[dataflow] Unified Systolic Array #232
Conversation
| for ri in range(Rtimes, name="row_loop"): | ||
| for ci in range(Ctimes, name="column_loop"): | ||
| # corner | ||
| with allo.meta_if(i in [0, Rt + 1] and j in [0, Ct + 1]): |
There was a problem hiding this comment.
This statement is a bit confusing. For example, for i in [0, Rt + 1], are you saying i is either 0 or Rt+1? In this case, it would be better to write it as a set (to be consistent with the math notation), i.e., i in {0, Rt+1}.
| r: int32 = in_R.get() | ||
| c: int32 = in_C.get() | ||
| # Core MAC | ||
| acti: int32 = r |
There was a problem hiding this comment.
Write the full name (activation) or use a proper shorthand (act).
| acti: int32 = r | ||
| weight: int32 = c if flowtag else s | ||
| psum: int32 = s if flowtag else c | ||
| accu: int32 = acti * weight + psum |
There was a problem hiding this comment.
Similarly, use acc or accumulation
| if hls.is_available("vitis_hls"): | ||
| gemm(A_flat, B_flat, insts, C_flat) | ||
| print(C_flat) | ||
| C_tru = np.dot(A, B).flatten() |
|
I have fixed the naming issues. Let me know if there are any problems. |
|
|
||
|
|
||
| @df.kernel(mapping=[P0, P1]) | ||
| def gemm(A: int32[M, K], B: int32[K, N], inst: int8[2], C: int32[M, N]): |
There was a problem hiding this comment.
Can you pull the latest commit (need to rebuild) and see whether a single bool inst now works for both LLVM and HLS backend? If so, you can use inst: bool.
There was a problem hiding this comment.
Sure. I'll try this later.
…rray Update to upstream.
|
The test of scalar interface looks good for csim. I've changed the interface to a single bool argument. |
Description
The implementation of unified systolic array to achieve GEMM with output-stationary or weight-stationary dataflow.
Examples
Numpy tests are included.
Checklist