GraphicsMagick 的 OpenCL 开发记录（三十六）

最新推荐文章于 2025-12-01 13:12:33 发布

原创

最新推荐文章于 2025-12-01 13:12:33 发布 · 735 阅读

7 ·

CC 4.0 BY-SA版权

文章标签：

#ImageMagick #OpenCL #c++ #GraphicsMagick

本文详细描述了如何使用OpenCL在硬件上优化ScaleImage函数进行图像缩放，包括Y轴处理和工作流程。然而，作者提到在将代码从Linux移植到Windows时遇到问题，涉及到跨平台兼容性挑战和输出显示错误。

文章目录

如何写`ScaleImage()`的硬件加速函数（十）

<2022-05-05 周四>

如何写`ScaleImage()`的硬件加速函数（十）

难道就这么被我轻松实现了？

“如何写ScaleImage()的硬件加速函数（九）”是在“如何写ScaleImage()的硬件加速函数（八）”的基础上处理了图片放大变亮的问题，但是他们都只是X方向的处理，没有实现原始函数ScaleImage()的Y方向缩放。

目前先处理Y方向再处理X方向的代码都有了，如下：

static MagickBooleanType scaleFilter(MagickCLDevice device,
  cl_command_queue queue, const Image* image, Image* filteredImage,
  cl_mem imageBuffer, cl_uint matte_or_cmyk, cl_uint columns, cl_uint rows,
  cl_mem scaledImageBuffer, cl_uint scaledColumns, cl_uint scaledRows,
  ExceptionInfo* exception)
{
   
   
  cl_kernel
    scaleKernel;

  cl_int
    status;

  const unsigned int
    workgroupSize = 256;

  float
    scale;

  int
    numCachedPixels;

  MagickBooleanType
    outputReady;

  size_t
    gammaAccumulatorLocalMemorySize,
    gsize[2],
    i,
    imageCacheLocalMemorySize,
    pixelAccumulatorLocalMemorySize,
    pixelAccumulatorLocalMemorySize2,
    lsize[2],
    totalLocalMemorySize,
    weightAccumulatorLocalMemorySize;

  unsigned int
    chunkSize,
    pixelPerWorkgroup;

  scaleKernel = NULL;
  outputReady = MagickFalse;

  scale = (float)scaledColumns / columns; // TODO(ocl)

  unsigned int stop = 0;
  unsigned int next_row = 1;
  float y_span = 1.0;
  float y_scale = (float)scaledRows / rows;
  if (scaledRows == rows)
    stop++;
  else {
   
   
    while (y_scale < y_span) {
   
   
      if (next_row) {
   
   
        stop++;
      }
      y_span -= y_scale;
      y_scale = (float)scaledRows / rows;
      next_row = 1;
    }

    if (next_row) {
   
   
      stop++;
      next_row = 0;
    }
  }

  if (scaledColumns < workgroupSize)
  {
   
   
    chunkSize = 32;
    pixelPerWorkgroup = 32;
  }
  else
  {
   
   
    chunkSize = workgroupSize;
    pixelPerWorkgroup = workgroupSize;
  }

  DisableMSCWarning(4127)
    while (1)
      RestoreMSCWarning
    {
   
   
      /* calculate the local memory size needed per workgroup */
      numCachedPixels=(int) ceil((pixelPerWorkgroup-1)/scale+2*(0.5+MagickEpsilon)); // TODO(ocl)
      imageCacheLocalMemorySize = numCachedPixels * sizeof(CLQuantum) * 4 * stop;
      totalLocalMemorySize = imageCacheLocalMemorySize;

      /* local size for the pixel accumulator */
      pixelAccumulatorLocalMemorySize = chunkSize * sizeof(cl_float4);
      totalLocalMemorySize += pixelAccumulatorLocalMemorySize;

      pixelAccumulatorLocalMemorySize2 = pixelAccumulatorLocalMemorySize;
      totalLocalMemorySize += pixelAccumulatorLocalMemorySize2;

      /* local memory size for the weight accumulator */
      weightAccumulatorLocalMemorySize = chunkSize * sizeof(float);
      totalLocalMemorySize += weightAccumulatorLocalMemorySize;

      /* local memory size for the gamma accumulator */
      gammaAccumulatorLocalMemorySize = chunkSize * sizeof(float);
      totalLocalMemorySize += gammaAccumulatorLocalMemorySize;

      if (totalLocalMemorySize <= device->local_memory_size)
        break;
      else
      {
   
   
        pixelPerWorkgroup = pixelPerWorkgroup / 2;
        chunkSize = chunkSize / 2;
        if ((pixelPerWorkgroup == 0) || (chunkSize == 0))
        {
   
   
          /* quit, fallback to CPU */
          goto cleanup;
        }
      }
    }

  scaleKernel = AcquireOpenCLKernel(device, "ScaleFilter");
  if (scaleKernel == (cl_kernel)NULL)
  {
   
   
    (void)OpenCLThrowMagickException(device, exception, GetMagickModule(),
      ResourceLimitWarning, "AcquireOpenCLKernel failed.", ".");
    goto cleanup;
  }

  i = 0;
  status = SetOpenCLKernelArg(scaleKernel, i++, sizeof(cl_mem), (void*)&imageBuffer);
  status |= SetOpenCLKernelArg(scaleKernel, i++, sizeof(cl_uint