warp-lang 1.4.1__py3-none-macosx_10_13_universal2.whl → 1.5.0__py3-none-macosx_10_13_universal2.whl

warp/native/coloring.cpp

@@ -0,0 +1,600 @@
+/** Copyright (c) 2024 NVIDIA CORPORATION.  All rights reserved.
+ * NVIDIA CORPORATION and its licensors retain all intellectual property
+ * and proprietary rights in and to this software, related documentation
+ * and any modifications thereto.  Any use, reproduction, disclosure or
+ * distribution of this software and related documentation without an express
+ * license agreement from NVIDIA CORPORATION is strictly prohibited.
+ */
+
+ // The Apache 2 License
+
+// Copyright 2023 Anka He Chen
+// 
+// Licensed under the Apache License, Version 2.0 (the "License"); you may not
+// use this file except in compliance with the License.You may obtain a copy of the License at
+// http ://www.apache.org/licenses/LICENSE-2.0
+// Unless required by applicable law or agreed to in writing, software distributed under the 
+// License is distributed on an "AS IS" BASIS, WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, 
+// either express or implied.See the License for the specific language governing permissions
+// and limitations under the License.
+// 
+// Source: https://github.com/AnkaChan/Gaia/blob/main/Simulator/Modules/GraphColoring/ColoringAlgorithms.cpp
+//         https://github.com/AnkaChan/Gaia/blob/main/Simulator/Modules/GraphColoring/ColoringAlgorithms.h
+//         https://github.com/AnkaChan/Gaia/blob/main/Simulator/Modules/GraphColoring/Graph.h
+
+
+
+#include "warp.h"
+
+#include <iostream>
+#include <vector>
+#include <array>
+#include <queue>
+#include <queue>
+#include <unordered_set>
+#include <random>
+#include <algorithm>
+#include <numeric>
+
+#define SHRINK_GRAPH_PER_PERCENTAGE (5)
+#define NODE_WEIGHTS_PREALLOC (64)
+#define WEIGHT_BUCKET_PREALLOC (512)
+
+namespace wp
+{
+
+    struct Graph
+    {
+        Graph(int num_nodes_in, const wp::array_t<int>& edges)
+            : num_nodes(num_nodes_in)
+        {
+            node_offsets.resize(num_nodes + 1, 0);
+            node_colors.resize(num_nodes, -1);
+
+           std::vector<int> node_degrees(num_nodes, 0);
+
+            // count degrees
+            for (size_t edge_idx = 0; edge_idx < edges.shape[0]; edge_idx++)
+            {
+                int e0 = *address(edges, edge_idx, 0);
+                int e1 = *address(edges, edge_idx, 1);
+                node_degrees[e0] += 1;
+                node_degrees[e1] += 1;
+            }
+            
+            int offset = 0;
+            for (size_t node = 0; node < num_nodes; node++)
+            {
+                offset += node_degrees[node];
+                node_offsets[node + 1] = offset;
+            }
+
+            // fill adjacency list
+            std::vector<int> node_adjacency_fill_count(num_nodes, 0);
+            graph_flatten.resize(offset, -1);
+            for (size_t edge_idx = 0; edge_idx < edges.shape[0]; edge_idx++)
+            {
+                int e0 = *address(edges, edge_idx, 0);
+                int e1 = *address(edges, edge_idx, 1);
+
+                int fill_count_e0 = node_adjacency_fill_count[e0];
+                graph_flatten[node_offsets[e0] + fill_count_e0] = e1;
+
+                int fill_count_e1 = node_adjacency_fill_count[e1];
+                graph_flatten[node_offsets[e1] + fill_count_e1] = e0;
+
+                node_adjacency_fill_count[e0] = fill_count_e0 + 1;
+                node_adjacency_fill_count[e1] = fill_count_e1 + 1;
+            }
+
+        }
+
+        int get_node_neighbor(int node, int neighbor_index) const {
+            return graph_flatten[node_offsets[node] + neighbor_index];
+        }
+
+        int get_node_degree(int node) const {
+            return node_offsets[node + 1] - node_offsets[node];
+        }
+
+
+        int num_nodes;
+        std::vector<int> graph_flatten;
+        std::vector<int> node_offsets;
+        std::vector<int> node_colors;
+    };
+
+void convert_to_color_groups(const int num_colors, const std::vector<int>& node_colors, std::vector<std::vector<int>>& color_groups)
+{
+    color_groups.resize(num_colors);
+
+    for (int node_idx = 0; node_idx < node_colors.size(); node_idx++) {
+        int color = node_colors[node_idx];
+        color_groups[color].push_back(node_idx);
+    }
+}
+
+float find_largest_smallest_groups(const std::vector<std::vector<int>>& color_groups, int& biggest_group, int& smallest_group)
+{
+    if (color_groups.size() == 0)
+    {
+        biggest_group = -1;
+        smallest_group = -1;
+
+        return 1;
+    }
+
+    size_t max_size = color_groups[0].size();
+    biggest_group = 0;
+    size_t min_size = color_groups[0].size();
+    smallest_group = 0;
+
+    for (size_t color = 0; color < color_groups.size(); color++)
+    {
+        if (max_size < color_groups[color].size()) {
+            biggest_group = color;
+            max_size = color_groups[color].size();
+        }
+
+        if (min_size > color_groups[color].size())
+        {
+            smallest_group = color;
+            min_size = color_groups[color].size();
+        }
+    }
+
+    return float(color_groups[biggest_group].size()) / float(color_groups[smallest_group].size());
+}
+
+bool color_changeable(const Graph& graph, int node, int target_color){
+    // loop through node and see if it has target color
+    for (size_t i = 0; i < graph.get_node_degree(node); i++)
+    {
+        int nei_node_idx = graph.get_node_neighbor(node, i);
+        if (graph.node_colors[nei_node_idx] == target_color)
+        {
+            return false;
+        }
+    }
+    return true;
+}
+
+int find_changeable_node_in_category(
+    const Graph& graph, 
+    const std::vector<std::vector<int>>& color_groups, 
+    int source_color, 
+    int target_color
+)
+{
+    auto& source_group = color_groups[source_color];
+    for (size_t node_idx = 0; node_idx < source_group.size(); node_idx++)
+    {
+        if (color_changeable(graph, source_group[node_idx], target_color)) {
+            return node_idx;
+        }
+    }
+    return -1;
+}
+
+void change_color(int color, int node_idx_in_group, int target_color, std::vector<int>& node_colors, std::vector<std::vector<int>>& color_groups)
+{
+    int node_idx = color_groups[color][node_idx_in_group];
+    node_colors[node_idx] = target_color;
+
+    if (color_groups.size())
+    {
+        // O(1) erase
+        std::swap(color_groups[color][node_idx_in_group], color_groups[color].back());
+        color_groups[color].pop_back();
+
+        color_groups[target_color].push_back(node_idx);
+    }
+}
+
+float balance_color_groups(float target_max_min_ratio, 
+    Graph& graph, 
+    std::vector<std::vector<int>>& color_groups)
+{
+    float max_min_ratio = -1.f;
+
+    do
+    {
+        int biggest_group = -1, smallest_group = -1;
+
+        max_min_ratio = find_largest_smallest_groups(color_groups, biggest_group, smallest_group);
+
+        // graph is not optimizable anymore or target ratio reached
+        if (color_groups[biggest_group].size() - color_groups[smallest_group].size() <= 2 
+            || max_min_ratio < target_max_min_ratio)
+        {
+            return max_min_ratio;
+        }
+
+        // find a available vertex from the biggest category to move to the smallest category
+        int changeable_color_group_idx = biggest_group;
+        int changeable_node_idx = find_changeable_node_in_category(graph, color_groups, biggest_group, smallest_group);
+        if (changeable_node_idx == -1)
+        {
+            for (size_t color = 0; color < color_groups.size(); color++)
+            {
+                if (color == biggest_group || color == smallest_group)
+                {
+                    continue;
+                }
+
+                changeable_node_idx = find_changeable_node_in_category(graph, color_groups, color, smallest_group);
+
+                if (changeable_node_idx != -1)
+                {
+                    changeable_color_group_idx = color;
+
+                    break;
+                }
+            }
+        }
+
+
+        if (changeable_node_idx == -1)
+        {
+            // fprintf(stderr, "The graph is not optimizable anymore, terminated with a max/min ratio: %f without reaching the target ratio: %f\n", max_min_ratio, target_max_min_ratio);
+            return max_min_ratio;
+        }
+        // change the color of changeable_color_idx in group changeable_color_group_idx to 
+        change_color(changeable_color_group_idx, changeable_node_idx, smallest_group, graph.node_colors, color_groups);
+
+
+    } while (max_min_ratio > target_max_min_ratio);
+
+    return max_min_ratio;
+}
+
+int graph_coloring_ordered_greedy(const std::vector<int>& order, Graph& graph)
+{
+    // greedy coloring
+    int max_color = -1;
+    int num_colored = 0;
+    std::vector<bool> color_used;
+    color_used.reserve(128);
+
+    for (size_t i = 0; i < order.size(); i++)
+    {
+        int node = order[i];
+
+        // first one
+        if (max_color == -1)
+        {
+            ++max_color;
+            graph.node_colors[node] = max_color;
+        }
+        else {
+            color_used.resize(max_color + 1);
+
+            for (int color_counter = 0; color_counter < color_used.size(); color_counter++)
+            {
+                color_used[color_counter] = false;
+            }
+
+            // see its neighbor's color
+            for (int nei_counter = 0; nei_counter < graph.get_node_degree(node); nei_counter++)
+            {
+                int nei_node_idx = graph.get_node_neighbor(node, nei_counter);
+                if (graph.node_colors[nei_node_idx] >= 0)
+                {
+                    color_used[graph.node_colors[nei_node_idx]] = true;
+                }
+            }
+
+            // find the minimal usable color
+            int min_usable_color = -1;
+            for (int color_counter = 0; color_counter < color_used.size(); color_counter++)
+            {
+                if (!color_used[color_counter]) {
+                    min_usable_color = color_counter;
+                    break;
+                }
+            }
+            if (min_usable_color == -1)
+            {
+                ++max_color;
+                graph.node_colors[node] = max_color;
+            }
+            else
+            {
+                graph.node_colors[node] = min_usable_color;
+            }
+        }
+
+        num_colored++;
+    }
+    return (max_color + 1);
+}
+
+class NodeWeightBuckets
+{
+public:
+    NodeWeightBuckets(int num_nodes)
+        : node_weights(num_nodes, 0), node_indices_in_bucket(num_nodes, -1)
+    {
+        weight_buckets.resize(NODE_WEIGHTS_PREALLOC);
+        for (size_t i = 1; i < weight_buckets.size(); i++)
+        {
+            weight_buckets[i].reserve(WEIGHT_BUCKET_PREALLOC);
+        }
+        max_weight = 0;
+    }
+
+    int get_node_weight(int node_idx) 
+    {
+        return node_weights[node_idx];
+    }
+
+    void add_node(int weight, int node_idx)
+    {
+        if (weight >= weight_buckets.size()) 
+        {
+            weight_buckets.resize(weight + 1);
+        }
+        
+        node_indices_in_bucket[node_idx] = weight_buckets[weight].size();
+        node_weights[node_idx] = weight;
+        weight_buckets[weight].push_back(node_idx);
+
+        if (max_weight < weight)
+        {
+            max_weight = weight;
+        }
+    }
+
+    int pop_node_with_max_weight() {
+        int node_with_max_weight = weight_buckets[max_weight].front();
+        node_indices_in_bucket[node_with_max_weight] = -1;
+
+        // we pop the first element so it has a breadth-first like behavior, which is better than depth-first
+        if (weight_buckets[max_weight].size() > 1)
+        {
+            node_indices_in_bucket[weight_buckets[max_weight].back()] = 0;
+            weight_buckets[max_weight][0] = weight_buckets[max_weight].back();
+        }
+        weight_buckets[max_weight].pop_back();
+        // mark node deleted
+        node_weights[node_with_max_weight] = -1;
+
+        if (weight_buckets[max_weight].size() == 0)
+            // we need to update max_weight because weight_buckets[max_weight] became empty
+        {
+            int new_max_weight = 0;
+            for (size_t bucket_idx = max_weight - 1; bucket_idx >= 0; bucket_idx--)
+            {
+                if (weight_buckets[bucket_idx].size())
+                {
+                    new_max_weight = bucket_idx;
+                    break;
+                }
+            }
+
+            max_weight = new_max_weight;
+        }
+        // mark deleted
+        return node_with_max_weight;
+    }
+
+    void increase_node_weight(int node_idx)
+    {
+        int weight = node_weights[node_idx];
+        assert(weight < weight_buckets.size());
+        int node_idx_in_bucket = node_indices_in_bucket[node_idx];
+        assert(node_idx_in_bucket < weight_buckets[weight].size());
+
+        // swap index with the last element
+        node_indices_in_bucket[weight_buckets[weight].back()] = node_idx_in_bucket;
+        // O(1) erase
+        weight_buckets[weight][node_idx_in_bucket] = weight_buckets[weight].back();
+        weight_buckets[weight].pop_back();
+
+        add_node(weight + 1, node_idx);
+    }
+
+    bool empty()
+    {
+        return max_weight <= 0 && weight_buckets[0].size() == 0;
+    }
+
+
+private:
+    int max_weight;
+    std::vector<std::vector<int>> weight_buckets;
+    std::vector<int> node_indices_in_bucket;
+    std::vector<int> node_weights;
+};
+
+// Pereira, F. M. Q., & Palsberg, J. (2005, November). Register allocation via coloring of chordal graphs. In Asian Symposium on Programming Languages and Systems (pp. 315-329). Berlin, Heidelberg: Springer Berlin Heidelberg.
+int graph_coloring_mcs_vector(Graph& graph)
+{
+    // Initially set the weight of each node to 0
+    std::vector<int> ordering;
+    ordering.reserve(graph.num_nodes);
+
+    NodeWeightBuckets weight_buckets(graph.num_nodes);
+    // add the first node
+    weight_buckets.add_node(0, 0);
+
+    for (int node_idx = 0; node_idx < graph.num_nodes; node_idx++)
+    {
+        // this might look like it's O(N^2) but this only happens once per connected components
+        if (weight_buckets.empty())
+        {
+            int non_negative_node = -1;
+            for (size_t i = 0; i < graph.num_nodes; i++)
+            {
+                if (weight_buckets.get_node_weight(i) >= 0) {
+                    non_negative_node = i;
+                    break;
+                }
+            }
+            assert(weight_buckets.get_node_weight(non_negative_node) == 0);
+            weight_buckets.add_node(0, non_negative_node);
+        }
+
+        int max_node = weight_buckets.pop_node_with_max_weight();
+
+        // Add highest weight node to the queue and increment all of its neighbors weights by 1
+        ordering.push_back(max_node);
+
+        for (unsigned j = 0; j < graph.get_node_degree(max_node); j++) {
+            int neighbor_node = graph.get_node_neighbor(max_node, j);
+            int old_weight = weight_buckets.get_node_weight(neighbor_node);
+
+            if (old_weight == 0)
+                // 0-weighted node is not in buckets by default
+            {
+                weight_buckets.add_node(old_weight + 1, neighbor_node);
+
+            }
+            else if (old_weight > 0) {
+                weight_buckets.increase_node_weight(neighbor_node);
+            }
+            // skip neighbor nodes with negative weight because they are visited
+        }
+    }
+
+    return graph_coloring_ordered_greedy(ordering, graph);
+}
+
+int next_node(const int num_nodes, const std::vector<int>& degrees)
+{
+    int node_min_degrees = -1;
+    int min_degree = num_nodes + 1;
+    for (size_t node_idx = 0; node_idx < degrees.size(); node_idx++)
+    {
+        if (degrees[node_idx] == -1)
+        {
+            continue;
+        }
+        if (min_degree > degrees[node_idx]) {
+            min_degree = degrees[node_idx];
+            node_min_degrees = node_idx;
+        }
+    }
+    return node_min_degrees;
+}
+
+void reduce_degree(int node_idx, Graph& graph, std::vector<int>& degrees)
+{
+    degrees[node_idx] = -1;
+    for (size_t nei_node_counter = 0; nei_node_counter < graph.get_node_degree(node_idx); nei_node_counter++)
+    {
+        int nei_node_idx = graph.get_node_neighbor(node_idx, nei_node_counter);
+
+        if (degrees[nei_node_idx] != -1)
+        {
+            degrees[nei_node_idx]--;
+        }
+    }
+}
+
+
+// Fratarcangeli, Marco, and Fabio Pellacini. "Scalable partitioning for parallel position based dynamics." Computer Graphics Forum. Vol. 34. No. 2. 2015.
+int graph_coloring_degree_ordered_greedy(Graph& graph)
+{
+    // initialize the degree
+    std::vector<int> degrees(graph.num_nodes, 0);
+    for (int node_idx = 0; node_idx < graph.num_nodes; node_idx++) {
+        degrees[node_idx] = graph.get_node_degree(node_idx);
+    }
+
+    // order them in a descending order
+    std::vector<int> ordering(graph.num_nodes);
+    std::iota(std::begin(ordering), std::end(ordering), 0);
+    std::sort(std::begin(ordering), std::end(ordering),
+        [&degrees](const auto& lhs, const auto& rhs)
+        {
+            return degrees[lhs] > degrees[rhs];
+        }
+    );
+
+    return graph_coloring_ordered_greedy(ordering, graph);
+}
+
+int graph_coloring_naive_greedy(Graph& graph)
+{
+    std::vector<int> ordering(graph.num_nodes);
+    std::iota(std::begin(ordering), std::end(ordering), 0);
+    return graph_coloring_ordered_greedy(ordering, graph);
+}
+}
+using namespace wp;
+
+extern "C"
+{
+    int graph_coloring(int num_nodes, wp::array_t<int> edges, int algorithm, wp::array_t<int> node_colors)
+    {
+        if (node_colors.ndim != 1 || node_colors.shape[0] != num_nodes)
+        {
+            fprintf(stderr, "The node_colors array must have the preallocated shape of (num_nodes,)!\n");
+            return -1;
+        }
+
+        if (edges.ndim != 2)
+        {
+            fprintf(stderr, "The edges array must have 2 dimensions!\n");
+            return -1;
+        }
+
+        if (num_nodes == 0)
+        {
+            fprintf(stderr, "Empty graph!\n");
+            return -1;
+        }
+
+        // convert to a format that coloring algorithm can recognize
+
+        Graph graph(num_nodes, edges);
+
+        int num_colors = -1;
+        switch (algorithm)
+        {
+        case 0:
+            // mcs algorithm
+            num_colors = graph_coloring_mcs_vector(graph);
+            break;
+        case 1:
+            // greedy
+            num_colors = graph_coloring_degree_ordered_greedy(graph);
+            break;
+        //case 2:
+        //    // mcs algorithm
+        //    num_colors = graph_coloring_mcs_set(graph);
+        //    break;
+        //case 3:
+        //    // naive greedy
+        //    num_colors = graph_coloring_naive_greedy(graph);
+        //    break;
+        default:
+            fprintf(stderr, "Unrecognized coloring algorithm number: %d!\n", algorithm);
+            return -1;
+            break;
+        }
+
+        // copy the color info back
+        memcpy(node_colors.data, graph.node_colors.data(), num_nodes * sizeof(int));
+
+        return num_colors;
+    }
+
+    float balance_coloring(int num_nodes, wp::array_t<int> edges, int num_colors, 
+        float target_max_min_ratio, wp::array_t<int> node_colors)
+    {
+        Graph graph(num_nodes, edges);
+        // copy the color info to graph
+        memcpy(graph.node_colors.data(), node_colors.data, num_nodes * sizeof(int));
+        if (num_colors > 1) {
+            std::vector<std::vector<int>> color_groups;
+            convert_to_color_groups(num_colors, graph.node_colors, color_groups);
+            return balance_color_groups(target_max_min_ratio, graph, color_groups);
+        }
+        else
+        {
+            return 1.f;
+        }
+    }
+}

warp/native/cuda_util.cpp

@@ -100,6 +100,8 @@ static PFN_cuGraphicsUnmapResources_v3000 pfn_cuGraphicsUnmapResources;
 static PFN_cuGraphicsResourceGetMappedPointer_v3020 pfn_cuGraphicsResourceGetMappedPointer;
 static PFN_cuGraphicsGLRegisterBuffer_v3000 pfn_cuGraphicsGLRegisterBuffer;
 static PFN_cuGraphicsUnregisterResource_v3000 pfn_cuGraphicsUnregisterResource;
+static PFN_cuModuleGetGlobal_v3020 pfn_cuModuleGetGlobal;
+static PFN_cuFuncSetAttribute_v9000 pfn_cuFuncSetAttribute;
 
 static bool cuda_driver_initialized = false;
 
@@ -231,6 +233,8 @@ bool init_cuda_driver()
     get_driver_entry_point("cuGraphicsResourceGetMappedPointer", &(void*&)pfn_cuGraphicsResourceGetMappedPointer);
     get_driver_entry_point("cuGraphicsGLRegisterBuffer", &(void*&)pfn_cuGraphicsGLRegisterBuffer);
     get_driver_entry_point("cuGraphicsUnregisterResource", &(void*&)pfn_cuGraphicsUnregisterResource);
+    get_driver_entry_point("cuModuleGetGlobal", &(void*&)pfn_cuModuleGetGlobal);
+    get_driver_entry_point("cuFuncSetAttribute", &(void*&)pfn_cuFuncSetAttribute);
 
     if (pfn_cuInit)
         cuda_driver_initialized = check_cu(pfn_cuInit(0));
@@ -568,4 +572,14 @@ CUresult cuGraphicsUnregisterResource_f(CUgraphicsResource resource)
     return pfn_cuGraphicsUnregisterResource ? pfn_cuGraphicsUnregisterResource(resource) : DRIVER_ENTRY_POINT_ERROR;
 }
 
+CUresult cuModuleGetGlobal_f(CUdeviceptr* dptr, size_t* bytes, CUmodule hmod, const char* name )
+{
+    return pfn_cuModuleGetGlobal ? pfn_cuModuleGetGlobal(dptr, bytes, hmod, name) : DRIVER_ENTRY_POINT_ERROR;
+}
+
+CUresult cuFuncSetAttribute_f(CUfunction hfunc, CUfunction_attribute attrib, int value) 
+{
+    return pfn_cuFuncSetAttribute ? pfn_cuFuncSetAttribute(hfunc, attrib, value) : DRIVER_ENTRY_POINT_ERROR;
+}
+
 #endif // WP_ENABLE_CUDA

warp/native/cuda_util.h

@@ -99,7 +99,8 @@ CUresult cuGraphicsUnmapResources_f(unsigned int count, CUgraphicsResource* reso
 CUresult cuGraphicsResourceGetMappedPointer_f(CUdeviceptr* pDevPtr, size_t* pSize, CUgraphicsResource resource);
 CUresult cuGraphicsGLRegisterBuffer_f(CUgraphicsResource *pCudaResource, unsigned int buffer, unsigned int flags);
 CUresult cuGraphicsUnregisterResource_f(CUgraphicsResource resource);
-
+CUresult cuModuleGetGlobal_f(CUdeviceptr* dptr, size_t* bytes, CUmodule hmod, const char* name );
+CUresult cuFuncSetAttribute_f(CUfunction hfunc, CUfunction_attribute attrib, int value);
 
 bool init_cuda_driver();
 bool is_cuda_driver_initialized();

warp/native/fabric.h

@@ -1,3 +1,11 @@
+/** Copyright (c) 2023 NVIDIA CORPORATION.  All rights reserved.
+ * NVIDIA CORPORATION and its licensors retain all intellectual property
+ * and proprietary rights in and to this software, related documentation
+ * and any modifications thereto.  Any use, reproduction, disclosure or
+ * distribution of this software and related documentation without an express
+ * license agreement from NVIDIA CORPORATION is strictly prohibited.
+ */
+
 #pragma once
 
 #include "builtin.h"

warp/native/hashgrid.h

@@ -209,6 +209,10 @@ CUDA_CALLABLE inline hash_grid_query_t iter_reverse(const hash_grid_query_t& que
     return query;
 }
 
+CUDA_CALLABLE inline void adj_iter_reverse(const hash_grid_query_t& query, hash_grid_query_t& adj_query, hash_grid_query_t& adj_ret)
+{
+}
+
 
 
 CUDA_CALLABLE inline int hash_grid_point_id(uint64_t id, int& index)

warp/native/marching.cu

@@ -1,3 +1,11 @@
+/** Copyright (c) 2022 NVIDIA CORPORATION.  All rights reserved.
+ * NVIDIA CORPORATION and its licensors retain all intellectual property
+ * and proprietary rights in and to this software, related documentation
+ * and any modifications thereto.  Any use, reproduction, disclosure or
+ * distribution of this software and related documentation without an express
+ * license agreement from NVIDIA CORPORATION is strictly prohibited.
+ */
+
 #include "warp.h"
 #include "cuda_util.h"
 #include "scan.h"