Halide在归一化互相关期间挂起

我正在尝试在Halide中实现归一化互相关。

下面的代码构建，Halide JIT编译没有抛出任何错误。然而，Halide似乎在JIT编译后挂起了。无论我在不同的Func上进行了多少次trace_*调用，都只能打印一个跟踪（在Func output上）：

Begin realization normxcorr.0(0, 2028, 0, 2028)
Produce normxcorr.0(0, 2028, 0, 2028)

任何建议都会有帮助。

该算法等效于OpenCV中的CV_TM_CCOEFF_NORMED和MATLAB:中的normxcorr2

void normxcorr( Halide::ImageParam input,
                Halide::ImageParam kernel,
                Halide::Param<pixel_t> kernel_mean,
                Halide::Param<pixel_t> kernel_var,
                Halide::Func& output )
{
    Halide::Var x, y;
    Halide::RDom rk( kernel );
    // reduction domain for cumulative sums
    Halide::RDom ri( 1, input.width() - kernel.width() - 1, 
                     1, input.height() - kernel.height() - 1 );
    Halide::Func input_32( "input32" ),
             bounded_input( "bounded_input"),
             kernel_32( "kernel32" ),
             knorm( "knorm" ),
             conv( "conv" ),
             normxcorr( "normxcorr_internal" ),
             sq_sum_x( "sq_sum_x" ),
             sq_sum_x_local( "sq_sum_x_local" ),
             sq_sum_y( "sq_sum_y" ),
             sq_sum_y_local( "sq_sum_y_local" ),
             sum_x( "sum_x" ),
             sum_x_local( "sum_x_local" ),
             sum_y( "sum_y" ),
             sum_y_local( "sum_y_local" ),
             win_var( "win_var" ),
             win_mean( "win_mean" );
    Halide::Expr ksize = kernel.width() * kernel.height();
    // accessing outside the input image always returns 0
    bounded_input( x, y ) = Halide::BoundaryConditions::constant_exterior( input, 0 )( x, y );
    // cast to 32-bit to make room for multiplication
    input_32( x, y ) = Halide::cast<int32_t>( bounded_input( x, y ) );
    kernel_32( x, y ) = Halide::cast<int32_t>( kernel( x, y ) );
    // cumulative sum along each row
    sum_x( x, y ) = input_32( x, y );
    sum_x( ri.x, ri.y ) += sum_x( ri.x - 1, ri.y );
    // sum of 1 x W strips
    // (W is the width of the kernel)
    sum_x_local( x, y ) = sum_x( x + kernel.width() - 1, y );
    sum_x_local( x, y ) -= sum_x( x - 1, y );
    // cumulative sums of the 1 x W strips along each column
    sum_y( x, y ) = sum_x_local( x, y );
    sum_y( ri.x, ri.y ) += sum_y( ri.x, ri.y - 1);
    // sums up H strips (as above) to get the sum of an H x W rectangle
    // (H is the height of the kernel)
    sum_y_local( x, y ) = sum_y( x, y + kernel.height() - 1 );
    sum_y_local( x, y ) -= sum_y( x, y - 1 );
    // same as above, just with squared image values
    sq_sum_x( x, y ) = input_32( x, y ) * input_32( x, y );
    sq_sum_x( ri.x, ri.y ) += sq_sum_x( ri.x - 1, ri.y );
    sq_sum_x_local( x, y ) = sq_sum_x( x + kernel.width() - 1, y );
    sq_sum_x_local( x, y ) -= sq_sum_x( x - 1, y );
    sq_sum_y( x, y ) = sq_sum_x_local( x, y );
    sq_sum_y( ri.x, ri.y ) += sq_sum_y( ri.x, ri.y - 1);
    sq_sum_y_local( x, y ) = sq_sum_y( x, y + kernel.height() - 1 );
    sq_sum_y_local( x, y ) -= sq_sum_y( x, y - 1 );
    // the mean value of each window
    win_mean( x, y ) = sum_y_local( x, y ) / ksize;
    // the variance of each window
    win_var( x, y ) =  sq_sum_y_local( x, y ) / ksize;
    win_var( x, y) -= win_mean( x, y ) * win_mean( x, y );
    // partially normalize the kernel
    // (we'll divide by std. dev. at the end)
    knorm( x, y ) = kernel_32( x, y ) - kernel_mean;
    // convolve kernel and the input
    conv( x, y ) = Halide::sum( knorm( rk.x, rk.y ) * input_32( x + rk.x, y + rk.y ) );
    // calculate normxcorr, except scaled to 0 to 254 (for an 8-bit image)
    normxcorr( x, y ) = conv( x, y ) * 127 / Halide::sqrt( kernel_var * win_var( x, y ) ) + 127;
    // after scaling pixel values, it's safe to cast down to 8-bit
    output( x, y ) = Halide::cast<pixel_t>( normxcorr( x, y ) );
}