终于自己找到原因了
TMS320C5535/34/33/32 Fixed-Point DSP Silicon Errata 这个手册中的这句话：
The data and scratch buffers can be located in word addresses greater than 0x10000, but the buffers must not cross 16-bit address boundaries (that is, address bits 22–16 must not change).
很关键，说的是所用要到的buffers最好都在数据空间的同一页，即data_buf,data_br_buf,scratch_buf，均要设置在数据空间的同一页，最后FFT再IFFT计算出来的结果才基本一样