seems like merge to kernel together is faster.

tests/test_triton.py

@@ -5,6 +5,14 @@ import torch
 from loguru import logger
 
 logger.add("triton.log", rotation="1 MB", retention="10 days", level="DEBUG")
+autotune_results = []
+
+
+def log_autotune(config, timing):
+    autotune_results.append((config, timing))
+    logger.info(f"Tested: {config} → {timing * 1e6:.2f} us")
+
+
 configs = [
     # 针对小规模矩阵的配置（例如 M,N,K < 1024）
     triton.Config(
@@ -83,6 +91,7 @@ def matmul_kernel(
         b_val = tl.load(b_ptr, mask=b_mask, other=0.0)
         acc += tl.dot(a_val, b_val)
     acc.to(tl.bfloat16)
+    acc = tl.sigmoid(acc)
     c_ptr = c + (block_m[:, None] * stride_cm + block_n[None, :] * stride_cn)
     c_mask = (block_m[:, None] < m) & (block_n[None, :] < n)
     tl.store(c_ptr, acc, mask=c_mask)
@@ -146,13 +155,19 @@ def benchmark_triton(M, N, K):
     return time_ms
 
 
+def torch_matmul_sigmoid(a, b):
+    c = torch.matmul(a, b)
+    c = torch.sigmoid(c)
+    return c
+
+
 def benchmark_torch(M, N, K):
     # 假设你的 PyTorch 矩阵乘法函数名为 torch.matmul
     a = torch.randn(M, K, device="cuda", dtype=torch.bfloat16)
     b = torch.randn(K, N, device="cuda", dtype=torch.bfloat16)
 
     # 测速
-    time_ms = benchmark(torch.matmul, a, b)
+    time_ms = benchmark(torch_matmul_sigmoid, a, b)
     print(f"PyTorch 耗时: {time_ms:.3f} ms")
     return time_ms