diff --git a/src/asahi/libagx/geometry.h b/src/asahi/libagx/geometry.h
index a36d32db2f8..3c31b1bf989 100644
--- a/src/asahi/libagx/geometry.h
+++ b/src/asahi/libagx/geometry.h
@@ -140,10 +140,13 @@ struct agx_geometry_params {
    uint32_t xfb_prims[MAX_VERTEX_STREAMS];
 
    /* Within an indirect GS draw, the grids used to dispatch the VS/GS written
-    * out by the GS indirect setup kernel or the CPU for a direct draw.
+    * out by the GS indirect setup kernel or the CPU for a direct draw. This is
+    * the "indirect local" format: first 3 is in threads, second 3 is in grid
+    * blocks. This lets us use nontrivial workgroups with indirect draws without
+    * needing any predication.
     */
-   uint32_t vs_grid[3];
-   uint32_t gs_grid[3];
+   uint32_t vs_grid[6];
+   uint32_t gs_grid[6];
 
    /* Number of input primitives across all instances, calculated by the CPU for
     * a direct draw or the GS indirect setup kernel for an indirect draw.
@@ -167,7 +170,7 @@ struct agx_geometry_params {
     */
    uint32_t input_topology;
 } PACKED;
-static_assert(sizeof(struct agx_geometry_params) == 82 * 4);
+static_assert(sizeof(struct agx_geometry_params) == 88 * 4);
 
 /* TCS shared memory layout:
  *
diff --git a/src/asahi/vulkan/hk_cmd_draw.c b/src/asahi/vulkan/hk_cmd_draw.c
index d321d3ee033..ac75ff6d33d 100644
--- a/src/asahi/vulkan/hk_cmd_draw.c
+++ b/src/asahi/vulkan/hk_cmd_draw.c
@@ -1151,6 +1151,11 @@ hk_upload_geometry_params(struct hk_cmd_buffer *cmd, struct agx_draw draw)
       params.count_buffer = T.gpu;
    }
 
+   /* Workgroup size */
+   params.vs_grid[3] = params.gs_grid[3] = 64;
+   params.vs_grid[4] = params.gs_grid[4] = 1;
+   params.vs_grid[5] = params.gs_grid[5] = 1;
+
    if (indirect) {
       /* TODO: size */
       cmd->geom_indirect = hk_pool_alloc(cmd, 64, 4).gpu;
@@ -1425,7 +1430,7 @@ hk_launch_gs_prerast(struct hk_cmd_buffer *cmd, struct hk_cs *cs,
 
    uint64_t geometry_params = desc->root.draw.geometry_params;
    unsigned count_words = count->info.gs.count_words;
-   struct agx_workgroup wg;
+   struct agx_workgroup wg = agx_workgroup(64, 1, 1);
 
    if (false /* TODO */)
       perf_debug(cmd, "Transform feedbck");
@@ -1473,18 +1478,14 @@ hk_launch_gs_prerast(struct hk_cmd_buffer *cmd, struct hk_cs *cs,
       libagx_gs_setup_indirect_struct(cmd, agx_1d(1),
                                       AGX_BARRIER_ALL | AGX_PREGFX, gsi);
 
-      grid_vs = agx_grid_indirect(
+      grid_vs = agx_grid_indirect_local(
          geometry_params + offsetof(struct agx_geometry_params, vs_grid));
 
-      grid_gs = agx_grid_indirect(
+      grid_gs = agx_grid_indirect_local(
          geometry_params + offsetof(struct agx_geometry_params, gs_grid));
-
-      /* TODO: Optimize */
-      wg = agx_workgroup(1, 1, 1);
    } else {
       grid_vs = grid_gs = draw.b;
       grid_gs.count[0] = u_decomposed_prims_for_vertices(mode, draw.b.count[0]);
-      wg = agx_workgroup(64, 1, 1);
    }
 
    /* Launch the vertex shader first */
diff --git a/src/gallium/drivers/asahi/agx_state.c b/src/gallium/drivers/asahi/agx_state.c
index a5dae8f5b58..4b0ae802175 100644
--- a/src/gallium/drivers/asahi/agx_state.c
+++ b/src/gallium/drivers/asahi/agx_state.c
@@ -4039,6 +4039,11 @@ agx_batch_geometry_params(struct agx_batch *batch, uint64_t input_index_buffer,
       params.count_buffer = T.gpu;
    }
 
+   /* Workgroup size */
+   params.vs_grid[3] = params.gs_grid[3] = 64;
+   params.vs_grid[4] = params.gs_grid[4] = 1;
+   params.vs_grid[5] = params.gs_grid[5] = 1;
+
    if (indirect) {
       batch->uniforms.vertex_output_buffer_ptr =
          agx_pool_alloc_aligned(&batch->pool, 8, 8).gpu;
@@ -4133,7 +4138,7 @@ agx_launch_gs_prerast(struct agx_batch *batch,
 
    uint64_t gp = batch->uniforms.geometry_params;
    struct agx_grid grid_vs, grid_gs;
-   struct agx_workgroup wg;
+   struct agx_workgroup wg = agx_workgroup(64, 1, 1);
 
    /* Setup grids */
    if (indirect) {
@@ -4163,14 +4168,12 @@ agx_launch_gs_prerast(struct agx_batch *batch,
 
       libagx_gs_setup_indirect_struct(batch, agx_1d(1), AGX_BARRIER_ALL, gsi);
 
-      wg = agx_workgroup(1, 1, 1);
-      grid_vs =
-         agx_grid_indirect(gp + offsetof(struct agx_geometry_params, vs_grid));
+      grid_vs = agx_grid_indirect_local(
+         gp + offsetof(struct agx_geometry_params, vs_grid));
 
-      grid_gs =
-         agx_grid_indirect(gp + offsetof(struct agx_geometry_params, gs_grid));
+      grid_gs = agx_grid_indirect_local(
+         gp + offsetof(struct agx_geometry_params, gs_grid));
    } else {
-      wg = agx_workgroup(64, 1, 1);
       grid_vs = agx_3d(draws->count, info->instance_count, 1);
 
       grid_gs =