diff options
-rw-r--r-- | ethosu/vela/pass_packing.py | 19 | ||||
-rw-r--r-- | ethosu/vela/register_command_stream_generator.py | 17 |
2 files changed, 16 insertions, 20 deletions
diff --git a/ethosu/vela/pass_packing.py b/ethosu/vela/pass_packing.py index 8fb95f03..fab00e00 100644 --- a/ethosu/vela/pass_packing.py +++ b/ethosu/vela/pass_packing.py @@ -259,20 +259,6 @@ def pack_into_passes(nng, arch, verbose_packing=False): ofm_tensor = op.outputs[0] build_pass((op,), ofm_tensor) - def broadcast_input_check(ps): - if len(ps.inputs) == 1 or ps.inputs[0].shape == ps.inputs[1].shape: - return - - if ps.inputs[0].shape == [] or ps.inputs[1].shape == []: - return - - for idx in range(len(ps.inputs[1].shape)): - if ps.inputs[1].shape[idx] != ps.inputs[0].shape[idx] and ps.inputs[0].shape[idx] != 1: - return - - ps.inputs[0], ps.inputs[1] = ps.inputs[1], ps.inputs[0] - ps.primary_op.inputs[0], ps.primary_op.inputs[1] = ps.primary_op.inputs[1], ps.primary_op.inputs[0] - def build_pass(start_ops_to_process, ofm_tensor=None): reverse_ops_list = [] curr_flags = PassFlags.Empty @@ -413,14 +399,9 @@ def pack_into_passes(nng, arch, verbose_packing=False): ps.inputs = ordered_input_list ps.intermediates = intermediates ps.outputs = list(ops_list[-1].outputs) - ps.ifm_tensor = ifm_tensor # ElementWise operation, 2 IFMs if ps.primary_op and ps.primary_op.type in binary_elem_wise_main_ops: - # Swap broadcast input if applicable - broadcast_input_check(ps) - - # If only 1 input, IFM and IFM2 will be the same tensor ps.ifm_tensor = ps.inputs[0] ps.ifm2_tensor = ps.inputs[-1] diff --git a/ethosu/vela/register_command_stream_generator.py b/ethosu/vela/register_command_stream_generator.py index e0f114eb..471953d9 100644 --- a/ethosu/vela/register_command_stream_generator.py +++ b/ethosu/vela/register_command_stream_generator.py @@ -330,6 +330,21 @@ def get_op_padding_lt(cmd): return (explicit_padding[1], explicit_padding[0]) +def ifm_ifm2_correct_order(ifm_shape, ifm2_shape): + if ifm_shape == []: + # Scalar needs to be in IFM2 + return False + elif ifm2_shape == []: + return True + + for ifm, ifm2 in zip(ifm_shape, ifm2_shape): + if ifm != ifm2 and ifm == 1: + # Broadcasted FM needs to be in IFM2 + return False + + return True + + def generate_register_command_stream(nng, sg, arch, verbose=False): emit = CommandStreamEmitter() @@ -472,7 +487,7 @@ def generate_register_command_stream(nng, sg, arch, verbose=False): IFM2Broadcast.ReverseOperandOrder if primary_op.attrs.get("reverse_op_order", False) else 0 ) - if cmd.ifm_tensor.shape == []: + if not ifm_ifm2_correct_order(cmd.ifm_tensor.shape, cmd.ifm2_tensor.shape): # The scalar has to be the ifm2 tensor so switch the ifms cmd.ifm_tensor, cmd.ifm2_tensor = cmd.ifm2_tensor, cmd.ifm_tensor cmd.ifm_box, cmd.ifm2_box = cmd.ifm2_box, cmd.ifm_box |