我正在尝试在 Halide 中实现标准化互相关。
下面的代码构建，Halide JIT 编译不会抛出任何错误。但是，在 JIT 编译后，Halide 似乎挂起。无论trace_*我在不同的 Funcs 上进行了多少次调用，都只会打印一条迹线（在 上Func output）：
Begin realization normxcorr.0(0, 2028, 0, 2028)
Produce normxcorr.0(0, 2028, 0, 2028)
任何建议都会有所帮助。
该算法等同于 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 ) );
}