是否有可能通过gcc提高浮点算法的准确性？

Question

some program in C which does extensive floating point calculations get right results on a pc linux box, but wrong results on the SPE of the cell processor, but not on the PPU of the cell. I am using gcc compilers. I wonder if there is some gcc compilation option to increase rounding method or similar so I get single float precision calculations with more precision. I can not change to double, as on the SPE performance would drastic reduce

Thanks

Answer 1

Based on the IBM documentation for the differences from IEEE 754 on the SPU , it could be any number of things:

• Zero results from arithmetic operations are always +0, never -0.
• Denormal inputs from 2-149 to 2-126 to arithmetic operations are treated as zero with the same sign. Arithmetic operations never produce denormal results, but produce +0 instead.
• Arithmetic operations do not support IEEE Inf or NaN. These bit patterns represent valid numbers. Overflow results produce the maximum magnitude value of appropriate sign.
• Arithmetic operations use only the round-to-zero (chop, truncate) rounding mode, regardless of the setting of the rounding mode in the Floating-Point Status and Control Register (FPSCR), which affects only double-precision arithmetic operations.

Of course, on a related page, you can also compile SPU code for strict IEEE conformance :

By default, XL C/C++ follows most, but not all of the rules in the IEEE standard. If you compile with the -qnostrict option, which is enabled by default at optimization level -O3 or higher, some IEEE floating-point rules are violated in ways that can improve performance but might affect program correctness. To avoid this issue, and to compile for strict compliance with the IEEE standard, do the following:

• Use the -qfloat=nomaf compiler option.
• If the program changes the rounding mode at runtime, use the -qfloat=rrm option.
• If the data or program code contains signaling NaN values (NaNS), use the -qfloat=nans option. (A signaling NaN is different from a quiet NaN; you must explicitly code it into the program or data or create it by using the -qinitauto compiler option.)
• If you compile with -O3, -O4, or -O5, include the option -qstrict after it.