diff --git a/src/gallium/drivers/freedreno/a6xx/fd6_barrier.c b/src/gallium/drivers/freedreno/a6xx/fd6_barrier.c
index 176d3d537f5..bc13f7fdf27 100644
--- a/src/gallium/drivers/freedreno/a6xx/fd6_barrier.c
+++ b/src/gallium/drivers/freedreno/a6xx/fd6_barrier.c
@@ -64,7 +64,7 @@ event_write(struct fd_context *ctx, struct fd_ringbuffer *ring,
    return 0;
 }
 
-static void
+void
 fd6_emit_flushes(struct fd_context *ctx, struct fd_ringbuffer *ring,
                  unsigned flushes)
 {
diff --git a/src/gallium/drivers/freedreno/a6xx/fd6_barrier.h b/src/gallium/drivers/freedreno/a6xx/fd6_barrier.h
index 74217d851d3..7cdd17d835d 100644
--- a/src/gallium/drivers/freedreno/a6xx/fd6_barrier.h
+++ b/src/gallium/drivers/freedreno/a6xx/fd6_barrier.h
@@ -41,6 +41,9 @@ enum fd6_flush {
    FD6_WAIT_FOR_ME          = BIT(8),
 };
 
+void fd6_emit_flushes(struct fd_context *ctx, struct fd_ringbuffer *ring,
+                      unsigned flushes);
+
 void fd6_barrier_flush(struct fd_batch *batch) assert_dt;
 
 void fd6_barrier_init(struct pipe_context *pctx);
diff --git a/src/gallium/drivers/freedreno/a6xx/fd6_compute.c b/src/gallium/drivers/freedreno/a6xx/fd6_compute.c
index afb5754da8b..5d751cfe1b0 100644
--- a/src/gallium/drivers/freedreno/a6xx/fd6_compute.c
+++ b/src/gallium/drivers/freedreno/a6xx/fd6_compute.c
@@ -118,6 +118,9 @@ fd6_launch_grid(struct fd_context *ctx, const struct pipe_grid_info *info) in_dt
    if (!v)
       return;
 
+   if (ctx->batch->barrier)
+      fd6_barrier_flush(ctx->batch);
+
    if (ctx->dirty_shader[PIPE_SHADER_COMPUTE] & FD_DIRTY_SHADER_PROG)
       cs_program_emit(ctx, ring, v);
 
@@ -204,9 +207,6 @@ fd6_launch_grid(struct fd_context *ctx, const struct pipe_grid_info *info) in_dt
    OUT_RING(ring, 1); /* HLSQ_CS_KERNEL_GROUP_Y */
    OUT_RING(ring, 1); /* HLSQ_CS_KERNEL_GROUP_Z */
 
-   if (ctx->batch->barrier)
-      fd6_barrier_flush(ctx->batch);
-
    if (info->indirect) {
       struct fd_resource *rsc = fd_resource(info->indirect);
 
@@ -227,10 +227,6 @@ fd6_launch_grid(struct fd_context *ctx, const struct pipe_grid_info *info) in_dt
 
    trace_end_compute(&ctx->batch->trace, ring);
 
-   OUT_WFI5(ring);
-
-   fd6_cache_flush(ctx->batch, ring);
-
    fd_context_all_clean(ctx);
 }
 
diff --git a/src/gallium/drivers/freedreno/a6xx/fd6_gmem.c b/src/gallium/drivers/freedreno/a6xx/fd6_gmem.c
index 5542ff82f97..e6c0b36203a 100644
--- a/src/gallium/drivers/freedreno/a6xx/fd6_gmem.c
+++ b/src/gallium/drivers/freedreno/a6xx/fd6_gmem.c
@@ -40,6 +40,7 @@
 #include "freedreno_state.h"
 #include "freedreno_tracepoints.h"
 
+#include "fd6_barrier.h"
 #include "fd6_blitter.h"
 #include "fd6_context.h"
 #include "fd6_draw.h"
@@ -609,6 +610,8 @@ emit_common_fini(struct fd_batch *batch)
    struct fd_autotune *at = &batch->ctx->autotune;
    struct fd_batch_result *result = batch->autotune_result;
 
+   fd6_emit_flushes(batch->ctx, ring, batch->barrier);
+
    if (!result)
       return;