aboutsummaryrefslogtreecommitdiff
path: root/src/kernels.c
diff options
context:
space:
mode:
Diffstat (limited to 'src/kernels.c')
-rw-r--r--src/kernels.c1469
1 files changed, 1469 insertions, 0 deletions
diff --git a/src/kernels.c b/src/kernels.c
new file mode 100644
index 0000000..226da45
--- /dev/null
+++ b/src/kernels.c
@@ -0,0 +1,1469 @@
+#include "kernels.h"
+#include "kernels_impl.h"
+#include <math.h>
+#include <string.h>
+#include <stdlib.h>
+#include <stdio.h>
+#include <pthread.h>
+#if (defined(__AVX512F__) || defined(__AVX2__)) && \
+ (defined(__x86_64__) || defined(__i386__) || defined(_M_X64) || defined(_M_IX86))
+#include <immintrin.h>
+#endif
+#if defined(__ARM_NEON) || defined(__ARM_NEON__)
+#include <arm_neon.h>
+#endif
+#ifdef __APPLE__
+#include <sys/sysctl.h>
+#else
+#include <unistd.h>
+#endif
+
+#ifdef USE_BLAS
+#ifdef __APPLE__
+#include <Accelerate/Accelerate.h>
+#else
+#include <cblas.h>
+#endif
+#endif
+
+#if defined(USE_OPENBLAS)
+void openblas_set_num_threads(int num_threads);
+#endif
+
+#ifndef M_PI
+#define M_PI 3.14159265358979323846
+#endif
+
+/* ========================================================================
+ * Thread Pool
+ * ======================================================================== */
+
+#define EMBED_MAX_THREADS 16
+
+typedef void (*parallel_fn_t)(int tid, int n_threads, void *arg);
+
+static struct {
+ pthread_t threads[EMBED_MAX_THREADS - 1];
+ int tids[EMBED_MAX_THREADS - 1];
+ int n_threads;
+ int shutdown;
+
+ parallel_fn_t fn;
+ void *arg;
+ int generation;
+
+ pthread_mutex_t mutex;
+ pthread_cond_t cond_work;
+ pthread_cond_t cond_done;
+ int n_done;
+} tp = {
+ .n_threads = 1,
+ .shutdown = 0,
+ .generation = 0,
+ .mutex = PTHREAD_MUTEX_INITIALIZER,
+ .cond_work = PTHREAD_COND_INITIALIZER,
+ .cond_done = PTHREAD_COND_INITIALIZER,
+};
+
+static void *worker_loop(void *arg) {
+ int tid = *(int *)arg;
+ int my_gen = 0;
+
+ for (;;) {
+ pthread_mutex_lock(&tp.mutex);
+ while (tp.generation == my_gen && !tp.shutdown)
+ pthread_cond_wait(&tp.cond_work, &tp.mutex);
+ if (tp.shutdown) {
+ pthread_mutex_unlock(&tp.mutex);
+ return NULL;
+ }
+ my_gen = tp.generation;
+ parallel_fn_t fn = tp.fn;
+ void *a = tp.arg;
+ int nt = tp.n_threads;
+ pthread_mutex_unlock(&tp.mutex);
+
+ fn(tid, nt, a);
+
+ pthread_mutex_lock(&tp.mutex);
+ if (++tp.n_done >= tp.n_threads - 1)
+ pthread_cond_signal(&tp.cond_done);
+ pthread_mutex_unlock(&tp.mutex);
+ }
+}
+
+void embed_set_threads(int n) {
+ if (n < 1)
+ n = 1;
+ if (n > EMBED_MAX_THREADS)
+ n = EMBED_MAX_THREADS;
+
+#if defined(USE_OPENBLAS)
+ /* OpenBLAS owns the heavy F32 GEMM parallelism on Linux. Keep the
+ * auxiliary pool serial to avoid nested thread contention. */
+ int blas_threads = n;
+ openblas_set_num_threads(blas_threads);
+ n = 1;
+#endif
+
+ /* Shutdown existing workers */
+ if (tp.n_threads > 1) {
+ pthread_mutex_lock(&tp.mutex);
+ tp.shutdown = 1;
+ pthread_cond_broadcast(&tp.cond_work);
+ pthread_mutex_unlock(&tp.mutex);
+ for (int i = 0; i < tp.n_threads - 1; i++)
+ pthread_join(tp.threads[i], NULL);
+ tp.shutdown = 0;
+ tp.generation = 0;
+ }
+
+ tp.n_threads = n;
+ if (n <= 1) {
+ if (embed_verbose >= 2) {
+#if defined(USE_OPENBLAS)
+ fprintf(stderr, "Thread pool: %d threads, OpenBLAS: %d threads\n", n, blas_threads);
+#else
+ fprintf(stderr, "Thread pool: %d threads\n", n);
+#endif
+ }
+ return;
+ }
+
+ for (int i = 0; i < n - 1; i++) {
+ tp.tids[i] = i + 1;
+ pthread_create(&tp.threads[i], NULL, worker_loop, &tp.tids[i]);
+ }
+
+ if (embed_verbose >= 2) {
+#if defined(USE_OPENBLAS)
+ fprintf(stderr, "Thread pool: %d threads, OpenBLAS: %d threads\n", n, blas_threads);
+#else
+ fprintf(stderr, "Thread pool: %d threads\n", n);
+#endif
+ }
+}
+
+#ifndef __APPLE__
+/* CPU bandwidth this process may actually use, in whole cores, or 0 if there
+ * is no cgroup limit. Containers (Docker --cpus, Kubernetes, RunPod) expose
+ * every host CPU through sysconf()/cpuset, but the CFS quota caps the CPU time
+ * the process really gets. Sizing the BLAS/thread pool to the host count then
+ * oversubscribes that quota: the threads spend the period's budget early and
+ * the scheduler throttles them for the rest of each period, which adds large,
+ * bursty stalls (e.g. a 16-thread default against a 5.1-core quota ran the
+ * 0.6B model ~2x slower than a 5-thread pool). Floor, not ceil, so a fully
+ * busy pool stays under the quota and never trips throttling. */
+static int cgroup_cpu_quota(void) {
+ /* cgroup v2: "/sys/fs/cgroup/cpu.max" is "<quota> <period>", or
+ * "max <period>" when unlimited. */
+ FILE *f = fopen("/sys/fs/cgroup/cpu.max", "r");
+ if (f) {
+ char quota[32];
+ long period = 0;
+ int got = fscanf(f, "%31s %ld", quota, &period);
+ fclose(f);
+ if (got == 2 && period > 0 && strcmp(quota, "max") != 0) {
+ long q = atol(quota);
+ if (q > 0)
+ return q / period < 1 ? 1 : (int)(q / period);
+ }
+ }
+ /* cgroup v1: separate files; quota -1 means unlimited. */
+ long q = -1, period = 0;
+ f = fopen("/sys/fs/cgroup/cpu/cpu.cfs_quota_us", "r");
+ if (f) {
+ if (fscanf(f, "%ld", &q) != 1)
+ q = -1;
+ fclose(f);
+ }
+ f = fopen("/sys/fs/cgroup/cpu/cpu.cfs_period_us", "r");
+ if (f) {
+ if (fscanf(f, "%ld", &period) != 1)
+ period = 0;
+ fclose(f);
+ }
+ if (q > 0 && period > 0)
+ return q / period < 1 ? 1 : (int)(q / period);
+ return 0;
+}
+#endif
+
+int embed_get_num_cpus(void) {
+#ifdef __APPLE__
+ int n = 0;
+ size_t len = sizeof(n);
+ sysctlbyname("hw.ncpu", &n, &len, NULL, 0);
+ return n > 0 ? n : 1;
+#else
+ int online = (int)sysconf(_SC_NPROCESSORS_ONLN);
+ if (online < 1)
+ online = 1;
+ /* Honor a cgroup CPU quota when it is tighter than the visible core count,
+ * so the default thread pool does not oversubscribe a CPU-limited
+ * container and get throttled. */
+ int quota = cgroup_cpu_quota();
+ if (quota >= 1 && quota < online)
+ return quota;
+ return online;
+#endif
+}
+
+/* Dispatch work to all threads; main thread is tid=0 */
+static void parallel_for(parallel_fn_t fn, void *arg) {
+ if (tp.n_threads <= 1) {
+ fn(0, 1, arg);
+ return;
+ }
+
+ pthread_mutex_lock(&tp.mutex);
+ tp.fn = fn;
+ tp.arg = arg;
+ tp.n_done = 0;
+ tp.generation++;
+ pthread_cond_broadcast(&tp.cond_work);
+ pthread_mutex_unlock(&tp.mutex);
+
+ fn(0, tp.n_threads, arg);
+
+ pthread_mutex_lock(&tp.mutex);
+ while (tp.n_done < tp.n_threads - 1)
+ pthread_cond_wait(&tp.cond_done, &tp.mutex);
+ pthread_mutex_unlock(&tp.mutex);
+}
+
+/* ========================================================================
+ * Basic Element-wise Operations
+ * ======================================================================== */
+
+void embed_add_inplace(float *a, const float *b, int n) {
+ for (int i = 0; i < n; i++)
+ a[i] += b[i];
+}
+
+/* ========================================================================
+ * Matrix Operations
+ * ======================================================================== */
+
+void embed_matmul_t(float *C, const float *A, const float *B, int M, int K, int N) {
+#ifdef USE_BLAS
+ cblas_sgemm(CblasRowMajor, CblasNoTrans, CblasTrans, M, N, K, 1.0f, A, K, B, K, 0.0f, C, N);
+#else
+ for (int m = 0; m < M; m++) {
+ for (int n = 0; n < N; n++) {
+ float sum = 0.0f;
+ for (int k = 0; k < K; k++) {
+ sum += A[m * K + k] * B[n * K + k];
+ }
+ C[m * N + n] = sum;
+ }
+ }
+#endif
+}
+
+void embed_linear(float *y,
+ const float *x,
+ const float *W,
+ const float *b,
+ int seq_len,
+ int in_dim,
+ int out_dim) {
+#ifdef USE_BLAS
+ if (seq_len == 1) {
+ cblas_sgemv(CblasRowMajor, CblasNoTrans, out_dim, in_dim, 1.0f, W, in_dim, x, 1, 0.0f, y,
+ 1);
+ } else {
+ cblas_sgemm(CblasRowMajor, CblasNoTrans, CblasTrans, seq_len, out_dim, in_dim, 1.0f, x,
+ in_dim, W, in_dim, 0.0f, y, out_dim);
+ }
+ if (b != NULL) {
+ for (int s = 0; s < seq_len; s++) {
+ for (int o = 0; o < out_dim; o++) {
+ y[s * out_dim + o] += b[o];
+ }
+ }
+ }
+#else
+ for (int s = 0; s < seq_len; s++) {
+ const float *x_row = x + s * in_dim;
+ float *y_row = y + s * out_dim;
+ for (int o = 0; o < out_dim; o++) {
+ const float *w_row = W + o * in_dim;
+ float sum = (b != NULL) ? b[o] : 0.0f;
+ for (int i = 0; i < in_dim; i++) {
+ sum += x_row[i] * w_row[i];
+ }
+ y_row[o] = sum;
+ }
+ }
+#endif
+}
+
+void embed_linear_nobias(
+ float *y, const float *x, const float *W, int seq_len, int in_dim, int out_dim) {
+ embed_linear(y, x, W, NULL, seq_len, in_dim, out_dim);
+}
+
+/* Convert bf16 buffer to f32 buffer */
+void embed_bf16_to_f32_buf(float *dst, const uint16_t *src, size_t n) {
+#if defined(__ARM_NEON) || defined(__ARM_NEON__)
+ size_t i = 0;
+ uint32_t *d = (uint32_t *)(void *)dst;
+ for (; i + 8 <= n; i += 8) {
+ uint16x8_t v = vld1q_u16(src + i);
+ uint32x4_t lo = vshlq_n_u32(vmovl_u16(vget_low_u16(v)), 16);
+ uint32x4_t hi = vshlq_n_u32(vmovl_u16(vget_high_u16(v)), 16);
+ vst1q_u32(d + i, lo);
+ vst1q_u32(d + i + 4, hi);
+ }
+ for (; i < n; i++)
+ d[i] = ((uint32_t)src[i]) << 16;
+#else
+ uint32_t *d = (uint32_t *)(void *)dst;
+ for (size_t i = 0; i < n; i++)
+ d[i] = ((uint32_t)src[i]) << 16;
+#endif
+}
+
+/*
+ * Fused BF16 matvec: y[out_dim] = W_bf16[out_dim, in_dim] @ x[in_dim] + bias
+ * Processes 2 output rows at a time to amortize x vector loads.
+ */
+static void bf16_matvec_fused(
+ float *y, const float *x, const uint16_t *W_bf16, const float *bias, int in_dim, int out_dim) {
+ embed_bf16_matvec_fused_impl(y, x, W_bf16, bias, in_dim, out_dim);
+}
+
+/* Threaded matvec: split output rows across threads */
+typedef struct {
+ float *y;
+ const float *x;
+ const uint16_t *W_bf16;
+ const float *bias;
+ int in_dim;
+ int out_dim;
+} matvec_task_t;
+
+static void matvec_worker(int tid, int n_threads, void *arg) {
+ matvec_task_t *t = (matvec_task_t *)arg;
+ int chunk = (t->out_dim + n_threads - 1) / n_threads;
+ int start = tid * chunk;
+ int end = start + chunk;
+ if (end > t->out_dim)
+ end = t->out_dim;
+ if (start >= end)
+ return;
+
+ bf16_matvec_fused(t->y + start, t->x, t->W_bf16 + (size_t)start * t->in_dim,
+ t->bias ? t->bias + start : NULL, t->in_dim, end - start);
+}
+
+static void bf16_matvec_threaded(
+ float *y, const float *x, const uint16_t *W_bf16, const float *bias, int in_dim, int out_dim) {
+ if (tp.n_threads <= 1) {
+ bf16_matvec_fused(y, x, W_bf16, bias, in_dim, out_dim);
+ return;
+ }
+ matvec_task_t task = {y, x, W_bf16, bias, in_dim, out_dim};
+ parallel_for(matvec_worker, &task);
+}
+
+typedef struct {
+ float *y;
+ const float *x;
+ const uint16_t *W_bf16;
+ const float *bias;
+ int seq_len;
+ int in_dim;
+ int out_dim;
+} bf16_linear_rows_task_t;
+
+static void bf16_linear_rows_worker(int tid, int n_threads, void *arg) {
+ bf16_linear_rows_task_t *t = (bf16_linear_rows_task_t *)arg;
+ int chunk = (t->seq_len + n_threads - 1) / n_threads;
+ int start = tid * chunk;
+ int end = start + chunk;
+ if (end > t->seq_len)
+ end = t->seq_len;
+ if (start >= end)
+ return;
+
+ for (int s = start; s < end; s++) {
+ bf16_matvec_fused(t->y + (size_t)s * t->out_dim, t->x + (size_t)s * t->in_dim, t->W_bf16,
+ t->bias, t->in_dim, t->out_dim);
+ }
+}
+
+static void bf16_linear_rows(float *y,
+ const float *x,
+ const uint16_t *W_bf16,
+ const float *bias,
+ int seq_len,
+ int in_dim,
+ int out_dim) {
+ bf16_linear_rows_task_t task = {
+ .y = y,
+ .x = x,
+ .W_bf16 = W_bf16,
+ .bias = bias,
+ .seq_len = seq_len,
+ .in_dim = in_dim,
+ .out_dim = out_dim,
+ };
+
+ if (tp.n_threads > 1 && seq_len >= 2)
+ parallel_for(bf16_linear_rows_worker, &task);
+ else
+ bf16_linear_rows_worker(0, 1, &task);
+}
+
+typedef struct {
+ 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;
+ int total_dim;
+} pair_matvec_task_t;
+
+static void pair_matvec_range(const pair_matvec_task_t *t, int row, int start, int end) {
+ const float *x_row = t->x + (size_t)row * t->in_dim;
+ int a_end = t->a_dim;
+ int b_end = a_end + t->b_dim;
+
+ if (start < a_end) {
+ int s = start;
+ int e = end < a_end ? end : a_end;
+ if (s < e) {
+ bf16_matvec_fused(t->a + (size_t)row * t->a_dim + s, x_row,
+ t->Wa_bf16 + (size_t)s * t->in_dim, NULL, t->in_dim, e - s);
+ }
+ }
+
+ if (end > a_end && start < b_end) {
+ int s = start > a_end ? start - a_end : 0;
+ int e_abs = end < b_end ? end : b_end;
+ int e = e_abs - a_end;
+ if (s < e) {
+ bf16_matvec_fused(t->b + (size_t)row * t->b_dim + s, x_row,
+ t->Wb_bf16 + (size_t)s * t->in_dim, NULL, t->in_dim, e - s);
+ }
+ }
+}
+
+static void pair_matvec_worker(int tid, int n_threads, void *arg) {
+ pair_matvec_task_t *t = (pair_matvec_task_t *)arg;
+ int total_work = t->seq_len * t->total_dim;
+ int chunk = (total_work + n_threads - 1) / n_threads;
+ int start = tid * chunk;
+ int end = start + chunk;
+ if (end > total_work)
+ end = total_work;
+ if (start >= end)
+ return;
+
+ while (start < end) {
+ int row = start / t->total_dim;
+ int pos = start - row * t->total_dim;
+ int row_end = (row + 1) * t->total_dim;
+ int stop = end < row_end ? end : row_end;
+ pair_matvec_range(t, row, pos, stop - row * t->total_dim);
+ start = stop;
+ }
+}
+
+void embed_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) {
+ if (seq_len <= 0)
+ return;
+
+ if (tp.n_threads <= 1) {
+ for (int s = 0; s < seq_len; s++) {
+ const float *x_row = x + (size_t)s * in_dim;
+ bf16_matvec_fused(a + (size_t)s * a_dim, x_row, Wa_bf16, NULL, in_dim, a_dim);
+ bf16_matvec_fused(b + (size_t)s * b_dim, x_row, Wb_bf16, NULL, in_dim, b_dim);
+ }
+ return;
+ }
+
+ pair_matvec_task_t task = {
+ .a = a,
+ .b = b,
+ .x = x,
+ .Wa_bf16 = Wa_bf16,
+ .Wb_bf16 = Wb_bf16,
+ .seq_len = seq_len,
+ .in_dim = in_dim,
+ .a_dim = a_dim,
+ .b_dim = b_dim,
+ .total_dim = a_dim + b_dim,
+ };
+ parallel_for(pair_matvec_worker, &task);
+}
+
+typedef struct {
+ 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 in_dim;
+ int q_dim;
+ int kv_dim;
+ int seq_len;
+ int total_dim;
+} qkv_matvec_task_t;
+
+static void qkv_matvec_range(const qkv_matvec_task_t *t, int row, int start, int end) {
+ const float *x_row = t->x + (size_t)row * t->in_dim;
+ int q_end = t->q_dim;
+ int k_end = q_end + t->kv_dim;
+ int v_end = k_end + t->kv_dim;
+
+ if (start < q_end) {
+ int s = start;
+ int e = end < q_end ? end : q_end;
+ if (s < e) {
+ bf16_matvec_fused(t->q + (size_t)row * t->q_dim + s, x_row,
+ t->Wq_bf16 + (size_t)s * t->in_dim, NULL, t->in_dim, e - s);
+ }
+ }
+
+ if (end > q_end && start < k_end) {
+ int s = start > q_end ? start - q_end : 0;
+ int e_abs = end < k_end ? end : k_end;
+ int e = e_abs - q_end;
+ if (s < e) {
+ bf16_matvec_fused(t->k + (size_t)row * t->kv_dim + s, x_row,
+ t->Wk_bf16 + (size_t)s * t->in_dim, NULL, t->in_dim, e - s);
+ }
+ }
+
+ if (end > k_end && start < v_end) {
+ int s = start > k_end ? start - k_end : 0;
+ int e_abs = end < v_end ? end : v_end;
+ int e = e_abs - k_end;
+ if (s < e) {
+ bf16_matvec_fused(t->v + (size_t)row * t->kv_dim + s, x_row,
+ t->Wv_bf16 + (size_t)s * t->in_dim, NULL, t->in_dim, e - s);
+ }
+ }
+}
+
+static void qkv_matvec_worker(int tid, int n_threads, void *arg) {
+ qkv_matvec_task_t *t = (qkv_matvec_task_t *)arg;
+ int total_work = t->seq_len * t->total_dim;
+ int chunk = (total_work + n_threads - 1) / n_threads;
+ int start = tid * chunk;
+ int end = start + chunk;
+ if (end > total_work)
+ end = total_work;
+ if (start >= end)
+ return;
+
+ while (start < end) {
+ int row = start / t->total_dim;
+ int pos = start - row * t->total_dim;
+ int row_end = (row + 1) * t->total_dim;
+ int stop = end < row_end ? end : row_end;
+ qkv_matvec_range(t, row, pos, stop - row * t->total_dim);
+ start = stop;
+ }
+}
+
+void embed_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) {
+ if (seq_len <= 0)
+ return;
+
+ if (tp.n_threads <= 1) {
+ for (int s = 0; s < seq_len; s++) {
+ const float *x_row = x + (size_t)s * in_dim;
+ bf16_matvec_fused(q + (size_t)s * q_dim, x_row, Wq_bf16, NULL, in_dim, q_dim);
+ bf16_matvec_fused(k + (size_t)s * kv_dim, x_row, Wk_bf16, NULL, in_dim, kv_dim);
+ bf16_matvec_fused(v + (size_t)s * kv_dim, x_row, Wv_bf16, NULL, in_dim, kv_dim);
+ }
+ return;
+ }
+
+ qkv_matvec_task_t task = {
+ .q = q,
+ .k = k,
+ .v = v,
+ .x = x,
+ .Wq_bf16 = Wq_bf16,
+ .Wk_bf16 = Wk_bf16,
+ .Wv_bf16 = Wv_bf16,
+ .in_dim = in_dim,
+ .q_dim = q_dim,
+ .kv_dim = kv_dim,
+ .seq_len = seq_len,
+ .total_dim = q_dim + 2 * kv_dim,
+ };
+ parallel_for(qkv_matvec_worker, &task);
+}
+
+void embed_linear_nobias_bf16(
+ float *y, const float *x, const uint16_t *W_bf16, int seq_len, int in_dim, int out_dim) {
+ if (seq_len == 1) {
+ bf16_matvec_threaded(y, x, W_bf16, NULL, in_dim, out_dim);
+ return;
+ }
+ /* Callers route longer sequences through an F32-widened weight matrix in
+ * caller-owned scratch (see embed.c); this per-row path stays correct for
+ * any length. */
+ bf16_linear_rows(y, x, W_bf16, NULL, seq_len, in_dim, out_dim);
+}
+
+/* ========================================================================
+ * Normalization
+ * ======================================================================== */
+
+/* Rows split across the pool only when the tensor is big enough to pay
+ * for the dispatch. */
+#define EMBED_RMS_NORM_PARALLEL_ELEMS 262144
+
+static void embed_rms_norm_range(
+ float *out, const float *x, const float *weight, int start, int end, int hidden, float eps) {
+ for (int s = start; s < end; s++) {
+ const float *x_row = x + (size_t)s * hidden;
+ float *out_row = out + (size_t)s * hidden;
+
+#if defined(__AVX512F__) && defined(__FMA__)
+ __m512 accv = _mm512_setzero_ps();
+ int i = 0;
+ for (; i + 16 <= hidden; i += 16) {
+ __m512 v = _mm512_loadu_ps(x_row + i);
+ accv = _mm512_fmadd_ps(v, v, accv);
+ }
+ float sum_sq = _mm512_reduce_add_ps(accv);
+ for (; i < hidden; i++)
+ sum_sq += x_row[i] * x_row[i];
+#elif defined(__AVX2__) && defined(__FMA__)
+ __m256 accv = _mm256_setzero_ps();
+ int i = 0;
+ for (; i + 8 <= hidden; i += 8) {
+ __m256 v = _mm256_loadu_ps(x_row + i);
+ accv = _mm256_fmadd_ps(v, v, accv);
+ }
+ __m128 acc128 = _mm_add_ps(_mm256_castps256_ps128(accv), _mm256_extractf128_ps(accv, 1));
+ acc128 = _mm_hadd_ps(acc128, acc128);
+ acc128 = _mm_hadd_ps(acc128, acc128);
+ float sum_sq = _mm_cvtss_f32(acc128);
+ for (; i < hidden; i++)
+ sum_sq += x_row[i] * x_row[i];
+#else
+ float sum_sq = 0.0f;
+ for (int i = 0; i < hidden; i++) {
+ sum_sq += x_row[i] * x_row[i];
+ }
+#endif
+ float rms_inv = 1.0f / sqrtf(sum_sq / hidden + eps);
+
+#if defined(__AVX512F__)
+ __m512 scale = _mm512_set1_ps(rms_inv);
+ int j = 0;
+ for (; j + 16 <= hidden; j += 16) {
+ __m512 vx = _mm512_loadu_ps(x_row + j);
+ __m512 vw = _mm512_loadu_ps(weight + j);
+ _mm512_storeu_ps(out_row + j, _mm512_mul_ps(_mm512_mul_ps(vx, vw), scale));
+ }
+ for (; j < hidden; j++)
+ out_row[j] = x_row[j] * rms_inv * weight[j];
+#elif defined(__AVX2__)
+ __m256 scale = _mm256_set1_ps(rms_inv);
+ int j = 0;
+ for (; j + 8 <= hidden; j += 8) {
+ __m256 vx = _mm256_loadu_ps(x_row + j);
+ __m256 vw = _mm256_loadu_ps(weight + j);
+ _mm256_storeu_ps(out_row + j, _mm256_mul_ps(_mm256_mul_ps(vx, vw), scale));
+ }
+ for (; j < hidden; j++)
+ out_row[j] = x_row[j] * rms_inv * weight[j];
+#else
+ for (int i = 0; i < hidden; i++) {
+ out_row[i] = x_row[i] * rms_inv * weight[i];
+ }
+#endif
+ }
+}
+
+typedef struct {
+ float *out;
+ const float *x;
+ const float *weight;
+ int seq_len;
+ int hidden;
+ float eps;
+} rms_norm_task_t;
+
+static void rms_norm_worker(int tid, int n_threads, void *arg) {
+ rms_norm_task_t *t = (rms_norm_task_t *)arg;
+ int chunk = (t->seq_len + n_threads - 1) / n_threads;
+ int start = tid * chunk;
+ int end = start + chunk;
+ if (end > t->seq_len)
+ end = t->seq_len;
+ if (start >= end)
+ return;
+ embed_rms_norm_range(t->out, t->x, t->weight, start, end, t->hidden, t->eps);
+}
+
+void embed_rms_norm(
+ float *out, const float *x, const float *weight, int seq_len, int hidden, float eps) {
+ if (seq_len <= 0 || hidden <= 0)
+ return;
+
+ long long elems = (long long)seq_len * hidden;
+ if (tp.n_threads > 1 && elems >= EMBED_RMS_NORM_PARALLEL_ELEMS) {
+ rms_norm_task_t task = {
+ .out = out,
+ .x = x,
+ .weight = weight,
+ .seq_len = seq_len,
+ .hidden = hidden,
+ .eps = eps,
+ };
+ parallel_for(rms_norm_worker, &task);
+ } else {
+ embed_rms_norm_range(out, x, weight, 0, seq_len, hidden, eps);
+ }
+}
+
+void embed_rms_norm_per_head(
+ float *x, const float *weight, int seq_len, int n_heads, int head_dim, float eps) {
+ /* x is [seq, n_heads * head_dim] - normalize each [head_dim] segment */
+ int hidden = n_heads * head_dim;
+ for (int s = 0; s < seq_len; s++) {
+ for (int h = 0; h < n_heads; h++) {
+ float *vec = x + s * hidden + h * head_dim;
+
+#if defined(__AVX512F__) && defined(__FMA__)
+ __m512 accv = _mm512_setzero_ps();
+ int d = 0;
+ for (; d + 16 <= head_dim; d += 16) {
+ __m512 v = _mm512_loadu_ps(vec + d);
+ accv = _mm512_fmadd_ps(v, v, accv);
+ }
+ float sum_sq = _mm512_reduce_add_ps(accv);
+ for (; d < head_dim; d++)
+ sum_sq += vec[d] * vec[d];
+#elif defined(__AVX2__) && defined(__FMA__)
+ __m256 accv = _mm256_setzero_ps();
+ int d = 0;
+ for (; d + 8 <= head_dim; d += 8) {
+ __m256 v = _mm256_loadu_ps(vec + d);
+ accv = _mm256_fmadd_ps(v, v, accv);
+ }
+ __m128 acc128 =
+ _mm_add_ps(_mm256_castps256_ps128(accv), _mm256_extractf128_ps(accv, 1));
+ acc128 = _mm_hadd_ps(acc128, acc128);
+ acc128 = _mm_hadd_ps(acc128, acc128);
+ float sum_sq = _mm_cvtss_f32(acc128);
+ for (; d < head_dim; d++)
+ sum_sq += vec[d] * vec[d];
+#else
+ float sum_sq = 0.0f;
+ for (int d = 0; d < head_dim; d++) {
+ sum_sq += vec[d] * vec[d];
+ }
+#endif
+ float rms_inv = 1.0f / sqrtf(sum_sq / head_dim + eps);
+
+#if defined(__AVX512F__)
+ __m512 scale = _mm512_set1_ps(rms_inv);
+ int j = 0;
+ for (; j + 16 <= head_dim; j += 16) {
+ __m512 v = _mm512_loadu_ps(vec + j);
+ __m512 w = _mm512_loadu_ps(weight + j);
+ _mm512_storeu_ps(vec + j, _mm512_mul_ps(_mm512_mul_ps(v, w), scale));
+ }
+ for (; j < head_dim; j++)
+ vec[j] = vec[j] * rms_inv * weight[j];
+#elif defined(__AVX2__)
+ __m256 scale = _mm256_set1_ps(rms_inv);
+ int j = 0;
+ for (; j + 8 <= head_dim; j += 8) {
+ __m256 v = _mm256_loadu_ps(vec + j);
+ __m256 w = _mm256_loadu_ps(weight + j);
+ _mm256_storeu_ps(vec + j, _mm256_mul_ps(_mm256_mul_ps(v, w), scale));
+ }
+ for (; j < head_dim; j++)
+ vec[j] = vec[j] * rms_inv * weight[j];
+#else
+ for (int d = 0; d < head_dim; d++) {
+ vec[d] = vec[d] * rms_inv * weight[d];
+ }
+#endif
+ }
+ }
+}
+
+/* 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 embed_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
+ * ======================================================================== */
+
+void embed_silu_mul_inplace(float *gate, const float *up, int n) {
+#if defined(__APPLE__) && defined(USE_BLAS)
+ enum { CHUNK = 4096 };
+ float exp_neg[CHUNK];
+
+ for (int base = 0; base < n; base += CHUNK) {
+ int len = n - base;
+ if (len > CHUNK)
+ len = CHUNK;
+
+ for (int i = 0; i < len; i++)
+ exp_neg[i] = -gate[base + i];
+ vvexpf(exp_neg, exp_neg, &len);
+
+ for (int i = 0; i < len; i++) {
+ float g = gate[base + i];
+ gate[base + i] = (g / (1.0f + exp_neg[i])) * up[base + i];
+ }
+ }
+#else
+ for (int i = 0; i < n; i++) {
+ float g = gate[i];
+ gate[i] = (g / (1.0f + expf(-g))) * up[i];
+ }
+#endif
+}
+
+void embed_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 embed_gelu_tanh_inplace(float *x, int n) {
+ /* Tanh-approximation GeLU; 0.79788456... = sqrt(2/pi). */
+ for (int i = 0; i < n; i++) {
+ float v = x[i];
+ float inner = 0.79788456080286536f * (v + 0.044715f * v * v * v);
+ x[i] = 0.5f * v * (1.0f + tanhf(inner));
+ }
+}
+
+void embed_softmax(float *x, int rows, int cols) {
+ for (int r = 0; r < rows; r++) {
+ float *row = x + r * cols;
+ float max_val = row[0];
+ for (int c = 1; c < cols; c++) {
+ if (row[c] > max_val)
+ max_val = row[c];
+ }
+ float sum = 0.0f;
+ for (int c = 0; c < cols; c++) {
+ row[c] = expf(row[c] - max_val);
+ sum += row[c];
+ }
+ float inv_sum = 1.0f / sum;
+ for (int c = 0; c < cols; c++) {
+ row[c] *= inv_sum;
+ }
+ }
+}
+
+/* ========================================================================
+ * Attention Operations
+ * ======================================================================== */
+
+float embed_dot_f32(const float *a, const float *b, int n) { return embed_dot_f32_impl(a, b, n); }
+
+static inline float embed_dot_f32_fast(const float *a, const float *b, int n) {
+ return embed_dot_f32_impl(a, b, n);
+}
+
+/* dst = dst * scale */
+static inline void embed_vec_scale_inplace(float *dst, float scale, int n) {
+ embed_vec_scale_inplace_impl(dst, scale, n);
+}
+
+/* dst += alpha * src */
+static inline void embed_vec_axpy_inplace(float *dst, const float *src, float alpha, int n) {
+ embed_vec_axpy_inplace_impl(dst, src, alpha, n);
+}
+
+/* dst = dst * correction + src */
+static inline void embed_vec_scale_add(float *dst, const float *src, float correction, int n) {
+ embed_vec_scale_add_impl(dst, src, correction, n);
+}
+
+#define EMBED_PACKED_ATTN_QUERY_TILE 32
+#define EMBED_PACKED_ATTN_BLAS_MIN_SEQ 128
+
+static void embed_bidirectional_gqa_attention_packed_online_rows(float *out,
+ const float *Q,
+ const float *K,
+ const float *V,
+ int start,
+ int end,
+ int n_heads,
+ int n_kv_heads,
+ int head_dim,
+ float scale,
+ int h,
+ int query_start,
+ int query_end) {
+ int heads_per_kv = n_heads / n_kv_heads;
+ int q_hidden = n_heads * head_dim;
+ int kv_hidden = n_kv_heads * head_dim;
+ int kv_h = h / heads_per_kv;
+
+ for (int i = query_start; i < query_end; i++) {
+ const float *q_row = Q + (size_t)i * q_hidden + h * head_dim;
+ float *o_row = out + (size_t)i * q_hidden + h * head_dim;
+
+ float max_score = -1e30f;
+ float sum_exp = 0.0f;
+ for (int d = 0; d < head_dim; d++)
+ o_row[d] = 0.0f;
+
+ for (int j = start; j < end; j++) {
+ const float *k_row = K + (size_t)j * kv_hidden + kv_h * head_dim;
+ const float *v_row = V + (size_t)j * kv_hidden + kv_h * head_dim;
+
+ float score = embed_dot_f32_fast(q_row, k_row, head_dim) * scale;
+
+ if (score > max_score) {
+ float correction = expf(max_score - score);
+ sum_exp = sum_exp * correction + 1.0f;
+ embed_vec_scale_add(o_row, v_row, correction, head_dim);
+ max_score = score;
+ } else {
+ float wt = expf(score - max_score);
+ sum_exp += wt;
+ embed_vec_axpy_inplace(o_row, v_row, wt, head_dim);
+ }
+ }
+
+ if (sum_exp > 0.0f) {
+ float inv_sum = 1.0f / sum_exp;
+ embed_vec_scale_inplace(o_row, inv_sum, head_dim);
+ }
+ }
+}
+
+#ifdef USE_BLAS
+static void embed_bidirectional_gqa_attention_packed_blas_rows(float *out,
+ const float *Q,
+ const float *K,
+ const float *V,
+ int start,
+ int end,
+ int n_heads,
+ int n_kv_heads,
+ int head_dim,
+ float scale,
+ int h,
+ int query_start,
+ int query_end,
+ float *scores) {
+ int heads_per_kv = n_heads / n_kv_heads;
+ int q_hidden = n_heads * head_dim;
+ int kv_hidden = n_kv_heads * head_dim;
+ int kv_h = h / heads_per_kv;
+ int rows = query_end - query_start;
+ int keys = end - start;
+
+ const float *q = Q + (size_t)query_start * q_hidden + h * head_dim;
+ const float *k = K + (size_t)start * kv_hidden + kv_h * head_dim;
+ const float *v = V + (size_t)start * kv_hidden + kv_h * head_dim;
+ float *o = out + (size_t)query_start * q_hidden + h * head_dim;
+
+ cblas_sgemm(CblasRowMajor, CblasNoTrans, CblasTrans, rows, keys, head_dim, scale, q, q_hidden,
+ k, kv_hidden, 0.0f, scores, keys);
+ embed_softmax(scores, rows, keys);
+ cblas_sgemm(CblasRowMajor, CblasNoTrans, CblasNoTrans, rows, head_dim, keys, 1.0f, scores, keys,
+ v, kv_hidden, 0.0f, o, q_hidden);
+}
+#endif
+
+typedef struct {
+ float *out;
+ const float *Q;
+ const float *K;
+ const float *V;
+ const int *offsets;
+ int batch;
+ int n_heads, n_kv_heads;
+ int head_dim;
+ float scale;
+ float *scores;
+ int max_seq;
+} packed_gqa_attn_task_t;
+
+static void packed_gqa_attn_worker(int tid, int n_threads, void *arg) {
+ packed_gqa_attn_task_t *t = (packed_gqa_attn_task_t *)arg;
+ int task_id = 0;
+#ifdef USE_BLAS
+ float *scores =
+ t->scores ? t->scores + (size_t)tid * EMBED_PACKED_ATTN_QUERY_TILE * t->max_seq : NULL;
+#endif
+
+ for (int h = 0; h < t->n_heads; h++) {
+ for (int b = 0; b < t->batch; b++) {
+ int start = t->offsets[b];
+ int end = t->offsets[b + 1];
+ for (int q0 = start; q0 < end; q0 += EMBED_PACKED_ATTN_QUERY_TILE) {
+ int q1 = q0 + EMBED_PACKED_ATTN_QUERY_TILE;
+ if (q1 > end)
+ q1 = end;
+ if (task_id++ % n_threads != tid)
+ continue;
+#ifdef USE_BLAS
+ if (scores && end - start >= EMBED_PACKED_ATTN_BLAS_MIN_SEQ) {
+ embed_bidirectional_gqa_attention_packed_blas_rows(
+ t->out, t->Q, t->K, t->V, start, end, t->n_heads, t->n_kv_heads,
+ t->head_dim, t->scale, h, q0, q1, scores);
+ continue;
+ }
+#endif
+ embed_bidirectional_gqa_attention_packed_online_rows(
+ t->out, t->Q, t->K, t->V, start, end, t->n_heads, t->n_kv_heads, t->head_dim,
+ t->scale, h, q0, q1);
+ }
+ }
+ }
+}
+
+size_t embed_bidirectional_gqa_attention_packed_scratch_bytes(const int *offsets, int batch) {
+#ifdef USE_BLAS
+ if (!offsets || batch <= 0)
+ return 0;
+
+ int max_seq = 0;
+ for (int b = 0; b < batch; b++) {
+ int len = offsets[b + 1] - offsets[b];
+ if (len > max_seq)
+ max_seq = len;
+ }
+ if (max_seq < EMBED_PACKED_ATTN_BLAS_MIN_SEQ)
+ return 0;
+ if ((size_t)max_seq >
+ SIZE_MAX / (sizeof(float) * EMBED_PACKED_ATTN_QUERY_TILE * (size_t)tp.n_threads))
+ return 0;
+ return (size_t)tp.n_threads * EMBED_PACKED_ATTN_QUERY_TILE * (size_t)max_seq * sizeof(float);
+#else
+ (void)offsets;
+ (void)batch;
+ return 0;
+#endif
+}
+
+void embed_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) {
+ long long qk_work = 0;
+ int max_seq = 0;
+ for (int b = 0; b < batch; b++) {
+ int len = offsets[b + 1] - offsets[b];
+ qk_work += (long long)len * len * n_heads;
+ if (len > max_seq)
+ max_seq = len;
+ }
+
+ float *scores = NULL;
+#ifdef USE_BLAS
+ size_t required = embed_bidirectional_gqa_attention_packed_scratch_bytes(offsets, batch);
+ if (required != 0 && scratch && scratch_bytes >= required)
+ scores = scratch;
+#else
+ (void)scratch;
+ (void)scratch_bytes;
+#endif
+ packed_gqa_attn_task_t task = {.out = out,
+ .Q = Q,
+ .K = K,
+ .V = V,
+ .offsets = offsets,
+ .batch = batch,
+ .n_heads = n_heads,
+ .n_kv_heads = n_kv_heads,
+ .head_dim = head_dim,
+ .scale = scale,
+ .scores = scores,
+ .max_seq = max_seq};
+ if (tp.n_threads > 1 && n_heads >= 2 && qk_work >= 4096) {
+ parallel_for(packed_gqa_attn_worker, &task);
+ } else {
+ packed_gqa_attn_worker(0, 1, &task);
+ }
+}
+
+void embed_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) {
+ size_t scratch_bytes = embed_bidirectional_gqa_attention_packed_scratch_bytes(offsets, batch);
+ float *scores = scratch_bytes ? malloc(scratch_bytes) : NULL;
+ embed_bidirectional_gqa_attention_packed_with_scratch(
+ out, Q, K, V, offsets, batch, n_heads, n_kv_heads, head_dim, scale, scores, scratch_bytes);
+ free(scores);
+}
+
+static void embed_causal_gqa_attention_packed_online_rows(float *out,
+ const float *Q,
+ const float *K,
+ const float *V,
+ int start,
+ int n_heads,
+ int n_kv_heads,
+ int head_dim,
+ float scale,
+ int h,
+ int query_start,
+ int query_end) {
+ int heads_per_kv = n_heads / n_kv_heads;
+ int q_hidden = n_heads * head_dim;
+ int kv_hidden = n_kv_heads * head_dim;
+ int kv_h = h / heads_per_kv;
+
+ for (int i = query_start; i < query_end; i++) {
+ const float *q_row = Q + (size_t)i * q_hidden + h * head_dim;
+ float *o_row = out + (size_t)i * q_hidden + h * head_dim;
+
+ float max_score = -1e30f;
+ float sum_exp = 0.0f;
+ for (int d = 0; d < head_dim; d++)
+ o_row[d] = 0.0f;
+
+ for (int j = start; j <= i; j++) {
+ const float *k_row = K + (size_t)j * kv_hidden + kv_h * head_dim;
+ const float *v_row = V + (size_t)j * kv_hidden + kv_h * head_dim;
+ float score = embed_dot_f32_fast(q_row, k_row, head_dim) * scale;
+
+ if (score > max_score) {
+ float correction = expf(max_score - score);
+ sum_exp = sum_exp * correction + 1.0f;
+ embed_vec_scale_add(o_row, v_row, correction, head_dim);
+ max_score = score;
+ } else {
+ float wt = expf(score - max_score);
+ sum_exp += wt;
+ embed_vec_axpy_inplace(o_row, v_row, wt, head_dim);
+ }
+ }
+
+ if (sum_exp > 0.0f)
+ embed_vec_scale_inplace(o_row, 1.0f / sum_exp, head_dim);
+ }
+}
+
+#ifdef USE_BLAS
+static void embed_causal_gqa_attention_packed_blas_rows(float *out,
+ const float *Q,
+ const float *K,
+ const float *V,
+ int start,
+ int n_heads,
+ int n_kv_heads,
+ int head_dim,
+ float scale,
+ int h,
+ int query_start,
+ int query_end,
+ float *scores) {
+ int heads_per_kv = n_heads / n_kv_heads;
+ int q_hidden = n_heads * head_dim;
+ int kv_hidden = n_kv_heads * head_dim;
+ int kv_h = h / heads_per_kv;
+ int rows = query_end - query_start;
+ int keys = query_end - start;
+
+ const float *q = Q + (size_t)query_start * q_hidden + h * head_dim;
+ const float *k = K + (size_t)start * kv_hidden + kv_h * head_dim;
+ const float *v = V + (size_t)start * kv_hidden + kv_h * head_dim;
+ float *o = out + (size_t)query_start * q_hidden + h * head_dim;
+
+ cblas_sgemm(CblasRowMajor, CblasNoTrans, CblasTrans, rows, keys, head_dim, scale, q, q_hidden,
+ k, kv_hidden, 0.0f, scores, keys);
+ for (int r = 0; r < rows; r++) {
+ int allowed = query_start + r - start + 1;
+ float *row = scores + (size_t)r * keys;
+ for (int j = allowed; j < keys; j++)
+ row[j] = -INFINITY;
+ }
+ embed_softmax(scores, rows, keys);
+ cblas_sgemm(CblasRowMajor, CblasNoTrans, CblasNoTrans, rows, head_dim, keys, 1.0f, scores, keys,
+ v, kv_hidden, 0.0f, o, q_hidden);
+}
+#endif
+
+static void packed_causal_gqa_attn_worker(int tid, int n_threads, void *arg) {
+ packed_gqa_attn_task_t *t = (packed_gqa_attn_task_t *)arg;
+ int task_id = 0;
+#ifdef USE_BLAS
+ float *scores =
+ t->scores ? t->scores + (size_t)tid * EMBED_PACKED_ATTN_QUERY_TILE * t->max_seq : NULL;
+#endif
+
+ for (int h = 0; h < t->n_heads; h++) {
+ for (int b = 0; b < t->batch; b++) {
+ int start = t->offsets[b];
+ int end = t->offsets[b + 1];
+ for (int q0 = start; q0 < end; q0 += EMBED_PACKED_ATTN_QUERY_TILE) {
+ int q1 = q0 + EMBED_PACKED_ATTN_QUERY_TILE;
+ if (q1 > end)
+ q1 = end;
+ if (task_id++ % n_threads != tid)
+ continue;
+#ifdef USE_BLAS
+ if (scores && end - start >= EMBED_PACKED_ATTN_BLAS_MIN_SEQ) {
+ embed_causal_gqa_attention_packed_blas_rows(
+ t->out, t->Q, t->K, t->V, start, t->n_heads, t->n_kv_heads, t->head_dim,
+ t->scale, h, q0, q1, scores);
+ continue;
+ }
+#endif
+ embed_causal_gqa_attention_packed_online_rows(t->out, t->Q, t->K, t->V, start,
+ t->n_heads, t->n_kv_heads, t->head_dim,
+ t->scale, h, q0, q1);
+ }
+ }
+ }
+}
+
+void embed_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) {
+ long long qk_work = 0;
+ int max_seq = 0;
+ for (int b = 0; b < batch; b++) {
+ int len = offsets[b + 1] - offsets[b];
+ qk_work += (long long)len * (len + 1) / 2 * n_heads;
+ if (len > max_seq)
+ max_seq = len;
+ }
+
+ float *scores = NULL;
+#ifdef USE_BLAS
+ size_t required = embed_bidirectional_gqa_attention_packed_scratch_bytes(offsets, batch);
+ if (required != 0 && scratch && scratch_bytes >= required)
+ scores = scratch;
+#else
+ (void)scratch;
+ (void)scratch_bytes;
+#endif
+ packed_gqa_attn_task_t task = {.out = out,
+ .Q = Q,
+ .K = K,
+ .V = V,
+ .offsets = offsets,
+ .batch = batch,
+ .n_heads = n_heads,
+ .n_kv_heads = n_kv_heads,
+ .head_dim = head_dim,
+ .scale = scale,
+ .scores = scores,
+ .max_seq = max_seq};
+ if (tp.n_threads > 1 && n_heads >= 2 && qk_work >= 4096)
+ parallel_for(packed_causal_gqa_attn_worker, &task);
+ else
+ packed_causal_gqa_attn_worker(0, 1, &task);
+}
+
+void embed_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) {
+ size_t scratch_bytes = embed_bidirectional_gqa_attention_packed_scratch_bytes(offsets, batch);
+ float *scores = scratch_bytes ? malloc(scratch_bytes) : NULL;
+ embed_causal_gqa_attention_packed_with_scratch(out, Q, K, V, offsets, batch, n_heads, n_kv_heads,
+ head_dim, scale, scores, scratch_bytes);
+ free(scores);
+}
+
+/* ========================================================================
+ * Position Embeddings
+ * ======================================================================== */
+
+void embed_compute_rope_neox(
+ float *cos_out, float *sin_out, const int *positions, int seq, int head_dim, float theta) {
+ int half = head_dim / 2;
+
+ for (int s = 0; s < seq; s++) {
+ float pos = (float)positions[s];
+ for (int d = 0; d < half; d++) {
+ float freq = 1.0f / powf(theta, (float)(2 * d) / (float)head_dim);
+ float angle = pos * freq;
+ float c = cosf(angle);
+ float sn = sinf(angle);
+ /* Duplicate for full head_dim */
+ cos_out[s * head_dim + d] = c;
+ cos_out[s * head_dim + half + d] = c;
+ sin_out[s * head_dim + d] = sn;
+ sin_out[s * head_dim + half + d] = sn;
+ }
+ }
+}
+
+void embed_apply_rope_neox(
+ float *x, const float *cos_vals, const float *sin_vals, int seq, int n_heads, int head_dim) {
+ /*
+ * NeoX split-half style:
+ * x1 = x[..., :half], x2 = x[..., half:]
+ * rotated = cat(-x2, x1)
+ * result = x * cos + rotated * sin
+ */
+ int half = head_dim / 2;
+ int hidden = n_heads * head_dim;
+
+ for (int s = 0; s < seq; s++) {
+ const float *c = cos_vals + s * head_dim;
+ const float *sn = sin_vals + s * head_dim;
+
+ for (int h = 0; h < n_heads; h++) {
+ float *vec = x + s * hidden + h * head_dim;
+
+#if defined(__AVX512F__) && defined(__FMA__)
+ int d = 0;
+ for (; d + 16 <= half; d += 16) {
+ __m512 x1 = _mm512_loadu_ps(vec + d);
+ __m512 x2 = _mm512_loadu_ps(vec + half + d);
+ /* RoPE cache duplicates cos/sin across halves. */
+ __m512 cc = _mm512_loadu_ps(c + d);
+ __m512 ss = _mm512_loadu_ps(sn + d);
+ __m512 new1 = _mm512_fmsub_ps(x1, cc, _mm512_mul_ps(x2, ss));
+ __m512 new2 = _mm512_fmadd_ps(x2, cc, _mm512_mul_ps(x1, ss));
+ _mm512_storeu_ps(vec + d, new1);
+ _mm512_storeu_ps(vec + half + d, new2);
+ }
+ for (; d < half; d++) {
+ float x1 = vec[d];
+ float x2 = vec[half + d];
+ vec[d] = x1 * c[d] + (-x2) * sn[d];
+ vec[half + d] = x2 * c[half + d] + x1 * sn[half + d];
+ }
+#elif defined(__AVX2__) && defined(__FMA__)
+ int d = 0;
+ for (; d + 8 <= half; d += 8) {
+ __m256 x1 = _mm256_loadu_ps(vec + d);
+ __m256 x2 = _mm256_loadu_ps(vec + half + d);
+ __m256 cc = _mm256_loadu_ps(c + d);
+ __m256 ss = _mm256_loadu_ps(sn + d);
+ __m256 new1 = _mm256_fmsub_ps(x1, cc, _mm256_mul_ps(x2, ss));
+ __m256 new2 = _mm256_fmadd_ps(x2, cc, _mm256_mul_ps(x1, ss));
+ _mm256_storeu_ps(vec + d, new1);
+ _mm256_storeu_ps(vec + half + d, new2);
+ }
+ for (; d < half; d++) {
+ float x1 = vec[d];
+ float x2 = vec[half + d];
+ vec[d] = x1 * c[d] + (-x2) * sn[d];
+ vec[half + d] = x2 * c[half + d] + x1 * sn[half + d];
+ }
+#elif defined(__ARM_NEON) || defined(__ARM_NEON__)
+ int d = 0;
+ for (; d + 4 <= half; d += 4) {
+ float32x4_t x1 = vld1q_f32(vec + d);
+ float32x4_t x2 = vld1q_f32(vec + half + d);
+ float32x4_t cc = vld1q_f32(c + d);
+ float32x4_t ss = vld1q_f32(sn + d);
+ float32x4_t new1 = vmlsq_f32(vmulq_f32(x1, cc), x2, ss);
+ float32x4_t new2 = vfmaq_f32(vmulq_f32(x2, cc), x1, ss);
+ vst1q_f32(vec + d, new1);
+ vst1q_f32(vec + half + d, new2);
+ }
+ for (; d < half; d++) {
+ float x1 = vec[d];
+ float x2 = vec[half + d];
+ vec[d] = x1 * c[d] + (-x2) * sn[d];
+ vec[half + d] = x2 * c[half + d] + x1 * sn[half + d];
+ }
+#else
+ for (int d = 0; d < half; d++) {
+ float x1 = vec[d]; /* first half */
+ float x2 = vec[half + d]; /* second half */
+ vec[d] = x1 * c[d] + (-x2) * sn[d];
+ vec[half + d] = x2 * c[half + d] + x1 * sn[half + d];
+ }
+#endif
+ }
+ }
+}