diff --git a/src/amd/common/ac_nir.c b/src/amd/common/ac_nir.c
index f26866bb317..0b0f8b891d2 100644
--- a/src/amd/common/ac_nir.c
+++ b/src/amd/common/ac_nir.c
@@ -233,7 +233,7 @@ lower_intrinsic_to_arg(nir_builder *b, nir_instr *instr, void *state)
                              ac_nir_load_arg(b, s->args, s->args->frag_pos[3]));
       break;
    case nir_intrinsic_load_local_invocation_id:
-      if (s->args->args[s->args->local_invocation_ids.arg_index].size == 1) {
+      if (s->args->local_invocation_ids_packed.used) {
          /* Thread IDs are packed in VGPR0, 10 bits per component. */
          unsigned num_bits[3];
 
@@ -263,13 +263,17 @@ lower_intrinsic_to_arg(nir_builder *b, nir_instr *instr, void *state)
          nir_def *vec[3];
          for (unsigned i = 0; i < 3; i++) {
             vec[i] = !num_bits[i] ? nir_imm_int(b, 0) :
-                        ac_nir_unpack_arg(b, s->args, s->args->local_invocation_ids, i * 10,
+                        ac_nir_unpack_arg(b, s->args,
+                                          s->args->local_invocation_ids_packed, i * 10,
                                           num_bits[i]);
          }
 
          replacement = nir_vec(b, vec, 3);
       } else {
-         replacement = ac_nir_load_arg(b, s->args, s->args->local_invocation_ids);
+         replacement = nir_vec3(b,
+                                ac_nir_load_arg(b, s->args, s->args->local_invocation_id_x),
+                                ac_nir_load_arg(b, s->args, s->args->local_invocation_id_y),
+                                ac_nir_load_arg(b, s->args, s->args->local_invocation_id_z));
       }
       break;
    case nir_intrinsic_load_merged_wave_info_amd:
diff --git a/src/amd/common/ac_shader_args.h b/src/amd/common/ac_shader_args.h
index 62ac708c3d1..3854dccdc71 100644
--- a/src/amd/common/ac_shader_args.h
+++ b/src/amd/common/ac_shader_args.h
@@ -161,7 +161,10 @@ struct ac_shader_args {
    struct ac_arg pos_fixed_pt;
 
    /* CS */
-   struct ac_arg local_invocation_ids;
+   struct ac_arg local_invocation_id_x;
+   struct ac_arg local_invocation_id_y;
+   struct ac_arg local_invocation_id_z;
+   struct ac_arg local_invocation_ids_packed;
    struct ac_arg num_work_groups;
    /* GFX6-11 only. GFX12+ uses read only SGPRs {TTMP9[0:31], TTMP7[0:15], TTMP7[16:31]}. */
    struct ac_arg workgroup_ids[3];
diff --git a/src/amd/compiler/aco_instruction_selection.cpp b/src/amd/compiler/aco_instruction_selection.cpp
index 569b8913bed..cae2be6d8cd 100644
--- a/src/amd/compiler/aco_instruction_selection.cpp
+++ b/src/amd/compiler/aco_instruction_selection.cpp
@@ -12411,9 +12411,11 @@ select_rt_prolog(Program* program, ac_shader_config* config,
    }
    if (options->gfx_level < GFX11)
       in_scratch_offset = get_arg_reg(in_args, in_args->scratch_offset);
+   struct ac_arg arg_id = options->gfx_level >= GFX11 ? in_args->local_invocation_ids_packed
+                                                      : in_args->local_invocation_id_x;
    PhysReg in_local_ids[2] = {
-      get_arg_reg(in_args, in_args->local_invocation_ids),
-      get_arg_reg(in_args, in_args->local_invocation_ids).advance(4),
+      get_arg_reg(in_args, arg_id),
+      get_arg_reg(in_args, arg_id).advance(4),
    };
 
    /* Outputs:
diff --git a/src/amd/vulkan/radv_shader_args.c b/src/amd/vulkan/radv_shader_args.c
index 6d14e50ad12..c6c13bada39 100644
--- a/src/amd/vulkan/radv_shader_args.c
+++ b/src/amd/vulkan/radv_shader_args.c
@@ -249,7 +249,7 @@ declare_ms_input_vgprs(const struct radv_device *device, struct radv_shader_args
    const struct radv_physical_device *pdev = radv_device_physical(device);
 
    if (pdev->mesh_fast_launch_2) {
-      ac_add_arg(&args->ac, AC_ARG_VGPR, 1, AC_ARG_INT, &args->ac.local_invocation_ids);
+      ac_add_arg(&args->ac, AC_ARG_VGPR, 1, AC_ARG_INT, &args->ac.local_invocation_ids_packed);
    } else {
       ac_add_arg(&args->ac, AC_ARG_VGPR, 1, AC_ARG_INT, &args->ac.vertex_id);
       ac_add_arg(&args->ac, AC_ARG_VGPR, 1, AC_ARG_INT, NULL); /* user vgpr */
@@ -616,10 +616,13 @@ declare_shader_args(const struct radv_device *device, const struct radv_graphics
          ac_add_arg(&args->ac, AC_ARG_SGPR, 1, AC_ARG_INT, &args->ac.scratch_offset);
       }
 
-      if (gfx_level >= GFX11)
-         ac_add_arg(&args->ac, AC_ARG_VGPR, 1, AC_ARG_INT, &args->ac.local_invocation_ids);
-      else
-         ac_add_arg(&args->ac, AC_ARG_VGPR, 3, AC_ARG_INT, &args->ac.local_invocation_ids);
+      if (gfx_level >= GFX11) {
+         ac_add_arg(&args->ac, AC_ARG_VGPR, 1, AC_ARG_INT, &args->ac.local_invocation_ids_packed);
+      } else {
+         ac_add_arg(&args->ac, AC_ARG_VGPR, 1, AC_ARG_INT, &args->ac.local_invocation_id_x);
+         ac_add_arg(&args->ac, AC_ARG_VGPR, 1, AC_ARG_INT, &args->ac.local_invocation_id_y);
+         ac_add_arg(&args->ac, AC_ARG_VGPR, 1, AC_ARG_INT, &args->ac.local_invocation_id_z);
+      }
       break;
    case MESA_SHADER_VERTEX:
       /* NGG is handled by the GS case */
diff --git a/src/gallium/drivers/radeonsi/si_shader.c b/src/gallium/drivers/radeonsi/si_shader.c
index 5bf2a1e8478..3d12f11064f 100644
--- a/src/gallium/drivers/radeonsi/si_shader.c
+++ b/src/gallium/drivers/radeonsi/si_shader.c
@@ -741,10 +741,13 @@ static void si_init_shader_args(struct si_shader *shader, struct si_shader_args
       /* Hardware VGPRs. */
       /* Thread IDs are packed in VGPR0, 10 bits per component or stored in 3 separate VGPRs */
       if (sel->screen->info.gfx_level >= GFX11 ||
-          (!sel->screen->info.has_graphics && sel->screen->info.family >= CHIP_MI200))
-         ac_add_arg(&args->ac, AC_ARG_VGPR, 1, AC_ARG_INT, &args->ac.local_invocation_ids);
-      else
-         ac_add_arg(&args->ac, AC_ARG_VGPR, 3, AC_ARG_INT, &args->ac.local_invocation_ids);
+          (!sel->screen->info.has_graphics && sel->screen->info.family >= CHIP_MI200)) {
+         ac_add_arg(&args->ac, AC_ARG_VGPR, 1, AC_ARG_INT, &args->ac.local_invocation_ids_packed);
+      } else {
+         ac_add_arg(&args->ac, AC_ARG_VGPR, 1, AC_ARG_INT, &args->ac.local_invocation_id_x);
+         ac_add_arg(&args->ac, AC_ARG_VGPR, 1, AC_ARG_INT, &args->ac.local_invocation_id_y);
+         ac_add_arg(&args->ac, AC_ARG_VGPR, 1, AC_ARG_INT, &args->ac.local_invocation_id_z);
+      }
       break;
    default:
       assert(0 && "unimplemented shader");