tu/a7xx: Emit HLSQ_CS_LAST_LOCAL_SIZE dynamically

We'll need this to implement unaligned dispatches.

Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/28447>

src/freedreno/vulkan/tu_cmd_buffer.cc

@@ -6768,6 +6768,13 @@ tu_dispatch(struct tu_cmd_buffer *cmd,
                    HLSQ_CS_NDRANGE_5(CHIP, .globalsize_z = local_size[2] * num_groups[2]),
                    HLSQ_CS_NDRANGE_6(CHIP, .globaloff_z = 0));
 
+   if (CHIP >= A7XX) {
+      tu_cs_emit_regs(cs,
+                      A7XX_HLSQ_CS_LAST_LOCAL_SIZE(.localsizex = local_size[0] - 1,
+                                                   .localsizey = local_size[1] - 1,
+                                                   .localsizez = local_size[2] - 1));
+   }
+
    if (info->indirect) {
       uint64_t iova = info->indirect->iova + info->indirect_offset;
 

src/freedreno/vulkan/tu_shader.cc

@@ -1552,11 +1552,6 @@ tu6_emit_cs_config(struct tu_cs *cs,
                            WORKITEMRASTORDER_LINEAR :
                            WORKITEMRASTORDER_TILED, ));
 
-      tu_cs_emit_regs(
-         cs, A7XX_HLSQ_CS_LAST_LOCAL_SIZE(.localsizex = v->local_size[0] - 1,
-                                          .localsizey = v->local_size[1] - 1,
-                                          .localsizez = v->local_size[2] - 1, ));
-
       tu_cs_emit_regs(cs, A7XX_SP_CS_UNKNOWN_A9BE(0)); // Sometimes is 0x08000000
    }
 }