diff --git a/src/freedreno/ir3/ir3.c b/src/freedreno/ir3/ir3.c
index 9746d335a64..50e6cd239ce 100644
--- a/src/freedreno/ir3/ir3.c
+++ b/src/freedreno/ir3/ir3.c
@@ -449,6 +449,10 @@ ir3_collect_info(struct ir3_shader_variant *v)
       (compiler->gen >= 6 ? ((info->max_half_reg + 2) / 2) : 0);
 
    info->double_threadsize = ir3_should_double_threadsize(v, regs_count);
+
+   /* TODO this is different for earlier gens, but earlier gens don't use this */
+   info->subgroup_size = v->info.double_threadsize ? 128 : 64;
+
    unsigned reg_independent_max_waves =
       ir3_get_reg_independent_max_waves(v, info->double_threadsize);
    unsigned reg_dependent_max_waves = ir3_get_reg_dependent_max_waves(
diff --git a/src/freedreno/ir3/ir3.h b/src/freedreno/ir3/ir3.h
index 722b98f24f9..aace5df184d 100644
--- a/src/freedreno/ir3/ir3.h
+++ b/src/freedreno/ir3/ir3.h
@@ -72,6 +72,7 @@ struct ir3_info {
     * assuming that they are all executing this shader.
     */
    int8_t max_waves;
+   uint8_t subgroup_size;
    bool double_threadsize;
    bool multi_dword_ldp_stp;
 
diff --git a/src/freedreno/vulkan/tu_pipeline.c b/src/freedreno/vulkan/tu_pipeline.c
index 2fdb466c57d..b7bd9696318 100644
--- a/src/freedreno/vulkan/tu_pipeline.c
+++ b/src/freedreno/vulkan/tu_pipeline.c
@@ -5283,7 +5283,7 @@ tu_compute_pipeline_create(VkDevice device,
    for (int i = 0; i < 3; i++)
       pipeline->compute.local_size[i] = v->local_size[i];
 
-   pipeline->compute.subgroup_size = v->info.double_threadsize ? 128 : 64;
+   pipeline->compute.subgroup_size = v->info.subgroup_size;
 
    struct tu_cs prog_cs;
    uint32_t additional_reserve_size = tu_xs_get_additional_cs_size_dwords(v);