diff --git a/README.md b/README.md
index 12c3be0..ee9db88 100644
--- a/README.md
+++ b/README.md
@@ -33,7 +33,7 @@ BenchmarkJIT/rotate90                      : 0.02599ms
 BenchmarkJIT/rotate180                     : 0.00802ms
 BenchmarkJIT/rotate270                     : 0.02582ms
 BenchmarkJIT/crop                          : 0.06126ms
-BenchmarkJIT/scale                         : 0.14009ms
+BenchmarkJIT/scale                         : 0.13889ms
 BenchmarkJIT/scale_box                     : 0.20598ms
 BenchmarkJIT/scale_linear                  : 0.20440ms
 BenchmarkJIT/scale_gaussian                : 0.31444ms
@@ -621,7 +621,7 @@ ycbcr, err := blurry.ConvertToYUV444(rgba)
 | 420          | `blurry.ConvertToYUV420(*image.RGBA)` |
 | 444          | `blurry.ConvertToYUV444(*image.RGBA)` |
 
-#### PCM16 Decibel
+### PCM16 Decibel
 
 Gets the decibel of given PCM16.
 
@@ -662,7 +662,7 @@ USAGE:
    blurry [global options] command [command options] [arguments...]
 
 VERSION:
-   1.19.0
+   1.20.1
 
 COMMANDS:
      blend             
@@ -691,6 +691,7 @@ COMMANDS:
      laplacian         
      morphology        
      match_template    
+     pcm16             
      rotate            
      scale             
      sobel             
diff --git a/blurry.cpp b/blurry.cpp
index 5f7be33..4d4b3b5 100644
--- a/blurry.cpp
+++ b/blurry.cpp
@@ -1211,7 +1211,7 @@ Func scale(
   Var ti("ti");
 
   Region src_bounds = {{0, width},{0, height},{0, 4}};
-  Func in = readFloat(BoundaryConditions::constant_exterior(input, 0, src_bounds), "in");
+  Func in = BoundaryConditions::constant_exterior(readFloat(input, "in"), 0, src_bounds);
 
   Expr dx = cast<float>(width) / cast<float>(scale_width);
   Expr dy = cast<float>(height) / cast<float>(scale_height);
@@ -1220,23 +1220,28 @@ Func scale(
   scale(x, y, ch) = in(cast<int>((x + f05) * dx), cast<int>((y + f05) * dy), ch);
 
   Func f = Func("scale");
-  Expr value = select(
+  f(x, y, ch) = cast<uint8_t>(select(
     ch < 3, scale(x, y, ch),
     likely(float_255)
-  );
-  f(x, y, ch) = cast<uint8_t>(value);
-
-  scale.compute_at(f, ti)
-    .parallel(ch)
-    .vectorize(y, 8)
-    .vectorize(x, 8);
+  ));
+  f(x, y, 0) = cast<uint8_t>(scale(x, y, 0));
+  f(x, y, 1) = cast<uint8_t>(scale(x, y, 1));
+  f(x, y, 2) = cast<uint8_t>(scale(x, y, 2));
+  f(x, y, 3) = ui8_255;
 
-  f.compute_root()
+  scale.compute_root()
+    .parallel(y, 8)
+    .vectorize(x, 32);
+  f.compute_at(scale, ti)
+    .store_at(scale, ti)
     .tile(x, y, xo, yo, xi, yi, 32, 32)
     .fuse(xo, yo, ti)
-    .parallel(ch)
-    .parallel(ti)
+    .parallel(ti, 8)
     .vectorize(xi, 32);
+  f.update(0).unscheduled();
+  f.update(1).unscheduled();
+  f.update(2).unscheduled();
+  f.update(3).unscheduled();
   return f;
 }
 
@@ -1853,21 +1858,25 @@ Func boxblur(Func input, Expr width, Expr height, Expr size) {
 
   blur_x.compute_root()
     .parallel(ch)
-    .vectorize(x, 32);
+    .vectorize(x, 64);
   blur_y.compute_at(f, yi)
+    .store_at(f, yi)
     .parallel(ch)
-    .vectorize(x, 32);
+    .vectorize(x);
 
   f.compute_at(in, ti)
-    .tile(x, y, xo, yo, xi, yi, 32, 32)
+    .bound(ch, 0, 4).unroll(ch)
+    .bound(x, 0, width)
+    .bound(y, 0, height)
+    .store_root()
+    .tile(x, y, xo, yo, xi, yi, 64, 64)
     .fuse(xo, yo, ti)
-    .parallel(ch)
     .parallel(ti, 8)
-    .vectorize(xi, 32);
+    .vectorize(xi, 64);
 
   in.compute_at(blur_x, y)
     .parallel(ch)
-    .vectorize(x, 32);
+    .vectorize(x, 64);
   return f;
 }
 
diff --git a/include/libblend_diff_linux.h b/include/libblend_diff_linux.h
index 4d5226e..741973f 100644
--- a/include/libblend_diff_linux.h
+++ b/include/libblend_diff_linux.h
@@ -41,13 +41,13 @@ extern "C" {
 #endif
 
 HALIDE_FUNCTION_ATTRS
-int blend_diff_par_for_blend_diff_3_s0_ch_par_for_blend_diff_3_s0_x_xo_ti_ti__6(void *__user_context, int32_t _blend_diff__3_s0_x_xo_ti_ti, uint8_t *_closure_arg__118);
+int blend_diff_par_for_blend_diff_s0_ch_par_for_blend_diff_s0_x_xo_ti_ti__2(void *__user_context, int32_t _blend_diff_s0_x_xo_ti_ti, uint8_t *_closure_arg__118);
 HALIDE_FUNCTION_ATTRS
-int blend_diff_par_for_blend_diff_3_s0_ch_par_for_blend_diff_3_s0_x_xo_ti_ti__7(void *__user_context, int32_t _blend_diff__3_s0_x_xo_ti_ti, uint8_t *_closure_arg__119);
+int blend_diff_par_for_blend_diff_s0_ch_par_for_blend_diff_s0_x_xo_ti_ti__3(void *__user_context, int32_t _blend_diff_s0_x_xo_ti_ti, uint8_t *_closure_arg__119);
 HALIDE_FUNCTION_ATTRS
-int blend_diff_par_for_blend_diff_3_s0_ch__1(void *__user_context, int32_t _blend_diff__3_s0_ch, uint8_t *_closure_arg__117);
+int blend_diff_par_for_blend_diff_s0_ch__1(void *__user_context, int32_t _blend_diff_s0_ch, uint8_t *_closure_arg__117);
 HALIDE_FUNCTION_ATTRS
-int blend_diff(struct halide_buffer_t *_src0_buffer, int32_t _width0, int32_t _height0, struct halide_buffer_t *_src1_buffer, int32_t _width1, int32_t _height1, int32_t _px, int32_t _py, struct halide_buffer_t *_blend_diff__3_buffer);
+int blend_diff(struct halide_buffer_t *_src0_buffer, int32_t _width0, int32_t _height0, struct halide_buffer_t *_src1_buffer, int32_t _width1, int32_t _height1, int32_t _px, int32_t _py, struct halide_buffer_t *_blend_diff_buffer);
 
 HALIDE_FUNCTION_ATTRS
 int blend_diff_argv(void **args);
diff --git a/include/libblend_diff_osx.h b/include/libblend_diff_osx.h
index 8f76031..837c823 100644
--- a/include/libblend_diff_osx.h
+++ b/include/libblend_diff_osx.h
@@ -41,13 +41,13 @@ extern "C" {
 #endif
 
 HALIDE_FUNCTION_ATTRS
-int blend_diff_par_for_blend_diff_3_s0_ch_par_for_blend_diff_3_s0_x_xo_ti_ti__4(void *__user_context, int32_t _blend_diff__3_s0_x_xo_ti_ti, uint8_t *_closure_arg__115);
+int blend_diff_par_for_blend_diff_s0_ch_par_for_blend_diff_s0_x_xo_ti_ti(void *__user_context, int32_t _blend_diff_s0_x_xo_ti_ti, uint8_t *_closure_arg__115);
 HALIDE_FUNCTION_ATTRS
-int blend_diff_par_for_blend_diff_3_s0_ch_par_for_blend_diff_3_s0_x_xo_ti_ti__5(void *__user_context, int32_t _blend_diff__3_s0_x_xo_ti_ti, uint8_t *_closure_arg__116);
+int blend_diff_par_for_blend_diff_s0_ch_par_for_blend_diff_s0_x_xo_ti_ti__1(void *__user_context, int32_t _blend_diff_s0_x_xo_ti_ti, uint8_t *_closure_arg__116);
 HALIDE_FUNCTION_ATTRS
-int blend_diff_par_for_blend_diff__3_s0_ch(void *__user_context, int32_t _blend_diff__3_s0_ch, uint8_t *_closure_arg__114);
+int blend_diff_par_for_blend_diff_s0_ch(void *__user_context, int32_t _blend_diff_s0_ch, uint8_t *_closure_arg__114);
 HALIDE_FUNCTION_ATTRS
-int blend_diff(struct halide_buffer_t *_src0_buffer, int32_t _width0, int32_t _height0, struct halide_buffer_t *_src1_buffer, int32_t _width1, int32_t _height1, int32_t _px, int32_t _py, struct halide_buffer_t *_blend_diff__3_buffer);
+int blend_diff(struct halide_buffer_t *_src0_buffer, int32_t _width0, int32_t _height0, struct halide_buffer_t *_src1_buffer, int32_t _width1, int32_t _height1, int32_t _px, int32_t _py, struct halide_buffer_t *_blend_diff_buffer);
 
 HALIDE_FUNCTION_ATTRS
 int blend_diff_argv(void **args);
diff --git a/include/libblend_normal_linux.h b/include/libblend_normal_linux.h
index e54cab6..23968e7 100644
--- a/include/libblend_normal_linux.h
+++ b/include/libblend_normal_linux.h
@@ -41,9 +41,9 @@ extern "C" {
 #endif
 
 HALIDE_FUNCTION_ATTRS
-int blend_normal_par_for_blend_normal_s0_ch_par_for_blend_normal_s0_x_xo_ti_ti__2(void *__user_context, int32_t _blend_normal_s0_x_xo_ti_ti, uint8_t *_closure_arg__100);
+int blend_normal_par_for_blend_normal_s0_ch_par_for_blend_normal_s0_x_xo_ti_ti__6(void *__user_context, int32_t _blend_normal_s0_x_xo_ti_ti, uint8_t *_closure_arg__100);
 HALIDE_FUNCTION_ATTRS
-int blend_normal_par_for_blend_normal_s0_ch_par_for_blend_normal_s0_x_xo_ti_ti__3(void *__user_context, int32_t _blend_normal_s0_x_xo_ti_ti, uint8_t *_closure_arg__101);
+int blend_normal_par_for_blend_normal_s0_ch_par_for_blend_normal_s0_x_xo_ti_ti__7(void *__user_context, int32_t _blend_normal_s0_x_xo_ti_ti, uint8_t *_closure_arg__101);
 HALIDE_FUNCTION_ATTRS
 int blend_normal_par_for_blend_normal_s0_ch__1(void *__user_context, int32_t _blend_normal_s0_ch, uint8_t *_closure_arg__99);
 HALIDE_FUNCTION_ATTRS
diff --git a/include/libblend_normal_osx.h b/include/libblend_normal_osx.h
index d550d4d..28ac9e7 100644
--- a/include/libblend_normal_osx.h
+++ b/include/libblend_normal_osx.h
@@ -41,9 +41,9 @@ extern "C" {
 #endif
 
 HALIDE_FUNCTION_ATTRS
-int blend_normal_par_for_blend_normal_s0_ch_par_for_blend_normal_s0_x_xo_ti_ti(void *__user_context, int32_t _blend_normal_s0_x_xo_ti_ti, uint8_t *_closure_arg__97);
+int blend_normal_par_for_blend_normal_s0_ch_par_for_blend_normal_s0_x_xo_ti_ti__4(void *__user_context, int32_t _blend_normal_s0_x_xo_ti_ti, uint8_t *_closure_arg__97);
 HALIDE_FUNCTION_ATTRS
-int blend_normal_par_for_blend_normal_s0_ch_par_for_blend_normal_s0_x_xo_ti_ti__1(void *__user_context, int32_t _blend_normal_s0_x_xo_ti_ti, uint8_t *_closure_arg__98);
+int blend_normal_par_for_blend_normal_s0_ch_par_for_blend_normal_s0_x_xo_ti_ti__5(void *__user_context, int32_t _blend_normal_s0_x_xo_ti_ti, uint8_t *_closure_arg__98);
 HALIDE_FUNCTION_ATTRS
 int blend_normal_par_for_blend_normal_s0_ch(void *__user_context, int32_t _blend_normal_s0_ch, uint8_t *_closure_arg__96);
 HALIDE_FUNCTION_ATTRS
diff --git a/include/libblend_sub_linux.h b/include/libblend_sub_linux.h
index 7fb2a9a..d7bd6a3 100644
--- a/include/libblend_sub_linux.h
+++ b/include/libblend_sub_linux.h
@@ -41,9 +41,9 @@ extern "C" {
 #endif
 
 HALIDE_FUNCTION_ATTRS
-int blend_sub_par_for_blend_sub_s0_ch_par_for_blend_sub_s0_x_xo_ti_ti__6(void *__user_context, int32_t _blend_sub_s0_x_xo_ti_ti, uint8_t *_closure_arg__106);
+int blend_sub_par_for_blend_sub_s0_ch_par_for_blend_sub_s0_x_xo_ti_ti__2(void *__user_context, int32_t _blend_sub_s0_x_xo_ti_ti, uint8_t *_closure_arg__106);
 HALIDE_FUNCTION_ATTRS
-int blend_sub_par_for_blend_sub_s0_ch_par_for_blend_sub_s0_x_xo_ti_ti__7(void *__user_context, int32_t _blend_sub_s0_x_xo_ti_ti, uint8_t *_closure_arg__107);
+int blend_sub_par_for_blend_sub_s0_ch_par_for_blend_sub_s0_x_xo_ti_ti__3(void *__user_context, int32_t _blend_sub_s0_x_xo_ti_ti, uint8_t *_closure_arg__107);
 HALIDE_FUNCTION_ATTRS
 int blend_sub_par_for_blend_sub_s0_ch__1(void *__user_context, int32_t _blend_sub_s0_ch, uint8_t *_closure_arg__105);
 HALIDE_FUNCTION_ATTRS
diff --git a/include/libblend_sub_osx.h b/include/libblend_sub_osx.h
index dc838b8..20cb101 100644
--- a/include/libblend_sub_osx.h
+++ b/include/libblend_sub_osx.h
@@ -41,9 +41,9 @@ extern "C" {
 #endif
 
 HALIDE_FUNCTION_ATTRS
-int blend_sub_par_for_blend_sub_s0_ch_par_for_blend_sub_s0_x_xo_ti_ti__4(void *__user_context, int32_t _blend_sub_s0_x_xo_ti_ti, uint8_t *_closure_arg__103);
+int blend_sub_par_for_blend_sub_s0_ch_par_for_blend_sub_s0_x_xo_ti_ti(void *__user_context, int32_t _blend_sub_s0_x_xo_ti_ti, uint8_t *_closure_arg__103);
 HALIDE_FUNCTION_ATTRS
-int blend_sub_par_for_blend_sub_s0_ch_par_for_blend_sub_s0_x_xo_ti_ti__5(void *__user_context, int32_t _blend_sub_s0_x_xo_ti_ti, uint8_t *_closure_arg__104);
+int blend_sub_par_for_blend_sub_s0_ch_par_for_blend_sub_s0_x_xo_ti_ti__1(void *__user_context, int32_t _blend_sub_s0_x_xo_ti_ti, uint8_t *_closure_arg__104);
 HALIDE_FUNCTION_ATTRS
 int blend_sub_par_for_blend_sub_s0_ch(void *__user_context, int32_t _blend_sub_s0_ch, uint8_t *_closure_arg__102);
 HALIDE_FUNCTION_ATTRS
diff --git a/include/libblockmozaic_linux.h b/include/libblockmozaic_linux.h
index e97260d..bd9dc46 100644
--- a/include/libblockmozaic_linux.h
+++ b/include/libblockmozaic_linux.h
@@ -41,15 +41,15 @@ extern "C" {
 #endif
 
 HALIDE_FUNCTION_ATTRS
-int blockmozaic_par_for_avg_color_s0_y_y__6(void *__user_context, int32_t _avg_color_s0_y_y, uint8_t *_closure_arg__166);
+int blockmozaic_par_for_avg_color_s0_y_y__2(void *__user_context, int32_t _avg_color_s0_y_y, uint8_t *_closure_arg__170);
 HALIDE_FUNCTION_ATTRS
-int blockmozaic_par_for_avg_color_s0_y_y__7(void *__user_context, int32_t _avg_color_s0_y_y, uint8_t *_closure_arg__167);
+int blockmozaic_par_for_avg_color_s0_y_y__3(void *__user_context, int32_t _avg_color_s0_y_y, uint8_t *_closure_arg__171);
 HALIDE_FUNCTION_ATTRS
-int blockmozaic_par_for_blockmozaic_1_s0_ch_par_for_blockmozaic_1_s0_x_xo_ti_ti__1(void *__user_context, int32_t _blockmozaic__1_s0_x_xo_ti_ti, uint8_t *_closure_arg__169);
+int blockmozaic_par_for_blockmozaic_s0_ch_par_for_blockmozaic_s0_x_xo_ti_ti__1(void *__user_context, int32_t _blockmozaic_s0_x_xo_ti_ti, uint8_t *_closure_arg__173);
 HALIDE_FUNCTION_ATTRS
-int blockmozaic_par_for_blockmozaic_1_s0_ch__1(void *__user_context, int32_t _blockmozaic__1_s0_ch, uint8_t *_closure_arg__168);
+int blockmozaic_par_for_blockmozaic_s0_ch__1(void *__user_context, int32_t _blockmozaic_s0_ch, uint8_t *_closure_arg__172);
 HALIDE_FUNCTION_ATTRS
-int blockmozaic(struct halide_buffer_t *_src_buffer, int32_t _width, int32_t _height, int32_t _block, struct halide_buffer_t *_blockmozaic__1_buffer);
+int blockmozaic(struct halide_buffer_t *_src_buffer, int32_t _width, int32_t _height, int32_t _block, struct halide_buffer_t *_blockmozaic_buffer);
 
 HALIDE_FUNCTION_ATTRS
 int blockmozaic_argv(void **args);
diff --git a/include/libblockmozaic_osx.h b/include/libblockmozaic_osx.h
index f32bbcb..469a4aa 100644
--- a/include/libblockmozaic_osx.h
+++ b/include/libblockmozaic_osx.h
@@ -41,15 +41,15 @@ extern "C" {
 #endif
 
 HALIDE_FUNCTION_ATTRS
-int blockmozaic_par_for_avg_color_s0_y_y__4(void *__user_context, int32_t _avg_color_s0_y_y, uint8_t *_closure_arg__162);
+int blockmozaic_par_for_avg_color_s0_y_y(void *__user_context, int32_t _avg_color_s0_y_y, uint8_t *_closure_arg__166);
 HALIDE_FUNCTION_ATTRS
-int blockmozaic_par_for_avg_color_s0_y_y__5(void *__user_context, int32_t _avg_color_s0_y_y, uint8_t *_closure_arg__163);
+int blockmozaic_par_for_avg_color_s0_y_y__1(void *__user_context, int32_t _avg_color_s0_y_y, uint8_t *_closure_arg__167);
 HALIDE_FUNCTION_ATTRS
-int blockmozaic_par_for_blockmozaic__1_s0_ch_par_for_blockmozaic__1_s0_x_xo_ti_ti(void *__user_context, int32_t _blockmozaic__1_s0_x_xo_ti_ti, uint8_t *_closure_arg__165);
+int blockmozaic_par_for_blockmozaic_s0_ch_par_for_blockmozaic_s0_x_xo_ti_ti(void *__user_context, int32_t _blockmozaic_s0_x_xo_ti_ti, uint8_t *_closure_arg__169);
 HALIDE_FUNCTION_ATTRS
-int blockmozaic_par_for_blockmozaic__1_s0_ch(void *__user_context, int32_t _blockmozaic__1_s0_ch, uint8_t *_closure_arg__164);
+int blockmozaic_par_for_blockmozaic_s0_ch(void *__user_context, int32_t _blockmozaic_s0_ch, uint8_t *_closure_arg__168);
 HALIDE_FUNCTION_ATTRS
-int blockmozaic(struct halide_buffer_t *_src_buffer, int32_t _width, int32_t _height, int32_t _block, struct halide_buffer_t *_blockmozaic__1_buffer);
+int blockmozaic(struct halide_buffer_t *_src_buffer, int32_t _width, int32_t _height, int32_t _block, struct halide_buffer_t *_blockmozaic_buffer);
 
 HALIDE_FUNCTION_ATTRS
 int blockmozaic_argv(void **args);
diff --git a/include/libboxblur_linux.h b/include/libboxblur_linux.h
index 5e2d877..3c0e4df 100644
--- a/include/libboxblur_linux.h
+++ b/include/libboxblur_linux.h
@@ -41,11 +41,15 @@ extern "C" {
 #endif
 
 HALIDE_FUNCTION_ATTRS
-int boxblur_par_for_blur_x_s0_ch__1(void *__user_context, int32_t _blur_x_s0_ch, uint8_t *_closure_arg__155);
+int boxblur_par_for_blur_x_s0_ch__1(void *__user_context, int32_t _blur_x_s0_ch, uint8_t *_closure_arg__157);
 HALIDE_FUNCTION_ATTRS
-int boxblur_par_for_boxblur_s0_ch_par_for_boxblur_s0_x_xo_ti_ti__1(void *__user_context, int32_t _boxblur_s0_x_xo_ti_ti, uint8_t *_closure_arg__157);
+int boxblur_par_for_boxblur_s0_x_xo_ti_ti__4(void *__user_context, int32_t _boxblur_s0_x_xo_ti_ti, uint8_t *_closure_arg__158);
 HALIDE_FUNCTION_ATTRS
-int boxblur_par_for_boxblur_s0_ch__1(void *__user_context, int32_t _boxblur_s0_ch, uint8_t *_closure_arg__156);
+int boxblur_par_for_boxblur_s0_x_xo_ti_ti__5(void *__user_context, int32_t _boxblur_s0_x_xo_ti_ti, uint8_t *_closure_arg__159);
+HALIDE_FUNCTION_ATTRS
+int boxblur_par_for_boxblur_s0_x_xo_ti_ti__6(void *__user_context, int32_t _boxblur_s0_x_xo_ti_ti, uint8_t *_closure_arg__160);
+HALIDE_FUNCTION_ATTRS
+int boxblur_par_for_boxblur_s0_x_xo_ti_ti__7(void *__user_context, int32_t _boxblur_s0_x_xo_ti_ti, uint8_t *_closure_arg__161);
 HALIDE_FUNCTION_ATTRS
 int boxblur(struct halide_buffer_t *_src_buffer, int32_t _width, int32_t _height, uint8_t _size, struct halide_buffer_t *_boxblur_buffer);
 
diff --git a/include/libboxblur_osx.h b/include/libboxblur_osx.h
index 760ae9b..0af1b2e 100644
--- a/include/libboxblur_osx.h
+++ b/include/libboxblur_osx.h
@@ -43,9 +43,13 @@ extern "C" {
 HALIDE_FUNCTION_ATTRS
 int boxblur_par_for_blur_x_s0_ch(void *__user_context, int32_t _blur_x_s0_ch, uint8_t *_closure_arg__152);
 HALIDE_FUNCTION_ATTRS
-int boxblur_par_for_boxblur_s0_ch_par_for_boxblur_s0_x_xo_ti_ti(void *__user_context, int32_t _boxblur_s0_x_xo_ti_ti, uint8_t *_closure_arg__154);
+int boxblur_par_for_boxblur_s0_x_xo_ti_ti(void *__user_context, int32_t _boxblur_s0_x_xo_ti_ti, uint8_t *_closure_arg__153);
 HALIDE_FUNCTION_ATTRS
-int boxblur_par_for_boxblur_s0_ch(void *__user_context, int32_t _boxblur_s0_ch, uint8_t *_closure_arg__153);
+int boxblur_par_for_boxblur_s0_x_xo_ti_ti__1(void *__user_context, int32_t _boxblur_s0_x_xo_ti_ti, uint8_t *_closure_arg__154);
+HALIDE_FUNCTION_ATTRS
+int boxblur_par_for_boxblur_s0_x_xo_ti_ti__2(void *__user_context, int32_t _boxblur_s0_x_xo_ti_ti, uint8_t *_closure_arg__155);
+HALIDE_FUNCTION_ATTRS
+int boxblur_par_for_boxblur_s0_x_xo_ti_ti__3(void *__user_context, int32_t _boxblur_s0_x_xo_ti_ti, uint8_t *_closure_arg__156);
 HALIDE_FUNCTION_ATTRS
 int boxblur(struct halide_buffer_t *_src_buffer, int32_t _width, int32_t _height, uint8_t _size, struct halide_buffer_t *_boxblur_buffer);
 
diff --git a/include/libbrightness_linux.h b/include/libbrightness_linux.h
index 8bca3bf..8feb744 100644
--- a/include/libbrightness_linux.h
+++ b/include/libbrightness_linux.h
@@ -41,7 +41,7 @@ extern "C" {
 #endif
 
 HALIDE_FUNCTION_ATTRS
-int brightness_par_for_in_22_s0_ch__1(void *__user_context, int32_t _in__22_s0_ch, uint8_t *_closure_arg__139);
+int brightness_par_for_in_22_s0_ch__5(void *__user_context, int32_t _in__22_s0_ch, uint8_t *_closure_arg__139);
 HALIDE_FUNCTION_ATTRS
 int brightness_par_for_brightness_s0_ch_par_for_brightness_s0_x_xo_ti_ti__1(void *__user_context, int32_t _brightness_s0_x_xo_ti_ti, uint8_t *_closure_arg__141);
 HALIDE_FUNCTION_ATTRS
diff --git a/include/libbrightness_osx.h b/include/libbrightness_osx.h
index a445b9b..41e5bf1 100644
--- a/include/libbrightness_osx.h
+++ b/include/libbrightness_osx.h
@@ -41,7 +41,7 @@ extern "C" {
 #endif
 
 HALIDE_FUNCTION_ATTRS
-int brightness_par_for_in__22_s0_ch(void *__user_context, int32_t _in__22_s0_ch, uint8_t *_closure_arg__136);
+int brightness_par_for_in_22_s0_ch__4(void *__user_context, int32_t _in__22_s0_ch, uint8_t *_closure_arg__136);
 HALIDE_FUNCTION_ATTRS
 int brightness_par_for_brightness_s0_ch_par_for_brightness_s0_x_xo_ti_ti(void *__user_context, int32_t _brightness_s0_x_xo_ti_ti, uint8_t *_closure_arg__138);
 HALIDE_FUNCTION_ATTRS
diff --git a/include/libcanny_dilate_linux.h b/include/libcanny_dilate_linux.h
index 8929f9a..4143a28 100644
--- a/include/libcanny_dilate_linux.h
+++ b/include/libcanny_dilate_linux.h
@@ -41,9 +41,9 @@ extern "C" {
 #endif
 
 HALIDE_FUNCTION_ATTRS
-int canny_dilate_par_for_hysteresis_1_s0_x_xo_ti__1(void *__user_context, int32_t _hysteresis__1_s0_x_xo_ti, uint8_t *_closure_arg__258);
+int canny_dilate_par_for_hysteresis_1_s0_x_xo_ti__1(void *__user_context, int32_t _hysteresis__1_s0_x_xo_ti, uint8_t *_closure_arg__262);
 HALIDE_FUNCTION_ATTRS
-int canny_dilate_par_for_canny_dilate_s0_ch__1(void *__user_context, int32_t _canny_dilate_s0_ch, uint8_t *_closure_arg__259);
+int canny_dilate_par_for_canny_dilate_s0_ch__1(void *__user_context, int32_t _canny_dilate_s0_ch, uint8_t *_closure_arg__263);
 HALIDE_FUNCTION_ATTRS
 int canny_dilate(struct halide_buffer_t *_src_buffer, int32_t _width, int32_t _height, int32_t _threshold_max, int32_t _threshold_min, int32_t _dilate, struct halide_buffer_t *_canny_dilate_buffer);
 
diff --git a/include/libcanny_dilate_osx.h b/include/libcanny_dilate_osx.h
index 3e8fbd5..370db76 100644
--- a/include/libcanny_dilate_osx.h
+++ b/include/libcanny_dilate_osx.h
@@ -41,9 +41,9 @@ extern "C" {
 #endif
 
 HALIDE_FUNCTION_ATTRS
-int canny_dilate_par_for_hysteresis__1_s0_x_xo_ti(void *__user_context, int32_t _hysteresis__1_s0_x_xo_ti, uint8_t *_closure_arg__256);
+int canny_dilate_par_for_hysteresis__1_s0_x_xo_ti(void *__user_context, int32_t _hysteresis__1_s0_x_xo_ti, uint8_t *_closure_arg__260);
 HALIDE_FUNCTION_ATTRS
-int canny_dilate_par_for_canny_dilate_s0_ch(void *__user_context, int32_t _canny_dilate_s0_ch, uint8_t *_closure_arg__257);
+int canny_dilate_par_for_canny_dilate_s0_ch(void *__user_context, int32_t _canny_dilate_s0_ch, uint8_t *_closure_arg__261);
 HALIDE_FUNCTION_ATTRS
 int canny_dilate(struct halide_buffer_t *_src_buffer, int32_t _width, int32_t _height, int32_t _threshold_max, int32_t _threshold_min, int32_t _dilate, struct halide_buffer_t *_canny_dilate_buffer);
 
diff --git a/include/libcanny_linux.h b/include/libcanny_linux.h
index 9a95b6d..07852f4 100644
--- a/include/libcanny_linux.h
+++ b/include/libcanny_linux.h
@@ -41,11 +41,11 @@ extern "C" {
 #endif
 
 HALIDE_FUNCTION_ATTRS
-int canny_par_for_in_39_s0_y_rebased__7(void *__user_context, int32_t _in__39_s0_y_rebased, uint8_t *_closure_arg__253);
+int canny_par_for_in_39_s0_y_rebased__4(void *__user_context, int32_t _in__39_s0_y_rebased, uint8_t *_closure_arg__257);
 HALIDE_FUNCTION_ATTRS
-int canny_par_for_hysteresis_s0_x_xo_ti__1(void *__user_context, int32_t _hysteresis_s0_x_xo_ti, uint8_t *_closure_arg__254);
+int canny_par_for_hysteresis_s0_x_xo_ti__1(void *__user_context, int32_t _hysteresis_s0_x_xo_ti, uint8_t *_closure_arg__258);
 HALIDE_FUNCTION_ATTRS
-int canny_par_for_canny_s0_ch__1(void *__user_context, int32_t _canny_s0_ch, uint8_t *_closure_arg__255);
+int canny_par_for_canny_s0_ch__1(void *__user_context, int32_t _canny_s0_ch, uint8_t *_closure_arg__259);
 HALIDE_FUNCTION_ATTRS
 int canny(struct halide_buffer_t *_src_buffer, int32_t _width, int32_t _height, int32_t _threshold_max, int32_t _threshold_min, struct halide_buffer_t *_canny_buffer);
 
diff --git a/include/libcanny_morphology_close_linux.h b/include/libcanny_morphology_close_linux.h
index 44bef61..04232dc 100644
--- a/include/libcanny_morphology_close_linux.h
+++ b/include/libcanny_morphology_close_linux.h
@@ -41,13 +41,13 @@ extern "C" {
 #endif
 
 HALIDE_FUNCTION_ATTRS
-int canny_morphology_close_par_for_dilate_tmp_6_s0_y_y__1(void *__user_context, int32_t _dilate_tmp__6_s0_y_y, uint8_t *_closure_arg__272);
+int canny_morphology_close_par_for_dilate_tmp_3_s0_y_y__1(void *__user_context, int32_t _dilate_tmp__3_s0_y_y, uint8_t *_closure_arg__297);
 HALIDE_FUNCTION_ATTRS
-int canny_morphology_close_par_for_erode_tmp_3_s0_y_y__5(void *__user_context, int32_t _erode_tmp__3_s0_y_y, uint8_t *_closure_arg__273);
+int canny_morphology_close_par_for_erode_tmp_3_s0_y_y__1(void *__user_context, int32_t _erode_tmp__3_s0_y_y, uint8_t *_closure_arg__298);
 HALIDE_FUNCTION_ATTRS
-int canny_morphology_close_par_for_hysteresis_3_s0_x_xo_ti__1(void *__user_context, int32_t _hysteresis__3_s0_x_xo_ti, uint8_t *_closure_arg__274);
+int canny_morphology_close_par_for_hysteresis_3_s0_x_xo_ti__1(void *__user_context, int32_t _hysteresis__3_s0_x_xo_ti, uint8_t *_closure_arg__299);
 HALIDE_FUNCTION_ATTRS
-int canny_morphology_close_par_for_canny_morphology_close_s0_ch__1(void *__user_context, int32_t _canny_morphology_close_s0_ch, uint8_t *_closure_arg__275);
+int canny_morphology_close_par_for_canny_morphology_close_s0_ch__1(void *__user_context, int32_t _canny_morphology_close_s0_ch, uint8_t *_closure_arg__300);
 HALIDE_FUNCTION_ATTRS
 int canny_morphology_close(struct halide_buffer_t *_src_buffer, int32_t _width, int32_t _height, int32_t _threshold_max, int32_t _threshold_min, int32_t _morphology_size, int32_t _dilate, struct halide_buffer_t *_canny_morphology_close_buffer);
 
diff --git a/include/libcanny_morphology_close_osx.h b/include/libcanny_morphology_close_osx.h
index 6a0b036..bfff9eb 100644
--- a/include/libcanny_morphology_close_osx.h
+++ b/include/libcanny_morphology_close_osx.h
@@ -41,13 +41,13 @@ extern "C" {
 #endif
 
 HALIDE_FUNCTION_ATTRS
-int canny_morphology_close_par_for_dilate_tmp__6_s0_y_y(void *__user_context, int32_t _dilate_tmp__6_s0_y_y, uint8_t *_closure_arg__268);
+int canny_morphology_close_par_for_dilate_tmp__3_s0_y_y(void *__user_context, int32_t _dilate_tmp__3_s0_y_y, uint8_t *_closure_arg__293);
 HALIDE_FUNCTION_ATTRS
-int canny_morphology_close_par_for_erode_tmp_3_s0_y_y__4(void *__user_context, int32_t _erode_tmp__3_s0_y_y, uint8_t *_closure_arg__269);
+int canny_morphology_close_par_for_erode_tmp__3_s0_y_y(void *__user_context, int32_t _erode_tmp__3_s0_y_y, uint8_t *_closure_arg__294);
 HALIDE_FUNCTION_ATTRS
-int canny_morphology_close_par_for_hysteresis__3_s0_x_xo_ti(void *__user_context, int32_t _hysteresis__3_s0_x_xo_ti, uint8_t *_closure_arg__270);
+int canny_morphology_close_par_for_hysteresis__3_s0_x_xo_ti(void *__user_context, int32_t _hysteresis__3_s0_x_xo_ti, uint8_t *_closure_arg__295);
 HALIDE_FUNCTION_ATTRS
-int canny_morphology_close_par_for_canny_morphology_close_s0_ch(void *__user_context, int32_t _canny_morphology_close_s0_ch, uint8_t *_closure_arg__271);
+int canny_morphology_close_par_for_canny_morphology_close_s0_ch(void *__user_context, int32_t _canny_morphology_close_s0_ch, uint8_t *_closure_arg__296);
 HALIDE_FUNCTION_ATTRS
 int canny_morphology_close(struct halide_buffer_t *_src_buffer, int32_t _width, int32_t _height, int32_t _threshold_max, int32_t _threshold_min, int32_t _morphology_size, int32_t _dilate, struct halide_buffer_t *_canny_morphology_close_buffer);
 
diff --git a/include/libcanny_morphology_open_linux.h b/include/libcanny_morphology_open_linux.h
index c4efa4c..2c9442f 100644
--- a/include/libcanny_morphology_open_linux.h
+++ b/include/libcanny_morphology_open_linux.h
@@ -41,13 +41,13 @@ extern "C" {
 #endif
 
 HALIDE_FUNCTION_ATTRS
-int canny_morphology_open_par_for_erode_tmp_2_s0_y_y__5(void *__user_context, int32_t _erode_tmp__2_s0_y_y, uint8_t *_closure_arg__264);
+int canny_morphology_open_par_for_erode_tmp_2_s0_y_y__1(void *__user_context, int32_t _erode_tmp__2_s0_y_y, uint8_t *_closure_arg__268);
 HALIDE_FUNCTION_ATTRS
-int canny_morphology_open_par_for_dilate_tmp_5_s0_y_y__1(void *__user_context, int32_t _dilate_tmp__5_s0_y_y, uint8_t *_closure_arg__265);
+int canny_morphology_open_par_for_dilate_tmp_2_s0_y_y__1(void *__user_context, int32_t _dilate_tmp__2_s0_y_y, uint8_t *_closure_arg__269);
 HALIDE_FUNCTION_ATTRS
-int canny_morphology_open_par_for_hysteresis_2_s0_x_xo_ti__1(void *__user_context, int32_t _hysteresis__2_s0_x_xo_ti, uint8_t *_closure_arg__266);
+int canny_morphology_open_par_for_hysteresis_2_s0_x_xo_ti__1(void *__user_context, int32_t _hysteresis__2_s0_x_xo_ti, uint8_t *_closure_arg__270);
 HALIDE_FUNCTION_ATTRS
-int canny_morphology_open_par_for_canny_morphology_open_4_s0_ch__1(void *__user_context, int32_t _canny_morphology_open__4_s0_ch, uint8_t *_closure_arg__267);
+int canny_morphology_open_par_for_canny_morphology_open_4_s0_ch__1(void *__user_context, int32_t _canny_morphology_open__4_s0_ch, uint8_t *_closure_arg__271);
 HALIDE_FUNCTION_ATTRS
 int canny_morphology_open(struct halide_buffer_t *_src_buffer, int32_t _width, int32_t _height, int32_t _threshold_max, int32_t _threshold_min, int32_t _morphology_size, int32_t _dilate, struct halide_buffer_t *_canny_morphology_open__4_buffer);
 
diff --git a/include/libcanny_morphology_open_osx.h b/include/libcanny_morphology_open_osx.h
index 969cd8f..3f141b3 100644
--- a/include/libcanny_morphology_open_osx.h
+++ b/include/libcanny_morphology_open_osx.h
@@ -41,13 +41,13 @@ extern "C" {
 #endif
 
 HALIDE_FUNCTION_ATTRS
-int canny_morphology_open_par_for_erode_tmp_2_s0_y_y__4(void *__user_context, int32_t _erode_tmp__2_s0_y_y, uint8_t *_closure_arg__260);
+int canny_morphology_open_par_for_erode_tmp__2_s0_y_y(void *__user_context, int32_t _erode_tmp__2_s0_y_y, uint8_t *_closure_arg__264);
 HALIDE_FUNCTION_ATTRS
-int canny_morphology_open_par_for_dilate_tmp__5_s0_y_y(void *__user_context, int32_t _dilate_tmp__5_s0_y_y, uint8_t *_closure_arg__261);
+int canny_morphology_open_par_for_dilate_tmp__2_s0_y_y(void *__user_context, int32_t _dilate_tmp__2_s0_y_y, uint8_t *_closure_arg__265);
 HALIDE_FUNCTION_ATTRS
-int canny_morphology_open_par_for_hysteresis__2_s0_x_xo_ti(void *__user_context, int32_t _hysteresis__2_s0_x_xo_ti, uint8_t *_closure_arg__262);
+int canny_morphology_open_par_for_hysteresis__2_s0_x_xo_ti(void *__user_context, int32_t _hysteresis__2_s0_x_xo_ti, uint8_t *_closure_arg__266);
 HALIDE_FUNCTION_ATTRS
-int canny_morphology_open_par_for_canny_morphology_open__4_s0_ch(void *__user_context, int32_t _canny_morphology_open__4_s0_ch, uint8_t *_closure_arg__263);
+int canny_morphology_open_par_for_canny_morphology_open__4_s0_ch(void *__user_context, int32_t _canny_morphology_open__4_s0_ch, uint8_t *_closure_arg__267);
 HALIDE_FUNCTION_ATTRS
 int canny_morphology_open(struct halide_buffer_t *_src_buffer, int32_t _width, int32_t _height, int32_t _threshold_max, int32_t _threshold_min, int32_t _morphology_size, int32_t _dilate, struct halide_buffer_t *_canny_morphology_open__4_buffer);
 
diff --git a/include/libcanny_osx.h b/include/libcanny_osx.h
index b862012..5fffe38 100644
--- a/include/libcanny_osx.h
+++ b/include/libcanny_osx.h
@@ -41,11 +41,11 @@ extern "C" {
 #endif
 
 HALIDE_FUNCTION_ATTRS
-int canny_par_for_in_39_s0_y_rebased__6(void *__user_context, int32_t _in__39_s0_y_rebased, uint8_t *_closure_arg__250);
+int canny_par_for_in_39_s0_y_rebased__3(void *__user_context, int32_t _in__39_s0_y_rebased, uint8_t *_closure_arg__254);
 HALIDE_FUNCTION_ATTRS
-int canny_par_for_hysteresis_s0_x_xo_ti(void *__user_context, int32_t _hysteresis_s0_x_xo_ti, uint8_t *_closure_arg__251);
+int canny_par_for_hysteresis_s0_x_xo_ti(void *__user_context, int32_t _hysteresis_s0_x_xo_ti, uint8_t *_closure_arg__255);
 HALIDE_FUNCTION_ATTRS
-int canny_par_for_canny_s0_ch(void *__user_context, int32_t _canny_s0_ch, uint8_t *_closure_arg__252);
+int canny_par_for_canny_s0_ch(void *__user_context, int32_t _canny_s0_ch, uint8_t *_closure_arg__256);
 HALIDE_FUNCTION_ATTRS
 int canny(struct halide_buffer_t *_src_buffer, int32_t _width, int32_t _height, int32_t _threshold_max, int32_t _threshold_min, struct halide_buffer_t *_canny_buffer);
 
diff --git a/include/libcontour_linux.h b/include/libcontour_linux.h
index 555494f..0da045d 100644
--- a/include/libcontour_linux.h
+++ b/include/libcontour_linux.h
@@ -41,9 +41,9 @@ extern "C" {
 #endif
 
 HALIDE_FUNCTION_ATTRS
-int contour_par_for_sobel_5_s0_x_xo_ti_ti__1(void *__user_context, int32_t _sobel__5_s0_x_xo_ti_ti, uint8_t *_closure_arg__302);
+int contour_par_for_sobel_5_s0_x_xo_ti_ti__1(void *__user_context, int32_t _sobel__5_s0_x_xo_ti_ti, uint8_t *_closure_arg__327);
 HALIDE_FUNCTION_ATTRS
-int contour_par_for_countor_line_s0_x_xo_ti__1(void *__user_context, int32_t _countor_line_s0_x_xo_ti, uint8_t *_closure_arg__303);
+int contour_par_for_countor_line_s0_x_xo_ti__1(void *__user_context, int32_t _countor_line_s0_x_xo_ti, uint8_t *_closure_arg__328);
 HALIDE_FUNCTION_ATTRS
 int contour(struct halide_buffer_t *_src_buffer, int32_t _width, int32_t _height, uint8_t _threshold, uint8_t _size, struct halide_buffer_t *_contour_buffer);
 
diff --git a/include/libcontour_osx.h b/include/libcontour_osx.h
index 0c27451..5fe626a 100644
--- a/include/libcontour_osx.h
+++ b/include/libcontour_osx.h
@@ -41,9 +41,9 @@ extern "C" {
 #endif
 
 HALIDE_FUNCTION_ATTRS
-int contour_par_for_sobel__5_s0_x_xo_ti_ti(void *__user_context, int32_t _sobel__5_s0_x_xo_ti_ti, uint8_t *_closure_arg__300);
+int contour_par_for_sobel__5_s0_x_xo_ti_ti(void *__user_context, int32_t _sobel__5_s0_x_xo_ti_ti, uint8_t *_closure_arg__325);
 HALIDE_FUNCTION_ATTRS
-int contour_par_for_countor_line_s0_x_xo_ti(void *__user_context, int32_t _countor_line_s0_x_xo_ti, uint8_t *_closure_arg__301);
+int contour_par_for_countor_line_s0_x_xo_ti(void *__user_context, int32_t _countor_line_s0_x_xo_ti, uint8_t *_closure_arg__326);
 HALIDE_FUNCTION_ATTRS
 int contour(struct halide_buffer_t *_src_buffer, int32_t _width, int32_t _height, uint8_t _threshold, uint8_t _size, struct halide_buffer_t *_contour_buffer);
 
diff --git a/include/libcontrast_linux.h b/include/libcontrast_linux.h
index 2f9365d..4a7e123 100644
--- a/include/libcontrast_linux.h
+++ b/include/libcontrast_linux.h
@@ -41,9 +41,9 @@ extern "C" {
 #endif
 
 HALIDE_FUNCTION_ATTRS
-int contrast_par_for_contrast_4_s0_ch__1(void *__user_context, int32_t _contrast__4_s0_ch, uint8_t *_closure_arg__151);
+int contrast_par_for_contrast_s0_ch__3(void *__user_context, int32_t _contrast_s0_ch, uint8_t *_closure_arg__151);
 HALIDE_FUNCTION_ATTRS
-int contrast(struct halide_buffer_t *_src_buffer, int32_t _width, int32_t _height, float _factor, struct halide_buffer_t *_contrast__4_buffer);
+int contrast(struct halide_buffer_t *_src_buffer, int32_t _width, int32_t _height, float _factor, struct halide_buffer_t *_contrast_buffer);
 
 HALIDE_FUNCTION_ATTRS
 int contrast_argv(void **args);
diff --git a/include/libcontrast_osx.h b/include/libcontrast_osx.h
index 437321d..4b8ce81 100644
--- a/include/libcontrast_osx.h
+++ b/include/libcontrast_osx.h
@@ -41,9 +41,9 @@ extern "C" {
 #endif
 
 HALIDE_FUNCTION_ATTRS
-int contrast_par_for_contrast__4_s0_ch(void *__user_context, int32_t _contrast__4_s0_ch, uint8_t *_closure_arg__150);
+int contrast_par_for_contrast_s0_ch__2(void *__user_context, int32_t _contrast_s0_ch, uint8_t *_closure_arg__150);
 HALIDE_FUNCTION_ATTRS
-int contrast(struct halide_buffer_t *_src_buffer, int32_t _width, int32_t _height, float _factor, struct halide_buffer_t *_contrast__4_buffer);
+int contrast(struct halide_buffer_t *_src_buffer, int32_t _width, int32_t _height, float _factor, struct halide_buffer_t *_contrast_buffer);
 
 HALIDE_FUNCTION_ATTRS
 int contrast_argv(void **args);
diff --git a/include/libdilation_linux.h b/include/libdilation_linux.h
index 6265e34..b063a97 100644
--- a/include/libdilation_linux.h
+++ b/include/libdilation_linux.h
@@ -41,13 +41,13 @@ extern "C" {
 #endif
 
 HALIDE_FUNCTION_ATTRS
-int dilation_par_for_in_29_s0_ch__1(void *__user_context, int32_t _in__29_s0_ch, uint8_t *_closure_arg__179);
+int dilation_par_for_in_29_s0_ch__5(void *__user_context, int32_t _in__29_s0_ch, uint8_t *_closure_arg__183);
 HALIDE_FUNCTION_ATTRS
-int dilation_par_for_dilation_3_s0_ch_par_for_dilation_3_s0_x_xo_ti_ti__1(void *__user_context, int32_t _dilation__3_s0_x_xo_ti_ti, uint8_t *_closure_arg__181);
+int dilation_par_for_dilation_s0_ch_par_for_dilation_s0_x_xo_ti_ti__1(void *__user_context, int32_t _dilation_s0_x_xo_ti_ti, uint8_t *_closure_arg__185);
 HALIDE_FUNCTION_ATTRS
-int dilation_par_for_dilation_3_s0_ch__1(void *__user_context, int32_t _dilation__3_s0_ch, uint8_t *_closure_arg__180);
+int dilation_par_for_dilation_s0_ch__1(void *__user_context, int32_t _dilation_s0_ch, uint8_t *_closure_arg__184);
 HALIDE_FUNCTION_ATTRS
-int dilation(struct halide_buffer_t *_src_buffer, int32_t _width, int32_t _height, int32_t _size, struct halide_buffer_t *_dilation__3_buffer);
+int dilation(struct halide_buffer_t *_src_buffer, int32_t _width, int32_t _height, int32_t _size, struct halide_buffer_t *_dilation_buffer);
 
 HALIDE_FUNCTION_ATTRS
 int dilation_argv(void **args);
diff --git a/include/libdilation_osx.h b/include/libdilation_osx.h
index fca595e..1bba526 100644
--- a/include/libdilation_osx.h
+++ b/include/libdilation_osx.h
@@ -41,13 +41,13 @@ extern "C" {
 #endif
 
 HALIDE_FUNCTION_ATTRS
-int dilation_par_for_in__29_s0_ch(void *__user_context, int32_t _in__29_s0_ch, uint8_t *_closure_arg__176);
+int dilation_par_for_in_29_s0_ch__4(void *__user_context, int32_t _in__29_s0_ch, uint8_t *_closure_arg__180);
 HALIDE_FUNCTION_ATTRS
-int dilation_par_for_dilation__3_s0_ch_par_for_dilation__3_s0_x_xo_ti_ti(void *__user_context, int32_t _dilation__3_s0_x_xo_ti_ti, uint8_t *_closure_arg__178);
+int dilation_par_for_dilation_s0_ch_par_for_dilation_s0_x_xo_ti_ti(void *__user_context, int32_t _dilation_s0_x_xo_ti_ti, uint8_t *_closure_arg__182);
 HALIDE_FUNCTION_ATTRS
-int dilation_par_for_dilation__3_s0_ch(void *__user_context, int32_t _dilation__3_s0_ch, uint8_t *_closure_arg__177);
+int dilation_par_for_dilation_s0_ch(void *__user_context, int32_t _dilation_s0_ch, uint8_t *_closure_arg__181);
 HALIDE_FUNCTION_ATTRS
-int dilation(struct halide_buffer_t *_src_buffer, int32_t _width, int32_t _height, int32_t _size, struct halide_buffer_t *_dilation__3_buffer);
+int dilation(struct halide_buffer_t *_src_buffer, int32_t _width, int32_t _height, int32_t _size, struct halide_buffer_t *_dilation_buffer);
 
 HALIDE_FUNCTION_ATTRS
 int dilation_argv(void **args);
diff --git a/include/libedgedetect_linux.h b/include/libedgedetect_linux.h
index 66332c3..6aab7e5 100644
--- a/include/libedgedetect_linux.h
+++ b/include/libedgedetect_linux.h
@@ -41,11 +41,11 @@ extern "C" {
 #endif
 
 HALIDE_FUNCTION_ATTRS
-int edgedetect_par_for_edgedetect_s0_ch_par_for_edgedetect_s0_x_xo_ti_ti__2(void *__user_context, int32_t _edgedetect_s0_x_xo_ti_ti, uint8_t *_closure_arg__242);
+int edgedetect_par_for_edgedetect_s0_ch_par_for_edgedetect_s0_x_xo_ti_ti__2(void *__user_context, int32_t _edgedetect_s0_x_xo_ti_ti, uint8_t *_closure_arg__246);
 HALIDE_FUNCTION_ATTRS
-int edgedetect_par_for_edgedetect_s0_ch_par_for_edgedetect_s0_x_xo_ti_ti__3(void *__user_context, int32_t _edgedetect_s0_x_xo_ti_ti, uint8_t *_closure_arg__243);
+int edgedetect_par_for_edgedetect_s0_ch_par_for_edgedetect_s0_x_xo_ti_ti__3(void *__user_context, int32_t _edgedetect_s0_x_xo_ti_ti, uint8_t *_closure_arg__247);
 HALIDE_FUNCTION_ATTRS
-int edgedetect_par_for_edgedetect_s0_ch__1(void *__user_context, int32_t _edgedetect_s0_ch, uint8_t *_closure_arg__241);
+int edgedetect_par_for_edgedetect_s0_ch__4(void *__user_context, int32_t _edgedetect_s0_ch, uint8_t *_closure_arg__245);
 HALIDE_FUNCTION_ATTRS
 int edgedetect(struct halide_buffer_t *_src_buffer, int32_t _width, int32_t _height, struct halide_buffer_t *_edgedetect_buffer);
 
diff --git a/include/libedgedetect_osx.h b/include/libedgedetect_osx.h
index f1bc670..1027e19 100644
--- a/include/libedgedetect_osx.h
+++ b/include/libedgedetect_osx.h
@@ -41,11 +41,11 @@ extern "C" {
 #endif
 
 HALIDE_FUNCTION_ATTRS
-int edgedetect_par_for_edgedetect_s0_ch_par_for_edgedetect_s0_x_xo_ti_ti(void *__user_context, int32_t _edgedetect_s0_x_xo_ti_ti, uint8_t *_closure_arg__239);
+int edgedetect_par_for_edgedetect_s0_ch_par_for_edgedetect_s0_x_xo_ti_ti(void *__user_context, int32_t _edgedetect_s0_x_xo_ti_ti, uint8_t *_closure_arg__243);
 HALIDE_FUNCTION_ATTRS
-int edgedetect_par_for_edgedetect_s0_ch_par_for_edgedetect_s0_x_xo_ti_ti__1(void *__user_context, int32_t _edgedetect_s0_x_xo_ti_ti, uint8_t *_closure_arg__240);
+int edgedetect_par_for_edgedetect_s0_ch_par_for_edgedetect_s0_x_xo_ti_ti__1(void *__user_context, int32_t _edgedetect_s0_x_xo_ti_ti, uint8_t *_closure_arg__244);
 HALIDE_FUNCTION_ATTRS
-int edgedetect_par_for_edgedetect_s0_ch(void *__user_context, int32_t _edgedetect_s0_ch, uint8_t *_closure_arg__238);
+int edgedetect_par_for_edgedetect_s0_ch__3(void *__user_context, int32_t _edgedetect_s0_ch, uint8_t *_closure_arg__242);
 HALIDE_FUNCTION_ATTRS
 int edgedetect(struct halide_buffer_t *_src_buffer, int32_t _width, int32_t _height, struct halide_buffer_t *_edgedetect_buffer);
 
diff --git a/include/libemboss_linux.h b/include/libemboss_linux.h
index ed9f7aa..50119ee 100644
--- a/include/libemboss_linux.h
+++ b/include/libemboss_linux.h
@@ -41,13 +41,13 @@ extern "C" {
 #endif
 
 HALIDE_FUNCTION_ATTRS
-int emboss_par_for_in_33_s0_ch__1(void *__user_context, int32_t _in__33_s0_ch, uint8_t *_closure_arg__216);
+int emboss_par_for_in_33_s0_ch__1(void *__user_context, int32_t _in__33_s0_ch, uint8_t *_closure_arg__220);
 HALIDE_FUNCTION_ATTRS
-int emboss_par_for_emboss_s0_ch_par_for_emboss_s0_x_xo_ti_ti__2(void *__user_context, int32_t _emboss_s0_x_xo_ti_ti, uint8_t *_closure_arg__218);
+int emboss_par_for_emboss_s0_ch_par_for_emboss_s0_x_xo_ti_ti__4(void *__user_context, int32_t _emboss_s0_x_xo_ti_ti, uint8_t *_closure_arg__222);
 HALIDE_FUNCTION_ATTRS
-int emboss_par_for_emboss_s0_ch_par_for_emboss_s0_x_xo_ti_ti__3(void *__user_context, int32_t _emboss_s0_x_xo_ti_ti, uint8_t *_closure_arg__219);
+int emboss_par_for_emboss_s0_ch_par_for_emboss_s0_x_xo_ti_ti__5(void *__user_context, int32_t _emboss_s0_x_xo_ti_ti, uint8_t *_closure_arg__223);
 HALIDE_FUNCTION_ATTRS
-int emboss_par_for_emboss_s0_ch__1(void *__user_context, int32_t _emboss_s0_ch, uint8_t *_closure_arg__217);
+int emboss_par_for_emboss_s0_ch__1(void *__user_context, int32_t _emboss_s0_ch, uint8_t *_closure_arg__221);
 HALIDE_FUNCTION_ATTRS
 int emboss(struct halide_buffer_t *_src_buffer, int32_t _width, int32_t _height, struct halide_buffer_t *_emboss_buffer);
 
diff --git a/include/libemboss_osx.h b/include/libemboss_osx.h
index 954c332..bcfdc7e 100644
--- a/include/libemboss_osx.h
+++ b/include/libemboss_osx.h
@@ -41,13 +41,13 @@ extern "C" {
 #endif
 
 HALIDE_FUNCTION_ATTRS
-int emboss_par_for_in__33_s0_ch(void *__user_context, int32_t _in__33_s0_ch, uint8_t *_closure_arg__212);
+int emboss_par_for_in__33_s0_ch(void *__user_context, int32_t _in__33_s0_ch, uint8_t *_closure_arg__216);
 HALIDE_FUNCTION_ATTRS
-int emboss_par_for_emboss_s0_ch_par_for_emboss_s0_x_xo_ti_ti(void *__user_context, int32_t _emboss_s0_x_xo_ti_ti, uint8_t *_closure_arg__214);
+int emboss_par_for_emboss_s0_ch_par_for_emboss_s0_x_xo_ti_ti__2(void *__user_context, int32_t _emboss_s0_x_xo_ti_ti, uint8_t *_closure_arg__218);
 HALIDE_FUNCTION_ATTRS
-int emboss_par_for_emboss_s0_ch_par_for_emboss_s0_x_xo_ti_ti__1(void *__user_context, int32_t _emboss_s0_x_xo_ti_ti, uint8_t *_closure_arg__215);
+int emboss_par_for_emboss_s0_ch_par_for_emboss_s0_x_xo_ti_ti__3(void *__user_context, int32_t _emboss_s0_x_xo_ti_ti, uint8_t *_closure_arg__219);
 HALIDE_FUNCTION_ATTRS
-int emboss_par_for_emboss_s0_ch(void *__user_context, int32_t _emboss_s0_ch, uint8_t *_closure_arg__213);
+int emboss_par_for_emboss_s0_ch(void *__user_context, int32_t _emboss_s0_ch, uint8_t *_closure_arg__217);
 HALIDE_FUNCTION_ATTRS
 int emboss(struct halide_buffer_t *_src_buffer, int32_t _width, int32_t _height, struct halide_buffer_t *_emboss_buffer);
 
diff --git a/include/liberosion_linux.h b/include/liberosion_linux.h
index 7f389e9..4cb9d33 100644
--- a/include/liberosion_linux.h
+++ b/include/liberosion_linux.h
@@ -41,11 +41,11 @@ extern "C" {
 #endif
 
 HALIDE_FUNCTION_ATTRS
-int erosion_par_for_in_28_s0_ch__1(void *__user_context, int32_t _in__28_s0_ch, uint8_t *_closure_arg__173);
+int erosion_par_for_in_28_s0_ch__1(void *__user_context, int32_t _in__28_s0_ch, uint8_t *_closure_arg__177);
 HALIDE_FUNCTION_ATTRS
-int erosion_par_for_erosion_s0_ch_par_for_erosion_s0_x_xo_ti_ti__1(void *__user_context, int32_t _erosion_s0_x_xo_ti_ti, uint8_t *_closure_arg__175);
+int erosion_par_for_erosion_s0_ch_par_for_erosion_s0_x_xo_ti_ti__1(void *__user_context, int32_t _erosion_s0_x_xo_ti_ti, uint8_t *_closure_arg__179);
 HALIDE_FUNCTION_ATTRS
-int erosion_par_for_erosion_s0_ch__1(void *__user_context, int32_t _erosion_s0_ch, uint8_t *_closure_arg__174);
+int erosion_par_for_erosion_s0_ch__1(void *__user_context, int32_t _erosion_s0_ch, uint8_t *_closure_arg__178);
 HALIDE_FUNCTION_ATTRS
 int erosion(struct halide_buffer_t *_src_buffer, int32_t _width, int32_t _height, int32_t _size, struct halide_buffer_t *_erosion_buffer);
 
diff --git a/include/liberosion_osx.h b/include/liberosion_osx.h
index 560814f..dbfe647 100644
--- a/include/liberosion_osx.h
+++ b/include/liberosion_osx.h
@@ -41,11 +41,11 @@ extern "C" {
 #endif
 
 HALIDE_FUNCTION_ATTRS
-int erosion_par_for_in__28_s0_ch(void *__user_context, int32_t _in__28_s0_ch, uint8_t *_closure_arg__170);
+int erosion_par_for_in__28_s0_ch(void *__user_context, int32_t _in__28_s0_ch, uint8_t *_closure_arg__174);
 HALIDE_FUNCTION_ATTRS
-int erosion_par_for_erosion_s0_ch_par_for_erosion_s0_x_xo_ti_ti(void *__user_context, int32_t _erosion_s0_x_xo_ti_ti, uint8_t *_closure_arg__172);
+int erosion_par_for_erosion_s0_ch_par_for_erosion_s0_x_xo_ti_ti(void *__user_context, int32_t _erosion_s0_x_xo_ti_ti, uint8_t *_closure_arg__176);
 HALIDE_FUNCTION_ATTRS
-int erosion_par_for_erosion_s0_ch(void *__user_context, int32_t _erosion_s0_ch, uint8_t *_closure_arg__171);
+int erosion_par_for_erosion_s0_ch(void *__user_context, int32_t _erosion_s0_ch, uint8_t *_closure_arg__175);
 HALIDE_FUNCTION_ATTRS
 int erosion(struct halide_buffer_t *_src_buffer, int32_t _width, int32_t _height, int32_t _size, struct halide_buffer_t *_erosion_buffer);
 
diff --git a/include/libgaussianblur_linux.h b/include/libgaussianblur_linux.h
index 65b97c8..9176a25 100644
--- a/include/libgaussianblur_linux.h
+++ b/include/libgaussianblur_linux.h
@@ -41,9 +41,9 @@ extern "C" {
 #endif
 
 HALIDE_FUNCTION_ATTRS
-int gaussianblur_par_for_gaussianblur_s0_ch_par_for_gaussianblur_s0_x_xo_ti_ti__1(void *__user_context, int32_t _gaussianblur_s0_x_xo_ti_ti, uint8_t *_closure_arg__161);
+int gaussianblur_par_for_gaussianblur_s0_ch_par_for_gaussianblur_s0_x_xo_ti_ti__1(void *__user_context, int32_t _gaussianblur_s0_x_xo_ti_ti, uint8_t *_closure_arg__165);
 HALIDE_FUNCTION_ATTRS
-int gaussianblur_par_for_gaussianblur_s0_ch__65(void *__user_context, int32_t _gaussianblur_s0_ch, uint8_t *_closure_arg__160);
+int gaussianblur_par_for_gaussianblur_s0_ch__6(void *__user_context, int32_t _gaussianblur_s0_ch, uint8_t *_closure_arg__164);
 HALIDE_FUNCTION_ATTRS
 int gaussianblur(struct halide_buffer_t *_src_buffer, int32_t _width, int32_t _height, float _sigma, struct halide_buffer_t *_gaussianblur_buffer);
 
diff --git a/include/libgaussianblur_osx.h b/include/libgaussianblur_osx.h
index 4188a8a..6891022 100644
--- a/include/libgaussianblur_osx.h
+++ b/include/libgaussianblur_osx.h
@@ -41,9 +41,9 @@ extern "C" {
 #endif
 
 HALIDE_FUNCTION_ATTRS
-int gaussianblur_par_for_gaussianblur_s0_ch_par_for_gaussianblur_s0_x_xo_ti_ti(void *__user_context, int32_t _gaussianblur_s0_x_xo_ti_ti, uint8_t *_closure_arg__159);
+int gaussianblur_par_for_gaussianblur_s0_ch_par_for_gaussianblur_s0_x_xo_ti_ti(void *__user_context, int32_t _gaussianblur_s0_x_xo_ti_ti, uint8_t *_closure_arg__163);
 HALIDE_FUNCTION_ATTRS
-int gaussianblur_par_for_gaussianblur_s0_ch__64(void *__user_context, int32_t _gaussianblur_s0_ch, uint8_t *_closure_arg__158);
+int gaussianblur_par_for_gaussianblur_s0_ch__5(void *__user_context, int32_t _gaussianblur_s0_ch, uint8_t *_closure_arg__162);
 HALIDE_FUNCTION_ATTRS
 int gaussianblur(struct halide_buffer_t *_src_buffer, int32_t _width, int32_t _height, float _sigma, struct halide_buffer_t *_gaussianblur_buffer);
 
diff --git a/include/libgradient_linux.h b/include/libgradient_linux.h
index 870d620..ca23538 100644
--- a/include/libgradient_linux.h
+++ b/include/libgradient_linux.h
@@ -41,11 +41,11 @@ extern "C" {
 #endif
 
 HALIDE_FUNCTION_ATTRS
-int gradient_par_for_gradient_s0_ch_par_for_gradient_s0_x_xo_ti_ti__5(void *__user_context, int32_t _gradient_s0_x_xo_ti_ti, uint8_t *_closure_arg__236);
+int gradient_par_for_gradient_s0_ch_par_for_gradient_s0_x_xo_ti_ti__2(void *__user_context, int32_t _gradient_s0_x_xo_ti_ti, uint8_t *_closure_arg__240);
 HALIDE_FUNCTION_ATTRS
-int gradient_par_for_gradient_s0_ch_par_for_gradient_s0_x_xo_ti_ti__6(void *__user_context, int32_t _gradient_s0_x_xo_ti_ti, uint8_t *_closure_arg__237);
+int gradient_par_for_gradient_s0_ch_par_for_gradient_s0_x_xo_ti_ti__3(void *__user_context, int32_t _gradient_s0_x_xo_ti_ti, uint8_t *_closure_arg__241);
 HALIDE_FUNCTION_ATTRS
-int gradient_par_for_gradient_s0_ch__1(void *__user_context, int32_t _gradient_s0_ch, uint8_t *_closure_arg__235);
+int gradient_par_for_gradient_s0_ch__2(void *__user_context, int32_t _gradient_s0_ch, uint8_t *_closure_arg__239);
 HALIDE_FUNCTION_ATTRS
 int gradient(struct halide_buffer_t *_src_buffer, int32_t _width, int32_t _height, struct halide_buffer_t *_gradient_buffer);
 
diff --git a/include/libgradient_osx.h b/include/libgradient_osx.h
index 3ecfb72..6fd315a 100644
--- a/include/libgradient_osx.h
+++ b/include/libgradient_osx.h
@@ -41,11 +41,11 @@ extern "C" {
 #endif
 
 HALIDE_FUNCTION_ATTRS
-int gradient_par_for_gradient_s0_ch_par_for_gradient_s0_x_xo_ti_ti__3(void *__user_context, int32_t _gradient_s0_x_xo_ti_ti, uint8_t *_closure_arg__233);
+int gradient_par_for_gradient_s0_ch_par_for_gradient_s0_x_xo_ti_ti(void *__user_context, int32_t _gradient_s0_x_xo_ti_ti, uint8_t *_closure_arg__237);
 HALIDE_FUNCTION_ATTRS
-int gradient_par_for_gradient_s0_ch_par_for_gradient_s0_x_xo_ti_ti__4(void *__user_context, int32_t _gradient_s0_x_xo_ti_ti, uint8_t *_closure_arg__234);
+int gradient_par_for_gradient_s0_ch_par_for_gradient_s0_x_xo_ti_ti__1(void *__user_context, int32_t _gradient_s0_x_xo_ti_ti, uint8_t *_closure_arg__238);
 HALIDE_FUNCTION_ATTRS
-int gradient_par_for_gradient_s0_ch(void *__user_context, int32_t _gradient_s0_ch, uint8_t *_closure_arg__232);
+int gradient_par_for_gradient_s0_ch__1(void *__user_context, int32_t _gradient_s0_ch, uint8_t *_closure_arg__236);
 HALIDE_FUNCTION_ATTRS
 int gradient(struct halide_buffer_t *_src_buffer, int32_t _width, int32_t _height, struct halide_buffer_t *_gradient_buffer);
 
diff --git a/include/libhighpass_linux.h b/include/libhighpass_linux.h
index d50485a..8b508ce 100644
--- a/include/libhighpass_linux.h
+++ b/include/libhighpass_linux.h
@@ -41,11 +41,11 @@ extern "C" {
 #endif
 
 HALIDE_FUNCTION_ATTRS
-int highpass_par_for_highpass_s0_ch_par_for_highpass_s0_x_xo_ti_ti__2(void *__user_context, int32_t _highpass_s0_x_xo_ti_ti, uint8_t *_closure_arg__230);
+int highpass_par_for_highpass_s0_ch_par_for_highpass_s0_x_xo_ti_ti__2(void *__user_context, int32_t _highpass_s0_x_xo_ti_ti, uint8_t *_closure_arg__234);
 HALIDE_FUNCTION_ATTRS
-int highpass_par_for_highpass_s0_ch_par_for_highpass_s0_x_xo_ti_ti__3(void *__user_context, int32_t _highpass_s0_x_xo_ti_ti, uint8_t *_closure_arg__231);
+int highpass_par_for_highpass_s0_ch_par_for_highpass_s0_x_xo_ti_ti__3(void *__user_context, int32_t _highpass_s0_x_xo_ti_ti, uint8_t *_closure_arg__235);
 HALIDE_FUNCTION_ATTRS
-int highpass_par_for_highpass_s0_ch__1(void *__user_context, int32_t _highpass_s0_ch, uint8_t *_closure_arg__229);
+int highpass_par_for_highpass_s0_ch__1(void *__user_context, int32_t _highpass_s0_ch, uint8_t *_closure_arg__233);
 HALIDE_FUNCTION_ATTRS
 int highpass(struct halide_buffer_t *_src_buffer, int32_t _width, int32_t _height, struct halide_buffer_t *_highpass_buffer);
 
diff --git a/include/libhighpass_osx.h b/include/libhighpass_osx.h
index dd35e58..f7edf2c 100644
--- a/include/libhighpass_osx.h
+++ b/include/libhighpass_osx.h
@@ -41,11 +41,11 @@ extern "C" {
 #endif
 
 HALIDE_FUNCTION_ATTRS
-int highpass_par_for_highpass_s0_ch_par_for_highpass_s0_x_xo_ti_ti(void *__user_context, int32_t _highpass_s0_x_xo_ti_ti, uint8_t *_closure_arg__227);
+int highpass_par_for_highpass_s0_ch_par_for_highpass_s0_x_xo_ti_ti(void *__user_context, int32_t _highpass_s0_x_xo_ti_ti, uint8_t *_closure_arg__231);
 HALIDE_FUNCTION_ATTRS
-int highpass_par_for_highpass_s0_ch_par_for_highpass_s0_x_xo_ti_ti__1(void *__user_context, int32_t _highpass_s0_x_xo_ti_ti, uint8_t *_closure_arg__228);
+int highpass_par_for_highpass_s0_ch_par_for_highpass_s0_x_xo_ti_ti__1(void *__user_context, int32_t _highpass_s0_x_xo_ti_ti, uint8_t *_closure_arg__232);
 HALIDE_FUNCTION_ATTRS
-int highpass_par_for_highpass_s0_ch(void *__user_context, int32_t _highpass_s0_ch, uint8_t *_closure_arg__226);
+int highpass_par_for_highpass_s0_ch(void *__user_context, int32_t _highpass_s0_ch, uint8_t *_closure_arg__230);
 HALIDE_FUNCTION_ATTRS
 int highpass(struct halide_buffer_t *_src_buffer, int32_t _width, int32_t _height, struct halide_buffer_t *_highpass_buffer);
 
diff --git a/include/libinvert_linux.h b/include/libinvert_linux.h
index d568500..8ce6d1a 100644
--- a/include/libinvert_linux.h
+++ b/include/libinvert_linux.h
@@ -43,11 +43,11 @@ extern "C" {
 HALIDE_FUNCTION_ATTRS
 int invert_par_for_in_21_s0_ch__1(void *__user_context, int32_t _in__21_s0_ch, uint8_t *_closure_arg__132);
 HALIDE_FUNCTION_ATTRS
-int invert_par_for_invert_s0_ch_par_for_invert_s0_x_xo_ti_ti__6(void *__user_context, int32_t _invert_s0_x_xo_ti_ti, uint8_t *_closure_arg__134);
+int invert_par_for_invert_s0_ch_par_for_invert_s0_x_xo_ti_ti__2(void *__user_context, int32_t _invert_s0_x_xo_ti_ti, uint8_t *_closure_arg__134);
 HALIDE_FUNCTION_ATTRS
-int invert_par_for_invert_s0_ch_par_for_invert_s0_x_xo_ti_ti__7(void *__user_context, int32_t _invert_s0_x_xo_ti_ti, uint8_t *_closure_arg__135);
+int invert_par_for_invert_s0_ch_par_for_invert_s0_x_xo_ti_ti__3(void *__user_context, int32_t _invert_s0_x_xo_ti_ti, uint8_t *_closure_arg__135);
 HALIDE_FUNCTION_ATTRS
-int invert_par_for_invert_s0_ch__5(void *__user_context, int32_t _invert_s0_ch, uint8_t *_closure_arg__133);
+int invert_par_for_invert_s0_ch__1(void *__user_context, int32_t _invert_s0_ch, uint8_t *_closure_arg__133);
 HALIDE_FUNCTION_ATTRS
 int invert(struct halide_buffer_t *_src_buffer, int32_t _width, int32_t _height, struct halide_buffer_t *_invert_buffer);
 
diff --git a/include/libinvert_osx.h b/include/libinvert_osx.h
index 6e66f87..a7624a7 100644
--- a/include/libinvert_osx.h
+++ b/include/libinvert_osx.h
@@ -43,9 +43,9 @@ extern "C" {
 HALIDE_FUNCTION_ATTRS
 int invert_par_for_in__21_s0_ch(void *__user_context, int32_t _in__21_s0_ch, uint8_t *_closure_arg__128);
 HALIDE_FUNCTION_ATTRS
-int invert_par_for_invert_s0_ch_par_for_invert_s0_x_xo_ti_ti__4(void *__user_context, int32_t _invert_s0_x_xo_ti_ti, uint8_t *_closure_arg__130);
+int invert_par_for_invert_s0_ch_par_for_invert_s0_x_xo_ti_ti(void *__user_context, int32_t _invert_s0_x_xo_ti_ti, uint8_t *_closure_arg__130);
 HALIDE_FUNCTION_ATTRS
-int invert_par_for_invert_s0_ch_par_for_invert_s0_x_xo_ti_ti__5(void *__user_context, int32_t _invert_s0_x_xo_ti_ti, uint8_t *_closure_arg__131);
+int invert_par_for_invert_s0_ch_par_for_invert_s0_x_xo_ti_ti__1(void *__user_context, int32_t _invert_s0_x_xo_ti_ti, uint8_t *_closure_arg__131);
 HALIDE_FUNCTION_ATTRS
 int invert_par_for_invert_s0_ch(void *__user_context, int32_t _invert_s0_ch, uint8_t *_closure_arg__129);
 HALIDE_FUNCTION_ATTRS
diff --git a/include/liblaplacian_linux.h b/include/liblaplacian_linux.h
index 9901356..1179cd0 100644
--- a/include/liblaplacian_linux.h
+++ b/include/liblaplacian_linux.h
@@ -41,11 +41,11 @@ extern "C" {
 #endif
 
 HALIDE_FUNCTION_ATTRS
-int laplacian_par_for_laplacian_s0_ch_par_for_laplacian_s0_x_xo_ti_ti__2(void *__user_context, int32_t _laplacian_s0_x_xo_ti_ti, uint8_t *_closure_arg__224);
+int laplacian_par_for_laplacian_s0_ch_par_for_laplacian_s0_x_xo_ti_ti__2(void *__user_context, int32_t _laplacian_s0_x_xo_ti_ti, uint8_t *_closure_arg__228);
 HALIDE_FUNCTION_ATTRS
-int laplacian_par_for_laplacian_s0_ch_par_for_laplacian_s0_x_xo_ti_ti__3(void *__user_context, int32_t _laplacian_s0_x_xo_ti_ti, uint8_t *_closure_arg__225);
+int laplacian_par_for_laplacian_s0_ch_par_for_laplacian_s0_x_xo_ti_ti__3(void *__user_context, int32_t _laplacian_s0_x_xo_ti_ti, uint8_t *_closure_arg__229);
 HALIDE_FUNCTION_ATTRS
-int laplacian_par_for_laplacian_s0_ch__1(void *__user_context, int32_t _laplacian_s0_ch, uint8_t *_closure_arg__223);
+int laplacian_par_for_laplacian_s0_ch__1(void *__user_context, int32_t _laplacian_s0_ch, uint8_t *_closure_arg__227);
 HALIDE_FUNCTION_ATTRS
 int laplacian(struct halide_buffer_t *_src_buffer, int32_t _width, int32_t _height, struct halide_buffer_t *_laplacian_buffer);
 
diff --git a/include/liblaplacian_osx.h b/include/liblaplacian_osx.h
index fac314c..478cb38 100644
--- a/include/liblaplacian_osx.h
+++ b/include/liblaplacian_osx.h
@@ -41,11 +41,11 @@ extern "C" {
 #endif
 
 HALIDE_FUNCTION_ATTRS
-int laplacian_par_for_laplacian_s0_ch_par_for_laplacian_s0_x_xo_ti_ti(void *__user_context, int32_t _laplacian_s0_x_xo_ti_ti, uint8_t *_closure_arg__221);
+int laplacian_par_for_laplacian_s0_ch_par_for_laplacian_s0_x_xo_ti_ti(void *__user_context, int32_t _laplacian_s0_x_xo_ti_ti, uint8_t *_closure_arg__225);
 HALIDE_FUNCTION_ATTRS
-int laplacian_par_for_laplacian_s0_ch_par_for_laplacian_s0_x_xo_ti_ti__1(void *__user_context, int32_t _laplacian_s0_x_xo_ti_ti, uint8_t *_closure_arg__222);
+int laplacian_par_for_laplacian_s0_ch_par_for_laplacian_s0_x_xo_ti_ti__1(void *__user_context, int32_t _laplacian_s0_x_xo_ti_ti, uint8_t *_closure_arg__226);
 HALIDE_FUNCTION_ATTRS
-int laplacian_par_for_laplacian_s0_ch(void *__user_context, int32_t _laplacian_s0_ch, uint8_t *_closure_arg__220);
+int laplacian_par_for_laplacian_s0_ch(void *__user_context, int32_t _laplacian_s0_ch, uint8_t *_closure_arg__224);
 HALIDE_FUNCTION_ATTRS
 int laplacian(struct halide_buffer_t *_src_buffer, int32_t _width, int32_t _height, struct halide_buffer_t *_laplacian_buffer);
 
diff --git a/include/libmatch_template_ncc_linux.h b/include/libmatch_template_ncc_linux.h
index b7b0268..3a960d9 100644
--- a/include/libmatch_template_ncc_linux.h
+++ b/include/libmatch_template_ncc_linux.h
@@ -41,9 +41,9 @@ extern "C" {
 #endif
 
 HALIDE_FUNCTION_ATTRS
-int match_template_ncc_par_for_match_template_ncc_s0_x_xo_ti__22(void *__user_context, int32_t _match_template_ncc_s0_x_xo_ti, uint8_t *_closure_arg__281);
+int match_template_ncc_par_for_match_template_ncc_3_s0_x_xo_ti__1(void *__user_context, int32_t _match_template_ncc__3_s0_x_xo_ti, uint8_t *_closure_arg__306);
 HALIDE_FUNCTION_ATTRS
-int match_template_ncc(struct halide_buffer_t *_src_buffer, int32_t _width, int32_t _height, struct halide_buffer_t *_tpl_buffer, int32_t _tpl_width, int32_t _tpl_height, struct halide_buffer_t *_match_template_ncc_buffer);
+int match_template_ncc(struct halide_buffer_t *_src_buffer, int32_t _width, int32_t _height, struct halide_buffer_t *_tpl_buffer, int32_t _tpl_width, int32_t _tpl_height, struct halide_buffer_t *_match_template_ncc__3_buffer);
 
 HALIDE_FUNCTION_ATTRS
 int match_template_ncc_argv(void **args);
diff --git a/include/libmatch_template_ncc_osx.h b/include/libmatch_template_ncc_osx.h
index 0c8694e..a1d2ac9 100644
--- a/include/libmatch_template_ncc_osx.h
+++ b/include/libmatch_template_ncc_osx.h
@@ -41,9 +41,9 @@ extern "C" {
 #endif
 
 HALIDE_FUNCTION_ATTRS
-int match_template_ncc_par_for_match_template_ncc_s0_x_xo_ti__21(void *__user_context, int32_t _match_template_ncc_s0_x_xo_ti, uint8_t *_closure_arg__280);
+int match_template_ncc_par_for_match_template_ncc__3_s0_x_xo_ti(void *__user_context, int32_t _match_template_ncc__3_s0_x_xo_ti, uint8_t *_closure_arg__305);
 HALIDE_FUNCTION_ATTRS
-int match_template_ncc(struct halide_buffer_t *_src_buffer, int32_t _width, int32_t _height, struct halide_buffer_t *_tpl_buffer, int32_t _tpl_width, int32_t _tpl_height, struct halide_buffer_t *_match_template_ncc_buffer);
+int match_template_ncc(struct halide_buffer_t *_src_buffer, int32_t _width, int32_t _height, struct halide_buffer_t *_tpl_buffer, int32_t _tpl_width, int32_t _tpl_height, struct halide_buffer_t *_match_template_ncc__3_buffer);
 
 HALIDE_FUNCTION_ATTRS
 int match_template_ncc_argv(void **args);
diff --git a/include/libmatch_template_sad_linux.h b/include/libmatch_template_sad_linux.h
index b681ba1..dc29c3a 100644
--- a/include/libmatch_template_sad_linux.h
+++ b/include/libmatch_template_sad_linux.h
@@ -41,7 +41,7 @@ extern "C" {
 #endif
 
 HALIDE_FUNCTION_ATTRS
-int match_template_sad_par_for_match_template_sad_s0_x_xo_ti__1(void *__user_context, int32_t _match_template_sad_s0_x_xo_ti, uint8_t *_closure_arg__277);
+int match_template_sad_par_for_match_template_sad_s0_x_xo_ti__1(void *__user_context, int32_t _match_template_sad_s0_x_xo_ti, uint8_t *_closure_arg__302);
 HALIDE_FUNCTION_ATTRS
 int match_template_sad(struct halide_buffer_t *_src_buffer, int32_t _width, int32_t _height, struct halide_buffer_t *_tpl_buffer, int32_t _tpl_width, int32_t _tpl_height, struct halide_buffer_t *_match_template_sad_buffer);
 
diff --git a/include/libmatch_template_sad_osx.h b/include/libmatch_template_sad_osx.h
index 880de65..a81c66d 100644
--- a/include/libmatch_template_sad_osx.h
+++ b/include/libmatch_template_sad_osx.h
@@ -41,7 +41,7 @@ extern "C" {
 #endif
 
 HALIDE_FUNCTION_ATTRS
-int match_template_sad_par_for_match_template_sad_s0_x_xo_ti(void *__user_context, int32_t _match_template_sad_s0_x_xo_ti, uint8_t *_closure_arg__276);
+int match_template_sad_par_for_match_template_sad_s0_x_xo_ti(void *__user_context, int32_t _match_template_sad_s0_x_xo_ti, uint8_t *_closure_arg__301);
 HALIDE_FUNCTION_ATTRS
 int match_template_sad(struct halide_buffer_t *_src_buffer, int32_t _width, int32_t _height, struct halide_buffer_t *_tpl_buffer, int32_t _tpl_width, int32_t _tpl_height, struct halide_buffer_t *_match_template_sad_buffer);
 
diff --git a/include/libmatch_template_ssd_linux.h b/include/libmatch_template_ssd_linux.h
index dc9b2db..d8c609d 100644
--- a/include/libmatch_template_ssd_linux.h
+++ b/include/libmatch_template_ssd_linux.h
@@ -41,9 +41,9 @@ extern "C" {
 #endif
 
 HALIDE_FUNCTION_ATTRS
-int match_template_ssd_par_for_match_template_ssd_4_s0_x_xo_ti__3(void *__user_context, int32_t _match_template_ssd__4_s0_x_xo_ti, uint8_t *_closure_arg__279);
+int match_template_ssd_par_for_match_template_ssd_s0_x_xo_ti__1(void *__user_context, int32_t _match_template_ssd_s0_x_xo_ti, uint8_t *_closure_arg__304);
 HALIDE_FUNCTION_ATTRS
-int match_template_ssd(struct halide_buffer_t *_src_buffer, int32_t _width, int32_t _height, struct halide_buffer_t *_tpl_buffer, int32_t _tpl_width, int32_t _tpl_height, struct halide_buffer_t *_match_template_ssd__4_buffer);
+int match_template_ssd(struct halide_buffer_t *_src_buffer, int32_t _width, int32_t _height, struct halide_buffer_t *_tpl_buffer, int32_t _tpl_width, int32_t _tpl_height, struct halide_buffer_t *_match_template_ssd_buffer);
 
 HALIDE_FUNCTION_ATTRS
 int match_template_ssd_argv(void **args);
diff --git a/include/libmatch_template_ssd_osx.h b/include/libmatch_template_ssd_osx.h
index b55a8bc..830f292 100644
--- a/include/libmatch_template_ssd_osx.h
+++ b/include/libmatch_template_ssd_osx.h
@@ -41,9 +41,9 @@ extern "C" {
 #endif
 
 HALIDE_FUNCTION_ATTRS
-int match_template_ssd_par_for_match_template_ssd_4_s0_x_xo_ti__2(void *__user_context, int32_t _match_template_ssd__4_s0_x_xo_ti, uint8_t *_closure_arg__278);
+int match_template_ssd_par_for_match_template_ssd_s0_x_xo_ti(void *__user_context, int32_t _match_template_ssd_s0_x_xo_ti, uint8_t *_closure_arg__303);
 HALIDE_FUNCTION_ATTRS
-int match_template_ssd(struct halide_buffer_t *_src_buffer, int32_t _width, int32_t _height, struct halide_buffer_t *_tpl_buffer, int32_t _tpl_width, int32_t _tpl_height, struct halide_buffer_t *_match_template_ssd__4_buffer);
+int match_template_ssd(struct halide_buffer_t *_src_buffer, int32_t _width, int32_t _height, struct halide_buffer_t *_tpl_buffer, int32_t _tpl_width, int32_t _tpl_height, struct halide_buffer_t *_match_template_ssd_buffer);
 
 HALIDE_FUNCTION_ATTRS
 int match_template_ssd_argv(void **args);
diff --git a/include/libmatch_template_zncc_linux.h b/include/libmatch_template_zncc_linux.h
index c547955..31b3a16 100644
--- a/include/libmatch_template_zncc_linux.h
+++ b/include/libmatch_template_zncc_linux.h
@@ -41,11 +41,11 @@ extern "C" {
 #endif
 
 HALIDE_FUNCTION_ATTRS
-int match_template_zncc_par_for_zncc_avg_s0_y_rebased__1(void *__user_context, int32_t _zncc_avg_s0_y_rebased, uint8_t *_closure_arg__285);
+int match_template_zncc_par_for_zncc_avg_s0_y_rebased__1(void *__user_context, int32_t _zncc_avg_s0_y_rebased, uint8_t *_closure_arg__310);
 HALIDE_FUNCTION_ATTRS
-int match_template_zncc_par_for_zncc_stddev_s0_y_rebased__2(void *__user_context, int32_t _zncc_stddev_s0_y_rebased, uint8_t *_closure_arg__286);
+int match_template_zncc_par_for_zncc_stddev_s0_y_rebased__3(void *__user_context, int32_t _zncc_stddev_s0_y_rebased, uint8_t *_closure_arg__311);
 HALIDE_FUNCTION_ATTRS
-int match_template_zncc_par_for_match_template_zncc_s0_x_xo_ti__1(void *__user_context, int32_t _match_template_zncc_s0_x_xo_ti, uint8_t *_closure_arg__287);
+int match_template_zncc_par_for_match_template_zncc_s0_x_xo_ti__1(void *__user_context, int32_t _match_template_zncc_s0_x_xo_ti, uint8_t *_closure_arg__312);
 HALIDE_FUNCTION_ATTRS
 int match_template_zncc(struct halide_buffer_t *_src_buffer, int32_t _width, int32_t _height, struct halide_buffer_t *_tpl_buffer, int32_t _tpl_width, int32_t _tpl_height, struct halide_buffer_t *_match_template_zncc_buffer);
 
diff --git a/include/libmatch_template_zncc_osx.h b/include/libmatch_template_zncc_osx.h
index c0ccd3d..87b1a6d 100644
--- a/include/libmatch_template_zncc_osx.h
+++ b/include/libmatch_template_zncc_osx.h
@@ -41,11 +41,11 @@ extern "C" {
 #endif
 
 HALIDE_FUNCTION_ATTRS
-int match_template_zncc_par_for_zncc_avg_s0_y_rebased(void *__user_context, int32_t _zncc_avg_s0_y_rebased, uint8_t *_closure_arg__282);
+int match_template_zncc_par_for_zncc_avg_s0_y_rebased(void *__user_context, int32_t _zncc_avg_s0_y_rebased, uint8_t *_closure_arg__307);
 HALIDE_FUNCTION_ATTRS
-int match_template_zncc_par_for_zncc_stddev_s0_y_rebased__1(void *__user_context, int32_t _zncc_stddev_s0_y_rebased, uint8_t *_closure_arg__283);
+int match_template_zncc_par_for_zncc_stddev_s0_y_rebased__2(void *__user_context, int32_t _zncc_stddev_s0_y_rebased, uint8_t *_closure_arg__308);
 HALIDE_FUNCTION_ATTRS
-int match_template_zncc_par_for_match_template_zncc_s0_x_xo_ti(void *__user_context, int32_t _match_template_zncc_s0_x_xo_ti, uint8_t *_closure_arg__284);
+int match_template_zncc_par_for_match_template_zncc_s0_x_xo_ti(void *__user_context, int32_t _match_template_zncc_s0_x_xo_ti, uint8_t *_closure_arg__309);
 HALIDE_FUNCTION_ATTRS
 int match_template_zncc(struct halide_buffer_t *_src_buffer, int32_t _width, int32_t _height, struct halide_buffer_t *_tpl_buffer, int32_t _tpl_width, int32_t _tpl_height, struct halide_buffer_t *_match_template_zncc_buffer);
 
diff --git a/include/libmorphology_close_linux.h b/include/libmorphology_close_linux.h
index 9c3dc63..b226573 100644
--- a/include/libmorphology_close_linux.h
+++ b/include/libmorphology_close_linux.h
@@ -41,17 +41,17 @@ extern "C" {
 #endif
 
 HALIDE_FUNCTION_ATTRS
-int morphology_close_par_for_in_31_s0_y_rebased__2(void *__user_context, int32_t _in__31_s0_y_rebased, uint8_t *_closure_arg__200);
+int morphology_close_par_for_in_31_s0_y_rebased__1(void *__user_context, int32_t _in__31_s0_y_rebased, uint8_t *_closure_arg__204);
 HALIDE_FUNCTION_ATTRS
-int morphology_close_par_for_dilate_tmp_4_s0_y_y__1(void *__user_context, int32_t _dilate_tmp__4_s0_y_y, uint8_t *_closure_arg__201);
+int morphology_close_par_for_dilate_tmp_1_s0_y_y__1(void *__user_context, int32_t _dilate_tmp__1_s0_y_y, uint8_t *_closure_arg__205);
 HALIDE_FUNCTION_ATTRS
-int morphology_close_par_for_erode_tmp_1_s0_y_y__1(void *__user_context, int32_t _erode_tmp__1_s0_y_y, uint8_t *_closure_arg__202);
+int morphology_close_par_for_erode_tmp_1_s0_y_y__1(void *__user_context, int32_t _erode_tmp__1_s0_y_y, uint8_t *_closure_arg__206);
 HALIDE_FUNCTION_ATTRS
-int morphology_close_par_for_morphology_close_s0_ch_par_for_morphology_close_s0_x_xo_ti_ti__2(void *__user_context, int32_t _morphology_close_s0_x_xo_ti_ti, uint8_t *_closure_arg__204);
+int morphology_close_par_for_morphology_close_s0_ch_par_for_morphology_close_s0_x_xo_ti_ti__2(void *__user_context, int32_t _morphology_close_s0_x_xo_ti_ti, uint8_t *_closure_arg__208);
 HALIDE_FUNCTION_ATTRS
-int morphology_close_par_for_morphology_close_s0_ch_par_for_morphology_close_s0_x_xo_ti_ti__3(void *__user_context, int32_t _morphology_close_s0_x_xo_ti_ti, uint8_t *_closure_arg__205);
+int morphology_close_par_for_morphology_close_s0_ch_par_for_morphology_close_s0_x_xo_ti_ti__3(void *__user_context, int32_t _morphology_close_s0_x_xo_ti_ti, uint8_t *_closure_arg__209);
 HALIDE_FUNCTION_ATTRS
-int morphology_close_par_for_morphology_close_s0_ch__5(void *__user_context, int32_t _morphology_close_s0_ch, uint8_t *_closure_arg__203);
+int morphology_close_par_for_morphology_close_s0_ch__1(void *__user_context, int32_t _morphology_close_s0_ch, uint8_t *_closure_arg__207);
 HALIDE_FUNCTION_ATTRS
 int morphology_close(struct halide_buffer_t *_src_buffer, int32_t _width, int32_t _height, int32_t _size, struct halide_buffer_t *_morphology_close_buffer);
 
diff --git a/include/libmorphology_close_osx.h b/include/libmorphology_close_osx.h
index 78245d6..9a92ebf 100644
--- a/include/libmorphology_close_osx.h
+++ b/include/libmorphology_close_osx.h
@@ -41,17 +41,17 @@ extern "C" {
 #endif
 
 HALIDE_FUNCTION_ATTRS
-int morphology_close_par_for_in_31_s0_y_rebased__1(void *__user_context, int32_t _in__31_s0_y_rebased, uint8_t *_closure_arg__194);
+int morphology_close_par_for_in__31_s0_y_rebased(void *__user_context, int32_t _in__31_s0_y_rebased, uint8_t *_closure_arg__198);
 HALIDE_FUNCTION_ATTRS
-int morphology_close_par_for_dilate_tmp__4_s0_y_y(void *__user_context, int32_t _dilate_tmp__4_s0_y_y, uint8_t *_closure_arg__195);
+int morphology_close_par_for_dilate_tmp__1_s0_y_y(void *__user_context, int32_t _dilate_tmp__1_s0_y_y, uint8_t *_closure_arg__199);
 HALIDE_FUNCTION_ATTRS
-int morphology_close_par_for_erode_tmp__1_s0_y_y(void *__user_context, int32_t _erode_tmp__1_s0_y_y, uint8_t *_closure_arg__196);
+int morphology_close_par_for_erode_tmp__1_s0_y_y(void *__user_context, int32_t _erode_tmp__1_s0_y_y, uint8_t *_closure_arg__200);
 HALIDE_FUNCTION_ATTRS
-int morphology_close_par_for_morphology_close_s0_ch_par_for_morphology_close_s0_x_xo_ti_ti(void *__user_context, int32_t _morphology_close_s0_x_xo_ti_ti, uint8_t *_closure_arg__198);
+int morphology_close_par_for_morphology_close_s0_ch_par_for_morphology_close_s0_x_xo_ti_ti(void *__user_context, int32_t _morphology_close_s0_x_xo_ti_ti, uint8_t *_closure_arg__202);
 HALIDE_FUNCTION_ATTRS
-int morphology_close_par_for_morphology_close_s0_ch_par_for_morphology_close_s0_x_xo_ti_ti__1(void *__user_context, int32_t _morphology_close_s0_x_xo_ti_ti, uint8_t *_closure_arg__199);
+int morphology_close_par_for_morphology_close_s0_ch_par_for_morphology_close_s0_x_xo_ti_ti__1(void *__user_context, int32_t _morphology_close_s0_x_xo_ti_ti, uint8_t *_closure_arg__203);
 HALIDE_FUNCTION_ATTRS
-int morphology_close_par_for_morphology_close_s0_ch__4(void *__user_context, int32_t _morphology_close_s0_ch, uint8_t *_closure_arg__197);
+int morphology_close_par_for_morphology_close_s0_ch(void *__user_context, int32_t _morphology_close_s0_ch, uint8_t *_closure_arg__201);
 HALIDE_FUNCTION_ATTRS
 int morphology_close(struct halide_buffer_t *_src_buffer, int32_t _width, int32_t _height, int32_t _size, struct halide_buffer_t *_morphology_close_buffer);
 
diff --git a/include/libmorphology_gradient_linux.h b/include/libmorphology_gradient_linux.h
index d6dd6e4..37bcb0c 100644
--- a/include/libmorphology_gradient_linux.h
+++ b/include/libmorphology_gradient_linux.h
@@ -41,13 +41,13 @@ extern "C" {
 #endif
 
 HALIDE_FUNCTION_ATTRS
-int morphology_gradient_par_for_in_32_s0_y_rebased__1(void *__user_context, int32_t _in__32_s0_y_rebased, uint8_t *_closure_arg__209);
+int morphology_gradient_par_for_in_32_s0_y_rebased__1(void *__user_context, int32_t _in__32_s0_y_rebased, uint8_t *_closure_arg__213);
 HALIDE_FUNCTION_ATTRS
-int morphology_gradient_par_for_morphology_gradient_s0_ch_par_for_morphology_gradient_s0_x_xo_ti_ti__1(void *__user_context, int32_t _morphology_gradient_s0_x_xo_ti_ti, uint8_t *_closure_arg__211);
+int morphology_gradient_par_for_morphology_gradient_4_s0_ch_par_for_morphology_gradient_4_s0_x_xo_ti_ti__1(void *__user_context, int32_t _morphology_gradient__4_s0_x_xo_ti_ti, uint8_t *_closure_arg__215);
 HALIDE_FUNCTION_ATTRS
-int morphology_gradient_par_for_morphology_gradient_s0_ch__1(void *__user_context, int32_t _morphology_gradient_s0_ch, uint8_t *_closure_arg__210);
+int morphology_gradient_par_for_morphology_gradient_4_s0_ch__1(void *__user_context, int32_t _morphology_gradient__4_s0_ch, uint8_t *_closure_arg__214);
 HALIDE_FUNCTION_ATTRS
-int morphology_gradient(struct halide_buffer_t *_src_buffer, int32_t _width, int32_t _height, int32_t _size, struct halide_buffer_t *_morphology_gradient_buffer);
+int morphology_gradient(struct halide_buffer_t *_src_buffer, int32_t _width, int32_t _height, int32_t _size, struct halide_buffer_t *_morphology_gradient__4_buffer);
 
 HALIDE_FUNCTION_ATTRS
 int morphology_gradient_argv(void **args);
diff --git a/include/libmorphology_gradient_osx.h b/include/libmorphology_gradient_osx.h
index c3c6e81..57ffc18 100644
--- a/include/libmorphology_gradient_osx.h
+++ b/include/libmorphology_gradient_osx.h
@@ -41,13 +41,13 @@ extern "C" {
 #endif
 
 HALIDE_FUNCTION_ATTRS
-int morphology_gradient_par_for_in__32_s0_y_rebased(void *__user_context, int32_t _in__32_s0_y_rebased, uint8_t *_closure_arg__206);
+int morphology_gradient_par_for_in__32_s0_y_rebased(void *__user_context, int32_t _in__32_s0_y_rebased, uint8_t *_closure_arg__210);
 HALIDE_FUNCTION_ATTRS
-int morphology_gradient_par_for_morphology_gradient_s0_ch_par_for_morphology_gradient_s0_x_xo_ti_ti(void *__user_context, int32_t _morphology_gradient_s0_x_xo_ti_ti, uint8_t *_closure_arg__208);
+int morphology_gradient_par_for_morphology_gradient__4_s0_ch_par_for_morphology_gradient__4_s0_x_xo_ti_ti(void *__user_context, int32_t _morphology_gradient__4_s0_x_xo_ti_ti, uint8_t *_closure_arg__212);
 HALIDE_FUNCTION_ATTRS
-int morphology_gradient_par_for_morphology_gradient_s0_ch(void *__user_context, int32_t _morphology_gradient_s0_ch, uint8_t *_closure_arg__207);
+int morphology_gradient_par_for_morphology_gradient__4_s0_ch(void *__user_context, int32_t _morphology_gradient__4_s0_ch, uint8_t *_closure_arg__211);
 HALIDE_FUNCTION_ATTRS
-int morphology_gradient(struct halide_buffer_t *_src_buffer, int32_t _width, int32_t _height, int32_t _size, struct halide_buffer_t *_morphology_gradient_buffer);
+int morphology_gradient(struct halide_buffer_t *_src_buffer, int32_t _width, int32_t _height, int32_t _size, struct halide_buffer_t *_morphology_gradient__4_buffer);
 
 HALIDE_FUNCTION_ATTRS
 int morphology_gradient_argv(void **args);
diff --git a/include/libmorphology_open_linux.h b/include/libmorphology_open_linux.h
index af50105..57bf5a8 100644
--- a/include/libmorphology_open_linux.h
+++ b/include/libmorphology_open_linux.h
@@ -41,17 +41,17 @@ extern "C" {
 #endif
 
 HALIDE_FUNCTION_ATTRS
-int morphology_open_par_for_in_30_s0_y_rebased__1(void *__user_context, int32_t _in__30_s0_y_rebased, uint8_t *_closure_arg__188);
+int morphology_open_par_for_in_30_s0_y_rebased__1(void *__user_context, int32_t _in__30_s0_y_rebased, uint8_t *_closure_arg__192);
 HALIDE_FUNCTION_ATTRS
-int morphology_open_par_for_erode_tmp_s0_y_y__1(void *__user_context, int32_t _erode_tmp_s0_y_y, uint8_t *_closure_arg__189);
+int morphology_open_par_for_erode_tmp_s0_y_y__1(void *__user_context, int32_t _erode_tmp_s0_y_y, uint8_t *_closure_arg__193);
 HALIDE_FUNCTION_ATTRS
-int morphology_open_par_for_dilate_tmp_3_s0_y_y__1(void *__user_context, int32_t _dilate_tmp__3_s0_y_y, uint8_t *_closure_arg__190);
+int morphology_open_par_for_dilate_tmp_s0_y_y__1(void *__user_context, int32_t _dilate_tmp_s0_y_y, uint8_t *_closure_arg__194);
 HALIDE_FUNCTION_ATTRS
-int morphology_open_par_for_morphology_open_s0_ch_par_for_morphology_open_s0_x_xo_ti_ti__2(void *__user_context, int32_t _morphology_open_s0_x_xo_ti_ti, uint8_t *_closure_arg__192);
+int morphology_open_par_for_morphology_open_s0_ch_par_for_morphology_open_s0_x_xo_ti_ti__2(void *__user_context, int32_t _morphology_open_s0_x_xo_ti_ti, uint8_t *_closure_arg__196);
 HALIDE_FUNCTION_ATTRS
-int morphology_open_par_for_morphology_open_s0_ch_par_for_morphology_open_s0_x_xo_ti_ti__3(void *__user_context, int32_t _morphology_open_s0_x_xo_ti_ti, uint8_t *_closure_arg__193);
+int morphology_open_par_for_morphology_open_s0_ch_par_for_morphology_open_s0_x_xo_ti_ti__3(void *__user_context, int32_t _morphology_open_s0_x_xo_ti_ti, uint8_t *_closure_arg__197);
 HALIDE_FUNCTION_ATTRS
-int morphology_open_par_for_morphology_open_s0_ch__1(void *__user_context, int32_t _morphology_open_s0_ch, uint8_t *_closure_arg__191);
+int morphology_open_par_for_morphology_open_s0_ch__1(void *__user_context, int32_t _morphology_open_s0_ch, uint8_t *_closure_arg__195);
 HALIDE_FUNCTION_ATTRS
 int morphology_open(struct halide_buffer_t *_src_buffer, int32_t _width, int32_t _height, int32_t _size, struct halide_buffer_t *_morphology_open_buffer);
 
diff --git a/include/libmorphology_open_osx.h b/include/libmorphology_open_osx.h
index 92056db..81e325a 100644
--- a/include/libmorphology_open_osx.h
+++ b/include/libmorphology_open_osx.h
@@ -41,17 +41,17 @@ extern "C" {
 #endif
 
 HALIDE_FUNCTION_ATTRS
-int morphology_open_par_for_in__30_s0_y_rebased(void *__user_context, int32_t _in__30_s0_y_rebased, uint8_t *_closure_arg__182);
+int morphology_open_par_for_in__30_s0_y_rebased(void *__user_context, int32_t _in__30_s0_y_rebased, uint8_t *_closure_arg__186);
 HALIDE_FUNCTION_ATTRS
-int morphology_open_par_for_erode_tmp_s0_y_y(void *__user_context, int32_t _erode_tmp_s0_y_y, uint8_t *_closure_arg__183);
+int morphology_open_par_for_erode_tmp_s0_y_y(void *__user_context, int32_t _erode_tmp_s0_y_y, uint8_t *_closure_arg__187);
 HALIDE_FUNCTION_ATTRS
-int morphology_open_par_for_dilate_tmp__3_s0_y_y(void *__user_context, int32_t _dilate_tmp__3_s0_y_y, uint8_t *_closure_arg__184);
+int morphology_open_par_for_dilate_tmp_s0_y_y(void *__user_context, int32_t _dilate_tmp_s0_y_y, uint8_t *_closure_arg__188);
 HALIDE_FUNCTION_ATTRS
-int morphology_open_par_for_morphology_open_s0_ch_par_for_morphology_open_s0_x_xo_ti_ti(void *__user_context, int32_t _morphology_open_s0_x_xo_ti_ti, uint8_t *_closure_arg__186);
+int morphology_open_par_for_morphology_open_s0_ch_par_for_morphology_open_s0_x_xo_ti_ti(void *__user_context, int32_t _morphology_open_s0_x_xo_ti_ti, uint8_t *_closure_arg__190);
 HALIDE_FUNCTION_ATTRS
-int morphology_open_par_for_morphology_open_s0_ch_par_for_morphology_open_s0_x_xo_ti_ti__1(void *__user_context, int32_t _morphology_open_s0_x_xo_ti_ti, uint8_t *_closure_arg__187);
+int morphology_open_par_for_morphology_open_s0_ch_par_for_morphology_open_s0_x_xo_ti_ti__1(void *__user_context, int32_t _morphology_open_s0_x_xo_ti_ti, uint8_t *_closure_arg__191);
 HALIDE_FUNCTION_ATTRS
-int morphology_open_par_for_morphology_open_s0_ch(void *__user_context, int32_t _morphology_open_s0_ch, uint8_t *_closure_arg__185);
+int morphology_open_par_for_morphology_open_s0_ch(void *__user_context, int32_t _morphology_open_s0_ch, uint8_t *_closure_arg__189);
 HALIDE_FUNCTION_ATTRS
 int morphology_open(struct halide_buffer_t *_src_buffer, int32_t _width, int32_t _height, int32_t _size, struct halide_buffer_t *_morphology_open_buffer);
 
diff --git a/include/libprepare_ncc_template_linux.h b/include/libprepare_ncc_template_linux.h
index acff73b..080a687 100644
--- a/include/libprepare_ncc_template_linux.h
+++ b/include/libprepare_ncc_template_linux.h
@@ -41,7 +41,7 @@ extern "C" {
 #endif
 
 HALIDE_FUNCTION_ATTRS
-int prepare_ncc_template_par_for_prepare_ncc_template_s0_x_xo_ti__8(void *__user_context, int32_t _prepare_ncc_template_s0_x_xo_ti, uint8_t *_closure_arg__289);
+int prepare_ncc_template_par_for_prepare_ncc_template_s0_x_xo_ti__8(void *__user_context, int32_t _prepare_ncc_template_s0_x_xo_ti, uint8_t *_closure_arg__314);
 HALIDE_FUNCTION_ATTRS
 int prepare_ncc_template(struct halide_buffer_t *_tpl_buffer, int32_t _tpl_width, int32_t _tpl_height, struct halide_buffer_t *_prepare_ncc_template_0_buffer, struct halide_buffer_t *_prepare_ncc_template_1_buffer);
 
diff --git a/include/libprepare_ncc_template_osx.h b/include/libprepare_ncc_template_osx.h
index 2ec6c13..2fe5ddd 100644
--- a/include/libprepare_ncc_template_osx.h
+++ b/include/libprepare_ncc_template_osx.h
@@ -41,7 +41,7 @@ extern "C" {
 #endif
 
 HALIDE_FUNCTION_ATTRS
-int prepare_ncc_template_par_for_prepare_ncc_template_s0_x_xo_ti__7(void *__user_context, int32_t _prepare_ncc_template_s0_x_xo_ti, uint8_t *_closure_arg__288);
+int prepare_ncc_template_par_for_prepare_ncc_template_s0_x_xo_ti__7(void *__user_context, int32_t _prepare_ncc_template_s0_x_xo_ti, uint8_t *_closure_arg__313);
 HALIDE_FUNCTION_ATTRS
 int prepare_ncc_template(struct halide_buffer_t *_tpl_buffer, int32_t _tpl_width, int32_t _tpl_height, struct halide_buffer_t *_prepare_ncc_template_0_buffer, struct halide_buffer_t *_prepare_ncc_template_1_buffer);
 
diff --git a/include/libprepare_zncc_template_linux.h b/include/libprepare_zncc_template_linux.h
index b123c24..60bc1ba 100644
--- a/include/libprepare_zncc_template_linux.h
+++ b/include/libprepare_zncc_template_linux.h
@@ -41,7 +41,7 @@ extern "C" {
 #endif
 
 HALIDE_FUNCTION_ATTRS
-int prepare_zncc_template_par_for_prepare_zncc_template_s0_x_xo_ti__1(void *__user_context, int32_t _prepare_zncc_template_s0_x_xo_ti, uint8_t *_closure_arg__293);
+int prepare_zncc_template_par_for_prepare_zncc_template_s0_x_xo_ti__1(void *__user_context, int32_t _prepare_zncc_template_s0_x_xo_ti, uint8_t *_closure_arg__318);
 HALIDE_FUNCTION_ATTRS
 int prepare_zncc_template(struct halide_buffer_t *_tpl_buffer, int32_t _tpl_width, int32_t _tpl_height, struct halide_buffer_t *_prepare_zncc_template_0_buffer, struct halide_buffer_t *_prepare_zncc_template_1_buffer);
 
diff --git a/include/libprepare_zncc_template_osx.h b/include/libprepare_zncc_template_osx.h
index 7f93fdb..a31a3b3 100644
--- a/include/libprepare_zncc_template_osx.h
+++ b/include/libprepare_zncc_template_osx.h
@@ -41,7 +41,7 @@ extern "C" {
 #endif
 
 HALIDE_FUNCTION_ATTRS
-int prepare_zncc_template_par_for_prepare_zncc_template_s0_x_xo_ti(void *__user_context, int32_t _prepare_zncc_template_s0_x_xo_ti, uint8_t *_closure_arg__292);
+int prepare_zncc_template_par_for_prepare_zncc_template_s0_x_xo_ti(void *__user_context, int32_t _prepare_zncc_template_s0_x_xo_ti, uint8_t *_closure_arg__317);
 HALIDE_FUNCTION_ATTRS
 int prepare_zncc_template(struct halide_buffer_t *_tpl_buffer, int32_t _tpl_width, int32_t _tpl_height, struct halide_buffer_t *_prepare_zncc_template_0_buffer, struct halide_buffer_t *_prepare_zncc_template_1_buffer);
 
diff --git a/include/libprepared_match_template_ncc_linux.h b/include/libprepared_match_template_ncc_linux.h
index e20a38b..8bbf0d1 100644
--- a/include/libprepared_match_template_ncc_linux.h
+++ b/include/libprepared_match_template_ncc_linux.h
@@ -41,9 +41,9 @@ extern "C" {
 #endif
 
 HALIDE_FUNCTION_ATTRS
-int prepared_match_template_ncc_par_for_prepared_match_template_ncc_5_s0_x_xo_ti__1(void *__user_context, int32_t _prepared_match_template_ncc__5_s0_x_xo_ti, uint8_t *_closure_arg__291);
+int prepared_match_template_ncc_par_for_prepared_match_template_ncc_s0_x_xo_ti__2(void *__user_context, int32_t _prepared_match_template_ncc_s0_x_xo_ti, uint8_t *_closure_arg__316);
 HALIDE_FUNCTION_ATTRS
-int prepared_match_template_ncc(struct halide_buffer_t *_src_buffer, int32_t _width, int32_t _height, struct halide_buffer_t *_tpl_val_buffer, struct halide_buffer_t *_tpl_sum_buffer, int32_t _tpl_width, int32_t _tpl_height, struct halide_buffer_t *_prepared_match_template_ncc__5_buffer);
+int prepared_match_template_ncc(struct halide_buffer_t *_src_buffer, int32_t _width, int32_t _height, struct halide_buffer_t *_tpl_val_buffer, struct halide_buffer_t *_tpl_sum_buffer, int32_t _tpl_width, int32_t _tpl_height, struct halide_buffer_t *_prepared_match_template_ncc_buffer);
 
 HALIDE_FUNCTION_ATTRS
 int prepared_match_template_ncc_argv(void **args);
diff --git a/include/libprepared_match_template_ncc_osx.h b/include/libprepared_match_template_ncc_osx.h
index fa30bbc..8501a53 100644
--- a/include/libprepared_match_template_ncc_osx.h
+++ b/include/libprepared_match_template_ncc_osx.h
@@ -41,9 +41,9 @@ extern "C" {
 #endif
 
 HALIDE_FUNCTION_ATTRS
-int prepared_match_template_ncc_par_for_prepared_match_template_ncc__5_s0_x_xo_ti(void *__user_context, int32_t _prepared_match_template_ncc__5_s0_x_xo_ti, uint8_t *_closure_arg__290);
+int prepared_match_template_ncc_par_for_prepared_match_template_ncc_s0_x_xo_ti__1(void *__user_context, int32_t _prepared_match_template_ncc_s0_x_xo_ti, uint8_t *_closure_arg__315);
 HALIDE_FUNCTION_ATTRS
-int prepared_match_template_ncc(struct halide_buffer_t *_src_buffer, int32_t _width, int32_t _height, struct halide_buffer_t *_tpl_val_buffer, struct halide_buffer_t *_tpl_sum_buffer, int32_t _tpl_width, int32_t _tpl_height, struct halide_buffer_t *_prepared_match_template_ncc__5_buffer);
+int prepared_match_template_ncc(struct halide_buffer_t *_src_buffer, int32_t _width, int32_t _height, struct halide_buffer_t *_tpl_val_buffer, struct halide_buffer_t *_tpl_sum_buffer, int32_t _tpl_width, int32_t _tpl_height, struct halide_buffer_t *_prepared_match_template_ncc_buffer);
 
 HALIDE_FUNCTION_ATTRS
 int prepared_match_template_ncc_argv(void **args);
diff --git a/include/libprepared_match_template_zncc_linux.h b/include/libprepared_match_template_zncc_linux.h
index c5f6853..e21cc4f 100644
--- a/include/libprepared_match_template_zncc_linux.h
+++ b/include/libprepared_match_template_zncc_linux.h
@@ -41,13 +41,13 @@ extern "C" {
 #endif
 
 HALIDE_FUNCTION_ATTRS
-int prepared_match_template_zncc_par_for_zncc_avg_1_s0_y_rebased__1(void *__user_context, int32_t _zncc_avg__1_s0_y_rebased, uint8_t *_closure_arg__297);
+int prepared_match_template_zncc_par_for_zncc_avg_1_s0_y_rebased__1(void *__user_context, int32_t _zncc_avg__1_s0_y_rebased, uint8_t *_closure_arg__322);
 HALIDE_FUNCTION_ATTRS
-int prepared_match_template_zncc_par_for_zncc_stddev_1_s0_y_rebased__1(void *__user_context, int32_t _zncc_stddev__1_s0_y_rebased, uint8_t *_closure_arg__298);
+int prepared_match_template_zncc_par_for_zncc_stddev_1_s0_y_rebased__1(void *__user_context, int32_t _zncc_stddev__1_s0_y_rebased, uint8_t *_closure_arg__323);
 HALIDE_FUNCTION_ATTRS
-int prepared_match_template_zncc_par_for_prepared_match_template_zncc_s0_x_xo_ti__1(void *__user_context, int32_t _prepared_match_template_zncc_s0_x_xo_ti, uint8_t *_closure_arg__299);
+int prepared_match_template_zncc_par_for_prepared_match_template_zncc_1_s0_x_xo_ti__1(void *__user_context, int32_t _prepared_match_template_zncc__1_s0_x_xo_ti, uint8_t *_closure_arg__324);
 HALIDE_FUNCTION_ATTRS
-int prepared_match_template_zncc(struct halide_buffer_t *_src_buffer, int32_t _width, int32_t _height, struct halide_buffer_t *_tpl_val_buffer, struct halide_buffer_t *_tpl_stddev_buffer, int32_t _tpl_width, int32_t _tpl_height, struct halide_buffer_t *_prepared_match_template_zncc_buffer);
+int prepared_match_template_zncc(struct halide_buffer_t *_src_buffer, int32_t _width, int32_t _height, struct halide_buffer_t *_tpl_val_buffer, struct halide_buffer_t *_tpl_stddev_buffer, int32_t _tpl_width, int32_t _tpl_height, struct halide_buffer_t *_prepared_match_template_zncc__1_buffer);
 
 HALIDE_FUNCTION_ATTRS
 int prepared_match_template_zncc_argv(void **args);
diff --git a/include/libprepared_match_template_zncc_osx.h b/include/libprepared_match_template_zncc_osx.h
index 37ef996..b343348 100644
--- a/include/libprepared_match_template_zncc_osx.h
+++ b/include/libprepared_match_template_zncc_osx.h
@@ -41,13 +41,13 @@ extern "C" {
 #endif
 
 HALIDE_FUNCTION_ATTRS
-int prepared_match_template_zncc_par_for_zncc_avg__1_s0_y_rebased(void *__user_context, int32_t _zncc_avg__1_s0_y_rebased, uint8_t *_closure_arg__294);
+int prepared_match_template_zncc_par_for_zncc_avg__1_s0_y_rebased(void *__user_context, int32_t _zncc_avg__1_s0_y_rebased, uint8_t *_closure_arg__319);
 HALIDE_FUNCTION_ATTRS
-int prepared_match_template_zncc_par_for_zncc_stddev__1_s0_y_rebased(void *__user_context, int32_t _zncc_stddev__1_s0_y_rebased, uint8_t *_closure_arg__295);
+int prepared_match_template_zncc_par_for_zncc_stddev__1_s0_y_rebased(void *__user_context, int32_t _zncc_stddev__1_s0_y_rebased, uint8_t *_closure_arg__320);
 HALIDE_FUNCTION_ATTRS
-int prepared_match_template_zncc_par_for_prepared_match_template_zncc_s0_x_xo_ti(void *__user_context, int32_t _prepared_match_template_zncc_s0_x_xo_ti, uint8_t *_closure_arg__296);
+int prepared_match_template_zncc_par_for_prepared_match_template_zncc__1_s0_x_xo_ti(void *__user_context, int32_t _prepared_match_template_zncc__1_s0_x_xo_ti, uint8_t *_closure_arg__321);
 HALIDE_FUNCTION_ATTRS
-int prepared_match_template_zncc(struct halide_buffer_t *_src_buffer, int32_t _width, int32_t _height, struct halide_buffer_t *_tpl_val_buffer, struct halide_buffer_t *_tpl_stddev_buffer, int32_t _tpl_width, int32_t _tpl_height, struct halide_buffer_t *_prepared_match_template_zncc_buffer);
+int prepared_match_template_zncc(struct halide_buffer_t *_src_buffer, int32_t _width, int32_t _height, struct halide_buffer_t *_tpl_val_buffer, struct halide_buffer_t *_tpl_stddev_buffer, int32_t _tpl_width, int32_t _tpl_height, struct halide_buffer_t *_prepared_match_template_zncc__1_buffer);
 
 HALIDE_FUNCTION_ATTRS
 int prepared_match_template_zncc_argv(void **args);
diff --git a/include/libscale_gaussian_linux.h b/include/libscale_gaussian_linux.h
index fb350d3..2901f9f 100644
--- a/include/libscale_gaussian_linux.h
+++ b/include/libscale_gaussian_linux.h
@@ -47,7 +47,7 @@ int scale_gaussian_par_for_scale_gaussian_s0_ch_par_for_scale_gaussian_s0_x_xo_t
 HALIDE_FUNCTION_ATTRS
 int scale_gaussian_par_for_scale_gaussian_s0_ch_par_for_scale_gaussian_s0_x_xo_ti_ti__3(void *__user_context, int32_t _scale_gaussian_s0_x_xo_ti_ti, uint8_t *_closure_arg__95);
 HALIDE_FUNCTION_ATTRS
-int scale_gaussian_par_for_scale_gaussian_s0_ch__1(void *__user_context, int32_t _scale_gaussian_s0_ch, uint8_t *_closure_arg__93);
+int scale_gaussian_par_for_scale_gaussian_s0_ch__5(void *__user_context, int32_t _scale_gaussian_s0_ch, uint8_t *_closure_arg__93);
 HALIDE_FUNCTION_ATTRS
 int scale_gaussian(struct halide_buffer_t *_src_buffer, int32_t _width, int32_t _height, int32_t _scale_width, int32_t _scale_height, struct halide_buffer_t *_scale_gaussian_buffer);
 
diff --git a/include/libscale_gaussian_osx.h b/include/libscale_gaussian_osx.h
index 4d0c5bf..2811566 100644
--- a/include/libscale_gaussian_osx.h
+++ b/include/libscale_gaussian_osx.h
@@ -47,7 +47,7 @@ int scale_gaussian_par_for_scale_gaussian_s0_ch_par_for_scale_gaussian_s0_x_xo_t
 HALIDE_FUNCTION_ATTRS
 int scale_gaussian_par_for_scale_gaussian_s0_ch_par_for_scale_gaussian_s0_x_xo_ti_ti__1(void *__user_context, int32_t _scale_gaussian_s0_x_xo_ti_ti, uint8_t *_closure_arg__91);
 HALIDE_FUNCTION_ATTRS
-int scale_gaussian_par_for_scale_gaussian_s0_ch(void *__user_context, int32_t _scale_gaussian_s0_ch, uint8_t *_closure_arg__89);
+int scale_gaussian_par_for_scale_gaussian_s0_ch__4(void *__user_context, int32_t _scale_gaussian_s0_ch, uint8_t *_closure_arg__89);
 HALIDE_FUNCTION_ATTRS
 int scale_gaussian(struct halide_buffer_t *_src_buffer, int32_t _width, int32_t _height, int32_t _scale_width, int32_t _scale_height, struct halide_buffer_t *_scale_gaussian_buffer);
 
diff --git a/include/libscale_linux.h b/include/libscale_linux.h
index 9f6c364..336050b 100644
--- a/include/libscale_linux.h
+++ b/include/libscale_linux.h
@@ -41,11 +41,11 @@ extern "C" {
 #endif
 
 HALIDE_FUNCTION_ATTRS
-int scale_par_for_scale_s0_ch_par_for_scale_s0_x_xo_ti__2(void *__user_context, int32_t _scale_s0_x_xo_ti, uint8_t *_closure_arg__70);
+int scale_par_for_scale_normal_s0_y_y__1(void *__user_context, int32_t _scale_normal_s0_y_y, uint8_t *_closure_arg__69);
 HALIDE_FUNCTION_ATTRS
-int scale_par_for_scale_s0_ch_par_for_scale_s0_x_xo_ti__3(void *__user_context, int32_t _scale_s0_x_xo_ti, uint8_t *_closure_arg__71);
+int scale_par_for_scale_s0_x_xo_ti_ti__2(void *__user_context, int32_t _scale_s0_x_xo_ti_ti, uint8_t *_closure_arg__70);
 HALIDE_FUNCTION_ATTRS
-int scale_par_for_scale_s0_ch__1(void *__user_context, int32_t _scale_s0_ch, uint8_t *_closure_arg__69);
+int scale_par_for_scale_s0_x_xo_ti_ti__3(void *__user_context, int32_t _scale_s0_x_xo_ti_ti, uint8_t *_closure_arg__71);
 HALIDE_FUNCTION_ATTRS
 int scale(struct halide_buffer_t *_src_buffer, int32_t _width, int32_t _height, int32_t _scale_width, int32_t _scale_height, struct halide_buffer_t *_scale_buffer);
 
diff --git a/include/libscale_osx.h b/include/libscale_osx.h
index deab8a6..9edfbdc 100644
--- a/include/libscale_osx.h
+++ b/include/libscale_osx.h
@@ -41,11 +41,11 @@ extern "C" {
 #endif
 
 HALIDE_FUNCTION_ATTRS
-int scale_par_for_scale_s0_ch_par_for_scale_s0_x_xo_ti(void *__user_context, int32_t _scale_s0_x_xo_ti, uint8_t *_closure_arg__67);
+int scale_par_for_scale_normal_s0_y_y(void *__user_context, int32_t _scale_normal_s0_y_y, uint8_t *_closure_arg__66);
 HALIDE_FUNCTION_ATTRS
-int scale_par_for_scale_s0_ch_par_for_scale_s0_x_xo_ti__1(void *__user_context, int32_t _scale_s0_x_xo_ti, uint8_t *_closure_arg__68);
+int scale_par_for_scale_s0_x_xo_ti_ti(void *__user_context, int32_t _scale_s0_x_xo_ti_ti, uint8_t *_closure_arg__67);
 HALIDE_FUNCTION_ATTRS
-int scale_par_for_scale_s0_ch(void *__user_context, int32_t _scale_s0_ch, uint8_t *_closure_arg__66);
+int scale_par_for_scale_s0_x_xo_ti_ti__1(void *__user_context, int32_t _scale_s0_x_xo_ti_ti, uint8_t *_closure_arg__68);
 HALIDE_FUNCTION_ATTRS
 int scale(struct halide_buffer_t *_src_buffer, int32_t _width, int32_t _height, int32_t _scale_width, int32_t _scale_height, struct halide_buffer_t *_scale_buffer);
 
diff --git a/include/libsobel_linux.h b/include/libsobel_linux.h
index b55000a..c076b9a 100644
--- a/include/libsobel_linux.h
+++ b/include/libsobel_linux.h
@@ -41,11 +41,11 @@ extern "C" {
 #endif
 
 HALIDE_FUNCTION_ATTRS
-int sobel_par_for_sobel_s0_ch_par_for_sobel_s0_x_xo_ti_ti__2(void *__user_context, int32_t _sobel_s0_x_xo_ti_ti, uint8_t *_closure_arg__248);
+int sobel_par_for_sobel_s0_ch_par_for_sobel_s0_x_xo_ti_ti__6(void *__user_context, int32_t _sobel_s0_x_xo_ti_ti, uint8_t *_closure_arg__252);
 HALIDE_FUNCTION_ATTRS
-int sobel_par_for_sobel_s0_ch_par_for_sobel_s0_x_xo_ti_ti__3(void *__user_context, int32_t _sobel_s0_x_xo_ti_ti, uint8_t *_closure_arg__249);
+int sobel_par_for_sobel_s0_ch_par_for_sobel_s0_x_xo_ti_ti__7(void *__user_context, int32_t _sobel_s0_x_xo_ti_ti, uint8_t *_closure_arg__253);
 HALIDE_FUNCTION_ATTRS
-int sobel_par_for_sobel_s0_ch__3(void *__user_context, int32_t _sobel_s0_ch, uint8_t *_closure_arg__247);
+int sobel_par_for_sobel_s0_ch__3(void *__user_context, int32_t _sobel_s0_ch, uint8_t *_closure_arg__251);
 HALIDE_FUNCTION_ATTRS
 int sobel(struct halide_buffer_t *_src_buffer, int32_t _width, int32_t _height, struct halide_buffer_t *_sobel_buffer);
 
diff --git a/include/libsobel_osx.h b/include/libsobel_osx.h
index 4d53faf..f2e2e52 100644
--- a/include/libsobel_osx.h
+++ b/include/libsobel_osx.h
@@ -41,11 +41,11 @@ extern "C" {
 #endif
 
 HALIDE_FUNCTION_ATTRS
-int sobel_par_for_sobel_s0_ch_par_for_sobel_s0_x_xo_ti_ti(void *__user_context, int32_t _sobel_s0_x_xo_ti_ti, uint8_t *_closure_arg__245);
+int sobel_par_for_sobel_s0_ch_par_for_sobel_s0_x_xo_ti_ti__4(void *__user_context, int32_t _sobel_s0_x_xo_ti_ti, uint8_t *_closure_arg__249);
 HALIDE_FUNCTION_ATTRS
-int sobel_par_for_sobel_s0_ch_par_for_sobel_s0_x_xo_ti_ti__1(void *__user_context, int32_t _sobel_s0_x_xo_ti_ti, uint8_t *_closure_arg__246);
+int sobel_par_for_sobel_s0_ch_par_for_sobel_s0_x_xo_ti_ti__5(void *__user_context, int32_t _sobel_s0_x_xo_ti_ti, uint8_t *_closure_arg__250);
 HALIDE_FUNCTION_ATTRS
-int sobel_par_for_sobel_s0_ch__2(void *__user_context, int32_t _sobel_s0_ch, uint8_t *_closure_arg__244);
+int sobel_par_for_sobel_s0_ch__2(void *__user_context, int32_t _sobel_s0_ch, uint8_t *_closure_arg__248);
 HALIDE_FUNCTION_ATTRS
 int sobel(struct halide_buffer_t *_src_buffer, int32_t _width, int32_t _height, struct halide_buffer_t *_sobel_buffer);
 
diff --git a/lib/libblend_add_linux.a b/lib/libblend_add_linux.a
index 1dd1301..4cfe7fe 100644
Binary files a/lib/libblend_add_linux.a and b/lib/libblend_add_linux.a differ
diff --git a/lib/libblend_add_osx.a b/lib/libblend_add_osx.a
index 7bc7baa..d426920 100644
Binary files a/lib/libblend_add_osx.a and b/lib/libblend_add_osx.a differ
diff --git a/lib/libblend_diff_linux.a b/lib/libblend_diff_linux.a
index 82af48a..d2823b3 100644
Binary files a/lib/libblend_diff_linux.a and b/lib/libblend_diff_linux.a differ
diff --git a/lib/libblend_diff_osx.a b/lib/libblend_diff_osx.a
index 33308f5..71fd513 100644
Binary files a/lib/libblend_diff_osx.a and b/lib/libblend_diff_osx.a differ
diff --git a/lib/libblend_normal_linux.a b/lib/libblend_normal_linux.a
index 63cd06b..2e0dce1 100644
Binary files a/lib/libblend_normal_linux.a and b/lib/libblend_normal_linux.a differ
diff --git a/lib/libblend_normal_osx.a b/lib/libblend_normal_osx.a
index 30395b7..92c595e 100644
Binary files a/lib/libblend_normal_osx.a and b/lib/libblend_normal_osx.a differ
diff --git a/lib/libblend_sub_linux.a b/lib/libblend_sub_linux.a
index 6d0c18a..ba2551c 100644
Binary files a/lib/libblend_sub_linux.a and b/lib/libblend_sub_linux.a differ
diff --git a/lib/libblend_sub_osx.a b/lib/libblend_sub_osx.a
index 291a693..e589fed 100644
Binary files a/lib/libblend_sub_osx.a and b/lib/libblend_sub_osx.a differ
diff --git a/lib/libblockmozaic_linux.a b/lib/libblockmozaic_linux.a
index 1db2f1a..2b46b64 100644
Binary files a/lib/libblockmozaic_linux.a and b/lib/libblockmozaic_linux.a differ
diff --git a/lib/libblockmozaic_osx.a b/lib/libblockmozaic_osx.a
index 84c1f88..9818fb7 100644
Binary files a/lib/libblockmozaic_osx.a and b/lib/libblockmozaic_osx.a differ
diff --git a/lib/libboxblur_linux.a b/lib/libboxblur_linux.a
index 269f569..ed79a59 100644
Binary files a/lib/libboxblur_linux.a and b/lib/libboxblur_linux.a differ
diff --git a/lib/libboxblur_osx.a b/lib/libboxblur_osx.a
index b8aebfe..10a531c 100644
Binary files a/lib/libboxblur_osx.a and b/lib/libboxblur_osx.a differ
diff --git a/lib/libbrightness_linux.a b/lib/libbrightness_linux.a
index 18073b1..1d315d8 100644
Binary files a/lib/libbrightness_linux.a and b/lib/libbrightness_linux.a differ
diff --git a/lib/libbrightness_osx.a b/lib/libbrightness_osx.a
index bfa2b3b..cd2696b 100644
Binary files a/lib/libbrightness_osx.a and b/lib/libbrightness_osx.a differ
diff --git a/lib/libcanny_dilate_linux.a b/lib/libcanny_dilate_linux.a
index 7109748..ab8650e 100644
Binary files a/lib/libcanny_dilate_linux.a and b/lib/libcanny_dilate_linux.a differ
diff --git a/lib/libcanny_dilate_osx.a b/lib/libcanny_dilate_osx.a
index e1fcbf0..f655c1a 100644
Binary files a/lib/libcanny_dilate_osx.a and b/lib/libcanny_dilate_osx.a differ
diff --git a/lib/libcanny_linux.a b/lib/libcanny_linux.a
index 4ca1fb3..1730694 100644
Binary files a/lib/libcanny_linux.a and b/lib/libcanny_linux.a differ
diff --git a/lib/libcanny_morphology_close_linux.a b/lib/libcanny_morphology_close_linux.a
index 0982cb1..db257b2 100644
Binary files a/lib/libcanny_morphology_close_linux.a and b/lib/libcanny_morphology_close_linux.a differ
diff --git a/lib/libcanny_morphology_close_osx.a b/lib/libcanny_morphology_close_osx.a
index 0e9c1f5..02d9bbc 100644
Binary files a/lib/libcanny_morphology_close_osx.a and b/lib/libcanny_morphology_close_osx.a differ
diff --git a/lib/libcanny_morphology_open_linux.a b/lib/libcanny_morphology_open_linux.a
index 1972222..3794317 100644
Binary files a/lib/libcanny_morphology_open_linux.a and b/lib/libcanny_morphology_open_linux.a differ
diff --git a/lib/libcanny_morphology_open_osx.a b/lib/libcanny_morphology_open_osx.a
index e73bbd0..fa085c8 100644
Binary files a/lib/libcanny_morphology_open_osx.a and b/lib/libcanny_morphology_open_osx.a differ
diff --git a/lib/libcanny_osx.a b/lib/libcanny_osx.a
index b27ed07..6f8e218 100644
Binary files a/lib/libcanny_osx.a and b/lib/libcanny_osx.a differ
diff --git a/lib/libcontour_linux.a b/lib/libcontour_linux.a
index 35fa8db..2cb24bb 100644
Binary files a/lib/libcontour_linux.a and b/lib/libcontour_linux.a differ
diff --git a/lib/libcontour_osx.a b/lib/libcontour_osx.a
index 2bcb9de..f69aa47 100644
Binary files a/lib/libcontour_osx.a and b/lib/libcontour_osx.a differ
diff --git a/lib/libcontrast_linux.a b/lib/libcontrast_linux.a
index 3a156c1..d2ee008 100644
Binary files a/lib/libcontrast_linux.a and b/lib/libcontrast_linux.a differ
diff --git a/lib/libcontrast_osx.a b/lib/libcontrast_osx.a
index 5f50971..8f41e81 100644
Binary files a/lib/libcontrast_osx.a and b/lib/libcontrast_osx.a differ
diff --git a/lib/libdilation_linux.a b/lib/libdilation_linux.a
index 5250275..ee49d46 100644
Binary files a/lib/libdilation_linux.a and b/lib/libdilation_linux.a differ
diff --git a/lib/libdilation_osx.a b/lib/libdilation_osx.a
index 0e0b88e..0771884 100644
Binary files a/lib/libdilation_osx.a and b/lib/libdilation_osx.a differ
diff --git a/lib/libedgedetect_linux.a b/lib/libedgedetect_linux.a
index be2accd..95651a5 100644
Binary files a/lib/libedgedetect_linux.a and b/lib/libedgedetect_linux.a differ
diff --git a/lib/libedgedetect_osx.a b/lib/libedgedetect_osx.a
index 25e7d03..1a994ef 100644
Binary files a/lib/libedgedetect_osx.a and b/lib/libedgedetect_osx.a differ
diff --git a/lib/libemboss_linux.a b/lib/libemboss_linux.a
index 5930eaa..5d13783 100644
Binary files a/lib/libemboss_linux.a and b/lib/libemboss_linux.a differ
diff --git a/lib/libemboss_osx.a b/lib/libemboss_osx.a
index da8a3a4..61dd798 100644
Binary files a/lib/libemboss_osx.a and b/lib/libemboss_osx.a differ
diff --git a/lib/liberosion_linux.a b/lib/liberosion_linux.a
index 8fff69b..495fbb7 100644
Binary files a/lib/liberosion_linux.a and b/lib/liberosion_linux.a differ
diff --git a/lib/liberosion_osx.a b/lib/liberosion_osx.a
index b921f49..538e3e0 100644
Binary files a/lib/liberosion_osx.a and b/lib/liberosion_osx.a differ
diff --git a/lib/libgammacorrection_linux.a b/lib/libgammacorrection_linux.a
index 2987d83..477517f 100644
Binary files a/lib/libgammacorrection_linux.a and b/lib/libgammacorrection_linux.a differ
diff --git a/lib/libgammacorrection_osx.a b/lib/libgammacorrection_osx.a
index 4cdc31f..c25d089 100644
Binary files a/lib/libgammacorrection_osx.a and b/lib/libgammacorrection_osx.a differ
diff --git a/lib/libgaussianblur_linux.a b/lib/libgaussianblur_linux.a
index 7b0feb7..22b392e 100644
Binary files a/lib/libgaussianblur_linux.a and b/lib/libgaussianblur_linux.a differ
diff --git a/lib/libgaussianblur_osx.a b/lib/libgaussianblur_osx.a
index 819f5fe..66a5d40 100644
Binary files a/lib/libgaussianblur_osx.a and b/lib/libgaussianblur_osx.a differ
diff --git a/lib/libgradient_linux.a b/lib/libgradient_linux.a
index dfd387c..a02ca3e 100644
Binary files a/lib/libgradient_linux.a and b/lib/libgradient_linux.a differ
diff --git a/lib/libgradient_osx.a b/lib/libgradient_osx.a
index fcdb037..04c0d5f 100644
Binary files a/lib/libgradient_osx.a and b/lib/libgradient_osx.a differ
diff --git a/lib/libgrayscale_linux.a b/lib/libgrayscale_linux.a
index a036f41..4ab0b38 100644
Binary files a/lib/libgrayscale_linux.a and b/lib/libgrayscale_linux.a differ
diff --git a/lib/libgrayscale_osx.a b/lib/libgrayscale_osx.a
index 623009b..a31063a 100644
Binary files a/lib/libgrayscale_osx.a and b/lib/libgrayscale_osx.a differ
diff --git a/lib/libhighpass_linux.a b/lib/libhighpass_linux.a
index 7736a24..0c2fb6e 100644
Binary files a/lib/libhighpass_linux.a and b/lib/libhighpass_linux.a differ
diff --git a/lib/libhighpass_osx.a b/lib/libhighpass_osx.a
index 81db458..dd4c83b 100644
Binary files a/lib/libhighpass_osx.a and b/lib/libhighpass_osx.a differ
diff --git a/lib/libinvert_linux.a b/lib/libinvert_linux.a
index bc1418a..abf05ba 100644
Binary files a/lib/libinvert_linux.a and b/lib/libinvert_linux.a differ
diff --git a/lib/libinvert_osx.a b/lib/libinvert_osx.a
index b0bcc83..5cced69 100644
Binary files a/lib/libinvert_osx.a and b/lib/libinvert_osx.a differ
diff --git a/lib/liblaplacian_linux.a b/lib/liblaplacian_linux.a
index 0a22796..df86b75 100644
Binary files a/lib/liblaplacian_linux.a and b/lib/liblaplacian_linux.a differ
diff --git a/lib/liblaplacian_osx.a b/lib/liblaplacian_osx.a
index 1db1608..01ec1f4 100644
Binary files a/lib/liblaplacian_osx.a and b/lib/liblaplacian_osx.a differ
diff --git a/lib/libmatch_template_ncc_linux.a b/lib/libmatch_template_ncc_linux.a
index 780739e..0eed660 100644
Binary files a/lib/libmatch_template_ncc_linux.a and b/lib/libmatch_template_ncc_linux.a differ
diff --git a/lib/libmatch_template_ncc_osx.a b/lib/libmatch_template_ncc_osx.a
index 012092e..2d716ee 100644
Binary files a/lib/libmatch_template_ncc_osx.a and b/lib/libmatch_template_ncc_osx.a differ
diff --git a/lib/libmatch_template_sad_linux.a b/lib/libmatch_template_sad_linux.a
index f7a844d..fafdb7f 100644
Binary files a/lib/libmatch_template_sad_linux.a and b/lib/libmatch_template_sad_linux.a differ
diff --git a/lib/libmatch_template_sad_osx.a b/lib/libmatch_template_sad_osx.a
index f464c3f..238ecf9 100644
Binary files a/lib/libmatch_template_sad_osx.a and b/lib/libmatch_template_sad_osx.a differ
diff --git a/lib/libmatch_template_ssd_linux.a b/lib/libmatch_template_ssd_linux.a
index d98e6a0..8c27c39 100644
Binary files a/lib/libmatch_template_ssd_linux.a and b/lib/libmatch_template_ssd_linux.a differ
diff --git a/lib/libmatch_template_ssd_osx.a b/lib/libmatch_template_ssd_osx.a
index 68f0509..31b0f06 100644
Binary files a/lib/libmatch_template_ssd_osx.a and b/lib/libmatch_template_ssd_osx.a differ
diff --git a/lib/libmatch_template_zncc_linux.a b/lib/libmatch_template_zncc_linux.a
index c785abd..22f2165 100644
Binary files a/lib/libmatch_template_zncc_linux.a and b/lib/libmatch_template_zncc_linux.a differ
diff --git a/lib/libmatch_template_zncc_osx.a b/lib/libmatch_template_zncc_osx.a
index 560f79a..384cfb1 100644
Binary files a/lib/libmatch_template_zncc_osx.a and b/lib/libmatch_template_zncc_osx.a differ
diff --git a/lib/libmorphology_close_linux.a b/lib/libmorphology_close_linux.a
index 3502a57..f0117f0 100644
Binary files a/lib/libmorphology_close_linux.a and b/lib/libmorphology_close_linux.a differ
diff --git a/lib/libmorphology_close_osx.a b/lib/libmorphology_close_osx.a
index 271ecfd..28ee23f 100644
Binary files a/lib/libmorphology_close_osx.a and b/lib/libmorphology_close_osx.a differ
diff --git a/lib/libmorphology_gradient_linux.a b/lib/libmorphology_gradient_linux.a
index 2d4c0b0..6e5fd52 100644
Binary files a/lib/libmorphology_gradient_linux.a and b/lib/libmorphology_gradient_linux.a differ
diff --git a/lib/libmorphology_gradient_osx.a b/lib/libmorphology_gradient_osx.a
index df39ec3..cd87e3f 100644
Binary files a/lib/libmorphology_gradient_osx.a and b/lib/libmorphology_gradient_osx.a differ
diff --git a/lib/libmorphology_open_linux.a b/lib/libmorphology_open_linux.a
index a243357..aa25ced 100644
Binary files a/lib/libmorphology_open_linux.a and b/lib/libmorphology_open_linux.a differ
diff --git a/lib/libmorphology_open_osx.a b/lib/libmorphology_open_osx.a
index 793b96a..e735765 100644
Binary files a/lib/libmorphology_open_osx.a and b/lib/libmorphology_open_osx.a differ
diff --git a/lib/libpcm16_decibel_linux.a b/lib/libpcm16_decibel_linux.a
index f3ff5c8..292628b 100644
Binary files a/lib/libpcm16_decibel_linux.a and b/lib/libpcm16_decibel_linux.a differ
diff --git a/lib/libpcm16_decibel_osx.a b/lib/libpcm16_decibel_osx.a
index 9d4e483..5c4391e 100644
Binary files a/lib/libpcm16_decibel_osx.a and b/lib/libpcm16_decibel_osx.a differ
diff --git a/lib/libprepare_ncc_template_linux.a b/lib/libprepare_ncc_template_linux.a
index f7f6b7a..c6b47ca 100644
Binary files a/lib/libprepare_ncc_template_linux.a and b/lib/libprepare_ncc_template_linux.a differ
diff --git a/lib/libprepare_ncc_template_osx.a b/lib/libprepare_ncc_template_osx.a
index fa7d0b3..e7ee11d 100644
Binary files a/lib/libprepare_ncc_template_osx.a and b/lib/libprepare_ncc_template_osx.a differ
diff --git a/lib/libprepare_zncc_template_linux.a b/lib/libprepare_zncc_template_linux.a
index 87d8368..2a943cc 100644
Binary files a/lib/libprepare_zncc_template_linux.a and b/lib/libprepare_zncc_template_linux.a differ
diff --git a/lib/libprepare_zncc_template_osx.a b/lib/libprepare_zncc_template_osx.a
index 29df443..39ab2be 100644
Binary files a/lib/libprepare_zncc_template_osx.a and b/lib/libprepare_zncc_template_osx.a differ
diff --git a/lib/libprepared_match_template_ncc_linux.a b/lib/libprepared_match_template_ncc_linux.a
index 76a1cfb..a32c0b4 100644
Binary files a/lib/libprepared_match_template_ncc_linux.a and b/lib/libprepared_match_template_ncc_linux.a differ
diff --git a/lib/libprepared_match_template_ncc_osx.a b/lib/libprepared_match_template_ncc_osx.a
index c665854..a9c1708 100644
Binary files a/lib/libprepared_match_template_ncc_osx.a and b/lib/libprepared_match_template_ncc_osx.a differ
diff --git a/lib/libprepared_match_template_zncc_linux.a b/lib/libprepared_match_template_zncc_linux.a
index a2c84d9..c3fe86c 100644
Binary files a/lib/libprepared_match_template_zncc_linux.a and b/lib/libprepared_match_template_zncc_linux.a differ
diff --git a/lib/libprepared_match_template_zncc_osx.a b/lib/libprepared_match_template_zncc_osx.a
index 7420b44..e894a81 100644
Binary files a/lib/libprepared_match_template_zncc_osx.a and b/lib/libprepared_match_template_zncc_osx.a differ
diff --git a/lib/libscale_box_linux.a b/lib/libscale_box_linux.a
index c46c686..4033ccc 100644
Binary files a/lib/libscale_box_linux.a and b/lib/libscale_box_linux.a differ
diff --git a/lib/libscale_box_osx.a b/lib/libscale_box_osx.a
index 6ba7615..7cfe00c 100644
Binary files a/lib/libscale_box_osx.a and b/lib/libscale_box_osx.a differ
diff --git a/lib/libscale_gaussian_linux.a b/lib/libscale_gaussian_linux.a
index 6f84de7..449bdb9 100644
Binary files a/lib/libscale_gaussian_linux.a and b/lib/libscale_gaussian_linux.a differ
diff --git a/lib/libscale_gaussian_osx.a b/lib/libscale_gaussian_osx.a
index fa452c3..3781178 100644
Binary files a/lib/libscale_gaussian_osx.a and b/lib/libscale_gaussian_osx.a differ
diff --git a/lib/libscale_linear_linux.a b/lib/libscale_linear_linux.a
index 226cc75..588a85a 100644
Binary files a/lib/libscale_linear_linux.a and b/lib/libscale_linear_linux.a differ
diff --git a/lib/libscale_linear_osx.a b/lib/libscale_linear_osx.a
index dd02881..bdbd67e 100644
Binary files a/lib/libscale_linear_osx.a and b/lib/libscale_linear_osx.a differ
diff --git a/lib/libscale_linux.a b/lib/libscale_linux.a
index 1bfd9ba..6e3e527 100644
Binary files a/lib/libscale_linux.a and b/lib/libscale_linux.a differ
diff --git a/lib/libscale_osx.a b/lib/libscale_osx.a
index 43c33d7..c3969a6 100644
Binary files a/lib/libscale_osx.a and b/lib/libscale_osx.a differ
diff --git a/lib/libsobel_linux.a b/lib/libsobel_linux.a
index cf73136..83ba640 100644
Binary files a/lib/libsobel_linux.a and b/lib/libsobel_linux.a differ
diff --git a/lib/libsobel_osx.a b/lib/libsobel_osx.a
index dde06d8..bd236a9 100644
Binary files a/lib/libsobel_osx.a and b/lib/libsobel_osx.a differ
diff --git a/version.go b/version.go
index 05a23a7..f6badfe 100644
--- a/version.go
+++ b/version.go
@@ -2,5 +2,5 @@ package blurry
 
 const (
 	AppName string = "blurry"
-	Version string = "1.20.0"
+	Version string = "1.20.1"
 )