diff --git a/src/asahi/libagx/geometry.h b/src/asahi/libagx/geometry.h
index 103f6b0fcd1..54ef991396b 100644
--- a/src/asahi/libagx/geometry.h
+++ b/src/asahi/libagx/geometry.h
@@ -113,10 +113,17 @@ static_assert(sizeof(struct agx_heap) == 4 * 4);
 
 #ifdef __OPENCL_VERSION__
 static inline uint
-agx_heap_alloc_nonatomic_offs(global struct agx_heap *heap, uint size_B)
+_agx_heap_alloc_offs(global struct agx_heap *heap, uint size_B, bool atomic)
 {
-   uint offs = heap->bottom;
-   heap->bottom += align(size_B, 16);
+   size_B = align(size_B, 16);
+
+   uint offs;
+   if (atomic) {
+      offs = atomic_fetch_add((volatile atomic_uint *)(&heap->bottom), size_B);
+   } else {
+      offs = heap->bottom;
+      heap->bottom = offs + size_B;
+   }
 
    /* Use printf+abort because assert is stripped from release builds. */
    if (heap->bottom >= heap->size) {
@@ -130,6 +137,18 @@ agx_heap_alloc_nonatomic_offs(global struct agx_heap *heap, uint size_B)
    return offs;
 }
 
+static inline uint
+agx_heap_alloc_nonatomic_offs(global struct agx_heap *heap, uint size_B)
+{
+   return _agx_heap_alloc_offs(heap, size_B, false);
+}
+
+static inline uint
+agx_heap_alloc_atomic_offs(global struct agx_heap *heap, uint size_B)
+{
+   return _agx_heap_alloc_offs(heap, size_B, true);
+}
+
 static inline global void *
 agx_heap_alloc_nonatomic(global struct agx_heap *heap, uint size_B)
 {
diff --git a/src/asahi/libagx/tessellator.cl b/src/asahi/libagx/tessellator.cl
index 090064b7cf3..957230e422d 100644
--- a/src/asahi/libagx/tessellator.cl
+++ b/src/asahi/libagx/tessellator.cl
@@ -117,14 +117,6 @@ tess_factors(constant struct libagx_tess_args *p, uint patch)
    return p->tcs_buffer + (patch * p->tcs_stride_el);
 }
 
-static inline uint
-libagx_heap_alloc(global struct agx_heap *heap, uint size_B)
-{
-   // TODO: drop align to 4 I think
-   return atomic_fetch_add((volatile atomic_uint *)(&heap->bottom),
-                           align(size_B, 8));
-}
-
 /*
  * Generate an indexed draw for a patch with the computed number of indices.
  * This allocates heap memory for the index buffer, returning the allocated
@@ -196,7 +188,7 @@ libagx_heap_alloc_points(constant struct libagx_tess_args *p, uint patch,
    }
 
    uint32_t elsize_B = sizeof(struct libagx_tess_point);
-   uint32_t alloc_B = libagx_heap_alloc(p->heap, elsize_B * count);
+   uint32_t alloc_B = agx_heap_alloc_atomic_offs(p->heap, elsize_B * count);
    uint32_t alloc_el = alloc_B / elsize_B;
 
    p->coord_allocs[patch] = alloc_el;