diff --git a/docker/finn_entrypoint.sh b/docker/finn_entrypoint.sh
index 80e5261c9b6568e0f340ae0add86d269c036ff16..b312737c317517ca0ab19c74cf22284b5977b661 100644
--- a/docker/finn_entrypoint.sh
+++ b/docker/finn_entrypoint.sh
@@ -13,9 +13,9 @@ gecho () {
# checkout the correct dependency repo commits
# the repos themselves are cloned in the Dockerfile
-BREVITAS_COMMIT=026a509186b7e7b0b65d46a2f905043d41069306
+BREVITAS_COMMIT=f9a27226d4acf1661dd38bc449f71f89e0983cce
CNPY_COMMIT=4e8810b1a8637695171ed346ce68f6984e585ef4
-HLSLIB_COMMIT=8aed899c278c36c977a249558d71795086cf852c
+HLSLIB_COMMIT=8f9f2018762f654f196b666838aeaf6fc730ad9a
PYVERILATOR_COMMIT=c97a5ba41bbc7c419d6f25c74cdf3bdc3393174f
PYNQSHELL_COMMIT=0c82a61b0ec1a07fa275a14146233824ded7a13d
OMX_COMMIT=1bae737669901e762f581af73348332b5c4b2ada
diff --git a/docs/finn/example_networks.rst b/docs/finn/example_networks.rst
index 9f221871f09bf655db9d81988d6fa83e53473634..86bb2bd11fd805a23a3bdf6da8a8ed686259ecc1 100644
--- a/docs/finn/example_networks.rst
+++ b/docs/finn/example_networks.rst
@@ -20,17 +20,17 @@ version, this is indicated by an x mark in the table.
+-----------------------+------------+----------+----------+----------+----------+----------+----------+
| Export/Import | x | x | x | x | x | x | x |
+-----------------------+------------+----------+----------+----------+----------+----------+----------+
-| Streamlining | x | x | x | x | x | | |
+| Streamlining | x | x | x | x | x | | x |
+-----------------------+------------+----------+----------+----------+----------+----------+----------+
-| Convert to HLS layers | x | x | x | x | x | | |
+| Convert to HLS layers | x | x | x | x | x | | x |
+-----------------------+------------+----------+----------+----------+----------+----------+----------+
-| Stitched IP | x | x | x | x | x | | |
+| Stitched IP | x | x | x | x | x | | x |
+-----------------------+------------+----------+----------+----------+----------+----------+----------+
-| Hardware test | x | x | x | | x | | |
+| Hardware test | x | x | x | | x | | x |
+-----------------------+------------+----------+----------+----------+----------+----------+----------+
-| cppsim | x | x | x | x | x | | |
+| cppsim | x | x | x | x | x | | x |
+-----------------------+------------+----------+----------+----------+----------+----------+----------+
-| rtlsim node-by-node | x | x | x | x | x | | |
+| rtlsim node-by-node | x | x | x | x | x | | x |
+-----------------------+------------+----------+----------+----------+----------+----------+----------+
-| rtlsim stitched IP | x | x | x | x | x | | |
+| rtlsim stitched IP | x | x | x | x | x | | x |
+-----------------------+------------+----------+----------+----------+----------+----------+----------+
diff --git a/docs/finn/getting_started.rst b/docs/finn/getting_started.rst
index 5beabeb2980840fd1dbd2ee5b058738fa8553152..f4fa7a13dcbe4fe8ab9667a111df00c605747710 100644
--- a/docs/finn/getting_started.rst
+++ b/docs/finn/getting_started.rst
@@ -31,6 +31,7 @@ Getting an interactive shell for development or experimentation
sh run_docker.sh
Simply running sh run-docker.sh without any additional arguments will clone the dependency repos, create a Docker container and give you a terminal with you can use for development for experimentation.
+If you want a new terminal on an already-running container, you can do this with `docker exec -it finn_dev_<username> bash`.
.. warning:: The Docker container is spawned with the `--rm` option, so make sure that any important files you created inside the container are either in the /workspace/finn folder (which is mounted from the host computer) or otherwise backed up.
diff --git a/docs/finn/verification.rst b/docs/finn/verification.rst
index 391c6f999312839daca0d4161336c7c0ae822f89..c52c0840aa40566d930164490b1fd249d7c07757 100644
--- a/docs/finn/verification.rst
+++ b/docs/finn/verification.rst
@@ -28,4 +28,15 @@ This simulation can be used for a model containing several HLS custom operations
Emulation using PyVerilator
===========================
-The emulation using PyVerilator can be used when IP blocks were generated, either node by node or of a whole design. For that purpose PyVerilator gets the generated verilog files.
+The emulation using PyVerilator can be used when IP blocks were generated, either node by node or of a whole (IP-stitched) design. For that purpose PyVerilator gets the generated verilog files.
+
+For debugging purposes, it's possible to generate .vcd trace files that show the value of external & internal signals as the emuation is running. To enable this:
+ - for node-by-node rtlsim, set the `rtlsim_trace` attribute of each node of interest to either a file name for the vcd or `default` to use the node name as the filename.
+ - for IP-stitched rtlsim, set the `rtlsim_trace` metadata_prop for the graph as per above.
+
+To control the tracing depth in the module hierarchy, use the `RTLSIM_TRACE_DEPTH` environment variable (default is 1):
+ - level 1 shows top-level input/output streams
+ - level 2 shows per-layer input/output streams
+ - level 3 shows per full-layer I/O including FIFO count signals
+
+Note that deeper tracing will take longer to execute and may produce very large .vcd files.
diff --git a/src/finn/core/onnx_exec.py b/src/finn/core/onnx_exec.py
index efdfaa19d9f9e5dfa41911a2184e989337b3d9c2..7c3123cd5eb29a54dc5cbfb912225ad3fdb0f219 100644
--- a/src/finn/core/onnx_exec.py
+++ b/src/finn/core/onnx_exec.py
@@ -108,7 +108,9 @@ def execute_node(node, context, graph):
context[outp] = output_list[list_ind]
-def execute_onnx(model, input_dict, return_full_exec_context=False):
+def execute_onnx(
+ model, input_dict, return_full_exec_context=False, start_node=None, end_node=None
+):
"""Executes given ONNX ModelWrapper with given named inputs.
If return_full_exec_context is False, a dict of named outputs is returned
@@ -116,7 +118,12 @@ def execute_onnx(model, input_dict, return_full_exec_context=False):
If return return_full_exec_context is True, the full set of tensors used by
the execution (including inputs, weights, activations and final outputs)
- will be returned as a dict."""
+ will be returned as a dict.
+
+ When start_node and end_node are set to None, the whole graph is executed.
+ If they are set to particular ONNX nodes, only the subgraph between (and
+ including) those nodes is executed.
+ """
if not model.check_all_tensor_shapes_specified():
raise Exception("Found unspecified tensor shapes, try infer_shapes")
@@ -159,7 +166,17 @@ def execute_onnx(model, input_dict, return_full_exec_context=False):
# execute the model node by node
# we can simply walk down the list since the ONNX spec guarantees that it is
# topologically sorted
- for node in graph.node:
+ subgraph = []
+ if start_node is None:
+ start_node = model.graph.node[0]
+ if end_node is None:
+ end_node = model.graph.node[-1]
+ # select the nodes between specified start/end nodes
+ start_ind = model.get_node_index(start_node)
+ end_ind = model.get_node_index(end_node) + 1
+ assert end_ind >= start_ind, "Start/end nodes must define valid subgraph"
+ subgraph = graph.node[start_ind:end_ind]
+ for node in subgraph:
if get_sanitize_quant_tensors() != 0:
# round input values to match quantization annotation
execution_context = sanitize_quant_values(
diff --git a/src/finn/custom_op/__init__.py b/src/finn/custom_op/__init__.py
index ab6e03bee65b8bf5c4041dd8021b1a561e7673d2..4ae7b9ebffaab6ca6be04b8d73f647b2db22dc78 100644
--- a/src/finn/custom_op/__init__.py
+++ b/src/finn/custom_op/__init__.py
@@ -56,8 +56,15 @@ class CustomOp(ABC):
ret = ret.decode("utf-8")
return ret
else:
- # not set, return default value
- return def_val
+ if req:
+ raise Exception(
+ """Required attribute %s unspecified in
+ a %s node"""
+ % (name, self.onnx_node.op_type)
+ )
+ else:
+ # not set, return default value
+ return def_val
except KeyError:
raise AttributeError("Op has no such attribute: " + name)
diff --git a/src/finn/custom_op/fpgadataflow/__init__.py b/src/finn/custom_op/fpgadataflow/__init__.py
index a688898f4a43b33fd3f07cda12144b84829e451f..71c731f96ca45519c443a5f932ead050770e17de 100644
--- a/src/finn/custom_op/fpgadataflow/__init__.py
+++ b/src/finn/custom_op/fpgadataflow/__init__.py
@@ -88,6 +88,8 @@ class HLSCustomOp(CustomOp):
"res_hls": ("s", False, ""),
"res_synth": ("s", False, ""),
"rtlsim_so": ("s", False, ""),
+ # partitioning info
+ "partition_id": ("i", False, 0),
# input and output FIFO depths
"inFIFODepth": ("i", False, 2),
"outFIFODepth": ("i", False, 2),
@@ -171,9 +173,15 @@ class HLSCustomOp(CustomOp):
of the node as a dictionary."""
ret = dict()
ret["BRAM_18K"] = self.bram_estimation()
+ ret["BRAM_efficiency"] = self.bram_efficiency_estimation()
ret["LUT"] = self.lut_estimation()
return ret
+ def bram_efficiency_estimation(self):
+ """Function for BRAM efficiency estimation: actual parameter storage
+ needed divided by the allocated BRAM storage (from estimation)"""
+ return 1
+
def bram_estimation(self):
"""Function for BRAM resource estimation, is member function of
HLSCustomOp class but has to be filled by every node"""
@@ -219,7 +227,6 @@ class HLSCustomOp(CustomOp):
self.code_gen_dict["$CLKPERIOD$"] = [str(clk)]
self.code_gen_dict["$EXTRA_DIRECTIVES$"] = self.ipgen_extra_directives()
-
template = self.ipgentcl_template
for key in self.code_gen_dict:
@@ -235,7 +242,7 @@ class HLSCustomOp(CustomOp):
def ipgen_extra_directives(self):
"Return a list of extra tcl directives for HLS synthesis."
return []
-
+
def ipgen_singlenode_code(self):
"""Builds the bash script for ip generation using the IPGenBuilder from
finn.util.fpgadataflow."""
diff --git a/src/finn/custom_op/fpgadataflow/channelwise_op_batch.py b/src/finn/custom_op/fpgadataflow/channelwise_op_batch.py
new file mode 100644
index 0000000000000000000000000000000000000000..ad68a4bde29123b2498ac7789048bcd2e13bf3bc
--- /dev/null
+++ b/src/finn/custom_op/fpgadataflow/channelwise_op_batch.py
@@ -0,0 +1,576 @@
+# Copyright (c) 2020, Xilinx
+# All rights reserved.
+#
+# Redistribution and use in source and binary forms, with or without
+# modification, are permitted provided that the following conditions are met:
+#
+# * Redistributions of source code must retain the above copyright notice, this
+# list of conditions and the following disclaimer.
+#
+# * Redistributions in binary form must reproduce the above copyright notice,
+# this list of conditions and the following disclaimer in the documentation
+# and/or other materials provided with the distribution.
+#
+# * Neither the name of FINN nor the names of its
+# contributors may be used to endorse or promote products derived from
+# this software without specific prior written permission.
+#
+# THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS "AS IS"
+# AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
+# IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE ARE
+# DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT HOLDER OR CONTRIBUTORS BE LIABLE
+# FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL
+# DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR
+# SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER
+# CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY,
+# OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE
+# OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
+
+from math import ceil
+import os
+
+import numpy as np
+
+from onnx import TensorProto, helper
+from finn.core.datatype import DataType
+from finn.custom_op.fpgadataflow import HLSCustomOp
+from finn.util.data_packing import (
+ npy_to_rtlsim_input,
+ numpy_to_hls_code,
+ rtlsim_output_to_npy,
+)
+from . import templates
+
+# ONNX i/o tensor shape assumptions for channelwise ops:
+# input 0 is the input tensor, shape (..., NumChannels)
+# input 1 is the channelwise parameter tensor, shape (NumChannels, params_per_channel)
+# output 0 is the output tensor, shape (..., NumChannels) - same as input
+# the ... here can be any shape (representing groups of vectors)
+
+
+class ChannelwiseOp_Batch(HLSCustomOp):
+ """Class that corresponds to finn-hls Thresholding_Batch function.
+ It can implement a variety of channel-wise parametrized operations,
+ including Add, Mul and multi-thresholding.
+ """
+
+ def __init__(self, onnx_node):
+ super().__init__(onnx_node)
+ self.decoupled_wrapper = templates.decoupled_wrapper
+
+ def get_nodeattr_types(self):
+ my_attrs = {
+ # channelwise "map" function to apply:
+ # one of cmp_le, cmp_ge, add, mul
+ "Func": ("s", False, "cmp_le"),
+ "PE": ("i", True, 0),
+ "NumChannels": ("i", True, 0),
+ # string defining memory resource type for parameters
+ "ram_style": ("s", False, "distributed"),
+ # FINN DataTypes for inputs, weights, outputs
+ "inputDataType": ("s", True, ""),
+ "paramDataType": ("s", True, ""),
+ "outputDataType": ("s", True, ""),
+ # input and output FIFO depths
+ "inFIFODepth": ("i", False, 0),
+ "outFIFODepth": ("i", False, 0),
+ # number of input vectors, examples:
+ # [1] is a single vector (like a FC layer with batch=1)
+ # [4] is four vectors (like a FC layer with batch=4)
+ # [1, 4, 4] is four * four vectors (like a conv layer with batch=1)
+ "numInputVectors": ("ints", False, [1]),
+ }
+ my_attrs.update(super().get_nodeattr_types())
+ return my_attrs
+
+ def calc_tmem(self):
+ """Calculates and returns TMEM, the depth of the memory used
+ to store the channelwise op parameters."""
+ chn = self.get_nodeattr("NumChannels")
+ pe = self.get_nodeattr("PE")
+ return chn // pe
+
+ def make_shape_compatible_op(self, model):
+ oshape = self.get_normal_output_shape()
+ # implement tensor with correct shape
+ values = np.random.randn(*oshape).astype(np.float32)
+ return helper.make_node(
+ "Constant",
+ inputs=[],
+ outputs=[self.onnx_node.output[0]],
+ value=helper.make_tensor(
+ name="const_tensor",
+ data_type=TensorProto.FLOAT,
+ dims=values.shape,
+ vals=values.flatten().astype(float),
+ ),
+ )
+
+ def infer_node_datatype(self, model):
+ node = self.onnx_node
+ # check input datatype against property
+ idt_name = self.get_input_datatype().name
+ exp_idt_name = self.get_nodeattr("inputDataType")
+ assert exp_idt_name == idt_name, "Bad input DataType for ChannelwiseOp layer"
+ # TODO: dynamically infer/update odt based on idt as done in ConvertToHLSLayers?
+ # set output datatype from property
+ odt = self.get_output_datatype()
+ model.set_tensor_datatype(node.output[0], odt)
+
+ def verify_node(self):
+ info_messages = []
+ # verify that "domain" is set to "finn"
+ domain_value = self.onnx_node.domain
+ if domain_value == "finn":
+ info_messages.append("Attribute domain is set correctly")
+ else:
+ info_messages.append('Attribute domain should be set to "finn"')
+
+ # verify that "backend" is set to "fpgadataflow"
+ backend_value = self.get_nodeattr("backend")
+ if backend_value == "fpgadataflow":
+ info_messages.append("Attribute backend is set correctly")
+ else:
+ info_messages.append('Attribute backend should be set to "fpgadataflow"')
+
+ # verify that all necessary attributes exist
+ # TODO collect automatically from get_nodeattr_types
+ try:
+ self.get_nodeattr("code_gen_dir_cppsim")
+ self.get_nodeattr("executable_path")
+ self.get_nodeattr("NumChannels")
+ self.get_nodeattr("PE")
+ self.get_nodeattr("inputDataType")
+ self.get_nodeattr("paramDataType")
+ self.get_nodeattr("outputDataType")
+ info_messages.append("All necessary attributes exist")
+ except Exception:
+ info_messages.append(
+ """The required Threshold_Batch attributes do not exist."""
+ )
+
+ return info_messages
+
+ def bram_estimation(self):
+ """Calculates BRAM cost if resource set to BRAM"""
+ style = self.get_nodeattr("ram_style")
+ P = self.get_nodeattr("PE")
+ idt = self.get_input_datatype()
+ A = idt.bitwidth()
+ tmem = self.calc_tmem()
+
+ if style == "block" and tmem > 1:
+ return int(ceil(A * P / 16)) * int(ceil(tmem / 1024))
+ else:
+ return 0
+
+ def lut_estimation(self):
+ """Calculates LUT cost, taking memory resource type into account """
+ # TODO add in/out FIFO contributions
+ style = self.get_nodeattr("ram_style")
+ P = self.get_nodeattr("PE")
+ idt = self.get_input_datatype()
+ A = idt.bitwidth()
+ tmem = self.calc_tmem()
+ # cost of comparators
+ comparator_cost = A * P
+ # cost of LUTRAM
+ if style == "distributed" and tmem > 1:
+ lutram_cost = P * A * int(ceil(tmem / 64))
+ else:
+ lutram_cost = 0
+ # total cost
+ return comparator_cost + lutram_cost
+
+ def get_input_datatype(self):
+ """Returns FINN DataType of input."""
+ return DataType[self.get_nodeattr("inputDataType")]
+
+ def get_output_datatype(self):
+ """Returns FINN DataType of output."""
+ return DataType[self.get_nodeattr("outputDataType")]
+
+ def get_instream_width(self):
+ i_bits = self.get_input_datatype().bitwidth()
+ return i_bits * self.get_nodeattr("PE")
+
+ def get_outstream_width(self):
+ o_bits = self.get_output_datatype().bitwidth()
+ return o_bits * self.get_nodeattr("PE")
+
+ def get_folded_input_shape(self):
+ ich = self.get_nodeattr("NumChannels")
+ pe = self.get_nodeattr("PE")
+ fold = ich // pe
+ vecs = list(self.get_nodeattr("numInputVectors"))
+ folded_input_shape = tuple(vecs + [fold, pe])
+ return folded_input_shape
+
+ def get_folded_output_shape(self):
+ # same shape as input
+ return self.get_folded_input_shape()
+
+ def get_normal_input_shape(self):
+ ich = self.get_nodeattr("NumChannels")
+ vecs = list(self.get_nodeattr("numInputVectors"))
+ normal_input_shape = tuple(vecs + [ich])
+ return normal_input_shape
+
+ def get_normal_output_shape(self):
+ # same shape as input
+ return self.get_normal_input_shape()
+
+ def get_number_output_values(self):
+ nf = np.prod(self.get_folded_output_shape()[:-1])
+ return nf
+
+ def get_template_param_values(self):
+ """Returns the template parameter values according to input, output and weight
+ data types."""
+ ret = dict()
+ inp_hls_str = self.get_input_datatype().get_hls_datatype_str()
+ out_hls_str = self.get_output_datatype().get_hls_datatype_str()
+ # fill in TSrcI
+ ret["TSrcI"] = "Slice<%s>" % inp_hls_str
+ # fill in TDstI
+ ret["TDstI"] = "Slice<%s>" % out_hls_str
+
+ return ret
+
+ def get_hls_compatible_parameter_tensor(self, orig_param_vector):
+ """Convert the original numpy weight matrix orig_weight_matrix into
+ a form suitable for passing to the hlslib call:
+ * ensure chn % PE == 0
+ * interleave rows between PEs
+ * reshape into (PE, TMEM) and return
+ """
+ chn = self.get_nodeattr("NumChannels")
+ pe = self.get_nodeattr("PE")
+ tmem = chn // pe
+ assert chn % pe == 0, "Requirement NumChannels divisable by PE is violated."
+ assert (
+ orig_param_vector.ndim == 1
+ ), """Parameter vector dimension is {}.
+ Expected dimension: 1.""".format(
+ orig_param_vector.ndim
+ )
+
+ # if not self.get_input_datatype().signed():
+ # # ensure all thresholds are nonnegative
+ # assert (orig_param_vector >= 0).all()
+
+ # ensure all thresholds are integer
+ assert (orig_param_vector.astype(np.int32) == orig_param_vector).all()
+ ret = orig_param_vector
+
+ assert (
+ ret.shape[0] == chn
+ ), "Cardinality of parameter vector is not as expected (chn)"
+
+ # distribute rows between PEs
+ ret = ret.reshape(tmem, pe).transpose()
+ assert (
+ ret.shape[0] == pe
+ ), """First dimension after distribution of the
+ rows between PEs is not as expected (pe)"""
+ assert (
+ ret.shape[1] == tmem
+ ), """Second dimension after distribution of the
+ rows between PEs is not as expected (tmem)"""
+
+ return ret.reshape(1, pe, tmem)
+
+ def generate_params(self, model, path):
+ code_gen_dir = path
+ # save thresholds in params.h
+ parameters = model.get_initializer(self.onnx_node.input[1])
+ parameter_tensor = self.get_hls_compatible_parameter_tensor(parameters)
+ pdt = DataType[self.get_nodeattr("paramDataType")]
+
+ parameters_hls_code = numpy_to_hls_code(
+ parameter_tensor, pdt, "parameters", False, True
+ )
+ # get input data type
+ export_idt = self.get_input_datatype()
+ if self.get_input_datatype() == DataType.BIPOLAR:
+ export_idt = DataType.BINARY
+ idt_hls = export_idt.get_hls_datatype_str()
+
+ # write parameters into params.h
+ f_params = open("{}/params.h".format(code_gen_dir), "w")
+ pdt_hls = pdt.get_hls_datatype_str()
+ # use binary to export bipolar activations
+ export_odt = self.get_output_datatype()
+ if self.get_output_datatype() == DataType.BIPOLAR:
+ export_odt = DataType.BINARY
+ odt_hls = export_odt.get_hls_datatype_str()
+ # get desired function
+ func = self.get_nodeattr("Func")
+ if func == "cmp_le":
+ func_str = "std::less_equal"
+ elif func == "cmp_ge":
+ func_str = "std::greater_equal"
+ elif func == "add":
+ func_str = "std::plus"
+ elif func == "mul":
+ func_str = "std::multiplies"
+ else:
+ raise Exception(
+ """Invalid value for attribute Func! Is currently set to: {}
+ has to be set to one of the following value
+ ("cmp_le", "cmp_ge", "add", "mul")""".format(
+ func
+ )
+ )
+ f_params.write(
+ "static ChannelWiseOperation<{},{},{},{},{},{}> threshs \
+ = ".format(
+ self.calc_tmem(),
+ self.get_nodeattr("PE"),
+ idt_hls,
+ pdt_hls,
+ odt_hls,
+ "%s<%s>" % (func_str, odt_hls),
+ )
+ )
+ f_params.write(parameters_hls_code)
+ f_params.close()
+
+ def execute_node(self, context, graph):
+ mode = self.get_nodeattr("exec_mode")
+ node = self.onnx_node
+
+ # TODO ensure codegen dir exists
+ if mode == "cppsim":
+ code_gen_dir = self.get_nodeattr("code_gen_dir_cppsim")
+ elif mode == "rtlsim":
+ code_gen_dir = self.get_nodeattr("code_gen_dir_ipgen")
+ else:
+ raise Exception(
+ """Invalid value for attribute exec_mode! Is currently set to: {}
+ has to be set to one of the following value ("cppsim", "rtlsim")""".format(
+ mode
+ )
+ )
+
+ # create a npy file fore each input of the node (in_ind is input index)
+ in_ind = 0
+ for inputs in node.input:
+ # it is assumed that the first input of the node is the data input
+ # the second input are the weights
+ # the third input are the thresholds
+ if in_ind == 0:
+ assert (
+ str(context[inputs].dtype) == "float32"
+ ), """Input datatype is
+ not float32 as expected."""
+ expected_inp_shape = self.get_folded_input_shape()
+ reshaped_input = context[inputs].reshape(expected_inp_shape)
+ export_idt = self.get_input_datatype()
+ # make copy before saving the array
+ reshaped_input = reshaped_input.copy()
+ np.save(
+ os.path.join(code_gen_dir, "input_{}.npy".format(in_ind)),
+ reshaped_input,
+ )
+ elif in_ind > 2:
+ raise Exception("Unexpected input found for ChannelwiseOp_Batch")
+ in_ind += 1
+
+ if mode == "cppsim":
+ # execute the precompiled model
+ super().exec_precompiled_singlenode_model()
+ # load output npy file
+ super().npy_to_dynamic_output(context)
+ # reinterpret binary output as bipolar where needed
+ if self.get_output_datatype() == DataType.BIPOLAR:
+ out = context[node.output[0]]
+ out = 2 * out - 1
+ context[node.output[0]] = out
+ assert (
+ context[node.output[0]].shape == self.get_folded_output_shape()
+ ), """Output shape is not as expected"""
+ # reshape output to have expected shape
+ oshape = self.get_normal_output_shape()
+ context[node.output[0]] = context[node.output[0]].reshape(*oshape)
+ elif mode == "rtlsim":
+ sim = self.get_rtlsim()
+ nbits = self.get_instream_width()
+ inp = npy_to_rtlsim_input(
+ "{}/input_0.npy".format(code_gen_dir), export_idt, nbits
+ )
+ super().reset_rtlsim(sim)
+ super().toggle_clk(sim)
+ output = self.rtlsim(sim, inp)
+ odt = self.get_output_datatype()
+ target_bits = odt.bitwidth()
+ packed_bits = self.get_outstream_width()
+ out_npy_path = "{}/output.npy".format(code_gen_dir)
+ out_shape = self.get_folded_output_shape()
+ rtlsim_output_to_npy(
+ output, out_npy_path, odt, out_shape, packed_bits, target_bits
+ )
+
+ # load and reshape output
+ output = np.load(out_npy_path)
+ oshape = self.get_normal_output_shape()
+ output = np.asarray([output], dtype=np.float32).reshape(*oshape)
+ context[node.output[0]] = output
+ else:
+ raise Exception(
+ """Invalid value for attribute exec_mode! Is currently set to: {}
+ has to be set to one of the following value ("cppsim", "rtlsim")""".format(
+ mode
+ )
+ )
+
+ def global_includes(self):
+ self.code_gen_dict["$GLOBALS$"] = ['#include "activations.hpp"']
+ self.code_gen_dict["$GLOBALS$"] += ['#include "params.h"']
+
+ # TODO check and add whatever missing
+ def defines(self, var):
+ numInputVectors = list(self.get_nodeattr("numInputVectors"))
+ numReps = numInputVectors[0]
+ self.code_gen_dict["$DEFINES$"] = [
+ """#define NumChannels1 {}\n#define PE1 {}\n#define numReps {}""".format(
+ self.get_nodeattr("NumChannels"), self.get_nodeattr("PE"), numReps,
+ )
+ ]
+
+ def read_npy_data(self):
+ code_gen_dir = self.get_nodeattr("code_gen_dir_cppsim")
+ dtype = self.get_input_datatype()
+ elem_bits = dtype.bitwidth()
+ packed_bits = self.get_instream_width()
+ packed_hls_type = "ap_uint<%d>" % packed_bits
+ elem_hls_type = dtype.get_hls_datatype_str()
+ npy_type = "float"
+ npy_in = "%s/input_0.npy" % code_gen_dir
+ self.code_gen_dict["$READNPYDATA$"] = []
+ # note: the innermost dim is reversed for the input
+ self.code_gen_dict["$READNPYDATA$"].append(
+ 'npy2apintstream<%s, %s, %d, %s>("%s", in0, false);'
+ % (packed_hls_type, elem_hls_type, elem_bits, npy_type, npy_in)
+ )
+
+ def strm_decl(self):
+ self.code_gen_dict["$STREAMDECLARATIONS$"] = []
+ self.code_gen_dict["$STREAMDECLARATIONS$"].append(
+ 'hls::stream<ap_uint<{}>> in0 ("in0");'.format(self.get_instream_width())
+ )
+ self.code_gen_dict["$STREAMDECLARATIONS$"].append(
+ 'hls::stream<ap_uint<{}>> out ("out");'.format(self.get_outstream_width())
+ )
+
+ def docompute(self):
+ tmpl_args = self.get_template_param_values()
+ # TODO: why put some template parameters into defines and not others?
+ # should ImgDim be defined or just filled in here like we do now?
+ ishape = self.get_folded_input_shape()
+ if len(ishape) == 3:
+ imgdim = 1
+ elif len(ishape) == 5:
+ imgdim = ishape[1]
+ else:
+ raise Exception("""Unexpeted input shape""")
+ self.code_gen_dict["$DOCOMPUTE$"] = [
+ """Thresholding_Batch<{}, NumChannels1, PE1, {}, {}>
+ (in0, out, threshs, numReps);""".format(
+ imgdim, tmpl_args["TSrcI"], tmpl_args["TDstI"],
+ )
+ ]
+
+ def dataoutstrm(self):
+ code_gen_dir = self.get_nodeattr("code_gen_dir_cppsim")
+ dtype = self.get_output_datatype()
+ if dtype == DataType.BIPOLAR:
+ # use binary for bipolar storage
+ dtype = DataType.BINARY
+ elem_bits = dtype.bitwidth()
+ packed_bits = self.get_outstream_width()
+ packed_hls_type = "ap_uint<%d>" % packed_bits
+ elem_hls_type = dtype.get_hls_datatype_str()
+ npy_type = "float"
+ npy_out = "%s/output.npy" % code_gen_dir
+ shape = self.get_folded_output_shape()
+ shape_cpp_str = str(shape).replace("(", "{").replace(")", "}")
+
+ # note: the innermost dim is not reversed for the output
+ self.code_gen_dict["$DATAOUTSTREAM$"] = [
+ 'apintstream2npy<%s, %s, %d, %s>(out, %s, "%s", false);'
+ % (
+ packed_hls_type,
+ elem_hls_type,
+ elem_bits,
+ npy_type,
+ shape_cpp_str,
+ npy_out,
+ )
+ ]
+
+ def save_as_npy(self):
+ self.code_gen_dict["$SAVEASCNPY$"] = []
+
+ def blackboxfunction(self):
+ self.code_gen_dict["$BLACKBOXFUNCTION$"] = [
+ """void {}(hls::stream<ap_uint<{}>> &in0,
+ hls::stream<ap_uint<{}>> &out
+ )""".format(
+ self.onnx_node.name,
+ self.get_instream_width(),
+ self.get_outstream_width(),
+ )
+ ]
+
+ def pragmas(self):
+ self.code_gen_dict["$PRAGMAS$"] = ["#pragma HLS INTERFACE axis port=in0"]
+ self.code_gen_dict["$PRAGMAS$"].append("#pragma HLS INTERFACE axis port=out")
+ self.code_gen_dict["$PRAGMAS$"].append(
+ "#pragma HLS INTERFACE ap_ctrl_none port=return"
+ )
+
+ # the channelwise parameter tensor is acc_type [PE][TMEM][N_PARAMS_PER_CHANNEL]
+ # partition for parallel access along PE and N_PARAMS_PER_CHANNEL
+ # dimensions (dims 1 and 3)
+ self.code_gen_dict["$PRAGMAS$"].append(
+ (
+ "#pragma HLS ARRAY_PARTITION variable=threshs.parameters "
+ "complete dim=1"
+ )
+ )
+ # self.code_gen_dict["$PRAGMAS$"].append(
+ # (
+ # "#pragma HLS ARRAY_PARTITION variable=threshs.parameters "
+ # "complete dim=3"
+ # )
+ # )
+
+ # set resource type
+ ram_style = self.get_nodeattr("ram_style")
+ pe = self.get_nodeattr("PE")
+ ich = self.get_nodeattr("NumChannels")
+ # if PE less than NumChannels, assign cores according to ram_style;
+ # otherwise if PE == NumChannels, Vivado HLS will unroll to FFs
+ if pe < ich:
+ if ram_style == "distributed":
+ self.code_gen_dict["$PRAGMAS$"].append(
+ (
+ "#pragma HLS RESOURCE variable=threshs.parameters "
+ "core=ROM_2P_LUTRAM"
+ )
+ )
+ elif ram_style == "block":
+ self.code_gen_dict["$PRAGMAS$"].append(
+ (
+ "#pragma HLS RESOURCE variable=threshs.parameters "
+ "core=ROM_2P_BRAM"
+ )
+ )
+ else:
+ raise Exception(
+ """Invalid value for attribute ram_style! Is currently set to: {}
+ has to be set to one of ("block", "distributed")""".format(
+ ram_style
+ )
+ )
diff --git a/src/finn/custom_op/fpgadataflow/downsampler.py b/src/finn/custom_op/fpgadataflow/downsampler.py
new file mode 100644
index 0000000000000000000000000000000000000000..0ce4379a2c41baa5bc009e9df7623d133ee89a09
--- /dev/null
+++ b/src/finn/custom_op/fpgadataflow/downsampler.py
@@ -0,0 +1,297 @@
+import os
+import numpy as np
+from onnx import TensorProto, helper
+from finn.core.datatype import DataType
+from finn.custom_op.fpgadataflow import HLSCustomOp
+from finn.util.data_packing import npy_to_rtlsim_input, rtlsim_output_to_npy
+
+
+class DownSampler(HLSCustomOp):
+ """Corresponds to finn-hlslib ConvolutionInputGenerator_kernel1 function.
+ Basically performs a down sampling of the image removing rows and columns."""
+
+ def __init__(self, onnx_node):
+ super().__init__(onnx_node)
+
+ def get_nodeattr_types(self):
+ my_attrs = {
+ # spatial size of input images
+ "ImgDim": ("i", True, 0),
+ # number of channels in input image
+ "NumChannels": ("i", True, 0),
+ # Number of input columns computed in parallel
+ "SIMD": ("i", False, 1),
+ "Stride": ("i", True, 2),
+ # FINN input datatype
+ "inputDataType": ("s", True, ""),
+ # Batch size
+ "numInputVectors": ("i", False, 1),
+ }
+ my_attrs.update(super().get_nodeattr_types())
+ return my_attrs
+
+ def get_downsampled_odim(self):
+ "Return the down sampled spatial size of the output."
+ idim = self.get_nodeattr("ImgDim")
+ stride = self.get_nodeattr("Stride")
+ return int(np.floor((idim - 1) / stride) + 1)
+
+ def get_normal_input_shape(self):
+ idim = self.get_nodeattr("ImgDim")
+ num_ch = self.get_nodeattr("NumChannels")
+ batch = self.get_nodeattr("numInputVectors")
+ ishape = (batch, idim, idim, num_ch)
+ return ishape
+
+ def get_normal_output_shape(self):
+ odim = self.get_downsampled_odim()
+ num_ch = self.get_nodeattr("NumChannels")
+ batch = self.get_nodeattr("numInputVectors")
+ oshape = (batch, odim, odim, num_ch)
+ return oshape
+
+ def get_folded_input_shape(self):
+ normal_ishape = list(self.get_normal_input_shape())
+ ifm_ch = self.get_nodeattr("NumChannels")
+ simd = self.get_nodeattr("SIMD")
+ assert ifm_ch % simd == 0, "SIMD must divide input channels"
+ fold = int(normal_ishape[-1] / simd)
+ folded_ishape = normal_ishape[:-1] + [fold, simd]
+ return tuple(folded_ishape)
+
+ def get_folded_output_shape(self):
+ normal_oshape = list(self.get_normal_output_shape())
+ ifm_ch = self.get_nodeattr("NumChannels")
+ simd = self.get_nodeattr("SIMD")
+ assert ifm_ch % simd == 0, "SIMD must divide input channels"
+ fold = int(normal_oshape[-1] / simd)
+ folded_oshape = normal_oshape[:-1] + [fold, simd]
+ return tuple(folded_oshape)
+
+ def make_shape_compatible_op(self, model):
+ exp_ishape = self.get_normal_input_shape()
+ oshape = self.get_normal_output_shape()
+ ishape = tuple(model.get_tensor_shape(self.onnx_node.input[0]))
+ assert ishape == exp_ishape, "Unexpect input shape for DownSampler."
+ # implement tensor with correct shape
+ values = np.random.randn(*oshape).astype(np.float32)
+ return helper.make_node(
+ "Constant",
+ inputs=[],
+ outputs=[self.onnx_node.output[0]],
+ value=helper.make_tensor(
+ name="const_tensor",
+ data_type=TensorProto.FLOAT,
+ dims=values.shape,
+ vals=values.flatten().astype(float),
+ ),
+ )
+
+ def infer_node_datatype(self, model):
+ node = self.onnx_node
+ # data type stays the same
+ dtype = model.get_tensor_datatype(node.input[0])
+ exp_idtype = self.get_input_datatype()
+ assert dtype == exp_idtype, "Unexpected datatype for DownSampler"
+ model.set_tensor_datatype(node.output[0], dtype)
+
+ def verify_node(self):
+ pass
+
+ def get_input_datatype(self):
+ """Returns FINN DataType of input."""
+ ret = DataType[self.get_nodeattr("inputDataType")]
+ return ret
+
+ def get_output_datatype(self):
+ """Returns FINN DataType of output. (Same as input datatype)"""
+ return self.get_input_datatype()
+
+ def get_instream_width(self):
+ ibits = self.get_input_datatype().bitwidth()
+ simd = self.get_nodeattr("SIMD")
+ return ibits * simd
+
+ def get_outstream_width(self):
+ obits = self.get_output_datatype().bitwidth()
+ simd = self.get_nodeattr("SIMD")
+ return obits * simd
+
+ def get_number_output_values(self):
+ folded_oshape = self.get_folded_output_shape()
+ return np.prod(folded_oshape[:-1])
+
+ def global_includes(self):
+ self.code_gen_dict["$GLOBALS$"] = ['#include "slidingwindow.h"']
+
+ def defines(self, var):
+ self.code_gen_dict["$DEFINES$"] = []
+
+ ifm_ch = self.get_nodeattr("NumChannels")
+ self.code_gen_dict["$DEFINES$"] += ["#define IFMChannels {}".format(ifm_ch)]
+
+ ibits = self.get_input_datatype().bitwidth()
+ self.code_gen_dict["$DEFINES$"] += ["#define Input_precision {}".format(ibits)]
+
+ idim = self.get_nodeattr("ImgDim")
+ self.code_gen_dict["$DEFINES$"] += ["#define IFMDim {}".format(idim)]
+
+ simd = self.get_nodeattr("SIMD")
+ self.code_gen_dict["$DEFINES$"] += ["#define SIMD {}".format(simd)]
+
+ stride = self.get_nodeattr("Stride")
+ self.code_gen_dict["$DEFINES$"] += ["#define Stride {}".format(stride)]
+
+ batch_size = self.get_nodeattr("numInputVectors")
+ self.code_gen_dict["$DEFINES$"] += ["#define numReps {}".format(batch_size)]
+
+ def read_npy_data(self):
+ code_gen_dir = self.get_nodeattr("code_gen_dir_cppsim")
+ dtype = self.get_input_datatype()
+ if dtype == DataType.BIPOLAR:
+ # use binary for bipolar storage
+ dtype = DataType.BINARY
+ elem_bits = dtype.bitwidth()
+ packed_bits = self.get_instream_width()
+ packed_hls_type = "ap_uint<%d>" % packed_bits
+ elem_hls_type = dtype.get_hls_datatype_str()
+ npy_type = "float"
+ npy_in = "%s/input_0.npy" % code_gen_dir
+ self.code_gen_dict["$READNPYDATA$"] = []
+ self.code_gen_dict["$READNPYDATA$"].append(
+ 'npy2apintstream<%s, %s, %d, %s>("%s", in0);'
+ % (packed_hls_type, elem_hls_type, elem_bits, npy_type, npy_in)
+ )
+
+ def strm_decl(self):
+ self.code_gen_dict["$STREAMDECLARATIONS$"] = []
+ self.code_gen_dict["$STREAMDECLARATIONS$"].append(
+ 'hls::stream<ap_uint<{}>> in0 ("in0");'.format(self.get_instream_width())
+ )
+ self.code_gen_dict["$STREAMDECLARATIONS$"].append(
+ 'hls::stream<ap_uint<{}>> out ("out");'.format(self.get_outstream_width())
+ )
+
+ def docompute(self):
+ self.code_gen_dict["$DOCOMPUTE$"] = [
+ """ConvolutionInputGenerator_kernel1<IFMChannels, Input_precision,
+ IFMDim, SIMD,Stride> (in0, out, numReps);"""
+ ]
+
+ def dataoutstrm(self):
+ code_gen_dir = self.get_nodeattr("code_gen_dir_cppsim")
+ dtype = self.get_output_datatype()
+ if dtype == DataType.BIPOLAR:
+ # use binary for bipolar storage
+ dtype = DataType.BINARY
+ elem_bits = dtype.bitwidth()
+ packed_bits = self.get_outstream_width()
+ packed_hls_type = "ap_uint<%d>" % packed_bits
+ elem_hls_type = dtype.get_hls_datatype_str()
+ npy_type = "float"
+ npy_out = "%s/output.npy" % code_gen_dir
+ oshape = self.get_folded_output_shape()
+ oshape_cpp_str = str(oshape).replace("(", "{").replace(")", "}")
+
+ self.code_gen_dict["$DATAOUTSTREAM$"] = [
+ 'apintstream2npy<%s, %s, %d, %s>(out, %s, "%s");'
+ % (
+ packed_hls_type,
+ elem_hls_type,
+ elem_bits,
+ npy_type,
+ oshape_cpp_str,
+ npy_out,
+ )
+ ]
+
+ def save_as_npy(self):
+ self.code_gen_dict["$SAVEASCNPY$"] = []
+
+ def blackboxfunction(self):
+ packed_bits = self.get_instream_width()
+ packed_hls_type = "ap_uint<%d>" % packed_bits
+ self.code_gen_dict["$BLACKBOXFUNCTION$"] = [
+ "void %s(hls::stream<%s > &in0, hls::stream<%s > &out)"
+ % (self.onnx_node.name, packed_hls_type, packed_hls_type)
+ ]
+
+ def pragmas(self):
+ self.code_gen_dict["$PRAGMAS$"] = ["#pragma HLS INTERFACE axis port=in0"]
+ self.code_gen_dict["$PRAGMAS$"].append("#pragma HLS INTERFACE axis port=out")
+ self.code_gen_dict["$PRAGMAS$"].append(
+ "#pragma HLS INTERFACE ap_ctrl_none port=return"
+ )
+
+ def execute_node(self, context, graph):
+ mode = self.get_nodeattr("exec_mode")
+ node = self.onnx_node
+ exp_ishape = self.get_normal_input_shape()
+ exp_oshape = self.get_normal_output_shape()
+ folded_ishape = self.get_folded_input_shape()
+ folded_oshape = self.get_folded_output_shape()
+
+ if mode == "cppsim":
+ code_gen_dir = self.get_nodeattr("code_gen_dir_cppsim")
+ elif mode == "rtlsim":
+ code_gen_dir = self.get_nodeattr("code_gen_dir_ipgen")
+ else:
+ raise Exception(
+ """Invalid value for attribute exec_mode! Is currently set to: {}
+ has to be set to one of the following value ("cppsim", "rtlsim")""".format(
+ mode
+ )
+ )
+
+ inp = context[node.input[0]]
+ assert str(inp.dtype) == "float32", "Input datatype is not float32"
+ assert (
+ inp.shape == exp_ishape
+ ), """Input shape doesn't
+ match expected shape (numInputVectors, ImgDim, ImgDim, NumChannels)."""
+ export_idt = self.get_input_datatype()
+
+ reshaped_input = inp.reshape(folded_ishape)
+ np.save(os.path.join(code_gen_dir, "input_0.npy"), reshaped_input)
+
+ if mode == "cppsim":
+ # execute the precompiled model
+ super().exec_precompiled_singlenode_model()
+ # load output npy file
+ super().npy_to_dynamic_output(context)
+ assert (
+ context[node.output[0]].shape == folded_oshape
+ ), "cppsim did not produce expected folded output shape"
+ context[node.output[0]] = context[node.output[0]].reshape(*exp_oshape)
+ elif mode == "rtlsim":
+ sim = self.get_rtlsim()
+ nbits = self.get_instream_width()
+ rtlsim_inp = npy_to_rtlsim_input(
+ "{}/input_0.npy".format(code_gen_dir), export_idt, nbits
+ )
+ super().reset_rtlsim(sim)
+ super().toggle_clk(sim)
+ rtlsim_output = self.rtlsim(sim, rtlsim_inp)
+ odt = export_idt
+ target_bits = odt.bitwidth()
+ packed_bits = self.get_outstream_width()
+ out_npy_path = "{}/output.npy".format(code_gen_dir)
+ out_shape = self.get_folded_output_shape()
+ rtlsim_output_to_npy(
+ rtlsim_output, out_npy_path, odt, out_shape, packed_bits, target_bits
+ )
+ # load and reshape output
+ output = np.load(out_npy_path)
+ output = np.asarray([output], dtype=np.float32).reshape(*exp_oshape)
+ context[node.output[0]] = output
+ else:
+ raise Exception(
+ """Invalid value for attribute exec_mode! Is currently set to: {}
+ has to be set to one of the following value ("cppsim", "rtlsim")""".format(
+ mode
+ )
+ )
+ assert (
+ context[node.output[0]].shape == exp_oshape
+ ), """Output shape doesn't match expected shape
+ (1, OutputDim, OutputDim, NumChannels)."""
diff --git a/src/finn/custom_op/fpgadataflow/globalaccpool_batch.py b/src/finn/custom_op/fpgadataflow/globalaccpool_batch.py
index 9e6c63dc510aab5f6baff9cb6326a2d0476f67a9..83152dea6cc494b8464c78605399b21b38d48b80 100644
--- a/src/finn/custom_op/fpgadataflow/globalaccpool_batch.py
+++ b/src/finn/custom_op/fpgadataflow/globalaccpool_batch.py
@@ -75,16 +75,19 @@ class GlobalAccPool_Batch(HLSCustomOp):
def get_normal_output_shape(self):
ch = self.get_nodeattr("NumChannels")
vecs = list(self.get_nodeattr("numInputVectors"))
- oshape = tuple([vecs[0]] + [ch])
+ if len(vecs) == 1:
+ oshape = tuple(vecs + [ch])
+ elif len(vecs) == 3:
+ oshape = tuple([vecs[0]] + [1, 1, ch])
return oshape
def get_folded_output_shape(self):
ch = self.get_nodeattr("NumChannels")
pe = self.get_nodeattr("PE")
- vecs = list(self.get_nodeattr("numInputVectors"))
+ unfolded_shape = list(self.get_normal_output_shape())
assert ch % pe == 0, "PE must divide NumChannels"
folds = int(ch / pe)
- oshape = tuple([vecs[0]] + [folds, pe])
+ oshape = tuple(unfolded_shape[:-1] + [folds, pe])
return oshape
def make_shape_compatible_op(self, model):
diff --git a/src/finn/custom_op/fpgadataflow/iodma.py b/src/finn/custom_op/fpgadataflow/iodma.py
new file mode 100644
index 0000000000000000000000000000000000000000..9b718ecbbc490610790b68871080de23a54f4891
--- /dev/null
+++ b/src/finn/custom_op/fpgadataflow/iodma.py
@@ -0,0 +1,346 @@
+# Copyright (c) 2020, Xilinx
+# All rights reserved.
+#
+# Redistribution and use in source and binary forms, with or without
+# modification, are permitted provided that the following conditions are met:
+#
+# * Redistributions of source code must retain the above copyright notice, this
+# list of conditions and the following disclaimer.
+#
+# * Redistributions in binary form must reproduce the above copyright notice,
+# this list of conditions and the following disclaimer in the documentation
+# and/or other materials provided with the distribution.
+#
+# * Neither the name of FINN nor the names of its
+# contributors may be used to endorse or promote products derived from
+# this software without specific prior written permission.
+#
+# THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS "AS IS"
+# AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
+# IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE ARE
+# DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT HOLDER OR CONTRIBUTORS BE LIABLE
+# FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL
+# DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR
+# SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER
+# CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY,
+# OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE
+# OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
+
+import numpy as np
+import math
+from onnx import TensorProto, helper
+from finn.core.datatype import DataType
+from finn.custom_op.fpgadataflow import HLSCustomOp
+
+
+# the IODMA inerfaces a memory-mapped AXI interface and an AXI stream
+# direction "in": pulls data from AXI-MM to AXI stream
+# direction "out": pushes data from AXI stream to AXI-MM
+
+# DMA Addressing
+# - burst mode can be "wrap" or "increment"
+# - "increment" bursts will increment the address when moving to the next image
+# - "wrap" bursts will reinitialize the address to the start address,
+# and are useful for e.g. streaming weights, where the same buffer is
+# repeatedly read into the FPGA
+# - no additional alignment restrictions beyond anything specified in the AXI spec
+
+# Interfaces
+# - AXI-MM name specified by intfName unless this is set to "" (empty, the default)
+# in which case output AXI-MM are named "out" and input AXI-MM are named "in0"
+# - AXI-MM interface width (in bits) is specified by intfWidth
+# - AXI-Stream interface width (in bits) is specified by streamWidth
+# - If inftWidth and streamWidth are not equal, the DMA core performs
+# width conversion by going up to the least common multiple of bitwidths
+# e.g. intfWidth=32b -> 96b -> sreamWidth=24b
+# - transfers occur in multiples of the AXI-MM interface width, therefore
+# the total number of bits in the tensor must be a multiple of intfWidth
+# - transfers occur in multiples of the AXI-Stream interface width, therefore
+# the total number of bits in the tensor must be a multiple of streamWidth
+# - both interface widths must be a multiple of 8b (AXI protocol requirement)
+# - in most systems, intfWidth is also restricted to a power of 2 (e.g. Vitis)
+# but this is not universal so we don't check here explicitly
+
+# Input/output tensor sizes shapes
+# - The data being moved is a tensor of shape numInputVectors+[NumChannels]
+# - The data type of the tensor elements is specified by dataType
+# - on the stream side
+# -the normal shape is the same as the ONNX tensor attached to it
+# -the folded shape is computed from the stream width and normal shape
+# - on the AXI-MM side
+# -the normal shape is the same as the one on the stream side
+# -the folded shape is not defined
+
+
+class IODMA(HLSCustomOp):
+ """Class that corresponds to finn-hlslib DMA function(s)."""
+
+ def __init__(self, onnx_node):
+ super().__init__(onnx_node)
+
+ def get_nodeattr_types(self):
+ my_attrs = {
+ "NumChannels": ("i", True, 0),
+ # FINN input datatype
+ "dataType": ("s", True, ""),
+ # Stream parameters
+ "streamWidth": ("i", False, 32),
+ # DMA-specific parameters
+ "intfWidth": ("i", False, 32),
+ "burstMode": ("s", False, "increment"),
+ "direction": ("s", False, "in"),
+ # shape describing input vecs per execution
+ "numInputVectors": ("ints", False, [1]),
+ # name of axi-mm interface
+ "intfName": ("s", False, ""),
+ }
+ my_attrs.update(super().get_nodeattr_types())
+ return my_attrs
+
+ def get_normal_input_shape(self):
+ vecs = list(self.get_nodeattr("numInputVectors"))
+ num_ch = self.get_nodeattr("NumChannels")
+ ishape = tuple(vecs + [num_ch])
+ return ishape
+
+ def get_normal_output_shape(self):
+ return self.get_normal_input_shape()
+
+ def get_folded_input_shape(self):
+ if self.get_nodeattr("direction") == "in":
+ raise ValueError("Folded input shape not defined for input IODMA")
+ else:
+ shape = list(self.get_normal_input_shape())
+ itype_bits = self.get_input_datatype().bitwidth()
+ intfw = self.get_nodeattr("streamWidth")
+ assert (
+ intfw % itype_bits == 0
+ ), "Input stream width must be a multiple of datatype bits"
+ elems_per_word = intfw // itype_bits
+ assert shape[-1] % elems_per_word == 0, "Fold depth must be integer"
+ fold_depth = shape[-1] // elems_per_word
+ shape[-1] = fold_depth
+ shape.append(elems_per_word)
+ return tuple(shape)
+
+ def get_folded_output_shape(self):
+ if self.get_nodeattr("direction") == "out":
+ raise ValueError("Folded output shape not defined for output IODMA")
+ else:
+ shape = list(self.get_normal_output_shape())
+ itype_bits = self.get_output_datatype().bitwidth()
+ intfw = self.get_nodeattr("streamWidth")
+ assert (
+ intfw % itype_bits == 0
+ ), "Input stream width must be a multiple of datatype bits"
+ elems_per_word = intfw // itype_bits
+ assert shape[-1] % elems_per_word == 0, "Fold depth must be integer"
+ fold_depth = shape[-1] // elems_per_word
+ shape[-1] = fold_depth
+ shape.append(elems_per_word)
+ return tuple(shape)
+
+ def make_shape_compatible_op(self, model):
+ exp_ishape = self.get_normal_input_shape()
+ oshape = self.get_normal_output_shape()
+ ishape = tuple(model.get_tensor_shape(self.onnx_node.input[0]))
+ assert ishape == exp_ishape, "Unexpected input shape."
+ # implement tensor with correct shape
+ values = np.random.randn(*oshape).astype(np.float32)
+ return helper.make_node(
+ "Constant",
+ inputs=[],
+ outputs=[self.onnx_node.output[0]],
+ value=helper.make_tensor(
+ name="const_tensor",
+ data_type=TensorProto.FLOAT,
+ dims=values.shape,
+ vals=values.flatten().astype(float),
+ ),
+ )
+
+ def infer_node_datatype(self, model):
+ node = self.onnx_node
+ # data type stays the same
+ dtype = model.get_tensor_datatype(node.input[0])
+ exp_idtype = self.get_input_datatype()
+ assert dtype == exp_idtype, "Unexpected datatype."
+ model.set_tensor_datatype(node.output[0], dtype)
+
+ def verify_node(self):
+ pass
+
+ def get_input_datatype(self):
+ """Returns FINN DataType of input."""
+ return DataType[self.get_nodeattr("dataType")]
+
+ def get_output_datatype(self):
+ """Returns FINN DataType of output. (Same as input datatype)"""
+ return self.get_input_datatype()
+
+ def get_instream_width(self):
+ if self.get_nodeattr("direction") == "in":
+ return self.get_nodeattr("intfWidth")
+ elif self.get_nodeattr("direction") == "out":
+ return self.get_nodeattr("streamWidth")
+ else:
+ raise ValueError("Invalid IODMA direction, please set to in or out")
+
+ def get_outstream_width(self):
+ if self.get_nodeattr("direction") == "out":
+ return self.get_nodeattr("intfWidth")
+ elif self.get_nodeattr("direction") == "in":
+ return self.get_nodeattr("streamWidth")
+ else:
+ raise ValueError("Invalid IODMA direction, please set to in or out")
+
+ def get_number_output_values(self):
+ oshape = self.get_normal_output_shape()
+ itype_bits = self.get_input_datatype().bitwidth()
+ intfw = self.get_nodeattr("intfWidth")
+ nelems = np.prod(oshape)
+ nbits = nelems * itype_bits
+ assert nbits % intfw == 0, "DMA: total transfer size must be word multiple"
+ ovalues = nbits // intfw
+ return ovalues
+
+ def global_includes(self):
+ self.code_gen_dict["$GLOBALS$"] = ['#include "dma.h"']
+ self.code_gen_dict["$GLOBALS$"].append('#include "streamtools.h"')
+
+ def defines(self, var):
+ itype_bits = self.get_input_datatype().bitwidth()
+ total_bits = itype_bits * np.prod(self.get_normal_input_shape())
+ assert total_bits % 8 == 0, "DMA input not a multiple of 1 Byte"
+ total_bytes = total_bits // 8
+ self.code_gen_dict["$DEFINES$"] = [
+ """#define NumBytes1 {}\n#define DataWidth1 {}\n""".format(
+ total_bytes, self.get_nodeattr("intfWidth")
+ )
+ ]
+
+ def get_ap_int_max_w(self):
+ "Return the maximum width of any ap_int used in this module."
+ instream = self.get_instream_width()
+ outstream = self.get_outstream_width()
+ width_lcm = (instream * outstream) // math.gcd(instream, outstream)
+ return width_lcm
+
+ def docompute(self):
+ direction = self.get_nodeattr("direction")
+ mode = self.get_nodeattr("burstMode")
+ if direction == "in":
+ if mode == "wrap":
+ func = "Mem2Stream_Batch_external_wmem"
+ else:
+ func = "Mem2Stream_Batch"
+ dwc_func = "WidthAdjustedOutputStream"
+ elif direction == "out":
+ func = "Stream2Mem_Batch"
+ dwc_func = "WidthAdjustedInputStream"
+ else:
+ raise ValueError("Invalid IODMA direction, please set to in or out")
+ # define templates for instantiation
+ dma_inst_template = func + "<DataWidth1, NumBytes1>(%s, %s, numReps);"
+ dwc_inst_template = dwc_func + "<%d, %d, %d> %s(%s, numReps);"
+ # do stream infrastructure and instantiations
+ intfw = self.get_nodeattr("intfWidth")
+ strmw = self.get_nodeattr("streamWidth")
+ width_lcm = (strmw * intfw) // math.gcd(strmw, intfw)
+ # we always need two streams: one of width_lcm, and one of intfw width
+ # because we use WidthAdjustedInputStream,
+ dtype_bits = self.get_input_datatype().bitwidth()
+ total_bits = dtype_bits * np.prod(self.get_normal_input_shape())
+ if direction == "in":
+ self.code_gen_dict["$DOCOMPUTE$"] = [
+ dwc_inst_template
+ % (width_lcm, strmw, total_bits // width_lcm, "dwc_lcm", "out"),
+ dwc_inst_template
+ % (intfw, width_lcm, total_bits // intfw, "dwc_intfw", "dwc_lcm"),
+ dma_inst_template % ("in0", "dwc_intfw"),
+ ]
+ else:
+ self.code_gen_dict["$DOCOMPUTE$"] = [
+ dwc_inst_template
+ % (strmw, width_lcm, total_bits // strmw, "dwc_lcm", "in0"),
+ dwc_inst_template
+ % (width_lcm, intfw, total_bits // width_lcm, "dwc_intfw", "dwc_lcm"),
+ dma_inst_template % ("dwc_intfw", "out"),
+ ]
+
+ def blackboxfunction(self):
+ packed_ibits = self.get_instream_width()
+ packed_hls_type_in = "ap_uint<%d>" % packed_ibits
+ packed_obits = self.get_outstream_width()
+ packed_hls_type_out = "ap_uint<%d>" % packed_obits
+ direction = self.get_nodeattr("direction")
+ if direction == "in":
+ self.code_gen_dict["$BLACKBOXFUNCTION$"] = [
+ "void %s(%s *in0, hls::stream<%s > &out, unsigned int numReps)"
+ % (self.onnx_node.name, packed_hls_type_in, packed_hls_type_out)
+ ]
+ elif direction == "out":
+ self.code_gen_dict["$BLACKBOXFUNCTION$"] = [
+ "void %s(hls::stream<%s > &in0, %s *out, unsigned int numReps)"
+ % (self.onnx_node.name, packed_hls_type_in, packed_hls_type_out)
+ ]
+ else:
+ raise ValueError("Invalid IODMA direction, please set to in or out")
+
+ def pragmas(self):
+ self.code_gen_dict["$PRAGMAS$"] = [
+ "#pragma HLS INTERFACE s_axilite port=numReps bundle=control"
+ ]
+ self.code_gen_dict["$PRAGMAS$"].append(
+ "#pragma HLS INTERFACE s_axilite port=return bundle=control"
+ )
+ direction = self.get_nodeattr("direction")
+ intfname = self.get_nodeattr("intfName")
+ if direction == "in":
+ if intfname == "":
+ self.code_gen_dict["$PRAGMAS$"].append(
+ "#pragma HLS INTERFACE m_axi offset=slave port=in0"
+ )
+ else:
+ self.code_gen_dict["$PRAGMAS$"].append(
+ "#pragma HLS INTERFACE m_axi offset=slave port=%s" % (intfname)
+ )
+ self.code_gen_dict["$PRAGMAS$"].append(
+ "#pragma HLS INTERFACE s_axilite port=in0 bundle=control"
+ )
+ self.code_gen_dict["$PRAGMAS$"].append(
+ "#pragma HLS INTERFACE axis port=out"
+ )
+ elif direction == "out":
+ self.code_gen_dict["$PRAGMAS$"].append(
+ "#pragma HLS INTERFACE axis port=in0"
+ )
+ if intfname == "":
+ self.code_gen_dict["$PRAGMAS$"].append(
+ "#pragma HLS INTERFACE m_axi offset=slave port=out"
+ )
+ else:
+ self.code_gen_dict["$PRAGMAS$"].append(
+ "#pragma HLS INTERFACE m_axi offset=slave port=%s" % (intfname)
+ )
+ self.code_gen_dict["$PRAGMAS$"].append(
+ "#pragma HLS INTERFACE s_axilite port=out bundle=control"
+ )
+ else:
+ raise ValueError("Invalid IODMA direction, please set to in or out")
+ self.code_gen_dict["$PRAGMAS$"].append("#pragma HLS DATAFLOW")
+
+ def execute_node(self, context, graph):
+ pass
+
+ def dataoutstrm(self):
+ pass
+
+ def read_npy_data(self):
+ pass
+
+ def save_as_npy(self):
+ pass
+
+ def strm_decl(self):
+ pass
diff --git a/src/finn/custom_op/fpgadataflow/pool_batch.py b/src/finn/custom_op/fpgadataflow/pool_batch.py
new file mode 100644
index 0000000000000000000000000000000000000000..c7edc24d0e24eef1154293caca2519ab3aa68358
--- /dev/null
+++ b/src/finn/custom_op/fpgadataflow/pool_batch.py
@@ -0,0 +1,395 @@
+# Copyright (c) 2020, Xilinx
+# All rights reserved.
+#
+# Redistribution and use in source and binary forms, with or without
+# modification, are permitted provided that the following conditions are met:
+#
+# * Redistributions of source code must retain the above copyright notice, this
+# list of conditions and the following disclaimer.
+#
+# * Redistributions in binary form must reproduce the above copyright notice,
+# this list of conditions and the following disclaimer in the documentation
+# and/or other materials provided with the distribution.
+#
+# * Neither the name of FINN nor the names of its
+# contributors may be used to endorse or promote products derived from
+# this software without specific prior written permission.
+#
+# THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS "AS IS"
+# AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
+# IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE ARE
+# DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT HOLDER OR CONTRIBUTORS BE LIABLE
+# FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL
+# DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR
+# SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER
+# CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY,
+# OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE
+# OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
+
+import os
+import numpy as np
+
+from finn.custom_op.fpgadataflow import HLSCustomOp
+from finn.core.datatype import DataType
+from onnx import TensorProto, helper
+from finn.util.data_packing import npy_to_rtlsim_input, rtlsim_output_to_npy
+
+
+class Pool_Batch(HLSCustomOp):
+ """Class that corresponds to finn-hlslib Pool_batch function.
+ Requires ConvolutionInputGenerator(depthwise == 1) to format its input
+
+ TODO: explain input shape (to reuse im2col code)
+ Input shape (BatchSize,OutImgDim,OutImgDim,KernelSize^2*Channels)
+ Output shape (BatchSize,OutImgDim,OutImgDim,Channels)
+
+ # note: the actual data layout produced by the hlslib kernels is different
+ # for depthwise ops.
+ # * depthwise SWG: (1, OFMDim, OFMDim, IFMChannels/PE, K, K, PE)
+
+ Channels can be folded using PE (SIMD from the input perspective)
+ TODO: doc
+ """
+
+ def get_nodeattr_types(self):
+ my_attrs = {
+ "Channels": ("i", True, 0),
+ "PE": ("i", True, 1),
+ "KernelSize": ("i", True, 0),
+ # Function:
+ # - MaxPool
+ # - AvgPool (not yet supported, but HLSLIB does)
+ # - AccPool (not yet supported, but HLSLIB does)
+ "Function": ("s", True, ""),
+ "OutImgDim": ("i", True, 0),
+ # FINN DataTypes for inputs/outputs
+ "dataType": ("s", True, ""),
+ "BatchSize": ("i", False, 1),
+ }
+
+ my_attrs.update(super().get_nodeattr_types())
+ return my_attrs
+
+ def get_input_datatype(self):
+ """Returns FINN DataType of input."""
+ return DataType[self.get_nodeattr("dataType")]
+
+ def get_output_datatype(self):
+ """Returns FINN DataType of output."""
+ fxn = self.get_nodeattr("Function")
+ if fxn == "MaxPool":
+ # Same as input
+ return DataType[self.get_nodeattr("dataType")]
+ else:
+ raise Exception("Pool_Batch doesn't currently support " + fxn)
+
+ def get_normal_input_shape(self):
+ ifm_ch = self.get_nodeattr("Channels")
+ odim = self.get_nodeattr("OutImgDim")
+ batch_size = self.get_nodeattr("BatchSize")
+ k = self.get_nodeattr("KernelSize")
+ ishape = (batch_size, odim, odim, k * k * ifm_ch)
+ return ishape
+
+ def get_folded_input_shape(self):
+ normal_ishape = list(self.get_normal_input_shape())
+ ifm_ch = self.get_nodeattr("Channels")
+ pe = self.get_nodeattr("PE")
+ assert ifm_ch % pe == 0, "PE must divide input channels"
+ fold = int(normal_ishape[-1] / pe)
+ folded_ishape = normal_ishape[:-1] + [fold, pe]
+ return tuple(folded_ishape)
+
+ def get_normal_output_shape(self):
+ ofm_ch = self.get_nodeattr("Channels")
+ odim = self.get_nodeattr("OutImgDim")
+ batch_size = self.get_nodeattr("BatchSize")
+ oshape = (batch_size, odim, odim, ofm_ch)
+ return oshape
+
+ def get_folded_output_shape(self):
+ normal_oshape = list(self.get_normal_output_shape())
+ ifm_ch = self.get_nodeattr("Channels")
+ pe = self.get_nodeattr("PE")
+ assert ifm_ch % pe == 0, "PE must divide input channels"
+ fold = int(ifm_ch / pe)
+ folded_oshape = normal_oshape[:-1] + [fold, pe]
+ return tuple(folded_oshape)
+
+ def get_number_output_values(self):
+ folded_oshape = self.get_folded_output_shape()
+ return np.prod(folded_oshape[1:-1])
+
+ def get_instream_width(self):
+ dt_bits = self.get_input_datatype().bitwidth()
+ pe = self.get_nodeattr("PE")
+ # ofm_ch = self.get_nodeattr("Channels")
+ # k = self.get_nodeattr("KernelSize")
+ # assert ifm_ch % pe == 0, "PE must divide input channels"
+ # simd = int(ifm_ch/pe)
+ in_width = int(dt_bits * pe)
+ return in_width
+
+ def get_outstream_width(self):
+ fxn = self.get_nodeattr("Function")
+ if fxn == "MaxPool":
+ return self.get_instream_width()
+ else:
+ raise Exception("Pool_Batch doesn't currently support " + fxn)
+
+ def make_shape_compatible_op(self, model):
+ exp_ishape = self.get_normal_input_shape()
+ oshape = self.get_normal_output_shape()
+ ishape = tuple(model.get_tensor_shape(self.onnx_node.input[0]))
+ assert ishape == exp_ishape, "Unexpected input shape for Pool_Batch."
+ # implement tensor with correct shape
+ values = np.random.randn(*oshape).astype(np.float32)
+ return helper.make_node(
+ "Constant",
+ inputs=[],
+ outputs=[self.onnx_node.output[0]],
+ value=helper.make_tensor(
+ name="const_tensor",
+ data_type=TensorProto.FLOAT,
+ dims=values.shape,
+ vals=values.flatten().astype(float),
+ ),
+ )
+
+ def infer_node_datatype(self, model):
+ node = self.onnx_node
+ # data type stays the same
+ dtype = self.get_output_datatype()
+ model.set_tensor_datatype(node.output[0], dtype)
+
+ def verify_node(self):
+ info_messages = []
+
+ # verify that "domain" is set to "finn"
+ domain_value = self.onnx_node.domain
+ if domain_value == "finn":
+ info_messages.append("Attribute domain is set correctly")
+ else:
+ info_messages.append('Attribute domain should be set to "finn"')
+
+ # verify that "backend" is set to "fpgadataflow"
+ backend_value = self.get_nodeattr("backend")
+ if backend_value == "fpgadataflow":
+ info_messages.append("Attribute backend is set correctly")
+ else:
+ info_messages.append('Attribute backend should be set to "fpgadataflow"')
+
+ # verify the number of inputs
+ if len(self.onnx_node.input) == 1:
+ info_messages.append("The number of inputs is correct")
+ else:
+ info_messages.append("""Pool_Batch needs 1 data input""")
+
+ # check supported function
+ fnx = self.get_nodeattr("Function")
+ if fnx == "MaxPool":
+ info_messages.append(
+ "Attribute Function contains a supported pool function"
+ )
+ else:
+ info_messages.append(
+ "Attribute Function contains an unsupported pool function"
+ )
+ return info_messages
+
+ def global_includes(self):
+ self.code_gen_dict["$GLOBALS$"] = ['#include "maxpool.h"']
+ self.code_gen_dict["$GLOBALS$"] += ['#include "pool.hpp"']
+
+ def defines(self, var):
+ self.code_gen_dict["$DEFINES$"] = []
+
+ ifm_ch = self.get_nodeattr("Channels")
+ self.code_gen_dict["$DEFINES$"] += ["#define Channels {}".format(ifm_ch)]
+
+ pe = self.get_nodeattr("PE")
+ self.code_gen_dict["$DEFINES$"] += ["#define PE {}".format(pe)]
+
+ k = self.get_nodeattr("KernelSize")
+ self.code_gen_dict["$DEFINES$"] += ["#define KernelSize {}".format(k)]
+
+ odim = self.get_nodeattr("OutImgDim")
+ self.code_gen_dict["$DEFINES$"] += ["#define OFMDim {}".format(odim)]
+
+ numReps = self.get_nodeattr("BatchSize")
+ self.code_gen_dict["$DEFINES$"] += ["#define numReps {}".format(numReps)]
+
+ def read_npy_data(self):
+ code_gen_dir = self.get_nodeattr("code_gen_dir_cppsim")
+ dtype = self.get_input_datatype()
+ if dtype == DataType.BIPOLAR:
+ # use binary for bipolar storage
+ dtype = DataType.BINARY
+ elem_bits = dtype.bitwidth()
+ packed_bits = self.get_instream_width()
+ packed_hls_type = "ap_uint<%d>" % packed_bits
+ elem_hls_type = dtype.get_hls_datatype_str()
+ npy_type = "float"
+ npy_in = "%s/input_0.npy" % code_gen_dir
+ self.code_gen_dict["$READNPYDATA$"] = []
+ self.code_gen_dict["$READNPYDATA$"].append(
+ 'npy2apintstream<%s, %s, %d, %s>("%s", in0,false);'
+ % (packed_hls_type, elem_hls_type, elem_bits, npy_type, npy_in)
+ )
+
+ def strm_decl(self):
+ self.code_gen_dict["$STREAMDECLARATIONS$"] = []
+ self.code_gen_dict["$STREAMDECLARATIONS$"].append(
+ 'hls::stream<ap_uint<{}>> in0 ("in0");'.format(self.get_instream_width())
+ )
+ self.code_gen_dict["$STREAMDECLARATIONS$"].append(
+ 'hls::stream<ap_uint<{}>> out ("out");'.format(self.get_outstream_width())
+ )
+
+ def docompute(self):
+ idt = self.get_input_datatype()
+ i_hls_dt = idt.get_hls_datatype_str()
+ odt = self.get_output_datatype()
+ o_hls_dt = odt.get_hls_datatype_str()
+
+ self.code_gen_dict["$DOCOMPUTE$"] = []
+
+ fxn = self.get_nodeattr("Function")
+ if fxn == "MaxPool":
+ self.code_gen_dict["$DOCOMPUTE$"] += [
+ "MaxPoolFunction<{},KernelSize> pool_fxn;".format(i_hls_dt)
+ ]
+ else:
+ raise Exception("Pool_Batch doesn't currently support " + fxn)
+
+ self.code_gen_dict["$DOCOMPUTE$"] += [
+ """Pool_batch<Channels, PE, KernelSize,Slice<{} >, Slice< {} > >
+ (in0,out, pool_fxn, OFMDim*OFMDim*numReps);""".format(
+ i_hls_dt, o_hls_dt
+ )
+ ]
+
+ def dataoutstrm(self):
+ code_gen_dir = self.get_nodeattr("code_gen_dir_cppsim")
+ dtype = self.get_output_datatype()
+ if dtype == DataType.BIPOLAR:
+ # use binary for bipolar storage
+ dtype = DataType.BINARY
+ elem_bits = dtype.bitwidth()
+ packed_bits = self.get_outstream_width()
+ packed_hls_type = "ap_uint<%d>" % packed_bits
+ elem_hls_type = dtype.get_hls_datatype_str()
+ npy_type = "float"
+ npy_out = "%s/output.npy" % code_gen_dir
+ oshape = self.get_folded_output_shape()
+ oshape_cpp_str = str(oshape).replace("(", "{").replace(")", "}")
+
+ self.code_gen_dict["$DATAOUTSTREAM$"] = [
+ 'apintstream2npy<%s, %s, %d, %s>(out, %s, "%s",false);'
+ % (
+ packed_hls_type,
+ elem_hls_type,
+ elem_bits,
+ npy_type,
+ oshape_cpp_str,
+ npy_out,
+ )
+ ]
+
+ def save_as_npy(self):
+ self.code_gen_dict["$SAVEASCNPY$"] = []
+
+ def blackboxfunction(self):
+ packed_ibits = self.get_instream_width()
+ packed_in_hls_type = "ap_uint<%d>" % packed_ibits
+
+ packed_obits = self.get_outstream_width()
+ packed_out_hls_type = "ap_uint<%d>" % packed_obits
+ self.code_gen_dict["$BLACKBOXFUNCTION$"] = [
+ "void %s(hls::stream<%s > &in0, hls::stream<%s > &out)"
+ % (self.onnx_node.name, packed_in_hls_type, packed_out_hls_type)
+ ]
+
+ def pragmas(self):
+ self.code_gen_dict["$PRAGMAS$"] = ["#pragma HLS INTERFACE axis port=in0"]
+ self.code_gen_dict["$PRAGMAS$"].append("#pragma HLS INTERFACE axis port=out")
+ self.code_gen_dict["$PRAGMAS$"].append(
+ "#pragma HLS INTERFACE ap_ctrl_none port=return"
+ )
+
+ def execute_node(self, context, graph):
+ mode = self.get_nodeattr("exec_mode")
+ node = self.onnx_node
+ exp_ishape = self.get_normal_input_shape()
+ folded_ishape = self.get_folded_input_shape()
+ exp_oshape = self.get_normal_output_shape()
+ folded_oshape = self.get_folded_output_shape()
+
+ # TODO ensure codegen dir exists
+ if mode == "cppsim":
+ code_gen_dir = self.get_nodeattr("code_gen_dir_cppsim")
+ elif mode == "rtlsim":
+ code_gen_dir = self.get_nodeattr("code_gen_dir_ipgen")
+ else:
+ raise Exception(
+ """Invalid value for attribute exec_mode! Is currently set to: {}
+ has to be set to one of the following value ("cppsim", "rtlsim")""".format(
+ mode
+ )
+ )
+
+ inp = context[node.input[0]]
+
+ assert str(inp.dtype) == "float32", "Input datatype is not float32"
+ assert (
+ inp.shape == exp_ishape
+ ), """Input shape doesn't
+ match expected shape (batch_size,odim,odim,k*k*ifm_ch)."""
+
+ export_idt = self.get_input_datatype()
+ reshaped_input = inp.reshape(folded_ishape)
+
+ np.save(os.path.join(code_gen_dir, "input_0.npy"), reshaped_input)
+
+ if mode == "cppsim":
+ # execute the precompiled model
+ super().exec_precompiled_singlenode_model()
+ # load output npy file
+ super().npy_to_dynamic_output(context)
+ assert (
+ context[node.output[0]].shape == folded_oshape
+ ), "cppsim did not produce expected folded output shape"
+ context[node.output[0]] = context[node.output[0]].reshape(*exp_oshape)
+ elif mode == "rtlsim":
+ sim = self.get_rtlsim()
+ nbits = self.get_instream_width()
+ rtlsim_inp = npy_to_rtlsim_input(
+ "{}/input_0.npy".format(code_gen_dir), export_idt, nbits
+ )
+ super().reset_rtlsim(sim)
+ super().toggle_clk(sim)
+ rtlsim_output = self.rtlsim(sim, rtlsim_inp)
+ odt = export_idt
+ target_bits = odt.bitwidth()
+ packed_bits = self.get_outstream_width()
+ out_npy_path = "{}/output.npy".format(code_gen_dir)
+ out_shape = self.get_folded_output_shape()
+ rtlsim_output_to_npy(
+ rtlsim_output, out_npy_path, odt, out_shape, packed_bits, target_bits
+ )
+ # load and reshape output
+ output = np.load(out_npy_path)
+ output = np.asarray([output], dtype=np.float32).reshape(*exp_oshape)
+ context[node.output[0]] = output
+ else:
+ raise Exception(
+ """Invalid value for attribute exec_mode! Is currently set to: {}
+ has to be set to one of the following value ("cppsim", "rtlsim")""".format(
+ mode
+ )
+ )
+
+ assert (
+ context[node.output[0]].shape == exp_oshape
+ ), """Output
+ shape doesn't match expected shape (1, ofm_dim, ofm_dim, k*k*ifm_ch)."""
diff --git a/src/finn/custom_op/fpgadataflow/streamingfclayer_batch.py b/src/finn/custom_op/fpgadataflow/streamingfclayer_batch.py
index b6c992e4b5ea1ced088928b3d9a4db381d82db22..a7ebff68749120868cae9ce5ac18d2856fe2cb8a 100644
--- a/src/finn/custom_op/fpgadataflow/streamingfclayer_batch.py
+++ b/src/finn/custom_op/fpgadataflow/streamingfclayer_batch.py
@@ -240,11 +240,21 @@ class StreamingFCLayer_Batch(HLSCustomOp):
Q = self.get_nodeattr("SIMD")
wdt = self.get_weight_datatype()
W = wdt.bitwidth()
- D_in = self.get_instream_width()
- D_out = self.get_outstream_width()
+ D_in = self.get_nodeattr("MW")
+ D_out = self.get_nodeattr("MH")
omega = (D_in * D_out) / (Q * P)
return P * (math.ceil(omega / 512)) * (math.ceil((Q * W) / 36))
+ def bram_efficiency_estimation(self):
+ wdt = self.get_weight_datatype()
+ W = wdt.bitwidth()
+ D_in = self.get_nodeattr("MW")
+ D_out = self.get_nodeattr("MH")
+ bram16_est = self.bram_estimation()
+ wbits = W * D_in * D_out
+ bram16_est_capacity = bram16_est * 36 * 512
+ return wbits / bram16_est_capacity
+
def lut_estimation(self):
"""Calculates resource estimations for LUTs based on:
- FINN-R: An End-to-End Deep-Learning Framework for Fast
diff --git a/src/finn/custom_op/im2col.py b/src/finn/custom_op/im2col.py
index 82a6b140f7af1be4e5c0f429d077b99c7865383e..8ed0041704d421dab587f08bcbcd9e739e8434e9 100644
--- a/src/finn/custom_op/im2col.py
+++ b/src/finn/custom_op/im2col.py
@@ -80,6 +80,8 @@ class Im2Col(CustomOp):
"input_shape": ("s", True, ""),
"pad_amount": ("i", False, 0),
"pad_value": ("i", False, 0),
+ # depthwise: if != 0, infer ConvolutionInputGenerator with depthwise == 1
+ "depthwise": ("i", False, 0),
}
def make_shape_compatible_op(self, model):
diff --git a/src/finn/custom_op/registry.py b/src/finn/custom_op/registry.py
index 3bfdebed58b3995cfeb6e97ed2836d9ee6a0da79..e4317e02d46df90c8fd0c8854262ca6eb0ea4f48 100644
--- a/src/finn/custom_op/registry.py
+++ b/src/finn/custom_op/registry.py
@@ -31,6 +31,7 @@
from finn.custom_op.fpgadataflow.convolutioninputgenerator import (
ConvolutionInputGenerator,
)
+from finn.custom_op.fpgadataflow.downsampler import DownSampler
from finn.custom_op.fpgadataflow.streamingfclayer_batch import StreamingFCLayer_Batch
from finn.custom_op.fpgadataflow.streamingmaxpool_batch import StreamingMaxPool_Batch
from finn.custom_op.fpgadataflow.streamingfifo import StreamingFIFO
@@ -44,17 +45,21 @@ from finn.custom_op.fpgadataflow.streamingdatawidthconverter_batch import (
StreamingDataWidthConverter_Batch,
)
from finn.custom_op.fpgadataflow.globalaccpool_batch import GlobalAccPool_Batch
+from finn.custom_op.fpgadataflow.pool_batch import Pool_Batch
from finn.custom_op.fpgadataflow.fmpadding_batch import FMPadding_Batch
from finn.custom_op.fpgadataflow.thresholding_batch import Thresholding_Batch
from finn.custom_op.fpgadataflow.addstreams_batch import AddStreams_Batch
from finn.custom_op.fpgadataflow.labelselect_batch import LabelSelect_Batch
from finn.custom_op.quantavgpool2d import QuantAvgPool2d
from finn.custom_op.fpgadataflow.duplicatestreams_batch import DuplicateStreams_Batch
+from finn.custom_op.fpgadataflow.channelwise_op_batch import ChannelwiseOp_Batch
+from finn.custom_op.fpgadataflow.iodma import IODMA
# create a mapping of all known CustomOp names and classes
custom_op = {}
custom_op["MultiThreshold"] = MultiThreshold
+custom_op["DownSampler"] = DownSampler
custom_op["XnorPopcountMatMul"] = XnorPopcountMatMul
custom_op["Im2Col"] = Im2Col
custom_op["StreamingMaxPool_Batch"] = StreamingMaxPool_Batch
@@ -66,12 +71,15 @@ custom_op["MaxPoolNHWC"] = MaxPoolNHWC
custom_op["StreamingDataWidthConverter_Batch"] = StreamingDataWidthConverter_Batch
custom_op["StreamingFIFO"] = StreamingFIFO
custom_op["GlobalAccPool_Batch"] = GlobalAccPool_Batch
+custom_op["Pool_Batch"] = Pool_Batch
custom_op["FMPadding_Batch"] = FMPadding_Batch
custom_op["Thresholding_Batch"] = Thresholding_Batch
custom_op["AddStreams_Batch"] = AddStreams_Batch
custom_op["LabelSelect_Batch"] = LabelSelect_Batch
custom_op["QuantAvgPool2d"] = QuantAvgPool2d
custom_op["DuplicateStreams_Batch"] = DuplicateStreams_Batch
+custom_op["ChannelwiseOp_Batch"] = ChannelwiseOp_Batch
+custom_op["IODMA"] = IODMA
def getCustomOp(node):
diff --git a/src/finn/transformation/bipolar_to_xnor.py b/src/finn/transformation/bipolar_to_xnor.py
index 8b65cfee17edd5d89fcca0bd86da12415d38fe78..80f2a73351f8548c99efd8dedd8a04d44c8558a3 100644
--- a/src/finn/transformation/bipolar_to_xnor.py
+++ b/src/finn/transformation/bipolar_to_xnor.py
@@ -27,6 +27,7 @@
# OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
import numpy as np
+import warnings
from onnx import TensorProto
from onnx import helper as oh
@@ -66,26 +67,40 @@ class ConvertBipolarMatMulToXnorPopcount(Transformation):
mt_chain = model.find_upstream(mm_input, find_prod_mt)
if len(mt_chain) == 0:
- raise Exception(
- """Could not find upstream bipolar
- MultiThreshold"""
- )
- graph_modified = True
- mt = mt_chain[-1]
- mt_inst = getCustomOp(mt)
- # ensure old scale/bias were correct for BIPOLAR
- scale_ok = mt_inst.get_nodeattr("out_scale") == 2.0
- bias_ok = mt_inst.get_nodeattr("out_bias") == -1.0
- assert (
- scale_ok and bias_ok
- ), """Unexpected scale/bias
- attributes for BIPOLAR MultiThreshold node."""
- # start conversion, set MT output to binary
- # (this is what XnorPopcountMatMul expects)
- mt_inst.set_nodeattr("out_dtype", "BINARY")
- mt_inst.set_nodeattr("out_scale", 1.0)
- mt_inst.set_nodeattr("out_bias", 0.0)
- model.set_tensor_datatype(mm_input, DataType.BINARY)
+ if mm_input == graph.input[0].name:
+ # change input datatype to BINARY
+ model.set_tensor_datatype(mm_input, DataType.BINARY)
+ graph_modified = True
+ warnings.warn(
+ """IMPORTANT: Changing graph input DataType
+ to BINARY instead of BIPOLAR. Ensure this is respected
+ when checking for correctness.
+ """
+ )
+ else:
+ raise Exception(
+ """Could not find upstream bipolar
+ MultiThreshold, and the MatMul is not the
+ first node on graph input. Unable to convert
+ input tensor from BIPOLAR to BINARY."""
+ )
+ else:
+ graph_modified = True
+ mt = mt_chain[-1]
+ mt_inst = getCustomOp(mt)
+ # ensure old scale/bias were correct for BIPOLAR
+ scale_ok = mt_inst.get_nodeattr("out_scale") == 2.0
+ bias_ok = mt_inst.get_nodeattr("out_bias") == -1.0
+ assert (
+ scale_ok and bias_ok
+ ), """Unexpected scale/bias
+ attributes for BIPOLAR MultiThreshold node."""
+ # start conversion, set MT output to binary
+ # (this is what XnorPopcountMatMul expects)
+ mt_inst.set_nodeattr("out_dtype", "BINARY")
+ mt_inst.set_nodeattr("out_scale", 1.0)
+ mt_inst.set_nodeattr("out_bias", 0.0)
+ model.set_tensor_datatype(mm_input, DataType.BINARY)
# change node type and domain
n.op_type = "XnorPopcountMatMul"
n.domain = "finn"
diff --git a/src/finn/transformation/fpgadataflow/annotate_resources.py b/src/finn/transformation/fpgadataflow/annotate_resources.py
index 207075b00de1871da19ea78472125d435449ed6e..62ee92df54eee2b63d84657515d7fbc3a8808b81 100644
--- a/src/finn/transformation/fpgadataflow/annotate_resources.py
+++ b/src/finn/transformation/fpgadataflow/annotate_resources.py
@@ -69,6 +69,9 @@ class AnnotateResources(Transformation):
total_dict[r_type] += r_amount
else:
total_dict[r_type] = r_amount
+ for k in total_dict.keys():
+ if "efficiency" in k:
+ total_dict[k] = total_dict[k] / len(graph.node)
model.set_metadata_prop("res_total_" + self.mode, str(total_dict))
for node in graph.node:
if _is_fpgadataflow_node(node) and node.name in res_dict.keys():
diff --git a/src/finn/transformation/fpgadataflow/convert_to_hls_layers.py b/src/finn/transformation/fpgadataflow/convert_to_hls_layers.py
index d421a5f3ef8ca980b399087de1482b2ae913da1b..34a697a43426aae0f984770689552063aa35b9e8 100644
--- a/src/finn/transformation/fpgadataflow/convert_to_hls_layers.py
+++ b/src/finn/transformation/fpgadataflow/convert_to_hls_layers.py
@@ -27,6 +27,7 @@
# OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
from onnx import helper, TensorProto
+import numpy as np
from finn.core.datatype import DataType
from finn.transformation import Transformation
@@ -34,6 +35,10 @@ from finn.custom_op.registry import getCustomOp
from finn.transformation.infer_shapes import InferShapes
from finn.transformation.infer_datatypes import InferDataTypes
import finn.core.data_layout as DataLayout
+from finn.util.onnx import nchw_to_nhwc
+import warnings
+from finn.util.basic import get_by_name
+import warnings
class InferConvInpGen(Transformation):
@@ -51,11 +56,15 @@ class InferConvInpGen(Transformation):
i2c_in_shape = model.get_tensor_shape(i2c_input)
i2c_out_shape = model.get_tensor_shape(i2c_output)
dt = model.get_tensor_datatype(i2c_input)
+ if not dt.is_integer():
+ warnings.warn("Input is not int. Can't infer ConvInpGen")
+ continue
i2c_inst = getCustomOp(n)
stride = i2c_inst.get_nodeattr("stride")
k = i2c_inst.get_nodeattr("kernel_size")
pad = i2c_inst.get_nodeattr("pad_amount")
pad_val = i2c_inst.get_nodeattr("pad_value")
+ depthwise = i2c_inst.get_nodeattr("depthwise")
ifm_ch = i2c_in_shape[-1]
ifm_dim = i2c_in_shape[1]
ofm_dim = i2c_out_shape[1]
@@ -67,7 +76,11 @@ class InferConvInpGen(Transformation):
if pad > 0:
# if padding enabled, ensure pad_val supported by DataType
- assert dt.allowed(pad_val), "Im2Col DataType must support pad_val"
+ # assert dt.allowed(pad_val),"""FMPadding_Batch DataType
+ # must support pad_val"""
+ assert (
+ pad_val == 0
+ ), "FMPadding_Batch doesn't currently support pad_val!= 0"
odim_padding = ifm_dim + 2 * pad
@@ -97,23 +110,40 @@ class InferConvInpGen(Transformation):
)
graph.node.insert(node_ind, padding_node)
- # create equivalent ConvolutionInputGenerator node
- ConvInpGen_node = helper.make_node(
- "ConvolutionInputGenerator",
- [ConvInpGen_input],
- [i2c_output],
- domain="finn",
- backend="fpgadataflow",
- ConvKernelDim=k,
- IFMChannels=ifm_ch,
- IFMDim=ConvInpGen_idim,
- OFMDim=ofm_dim,
- SIMD=ifm_ch,
- Stride=stride,
- inputDataType=dt.name,
- outputDataType=dt.name,
- )
- graph.node.insert(ConvInpGen_node_idx, ConvInpGen_node)
+ if stride > 1 and k == 1:
+ # create DownSampler node
+ ConvInpGen_node = helper.make_node(
+ "DownSampler",
+ [ConvInpGen_input],
+ [i2c_output],
+ domain="finn",
+ backend="fpgadataflow",
+ ImgDim=ConvInpGen_idim,
+ NumChannels=ifm_ch,
+ SIMD=ifm_ch,
+ Stride=stride,
+ inputDataType=dt.name,
+ )
+ graph.node.insert(ConvInpGen_node_idx, ConvInpGen_node)
+ else:
+ # create equivalent ConvolutionInputGenerator node
+ ConvInpGen_node = helper.make_node(
+ "ConvolutionInputGenerator",
+ [ConvInpGen_input],
+ [i2c_output],
+ domain="finn",
+ backend="fpgadataflow",
+ ConvKernelDim=k,
+ IFMChannels=ifm_ch,
+ IFMDim=ConvInpGen_idim,
+ OFMDim=ofm_dim,
+ SIMD=ifm_ch,
+ Stride=stride,
+ inputDataType=dt.name,
+ outputDataType=dt.name,
+ depthwise=depthwise,
+ )
+ graph.node.insert(ConvInpGen_node_idx, ConvInpGen_node)
# remove old nodes
graph.node.remove(n)
graph_modified = True
@@ -169,6 +199,137 @@ class InferStreamingMaxPool(Transformation):
return (model, graph_modified)
+class InferPool_Batch(Transformation):
+ """If kernel_shape > strides, replace Pool layer with with of Im2col
+ + pool(with kernel_shape == strides), plus Transpose layers to keep the original
+ data layout."""
+
+ def apply(self, model):
+ graph = model.graph
+ node_ind = 0
+ graph_modified = False
+ for n in graph.node:
+ node_ind += 1
+ if n.op_type in ["MaxPool"]:
+ # extract pool parameters
+ k = get_by_name(n.attribute, "kernel_shape").ints[-1]
+ stride = get_by_name(n.attribute, "strides").ints[-1]
+
+ if k <= stride:
+ continue
+
+ try:
+ pad = get_by_name(n.attribute, "pads").ints[-1]
+ except AttributeError:
+ pad = 0
+
+ node_input = n.input[0]
+ node_output = n.output[0]
+ idt = model.get_tensor_datatype(node_input)
+ if not idt.is_integer():
+ continue
+
+ # odt = model.get_tensor_datatype(node_output)
+
+ ifm_ch = model.get_tensor_shape(n.input[0])[1] # assume NCHW
+ ofm_ch = ifm_ch
+ ifm_dim = model.get_tensor_shape(n.input[0])[-1] # assume NCHW
+ ofm_dim = model.get_tensor_shape(n.output[0])[-1] # assume NCHW
+ # create new intermediate values
+ inp_trans_out = helper.make_tensor_value_info(
+ model.make_new_valueinfo_name(),
+ TensorProto.FLOAT,
+ (1, ifm_dim, ifm_dim, ifm_ch), # NHWC
+ )
+ graph.value_info.append(inp_trans_out)
+ inp_trans_out = inp_trans_out.name
+ model.set_tensor_datatype(inp_trans_out, idt)
+
+ im2col_out = helper.make_tensor_value_info(
+ model.make_new_valueinfo_name(),
+ TensorProto.FLOAT,
+ (1, ofm_dim, ofm_dim, ifm_ch * k * k),
+ )
+ graph.value_info.append(im2col_out)
+ im2col_out = im2col_out.name
+ model.set_tensor_datatype(im2col_out, idt)
+
+ pool_output = helper.make_tensor_value_info(
+ model.make_new_valueinfo_name(),
+ TensorProto.FLOAT,
+ (1, ofm_dim, ofm_dim, ofm_ch),
+ )
+ graph.value_info.append(pool_output)
+ pool_output = pool_output.name
+ # model.set_tensor_datatype(pool_output, odt)
+
+ # create new nodes
+ # NCHW -> NHWC
+ inp_trans_node = helper.make_node(
+ "Transpose", [node_input], [inp_trans_out], perm=[0, 2, 3, 1]
+ )
+
+ if n.op_type == "MaxPool":
+ pool_fxn = "MaxPool"
+ pad_value = idt.min()
+ else:
+ raise Exception(
+ "pad_value and pool_fxn not configured for {}".format(n.op_type)
+ )
+
+ # format input tensor
+ im2col_node = helper.make_node(
+ "Im2Col",
+ [inp_trans_out],
+ [im2col_out],
+ domain="finn",
+ stride=stride,
+ kernel_size=k,
+ pad_amount=pad,
+ pad_value=pad_value,
+ depthwise=1,
+ input_shape="(1,{},{},{})".format(ifm_dim, ifm_dim, ifm_ch),
+ )
+
+ # Warning PE has to be equal to ifm_ch until Im2Col is replaced by
+ # ConvolutionInputGenerator with depthwise=1.
+ # For other settings the output will be incorrect due to incorrect input
+ # data layout
+ pool_node = helper.make_node(
+ "Pool_Batch",
+ [im2col_out],
+ [pool_output],
+ domain="finn",
+ backend="fpgadataflow",
+ dataType=idt.name,
+ Channels=ifm_ch,
+ PE=ifm_ch,
+ KernelSize=k,
+ Function=pool_fxn,
+ OutImgDim=ofm_dim,
+ BatchSize=1,
+ )
+
+ # NHWC -> NCHW
+ out_trans_node = helper.make_node(
+ "Transpose", [pool_output], [node_output], perm=[0, 3, 1, 2]
+ )
+
+ # insert nodes where the conv is to preserve topological ordering
+ graph.node.insert(node_ind, inp_trans_node)
+ graph.node.insert(node_ind + 1, im2col_node)
+ graph.node.insert(node_ind + 2, pool_node)
+ graph.node.insert(node_ind + 3, out_trans_node)
+ # remove old node
+ graph.node.remove(n)
+ graph_modified = True
+
+ if graph_modified:
+ model = model.transform(InferShapes())
+ model = model.transform(InferDataTypes())
+ return (model, graph_modified)
+
+
class InferBinaryStreamingFCLayer(Transformation):
"""Convert XnorPopcountMatMul layers to
StreamingFCLayer_Batch layers. Any immediately following MultiThreshold
@@ -489,3 +650,243 @@ class InferThresholdingLayer(Transformation):
model = model.transform(InferShapes())
model = model.transform(InferDataTypes())
return (model, graph_modified)
+
+
+class InferChannelwiseLinearLayer(Transformation):
+ """Convert any channel-wise Add/Mul into a HLS layer."""
+
+ def get_smallest_possible(self, vals):
+ """Returns smallest (fewest bits) possible DataType that can represent
+ value. Prefers unsigned integers where possible."""
+ vals = np.array(vals)
+ for v in vals:
+ assert int(v) == v, "Error float value"
+
+ for k in DataType.__members__:
+ dt = DataType[k]
+
+ if dt in [DataType.BIPOLAR, DataType.TERNARY, DataType.FLOAT32]:
+ # not currently supported
+ continue
+
+ if (dt.min() <= vals).all() and (vals <= dt.max()).all():
+ return dt
+
+ warnings.warn(
+ """InferChannelwiseLinearLayer: Output values may not be
+ representable with supported data types.
+ Setting maximum width data type available.
+ This will lead to errors if there are no constrains on the input
+ """
+ )
+
+ if (0 <= vals).all():
+ return DataType.UINT32
+ else:
+ return DataType.INT32
+
+ def apply(self, model):
+ graph = model.graph
+ node_ind = 0
+ graph_modified = False
+ for node in graph.node:
+ node_ind += 1
+ if node.op_type == "Add" or node.op_type == "Mul":
+ # assuming input[0] is dynamic
+ ll_input = node.input[0]
+ ll_output = node.output[0]
+ ll_in_shape = model.get_tensor_shape(ll_input)
+
+ # check if input 1 has an initializer
+ ll_const = node.input[1]
+ if ll_const is not None:
+ ll_cinit = model.get_initializer(ll_const)
+ if ll_cinit is None:
+ # input 1 is also dynamic
+ continue
+ else:
+ continue
+
+ # get number of channels and channel index from input
+ ll_in_layout = model.get_tensor_layout(ll_input)
+ if ll_in_layout == DataLayout.NHWC or ll_in_layout == DataLayout.NC:
+ ch_index = -1
+ ch = ll_in_shape[-1]
+ elif ll_in_layout == DataLayout.NCHW:
+ ch_index = 1
+ ch = ll_in_shape[1]
+ else:
+ continue
+
+ # check if the shape of initializer is compatible
+ ll_cinit_shape = list(ll_cinit.shape)
+ if np.prod(ll_cinit_shape) == 1:
+ warnings.warn(
+ "Broadcasting " + str(node.op_type) + "(" + node.name + ")"
+ )
+ ll_cinit = np.full((ch), ll_cinit.flatten()[0])
+ elif np.prod(ll_cinit_shape) != ch or ll_cinit_shape[ch_index] != ch:
+ # parameter shape not compatible with Channelwise_batch
+ continue
+
+ # check initializer contains integers as floats
+ if not (ll_cinit.astype(np.int32) == ll_cinit).all():
+ continue
+ # all initializer conditions are met
+
+ # check inputs
+ idt = model.get_tensor_datatype(ll_input)
+ if not idt.is_integer():
+ # skip conversion for layers with float input
+ continue
+
+ # check layout of inputs/outputs, and convert if needed
+ # check layout and convert if necessary
+ if ll_in_layout == DataLayout.NCHW:
+ ll_input = nchw_to_nhwc(ll_input, model, node_ind)
+ node_ind += 1
+ ll_in_shape = model.get_tensor_shape(ll_input)
+
+ # keep track of where we need to insert the HLS Op
+ # it has to be ahead of the output transform
+ insert_point = node_ind
+ ll_output_layout = model.get_tensor_layout(ll_output)
+ if ll_output_layout == DataLayout.NCHW:
+ ll_output = nchw_to_nhwc(ll_output, model, node_ind, reverse=True)
+ node_ind += 1
+
+ # get parameter data type
+ param_min = min(ll_cinit.flatten())
+ param_max = max(ll_cinit.flatten())
+ pdt = self.get_smallest_possible([param_min, param_max])
+
+ # set function and determine output data type
+ if node.op_type == "Add":
+ func = "add"
+ out_min = idt.min() + param_min
+ out_max = idt.max() + param_max
+ odt = self.get_smallest_possible([out_min, out_max])
+ elif node.op_type == "Mul":
+ func = "mul"
+ possible_limits = []
+ possible_limits += [idt.min() * param_min]
+ possible_limits += [idt.min() * param_max]
+ possible_limits += [idt.max() * param_min]
+ possible_limits += [idt.max() * param_max]
+ odt = self.get_smallest_possible(possible_limits)
+
+ model.set_initializer(ll_const, ll_cinit.reshape(ch))
+ model.set_tensor_datatype(ll_output, odt)
+
+ # create node with no parallelization first
+ pe = 1
+ assert ch % pe == 0, "Requirement IFC divisable by PE is violated."
+ # create and insert node
+ new_node = helper.make_node(
+ "ChannelwiseOp_Batch",
+ [ll_input, ll_const],
+ [ll_output],
+ domain="finn",
+ backend="fpgadataflow",
+ Func=func,
+ NumChannels=ch,
+ PE=pe,
+ inputDataType=idt.name,
+ paramDataType=pdt.name,
+ outputDataType=odt.name,
+ numInputVectors=list(ll_in_shape[:-1]),
+ )
+ graph.node.insert(insert_point, new_node)
+ # remove old node
+ graph.node.remove(node)
+ graph_modified = True
+
+ if graph_modified:
+ model = model.transform(InferShapes())
+ model = model.transform(InferDataTypes())
+ return (model, graph_modified)
+
+
+class InferGlobalAccPoolLayer(Transformation):
+ """Convert any GlobalAveragePool into a GlobalAccPool HLS layer and a scalar Mul."""
+
+ def apply(self, model):
+ graph = model.graph
+ node_ind = 0
+ graph_modified = False
+ for node in graph.node:
+ node_ind += 1
+ if node.op_type == "GlobalAveragePool":
+ in0 = node.input[0]
+ result = node.output[0]
+ in0_shape = model.get_tensor_shape(in0)
+
+ idt = model.get_tensor_datatype(in0)
+
+ # skip conversion for layers with float input
+ if not idt.is_integer():
+ continue
+
+ # check layout and convert if necessary
+ in0_layout = model.get_tensor_layout(in0)
+ result_layout = model.get_tensor_layout(result)
+
+ if in0_layout == DataLayout.NCHW:
+ in0 = nchw_to_nhwc(in0, model, node_ind)
+ node_ind += 1
+ in0_shape = model.get_tensor_shape(in0)
+
+ # keep track of where we need to insert the HLS Op
+ # it has to be ahead of the output transform
+ insert_point = node_ind
+
+ if result_layout == DataLayout.NCHW:
+ result = nchw_to_nhwc(result, model, node_ind, reverse=True)
+ node_ind += 1
+
+ num_ch = int(in0_shape[-1])
+ vecs = in0_shape[:-1]
+ # create node with no parallelization first
+ pe = 1
+ assert (
+ num_ch % pe == 0
+ ), "Requirement Labels divisable by PE is violated."
+
+ # create an additional tensor of the same shape and layout as result
+ out_shape = model.get_tensor_shape(result)
+ pool_out = helper.make_tensor_value_info(
+ model.make_new_valueinfo_name(), TensorProto.FLOAT, out_shape
+ )
+ model.graph.value_info.append(pool_out)
+ pool_out = pool_out.name
+ model.set_tensor_layout(pool_out, model.get_tensor_layout(result))
+
+ new_pool = helper.make_node(
+ "GlobalAccPool_Batch",
+ [in0],
+ [pool_out],
+ domain="finn",
+ backend="fpgadataflow",
+ NumChannels=num_ch,
+ PE=pe,
+ inputDataType=idt.name,
+ numInputVectors=vecs,
+ )
+
+ mul_value = helper.make_tensor_value_info(
+ model.make_new_valueinfo_name(), TensorProto.FLOAT, [1]
+ )
+ model.graph.value_info.append(mul_value)
+ model.set_initializer(mul_value.name, np.array(1 / (vecs[1] * vecs[2])))
+ new_mul = helper.make_node("Mul", [pool_out, mul_value.name], [result],)
+ graph.node.insert(insert_point, new_pool)
+ graph.node.insert(insert_point + 1, new_mul)
+ node_ind += 1
+ # remove old node
+ graph.node.remove(node)
+ graph_modified = True
+
+ if graph_modified:
+ model = model.transform(InferShapes())
+ model = model.transform(InferDataTypes())
+ return (model, graph_modified)
diff --git a/src/finn/transformation/fpgadataflow/floorplan.py b/src/finn/transformation/fpgadataflow/floorplan.py
new file mode 100644
index 0000000000000000000000000000000000000000..1d9a51875499d77f384c03f54009a9dd1001dea0
--- /dev/null
+++ b/src/finn/transformation/fpgadataflow/floorplan.py
@@ -0,0 +1,80 @@
+# Copyright (c) 2020, Xilinx
+# All rights reserved.
+#
+# Redistribution and use in source and binary forms, with or without
+# modification, are permitted provided that the following conditions are met:
+#
+# * Redistributions of source code must retain the above copyright notice, this
+# list of conditions and the following disclaimer.
+#
+# * Redistributions in binary form must reproduce the above copyright notice,
+# this list of conditions and the following disclaimer in the documentation
+# and/or other materials provided with the distribution.
+#
+# * Neither the name of FINN nor the names of its
+# contributors may be used to endorse or promote products derived from
+# this software without specific prior written permission.
+#
+# THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS "AS IS"
+# AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
+# IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE ARE
+# DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT HOLDER OR CONTRIBUTORS BE LIABLE
+# FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL
+# DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR
+# SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER
+# CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY,
+# OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE
+# OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
+
+from finn.custom_op.registry import getCustomOp
+from finn.transformation import Transformation
+from finn.util.basic import get_by_name
+
+
+class Floorplan(Transformation):
+ """Perform Floorplanning of the dataflow design. Separate DMAs into their own
+ partitions IDs, and TODO: split the design into sections of defined size"""
+
+ def __init__(self, limits=None):
+ super().__init__()
+ self.resource_limits = limits
+
+ def apply(self, model):
+ target_partition_id = 0
+ # we currently assume that all dataflow nodes belonging to the same partition
+ # are connected to each other and there is a single input/output to/from each.
+ all_nodes = list(model.graph.node)
+ df_nodes = list(
+ filter(lambda x: get_by_name(x.attribute, "backend") is not None, all_nodes)
+ )
+ dma_nodes = list(filter(lambda x: x.op_type == "IODMA", df_nodes))
+
+ non_dma_nodes = list(filter(lambda x: x not in dma_nodes, df_nodes))
+ dyn_tlastmarker_nodes = list(
+ filter(
+ lambda x: x.op_type == "TLastMarker"
+ and getCustomOp(x).get_nodeattr("DynIters") == "true",
+ non_dma_nodes,
+ )
+ )
+
+ non_dma_nodes = list(
+ filter(lambda x: x not in dyn_tlastmarker_nodes, non_dma_nodes)
+ )
+
+ for node in dma_nodes:
+ node_inst = getCustomOp(node)
+ node_inst.set_nodeattr("partition_id", target_partition_id)
+ target_partition_id += 1
+
+ for node in dyn_tlastmarker_nodes:
+ node_inst = getCustomOp(node)
+ node_inst.set_nodeattr("partition_id", target_partition_id)
+ target_partition_id += 1
+
+ for node in non_dma_nodes:
+ # TODO: implement proper floorplanning; for now just a single partition
+ node_inst = getCustomOp(node)
+ node_inst.set_nodeattr("partition_id", target_partition_id)
+
+ return (model, False)
diff --git a/src/finn/transformation/fpgadataflow/insert_iodma.py b/src/finn/transformation/fpgadataflow/insert_iodma.py
new file mode 100644
index 0000000000000000000000000000000000000000..e4368edea717f7499481e9b1c6ac20f7d5bb5f58
--- /dev/null
+++ b/src/finn/transformation/fpgadataflow/insert_iodma.py
@@ -0,0 +1,198 @@
+# Copyright (c) 2020, Xilinx
+# All rights reserved.
+#
+# Redistribution and use in source and binary forms, with or without
+# modification, are permitted provided that the following conditions are met:
+#
+# * Redistributions of source code must retain the above copyright notice, this
+# list of conditions and the following disclaimer.
+#
+# * Redistributions in binary form must reproduce the above copyright notice,
+# this list of conditions and the following disclaimer in the documentation
+# and/or other materials provided with the distribution.
+#
+# * Neither the name of FINN nor the names of its
+# contributors may be used to endorse or promote products derived from
+# this software without specific prior written permission.
+#
+# THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS "AS IS"
+# AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
+# IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE ARE
+# DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT HOLDER OR CONTRIBUTORS BE LIABLE
+# FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL
+# DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR
+# SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER
+# CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY,
+# OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE
+# OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
+
+from onnx import TensorProto
+from onnx import helper as oh
+
+from finn.util.basic import get_by_name
+from finn.custom_op.registry import getCustomOp
+from finn.transformation import Transformation
+from finn.transformation.general import SortGraph
+import finn.core.data_layout as DataLayout
+import math
+import numpy as np
+
+
+class InsertIODMA(Transformation):
+ """Insert DMA nodes on all inputs and outputs."""
+
+ def __init__(self, max_intfwidth=32):
+ super().__init__()
+ assert (
+ 2 ** math.log2(max_intfwidth) == max_intfwidth
+ ), "max_intfwidth must be a power of 2"
+ self.max_intfwidth = max_intfwidth
+
+ def apply(self, model):
+ # only makes sense for a pure fpgadataflow graph -- so we check!
+ all_nodes = list(model.graph.node)
+ assert all(
+ get_by_name(x.attribute, "backend").s.decode("UTF-8") == "fpgadataflow"
+ for x in all_nodes
+ )
+ # parse streamingfclayers looking for external weights with no attached IODMA
+ fc_extw_nodes = list(
+ filter(
+ lambda x: x.op_type == "StreamingFCLayer_Batch"
+ and get_by_name(x.attribute, "mem_mode") is not None
+ and get_by_name(x.attribute, "mem_mode").s.decode("UTF-8") == "external"
+ and model.find_producer(x.input[1]) is None,
+ all_nodes,
+ )
+ )
+ graph_in_name = model.graph.input[0].name
+ first_node = model.find_consumer(graph_in_name)
+ graph_out_name = model.graph.output[0].name
+ final_node = model.find_producer(graph_out_name)
+ if (
+ final_node.op_type == "IODMA"
+ and first_node.op_type == "IODMA"
+ and len(fc_extw_nodes) == 0
+ ):
+ # TODO maybe check the correctness of properties
+ return (model, False)
+ else:
+ if final_node.op_type != "IODMA":
+ # check if tensor is NHWC
+ assert (
+ model.get_tensor_layout(graph_out_name) == DataLayout.NHWC
+ or model.get_tensor_layout(graph_in_name) == DataLayout.NC
+ ), "Data layout of tensors must be NHWC or NC"
+ out_shape = model.get_tensor_shape(graph_out_name)
+ out_dtype = model.get_tensor_datatype(graph_out_name)
+ # determine the feasible interface width
+ transfer_bits = np.prod(out_shape) * out_dtype.bitwidth()
+ intfwidth = math.gcd(transfer_bits, self.max_intfwidth)
+ assert (
+ intfwidth % 8 == 0
+ ), "No feasible interface width for transfer size"
+ # get width of stream input to DMA
+ streamWidth = getCustomOp(final_node).get_outstream_width()
+ # make new buffer
+ final_node_out = oh.make_tensor_value_info(
+ model.make_new_valueinfo_name(), TensorProto.FLOAT, out_shape
+ )
+ model.graph.value_info.append(final_node_out)
+ model.set_tensor_datatype(final_node_out.name, out_dtype)
+ # reroute final node output to final_node_out_name
+ final_node.output[0] = final_node_out.name
+ dma_node = oh.make_node(
+ "IODMA",
+ [final_node_out.name],
+ [graph_out_name],
+ numInputVectors=out_shape[:-1],
+ NumChannels=out_shape[-1],
+ dataType=str(out_dtype.name),
+ intfWidth=intfwidth,
+ streamWidth=streamWidth,
+ direction="out",
+ domain="finn",
+ backend="fpgadataflow",
+ )
+ model.graph.node.append(dma_node)
+ if first_node.op_type != "IODMA":
+ # check if tensor is NHWC
+ assert (
+ model.get_tensor_layout(graph_in_name) == DataLayout.NHWC
+ or model.get_tensor_layout(graph_in_name) == DataLayout.NC
+ ), "Data layout of tensors must be NHWC or NC"
+ in_shape = model.get_tensor_shape(graph_in_name)
+ in_dtype = model.get_tensor_datatype(graph_in_name)
+ # determine the feasible interface width
+ transfer_bits = np.prod(in_shape) * in_dtype.bitwidth()
+ intfwidth = math.gcd(transfer_bits, self.max_intfwidth)
+ assert (
+ intfwidth % 8 == 0
+ ), "No feasible interface width for transfer size"
+ # get width of stream output from DMA
+ streamWidth = getCustomOp(first_node).get_instream_width()
+ # make new buffer
+ first_node_in = oh.make_tensor_value_info(
+ model.make_new_valueinfo_name(), TensorProto.FLOAT, in_shape
+ )
+ model.graph.value_info.append(first_node_in)
+ model.set_tensor_datatype(first_node_in.name, in_dtype)
+ # reroute final node output to final_node_out_name
+ first_node.input[0] = first_node_in.name
+ dma_node = oh.make_node(
+ "IODMA",
+ [graph_in_name],
+ [first_node_in.name],
+ numInputVectors=in_shape[:-1],
+ NumChannels=in_shape[-1],
+ dataType=str(in_dtype.name),
+ intfWidth=intfwidth,
+ streamWidth=streamWidth,
+ direction="in",
+ domain="finn",
+ backend="fpgadataflow",
+ )
+ model.graph.node.insert(0, dma_node)
+ for fc_node in fc_extw_nodes:
+ # check if tensor is NHWC
+ assert (
+ model.get_tensor_layout(fc_node.input[1]) == DataLayout.NHWC
+ or model.get_tensor_layout(graph_in_name) == DataLayout.NC
+ ), "Data layout of tensors must be NHWC or NC"
+ fc_w_name = fc_node.input[1]
+ w_shape = model.get_tensor_shape(fc_w_name)
+ w_dtype = model.get_tensor_datatype(fc_w_name)
+ # determine the feasible interface width
+ transfer_bits = np.prod(w_shape) * w_dtype.bitwidth()
+ intfwidth = math.gcd(transfer_bits, self.max_intfwidth)
+ assert (
+ intfwidth % 8 == 0
+ ), "No feasible interface width for transfer size"
+ # calculate width of stream output from DMA
+ pe = get_by_name(fc_node.attribute, "PE").i
+ simd = get_by_name(fc_node.attribute, "SIMD").i
+ streamWidth = simd * pe * w_dtype.bitwidth()
+ # make new buffer
+ fc_node_in = oh.make_tensor_value_info(
+ model.make_new_valueinfo_name(), TensorProto.FLOAT, w_shape
+ )
+ model.graph.value_info.append(fc_node_in)
+ model.set_tensor_datatype(fc_node_in.name, w_dtype)
+ dma_node = oh.make_node(
+ "IODMA",
+ [fc_w_name],
+ [fc_node_in.name],
+ numInputVectors=w_shape[:-1],
+ NumChannels=w_shape[-1],
+ dataType=str(w_dtype.name),
+ intfWidth=intfwidth,
+ streamWidth=streamWidth,
+ direction="in",
+ burstMode="wrap",
+ domain="finn",
+ backend="fpgadataflow",
+ )
+ fc_node.input[1] = fc_node_in.name
+ model.graph.node.insert(0, dma_node)
+ model = model.transform(SortGraph())
+ return (model, True)
diff --git a/src/finn/transformation/fpgadataflow/prepare_cppsim.py b/src/finn/transformation/fpgadataflow/prepare_cppsim.py
index 4f050be8540ddf5ef48699d1658b571852ff4510..6eae560e1191642cfaf85d92c6d0fcf644630973 100644
--- a/src/finn/transformation/fpgadataflow/prepare_cppsim.py
+++ b/src/finn/transformation/fpgadataflow/prepare_cppsim.py
@@ -80,7 +80,6 @@ class PrepareCppSim(Transformation):
self._num_workers = mp.cpu_count()
def prepareCppSim_node(self, node):
- print(node.name)
if is_fpgadataflow_node(node) is True:
_codegen_single_node(node, self.model)
return (node, False)
diff --git a/src/finn/transformation/remove_identity.py b/src/finn/transformation/remove_identity.py
new file mode 100644
index 0000000000000000000000000000000000000000..d7a58d59c1bb8ff643e691442e7eda3c0516aa5c
--- /dev/null
+++ b/src/finn/transformation/remove_identity.py
@@ -0,0 +1,62 @@
+# Copyright (c) 2020, Xilinx
+# All rights reserved.
+#
+# Redistribution and use in source and binary forms, with or without
+# modification, are permitted provided that the following conditions are met:
+#
+# * Redistributions of source code must retain the above copyright notice, this
+# list of conditions and the following disclaimer.
+#
+# * Redistributions in binary form must reproduce the above copyright notice,
+# this list of conditions and the following disclaimer in the documentation
+# and/or other materials provided with the distribution.
+#
+# * Neither the name of FINN nor the names of its
+# contributors may be used to endorse or promote products derived from
+# this software without specific prior written permission.
+#
+# THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS "AS IS"
+# AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
+# IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE ARE
+# DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT HOLDER OR CONTRIBUTORS BE LIABLE
+# FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL
+# DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR
+# SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER
+# CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY,
+# OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE
+# OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
+
+from finn.transformation import Transformation
+
+
+def _is_identity(node, model):
+ if node.op_type == "Mul":
+ scale = model.get_initializer(node.input[1])
+ if scale is not None:
+ return (scale == 1).all()
+ elif node.op_type == "Add":
+ bias = model.get_initializer(node.input[1])
+ if bias is not None:
+ return (bias == 0).all()
+ return False
+
+
+class RemoveIdentity(Transformation):
+ """Remove nodes that apply identity ops from the graph, including:
+ * Multiply by 1
+ * Add 0
+ ."""
+
+ def apply(self, model):
+ graph = model.graph
+ node_ind = 0
+ graph_modified = False
+ for node in graph.node:
+ node_ind += 1
+ if _is_identity(node, model):
+ node_src = node.input[0]
+ node_dst = node.output[0]
+ graph.node.remove(node)
+ model.rename_tensor(node_dst, node_src)
+ graph_modified = True
+ return (model, graph_modified)
diff --git a/src/finn/transformation/streamline/__init__.py b/src/finn/transformation/streamline/__init__.py
index c9c73fa4c8303ee28bc1cc6aee879d633740e01e..d9c12a20975084705b801c0ff027d4b99aff9490 100644
--- a/src/finn/transformation/streamline/__init__.py
+++ b/src/finn/transformation/streamline/__init__.py
@@ -41,6 +41,7 @@ from finn.transformation.streamline.absorb import (
FactorOutMulSignMagnitude,
Absorb1BitMulIntoMatMul,
Absorb1BitMulIntoConv,
+ AbsorbSignBiasIntoMultiThreshold,
)
from finn.transformation.streamline.collapse_repeated import (
@@ -59,6 +60,7 @@ from finn.transformation.streamline.reorder import (
from finn.transformation.streamline.round_thresholds import RoundAndClipThresholds
from finn.transformation.streamline.sign_to_thres import ConvertSignToThres
from finn.transformation.batchnorm_to_affine import BatchNormToAffine
+from finn.transformation.streamline.remove import RemoveIdentityOps
class Streamline(Transformation):
@@ -70,6 +72,7 @@ class Streamline(Transformation):
ConvertDivToMul(),
BatchNormToAffine(),
ConvertSignToThres(),
+ AbsorbSignBiasIntoMultiThreshold(),
MoveAddPastMul(),
MoveScalarAddPastMatMul(),
MoveScalarAddPastConv(),
@@ -87,6 +90,7 @@ class Streamline(Transformation):
]
for trn in streamline_transformations:
model = model.transform(trn)
+ model = model.transform(RemoveIdentityOps())
model = model.transform(GiveUniqueNodeNames())
model = model.transform(GiveReadableTensorNames())
model = model.transform(InferDataTypes())
diff --git a/src/finn/transformation/streamline/absorb.py b/src/finn/transformation/streamline/absorb.py
index f426a2f631e3fd65bb3bb6a7efc8662e0375a615..f089275c221f769daace3e9628a00bf87b4e5457 100644
--- a/src/finn/transformation/streamline/absorb.py
+++ b/src/finn/transformation/streamline/absorb.py
@@ -35,9 +35,74 @@ import finn.core.data_layout as DataLayout
from finn.transformation import Transformation
from finn.util.basic import get_by_name
from finn.custom_op.registry import getCustomOp
+from finn.transformation.infer_shapes import InferShapes
from finn.transformation.infer_datatypes import InferDataTypes
+class AbsorbSignBiasIntoMultiThreshold(Transformation):
+ """Absorb scalar bias originating from signed int export back into
+ MultiThreshold and re-evaluate the output datatype."""
+
+ def apply(self, model):
+ graph = model.graph
+ node_ind = 0
+ graph_modified = False
+ for n in graph.node:
+ # search for (MultiThreshold, Add) pair
+ node_ind += 1
+ if (
+ n.op_type == "MultiThreshold"
+ and not model.is_fork_node(n)
+ and not model.is_join_node(n)
+ ):
+ consumer = model.find_consumer(n.output[0])
+ if consumer is not None and consumer.op_type == "Add":
+ mt_node = n
+ add_node = consumer
+ threshold_name = mt_node.input[1]
+ add_weight_name = add_node.input[1]
+ T = model.get_initializer(threshold_name)
+ A = model.get_initializer(add_weight_name)
+ if (A is None) or (T is None):
+ warnings.warn("Threshold or add bias not constant, skipping")
+ continue
+ end_name = add_node.output[0]
+ # we can only absorb scalar adds
+ is_scalar = A.ndim == 0 or all(x == 1 for x in A.shape)
+ if not is_scalar:
+ continue
+ bias = A.flatten()[0]
+ # set MultiThreshold bias property
+ mt_inst = getCustomOp(mt_node)
+ bias += mt_inst.get_nodeattr("out_bias")
+ mt_inst.set_nodeattr("out_bias", bias)
+ graph_modified = True
+ # compute new DataType for MultiThreshold output
+ steps = T.shape[-1]
+ new_min = bias
+ new_max = steps + bias
+ odt = DataType.get_smallest_possible(steps).name.replace(
+ "UINT", "INT"
+ )
+ odt = DataType[odt]
+ assert odt.allowed(new_max) and odt.allowed(
+ new_min
+ ), """Could
+ not compute new MultiThreshold DataType (min = %d max = %d)""" % (
+ new_min,
+ new_max,
+ )
+ mt_inst.set_nodeattr("out_dtype", odt.name)
+ # remove Add node, rewire MultiThreshold
+ graph.node.remove(add_node)
+ mt_node.output[0] = end_name
+ # set datatype
+ model.set_tensor_datatype(end_name, odt)
+ if graph_modified:
+ model = model.transform(InferDataTypes())
+ return (model, graph_modified)
+
+
class AbsorbAddIntoMultiThreshold(Transformation):
"""Absorb preceding Add ops into MultiThreshold by updating the threshold
values. Only scalar/1D add vectors can be absorbed."""
@@ -293,7 +358,6 @@ class AbsorbTransposeIntoMultiThreshold(Transformation):
model = model.transform(InferDataTypes())
return (model, graph_modified)
-
class AbsorbTransposeIntoFlatten(Transformation):
"""Absorb transpose node into succeeding flatten node, if H=W=1 and the first
dimension stays the same. Can also be applied if flatten is implemented implicitly
@@ -321,7 +385,7 @@ class AbsorbTransposeIntoFlatten(Transformation):
# check for the data layout to interpret input shape correctly
if data_layout is None:
warnings.warn(
- """Datalayout for input tensor of transpose node is not set.
+ """Data layout for input tensor of Transpose node is not set.
To use AbsorbTransposeIntoFlatten transformation
please set tensor data layout."""
)
@@ -334,7 +398,7 @@ class AbsorbTransposeIntoFlatten(Transformation):
continue
# the flatten node from onnx keeps by default the first
# dim and flattens the rest, that is why this transformation
- # can only works with b != 1 if the model contains already a
+ # can only work with b != 1 if the model contains already a
# flatten node and not a reshape node with shape = [1, -1].
# If the first dim of the input tensor is not 1, flatten and
# reshape (with shape = [1, -1]) would lead to different results
@@ -352,5 +416,40 @@ class AbsorbTransposeIntoFlatten(Transformation):
graph.node.remove(prod)
graph.node.insert(node_ind, node)
graph_modified = True
- model = model.transform(InferDataTypes())
+ if graph_modified:
+ model = model.transform(InferDataTypes())
+ return (model, graph_modified)
+
+class AbsorbScalarMulIntoTopK(Transformation):
+ """Absorb a mul node into a suceeding topk node if the mul is scalar."""
+
+ def apply(self, model):
+ graph = model.graph
+ node_ind = 0
+ graph_modified = False
+ for n in graph.node:
+ node_ind += 1
+ if n.op_type == "TopK":
+ prod = model.find_producer(n.input[0])
+ if prod is not None and prod.op_type == "Mul":
+ prod_input = prod.input[0]
+ param_name = prod.input[1]
+ A = model.get_initializer(param_name)
+ if A is None:
+ warnings.warn("Param is not constant, skipping")
+ continue
+ if all(x == 1 for x in A.shape) and A > 0:
+ # if the mul is scalar and positive, we can just delete the
+ # mul node and rewire the top k node. Because the top k node
+ # works with probabilities and their relation to each other
+ # the relation doesn't change if every value is multiplied
+ # with a scalar
+ graph.node.remove(prod)
+ n.input[0] = prod_input
+ # to avoid error the dataype is set to float32
+ model.set_tensor_datatype(n.input[0], DataType.FLOAT32)
+ graph_modified = True
+ if graph_modified:
+ model = model.transform(InferShapes())
+ model = model.transform(InferDataTypes())
return (model, graph_modified)
diff --git a/src/finn/transformation/streamline/reorder.py b/src/finn/transformation/streamline/reorder.py
index b46b82c77a3f1b70a3b05d87cd3c48fc1d94fd45..cc95d34b784b47c9baeb6c1076915db8b1d09d57 100644
--- a/src/finn/transformation/streamline/reorder.py
+++ b/src/finn/transformation/streamline/reorder.py
@@ -32,6 +32,8 @@ from onnx import helper as oh
from finn.transformation import Transformation
from finn.transformation.infer_shapes import InferShapes
+from finn.transformation.infer_data_layouts import InferDataLayouts
+from finn.core.datatype import DataType
from finn.core.onnx_exec import execute_node
from finn.util.basic import get_by_name
from finn.custom_op.registry import getCustomOp
@@ -67,8 +69,11 @@ class MoveAddPastMul(Transformation):
add_weight_name = n.input[1]
A = model.get_initializer(mul_weight_name)
B = model.get_initializer(add_weight_name)
- assert A is not None, "Initializer for mul weights is not set."
- assert B is not None, "Initializer for add weights is not set."
+ if (A is None) or (B is None):
+ warnings.warn(
+ "Mul or add does not have constant params, skipping"
+ )
+ continue
start_name = n.input[0]
middle_name = n.output[0]
end_name = consumer.output[0]
@@ -123,8 +128,9 @@ class MoveScalarMulPastMatMul(Transformation):
matmul_weight_name = consumer.input[1]
A = model.get_initializer(mul_weight_name)
W = model.get_initializer(matmul_weight_name)
- assert A is not None, "Initializer for mul weights is not set."
- assert W is not None, "Initializer for matmul weights is not set."
+ if (A is None) or (W is None):
+ warnings.warn("MatMul or Mul params are not constant, skipping")
+ continue
start_name = n.input[0]
middle_name = n.output[0]
end_name = consumer.output[0]
@@ -180,8 +186,9 @@ class MoveScalarAddPastMatMul(Transformation):
matmul_weight_name = consumer.input[1]
A = model.get_initializer(add_weight_name)
W = model.get_initializer(matmul_weight_name)
- assert A is not None, "Initializer for add weights is not set."
- assert W is not None, "Initializer for matmul weights is not set."
+ if (A is None) or (W is None):
+ warnings.warn("MatMul or Add params are not constant, skipping")
+ continue
start_name = n.input[0]
middle_name = n.output[0]
end_name = consumer.output[0]
@@ -242,7 +249,9 @@ class MoveScalarAddPastConv(Transformation):
conv_in_name = consumer.input[0]
conv_in_shape = model.get_tensor_shape(conv_in_name)
A = model.get_initializer(add_weight_name)
- assert A is not None, "Initializer for add weights is not set."
+ if A is None:
+ warnings.warn("Add param is not constant, skipping")
+ continue
start_name = n.input[0]
end_name = consumer.output[0]
conv_out_shape = model.get_tensor_shape(end_name)
@@ -310,7 +319,9 @@ class MoveScalarMulPastConv(Transformation):
):
mul_weight_name = n.input[1]
A = model.get_initializer(mul_weight_name)
- assert A is not None, "Initializer for mul weights is not set."
+ if A is None:
+ warnings.warn("Mul param is not constant, skipping")
+ continue
conv_node = consumer
mul_node = n
start_name = mul_node.input[0]
@@ -338,6 +349,71 @@ class MoveScalarMulPastConv(Transformation):
return (model, graph_modified)
+class MoveMulPastDWConv(Transformation):
+ """Move channelwise mul operations past depthwise conv operations. We want to have muls
+ next to each other such that they can be collapsed into a single mul."""
+
+ def apply(self, model):
+ graph = model.graph
+ node_ind = 0
+ graph_modified = False
+ for n in graph.node:
+ node_ind += 1
+ if (
+ n.op_type == "Mul"
+ and not model.is_fork_node(n)
+ and not model.is_join_node(n)
+ ):
+ consumer = model.find_consumer(n.output[0])
+ if (
+ consumer is not None
+ and consumer.op_type == "Conv"
+ and not model.is_join_node(consumer)
+ ):
+ mul_weight_name = n.input[1]
+ A = model.get_initializer(mul_weight_name)
+ if A is None:
+ warnings.warn(
+ """Mul weight tensor is not set. If it is a constant,
+ please use set_initializer to set the tensor."""
+ )
+ continue
+ conv_node = consumer
+ mul_node = n
+ start_name = mul_node.input[0]
+ conv_in_name = conv_node.input[0]
+ conv_in_shape = model.get_tensor_shape(conv_in_name)
+ ifm_ch = conv_in_shape[1]
+ group_attribute = get_by_name(consumer.attribute, "group")
+ if group_attribute is None:
+ continue
+ group_attribute = group_attribute.i
+ conv_out_name = conv_node.output[0]
+ conv_out_shape = model.get_tensor_shape(conv_out_name)
+ if A.shape == (1, ifm_ch, 1, 1) and ifm_ch == group_attribute:
+ # if the mul is channelwise and conv is depthwise,
+ # we can simply swap the order of ops
+ # rewire mul input to be conv input
+ conv_node.input[0] = start_name
+ model.set_tensor_shape(start_name, conv_in_shape)
+ model.set_tensor_datatype(start_name, DataType.FLOAT32)
+ # use old conv input tensor as conv output
+ conv_node.output[0] = conv_in_name
+ model.set_tensor_shape(conv_in_name, conv_out_shape)
+ model.set_tensor_datatype(conv_in_name, DataType.FLOAT32)
+ # use new conv output as new mul node input
+ mul_node.input[0] = conv_in_name
+ # use old conv output as new mul node output
+ mul_node.output[0] = conv_out_name
+ model.set_tensor_datatype(conv_out_name, DataType.FLOAT32)
+ # move mul node past conv node
+ graph.node.remove(mul_node)
+ graph.node.insert(node_ind, mul_node)
+ graph_modified = True
+ model = model.transform(InferShapes())
+ return (model, graph_modified)
+
+
class MoveLinearPastEltwiseAdd(Transformation):
"""Move linear operations (mul, add) past elementwise add operations where possible.
Specifically,matches and transforms the following patterns:
@@ -597,3 +673,66 @@ class MoveMaxPoolPastMultiThreshold(Transformation):
model = model.transform(InferShapes())
return (model, graph_modified)
+
+
+class MoveTransposePastScalarMul(Transformation):
+ """Moves a Transpose node past a scalar Mul node"""
+
+ def apply(self, model):
+ graph = model.graph
+ node_ind = 0
+ graph_modified = False
+ for n in graph.node:
+ node_ind += 1
+ if (
+ n.op_type == "Transpose"
+ and not model.is_fork_node(n)
+ and not model.is_join_node(n)
+ ):
+ consumer = model.find_consumer(n.output[0])
+ if (
+ consumer is not None
+ and consumer.op_type == "Mul"
+ and not model.is_join_node(consumer)
+ ):
+ mul_weight_name = consumer.input[1]
+ A = model.get_initializer(mul_weight_name)
+ if A is None:
+ warnings.warn("Mul param is not constant, skipping")
+ continue
+ transp_node = n
+ mul_node = consumer
+ start_name = transp_node.input[0]
+ middle_name = transp_node.output[0]
+ end_name = mul_node.output[0]
+ transp_in_shape = model.get_tensor_shape(start_name)
+ transp_out_shape = model.get_tensor_shape(middle_name)
+ transp_in_layout = model.get_tensor_layout(start_name)
+ transp_out_layout = model.get_tensor_layout(middle_name)
+ if transp_in_layout is None or transp_out_layout is None:
+ warnings.warn(
+ """Datalayout is not set for tensors.
+ Transformation can't be applied."""
+ )
+ continue
+ if all(x == 1 for x in A.shape):
+ # if the mul is scalar, we can simply swap the order of ops
+ # rewire transpose input to be mul input
+ mul_node.input[0] = start_name
+ model.set_tensor_shape(start_name, transp_in_shape)
+ model.set_tensor_layout(start_name, transp_in_layout)
+ mul_node.output[0] = middle_name
+ model.set_tensor_shape(middle_name, transp_in_shape)
+ model.set_tensor_layout(middle_name, transp_in_layout)
+ transp_node.input[0] = middle_name
+ transp_node.output[0] = end_name
+ model.set_tensor_shape(end_name, transp_out_shape)
+ model.set_tensor_layout(end_name, transp_out_layout)
+ graph.node.remove(transp_node)
+ graph.node.insert(node_ind, transp_node)
+ graph_modified = True
+
+ if graph_modified is True:
+ model = model.transform(InferDataLayouts())
+ model = model.transform(InferShapes())
+ return (model, graph_modified)
diff --git a/src/finn/util/onnx.py b/src/finn/util/onnx.py
index b9932111d86d7206b23e1d0e49a6aa8451f8ba24..4d7cdd126ededac887639a932c2021ef5f081c02 100644
--- a/src/finn/util/onnx.py
+++ b/src/finn/util/onnx.py
@@ -28,6 +28,7 @@
import numpy as np
import onnx
+import finn.core.data_layout as DataLayout
def valueinfo_to_tensor(vi):
@@ -37,3 +38,38 @@ def valueinfo_to_tensor(vi):
return np.zeros(
dims, dtype=onnx.mapping.TENSOR_TYPE_TO_NP_TYPE[vi.type.tensor_type.elem_type]
)
+
+
+def nchw_to_nhwc(t, model, idx, reverse=False):
+ """Converts between NCHW <-> NHWC layouts for tensor t by inserting a transpose.
+ If reverse=False, t is assumed NCHW and we insert transpose to convert NCHW -> NHWC
+ If reverse=True, t is assumed NHWC and we insert transpose to convert NHWC -> NCHW.
+ """
+ graph = model.graph
+ # create new NHWC tensor
+ t_shape = model.get_tensor_shape(t)
+ bs = t_shape[0]
+ ch = t_shape[1]
+ height = t_shape[2]
+ width = t_shape[3]
+ t_trans = onnx.helper.make_tensor_value_info(
+ model.make_new_valueinfo_name(),
+ onnx.TensorProto.FLOAT,
+ (bs, height, width, ch), # NHWC
+ )
+ graph.value_info.append(t_trans)
+ dt = model.get_tensor_datatype(t)
+ t_trans = t_trans.name
+ model.set_tensor_datatype(t_trans, dt)
+ model.set_tensor_layout(t_trans, DataLayout.NHWC)
+ # NCHW <-> NHWC transpose
+ if reverse:
+ t_trans_node = onnx.helper.make_node(
+ "Transpose", [t_trans], [t], perm=[0, 3, 1, 2]
+ )
+ else:
+ t_trans_node = onnx.helper.make_node(
+ "Transpose", [t], [t_trans], perm=[0, 2, 3, 1]
+ )
+ graph.node.insert(idx, t_trans_node)
+ return t_trans
diff --git a/tests/core/test_basic_onnx_exec.py b/tests/core/test_basic_onnx_exec.py
index 7b0412432cc6360cb9c42d66417bd187ed142563..ddb2cbfc40c7647970f0c51ecb95340e7d1dddae 100644
--- a/tests/core/test_basic_onnx_exec.py
+++ b/tests/core/test_basic_onnx_exec.py
@@ -49,19 +49,33 @@ def test_mnist_onnx_download_extract_run():
raw_o = get_data("finn", "data/onnx/mnist-conv/test_data_set_0/output_0.pb")
input_tensor = onnx.load_tensor_from_string(raw_i)
output_tensor = onnx.load_tensor_from_string(raw_o)
- # run using FINN-based execution
+ # run using FINN-based execution (full graph)
input_dict = {"Input3": np_helper.to_array(input_tensor)}
- output_dict = oxe.execute_onnx(model, input_dict)
+ output_dict = oxe.execute_onnx(model, input_dict, return_full_exec_context=True)
assert np.isclose(
np_helper.to_array(output_tensor), output_dict["Plus214_Output_0"], atol=1e-3
).all()
+ # test subgraph execution
+ start_node = model.graph.node[1]
+ end_node = model.graph.node[3]
+ subgraph_i_dict = {start_node.input[0]: output_dict[start_node.input[0]]}
+ subgraph_o_dict = oxe.execute_onnx(
+ model,
+ subgraph_i_dict,
+ return_full_exec_context=True,
+ start_node=start_node,
+ end_node=end_node,
+ )
+ assert np.isclose(
+ subgraph_o_dict[end_node.output[0]], output_dict[end_node.output[0]], atol=1e-3
+ ).all()
def test_onnx_exec_internal_rounding():
inp0 = onnx.helper.make_tensor_value_info("inp0", onnx.TensorProto.FLOAT, [2, 2])
inp1 = onnx.helper.make_tensor_value_info("inp1", onnx.TensorProto.FLOAT, [1])
outp = onnx.helper.make_tensor_value_info("outp", onnx.TensorProto.FLOAT, [2, 2])
- mul_node = onnx.helper.make_node("Mul", inputs=["inp0", "inp1"], outputs=["outp"],)
+ mul_node = onnx.helper.make_node("Mul", inputs=["inp0", "inp1"], outputs=["outp"])
graph = onnx.helper.make_graph(
nodes=[mul_node], name="mul_graph", inputs=[inp0, inp1], outputs=[outp]
)
diff --git a/tests/end2end/test_end2end_cnv_w2a2.py b/tests/end2end/test_end2end_cnv_w2a2.py
new file mode 100644
index 0000000000000000000000000000000000000000..31ccebd4c175ad2badef17499bf113d978b637f7
--- /dev/null
+++ b/tests/end2end/test_end2end_cnv_w2a2.py
@@ -0,0 +1,377 @@
+# Copyright (c) 2020, Xilinx
+# All rights reserved.
+#
+# Redistribution and use in source and binary forms, with or without
+# modification, are permitted provided that the following conditions are met:
+#
+# * Redistributions of source code must retain the above copyright notice, this
+# list of conditions and the following disclaimer.
+#
+# * Redistributions in binary form must reproduce the above copyright notice,
+# this list of conditions and the following disclaimer in the documentation
+# and/or other materials provided with the distribution.
+#
+# * Neither the name of FINN nor the names of its
+# contributors may be used to endorse or promote products derived from
+# this software without specific prior written permission.
+#
+# THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS "AS IS"
+# AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
+# IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE ARE
+# DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT HOLDER OR CONTRIBUTORS BE LIABLE
+# FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL
+# DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR
+# SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER
+# CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY,
+# OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE
+# OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
+
+import os
+
+import numpy as np
+
+# as of Feb'20 there is a bug that segfaults ONNX shape inference if we
+# import pytorch before onnx, so we make sure to import onnx first
+import onnx # NOQA
+
+import pytest
+import pkg_resources as pk
+from finn.custom_op.registry import getCustomOp
+from finn.core.onnx_exec import execute_onnx
+from finn.transformation.double_to_single_float import DoubleToSingleFloat
+from finn.transformation.infer_shapes import InferShapes
+from finn.transformation.move_reshape import RemoveCNVtoFCFlatten
+from finn.transformation.fold_constants import FoldConstants
+from finn.transformation.general import GiveReadableTensorNames, GiveUniqueNodeNames
+from finn.transformation.streamline import Streamline
+from finn.transformation.lower_convs_to_matmul import LowerConvsToMatMul
+import finn.transformation.streamline.absorb as absorb
+from finn.transformation.streamline.reorder import MakeMaxPoolNHWC
+import finn.transformation.fpgadataflow.convert_to_hls_layers as to_hls
+from finn.transformation.fpgadataflow.create_dataflow_partition import (
+ CreateDataflowPartition,
+)
+from finn.transformation.fpgadataflow.insert_dwc import InsertDWC
+from finn.transformation.fpgadataflow.insert_tlastmarker import InsertTLastMarker
+from finn.transformation.fpgadataflow.prepare_ip import PrepareIP
+from finn.transformation.fpgadataflow.hlssynth_ip import HLSSynthIP
+from finn.transformation.fpgadataflow.replace_verilog_relpaths import (
+ ReplaceVerilogRelPaths,
+)
+from finn.transformation.fpgadataflow.create_stitched_ip import CreateStitchedIP
+from finn.transformation.fpgadataflow.set_exec_mode import SetExecMode
+from finn.transformation.fpgadataflow.prepare_cppsim import PrepareCppSim
+from finn.transformation.fpgadataflow.compile_cppsim import CompileCppSim
+from finn.transformation.fpgadataflow.make_pynq_driver import MakePYNQDriver
+from finn.transformation.fpgadataflow.make_pynq_proj import MakePYNQProject
+from finn.transformation.fpgadataflow.synth_pynq_proj import SynthPYNQProject
+from finn.transformation.fpgadataflow.make_deployment import DeployToPYNQ
+from finn.util.basic import pynq_part_map
+from finn.util.test import get_test_model_trained, load_test_checkpoint_or_skip
+from finn.transformation.fpgadataflow.annotate_resources import AnnotateResources
+from finn.transformation.fpgadataflow.prepare_rtlsim import PrepareRTLSim
+from finn.transformation.fpgadataflow.insert_fifo import InsertFIFO
+from finn.core.throughput_test import throughput_test_rtlsim
+
+build_dir = "/tmp/" + os.environ["FINN_INST_NAME"]
+test_pynq_board = os.getenv("PYNQ_BOARD", default="Pynq-Z1")
+test_fpga_part = pynq_part_map[test_pynq_board]
+target_clk_ns = 10
+mem_mode = "decoupled"
+
+
+def test_end2end_cnv_w2a2_export():
+ import brevitas.onnx as bo
+
+ cnv = get_test_model_trained("CNV", 2, 2)
+ bo.export_finn_onnx(
+ cnv, (1, 3, 32, 32), build_dir + "/end2end_cnv_w2a2_export.onnx"
+ )
+
+
+def test_end2end_cnv_w2a2_import_and_tidy():
+ model = load_test_checkpoint_or_skip(build_dir + "/end2end_cnv_w2a2_export.onnx")
+ model = model.transform(DoubleToSingleFloat())
+ model = model.transform(InferShapes())
+ model = model.transform(FoldConstants())
+ model = model.transform(GiveUniqueNodeNames())
+ model = model.transform(GiveReadableTensorNames())
+ model.save(build_dir + "/end2end_cnv_w2a2_tidy.onnx")
+
+
+def test_end2end_cnv_w2a2_streamline():
+ model = load_test_checkpoint_or_skip(build_dir + "/end2end_cnv_w2a2_tidy.onnx")
+ model = model.transform(Streamline())
+ model = model.transform(LowerConvsToMatMul())
+ model = model.transform(MakeMaxPoolNHWC())
+ model = model.transform(absorb.AbsorbTransposeIntoMultiThreshold())
+ model = model.transform(Streamline())
+ model.save(build_dir + "/end2end_cnv_w2a2_streamlined.onnx")
+
+
+def test_end2end_cnv_w2a2_convert_to_hls_layers():
+ model = load_test_checkpoint_or_skip(
+ build_dir + "/end2end_cnv_w2a2_streamlined.onnx"
+ )
+ model = model.transform(to_hls.InferQuantizedStreamingFCLayer(mem_mode))
+ model = model.transform(to_hls.InferConvInpGen())
+ model = model.transform(to_hls.InferStreamingMaxPool())
+ model = model.transform(RemoveCNVtoFCFlatten())
+ model.save(build_dir + "/end2end_cnv_w2a2_hls_layers.onnx")
+
+
+def test_end2end_cnv_w2a2_create_dataflow_partition():
+ model = load_test_checkpoint_or_skip(
+ build_dir + "/end2end_cnv_w2a2_hls_layers.onnx"
+ )
+ parent_model = model.transform(CreateDataflowPartition())
+ parent_model.save(build_dir + "/end2end_cnv_w2a2_dataflow_parent.onnx")
+ sdp_node = parent_model.get_nodes_by_op_type("StreamingDataflowPartition")[0]
+ sdp_node = getCustomOp(sdp_node)
+ dataflow_model_filename = sdp_node.get_nodeattr("model")
+ dataflow_model = load_test_checkpoint_or_skip(dataflow_model_filename)
+ dataflow_model.save(build_dir + "/end2end_cnv_w2a2_dataflow_model.onnx")
+
+
+def test_end2end_cnv_w2a2_fold_and_tlastmarker():
+ model = load_test_checkpoint_or_skip(
+ build_dir + "/end2end_cnv_w2a2_dataflow_model.onnx"
+ )
+ fc_layers = model.get_nodes_by_op_type("StreamingFCLayer_Batch")
+ # each tuple is (PE, SIMD, in_fifo_depth) for a layer
+ folding = [
+ (8, 3, 256, "auto"),
+ (16, 16, 256, "auto"),
+ (8, 16, 256, "auto"),
+ (8, 16, 256, "block"),
+ (4, 8, 214, "auto"),
+ (1, 8, 2, "auto"),
+ (1, 2, 126, "distributed"),
+ (2, 2, 62, "block"),
+ (5, 1, 6, "distributed"),
+ ]
+ for fcl, (pe, simd, ififodepth, ramstyle) in zip(fc_layers, folding):
+ fcl_inst = getCustomOp(fcl)
+ fcl_inst.set_nodeattr("PE", pe)
+ fcl_inst.set_nodeattr("SIMD", simd)
+ fcl_inst.set_nodeattr("inFIFODepth", ififodepth)
+ fcl_inst.set_nodeattr("ram_style", ramstyle)
+
+ swg_layers = model.get_nodes_by_op_type("ConvolutionInputGenerator")
+ swg_idepth = [2, 51, 9, 106, 2, 2]
+ for i in range(len(swg_layers)):
+ swg_inst = getCustomOp(swg_layers[i])
+ simd = folding[i][1]
+ swg_inst.set_nodeattr("SIMD", simd)
+ swg_inst.set_nodeattr("inFIFODepth", swg_idepth[i])
+
+ model = model.transform(InsertDWC())
+ model = model.transform(InsertFIFO())
+ model = model.transform(InsertTLastMarker())
+ model = model.transform(GiveUniqueNodeNames())
+ model = model.transform(AnnotateResources("estimate"))
+ model.save(build_dir + "/end2end_cnv_w2a2_folded.onnx")
+
+
+@pytest.mark.slow
+@pytest.mark.vivado
+def test_end2end_cnv_w2a2_gen_hls_ip():
+ model = load_test_checkpoint_or_skip(build_dir + "/end2end_cnv_w2a2_folded.onnx")
+ model = model.transform(PrepareIP(test_fpga_part, target_clk_ns))
+ model = model.transform(HLSSynthIP())
+ model = model.transform(AnnotateResources("hls"))
+ model.save(build_dir + "/end2end_cnv_w2a2_ipgen.onnx")
+
+
+@pytest.mark.vivado
+def test_end2end_cnv_w2a2_ip_stitch():
+ model = load_test_checkpoint_or_skip(build_dir + "/end2end_cnv_w2a2_ipgen.onnx")
+ model = model.transform(ReplaceVerilogRelPaths())
+ model = model.transform(CreateStitchedIP(test_fpga_part, target_clk_ns))
+ model.save(build_dir + "/end2end_cnv_w2a2_ipstitch.onnx")
+
+
+@pytest.mark.vivado
+def test_end2end_cnv_w2a2_verify_dataflow_part():
+ model = load_test_checkpoint_or_skip(build_dir + "/end2end_cnv_w2a2_ipstitch.onnx")
+ x = np.zeros((1, 32, 32, 3), dtype=np.float32)
+ inp_name = model.graph.input[0].name
+ out_name = model.graph.output[0].name
+ inp_dict = {inp_name: x}
+ # cppsim
+ model = model.transform(PrepareCppSim())
+ model = model.transform(CompileCppSim())
+ model = model.transform(SetExecMode("cppsim"))
+ model.save(build_dir + "/end2end_cnv_w2a2_ipgen_cppsim.onnx")
+ ret_cppsim = execute_onnx(model, inp_dict, True)
+ res_cppsim = ret_cppsim[out_name]
+ # node-by-node rtlsim
+ model = model.transform(SetExecMode("rtlsim"))
+ model = model.transform(PrepareRTLSim())
+ model.save(build_dir + "/end2end_cnv_w2a2_ipgen_nodebynode_rtlsim.onnx")
+ ret_rtlsim_nodebynode = execute_onnx(model, inp_dict, True)
+ res_rtlsim_nodebynode = ret_rtlsim_nodebynode[out_name]
+ # whole-network (ip-stitched) rtlsim
+ model.set_metadata_prop("exec_mode", "rtlsim")
+ model.save(build_dir + "/end2end_cnv_w2a2_ipstitch_whole_rtlsim.onnx")
+ # this is a particularly long-running test, set liveness thr. to unlimited
+ os.environ["LIVENESS_THRESHOLD"] = "-1"
+ ret_rtlsim_whole = execute_onnx(model, inp_dict, True)
+ res_rtlsim_whole = ret_rtlsim_whole[out_name]
+ assert np.isclose(res_cppsim, res_rtlsim_nodebynode).all()
+ assert np.isclose(res_cppsim, res_rtlsim_whole).all()
+
+
+@pytest.mark.vivado
+def test_end2end_cnv_w2a2_throughput_test_rtlsim():
+ model = load_test_checkpoint_or_skip(
+ build_dir + "/end2end_cnv_w2a2_ipstitch_whole_rtlsim.onnx"
+ )
+ model.set_metadata_prop("rtlsim_trace", "rtlsim_trace.vcd")
+ # os.environ["RTLSIM_TRACE_DEPTH"] = "4"
+ # run through IP-stitched rtlsim with increasing batch sizes and
+ # check the number of cycles it takes to execute
+ ret = throughput_test_rtlsim(model, 10)
+ # TODO check for expected performance
+ assert ret["cycles"] > 0
+
+
+@pytest.mark.vivado
+def test_end2end_cnv_w2a2_verify_all():
+ # use the streamlined model as the "golden" model for right answers
+ golden = load_test_checkpoint_or_skip(
+ build_dir + "/end2end_cnv_w2a2_streamlined.onnx"
+ )
+ iname = golden.graph.input[0].name
+ oname = golden.graph.output[0].name
+ # load one of the test vectors
+ fn = pk.resource_filename("finn", "data/cifar10/cifar10-test-data-class3.npz")
+ input_tensor = np.load(fn)["arr_0"].astype(np.float32)
+ input_tensor = input_tensor / 255
+ assert input_tensor.shape == (1, 3, 32, 32)
+ x = input_tensor
+ # x = np.zeros(ishape, dtype=np.float32)
+ ret_golden = execute_onnx(golden, {iname: x}, True)
+ y_golden = ret_golden[oname]
+ # set up parent+child graph to test
+ # we'll use models from the previous step as the child model
+ parent_model = load_test_checkpoint_or_skip(
+ build_dir + "/end2end_cnv_w2a2_dataflow_parent.onnx"
+ )
+ iname = parent_model.graph.input[0].name
+ oname = parent_model.graph.output[0].name
+ # produce results with cppsim
+ sdp_node = parent_model.get_nodes_by_op_type("StreamingDataflowPartition")[0]
+ sdp_node = getCustomOp(sdp_node)
+ load_test_checkpoint_or_skip(build_dir + "/end2end_cnv_w2a2_ipgen_cppsim.onnx")
+ sdp_node.set_nodeattr("model", build_dir + "/end2end_cnv_w2a2_ipgen_cppsim.onnx")
+ ret_cppsim = execute_onnx(parent_model, {iname: x}, True)
+ y_cppsim = ret_cppsim[oname]
+ # produce results with node-by-node rtlsim
+ load_test_checkpoint_or_skip(
+ build_dir + "/end2end_cnv_w2a2_ipgen_nodebynode_rtlsim.onnx"
+ )
+ sdp_node.set_nodeattr(
+ "model", build_dir + "/end2end_cnv_w2a2_ipgen_nodebynode_rtlsim.onnx"
+ )
+ ret_nodebynode_rtlsim = execute_onnx(parent_model, {iname: x}, True)
+ y_nodebynode_rtlsim = ret_nodebynode_rtlsim[oname]
+ # produce results with whole-network (stitched ip) rtlsim
+ load_test_checkpoint_or_skip(
+ build_dir + "/end2end_cnv_w2a2_ipstitch_whole_rtlsim.onnx"
+ )
+ sdp_node.set_nodeattr(
+ "model", build_dir + "/end2end_cnv_w2a2_ipstitch_whole_rtlsim.onnx"
+ )
+ # this is a particularly long-running test, set liveness thr. to unlimited
+ os.environ["LIVENESS_THRESHOLD"] = "-1"
+ ret_whole_rtlsim = execute_onnx(parent_model, {iname: x}, True)
+ y_whole_rtlsim = ret_whole_rtlsim[oname]
+ assert np.isclose(y_golden, y_cppsim).all()
+ assert np.isclose(y_golden, y_nodebynode_rtlsim).all()
+ assert np.isclose(y_golden, y_whole_rtlsim).all()
+ assert np.argmax(y_golden) == 3
+
+
+@pytest.mark.vivado
+def test_end2end_cnv_w2a2_make_pynq_proj():
+ model = load_test_checkpoint_or_skip(build_dir + "/end2end_cnv_w2a2_ipstitch.onnx")
+ model = model.transform(MakePYNQProject(test_pynq_board))
+ model.save(build_dir + "/end2end_cnv_w2a2_pynq_project.onnx")
+
+
+@pytest.mark.slow
+@pytest.mark.vivado
+def test_end2end_cnv_w2a2_synth_pynq_project():
+ model = load_test_checkpoint_or_skip(
+ build_dir + "/end2end_cnv_w2a2_pynq_project.onnx"
+ )
+ model = model.transform(SynthPYNQProject())
+ model = model.transform(AnnotateResources("synth"))
+ model.save(build_dir + "/end2end_cnv_w2a2_synth.onnx")
+
+
+def test_end2end_cnv_w2a2_make_driver():
+ model = load_test_checkpoint_or_skip(build_dir + "/end2end_cnv_w2a2_synth.onnx")
+ model = model.transform(MakePYNQDriver())
+ model.save(build_dir + "/end2end_cnv_w2a2_pynq_driver.onnx")
+
+
+def test_end2end_cnv_w2a2_deploy_on_pynq():
+ model = load_test_checkpoint_or_skip(
+ build_dir + "/end2end_cnv_w2a2_pynq_driver.onnx"
+ )
+ try:
+ ip = os.environ["PYNQ_IP"] # no fault for this one; skip if not defined
+ if ip == "":
+ pytest.skip("PYNQ board IP address not specified")
+ username = os.getenv("PYNQ_USERNAME", "xilinx")
+ password = os.getenv("PYNQ_PASSWORD", "xilinx")
+ port = os.getenv("PYNQ_PORT", 22)
+ target_dir = os.getenv("PYNQ_TARGET_DIR", "/home/xilinx/finn")
+ model = model.transform(DeployToPYNQ(ip, port, username, password, target_dir))
+ # save the model to be able to link it to the parent
+ model.save(build_dir + "/end2end_cnv_w2a2_pynq_deploy.onnx")
+ except KeyError:
+ pytest.skip("PYNQ board IP address not specified")
+
+
+def test_end2end_cnv_w2a2_run_on_pynq():
+ # use the streamlined model as the "golden" model for right answers
+ golden = load_test_checkpoint_or_skip(
+ build_dir + "/end2end_cnv_w2a2_streamlined.onnx"
+ )
+ iname = golden.graph.input[0].name
+ oname = golden.graph.output[0].name
+ # load one of the test vectors
+ fn = pk.resource_filename("finn", "data/cifar10/cifar10-test-data-class3.npz")
+ input_tensor = np.load(fn)["arr_0"].astype(np.float32)
+ input_tensor = input_tensor / 255
+ assert input_tensor.shape == (1, 3, 32, 32)
+ x = input_tensor
+ # run using FINN-based execution
+ ret_golden = execute_onnx(golden, {iname: x}, True)
+ y_golden = ret_golden[oname]
+ # set up parent+child graph to test
+ # we'll use models from the previous step as the child model
+ parent_model = load_test_checkpoint_or_skip(
+ build_dir + "/end2end_cnv_w2a2_dataflow_parent.onnx"
+ )
+ iname = parent_model.graph.input[0].name
+ oname = parent_model.graph.output[0].name
+ try:
+ ip = os.environ["PYNQ_IP"] # NOQA
+ if ip == "":
+ pytest.skip("PYNQ board IP address not specified")
+ # produce results with cppsim
+ sdp_node = parent_model.get_nodes_by_op_type("StreamingDataflowPartition")[0]
+ sdp_node = getCustomOp(sdp_node)
+ load_test_checkpoint_or_skip(build_dir + "/end2end_cnv_w2a2_pynq_deploy.onnx")
+ sdp_node.set_nodeattr("model", build_dir + "/end2end_cnv_w2a2_pynq_deploy.onnx")
+ ret = execute_onnx(parent_model, {iname: x}, True)
+ y = ret[oname]
+ assert np.isclose(y, y_golden).all()
+ assert np.argmax(y) == 3
+
+ except KeyError:
+ pytest.skip("PYNQ board IP address not specified")
diff --git a/tests/fpgadataflow/test_convert_to_hls_channelwise_layer.py b/tests/fpgadataflow/test_convert_to_hls_channelwise_layer.py
new file mode 100644
index 0000000000000000000000000000000000000000..d09c64a1250f78604c1a0a362cf234712de2cf57
--- /dev/null
+++ b/tests/fpgadataflow/test_convert_to_hls_channelwise_layer.py
@@ -0,0 +1,115 @@
+import pytest
+
+from onnx import TensorProto, helper
+
+import finn.core.onnx_exec as oxe
+from finn.core.datatype import DataType
+from finn.core.modelwrapper import ModelWrapper
+import finn.transformation.fpgadataflow.convert_to_hls_layers as to_hls
+from finn.transformation.fpgadataflow.prepare_ip import PrepareIP
+from finn.transformation.fpgadataflow.prepare_cppsim import PrepareCppSim
+from finn.transformation.fpgadataflow.compile_cppsim import CompileCppSim
+from finn.transformation.fpgadataflow.hlssynth_ip import HLSSynthIP
+from finn.transformation.fpgadataflow.replace_verilog_relpaths import (
+ ReplaceVerilogRelPaths,
+)
+from finn.transformation.fpgadataflow.set_exec_mode import SetExecMode
+
+from finn.transformation.fpgadataflow.prepare_rtlsim import PrepareRTLSim
+from finn.transformation.infer_data_layouts import InferDataLayouts
+from finn.transformation.general import GiveUniqueNodeNames
+from finn.util.basic import gen_finn_dt_tensor
+from finn.transformation.infer_shapes import InferShapes
+import numpy as np
+
+
+def prepare_inputs(input_tensor):
+ return {"inp": input_tensor}
+
+
+def make_single_maxpool_modelwrapper(onnx_op_name, ishape, idt, pdt, pshape):
+
+ inp = helper.make_tensor_value_info("inp", TensorProto.FLOAT, ishape)
+ outp = helper.make_tensor_value_info("outp", TensorProto.FLOAT, ishape)
+ p0 = helper.make_tensor_value_info("p0", TensorProto.FLOAT, pshape)
+
+ model = helper.make_model(
+ helper.make_graph(
+ name="test",
+ inputs=[inp],
+ outputs=[outp],
+ value_info=[p0],
+ nodes=[helper.make_node(onnx_op_name, ["inp", "p0"], ["outp"])],
+ )
+ )
+
+ model = ModelWrapper(model)
+ model.set_initializer("p0", gen_finn_dt_tensor(pdt, pshape))
+ model.set_tensor_datatype("inp", idt)
+ model.transform(InferDataLayouts(), make_deepcopy=False)
+ model.transform(InferShapes(), make_deepcopy=False)
+ return model
+
+
+# parameter datatype
+@pytest.mark.parametrize("pdt", [DataType.BIPOLAR, DataType.UINT4, DataType.INT2])
+# input datatype
+@pytest.mark.parametrize("idt", [DataType.INT32, DataType.UINT4, DataType.INT4])
+# function
+@pytest.mark.parametrize("onnx_op_name", ["Add", "Mul"])
+# vector parameter or scalar parameter (broadcast)
+@pytest.mark.parametrize("scalar_param", [True, False])
+# execution mode
+@pytest.mark.parametrize("exec_mode", ["cppsim", "rtlsim"])
+@pytest.mark.vivado
+@pytest.mark.slow
+def test_convert_to_hls_channelwise_layer(
+ pdt, idt, onnx_op_name, scalar_param, exec_mode
+):
+ ifm_ch = 16
+ ifm_dim = 5
+ ishape = (1, ifm_ch, ifm_dim, ifm_dim)
+ if scalar_param:
+ pshape = (1,)
+ else:
+ pshape = (1, ifm_ch, 1, 1)
+
+ np.random.seed(0)
+ model = make_single_maxpool_modelwrapper(onnx_op_name, ishape, idt, pdt, pshape)
+
+ # Since the aren't Data types with a bit width of a non power of 2,
+ # there are cases where the input won't use it full range.
+ if idt == DataType.INT32:
+ x = gen_finn_dt_tensor(DataType.INT16, (1, ifm_ch, ifm_dim, ifm_dim))
+ elif idt == DataType.UINT32:
+ x = gen_finn_dt_tensor(DataType.UINT16, (1, ifm_ch, ifm_dim, ifm_dim))
+ else:
+ x = gen_finn_dt_tensor(idt, (1, ifm_ch, ifm_dim, ifm_dim))
+
+ input_dict = prepare_inputs(x)
+ y_expected = oxe.execute_onnx(model, input_dict)["outp"]
+
+ new_model = model.transform(to_hls.InferChannelwiseLinearLayer())
+ new_model = new_model.transform(GiveUniqueNodeNames())
+
+ if exec_mode == "cppsim":
+ new_model = new_model.transform(PrepareCppSim())
+ new_model = new_model.transform(CompileCppSim())
+ new_model = new_model.transform(SetExecMode("cppsim"))
+ elif exec_mode == "rtlsim":
+ new_model = new_model.transform(SetExecMode("rtlsim"))
+ new_model = new_model.transform(GiveUniqueNodeNames())
+ new_model = new_model.transform(PrepareIP("xc7z020clg400-1", 5))
+ new_model = new_model.transform(HLSSynthIP())
+ new_model = new_model.transform(ReplaceVerilogRelPaths())
+ new_model = new_model.transform(PrepareRTLSim())
+ else:
+ raise Exception("Unknown exec_mode")
+
+ ctx_produced = oxe.execute_onnx(
+ new_model, input_dict, return_full_exec_context=True
+ )
+ y_produced = ctx_produced["outp"]
+
+ assert (y_produced == y_expected).all()
+ assert new_model.graph.node[1].op_type == "ChannelwiseOp_Batch"
diff --git a/tests/fpgadataflow/test_convert_to_hls_conv_layer.py b/tests/fpgadataflow/test_convert_to_hls_conv_layer.py
index ee65326ec57fb7fa7fa0490a8980dbabb8efc13c..22c356a5869b25fcc7ae3ef0164ed61b53ef232c 100644
--- a/tests/fpgadataflow/test_convert_to_hls_conv_layer.py
+++ b/tests/fpgadataflow/test_convert_to_hls_conv_layer.py
@@ -5,10 +5,15 @@ import pytest
from finn.core.datatype import DataType
from finn.transformation.infer_shapes import InferShapes
from finn.transformation.infer_datatypes import InferDataTypes
-from finn.transformation.general import GiveReadableTensorNames, GiveUniqueNodeNames
-from finn.transformation.infer_data_layouts import InferDataLayouts
+from finn.transformation.general import GiveUniqueNodeNames
from finn.transformation.lower_convs_to_matmul import LowerConvsToMatMul
+from finn.transformation.fpgadataflow.prepare_ip import PrepareIP
+from finn.transformation.fpgadataflow.prepare_rtlsim import PrepareRTLSim
+from finn.transformation.fpgadataflow.hlssynth_ip import HLSSynthIP
+from finn.transformation.fpgadataflow.replace_verilog_relpaths import (
+ ReplaceVerilogRelPaths,
+)
import finn.core.onnx_exec as oxe
from finn.core.modelwrapper import ModelWrapper
from finn.util.basic import gen_finn_dt_tensor
@@ -17,47 +22,40 @@ import finn.transformation.fpgadataflow.convert_to_hls_layers as to_hls
from finn.transformation.fpgadataflow.prepare_cppsim import PrepareCppSim
from finn.transformation.fpgadataflow.compile_cppsim import CompileCppSim
from finn.transformation.fpgadataflow.set_exec_mode import SetExecMode
+from finn.custom_op.im2col import compute_conv_output_dim
+# conv_config kernel_size,stride, pad
-@pytest.mark.parametrize("padding", [True, False])
-@pytest.mark.parametrize("kernel_size", [3, 5])
+
+@pytest.mark.parametrize(
+ "conv_config", [(1, 2, 0), (1, 3, 0), (3, 2, 1), (3, 1, 0), (3, 1, 1), (5, 2, 1)]
+)
+@pytest.mark.parametrize("exec_mode", ["cppsim", "rtlsim"])
@pytest.mark.slow
@pytest.mark.vivado
-def test_convert_to_hls_conv_layer(padding, kernel_size):
-
- assert (
- kernel_size % 2 != 0
- ), """test_convert_to_hls_conv_layer test only
- supports odd kernel_size"""
-
+def test_convert_to_hls_conv_layer(conv_config, exec_mode):
+ kernel_size, stride, pad = conv_config
np.random.seed(0)
- padding = True
idt = DataType.UINT4
in_feature_dim = 7
- in_chn = 3
+ in_chn = 16
+ out_chn = 20
- stages = 1 # just one convolution
-
- out_feature_dim = (
- in_feature_dim if padding else in_feature_dim - (kernel_size // 2 * 2) * stages
- )
+ out_feature_dim = compute_conv_output_dim(in_feature_dim, kernel_size, stride, pad)
input_shape = [1, in_chn, in_feature_dim, in_feature_dim]
- output_shape = [1, in_chn, out_feature_dim, out_feature_dim]
+ output_shape = [1, out_chn, out_feature_dim, out_feature_dim]
- conv_param_shape = [in_chn, in_chn, kernel_size, kernel_size]
+ conv_param_shape = [out_chn, in_chn, kernel_size, kernel_size]
+ conv_weight_dt = DataType.UINT4
conv_config = {}
conv_config["dilations"] = [1, 1]
conv_config["group"] = 1
conv_config["kernel_shape"] = [kernel_size, kernel_size]
- if padding:
- pad = kernel_size // 2
- conv_config["pads"] = [pad, pad, pad, pad]
- else:
- conv_config["pads"] = [0, 0, 0, 0]
- conv_config["strides"] = [1, 1]
+ conv_config["pads"] = [pad, pad, pad, pad]
+ conv_config["strides"] = [stride, stride]
top_in = helper.make_tensor_value_info("top_in", TensorProto.FLOAT, input_shape)
top_out = helper.make_tensor_value_info("top_out", TensorProto.FLOAT, output_shape)
@@ -80,27 +78,35 @@ def test_convert_to_hls_conv_layer(padding, kernel_size):
model = ModelWrapper(modelproto)
model.set_tensor_datatype("top_in", idt)
model.set_tensor_datatype("top_out", idt)
- model.set_tensor_datatype("p1", DataType.UINT4)
+ model.set_tensor_datatype("p1", conv_weight_dt)
+ model.set_initializer("p1", gen_finn_dt_tensor(conv_weight_dt, conv_param_shape))
model = model.transform(InferShapes())
- model.set_initializer(
- "p1", np.round(np.random.rand(*conv_param_shape).astype(np.float32) * 16)
- )
-
- model.set_tensor_datatype(model.graph.input[0].name, idt)
- model = model.transform(InferShapes())
- model = model.transform(InferDataLayouts())
- model = model.transform(GiveUniqueNodeNames())
- model = model.transform(GiveReadableTensorNames())
model = model.transform(InferDataTypes())
new_model = model.transform(LowerConvsToMatMul())
new_model = new_model.transform(to_hls.InferConvInpGen())
- new_model = new_model.transform(PrepareCppSim())
- new_model = new_model.transform(CompileCppSim())
- new_model = new_model.transform(SetExecMode("cppsim"))
+ new_model = new_model.transform(GiveUniqueNodeNames())
+ new_model = new_model.transform(InferShapes())
+ new_model = new_model.transform(InferDataTypes())
+
+ if exec_mode == "cppsim":
+ new_model = new_model.transform(PrepareCppSim())
+ new_model = new_model.transform(CompileCppSim())
+ new_model = new_model.transform(SetExecMode("cppsim"))
+ elif exec_mode == "rtlsim":
+ new_model = new_model.transform(SetExecMode("rtlsim"))
+ new_model = new_model.transform(GiveUniqueNodeNames())
+ new_model = new_model.transform(PrepareIP("xc7z020clg400-1", 5))
+ new_model = new_model.transform(HLSSynthIP())
+ new_model = new_model.transform(ReplaceVerilogRelPaths())
+ new_model = new_model.transform(PrepareRTLSim())
+ else:
+ raise Exception("Unknown exec_mode")
x = gen_finn_dt_tensor(idt, input_shape)
inp_dict = {model.graph.input[0].name: x}
assert oxe.compare_execution(model, new_model, inp_dict)
+ if kernel_size == 1 and stride > 1 and pad == 0:
+ assert new_model.graph.node[1].op_type == "DownSampler"
diff --git a/tests/fpgadataflow/test_convert_to_hls_pool_batch.py b/tests/fpgadataflow/test_convert_to_hls_pool_batch.py
new file mode 100644
index 0000000000000000000000000000000000000000..c9f78dcea1a1ce364d0657ad64de7d440d41b822
--- /dev/null
+++ b/tests/fpgadataflow/test_convert_to_hls_pool_batch.py
@@ -0,0 +1,160 @@
+# Copyright (c) 2020, Xilinx
+# All rights reserved.
+#
+# Redistribution and use in source and binary forms, with or without
+# modification, are permitted provided that the following conditions are met:
+#
+# * Redistributions of source code must retain the above copyright notice, this
+# list of conditions and the following disclaimer.
+#
+# * Redistributions in binary form must reproduce the above copyright notice,
+# this list of conditions and the following disclaimer in the documentation
+# and/or other materials provided with the distribution.
+#
+# * Neither the name of FINN nor the names of its
+# contributors may be used to endorse or promote products derived from
+# this software without specific prior written permission.
+#
+# THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS "AS IS"
+# AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
+# IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE ARE
+# DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT HOLDER OR CONTRIBUTORS BE LIABLE
+# FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL
+# DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR
+# SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER
+# CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY,
+# OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE
+# OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
+
+import pytest
+
+from onnx import TensorProto, helper
+import numpy as np
+import finn.core.onnx_exec as oxe
+from finn.core.datatype import DataType
+from finn.core.modelwrapper import ModelWrapper
+from finn.transformation.fpgadataflow.prepare_ip import PrepareIP
+from finn.transformation.fpgadataflow.prepare_cppsim import PrepareCppSim
+from finn.transformation.fpgadataflow.compile_cppsim import CompileCppSim
+from finn.transformation.fpgadataflow.hlssynth_ip import HLSSynthIP
+from finn.transformation.fpgadataflow.set_exec_mode import SetExecMode
+from finn.transformation.fpgadataflow.prepare_rtlsim import PrepareRTLSim
+import finn.transformation.fpgadataflow.convert_to_hls_layers as to_hls
+from finn.transformation.general import GiveUniqueNodeNames
+from finn.custom_op.registry import getCustomOp
+from finn.util.basic import gen_finn_dt_tensor
+from finn.transformation.infer_shapes import InferShapes
+
+
+def make_single_maxpool_modelwrapper(k, stride, pad, ifm_ch, ifm_dim, ofm_dim, idt):
+ odt = idt
+ inp = helper.make_tensor_value_info(
+ "inp", TensorProto.FLOAT, [1, ifm_ch, ifm_dim, ifm_dim]
+ )
+ outp = helper.make_tensor_value_info(
+ "outp", TensorProto.FLOAT, [1, ifm_ch, ofm_dim, ofm_dim]
+ )
+
+ mp_node = helper.make_node(
+ "MaxPool",
+ ["inp"],
+ ["outp"],
+ kernel_shape=[k, k],
+ pads=[pad, pad, pad, pad],
+ strides=[stride, stride],
+ )
+ graph = helper.make_graph(
+ nodes=[mp_node], name="mp_graph", inputs=[inp], outputs=[outp]
+ )
+
+ model = helper.make_model(graph, producer_name="mp-model")
+ model = ModelWrapper(model)
+
+ model.set_tensor_datatype("inp", idt)
+ model.set_tensor_datatype("outp", odt)
+ model = model.transform(InferShapes())
+
+ return model
+
+
+def prepare_inputs(input_tensor):
+ return {"inp": input_tensor}
+
+
+# input datatype
+@pytest.mark.parametrize("idt", [DataType.UINT4, DataType.INT4])
+# pool configuration: ( k,stride, pad, ifm_dim )
+@pytest.mark.parametrize(
+ "pool_config", [(3, 2, 0, 5), (3, 2, 1, 5), (2, 2, 0, 8), (5, 2, 2, 7)]
+)
+# input channels
+@pytest.mark.parametrize("ifm_ch", [1, 4, 20])
+# number of out channel computed in parallel
+@pytest.mark.parametrize("pe", [1, 4, 20])
+# execution mode
+@pytest.mark.parametrize("exec_mode", ["cppsim", "rtlsim"])
+# pool type
+@pytest.mark.parametrize("op_type", ["MaxPool"])
+@pytest.mark.slow
+@pytest.mark.vivado
+def test_convert_to_hls_pool_batch(idt, pool_config, ifm_ch, pe, exec_mode, op_type):
+ k, stride, pad, ifm_dim = pool_config
+
+ if ifm_ch % pe != 0:
+ pytest.skip("ifm_ch%pe != 0. Skipping")
+
+ if pad != 0 and idt.signed():
+ pytest.skip("No support for pal_val != 0. Skipping")
+
+ np.random.seed(0)
+ ofm_dim = int(((ifm_dim + 2 * pad - k) / stride) + 1)
+
+ x = gen_finn_dt_tensor(idt, (1, ifm_ch, ifm_dim, ifm_dim))
+ # prepare input data
+ input_dict = prepare_inputs(x)
+ if op_type == "MaxPool":
+ model = make_single_maxpool_modelwrapper(
+ k, stride, pad, ifm_ch, ifm_dim, ofm_dim, idt
+ )
+ else:
+ assert False, "{} is not a supported op_type".format(op_type)
+
+ y_expected = oxe.execute_onnx(model, input_dict)["outp"]
+
+ new_model = model.transform(to_hls.InferPool_Batch())
+ new_model = new_model.transform(GiveUniqueNodeNames())
+
+ if ifm_ch != pe:
+ new_model = new_model.transform(to_hls.InferConvInpGen())
+ # Folding
+ for n in new_model.graph.node:
+ if n.op_type == "ConvolutionInputGenerator":
+ inst = getCustomOp(n)
+ inst.set_nodeattr("SIMD", pe)
+ elif n.op_type == "Pool_Batch":
+ inst = getCustomOp(n)
+ inst.set_nodeattr("PE", pe)
+
+ if exec_mode == "cppsim":
+ new_model = new_model.transform(SetExecMode("cppsim"))
+ new_model = new_model.transform(PrepareCppSim())
+ new_model = new_model.transform(CompileCppSim())
+ elif exec_mode == "rtlsim":
+ new_model = new_model.transform(SetExecMode("rtlsim"))
+ new_model = new_model.transform(GiveUniqueNodeNames())
+ new_model = new_model.transform(PrepareIP("xc7z020clg400-1", 5))
+ new_model = new_model.transform(HLSSynthIP())
+ new_model = new_model.transform(PrepareRTLSim())
+ else:
+ raise Exception("Unknown exec_mode")
+
+ # execute new_model
+ y_produced = oxe.execute_onnx(new_model, input_dict)["outp"]
+ assert (y_produced == y_expected).all()
+ if stride != k:
+ if pad == 0 or ifm_ch == pe:
+ assert len(new_model.graph.node) == 4
+ else:
+ assert len(new_model.graph.node) == 5
+ else:
+ assert len(new_model.graph.node) == 1
diff --git a/tests/fpgadataflow/test_fpgadataflow_channelwise_ops.py b/tests/fpgadataflow/test_fpgadataflow_channelwise_ops.py
new file mode 100644
index 0000000000000000000000000000000000000000..2ed352e28981552b186bb778b94dcbc07471e14b
--- /dev/null
+++ b/tests/fpgadataflow/test_fpgadataflow_channelwise_ops.py
@@ -0,0 +1,156 @@
+# Copyright (c) 2020, Xilinx
+# All rights reserved.
+#
+# Redistribution and use in source and binary forms, with or without
+# modification, are permitted provided that the following conditions are met:
+#
+# * Redistributions of source code must retain the above copyright notice, this
+# list of conditions and the following disclaimer.
+#
+# * Redistributions in binary form must reproduce the above copyright notice,
+# this list of conditions and the following disclaimer in the documentation
+# and/or other materials provided with the distribution.
+#
+# * Neither the name of FINN nor the names of its
+# contributors may be used to endorse or promote products derived from
+# this software without specific prior written permission.
+#
+# THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS "AS IS"
+# AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
+# IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE ARE
+# DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT HOLDER OR CONTRIBUTORS BE LIABLE
+# FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL
+# DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR
+# SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER
+# CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY,
+# OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE
+# OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
+
+import pytest
+
+import numpy as np
+from onnx import TensorProto, helper
+
+import finn.core.onnx_exec as oxe
+from finn.analysis.fpgadataflow.hls_synth_res_estimation import hls_synth_res_estimation
+from finn.core.datatype import DataType
+from finn.core.modelwrapper import ModelWrapper
+from finn.transformation.fpgadataflow.prepare_ip import PrepareIP
+from finn.transformation.fpgadataflow.prepare_cppsim import PrepareCppSim
+from finn.transformation.fpgadataflow.compile_cppsim import CompileCppSim
+from finn.transformation.fpgadataflow.hlssynth_ip import HLSSynthIP
+from finn.transformation.fpgadataflow.set_exec_mode import SetExecMode
+from finn.transformation.general import GiveUniqueNodeNames
+from finn.transformation.fpgadataflow.prepare_rtlsim import PrepareRTLSim
+from finn.util.basic import gen_finn_dt_tensor
+from finn.transformation.fpgadataflow.replace_verilog_relpaths import (
+ ReplaceVerilogRelPaths,
+)
+
+
+def make_modelwrapper(C, pe, idt, odt, pdt, func, vecs):
+ NumChannels = C.shape[0]
+
+ inp = helper.make_tensor_value_info("inp", TensorProto.FLOAT, vecs + [NumChannels])
+ outp = helper.make_tensor_value_info(
+ "outp", TensorProto.FLOAT, vecs + [NumChannels]
+ )
+
+ node_inp_list = ["inp", "const"]
+
+ node = helper.make_node(
+ "ChannelwiseOp_Batch",
+ node_inp_list,
+ ["outp"],
+ domain="finn",
+ backend="fpgadataflow",
+ NumChannels=NumChannels,
+ Func=func,
+ PE=pe,
+ inputDataType=idt.name,
+ outputDataType=odt.name,
+ paramDataType=pdt.name,
+ numInputVectors=vecs,
+ )
+ graph = helper.make_graph(nodes=[node], name="graph", inputs=[inp], outputs=[outp])
+
+ model = helper.make_model(graph, producer_name="model")
+ model = ModelWrapper(model)
+
+ model.set_tensor_datatype("inp", idt)
+ model.set_tensor_datatype("outp", odt)
+
+ model.set_tensor_datatype("const", idt)
+ model.set_initializer("const", C)
+ return model
+
+
+# activation: None or DataType
+@pytest.mark.parametrize("act", [DataType.INT8])
+# input datatype
+@pytest.mark.parametrize("idt", [DataType.INT4])
+# param datatype
+@pytest.mark.parametrize("pdt", [DataType.INT4])
+# folding, -1 is maximum possible
+@pytest.mark.parametrize("nf", [-1, 2])
+# number of input features
+@pytest.mark.parametrize("ich", [16])
+# vecs
+@pytest.mark.parametrize("vecs", [[1], [1, 7, 7]])
+# function
+@pytest.mark.parametrize("func", ["add", "mul"])
+# execution mode
+@pytest.mark.parametrize("exec_mode", ["cppsim", "rtlsim"])
+@pytest.mark.vivado
+@pytest.mark.slow
+def test_fpgadataflow_channelwise_ops(idt, act, pdt, nf, ich, func, vecs, exec_mode):
+ if nf == -1:
+ nf = ich
+ pe = ich // nf
+ assert ich % pe == 0
+
+ # generate input and param data
+ x = gen_finn_dt_tensor(idt, tuple(vecs + [ich]))
+ # C = np.random.randint(idt.min(), idt.max() + 1, ich).astype(np.float32)
+ C = gen_finn_dt_tensor(pdt, (ich))
+
+ odt = act
+
+ model = make_modelwrapper(C, pe, idt, odt, pdt, func, vecs)
+
+ if exec_mode == "cppsim":
+ model = model.transform(PrepareCppSim())
+ model = model.transform(CompileCppSim())
+ model = model.transform(SetExecMode("cppsim"))
+ elif exec_mode == "rtlsim":
+ model = model.transform(SetExecMode("rtlsim"))
+ model = model.transform(GiveUniqueNodeNames())
+ model = model.transform(PrepareIP("xc7z020clg400-1", 5))
+ model = model.transform(HLSSynthIP())
+ model = model.transform(ReplaceVerilogRelPaths())
+ model = model.transform(PrepareRTLSim())
+ else:
+ raise Exception("Unknown exec_mode")
+
+ # package input data as dictionary
+ input_dict = {"inp": x}
+
+ oshape = model.get_tensor_shape("outp")
+
+ C_reshaped = np.broadcast_to(C.flatten(), x.shape)
+ if func == "add":
+ y = x + C_reshaped
+ elif func == "mul":
+ y = x * C_reshaped
+
+ y_expected = y.reshape(oshape)
+ # execute model
+ y_produced = oxe.execute_onnx(model, input_dict)["outp"]
+
+ y_produced = y_produced.reshape(y_expected.shape)
+
+ assert (y_produced == y_expected).all(), "cppsim failed"
+
+ if exec_mode == "rtlsim":
+ hls_synt_res_est = model.analysis(hls_synth_res_estimation)
+ assert "ChannelwiseOp_Batch_0" in hls_synt_res_est
diff --git a/tests/fpgadataflow/test_fpgadataflow_ipstitch.py b/tests/fpgadataflow/test_fpgadataflow_ipstitch.py
index b830693c32afe629dd6fc70868d0bddacac4c887..a9f5bf5ffa1f816b82ef701800e92249056b7c74 100644
--- a/tests/fpgadataflow/test_fpgadataflow_ipstitch.py
+++ b/tests/fpgadataflow/test_fpgadataflow_ipstitch.py
@@ -54,6 +54,10 @@ from finn.util.basic import gen_finn_dt_tensor, pynq_part_map
from finn.util.fpgadataflow import pyverilate_stitched_ip
from finn.util.test import load_test_checkpoint_or_skip
from finn.transformation.fpgadataflow.synth_ooc import SynthOutOfContext
+from finn.transformation.infer_data_layouts import InferDataLayouts
+from finn.transformation.fpgadataflow.insert_iodma import InsertIODMA
+from finn.transformation.fpgadataflow.floorplan import Floorplan
+
test_pynq_board = os.getenv("PYNQ_BOARD", default="Pynq-Z1")
test_fpga_part = pynq_part_map[test_pynq_board]
@@ -390,3 +394,19 @@ def test_fpgadataflow_ipstitch_remote_execution():
assert np.isclose(outp["outp"], x).all()
except KeyError:
pytest.skip("PYNQ board IP address not specified")
+
+
+def test_fpgadataflow_ipstitch_iodma_floorplan():
+ model = create_one_fc_model()
+ if model.graph.node[0].op_type == "StreamingDataflowPartition":
+ sdp_node = getCustomOp(model.graph.node[0])
+ assert sdp_node.__class__.__name__ == "StreamingDataflowPartition"
+ assert os.path.isfile(sdp_node.get_nodeattr("model"))
+ model = load_test_checkpoint_or_skip(sdp_node.get_nodeattr("model"))
+ model = model.transform(InferDataLayouts())
+ model = model.transform(InsertIODMA())
+ model = model.transform(Floorplan())
+ assert getCustomOp(model.graph.node[0]).get_nodeattr("partition_id") == 0
+ assert getCustomOp(model.graph.node[1]).get_nodeattr("partition_id") == 2
+ assert getCustomOp(model.graph.node[2]).get_nodeattr("partition_id") == 1
+ model.save(ip_stitch_model_dir + "/test_fpgadataflow_ipstitch_iodma_floorplan.onnx")
diff --git a/tests/fpgadataflow/test_fpgadataflow_res_estimate.py b/tests/fpgadataflow/test_fpgadataflow_res_estimate.py
index 38f792ed3cdd52044b28b4c19ac0603da4e502e6..398a17132a2ef6c92e600102ff5c0b71a1f65aaa 100644
--- a/tests/fpgadataflow/test_fpgadataflow_res_estimate.py
+++ b/tests/fpgadataflow/test_fpgadataflow_res_estimate.py
@@ -92,7 +92,7 @@ def test_res_estimate():
model = model.transform(GiveUniqueNodeNames())
prod_resource_estimation = model.analysis(res_estimation)
expect_resource_estimation = {
- "StreamingFCLayer_Batch_0": {"BRAM_18K": 1, "LUT": 304.4}
+ "StreamingFCLayer_Batch_0": {"BRAM_18K": 1, 'BRAM_efficiency': 0.001736111111111111, "LUT": 304.4}
}
assert check_two_dict_for_equality(
diff --git a/tests/pynq/test_pynq_performance_end2end.py b/tests/pynq/test_pynq_performance_end2end.py
index 66a93a190061e0142637be19bb2ea841d192745a..3b6ea86741b8adefce4faaa65b791f1d213cf3ae 100644
--- a/tests/pynq/test_pynq_performance_end2end.py
+++ b/tests/pynq/test_pynq_performance_end2end.py
@@ -10,7 +10,7 @@ from finn.core.throughput_test import throughput_test
build_dir = "/tmp/" + os.environ["FINN_INST_NAME"]
-@pytest.mark.parametrize("end2end_example", ["tfc_w1a1", "cnv_w1a1"])
+@pytest.mark.parametrize("end2end_example", ["tfc_w1a1", "cnv_w1a1", "cnv_w2a2"])
@pytest.mark.slow
def test_pynq_performance_end2end(end2end_example):
model = load_test_checkpoint_or_skip(
diff --git a/tests/transformation/streamline/test_streamline_cnv.py b/tests/transformation/streamline/test_streamline_cnv.py
index 56dcd26076ec0a5fba6e9be6acac7f5e13572c3d..103967dfb6b86cc6e2ce2bc9ab78249d8945d47d 100644
--- a/tests/transformation/streamline/test_streamline_cnv.py
+++ b/tests/transformation/streamline/test_streamline_cnv.py
@@ -44,9 +44,9 @@ from finn.transformation.double_to_single_float import DoubleToSingleFloat
export_onnx_path = make_build_dir("test_streamline_cnv_")
# act bits
-@pytest.mark.parametrize("abits", [1])
+@pytest.mark.parametrize("abits", [1, 2])
# weight bits
-@pytest.mark.parametrize("wbits", [1])
+@pytest.mark.parametrize("wbits", [1, 2])
# network topology / size
@pytest.mark.parametrize("size", ["CNV"])
def test_streamline_cnv(size, wbits, abits):
@@ -74,6 +74,7 @@ def test_streamline_cnv(size, wbits, abits):
# model.save("orig_cnv.onnx")
model = model.transform(Streamline())
# model.save("streamlined_cnv.onnx")
+ assert len(model.graph.node) == 23
produced_ctx = oxe.execute_onnx(model, input_dict, True)
produced = produced_ctx[model.graph.output[0].name]
assert np.isclose(expected, produced, atol=1e-3).all()
diff --git a/tests/transformation/test_absorb_mul_into_topk.py b/tests/transformation/test_absorb_mul_into_topk.py
new file mode 100644
index 0000000000000000000000000000000000000000..1394220f7c336ccea8fe9c494734c4175bf2e847
--- /dev/null
+++ b/tests/transformation/test_absorb_mul_into_topk.py
@@ -0,0 +1,108 @@
+# Copyright (c) 2020, Xilinx
+# All rights reserved.
+#
+# Redistribution and use in source and binary forms, with or without
+# modification, are permitted provided that the following conditions are met:
+#
+# * Redistributions of source code must retain the above copyright notice, this
+# list of conditions and the following disclaimer.
+#
+# * Redistributions in binary form must reproduce the above copyright notice,
+# this list of conditions and the following disclaimer in the documentation
+# and/or other materials provided with the distribution.
+#
+# * Neither the name of FINN nor the names of its
+# contributors may be used to endorse or promote products derived from
+# this software without specific prior written permission.
+#
+# THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS "AS IS"
+# AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
+# IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE ARE
+# DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT HOLDER OR CONTRIBUTORS BE LIABLE
+# FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL
+# DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR
+# SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER
+# CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY,
+# OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE
+# OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
+import pytest
+
+import numpy as np
+from onnx import TensorProto, helper
+
+from finn.core.modelwrapper import ModelWrapper
+from finn.transformation.infer_shapes import InferShapes
+from finn.transformation.infer_datatypes import InferDataTypes
+from finn.transformation.general import GiveUniqueNodeNames, GiveReadableTensorNames
+from finn.transformation.insert_topk import InsertTopK
+from finn.transformation.streamline.absorb import AbsorbScalarMulIntoTopK
+import finn.core.onnx_exec as oxe
+
+# parameter to indicate if mul parameter is negative or positive
+@pytest.mark.parametrize("mul_positive", [True, False])
+# parameter to indicate if mul parameter is scalar or not
+@pytest.mark.parametrize("scalar", [True, False])
+def test_absorb_mul_into_topk(mul_positive, scalar):
+ if scalar is True:
+ shape = [1]
+ else:
+ shape = [1, 1, 1, 1000]
+ inp = helper.make_tensor_value_info("inp", TensorProto.FLOAT, [1, 1, 1, 1000])
+ a0 = helper.make_tensor_value_info("a0", TensorProto.FLOAT, shape)
+ outp = helper.make_tensor_value_info("outp", TensorProto.FLOAT, [1, 1, 1, 1000])
+
+ mul_node = helper.make_node("Mul", ["inp", "a0"], ["outp"])
+ mul_graph = helper.make_graph(
+ nodes=[mul_node],
+ name="mul-graph",
+ inputs=[inp],
+ outputs=[outp],
+ value_info=[a0],
+ )
+
+ model = helper.make_model(mul_graph, producer_name="mul_model")
+ model = ModelWrapper(model)
+ # initialize values
+ if mul_positive is True:
+ a0_values = np.random.uniform(low=0.1, high=1, size=tuple(shape)).astype(
+ np.float32
+ )
+ else:
+ a0_values = np.random.uniform(low=-1, high=-0.1, size=tuple(shape)).astype(
+ np.float32
+ )
+ model.set_initializer("a0", a0_values)
+ model = model.transform(InsertTopK())
+ model = model.transform(InferShapes())
+ model = model.transform(InferDataTypes())
+ model = model.transform(GiveUniqueNodeNames())
+ model = model.transform(GiveReadableTensorNames())
+ model_transformed = model.transform(AbsorbScalarMulIntoTopK())
+
+ # compare execution results
+ inp_values = np.random.uniform(low=-10, high=10, size=(1, 1, 1, 1000)).astype(
+ np.float32
+ )
+ idict = {"global_in": inp_values}
+ odict = oxe.execute_onnx(model, idict, True)
+ y_indices = odict["global_out"]
+ y_values = odict["TopK_0_out0"]
+ odict = oxe.execute_onnx(model_transformed, idict, True)
+ y_tr_indices = odict["global_out"]
+ y_tr_values = odict["TopK_0_out0"]
+
+ # the indices stay the same, if the model is transformed or not
+ assert (y_indices == y_tr_indices).all()
+
+ if scalar is True and mul_positive is True:
+ # the values change if the model was transformed
+ assert (y_values != y_tr_values).all()
+
+ # check for new order
+ assert model.graph != model_transformed.graph
+ assert len(model.graph.node) - 1 == len(model_transformed.graph.node)
+ assert model_transformed.graph.node[0].op_type == "TopK"
+
+ else:
+ assert (y_values == y_tr_values).all()
+ assert model.graph == model_transformed.graph
diff --git a/tests/transformation/test_move_mul_past_dw_conv.py b/tests/transformation/test_move_mul_past_dw_conv.py
new file mode 100644
index 0000000000000000000000000000000000000000..1ae8fbfe89986d58d3d71f5f8735a98469d9d1e3
--- /dev/null
+++ b/tests/transformation/test_move_mul_past_dw_conv.py
@@ -0,0 +1,93 @@
+import pytest
+
+from onnx import helper, TensorProto
+from finn.custom_op.im2col import compute_conv_output_dim
+import finn.core.onnx_exec as oxe
+from finn.core.datatype import DataType
+from finn.core.modelwrapper import ModelWrapper
+from finn.transformation.infer_datatypes import InferDataTypes
+from finn.transformation.infer_shapes import InferShapes
+from finn.util.basic import gen_finn_dt_tensor
+from finn.transformation.streamline.reorder import MoveMulPastDWConv
+
+
+# input dimension
+@pytest.mark.parametrize("ifm_dim", [4, 7])
+# input channels
+@pytest.mark.parametrize("ifm_ch", [2, 3])
+# kernel size
+@pytest.mark.parametrize("k", [2, 3])
+# stride
+@pytest.mark.parametrize("stride", [1, 2])
+# padding
+@pytest.mark.parametrize("pad_amt", [0, 1])
+# depthwise
+@pytest.mark.parametrize("dw", [0, 1])
+def test_move_mul_past_dw_conv(ifm_dim, ifm_ch, k, stride, pad_amt, dw):
+ if dw == 1:
+ ofm_ch = ifm_ch
+ groups = ifm_ch
+ W_shape = [ofm_ch, 1, k, k]
+ else:
+ ofm_ch = ifm_ch + 2
+ groups = 1
+ W_shape = [ofm_ch, ifm_ch, k, k]
+
+ ofm_dim = compute_conv_output_dim(ifm_dim, k, stride, pad_amt)
+
+ # set up onnx model
+ inp = helper.make_tensor_value_info(
+ "inp", TensorProto.FLOAT, [1, ifm_ch, ifm_dim, ifm_dim]
+ )
+ mul = helper.make_tensor_value_info("mul", TensorProto.FLOAT, [1, ifm_ch, 1, 1])
+ W = helper.make_tensor_value_info("W", TensorProto.FLOAT, W_shape)
+ outp = helper.make_tensor_value_info(
+ "outp", TensorProto.FLOAT, [1, ofm_ch, ofm_dim, ofm_dim]
+ )
+
+ Mul_node = helper.make_node("Mul", ["inp", "mul"], ["mul_out"])
+
+ Conv_node = helper.make_node(
+ "Conv",
+ ["mul_out", "W"],
+ ["outp"],
+ group=groups,
+ kernel_shape=[k, k],
+ pads=[pad_amt, pad_amt, pad_amt, pad_amt],
+ strides=[stride, stride],
+ )
+
+ graph = helper.make_graph(
+ nodes=[Mul_node, Conv_node],
+ name="mulpastconv_graph",
+ inputs=[inp],
+ outputs=[outp],
+ value_info=[mul, W],
+ )
+
+ model = helper.make_model(graph, producer_name="mulpastconv-model")
+ model = ModelWrapper(model)
+ inp_values = gen_finn_dt_tensor(DataType.INT2, [1, ifm_ch, ifm_dim, ifm_dim])
+ mul_values = gen_finn_dt_tensor(DataType.INT2, [1, ifm_ch, 1, 1])
+ W_values = gen_finn_dt_tensor(DataType.INT2, W_shape)
+ model.set_initializer("W", W_values)
+ model.set_initializer("mul", mul_values)
+ model = model.transform(InferShapes())
+ model = model.transform(InferDataTypes())
+ idict = {"inp": inp_values}
+ odict = oxe.execute_onnx(model, idict, True)
+ out_before = odict["outp"]
+
+ # move channelwise multiplication past depthwise conv
+ model_transformed = model.transform(MoveMulPastDWConv())
+ odict = oxe.execute_onnx(model_transformed, idict, True)
+ out_after = odict["outp"]
+
+ assert (out_before == out_after).all()
+
+ if dw == 0:
+ assert model.graph.node[0].op_type == model_transformed.graph.node[0].op_type
+ assert model.graph.node[1].op_type == model_transformed.graph.node[1].op_type
+ else:
+ assert model.graph.node[0].op_type == model_transformed.graph.node[1].op_type
+ assert model.graph.node[1].op_type == model_transformed.graph.node[0].op_type
diff --git a/tests/transformation/test_move_transpose_past_scalar_mul.py b/tests/transformation/test_move_transpose_past_scalar_mul.py
new file mode 100644
index 0000000000000000000000000000000000000000..e434fc7d4f683120176e18a2bfa9da99d9ee0b0e
--- /dev/null
+++ b/tests/transformation/test_move_transpose_past_scalar_mul.py
@@ -0,0 +1,82 @@
+import pytest
+
+import numpy as np
+from onnx import TensorProto, helper
+
+from finn.core.modelwrapper import ModelWrapper
+import finn.core.data_layout as DataLayout
+from finn.transformation.infer_shapes import InferShapes
+from finn.transformation.infer_datatypes import InferDataTypes
+from finn.transformation.infer_data_layouts import InferDataLayouts
+from finn.transformation.general import GiveUniqueNodeNames, GiveReadableTensorNames
+from finn.transformation.streamline.reorder import MoveTransposePastScalarMul
+import finn.core.onnx_exec as oxe
+
+# permutation of transpose node
+@pytest.mark.parametrize("perm", [[0, 2, 3, 1], [0, 1, 3, 2], [3, 2, 0, 1]])
+# scalar mul
+@pytest.mark.parametrize("scalar", [True, False])
+# data layout
+@pytest.mark.parametrize("data_layout", [None, DataLayout.NHWC, DataLayout.NCHW])
+def test_move_transpose_past_scalar_mul(perm, scalar, data_layout):
+ inp = helper.make_tensor_value_info("inp", TensorProto.FLOAT, [1, 2, 3, 4])
+ # to determine out_size we need to calculate with "perm" for this test case
+ dummy_in = np.random.uniform(low=0, high=1, size=(1, 2, 3, 4)).astype(np.float32)
+ out_size = dummy_in.transpose(tuple(perm)).shape
+
+ if scalar is True:
+ a0_size = []
+ else:
+ a0_size = out_size
+ a0 = helper.make_tensor_value_info("a0", TensorProto.FLOAT, a0_size)
+ outp = helper.make_tensor_value_info("outp", TensorProto.FLOAT, out_size)
+ transp_node = helper.make_node("Transpose", ["inp"], ["transp_out"], perm=perm)
+ mul_node = helper.make_node("Mul", ["transp_out", "a0"], ["outp"])
+
+ graph = helper.make_graph(
+ nodes=[transp_node, mul_node],
+ name="mv-transpose-graph",
+ inputs=[inp],
+ outputs=[outp],
+ value_info=[a0],
+ )
+
+ model = helper.make_model(graph, producer_name="mv_transpose_model")
+ model = ModelWrapper(model)
+
+ # initialize values
+ a0_values = np.random.uniform(low=0, high=1, size=tuple(a0_size)).astype(np.float32)
+ model.set_initializer("a0", a0_values)
+ if data_layout is not None:
+ model.set_tensor_layout("inp", data_layout)
+ model = model.transform(InferDataLayouts())
+
+ model = model.transform(InferShapes())
+ model = model.transform(InferDataTypes())
+ model = model.transform(GiveUniqueNodeNames())
+ model = model.transform(GiveReadableTensorNames())
+
+ # compare execution before and after transformation
+ inp_values = np.random.uniform(low=0, high=1, size=(1, 2, 3, 4)).astype(np.float32)
+ idict = {model.graph.input[0].name: inp_values}
+ model_transformed = model.transform(MoveTransposePastScalarMul())
+ assert oxe.compare_execution(model, model_transformed, idict)
+
+ # check if order changed
+ if scalar is True and data_layout is not None:
+ assert model_transformed.graph.node[0] != model.graph.node[0]
+ assert model_transformed.graph.node[1] != model.graph.node[1]
+ assert model_transformed.graph.node[0].op_type == "Mul"
+ assert model_transformed.graph.node[1].op_type == "Transpose"
+ mul_input = model_transformed.graph.node[0].input[0]
+ mul_output = model_transformed.graph.node[0].output[0]
+ assert model_transformed.get_tensor_layout(mul_input) == data_layout
+ assert model_transformed.get_tensor_layout(mul_output) == data_layout
+ else:
+ assert model_transformed.graph.node[0] == model.graph.node[0]
+ assert model_transformed.graph.node[1] == model.graph.node[1]
+ if data_layout is not None:
+ mul_input = model_transformed.graph.node[1].input[0]
+ mul_output = model_transformed.graph.node[1].output[0]
+ assert model_transformed.get_tensor_layout(mul_input) != data_layout
+ assert model_transformed.get_tensor_layout(mul_output) != data_layout