GraphicsMagick 的 OpenCL 开发记录（三十三）

<2022-04-28 周四>

如何写ScaleImage()的硬件加速函数（七）

其实“如何写ScaleImage()的硬件加速函数（六）”的实现就是一个ResizeHorizontalFilter()将y改成y / xFactor的精简版，并不是ScaleImage()的硬件加速函数。虽然它不是，但至少省掉了ResizeVerticalFilter()的调用，速度上更快了。

但是目前发现的问题还是竖条纹，连续多次缩小一倍，最终图片被黑色竖条纹全部覆盖住，不断缩小或者放大，右侧会出现密集竖条纹，等等等的问题啦。

经过分析，黑色竖纹的产生原因是因为kernel函数ScaleFilter()的最内层的循环没有执行，导致将初始值0.0f赋进了目标地址。

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;
}

可以这样解决：

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) {
      outputPixelCache[itemID] = (float4)0.0f;
    }
    barrier(CLK_LOCAL_MEM_FENCE);

    for (unsigned int i = 0; i < numItems; i++) {
      if (pixelIndex != -1) {
        if (itemID % numItems == i) {
          outputPixelCache[pixelIndex] += filteredPixel;
        }
      }
      barrier(CLK_LOCAL_MEM_FENCE);
    }

    if (itemID < actualNumPixelInThisChunk)
    {
      float4 filteredPixel = outputPixelCache[itemID];
      WriteAllChannels(filteredImage, 4, filteredColumns, chunkStartX + itemID, y, filteredPixel);
    }
  }
}
)

测试了一下性能，感觉提升不少（原图缩小一半，共三次操作，原图连续放大一倍两次，共三次操作）：

ScaleImage()加速版本：

20220428104719 0:3.229821  1.672 11552 opencl.c AcquireOpenCLKernel 744 Accelerate Event Using kernel: ScaleFilter
20220428104719 0:3.230185  1.672 11552 resize.c ScaleImage 1764 Accelerate Event accelerate scale: 1360
20220428104725 0:9.628057  1.875 11552 opencl.c AcquireOpenCLKernel 744 Accelerate Event Using kernel: ScaleFilter
20220428104725 0:9.628288  1.875 11552 resize.c ScaleImage 1764 Accelerate Event accelerate scale: 0
20220428104732 0:16.078872 2.234 11552 opencl.c AcquireOpenCLKernel 744 Accelerate Event Using kernel: ScaleFilter
20220428104732 0:16.079057 2.234 11552 resize.c ScaleImage 1764 Accelerate Event accelerate scale: 0
20220428104740 0:24.253815 2.484 11552 opencl.c AcquireOpenCLKernel 744 Accelerate Event Using kernel: ScaleFilter
20220428104740 0:24.254118 2.484 11552 resize.c ScaleImage 1764 Accelerate Event accelerate scale: 0
20220428104749 0:33.888819 2.875 11552 opencl.c AcquireOpenCLKernel 744 Accelerate Event Using kernel: ScaleFilter
20220428104749 0:33.889007 2.875 11552 resize.c ScaleImage 1764 Accelerate Event accelerate scale: 31
20220428104752 0:36.173104 3.047 11552 opencl.c AcquireOpenCLKernel 744 Accelerate Event Using kernel: ScaleFilter
20220428104752 0:36.173301 3.047 11552 resize.c ScaleImage 1764 Accelerate Event accelerate scale: 156
20220428104800 0:44.287153 3.469 11552 opencl.c AcquireOpenCLKernel 744 Accelerate Event Using kernel: ScaleFilter
20220428104800 0:44.287372 3.469 11552 resize.c ScaleImage 1764 Accelerate Event accelerate scale: 47
20220428104801 0:45.546271 3.656 11552 opencl.c AcquireOpenCLKernel 744 Accelerate Event Using kernel: ScaleFilter
20220428104801 0:45.546588 3.656 11552 resize.c ScaleImage 1764 Accelerate Event accelerate scale: 140
20220428104806 0:49.973027 4.047 11552 opencl.c AcquireOpenCLKernel 744 Accelerate Event Using kernel: ScaleFilter
20220428104806 0:49.973217 4.047 11552 resize.c ScaleImage 1764 Accelerate Event accelerate scale: 31
20220428104806 0:50.640522 4.250 11552 opencl.c AcquireOpenCLKernel 744 Accelerate Event Using kernel: ScaleFilter
20220428104806 0:50.640730 4.250 11552 resize.c ScaleImage 1764 Accelerate Event accelerate scale: 141

ScaleImage()原先版本：

20220428104934 0:1.982873  0.266 10052 resize.c ScaleImage 1770 Accelerate Event AccelerateScaleImage null
20220428104934 0:2.040677  0.328 10052 resize.c ScaleImage 2116 Accelerate Event normal scale: 63
20220428104940 0:7.854823  0.578 10052 resize.c ScaleImage 1770 Accelerate Event AccelerateScaleImage null
20220428104940 0:7.913365  0.625 10052 resize.c ScaleImage 2116 Accelerate Event normal scale: 47
20220428104944 0:11.896725 0.875 10052 resize.c ScaleImage 1770 Accelerate Event AccelerateScaleImage null
20220428104944 0:11.956722 0.938 10052 resize.c ScaleImage 2116 Accelerate Event normal scale: 63
20220428104951 0:18.070817 1.219 10052 resize.c ScaleImage 1770 Accelerate Event AccelerateScaleImage null
20220428104951 0:18.378405 1.516 10052 resize.c ScaleImage 2116 Accelerate Event normal scale: 297
20220428104952 0:19.394056 1.531 10052 resize.c ScaleImage 1770 Accelerate Event AccelerateScaleImage null
20220428104953 0:20.634341 2.781 10052 resize.c ScaleImage 2116 Accelerate Event normal scale: 1250
20220428104958 0:25.534006 3.063 10052 resize.c ScaleImage 1770 Accelerate Event AccelerateScaleImage null
20220428104958 0:25.836584 3.375 10052 resize.c ScaleImage 2116 Accelerate Event normal scale: 312
20220428104959 0:26.729520 3.406 10052 resize.c ScaleImage 1770 Accelerate Event AccelerateScaleImage null
20220428105000 0:27.930533 4.609 10052 resize.c ScaleImage 2116 Accelerate Event normal scale: 1203
20220428105011 0:38.879392 5.438 10052 resize.c ScaleImage 1770 Accelerate Event AccelerateScaleImage null
20220428105012 0:39.210382 5.766 10052 resize.c ScaleImage 2116 Accelerate Event normal scale: 328
20220428105012 0:39.872525 5.797 10052 resize.c ScaleImage 1770 Accelerate Event AccelerateScaleImage null
20220428105014 0:41.176969 7.094 10052 resize.c ScaleImage 2116 Accelerate Event normal scale: 1297