kmill / lean4-raytracer Goto Github PK

I'm not sure if the C benchmark here is meant to be "optimised to the hilt" or just reasonably representative. Obviously you want it to somewhat correspond to the Lean implementation, but it's possible to teach the compiler more about this code.

All the functions except main in render.c could be marked static. Alternatively, -fwhole-program has the same effect in GCC. This improves the compiler's ability (or maybe willingness) to inline and specialize functions. Link-time optimisation could also be added on top of this with -flto that might further accelerate it.

Maybe these comments are not relevant and you just wanted to compare against reasonably optimised C. But if not, I hope these comments are helpful.

The following numbers were produced on my personal laptop with various programs running in the background (<5% noise). ASUS ZenBook UX434FAC (i7-10510U).

I've only run the default test once per version. I recommend using https://docs.rs/hyperfine for accurate benchmark results on a dedicated server with intentionally minimal noise and a static clock.

All single-threaded implementations clock around 3.7 GHz (+/- 0.1GHz), while the multi-threaded C clocked around 2.1 GHz (+/- 0.1 GHz) per core.

(Normalized to single threaded C time. Lower is better.)

Master branch Lean. 3.53x (~34m 30s)
Optimize branch Lean. 2.89x (~28m 16s) (I guess this implementation differs too much to be a fair comparison?)
Single threaded C. 1x (~9m 46s)
Multi-threaded C. 0.42x (~4m 9s) (Obviously flawed comparison)

--- render.c    2021-10-20 +0000
+++ render-single-core.c        2021-10-20 +0000
@@ -592,7 +592,7 @@

   printf("Starting %d threads.\n", num_threads);

-  pthread_t threads[num_threads];
+  struct RenderData *threads[num_threads];

   for (int i = 0; i < num_threads; i++) {
     struct RenderData *rd = malloc(sizeof(struct RenderData));
@@ -606,7 +606,8 @@
     rd->nobj = nobj;
     Rand_split(&rand, &rd->rand);
     rd->show_progress = (i == 0);
-    pthread_create(&threads[i], NULL, &render_task, (void*)rd);
+    render_task(rd);
+    threads[i] = rd;
   }

   Color *pixels = malloc(height * width * sizeof(Color));
@@ -615,8 +616,7 @@
   }

   for (int i = 0; i < num_threads; i++) {
-    struct RenderData *rd = NULL;
-    pthread_join(threads[i], (void **)&rd);
+    struct RenderData *rd = threads[i];
     Color *new_pixels = rd->pixels;
     for (int i = 0; i < height * width; i++) {
       pixels[i] = Vec3_add(pixels[i], new_pixels[i]);

There is no license file and the readme does not mention a license.

The repo should have a license file.