Merge pull request #69 from QuState/determine-level-once

Determine SIMD level in the planner

Author: Shnatsel

src/algorithms/dit.rs

@@ -14,7 +14,7 @@
 //! DIT starts with fine-grained memory access and progressively works with
 //! larger contiguous chunks.
 //!
-use fearless_simd::{dispatch, Level, Simd};
+use fearless_simd::{dispatch, Simd};
 
 use crate::algorithms::bravo::{bit_rev_bravo_f32, bit_rev_bravo_f64};
 use crate::kernels::dit::*;
@@ -251,7 +251,7 @@ pub fn fft_64_dit_with_planner_and_opts(
     let log_n = n.ilog2() as usize;
     assert_eq!(log_n, planner.log_n);
 
-    let simd_level = Level::new();
+    let simd_level = planner.simd_level;
 
     // DIT requires bit-reversed input
     run_maybe_in_parallel(
@@ -296,7 +296,7 @@ pub fn fft_32_dit_with_planner_and_opts(
     let log_n = n.ilog2() as usize;
     assert_eq!(log_n, planner.log_n);
 
-    let simd_level = Level::new();
+    let simd_level = planner.simd_level;
 
     // DIT requires bit-reversed input
     run_maybe_in_parallel(

src/planner.rs

@@ -87,13 +87,17 @@ macro_rules! impl_planner_dit_for {
             pub direction: Direction,
             /// The log2 of the FFT size
             pub log_n: usize,
+            /// The level of SIMD instruction support, detected at runtime on x86 and hardcoded elsewhere
+            pub simd_level: fearless_simd::Level,
         }
 
         impl $struct_name {
             /// Create a DIT planner for an FFT of size `num_points`
             pub fn new(num_points: usize, direction: Direction) -> Self {
                 assert!(num_points > 0 && num_points.is_power_of_two());
 
+                let simd_level = fearless_simd::Level::new();
+
                 let log_n = num_points.ilog2() as usize;
                 let mut stage_twiddles = Vec::new();
 
@@ -123,6 +127,7 @@ macro_rules! impl_planner_dit_for {
                     stage_twiddles,
                     direction,
                     log_n,
+                    simd_level,
                 }
             }
         }