aboutsummaryrefslogtreecommitdiff
path: root/qwen_kernels_generic.c
diff options
context:
space:
mode:
authorOnur Gündüz <ogunduz@gmail.com>2026-06-10 08:02:03 +0200
committerOnur Gündüz <ogunduz@gmail.com>2026-06-14 08:27:03 +0200
commitec417bcb52f2aa8e24aeec312e8ddbf3a0798056 (patch)
tree99c94ce75bbb815d50e3da320031ea732d7588b3 /qwen_kernels_generic.c
parentfe2c526e2c219c24ab8379c557ffcd3b107542af (diff)
downloadffwd-ec417bcb52f2aa8e24aeec312e8ddbf3a0798056.tar.gz
ffwd-ec417bcb52f2aa8e24aeec312e8ddbf3a0798056.zip
kernels: drop asr_ prefix naming
Diffstat (limited to 'qwen_kernels_generic.c')
-rw-r--r--qwen_kernels_generic.c65
1 files changed, 65 insertions, 0 deletions
diff --git a/qwen_kernels_generic.c b/qwen_kernels_generic.c
new file mode 100644
index 0000000..cc25d3a
--- /dev/null
+++ b/qwen_kernels_generic.c
@@ -0,0 +1,65 @@
+/*
+ * qwen_kernels_generic.c - architecture-generic hot kernels
+ */
+
+#include "qwen_kernels_impl.h"
+
+#include <string.h>
+
+void qwen_bf16_matvec_fused_generic(float *y, const float *x, const uint16_t *W_bf16,
+ const float *bias, int in_dim, int out_dim) {
+ for (int o = 0; o < out_dim; o++) {
+ const uint16_t *w_row = W_bf16 + (size_t)o * in_dim;
+ float sum = bias ? bias[o] : 0.0f;
+ for (int k = 0; k < in_dim; k++) {
+ uint32_t f32_bits = ((uint32_t)w_row[k]) << 16;
+ float w_val;
+ memcpy(&w_val, &f32_bits, sizeof(float));
+ sum += w_val * x[k];
+ }
+ y[o] = sum;
+ }
+}
+
+void qwen_argmax_bf16_range_generic(const float *x, const uint16_t *W_bf16,
+ int in_dim, int start, int end,
+ int *best_out, float *best_val_out) {
+ int best = start;
+ float best_val = -1e30f;
+
+ for (int o = start; o < end; o++) {
+ const uint16_t *w_row = W_bf16 + (size_t)o * in_dim;
+ float sum = 0.0f;
+ for (int k = 0; k < in_dim; k++) {
+ uint32_t f32_bits = ((uint32_t)w_row[k]) << 16;
+ float w_val;
+ memcpy(&w_val, &f32_bits, sizeof(float));
+ sum += w_val * x[k];
+ }
+ if (sum > best_val) {
+ best_val = sum;
+ best = o;
+ }
+ }
+
+ *best_out = best;
+ *best_val_out = best_val;
+}
+
+float qwen_dot_f32_generic(const float *a, const float *b, int n) {
+ float sum = 0.0f;
+ for (int i = 0; i < n; i++) sum += a[i] * b[i];
+ return sum;
+}
+
+void qwen_vec_scale_inplace_generic(float *dst, float scale, int n) {
+ for (int i = 0; i < n; i++) dst[i] *= scale;
+}
+
+void qwen_vec_axpy_inplace_generic(float *dst, const float *src, float alpha, int n) {
+ for (int i = 0; i < n; i++) dst[i] += alpha * src[i];
+}
+
+void qwen_vec_scale_add_generic(float *dst, const float *src, float correction, int n) {
+ for (int i = 0; i < n; i++) dst[i] = dst[i] * correction + src[i];
+}