如何让卤化物使用滑动窗口优化?

我正在学习卤化德，但我正在为调度部分而苦苦挣扎。我试图让 Halide 做与算法的手动编码实现相同的事情。我不想并行化它，而是矢量化它，但首先我想了解如何让 Halide 做一个简单的滑动窗口。我尝试了不同的事情：

• 将总和拆分为 x 和 y，然后进行不同的调度
• 以递归形式定义总和(末尾的代码片段)

但我无法让它产生类似的东西。它归结为平均滤波器的简单变体。那么，如何安排卤化物代码像原始代码一样实际执行滑动窗口？

这是代码：

void lmrCPU(const cv::Mat& image, std::vector<uint16_t>& vertSum, int xMin, int xMax, int yMin, int yMax, int lmrSize, cv::Mat& lmr ) {
int lmrWidth = 2*lmrSize + 1;
int area = lmrWidth*lmrWidth;
vertSum.resize(image.cols);
// Zero contents of vertSum
memset(vertSum.data(), 0, vertSum.size()*sizeof(uint16_t));
for (int yy = yMin; yy < yMin + lmrWidth; ++yy)
{
for (int x=0; x<image.cols; x++) vertSum[x] += unsigned(image.ptr<uint8_t>(yy)[x]);
}
for (int y = yMin; y <= yMax; y++)
{
if (y > yMin + lmrSize && y <= yMax - lmrSize)
{
for (int x = 0; x < image.cols; x++)
vertSum[x] += unsigned(image.at<uint8_t>(y + lmrSize, x)) - unsigned(image.at<uint8_t>(y - lmrSize - 1, x));
}
unsigned sum = 0;
int xx, x, xxBack;
for (xx = xMin; xx < xMin + lmrWidth; xx++) sum += vertSum[xx];
for (x = xMin; x < xMin + lmrSize; x++)
lmr.at<int>(y, x) = int(image.at<uint8_t>(y, x)) * area - sum;
sum -= vertSum[xMin + 2*lmrSize]; // take off ready for next loop
for (x = xMin + lmrSize, xxBack = x - lmrSize, xx = x + lmrSize; x < xMax - lmrSize; x++, xx++, xxBack++)
{
sum += vertSum[xx];
lmr.at<int>(y, x) = int(image.at<uint8_t>(y,x))*area - sum;
sum -= vertSum[xxBack];
}
sum += vertSum[xx];
for ( ; x <= xMax; x++)
lmr.at<int>(y, x) = int(image.at<uint8_t>(y,x))*area - sum;
}
}

#include "Halide.h"
namespace {
class LMR : public Halide::Generator<LMR> {
public:
ImageParam input{UInt(8), 2, "input"};
Param<int32_t> xMin{"xMin"}, xMax{"xMax"};
Param<int32_t> yMin{"yMin"}, yMax{"yMax"};
Param<int32_t> lmrSize{"lmrSize"};
Var x{"x"}, y{"y"};
Func build() {
auto lmrWidth = 2*lmrSize + 1;
auto area = lmrWidth*lmrWidth;
Halide::Func input_int32 ("input_int32");
input_int32(x, y) = Halide::cast<int32_t>(input(x, y));
Halide::Func input_uint16 ("input_uint16");
input_uint16(x, y) = Halide::cast<uint16_t>(input(x, y));
Halide::Expr clamped_x = Halide::clamp(x, xMin, xMax);
Halide::Expr clamped_y = Halide::clamp(y, yMin, yMax);
Halide::Expr lmr_x = Halide::clamp(x, xMin+lmrSize, xMax-lmrSize);
Halide::Expr lmr_y = Halide::clamp(y, yMin+lmrSize, yMax-lmrSize);
Halide::RDom box (-lmrSize, lmrWidth, "box");
Halide::Func vertSum ("vertSum");
vertSum(x, y) = Halide::undef<uint16_t>();
{
Halide::RDom ry (yMin+lmrSize+1, yMax-yMin-2*lmrSize, "ry");
vertSum(x, yMin+lmrSize) = Halide::cast<uint16_t>(0);//Halide::sum(input_uint16(x, yMin+lmrSize+box), "sum_y");
vertSum(x, yMin+lmrSize) += input(x, yMin+lmrSize+box);
vertSum(x, ry) = vertSum(x, ry-1) + input_uint16(x, ry+lmrSize) - input_uint16(x, ry-1-lmrSize);
}
Halide::Func sumLmr ("sumLmr");
sumLmr(x, y) = Halide::undef<uint16_t>();
{
Halide::RDom rx (xMin+lmrSize+1, xMax-xMin-2*lmrSize, "rx");
sumLmr(xMin+lmrSize, y) = Halide::cast<uint16_t>(0);//Halide::sum(vertSum(xMin+lmrSize+box, y), "sum_x");
sumLmr(xMin+lmrSize, y) += vertSum(xMin+lmrSize+box, y);
sumLmr(rx, y) = sumLmr(rx-1, y) + vertSum(rx+lmrSize, y) - vertSum(rx-1-lmrSize, y);
}
Halide::Func lmr ("lmr");
lmr(x, y) = input_int32(clamped_x, clamped_y)*area - Halide::cast<int32_t>(sumLmr(lmr_x, lmr_y));
vertSum
.fold_storage(y, 1)
.store_root()
.compute_at(lmr, y);
sumLmr
.fold_storage(x, 1)
.store_at(lmr, y)
.compute_at(lmr, x);
return lmr;
}
};
HALIDE_REGISTER_GENERATOR(LMR, "lmr")
}

这是lmr.print_loop_nest();的输出

store vertSum:
produce lmr:
for y:
produce vertSum:
for y:
for x:
vertSum(...) = ...
for x:
vertSum(...) = ...
for x:
for box:
vertSum(...) = ...
for x:
for ry:
vertSum(...) = ...
consume vertSum:
store sumLmr:
for x:
produce sumLmr:
for y:
for x:
sumLmr(...) = ...
for y:
sumLmr(...) = ...
for y:
for box:
sumLmr(...) = ...
for y:
for rx:
sumLmr(...) = ...
consume sumLmr:
lmr(...) = ...