From a84ee9dd685e67b72b8ab0f1b51ecec95bd1a41b Mon Sep 17 00:00:00 2001
From: Federico Ficarelli <1379149+nazavode@users.noreply.github.com>
Date: Mon, 6 Nov 2023 13:19:21 +0100
Subject: [PATCH] Add kernels/dense/fused

---
 kernels/dense/8x8xf64/Makefile |  1 +
 kernels/dense/8x8xf64/fused.c  | 24 ++++++++++++++++++++++++
 2 files changed, 25 insertions(+)
 create mode 100644 kernels/dense/8x8xf64/fused.c

diff --git a/kernels/dense/8x8xf64/Makefile b/kernels/dense/8x8xf64/Makefile
index 26ef0923..d51bb71b 100644
--- a/kernels/dense/8x8xf64/Makefile
+++ b/kernels/dense/8x8xf64/Makefile
@@ -4,6 +4,7 @@ include ../../../snitch/Makefile.rules
 
 TESTS =
 TESTS += baseline.x
+TESTS += fused.x
 TESTS += linalg.x
 
 include ../../Makefile.kernels
diff --git a/kernels/dense/8x8xf64/fused.c b/kernels/dense/8x8xf64/fused.c
new file mode 100644
index 00000000..05a29b4c
--- /dev/null
+++ b/kernels/dense/8x8xf64/fused.c
@@ -0,0 +1,24 @@
+#include "data.h"
+
+#include <stdint.h>
+
+// Baseline version, fused, unaliased ptrs allow for reduntant loads to
+// be elided
+// * Inputs:  x[ M x K ]
+// * Weights: w[ K x N ]
+// * Biases:  b[ M x N ]
+// * Outputs: y[ M x N ]
+void dense(const double* restrict x, const double* restrict w, const double* restrict b,
+           double* restrict y) {
+    // Y = relu(X W + B)
+    for (uint32_t i = 0; i < M; ++i) {
+        for (uint32_t j = 0; j < N; ++j) {
+            double v = y[i * N + j];
+            for (uint32_t k = 0; k < K; ++k) {
+                v += x[i * K + k] * w[k * N + j];
+            }
+            v += b[i * N + j];
+            y[i * N + j] = v > 0. ? v : 0.;
+        }
+    }
+}