refactor: parallel (#30)

dkurt · Feb 8, 2024 · beb96ce · beb96ce
1 parent 14b850d
commit beb96ce
Show file tree

Hide file tree

Showing 2 changed files with 12 additions and 4 deletions.
diff --git a/src/julia.cpp b/src/julia.cpp
@@ -20,6 +20,9 @@ static const float cr = 0.0;
 static const uint8_t step_bound = 255;
 static const uint8_t upper_bound = 4;
 
+static const int parallel_factor = 4;
+static const int vector_factor = 4;
+
 // no Expr type in riscv build
 #ifndef __riscv
 // src :: Int -> fraction :: Float -> Float
@@ -67,7 +70,6 @@ void halide_julia(uint8_t* dst, int height, int width) {
     halide_julia_rv(output);
 #else
     static Func julia("julia");
-    static const int vector_factor = 16;
 
     if (!julia.defined()) {
         Var x, y;
@@ -91,12 +93,18 @@ void halide_julia(uint8_t* dst, int height, int width) {
         Expr esc_cond = HComplex(julia(x, y, index)).magnitude_squared() < upper_bound;
         Tuple first_escape = argmin(esc_cond);
 
+        julia.update(0).vectorize(x, 4);
+
         // proj to result
         Func result;
         result(x, y) = cast<uint8_t>(first_escape[0]);
 
-        julia.compute_at(result, y);
-        julia.update(0).vectorize(x, 4);
+        Var yi, yo;
+        result.split(y, yo, yi, parallel_factor);
+        result.parallel(yo);
+
+        julia.store_at(result, yo);
+        julia.compute_at(result, yi);
 
         Target target;
         target.os = Target::OS::Linux;

diff --git a/test/test_main.cpp b/test/test_main.cpp
@@ -165,7 +165,7 @@ TEST(convolution_nhwc, halide) {
 
 
 TEST(idw, halide) {
-    Mat src(height, width, CV_8U); 
+    Mat src(height, width, CV_8U);
     Mat dst(height, width, CV_8U), cl_dst(height, width, CV_8UC3), dst_h(height, width, CV_8U), cl_dst_h(height, width, CV_8UC3);
     // randu(src, 0, 256);