cornell-zhang/heterocl

Inter-stage streaming cannot pass synthesis

Closed this issue · 4 comments

chhzh123 commented

This program uses .to to perform inter-stage data streaming.

def test():
    A = hcl.placeholder((10,), "A")

    def kernel(A):
        B = hcl.compute(A.shape, lambda i: A[i] + 1, "B")
        C = hcl.compute(B.shape, lambda i: A[i] + B[i], "C")
        return C

    target = hcl.platform.zc706
    target.config(compile="vivado_hls", mode="csyn")
    s = hcl.create_schedule([A], kernel)
    s.to([A], target.xcel)
    s.to(kernel.C, target.host)
    s.to(kernel.B, s[kernel.C]) # inter-stage streaming
    f = hcl.build(s, target)
    np_A = np.zeros((10,))
    np_C = np.zeros((10,))
    hcl_A = hcl.asarray(np_A)
    hcl_C = hcl.asarray(np_C)
    f(hcl_A, hcl_C)

The generated code is shown below, which do not specify the dataflow region or the FIFO size, causing synthesis errors.

  void test(hls::stream<bit32 >& A_channel, hls::stream<bit32 >& C_channel) {
  #pragma HLS INTERFACE axis port=A_channel offset=slave bundle=gmem0
  #pragma HLS INTERFACE axis port=C_channel offset=slave bundle=gmem1
  #pragma HLS INTERFACE s_axilite port=return bundle=control
    bit32 C[10];
    bit32 A[10];
    for (bit32 A0 = 0; A0 < 10; ++A0) {
      A[A0] = A_channel.read();
    }
    bit32 B[10];
    hls::stream<bit32 > B_pipe1;
    for (bit32 i = 0; i < 10; ++i) {
      B_pipe1.write((A[i] + 1));
    }
    for (bit32 i1 = 0; i1 < 10; ++i1) {
      C[i1] = ((bit32)(((ap_int<33>)A[i1]) + ((ap_int<33>)B_pipe1.read())));
    }
    for (bit32 C0 = 0; C0 < 10; ++C0) {
      C_channel.write(C[C0]);
    }
  }
ERROR: [XFORM 203-733] An internal stream 'B_pipe1.V.V' (kernel.cpp:22) with default size is used in a non-dataflow region, which may result in deadlock. Please consider to resize the stream using the directive 'set_directive_stream' or the 'HLS stream' pragma.
zhangzhiru commented

This underscores the importance of doing automatic inference of the FIFO sizing.
Can we manually specify the FIFO size in HeteroCL to workaround this issue?

chhzh123 commented

Yes, I am just trying to tackle this problem, and this program is a test example. I discussed with @hecmay about the implementation of .to yesterday. Seems there're still several issues need to be fixed.

chhzh123 commented

Two more issues here:

  1. s.to(A, s[kernel.B]) may cause Segmentation Fault.
  2. Consecutive data streaming cannot work correctly. Stream buffer for Stage C cannot be generated properly in the following code.
s.to(kernel.B, s[kernel.C])
s.to(kernel.C, s[kernel.D])