v3d: Fix predication with atomic image operations

[mesa.git] / src / broadcom / compiler / v3d40_tex.c

diff --git a/src/broadcom/compiler/v3d40_tex.c b/src/broadcom/compiler/v3d40_tex.c
index 9ee1692a77ce43d08338c59ed39bd7a041070aaa..379f90c146a27f465189c55fb9561b71bb6cd9e8 100644 (file)
--- a/src/broadcom/compiler/v3d40_tex.c
+++ b/src/broadcom/compiler/v3d40_tex.c
@@ -253,9 +253,11 @@ v3d40_image_load_store_tmu_op(nir_intrinsic_instr *instr)
         case nir_intrinsic_image_deref_atomic_add:
                 return v3d_get_op_for_atomic_add(instr, 3);
         case nir_intrinsic_image_deref_atomic_imin:
+                return V3D_TMU_OP_WRITE_SMIN;
         case nir_intrinsic_image_deref_atomic_umin:
                 return V3D_TMU_OP_WRITE_UMIN_FULL_L1_CLEAR;
         case nir_intrinsic_image_deref_atomic_imax:
+                return V3D_TMU_OP_WRITE_SMAX;
         case nir_intrinsic_image_deref_atomic_umax:
                 return V3D_TMU_OP_WRITE_UMAX;
         case nir_intrinsic_image_deref_atomic_and:
@@ -390,9 +392,21 @@ v3d40_vir_emit_image_load_store(struct v3d_compile *c,
                 }
         }
 
+        if (vir_in_nonuniform_control_flow(c) &&
+            instr->intrinsic != nir_intrinsic_image_deref_load) {
+           vir_set_pf(vir_MOV_dest(c, vir_nop_reg(), c->execute),
+                      V3D_QPU_PF_PUSHZ);
+        }
+
         vir_TMU_WRITE(c, V3D_QPU_WADDR_TMUSF, ntq_get_src(c, instr->src[1], 0),
                       &tmu_writes);
 
+        if (vir_in_nonuniform_control_flow(c) &&
+            instr->intrinsic != nir_intrinsic_image_deref_load) {
+           struct qinst *last_inst= (struct  qinst *)c->cur_block->instructions.prev;
+           vir_set_cond(last_inst, V3D_QPU_COND_IFA);
+        }
+
         vir_emit_thrsw(c);
 
         /* The input FIFO has 16 slots across all threads, so make sure we
@@ -408,4 +422,7 @@ v3d40_vir_emit_image_load_store(struct v3d_compile *c,
 
         if (nir_intrinsic_dest_components(instr) == 0)
                 vir_TMUWT(c);
+
+        if (instr->intrinsic != nir_intrinsic_image_deref_load)
+                c->tmu_dirty_rcl = true;
 }