diff --git a/src/panfrost/bifrost/valhall/test/test-insert-flow.cpp b/src/panfrost/bifrost/valhall/test/test-insert-flow.cpp
index 410fe4ec1b6..5ee55f1fb1d 100644
--- a/src/panfrost/bifrost/valhall/test/test-insert-flow.cpp
+++ b/src/panfrost/bifrost/valhall/test/test-insert-flow.cpp
@@ -104,12 +104,13 @@ TEST_F(InsertFlow, TilebufferWait7) {
    });
 }
 
-TEST_F(InsertFlow, AtestWait6) {
+TEST_F(InsertFlow, AtestWait6AndWait0After) {
    CASE(FRAGMENT, {
         flow(DISCARD);
         bi_fadd_f32_to(b, bi_register(0), bi_register(0), bi_register(0));
         flow(WAIT0126);
         bi_atest_to(b, bi_register(0), bi_register(4), bi_register(5));
+        flow(WAIT0);
         flow(END);
    });
 }
diff --git a/src/panfrost/bifrost/valhall/va_insert_flow.c b/src/panfrost/bifrost/valhall/va_insert_flow.c
index 9d666821e1b..0a2e886f518 100644
--- a/src/panfrost/bifrost/valhall/va_insert_flow.c
+++ b/src/panfrost/bifrost/valhall/va_insert_flow.c
@@ -429,6 +429,10 @@ va_insert_flow_control_nops(bi_context *ctx)
          /* Insert waits for tilebuffer and depth/stencil instructions. These
           * only happen in regular fragment shaders, as the required waits are
           * assumed to already have happened in blend shaders.
+          *
+          * For discarded thread handling, ATEST must be serialized against all
+          * other asynchronous instructions and should be serialized against all
+          * instructions. Wait for slot 0 immediately after the ATEST.
           */
          case BI_OPCODE_BLEND:
          case BI_OPCODE_LD_TILE:
@@ -437,6 +441,9 @@ va_insert_flow_control_nops(bi_context *ctx)
                bi_flow(ctx, bi_before_instr(I), VA_FLOW_WAIT);
             break;
          case BI_OPCODE_ATEST:
+            bi_flow(ctx, bi_before_instr(I), VA_FLOW_WAIT0126);
+            bi_flow(ctx, bi_after_instr(I), VA_FLOW_WAIT0);
+            break;
          case BI_OPCODE_ZS_EMIT:
             if (!ctx->inputs->is_blend)
                bi_flow(ctx, bi_before_instr(I), VA_FLOW_WAIT0126);