diff --git a/src/amd/compiler/aco_lower_to_hw_instr.cpp b/src/amd/compiler/aco_lower_to_hw_instr.cpp
index 3065f283549..592c3868e20 100644
--- a/src/amd/compiler/aco_lower_to_hw_instr.cpp
+++ b/src/amd/compiler/aco_lower_to_hw_instr.cpp
@@ -545,11 +545,44 @@ void emit_reduction(lower_context *ctx, aco_opcode op, ReduceOp reduce_op, unsig
             }
          }
          std::swap(tmp, vtmp);
-      } else {
+      } else if (ctx->program->chip_class >= GFX8) {
          emit_dpp_mov(ctx, tmp, tmp, src.size(), dpp_wf_sr1, 0xf, 0xf, true);
+      } else {
+         // TODO: use LDS on CS with a single write and shifted read
+         /* wavefront shift_right by 1 on SI/CI */
+         emit_ds_swizzle(bld, vtmp, tmp, src.size(), (1 << 15) | dpp_quad_perm(0, 0, 1, 2));
+         emit_ds_swizzle(bld, tmp, tmp, src.size(), ds_pattern_bitmode(0x1F, 0x00, 0x07)); /* mirror(8) */
+         bld.sop1(aco_opcode::s_mov_b32, Definition(exec_lo, s1), Operand(0x10101010u));
+         bld.sop1(aco_opcode::s_mov_b32, Definition(exec_hi, s1), Operand(exec_lo, s1));
+         for (unsigned i = 0; i < src.size(); i++)
+            bld.vop1(aco_opcode::v_mov_b32, Definition(PhysReg{vtmp+i}, v1), Operand(PhysReg{tmp+i}, v1));
+
+         bld.sop1(aco_opcode::s_mov_b64, Definition(exec, s2), Operand(UINT64_MAX));
+         emit_ds_swizzle(bld, tmp, tmp, src.size(), ds_pattern_bitmode(0x1F, 0x00, 0x08)); /* swap(8) */
+         bld.sop1(aco_opcode::s_mov_b32, Definition(exec_lo, s1), Operand(0x01000100u));
+         bld.sop1(aco_opcode::s_mov_b32, Definition(exec_hi, s1), Operand(exec_lo, s1));
+         for (unsigned i = 0; i < src.size(); i++)
+            bld.vop1(aco_opcode::v_mov_b32, Definition(PhysReg{vtmp+i}, v1), Operand(PhysReg{tmp+i}, v1));
+
+         bld.sop1(aco_opcode::s_mov_b64, Definition(exec, s2), Operand(UINT64_MAX));
+         emit_ds_swizzle(bld, tmp, tmp, src.size(), ds_pattern_bitmode(0x1F, 0x00, 0x10)); /* swap(16) */
+         bld.sop2(aco_opcode::s_bfm_b32, Definition(exec_lo, s1), Operand(1u), Operand(16u));
+         bld.sop2(aco_opcode::s_bfm_b32, Definition(exec_hi, s1), Operand(1u), Operand(16u));
+         for (unsigned i = 0; i < src.size(); i++)
+            bld.vop1(aco_opcode::v_mov_b32, Definition(PhysReg{vtmp+i}, v1), Operand(PhysReg{tmp+i}, v1));
+
+         bld.sop1(aco_opcode::s_mov_b64, Definition(exec, s2), Operand(UINT64_MAX));
+         for (unsigned i = 0; i < src.size(); i++) {
+            bld.writelane(Definition(PhysReg{vtmp+i}, v1), identity[i], Operand(0u), Operand(PhysReg{vtmp+i}, v1));
+            bld.readlane(Definition(PhysReg{sitmp+i}, s1), Operand(PhysReg{tmp+i}, v1), Operand(0u));
+            bld.writelane(Definition(PhysReg{vtmp+i}, v1), Operand(PhysReg{sitmp+i}, s1), Operand(32u), Operand(PhysReg{vtmp+i}, v1));
+            identity[i] = Operand(0u); /* prevent further uses of identity */
+         }
+         std::swap(tmp, vtmp);
       }
+
       for (unsigned i = 0; i < src.size(); i++) {
-         if (!identity[i].isConstant() || identity[i].constantValue()) { /* bound_ctrl should take case of this overwise */
+         if (!identity[i].isConstant() || identity[i].constantValue()) { /* bound_ctrl should take care of this overwise */
             if (ctx->program->chip_class < GFX10)
                assert((identity[i].isConstant() && !identity[i].isLiteral()) || identity[i].physReg() == PhysReg{sitmp+i});
             bld.writelane(Definition(PhysReg{tmp+i}, v1), identity[i], Operand(0u), Operand(PhysReg{tmp+i}, v1));