histogram64.cu

/* Copyright (c) 2021, NVIDIA CORPORATION. All rights reserved.
 *
 * Redistribution and use in source and binary forms, with or without
 * modification, are permitted provided that the following conditions
 * are met:
 *  * Redistributions of source code must retain the above copyright
 *    notice, this list of conditions and the following disclaimer.
 *  * Redistributions in binary form must reproduce the above copyright
 *    notice, this list of conditions and the following disclaimer in the
 *    documentation and/or other materials provided with the distribution.
 *  * Neither the name of NVIDIA CORPORATION nor the names of its
 *    contributors may be used to endorse or promote products derived
 *    from this software without specific prior written permission.
 *
 * THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS ``AS IS'' AND ANY
 * EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
 * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR
 * PURPOSE ARE DISCLAIMED.  IN NO EVENT SHALL THE COPYRIGHT OWNER OR
 * CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL,
 * EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT LIMITED TO,
 * PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE, DATA, OR
 * PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY THEORY
 * OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT
 * (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE
 * OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
 */

#include <assert.h>
#include <stdio.h>
#include <stdlib.h>
#include <string.h>
#include <cooperative_groups.h>

namespace cg = cooperative_groups;
#include <helper_cuda.h>
#include "histogram_common.h"

////////////////////////////////////////////////////////////////////////////////
// GPU-specific common definitions
////////////////////////////////////////////////////////////////////////////////
// Data type used for input data fetches
typedef uint4 data_t;

// May change on future hardware, so better parametrize the code
#define SHARED_MEMORY_BANKS 16

////////////////////////////////////////////////////////////////////////////////
// Main computation pass: compute gridDim.x partial histograms
////////////////////////////////////////////////////////////////////////////////
// Count a byte into shared-memory storage
inline __device__ void addByte(uchar *s_ThreadBase, uint data) {
  s_ThreadBase[UMUL(data, HISTOGRAM64_THREADBLOCK_SIZE)]++;
}

// Count four bytes of a word
inline __device__ void addWord(uchar *s_ThreadBase, uint data) {
  // Only higher 6 bits of each byte matter, as this is a 64-bin histogram
  addByte(s_ThreadBase, (data >> 2) & 0x3FU);
  addByte(s_ThreadBase, (data >> 10) & 0x3FU);
  addByte(s_ThreadBase, (data >> 18) & 0x3FU);
  addByte(s_ThreadBase, (data >> 26) & 0x3FU);
}

__global__ void histogram64Kernel(uint *d_PartialHistograms, data_t *d_Data,
                                  uint dataCount) {
  // Handle to thread block group
  cg::thread_block cta = cg::this_thread_block();
  // Encode thread index in order to avoid bank conflicts in s_Hist[] access:
  // each group of SHARED_MEMORY_BANKS threads accesses consecutive shared
  // memory banks
  // and the same bytes [0..3] within the banks
  // Because of this permutation block size should be a multiple of 4 *
  // SHARED_MEMORY_BANKS
  const uint threadPos = ((threadIdx.x & ~(SHARED_MEMORY_BANKS * 4 - 1)) << 0) |
                         ((threadIdx.x & (SHARED_MEMORY_BANKS - 1)) << 2) |
                         ((threadIdx.x & (SHARED_MEMORY_BANKS * 3)) >> 4);

  // Per-thread histogram storage
  __shared__ uchar s_Hist[HISTOGRAM64_THREADBLOCK_SIZE * HISTOGRAM64_BIN_COUNT];
  uchar *s_ThreadBase = s_Hist + threadPos;

// Initialize shared memory (writing 32-bit words)
#pragma unroll

  for (uint i = 0; i < (HISTOGRAM64_BIN_COUNT / 4); i++) {
    ((uint *)s_Hist)[threadIdx.x + i * HISTOGRAM64_THREADBLOCK_SIZE] = 0;
  }

  // Read data from global memory and submit to the shared-memory histogram
  // Since histogram counters are byte-sized, every single thread can't do more
  // than 255 submission
  cg::sync(cta);

  for (uint pos = UMAD(blockIdx.x, blockDim.x, threadIdx.x); pos < dataCount;
       pos += UMUL(blockDim.x, gridDim.x)) {
    data_t data = d_Data[pos];
    addWord(s_ThreadBase, data.x);
    addWord(s_ThreadBase, data.y);
    addWord(s_ThreadBase, data.z);
    addWord(s_ThreadBase, data.w);
  }

  // Accumulate per-thread histograms into per-block and write to global memory
  cg::sync(cta);

  if (threadIdx.x < HISTOGRAM64_BIN_COUNT) {
    uchar *s_HistBase =
        s_Hist + UMUL(threadIdx.x, HISTOGRAM64_THREADBLOCK_SIZE);

    uint sum = 0;
    uint pos = 4 * (threadIdx.x & (SHARED_MEMORY_BANKS - 1));

#pragma unroll

    for (uint i = 0; i < (HISTOGRAM64_THREADBLOCK_SIZE / 4); i++) {
      sum += s_HistBase[pos + 0] + s_HistBase[pos + 1] + s_HistBase[pos + 2] +
             s_HistBase[pos + 3];
      pos = (pos + 4) & (HISTOGRAM64_THREADBLOCK_SIZE - 1);
    }

    d_PartialHistograms[blockIdx.x * HISTOGRAM64_BIN_COUNT + threadIdx.x] = sum;
  }
}

////////////////////////////////////////////////////////////////////////////////
// Merge histogram64() output
// Run one threadblock per bin; each threadbock adds up the same bin counter
// from every partial histogram. Reads are uncoalesced, but mergeHistogram64
// takes only a fraction of total processing time
////////////////////////////////////////////////////////////////////////////////
#define MERGE_THREADBLOCK_SIZE 256

__global__ void mergeHistogram64Kernel(uint *d_Histogram,
                                       uint *d_PartialHistograms,
                                       uint histogramCount) {
  // Handle to thread block group
  cg::thread_block cta = cg::this_thread_block();
  __shared__ uint data[MERGE_THREADBLOCK_SIZE];

  uint sum = 0;

  for (uint i = threadIdx.x; i < histogramCount; i += MERGE_THREADBLOCK_SIZE) {
    sum += d_PartialHistograms[blockIdx.x + i * HISTOGRAM64_BIN_COUNT];
  }

  data[threadIdx.x] = sum;

  for (uint stride = MERGE_THREADBLOCK_SIZE / 2; stride > 0; stride >>= 1) {
    cg::sync(cta);

    if (threadIdx.x < stride) {
      data[threadIdx.x] += data[threadIdx.x + stride];
    }
  }

  if (threadIdx.x == 0) {
    d_Histogram[blockIdx.x] = data[0];
  }
}

////////////////////////////////////////////////////////////////////////////////
// CPU interface to GPU histogram calculator
////////////////////////////////////////////////////////////////////////////////
// histogram64kernel() intermediate results buffer
// MAX_PARTIAL_HISTOGRAM64_COUNT == 32768 and HISTOGRAM64_THREADBLOCK_SIZE == 64
// amounts to max. 480MB of input data
static const uint MAX_PARTIAL_HISTOGRAM64_COUNT = 32768;
static uint *d_PartialHistograms;

// Internal memory allocation
extern "C" void initHistogram64(void) {
  assert(HISTOGRAM64_THREADBLOCK_SIZE % (4 * SHARED_MEMORY_BANKS) == 0);
  checkCudaErrors(cudaMalloc(
      (void **)&d_PartialHistograms,
      MAX_PARTIAL_HISTOGRAM64_COUNT * HISTOGRAM64_BIN_COUNT * sizeof(uint)));
}

// Internal memory deallocation
extern "C" void closeHistogram64(void) {
  checkCudaErrors(cudaFree(d_PartialHistograms));
}

// Round a / b to nearest higher integer value
inline uint iDivUp(uint a, uint b) {
  return (a % b != 0) ? (a / b + 1) : (a / b);
}

// Snap a to nearest lower multiple of b
inline uint iSnapDown(uint a, uint b) { return a - a % b; }

extern "C" void histogram64(uint *d_Histogram, void *d_Data, uint byteCount) {
  const uint histogramCount = iDivUp(
      byteCount, HISTOGRAM64_THREADBLOCK_SIZE * iSnapDown(255, sizeof(data_t)));

  assert(byteCount % sizeof(data_t) == 0);
  assert(histogramCount <= MAX_PARTIAL_HISTOGRAM64_COUNT);

  histogram64Kernel<<<histogramCount, HISTOGRAM64_THREADBLOCK_SIZE>>>(
      d_PartialHistograms, (data_t *)d_Data, byteCount / sizeof(data_t));
  getLastCudaError("histogram64Kernel() execution failed\n");

  mergeHistogram64Kernel<<<HISTOGRAM64_BIN_COUNT, MERGE_THREADBLOCK_SIZE>>>(
      d_Histogram, d_PartialHistograms, histogramCount);
  getLastCudaError("mergeHistogram64() execution failed\n");
}