SBUF Data Dependency Issue

[MBshara]

Hello, I was loading a tensor into SBUF in this tiling fashion but was encountering some issues where only the last "out_" index returned the correctly loaded values in "weights_mat_mul_2":
I got no compiler issue and yet, since I am loading in sequentially, I expected w_temp to always load correct data into weights_mat_mul_2, as it w_temp was overwritten after every iteration.

weights_mat_mul_2 = nl.ndarray((filter_height, filter_width, n_tiles_c_out, nl.par_dim(c_in_pmax), n_tiles_c_in, c_in_pmax), dtype=W.dtype, buffer=nl.sbuf)
w_temp = nl.ndarray((nl.par_dim(c_in_pmax), n_tiles_c_in, c_in_pmax, filter_height, filter_width), dtype=W.dtype, buffer=nl.sbuf)
for out_ in nl.sequential_range(n_tiles_c_out):
    w_temp[...] = nl.load(W_reshaped[out_,:,:,:,:,:])
    for fH in nl.affine_range(filter_height):
        for fW in nl.affine_range(filter_width):
            for in_ in nl.affine_range(n_tiles_c_in):
                weights_mat_mul_2[fH,fW,out_,:,in_,:] = w_temp[:,in_,:,fH,fW]

I then tested simulation and every tile in weights_mat_mul_2 contained the correct values. I then tried moving the array "w_temp" declaration into the outermost for loop and all problems were resolved:

weights_mat_mul_2 = nl.ndarray((filter_height, filter_width, n_tiles_c_out, nl.par_dim(c_in_pmax), n_tiles_c_in, c_in_pmax), dtype=W.dtype, buffer=nl.sbuf)
for out_ in nl.sequential_range(n_tiles_c_out):
    w_temp = nl.ndarray((nl.par_dim(c_in_pmax), n_tiles_c_in, c_in_pmax, filter_height, filter_width), dtype=W.dtype, buffer=nl.sbuf)
    w_temp[...] = nl.load(W_reshaped[out_,:,:,:,:,:])
    for fH in nl.affine_range(filter_height):
        for fW in nl.affine_range(filter_width):
            for in_ in nl.affine_range(n_tiles_c_in):
                weights_mat_mul_2[fH,fW,out_,:,in_,:] = w_temp[:,in_,:,fH,fW]

[fayyadd]

Thanks for reaching out, we are looking into the issue