diff --git a/src/broadcom/compiler/qpu_schedule.c b/src/broadcom/compiler/qpu_schedule.c
index 9f11fe27eb8..9fe842c3e84 100644
--- a/src/broadcom/compiler/qpu_schedule.c
+++ b/src/broadcom/compiler/qpu_schedule.c
@@ -658,8 +658,8 @@ qpu_compatible_peripheral_access(const struct v3d_device_info *devinfo,
         /* V3D 4.1 and later allow TMU read along with a VPM read or write, and
          * WRTMUC with a TMU magic register write (other than tmuc).
          */
-        if ((a->sig.ldtmu && v3d_qpu_uses_vpm(b)) ||
-            (b->sig.ldtmu && v3d_qpu_uses_vpm(a))) {
+        if ((a->sig.ldtmu && v3d_qpu_reads_or_writes_vpm(b)) ||
+            (b->sig.ldtmu && v3d_qpu_reads_or_writes_vpm(a))) {
                 return true;
         }
 
diff --git a/src/broadcom/qpu/qpu_instr.c b/src/broadcom/qpu/qpu_instr.c
index 09d06b3fa3e..5c86daecdf7 100644
--- a/src/broadcom/qpu/qpu_instr.c
+++ b/src/broadcom/qpu/qpu_instr.c
@@ -591,7 +591,6 @@ v3d_qpu_add_op_reads_vpm(enum  v3d_qpu_add_op op)
 {
         switch (op) {
         case V3D_QPU_A_VPMSETUP:
-        case V3D_QPU_A_VPMWT:
         case V3D_QPU_A_LDVPMV_IN:
         case V3D_QPU_A_LDVPMV_OUT:
         case V3D_QPU_A_LDVPMD_IN:
@@ -610,7 +609,6 @@ v3d_qpu_add_op_writes_vpm(enum  v3d_qpu_add_op op)
 {
         switch (op) {
         case V3D_QPU_A_VPMSETUP:
-        case V3D_QPU_A_VPMWT:
         case V3D_QPU_A_STVPMV:
         case V3D_QPU_A_STVPMD:
         case V3D_QPU_A_STVPMP:
@@ -737,10 +735,25 @@ v3d_qpu_writes_vpm(const struct v3d_qpu_instr *inst)
         return false;
 }
 
+static bool
+v3d_qpu_waits_vpm(const struct v3d_qpu_instr *inst)
+{
+        return inst->type == V3D_QPU_INSTR_TYPE_ALU &&
+               inst->alu.add.op == V3D_QPU_A_VPMWT;
+}
+
+bool
+v3d_qpu_reads_or_writes_vpm(const struct v3d_qpu_instr *inst)
+{
+        return v3d_qpu_reads_vpm(inst) || v3d_qpu_writes_vpm(inst);
+}
+
 bool
 v3d_qpu_uses_vpm(const struct v3d_qpu_instr *inst)
 {
-        return v3d_qpu_reads_vpm(inst) || v3d_qpu_writes_vpm(inst);
+        return v3d_qpu_reads_vpm(inst) ||
+               v3d_qpu_writes_vpm(inst) ||
+               v3d_qpu_waits_vpm(inst);
 }
 
 bool
diff --git a/src/broadcom/qpu/qpu_instr.h b/src/broadcom/qpu/qpu_instr.h
index ad2d37b6051..76329c8864b 100644
--- a/src/broadcom/qpu/qpu_instr.h
+++ b/src/broadcom/qpu/qpu_instr.h
@@ -462,6 +462,7 @@ bool v3d_qpu_uses_mux(const struct v3d_qpu_instr *inst, enum v3d_qpu_mux mux);
 bool v3d_qpu_uses_vpm(const struct v3d_qpu_instr *inst) ATTRIBUTE_CONST;
 bool v3d_qpu_reads_vpm(const struct v3d_qpu_instr *inst) ATTRIBUTE_CONST;
 bool v3d_qpu_writes_vpm(const struct v3d_qpu_instr *inst) ATTRIBUTE_CONST;
+bool v3d_qpu_reads_or_writes_vpm(const struct v3d_qpu_instr *inst) ATTRIBUTE_CONST;
 bool v3d_qpu_reads_flags(const struct v3d_qpu_instr *inst) ATTRIBUTE_CONST;
 bool v3d_qpu_writes_flags(const struct v3d_qpu_instr *inst) ATTRIBUTE_CONST;
 bool v3d_qpu_sig_writes_address(const struct v3d_device_info *devinfo,