Visible to Intel only — GUID: GUID-FC57CAD3-7B6C-4CC1-9FB0-682E159BDEFD
Visible to Intel only — GUID: GUID-FC57CAD3-7B6C-4CC1-9FB0-682E159BDEFD
Modify the Rounding Mode of Floating-point Operations (-Xsrounding=<rounding_type>)
To modify the rounding mode of floating-point elementary operations in your design, use the -Xsrounding=<rounding_type> option in your icpx command. You can set the -Xsrounding option to one of the following values:
- ieee: All elementary operations (+, -, *, /) of both single-precision and double-precision floating-point use IEEE-754 round nearest ties to even (RNE) mode, which has a 0.5 unit of least precision (ULP) error at most.
- faithful: All elementary operations of double-precision floating-point and multiplication and division of single-precision floating-point use faithful rounding mode, which has a 1 ULP error at most. This rounding mode leads to more efficient hardware at the expense of numerical variation in results. Addition and subtraction of single-precision floating-point still have to use IEEE-754 RNE rounding mode.
The following tables summarize the rounding modes:
Single-precision Floating-point
Addition |
Subtraction |
Multiplication |
Division |
|
---|---|---|---|---|
Default |
IEEE-754 RNE |
IEEE-754 RNE |
IEEE-754 RNE |
Faithful |
-Xsrounding=ieee |
IEEE-754 RNE |
IEEE-754 RNE |
IEEE-754 RNE |
IEEE-754 RNE |
-Xsrounding=faithful |
IEEE-754 RNE |
IEEE-754 RNE |
Faithful |
Faithful |
Double-precision Floating-point
Addition | Subtraction | Multiplication | Division | |
---|---|---|---|---|
Default |
IEEE-754 RNE |
IEEE-754 RNE |
IEEE-754 RNE |
IEEE-754 RNE |
-Xsrounding=ieee |
IEEE-754 RNE |
IEEE-754 RNE |
IEEE-754 RNE |
IEEE-754 RNE |
-Xsrounding=faithful |
Faithful | Faithful | Faithful | Faithful |
Examples
icpx -fsycl -fintelfpga -Xshardware -Xsrounding=ieee <source_file>.cpp
icpx -fsycl -fintelfpga -Xshardware -Xsrounding=faithful <source_file>.cpp