diff options
| author | Dwight Lidman <dwight.lidman@arm.com> | 2020-05-13 12:00:08 +0200 |
|---|---|---|
| committer | Tim Hall <tim.hall@arm.com> | 2020-06-18 17:53:52 +0100 |
| commit | a9390f7fbd35dca75e80710835f67bb1d75d7c93 (patch) | |
| tree | df3606c28425e3a3504b29fe79ac8ddd0e2b8661 /ethosu/vela/architecture_features.py | |
| parent | 388e9c230898385df59e6175aa45012e5864c09a (diff) | |
| download | ethos-u-vela-a9390f7fbd35dca75e80710835f67bb1d75d7c93.tar.gz | |
MLBEDSW-1971: Verify ifm block size calculation against specification
This commit ensures the IFM block size calculation
in architecture_features.py matches the specification
by correctly setting the ifm upscaling factor based on
the upscaling mode.
This requires adding an attribute to the Tensor object
which stores the upscaling mode for that specific
tensor and making sure that information is correctly
carried over to shared_buffer_allocation.py.
Signed-off-by: Dwight Lidman <dwight.lidman@arm.com>
Change-Id: I4ab56086f4c694d3bf759bbad30cdb969b4a26db
Diffstat (limited to 'ethosu/vela/architecture_features.py')
| -rw-r--r-- | ethosu/vela/architecture_features.py | 9 |
1 files changed, 5 insertions, 4 deletions
diff --git a/ethosu/vela/architecture_features.py b/ethosu/vela/architecture_features.py index 1bf9d950..b59122ea 100644 --- a/ethosu/vela/architecture_features.py +++ b/ethosu/vela/architecture_features.py @@ -22,6 +22,7 @@ from configparser import ConfigParser import numpy as np from .errors import OptionError +from .ethos_u55_regs.ethos_u55_regs import resampling_mode from .numeric_util import round_up from .numeric_util import round_up_divide from .operation import NpuBlockType @@ -347,10 +348,10 @@ Note the difference between ArchitectureFeatures and CompilerOptions return min(max_block_depth, ifm_depth) # Calculate the size of the IFM block given a depth, target OFM block and a kernel - def get_ifm_block_size( - self, ifm_block_depth, ofm_block: Block, kernel: Kernel, subkernel: Block = Block(8, 8, 65536) - ): - upscaling = 1 + def get_ifm_block_size(self, ifm_block_depth, ofm_block: Block, + kernel: Kernel, subkernel: Block = Block(8, 8, 65536), + 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 |