Copy pixels perf

1 files changed, 12 insertions, 14 deletions

diff --git a/graphics.c b/graphics.c
index a14bedd..17afe97 100755
--- a/graphics.c
+++ b/graphics.c
@@ -311,20 +311,18 @@ void draw_pixels_to_buffer(Pixel_Buffer dst, Box area, Pixel_Buffer src) {
   if (j0 < 0)          j0 = 0;
   if (j1 > dst.height) j1 = dst.height;
 
-  f64 w_coeff = src.width  / area.width;
-  f64 h_coeff = src.height / area.height;
-
-  for (i64 j = j0; j < j1; ++j) {
-    i64 src_j = (i64) floor((j - area.y + .5) *  h_coeff);
-    if (src_j < 0 || src_j >= src.height) continue;
-    vec4_f32 *q = src.pixels + src_j * src.stride;
-    vec4_f32 *p = dst.pixels + j     * dst.stride + i0;
-    for (i64 i = i0; i < i1; ++i) {
-      i64 src_i = (i64) floor((i - area.x + .5) * w_coeff);
-      if (src_i < 0 || src_i >= src.width) continue;
-      put_pixel_(p, q[src_i]);
-      ++p;
-    }
+  f64 di = src.width  / area.width;
+  f64 dj = src.height / area.height;
+  f64 jj = (j0 - area.y) * dj;
+  for (i64 j = j0; j < j1; ++j, jj += dj) {
+    if (jj < 0 || jj >= src.height) continue;
+    vec4_f32 *d     = dst.pixels + j * dst.stride + i0;
+    vec4_f32 *d_end = d + i1 - i0;
+    vec4_f32 *s     = src.pixels + (i64) jj * src.stride;
+    f64       ii    = (i0 - area.x) * di;
+    if (ii < 0 || ii >= src.width) continue;
+    for (; d < d_end; ++d, ii += di)
+      put_pixel_(d, s[(i64) ii]);
   }
 }