{children};
}
diff --git a/web/src/layouts/Doc.astro b/web/src/layouts/Doc.astro
new file mode 100644
index 0000000..710e525
--- /dev/null
+++ b/web/src/layouts/Doc.astro
@@ -0,0 +1,20 @@
+---
+import Base from "./Base.astro";
+const { frontmatter } = Astro.props;
+---
+
- One tiny tensor language (
+ TinyCompiler turns
+ Per-pass numbers, and why -O2 is not simply better, are in the architecture notes.
+ C = relu(A @ B)) compiled to
- tiny-gpu SIMT kernels and to a
- tiny-tpu tile schedule, and run on the
- real simulator and RTL, bit-exact against numpy.
+ One tensor program. Two real chips. Bit-exact.
+ C = relu(A @ B) into
+ tiny-gpu SIMT kernels and into a
+ tiny-tpu tile schedule. Both run on the real
+ simulator and the real RTL, and both match numpy exactly.
+ What is checked, on every commit
+
+
+ Optimizing the tiny-gpu kernel (8×8 matmul + relu, measured)
+
+
+ level instructions cycles global loads -O0 naive 7320 10 1088 -O1 licm + strength reduction 5208 (−29%) 9 1024 -O2 tiling, unrolling, folding 3288 (−55%) 3 576 (−47%)
+ TinyCompiler is the fourth project in a series with + tiny-gpu, + tiny-tpu and tiny-cpu.
-