warp-lang 1.0.2__py3-none-manylinux2014_x86_64.whl → 1.1.0__py3-none-manylinux2014_x86_64.whl

warp/native/sparse.cu

@@ -1,544 +1,544 @@
-#include "cuda_util.h"
-#include "warp.h"
-
-#define THRUST_IGNORE_CUB_VERSION_CHECK
-
-#include <cub/device/device_radix_sort.cuh>
-#include <cub/device/device_run_length_encode.cuh>
-#include <cub/device/device_scan.cuh>
-#include <cub/device/device_select.cuh>
-
-namespace {
-
-// Combined row+column value that can be radix-sorted with CUB
-using BsrRowCol = uint64_t;
-
-CUDA_CALLABLE BsrRowCol bsr_combine_row_col(uint32_t row, uint32_t col) {
-  return (static_cast<uint64_t>(row) << 32) | col;
-}
-
-CUDA_CALLABLE uint32_t bsr_get_row(const BsrRowCol &row_col) {
-  return row_col >> 32;
-}
-
-CUDA_CALLABLE uint32_t bsr_get_col(const BsrRowCol &row_col) {
-  return row_col & INT_MAX;
-}
-
-// Cached temporary storage
-struct BsrFromTripletsTemp {
-  
-  int *count_buffer = NULL;
-  cudaEvent_t host_sync_event = NULL;
-
-  BsrFromTripletsTemp()
-    : count_buffer(static_cast<int*>(alloc_pinned(sizeof(int))))
-  {
-    cudaEventCreateWithFlags(&host_sync_event, cudaEventDisableTiming);
-  }
-  
-  ~BsrFromTripletsTemp()
-  {
-    cudaEventDestroy(host_sync_event);
-    free_pinned(count_buffer);
-  }
-
-  BsrFromTripletsTemp(const BsrFromTripletsTemp&) = delete;
-  BsrFromTripletsTemp& operator=(const BsrFromTripletsTemp&) = delete;
-
-};
-
-// map temp buffers to CUDA contexts
-static std::unordered_map<void *, BsrFromTripletsTemp> g_bsr_from_triplets_temp_map;
-
-template <typename T> struct BsrBlockIsNotZero {
-  int block_size;
-  const T *values;
-
-  CUDA_CALLABLE_DEVICE bool operator()(int i) const {
-    const T *val = values + i * block_size;
-    for (int i = 0; i < block_size; ++i, ++val) {
-      if (*val != T(0))
-        return true;
-    }
-    return false;
-  }
-};
-
-__global__ void bsr_fill_block_indices(int nnz, int *block_indices) {
-  int i = blockIdx.x * blockDim.x + threadIdx.x;
-  if (i >= nnz)
-    return;
-
-  block_indices[i] = i;
-}
-
-__global__ void bsr_fill_row_col(const int *nnz, const int *block_indices,
-                                 const int *tpl_rows, const int *tpl_columns,
-                                 BsrRowCol *tpl_row_col) {
-  int i = blockIdx.x * blockDim.x + threadIdx.x;
-  if (i >= *nnz)
-    return;
-
-  const int block = block_indices[i];
-
-  BsrRowCol row_col = bsr_combine_row_col(tpl_rows[block], tpl_columns[block]);
-  tpl_row_col[i] = row_col;
-}
-
-template <typename T>
-__global__ void
-bsr_merge_blocks(int nnz, int block_size, const int *block_offsets,
-                 const int *sorted_block_indices,
-                 const BsrRowCol *unique_row_cols, const T *tpl_values,
-                 int *bsr_row_counts, int *bsr_cols, T *bsr_values)
-
-{
-  int i = blockIdx.x * blockDim.x + threadIdx.x;
-  if (i >= nnz)
-    return;
-
-  const int beg = i ? block_offsets[i - 1] : 0;
-  const int end = block_offsets[i];
-
-  BsrRowCol row_col = unique_row_cols[i];
-
-  bsr_cols[i] = bsr_get_col(row_col);
-  atomicAdd(bsr_row_counts + bsr_get_row(row_col) + 1, 1);
-
-  if (bsr_values == nullptr)
-    return;
-
-  T *bsr_val = bsr_values + i * block_size;
-  const T *tpl_val = tpl_values + sorted_block_indices[beg] * block_size;
-
-  for (int k = 0; k < block_size; ++k) {
-    bsr_val[k] = tpl_val[k];
-  }
-
-  for (int cur = beg + 1; cur != end; ++cur) {
-    const T *tpl_val = tpl_values + sorted_block_indices[cur] * block_size;
-    for (int k = 0; k < block_size; ++k) {
-      bsr_val[k] += tpl_val[k];
-    }
-  }
-}
-
-template <typename T>
-int bsr_matrix_from_triplets_device(const int rows_per_block,
-                                    const int cols_per_block,
-                                    const int row_count, const int nnz,
-                                    const int *tpl_rows, const int *tpl_columns,
-                                    const T *tpl_values, int *bsr_offsets,
-                                    int *bsr_columns, T *bsr_values) {
-  const int block_size = rows_per_block * cols_per_block;
-
-  void *context = cuda_context_get_current();
-  ContextGuard guard(context);
-
-  // Per-context cached temporary buffers
-  BsrFromTripletsTemp &bsr_temp = g_bsr_from_triplets_temp_map[context];
-
-  cudaStream_t stream = static_cast<cudaStream_t>(cuda_stream_get_current());
-
-  ScopedTemporary<int> block_indices(context, 2*nnz);
-  ScopedTemporary<BsrRowCol> combined_row_col(context, 2*nnz);
-
-  cub::DoubleBuffer<int> d_keys(block_indices.buffer(),
-                                block_indices.buffer() + nnz);
-  cub::DoubleBuffer<BsrRowCol> d_values(combined_row_col.buffer(),
-                                        combined_row_col.buffer() + nnz);
-
-  int *p_nz_triplet_count = bsr_temp.count_buffer;
-
-  wp_launch_device(WP_CURRENT_CONTEXT, bsr_fill_block_indices, nnz,
-                   (nnz, d_keys.Current()));
-
-  if (tpl_values) {
-
-    // Remove zero blocks
-    {
-      size_t buff_size = 0;
-      BsrBlockIsNotZero<T> isNotZero{block_size, tpl_values};
-      check_cuda(cub::DeviceSelect::If(nullptr, buff_size, d_keys.Current(),
-                                      d_keys.Alternate(), p_nz_triplet_count,
-                                      nnz, isNotZero, stream));
-      ScopedTemporary<> temp(context, buff_size);
-      check_cuda(cub::DeviceSelect::If(
-          temp.buffer(), buff_size, d_keys.Current(), d_keys.Alternate(),
-          p_nz_triplet_count, nnz, isNotZero, stream));
-    }
-    cudaEventRecord(bsr_temp.host_sync_event, stream);
-
-    // switch current/alternate in double buffer
-    d_keys.selector ^= 1;
-
-  } else {
-    *p_nz_triplet_count = nnz;
-  }
-
-  // Combine rows and columns so we can sort on them both
-  wp_launch_device(WP_CURRENT_CONTEXT, bsr_fill_row_col, nnz,
-                   (p_nz_triplet_count, d_keys.Current(), tpl_rows, tpl_columns,
-                    d_values.Current()));
-
-  if (tpl_values) {
-    // Make sure count is available on host
-    cudaEventSynchronize(bsr_temp.host_sync_event);
-  }
-
-  const int nz_triplet_count = *p_nz_triplet_count;
-
-  // Sort
-  {
-    size_t buff_size = 0;
-    check_cuda(cub::DeviceRadixSort::SortPairs(
-        nullptr, buff_size, d_values, d_keys, nz_triplet_count, 0, 64, stream));
-    ScopedTemporary<> temp(context, buff_size);
-    check_cuda(cub::DeviceRadixSort::SortPairs(temp.buffer(), buff_size,
-                                              d_values, d_keys, nz_triplet_count,
-                                              0, 64, stream));
-  }
-
-  // Runlength encode row-col sequences
-  {
-    size_t buff_size = 0;
-    check_cuda(cub::DeviceRunLengthEncode::Encode(
-        nullptr, buff_size, d_values.Current(), d_values.Alternate(),
-        d_keys.Alternate(), p_nz_triplet_count, nz_triplet_count, stream));
-    ScopedTemporary<> temp(context, buff_size);
-    check_cuda(cub::DeviceRunLengthEncode::Encode(
-        temp.buffer(), buff_size, d_values.Current(), d_values.Alternate(),
-        d_keys.Alternate(), p_nz_triplet_count, nz_triplet_count, stream));
-  }
-
-  cudaEventRecord(bsr_temp.host_sync_event, stream);
-
-  // Now we have the following:
-  // d_values.Current(): sorted block row-col
-  // d_values.Alternate(): sorted unique row-col
-  // d_keys.Current(): sorted block indices
-  // d_keys.Alternate(): repeated block-row count
-
-  // Scan repeated block counts
-  {
-    size_t buff_size = 0;
-    check_cuda(cub::DeviceScan::InclusiveSum(
-        nullptr, buff_size, d_keys.Alternate(), d_keys.Alternate(),
-        nz_triplet_count, stream));
-    ScopedTemporary<> temp(context, buff_size);
-    check_cuda(cub::DeviceScan::InclusiveSum(
-        temp.buffer(), buff_size, d_keys.Alternate(), d_keys.Alternate(),
-        nz_triplet_count, stream));
-  }
-
-  // While we're at it, zero the bsr offsets buffer
-  memset_device(WP_CURRENT_CONTEXT, bsr_offsets, 0,
-                (row_count + 1) * sizeof(int));
-
-  // Wait for number of compressed blocks
-  cudaEventSynchronize(bsr_temp.host_sync_event);
-  const int compressed_nnz = *p_nz_triplet_count;
-
-  // We have all we need to accumulate our repeated blocks
-  wp_launch_device(WP_CURRENT_CONTEXT, bsr_merge_blocks, compressed_nnz,
-                   (compressed_nnz, block_size, d_keys.Alternate(),
-                    d_keys.Current(), d_values.Alternate(), tpl_values,
-                    bsr_offsets, bsr_columns, bsr_values));
-
-  // Last, prefix sum the row block counts
-  {
-    size_t buff_size = 0;
-    check_cuda(cub::DeviceScan::InclusiveSum(nullptr, buff_size, bsr_offsets,
-                                            bsr_offsets, row_count + 1, stream));
-    ScopedTemporary<> temp(context, buff_size);
-    check_cuda(cub::DeviceScan::InclusiveSum(temp.buffer(), buff_size,
-                                            bsr_offsets, bsr_offsets,
-                                            row_count + 1, stream));
-  }
-
-  return compressed_nnz;
-}
-
-__global__ void bsr_transpose_fill_row_col(const int nnz, const int row_count,
-                                           const int *bsr_offsets,
-                                           const int *bsr_columns,
-                                           int *block_indices,
-                                           BsrRowCol *transposed_row_col,
-                                           int *transposed_bsr_offsets) {
-  int i = blockIdx.x * blockDim.x + threadIdx.x;
-  if (i >= nnz)
-    return;
-
-  block_indices[i] = i;
-
-  // Binary search for row
-  int lower = 0;
-  int upper = row_count - 1;
-
-  while (lower < upper) {
-    int mid = lower + (upper - lower) / 2;
-
-    if (bsr_offsets[mid + 1] <= i) {
-      lower = mid + 1;
-    } else {
-      upper = mid;
-    }
-  }
-
-  const int row = lower;
-  const int col = bsr_columns[i];
-  BsrRowCol row_col = bsr_combine_row_col(col, row);
-  transposed_row_col[i] = row_col;
-
-  atomicAdd(transposed_bsr_offsets + col + 1, 1);
-}
-
-template <int Rows, int Cols, typename T> struct BsrBlockTransposer {
-  void CUDA_CALLABLE_DEVICE operator()(const T *src, T *dest) const {
-    for (int r = 0; r < Rows; ++r) {
-      for (int c = 0; c < Cols; ++c) {
-        dest[c * Rows + r] = src[r * Cols + c];
-      }
-    }
-  }
-};
-
-template <typename T> struct BsrBlockTransposer<-1, -1, T> {
-
-  int row_count;
-  int col_count;
-
-  void CUDA_CALLABLE_DEVICE operator()(const T *src, T *dest) const {
-    for (int r = 0; r < row_count; ++r) {
-      for (int c = 0; c < col_count; ++c) {
-        dest[c * row_count + r] = src[r * col_count + c];
-      }
-    }
-  }
-};
-
-template <int Rows, int Cols, typename T>
-__global__ void
-bsr_transpose_blocks(const int nnz, const int block_size,
-                     BsrBlockTransposer<Rows, Cols, T> transposer,
-                     const int *block_indices,
-                     const BsrRowCol *transposed_indices, const T *bsr_values,
-                     int *transposed_bsr_columns, T *transposed_bsr_values) {
-  int i = blockIdx.x * blockDim.x + threadIdx.x;
-  if (i >= nnz)
-    return;
-
-  const int src_idx = block_indices[i];
-
-  transposer(bsr_values + src_idx * block_size,
-             transposed_bsr_values + i * block_size);
-
-  transposed_bsr_columns[i] = bsr_get_col(transposed_indices[i]);
-}
-
-template <typename T>
-void
-launch_bsr_transpose_blocks(const int nnz, const int block_size,
-                     const int rows_per_block, const int cols_per_block,
-                     const int *block_indices,
-                     const BsrRowCol *transposed_indices, 
-                     const T *bsr_values,
-                     int *transposed_bsr_columns, T *transposed_bsr_values) {
-
-  switch (rows_per_block) {
-  case 1:
-    switch (cols_per_block) {
-    case 1:
-      wp_launch_device(WP_CURRENT_CONTEXT, bsr_transpose_blocks, nnz,
-                       (nnz, block_size, BsrBlockTransposer<1, 1, T>{},
-                        block_indices, transposed_indices, bsr_values,
-                        transposed_bsr_columns, transposed_bsr_values));
-      return;
-    case 2:
-      wp_launch_device(WP_CURRENT_CONTEXT, bsr_transpose_blocks, nnz,
-                       (nnz, block_size, BsrBlockTransposer<1, 2, T>{},
-                        block_indices, transposed_indices, bsr_values,
-                        transposed_bsr_columns, transposed_bsr_values));
-      return;
-    case 3:
-      wp_launch_device(WP_CURRENT_CONTEXT, bsr_transpose_blocks, nnz,
-                       (nnz, block_size, BsrBlockTransposer<1, 3, T>{},
-                        block_indices, transposed_indices, bsr_values,
-                        transposed_bsr_columns, transposed_bsr_values));
-      return;
-    }
-  case 2:
-    switch (cols_per_block) {
-    case 1:
-      wp_launch_device(WP_CURRENT_CONTEXT, bsr_transpose_blocks, nnz,
-                       (nnz, block_size, BsrBlockTransposer<2, 1, T>{},
-                        block_indices, transposed_indices, bsr_values,
-                        transposed_bsr_columns, transposed_bsr_values));
-      return;
-    case 2:
-      wp_launch_device(WP_CURRENT_CONTEXT, bsr_transpose_blocks, nnz,
-                       (nnz, block_size, BsrBlockTransposer<2, 2, T>{},
-                        block_indices, transposed_indices, bsr_values,
-                        transposed_bsr_columns, transposed_bsr_values));
-      return;
-    case 3:
-      wp_launch_device(WP_CURRENT_CONTEXT, bsr_transpose_blocks, nnz,
-                       (nnz, block_size, BsrBlockTransposer<2, 3, T>{},
-                        block_indices, transposed_indices, bsr_values,
-                        transposed_bsr_columns, transposed_bsr_values));
-      return;
-    }
-  case 3:
-    switch (cols_per_block) {
-    case 1:
-      wp_launch_device(WP_CURRENT_CONTEXT, bsr_transpose_blocks, nnz,
-                       (nnz, block_size, BsrBlockTransposer<3, 1, T>{},
-                        block_indices, transposed_indices, bsr_values,
-                        transposed_bsr_columns, transposed_bsr_values));
-      return;
-    case 2:
-      wp_launch_device(WP_CURRENT_CONTEXT, bsr_transpose_blocks, nnz,
-                       (nnz, block_size, BsrBlockTransposer<3, 2, T>{},
-                        block_indices, transposed_indices, bsr_values,
-                        transposed_bsr_columns, transposed_bsr_values));
-      return;
-    case 3:
-      wp_launch_device(WP_CURRENT_CONTEXT, bsr_transpose_blocks, nnz,
-                       (nnz, block_size, BsrBlockTransposer<3, 3, T>{},
-                        block_indices, transposed_indices, bsr_values,
-                        transposed_bsr_columns, transposed_bsr_values));
-      return;
-    }
-  }
-
-  wp_launch_device(
-      WP_CURRENT_CONTEXT, bsr_transpose_blocks, nnz,
-      (nnz, block_size,
-       BsrBlockTransposer<-1, -1, T>{rows_per_block, cols_per_block},
-       block_indices, transposed_indices, bsr_values, transposed_bsr_columns,
-       transposed_bsr_values));
-}
-
-template <typename T>
-void bsr_transpose_device(int rows_per_block, int cols_per_block, int row_count,
-                          int col_count, int nnz, const int *bsr_offsets,
-                          const int *bsr_columns, const T *bsr_values,
-                          int *transposed_bsr_offsets,
-                          int *transposed_bsr_columns,
-                          T *transposed_bsr_values) {
-
-  const int block_size = rows_per_block * cols_per_block;
-
-  void *context = cuda_context_get_current();
-  ContextGuard guard(context);
-
-  cudaStream_t stream = static_cast<cudaStream_t>(cuda_stream_get_current());
-
-  // Zero the transposed offsets
-  memset_device(WP_CURRENT_CONTEXT, transposed_bsr_offsets, 0,
-                (col_count + 1) * sizeof(int));
-
-  ScopedTemporary<int> block_indices(context, 2*nnz);
-  ScopedTemporary<BsrRowCol> combined_row_col(context, 2*nnz);
-
-  cub::DoubleBuffer<int> d_keys(block_indices.buffer(),
-                                block_indices.buffer() + nnz);
-  cub::DoubleBuffer<BsrRowCol> d_values(combined_row_col.buffer(),
-                                        combined_row_col.buffer() + nnz);
-
-  wp_launch_device(WP_CURRENT_CONTEXT, bsr_transpose_fill_row_col, nnz,
-                   (nnz, row_count, bsr_offsets, bsr_columns, d_keys.Current(),
-                    d_values.Current(), transposed_bsr_offsets));
-
-  // Sort blocks
-  {
-    size_t buff_size = 0;
-    check_cuda(cub::DeviceRadixSort::SortPairs(nullptr, buff_size, d_values,
-                                              d_keys, nnz, 0, 64, stream));
-    ScopedTemporary<> temp(context, buff_size);
-    check_cuda(cub::DeviceRadixSort::SortPairs(
-        temp.buffer(), buff_size, d_values, d_keys, nnz, 0, 64, stream));
-  }
-
-  // Prefix sum the transposed row block counts
-  {
-    size_t buff_size = 0;
-    check_cuda(cub::DeviceScan::InclusiveSum(
-        nullptr, buff_size, transposed_bsr_offsets, transposed_bsr_offsets,
-        col_count + 1, stream));
-    ScopedTemporary<> temp(context, buff_size);
-    check_cuda(cub::DeviceScan::InclusiveSum(
-        temp.buffer(), buff_size, transposed_bsr_offsets,
-        transposed_bsr_offsets, col_count + 1, stream));
-  }
-
-  // Move and transpose individual blocks
-  launch_bsr_transpose_blocks(
-       nnz, block_size,
-       rows_per_block, cols_per_block,
-       d_keys.Current(), d_values.Current(), bsr_values, transposed_bsr_columns,
-       transposed_bsr_values);
-}
-
-} // namespace
-
-int bsr_matrix_from_triplets_float_device(
-    int rows_per_block, int cols_per_block, int row_count, int nnz,
-    uint64_t tpl_rows, uint64_t tpl_columns, uint64_t tpl_values,
-    uint64_t bsr_offsets, uint64_t bsr_columns, uint64_t bsr_values) {
-  return bsr_matrix_from_triplets_device<float>(
-      rows_per_block, cols_per_block, row_count, nnz,
-      reinterpret_cast<const int *>(tpl_rows),
-      reinterpret_cast<const int *>(tpl_columns),
-      reinterpret_cast<const float *>(tpl_values),
-      reinterpret_cast<int *>(bsr_offsets),
-      reinterpret_cast<int *>(bsr_columns),
-      reinterpret_cast<float *>(bsr_values));
-}
-
-int bsr_matrix_from_triplets_double_device(
-    int rows_per_block, int cols_per_block, int row_count, int nnz,
-    uint64_t tpl_rows, uint64_t tpl_columns, uint64_t tpl_values,
-    uint64_t bsr_offsets, uint64_t bsr_columns, uint64_t bsr_values) {
-  return bsr_matrix_from_triplets_device<double>(
-      rows_per_block, cols_per_block, row_count, nnz,
-      reinterpret_cast<const int *>(tpl_rows),
-      reinterpret_cast<const int *>(tpl_columns),
-      reinterpret_cast<const double *>(tpl_values),
-      reinterpret_cast<int *>(bsr_offsets),
-      reinterpret_cast<int *>(bsr_columns),
-      reinterpret_cast<double *>(bsr_values));
-}
-
-void bsr_transpose_float_device(int rows_per_block, int cols_per_block,
-                                int row_count, int col_count, int nnz,
-                                uint64_t bsr_offsets, uint64_t bsr_columns,
-                                uint64_t bsr_values,
-                                uint64_t transposed_bsr_offsets,
-                                uint64_t transposed_bsr_columns,
-                                uint64_t transposed_bsr_values) {
-  bsr_transpose_device(rows_per_block, cols_per_block, row_count, col_count,
-                       nnz, reinterpret_cast<const int *>(bsr_offsets),
-                       reinterpret_cast<const int *>(bsr_columns),
-                       reinterpret_cast<const float *>(bsr_values),
-                       reinterpret_cast<int *>(transposed_bsr_offsets),
-                       reinterpret_cast<int *>(transposed_bsr_columns),
-                       reinterpret_cast<float *>(transposed_bsr_values));
-}
-
-void bsr_transpose_double_device(int rows_per_block, int cols_per_block,
-                                 int row_count, int col_count, int nnz,
-                                 uint64_t bsr_offsets, uint64_t bsr_columns,
-                                 uint64_t bsr_values,
-                                 uint64_t transposed_bsr_offsets,
-                                 uint64_t transposed_bsr_columns,
-                                 uint64_t transposed_bsr_values) {
-  bsr_transpose_device(rows_per_block, cols_per_block, row_count, col_count,
-                       nnz, reinterpret_cast<const int *>(bsr_offsets),
-                       reinterpret_cast<const int *>(bsr_columns),
-                       reinterpret_cast<const double *>(bsr_values),
-                       reinterpret_cast<int *>(transposed_bsr_offsets),
-                       reinterpret_cast<int *>(transposed_bsr_columns),
-                       reinterpret_cast<double *>(transposed_bsr_values));
-}
+#include "cuda_util.h"
+#include "warp.h"
+
+#define THRUST_IGNORE_CUB_VERSION_CHECK
+
+#include <cub/device/device_radix_sort.cuh>
+#include <cub/device/device_run_length_encode.cuh>
+#include <cub/device/device_scan.cuh>
+#include <cub/device/device_select.cuh>
+
+namespace {
+
+// Combined row+column value that can be radix-sorted with CUB
+using BsrRowCol = uint64_t;
+
+CUDA_CALLABLE BsrRowCol bsr_combine_row_col(uint32_t row, uint32_t col) {
+  return (static_cast<uint64_t>(row) << 32) | col;
+}
+
+CUDA_CALLABLE uint32_t bsr_get_row(const BsrRowCol &row_col) {
+  return row_col >> 32;
+}
+
+CUDA_CALLABLE uint32_t bsr_get_col(const BsrRowCol &row_col) {
+  return row_col & INT_MAX;
+}
+
+// Cached temporary storage
+struct BsrFromTripletsTemp {
+  
+  int *count_buffer = NULL;
+  cudaEvent_t host_sync_event = NULL;
+
+  BsrFromTripletsTemp()
+    : count_buffer(static_cast<int*>(alloc_pinned(sizeof(int))))
+  {
+    cudaEventCreateWithFlags(&host_sync_event, cudaEventDisableTiming);
+  }
+  
+  ~BsrFromTripletsTemp()
+  {
+    cudaEventDestroy(host_sync_event);
+    free_pinned(count_buffer);
+  }
+
+  BsrFromTripletsTemp(const BsrFromTripletsTemp&) = delete;
+  BsrFromTripletsTemp& operator=(const BsrFromTripletsTemp&) = delete;
+
+};
+
+// map temp buffers to CUDA contexts
+static std::unordered_map<void *, BsrFromTripletsTemp> g_bsr_from_triplets_temp_map;
+
+template <typename T> struct BsrBlockIsNotZero {
+  int block_size;
+  const T *values;
+
+  CUDA_CALLABLE_DEVICE bool operator()(int i) const {
+    const T *val = values + i * block_size;
+    for (int i = 0; i < block_size; ++i, ++val) {
+      if (*val != T(0))
+        return true;
+    }
+    return false;
+  }
+};
+
+__global__ void bsr_fill_block_indices(int nnz, int *block_indices) {
+  int i = blockIdx.x * blockDim.x + threadIdx.x;
+  if (i >= nnz)
+    return;
+
+  block_indices[i] = i;
+}
+
+__global__ void bsr_fill_row_col(const int *nnz, const int *block_indices,
+                                 const int *tpl_rows, const int *tpl_columns,
+                                 BsrRowCol *tpl_row_col) {
+  int i = blockIdx.x * blockDim.x + threadIdx.x;
+  if (i >= *nnz)
+    return;
+
+  const int block = block_indices[i];
+
+  BsrRowCol row_col = bsr_combine_row_col(tpl_rows[block], tpl_columns[block]);
+  tpl_row_col[i] = row_col;
+}
+
+template <typename T>
+__global__ void
+bsr_merge_blocks(int nnz, int block_size, const int *block_offsets,
+                 const int *sorted_block_indices,
+                 const BsrRowCol *unique_row_cols, const T *tpl_values,
+                 int *bsr_row_counts, int *bsr_cols, T *bsr_values)
+
+{
+  int i = blockIdx.x * blockDim.x + threadIdx.x;
+  if (i >= nnz)
+    return;
+
+  const int beg = i ? block_offsets[i - 1] : 0;
+  const int end = block_offsets[i];
+
+  BsrRowCol row_col = unique_row_cols[i];
+
+  bsr_cols[i] = bsr_get_col(row_col);
+  atomicAdd(bsr_row_counts + bsr_get_row(row_col) + 1, 1);
+
+  if (bsr_values == nullptr)
+    return;
+
+  T *bsr_val = bsr_values + i * block_size;
+  const T *tpl_val = tpl_values + sorted_block_indices[beg] * block_size;
+
+  for (int k = 0; k < block_size; ++k) {
+    bsr_val[k] = tpl_val[k];
+  }
+
+  for (int cur = beg + 1; cur != end; ++cur) {
+    const T *tpl_val = tpl_values + sorted_block_indices[cur] * block_size;
+    for (int k = 0; k < block_size; ++k) {
+      bsr_val[k] += tpl_val[k];
+    }
+  }
+}
+
+template <typename T>
+int bsr_matrix_from_triplets_device(const int rows_per_block,
+                                    const int cols_per_block,
+                                    const int row_count, const int nnz,
+                                    const int *tpl_rows, const int *tpl_columns,
+                                    const T *tpl_values, int *bsr_offsets,
+                                    int *bsr_columns, T *bsr_values) {
+  const int block_size = rows_per_block * cols_per_block;
+
+  void *context = cuda_context_get_current();
+  ContextGuard guard(context);
+
+  // Per-context cached temporary buffers
+  BsrFromTripletsTemp &bsr_temp = g_bsr_from_triplets_temp_map[context];
+
+  cudaStream_t stream = static_cast<cudaStream_t>(cuda_stream_get_current());
+
+  ScopedTemporary<int> block_indices(context, 2*nnz);
+  ScopedTemporary<BsrRowCol> combined_row_col(context, 2*nnz);
+
+  cub::DoubleBuffer<int> d_keys(block_indices.buffer(),
+                                block_indices.buffer() + nnz);
+  cub::DoubleBuffer<BsrRowCol> d_values(combined_row_col.buffer(),
+                                        combined_row_col.buffer() + nnz);
+
+  int *p_nz_triplet_count = bsr_temp.count_buffer;
+
+  wp_launch_device(WP_CURRENT_CONTEXT, bsr_fill_block_indices, nnz,
+                   (nnz, d_keys.Current()));
+
+  if (tpl_values) {
+
+    // Remove zero blocks
+    {
+      size_t buff_size = 0;
+      BsrBlockIsNotZero<T> isNotZero{block_size, tpl_values};
+      check_cuda(cub::DeviceSelect::If(nullptr, buff_size, d_keys.Current(),
+                                      d_keys.Alternate(), p_nz_triplet_count,
+                                      nnz, isNotZero, stream));
+      ScopedTemporary<> temp(context, buff_size);
+      check_cuda(cub::DeviceSelect::If(
+          temp.buffer(), buff_size, d_keys.Current(), d_keys.Alternate(),
+          p_nz_triplet_count, nnz, isNotZero, stream));
+    }
+    cudaEventRecord(bsr_temp.host_sync_event, stream);
+
+    // switch current/alternate in double buffer
+    d_keys.selector ^= 1;
+
+  } else {
+    *p_nz_triplet_count = nnz;
+  }
+
+  // Combine rows and columns so we can sort on them both
+  wp_launch_device(WP_CURRENT_CONTEXT, bsr_fill_row_col, nnz,
+                   (p_nz_triplet_count, d_keys.Current(), tpl_rows, tpl_columns,
+                    d_values.Current()));
+
+  if (tpl_values) {
+    // Make sure count is available on host
+    cudaEventSynchronize(bsr_temp.host_sync_event);
+  }
+
+  const int nz_triplet_count = *p_nz_triplet_count;
+
+  // Sort
+  {
+    size_t buff_size = 0;
+    check_cuda(cub::DeviceRadixSort::SortPairs(
+        nullptr, buff_size, d_values, d_keys, nz_triplet_count, 0, 64, stream));
+    ScopedTemporary<> temp(context, buff_size);
+    check_cuda(cub::DeviceRadixSort::SortPairs(temp.buffer(), buff_size,
+                                              d_values, d_keys, nz_triplet_count,
+                                              0, 64, stream));
+  }
+
+  // Runlength encode row-col sequences
+  {
+    size_t buff_size = 0;
+    check_cuda(cub::DeviceRunLengthEncode::Encode(
+        nullptr, buff_size, d_values.Current(), d_values.Alternate(),
+        d_keys.Alternate(), p_nz_triplet_count, nz_triplet_count, stream));
+    ScopedTemporary<> temp(context, buff_size);
+    check_cuda(cub::DeviceRunLengthEncode::Encode(
+        temp.buffer(), buff_size, d_values.Current(), d_values.Alternate(),
+        d_keys.Alternate(), p_nz_triplet_count, nz_triplet_count, stream));
+  }
+
+  cudaEventRecord(bsr_temp.host_sync_event, stream);
+
+  // Now we have the following:
+  // d_values.Current(): sorted block row-col
+  // d_values.Alternate(): sorted unique row-col
+  // d_keys.Current(): sorted block indices
+  // d_keys.Alternate(): repeated block-row count
+
+  // Scan repeated block counts
+  {
+    size_t buff_size = 0;
+    check_cuda(cub::DeviceScan::InclusiveSum(
+        nullptr, buff_size, d_keys.Alternate(), d_keys.Alternate(),
+        nz_triplet_count, stream));
+    ScopedTemporary<> temp(context, buff_size);
+    check_cuda(cub::DeviceScan::InclusiveSum(
+        temp.buffer(), buff_size, d_keys.Alternate(), d_keys.Alternate(),
+        nz_triplet_count, stream));
+  }
+
+  // While we're at it, zero the bsr offsets buffer
+  memset_device(WP_CURRENT_CONTEXT, bsr_offsets, 0,
+                (row_count + 1) * sizeof(int));
+
+  // Wait for number of compressed blocks
+  cudaEventSynchronize(bsr_temp.host_sync_event);
+  const int compressed_nnz = *p_nz_triplet_count;
+
+  // We have all we need to accumulate our repeated blocks
+  wp_launch_device(WP_CURRENT_CONTEXT, bsr_merge_blocks, compressed_nnz,
+                   (compressed_nnz, block_size, d_keys.Alternate(),
+                    d_keys.Current(), d_values.Alternate(), tpl_values,
+                    bsr_offsets, bsr_columns, bsr_values));
+
+  // Last, prefix sum the row block counts
+  {
+    size_t buff_size = 0;
+    check_cuda(cub::DeviceScan::InclusiveSum(nullptr, buff_size, bsr_offsets,
+                                            bsr_offsets, row_count + 1, stream));
+    ScopedTemporary<> temp(context, buff_size);
+    check_cuda(cub::DeviceScan::InclusiveSum(temp.buffer(), buff_size,
+                                            bsr_offsets, bsr_offsets,
+                                            row_count + 1, stream));
+  }
+
+  return compressed_nnz;
+}
+
+__global__ void bsr_transpose_fill_row_col(const int nnz, const int row_count,
+                                           const int *bsr_offsets,
+                                           const int *bsr_columns,
+                                           int *block_indices,
+                                           BsrRowCol *transposed_row_col,
+                                           int *transposed_bsr_offsets) {
+  int i = blockIdx.x * blockDim.x + threadIdx.x;
+  if (i >= nnz)
+    return;
+
+  block_indices[i] = i;
+
+  // Binary search for row
+  int lower = 0;
+  int upper = row_count - 1;
+
+  while (lower < upper) {
+    int mid = lower + (upper - lower) / 2;
+
+    if (bsr_offsets[mid + 1] <= i) {
+      lower = mid + 1;
+    } else {
+      upper = mid;
+    }
+  }
+
+  const int row = lower;
+  const int col = bsr_columns[i];
+  BsrRowCol row_col = bsr_combine_row_col(col, row);
+  transposed_row_col[i] = row_col;
+
+  atomicAdd(transposed_bsr_offsets + col + 1, 1);
+}
+
+template <int Rows, int Cols, typename T> struct BsrBlockTransposer {
+  void CUDA_CALLABLE_DEVICE operator()(const T *src, T *dest) const {
+    for (int r = 0; r < Rows; ++r) {
+      for (int c = 0; c < Cols; ++c) {
+        dest[c * Rows + r] = src[r * Cols + c];
+      }
+    }
+  }
+};
+
+template <typename T> struct BsrBlockTransposer<-1, -1, T> {
+
+  int row_count;
+  int col_count;
+
+  void CUDA_CALLABLE_DEVICE operator()(const T *src, T *dest) const {
+    for (int r = 0; r < row_count; ++r) {
+      for (int c = 0; c < col_count; ++c) {
+        dest[c * row_count + r] = src[r * col_count + c];
+      }
+    }
+  }
+};
+
+template <int Rows, int Cols, typename T>
+__global__ void
+bsr_transpose_blocks(const int nnz, const int block_size,
+                     BsrBlockTransposer<Rows, Cols, T> transposer,
+                     const int *block_indices,
+                     const BsrRowCol *transposed_indices, const T *bsr_values,
+                     int *transposed_bsr_columns, T *transposed_bsr_values) {
+  int i = blockIdx.x * blockDim.x + threadIdx.x;
+  if (i >= nnz)
+    return;
+
+  const int src_idx = block_indices[i];
+
+  transposer(bsr_values + src_idx * block_size,
+             transposed_bsr_values + i * block_size);
+
+  transposed_bsr_columns[i] = bsr_get_col(transposed_indices[i]);
+}
+
+template <typename T>
+void
+launch_bsr_transpose_blocks(const int nnz, const int block_size,
+                     const int rows_per_block, const int cols_per_block,
+                     const int *block_indices,
+                     const BsrRowCol *transposed_indices, 
+                     const T *bsr_values,
+                     int *transposed_bsr_columns, T *transposed_bsr_values) {
+
+  switch (rows_per_block) {
+  case 1:
+    switch (cols_per_block) {
+    case 1:
+      wp_launch_device(WP_CURRENT_CONTEXT, bsr_transpose_blocks, nnz,
+                       (nnz, block_size, BsrBlockTransposer<1, 1, T>{},
+                        block_indices, transposed_indices, bsr_values,
+                        transposed_bsr_columns, transposed_bsr_values));
+      return;
+    case 2:
+      wp_launch_device(WP_CURRENT_CONTEXT, bsr_transpose_blocks, nnz,
+                       (nnz, block_size, BsrBlockTransposer<1, 2, T>{},
+                        block_indices, transposed_indices, bsr_values,
+                        transposed_bsr_columns, transposed_bsr_values));
+      return;
+    case 3:
+      wp_launch_device(WP_CURRENT_CONTEXT, bsr_transpose_blocks, nnz,
+                       (nnz, block_size, BsrBlockTransposer<1, 3, T>{},
+                        block_indices, transposed_indices, bsr_values,
+                        transposed_bsr_columns, transposed_bsr_values));
+      return;
+    }
+  case 2:
+    switch (cols_per_block) {
+    case 1:
+      wp_launch_device(WP_CURRENT_CONTEXT, bsr_transpose_blocks, nnz,
+                       (nnz, block_size, BsrBlockTransposer<2, 1, T>{},
+                        block_indices, transposed_indices, bsr_values,
+                        transposed_bsr_columns, transposed_bsr_values));
+      return;
+    case 2:
+      wp_launch_device(WP_CURRENT_CONTEXT, bsr_transpose_blocks, nnz,
+                       (nnz, block_size, BsrBlockTransposer<2, 2, T>{},
+                        block_indices, transposed_indices, bsr_values,
+                        transposed_bsr_columns, transposed_bsr_values));
+      return;
+    case 3:
+      wp_launch_device(WP_CURRENT_CONTEXT, bsr_transpose_blocks, nnz,
+                       (nnz, block_size, BsrBlockTransposer<2, 3, T>{},
+                        block_indices, transposed_indices, bsr_values,
+                        transposed_bsr_columns, transposed_bsr_values));
+      return;
+    }
+  case 3:
+    switch (cols_per_block) {
+    case 1:
+      wp_launch_device(WP_CURRENT_CONTEXT, bsr_transpose_blocks, nnz,
+                       (nnz, block_size, BsrBlockTransposer<3, 1, T>{},
+                        block_indices, transposed_indices, bsr_values,
+                        transposed_bsr_columns, transposed_bsr_values));
+      return;
+    case 2:
+      wp_launch_device(WP_CURRENT_CONTEXT, bsr_transpose_blocks, nnz,
+                       (nnz, block_size, BsrBlockTransposer<3, 2, T>{},
+                        block_indices, transposed_indices, bsr_values,
+                        transposed_bsr_columns, transposed_bsr_values));
+      return;
+    case 3:
+      wp_launch_device(WP_CURRENT_CONTEXT, bsr_transpose_blocks, nnz,
+                       (nnz, block_size, BsrBlockTransposer<3, 3, T>{},
+                        block_indices, transposed_indices, bsr_values,
+                        transposed_bsr_columns, transposed_bsr_values));
+      return;
+    }
+  }
+
+  wp_launch_device(
+      WP_CURRENT_CONTEXT, bsr_transpose_blocks, nnz,
+      (nnz, block_size,
+       BsrBlockTransposer<-1, -1, T>{rows_per_block, cols_per_block},
+       block_indices, transposed_indices, bsr_values, transposed_bsr_columns,
+       transposed_bsr_values));
+}
+
+template <typename T>
+void bsr_transpose_device(int rows_per_block, int cols_per_block, int row_count,
+                          int col_count, int nnz, const int *bsr_offsets,
+                          const int *bsr_columns, const T *bsr_values,
+                          int *transposed_bsr_offsets,
+                          int *transposed_bsr_columns,
+                          T *transposed_bsr_values) {
+
+  const int block_size = rows_per_block * cols_per_block;
+
+  void *context = cuda_context_get_current();
+  ContextGuard guard(context);
+
+  cudaStream_t stream = static_cast<cudaStream_t>(cuda_stream_get_current());
+
+  // Zero the transposed offsets
+  memset_device(WP_CURRENT_CONTEXT, transposed_bsr_offsets, 0,
+                (col_count + 1) * sizeof(int));
+
+  ScopedTemporary<int> block_indices(context, 2*nnz);
+  ScopedTemporary<BsrRowCol> combined_row_col(context, 2*nnz);
+
+  cub::DoubleBuffer<int> d_keys(block_indices.buffer(),
+                                block_indices.buffer() + nnz);
+  cub::DoubleBuffer<BsrRowCol> d_values(combined_row_col.buffer(),
+                                        combined_row_col.buffer() + nnz);
+
+  wp_launch_device(WP_CURRENT_CONTEXT, bsr_transpose_fill_row_col, nnz,
+                   (nnz, row_count, bsr_offsets, bsr_columns, d_keys.Current(),
+                    d_values.Current(), transposed_bsr_offsets));
+
+  // Sort blocks
+  {
+    size_t buff_size = 0;
+    check_cuda(cub::DeviceRadixSort::SortPairs(nullptr, buff_size, d_values,
+                                              d_keys, nnz, 0, 64, stream));
+    ScopedTemporary<> temp(context, buff_size);
+    check_cuda(cub::DeviceRadixSort::SortPairs(
+        temp.buffer(), buff_size, d_values, d_keys, nnz, 0, 64, stream));
+  }
+
+  // Prefix sum the transposed row block counts
+  {
+    size_t buff_size = 0;
+    check_cuda(cub::DeviceScan::InclusiveSum(
+        nullptr, buff_size, transposed_bsr_offsets, transposed_bsr_offsets,
+        col_count + 1, stream));
+    ScopedTemporary<> temp(context, buff_size);
+    check_cuda(cub::DeviceScan::InclusiveSum(
+        temp.buffer(), buff_size, transposed_bsr_offsets,
+        transposed_bsr_offsets, col_count + 1, stream));
+  }
+
+  // Move and transpose individual blocks
+  launch_bsr_transpose_blocks(
+       nnz, block_size,
+       rows_per_block, cols_per_block,
+       d_keys.Current(), d_values.Current(), bsr_values, transposed_bsr_columns,
+       transposed_bsr_values);
+}
+
+} // namespace
+
+int bsr_matrix_from_triplets_float_device(
+    int rows_per_block, int cols_per_block, int row_count, int nnz,
+    uint64_t tpl_rows, uint64_t tpl_columns, uint64_t tpl_values,
+    uint64_t bsr_offsets, uint64_t bsr_columns, uint64_t bsr_values) {
+  return bsr_matrix_from_triplets_device<float>(
+      rows_per_block, cols_per_block, row_count, nnz,
+      reinterpret_cast<const int *>(tpl_rows),
+      reinterpret_cast<const int *>(tpl_columns),
+      reinterpret_cast<const float *>(tpl_values),
+      reinterpret_cast<int *>(bsr_offsets),
+      reinterpret_cast<int *>(bsr_columns),
+      reinterpret_cast<float *>(bsr_values));
+}
+
+int bsr_matrix_from_triplets_double_device(
+    int rows_per_block, int cols_per_block, int row_count, int nnz,
+    uint64_t tpl_rows, uint64_t tpl_columns, uint64_t tpl_values,
+    uint64_t bsr_offsets, uint64_t bsr_columns, uint64_t bsr_values) {
+  return bsr_matrix_from_triplets_device<double>(
+      rows_per_block, cols_per_block, row_count, nnz,
+      reinterpret_cast<const int *>(tpl_rows),
+      reinterpret_cast<const int *>(tpl_columns),
+      reinterpret_cast<const double *>(tpl_values),
+      reinterpret_cast<int *>(bsr_offsets),
+      reinterpret_cast<int *>(bsr_columns),
+      reinterpret_cast<double *>(bsr_values));
+}
+
+void bsr_transpose_float_device(int rows_per_block, int cols_per_block,
+                                int row_count, int col_count, int nnz,
+                                uint64_t bsr_offsets, uint64_t bsr_columns,
+                                uint64_t bsr_values,
+                                uint64_t transposed_bsr_offsets,
+                                uint64_t transposed_bsr_columns,
+                                uint64_t transposed_bsr_values) {
+  bsr_transpose_device(rows_per_block, cols_per_block, row_count, col_count,
+                       nnz, reinterpret_cast<const int *>(bsr_offsets),
+                       reinterpret_cast<const int *>(bsr_columns),
+                       reinterpret_cast<const float *>(bsr_values),
+                       reinterpret_cast<int *>(transposed_bsr_offsets),
+                       reinterpret_cast<int *>(transposed_bsr_columns),
+                       reinterpret_cast<float *>(transposed_bsr_values));
+}
+
+void bsr_transpose_double_device(int rows_per_block, int cols_per_block,
+                                 int row_count, int col_count, int nnz,
+                                 uint64_t bsr_offsets, uint64_t bsr_columns,
+                                 uint64_t bsr_values,
+                                 uint64_t transposed_bsr_offsets,
+                                 uint64_t transposed_bsr_columns,
+                                 uint64_t transposed_bsr_values) {
+  bsr_transpose_device(rows_per_block, cols_per_block, row_count, col_count,
+                       nnz, reinterpret_cast<const int *>(bsr_offsets),
+                       reinterpret_cast<const int *>(bsr_columns),
+                       reinterpret_cast<const double *>(bsr_values),
+                       reinterpret_cast<int *>(transposed_bsr_offsets),
+                       reinterpret_cast<int *>(transposed_bsr_columns),
+                       reinterpret_cast<double *>(transposed_bsr_values));
+}