Optimize edge

juno_samples/edge_detection/benches/edge_detection_bench.rs

@@ -87,9 +87,6 @@ fn edge_detection_bench(c: &mut Criterion) {
                     r.run(
                         height as u64,
                         width as u64,
-                        gs as u64,
-                        sz as u64,
-                        sb as u64,
                         input_h.to(),
                         gaussian_filter_h,
                         structure_h,

juno_samples/edge_detection/src/cpu.sch

@@ -134,7 +134,9 @@ if !feature("seq") {
   reject_zero_crossings = reject_zero_crossings_body;
 }
 
-async-call(edge_detection@le, edge_detection@zc);
+if !feature("seq") {
+  async-call(edge_detection@le, edge_detection@zc);
+}
 
 fork-split(gaussian_smoothing, laplacian_estimate, zero_crossings, gradient, reject_zero_crossings);
 unforkify(gaussian_smoothing, laplacian_estimate, zero_crossings, gradient, reject_zero_crossings);

juno_samples/edge_detection/src/edge_detection.jn

-fn gaussian_smoothing<n, m, gs : usize>(
+const gs : usize = 7;
+const sz : usize = 3;
+const sb : usize = 3;
+
+fn gaussian_smoothing<n, m : usize>(
   input: f32[n, m],
-  filter: f32[gs, gs],
+  filter: f32[7, 7],
 ) -> f32[n, m] {
   @res let result : f32[n, m];
 
@@ -13,12 +17,9 @@ fn gaussian_smoothing<n, m, gs : usize>(
 
       @filter_loop for i = 0 to gs {
         for j = 0 to gs {
-          let val = input[if row + i < gr               then 0
-                          else if row + i - gr > n - 1  then n - 1
-                                                        else row + i - gr,
-                          if col + j < gr               then 0
-                          else if col + j - gr > m - 1  then m - 1
-                                                        else col + j - gr];
+          let br = min!(max!(row + i, gr) - gr, n - 1);
+          let bc = min!(max!(col + j, gr) - gr, m - 1);
+          let val = input[br, bc];
           smoothed += val * filter[i, j];
         }
       }
@@ -33,9 +34,9 @@ fn gaussian_smoothing<n, m, gs : usize>(
 const MIN_BR : f32 = 0;
 const MAX_BR : f32 = 1;
 
-fn laplacian_estimate<n, m, sz: usize>(
+fn laplacian_estimate<n, m : usize>(
   input: f32[n, m],
-  structure: f32[sz, sz],
+  structure: f32[3, 3],
 ) -> f32[n, m] {
   const r = sz / 2;
 
@@ -77,9 +78,9 @@ fn laplacian_estimate<n, m, sz: usize>(
   return result;
 }
 
-fn zero_crossings<n, m, sz: usize>(
+fn zero_crossings<n, m : usize>(
   input: f32[n, m],
-  structure: f32[sz, sz],
+  structure: f32[3, 3],
 ) -> f32[n, m] {
   const r = sz / 2;
 
@@ -123,10 +124,10 @@ fn zero_crossings<n, m, sz: usize>(
   return result;
 }
 
-fn gradient<n, m, sb: usize>(
+fn gradient<n, m : usize>(
   input: f32[n, m],
-  sx: f32[sb, sb],
-  sy: f32[sb, sb],
+  sx: f32[3, 3],
+  sy: f32[3, 3],
 ) -> f32[n, m] {
   const sbr = sb / 2;
 
@@ -191,18 +192,18 @@ fn reject_zero_crossings<n, m: usize>(
 }
 
 #[entry]
-fn edge_detection<n, m, gs, sz, sb: usize>(
+fn edge_detection<n, m : usize>(
   input: f32[n, m],
   gaussian_filter: f32[gs, gs],
-  structure: f32[sz, sz],
-  sx: f32[sb, sb],
-  sy: f32[sb, sb],
+  structure: f32[3, 3],
+  sx: f32[3, 3],
+  sy: f32[3, 3],
   theta: f32,
 ) -> f32[n, m] {
-  let smoothed = gaussian_smoothing::<n, m, gs>(input, gaussian_filter);
-  @le let laplacian = laplacian_estimate::<n, m, sz>(smoothed, structure);
-  @zc let zcs = zero_crossings::<n, m, sz>(laplacian, structure);
-  let gradient = gradient::<n, m, sb>(smoothed, sx, sy);
+  let smoothed = gaussian_smoothing::<n, m>(input, gaussian_filter);
+  @le let laplacian = laplacian_estimate::<n, m>(smoothed, structure);
+  @zc let zcs = zero_crossings::<n, m>(laplacian, structure);
+  let gradient = gradient::<n, m>(smoothed, sx, sy);
   let maxgrad = max_gradient::<n, m>(gradient);
   return reject_zero_crossings::<n, m>(zcs, gradient, maxgrad, theta);
 }

juno_samples/edge_detection/src/lib.rs

@@ -194,9 +194,6 @@ pub fn edge_detection_harness(args: EdgeDetectionInputs) {
                 r.run(
                     height as u64,
                     width as u64,
-                    gs as u64,
-                    sz as u64,
-                    sb as u64,
                     input_h.to(),
                     gaussian_filter_h.to(),
                     structure_h.to(),