From 6e763dc446d8d31fb72f686ca547653b33140791 Mon Sep 17 00:00:00 2001
From: George Hotz <geohot@gmail.com>
Date: Mon, 6 Mar 2023 08:25:13 -0800
Subject: [PATCH] matmul example in readme

---
 README.md               | 11 +++++++++++
 tinygrad/codegen/gpu.py |  2 +-
 2 files changed, 12 insertions(+), 1 deletion(-)

diff --git a/README.md b/README.md
index 8360517ea7..a5fdad1b1f 100644
--- a/README.md
+++ b/README.md
@@ -66,6 +66,17 @@ print(x.grad)  # dz/dx
 print(y.grad)  # dz/dy
 ```
 
+## Is tinygrad fast?
+
+Try a matmul. See how, despite the style, it is fused into one kernel with the power of laziness. Currently getting 2.2 TFLOPS on my M1 Max, and will be 8 TFLOPS once it's using the M1 Tensor Cores.
+
+```python
+OPTLOCAL=1 GPU=1 DEBUG=3 python3 -c "from tinygrad.tensor import Tensor;
+N = 1024; a, b = Tensor.randn(N, N), Tensor.randn(N, N);
+c = (a.reshape(N, 1, N) * b.permute(1,0).reshape(1, N, N)).sum(axis=2);
+print((c.numpy() - (a.numpy() @ b.numpy())).mean())"
+```
+
 ## Neural networks?
 
 It turns out, a decent autograd tensor library is 90% of what you need for neural networks. Add an optimizer (SGD, RMSprop, and Adam implemented) from tinygrad.nn.optim, write some boilerplate minibatching code, and you have all you need.
diff --git a/tinygrad/codegen/gpu.py b/tinygrad/codegen/gpu.py
index 1bf8bdd2d7..88beafc754 100644
--- a/tinygrad/codegen/gpu.py
+++ b/tinygrad/codegen/gpu.py
@@ -275,7 +275,7 @@ class GPUCodegen(ASTKernel):
         for i in range(final_dimension-1, -1, -1):
           self.kernel += [f"int idx{i} = idx{final_dimension} % {self.output_shape[i]};", f"idx{final_dimension} = idx{final_dimension} / {self.output_shape[i]};\n"]
         self.output_shape = (prod(self.output_shape[0:final_dimension+1]), ) + self.output_shape[final_dimension+1:]
-        if DEBUG >= 3: print(f"replaced output shape with {self.output_shape}")
+        if DEBUG >= 4: print(f"replaced output shape with {self.output_shape}")
 
     # early ast
     accumulators : List[Token] = []