FFTW使用说明(PPT-24).ppt

3.1 The Planner For a given N, there are many factorizations not clear a priori which is best Eg: 32768 =16x8x8x32=64x16x32 The planner tries them “all” and picks the best one uses actual runtime timing measurements result is encoded in a “plan” Uses dynamic programming to reduce no. of possible plans remembers optimal sub-plans for small sizes The Runtime Planner optimizes FFTW for your CPU, your cache size, etc. Ideas Modern architectures are invalidating conventional wisdom about what is fast no new wisdom is emerging In the name of performance, designers have sacrificed: predictability, repeatability, composability Hand-optimization of programs is becoming impractical 3.2 The Codelet Generator Generates highly-optimized transforms of small sizes form the base cases of the FFT recursion Manipulates abstract syntax tree which is unparsed to C knows about complex arithmetic, etc. The codelets composable blocks of optimized code computer generated Advantages Long, unrolled code takes advantage of: optimizing compilers (instruction scheduling, etc.), large register sets Applies tedious optimizations Easy to experiment with different algorithms prime factor, split-radix (transform sizes not a power of 2) various optimization schemes 3.3 The Executor Executes the plan by composing codelets Explicit recursion divide-and-conquer uses all levels of the memory hierarchy fit in cache Novel storage for the twiddle factors store them in the order they are used 3.4 FFTW is Easy to Use COMPLEX A[n], B[n]; fftw_plan plan; plan = fftw_create_plan(n); /* create the plan */ fftw(plan, A); /* use the plan */ fftw(plan, B); /* re-use the plan */ fftw_destroy_plan(plan) /* destroy the plan*/ 4.1 调用说明 FFTW 的C函数允许Fortran程序调用 . fftw/fftwnd/rfftw/rfftwnd 由 fftw_f77/fftwnd_f77/rfftw_f77/rfftwnd_f77所替代 3. 函数的大多数参数相同, 少数例外: plan变量在C中为fftw_plan, rfftw_plan等类型， Fortran上对64位机器用integer*8 类型 Fortran数组列存储,C为行存储, integer对应于int, real对应于float fortran/fftw_f77.