diff --git a/drivers/gpu/nvgpu/common/pmu/pg/pmu_pg.c b/drivers/gpu/nvgpu/common/pmu/pg/pmu_pg.c
index 5bc8be27e..cfd08fab8 100644
--- a/drivers/gpu/nvgpu/common/pmu/pg/pmu_pg.c
+++ b/drivers/gpu/nvgpu/common/pmu/pg/pmu_pg.c
@@ -29,7 +29,6 @@
 #include <nvgpu/gk20a.h>
 #include <nvgpu/string.h>
 #include <nvgpu/engines.h>
-#include <nvgpu/gr/gr.h>
 
 /* state transition :
  * OFF => [OFF_ON_PENDING optional] => ON_PENDING => ON => OFF
@@ -536,7 +535,7 @@ int nvgpu_pmu_init_powergating(struct gk20a *g)
 		pg_engine_id_list = g->ops.pmu.pmu_pg_supported_engines_list(g);
 	}
 
-	nvgpu_gr_wait_initialized(g);
+	g->ops.gr.init.wait_initialized(g);
 
 	for (pg_engine_id = PMU_PG_ELPG_ENGINE_ID_GRAPHICS;
 		pg_engine_id < PMU_PG_ELPG_ENGINE_ID_INVALID_ENGINE;
diff --git a/drivers/gpu/nvgpu/common/power_features/cg/cg.c b/drivers/gpu/nvgpu/common/power_features/cg/cg.c
index 5bb75efda..57a501105 100644
--- a/drivers/gpu/nvgpu/common/power_features/cg/cg.c
+++ b/drivers/gpu/nvgpu/common/power_features/cg/cg.c
@@ -21,7 +21,6 @@
  */
 
 #include <nvgpu/gk20a.h>
-#include <nvgpu/gr/gr.h>
 #include <nvgpu/enabled.h>
 #include <nvgpu/power_features/cg.h>
 
@@ -94,7 +93,7 @@ void nvgpu_cg_elcg_enable(struct gk20a *g)
 		return;
 	}
 
-	nvgpu_gr_wait_initialized(g);
+	g->ops.gr.init.wait_initialized(g);
 
 	nvgpu_mutex_acquire(&g->cg_pg_lock);
 	if (g->elcg_enabled) {
@@ -111,7 +110,7 @@ void nvgpu_cg_elcg_disable(struct gk20a *g)
 		return;
 	}
 
-	nvgpu_gr_wait_initialized(g);
+	g->ops.gr.init.wait_initialized(g);
 
 	nvgpu_mutex_acquire(&g->cg_pg_lock);
 	if (g->elcg_enabled) {
@@ -129,7 +128,7 @@ void nvgpu_cg_blcg_mode_enable(struct gk20a *g)
 		return;
 	}
 
-	nvgpu_gr_wait_initialized(g);
+	g->ops.gr.init.wait_initialized(g);
 
 	nvgpu_mutex_acquire(&g->cg_pg_lock);
 	if (g->blcg_enabled) {
@@ -147,7 +146,7 @@ void nvgpu_cg_blcg_mode_disable(struct gk20a *g)
 		return;
 	}
 
-	nvgpu_gr_wait_initialized(g);
+	g->ops.gr.init.wait_initialized(g);
 
 	nvgpu_mutex_acquire(&g->cg_pg_lock);
 	if (g->blcg_enabled) {
@@ -305,7 +304,7 @@ void nvgpu_cg_slcg_gr_perf_ltc_load_enable(struct gk20a *g)
 		return;
 	}
 
-	nvgpu_gr_wait_initialized(g);
+	g->ops.gr.init.wait_initialized(g);
 
 	nvgpu_mutex_acquire(&g->cg_pg_lock);
 	if (!g->slcg_enabled) {
@@ -332,7 +331,7 @@ void nvgpu_cg_slcg_gr_perf_ltc_load_disable(struct gk20a *g)
 		return;
 	}
 
-	nvgpu_gr_wait_initialized(g);
+	g->ops.gr.init.wait_initialized(g);
 
 	nvgpu_mutex_acquire(&g->cg_pg_lock);
 	if (!g->slcg_enabled) {
@@ -476,7 +475,7 @@ void nvgpu_cg_elcg_set_elcg_enabled(struct gk20a *g, bool enable)
 		return;
 	}
 
-	nvgpu_gr_wait_initialized(g);
+	g->ops.gr.init.wait_initialized(g);
 
 	nvgpu_mutex_release(&g->cg_pg_lock);
 	if (enable) {
@@ -503,7 +502,7 @@ void nvgpu_cg_blcg_set_blcg_enabled(struct gk20a *g, bool enable)
 		return;
 	}
 
-	nvgpu_gr_wait_initialized(g);
+	g->ops.gr.init.wait_initialized(g);
 
 	nvgpu_mutex_acquire(&g->cg_pg_lock);
 	if (enable) {
@@ -564,7 +563,7 @@ void nvgpu_cg_slcg_set_slcg_enabled(struct gk20a *g, bool enable)
 		return;
 	}
 
-	nvgpu_gr_wait_initialized(g);
+	g->ops.gr.init.wait_initialized(g);
 
 	nvgpu_mutex_acquire(&g->cg_pg_lock);
 	if (enable) {
diff --git a/drivers/gpu/nvgpu/common/power_features/pg/pg.c b/drivers/gpu/nvgpu/common/power_features/pg/pg.c
index 6758569cb..ad7592eb6 100644
--- a/drivers/gpu/nvgpu/common/power_features/pg/pg.c
+++ b/drivers/gpu/nvgpu/common/power_features/pg/pg.c
@@ -21,7 +21,6 @@
  */
 
 #include <nvgpu/gk20a.h>
-#include <nvgpu/gr/gr.h>
 #include <nvgpu/pmu.h>
 #include <nvgpu/power_features/pg.h>
 
@@ -47,7 +46,7 @@ int nvgpu_pg_elpg_enable(struct gk20a *g)
 		return 0;
 	}
 
-	nvgpu_gr_wait_initialized(g);
+	g->ops.gr.init.wait_initialized(g);
 
 	nvgpu_mutex_acquire(&g->cg_pg_lock);
 	if (g->elpg_enabled) {
@@ -67,7 +66,7 @@ int nvgpu_pg_elpg_disable(struct gk20a *g)
 		return 0;
 	}
 
-	nvgpu_gr_wait_initialized(g);
+	g->ops.gr.init.wait_initialized(g);
 
 	nvgpu_mutex_acquire(&g->cg_pg_lock);
 	if (g->elpg_enabled) {
@@ -88,7 +87,7 @@ int nvgpu_pg_elpg_set_elpg_enabled(struct gk20a *g, bool enable)
 		return 0;
 	}
 
-	nvgpu_gr_wait_initialized(g);
+	g->ops.gr.init.wait_initialized(g);
 
 	nvgpu_mutex_acquire(&g->cg_pg_lock);
 	if (enable) {
diff --git a/drivers/gpu/nvgpu/common/power_features/power_features.c b/drivers/gpu/nvgpu/common/power_features/power_features.c
index 43e557dbe..a357b401b 100644
--- a/drivers/gpu/nvgpu/common/power_features/power_features.c
+++ b/drivers/gpu/nvgpu/common/power_features/power_features.c
@@ -21,7 +21,6 @@
  */
 
 #include <nvgpu/gk20a.h>
-#include <nvgpu/gr/gr.h>
 #include <nvgpu/power_features/cg.h>
 #include <nvgpu/power_features/pg.h>
 #include <nvgpu/power_features/power_features.h>
@@ -32,7 +31,7 @@ int nvgpu_cg_pg_disable(struct gk20a *g)
 
 	nvgpu_log_fn(g, " ");
 
-	nvgpu_gr_wait_initialized(g);
+	g->ops.gr.init.wait_initialized(g);
 
 	/* disable elpg before clock gating */
 	err = nvgpu_pg_elpg_disable(g);
@@ -54,7 +53,7 @@ int nvgpu_cg_pg_enable(struct gk20a *g)
 
 	nvgpu_log_fn(g, " ");
 
-	nvgpu_gr_wait_initialized(g);
+	g->ops.gr.init.wait_initialized(g);
 
 	nvgpu_cg_elcg_enable(g);
 
diff --git a/drivers/gpu/nvgpu/gm20b/hal_gm20b.c b/drivers/gpu/nvgpu/gm20b/hal_gm20b.c
index 083105a5c..47bb44b4f 100644
--- a/drivers/gpu/nvgpu/gm20b/hal_gm20b.c
+++ b/drivers/gpu/nvgpu/gm20b/hal_gm20b.c
@@ -38,6 +38,7 @@
 #include <nvgpu/regops.h>
 #include <nvgpu/gr/zbc.h>
 #include <nvgpu/gr/zcull.h>
+#include <nvgpu/gr/gr.h>
 #include <nvgpu/gr/gr_falcon.h>
 #include <nvgpu/gr/setup.h>
 
@@ -411,6 +412,7 @@ static const struct gpu_ops gm20b_ops = {
 			.program_zcull_mapping = gm20b_gr_program_zcull_mapping,
 		},
 		.init = {
+			.wait_initialized = nvgpu_gr_wait_initialized,
 			.ecc_scrub_reg = NULL,
 			.get_fbp_en_mask = gm20b_gr_init_get_fbp_en_mask,
 			.lg_coalesce = gm20b_gr_init_lg_coalesce,
diff --git a/drivers/gpu/nvgpu/gp10b/hal_gp10b.c b/drivers/gpu/nvgpu/gp10b/hal_gp10b.c
index 916820654..da49e201f 100644
--- a/drivers/gpu/nvgpu/gp10b/hal_gp10b.c
+++ b/drivers/gpu/nvgpu/gp10b/hal_gp10b.c
@@ -41,6 +41,7 @@
 #include <nvgpu/gr/gr_falcon.h>
 #include <nvgpu/gr/setup.h>
 #include <nvgpu/gr/fecs_trace.h>
+#include <nvgpu/gr/gr.h>
 
 #include "hal/mc/mc_gm20b.h"
 #include "hal/mc/mc_gp10b.h"
@@ -481,6 +482,7 @@ static const struct gpu_ops gp10b_ops = {
 			.program_zcull_mapping = gm20b_gr_program_zcull_mapping,
 		},
 		.init = {
+			.wait_initialized = nvgpu_gr_wait_initialized,
 			.ecc_scrub_reg = NULL,
 			.get_fbp_en_mask = gm20b_gr_init_get_fbp_en_mask,
 			.lg_coalesce = gm20b_gr_init_lg_coalesce,
diff --git a/drivers/gpu/nvgpu/gv100/hal_gv100.c b/drivers/gpu/nvgpu/gv100/hal_gv100.c
index 8791716dc..8178c4819 100644
--- a/drivers/gpu/nvgpu/gv100/hal_gv100.c
+++ b/drivers/gpu/nvgpu/gv100/hal_gv100.c
@@ -170,6 +170,7 @@
 #include <nvgpu/gr/gr_falcon.h>
 #include <nvgpu/gr/setup.h>
 #include <nvgpu/gr/fecs_trace.h>
+#include <nvgpu/gr/gr.h>
 
 #include <nvgpu/hw/gv100/hw_proj_gv100.h>
 #include <nvgpu/hw/gv100/hw_top_gv100.h>
@@ -622,6 +623,7 @@ static const struct gpu_ops gv100_ops = {
 				gv100_gr_hwpm_map_get_active_fbpa_mask,
 		},
 		.init = {
+			.wait_initialized = nvgpu_gr_wait_initialized,
 			.ecc_scrub_reg = NULL,
 			.get_fbp_en_mask = gm20b_gr_init_get_fbp_en_mask,
 			.lg_coalesce = gm20b_gr_init_lg_coalesce,
diff --git a/drivers/gpu/nvgpu/gv11b/hal_gv11b.c b/drivers/gpu/nvgpu/gv11b/hal_gv11b.c
index 0e919478b..e66a5ded7 100644
--- a/drivers/gpu/nvgpu/gv11b/hal_gv11b.c
+++ b/drivers/gpu/nvgpu/gv11b/hal_gv11b.c
@@ -145,6 +145,7 @@
 #include <nvgpu/gr/zcull.h>
 #include <nvgpu/gr/setup.h>
 #include <nvgpu/gr/fecs_trace.h>
+#include <nvgpu/gr/gr.h>
 
 #include <nvgpu/hw/gv11b/hw_proj_gv11b.h>
 #include <nvgpu/hw/gv11b/hw_top_gv11b.h>
@@ -580,6 +581,7 @@ static const struct gpu_ops gv11b_ops = {
 				gv100_gr_hwpm_map_align_regs_perf_pma,
 		},
 		.init = {
+			.wait_initialized = nvgpu_gr_wait_initialized,
 			.ecc_scrub_reg = gv11b_gr_init_ecc_scrub_reg,
 			.get_fbp_en_mask = gm20b_gr_init_get_fbp_en_mask,
 			.lg_coalesce = gm20b_gr_init_lg_coalesce,
diff --git a/drivers/gpu/nvgpu/include/nvgpu/gk20a.h b/drivers/gpu/nvgpu/include/nvgpu/gk20a.h
index 4a732492b..b51b4af72 100644
--- a/drivers/gpu/nvgpu/include/nvgpu/gk20a.h
+++ b/drivers/gpu/nvgpu/include/nvgpu/gk20a.h
@@ -680,6 +680,7 @@ struct gpu_ops {
 		} hwpm_map;
 
 		struct {
+			void (*wait_initialized)(struct gk20a *g);
 			void (*ecc_scrub_reg)(struct gk20a *g,
 				struct nvgpu_gr_config *gr_config);
 			u32 (*get_fbp_en_mask)(struct gk20a *g);
diff --git a/drivers/gpu/nvgpu/tu104/hal_tu104.c b/drivers/gpu/nvgpu/tu104/hal_tu104.c
index 8b9449b8b..f7d281108 100644
--- a/drivers/gpu/nvgpu/tu104/hal_tu104.c
+++ b/drivers/gpu/nvgpu/tu104/hal_tu104.c
@@ -189,6 +189,7 @@
 #include <nvgpu/gr/fecs_trace.h>
 #include <nvgpu/pmu/perf.h>
 #include <nvgpu/gr/gr_falcon.h>
+#include <nvgpu/gr/gr.h>
 
 #include <nvgpu/hw/tu104/hw_proj_tu104.h>
 #include <nvgpu/hw/tu104/hw_top_tu104.h>
@@ -650,6 +651,7 @@ static const struct gpu_ops tu104_ops = {
 				gv100_gr_hwpm_map_get_active_fbpa_mask,
 		},
 		.init = {
+			.wait_initialized = nvgpu_gr_wait_initialized,
 			.ecc_scrub_reg = NULL,
 			.get_fbp_en_mask = gm20b_gr_init_get_fbp_en_mask,
 			.lg_coalesce = gm20b_gr_init_lg_coalesce,