;
+      sum_right_gradient = static_cast<double>(static_cast<int32_t>((sum_right_gradient_hessian & 0xffffffff00000000) >> 32)) * grad_scale;
+      sum_right_hessian = static_cast<double>(static_cast<int32_t>(sum_right_gradient_hessian & 0x00000000ffffffff)) * hess_scale;
+      right_count = num_data - left_count;
+      if (sum_left_hessian >= min_sum_hessian_in_leaf && left_count >= min_data_in_leaf &&
+        sum_right_hessian >= min_sum_hessian_in_leaf && right_count >= min_data_in_leaf &&
+        (!USE_RAND || static_cast<int>(threadIdx_x + task->mfb_offset) == rand_threshold)) {
+        double current_gain = CUDALeafSplits::GetSplitGains<USE_L1, USE_SMOOTHING>(
+          sum_left_gradient, sum_left_hessian + kEpsilon, sum_right_gradient,
+          sum_right_hessian + kEpsilon, lambda_l1,
+          lambda_l2, path_smooth, left_count, right_count, parent_output);
+        // gain with split is worse than without split
+        if (current_gain > min_gain_shift) {
+          local_gain = current_gain - min_gain_shift;
+          threshold_value = static_cast<uint32_t>(threadIdx_x + task->mfb_offset);
+          threshold_found = true;
+        }
+      }
+    }
+  }
+  __syncthreads();
+  const uint32_t result = ReduceBestGain(local_gain, threshold_found, threadIdx_x, shared_double_buffer, shared_bool_buffer, shared_int_buffer);
+  if (threadIdx_x == 0) {
+    best_thread_index = result;
+  }
+  __syncthreads();
+  if (threshold_found && threadIdx_x == best_thread_index) {
+    cuda_best_split_info->is_valid = true;
+    cuda_best_split_info->threshold = threshold_value;
+    cuda_best_split_info->gain = local_gain;
+    cuda_best_split_info->default_left = task->assume_out_default_left;
+    const double left_output = CUDALeafSplits::CalculateSplittedLeafOutput<USE_L1, USE_SMOOTHING>(sum_left_gradient,
+      sum_left_hessian, lambda_l1, lambda_l2, path_smooth, left_count, parent_output);
+    const double right_output = CUDALeafSplits::CalculateSplittedLeafOutput<USE_L1, USE_SMOOTHING>(sum_right_gradient,
+      sum_right_hessian, lambda_l1, lambda_l2, path_smooth, right_count, parent_output);
+    cuda_best_split_info->left_sum_gradients = sum_left_gradient;
+    cuda_best_split_info->left_sum_hessians = sum_left_hessian;
+    cuda_best_split_info->left_sum_of_gradients_hessians = sum_left_gradient_hessian;
+    cuda_best_split_info->left_count = left_count;
+    cuda_best_split_info->right_sum_gradients = sum_right_gradient;
+    cuda_best_split_info->right_sum_hessians = sum_right_hessian;
+    cuda_best_split_info->right_sum_of_gradients_hessians = sum_right_gradient_hessian;
+    cuda_best_split_info->right_count = right_count;
+    cuda_best_split_info->left_value = left_output;
+    cuda_best_split_info->left_gain = CUDALeafSplits::GetLeafGainGivenOutput<USE_L1>(sum_left_gradient,
+      sum_left_hessian, lambda_l1, lambda_l2, left_output);
+    cuda_best_split_info->right_value = right_output;
+    cuda_best_split_info->right_gain = CUDALeafSplits::GetLeafGainGivenOutput<USE_L1>(sum_right_gradient,
+      sum_right_hessian, lambda_l1, lambda_l2, right_output);
+  }
+}
+
 template <bool USE_RAND, bool USE_L1, bool USE_SMOOTHING>
 __device__ void FindBestSplitsForLeafKernelCategoricalInner(
   // input feature information
@@ -715,6 +884,169 @@ __global__ void FindBestSplitsForLeafKernel(
   }
 }
 
+
+template <bool USE_RAND, bool USE_L1, bool USE_SMOOTHING, bool IS_LARGER>
+__global__ void FindBestSplitsDiscretizedForLeafKernel(
+  // input feature information
+  const int8_t* is_feature_used_bytree,
+  // input task information
+  const int num_tasks,
+  const SplitFindTask* tasks,
+  CUDARandom* cuda_randoms,
+  // input leaf information
+  const CUDALeafSplitsStruct* smaller_leaf_splits,
+  const CUDALeafSplitsStruct* larger_leaf_splits,
+  const uint8_t smaller_leaf_num_bits_in_histogram_bin,
+  const uint8_t larger_leaf_num_bits_in_histogram_bin,
+  // input config parameter values
+  const data_size_t min_data_in_leaf,
+  const double min_sum_hessian_in_leaf,
+  const double min_gain_to_split,
+  const double lambda_l1,
+  const double lambda_l2,
+  const double path_smooth,
+  const double cat_smooth,
+  const double cat_l2,
+  const int max_cat_threshold,
+  const int min_data_per_group,
+  const int max_cat_to_onehot,
+  // gradient scale
+  const score_t* grad_scale,
+  const score_t* hess_scale,
+  // output
+  CUDASplitInfo* cuda_best_split_info) {
+  const unsigned int task_index = blockIdx.x;
+  const SplitFindTask* task = tasks + task_index;
+  const int inner_feature_index = task->inner_feature_index;
+  const double parent_gain = IS_LARGER ? larger_leaf_splits->gain : smaller_leaf_splits->gain;
+  const int64_t sum_gradients_hessians = IS_LARGER ? larger_leaf_splits->sum_of_gradients_hessians : smaller_leaf_splits->sum_of_gradients_hessians;
+  const data_size_t num_data = IS_LARGER ? larger_leaf_splits->num_data_in_leaf : smaller_leaf_splits->num_data_in_leaf;
+  const double parent_output = IS_LARGER ? larger_leaf_splits->leaf_value : smaller_leaf_splits->leaf_value;
+  const unsigned int output_offset = IS_LARGER ? (task_index + num_tasks) : task_index;
+  CUDASplitInfo* out = cuda_best_split_info + output_offset;
+  CUDARandom* cuda_random = USE_RAND ?
+    (IS_LARGER ? cuda_randoms + task_index * 2 + 1 : cuda_randoms + task_index * 2) : nullptr;
+  const bool use_16bit_bin = IS_LARGER ? (larger_leaf_num_bits_in_histogram_bin <= 16) : (smaller_leaf_num_bits_in_histogram_bin <= 16);
+  if (is_feature_used_bytree[inner_feature_index]) {
+    if (task->is_categorical) {
+      __threadfence();  // ensure store issued before trap
+      asm("trap;");
+    } else {
+      if (!task->reverse) {
+        if (use_16bit_bin) {
+          const int32_t* hist_ptr =
+            reinterpret_cast<const int32_t*>(IS_LARGER ? larger_leaf_splits->hist_in_leaf : smaller_leaf_splits->hist_in_leaf) + task->hist_offset;
+          FindBestSplitsDiscretizedForLeafKernelInner<USE_RAND, USE_L1, USE_SMOOTHING, false, int32_t, int32_t, true, true>(
+            // input feature information
+            hist_ptr,
+            // input task information
+            task,
+            cuda_random,
+            // input config parameter values
+            lambda_l1,
+            lambda_l2,
+            path_smooth,
+            min_data_in_leaf,
+            min_sum_hessian_in_leaf,
+            min_gain_to_split,
+            // input parent node information
+            parent_gain,
+            sum_gradients_hessians,
+            num_data,
+            parent_output,
+            // gradient scale
+            *grad_scale,
+            *hess_scale,
+            // output parameters
+            out);
+        } else {
+          const int32_t* hist_ptr =
+            reinterpret_cast<const int32_t*>(IS_LARGER ? larger_leaf_splits->hist_in_leaf : smaller_leaf_splits->hist_in_leaf) + task->hist_offset;
+          FindBestSplitsDiscretizedForLeafKernelInner<USE_RAND, USE_L1, USE_SMOOTHING, false, int32_t, int64_t, false, false>(
+            // input feature information
+            hist_ptr,
+            // input task information
+            task,
+            cuda_random,
+            // input config parameter values
+            lambda_l1,
+            lambda_l2,
+            path_smooth,
+            min_data_in_leaf,
+            min_sum_hessian_in_leaf,
+            min_gain_to_split,
+            // input parent node information
+            parent_gain,
+            sum_gradients_hessians,
+            num_data,
+            parent_output,
+            // gradient scale
+            *grad_scale,
+            *hess_scale,
+            // output parameters
+            out);
+        }
+      } else {
+        if (use_16bit_bin) {
+          const int32_t* hist_ptr =
+            reinterpret_cast<const int32_t*>(IS_LARGER ? larger_leaf_splits->hist_in_leaf : smaller_leaf_splits->hist_in_leaf) + task->hist_offset;
+          FindBestSplitsDiscretizedForLeafKernelInner<USE_RAND, USE_L1, USE_SMOOTHING, true, int32_t, int32_t, true, true>(
+            // input feature information
+            hist_ptr,
+            // input task information
+            task,
+            cuda_random,
+            // input config parameter values
+            lambda_l1,
+            lambda_l2,
+            path_smooth,
+            min_data_in_leaf,
+            min_sum_hessian_in_leaf,
+            min_gain_to_split,
+            // input parent node information
+            parent_gain,
+            sum_gradients_hessians,
+            num_data,
+            parent_output,
+            // gradient scale
+            *grad_scale,
+            *hess_scale,
+            // output parameters
+            out);
+        } else {
+          const int32_t* hist_ptr =
+            reinterpret_cast<const int32_t*>(IS_LARGER ? larger_leaf_splits->hist_in_leaf : smaller_leaf_splits->hist_in_leaf) + task->hist_offset;
+          FindBestSplitsDiscretizedForLeafKernelInner<USE_RAND, USE_L1, USE_SMOOTHING, true, int32_t, int64_t, false, false>(
+            // input feature information
+            hist_ptr,
+            // input task information
+            task,
+            cuda_random,
+            // input config parameter values
+            lambda_l1,
+            lambda_l2,
+            path_smooth,
+            min_data_in_leaf,
+            min_sum_hessian_in_leaf,
+            min_gain_to_split,
+            // input parent node information
+            parent_gain,
+            sum_gradients_hessians,
+            num_data,
+            parent_output,
+            // gradient scale
+            *grad_scale,
+            *hess_scale,
+            // output parameters
+            out);
+        }
+      }
+    }
+  } else {
+    out->is_valid = false;
+  }
+}
+
 template <bool USE_RAND, bool USE_L1, bool USE_SMOOTHING, bool REVERSE>
 __device__ void FindBestSplitsForLeafKernelInner_GlobalMemory(
   // input feature information
@@ -1466,6 +1798,108 @@ void CUDABestSplitFinder::LaunchFindBestSplitsForLeafKernelInner2(LaunchFindBest
 #undef FindBestSplitsForLeafKernel_ARGS
 #undef GlobalMemory_Buffer_ARGS
 
+
+#define LaunchFindBestSplitsDiscretizedForLeafKernel_PARAMS \
+  const CUDALeafSplitsStruct* smaller_leaf_splits, \
+  const CUDALeafSplitsStruct* larger_leaf_splits, \
+  const int smaller_leaf_index, \
+  const int larger_leaf_index, \
+  const bool is_smaller_leaf_valid, \
+  const bool is_larger_leaf_valid, \
+  const score_t* grad_scale, \
+  const score_t* hess_scale, \
+  const uint8_t smaller_num_bits_in_histogram_bins, \
+  const uint8_t larger_num_bits_in_histogram_bins
+
+#define LaunchFindBestSplitsDiscretizedForLeafKernel_ARGS \
+  smaller_leaf_splits, \
+  larger_leaf_splits, \
+  smaller_leaf_index, \
+  larger_leaf_index, \
+  is_smaller_leaf_valid, \
+  is_larger_leaf_valid, \
+  grad_scale, \
+  hess_scale, \
+  smaller_num_bits_in_histogram_bins, \
+  larger_num_bits_in_histogram_bins
+
+#define FindBestSplitsDiscretizedForLeafKernel_ARGS \
+    cuda_is_feature_used_bytree_, \
+    num_tasks_, \
+    cuda_split_find_tasks_.RawData(), \
+    cuda_randoms_.RawData(), \
+    smaller_leaf_splits, \
+    larger_leaf_splits, \
+    smaller_num_bits_in_histogram_bins, \
+    larger_num_bits_in_histogram_bins, \
+    min_data_in_leaf_, \
+    min_sum_hessian_in_leaf_, \
+    min_gain_to_split_, \
+    lambda_l1_, \
+    lambda_l2_, \
+    path_smooth_, \
+    cat_smooth_, \
+    cat_l2_, \
+    max_cat_threshold_, \
+    min_data_per_group_, \
+    max_cat_to_onehot_, \
+    grad_scale, \
+    hess_scale, \
+    cuda_best_split_info_
+
+void CUDABestSplitFinder::LaunchFindBestSplitsDiscretizedForLeafKernel(LaunchFindBestSplitsDiscretizedForLeafKernel_PARAMS) {
+  if (!is_smaller_leaf_valid && !is_larger_leaf_valid) {
+    return;
+  }
+  if (!extra_trees_) {
+    LaunchFindBestSplitsDiscretizedForLeafKernelInner0<false>(LaunchFindBestSplitsDiscretizedForLeafKernel_ARGS);
+  } else {
+    LaunchFindBestSplitsDiscretizedForLeafKernelInner0<true>(LaunchFindBestSplitsDiscretizedForLeafKernel_ARGS);
+  }
+}
+
+template <bool USE_RAND>
+void CUDABestSplitFinder::LaunchFindBestSplitsDiscretizedForLeafKernelInner0(LaunchFindBestSplitsDiscretizedForLeafKernel_PARAMS) {
+  if (lambda_l1_ <= 0.0f) {
+    LaunchFindBestSplitsDiscretizedForLeafKernelInner1<USE_RAND, false>(LaunchFindBestSplitsDiscretizedForLeafKernel_ARGS);
+  } else {
+    LaunchFindBestSplitsDiscretizedForLeafKernelInner1<USE_RAND, true>(LaunchFindBestSplitsDiscretizedForLeafKernel_ARGS);
+  }
+}
+
+template <bool USE_RAND, bool USE_L1>
+void CUDABestSplitFinder::LaunchFindBestSplitsDiscretizedForLeafKernelInner1(LaunchFindBestSplitsDiscretizedForLeafKernel_PARAMS) {
+  if (!use_smoothing_) {
+    LaunchFindBestSplitsDiscretizedForLeafKernelInner2<USE_RAND, USE_L1, false>(LaunchFindBestSplitsDiscretizedForLeafKernel_ARGS);
+  } else {
+    LaunchFindBestSplitsDiscretizedForLeafKernelInner2<USE_RAND, USE_L1, true>(LaunchFindBestSplitsDiscretizedForLeafKernel_ARGS);
+  }
+}
+
+template <bool USE_RAND, bool USE_L1, bool USE_SMOOTHING>
+void CUDABestSplitFinder::LaunchFindBestSplitsDiscretizedForLeafKernelInner2(LaunchFindBestSplitsDiscretizedForLeafKernel_PARAMS) {
+  if (!use_global_memory_) {
+    if (is_smaller_leaf_valid) {
+      FindBestSplitsDiscretizedForLeafKernel<USE_RAND, USE_L1, USE_SMOOTHING, false>
+        <<<num_tasks_, NUM_THREADS_PER_BLOCK_BEST_SPLIT_FINDER, 0, cuda_streams_[0]>>>
+        (FindBestSplitsDiscretizedForLeafKernel_ARGS);
+    }
+    SynchronizeCUDADevice(__FILE__, __LINE__);
+    if (is_larger_leaf_valid) {
+      FindBestSplitsDiscretizedForLeafKernel<USE_RAND, USE_L1, USE_SMOOTHING, true>
+        <<<num_tasks_, NUM_THREADS_PER_BLOCK_BEST_SPLIT_FINDER, 0, cuda_streams_[1]>>>
+        (FindBestSplitsDiscretizedForLeafKernel_ARGS);
+    }
+  } else {
+    // TODO(shiyu1994)
+  }
+}
+
+#undef LaunchFindBestSplitsDiscretizedForLeafKernel_PARAMS
+#undef LaunchFindBestSplitsDiscretizedForLeafKernel_ARGS
+#undef FindBestSplitsDiscretizedForLeafKernel_ARGS
+
+
 __device__ void ReduceBestSplit(bool* found, double* gain, uint32_t* shared_read_index,
   uint32_t num_features_aligned) {
   const uint32_t threadIdx_x = threadIdx.x;

src/treelearner/cuda/cuda_best_split_finder.hpp

@@ -67,7 +67,11 @@ class CUDABestSplitFinder {
     const data_size_t num_data_in_smaller_leaf,
     const data_size_t num_data_in_larger_leaf,
     const double sum_hessians_in_smaller_leaf,
-    const double sum_hessians_in_larger_leaf);
+    const double sum_hessians_in_larger_leaf,
+    const score_t* grad_scale,
+    const score_t* hess_scale,
+    const uint8_t smaller_num_bits_in_histogram_bins,
+    const uint8_t larger_num_bits_in_histogram_bins);
 
   const CUDASplitInfo* FindBestFromAllSplits(
     const int cur_num_leaves,
@@ -114,6 +118,31 @@ class CUDABestSplitFinder {
 
   #undef LaunchFindBestSplitsForLeafKernel_PARAMS
 
+  #define LaunchFindBestSplitsDiscretizedForLeafKernel_PARAMS \
+  const CUDALeafSplitsStruct* smaller_leaf_splits, \
+  const CUDALeafSplitsStruct* larger_leaf_splits, \
+  const int smaller_leaf_index, \
+  const int larger_leaf_index, \
+  const bool is_smaller_leaf_valid, \
+  const bool is_larger_leaf_valid, \
+  const score_t* grad_scale, \
+  const score_t* hess_scale, \
+  const uint8_t smaller_num_bits_in_histogram_bins, \
+  const uint8_t larger_num_bits_in_histogram_bins
+
+  void LaunchFindBestSplitsDiscretizedForLeafKernel(LaunchFindBestSplitsDiscretizedForLeafKernel_PARAMS);
+
+  template <bool USE_RAND>
+  void LaunchFindBestSplitsDiscretizedForLeafKernelInner0(LaunchFindBestSplitsDiscretizedForLeafKernel_PARAMS);
+
+  template <bool USE_RAND, bool USE_L1>
+  void LaunchFindBestSplitsDiscretizedForLeafKernelInner1(LaunchFindBestSplitsDiscretizedForLeafKernel_PARAMS);
+
+  template <bool USE_RAND, bool USE_L1, bool USE_SMOOTHING>
+  void LaunchFindBestSplitsDiscretizedForLeafKernelInner2(LaunchFindBestSplitsDiscretizedForLeafKernel_PARAMS);
+
+  #undef LaunchFindBestSplitsDiscretizedForLeafKernel_PARAMS
+
   void LaunchSyncBestSplitForLeafKernel(
     const int host_smaller_leaf_index,
     const int host_larger_leaf_index,

src/treelearner/cuda/cuda_data_partition.cpp

@@ -368,6 +368,12 @@ void CUDADataPartition::ResetByLeafPred(const std::vector<int>& leaf_pred, int n
   cur_num_leaves_ = num_leaves;
 }
 
+void CUDADataPartition::ReduceLeafGradStat(
+  const score_t* gradients, const score_t* hessians,
+  CUDATree* tree, double* leaf_grad_stat_buffer, double* leaf_hess_state_buffer) const {
+  LaunchReduceLeafGradStat(gradients, hessians, tree, leaf_grad_stat_buffer, leaf_hess_state_buffer);
+}
+
 }  // namespace LightGBM
 
 #endif  // USE_CUDA

src/treelearner/cuda/cuda_data_partition.cu

@@ -1069,6 +1069,53 @@ void CUDADataPartition::LaunchAddPredictionToScoreKernel(const double* leaf_valu
   global_timer.Stop("CUDADataPartition::AddPredictionToScoreKernel");
 }
 
+__global__ void RenewDiscretizedTreeLeavesKernel(
+  const score_t* gradients,
+  const score_t* hessians,
+  const data_size_t* data_indices,
+  const data_size_t* leaf_data_start,
+  const data_size_t* leaf_num_data,
+  double* leaf_grad_stat_buffer,
+  double* leaf_hess_stat_buffer,
+  double* leaf_values) {
+  __shared__ double shared_mem_buffer[32];
+  const int leaf_index = static_cast<int>(blockIdx.x);
+  const data_size_t* data_indices_in_leaf = data_indices + leaf_data_start[leaf_index];
+  const data_size_t num_data_in_leaf = leaf_num_data[leaf_index];
+  double sum_gradients = 0.0f;
+  double sum_hessians = 0.0f;
+  for (data_size_t inner_data_index = static_cast<int>(threadIdx.x);
+    inner_data_index < num_data_in_leaf; inner_data_index += static_cast<int>(blockDim.x)) {
+    const data_size_t data_index = data_indices_in_leaf[inner_data_index];
+    const score_t gradient = gradients[data_index];
+    const score_t hessian = hessians[data_index];
+    sum_gradients += static_cast<double>(gradient);
+    sum_hessians += static_cast<double>(hessian);
+  }
+  sum_gradients = ShuffleReduceSum<double>(sum_gradients, shared_mem_buffer, blockDim.x);
+  __syncthreads();
+  sum_hessians = ShuffleReduceSum<double>(sum_hessians, shared_mem_buffer, blockDim.x);
+  if (threadIdx.x == 0) {
+    leaf_grad_stat_buffer[leaf_index] = sum_gradients;
+    leaf_hess_stat_buffer[leaf_index] = sum_hessians;
+  }
+}
+
+void CUDADataPartition::LaunchReduceLeafGradStat(
+  const score_t* gradients, const score_t* hessians,
+  CUDATree* tree, double* leaf_grad_stat_buffer, double* leaf_hess_state_buffer) const {
+  const int num_blocks = tree->num_leaves();
+  RenewDiscretizedTreeLeavesKernel<<<num_blocks, FILL_INDICES_BLOCK_SIZE_DATA_PARTITION>>>(
+    gradients,
+    hessians,
+    cuda_data_indices_,
+    cuda_leaf_data_start_,
+    cuda_leaf_num_data_,
+    leaf_grad_stat_buffer,
+    leaf_hess_state_buffer,
+    tree->cuda_leaf_value_ref());
+}
+
 }  // namespace LightGBM
 
 #endif  // USE_CUDA

src/treelearner/cuda/cuda_data_partition.hpp

@@ -78,6 +78,10 @@ class CUDADataPartition {
 
   void ResetByLeafPred(const std::vector<int>& leaf_pred, int num_leaves);
 
+  void ReduceLeafGradStat(
+    const score_t* gradients, const score_t* hessians,
+    CUDATree* tree, double* leaf_grad_stat_buffer, double* leaf_hess_state_buffer) const;
+
   data_size_t root_num_data() const {
     if (use_bagging_) {
       return num_used_indices_;
@@ -292,6 +296,10 @@ class CUDADataPartition {
 
   void LaunchFillDataIndexToLeafIndex();
 
+  void LaunchReduceLeafGradStat(
+    const score_t* gradients, const score_t* hessians,
+    CUDATree* tree, double* leaf_grad_stat_buffer, double* leaf_hess_state_buffer) const;
+
   // Host memory
 
   // dataset information

src/treelearner/cuda/cuda_gradient_discretizer.cu

@@ -0,0 +1,171 @@
+/*!
+ * Copyright (c) 2021 Microsoft Corporation. All rights reserved.
+ * Licensed under the MIT License. See LICENSE file in the project root for
+ * license information.
+ */
+
+#ifdef USE_CUDA
+
+#include <algorithm>
+
+#include <LightGBM/cuda/cuda_algorithms.hpp>
+
+#include "cuda_gradient_discretizer.hpp"
+
+namespace LightGBM {
+
+__global__ void ReduceMinMaxKernel(
+  const data_size_t num_data,
+  const score_t* input_gradients,
+  const score_t* input_hessians,
+  score_t* grad_min_block_buffer,
+  score_t* grad_max_block_buffer,
+  score_t* hess_min_block_buffer,
+  score_t* hess_max_block_buffer) {
+  __shared__ score_t shared_mem_buffer[32];
+  const data_size_t index = static_cast<data_size_t>(threadIdx.x + blockIdx.x * blockDim.x);
+  score_t grad_max_val = kMinScore;
+  score_t grad_min_val = kMaxScore;
+  score_t hess_max_val = kMinScore;
+  score_t hess_min_val = kMaxScore;
+  if (index < num_data) {
+    grad_max_val = input_gradients[index];
+    grad_min_val = input_gradients[index];
+    hess_max_val = input_hessians[index];
+    hess_min_val = input_hessians[index];
+  }
+  grad_min_val = ShuffleReduceMin<score_t>(grad_min_val, shared_mem_buffer, blockDim.x);
+  __syncthreads();
+  grad_max_val = ShuffleReduceMax<score_t>(grad_max_val, shared_mem_buffer, blockDim.x);
+  __syncthreads();
+  hess_min_val = ShuffleReduceMin<score_t>(hess_min_val, shared_mem_buffer, blockDim.x);
+  __syncthreads();
+  hess_max_val = ShuffleReduceMax<score_t>(hess_max_val, shared_mem_buffer, blockDim.x);
+  if (threadIdx.x == 0) {
+    grad_min_block_buffer[blockIdx.x] = grad_min_val;
+    grad_max_block_buffer[blockIdx.x] = grad_max_val;
+    hess_min_block_buffer[blockIdx.x] = hess_min_val;
+    hess_max_block_buffer[blockIdx.x] = hess_max_val;
+  }
+}
+
+__global__ void ReduceBlockMinMaxKernel(
+  const int num_blocks,
+  const int grad_discretize_bins,
+  score_t* grad_min_block_buffer,
+  score_t* grad_max_block_buffer,
+  score_t* hess_min_block_buffer,
+  score_t* hess_max_block_buffer) {
+  __shared__ score_t shared_mem_buffer[32];
+  score_t grad_max_val = kMinScore;
+  score_t grad_min_val = kMaxScore;
+  score_t hess_max_val = kMinScore;
+  score_t hess_min_val = kMaxScore;
+  for (int block_index = static_cast<int>(threadIdx.x); block_index < num_blocks; block_index += static_cast<int>(blockDim.x)) {
+    grad_min_val = min(grad_min_val, grad_min_block_buffer[block_index]);
+    grad_max_val = max(grad_max_val, grad_max_block_buffer[block_index]);
+    hess_min_val = min(hess_min_val, hess_min_block_buffer[block_index]);
+    hess_max_val = max(hess_max_val, hess_max_block_buffer[block_index]);
+  }
+  grad_min_val = ShuffleReduceMin<score_t>(grad_min_val, shared_mem_buffer, blockDim.x);
+  __syncthreads();
+  grad_max_val = ShuffleReduceMax<score_t>(grad_max_val, shared_mem_buffer, blockDim.x);
+  __syncthreads();
+  hess_max_val = ShuffleReduceMax<score_t>(hess_max_val, shared_mem_buffer, blockDim.x);
+  __syncthreads();
+  hess_max_val = ShuffleReduceMax<score_t>(hess_max_val, shared_mem_buffer, blockDim.x);
+  if (threadIdx.x == 0) {
+    const score_t grad_abs_max = max(fabs(grad_min_val), fabs(grad_max_val));
+    const score_t hess_abs_max = max(fabs(hess_min_val), fabs(hess_max_val));
+    grad_min_block_buffer[0] = 1.0f / (grad_abs_max / (grad_discretize_bins / 2));
+    grad_max_block_buffer[0] = (grad_abs_max / (grad_discretize_bins / 2));
+    hess_min_block_buffer[0] = 1.0f / (hess_abs_max / (grad_discretize_bins));
+    hess_max_block_buffer[0] = (hess_abs_max / (grad_discretize_bins));
+  }
+}
+
+template <bool STOCHASTIC_ROUNDING>
+__global__ void DiscretizeGradientsKernel(
+  const data_size_t num_data,
+  const score_t* input_gradients,
+  const score_t* input_hessians,
+  const score_t* grad_scale_ptr,
+  const score_t* hess_scale_ptr,
+  const int iter,
+  const int* random_values_use_start,
+  const score_t* gradient_random_values,
+  const score_t* hessian_random_values,
+  const int grad_discretize_bins,
+  int8_t* output_gradients_and_hessians) {
+  const int start = random_values_use_start[iter];
+  const data_size_t index = static_cast<data_size_t>(threadIdx.x + blockIdx.x * blockDim.x);
+  const score_t grad_scale = *grad_scale_ptr;
+  const score_t hess_scale = *hess_scale_ptr;
+  int16_t* output_gradients_and_hessians_ptr = reinterpret_cast<int16_t*>(output_gradients_and_hessians);
+  if (index < num_data) {
+    if (STOCHASTIC_ROUNDING) {
+      const data_size_t index_offset = (index + start) % num_data;
+      const score_t gradient = input_gradients[index];
+      const score_t hessian = input_hessians[index];
+      const score_t gradient_random_value = gradient_random_values[index_offset];
+      const score_t hessian_random_value = hessian_random_values[index_offset];
+      output_gradients_and_hessians_ptr[2 * index + 1] = gradient > 0.0f ?
+        static_cast<int16_t>(gradient * grad_scale + gradient_random_value) :
+        static_cast<int16_t>(gradient * grad_scale - gradient_random_value);
+      output_gradients_and_hessians_ptr[2 * index] = static_cast<int16_t>(hessian * hess_scale + hessian_random_value);
+    } else {
+      const score_t gradient = input_gradients[index];
+      const score_t hessian = input_hessians[index];
+      output_gradients_and_hessians_ptr[2 * index + 1] = gradient > 0.0f ?
+        static_cast<int16_t>(gradient * grad_scale + 0.5) :
+        static_cast<int16_t>(gradient * grad_scale - 0.5);
+      output_gradients_and_hessians_ptr[2 * index] = static_cast<int16_t>(hessian * hess_scale + 0.5);
+    }
+  }
+}
+
+void CUDAGradientDiscretizer::DiscretizeGradients(
+  const data_size_t num_data,
+  const score_t* input_gradients,
+  const score_t* input_hessians) {
+  ReduceMinMaxKernel<<<num_reduce_blocks_, CUDA_GRADIENT_DISCRETIZER_BLOCK_SIZE>>>(
+    num_data, input_gradients, input_hessians,
+    grad_min_block_buffer_.RawData(),
+    grad_max_block_buffer_.RawData(),
+    hess_min_block_buffer_.RawData(),
+    hess_max_block_buffer_.RawData());
+    SynchronizeCUDADevice(__FILE__, __LINE__);
+  ReduceBlockMinMaxKernel<<<1, CUDA_GRADIENT_DISCRETIZER_BLOCK_SIZE>>>(
+    num_reduce_blocks_,
+    num_grad_quant_bins_,
+    grad_min_block_buffer_.RawData(),
+    grad_max_block_buffer_.RawData(),
+    hess_min_block_buffer_.RawData(),
+    hess_max_block_buffer_.RawData());
+    SynchronizeCUDADevice(__FILE__, __LINE__);
+
+  #define DiscretizeGradientsKernel_ARGS \
+    num_data, \
+    input_gradients, \
+    input_hessians, \
+    grad_min_block_buffer_.RawData(), \
+    hess_min_block_buffer_.RawData(), \
+    iter_, \
+    random_values_use_start_.RawData(), \
+    gradient_random_values_.RawData(), \
+    hessian_random_values_.RawData(), \
+    num_grad_quant_bins_, \
+    discretized_gradients_and_hessians_.RawData()
+
+  if (stochastic_rounding_) {
+    DiscretizeGradientsKernel<true><<<num_reduce_blocks_, CUDA_GRADIENT_DISCRETIZER_BLOCK_SIZE>>>(DiscretizeGradientsKernel_ARGS);
+  } else {
+    DiscretizeGradientsKernel<false><<<num_reduce_blocks_, CUDA_GRADIENT_DISCRETIZER_BLOCK_SIZE>>>(DiscretizeGradientsKernel_ARGS);
+  }
+  SynchronizeCUDADevice(__FILE__, __LINE__);
+  ++iter_;
+}
+
+}  // namespace LightGBM
+
+#endif  // USE_CUDA

src/treelearner/cuda/cuda_gradient_discretizer.hpp

@@ -0,0 +1,118 @@
+/*!
+ * Copyright (c) 2021 Microsoft Corporation. All rights reserved.
+ * Licensed under the MIT License. See LICENSE file in the project root for
+ * license information.
+ */
+
+#ifndef LIGHTGBM_TREELEARNER_CUDA_CUDA_GRADIENT_DISCRETIZER_HPP_
+#define LIGHTGBM_TREELEARNER_CUDA_CUDA_GRADIENT_DISCRETIZER_HPP_
+
+#ifdef USE_CUDA
+
+#include <LightGBM/bin.h>
+#include <LightGBM/meta.h>
+#include <LightGBM/cuda/cuda_utils.hu>
+#include <LightGBM/utils/threading.h>
+
+#include <algorithm>
+#include <random>
+#include <vector>
+
+#include "cuda_leaf_splits.hpp"
+#include "../gradient_discretizer.hpp"
+
+namespace LightGBM {
+
+#define CUDA_GRADIENT_DISCRETIZER_BLOCK_SIZE (1024)
+
+class CUDAGradientDiscretizer: public GradientDiscretizer {
+ public:
+  CUDAGradientDiscretizer(int num_grad_quant_bins, int num_trees, int random_seed, bool is_constant_hessian, bool stochastic_roudning):
+    GradientDiscretizer(num_grad_quant_bins, num_trees, random_seed, is_constant_hessian, stochastic_roudning) {
+  }
+
+  void DiscretizeGradients(
+    const data_size_t num_data,
+    const score_t* input_gradients,
+    const score_t* input_hessians) override;
+
+  const int8_t* discretized_gradients_and_hessians() const override { return discretized_gradients_and_hessians_.RawData(); }
+
+  double grad_scale() const override {
+    Log::Fatal("grad_scale() of CUDAGradientDiscretizer should not be called.");
+    return 0.0;
+  }
+
+  double hess_scale() const override {
+    Log::Fatal("hess_scale() of CUDAGradientDiscretizer should not be called.");
+    return 0.0;
+  }
+
+  const score_t* grad_scale_ptr() const { return grad_max_block_buffer_.RawData(); }
+
+  const score_t* hess_scale_ptr() const { return hess_max_block_buffer_.RawData(); }
+
+  void Init(const data_size_t num_data, const int num_leaves,
+    const int num_features, const Dataset* train_data) override {
+    GradientDiscretizer::Init(num_data, num_leaves, num_features, train_data);
+    discretized_gradients_and_hessians_.Resize(num_data * 2);
+    num_reduce_blocks_ = (num_data + CUDA_GRADIENT_DISCRETIZER_BLOCK_SIZE - 1) / CUDA_GRADIENT_DISCRETIZER_BLOCK_SIZE;
+    grad_min_block_buffer_.Resize(num_reduce_blocks_);
+    grad_max_block_buffer_.Resize(num_reduce_blocks_);
+    hess_min_block_buffer_.Resize(num_reduce_blocks_);
+    hess_max_block_buffer_.Resize(num_reduce_blocks_);
+    random_values_use_start_.Resize(num_trees_);
+    gradient_random_values_.Resize(num_data);
+    hessian_random_values_.Resize(num_data);
+
+    std::vector<score_t> gradient_random_values(num_data, 0.0f);
+    std::vector<score_t> hessian_random_values(num_data, 0.0f);
+    std::vector<int> random_values_use_start(num_trees_, 0);
+
+    const int num_threads = OMP_NUM_THREADS();
+
+    std::mt19937 random_values_use_start_eng = std::mt19937(random_seed_);
+    std::uniform_int_distribution<data_size_t> random_values_use_start_dist = std::uniform_int_distribution<data_size_t>(0, num_data);
+    for (int tree_index = 0; tree_index < num_trees_; ++tree_index) {
+      random_values_use_start[tree_index] = random_values_use_start_dist(random_values_use_start_eng);
+    }
+
+    int num_blocks = 0;
+    data_size_t block_size = 0;
+    Threading::BlockInfo<data_size_t>(num_data, 512, &num_blocks, &block_size);
+    #pragma omp parallel for schedule(static, 1) num_threads(num_threads)
+    for (int thread_id = 0; thread_id < num_blocks; ++thread_id) {
+      const data_size_t start = thread_id * block_size;
+      const data_size_t end = std::min(start + block_size, num_data);
+      std::mt19937 gradient_random_values_eng(random_seed_ + thread_id);
+      std::uniform_real_distribution<double> gradient_random_values_dist(0.0f, 1.0f);
+      std::mt19937 hessian_random_values_eng(random_seed_ + thread_id + num_threads);
+      std::uniform_real_distribution<double> hessian_random_values_dist(0.0f, 1.0f);
+      for (data_size_t i = start; i < end; ++i) {
+        gradient_random_values[i] = gradient_random_values_dist(gradient_random_values_eng);
+        hessian_random_values[i] = hessian_random_values_dist(hessian_random_values_eng);
+      }
+    }
+
+    CopyFromHostToCUDADevice<score_t>(gradient_random_values_.RawData(), gradient_random_values.data(), gradient_random_values.size(), __FILE__, __LINE__);
+    CopyFromHostToCUDADevice<score_t>(hessian_random_values_.RawData(), hessian_random_values.data(), hessian_random_values.size(), __FILE__, __LINE__);
+    CopyFromHostToCUDADevice<int>(random_values_use_start_.RawData(), random_values_use_start.data(), random_values_use_start.size(), __FILE__, __LINE__);
+    iter_ = 0;
+  }
+
+ protected:
+  mutable CUDAVector<int8_t> discretized_gradients_and_hessians_;
+  mutable CUDAVector<score_t> grad_min_block_buffer_;
+  mutable CUDAVector<score_t> grad_max_block_buffer_;
+  mutable CUDAVector<score_t> hess_min_block_buffer_;
+  mutable CUDAVector<score_t> hess_max_block_buffer_;
+  CUDAVector<int> random_values_use_start_;
+  CUDAVector<score_t> gradient_random_values_;
+  CUDAVector<score_t> hessian_random_values_;
+  int num_reduce_blocks_;
+};
+
+}  // namespace LightGBM
+
+#endif  // USE_CUDA
+#endif  // LIGHTGBM_TREELEARNER_CUDA_CUDA_GRADIENT_DISCRETIZER_HPP_

src/treelearner/cuda/cuda_histogram_constructor.cpp

@@ -20,7 +20,9 @@ CUDAHistogramConstructor::CUDAHistogramConstructor(
   const int min_data_in_leaf,
   const double min_sum_hessian_in_leaf,
   const int gpu_device_id,
-  const bool gpu_use_dp):
+  const bool gpu_use_dp,
+  const bool use_quantized_grad,
+  const int num_grad_quant_bins):
   num_data_(train_data->num_data()),
   num_features_(train_data->num_features()),
   num_leaves_(num_leaves),
@@ -28,24 +30,14 @@ CUDAHistogramConstructor::CUDAHistogramConstructor(
   min_data_in_leaf_(min_data_in_leaf),
   min_sum_hessian_in_leaf_(min_sum_hessian_in_leaf),
   gpu_device_id_(gpu_device_id),
-  gpu_use_dp_(gpu_use_dp) {
+  gpu_use_dp_(gpu_use_dp),
+  use_quantized_grad_(use_quantized_grad),
+  num_grad_quant_bins_(num_grad_quant_bins) {
   InitFeatureMetaInfo(train_data, feature_hist_offsets);
   cuda_row_data_.reset(nullptr);
-  cuda_feature_num_bins_ = nullptr;
-  cuda_feature_hist_offsets_ = nullptr;
-  cuda_feature_most_freq_bins_ = nullptr;
-  cuda_hist_ = nullptr;
-  cuda_need_fix_histogram_features_ = nullptr;
-  cuda_need_fix_histogram_features_num_bin_aligned_ = nullptr;
 }
 
 CUDAHistogramConstructor::~CUDAHistogramConstructor() {
-  DeallocateCUDAMemory<uint32_t>(&cuda_feature_num_bins_, __FILE__, __LINE__);
-  DeallocateCUDAMemory<uint32_t>(&cuda_feature_hist_offsets_, __FILE__, __LINE__);
-  DeallocateCUDAMemory<uint32_t>(&cuda_feature_most_freq_bins_, __FILE__, __LINE__);
-  DeallocateCUDAMemory<hist_t>(&cuda_hist_, __FILE__, __LINE__);
-  DeallocateCUDAMemory<int>(&cuda_need_fix_histogram_features_, __FILE__, __LINE__);
-  DeallocateCUDAMemory<uint32_t>(&cuda_need_fix_histogram_features_num_bin_aligned_, __FILE__, __LINE__);
   gpuAssert(cudaStreamDestroy(cuda_stream_), __FILE__, __LINE__);
 }
 
@@ -84,54 +76,70 @@ void CUDAHistogramConstructor::InitFeatureMetaInfo(const Dataset* train_data, co
 void CUDAHistogramConstructor::BeforeTrain(const score_t* gradients, const score_t* hessians) {
   cuda_gradients_ = gradients;
   cuda_hessians_ = hessians;
-  SetCUDAMemory<hist_t>(cuda_hist_, 0, num_total_bin_ * 2 * num_leaves_, __FILE__, __LINE__);
+  cuda_hist_.SetValue(0);
 }
 
 void CUDAHistogramConstructor::Init(const Dataset* train_data, TrainingShareStates* share_state) {
-  AllocateCUDAMemory<hist_t>(&cuda_hist_, num_total_bin_ * 2 * num_leaves_, __FILE__, __LINE__);
-  SetCUDAMemory<hist_t>(cuda_hist_, 0, num_total_bin_ * 2 * num_leaves_, __FILE__, __LINE__);
+  cuda_hist_.Resize(static_cast<size_t>(num_total_bin_ * 2 * num_leaves_));
+  cuda_hist_.SetValue(0);
 
-  InitCUDAMemoryFromHostMemory<uint32_t>(&cuda_feature_num_bins_,
-    feature_num_bins_.data(), feature_num_bins_.size(), __FILE__, __LINE__);
-
-  InitCUDAMemoryFromHostMemory<uint32_t>(&cuda_feature_hist_offsets_,
-    feature_hist_offsets_.data(), feature_hist_offsets_.size(), __FILE__, __LINE__);
-
-  InitCUDAMemoryFromHostMemory<uint32_t>(&cuda_feature_most_freq_bins_,
-    feature_most_freq_bins_.data(), feature_most_freq_bins_.size(), __FILE__, __LINE__);
+  cuda_feature_num_bins_.InitFromHostVector(feature_num_bins_);
+  cuda_feature_hist_offsets_.InitFromHostVector(feature_hist_offsets_);
+  cuda_feature_most_freq_bins_.InitFromHostVector(feature_most_freq_bins_);
 
   cuda_row_data_.reset(new CUDARowData(train_data, share_state, gpu_device_id_, gpu_use_dp_));
   cuda_row_data_->Init(train_data, share_state);
 
   CUDASUCCESS_OR_FATAL(cudaStreamCreate(&cuda_stream_));
 
-  InitCUDAMemoryFromHostMemory<int>(&cuda_need_fix_histogram_features_, need_fix_histogram_features_.data(), need_fix_histogram_features_.size(), __FILE__, __LINE__);
-  InitCUDAMemoryFromHostMemory<uint32_t>(&cuda_need_fix_histogram_features_num_bin_aligned_, need_fix_histogram_features_num_bin_aligend_.data(),
-    need_fix_histogram_features_num_bin_aligend_.size(), __FILE__, __LINE__);
+  cuda_need_fix_histogram_features_.InitFromHostVector(need_fix_histogram_features_);
+  cuda_need_fix_histogram_features_num_bin_aligned_.InitFromHostVector(need_fix_histogram_features_num_bin_aligend_);
 
   if (cuda_row_data_->NumLargeBinPartition() > 0) {
     int grid_dim_x = 0, grid_dim_y = 0, block_dim_x = 0, block_dim_y = 0;
     CalcConstructHistogramKernelDim(&grid_dim_x, &grid_dim_y, &block_dim_x, &block_dim_y, num_data_);
-    const size_t buffer_size = static_cast<size_t>(grid_dim_y) * static_cast<size_t>(num_total_bin_) * 2;
-    AllocateCUDAMemory<float>(&cuda_hist_buffer_, buffer_size, __FILE__, __LINE__);
+    const size_t buffer_size = static_cast<size_t>(grid_dim_y) * static_cast<size_t>(num_total_bin_);
+    if (!use_quantized_grad_) {
+      if (gpu_use_dp_) {
+        // need to double the size of histogram buffer in global memory when using double precision in histogram construction
+        cuda_hist_buffer_.Resize(buffer_size * 4);
+      } else {
+        cuda_hist_buffer_.Resize(buffer_size * 2);
       }
+    } else {
+      // use only half the size of histogram buffer in global memory when quantized training since each gradient and hessian takes only 2 bytes
+      cuda_hist_buffer_.Resize(buffer_size);
+    }
+  }
+  hist_buffer_for_num_bit_change_.Resize(num_total_bin_ * 2);
 }
 
 void CUDAHistogramConstructor::ConstructHistogramForLeaf(
   const CUDALeafSplitsStruct* cuda_smaller_leaf_splits,
-  const CUDALeafSplitsStruct* cuda_larger_leaf_splits,
+  const CUDALeafSplitsStruct* /*cuda_larger_leaf_splits*/,
   const data_size_t num_data_in_smaller_leaf,
   const data_size_t num_data_in_larger_leaf,
   const double sum_hessians_in_smaller_leaf,
-  const double sum_hessians_in_larger_leaf) {
+  const double sum_hessians_in_larger_leaf,
+  const uint8_t num_bits_in_histogram_bins) {
   if ((num_data_in_smaller_leaf <= min_data_in_leaf_ || sum_hessians_in_smaller_leaf <= min_sum_hessian_in_leaf_) &&
     (num_data_in_larger_leaf <= min_data_in_leaf_ || sum_hessians_in_larger_leaf <= min_sum_hessian_in_leaf_)) {
     return;
   }
-  LaunchConstructHistogramKernel(cuda_smaller_leaf_splits, num_data_in_smaller_leaf);
+  LaunchConstructHistogramKernel(cuda_smaller_leaf_splits, num_data_in_smaller_leaf, num_bits_in_histogram_bins);
   SynchronizeCUDADevice(__FILE__, __LINE__);
+}
+
+void CUDAHistogramConstructor::SubtractHistogramForLeaf(
+  const CUDALeafSplitsStruct* cuda_smaller_leaf_splits,
+  const CUDALeafSplitsStruct* cuda_larger_leaf_splits,
+  const bool use_quantized_grad,
+  const uint8_t parent_num_bits_in_histogram_bins,
+  const uint8_t smaller_num_bits_in_histogram_bins,
+  const uint8_t larger_num_bits_in_histogram_bins) {
   global_timer.Start("CUDAHistogramConstructor::ConstructHistogramForLeaf::LaunchSubtractHistogramKernel");
-  LaunchSubtractHistogramKernel(cuda_smaller_leaf_splits, cuda_larger_leaf_splits);
+  LaunchSubtractHistogramKernel(cuda_smaller_leaf_splits, cuda_larger_leaf_splits, use_quantized_grad,
+                                parent_num_bits_in_histogram_bins, smaller_num_bits_in_histogram_bins, larger_num_bits_in_histogram_bins);
   global_timer.Stop("CUDAHistogramConstructor::ConstructHistogramForLeaf::LaunchSubtractHistogramKernel");
 }
 
@@ -152,33 +160,18 @@ void CUDAHistogramConstructor::ResetTrainingData(const Dataset* train_data, Trai
   num_data_ = train_data->num_data();
   num_features_ = train_data->num_features();
   InitFeatureMetaInfo(train_data, share_states->feature_hist_offsets());
-  if (feature_num_bins_.size() > 0) {
-    DeallocateCUDAMemory<uint32_t>(&cuda_feature_num_bins_, __FILE__, __LINE__);
-    DeallocateCUDAMemory<uint32_t>(&cuda_feature_hist_offsets_, __FILE__, __LINE__);
-    DeallocateCUDAMemory<uint32_t>(&cuda_feature_most_freq_bins_, __FILE__, __LINE__);
-    DeallocateCUDAMemory<int>(&cuda_need_fix_histogram_features_, __FILE__, __LINE__);
-    DeallocateCUDAMemory<uint32_t>(&cuda_need_fix_histogram_features_num_bin_aligned_, __FILE__, __LINE__);
-    DeallocateCUDAMemory<hist_t>(&cuda_hist_, __FILE__, __LINE__);
-  }
 
-  AllocateCUDAMemory<hist_t>(&cuda_hist_, num_total_bin_ * 2 * num_leaves_, __FILE__, __LINE__);
-  SetCUDAMemory<hist_t>(cuda_hist_, 0, num_total_bin_ * 2 * num_leaves_, __FILE__, __LINE__);
-
-  InitCUDAMemoryFromHostMemory<uint32_t>(&cuda_feature_num_bins_,
-    feature_num_bins_.data(), feature_num_bins_.size(), __FILE__, __LINE__);
-
-  InitCUDAMemoryFromHostMemory<uint32_t>(&cuda_feature_hist_offsets_,
-    feature_hist_offsets_.data(), feature_hist_offsets_.size(), __FILE__, __LINE__);
-
-  InitCUDAMemoryFromHostMemory<uint32_t>(&cuda_feature_most_freq_bins_,
-    feature_most_freq_bins_.data(), feature_most_freq_bins_.size(), __FILE__, __LINE__);
+  cuda_hist_.Resize(static_cast<size_t>(num_total_bin_ * 2 * num_leaves_));
+  cuda_hist_.SetValue(0);
+  cuda_feature_num_bins_.InitFromHostVector(feature_num_bins_);
+  cuda_feature_hist_offsets_.InitFromHostVector(feature_hist_offsets_);
+  cuda_feature_most_freq_bins_.InitFromHostVector(feature_most_freq_bins_);
 
   cuda_row_data_.reset(new CUDARowData(train_data, share_states, gpu_device_id_, gpu_use_dp_));
   cuda_row_data_->Init(train_data, share_states);
 
-  InitCUDAMemoryFromHostMemory<int>(&cuda_need_fix_histogram_features_, need_fix_histogram_features_.data(), need_fix_histogram_features_.size(), __FILE__, __LINE__);
-  InitCUDAMemoryFromHostMemory<uint32_t>(&cuda_need_fix_histogram_features_num_bin_aligned_, need_fix_histogram_features_num_bin_aligend_.data(),
-    need_fix_histogram_features_num_bin_aligend_.size(), __FILE__, __LINE__);
+  cuda_need_fix_histogram_features_.InitFromHostVector(need_fix_histogram_features_);
+  cuda_need_fix_histogram_features_num_bin_aligned_.InitFromHostVector(need_fix_histogram_features_num_bin_aligend_);
 }
 
 void CUDAHistogramConstructor::ResetConfig(const Config* config) {
@@ -186,9 +179,8 @@ void CUDAHistogramConstructor::ResetConfig(const Config* config) {
   num_leaves_ = config->num_leaves;
   min_data_in_leaf_ = config->min_data_in_leaf;
   min_sum_hessian_in_leaf_ = config->min_sum_hessian_in_leaf;
-  DeallocateCUDAMemory<hist_t>(&cuda_hist_, __FILE__, __LINE__);
-  AllocateCUDAMemory<hist_t>(&cuda_hist_, num_total_bin_ * 2 * num_leaves_, __FILE__, __LINE__);
-  SetCUDAMemory<hist_t>(cuda_hist_, 0, num_total_bin_ * 2 * num_leaves_, __FILE__, __LINE__);
+  cuda_hist_.Resize(static_cast<size_t>(num_total_bin_ * 2 * num_leaves_));
+  cuda_hist_.SetValue(0);
 }
 
 }  // namespace LightGBM

src/treelearner/cuda/cuda_histogram_constructor.cu

@@ -125,7 +125,7 @@ __global__ void CUDAConstructHistogramSparseKernel(
   }
 }
 
-template <typename BIN_TYPE>
+template <typename BIN_TYPE, typename HIST_TYPE>
 __global__ void CUDAConstructHistogramDenseKernel_GlobalMemory(
   const CUDALeafSplitsStruct* smaller_leaf_splits,
   const score_t* cuda_gradients,
@@ -135,7 +135,7 @@ __global__ void CUDAConstructHistogramDenseKernel_GlobalMemory(
   const uint32_t* column_hist_offsets_full,
   const int* feature_partition_column_index_offsets,
   const data_size_t num_data,
-  float* global_hist_buffer) {
+  HIST_TYPE* global_hist_buffer) {
   const int dim_y = static_cast<int>(gridDim.y * blockDim.y);
   const data_size_t num_data_in_smaller_leaf = smaller_leaf_splits->num_data_in_leaf;
   const data_size_t num_data_per_thread = (num_data_in_smaller_leaf + dim_y - 1) / dim_y;
@@ -150,7 +150,7 @@ __global__ void CUDAConstructHistogramDenseKernel_GlobalMemory(
   const uint32_t num_items_in_partition = (partition_hist_end - partition_hist_start) << 1;
   const unsigned int thread_idx = threadIdx.x + threadIdx.y * blockDim.x;
   const int num_total_bin = column_hist_offsets_full[gridDim.x];
-  float* shared_hist = global_hist_buffer + (blockIdx.y * num_total_bin + partition_hist_start) * 2;
+  HIST_TYPE* shared_hist = global_hist_buffer + (blockIdx.y * num_total_bin + partition_hist_start) * 2;
   for (unsigned int i = thread_idx; i < num_items_in_partition; i += num_threads_per_block) {
     shared_hist[i] = 0.0f;
   }
@@ -166,14 +166,14 @@ __global__ void CUDAConstructHistogramDenseKernel_GlobalMemory(
   data_size_t inner_data_index = static_cast<data_size_t>(threadIdx_y);
   const int column_index = static_cast<int>(threadIdx.x) + partition_column_start;
   if (threadIdx.x < static_cast<unsigned int>(num_columns_in_partition)) {
-    float* shared_hist_ptr = shared_hist + (column_hist_offsets[column_index] << 1);
+    HIST_TYPE* shared_hist_ptr = shared_hist + (column_hist_offsets[column_index] << 1);
     for (data_size_t i = 0; i < num_iteration_this; ++i) {
       const data_size_t data_index = data_indices_ref_this_block[inner_data_index];
       const score_t grad = cuda_gradients[data_index];
       const score_t hess = cuda_hessians[data_index];
       const uint32_t bin = static_cast<uint32_t>(data_ptr[static_cast<size_t>(data_index) * num_columns_in_partition + threadIdx.x]);
       const uint32_t pos = bin << 1;
-      float* pos_ptr = shared_hist_ptr + pos;
+      HIST_TYPE* pos_ptr = shared_hist_ptr + pos;
       atomicAdd_block(pos_ptr, grad);
       atomicAdd_block(pos_ptr + 1, hess);
       inner_data_index += blockDim.y;
@@ -186,7 +186,7 @@ __global__ void CUDAConstructHistogramDenseKernel_GlobalMemory(
   }
 }
 
-template <typename BIN_TYPE, typename DATA_PTR_TYPE>
+template <typename BIN_TYPE, typename HIST_TYPE, typename DATA_PTR_TYPE>
 __global__ void CUDAConstructHistogramSparseKernel_GlobalMemory(
   const CUDALeafSplitsStruct* smaller_leaf_splits,
   const score_t* cuda_gradients,
@@ -196,7 +196,7 @@ __global__ void CUDAConstructHistogramSparseKernel_GlobalMemory(
   const DATA_PTR_TYPE* partition_ptr,
   const uint32_t* column_hist_offsets_full,
   const data_size_t num_data,
-  float* global_hist_buffer) {
+  HIST_TYPE* global_hist_buffer) {
   const int dim_y = static_cast<int>(gridDim.y * blockDim.y);
   const data_size_t num_data_in_smaller_leaf = smaller_leaf_splits->num_data_in_leaf;
   const data_size_t num_data_per_thread = (num_data_in_smaller_leaf + dim_y - 1) / dim_y;
@@ -209,7 +209,7 @@ __global__ void CUDAConstructHistogramSparseKernel_GlobalMemory(
   const uint32_t num_items_in_partition = (partition_hist_end - partition_hist_start) << 1;
   const unsigned int thread_idx = threadIdx.x + threadIdx.y * blockDim.x;
   const int num_total_bin = column_hist_offsets_full[gridDim.x];
-  float* shared_hist = global_hist_buffer + (blockIdx.y * num_total_bin + partition_hist_start) * 2;
+  HIST_TYPE* shared_hist = global_hist_buffer + (blockIdx.y * num_total_bin + partition_hist_start) * 2;
   for (unsigned int i = thread_idx; i < num_items_in_partition; i += num_threads_per_block) {
     shared_hist[i] = 0.0f;
   }
@@ -233,7 +233,7 @@ __global__ void CUDAConstructHistogramSparseKernel_GlobalMemory(
       const score_t hess = cuda_hessians[data_index];
       const uint32_t bin = static_cast<uint32_t>(data_ptr[row_start + threadIdx.x]);
       const uint32_t pos = bin << 1;
-      float* pos_ptr = shared_hist + pos;
+      HIST_TYPE* pos_ptr = shared_hist + pos;
       atomicAdd_block(pos_ptr, grad);
       atomicAdd_block(pos_ptr + 1, hess);
     }
@@ -246,13 +246,278 @@ __global__ void CUDAConstructHistogramSparseKernel_GlobalMemory(
   }
 }
 
+template <typename BIN_TYPE, int SHARED_HIST_SIZE, bool USE_16BIT_HIST>
+__global__ void CUDAConstructDiscretizedHistogramDenseKernel(
+  const CUDALeafSplitsStruct* smaller_leaf_splits,
+  const int32_t* cuda_gradients_and_hessians,
+  const BIN_TYPE* data,
+  const uint32_t* column_hist_offsets,
+  const uint32_t* column_hist_offsets_full,
+  const int* feature_partition_column_index_offsets,
+  const data_size_t num_data) {
+  const int dim_y = static_cast<int>(gridDim.y * blockDim.y);
+  const data_size_t num_data_in_smaller_leaf = smaller_leaf_splits->num_data_in_leaf;
+  const data_size_t num_data_per_thread = (num_data_in_smaller_leaf + dim_y - 1) / dim_y;
+  const data_size_t* data_indices_ref = smaller_leaf_splits->data_indices_in_leaf;
+  __shared__ int16_t shared_hist[SHARED_HIST_SIZE];
+  int32_t* shared_hist_packed = reinterpret_cast<int32_t*>(shared_hist);
+  const unsigned int num_threads_per_block = blockDim.x * blockDim.y;
+  const int partition_column_start = feature_partition_column_index_offsets[blockIdx.x];
+  const int partition_column_end = feature_partition_column_index_offsets[blockIdx.x + 1];
+  const BIN_TYPE* data_ptr = data + partition_column_start * num_data;
+  const int num_columns_in_partition = partition_column_end - partition_column_start;
+  const uint32_t partition_hist_start = column_hist_offsets_full[blockIdx.x];
+  const uint32_t partition_hist_end = column_hist_offsets_full[blockIdx.x + 1];
+  const uint32_t num_items_in_partition = (partition_hist_end - partition_hist_start);
+  const unsigned int thread_idx = threadIdx.x + threadIdx.y * blockDim.x;
+  for (unsigned int i = thread_idx; i < num_items_in_partition; i += num_threads_per_block) {
+    shared_hist_packed[i] = 0;
+  }
+  __syncthreads();
+  const unsigned int threadIdx_y = threadIdx.y;
+  const unsigned int blockIdx_y = blockIdx.y;
+  const data_size_t block_start = (blockIdx_y * blockDim.y) * num_data_per_thread;
+  const data_size_t* data_indices_ref_this_block = data_indices_ref + block_start;
+  data_size_t block_num_data = max(0, min(num_data_in_smaller_leaf - block_start, num_data_per_thread * static_cast<data_size_t>(blockDim.y)));
+  const data_size_t num_iteration_total = (block_num_data + blockDim.y - 1) / blockDim.y;
+  const data_size_t remainder = block_num_data % blockDim.y;
+  const data_size_t num_iteration_this = remainder == 0 ? num_iteration_total : num_iteration_total - static_cast<data_size_t>(threadIdx_y >= remainder);
+  data_size_t inner_data_index = static_cast<data_size_t>(threadIdx_y);
+  const int column_index = static_cast<int>(threadIdx.x) + partition_column_start;
+  if (threadIdx.x < static_cast<unsigned int>(num_columns_in_partition)) {
+    int32_t* shared_hist_ptr = shared_hist_packed + (column_hist_offsets[column_index]);
+    for (data_size_t i = 0; i < num_iteration_this; ++i) {
+      const data_size_t data_index = data_indices_ref_this_block[inner_data_index];
+      const int32_t grad_and_hess = cuda_gradients_and_hessians[data_index];
+      const uint32_t bin = static_cast<uint32_t>(data_ptr[data_index * num_columns_in_partition + threadIdx.x]);
+      int32_t* pos_ptr = shared_hist_ptr + bin;
+      atomicAdd_block(pos_ptr, grad_and_hess);
+      inner_data_index += blockDim.y;
+    }
+  }
+  __syncthreads();
+  if (USE_16BIT_HIST) {
+    int32_t* feature_histogram_ptr = reinterpret_cast<int32_t*>(smaller_leaf_splits->hist_in_leaf) + partition_hist_start;
+    for (unsigned int i = thread_idx; i < num_items_in_partition; i += num_threads_per_block) {
+      const int32_t packed_grad_hess = shared_hist_packed[i];
+      atomicAdd_system(feature_histogram_ptr + i, packed_grad_hess);
+    }
+  } else {
+    atomic_add_long_t* feature_histogram_ptr = reinterpret_cast<atomic_add_long_t*>(smaller_leaf_splits->hist_in_leaf) + partition_hist_start;
+    for (unsigned int i = thread_idx; i < num_items_in_partition; i += num_threads_per_block) {
+      const int32_t packed_grad_hess = shared_hist_packed[i];
+      const int64_t packed_grad_hess_int64 = (static_cast<int64_t>(static_cast<int16_t>(packed_grad_hess >> 16)) << 32) | (static_cast<int64_t>(packed_grad_hess & 0x0000ffff));
+      atomicAdd_system(feature_histogram_ptr + i, (atomic_add_long_t)(packed_grad_hess_int64));
+    }
+  }
+}
+
+template <typename BIN_TYPE, typename DATA_PTR_TYPE, int SHARED_HIST_SIZE, bool USE_16BIT_HIST>
+__global__ void CUDAConstructDiscretizedHistogramSparseKernel(
+  const CUDALeafSplitsStruct* smaller_leaf_splits,
+  const int32_t* cuda_gradients_and_hessians,
+  const BIN_TYPE* data,
+  const DATA_PTR_TYPE* row_ptr,
+  const DATA_PTR_TYPE* partition_ptr,
+  const uint32_t* column_hist_offsets_full,
+  const data_size_t num_data) {
+  const int dim_y = static_cast<int>(gridDim.y * blockDim.y);
+  const data_size_t num_data_in_smaller_leaf = smaller_leaf_splits->num_data_in_leaf;
+  const data_size_t num_data_per_thread = (num_data_in_smaller_leaf + dim_y - 1) / dim_y;
+  const data_size_t* data_indices_ref = smaller_leaf_splits->data_indices_in_leaf;
+  __shared__ int16_t shared_hist[SHARED_HIST_SIZE];
+  int32_t* shared_hist_packed = reinterpret_cast<int32_t*>(shared_hist);
+  const unsigned int num_threads_per_block = blockDim.x * blockDim.y;
+  const DATA_PTR_TYPE* block_row_ptr = row_ptr + blockIdx.x * (num_data + 1);
+  const BIN_TYPE* data_ptr = data + partition_ptr[blockIdx.x];
+  const uint32_t partition_hist_start = column_hist_offsets_full[blockIdx.x];
+  const uint32_t partition_hist_end = column_hist_offsets_full[blockIdx.x + 1];
+  const uint32_t num_items_in_partition = (partition_hist_end - partition_hist_start);
+  const unsigned int thread_idx = threadIdx.x + threadIdx.y * blockDim.x;
+  for (unsigned int i = thread_idx; i < num_items_in_partition; i += num_threads_per_block) {
+    shared_hist_packed[i] = 0.0f;
+  }
+  __syncthreads();
+  const unsigned int threadIdx_y = threadIdx.y;
+  const unsigned int blockIdx_y = blockIdx.y;
+  const data_size_t block_start = (blockIdx_y * blockDim.y) * num_data_per_thread;
+  const data_size_t* data_indices_ref_this_block = data_indices_ref + block_start;
+  data_size_t block_num_data = max(0, min(num_data_in_smaller_leaf - block_start, num_data_per_thread * static_cast<data_size_t>(blockDim.y)));
+  const data_size_t num_iteration_total = (block_num_data + blockDim.y - 1) / blockDim.y;
+  const data_size_t remainder = block_num_data % blockDim.y;
+  const data_size_t num_iteration_this = remainder == 0 ? num_iteration_total : num_iteration_total - static_cast<data_size_t>(threadIdx_y >= remainder);
+  data_size_t inner_data_index = static_cast<data_size_t>(threadIdx_y);
+  for (data_size_t i = 0; i < num_iteration_this; ++i) {
+    const data_size_t data_index = data_indices_ref_this_block[inner_data_index];
+    const DATA_PTR_TYPE row_start = block_row_ptr[data_index];
+    const DATA_PTR_TYPE row_end = block_row_ptr[data_index + 1];
+    const DATA_PTR_TYPE row_size = row_end - row_start;
+    if (threadIdx.x < row_size) {
+      const int32_t grad_and_hess = cuda_gradients_and_hessians[data_index];
+      const uint32_t bin = static_cast<uint32_t>(data_ptr[row_start + threadIdx.x]);
+      int32_t* pos_ptr = shared_hist_packed + bin;
+      atomicAdd_block(pos_ptr, grad_and_hess);
+    }
+    inner_data_index += blockDim.y;
+  }
+  __syncthreads();
+  if (USE_16BIT_HIST) {
+    int32_t* feature_histogram_ptr = reinterpret_cast<int32_t*>(smaller_leaf_splits->hist_in_leaf) + partition_hist_start;
+    for (unsigned int i = thread_idx; i < num_items_in_partition; i += num_threads_per_block) {
+      const int32_t packed_grad_hess = shared_hist_packed[i];
+      atomicAdd_system(feature_histogram_ptr + i, packed_grad_hess);
+    }
+  } else {
+    atomic_add_long_t* feature_histogram_ptr = reinterpret_cast<atomic_add_long_t*>(smaller_leaf_splits->hist_in_leaf) + partition_hist_start;
+    for (unsigned int i = thread_idx; i < num_items_in_partition; i += num_threads_per_block) {
+      const int32_t packed_grad_hess = shared_hist_packed[i];
+      const int64_t packed_grad_hess_int64 = (static_cast<int64_t>(static_cast<int16_t>(packed_grad_hess >> 16)) << 32) | (static_cast<int64_t>(packed_grad_hess & 0x0000ffff));
+      atomicAdd_system(feature_histogram_ptr + i, (atomic_add_long_t)(packed_grad_hess_int64));
+    }
+  }
+}
+
+template <typename BIN_TYPE, int SHARED_HIST_SIZE, bool USE_16BIT_HIST>
+__global__ void CUDAConstructDiscretizedHistogramDenseKernel_GlobalMemory(
+  const CUDALeafSplitsStruct* smaller_leaf_splits,
+  const int32_t* cuda_gradients_and_hessians,
+  const BIN_TYPE* data,
+  const uint32_t* column_hist_offsets,
+  const uint32_t* column_hist_offsets_full,
+  const int* feature_partition_column_index_offsets,
+  const data_size_t num_data,
+  int32_t* global_hist_buffer) {
+  const int dim_y = static_cast<int>(gridDim.y * blockDim.y);
+  const data_size_t num_data_in_smaller_leaf = smaller_leaf_splits->num_data_in_leaf;
+  const data_size_t num_data_per_thread = (num_data_in_smaller_leaf + dim_y - 1) / dim_y;
+  const data_size_t* data_indices_ref = smaller_leaf_splits->data_indices_in_leaf;
+  const unsigned int num_threads_per_block = blockDim.x * blockDim.y;
+  const int partition_column_start = feature_partition_column_index_offsets[blockIdx.x];
+  const int partition_column_end = feature_partition_column_index_offsets[blockIdx.x + 1];
+  const BIN_TYPE* data_ptr = data + partition_column_start * num_data;
+  const int num_columns_in_partition = partition_column_end - partition_column_start;
+  const uint32_t partition_hist_start = column_hist_offsets_full[blockIdx.x];
+  const uint32_t partition_hist_end = column_hist_offsets_full[blockIdx.x + 1];
+  const uint32_t num_items_in_partition = (partition_hist_end - partition_hist_start);
+  const int num_total_bin = column_hist_offsets_full[gridDim.x];
+  int32_t* shared_hist_packed = global_hist_buffer + (blockIdx.y * num_total_bin + partition_column_start);
+  const unsigned int thread_idx = threadIdx.x + threadIdx.y * blockDim.x;
+  for (unsigned int i = thread_idx; i < num_items_in_partition; i += num_threads_per_block) {
+    shared_hist_packed[i] = 0;
+  }
+  __syncthreads();
+  const unsigned int threadIdx_y = threadIdx.y;
+  const unsigned int blockIdx_y = blockIdx.y;
+  const data_size_t block_start = (blockIdx_y * blockDim.y) * num_data_per_thread;
+  const data_size_t* data_indices_ref_this_block = data_indices_ref + block_start;
+  data_size_t block_num_data = max(0, min(num_data_in_smaller_leaf - block_start, num_data_per_thread * static_cast<data_size_t>(blockDim.y)));
+  const data_size_t num_iteration_total = (block_num_data + blockDim.y - 1) / blockDim.y;
+  const data_size_t remainder = block_num_data % blockDim.y;
+  const data_size_t num_iteration_this = remainder == 0 ? num_iteration_total : num_iteration_total - static_cast<data_size_t>(threadIdx_y >= remainder);
+  data_size_t inner_data_index = static_cast<data_size_t>(threadIdx_y);
+  const int column_index = static_cast<int>(threadIdx.x) + partition_column_start;
+  if (threadIdx.x < static_cast<unsigned int>(num_columns_in_partition)) {
+    int32_t* shared_hist_ptr = shared_hist_packed + (column_hist_offsets[column_index]);
+    for (data_size_t i = 0; i < num_iteration_this; ++i) {
+      const data_size_t data_index = data_indices_ref_this_block[inner_data_index];
+      const int32_t grad_and_hess = cuda_gradients_and_hessians[data_index];
+      const uint32_t bin = static_cast<uint32_t>(data_ptr[data_index * num_columns_in_partition + threadIdx.x]);
+      int32_t* pos_ptr = shared_hist_ptr + bin;
+      atomicAdd_block(pos_ptr, grad_and_hess);
+      inner_data_index += blockDim.y;
+    }
+  }
+  __syncthreads();
+  if (USE_16BIT_HIST) {
+    int32_t* feature_histogram_ptr = reinterpret_cast<int32_t*>(smaller_leaf_splits->hist_in_leaf) + partition_hist_start;
+    for (unsigned int i = thread_idx; i < num_items_in_partition; i += num_threads_per_block) {
+      const int32_t packed_grad_hess = shared_hist_packed[i];
+      atomicAdd_system(feature_histogram_ptr + i, packed_grad_hess);
+    }
+  } else {
+    atomic_add_long_t* feature_histogram_ptr = reinterpret_cast<atomic_add_long_t*>(smaller_leaf_splits->hist_in_leaf) + partition_hist_start;
+    for (unsigned int i = thread_idx; i < num_items_in_partition; i += num_threads_per_block) {
+      const int32_t packed_grad_hess = shared_hist_packed[i];
+      const int64_t packed_grad_hess_int64 = (static_cast<int64_t>(static_cast<int16_t>(packed_grad_hess >> 16)) << 32) | (static_cast<int64_t>(packed_grad_hess & 0x0000ffff));
+      atomicAdd_system(feature_histogram_ptr + i, (atomic_add_long_t)(packed_grad_hess_int64));
+    }
+  }
+}
+
+template <typename BIN_TYPE, typename DATA_PTR_TYPE, int SHARED_HIST_SIZE, bool USE_16BIT_HIST>
+__global__ void CUDAConstructDiscretizedHistogramSparseKernel_GlobalMemory(
+  const CUDALeafSplitsStruct* smaller_leaf_splits,
+  const int32_t* cuda_gradients_and_hessians,
+  const BIN_TYPE* data,
+  const DATA_PTR_TYPE* row_ptr,
+  const DATA_PTR_TYPE* partition_ptr,
+  const uint32_t* column_hist_offsets_full,
+  const data_size_t num_data,
+  int32_t* global_hist_buffer) {
+  const int dim_y = static_cast<int>(gridDim.y * blockDim.y);
+  const data_size_t num_data_in_smaller_leaf = smaller_leaf_splits->num_data_in_leaf;
+  const data_size_t num_data_per_thread = (num_data_in_smaller_leaf + dim_y - 1) / dim_y;
+  const data_size_t* data_indices_ref = smaller_leaf_splits->data_indices_in_leaf;
+  const int num_total_bin = column_hist_offsets_full[gridDim.x];
+  const unsigned int num_threads_per_block = blockDim.x * blockDim.y;
+  const DATA_PTR_TYPE* block_row_ptr = row_ptr + blockIdx.x * (num_data + 1);
+  const BIN_TYPE* data_ptr = data + partition_ptr[blockIdx.x];
+  const uint32_t partition_hist_start = column_hist_offsets_full[blockIdx.x];
+  const uint32_t partition_hist_end = column_hist_offsets_full[blockIdx.x + 1];
+  const uint32_t num_items_in_partition = (partition_hist_end - partition_hist_start);
+  const unsigned int thread_idx = threadIdx.x + threadIdx.y * blockDim.x;
+  int32_t* shared_hist_packed = global_hist_buffer + (blockIdx.y * num_total_bin + partition_hist_start);
+  for (unsigned int i = thread_idx; i < num_items_in_partition; i += num_threads_per_block) {
+    shared_hist_packed[i] = 0.0f;
+  }
+  __syncthreads();
+  const unsigned int threadIdx_y = threadIdx.y;
+  const unsigned int blockIdx_y = blockIdx.y;
+  const data_size_t block_start = (blockIdx_y * blockDim.y) * num_data_per_thread;
+  const data_size_t* data_indices_ref_this_block = data_indices_ref + block_start;
+  data_size_t block_num_data = max(0, min(num_data_in_smaller_leaf - block_start, num_data_per_thread * static_cast<data_size_t>(blockDim.y)));
+  const data_size_t num_iteration_total = (block_num_data + blockDim.y - 1) / blockDim.y;
+  const data_size_t remainder = block_num_data % blockDim.y;
+  const data_size_t num_iteration_this = remainder == 0 ? num_iteration_total : num_iteration_total - static_cast<data_size_t>(threadIdx_y >= remainder);
+  data_size_t inner_data_index = static_cast<data_size_t>(threadIdx_y);
+  for (data_size_t i = 0; i < num_iteration_this; ++i) {
+    const data_size_t data_index = data_indices_ref_this_block[inner_data_index];
+    const DATA_PTR_TYPE row_start = block_row_ptr[data_index];
+    const DATA_PTR_TYPE row_end = block_row_ptr[data_index + 1];
+    const DATA_PTR_TYPE row_size = row_end - row_start;
+    if (threadIdx.x < row_size) {
+      const int32_t grad_and_hess = cuda_gradients_and_hessians[data_index];
+      const uint32_t bin = static_cast<uint32_t>(data_ptr[row_start + threadIdx.x]);
+      int32_t* pos_ptr = shared_hist_packed + bin;
+      atomicAdd_block(pos_ptr, grad_and_hess);
+    }
+    inner_data_index += blockDim.y;
+  }
+  __syncthreads();
+  if (USE_16BIT_HIST) {
+    int32_t* feature_histogram_ptr = reinterpret_cast<int32_t*>(smaller_leaf_splits->hist_in_leaf) + partition_hist_start;
+    for (unsigned int i = thread_idx; i < num_items_in_partition; i += num_threads_per_block) {
+      const int32_t packed_grad_hess = shared_hist_packed[i];
+      atomicAdd_system(feature_histogram_ptr + i, packed_grad_hess);
+    }
+  } else {
+    atomic_add_long_t* feature_histogram_ptr = reinterpret_cast<atomic_add_long_t*>(smaller_leaf_splits->hist_in_leaf) + partition_hist_start;
+    for (unsigned int i = thread_idx; i < num_items_in_partition; i += num_threads_per_block) {
+      const int32_t packed_grad_hess = shared_hist_packed[i];
+      const int64_t packed_grad_hess_int64 = (static_cast<int64_t>(static_cast<int16_t>(packed_grad_hess >> 16)) << 32) | (static_cast<int64_t>(packed_grad_hess & 0x0000ffff));
+      atomicAdd_system(feature_histogram_ptr + i, (atomic_add_long_t)(packed_grad_hess_int64));
+    }
+  }
+}
+
 void CUDAHistogramConstructor::LaunchConstructHistogramKernel(
   const CUDALeafSplitsStruct* cuda_smaller_leaf_splits,
-  const data_size_t num_data_in_smaller_leaf) {
+  const data_size_t num_data_in_smaller_leaf,
+  const uint8_t num_bits_in_histogram_bins) {
   if (cuda_row_data_->shared_hist_size() == DP_SHARED_HIST_SIZE && gpu_use_dp_) {
-    LaunchConstructHistogramKernelInner<double, DP_SHARED_HIST_SIZE>(cuda_smaller_leaf_splits, num_data_in_smaller_leaf);
+    LaunchConstructHistogramKernelInner<double, DP_SHARED_HIST_SIZE>(cuda_smaller_leaf_splits, num_data_in_smaller_leaf, num_bits_in_histogram_bins);
   } else if (cuda_row_data_->shared_hist_size() == SP_SHARED_HIST_SIZE && !gpu_use_dp_) {
-    LaunchConstructHistogramKernelInner<float, SP_SHARED_HIST_SIZE>(cuda_smaller_leaf_splits, num_data_in_smaller_leaf);
+    LaunchConstructHistogramKernelInner<float, SP_SHARED_HIST_SIZE>(cuda_smaller_leaf_splits, num_data_in_smaller_leaf, num_bits_in_histogram_bins);
   } else {
     Log::Fatal("Unknown shared histogram size %d", cuda_row_data_->shared_hist_size());
   }
@@ -261,13 +526,14 @@ void CUDAHistogramConstructor::LaunchConstructHistogramKernel(
 template <typename HIST_TYPE, size_t SHARED_HIST_SIZE>
 void CUDAHistogramConstructor::LaunchConstructHistogramKernelInner(
   const CUDALeafSplitsStruct* cuda_smaller_leaf_splits,
-  const data_size_t num_data_in_smaller_leaf) {
+  const data_size_t num_data_in_smaller_leaf,
+  const uint8_t num_bits_in_histogram_bins) {
   if (cuda_row_data_->bit_type() == 8) {
-    LaunchConstructHistogramKernelInner0<HIST_TYPE, SHARED_HIST_SIZE, uint8_t>(cuda_smaller_leaf_splits, num_data_in_smaller_leaf);
+    LaunchConstructHistogramKernelInner0<HIST_TYPE, SHARED_HIST_SIZE, uint8_t>(cuda_smaller_leaf_splits, num_data_in_smaller_leaf, num_bits_in_histogram_bins);
   } else if (cuda_row_data_->bit_type() == 16) {
-    LaunchConstructHistogramKernelInner0<HIST_TYPE, SHARED_HIST_SIZE, uint16_t>(cuda_smaller_leaf_splits, num_data_in_smaller_leaf);
+    LaunchConstructHistogramKernelInner0<HIST_TYPE, SHARED_HIST_SIZE, uint16_t>(cuda_smaller_leaf_splits, num_data_in_smaller_leaf, num_bits_in_histogram_bins);
   } else if (cuda_row_data_->bit_type() == 32) {
-    LaunchConstructHistogramKernelInner0<HIST_TYPE, SHARED_HIST_SIZE, uint32_t>(cuda_smaller_leaf_splits, num_data_in_smaller_leaf);
+    LaunchConstructHistogramKernelInner0<HIST_TYPE, SHARED_HIST_SIZE, uint32_t>(cuda_smaller_leaf_splits, num_data_in_smaller_leaf, num_bits_in_histogram_bins);
   } else {
     Log::Fatal("Unknown bit_type = %d", cuda_row_data_->bit_type());
   }
@@ -276,16 +542,17 @@ void CUDAHistogramConstructor::LaunchConstructHistogramKernelInner(
 template <typename HIST_TYPE, size_t SHARED_HIST_SIZE, typename BIN_TYPE>
 void CUDAHistogramConstructor::LaunchConstructHistogramKernelInner0(
   const CUDALeafSplitsStruct* cuda_smaller_leaf_splits,
-  const data_size_t num_data_in_smaller_leaf) {
+  const data_size_t num_data_in_smaller_leaf,
+  const uint8_t num_bits_in_histogram_bins) {
   if (cuda_row_data_->row_ptr_bit_type() == 16) {
-    LaunchConstructHistogramKernelInner1<HIST_TYPE, SHARED_HIST_SIZE, BIN_TYPE, uint16_t>(cuda_smaller_leaf_splits, num_data_in_smaller_leaf);
+    LaunchConstructHistogramKernelInner1<HIST_TYPE, SHARED_HIST_SIZE, BIN_TYPE, uint16_t>(cuda_smaller_leaf_splits, num_data_in_smaller_leaf, num_bits_in_histogram_bins);
   } else if (cuda_row_data_->row_ptr_bit_type() == 32) {
-    LaunchConstructHistogramKernelInner1<HIST_TYPE, SHARED_HIST_SIZE, BIN_TYPE, uint32_t>(cuda_smaller_leaf_splits, num_data_in_smaller_leaf);
+    LaunchConstructHistogramKernelInner1<HIST_TYPE, SHARED_HIST_SIZE, BIN_TYPE, uint32_t>(cuda_smaller_leaf_splits, num_data_in_smaller_leaf, num_bits_in_histogram_bins);
   } else if (cuda_row_data_->row_ptr_bit_type() == 64) {
-    LaunchConstructHistogramKernelInner1<HIST_TYPE, SHARED_HIST_SIZE, BIN_TYPE, uint64_t>(cuda_smaller_leaf_splits, num_data_in_smaller_leaf);
+    LaunchConstructHistogramKernelInner1<HIST_TYPE, SHARED_HIST_SIZE, BIN_TYPE, uint64_t>(cuda_smaller_leaf_splits, num_data_in_smaller_leaf, num_bits_in_histogram_bins);
   } else {
     if (!cuda_row_data_->is_sparse()) {
-      LaunchConstructHistogramKernelInner1<HIST_TYPE, SHARED_HIST_SIZE, BIN_TYPE, uint16_t>(cuda_smaller_leaf_splits, num_data_in_smaller_leaf);
+      LaunchConstructHistogramKernelInner1<HIST_TYPE, SHARED_HIST_SIZE, BIN_TYPE, uint16_t>(cuda_smaller_leaf_splits, num_data_in_smaller_leaf, num_bits_in_histogram_bins);
     } else {
       Log::Fatal("Unknown row_ptr_bit_type = %d", cuda_row_data_->row_ptr_bit_type());
     }
@@ -295,18 +562,20 @@ void CUDAHistogramConstructor::LaunchConstructHistogramKernelInner0(
 template <typename HIST_TYPE, size_t SHARED_HIST_SIZE, typename BIN_TYPE, typename PTR_TYPE>
 void CUDAHistogramConstructor::LaunchConstructHistogramKernelInner1(
   const CUDALeafSplitsStruct* cuda_smaller_leaf_splits,
-  const data_size_t num_data_in_smaller_leaf) {
+  const data_size_t num_data_in_smaller_leaf,
+  const uint8_t num_bits_in_histogram_bins) {
   if (cuda_row_data_->NumLargeBinPartition() == 0) {
-    LaunchConstructHistogramKernelInner2<HIST_TYPE, SHARED_HIST_SIZE, BIN_TYPE, PTR_TYPE, false>(cuda_smaller_leaf_splits, num_data_in_smaller_leaf);
+    LaunchConstructHistogramKernelInner2<HIST_TYPE, SHARED_HIST_SIZE, BIN_TYPE, PTR_TYPE, false>(cuda_smaller_leaf_splits, num_data_in_smaller_leaf, num_bits_in_histogram_bins);
   } else {
-    LaunchConstructHistogramKernelInner2<HIST_TYPE, SHARED_HIST_SIZE, BIN_TYPE, PTR_TYPE, true>(cuda_smaller_leaf_splits, num_data_in_smaller_leaf);
+    LaunchConstructHistogramKernelInner2<HIST_TYPE, SHARED_HIST_SIZE, BIN_TYPE, PTR_TYPE, true>(cuda_smaller_leaf_splits, num_data_in_smaller_leaf, num_bits_in_histogram_bins);
   }
 }
 
 template <typename HIST_TYPE, size_t SHARED_HIST_SIZE, typename BIN_TYPE, typename PTR_TYPE, bool USE_GLOBAL_MEM_BUFFER>
 void CUDAHistogramConstructor::LaunchConstructHistogramKernelInner2(
   const CUDALeafSplitsStruct* cuda_smaller_leaf_splits,
-  const data_size_t num_data_in_smaller_leaf) {
+  const data_size_t num_data_in_smaller_leaf,
+  const uint8_t num_bits_in_histogram_bins) {
   int grid_dim_x = 0;
   int grid_dim_y = 0;
   int block_dim_x = 0;
@@ -314,6 +583,97 @@ void CUDAHistogramConstructor::LaunchConstructHistogramKernelInner2(
   CalcConstructHistogramKernelDim(&grid_dim_x, &grid_dim_y, &block_dim_x, &block_dim_y, num_data_in_smaller_leaf);
   dim3 grid_dim(grid_dim_x, grid_dim_y);
   dim3 block_dim(block_dim_x, block_dim_y);
+  if (use_quantized_grad_) {
+    if (USE_GLOBAL_MEM_BUFFER) {
+      if (cuda_row_data_->is_sparse()) {
+        if (num_bits_in_histogram_bins <= 16) {
+          CUDAConstructDiscretizedHistogramSparseKernel_GlobalMemory<BIN_TYPE, PTR_TYPE, SHARED_HIST_SIZE, true><<<grid_dim, block_dim, 0, cuda_stream_>>>(
+            cuda_smaller_leaf_splits,
+            reinterpret_cast<const int32_t*>(cuda_gradients_),
+            cuda_row_data_->GetBin<BIN_TYPE>(),
+            cuda_row_data_->GetRowPtr<PTR_TYPE>(),
+            cuda_row_data_->GetPartitionPtr<PTR_TYPE>(),
+            cuda_row_data_->cuda_partition_hist_offsets(),
+            num_data_,
+            reinterpret_cast<int32_t*>(cuda_hist_buffer_.RawData()));
+        } else {
+          CUDAConstructDiscretizedHistogramSparseKernel_GlobalMemory<BIN_TYPE, PTR_TYPE, SHARED_HIST_SIZE, false><<<grid_dim, block_dim, 0, cuda_stream_>>>(
+            cuda_smaller_leaf_splits,
+            reinterpret_cast<const int32_t*>(cuda_gradients_),
+            cuda_row_data_->GetBin<BIN_TYPE>(),
+            cuda_row_data_->GetRowPtr<PTR_TYPE>(),
+            cuda_row_data_->GetPartitionPtr<PTR_TYPE>(),
+            cuda_row_data_->cuda_partition_hist_offsets(),
+            num_data_,
+            reinterpret_cast<int32_t*>(cuda_hist_buffer_.RawData()));
+        }
+      } else {
+        if (num_bits_in_histogram_bins <= 16) {
+          CUDAConstructDiscretizedHistogramDenseKernel_GlobalMemory<BIN_TYPE, SHARED_HIST_SIZE, true><<<grid_dim, block_dim, 0, cuda_stream_>>>(
+            cuda_smaller_leaf_splits,
+            reinterpret_cast<const int32_t*>(cuda_gradients_),
+            cuda_row_data_->GetBin<BIN_TYPE>(),
+            cuda_row_data_->cuda_column_hist_offsets(),
+            cuda_row_data_->cuda_partition_hist_offsets(),
+            cuda_row_data_->cuda_feature_partition_column_index_offsets(),
+            num_data_,
+            reinterpret_cast<int32_t*>(cuda_hist_buffer_.RawData()));
+        } else {
+          CUDAConstructDiscretizedHistogramDenseKernel_GlobalMemory<BIN_TYPE, SHARED_HIST_SIZE, false><<<grid_dim, block_dim, 0, cuda_stream_>>>(
+            cuda_smaller_leaf_splits,
+            reinterpret_cast<const int32_t*>(cuda_gradients_),
+            cuda_row_data_->GetBin<BIN_TYPE>(),
+            cuda_row_data_->cuda_column_hist_offsets(),
+            cuda_row_data_->cuda_partition_hist_offsets(),
+            cuda_row_data_->cuda_feature_partition_column_index_offsets(),
+            num_data_,
+            reinterpret_cast<int32_t*>(cuda_hist_buffer_.RawData()));
+        }
+      }
+    } else {
+      if (cuda_row_data_->is_sparse()) {
+        if (num_bits_in_histogram_bins <= 16) {
+          CUDAConstructDiscretizedHistogramSparseKernel<BIN_TYPE, PTR_TYPE, SHARED_HIST_SIZE, true><<<grid_dim, block_dim, 0, cuda_stream_>>>(
+            cuda_smaller_leaf_splits,
+            reinterpret_cast<const int32_t*>(cuda_gradients_),
+            cuda_row_data_->GetBin<BIN_TYPE>(),
+            cuda_row_data_->GetRowPtr<PTR_TYPE>(),
+            cuda_row_data_->GetPartitionPtr<PTR_TYPE>(),
+            cuda_row_data_->cuda_partition_hist_offsets(),
+            num_data_);
+        } else {
+          CUDAConstructDiscretizedHistogramSparseKernel<BIN_TYPE, PTR_TYPE, SHARED_HIST_SIZE, false><<<grid_dim, block_dim, 0, cuda_stream_>>>(
+            cuda_smaller_leaf_splits,
+            reinterpret_cast<const int32_t*>(cuda_gradients_),
+            cuda_row_data_->GetBin<BIN_TYPE>(),
+            cuda_row_data_->GetRowPtr<PTR_TYPE>(),
+            cuda_row_data_->GetPartitionPtr<PTR_TYPE>(),
+            cuda_row_data_->cuda_partition_hist_offsets(),
+            num_data_);
+        }
+      } else {
+        if (num_bits_in_histogram_bins <= 16) {
+          CUDAConstructDiscretizedHistogramDenseKernel<BIN_TYPE, SHARED_HIST_SIZE, true><<<grid_dim, block_dim, 0, cuda_stream_>>>(
+            cuda_smaller_leaf_splits,
+            reinterpret_cast<const int32_t*>(cuda_gradients_),
+            cuda_row_data_->GetBin<BIN_TYPE>(),
+            cuda_row_data_->cuda_column_hist_offsets(),
+            cuda_row_data_->cuda_partition_hist_offsets(),
+            cuda_row_data_->cuda_feature_partition_column_index_offsets(),
+            num_data_);
+        } else {
+          CUDAConstructDiscretizedHistogramDenseKernel<BIN_TYPE, SHARED_HIST_SIZE, false><<<grid_dim, block_dim, 0, cuda_stream_>>>(
+            cuda_smaller_leaf_splits,
+            reinterpret_cast<const int32_t*>(cuda_gradients_),
+            cuda_row_data_->GetBin<BIN_TYPE>(),
+            cuda_row_data_->cuda_column_hist_offsets(),
+            cuda_row_data_->cuda_partition_hist_offsets(),
+            cuda_row_data_->cuda_feature_partition_column_index_offsets(),
+            num_data_);
+        }
+      }
+    }
+  } else {
     if (!USE_GLOBAL_MEM_BUFFER) {
       if (cuda_row_data_->is_sparse()) {
         CUDAConstructHistogramSparseKernel<BIN_TYPE, PTR_TYPE, HIST_TYPE, SHARED_HIST_SIZE><<<grid_dim, block_dim, 0, cuda_stream_>>>(
@@ -336,7 +696,7 @@ void CUDAHistogramConstructor::LaunchConstructHistogramKernelInner2(
       }
     } else {
       if (cuda_row_data_->is_sparse()) {
-      CUDAConstructHistogramSparseKernel_GlobalMemory<BIN_TYPE, PTR_TYPE><<<grid_dim, block_dim, 0, cuda_stream_>>>(
+        CUDAConstructHistogramSparseKernel_GlobalMemory<BIN_TYPE, HIST_TYPE, PTR_TYPE><<<grid_dim, block_dim, 0, cuda_stream_>>>(
           cuda_smaller_leaf_splits,
           cuda_gradients_, cuda_hessians_,
           cuda_row_data_->GetBin<BIN_TYPE>(),
@@ -344,9 +704,9 @@ void CUDAHistogramConstructor::LaunchConstructHistogramKernelInner2(
           cuda_row_data_->GetPartitionPtr<PTR_TYPE>(),
           cuda_row_data_->cuda_partition_hist_offsets(),
           num_data_,
-        cuda_hist_buffer_);
+          reinterpret_cast<HIST_TYPE*>(cuda_hist_buffer_.RawData()));
       } else {
-      CUDAConstructHistogramDenseKernel_GlobalMemory<BIN_TYPE><<<grid_dim, block_dim, 0, cuda_stream_>>>(
+        CUDAConstructHistogramDenseKernel_GlobalMemory<BIN_TYPE, HIST_TYPE><<<grid_dim, block_dim, 0, cuda_stream_>>>(
           cuda_smaller_leaf_splits,
           cuda_gradients_, cuda_hessians_,
           cuda_row_data_->GetBin<BIN_TYPE>(),
@@ -354,7 +714,8 @@ void CUDAHistogramConstructor::LaunchConstructHistogramKernelInner2(
           cuda_row_data_->cuda_partition_hist_offsets(),
           cuda_row_data_->cuda_feature_partition_column_index_offsets(),
           num_data_,
-        cuda_hist_buffer_);
+          reinterpret_cast<HIST_TYPE*>(cuda_hist_buffer_.RawData()));
+      }
     }
   }
 }
@@ -403,19 +764,127 @@ __global__ void FixHistogramKernel(
   }
 }
 
+template <bool SMALLER_USE_16BIT_HIST, bool LARGER_USE_16BIT_HIST, bool PARENT_USE_16BIT_HIST>
+__global__ void SubtractHistogramDiscretizedKernel(
+  const int num_total_bin,
+  const CUDALeafSplitsStruct* cuda_smaller_leaf_splits,
+  const CUDALeafSplitsStruct* cuda_larger_leaf_splits,
+  hist_t* num_bit_change_buffer) {
+  const unsigned int global_thread_index = threadIdx.x + blockIdx.x * blockDim.x;
+  const int cuda_larger_leaf_index_ref = cuda_larger_leaf_splits->leaf_index;
+  if (cuda_larger_leaf_index_ref >= 0) {
+    if (PARENT_USE_16BIT_HIST) {
+      const int32_t* smaller_leaf_hist = reinterpret_cast<const int32_t*>(cuda_smaller_leaf_splits->hist_in_leaf);
+      int32_t* larger_leaf_hist = reinterpret_cast<int32_t*>(cuda_larger_leaf_splits->hist_in_leaf);
+      if (global_thread_index < num_total_bin) {
+        larger_leaf_hist[global_thread_index] -= smaller_leaf_hist[global_thread_index];
+      }
+    } else if (LARGER_USE_16BIT_HIST) {
+      int32_t* buffer = reinterpret_cast<int32_t*>(num_bit_change_buffer);
+      const int32_t* smaller_leaf_hist = reinterpret_cast<const int32_t*>(cuda_smaller_leaf_splits->hist_in_leaf);
+      int64_t* larger_leaf_hist = reinterpret_cast<int64_t*>(cuda_larger_leaf_splits->hist_in_leaf);
+      if (global_thread_index < num_total_bin) {
+        const int64_t parent_hist_item = larger_leaf_hist[global_thread_index];
+        const int32_t smaller_hist_item = smaller_leaf_hist[global_thread_index];
+        const int64_t smaller_hist_item_int64 = (static_cast<int64_t>(static_cast<int16_t>(smaller_hist_item >> 16)) << 32) |
+          static_cast<int64_t>(smaller_hist_item & 0x0000ffff);
+        const int64_t larger_hist_item = parent_hist_item - smaller_hist_item_int64;
+        buffer[global_thread_index] = static_cast<int32_t>(static_cast<int16_t>(larger_hist_item >> 32) << 16) |
+          static_cast<int32_t>(larger_hist_item & 0x000000000000ffff);
+      }
+    } else if (SMALLER_USE_16BIT_HIST) {
+        const int32_t* smaller_leaf_hist = reinterpret_cast<const int32_t*>(cuda_smaller_leaf_splits->hist_in_leaf);
+        int64_t* larger_leaf_hist = reinterpret_cast<int64_t*>(cuda_larger_leaf_splits->hist_in_leaf);
+        if (global_thread_index < num_total_bin) {
+          const int64_t parent_hist_item = larger_leaf_hist[global_thread_index];
+          const int32_t smaller_hist_item = smaller_leaf_hist[global_thread_index];
+          const int64_t smaller_hist_item_int64 = (static_cast<int64_t>(static_cast<int16_t>(smaller_hist_item >> 16)) << 32) |
+            static_cast<int64_t>(smaller_hist_item & 0x0000ffff);
+          const int64_t larger_hist_item = parent_hist_item - smaller_hist_item_int64;
+          larger_leaf_hist[global_thread_index] = larger_hist_item;
+        }
+    } else {
+      const int64_t* smaller_leaf_hist = reinterpret_cast<const int64_t*>(cuda_smaller_leaf_splits->hist_in_leaf);
+      int64_t* larger_leaf_hist = reinterpret_cast<int64_t*>(cuda_larger_leaf_splits->hist_in_leaf);
+      if (global_thread_index < num_total_bin) {
+        larger_leaf_hist[global_thread_index] -= smaller_leaf_hist[global_thread_index];
+      }
+    }
+  }
+}
+
+__global__ void CopyChangedNumBitHistogram(
+  const int num_total_bin,
+  const CUDALeafSplitsStruct* cuda_larger_leaf_splits,
+  hist_t* num_bit_change_buffer) {
+  int32_t* hist_dst = reinterpret_cast<int32_t*>(cuda_larger_leaf_splits->hist_in_leaf);
+  const int32_t* hist_src = reinterpret_cast<const int32_t*>(num_bit_change_buffer);
+  const unsigned int global_thread_index = threadIdx.x + blockIdx.x * blockDim.x;
+  if (global_thread_index < static_cast<unsigned int>(num_total_bin)) {
+    hist_dst[global_thread_index] = hist_src[global_thread_index];
+  }
+}
+
+template <bool USE_16BIT_HIST>
+__global__ void FixHistogramDiscretizedKernel(
+  const uint32_t* cuda_feature_num_bins,
+  const uint32_t* cuda_feature_hist_offsets,
+  const uint32_t* cuda_feature_most_freq_bins,
+  const int* cuda_need_fix_histogram_features,
+  const uint32_t* cuda_need_fix_histogram_features_num_bin_aligned,
+  const CUDALeafSplitsStruct* cuda_smaller_leaf_splits) {
+  __shared__ int64_t shared_mem_buffer[32];
+  const unsigned int blockIdx_x = blockIdx.x;
+  const int feature_index = cuda_need_fix_histogram_features[blockIdx_x];
+  const uint32_t num_bin_aligned = cuda_need_fix_histogram_features_num_bin_aligned[blockIdx_x];
+  const uint32_t feature_hist_offset = cuda_feature_hist_offsets[feature_index];
+  const uint32_t most_freq_bin = cuda_feature_most_freq_bins[feature_index];
+  if (USE_16BIT_HIST) {
+    const int64_t leaf_sum_gradients_hessians_int64 = cuda_smaller_leaf_splits->sum_of_gradients_hessians;
+    const int32_t leaf_sum_gradients_hessians =
+      (static_cast<int32_t>(leaf_sum_gradients_hessians_int64 >> 32) << 16) | static_cast<int32_t>(leaf_sum_gradients_hessians_int64 & 0x000000000000ffff);
+    int32_t* feature_hist = reinterpret_cast<int32_t*>(cuda_smaller_leaf_splits->hist_in_leaf) + feature_hist_offset;
+    const unsigned int threadIdx_x = threadIdx.x;
+    const uint32_t num_bin = cuda_feature_num_bins[feature_index];
+    const int32_t bin_gradient_hessian = (threadIdx_x < num_bin && threadIdx_x != most_freq_bin) ? feature_hist[threadIdx_x] : 0;
+    const int32_t sum_gradient_hessian = ShuffleReduceSum<int32_t>(
+      bin_gradient_hessian,
+      reinterpret_cast<int32_t*>(shared_mem_buffer),
+      num_bin_aligned);
+    if (threadIdx_x == 0) {
+      feature_hist[most_freq_bin] = leaf_sum_gradients_hessians - sum_gradient_hessian;
+    }
+  } else {
+    const int64_t leaf_sum_gradients_hessians = cuda_smaller_leaf_splits->sum_of_gradients_hessians;
+    int64_t* feature_hist = reinterpret_cast<int64_t*>(cuda_smaller_leaf_splits->hist_in_leaf) + feature_hist_offset;
+    const unsigned int threadIdx_x = threadIdx.x;
+    const uint32_t num_bin = cuda_feature_num_bins[feature_index];
+    const int64_t bin_gradient_hessian = (threadIdx_x < num_bin && threadIdx_x != most_freq_bin) ? feature_hist[threadIdx_x] : 0;
+    const int64_t sum_gradient_hessian = ShuffleReduceSum<int64_t>(bin_gradient_hessian, shared_mem_buffer, num_bin_aligned);
+    if (threadIdx_x == 0) {
+      feature_hist[most_freq_bin] = leaf_sum_gradients_hessians - sum_gradient_hessian;
+    }
+  }
+}
+
 void CUDAHistogramConstructor::LaunchSubtractHistogramKernel(
   const CUDALeafSplitsStruct* cuda_smaller_leaf_splits,
-  const CUDALeafSplitsStruct* cuda_larger_leaf_splits) {
+  const CUDALeafSplitsStruct* cuda_larger_leaf_splits,
+  const bool use_discretized_grad,
+  const uint8_t parent_num_bits_in_histogram_bins,
+  const uint8_t smaller_num_bits_in_histogram_bins,
+  const uint8_t larger_num_bits_in_histogram_bins) {
+    if (!use_discretized_grad) {
       const int num_subtract_threads = 2 * num_total_bin_;
       const int num_subtract_blocks = (num_subtract_threads + SUBTRACT_BLOCK_SIZE - 1) / SUBTRACT_BLOCK_SIZE;
       global_timer.Start("CUDAHistogramConstructor::FixHistogramKernel");
       if (need_fix_histogram_features_.size() > 0) {
         FixHistogramKernel<<<need_fix_histogram_features_.size(), FIX_HISTOGRAM_BLOCK_SIZE, 0, cuda_stream_>>>(
-      cuda_feature_num_bins_,
-      cuda_feature_hist_offsets_,
-      cuda_feature_most_freq_bins_,
-      cuda_need_fix_histogram_features_,
-      cuda_need_fix_histogram_features_num_bin_aligned_,
+          cuda_feature_num_bins_.RawData(),
+          cuda_feature_hist_offsets_.RawData(),
+          cuda_feature_most_freq_bins_.RawData(),
+          cuda_need_fix_histogram_features_.RawData(),
+          cuda_need_fix_histogram_features_num_bin_aligned_.RawData(),
           cuda_smaller_leaf_splits);
       }
       global_timer.Stop("CUDAHistogramConstructor::FixHistogramKernel");
@@ -425,6 +894,65 @@ void CUDAHistogramConstructor::LaunchSubtractHistogramKernel(
         cuda_smaller_leaf_splits,
         cuda_larger_leaf_splits);
       global_timer.Stop("CUDAHistogramConstructor::SubtractHistogramKernel");
+    } else {
+      const int num_subtract_threads = num_total_bin_;
+      const int num_subtract_blocks = (num_subtract_threads + SUBTRACT_BLOCK_SIZE - 1) / SUBTRACT_BLOCK_SIZE;
+      global_timer.Start("CUDAHistogramConstructor::FixHistogramDiscretizedKernel");
+      if (need_fix_histogram_features_.size() > 0) {
+        if (smaller_num_bits_in_histogram_bins <= 16) {
+          FixHistogramDiscretizedKernel<true><<<need_fix_histogram_features_.size(), FIX_HISTOGRAM_BLOCK_SIZE, 0, cuda_stream_>>>(
+            cuda_feature_num_bins_.RawData(),
+            cuda_feature_hist_offsets_.RawData(),
+            cuda_feature_most_freq_bins_.RawData(),
+            cuda_need_fix_histogram_features_.RawData(),
+            cuda_need_fix_histogram_features_num_bin_aligned_.RawData(),
+            cuda_smaller_leaf_splits);
+        } else {
+          FixHistogramDiscretizedKernel<false><<<need_fix_histogram_features_.size(), FIX_HISTOGRAM_BLOCK_SIZE, 0, cuda_stream_>>>(
+            cuda_feature_num_bins_.RawData(),
+            cuda_feature_hist_offsets_.RawData(),
+            cuda_feature_most_freq_bins_.RawData(),
+            cuda_need_fix_histogram_features_.RawData(),
+            cuda_need_fix_histogram_features_num_bin_aligned_.RawData(),
+            cuda_smaller_leaf_splits);
+        }
+      }
+      global_timer.Stop("CUDAHistogramConstructor::FixHistogramDiscretizedKernel");
+      global_timer.Start("CUDAHistogramConstructor::SubtractHistogramDiscretizedKernel");
+      if (parent_num_bits_in_histogram_bins <= 16) {
+        CHECK_LE(smaller_num_bits_in_histogram_bins, 16);
+        CHECK_LE(larger_num_bits_in_histogram_bins, 16);
+        SubtractHistogramDiscretizedKernel<true, true, true><<<num_subtract_blocks, SUBTRACT_BLOCK_SIZE, 0, cuda_stream_>>>(
+          num_total_bin_,
+          cuda_smaller_leaf_splits,
+          cuda_larger_leaf_splits,
+          hist_buffer_for_num_bit_change_.RawData());
+      } else if (larger_num_bits_in_histogram_bins <= 16) {
+        CHECK_LE(smaller_num_bits_in_histogram_bins, 16);
+        SubtractHistogramDiscretizedKernel<true, true, false><<<num_subtract_blocks, SUBTRACT_BLOCK_SIZE, 0, cuda_stream_>>>(
+          num_total_bin_,
+          cuda_smaller_leaf_splits,
+          cuda_larger_leaf_splits,
+          hist_buffer_for_num_bit_change_.RawData());
+        CopyChangedNumBitHistogram<<<num_subtract_blocks, SUBTRACT_BLOCK_SIZE, 0, cuda_stream_>>>(
+          num_total_bin_,
+          cuda_larger_leaf_splits,
+          hist_buffer_for_num_bit_change_.RawData());
+      } else if (smaller_num_bits_in_histogram_bins <= 16) {
+        SubtractHistogramDiscretizedKernel<true, false, false><<<num_subtract_blocks, SUBTRACT_BLOCK_SIZE, 0, cuda_stream_>>>(
+          num_total_bin_,
+          cuda_smaller_leaf_splits,
+          cuda_larger_leaf_splits,
+          hist_buffer_for_num_bit_change_.RawData());
+      } else {
+        SubtractHistogramDiscretizedKernel<false, false, false><<<num_subtract_blocks, SUBTRACT_BLOCK_SIZE, 0, cuda_stream_>>>(
+          num_total_bin_,
+          cuda_smaller_leaf_splits,
+          cuda_larger_leaf_splits,
+          hist_buffer_for_num_bit_change_.RawData());
+      }
+      global_timer.Stop("CUDAHistogramConstructor::SubtractHistogramDiscretizedKernel");
+    }
 }
 
 }  // namespace LightGBM

src/treelearner/cuda/cuda_histogram_constructor.hpp

@@ -9,6 +9,7 @@
 #ifdef USE_CUDA
 
 #include <LightGBM/cuda/cuda_row_data.hpp>
+#include <LightGBM/cuda/cuda_utils.hu>
 #include <LightGBM/feature_group.h>
 #include <LightGBM/tree.h>
 
@@ -37,7 +38,9 @@ class CUDAHistogramConstructor {
     const int min_data_in_leaf,
     const double min_sum_hessian_in_leaf,
     const int gpu_device_id,
-    const bool gpu_use_dp);
+    const bool gpu_use_dp,
+    const bool use_discretized_grad,
+    const int grad_discretized_bins);
 
   ~CUDAHistogramConstructor();
 
@@ -49,7 +52,16 @@ class CUDAHistogramConstructor {
     const data_size_t num_data_in_smaller_leaf,
     const data_size_t num_data_in_larger_leaf,
     const double sum_hessians_in_smaller_leaf,
-    const double sum_hessians_in_larger_leaf);
+    const double sum_hessians_in_larger_leaf,
+    const uint8_t num_bits_in_histogram_bins);
+
+  void SubtractHistogramForLeaf(
+    const CUDALeafSplitsStruct* cuda_smaller_leaf_splits,
+    const CUDALeafSplitsStruct* cuda_larger_leaf_splits,
+    const bool use_discretized_grad,
+    const uint8_t parent_num_bits_in_histogram_bins,
+    const uint8_t smaller_num_bits_in_histogram_bins,
+    const uint8_t larger_num_bits_in_histogram_bins);
 
   void ResetTrainingData(const Dataset* train_data, TrainingShareStates* share_states);
 
@@ -57,9 +69,9 @@ class CUDAHistogramConstructor {
 
   void BeforeTrain(const score_t* gradients, const score_t* hessians);
 
-  const hist_t* cuda_hist() const { return cuda_hist_; }
+  const hist_t* cuda_hist() const { return cuda_hist_.RawData(); }
 
-  hist_t* cuda_hist_pointer() { return cuda_hist_; }
+  hist_t* cuda_hist_pointer() { return cuda_hist_.RawData(); }
 
  private:
   void InitFeatureMetaInfo(const Dataset* train_data, const std::vector<uint32_t>& feature_hist_offsets);
@@ -74,30 +86,39 @@ class CUDAHistogramConstructor {
   template <typename HIST_TYPE, size_t SHARED_HIST_SIZE>
   void LaunchConstructHistogramKernelInner(
     const CUDALeafSplitsStruct* cuda_smaller_leaf_splits,
-    const data_size_t num_data_in_smaller_leaf);
+    const data_size_t num_data_in_smaller_leaf,
+    const uint8_t num_bits_in_histogram_bins);
 
   template <typename HIST_TYPE, size_t SHARED_HIST_SIZE, typename BIN_TYPE>
   void LaunchConstructHistogramKernelInner0(
     const CUDALeafSplitsStruct* cuda_smaller_leaf_splits,
-    const data_size_t num_data_in_smaller_leaf);
+    const data_size_t num_data_in_smaller_leaf,
+    const uint8_t num_bits_in_histogram_bins);
 
   template <typename HIST_TYPE, size_t SHARED_HIST_SIZE, typename BIN_TYPE, typename PTR_TYPE>
   void LaunchConstructHistogramKernelInner1(
     const CUDALeafSplitsStruct* cuda_smaller_leaf_splits,
-    const data_size_t num_data_in_smaller_leaf);
+    const data_size_t num_data_in_smaller_leaf,
+    const uint8_t num_bits_in_histogram_bins);
 
   template <typename HIST_TYPE, size_t SHARED_HIST_SIZE, typename BIN_TYPE, typename PTR_TYPE, bool USE_GLOBAL_MEM_BUFFER>
   void LaunchConstructHistogramKernelInner2(
     const CUDALeafSplitsStruct* cuda_smaller_leaf_splits,
-    const data_size_t num_data_in_smaller_leaf);
+    const data_size_t num_data_in_smaller_leaf,
+    const uint8_t num_bits_in_histogram_bins);
 
   void LaunchConstructHistogramKernel(
     const CUDALeafSplitsStruct* cuda_smaller_leaf_splits,
-    const data_size_t num_data_in_smaller_leaf);
+    const data_size_t num_data_in_smaller_leaf,
+    const uint8_t num_bits_in_histogram_bins);
 
   void LaunchSubtractHistogramKernel(
     const CUDALeafSplitsStruct* cuda_smaller_leaf_splits,
-    const CUDALeafSplitsStruct* cuda_larger_leaf_splits);
+    const CUDALeafSplitsStruct* cuda_larger_leaf_splits,
+    const bool use_discretized_grad,
+    const uint8_t parent_num_bits_in_histogram_bins,
+    const uint8_t smaller_num_bits_in_histogram_bins,
+    const uint8_t larger_num_bits_in_histogram_bins);
 
   // Host memory
 
@@ -136,19 +157,21 @@ class CUDAHistogramConstructor {
   /*! \brief CUDA row wise data */
   std::unique_ptr<CUDARowData> cuda_row_data_;
   /*! \brief number of bins per feature */
-  uint32_t* cuda_feature_num_bins_;
+  CUDAVector<uint32_t> cuda_feature_num_bins_;
   /*! \brief offsets in histogram of all features */
-  uint32_t* cuda_feature_hist_offsets_;
+  CUDAVector<uint32_t> cuda_feature_hist_offsets_;
   /*! \brief most frequent bins in each feature */
-  uint32_t* cuda_feature_most_freq_bins_;
+  CUDAVector<uint32_t> cuda_feature_most_freq_bins_;
   /*! \brief CUDA histograms */
-  hist_t* cuda_hist_;
+  CUDAVector<hist_t> cuda_hist_;
   /*! \brief CUDA histograms buffer for each block */
-  float* cuda_hist_buffer_;
+  CUDAVector<float> cuda_hist_buffer_;
   /*! \brief indices of feature whose histograms need to be fixed */
-  int* cuda_need_fix_histogram_features_;
+  CUDAVector<int> cuda_need_fix_histogram_features_;
   /*! \brief aligned number of bins of the features whose histograms need to be fixed */
-  uint32_t* cuda_need_fix_histogram_features_num_bin_aligned_;
+  CUDAVector<uint32_t> cuda_need_fix_histogram_features_num_bin_aligned_;
+  /*! \brief histogram buffer used in histogram subtraction with different number of bits for histogram bins */
+  CUDAVector<hist_t> hist_buffer_for_num_bit_change_;
 
   // CUDA memory, held by other object
 
@@ -161,6 +184,10 @@ class CUDAHistogramConstructor {
   const int gpu_device_id_;
   /*! \brief use double precision histogram per block */
   const bool gpu_use_dp_;
+  /*! \brief whether to use quantized gradients */
+  const bool use_quantized_grad_;
+  /*! \brief the number of bins to quantized gradients */
+  const int num_grad_quant_bins_;
 };
 
 }  // namespace LightGBM

src/treelearner/cuda/cuda_leaf_splits.cpp

@@ -11,27 +11,22 @@
 namespace LightGBM {
 
 CUDALeafSplits::CUDALeafSplits(const data_size_t num_data):
-num_data_(num_data) {
-  cuda_struct_ = nullptr;
-  cuda_sum_of_gradients_buffer_ = nullptr;
-  cuda_sum_of_hessians_buffer_ = nullptr;
-}
+num_data_(num_data) {}
 
-CUDALeafSplits::~CUDALeafSplits() {
-  DeallocateCUDAMemory<CUDALeafSplitsStruct>(&cuda_struct_, __FILE__, __LINE__);
-  DeallocateCUDAMemory<double>(&cuda_sum_of_gradients_buffer_, __FILE__, __LINE__);
-  DeallocateCUDAMemory<double>(&cuda_sum_of_hessians_buffer_, __FILE__, __LINE__);
-}
+CUDALeafSplits::~CUDALeafSplits() {}
 
-void CUDALeafSplits::Init() {
+void CUDALeafSplits::Init(const bool use_quantized_grad) {
   num_blocks_init_from_gradients_ = (num_data_ + NUM_THRADS_PER_BLOCK_LEAF_SPLITS - 1) / NUM_THRADS_PER_BLOCK_LEAF_SPLITS;
 
   // allocate more memory for sum reduction in CUDA
   // only the first element records the final sum
-  AllocateCUDAMemory<double>(&cuda_sum_of_gradients_buffer_, num_blocks_init_from_gradients_, __FILE__, __LINE__);
-  AllocateCUDAMemory<double>(&cuda_sum_of_hessians_buffer_, num_blocks_init_from_gradients_, __FILE__, __LINE__);
+  cuda_sum_of_gradients_buffer_.Resize(static_cast<size_t>(num_blocks_init_from_gradients_));
+  cuda_sum_of_hessians_buffer_.Resize(static_cast<size_t>(num_blocks_init_from_gradients_));
+  if (use_quantized_grad) {
+    cuda_sum_of_gradients_hessians_buffer_.Resize(static_cast<size_t>(num_blocks_init_from_gradients_));
+  }
 
-  AllocateCUDAMemory<CUDALeafSplitsStruct>(&cuda_struct_, 1, __FILE__, __LINE__);
+  cuda_struct_.Resize(1);
 }
 
 void CUDALeafSplits::InitValues() {
@@ -46,24 +41,33 @@ void CUDALeafSplits::InitValues(
   const data_size_t num_used_indices, hist_t* cuda_hist_in_leaf, double* root_sum_hessians) {
   cuda_gradients_ = cuda_gradients;
   cuda_hessians_ = cuda_hessians;
-  SetCUDAMemory<double>(cuda_sum_of_gradients_buffer_, 0, num_blocks_init_from_gradients_, __FILE__, __LINE__);
-  SetCUDAMemory<double>(cuda_sum_of_hessians_buffer_, 0, num_blocks_init_from_gradients_, __FILE__, __LINE__);
+  cuda_sum_of_gradients_buffer_.SetValue(0);
+  cuda_sum_of_hessians_buffer_.SetValue(0);
   LaunchInitValuesKernal(lambda_l1, lambda_l2, cuda_bagging_data_indices, cuda_data_indices_in_leaf, num_used_indices, cuda_hist_in_leaf);
-  CopyFromCUDADeviceToHost<double>(root_sum_hessians, cuda_sum_of_hessians_buffer_, 1, __FILE__, __LINE__);
+  CopyFromCUDADeviceToHost<double>(root_sum_hessians, cuda_sum_of_hessians_buffer_.RawData(), 1, __FILE__, __LINE__);
+  SynchronizeCUDADevice(__FILE__, __LINE__);
+}
+
+void CUDALeafSplits::InitValues(
+  const double lambda_l1, const double lambda_l2,
+  const int16_t* cuda_gradients_and_hessians,
+  const data_size_t* cuda_bagging_data_indices,
+  const data_size_t* cuda_data_indices_in_leaf, const data_size_t num_used_indices,
+  hist_t* cuda_hist_in_leaf, double* root_sum_hessians,
+  const score_t* grad_scale, const score_t* hess_scale) {
+  cuda_gradients_ = reinterpret_cast<const score_t*>(cuda_gradients_and_hessians);
+  cuda_hessians_ = nullptr;
+  LaunchInitValuesKernal(lambda_l1, lambda_l2, cuda_bagging_data_indices, cuda_data_indices_in_leaf, num_used_indices, cuda_hist_in_leaf, grad_scale, hess_scale);
+  CopyFromCUDADeviceToHost<double>(root_sum_hessians, cuda_sum_of_hessians_buffer_.RawData(), 1, __FILE__, __LINE__);
   SynchronizeCUDADevice(__FILE__, __LINE__);
 }
 
 void CUDALeafSplits::Resize(const data_size_t num_data) {
-  if (num_data > num_data_) {
-    DeallocateCUDAMemory<double>(&cuda_sum_of_gradients_buffer_, __FILE__, __LINE__);
-    DeallocateCUDAMemory<double>(&cuda_sum_of_hessians_buffer_, __FILE__, __LINE__);
-    num_blocks_init_from_gradients_ = (num_data + NUM_THRADS_PER_BLOCK_LEAF_SPLITS - 1) / NUM_THRADS_PER_BLOCK_LEAF_SPLITS;
-    AllocateCUDAMemory<double>(&cuda_sum_of_gradients_buffer_, num_blocks_init_from_gradients_, __FILE__, __LINE__);
-    AllocateCUDAMemory<double>(&cuda_sum_of_hessians_buffer_, num_blocks_init_from_gradients_, __FILE__, __LINE__);
-  } else {
-    num_blocks_init_from_gradients_ = (num_data + NUM_THRADS_PER_BLOCK_LEAF_SPLITS - 1) / NUM_THRADS_PER_BLOCK_LEAF_SPLITS;
-  }
   num_data_ = num_data;
+  num_blocks_init_from_gradients_ = (num_data + NUM_THRADS_PER_BLOCK_LEAF_SPLITS - 1) / NUM_THRADS_PER_BLOCK_LEAF_SPLITS;
+  cuda_sum_of_gradients_buffer_.Resize(static_cast<size_t>(num_blocks_init_from_gradients_));
+  cuda_sum_of_hessians_buffer_.Resize(static_cast<size_t>(num_blocks_init_from_gradients_));
+  cuda_sum_of_gradients_hessians_buffer_.Resize(static_cast<size_t>(num_blocks_init_from_gradients_));
 }
 
 }  // namespace LightGBM

src/treelearner/cuda/cuda_leaf_splits.cu

@@ -81,6 +81,90 @@ __global__ void CUDAInitValuesKernel2(
   }
 }
 
+template <bool USE_INDICES>
+__global__ void CUDAInitValuesKernel3(const int16_t* cuda_gradients_and_hessians,
+  const data_size_t num_data, const data_size_t* cuda_bagging_data_indices,
+  double* cuda_sum_of_gradients, double* cuda_sum_of_hessians, int64_t* cuda_sum_of_hessians_hessians,
+  const score_t* grad_scale_pointer, const score_t* hess_scale_pointer) {
+  const score_t grad_scale = *grad_scale_pointer;
+  const score_t hess_scale = *hess_scale_pointer;
+  __shared__ int64_t shared_mem_buffer[32];
+  const data_size_t data_index = static_cast<data_size_t>(threadIdx.x + blockIdx.x * blockDim.x);
+  int64_t int_gradient = 0;
+  int64_t int_hessian = 0;
+  if (data_index < num_data) {
+    int_gradient = USE_INDICES ? cuda_gradients_and_hessians[2 * cuda_bagging_data_indices[data_index] + 1] :
+      cuda_gradients_and_hessians[2 * data_index + 1];
+    int_hessian = USE_INDICES ? cuda_gradients_and_hessians[2 * cuda_bagging_data_indices[data_index]] :
+      cuda_gradients_and_hessians[2 * data_index];
+  }
+  const int64_t block_sum_gradient = ShuffleReduceSum<int64_t>(int_gradient, shared_mem_buffer, blockDim.x);
+  __syncthreads();
+  const int64_t block_sum_hessian = ShuffleReduceSum<int64_t>(int_hessian, shared_mem_buffer, blockDim.x);
+  if (threadIdx.x == 0) {
+    cuda_sum_of_gradients[blockIdx.x] = block_sum_gradient * grad_scale;
+    cuda_sum_of_hessians[blockIdx.x] = block_sum_hessian * hess_scale;
+    cuda_sum_of_hessians_hessians[blockIdx.x] = ((block_sum_gradient << 32) | block_sum_hessian);
+  }
+}
+
+__global__ void CUDAInitValuesKernel4(
+  const double lambda_l1,
+  const double lambda_l2,
+  const int num_blocks_to_reduce,
+  double* cuda_sum_of_gradients,
+  double* cuda_sum_of_hessians,
+  int64_t* cuda_sum_of_gradients_hessians,
+  const data_size_t num_data,
+  const data_size_t* cuda_data_indices_in_leaf,
+  hist_t* cuda_hist_in_leaf,
+  CUDALeafSplitsStruct* cuda_struct) {
+  __shared__ double shared_mem_buffer[32];
+  double thread_sum_of_gradients = 0.0f;
+  double thread_sum_of_hessians = 0.0f;
+  int64_t thread_sum_of_gradients_hessians = 0;
+  for (int block_index = static_cast<int>(threadIdx.x); block_index < num_blocks_to_reduce; block_index += static_cast<int>(blockDim.x)) {
+    thread_sum_of_gradients += cuda_sum_of_gradients[block_index];
+    thread_sum_of_hessians += cuda_sum_of_hessians[block_index];
+    thread_sum_of_gradients_hessians += cuda_sum_of_gradients_hessians[block_index];
+  }
+  const double sum_of_gradients = ShuffleReduceSum<double>(thread_sum_of_gradients, shared_mem_buffer, blockDim.x);
+  __syncthreads();
+  const double sum_of_hessians = ShuffleReduceSum<double>(thread_sum_of_hessians, shared_mem_buffer, blockDim.x);
+  __syncthreads();
+  const double sum_of_gradients_hessians = ShuffleReduceSum<int64_t>(
+    thread_sum_of_gradients_hessians,
+    reinterpret_cast<int64_t*>(shared_mem_buffer),
+    blockDim.x);
+  if (threadIdx.x == 0) {
+    cuda_sum_of_hessians[0] = sum_of_hessians;
+    cuda_struct->leaf_index = 0;
+    cuda_struct->sum_of_gradients = sum_of_gradients;
+    cuda_struct->sum_of_hessians = sum_of_hessians;
+    cuda_struct->sum_of_gradients_hessians = sum_of_gradients_hessians;
+    cuda_struct->num_data_in_leaf = num_data;
+    const bool use_l1 = lambda_l1 > 0.0f;
+    if (!use_l1) {
+      // no smoothing on root node
+      cuda_struct->gain = CUDALeafSplits::GetLeafGain<false, false>(sum_of_gradients, sum_of_hessians, lambda_l1, lambda_l2, 0.0f, 0, 0.0f);
+    } else {
+      // no smoothing on root node
+      cuda_struct->gain = CUDALeafSplits::GetLeafGain<true, false>(sum_of_gradients, sum_of_hessians, lambda_l1, lambda_l2, 0.0f, 0, 0.0f);
+    }
+    if (!use_l1) {
+      // no smoothing on root node
+      cuda_struct->leaf_value =
+        CUDALeafSplits::CalculateSplittedLeafOutput<false, false>(sum_of_gradients, sum_of_hessians, lambda_l1, lambda_l2, 0.0f, 0, 0.0f);
+    } else {
+      // no smoothing on root node
+      cuda_struct->leaf_value =
+        CUDALeafSplits::CalculateSplittedLeafOutput<true, false>(sum_of_gradients, sum_of_hessians, lambda_l1, lambda_l2, 0.0f, 0, 0.0f);
+    }
+    cuda_struct->data_indices_in_leaf = cuda_data_indices_in_leaf;
+    cuda_struct->hist_in_leaf = cuda_hist_in_leaf;
+  }
+}
+
 __global__ void InitValuesEmptyKernel(CUDALeafSplitsStruct* cuda_struct) {
   cuda_struct->leaf_index = -1;
   cuda_struct->sum_of_gradients = 0.0f;
@@ -93,7 +177,7 @@ __global__ void InitValuesEmptyKernel(CUDALeafSplitsStruct* cuda_struct) {
 }
 
 void CUDALeafSplits::LaunchInitValuesEmptyKernel() {
-  InitValuesEmptyKernel<<<1, 1>>>(cuda_struct_);
+  InitValuesEmptyKernel<<<1, 1>>>(cuda_struct_.RawData());
 }
 
 void CUDALeafSplits::LaunchInitValuesKernal(
@@ -104,23 +188,55 @@ void CUDALeafSplits::LaunchInitValuesKernal(
   hist_t* cuda_hist_in_leaf) {
   if (cuda_bagging_data_indices == nullptr) {
     CUDAInitValuesKernel1<false><<<num_blocks_init_from_gradients_, NUM_THRADS_PER_BLOCK_LEAF_SPLITS>>>(
-      cuda_gradients_, cuda_hessians_, num_used_indices, nullptr, cuda_sum_of_gradients_buffer_,
-      cuda_sum_of_hessians_buffer_);
+      cuda_gradients_, cuda_hessians_, num_used_indices, nullptr, cuda_sum_of_gradients_buffer_.RawData(),
+      cuda_sum_of_hessians_buffer_.RawData());
   } else {
     CUDAInitValuesKernel1<true><<<num_blocks_init_from_gradients_, NUM_THRADS_PER_BLOCK_LEAF_SPLITS>>>(
-      cuda_gradients_, cuda_hessians_, num_used_indices, cuda_bagging_data_indices, cuda_sum_of_gradients_buffer_,
-      cuda_sum_of_hessians_buffer_);
+      cuda_gradients_, cuda_hessians_, num_used_indices, cuda_bagging_data_indices, cuda_sum_of_gradients_buffer_.RawData(),
+      cuda_sum_of_hessians_buffer_.RawData());
   }
   SynchronizeCUDADevice(__FILE__, __LINE__);
   CUDAInitValuesKernel2<<<1, NUM_THRADS_PER_BLOCK_LEAF_SPLITS>>>(
     lambda_l1, lambda_l2,
     num_blocks_init_from_gradients_,
-    cuda_sum_of_gradients_buffer_,
-    cuda_sum_of_hessians_buffer_,
+    cuda_sum_of_gradients_buffer_.RawData(),
+    cuda_sum_of_hessians_buffer_.RawData(),
     num_used_indices,
     cuda_data_indices_in_leaf,
     cuda_hist_in_leaf,
-    cuda_struct_);
+    cuda_struct_.RawData());
+  SynchronizeCUDADevice(__FILE__, __LINE__);
+}
+
+void CUDALeafSplits::LaunchInitValuesKernal(
+  const double lambda_l1, const double lambda_l2,
+  const data_size_t* cuda_bagging_data_indices,
+  const data_size_t* cuda_data_indices_in_leaf,
+  const data_size_t num_used_indices,
+  hist_t* cuda_hist_in_leaf,
+  const score_t* grad_scale,
+  const score_t* hess_scale) {
+  if (cuda_bagging_data_indices == nullptr) {
+    CUDAInitValuesKernel3<false><<<num_blocks_init_from_gradients_, NUM_THRADS_PER_BLOCK_LEAF_SPLITS>>>(
+      reinterpret_cast<const int16_t*>(cuda_gradients_), num_used_indices, nullptr, cuda_sum_of_gradients_buffer_.RawData(),
+      cuda_sum_of_hessians_buffer_.RawData(), cuda_sum_of_gradients_hessians_buffer_.RawData(), grad_scale, hess_scale);
+  } else {
+    CUDAInitValuesKernel3<true><<<num_blocks_init_from_gradients_, NUM_THRADS_PER_BLOCK_LEAF_SPLITS>>>(
+      reinterpret_cast<const int16_t*>(cuda_gradients_), num_used_indices, cuda_bagging_data_indices, cuda_sum_of_gradients_buffer_.RawData(),
+      cuda_sum_of_hessians_buffer_.RawData(), cuda_sum_of_gradients_hessians_buffer_.RawData(), grad_scale, hess_scale);
+  }
+
+  SynchronizeCUDADevice(__FILE__, __LINE__);
+  CUDAInitValuesKernel4<<<1, NUM_THRADS_PER_BLOCK_LEAF_SPLITS>>>(
+    lambda_l1, lambda_l2,
+    num_blocks_init_from_gradients_,
+    cuda_sum_of_gradients_buffer_.RawData(),
+    cuda_sum_of_hessians_buffer_.RawData(),
+    cuda_sum_of_gradients_hessians_buffer_.RawData(),
+    num_used_indices,
+    cuda_data_indices_in_leaf,
+    cuda_hist_in_leaf,
+    cuda_struct_.RawData());
   SynchronizeCUDADevice(__FILE__, __LINE__);
 }
 

src/treelearner/cuda/cuda_leaf_splits.hpp

@@ -8,7 +8,7 @@
 
 #ifdef USE_CUDA
 
-#include <LightGBM/cuda/cuda_utils.h>
+#include <LightGBM/cuda/cuda_utils.hu>
 #include <LightGBM/bin.h>
 #include <LightGBM/utils/log.h>
 #include <LightGBM/meta.h>
@@ -23,6 +23,7 @@ struct CUDALeafSplitsStruct {
   int leaf_index;
   double sum_of_gradients;
   double sum_of_hessians;
+  int64_t sum_of_gradients_hessians;
   data_size_t num_data_in_leaf;
   double gain;
   double leaf_value;
@@ -36,7 +37,7 @@ class CUDALeafSplits {
 
   ~CUDALeafSplits();
 
-  void Init();
+  void Init(const bool use_quantized_grad);
 
   void InitValues(
     const double lambda_l1, const double lambda_l2,
@@ -45,11 +46,19 @@ class CUDALeafSplits {
     const data_size_t* cuda_data_indices_in_leaf, const data_size_t num_used_indices,
     hist_t* cuda_hist_in_leaf, double* root_sum_hessians);
 
+  void InitValues(
+    const double lambda_l1, const double lambda_l2,
+    const int16_t* cuda_gradients_and_hessians,
+    const data_size_t* cuda_bagging_data_indices,
+    const data_size_t* cuda_data_indices_in_leaf, const data_size_t num_used_indices,
+    hist_t* cuda_hist_in_leaf, double* root_sum_hessians,
+    const score_t* grad_scale, const score_t* hess_scale);
+
   void InitValues();
 
-  const CUDALeafSplitsStruct* GetCUDAStruct() const { return cuda_struct_; }
+  const CUDALeafSplitsStruct* GetCUDAStruct() const { return cuda_struct_.RawDataReadOnly(); }
 
-  CUDALeafSplitsStruct* GetCUDAStructRef() { return cuda_struct_; }
+  CUDALeafSplitsStruct* GetCUDAStructRef() { return cuda_struct_.RawData(); }
 
   void Resize(const data_size_t num_data);
 
@@ -140,14 +149,24 @@ class CUDALeafSplits {
     const data_size_t num_used_indices,
     hist_t* cuda_hist_in_leaf);
 
+  void LaunchInitValuesKernal(
+    const double lambda_l1, const double lambda_l2,
+    const data_size_t* cuda_bagging_data_indices,
+    const data_size_t* cuda_data_indices_in_leaf,
+    const data_size_t num_used_indices,
+    hist_t* cuda_hist_in_leaf,
+    const score_t* grad_scale,
+    const score_t* hess_scale);
+
   // Host memory
   data_size_t num_data_;
   int num_blocks_init_from_gradients_;
 
   // CUDA memory, held by this object
-  CUDALeafSplitsStruct* cuda_struct_;
-  double* cuda_sum_of_gradients_buffer_;
-  double* cuda_sum_of_hessians_buffer_;
+  CUDAVector<CUDALeafSplitsStruct> cuda_struct_;
+  CUDAVector<double> cuda_sum_of_gradients_buffer_;
+  CUDAVector<double> cuda_sum_of_hessians_buffer_;
+  CUDAVector<int64_t> cuda_sum_of_gradients_hessians_buffer_;
 
   // CUDA memory, held by other object
   const score_t* cuda_gradients_;

src/treelearner/cuda/cuda_single_gpu_tree_learner.cpp

@@ -9,7 +9,7 @@
 #include "cuda_single_gpu_tree_learner.hpp"
 
 #include <LightGBM/cuda/cuda_tree.hpp>
-#include <LightGBM/cuda/cuda_utils.h>
+#include <LightGBM/cuda/cuda_utils.hu>
 #include <LightGBM/feature_group.h>
 #include <LightGBM/network.h>
 #include <LightGBM/objective_function.h>
@@ -39,13 +39,14 @@ void CUDASingleGPUTreeLearner::Init(const Dataset* train_data, bool is_constant_
   SetCUDADevice(gpu_device_id_, __FILE__, __LINE__);
 
   cuda_smaller_leaf_splits_.reset(new CUDALeafSplits(num_data_));
-  cuda_smaller_leaf_splits_->Init();
+  cuda_smaller_leaf_splits_->Init(config_->use_quantized_grad);
   cuda_larger_leaf_splits_.reset(new CUDALeafSplits(num_data_));
-  cuda_larger_leaf_splits_->Init();
+  cuda_larger_leaf_splits_->Init(config_->use_quantized_grad);
 
   cuda_histogram_constructor_.reset(new CUDAHistogramConstructor(train_data_, config_->num_leaves, num_threads_,
     share_state_->feature_hist_offsets(),
-    config_->min_data_in_leaf, config_->min_sum_hessian_in_leaf, gpu_device_id_, config_->gpu_use_dp));
+    config_->min_data_in_leaf, config_->min_sum_hessian_in_leaf, gpu_device_id_, config_->gpu_use_dp,
+    config_->use_quantized_grad, config_->num_grad_quant_bins));
   cuda_histogram_constructor_->Init(train_data_, share_state_.get());
 
   const auto& feature_hist_offsets = share_state_->feature_hist_offsets();
@@ -73,11 +74,19 @@ void CUDASingleGPUTreeLearner::Init(const Dataset* train_data, bool is_constant_
   }
   AllocateBitset();
 
-  cuda_leaf_gradient_stat_buffer_ = nullptr;
-  cuda_leaf_hessian_stat_buffer_ = nullptr;
   leaf_stat_buffer_size_ = 0;
   num_cat_threshold_ = 0;
 
+  if (config_->use_quantized_grad) {
+    cuda_leaf_gradient_stat_buffer_.Resize(config_->num_leaves);
+    cuda_leaf_hessian_stat_buffer_.Resize(config_->num_leaves);
+    cuda_gradient_discretizer_.reset(new CUDAGradientDiscretizer(
+      config_->num_grad_quant_bins, config_->num_iterations, config_->seed, is_constant_hessian, config_->stochastic_rounding));
+    cuda_gradient_discretizer_->Init(num_data_, config_->num_leaves, train_data_->num_features(), train_data_);
+  } else {
+    cuda_gradient_discretizer_.reset(nullptr);
+  }
+
   #ifdef DEBUG
   host_gradients_.resize(num_data_, 0.0f);
   host_hessians_.resize(num_data_, 0.0f);
@@ -101,6 +110,24 @@ void CUDASingleGPUTreeLearner::BeforeTrain() {
   const data_size_t* leaf_splits_init_indices =
     cuda_data_partition_->use_bagging() ? cuda_data_partition_->cuda_data_indices() : nullptr;
   cuda_data_partition_->BeforeTrain();
+  if (config_->use_quantized_grad) {
+    cuda_gradient_discretizer_->DiscretizeGradients(num_data_, gradients_, hessians_);
+    cuda_histogram_constructor_->BeforeTrain(
+      reinterpret_cast<const score_t*>(cuda_gradient_discretizer_->discretized_gradients_and_hessians()), nullptr);
+    cuda_smaller_leaf_splits_->InitValues(
+      config_->lambda_l1,
+      config_->lambda_l2,
+      reinterpret_cast<const int16_t*>(cuda_gradient_discretizer_->discretized_gradients_and_hessians()),
+      leaf_splits_init_indices,
+      cuda_data_partition_->cuda_data_indices(),
+      root_num_data,
+      cuda_histogram_constructor_->cuda_hist_pointer(),
+      &leaf_sum_hessians_[0],
+      cuda_gradient_discretizer_->grad_scale_ptr(),
+      cuda_gradient_discretizer_->hess_scale_ptr());
+      cuda_gradient_discretizer_->SetNumBitsInHistogramBin<false>(0, -1, root_num_data, 0);
+  } else {
+    cuda_histogram_constructor_->BeforeTrain(gradients_, hessians_);
     cuda_smaller_leaf_splits_->InitValues(
       config_->lambda_l1,
       config_->lambda_l2,
@@ -111,9 +138,9 @@ void CUDASingleGPUTreeLearner::BeforeTrain() {
       root_num_data,
       cuda_histogram_constructor_->cuda_hist_pointer(),
       &leaf_sum_hessians_[0]);
+  }
   leaf_num_data_[0] = root_num_data;
   cuda_larger_leaf_splits_->InitValues();
-  cuda_histogram_constructor_->BeforeTrain(gradients_, hessians_);
   col_sampler_.ResetByTree();
   cuda_best_split_finder_->BeforeTrain(col_sampler_.is_feature_used_bytree());
   leaf_data_start_[0] = 0;
@@ -141,24 +168,70 @@ Tree* CUDASingleGPUTreeLearner::Train(const score_t* gradients,
     const data_size_t num_data_in_larger_leaf = larger_leaf_index_ < 0 ? 0 : leaf_num_data_[larger_leaf_index_];
     const double sum_hessians_in_smaller_leaf = leaf_sum_hessians_[smaller_leaf_index_];
     const double sum_hessians_in_larger_leaf = larger_leaf_index_ < 0 ? 0 : leaf_sum_hessians_[larger_leaf_index_];
+    const uint8_t num_bits_in_histogram_bins = config_->use_quantized_grad ? cuda_gradient_discretizer_->GetHistBitsInLeaf<false>(smaller_leaf_index_) : 0;
     cuda_histogram_constructor_->ConstructHistogramForLeaf(
       cuda_smaller_leaf_splits_->GetCUDAStruct(),
       cuda_larger_leaf_splits_->GetCUDAStruct(),
       num_data_in_smaller_leaf,
       num_data_in_larger_leaf,
       sum_hessians_in_smaller_leaf,
-      sum_hessians_in_larger_leaf);
+      sum_hessians_in_larger_leaf,
+      num_bits_in_histogram_bins);
     global_timer.Stop("CUDASingleGPUTreeLearner::ConstructHistogramForLeaf");
     global_timer.Start("CUDASingleGPUTreeLearner::FindBestSplitsForLeaf");
 
+    uint8_t parent_num_bits_bin = 0;
+    uint8_t smaller_num_bits_bin = 0;
+    uint8_t larger_num_bits_bin = 0;
+    if (config_->use_quantized_grad) {
+      if (larger_leaf_index_ != -1) {
+        const int parent_leaf_index = std::min(smaller_leaf_index_, larger_leaf_index_);
+        parent_num_bits_bin = cuda_gradient_discretizer_->GetHistBitsInNode<false>(parent_leaf_index);
+        smaller_num_bits_bin = cuda_gradient_discretizer_->GetHistBitsInLeaf<false>(smaller_leaf_index_);
+        larger_num_bits_bin = cuda_gradient_discretizer_->GetHistBitsInLeaf<false>(larger_leaf_index_);
+      } else {
+        parent_num_bits_bin = cuda_gradient_discretizer_->GetHistBitsInLeaf<false>(0);
+        smaller_num_bits_bin = cuda_gradient_discretizer_->GetHistBitsInLeaf<false>(0);
+        larger_num_bits_bin = cuda_gradient_discretizer_->GetHistBitsInLeaf<false>(0);
+      }
+    } else {
+      parent_num_bits_bin = 0;
+      smaller_num_bits_bin = 0;
+      larger_num_bits_bin = 0;
+    }
+    cuda_histogram_constructor_->SubtractHistogramForLeaf(
+      cuda_smaller_leaf_splits_->GetCUDAStruct(),
+      cuda_larger_leaf_splits_->GetCUDAStruct(),
+      config_->use_quantized_grad,
+      parent_num_bits_bin,
+      smaller_num_bits_bin,
+      larger_num_bits_bin);
+
     SelectFeatureByNode(tree.get());
 
+    if (config_->use_quantized_grad) {
+      const uint8_t smaller_leaf_num_bits_bin = cuda_gradient_discretizer_->GetHistBitsInLeaf<false>(smaller_leaf_index_);
+      const uint8_t larger_leaf_num_bits_bin = larger_leaf_index_ < 0 ? 32 : cuda_gradient_discretizer_->GetHistBitsInLeaf<false>(larger_leaf_index_);
       cuda_best_split_finder_->FindBestSplitsForLeaf(
         cuda_smaller_leaf_splits_->GetCUDAStruct(),
         cuda_larger_leaf_splits_->GetCUDAStruct(),
         smaller_leaf_index_, larger_leaf_index_,
         num_data_in_smaller_leaf, num_data_in_larger_leaf,
-      sum_hessians_in_smaller_leaf, sum_hessians_in_larger_leaf);
+        sum_hessians_in_smaller_leaf, sum_hessians_in_larger_leaf,
+        cuda_gradient_discretizer_->grad_scale_ptr(),
+        cuda_gradient_discretizer_->hess_scale_ptr(),
+        smaller_leaf_num_bits_bin,
+        larger_leaf_num_bits_bin);
+    } else {
+      cuda_best_split_finder_->FindBestSplitsForLeaf(
+        cuda_smaller_leaf_splits_->GetCUDAStruct(),
+        cuda_larger_leaf_splits_->GetCUDAStruct(),
+        smaller_leaf_index_, larger_leaf_index_,
+        num_data_in_smaller_leaf, num_data_in_larger_leaf,
+        sum_hessians_in_smaller_leaf, sum_hessians_in_larger_leaf,
+        nullptr, nullptr, 0, 0);
+    }
+
     global_timer.Stop("CUDASingleGPUTreeLearner::FindBestSplitsForLeaf");
     global_timer.Start("CUDASingleGPUTreeLearner::FindBestFromAllSplits");
     const CUDASplitInfo* best_split_info = nullptr;
@@ -247,9 +320,19 @@ Tree* CUDASingleGPUTreeLearner::Train(const score_t* gradients,
     #endif  // DEBUG
     smaller_leaf_index_ = (leaf_num_data_[best_leaf_index_] < leaf_num_data_[right_leaf_index] ? best_leaf_index_ : right_leaf_index);
     larger_leaf_index_ = (smaller_leaf_index_ == best_leaf_index_ ? right_leaf_index : best_leaf_index_);
+
+    if (config_->use_quantized_grad) {
+      cuda_gradient_discretizer_->SetNumBitsInHistogramBin<false>(
+        best_leaf_index_, right_leaf_index, leaf_num_data_[best_leaf_index_], leaf_num_data_[right_leaf_index]);
+    }
     global_timer.Stop("CUDASingleGPUTreeLearner::Split");
   }
   SynchronizeCUDADevice(__FILE__, __LINE__);
+  if (config_->use_quantized_grad && config_->quant_train_renew_leaf) {
+    global_timer.Start("CUDASingleGPUTreeLearner::RenewDiscretizedTreeLeaves");
+    RenewDiscretizedTreeLeaves(tree.get());
+    global_timer.Stop("CUDASingleGPUTreeLearner::RenewDiscretizedTreeLeaves");
+  }
   tree->ToHost();
   return tree.release();
 }
@@ -357,8 +440,8 @@ void CUDASingleGPUTreeLearner::RenewTreeOutput(Tree* tree, const ObjectiveFuncti
 
 Tree* CUDASingleGPUTreeLearner::FitByExistingTree(const Tree* old_tree, const score_t* gradients, const score_t* hessians) const {
   std::unique_ptr<CUDATree> cuda_tree(new CUDATree(old_tree));
-  SetCUDAMemory<double>(cuda_leaf_gradient_stat_buffer_, 0, static_cast<size_t>(old_tree->num_leaves()), __FILE__, __LINE__);
-  SetCUDAMemory<double>(cuda_leaf_hessian_stat_buffer_, 0, static_cast<size_t>(old_tree->num_leaves()), __FILE__, __LINE__);
+  cuda_leaf_gradient_stat_buffer_.SetValue(0);
+  cuda_leaf_hessian_stat_buffer_.SetValue(0);
   ReduceLeafStat(cuda_tree.get(), gradients, hessians, cuda_data_partition_->cuda_data_indices());
   cuda_tree->SyncLeafOutputFromCUDAToHost();
   return cuda_tree.release();
@@ -373,13 +456,9 @@ Tree* CUDASingleGPUTreeLearner::FitByExistingTree(const Tree* old_tree, const st
     const int num_block = (refit_num_data_ + CUDA_SINGLE_GPU_TREE_LEARNER_BLOCK_SIZE - 1) / CUDA_SINGLE_GPU_TREE_LEARNER_BLOCK_SIZE;
     buffer_size *= static_cast<data_size_t>(num_block + 1);
   }
-  if (buffer_size != leaf_stat_buffer_size_) {
-    if (leaf_stat_buffer_size_ != 0) {
-      DeallocateCUDAMemory<double>(&cuda_leaf_gradient_stat_buffer_, __FILE__, __LINE__);
-      DeallocateCUDAMemory<double>(&cuda_leaf_hessian_stat_buffer_, __FILE__, __LINE__);
-    }
-    AllocateCUDAMemory<double>(&cuda_leaf_gradient_stat_buffer_, static_cast<size_t>(buffer_size), __FILE__, __LINE__);
-    AllocateCUDAMemory<double>(&cuda_leaf_hessian_stat_buffer_, static_cast<size_t>(buffer_size), __FILE__, __LINE__);
+  if (static_cast<size_t>(buffer_size) > cuda_leaf_gradient_stat_buffer_.Size()) {
+    cuda_leaf_gradient_stat_buffer_.Resize(buffer_size);
+    cuda_leaf_hessian_stat_buffer_.Resize(buffer_size);
   }
   return FitByExistingTree(old_tree, gradients, hessians);
 }
@@ -513,6 +592,15 @@ void CUDASingleGPUTreeLearner::CheckSplitValid(
 }
 #endif  // DEBUG
 
+void CUDASingleGPUTreeLearner::RenewDiscretizedTreeLeaves(CUDATree* cuda_tree) {
+  cuda_data_partition_->ReduceLeafGradStat(
+    gradients_, hessians_, cuda_tree,
+    cuda_leaf_gradient_stat_buffer_.RawData(),
+    cuda_leaf_hessian_stat_buffer_.RawData());
+  LaunchCalcLeafValuesGivenGradStat(cuda_tree, cuda_data_partition_->cuda_data_indices());
+  SynchronizeCUDADevice(__FILE__, __LINE__);
+}
+
 }  // namespace LightGBM
 
 #endif  // USE_CUDA

src/treelearner/cuda/cuda_single_gpu_tree_learner.cu

@@ -129,18 +129,18 @@ void CUDASingleGPUTreeLearner::LaunchReduceLeafStatKernel(
   if (num_leaves <= 2048) {
     ReduceLeafStatKernel_SharedMemory<<<num_block, CUDA_SINGLE_GPU_TREE_LEARNER_BLOCK_SIZE, 2 * num_leaves * sizeof(double)>>>(
       gradients, hessians, num_leaves, num_data, cuda_data_partition_->cuda_data_index_to_leaf_index(),
-      cuda_leaf_gradient_stat_buffer_, cuda_leaf_hessian_stat_buffer_);
+      cuda_leaf_gradient_stat_buffer_.RawData(), cuda_leaf_hessian_stat_buffer_.RawData());
   } else {
     ReduceLeafStatKernel_GlobalMemory<<<num_block, CUDA_SINGLE_GPU_TREE_LEARNER_BLOCK_SIZE>>>(
       gradients, hessians, num_leaves, num_data, cuda_data_partition_->cuda_data_index_to_leaf_index(),
-      cuda_leaf_gradient_stat_buffer_, cuda_leaf_hessian_stat_buffer_);
+      cuda_leaf_gradient_stat_buffer_.RawData(), cuda_leaf_hessian_stat_buffer_.RawData());
   }
   const bool use_l1 = config_->lambda_l1 > 0.0f;
   const bool use_smoothing = config_->path_smooth > 0.0f;
   num_block = (num_leaves + CUDA_SINGLE_GPU_TREE_LEARNER_BLOCK_SIZE - 1) / CUDA_SINGLE_GPU_TREE_LEARNER_BLOCK_SIZE;
 
   #define CalcRefitLeafOutputKernel_ARGS \
-    num_leaves, cuda_leaf_gradient_stat_buffer_, cuda_leaf_hessian_stat_buffer_, num_data_in_leaf, \
+    num_leaves, cuda_leaf_gradient_stat_buffer_.RawData(), cuda_leaf_hessian_stat_buffer_.RawData(), num_data_in_leaf, \
     leaf_parent, left_child, right_child, \
     config_->lambda_l1, config_->lambda_l2, config_->path_smooth, \
     shrinkage_rate, config_->refit_decay_rate, cuda_leaf_value
@@ -162,6 +162,7 @@ void CUDASingleGPUTreeLearner::LaunchReduceLeafStatKernel(
         <<<num_block, CUDA_SINGLE_GPU_TREE_LEARNER_BLOCK_SIZE>>>(CalcRefitLeafOutputKernel_ARGS);
     }
   }
+  #undef CalcRefitLeafOutputKernel_ARGS
 }
 
 template <typename T, bool IS_INNER>
@@ -256,6 +257,37 @@ void CUDASingleGPUTreeLearner::LaunchConstructBitsetForCategoricalSplitKernel(
   CUDAConstructBitset<int, false>(best_split_info, num_cat_threshold_, cuda_bitset_, cuda_bitset_len_);
 }
 
+void CUDASingleGPUTreeLearner::LaunchCalcLeafValuesGivenGradStat(
+  CUDATree* cuda_tree, const data_size_t* num_data_in_leaf) {
+  #define CalcRefitLeafOutputKernel_ARGS \
+    cuda_tree->num_leaves(), cuda_leaf_gradient_stat_buffer_.RawData(), cuda_leaf_hessian_stat_buffer_.RawData(), num_data_in_leaf, \
+    cuda_tree->cuda_leaf_parent(), cuda_tree->cuda_left_child(), cuda_tree->cuda_right_child(), \
+    config_->lambda_l1, config_->lambda_l2, config_->path_smooth, \
+    1.0f, config_->refit_decay_rate, cuda_tree->cuda_leaf_value_ref()
+  const bool use_l1 = config_->lambda_l1 > 0.0f;
+  const bool use_smoothing = config_->path_smooth > 0.0f;
+  const int num_block = (cuda_tree->num_leaves() + CUDA_SINGLE_GPU_TREE_LEARNER_BLOCK_SIZE - 1) / CUDA_SINGLE_GPU_TREE_LEARNER_BLOCK_SIZE;
+  if (!use_l1) {
+    if (!use_smoothing) {
+      CalcRefitLeafOutputKernel<false, false>
+        <<<num_block, CUDA_SINGLE_GPU_TREE_LEARNER_BLOCK_SIZE>>>(CalcRefitLeafOutputKernel_ARGS);
+    } else {
+      CalcRefitLeafOutputKernel<false, true>
+        <<<num_block, CUDA_SINGLE_GPU_TREE_LEARNER_BLOCK_SIZE>>>(CalcRefitLeafOutputKernel_ARGS);
+    }
+  } else {
+    if (!use_smoothing) {
+      CalcRefitLeafOutputKernel<true, false>
+        <<<num_block, CUDA_SINGLE_GPU_TREE_LEARNER_BLOCK_SIZE>>>(CalcRefitLeafOutputKernel_ARGS);
+    } else {
+      CalcRefitLeafOutputKernel<true, true>
+        <<<num_block, CUDA_SINGLE_GPU_TREE_LEARNER_BLOCK_SIZE>>>(CalcRefitLeafOutputKernel_ARGS);
+    }
+  }
+
+  #undef CalcRefitLeafOutputKernel_ARGS
+}
+
 }  // namespace LightGBM
 
 #endif  // USE_CUDA

src/treelearner/cuda/cuda_single_gpu_tree_learner.hpp

@@ -16,6 +16,7 @@
 #include "cuda_data_partition.hpp"
 #include "cuda_best_split_finder.hpp"
 
+#include "cuda_gradient_discretizer.hpp"
 #include "../serial_tree_learner.h"
 
 namespace LightGBM {
@@ -74,6 +75,10 @@ class CUDASingleGPUTreeLearner: public SerialTreeLearner {
     const double sum_left_gradients, const double sum_right_gradients);
   #endif  // DEBUG
 
+  void RenewDiscretizedTreeLeaves(CUDATree* cuda_tree);
+
+  void LaunchCalcLeafValuesGivenGradStat(CUDATree* cuda_tree, const data_size_t* num_data_in_leaf);
+
   // GPU device ID
   int gpu_device_id_;
   // number of threads on CPU
@@ -90,6 +95,8 @@ class CUDASingleGPUTreeLearner: public SerialTreeLearner {
   std::unique_ptr<CUDAHistogramConstructor> cuda_histogram_constructor_;
   // for best split information finding, given the histograms
   std::unique_ptr<CUDABestSplitFinder> cuda_best_split_finder_;
+  // gradient discretizer for quantized training
+  std::unique_ptr<CUDAGradientDiscretizer> cuda_gradient_discretizer_;
 
   std::vector<int> leaf_best_split_feature_;
   std::vector<uint32_t> leaf_best_split_threshold_;
@@ -108,8 +115,8 @@ class CUDASingleGPUTreeLearner: public SerialTreeLearner {
   std::vector<int> categorical_bin_to_value_;
   std::vector<int> categorical_bin_offsets_;
 
-  mutable double* cuda_leaf_gradient_stat_buffer_;
-  mutable double* cuda_leaf_hessian_stat_buffer_;
+  mutable CUDAVector<double> cuda_leaf_gradient_stat_buffer_;
+  mutable CUDAVector<double> cuda_leaf_hessian_stat_buffer_;
   mutable data_size_t leaf_stat_buffer_size_;
   mutable data_size_t refit_num_data_;
   uint32_t* cuda_bitset_;