Re:RustBelt: securing the foundations of the Rust programmin...
by Const-me on Monday February 05, 2018 @01:01PM
(#16310590)
One response I received to that argument was “not many projects need to multiply a lot of single-precision floating point matrices”.
Later additions, SSE2, SSE 4.1, SSSE3, add double-precision floating point, and 8-16-32-64 bit integers support. Some of them come with unsigned and saturated versions (the latter clips to min/max when overflown).
In C++, I use these for performance-critical functions doing integer computations, bit manipulation, image processing (saturated integer math is especially useful for the latter). While it adds some complexity to the code esp. when it benefits from SSE 4.1 or SSSE3 instructions (I must implement SSE2-only workarounds for older CPUs), the performance improvement IMO justifies that.
Later additions, SSE2, SSE 4.1, SSSE3, add double-precision floating point, and 8-16-32-64 bit integers support. Some of them come with unsigned and saturated versions (the latter clips to min/max when overflown).
In C++, I use these for performance-critical functions doing integer computations, bit manipulation, image processing (saturated integer math is especially useful for the latter). While it adds some complexity to the code esp. when it benefits from SSE 4.1 or SSSE3 instructions (I must implement SSE2-only workarounds for older CPUs), the performance improvement IMO justifies that.