aboutsummaryrefslogtreecommitdiff
diff options
context:
space:
mode:
authorOnur Gündüz <ogunduz@gmail.com>2026-06-15 15:19:53 +0200
committerOnur Gündüz <ogunduz@gmail.com>2026-06-15 15:19:53 +0200
commitceecbea701e3ac882ce5f1693204152cc75c0d79 (patch)
tree397fc18d66aa68235f008d8d54a85683adb1ff5b
parent8e8836796e7dd22820c820c0916e83ecdbf38368 (diff)
downloadffwd-ceecbea701e3ac882ce5f1693204152cc75c0d79.tar.gz
ffwd-ceecbea701e3ac882ce5f1693204152cc75c0d79.zip
kernels: add LayerNorm and exact-GeLU for the BERT family
-rw-r--r--qwen_kernels.c38
-rw-r--r--qwen_kernels.h15
-rw-r--r--tests/test_kernels.c52
3 files changed, 105 insertions, 0 deletions
diff --git a/qwen_kernels.c b/qwen_kernels.c
index c7f5762..e1da7ed 100644
--- a/qwen_kernels.c
+++ b/qwen_kernels.c
@@ -768,6 +768,38 @@ void qwen_rms_norm_per_head(
}
}
+/* Mean-subtracting LayerNorm with bias. Safe in place (out may alias x): each
+ * row's mean and variance are reduced before the row is rewritten. Scalar for
+ * now; LayerNorm is not GEMM-bound, so SIMD/threading is a later perf step. */
+void qwen_layer_norm(float *out,
+ const float *x,
+ const float *gamma,
+ const float *beta,
+ int seq_len,
+ int hidden,
+ float eps) {
+ if (seq_len <= 0 || hidden <= 0)
+ return;
+ float inv_h = 1.0f / (float)hidden;
+ for (int s = 0; s < seq_len; s++) {
+ const float *row = x + (size_t)s * hidden;
+ float *dst = out + (size_t)s * hidden;
+ float mean = 0.0f;
+ for (int d = 0; d < hidden; d++)
+ mean += row[d];
+ mean *= inv_h;
+ float var = 0.0f;
+ for (int d = 0; d < hidden; d++) {
+ float diff = row[d] - mean;
+ var += diff * diff;
+ }
+ var *= inv_h;
+ float inv_std = 1.0f / sqrtf(var + eps);
+ for (int d = 0; d < hidden; d++)
+ dst[d] = gamma[d] * (row[d] - mean) * inv_std + beta[d];
+ }
+}
+
/* ========================================================================
* Activation Functions
* ======================================================================== */
@@ -799,6 +831,12 @@ void qwen_silu_mul_inplace(float *gate, const float *up, int n) {
#endif
}
+void qwen_gelu_inplace(float *x, int n) {
+ /* Exact erf GeLU; 0.70710678... = 1/sqrt(2). */
+ for (int i = 0; i < n; i++)
+ x[i] = 0.5f * x[i] * (1.0f + erff(x[i] * 0.70710678118654752f));
+}
+
void qwen_softmax(float *x, int rows, int cols) {
for (int r = 0; r < rows; r++) {
float *row = x + r * cols;
diff --git a/qwen_kernels.h b/qwen_kernels.h
index 3a6a4d6..36d2913 100644
--- a/qwen_kernels.h
+++ b/qwen_kernels.h
@@ -88,6 +88,17 @@ void qwen_rms_norm(
void qwen_rms_norm_per_head(
float *x, const float *weight, int seq_len, int n_heads, int head_dim, float eps);
+/* Layer Normalization (BERT family): out = gamma * (x - mean) / sqrt(var + eps)
+ * + beta, with mean and biased (population) variance over the hidden axis per
+ * row. Unlike RMSNorm this subtracts the mean and adds a bias. */
+void qwen_layer_norm(float *out,
+ const float *x,
+ const float *gamma,
+ const float *beta,
+ int seq_len,
+ int hidden,
+ float eps);
+
/* ========================================================================
* Activation Functions
* ======================================================================== */
@@ -96,6 +107,10 @@ void qwen_softmax(float *x, int rows, int cols);
/* gate = SiLU(gate) * up */
void qwen_silu_mul_inplace(float *gate, const float *up, int n);
+/* Exact (erf) GeLU in place: x = 0.5 * x * (1 + erf(x / sqrt(2))). This is the
+ * "gelu" activation HF BERT/BGE/MiniLM use, not the tanh approximation. */
+void qwen_gelu_inplace(float *x, int n);
+
/* ========================================================================
* Attention Operations
* ======================================================================== */
diff --git a/tests/test_kernels.c b/tests/test_kernels.c
index fb3d2df..0f92df6 100644
--- a/tests/test_kernels.c
+++ b/tests/test_kernels.c
@@ -58,6 +58,56 @@ static void test_rms_norm(void) {
}
}
+static void test_layer_norm(void) {
+ const int seq = 2, hidden = 4;
+ const float x[8] = {1.0f, 2.0f, -3.0f, 4.0f, 0.5f, -1.0f, 2.0f, -0.5f};
+ const float g[4] = {1.0f, 0.5f, -1.0f, 2.0f};
+ const float b[4] = {0.1f, -0.2f, 0.3f, -0.4f};
+ float out[8];
+
+ qwen_layer_norm(out, x, g, b, seq, hidden, 1e-12f);
+
+ for (int s = 0; s < seq; s++) {
+ float mean = 0.0f;
+ for (int d = 0; d < hidden; d++)
+ mean += x[s * hidden + d];
+ mean /= hidden;
+ float var = 0.0f;
+ for (int d = 0; d < hidden; d++) {
+ float diff = x[s * hidden + d] - mean;
+ var += diff * diff;
+ }
+ var /= hidden;
+ float inv = 1.0f / sqrtf(var + 1e-12f);
+ for (int d = 0; d < hidden; d++) {
+ float want = g[d] * (x[s * hidden + d] - mean) * inv + b[d];
+ expect_close("layer_norm", out[s * hidden + d], want, 1e-5f);
+ }
+ }
+
+ /* In-place (out aliases x) must match the out-of-place result. */
+ float z[8];
+ memcpy(z, x, sizeof(z));
+ qwen_layer_norm(z, z, g, b, seq, hidden, 1e-12f);
+ for (int i = 0; i < 8; i++)
+ expect_close("layer_norm_inplace", z[i], out[i], 1e-6f);
+}
+
+static void test_gelu(void) {
+ float x[5] = {0.0f, 1.0f, -1.0f, 2.5f, -3.0f};
+ float out[5];
+ memcpy(out, x, sizeof(out));
+ qwen_gelu_inplace(out, 5);
+ for (int i = 0; i < 5; i++) {
+ float want = 0.5f * x[i] * (1.0f + erff(x[i] * 0.70710678118654752f));
+ expect_close("gelu", out[i], want, 1e-6f);
+ }
+ /* Known reference values for the exact erf GeLU. */
+ expect_close("gelu0", out[0], 0.0f, 1e-7f);
+ expect_close("gelu1", out[1], 0.841344746f, 1e-5f);
+ expect_close("gelu_neg1", out[2], -0.158655254f, 1e-5f);
+}
+
static void test_rope_neox(void) {
const int seq = 2, heads = 1, head_dim = 4;
const int positions[2] = {0, 3};
@@ -527,6 +577,8 @@ static void test_generic_vs_impl(void) {
int main(void) {
qwen_set_threads(1);
test_rms_norm();
+ test_layer_norm();
+ test_gelu();
test_rope_neox();
test_packed_gqa_attention();
test_packed_gqa_attention_long();