symbol not found (probably due to devectorization)

[ronghongbo]

In t2s/tests/performance/gemm/gemm.cpp

1. add another URE W(P)=matrixC(total_j, total_i), add another ImageParam matrixC.
2. Z(P) = select(kkk == 0 && kk == 0 && k == 0, W(P), ...
W.merge_ures(X, Y, Z, Out);
W.set_bounds(...);
W.space_time_transform(...);

Follow t2s/tests/performance/gemm/README to compile it on an FPGA emulator. We will see Internal Error at /home/u128292/t2sp/Halide/src/CodeGen_LLVM.cpp:1465 triggered by user code at : Symbol not found: W.s0.kkk

IR:

  .......
  Z.shreg.temp() = ... (float32)read_shift_reg("W.shreg", W.s0.jjj, W.s0.iii, W.s0.kkk) ...
  unrolled (W.s0.kkk, 0, 4) {... }

Note the operand 'read_shift_reg("W.shreg", W.s0.jjj, W.s0.iii, W.s0.kkk)' is before the kkk loop, but it refers to kkk. This looks like an issue with devectorization.