GraphicsMagick 的 OpenCL 开发记录（三十二）

<2022-04-27 周三>

如何写ScaleImage()的硬件加速函数（六）

不管什么事儿看来都怕琢磨，如果连做梦都能梦到你正在琢磨的事儿，估计离成功也就不远了。似乎目前已经达到了最好的效果，离目标越来越近了。

1. 要理解clEnqueueNDRangeKernel()函数的第五第六个参数意义，但目前为止只能说暂时理解了。

cl_int clEnqueueNDRangeKernel (cl_command_queue command_queue,
  cl_kernel kernel,
  cl_uint work_dim,
  const size_t *global_work_offset,
  const size_t *global_work_size,
  const size_t *local_work_size,
  cl_uint num_events_in_wait_list,
  const cl_event *event_wait_list,
  cl_event *event)

2. 第五第六个参数要结合__attribute__，否则无法调用kernel函数。

__kernel __attribute__((reqd_work_group_size(256, 1, 1)))

3. 来回理解ResizeImage()和ScaleImage()函数的实现，对它们的内部逻辑流程了如指掌了可以说。
4. 我选择以ResizeHorizontalFilter()为模板修改，之所以ResizeImage()处理速度慢，因为它的处理效果好，且有多种过滤效果可供选择，ResizeHorizontalFilter()是处理水平方向缩放，所以它将读入一整行原图像素，这正好和ScaleImage()的最内层循环处理方式相同。
5. ResizeHorizontalFilter()的最内层循环（如下），因为有累加操作，所以这正是处理水平缩放的操作。

for (unsigned int i = startStep; i < stopStep; i++, cacheIndex++)
{
  /* float weight = getResizeFilterWeight(resizeFilterCubicCoefficients,
    (ResizeWeightingFunctionType) resizeFilterType,
    (ResizeWeightingFunctionType) resizeWindowType,
    resizeFilterScale, resizeFilterWindowSupport,
    resizeFilterBlur, scale*(start + i - bisect + 0.5)); */

  float weight = getResizeFilterWeightForGM(resizeFilterType,
    scale*(start + i - bisect + 0.5), support);

  float4 cp = (float4)0.0f;

  __local CLQuantum *p = inputImageCache + (cacheIndex*4);
  cp.x = (float) *(p);
  cp.y = (float) *(p + 1);
  cp.z = (float) *(p + 2);

  if (matte_or_cmyk != 0)
  {
    cp.w = (float) *(p + 3);

    // float alpha = weight * QuantumScale * cp.w;
    // error: use of type 'double' requires cl_khr_fp64 support
    float alpha = weight * (1 - (float) cp.w / 255);

    filteredPixel.x += alpha * cp.x;
    filteredPixel.y += alpha * cp.y;
    filteredPixel.z += alpha * cp.z;
    filteredPixel.w += weight * cp.w;
    gamma += alpha;
  }
  else
    filteredPixel += ((float4) weight)*cp;

  density += weight;
}

6. “如何写ScaleImage()的硬件加速函数（五）”的问题在于没有办法处理图片下半部分（如何缩小一半的话），这里主要是因为y变量的限定（代码如下），因为传入kernel函数的gsize[1]=resizedRows;被限定的死死的。

const unsigned int pos = getPixelIndex(4, inputColumns, cacheRangeStartX, y);

7. 目前只考虑按比例缩放，所以这里的y需要除以缩放比。
8. 缩小后图片如果垂直方向相间着黑色宽竖条，那可能是因为numCachedPixels参数没有计算正确，这正可以修复“如何写ScaleImage()的硬件加速函数（五）”中的scale_ratio变量。

numCachedPixels=(int) ceil((pixelPerWorkgroup-1)/xFactor+2*support);

9. 附上目前代码：

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,
    lsize[2],
    totalLocalMemorySize,
    weightAccumulatorLocalMemorySize;

  unsigned int
    chunkSize,
    pixelPerWorkgroup;

  scaleKernel=NULL;
  outputReady=MagickFalse;

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

  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;
    totalLocalMemorySize=imageCacheLocalMemorySize;

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

    /* 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),(void*)&matte_or_cmyk);
  status|=SetOpenCLKernelArg(scaleKernel,i++,sizeof(cl_uint),(void*)&columns);
  status|=SetOpenCLKernelArg(scaleKernel,i++,sizeof(cl_uint),(void*)&rows);
  status|=SetOpenCLKernelArg(scaleKernel,i++,sizeof(cl_mem),(void*)&scaledImageBuffer);
  status|=SetOpenCLKernelArg(scaleKernel,i++,sizeof(cl_uint),(void*)&scaledColumns);
  status|=SetOpenCLKernelArg(scaleKernel,i++,sizeof(cl_uint),(void*)&scaledRows);
  status|=SetOpenCLKernelArg(scaleKernel,i++,sizeof(float),(void*)&scale);
  status|=SetOpenCLKernelArg(scaleKernel,i++,imageCacheLocalMemorySize,NULL);
  status|=SetOpenCLKernelArg(scaleKernel,i++,sizeof(int),&numCachedPixels);
  status|=SetOpenCLKernelArg(scaleKernel,i++,sizeof(unsigned int),&pixelPerWorkgroup);
  status|=SetOpenCLKernelArg(scaleKernel,i++,sizeof(unsigned int),&chunkSize);
  status|=SetOpenCLKernelArg(scaleKernel,i++,pixelAccumulatorLocalMemorySize,NULL);
  status|=SetOpenCLKernelArg(scaleKernel,i++,weightAccumulatorLocalMemorySize,NULL);
  status|=SetOpenCLKernelArg(scaleKernel,i++,gammaAccumulatorLocalMemorySize,NULL);

  if (status != CL_SUCCESS)
  {
    (void) OpenCLThrowMagickException(device,exception,GetMagickModule(),
      ResourceLimitWarning,"SetOpenCLKernelArg failed.",".");
    goto cleanup;
  }

  gsize[0] = (scaledColumns + pixelPerWorkgroup - 1) / pixelPerWorkgroup *
    workgroupSize;
  gsize[1] = scaledRows;
  lsize[0] = workgroupSize;
  lsize[1] = 1;
  outputReady=EnqueueOpenCLKernel(queue,scaleKernel,2,
    (const size_t *) NULL,gsize,lsize,image,filteredImage,MagickFalse,
    exception);

cleanup:

  if (scaleKernel != (cl_kernel) NULL)
    ReleaseOpenCLKernel(scaleKernel);

  return(outputReady);
}

STRINGIFY(
__kernel __attribute__((reqd_work_group_size(256, 1, 1)))
  void ScaleFilter(const __global CLQuantum *inputImage, const unsigned int matte_or_cmyk,
    const unsigned int inputColumns, const unsigned int inputRows, __global CLQuantum *filteredImage,
    const unsigned int filteredColumns, const unsigned int filteredRows,
    const float resizeFilterScale,
    __local CLQuantum *inputImageCache, const int numCachedPixels,
    const unsigned int pixelPerWorkgroup, const unsigned int pixelChunkSize,
    __local float4 *outputPixelCache, __local float *densityCache, __local float *gammaCache)
{
  // calculate the range of resized image pixels computed by this workgroup
  const unsigned int startX = get_group_id(0) * pixelPerWorkgroup;
  const unsigned int stopX = MagickMin(startX + pixelPerWorkgroup, filteredColumns);
  const unsigned int actualNumPixelToCompute = stopX - startX;

  float xFactor = (float)filteredColumns / inputColumns;

  // calculate the range of input image pixels to cache
  const int cacheRangeStartX = MagickMax((int)((startX + 0.5f) / xFactor), (int)(0));
  const int cacheRangeEndX = MagickMin((int)(cacheRangeStartX + numCachedPixels), (int)inputColumns);

  // cache the input pixels into local memory
  const unsigned int y = get_global_id(1);
  const unsigned int pos = getPixelIndex(4, inputColumns, cacheRangeStartX, y / xFactor);
  const unsigned int num_elements = (cacheRangeEndX - cacheRangeStartX) * 4;
  event_t e = async_work_group_copy(inputImageCache, inputImage + pos, num_elements, 0);
  wait_group_events(1, &e);

  unsigned int totalNumChunks = (actualNumPixelToCompute + pixelChunkSize - 1) / pixelChunkSize;
  for (unsigned int chunk = 0; chunk < totalNumChunks; chunk++)
  {
    const unsigned int chunkStartX = startX + chunk * pixelChunkSize;
    const unsigned int chunkStopX = MagickMin(chunkStartX + pixelChunkSize, stopX);
    const unsigned int actualNumPixelInThisChunk = chunkStopX - chunkStartX;

    // determine which resized pixel computed by this workitem
    const unsigned int itemID = get_local_id(0);
    const unsigned int numItems = getNumWorkItemsPerPixel(actualNumPixelInThisChunk, get_local_size(0));

    const int pixelIndex = pixelToCompute(itemID, actualNumPixelInThisChunk, get_local_size(0));

    float4 filteredPixel = (float4)0.0f;

    // -1 means this workitem doesn't participate in the computation
    if (pixelIndex != -1)
    {
      // x coordinated of the resized pixel computed by this workitem
      const int x = chunkStartX + pixelIndex;

      // calculate how many steps required for this pixel
      const float bisect = (x + 0.5) / xFactor + MagickEpsilon;
      const unsigned int start = (unsigned int)MagickMax(bisect, 0.0f);
      const unsigned int stop = (unsigned int)MagickMin(bisect + 1, (float)inputColumns);
      const unsigned int n = stop - start;

      // calculate how many steps this workitem will contribute
      unsigned int numStepsPerWorkItem = n / numItems;
      numStepsPerWorkItem += ((numItems * numStepsPerWorkItem) == n ? 0 : 1);

      const unsigned int startStep = (itemID % numItems) * numStepsPerWorkItem;
      if (startStep < n)
      {
        const unsigned int stopStep = MagickMin(startStep + numStepsPerWorkItem, n);

        unsigned int cacheIndex = start + startStep - cacheRangeStartX;
        for (unsigned int i = startStep; i < stopStep; i++, cacheIndex++)
        {
          float4 cp = (float4)0.0f;

          __local CLQuantum* p = inputImageCache + (cacheIndex * 4);
          cp.x = (float)*(p);
          cp.y = (float)*(p + 1);
          cp.z = (float)*(p + 2);
          cp.w = (float)*(p + 3);

          filteredPixel += cp;
        }
      }
    }

    if (itemID < actualNumPixelInThisChunk)
    {
      WriteAllChannels(filteredImage, 4, filteredColumns, chunkStartX + itemID, y, filteredPixel);
    }
  }
}
)