diff options
| author | erik.andersson@arm.com <erik.andersson@arm.com> | 2021-04-14 13:31:05 +0200 |
|---|---|---|
| committer | erik.andersson@arm.com <erik.andersson@arm.com> | 2021-04-14 14:11:42 +0200 |
| commit | 1d6d5c47c2000facc377620a64084738339ccda9 (patch) | |
| tree | e1ec489f9c2270dc1788b890a5d317004da6ee13 | |
| parent | 1c772e43434bd39aad1cb2e0036dcdf67b038633 (diff) | |
| download | ethos-u-vela-1d6d5c47c2000facc377620a64084738339ccda9.tar.gz | |
MLBEDSW-4103: Block config calc update
The previous calculation of the IFM block height and width
yielded incorrect block configs when running transpose_conv
networks with certain hardware constraints.
Signed-off-by: erik.andersson@arm.com <erik.andersson@arm.com>
Change-Id: I8b6936a3e8c37da640bdeac84ecfea8363f910f9
| -rw-r--r-- | ethosu/vela/architecture_features.py | 24 |
1 files changed, 8 insertions, 16 deletions
diff --git a/ethosu/vela/architecture_features.py b/ethosu/vela/architecture_features.py index c6ca9cdb..43e82655 100644 --- a/ethosu/vela/architecture_features.py +++ b/ethosu/vela/architecture_features.py @@ -28,6 +28,7 @@ from .ethos_u55_regs.ethos_u55_regs import resampling_mode from .numeric_util import full_shape from .numeric_util import round_up from .numeric_util import round_up_divide +from .numeric_util import round_up_to_int from .operation import Kernel from .operation import NpuBlockType from .operation import PointXYZ @@ -426,27 +427,18 @@ class ArchitectureFeatures: ifm_resampling_mode=resampling_mode.NONE, ): upscaling = 1 if ifm_resampling_mode == resampling_mode.NONE else 2 + # Height - ifm_odd_2x_height_enable = 0 dilated_kernel_height = ((kernel.height - 1) * kernel.dilation.y) + 1 - ifm_block_height = ( - (ofm_block.height - 1) * kernel.stride.y - + min(subkernel.height, dilated_kernel_height) - + ifm_odd_2x_height_enable - ) // upscaling - - ifm_block_height = round_up(ifm_block_height, self.ofm_ublock.height) + ifm_block_height = round_up_to_int( + ((ofm_block.height - 1) * kernel.stride.y + min(subkernel.height, dilated_kernel_height)) / upscaling + ) # Width - ifm_odd_2x_width_enable = 0 dilated_kernel_width = ((kernel.width - 1) * kernel.dilation.x) + 1 - ifm_block_width = ( - (ofm_block.width - 1) * kernel.stride.x - + min(subkernel.width, dilated_kernel_width) - + ifm_odd_2x_width_enable - ) // upscaling - - ifm_block_width = round_up(ifm_block_width, self.ofm_ublock.width) + ifm_block_width = round_up_to_int( + ((ofm_block.width - 1) * kernel.stride.x + min(subkernel.width, dilated_kernel_width)) / upscaling + ) return Block(ifm_block_width, ifm_block_height, ifm_block_depth) |