DSP教程附录浮点运算单元.ppt

DSP28 3 Floating-Point Unit Appendix E TMS320C28x? MCU Workshop Learning Objectives IEEE Single-Precision Floating-Point Format IEEE Single-Precision Floating-Point Format (IEEE 754) Most widely used standard for floating-point Standard number formats, Special values (NaN, Infinity) Rounding modes floating-point operations Used on many CPUs Simplifications for the C28x floating-point unit Flags Compare Operations: Negative zero is treated as positive zero Denormalized values are treated as zero Not-a-number (NaN) is treated as infinity Round-to-Zero Mode Supported (truncate) Round-to-Nearest Mode Supported (even) These formats are commonly handled this way on embedded processors C28x CPU and FPU Registers Floating-Point Unit Pipeline Floating-point operations are not pipeline protected Some instructions require delay slots for the operation to complete Insert NOPs or other non-conflicting instructions between operations Floating-Point Unit Pipeline Delay Slots Instructions requiring delay slots have a ‘p’ after their cycle count 2p stands for 2 pipelined cycles A new instruction can be started on each cycle The result is valid only 2 instructions later Assembler issues errors for pipeline conflicts Fill delay slots with non-conflicting instructions to improve performance 3 general guidelines: Floating-Point Instruction Format Same instruction format as the fixed-point instructions Destination operand is always on the left Types of Floating-Point Instructions (1 of 2) Types of Floating-Point Instructions (2 of 2) Math Operations Parallel Instructions Multiply and Store Parallel Instruction Multiply and Store Parallel Instruction Using Floating-Point Status Flags Repeat Block Improves performance of block algorithms (FFT, IIR) 0 cycles after the first iteration Repeat Block (RPTB) The RPTB is interruptible and allows for nested interrupts High Priority ISR: Save/Restore RB if RPTB is used Low Priority ISR: Always save/restore RB No other discontinuities are al