This slide feels pretty important. We looked at just how hairy a properly optimized two-pass blur routine can look when manually implementing it compared to the simplicity of equivalently declaring it in Halide code. So not only is Halide useful for more quickly testing different underlying implementations to tweak performance, but it also facilitates more rapid software development to the point where large image processing pipelines like the Google HDR+ pipeline with thousands of transformations wouldn't be justifiable from a product management standpoint if engineers had to write it in C/C++ instead of a higher-level abstraction like Halide.
This slide feels pretty important. We looked at just how hairy a properly optimized two-pass blur routine can look when manually implementing it compared to the simplicity of equivalently declaring it in Halide code. So not only is Halide useful for more quickly testing different underlying implementations to tweak performance, but it also facilitates more rapid software development to the point where large image processing pipelines like the Google HDR+ pipeline with thousands of transformations wouldn't be justifiable from a product management standpoint if engineers had to write it in C/C++ instead of a higher-level abstraction like Halide.