diff --git a/src/asahi/lib/agx_nir_lower_gs.c b/src/asahi/lib/agx_nir_lower_gs.c
index 0fcf6bb005c..4dedeedcf26 100644
--- a/src/asahi/lib/agx_nir_lower_gs.c
+++ b/src/asahi/lib/agx_nir_lower_gs.c
@@ -1461,22 +1461,25 @@ lower_vs_before_gs(nir_builder *b, nir_intrinsic_instr *intr, void *data)
     */
    nir_def *mask = nir_imm_int64(b, b->shader->info.outputs_written);
 
+   nir_def *buffer;
    nir_def *nr_verts;
    if (b->shader->info.stage == MESA_SHADER_VERTEX) {
+      buffer = nir_load_vs_output_buffer_agx(b);
       nr_verts =
          libagx_input_vertices(b, nir_load_input_assembly_buffer_agx(b));
    } else {
-      /* TODO: Do something similar for tessellation, load_num_workgroups is
-       * annoying in a software graphics shader.
-       */
-      nr_verts = nir_channel(b, nir_load_num_workgroups(b), 0);
+      assert(b->shader->info.stage == MESA_SHADER_TESS_EVAL);
+
+      /* Instancing is unrolled during tessellation so nr_verts is ignored. */
+      nr_verts = nir_imm_int(b, 0);
+      buffer = libagx_tes_buffer(b, nir_load_tess_param_buffer_agx(b));
    }
 
    nir_def *linear_id = nir_iadd(b, nir_imul(b, load_instance_id(b), nr_verts),
                                  load_primitive_id(b));
 
-   nir_def *addr = libagx_vertex_output_address(
-      b, nir_load_vs_output_buffer_agx(b), mask, linear_id, location);
+   nir_def *addr =
+      libagx_vertex_output_address(b, buffer, mask, linear_id, location);
 
    assert(nir_src_bit_size(intr->src[0]) == 32);
    addr = nir_iadd_imm(b, addr, nir_intrinsic_component(intr) * 4);
diff --git a/src/asahi/lib/shaders/tessellation.cl b/src/asahi/lib/shaders/tessellation.cl
index a001d71ce97..d0bcbb68017 100644
--- a/src/asahi/lib/shaders/tessellation.cl
+++ b/src/asahi/lib/shaders/tessellation.cl
@@ -25,6 +25,12 @@ libagx_tcs_unrolled_id(constant struct libagx_tess_args *p, uint3 wg_id)
    return (wg_id.y * p->patches_per_instance) + wg_id.x;
 }
 
+uint64_t
+libagx_tes_buffer(constant struct libagx_tess_args *p)
+{
+   return p->tes_buffer;
+}
+
 /*
  * Helper to lower indexing for a tess eval shader ran as a compute shader. This
  * handles the tess+geom case. This is simpler than the general input assembly
diff --git a/src/asahi/lib/shaders/tessellator.h b/src/asahi/lib/shaders/tessellator.h
index ec674f43764..a2547e51e38 100644
--- a/src/asahi/lib/shaders/tessellator.h
+++ b/src/asahi/lib/shaders/tessellator.h
@@ -91,6 +91,11 @@ struct libagx_tess_args {
     */
    GLOBAL(uint64_t) vertex_output_buffer_ptr;
 
+   /* When geom+tess used together, the buffer containing TES outputs (executed
+    * as a hardware compute shader).
+    */
+   uint64_t tes_buffer;
+
    /* For indirect draws, the bitfield of VS outputs */
    uint64_t vertex_outputs;
 
@@ -121,4 +126,4 @@ struct libagx_tess_args {
    /* Number of patches being tessellated */
    uint32_t nr_patches;
 } PACKED;
-AGX_STATIC_ASSERT(sizeof(struct libagx_tess_args) == 40 * 4);
+AGX_STATIC_ASSERT(sizeof(struct libagx_tess_args) == 42 * 4);