This is an automated email from the ASF dual-hosted git repository.
leandron pushed a commit to branch main
in repository https://gitbox.apache.org/repos/asf/tvm.git
The following commit(s) were added to refs/heads/main by this push:
new 1b3990808e [microNPU] Update existing microNPU tutorial for CMSIS-NN
(#11285)
1b3990808e is described below
commit 1b3990808ec11c0255609a62a2dfc37df9152172
Author: Grant Watson <[email protected]>
AuthorDate: Thu May 12 17:06:55 2022 +0100
[microNPU] Update existing microNPU tutorial for CMSIS-NN (#11285)
* [microNPU] Update existing microNPU tutorial for CMSIS-NN
* Added instructions to existing microNPU tutorial indicating how to
offload operators to CMSIS-NN.
Change-Id: I9faef1d92a2107e04cfc21b7bfd1b72dc1bd5489
* [microNPU] Update existing microNPU tutorial for CMSIS-NN
* Reformat micro_ethosu.py with black
Change-Id: Id189f333b5bd891232781d4eb58522a240146c95
---
gallery/how_to/work_with_microtvm/micro_ethosu.py | 41 ++++++++++++++++++-----
1 file changed, 32 insertions(+), 9 deletions(-)
diff --git a/gallery/how_to/work_with_microtvm/micro_ethosu.py
b/gallery/how_to/work_with_microtvm/micro_ethosu.py
index ef34c7d7cc..f55fad71dd 100644
--- a/gallery/how_to/work_with_microtvm/micro_ethosu.py
+++ b/gallery/how_to/work_with_microtvm/micro_ethosu.py
@@ -15,15 +15,16 @@
# specific language governing permissions and limitations
# under the License.
"""
-Running TVM on bare metal Arm(R) Cortex(R)-M55 CPU and Ethos(TM)-U55 NPU
-========================================================================
+Running TVM on bare metal Arm(R) Cortex(R)-M55 CPU and Ethos(TM)-U55 NPU with
CMSIS-NN
+======================================================================================
**Author**:
`Grant Watson <https://github.com/grant-arm>`_
This section contains an example of how to use TVM to run a model
-on an Arm(R) Cortex(R)-M55 CPU and Ethos(TM)-U55 NPU, using bare metal.
+on an Arm(R) Cortex(R)-M55 CPU and Ethos(TM)-U55 NPU with CMSIS-NN, using bare
metal.
The Cortex(R)-M55 is a small, low-power CPU designed for use in embedded
-devices. The Ethos(TM)-U55 is a microNPU, specifically designed to accelerate
+devices. CMSIS-NN is a collection of kernels optimized for Arm(R) Cortex(R)-M
CPUs.
+The Ethos(TM)-U55 is a microNPU, specifically designed to accelerate
ML inference in resource-constrained embedded devices.
In order to run the demo application without having access to a Cortex(R)-M55
@@ -121,9 +122,9 @@ TVM to offload operators to the Ethos(TM)-U55 where
possible.
# tar xvf mobilenet_v1_1.0_224_quant.tar
#
-################################################################################
-# Compiling the model for Arm(R) Cortex(R)-M55 CPU and Ethos(TM)-U55 NPU
-# ----------------------------------------------------------------------
+######################################################################################
+# Compiling the model for Arm(R) Cortex(R)-M55 CPU and Ethos(TM)-U55 NPU with
CMSIS-NN
+#
------------------------------------------------------------------------------------
#
# Once we've downloaded the MobileNet v1 model, the next step is to compile it.
# To accomplish that, we are going to use ``tvmc compile``. The output we get
from
@@ -133,12 +134,17 @@ TVM to offload operators to the Ethos(TM)-U55 where
possible.
#
# .. code-block:: bash
#
-# tvmc compile --target="ethos-u -accelerator_config=ethos-u55-256, c" \
+# tvmc compile --target=ethos-u,cmsis-nn,c \
+# --target-ethos-u-accelerator_config=ethos-u55-256 \
+# --target-cmsis-nn-mcpu=cortex-m55 \
# --target-c-mcpu=cortex-m55 \
# --runtime=crt \
# --executor=aot \
# --executor-aot-interface-api=c \
# --executor-aot-unpacked-api=1 \
+# --pass-config tir.usmp.enable=1 \
+# --pass-config tir.usmp.algorithm=hill_climb \
+# --pass-config tir.disable_storage_rewrite=1 \
# --pass-config tir.disable_vectorize=1 \
# ./mobilenet_v1_1.0_224_quant.tflite \
# --output-format=mlf
@@ -147,7 +153,9 @@ TVM to offload operators to the Ethos(TM)-U55 where
possible.
################################################################################
# .. note:: Explanation of tvmc compile arguments:
#
-# * ``--target="ethos-u -accelerator_config=ethos-u55-256, c"`` : offload
operators to the Ethos(TM)-U55 NPU where possible and fall back to using
generated C code on the Cortex(R)-M where an operator is not supported on the
NPU..
+# * ``--target=ethos-u,cmsis-nn,c`` : offload operators to the microNPU
where possible, falling back to CMSIS-NN and finally generated C code where an
operator is not supported on the microNPU..
+#
+# * ``--target-ethos-u-accelerator_config=ethos-u55-256`` : specifies the
microNPU configuration
#
# * ``--target-c-mcpu=cortex-m55`` : Cross-compile for the Cortex(R)-M55.
#
@@ -159,6 +167,12 @@ TVM to offload operators to the Ethos(TM)-U55 where
possible.
#
# * ``--executor-aot-unpacked-api=1`` : Use the unpacked API internally.
#
+# * ``--pass-config tir.usmp.enable=1`` : Enable Unified Static Memory
Planning
+#
+# * ``--pass-config tir.usmp.algorithm=hill_climb`` : Use the hill-climb
algorithm for USMP
+#
+# * ``--pass-config tir.disable_storage_rewrite=1`` : Disable storage rewrite
+#
# * ``--pass-config tir.disable_vectorize=1`` : Disable vectorize since
there are no standard vectorized types in C.
#
# * ``./mobilenet_v1_1.0_224_quant.tflite`` : The TFLite model that is being
compiled.
@@ -166,6 +180,15 @@ TVM to offload operators to the Ethos(TM)-U55 where
possible.
# * ``--output-format=mlf`` : Output should be generated in the Model
Library Format.
#
+################################################################################
+# .. note:: If you don't want to make use of the microNPU and want to offload
+# operators to CMSIS-NN only:
+#
+# * Use ``--target=cmsis-nn,c`` in place of ``--target=ethos-u,cmsis-nn,c``
+#
+# * Remove the microNPU config parameter
``--target-ethos-u-accelerator_config=ethos-u55-256``
+#
+
################################################################################
# Extracting the generated code into the current directory
# --------------------------------------------------------
