Added support for cost estimation for upsampling (#37)

* Update AUTHORS.rst

* Create python-publish.yml

* Add ZCU111 board to part map (#32)

* Update AUTHORS.rst

* GHA for PyPI, sdist only (#35)

* finn-base v0.0.2 (#34)

* Modified set_nodeattr to allow using it on repeated fields (#18)

* [base]: changed how the floats, ints, strings, tensors, graphs and sparse_tensors field of AttributeProto is set.

* [Core] restrict attributes to tested types

Co-authored-by: Yaman Umuroglu <[email protected]>

* Support for non-square input images and kernels for im2col node (#20)

* [im2col.py]: support for non-square input images and kernels, [test_im2col]: added/modified several test cases for (non-)square images and kernels

* [im2col]: support for non-square input images and kernels, [test_im2col]: added/modified several test cases for (non-)square images and kernels

* [test_general_transformation]: changed kernel_size attribute to list instead of integer as required by im2col node

* [base]: changed how the "ints" field of AttributeProto set.
[test_general_transformation]: changed the type of the kernel_size attribute to list of integers

* removed unused import

* [base]: added support for writing repeated fields in AttributeProto

* minor style changes

* [im2col, test_im2col]: added support for non-equal padding

* [lower_convs_to_matmul]: added support for non-square input images and kernels and non-equal padding.
[test_conv_lowering]: added/modified test cases for non-equal padding, depthwise convolution and 'standard' convolution.

* [test_conv_lowering]: included 1D depthwise and regular convolutions in tests

* Revert "[test_conv_lowering]: included 1D depthwise and regular convolutions in tests"

This reverts commit 3ff449c

* Revert "[lower_convs_to_matmul]: added support for non-square input images and kernels and non-equal padding."

This reverts commit 15e34ed.

* Revert "[im2col, test_im2col]: added support for non-equal padding"

This reverts commit c524020.

* [im2col]: changed how a square kernel is instantiated.
[lower_convs_to_matmul]: changed how the kernel size attribute is read (based on how a square kernel is instantiated).

* [im2col]: minor change in style.

* [Im2Col] style fixes and comments

Co-authored-by: Yaman Umuroglu <[email protected]>

* Update AUTHORS.rst

* Support for non-square input images and kernels in LowerConvsToMatMul transformation (#16)

* [im2col.py]: support for non-square input images and kernels, [test_im2col]: added/modified several test cases for (non-)square images and kernels

* [im2col]: support for non-square input images and kernels, [test_im2col]: added/modified several test cases for (non-)square images and kernels

* [test_general_transformation]: changed kernel_size attribute to list instead of integer as required by im2col node

* [base]: changed how the "ints" field of AttributeProto set.
[test_general_transformation]: changed the type of the kernel_size attribute to list of integers

* removed unused import

* [base]: added support for writing repeated fields in AttributeProto

* minor style changes

* [im2col, test_im2col]: added support for non-equal padding

* [lower_convs_to_matmul]: added support for non-square input images and kernels and non-equal padding.
[test_conv_lowering]: added/modified test cases for non-equal padding, depthwise convolution and 'standard' convolution.

* [test_conv_lowering]: included 1D depthwise and regular convolutions in tests

* Revert "[test_conv_lowering]: included 1D depthwise and regular convolutions in tests"

This reverts commit 3ff449c

* Revert "[lower_convs_to_matmul]: added support for non-square input images and kernels and non-equal padding."

This reverts commit 15e34ed.

* Revert "[im2col, test_im2col]: added support for non-equal padding"

This reverts commit c524020.

* [im2col] function compute_conv_output_dim can now be called in case non-equal and equal padding is assumed.
[test_im2col] changed function call to compute_conv_output_dim.
[test_conv_lowering] changed function call to compute_conv_output_dim.
[lower_convs_to_matmul] removed old assertion.

* [im2col]: changed how a square kernel is instantiated.
[lower_convs_to_matmul]: changed how the kernel size attribute is read (based on how a square kernel is instantiated).

* [im2col]: changed how a square kernel is instantiated.
[lower_convs_to_matmul]: changed how the kernel size attribute is read (based on how a square kernel is instantiated).

* [im2col]: minor change in style.

* [test_conv_lowering]: minor fix for test case depthwise and regular convolutions

* Support for non-square input images and kernels for im2col node (#20)

* [im2col.py]: support for non-square input images and kernels, [test_im2col]: added/modified several test cases for (non-)square images and kernels

* [im2col]: support for non-square input images and kernels, [test_im2col]: added/modified several test cases for (non-)square images and kernels

* [test_general_transformation]: changed kernel_size attribute to list instead of integer as required by im2col node

* [base]: changed how the "ints" field of AttributeProto set.
[test_general_transformation]: changed the type of the kernel_size attribute to list of integers

* removed unused import

* [base]: added support for writing repeated fields in AttributeProto

* minor style changes

* [im2col, test_im2col]: added support for non-equal padding

* [lower_convs_to_matmul]: added support for non-square input images and kernels and non-equal padding.
[test_conv_lowering]: added/modified test cases for non-equal padding, depthwise convolution and 'standard' convolution.

* [test_conv_lowering]: included 1D depthwise and regular convolutions in tests

* Revert "[test_conv_lowering]: included 1D depthwise and regular convolutions in tests"

This reverts commit 3ff449c

* Revert "[lower_convs_to_matmul]: added support for non-square input images and kernels and non-equal padding."

This reverts commit 15e34ed.

* Revert "[im2col, test_im2col]: added support for non-equal padding"

This reverts commit c524020.

* [im2col]: changed how a square kernel is instantiated.
[lower_convs_to_matmul]: changed how the kernel size attribute is read (based on how a square kernel is instantiated).

* [im2col]: minor change in style.

* [Im2Col] style fixes and comments

Co-authored-by: Yaman Umuroglu <[email protected]>

* Update AUTHORS.rst

Co-authored-by: Yaman Umuroglu <[email protected]>

* Added support for dilation value = 2 for 1D and 2D images/kernels (#17)

* [im2col.py]: support for non-square input images and kernels, [test_im2col]: added/modified several test cases for (non-)square images and kernels

* [im2col]: support for non-square input images and kernels, [test_im2col]: added/modified several test cases for (non-)square images and kernels

* [test_general_transformation]: changed kernel_size attribute to list instead of integer as required by im2col node

* [base]: changed how the "ints" field of AttributeProto set.
[test_general_transformation]: changed the type of the kernel_size attribute to list of integers

* removed unused import

* [base]: added support for writing repeated fields in AttributeProto

* minor style changes

* [im2col, test_im2col]: added support for non-equal padding

* [lower_convs_to_matmul]: added support for non-square input images and kernels and non-equal padding.
[test_conv_lowering]: added/modified test cases for non-equal padding, depthwise convolution and 'standard' convolution.

* [test_conv_lowering]: included 1D depthwise and regular convolutions in tests

* Revert "[test_conv_lowering]: included 1D depthwise and regular convolutions in tests"

This reverts commit 3ff449c

* Revert "[lower_convs_to_matmul]: added support for non-square input images and kernels and non-equal padding."

This reverts commit 15e34ed.

* Revert "[im2col, test_im2col]: added support for non-equal padding"

This reverts commit c524020.

* [im2col] added support for dilations = 2 for 2D and 1D images. Dilation value must be equal along each axis.
[test_im2col] added several test cases in case dilations = 2 (a.o. cases where image and kernel are 2D and 1D, with and without padding, and with stride = 2).
[lower_convs_to_matmul] added support for dilation value. Dilation value must be equal along each axis.
[test_conv_lowering] added test case for dilations = 2 for 2D and 1D images and kernels, with and without padding, and with stride = 2.

* [lower_convs_to_matmul] removed old assertion
[test_conv_lowering] added more dilation values to test cases. Dilation values of {1, 2, 3, 4} are tested.

* [im2col] function compute_conv_output_dim can now be called in case non-equal and equal padding is assumed.
[test_im2col] changed function call to compute_conv_output_dim
[test_conv_lowering] changed function call to compute_conv_output_dim

* [im2col] function compute_conv_output_dim can now be called in case non-equal and equal padding is assumed.
[test_im2col] changed function call to compute_conv_output_dim.
[test_conv_lowering] changed function call to compute_conv_output_dim.
[lower_convs_to_matmul] removed old assertion.

* [im2col]: changed how a square kernel is instantiated.
[lower_convs_to_matmul]: changed how the kernel size attribute is read (based on how a square kernel is instantiated).

* [im2col]: changed how a square kernel is instantiated.
[lower_convs_to_matmul]: changed how the kernel size attribute is read (based on how a square kernel is instantiated).

* [im2col]: changed how a square kernel is instantiated.
[lower_convs_to_matmul]: changed how the kernel size attribute is read (based on how a square kernel is instantiated).

* [im2col]: minor change in style.

* [test_conv_lowering]: minor fix for test case depthwise and regular convolutions

* [im2col]: minor style adjustment.
[test_conv_lowering]: merged test functions into one test function.

* Support for non-square input images and kernels for im2col node (#20)

* [im2col.py]: support for non-square input images and kernels, [test_im2col]: added/modified several test cases for (non-)square images and kernels

* [im2col]: support for non-square input images and kernels, [test_im2col]: added/modified several test cases for (non-)square images and kernels

* [test_general_transformation]: changed kernel_size attribute to list instead of integer as required by im2col node

* [base]: changed how the "ints" field of AttributeProto set.
[test_general_transformation]: changed the type of the kernel_size attribute to list of integers

* removed unused import

* [base]: added support for writing repeated fields in AttributeProto

* minor style changes

* [im2col, test_im2col]: added support for non-equal padding

* [lower_convs_to_matmul]: added support for non-square input images and kernels and non-equal padding.
[test_conv_lowering]: added/modified test cases for non-equal padding, depthwise convolution and 'standard' convolution.

* [test_conv_lowering]: included 1D depthwise and regular convolutions in tests

* Revert "[test_conv_lowering]: included 1D depthwise and regular convolutions in tests"

This reverts commit 3ff449c

* Revert "[lower_convs_to_matmul]: added support for non-square input images and kernels and non-equal padding."

This reverts commit 15e34ed.

* Revert "[im2col, test_im2col]: added support for non-equal padding"

This reverts commit c524020.

* [im2col]: changed how a square kernel is instantiated.
[lower_convs_to_matmul]: changed how the kernel size attribute is read (based on how a square kernel is instantiated).

* [im2col]: minor change in style.

* [Im2Col] style fixes and comments

Co-authored-by: Yaman Umuroglu <[email protected]>

* Update AUTHORS.rst

* Support for non-square input images and kernels in LowerConvsToMatMul transformation (#16)

* [im2col.py]: support for non-square input images and kernels, [test_im2col]: added/modified several test cases for (non-)square images and kernels

* [im2col]: support for non-square input images and kernels, [test_im2col]: added/modified several test cases for (non-)square images and kernels

* [test_general_transformation]: changed kernel_size attribute to list instead of integer as required by im2col node

* [base]: changed how the "ints" field of AttributeProto set.
[test_general_transformation]: changed the type of the kernel_size attribute to list of integers

* removed unused import

* [base]: added support for writing repeated fields in AttributeProto

* minor style changes

* [im2col, test_im2col]: added support for non-equal padding

* [lower_convs_to_matmul]: added support for non-square input images and kernels and non-equal padding.
[test_conv_lowering]: added/modified test cases for non-equal padding, depthwise convolution and 'standard' convolution.

* [test_conv_lowering]: included 1D depthwise and regular convolutions in tests

* Revert "[test_conv_lowering]: included 1D depthwise and regular convolutions in tests"

This reverts commit 3ff449c

* Revert "[lower_convs_to_matmul]: added support for non-square input images and kernels and non-equal padding."

This reverts commit 15e34ed.

* Revert "[im2col, test_im2col]: added support for non-equal padding"

This reverts commit c524020.

* [im2col] function compute_conv_output_dim can now be called in case non-equal and equal padding is assumed.
[test_im2col] changed function call to compute_conv_output_dim.
[test_conv_lowering] changed function call to compute_conv_output_dim.
[lower_convs_to_matmul] removed old assertion.

* [im2col]: changed how a square kernel is instantiated.
[lower_convs_to_matmul]: changed how the kernel size attribute is read (based on how a square kernel is instantiated).

* [im2col]: changed how a square kernel is instantiated.
[lower_convs_to_matmul]: changed how the kernel size attribute is read (based on how a square kernel is instantiated).

* [im2col]: minor change in style.

* [test_conv_lowering]: minor fix for test case depthwise and regular convolutions

* Support for non-square input images and kernels for im2col node (#20)

* [im2col.py]: support for non-square input images and kernels, [test_im2col]: added/modified several test cases for (non-)square images and kernels

* [im2col]: support for non-square input images and kernels, [test_im2col]: added/modified several test cases for (non-)square images and kernels

* [test_general_transformation]: changed kernel_size attribute to list instead of integer as required by im2col node

* [base]: changed how the "ints" field of AttributeProto set.
[test_general_transformation]: changed the type of the kernel_size attribute to list of integers

* removed unused import

* [base]: added support for writing repeated fields in AttributeProto

* minor style changes

* [im2col, test_im2col]: added support for non-equal padding

* [lower_convs_to_matmul]: added support for non-square input images and kernels and non-equal padding.
[test_conv_lowering]: added/modified test cases for non-equal padding, depthwise convolution and 'standard' convolution.

* [test_conv_lowering]: included 1D depthwise and regular convolutions in tests

* Revert "[test_conv_lowering]: included 1D depthwise and regular convolutions in tests"

This reverts commit 3ff449c

* Revert "[lower_convs_to_matmul]: added support for non-square input images and kernels and non-equal padding."

This reverts commit 15e34ed.

* Revert "[im2col, test_im2col]: added support for non-equal padding"

This reverts commit c524020.

* [im2col]: changed how a square kernel is instantiated.
[lower_convs_to_matmul]: changed how the kernel size attribute is read (based on how a square kernel is instantiated).

* [im2col]: minor change in style.

* [Im2Col] style fixes and comments

Co-authored-by: Yaman Umuroglu <[email protected]>

* Update AUTHORS.rst

Co-authored-by: Yaman Umuroglu <[email protected]>

Co-authored-by: Yaman Umuroglu <[email protected]>

* Update quantavgpool2d.py (#22)

Add  "Copyright (c) 2021 Xilinx, Inc" heather

* [batchnorm_to_affine]: epsilon value is now read out from the attributes. (#21)

[test_batchnorm_to_affine]: added a test case for various epsilon values.

* Added 3D to 4D (tensor) transformation (#19)

* [im2col.py]: support for non-square input images and kernels, [test_im2col]: added/modified several test cases for (non-)square images and kernels

* [im2col]: support for non-square input images and kernels, [test_im2col]: added/modified several test cases for (non-)square images and kernels

* [test_general_transformation]: changed kernel_size attribute to list instead of integer as required by im2col node

* [base]: changed how the "ints" field of AttributeProto set.
[test_general_transformation]: changed the type of the kernel_size attribute to list of integers

* removed unused import

* [base]: added support for writing repeated fields in AttributeProto

* minor style changes

* [im2col, test_im2col]: added support for non-equal padding

* [lower_convs_to_matmul]: added support for non-square input images and kernels and non-equal padding.
[test_conv_lowering]: added/modified test cases for non-equal padding, depthwise convolution and 'standard' convolution.

* [test_conv_lowering]: included 1D depthwise and regular convolutions in tests

* Revert "[test_conv_lowering]: included 1D depthwise and regular convolutions in tests"

This reverts commit 3ff449c

* Revert "[lower_convs_to_matmul]: added support for non-square input images and kernels and non-equal padding."

This reverts commit 15e34ed.

* Revert "[im2col, test_im2col]: added support for non-equal padding"

This reverts commit c524020.

* [im2col] added support for dilations = 2 for 2D and 1D images. Dilation value must be equal along each axis.
[test_im2col] added several test cases in case dilations = 2 (a.o. cases where image and kernel are 2D and 1D, with and without padding, and with stride = 2).
[lower_convs_to_matmul] added support for dilation value. Dilation value must be equal along each axis.
[test_conv_lowering] added test case for dilations = 2 for 2D and 1D images and kernels, with and without padding, and with stride = 2.

* [lower_convs_to_matmul] removed old assertion
[test_conv_lowering] added more dilation values to test cases. Dilation values of {1, 2, 3, 4} are tested.

* [im2col] function compute_conv_output_dim can now be called in case non-equal and equal padding is assumed.
[test_im2col] changed function call to compute_conv_output_dim
[test_conv_lowering] changed function call to compute_conv_output_dim

* [im2col] function compute_conv_output_dim can now be called in case non-equal and equal padding is assumed.
[test_im2col] changed function call to compute_conv_output_dim.
[test_conv_lowering] changed function call to compute_conv_output_dim.
[lower_convs_to_matmul] removed old assertion.

* [im2col]: minor fix with assumption on kernel dimension
[lower_convs_to_matmul]: minor fix with assumption on kernel dimension

* [change_3d_tensors_to_4d]: added new transformation that transforms 3D tensors to 4D and changes the nodes accordingly
[test_4d_conversion]: test function for 3D to 4D tensor transformation

* [change_3d_tensors_to_4d]: added new transformation that changes 3D tensors to 4D.
[test_4d_conversion]: added a test case for the 3D to 4D transformation.

* [change_3d_tensors_to_4d]: added 3D to 4D transformation (for QuartzNet).
[test_4d_conversion]: added test case for 3D to 4D transform.

* [change_3d_tensors_to_4d]: changed how an invalid graph is handled.
[test_4d_conversion]: changed the test case for an invalid graph.

* [im2col]: changed how a square kernel is instantiated.
[lower_convs_to_matmul]: changed how the kernel size attribute is read (based on how a square kernel is instantiated).

* [im2col]: changed how a square kernel is instantiated.
[lower_convs_to_matmul]: changed how the kernel size attribute is read (based on how a square kernel is instantiated).

* [im2col]: changed how a square kernel is instantiated.
[lower_convs_to_matmul]: changed how the kernel size attribute is read (based on how a square kernel is instantiated).

* [im2col]: minor change in style.

* [im2col]: minor style change and changed the way how a square kernel is instantiated.

* [test_conv_lowering]: merged tests for depthwise and standard convolutions.

* [test_conv_lowering]: minor fix for test case depthwise and regular convolutions

* [im2col]: minor style adjustment.
[test_conv_lowering]: merged test functions into one test function.

* [change_3d_tensors_to_4d]: style fixes and comments.
[test_4d_converions]: rearranged code.

* [Transform] check invalid node list length

* [change_3d_tensors_to_4d]: rearranged the code to make it more readable.

Co-authored-by: Yaman Umuroglu <[email protected]>

* [Docs] update tutorials

* [Util] experimental: fast_mode data packing for binary (#24)

* Generic partitioning feature (#23)

* Add basic partitioning functionality

* Mount build dir within docker container

* Support for non-linear models and multi in/out partitions

* Remove dataflowpartition custom op from finn-base

* Fix temporary build dir for CI

* Fix docstring

* [create_generic_partitions]: minor modification, removed redundant output value_info entries. (#26)

* [extend_partition]: added a new transformation ExtendPartition. (#27)

[test_extend_partition]: added a test case for the new transformation.

* Added support for non-equal strides along different axes (#25)

* [im2col]: added support for non-equal strides along different axes and cleaned up the code.
[lower_convs_to_matmul]: added support for non-equal strides along different axes and cleaned up the code.
[test_conv_lowering]: added test case for non-equal strides along different axes.

* [im2col]: minor fix.
[test_im2col]: added test case for non-equal strides along different axes.

* Changes for supporting vitis_hls (#28)

* [Refactor] split up RTL/HLS-related utils

* [Util] rename to CallHLS and allow specifying vivado_hls/vitis_hls

* [Util] more flexible stream naming in rtlsim_multi_io

* Changes for supporting non-equal dilation (#29)

* added support for non-equal dilation value along (H, W) dimension

* added test cases for non-equal dilation configurations

* appending dilation value along dummy dimension correctly (i.e. with a '1')

* changed tensor sparsity annotation for consistency

* Support infer_datatype for flatten layer (#30)

* Support infer_datatype for flatten layer

* [InferDT] add more identity op types for datatype inference

* [Lint] fix linting issues

Co-authored-by: Yaman Umuroglu <[email protected]>

* Update AUTHORS.rst

* Create python-publish.yml

* Add ZCU111 board to part map (#32)

* Update AUTHORS.rst

Co-authored-by: Mirza Mrahorovic <[email protected]>
Co-authored-by: jalezeta <[email protected]>
Co-authored-by: Felix Jentzsch <[email protected]>

* Update python-publish.yml

* Update python-publish.yml

* Lint

Co-authored-by: Mirza Mrahorovic <[email protected]>
Co-authored-by: jalezeta <[email protected]>
Co-authored-by: Felix Jentzsch <[email protected]>

* Inference cost: JSON export, parameter for output onnx

* Added support for cost estimation for upsampling

Co-authored-by: Yaman Umuroglu <[email protected]>
Co-authored-by: Felix Jentzsch <[email protected]>
Co-authored-by: Mirza Mrahorovic <[email protected]>
Co-authored-by: jalezeta <[email protected]>
Co-authored-by: Felix Jentzsch <[email protected]>

.github/workflows/python-publish.yml

@@ -0,0 +1,31 @@
+# This workflow will upload a Python Package using Twine when a release is created
+# For more information see: https://help.github.com/en/actions/language-and-framework-guides/using-python-with-github-actions#publishing-to-package-registries
+
+name: Upload Python Package
+
+on:
+  release:
+    types: [created]
+
+jobs:
+  deploy:
+
+    runs-on: ubuntu-latest
+
+    steps:
+    - uses: actions/checkout@v2
+    - name: Set up Python
+      uses: actions/setup-python@v2
+      with:
+        python-version: '3.x'
+    - name: Install dependencies
+      run: |
+        python -m pip install --upgrade pip
+        pip install setuptools wheel twine
+    - name: Build and publish
+      env:
+        TWINE_USERNAME: __token__
+        TWINE_PASSWORD: ${{ secrets.PYPI_API_TOKEN }}
+      run: |
+        python setup.py sdist
+        twine upload dist/*

AUTHORS.rst

@@ -2,10 +2,12 @@
 Contributors
 ============
 
-* Yaman Umuroglu <[email protected]> (maintainer)
-* Sambhav Jain <[email protected]> (maintainer)
+* Yaman Umuroglu (@maltanar) (maintainer)
+* Sambhav Jain (@sjain-stanford)
 * Jakoba Petri-Koenig (@auphelia)
 * Lucian Petrica (@quetric)
 * Tobias Alonso (@Tobi-Alonso)
 * Hendrik Borras (@HenniOVP)
 * Mirza Mrahorovic (@mmrahorovic)
+* Felix Paul Jentzsch (@felixpj)
+* Jon Ander Lezeta (@jalezeta)

src/finn/analysis/inference_cost.py

@@ -165,6 +165,42 @@ def inference_cost_matmul(model, node, discount_sparsity):
     return ret
 
 
+def inference_cost_upsample(model, node, discount_sparsity):
+    # extract info about the upsampling kernel attributes
+    mode = get_by_name(node.attribute, "mode").s.decode("utf-8")
+    scales_tensor = node.input[1]
+    scales_initializer = model.get_initializer(scales_tensor)
+
+    # extract info from tensor shapes and datatypes
+    (i_dtype, scale_dtype, o_dtype) = get_node_tensor_dtypes(model, node)
+    (i_shape, scale_shape, o_shape) = get_node_tensor_shapes(model, node)
+    bsize = i_shape[0]
+    ifm_ch = i_shape[1]
+    ofm_pix_total = np.prod(o_shape[2:])
+
+    # MAC calculation
+    if mode == "nearest":
+        # No calculation involved, since data is just copied over multiple times
+        n_macs = 0
+    elif mode == "linear":
+        # Data gets linearly interpolated in each dimension
+        # Two MACs per dimension and output pixel assumed
+        n_dim_scaling = np.sum(scales_initializer > 1)
+        n_macs = 2 * n_dim_scaling * ofm_pix_total * ifm_ch * bsize
+    else:
+        raise ValueError(f"Upsampling mode {mode} not supported for estimation.")
+
+    # Mem calculation
+    o_mem = np.prod(o_shape)
+    idt_name = i_dtype.name
+    odt_name = o_dtype.name
+    mac_op_type_str = "op_mac_%s_%s" % (idt_name, idt_name)
+    o_mem_type_str = "mem_o_%s" % (odt_name)
+
+    ret = {mac_op_type_str: n_macs, o_mem_type_str: o_mem}
+    return ret
+
+
 def inference_cost(model, discount_sparsity=True):
     "Ensure all nodes have unique names prior to calling this analysis pass."
 
@@ -193,6 +229,7 @@ def inference_cost(model, discount_sparsity=True):
         "Conv": inference_cost_conv,
         "MatMul": inference_cost_matmul,
         "Gemm": inference_cost_matmul,
+        "Upsample": inference_cost_upsample,
     }
     for node in model.graph.node:
         if node.op_type in inference_cost_fxn_map.keys():

src/finn/util/basic.py

@@ -44,6 +44,7 @@
 pynq_part_map["Pynq-Z2"] = "xc7z020clg400-1"
 pynq_part_map["ZCU102"] = "xczu9eg-ffvb1156-2-e"
 pynq_part_map["ZCU104"] = "xczu7ev-ffvc1156-2-e"
+pynq_part_map["ZCU111"] = "xczu28dr-ffvg1517-2-e"
 
 # native AXI HP port width (in bits) for PYNQ boards
 pynq_native_port_width = dict()
@@ -52,6 +53,7 @@
 pynq_native_port_width["Ultra96"] = 128
 pynq_native_port_width["ZCU102"] = 128
 pynq_native_port_width["ZCU104"] = 128
+pynq_native_port_width["ZCU111"] = 128
 
 # Alveo device and platform mappings
 alveo_part_map = dict()

src/finn/util/inference_cost.py

@@ -68,20 +68,20 @@ def compute_mem_bits(inf_cost_dict, filter_string="mem_w"):
 def inference_cost(
     model_filename,
     *,
-    output_json="inference_cost.json",
+    output_json=None,
+    output_onnx=None,
     preprocess=True,
-    save_final=True,
     discount_sparsity=True
 ):
     """Print the inference cost estimate metric for given ONNX model.
     Supports the Quant op for weight/activation quantization.
 
     :param model_filename: Filename for ONNX model
-    :param output_json: Filename for output JSON with detailed inference cost dict
+    :param output_json: Optional JSON filename to save the inference cost dict
+    :param output_onnx: Optional ONNX filename to save the final model after any
+        preprocessing
     :param preprocess: If set, run preprocessing steps such as shape inference,
         datatype inference and constant folding. Strongly recommended.
-    :param save_final: If set, save the final ONNX model after any preprocessing
-        as final.onnx
     :param discount_sparsity: If set, will discount op cost of MAC ops with a
         constant zero weight, and the mem cost of constant zero weights.
     """
@@ -100,8 +100,8 @@ def inference_cost(
         model = model.transform(InferDataTypes())
     model = model.transform(GiveUniqueNodeNames())
     model = model.transform(GiveReadableTensorNames())
-    if save_final:
-        model.save("final.onnx")
+    if output_onnx is not None:
+        model.save(output_onnx)
     ret = model.analysis(lambda x: infca.inference_cost(x, discount_sparsity))
     bops = compute_bops(ret)
     mem_w_bits = compute_mem_bits(ret, "mem_w")
@@ -114,6 +114,10 @@ def inference_cost(
         ret["unsupported"] = str(ret["unsupported"])
     print(json.dumps(ret, sort_keys=True, indent=2))
 
+    if output_json is not None:
+        with open(output_json, "w") as f:
+            json.dump(ret, f, sort_keys=True, indent=2)
+
 
 def main():
     clize.run(inference_cost)