hexagon: make vmem and buffer-size configurable (#22487)

max-krasnyansky · ServeurpersoCom · web-flow · commit 41a63be28eb9 · 2026-04-29T11:51:21.000-07:00
* hexagon: allow host to set max vmem size

We use a sane default but it's helpful to allow for an override if needed.

* hexagon: add support for measuring vmem space and move pinned mmaping management to host

* hexagon: update vmem checks to use uint64

* hexagon: bump op buffers to 16 (matches max mmaps)

* hexagon: bump default vmem to 3.2GB

* hexagon: add support for autodetecting vmem space and some logging cleanup in that area

* hexagon: fix whitespace warnings

* Update scripts/snapdragon/adb/run-cli.sh

Co-authored-by: Pascal &lt;admin@serveurperso.com&gt;

* hex-adb: fix run-completion script

---------

Co-authored-by: Pascal &lt;admin@serveurperso.com&gt;
diff --git a/ggml/src/ggml-hexagon/ggml-hexagon.cpp b/ggml/src/ggml-hexagon/ggml-hexagon.cpp
diff --git a/ggml/src/ggml-hexagon/htp/htp-ctx.h b/ggml/src/ggml-hexagon/htp/htp-ctx.h
@@ -20,7 +20,7 @@ struct htp_mmap {
     uint64_t size;
     uint64_t base;
     uint32_t fd;
-    uint32_t pinned;
+    uint32_t reserved;
 };
 
 // Scratchpad state
@@ -77,6 +77,8 @@ struct htp_context {
     atomic_bool            vtcm_valid;
     atomic_bool            vtcm_needs_release;
 
+    uint64_t               max_vmem;
+
     struct htp_ops_context octx;
 
 #ifdef HTP_HAS_HMX
diff --git a/ggml/src/ggml-hexagon/htp/htp-ops.h b/ggml/src/ggml-hexagon/htp/htp-ops.h
@@ -90,15 +90,11 @@ enum htp_op_code {
 #define HTP_OP_MAX_INPUTS  6    // aka GGML_MAX_SRCS
 #define HTP_OP_MAX_PARAMS  16   // aka GGML_MAX_OP_PARAMS
 
-#define HTP_OP_MAX_BUFS    8
+#define HTP_OP_MAX_BUFS    16
 #define HTP_OP_MAX_REQS    256
 #define HTP_OP_MAX_TENSORS (HTP_OP_MAX_REQS * HTP_OP_MAX_INPUTS + HTP_OP_MAX_REQS)
 
-#if __HVX_ARCH__ < 75
-#define HTP_OP_MAX_VMEM    (3167538380u)
-#else
-#define HTP_OP_MAX_VMEM    (3221225472u)
-#endif
+#define HTP_OP_MAX_VMEM_DEFAULT (3355443200u)
 
 #define HTP_MMAP_MAX_VMEM  (2147483648u)
 
diff --git a/ggml/src/ggml-hexagon/htp/htp_iface.idl b/ggml/src/ggml-hexagon/htp/htp_iface.idl
@@ -11,9 +11,9 @@ struct htp_iface_pmu_conf {
 };
 
 interface htp_iface : remote_handle64 {
-    AEEResult start(in uint32 sess_id, in uint64 dsp_queue_id, in uint32 n_hvx, in uint32 use_hmx);
+    AEEResult start(in uint32 sess_id, in uint64 dsp_queue_id, in uint32 n_hvx, in uint32 use_hmx, in uint64 max_vmem);
     AEEResult stop();
-    AEEResult mmap(in uint32 fd, in uint32 size, in uint32 pinned);
+    AEEResult mmap(in uint32 fd, in uint32 size);
     AEEResult munmap(in uint32 fd);
     AEEResult profiler(in uint32 mode, in htp_iface_pmu_conf pmu);
     AEEResult etm(in uint32 enable);
diff --git a/ggml/src/ggml-hexagon/htp/main.c b/ggml/src/ggml-hexagon/htp/main.c
@@ -210,7 +210,7 @@ AEEResult htp_iface_close(remote_handle64 handle) {
     return AEE_SUCCESS;
 }
 
-AEEResult htp_iface_mmap(remote_handle64 handle, uint32 fd, uint32 size, uint32 pinned) {
+AEEResult htp_iface_mmap(remote_handle64 handle, uint32_t fd, uint32_t size) {
     struct htp_context * ctx = (struct htp_context *) handle;
     if (!ctx) {
         return AEE_EBADPARM;
@@ -220,7 +220,6 @@ AEEResult htp_iface_mmap(remote_handle64 handle, uint32 fd, uint32 size, uint32
     for (uint32_t i=0; i<HTP_MAX_MMAPS; i++) {
         struct htp_mmap *m = &ctx->mmap[i];
         if (m->fd == fd) {
-            m->pinned = pinned;
             return AEE_SUCCESS;
         }
     }
@@ -229,7 +228,7 @@ AEEResult htp_iface_mmap(remote_handle64 handle, uint32 fd, uint32 size, uint32
     for (uint32_t i=0; i<HTP_MAX_MMAPS; i++) {
         struct htp_mmap *m = &ctx->mmap[i];
         if (!m->size) {
-            FARF(HIGH, "mmap : fd %u size %u pinned %u", fd, size, pinned);
+            FARF(HIGH, "mmap : fd %u size %u", fd, size);
 #if __HVX_ARCH__ > 73
             void *va = HAP_mmap2(NULL, size, HAP_PROT_READ | HAP_PROT_WRITE, 0, fd, 0);
 #else
@@ -248,7 +247,6 @@ AEEResult htp_iface_mmap(remote_handle64 handle, uint32 fd, uint32 size, uint32
             m->base   = (uint64_t) va;
             m->fd     = fd;
             m->size   = size;
-            m->pinned = pinned;
 
             return AEE_SUCCESS;
         }
@@ -275,7 +273,6 @@ AEEResult htp_iface_munmap(remote_handle64 handle, uint32 fd) {
             m->size   = 0;
             m->base   = NULL;
             m->fd     = -1;
-            m->pinned = 0;
         }
     }
 
@@ -358,7 +355,7 @@ static void vtcm_free(struct htp_context * ctx) {
 static void htp_packet_callback(dspqueue_t queue, int error, void * context);
 static void htp_error_callback(dspqueue_t queue, int error, void * context);
 
-AEEResult htp_iface_start(remote_handle64 handle, uint32 sess_id, uint64 dsp_queue_id, uint32 n_hvx, uint32 use_hmx) {
+AEEResult htp_iface_start(remote_handle64 handle, uint32 sess_id, uint64 dsp_queue_id, uint32 n_hvx, uint32 use_hmx, uint64_t max_vmem) {
     struct htp_context * ctx = (struct htp_context *) handle;
 
     if (!ctx) {
@@ -376,12 +373,12 @@ AEEResult htp_iface_start(remote_handle64 handle, uint32 sess_id, uint64 dsp_que
                               htp_error_callback,   // Error callback; no errors expected on the DSP
                               (void *) ctx,         // Callback context
                               &ctx->queue);
-
     if (err) {
         FARF(ERROR, "Queue import failed with 0x%08x", (unsigned) err);
         return err;
     }
 
+    ctx->max_vmem    = max_vmem;
     ctx->thread_id   = qurt_thread_get_id();
     ctx->thread_prio = qurt_thread_get_priority(ctx->thread_id);
 
@@ -622,8 +619,8 @@ static inline bool reuse_buf(struct htp_context *ctx, uint32_t *m_reuse, struct
 }
 
 static inline void drop_mmap(struct htp_context *ctx, struct htp_mmap *m) {
-    if (m->size && !m->pinned) {
-        FARF(HIGH, "unmap : fd %u base %p size %u pinned %u", m->fd, (void*) m->base, (uint32_t) m->size, m->pinned);
+    if (m->size) {
+        FARF(HIGH, "unmap : fd %u base %p size %u", m->fd, (void*) m->base, (uint32_t) m->size);
 #if __HVX_ARCH__ > 73
         HAP_munmap2((void *) m->base, m->size);
 #else
@@ -660,9 +657,8 @@ static inline void mmap_buf(struct htp_context *ctx, struct htp_buf_desc *b) {
             m->base   = b->base = (uint64_t) va;
             m->fd     = b->fd;
             m->size   = b->size;
-            m->pinned = 0;
 
-            FARF(HIGH, "mmap : fd %u base %p size %u pinned %u", m->fd, (void*) m->base, (uint32_t) m->size, m->pinned);
+            FARF(HIGH, "mmap : fd %u base %p size %u", m->fd, (void*) m->base, (uint32_t) m->size);
             return;
         }
     }
@@ -672,8 +668,8 @@ static void prep_op_bufs(struct htp_context *ctx, struct htp_buf_desc *bufs, uin
     uint32_t m_reuse = 0; // mmap reuse mask (index from ctx->mmap array)
     uint32_t b_reuse = 0; // buf reuse count
 
-    size_t   m_vmem  = 0; // mapped vmem
-    size_t   e_vmem  = 0; // extra  vmem
+    uint64_t m_vmem  = 0; // mapped vmem
+    uint64_t e_vmem  = 0; // extra  vmem
 
     // See what we can reuse
     for (uint32_t i=0; i < n_bufs; i++) {
@@ -687,9 +683,10 @@ static void prep_op_bufs(struct htp_context *ctx, struct htp_buf_desc *bufs, uin
     // See how much vmem we have mmaped right now
     for (uint32_t i=0; i<HTP_MAX_MMAPS; i++) { m_vmem += ctx->mmap[i].size; }
 
-    FARF(HIGH, "prep-bufs : pass1 mmap-vmem %zu extra-vmem %zu n-bufs %u b-reuse %u", m_vmem, e_vmem, n_bufs, b_reuse);
+    FARF(HIGH, "prep-bufs : pass1 mmap-vmem %zu extra-vmem %zu max-vmem %zu : n-bufs %u b-reuse %u",
+            (size_t) m_vmem, (size_t) e_vmem, (size_t) ctx->max_vmem, n_bufs, b_reuse);
 
-    if ((m_vmem + e_vmem) > HTP_OP_MAX_VMEM) {
+    if ((m_vmem + e_vmem) > ctx->max_vmem) {
         // Drop unused mappings
         for (uint32_t i=0; i < HTP_MAX_MMAPS; i++) {
             bool used = m_reuse & (1<<i);
diff --git a/scripts/snapdragon/adb/run-cli.sh b/scripts/snapdragon/adb/run-cli.sh
@@ -54,13 +54,23 @@ opqueue=
 opflt=
 [ "$OF" != "" ] && opflt="GGML_HEXAGON_OPFILTER=$OF"
 
+vmem=
+[ "$VM" != "" ] && opflt="GGML_HEXAGON_VMEM=$VM"
+
+mbuf=
+[ "$MB" != "" ] && opflt="GGML_HEXAGON_MBUF=$MB"
+vmem=
+[ "$VM" != "" ] && vmem="GGML_HEXAGON_VMEM=$VM"
+
+mbuf=
+[ "$MB" != "" ] && mbuf="GGML_HEXAGON_MBUF=$MB"
 set -x
 
 adb $adbserial $adbhost shell " \
   cd $basedir; ulimit -c unlimited;        \
     LD_LIBRARY_PATH=$basedir/$branch/lib   \
     ADSP_LIBRARY_PATH=$basedir/$branch/lib \
-    $verbose $sched $opmask $profile $nhvx $hmx $ndev $hb $opbatch $opqueue $opflt \
+    $verbose $sched $opmask $profile $nhvx $hmx $ndev $hb $opbatch $opqueue $opflt $vmem $mbuf \
       ./$branch/bin/llama-cli --no-mmap -m $basedir/../gguf/$model \
          --poll 1000 -t 6 --cpu-mask 0xfc --cpu-strict 1           \
          --ctx-size 8192 --ubatch-size 256 -fa on                  \
diff --git a/scripts/snapdragon/adb/run-completion.sh b/scripts/snapdragon/adb/run-completion.sh
@@ -54,13 +54,19 @@ opqueue=
 opflt=
 [ "$OF" != "" ] && opflt="GGML_HEXAGON_OPFILTER=$OF"
 
+vmem=
+[ "$VM" != "" ] && vmem="GGML_HEXAGON_VMEM=$VM"
+
+mbuf=
+[ "$MB" != "" ] && mbuf="GGML_HEXAGON_MBUF=$MB"
+
 set -x
 
 adb $adbserial $adbhost shell " \
   cd $basedir; ulimit -c unlimited;        \
     LD_LIBRARY_PATH=$basedir/$branch/lib   \
     ADSP_LIBRARY_PATH=$basedir/$branch/lib \
-    $verbose $sched $opmask $profile $nhvx $hmx $ndev $hb $opbatch $opqueue $opflt \
+    $verbose $sched $opmask $profile $nhvx $hmx $ndev $hb $opbatch $opqueue $opflt $vmem $mbuf \
       ./$branch/bin/llama-completion --no-mmap -m $basedir/../gguf/$model \
          --poll 1000 -t 6 --cpu-mask 0xfc --cpu-strict 1                  \
          --ctx-size 8192 --ubatch-size 256 -fa on                         \
