local-llm-rn 1.0.0

package/cpp/ggml/src/ggml-hexagon/htp/unary-ops.c

@@ -0,0 +1,384 @@
+#pragma clang diagnostic ignored "-Wunused-variable"
+#pragma clang diagnostic ignored "-Wunused-function"
+#pragma clang diagnostic ignored "-Wunused-but-set-variable"
+
+#include <HAP_farf.h>
+#include <HAP_perf.h>
+
+#include <math.h>
+#include <string.h>
+
+#include "hex-dma.h"
+#include "hvx-utils.h"
+
+#define GGML_COMMON_DECL_C
+#include "ggml-common.h"
+#include "htp-ctx.h"
+#include "htp-msg.h"
+#include "htp-ops.h"
+
+struct htp_unary_context {
+    struct htp_ops_context * octx;
+
+    // Precomputed values
+    const uint8_t *           data_src0;
+    uint8_t *                 data_dst;
+
+    size_t                    src0_row_size;
+    size_t                    dst_row_size;
+
+    size_t                    src0_row_size_aligned;
+    size_t                    dst_row_size_aligned;
+
+    size_t                    src0_spad_half_size;
+    size_t                    dst_spad_half_size;
+
+    uint32_t                  block;
+    uint32_t                  src0_nrows;
+    uint32_t                  src0_nrows_per_thread;
+    uint32_t                  nc;
+};
+
+#define htp_unary_preamble            \
+    const uint32_t ne00 = src->ne[0]; \
+    const uint32_t ne01 = src->ne[1]; \
+    const uint32_t ne02 = src->ne[2]; \
+    const uint32_t ne03 = src->ne[3]; \
+                                      \
+    const uint32_t ne0 = dst->ne[0];  \
+    const uint32_t ne1 = dst->ne[1];  \
+    const uint32_t ne2 = dst->ne[2];  \
+    const uint32_t ne3 = dst->ne[3];  \
+                                      \
+    const uint32_t nb00 = src->nb[0]; \
+    const uint32_t nb01 = src->nb[1]; \
+    const uint32_t nb02 = src->nb[2]; \
+    const uint32_t nb03 = src->nb[3]; \
+                                      \
+    const uint32_t nb0 = dst->nb[0];  \
+    const uint32_t nb1 = dst->nb[1];  \
+    const uint32_t nb2 = dst->nb[2];  \
+    const uint32_t nb3 = dst->nb[3];
+
+static void hvx_fast_rms_norm_f32(const uint8_t * restrict src,
+                                  uint8_t * restrict dst,
+                                  uint8_t * restrict pad,
+                                  const int num_elems,
+                                  float     epsilon) {
+    const HVX_Vector * restrict v_src = (HVX_Vector *) src;
+    HVX_Vector * restrict v_dst       = (HVX_Vector *) dst;
+
+    HVX_Vector sum_v     = Q6_V_vsplat_R(0x00000000);
+    HVX_Vector epsilon_v = hvx_vec_splat_f32(epsilon);
+
+    int step_of_1 = num_elems >> 5;
+    #pragma unroll(4)
+    for (int i = 0; i < step_of_1; i++) {
+        HVX_Vector v1 = v_src[i];
+        HVX_Vector v2 = Q6_Vqf32_vmpy_VsfVsf(v1, v1);
+        sum_v         = Q6_Vqf32_vadd_Vqf32Vqf32(sum_v, v2);
+    }
+
+    sum_v = hvx_vec_reduce_sum_f32(Q6_Vsf_equals_Vqf32(sum_v)); // replicated over all lanes
+
+    HVX_Vector t_v            = hvx_vec_splat_f32((float) num_elems);
+    HVX_Vector denom_v        = hvx_vec_inverse_f32(t_v);
+    HVX_Vector mean_v         = Q6_Vqf32_vmpy_VsfVsf(sum_v, denom_v);
+    HVX_Vector mean_epsilon_v = Q6_Vqf32_vadd_Vqf32Vsf(mean_v, epsilon_v);
+
+    HVX_Vector scale_v = hvx_vec_rsqrt_f32(Q6_Vsf_equals_Vqf32(mean_epsilon_v));
+
+    #pragma unroll(4)
+    for (int i = 0; i < step_of_1; i++) {
+        HVX_Vector v1 = v_src[i];
+        HVX_Vector v2 = Q6_Vqf32_vmpy_VsfVsf(v1, scale_v);
+        v_dst[i]      = Q6_Vsf_equals_Vqf32(v2);
+    }
+}
+
+static void scale_f32(const float * restrict src,
+                      float * restrict dst,
+                      uint8_t * restrict spad,
+                      const uint32_t num_rows,
+                      const uint32_t row_elems,
+                      const size_t   row_size,
+                      int32_t *      op_params) {
+    float scale = 0.f;
+    float bias  = 0.f;
+    memcpy(&scale, &op_params[0], sizeof(float));
+    memcpy(&bias,  &op_params[1], sizeof(float));
+
+    for (uint32_t ir = 0; ir < num_rows; ir++) {
+        const uint8_t * restrict src_local = (const uint8_t *)src + (ir * row_size);
+        uint8_t * restrict dst_local       = (uint8_t *)dst + (ir * row_size);
+
+        hvx_scale_offset_f32_aa((uint8_t *) dst_local, (const uint8_t *) src_local, row_elems, scale, bias);
+    }
+}
+
+static void rms_norm_f32(const float * restrict src,
+                         float * restrict dst,
+                         uint8_t * restrict spad,
+                         const uint32_t num_rows,
+                         const uint32_t row_elems,
+                         const size_t   row_size,
+                         int32_t *      op_params) {
+    float epsilon = 0.f;
+    memcpy(&epsilon, op_params, sizeof(float));
+
+    for (uint32_t ir = 0; ir < num_rows; ir++) {
+        const uint8_t * restrict src_local = (const uint8_t *)src + (ir * row_size);
+        uint8_t * restrict dst_local       = (uint8_t *)dst + (ir * row_size);
+
+        hvx_fast_rms_norm_f32((const uint8_t *) src_local, (uint8_t *) dst_local, spad, row_elems, epsilon);
+    }
+}
+
+static void sqr_f32(const float * restrict src,
+                    float * restrict dst,
+                    uint8_t * restrict spad,
+                    const uint32_t num_rows,
+                    const uint32_t row_elems,
+                    const size_t   row_size,
+                    int32_t *      op_params) {
+
+    for (uint32_t ir = 0; ir < num_rows; ir++) {
+        const uint8_t * restrict src_local = (const uint8_t *)src + (ir * row_size);
+        uint8_t * restrict dst_local       = (uint8_t *)dst + (ir * row_size);
+
+        hvx_sqr_f32_aa((uint8_t *) dst_local, (const uint8_t *) src_local, row_elems);
+    }
+}
+
+static void sqrt_f32(const float * restrict src,
+                     float * restrict dst,
+                     uint8_t * restrict spad,
+                     const uint32_t num_rows,
+                     const uint32_t row_elems,
+                     const size_t   row_size,
+                     int32_t *      op_params) {
+
+    for (uint32_t ir = 0; ir < num_rows; ir++) {
+        const uint8_t * restrict src_local = (const uint8_t *)src + (ir * row_size);
+        uint8_t * restrict dst_local       = (uint8_t *)dst + (ir * row_size);
+
+        hvx_sqrt_f32_aa((uint8_t *) dst_local, (const uint8_t *) src_local, row_elems);
+    }
+}
+
+static void unary_job_f32_per_thread(unsigned int nth, unsigned int ith, void * data) {
+    const struct htp_unary_context * uctx = (const struct htp_unary_context *) data;
+    struct htp_ops_context * octx = uctx->octx;
+    const struct htp_tensor * src = &octx->src0;
+    const struct htp_tensor * dst = &octx->dst;
+
+    htp_unary_preamble;
+
+    int                       htp_op = octx->op;
+    int32_t *                 op_params = octx->op_params;
+    uint32_t                  src0_nrows_per_thread = uctx->src0_nrows_per_thread;
+
+    const size_t src0_row_size = uctx->src0_row_size;
+    const size_t dst_row_size  = uctx->dst_row_size;
+
+    const size_t src0_row_size_aligned = uctx->src0_row_size_aligned;
+    const size_t dst_row_size_aligned  = uctx->dst_row_size_aligned;
+
+    const uint32_t src0_nrows = uctx->src0_nrows;
+    const uint32_t src0_start_row = src0_nrows_per_thread * ith;
+    const uint32_t src0_end_row   = MIN(src0_start_row + src0_nrows_per_thread, src0_nrows);
+
+    // no work for this thread
+    if (src0_start_row >= src0_end_row) {
+        return;
+    }
+
+    uint64_t t1, t2;
+    t1 = HAP_perf_get_qtimer_count();
+
+    const uint8_t * restrict data_src = uctx->data_src0;
+    uint8_t * restrict       data_dst = uctx->data_dst;
+
+    uint8_t * src0_spad_data = octx->src0_spad.data + (ith * octx->src0_spad.size_per_thread);
+    uint8_t * dst_spad_data  = octx->dst_spad.data  + (ith * octx->dst_spad.size_per_thread);
+
+    size_t src0_spad_half_size = uctx->src0_spad_half_size;
+    size_t dst_spad_half_size  = uctx->dst_spad_half_size;
+
+    const int BLOCK = uctx->block;
+    if (BLOCK == 0) {
+        FARF(ERROR, "unary-f32 : current VTCM reservation %zu is too small for even 1 row per thread, needed at least %zu\n",
+             octx->src0_spad.size_per_thread, src0_row_size_aligned);
+        return;
+    }
+
+    dma_queue * dma_queue = octx->ctx->dma[ith];
+
+    for (uint32_t ir = src0_start_row, spad_idx = 0; ir < src0_end_row && spad_idx < 2; ir += BLOCK, spad_idx++) {
+        const uint32_t block_size = MIN(BLOCK, src0_end_row - ir);
+
+        // Dummy DMA transation for sequencing (interleaving dst,src,dst,...)
+        dma_queue_push_vtcm_to_ddr(dma_queue,
+            dma_make_ptr(data_dst, dst_spad_data + (spad_idx * dst_spad_half_size)),
+            dst_row_size, dst_row_size_aligned, 0);
+
+        dma_queue_push_ddr_to_vtcm(dma_queue,
+            dma_make_ptr(src0_spad_data + (spad_idx * src0_spad_half_size), data_src + (ir * src0_row_size)),
+            src0_row_size_aligned, src0_row_size, block_size);
+    }
+
+    for (uint32_t ir = src0_start_row; ir < src0_end_row; ir += BLOCK) {
+        const uint32_t block_size = MIN(BLOCK, src0_end_row - ir);
+
+        float * dst_spad  = (float *) dma_queue_pop(dma_queue).src;
+        float * src0_spad = (float *) dma_queue_pop(dma_queue).dst;
+
+        // Process block in VTCM
+        switch (htp_op) {
+            case HTP_OP_RMS_NORM:
+                rms_norm_f32(src0_spad, dst_spad, NULL, block_size, ne0, src0_row_size_aligned, op_params);
+                break;
+            case HTP_OP_SCALE:
+                scale_f32(src0_spad, dst_spad, NULL, block_size, ne0, src0_row_size_aligned, op_params);
+                break;
+            case HTP_OP_SQR:
+                sqr_f32(src0_spad, dst_spad, NULL, block_size, ne0, src0_row_size_aligned, op_params);
+                break;
+            case HTP_OP_SQRT:
+                sqrt_f32(src0_spad, dst_spad, NULL, block_size, ne0, src0_row_size_aligned, op_params);
+                break;
+            default:
+                break;
+        }
+
+        dma_queue_push_vtcm_to_ddr(dma_queue,
+            dma_make_ptr(data_dst + (ir * dst_row_size), dst_spad),
+            dst_row_size, dst_row_size_aligned, block_size);
+
+        // prefetch N+2 loop iteration if any
+        const uint32_t pref_block = (ir + BLOCK * 2);
+        if (pref_block < src0_end_row) {
+            const uint32_t pref_block_size = MIN(BLOCK, src0_end_row - pref_block);
+            dma_queue_push_ddr_to_vtcm(dma_queue,
+                dma_make_ptr(src0_spad, data_src + (pref_block * src0_row_size)),
+                src0_row_size_aligned, src0_row_size, pref_block_size);
+        }
+    }
+
+    dma_queue_flush(dma_queue);
+
+    t2 = HAP_perf_get_qtimer_count();
+
+    FARF(HIGH, "unary-f32 %d/%d: %ux%ux%ux%u (%u:%u) -> %ux%ux%ux%u usec %u\n", ith, nth, src->ne[0],
+         src->ne[1], src->ne[2], src->ne[3], src0_start_row, src0_end_row, dst->ne[0], dst->ne[1], dst->ne[2],
+         dst->ne[3], (unsigned) HAP_perf_qtimer_count_to_us(t2 - t1));
+}
+
+static int execute_op_unary_f32(struct htp_ops_context * octx) {
+    int err = HTP_STATUS_OK;
+
+    const struct htp_tensor * src0 = &octx->src0;
+    struct htp_tensor *       dst  = &octx->dst;
+
+    const char * op_type = NULL;
+
+    switch (octx->op) {
+        case HTP_OP_RMS_NORM:
+            op_type = "rmsnorm-f32";
+            break;
+        case HTP_OP_SCALE:
+            op_type = "scale-f32";
+            break;
+        case HTP_OP_SQR:
+            op_type = "sqr-f32";
+            break;
+        case HTP_OP_SQRT:
+            op_type = "sqrt-f32";
+            break;
+
+        default:
+            FARF(ERROR, "Unsupported unary Op %u\n", octx->op);
+            return HTP_STATUS_NO_SUPPORT;
+    }
+
+    const int      n_threads  = octx->n_threads;
+    const uint32_t src0_nrows = src0->ne[1] * src0->ne[2] * src0->ne[3];
+
+    const size_t src0_row_size = src0->nb[1];
+    const size_t dst_row_size  = dst->nb[1];
+
+    const size_t src0_row_size_aligned = hex_round_up(src0_row_size, VLEN);
+    const size_t dst_row_size_aligned  = hex_round_up(dst_row_size, VLEN);
+
+    // VTCM scratchpads for all tensors
+    // N rows per thread, padded to HVX vector size
+    // Double buffering requires 2x size per buffer
+
+    size_t spad_size_per_row   = 2 * (src0_row_size_aligned + dst_row_size_aligned);
+    size_t vtcm_row_per_thread = (octx->ctx->vtcm_size)/ (n_threads * spad_size_per_row);
+
+    // Make sure the reserved vtcm size is sufficient
+    if (vtcm_row_per_thread == 0) {
+        FARF(ERROR, "unary-%s : current VTCM reservation %zu is too small, needed %zu\n", op_type, octx->ctx->vtcm_size,
+             spad_size_per_row * n_threads);
+        return HTP_STATUS_VTCM_TOO_SMALL;
+    }
+
+    octx->src0_spad.size_per_thread = src0_row_size_aligned * vtcm_row_per_thread * 2;
+    octx->dst_spad.size_per_thread  = dst_row_size_aligned * vtcm_row_per_thread * 2;
+
+    octx->src0_spad.size = n_threads * octx->src0_spad.size_per_thread;
+    octx->dst_spad.size  = n_threads * octx->dst_spad.size_per_thread;
+
+    octx->src0_spad.data = octx->ctx->vtcm_base;
+    octx->dst_spad.data  = octx->src0_spad.data + octx->src0_spad.size;
+
+    FARF(HIGH, "%s: (%ux%ux%ux%u) -> (%ux%ux%ux%u) : src0-spad-size %u src1-spad-size %u dst-spad-size %u\n", op_type,
+         src0->ne[0], src0->ne[1], src0->ne[2], src0->ne[3], dst->ne[0], dst->ne[1], dst->ne[2], dst->ne[3],
+         octx->src0_spad.size, octx->src1_spad.size, octx->dst_spad.size);
+
+    if (!(octx->flags & HTP_OPFLAGS_SKIP_COMPUTE)) {
+        uint32_t n_jobs = MIN(n_threads, src0_nrows);
+
+        struct htp_unary_context uctx = {
+            .octx                  = octx,
+            .src0_nrows_per_thread = (src0_nrows + n_jobs - 1) / n_jobs,
+            .src0_nrows            = src0_nrows,
+
+            .data_src0             = (const uint8_t *)src0->data,
+            .data_dst              = (uint8_t *)dst->data,
+
+            .src0_row_size         = src0_row_size,
+            .dst_row_size          = dst_row_size,
+
+            .src0_row_size_aligned = src0_row_size_aligned,
+            .dst_row_size_aligned  = dst_row_size_aligned,
+
+            .src0_spad_half_size   = octx->src0_spad.size_per_thread / 2,
+            .dst_spad_half_size    = octx->dst_spad.size_per_thread / 2,
+
+            .block                 = (octx->src0_spad.size_per_thread / 2) / src0_row_size_aligned,
+            .nc                    = src0->ne[0],
+        };
+
+        worker_pool_run_func(octx->ctx->worker_pool, unary_job_f32_per_thread, &uctx, n_jobs);
+    }
+
+    return err;
+}
+
+int op_unary(struct htp_ops_context * octx) {
+    int err = HTP_STATUS_OK;
+
+    switch (octx->src0.type) {
+        case HTP_TYPE_F32:
+            err = execute_op_unary_f32(octx);
+            break;
+
+        default:
+            err = HTP_STATUS_NO_SUPPORT;
+            break;
+    }
+
+    return err;
+}

package/cpp/ggml/src/ggml-hexagon/htp/worker-pool.c

@@ -0,0 +1,293 @@
+#include "worker-pool.h"
+
+#include <qurt.h>
+#include <stdatomic.h>
+#include <stdint.h>
+#include <stdio.h>
+#include <stdlib.h>
+#include <string.h>
+
+#include "HAP_farf.h"
+
+#define WORKER_THREAD_STACK_SZ  (2 * 16384)
+#define LOWEST_USABLE_QURT_PRIO (254)
+
+struct worker_pool_s;
+
+// internal structure kept in thread-local storage per instance of worker pool
+typedef struct {
+    struct worker_pool_s * pool;
+    unsigned int           id;
+} worker_context_t;
+
+// internal structure kept in thread-local storage per instance of worker pool
+typedef struct worker_pool_s {
+    worker_pool_job_t job[MAX_NUM_WORKERS];      // list of job descriptors
+    qurt_thread_t     thread[MAX_NUM_WORKERS];   // thread ID's of the workers
+    worker_context_t  context[MAX_NUM_WORKERS];  // worker contexts
+    void *            stack[MAX_NUM_WORKERS];    // thread stack pointers
+    unsigned int      n_threads;                 // number of workers in this pool
+
+    atomic_uint seqn;                            // seqno used to detect new jobs
+    atomic_uint next_job;                        // next job index
+    atomic_uint n_pending;                       // number of pending jobs
+    atomic_uint n_jobs;                          // number of current jobs
+    atomic_bool killed;                          // threads need to exit
+} worker_pool_t;
+
+static void worker_pool_main(void * context) {
+    worker_context_t * me   = (worker_context_t *) context;
+    worker_pool_t *    pool = me->pool;
+
+    FARF(HIGH, "worker-pool: thread %u started", me->id);
+
+    unsigned int prev_seqn = 0;
+    while (!atomic_load(&pool->killed)) {
+        unsigned int seqn = atomic_load(&pool->seqn);
+        if (seqn == prev_seqn) {
+            // Nothing to do
+            qurt_futex_wait(&pool->seqn, prev_seqn);
+            continue;
+        }
+
+        // New job
+        prev_seqn = seqn;
+
+        unsigned int n = atomic_load(&pool->n_jobs);
+        unsigned int i = atomic_fetch_add(&pool->next_job, 1);
+        if (i >= n) {
+            // Spurios wakeup
+            continue;
+        }
+
+        pool->job[i].func(n, i, pool->job[i].data);
+
+        atomic_fetch_sub(&pool->n_pending, 1);
+    }
+
+    FARF(HIGH, "worker-pool: thread %u stopped", me->id);
+}
+
+AEEResult worker_pool_init_with_stack_size(worker_pool_context_t * context, uint32_t n_threads, uint32_t stack_size) {
+    int err = 0;
+
+    if (NULL == context) {
+        FARF(ERROR, "NULL context passed to worker_pool_init().");
+        return AEE_EBADPARM;
+    }
+
+    // Allocations
+    int size = (stack_size * n_threads) + (sizeof(worker_pool_t));
+
+    unsigned char * mem_blob = (unsigned char *) malloc(size);
+    if (!mem_blob) {
+        FARF(ERROR, "Could not allocate memory for worker pool!!");
+        return AEE_ENOMEMORY;
+    }
+
+    worker_pool_t * me = (worker_pool_t *) (mem_blob + stack_size * n_threads);
+
+    // name for the first worker, useful in debugging threads
+    char name[19];
+    snprintf(name, 12, "0x%8x:", (int) me);
+    strcat(name, "worker0");
+    me->n_threads = n_threads;
+
+    // initializations
+    for (unsigned int i = 0; i < me->n_threads; i++) {
+        me->stack[i]  = NULL;
+        me->thread[i] = 0;
+
+        me->context[i].id   = i;
+        me->context[i].pool = me;
+    }
+
+    // initialize job queue
+    me->n_pending = 0;
+    me->n_jobs    = 0;
+    me->next_job  = 0;
+    me->seqn      = 0;
+    me->killed    = 0;
+
+    // launch the workers
+    qurt_thread_attr_t attr;
+    qurt_thread_attr_init(&attr);
+
+    for (unsigned int i = 0; i < me->n_threads; i++) {
+        // set up stack
+        me->stack[i] = mem_blob;
+        mem_blob += stack_size;
+        qurt_thread_attr_set_stack_addr(&attr, me->stack[i]);
+        qurt_thread_attr_set_stack_size(&attr, stack_size);
+
+        // set up name
+        qurt_thread_attr_set_name(&attr, name);
+        name[17] = (name[17] + 1);
+        // name threads context:worker0, context:worker1, .. (recycle at 9, but num threads should be less than that anyway)
+        if (name[17] > '9') {
+            name[17] = '0';
+        }
+
+        // set up priority - by default, match the creating thread's prio
+        int prio = qurt_thread_get_priority(qurt_thread_get_id());
+
+        if (prio < 1) {
+            prio = 1;
+        }
+        if (prio > LOWEST_USABLE_QURT_PRIO) {
+            prio = LOWEST_USABLE_QURT_PRIO;
+        }
+
+        qurt_thread_attr_set_priority(&attr, prio);
+
+        // launch
+        err = qurt_thread_create(&me->thread[i], &attr, worker_pool_main, (void *) &me->context[i]);
+        if (err) {
+            FARF(ERROR, "Could not launch worker threads!");
+            worker_pool_release((worker_pool_context_t *) &me);
+            return AEE_EQURTTHREADCREATE;
+        }
+    }
+    *context = (worker_pool_context_t *) me;
+    return AEE_SUCCESS;
+}
+
+AEEResult worker_pool_init(worker_pool_context_t * context, uint32_t n_threads) {
+    return worker_pool_init_with_stack_size(context, n_threads, WORKER_THREAD_STACK_SZ);
+}
+
+// clean up worker pool
+void worker_pool_release(worker_pool_context_t * context) {
+    worker_pool_t * me = (worker_pool_t *) *context;
+
+    // if no worker pool exists, return error.
+    if (NULL == me) {
+        return;
+    }
+
+    atomic_store(&me->killed, 1);
+    atomic_fetch_add(&me->seqn, 1);
+    qurt_futex_wake(&me->seqn, me->n_threads);
+
+    // de-initializations
+    for (unsigned int i = 0; i < me->n_threads; i++) {
+        if (me->thread[i]) {
+            int status;
+            (void) qurt_thread_join(me->thread[i], &status);
+        }
+    }
+
+    // free allocated memory (were allocated as a single buffer starting at stack[0])
+    if (me->stack[0]) {
+        free(me->stack[0]);
+    }
+
+    *context = NULL;
+}
+
+// run jobs
+AEEResult worker_pool_run_jobs(worker_pool_context_t context, worker_pool_job_t * job, unsigned int n) {
+    worker_pool_t * me = (worker_pool_t *) context;
+    if (NULL == me) {
+        FARF(ERROR, "worker-pool: invalid context");
+        return AEE_EBADPARM;
+    }
+
+    if (n > me->n_threads) {
+        FARF(ERROR, "worker-pool: invalid number of jobs %u for n-threads %u", n, me->n_threads);
+        return AEE_EBADPARM;
+    }
+
+    memcpy(me->job, job, sizeof(worker_pool_job_t) * n);
+
+    if (n > 1) {
+        atomic_store(&me->next_job, 1);
+        atomic_store(&me->n_jobs, n);
+        atomic_store(&me->n_pending, n - 1);
+
+        // wake up workers
+        atomic_fetch_add(&me->seqn, 1);
+        qurt_futex_wake(&me->seqn, n - 1);
+    }
+
+    // main thread runs job #0
+    me->job[0].func(n, 0, me->job[0].data);
+
+    if (n > 1) {
+        while (atomic_load(&me->n_pending))
+            ;
+    }
+
+    return 0;
+}
+
+// run func
+AEEResult worker_pool_run_func(worker_pool_context_t context, worker_callback_t func, void * data, unsigned int n) {
+    worker_pool_job_t job[n];
+
+    for (unsigned int i = 0; i < n; i++) {
+        job[i].func = func;
+        job[i].data = data;
+    }
+
+    return worker_pool_run_jobs(context, job, n);
+}
+
+AEEResult worker_pool_set_thread_priority(worker_pool_context_t context, unsigned int prio) {
+    worker_pool_t * me = (worker_pool_t *) context;
+
+    // if no worker pool exists, return error.
+    if (!me) {
+        return AEE_ENOMORE;
+    }
+
+    int result = AEE_SUCCESS;
+    if (prio < 1) {
+        prio = 1;
+    }
+    if (prio > LOWEST_USABLE_QURT_PRIO) {
+        prio = LOWEST_USABLE_QURT_PRIO;
+    }
+
+    for (unsigned int i = 0; i < me->n_threads; i++) {
+        int res = qurt_thread_set_priority(me->thread[i], (unsigned short) prio);
+        if (0 != res) {
+            result = AEE_EBADPARM;
+            FARF(ERROR, "QURT failed to set priority of thread %d, ERROR = %d", me->thread[i], res);
+        }
+    }
+
+    return result;
+}
+
+AEEResult worker_pool_retrieve_thread_id(worker_pool_context_t context, unsigned int * tids) {
+    worker_pool_t * me = (worker_pool_t *) context;
+    if (!me) {
+        FARF(ERROR, "worker-pool: invalid context");
+        return AEE_EBADPARM;
+        ;
+    }
+
+    for (int i = 0; i < me->n_threads; i++) {
+        tids[i] = me->thread[i];
+    }
+
+    return AEE_SUCCESS;
+}
+
+AEEResult worker_pool_get_thread_priority(worker_pool_context_t context, unsigned int * prio) {
+    worker_pool_t * me = (worker_pool_t *) context;
+    if (!me) {
+        FARF(ERROR, "worker-pool: invalid context");
+        return AEE_EBADPARM;
+    }
+
+    int priority = qurt_thread_get_priority(me->thread[0]);
+    if (priority > 0) {
+        *prio = priority;
+        return 0;
+    } else {
+        *prio = 0;
+        return AEE_EBADSTATE;
+    }
+}