diff options
| author | Jacob Bohlin <jacob.bohlin@arm.com> | 2021-06-15 14:06:23 +0200 |
|---|---|---|
| committer | Jacob Bohlin <jacob.bohlin@arm.com> | 2021-06-16 09:04:57 +0200 |
| commit | 4c00abb83487b3c8445f6db9140de44dec71efda (patch) | |
| tree | e3a68b3865fed12deaaec862819e5163c1dc9069 | |
| parent | 3e4168d741c167d2d52b1a3fc9a800c101bba09b (diff) | |
| download | ethos-u-vela-4c00abb83487b3c8445f6db9140de44dec71efda.tar.gz | |
MLBEDSW-4644 Removed unnecessary LUT DMA commands
Fixed a bug where a DMA command for the activation LUT would be issued
for every depth-slice of an operator. This caused multiple
unnecessary DMA commands.
Signed-off-by: Jacob Bohlin <jacob.bohlin@arm.com>
Change-Id: I9c291692d8002f05656bb88214836ab389a56cdb
| -rw-r--r-- | ethosu/vela/high_level_command_stream_generator.py | 10 |
1 files changed, 5 insertions, 5 deletions
diff --git a/ethosu/vela/high_level_command_stream_generator.py b/ethosu/vela/high_level_command_stream_generator.py index 5a838f88..6fcf80cb 100644 --- a/ethosu/vela/high_level_command_stream_generator.py +++ b/ethosu/vela/high_level_command_stream_generator.py @@ -128,6 +128,11 @@ def generate_high_level_commands_for_sched_op(sched_op, schedule): for start_width in range(ofm_start.width, ofm_end.width, ofm_step.width): end_width = min(start_width + ofm_step.width, ofm_end.width) + if parent_op.activation_lut: + lut_tensor = [tens for tens in parent_op.inputs if tens.purpose == TensorPurpose.LUT][0] + lut_box = Box([0] * len(lut_tensor.shape), list(lut_tensor.shape)) + yield from dma_if_necessary(sched_op.parent_ps, lut_box, lut_tensor) + for depth_idx, start_channel in enumerate(ofm_depth_slices[:-1]): start_channel = max(start_channel, ofm_start.depth) end_channel = min(ofm_depth_slices[depth_idx + 1], ofm_end.depth) @@ -196,11 +201,6 @@ def generate_high_level_commands_for_sched_op(sched_op, schedule): else: weight_box = None - if parent_op.activation_lut: - lut_tensor = [tens for tens in parent_op.inputs if tens.purpose == TensorPurpose.LUT][0] - lut_box = Box([0] * len(lut_tensor.shape), list(lut_tensor.shape)) - yield from dma_if_necessary(sched_op.parent_ps, lut_box, lut_tensor) - yield NpuStripe( sched_op.parent_ps, block_config.old_style_representation(), |