aboutsummaryrefslogtreecommitdiff
path: root/libffwd/kernels.h
diff options
context:
space:
mode:
authorOnur Gündüz <ogunduz@gmail.com>2026-06-25 17:58:46 +0200
committerOnur Gündüz <ogunduz@gmail.com>2026-06-25 17:58:46 +0200
commitf7ef995ff54da671f9d7665b3d92ff25934c1e8b (patch)
tree1401d04c0f464b32bffd63f787a63a82c6959e97 /libffwd/kernels.h
parent417f861f3661c2253258b5683a627d961e81fc68 (diff)
downloadffwd-f7ef995ff54da671f9d7665b3d92ff25934c1e8b.tar.gz
ffwd-f7ef995ff54da671f9d7665b3d92ff25934c1e8b.zip
chore: run clang-format
Diffstat (limited to 'libffwd/kernels.h')
-rw-r--r--libffwd/kernels.h134
1 files changed, 66 insertions, 68 deletions
diff --git a/libffwd/kernels.h b/libffwd/kernels.h
index f36d0d3..2a917e0 100644
--- a/libffwd/kernels.h
+++ b/libffwd/kernels.h
@@ -27,8 +27,7 @@ void matmul_t(float *C, const float *A, const float *B, int M, int K, int N);
void linear(
float *y, const float *x, const float *W, const float *b, int seq_len, int in_dim, int out_dim);
-void linear_nobias(
- float *y, const float *x, const float *W, int seq_len, int in_dim, int out_dim);
+void linear_nobias(float *y, const float *x, const float *W, int seq_len, int in_dim, int out_dim);
/* bf16 weight variants */
@@ -38,27 +37,27 @@ void linear_nobias_bf16(
/* Small-sequence BF16 path: compute two projections with one threaded dispatch. */
void linear_nobias_bf16_pair(float *a,
- float *b,
- const float *x,
- const uint16_t *Wa_bf16,
- const uint16_t *Wb_bf16,
- int seq_len,
- int in_dim,
- int a_dim,
- int b_dim);
+ float *b,
+ const float *x,
+ const uint16_t *Wa_bf16,
+ const uint16_t *Wb_bf16,
+ int seq_len,
+ int in_dim,
+ int a_dim,
+ int b_dim);
/* Small-sequence BF16 path: compute Q/K/V matvecs with one threaded dispatch. */
void linear_nobias_bf16_qkv(float *q,
- float *k,
- float *v,
- const float *x,
- const uint16_t *Wq_bf16,
- const uint16_t *Wk_bf16,
- const uint16_t *Wv_bf16,
- int seq_len,
- int in_dim,
- int q_dim,
- int kv_dim);
+ float *k,
+ float *v,
+ const float *x,
+ const uint16_t *Wq_bf16,
+ const uint16_t *Wk_bf16,
+ const uint16_t *Wv_bf16,
+ int seq_len,
+ int in_dim,
+ int q_dim,
+ int kv_dim);
/* Dot product using the best available local SIMD implementation. */
float dot_f32(const float *a, const float *b, int n);
@@ -68,8 +67,7 @@ float dot_f32(const float *a, const float *b, int n);
* ======================================================================== */
/* RMS Normalization: out = x / rms(x) * weight */
-void rms_norm(
- float *out, const float *x, const float *weight, int seq_len, int hidden, float eps);
+void rms_norm(float *out, const float *x, const float *weight, int seq_len, int hidden, float eps);
/* Per-head RMS Normalization for Q/K norms in decoder
* x: [seq, n_heads, head_dim], weight: [head_dim]
@@ -81,12 +79,12 @@ void rms_norm_per_head(
* + beta, with mean and biased (population) variance over the hidden axis per
* row. Unlike RMSNorm this subtracts the mean and adds a bias. */
void layer_norm(float *out,
- const float *x,
- const float *gamma,
- const float *beta,
- int seq_len,
- int hidden,
- float eps);
+ const float *x,
+ const float *gamma,
+ const float *beta,
+ int seq_len,
+ int hidden,
+ float eps);
/* ========================================================================
* Activation Functions
@@ -119,57 +117,57 @@ void gelu_tanh_inplace(float *x, int n);
* offsets[b]..offsets[b+1] span.
*/
void bidirectional_gqa_attention_packed(float *out,
- const float *Q,
- const float *K,
- const float *V,
- const int *offsets,
- int batch,
- int n_heads,
- int n_kv_heads,
- int head_dim,
- float scale);
+ const float *Q,
+ const float *K,
+ const float *V,
+ const int *offsets,
+ int batch,
+ int n_heads,
+ int n_kv_heads,
+ int head_dim,
+ float scale);
/* Optional reusable scratch for the BLAS-tiled packed attention path. */
size_t bidirectional_gqa_attention_packed_scratch_bytes(const int *offsets, int batch);
void bidirectional_gqa_attention_packed_with_scratch(float *out,
- const float *Q,
- const float *K,
- const float *V,
- const int *offsets,
- int batch,
- int n_heads,
- int n_kv_heads,
- int head_dim,
- float scale,
- float *scratch,
- size_t scratch_bytes);
+ const float *Q,
+ const float *K,
+ const float *V,
+ const int *offsets,
+ int batch,
+ int n_heads,
+ int n_kv_heads,
+ int head_dim,
+ float scale,
+ float *scratch,
+ size_t scratch_bytes);
/*
* Causal GQA attention over a packed/ragged batch. Query position i attends
* only to positions at or before i within its sequence.
*/
void causal_gqa_attention_packed(float *out,
- const float *Q,
- const float *K,
- const float *V,
- const int *offsets,
- int batch,
- int n_heads,
- int n_kv_heads,
- int head_dim,
- float scale);
+ const float *Q,
+ const float *K,
+ const float *V,
+ const int *offsets,
+ int batch,
+ int n_heads,
+ int n_kv_heads,
+ int head_dim,
+ float scale);
void causal_gqa_attention_packed_with_scratch(float *out,
- const float *Q,
- const float *K,
- const float *V,
- const int *offsets,
- int batch,
- int n_heads,
- int n_kv_heads,
- int head_dim,
- float scale,
- float *scratch,
- size_t scratch_bytes);
+ const float *Q,
+ const float *K,
+ const float *V,
+ const int *offsets,
+ int batch,
+ int n_heads,
+ int n_kv_heads,
+ int head_dim,
+ float scale,
+ float *scratch,
+ size_t scratch_bytes);
/* ========================================================================
* Position Embeddings