diff --git a/.gitignore b/.gitignore
index f838c1695130d232ac6a2b888aed0cea31aafaa7..8b3166a44070a4575aac86c445c4504b594cda08 100644
--- a/.gitignore
+++ b/.gitignore
@@ -78,3 +78,6 @@ MANIFEST
# Jenkins cfg dir
/docker/jenkins_home
+
+# SSH key dir mounted into Docker
+/ssh_keys/
diff --git a/docker/Dockerfile.finn_dev b/docker/Dockerfile.finn_dev
index 22e3eb623c7a5da19a5e3ae2284557577898ad23..0e12b504a26ccdb8fd78e162f04cfdeab5a186f1 100644
--- a/docker/Dockerfile.finn_dev
+++ b/docker/Dockerfile.finn_dev
@@ -42,7 +42,7 @@ WORKDIR /workspace
RUN apt-get update
RUN apt-get -y upgrade
RUN apt-get install -y build-essential libglib2.0-0 libsm6 libxext6 libxrender-dev
-RUN apt-get install -y verilator nano zsh
+RUN apt-get install -y verilator nano zsh rsync
RUN apt-get -y install sshpass
RUN echo "StrictHostKeyChecking no" >> /etc/ssh/ssh_config
diff --git a/docker/finn_entrypoint.sh b/docker/finn_entrypoint.sh
index 6ad4adc352ca5d71993ee8118568df20dba00810..2a6e41797307f350a8827050f9d027e72d547e57 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=989cdfdba4700fdd900ba0b25a820591d561c21a
+BREVITAS_COMMIT=026a509186b7e7b0b65d46a2f905043d41069306
CNPY_COMMIT=4e8810b1a8637695171ed346ce68f6984e585ef4
-HLSLIB_COMMIT=afcfe75f3404249bddeeb3f15df65bd1fcb1072e
+HLSLIB_COMMIT=8f9f2018762f654f196b666838aeaf6fc730ad9a
PYVERILATOR_COMMIT=c97a5ba41bbc7c419d6f25c74cdf3bdc3393174f
PYNQSHELL_COMMIT=0c82a61b0ec1a07fa275a14146233824ded7a13d
OMX_COMMIT=1bae737669901e762f581af73348332b5c4b2ada
diff --git a/docs/finn/getting_started.rst b/docs/finn/getting_started.rst
index 95594bb67a2be3a4c3fbba488c75a704f623c136..feba21ea20d6445b88f96cf8ed2ca57598e614f4 100644
--- a/docs/finn/getting_started.rst
+++ b/docs/finn/getting_started.rst
@@ -18,6 +18,7 @@ Requirements
* A working Vivado 2019.1 installation
* A `VIVADO_PATH` environment variable pointing to the Vivado installation directory (e.g. the directory where settings64.sh is located)
* (optional) A PYNQ board with a network connection
+ * the ``bitstring`` package must be installed on the PYNQ: ``sudo pip install bitstring``
Running FINN in Docker
======================
diff --git a/docs/finn/internals.rst b/docs/finn/internals.rst
index 7a4bc687eeb827320991f7d3f1ef8cc35e97f3da..dee62f09a9253380e05300dac8fa34915c20dab5 100644
--- a/docs/finn/internals.rst
+++ b/docs/finn/internals.rst
@@ -16,6 +16,10 @@ Custom Quantization Annotations
ONNX does not support datatypes smaller than 8-bit integers, whereas in FINN we are interested in smaller integers down to ternary and bipolar. To make this work, FINN uses the quantization_annotation field in ONNX to annotate tensors with their FINN DataType (:py:mod:`finn.core.datatype.DataType`) information. However, all tensors are expected to use single-precision floating point (float32) storage in FINN. This means we store even a 1-bit value as floating point for the purposes of representation. The FINN compiler flow is responsible for eventually producing a packed representation for the target hardware, where the 1-bit is actually stored as 1-bit.
+Note that FINN uses floating point tensors as a carrier data type to represent integers. Floating point arithmetic can introduce rounding errors, e.g. (int_num * float_scale) / float_scale is not always equal to int_num.
+When using the custom ONNX execution flow, FINN will attempt to sanitize any rounding errors for integer tensors. See (:py:mod:`finn.util.basic.sanitize_quant_values`) for more information.
+This behavior can be disabled (not recommended!) by setting the environment variable SANITIZE_QUANT_TENSORS=0.
+
Custom Operations/Nodes
=======================
diff --git a/run-docker.sh b/run-docker.sh
index e07556716db335421f57a390f1e6a17168ac058b..00ca8f86985a78d8f2af099c51dcd4b80cd2e974 100755
--- a/run-docker.sh
+++ b/run-docker.sh
@@ -65,6 +65,11 @@ DOCKER_INST_NAME="finn_dev_${DOCKER_UNAME}"
# ensure Docker tag and inst. name are all lowercase
DOCKER_TAG=$(echo "$DOCKER_TAG" | tr '[:upper:]' '[:lower:]')
DOCKER_INST_NAME=$(echo "$DOCKER_INST_NAME" | tr '[:upper:]' '[:lower:]')
+# Absolute path to this script, e.g. /home/user/bin/foo.sh
+SCRIPT=$(readlink -f "$0")
+# Absolute path this script is in, thus /home/user/bin
+SCRIPTPATH=$(dirname "$SCRIPT")
+
# the settings below will be taken from environment variables if available,
# otherwise the defaults below will be used
: ${JUPYTER_PORT=8888}
@@ -74,11 +79,7 @@ DOCKER_INST_NAME=$(echo "$DOCKER_INST_NAME" | tr '[:upper:]' '[:lower:]')
: ${PYNQ_BOARD="Pynq-Z1"}
: ${PYNQ_TARGET_DIR="/home/xilinx/$DOCKER_INST_NAME"}
: ${NUM_DEFAULT_WORKERS=1}
-
-# Absolute path to this script, e.g. /home/user/bin/foo.sh
-SCRIPT=$(readlink -f "$0")
-# Absolute path this script is in, thus /home/user/bin
-SCRIPTPATH=$(dirname "$SCRIPT")
+: ${FINN_SSH_KEY_DIR="$SCRIPTPATH/ssh_keys"}
BUILD_LOCAL=/tmp/$DOCKER_INST_NAME
VIVADO_HLS_LOCAL=$VIVADO_PATH
@@ -87,6 +88,7 @@ VIVADO_IP_CACHE=$BUILD_LOCAL/vivado_ip_cache
# ensure build dir exists locally
mkdir -p $BUILD_LOCAL
mkdir -p $VIVADO_IP_CACHE
+mkdir -p $FINN_SSH_KEY_DIR
gecho "Instance is named as $DOCKER_INST_NAME"
gecho "Mounting $BUILD_LOCAL into $BUILD_LOCAL"
@@ -133,6 +135,7 @@ docker run -t --rm --name $DOCKER_INST_NAME $DOCKER_INTERACTIVE --init \
-v $SCRIPTPATH:/workspace/finn \
-v $BUILD_LOCAL:$BUILD_LOCAL \
-v $VIVADO_PATH:$VIVADO_PATH \
+-v $FINN_SSH_KEY_DIR:/home/$DOCKER_UNAME/.ssh \
-e VIVADO_PATH=$VIVADO_PATH \
-e FINN_INST_NAME=$DOCKER_INST_NAME \
-e FINN_ROOT="/workspace/finn" \
diff --git a/src/finn/core/onnx_exec.py b/src/finn/core/onnx_exec.py
index c2f68a35076418e0cf2edb578bdb8d548772fc78..efdfaa19d9f9e5dfa41911a2184e989337b3d9c2 100644
--- a/src/finn/core/onnx_exec.py
+++ b/src/finn/core/onnx_exec.py
@@ -39,6 +39,7 @@ from finn.core.remote_exec import remote_exec
from finn.core.rtlsim_exec import rtlsim_exec
from finn.custom_op.registry import getCustomOp
import finn.analysis.topology as ta
+from finn.util.basic import sanitize_quant_values, get_sanitize_quant_tensors
def execute_node(node, context, graph):
@@ -102,10 +103,7 @@ def execute_node(node, context, graph):
raise Exception(
"""Output shapes disagree after node execution:
found %s vs expected %s"""
- % (
- str(output_list[list_ind].shape),
- str(context[outp].shape),
- )
+ % (str(output_list[list_ind].shape), str(context[outp].shape))
)
context[outp] = output_list[list_ind]
@@ -162,7 +160,17 @@ def execute_onnx(model, input_dict, return_full_exec_context=False):
# we can simply walk down the list since the ONNX spec guarantees that it is
# topologically sorted
for node in graph.node:
+ if get_sanitize_quant_tensors() != 0:
+ # round input values to match quantization annotation
+ execution_context = sanitize_quant_values(
+ model, node.input, execution_context
+ )
execute_node(node, execution_context, graph)
+ if get_sanitize_quant_tensors() != 0:
+ # round output values to quantization annotation
+ execution_context = sanitize_quant_values(
+ model, node.output, execution_context
+ )
elif model_exec_mode == "remote_pynq":
# use remote exec metadata built into model to execute on a remote PYNQ
remote_exec(model, execution_context)
diff --git a/src/finn/custom_op/fpgadataflow/fmpadding.py b/src/finn/custom_op/fpgadataflow/fmpadding_batch.py
similarity index 88%
rename from src/finn/custom_op/fpgadataflow/fmpadding.py
rename to src/finn/custom_op/fpgadataflow/fmpadding_batch.py
index fa321dfa65d14b67fa218fb6a49f602ddab8d57e..d326ae7dfc7830a0081c3b13233d67ef08b12eff 100644
--- a/src/finn/custom_op/fpgadataflow/fmpadding.py
+++ b/src/finn/custom_op/fpgadataflow/fmpadding_batch.py
@@ -21,6 +21,8 @@ class FMPadding_Batch(HLSCustomOp):
"Padding": ("i", True, 2),
# number of channels in input image
"NumChannels": ("i", True, 0),
+ # SIMD Input parallelism
+ "SIMD": ("i", False, 1),
# FINN input datatype
"inputDataType": ("s", True, ""),
# controls distribution of padded pixels
@@ -55,20 +57,22 @@ class FMPadding_Batch(HLSCustomOp):
return oshape
def get_folded_input_shape(self):
- # even though there is no folding in the current hlslib op,
- # insert a time multiplexing axis to remain compatible with the
- # shapes produced by the rest of the dataflow pipeline
- ret = list(self.get_normal_input_shape())
- ret.insert(-1, 1)
- return tuple(ret)
+ 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):
- # even though there is no folding in the current hlslib op,
- # insert a time multiplexing axis to remain compatible with the
- # shapes produced by the rest of the dataflow pipeline
- ret = list(self.get_normal_output_shape())
- ret.insert(-1, 1)
- return tuple(ret)
+ 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()
@@ -114,15 +118,13 @@ class FMPadding_Batch(HLSCustomOp):
def get_instream_width(self):
ibits = self.get_input_datatype().bitwidth()
- num_ch = self.get_nodeattr("NumChannels")
-
- return ibits * num_ch
+ simd = self.get_nodeattr("SIMD")
+ return ibits * simd
def get_outstream_width(self):
obits = self.get_output_datatype().bitwidth()
- num_ch = self.get_nodeattr("NumChannels")
-
- return obits * num_ch
+ simd = self.get_nodeattr("SIMD")
+ return obits * simd
def get_number_output_values(self):
folded_oshape = self.get_folded_output_shape()
@@ -135,13 +137,15 @@ class FMPadding_Batch(HLSCustomOp):
self.code_gen_dict["$DEFINES$"] = [
"""#define ImgDim1 {}\n#define OutputDim1 {}\n
#define Padding1 {}\n#define NumChannels1 {}\n
- #define PaddingStyle1 {}\n#define numReps {}\n""".format(
+ #define PaddingStyle1 {}\n#define numReps {}
+ #define SIMD1 {}\n""".format(
self.get_nodeattr("ImgDim"),
self.get_padded_odim(),
self.get_nodeattr("Padding"),
self.get_nodeattr("NumChannels"),
self.get_nodeattr("PaddingStyle"),
self.get_nodeattr("numInputVectors"),
+ self.get_nodeattr("SIMD"),
)
]
@@ -176,7 +180,7 @@ class FMPadding_Batch(HLSCustomOp):
in_t = self.get_input_datatype().get_hls_datatype_str()
node = self.onnx_node
self.code_gen_dict["$DOCOMPUTE$"] = [
- """{}<ImgDim1, OutputDim1, Padding1, NumChannels1,
+ """{}<ImgDim1, OutputDim1, Padding1, NumChannels1,SIMD1,
{}, PaddingStyle1> (in0, out, numReps);""".format(
node.op_type, in_t
)
@@ -232,6 +236,7 @@ class FMPadding_Batch(HLSCustomOp):
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":
@@ -254,10 +259,8 @@ class FMPadding_Batch(HLSCustomOp):
match expected shape (1, ImgDim, ImgDim, NumChannels)."""
export_idt = self.get_input_datatype()
- # no reshaping for input since assuming no folding on input
- # make copy before saving array
- inp = inp.copy()
- np.save(os.path.join(code_gen_dir, "input_0.npy"), inp)
+ 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
diff --git a/src/finn/custom_op/fpgadataflow/tlastmarker.py b/src/finn/custom_op/fpgadataflow/tlastmarker.py
index 25ea05e3607a52731ae1b64de421837bf137ee2b..17ba44b959577faf573d77ae222f7b2a3be6669d 100644
--- a/src/finn/custom_op/fpgadataflow/tlastmarker.py
+++ b/src/finn/custom_op/fpgadataflow/tlastmarker.py
@@ -30,20 +30,30 @@ from finn.custom_op.fpgadataflow import HLSCustomOp
class TLastMarker(HLSCustomOp):
- """Class that corresponds to the TLastMarker node that needs to be
- inserted at the end of the model for rtlsim with stitched IP.
- It marks the end of the current image/input sample."""
+ """Node that adds/removes AXI stream TLAST signals where needed. Its behavior
+ is transparent in node-by-node execution, only visible in IP-stitched rtlsim or
+ actual hardware.
+ This node may be needed at the end of the network to signal a DMA write (needed by the
+ FINN PYNQ shell) or at the beginning to remove the end-of-burst from DMA read."""
def __init__(self, onnx_node):
super().__init__(onnx_node)
def get_nodeattr_types(self):
my_attrs = {
+ # number of (static) iterations until TLAST=1 is generated for Direction=out
"NumIters": ("i", True, 0),
+ # whether static or dynamic (from AXI lite) number of iterations are used
+ "DynIters": ("i", False, 1),
+ # direction: whether to insert or remove TLAST
+ "Direction": ("s", False, "out"),
# width of input-output data streams, in bits
"StreamWidth": ("i", True, 0),
# width of individual element in stream, in bits
"ElemWidth": ("i", True, 0),
+ # Protocol: external or internal
+ # Vitis docs recommend using qdma_axis for external, ap_axiu for internal
+ "Protocol": ("s", False, "external"),
}
my_attrs.update(super().get_nodeattr_types())
return my_attrs
@@ -76,12 +86,33 @@ class TLastMarker(HLSCustomOp):
def defines(self, var):
stream_width = self.get_nodeattr("StreamWidth")
+ direction = self.get_nodeattr("Direction")
+ protocol = self.get_nodeattr("Protocol")
# output stream must have TLAST, so we use this stream data type:
# qdma_axis<stream_data_width,0,0,0 >
- out_stream_dtype = "qdma_axis<%d,0,0,0>" % stream_width
+ if direction == "out":
+ if protocol == "external":
+ out_stream_dtype = "qdma_axis<%d,0,0,0>" % stream_width
+ elif protocol == "internal":
+ out_stream_dtype = "ap_axiu<%d,0,0,0>" % stream_width
+ else:
+ raise Exception("Unrecognized Protocol in TLastMarker")
+ in_stream_dtype = "ap_uint<%d>" % stream_width
+ elif direction == "in":
+ out_stream_dtype = "ap_uint<%d>" % stream_width
+ if protocol == "external":
+ in_stream_dtype = "qdma_axis<%d,0,0,0>" % stream_width
+ elif protocol == "internal":
+ in_stream_dtype = "ap_axiu<%d,0,0,0>" % stream_width
+ else:
+ raise Exception("Unrecognized Protocol in TLastMarker")
+ else:
+ raise Exception("Unrecognized Direction in TLastMarker")
+
self.code_gen_dict["$DEFINES$"] = [
"#define StreamWidth %d" % stream_width,
"#define OutDType %s" % out_stream_dtype,
+ "#define InDType %s" % in_stream_dtype,
"#define NumItersPerImg %d" % self.get_nodeattr("NumIters"),
]
@@ -89,27 +120,60 @@ class TLastMarker(HLSCustomOp):
self.code_gen_dict["$READNPYDATA$"] = []
def docompute(self):
- self.code_gen_dict["$DOCOMPUTE$"] = [
- "unsigned int n = 1;",
- "OutDType t;",
- "t.set_keep(-1);",
- "io_section: { // start of cycle accurate region",
- "#pragma HLS protocol fixed",
- "// do a first read from stream before we decide on numIters",
- "// giving software a chance to set up the numIters prior to startup",
- "t.set_data(in0.read());",
- "n = (numIters == 0 ? NumItersPerImg : numIters);",
- "t.set_last(n==1);",
- "out.write(t);",
- "} // end of cycle accurate region",
- "// do one less iteration than spec since we already did one",
- "for(unsigned int i=1; i<n; i++) {",
- "#pragma HLS PIPELINE II=1",
- "t.set_data(in0.read());",
- "t.set_last(i==(n-1));",
- "out.write(t);",
- "}",
- ]
+ dyn_iters = self.get_nodeattr("DynIters")
+ direction = self.get_nodeattr("Direction")
+ use_qdma_axis = self.get_nodeattr("Protocol") == "external"
+ if direction == "in":
+ # read from input and just pass data along; ignore tlast
+ # no dyn iters on input, it doesnt make sense
+ self.code_gen_dict["$DOCOMPUTE$"] = [
+ "for(unsigned int i=0; i<NumItersPerImg; i++) {",
+ "#pragma HLS PIPELINE II=1",
+ "out.write(in0.read().get_data());"
+ if use_qdma_axis
+ else "out.write(in0.read().data);",
+ "}",
+ ]
+
+ elif dyn_iters == 1:
+ # output, with dynamic iteration counts
+ self.code_gen_dict["$DOCOMPUTE$"] = [
+ "unsigned int n = 1;",
+ "OutDType t;",
+ "t.set_keep(-1);" if use_qdma_axis else "t.keep = -1;",
+ "io_section: { // start of cycle accurate region",
+ "#pragma HLS protocol fixed",
+ "// do a first read from stream before we decide on numIters",
+ "// giving software a chance to set up the numIters prior to startup",
+ "t.set_data(in0.read());" if use_qdma_axis else "t.data = in0.read();",
+ "n = (numIters == 0 ? NumItersPerImg : numIters);",
+ "t.set_last(n==1);" if use_qdma_axis else "t.last = (n==1);",
+ "out.write(t);",
+ "} // end of cycle accurate region",
+ "// do one less iteration than spec since we already did one",
+ "for(unsigned int i=1; i<n; i++) {",
+ "#pragma HLS PIPELINE II=1",
+ "t.set_data(in0.read());" if use_qdma_axis else "t.data = in0.read();",
+ "t.set_last(i==(n-1));" if use_qdma_axis else "t.last = (i==(n-1));",
+ "out.write(t);",
+ "}",
+ ]
+
+ else:
+ # output, with static iteration counts
+ self.code_gen_dict["$DOCOMPUTE$"] = [
+ "unsigned int n = 1;",
+ "OutDType t;",
+ "t.set_keep(-1);" if use_qdma_axis else "t.keep = -1;",
+ "for(unsigned int i=0; i<NumItersPerImg; i++) {",
+ "#pragma HLS PIPELINE II=1",
+ "t.set_data(in0.read());" if use_qdma_axis else "t.data = in0.read();",
+ "t.set_last(i==(NumItersPerImg-1));"
+ if use_qdma_axis
+ else "t.last = (i==(NumItersPerImg-1));",
+ "out.write(t);",
+ "}",
+ ]
def dataoutstrm(self):
self.code_gen_dict["$DATAOUTSTREAM$"] = []
@@ -118,18 +182,30 @@ class TLastMarker(HLSCustomOp):
self.code_gen_dict["$SAVEASCNPY$"] = []
def blackboxfunction(self):
- self.code_gen_dict["$BLACKBOXFUNCTION$"] = [
- """void %s(hls::stream<ap_uint<StreamWidth> > &in0,
- hls::stream<OutDType> &out, unsigned int numIters)"""
- % self.onnx_node.name
- ]
+ dyn_iters = self.get_nodeattr("DynIters")
+
+ if dyn_iters == 1:
+ self.code_gen_dict["$BLACKBOXFUNCTION$"] = [
+ """void %s(hls::stream<InDType> &in0,
+ hls::stream<OutDType> &out, unsigned int numIters)"""
+ % self.onnx_node.name
+ ]
+ else:
+ self.code_gen_dict["$BLACKBOXFUNCTION$"] = [
+ """void %s(hls::stream<InDType> &in0, hls::stream<OutDType> &out)"""
+ % self.onnx_node.name
+ ]
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 s_axilite port=numIters bundle=control"
- )
+
+ dyn_iters = self.get_nodeattr("DynIters")
+ if dyn_iters == 1:
+ self.code_gen_dict["$PRAGMAS$"].append(
+ "#pragma HLS INTERFACE s_axilite port=numIters bundle=control"
+ )
+
self.code_gen_dict["$PRAGMAS$"].append(
"#pragma HLS INTERFACE ap_ctrl_none port=return"
)
@@ -158,7 +234,7 @@ class TLastMarker(HLSCustomOp):
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())
+ 'hls::stream<InDType> in0 ("in0");'
)
self.code_gen_dict["$STREAMDECLARATIONS$"].append(
'hls::stream<OutDType> out ("out");'
diff --git a/src/finn/custom_op/quantavgpool2d.py b/src/finn/custom_op/quantavgpool2d.py
new file mode 100644
index 0000000000000000000000000000000000000000..3bc328a9f4f6670041d33491d58af6c553bafac9
--- /dev/null
+++ b/src/finn/custom_op/quantavgpool2d.py
@@ -0,0 +1,83 @@
+import numpy as np
+from onnx import TensorProto, helper
+import onnxruntime as rt
+
+from finn.custom_op import CustomOp
+from finn.core.datatype import DataType
+
+
+class QuantAvgPool2d(CustomOp):
+ """Class that corresponds to the quantized average pooling
+ layer from brevitas"""
+
+ def get_nodeattr_types(self):
+ return {
+ "stride": ("i", True, 1),
+ "kernel": ("i", True, 1),
+ "ibits": ("i", True, 1),
+ "obits": ("i", True, 1),
+ "signed": ("i", True, 0),
+ }
+
+ def make_shape_compatible_op(self, model):
+ node = self.onnx_node
+ k = self.get_nodeattr("kernel")
+ s = self.get_nodeattr("stride")
+ return helper.make_node(
+ "AveragePool",
+ inputs=[node.input[0]],
+ outputs=[node.output[0]],
+ kernel_shape=[k, k],
+ strides=[s, s],
+ )
+
+ def infer_node_datatype(self, model):
+ node = self.onnx_node
+ bw = self.get_nodeattr("obits")
+ if bw in [2, 4, 8, 16, 32]:
+ if self.get_nodeattr("signed") == 0:
+ dtype = DataType["UINT%d" % bw]
+ else:
+ dtype = DataType["INT%d" % bw]
+ else:
+ raise Exception("Unsupported output datatype for QuantAvgPool2d")
+ model.set_tensor_datatype(node.output[0], dtype)
+
+ def execute_node(self, context, graph):
+ # create a standard average pooling node to help calculate the result
+ node = self.onnx_node
+ k = self.get_nodeattr("kernel")
+ s = self.get_nodeattr("stride")
+ ishape = context[node.input[0]].shape
+ oshape = context[node.output[0]].shape
+ inp = helper.make_tensor_value_info(node.input[0], TensorProto.FLOAT, ishape)
+ outp = helper.make_tensor_value_info(node.output[0], TensorProto.FLOAT, oshape)
+ node_avgpool = helper.make_node(
+ "AveragePool",
+ inputs=[node.input[0]],
+ outputs=[node.output[0]],
+ kernel_shape=[k, k],
+ strides=[s, s],
+ )
+ graph_avgpool = helper.make_graph(
+ nodes=[node_avgpool],
+ name="single-avgpool-exec",
+ inputs=[inp],
+ outputs=[outp],
+ )
+ model_avgpool = helper.make_model(graph_avgpool)
+ idict = {node.input[0]: context[node.input[0]]}
+ sess = rt.InferenceSession(model_avgpool.SerializeToString())
+ result_temp = sess.run(None, idict)
+ # remove scaling introduced by average
+ result_temp = result_temp[0] * (k * k)
+ ibits = self.get_nodeattr("ibits")
+ max_value = 2 ** ibits - 1
+ max_value = max_value * k * k
+ max_bit_width = int(max_value).bit_length()
+ shift_bits = max_bit_width - self.get_nodeattr("obits")
+ result = np.right_shift(result_temp.astype(int), shift_bits)
+ context[node.output[0]] = result.astype(np.float32)
+
+ def verify_node(self):
+ pass
diff --git a/src/finn/custom_op/registry.py b/src/finn/custom_op/registry.py
index 6105b1342595fb083a194b6d0fc4af3fedada7ba..46d27472a9802a4c2a9004bb28c8bd09be8fbfdb 100644
--- a/src/finn/custom_op/registry.py
+++ b/src/finn/custom_op/registry.py
@@ -44,10 +44,11 @@ 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.fmpadding import FMPadding_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
@@ -70,6 +71,7 @@ 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
diff --git a/src/finn/transformation/fpgadataflow/insert_tlastmarker.py b/src/finn/transformation/fpgadataflow/insert_tlastmarker.py
index 32f32ece585a93465ba32fede45d5eb606a2b0a3..04dd437af27b9fbe18b2255c20a8e4acda03b3d0 100644
--- a/src/finn/transformation/fpgadataflow/insert_tlastmarker.py
+++ b/src/finn/transformation/fpgadataflow/insert_tlastmarker.py
@@ -31,23 +31,34 @@ from onnx import helper as oh
from finn.custom_op.registry import getCustomOp
from finn.transformation import Transformation
+from finn.util.basic import get_by_name
+
+import numpy as np
class InsertTLastMarker(Transformation):
- """Ensure that the graph is terminated with a TLastMarker node, inserting
- one if necessary."""
+ """Ensure that the graph is started/terminated with a TLastMarker node, inserting
+ one if necessary. Use constructor args to determine type of TLastMarker to be inserted.
+ More information available on the TLastMarker documentation.
+ """
- def __init__(self):
+ def __init__(self, both=False, external=True, dynamic=True):
super().__init__()
+ self.dyniters = dynamic
+ self.external = external
+ self.both = both
def apply(self, model):
# TODO only makes sense for a pure fpgadataflow graph -- check!
graph_out_name = model.graph.output[0].name
final_node = model.find_producer(graph_out_name)
- if final_node.op_type == "TLastMarker":
- # TODO maybe check the correctness of properties
- return (model, False)
- else:
+ graph_modified = False
+ if final_node.op_type != "TLastMarker" and not (
+ final_node.op_type == "IODMA"
+ and get_by_name(final_node.attribute, "direction").s.decode("UTF-8")
+ == "out"
+ ):
+
custom_op = getCustomOp(final_node)
num_iters = int(custom_op.get_number_output_values())
stream_width = int(custom_op.get_outstream_width())
@@ -69,8 +80,51 @@ class InsertTLastMarker(Transformation):
NumIters=num_iters,
StreamWidth=stream_width,
ElemWidth=elem_width,
+ DynIters=(1 if self.dyniters else 0),
+ Direction="out",
+ Protocol=("external" if self.external else "internal"),
domain="finn",
backend="fpgadataflow",
)
model.graph.node.append(tlast_node)
- return (model, True)
+ graph_modified = True
+ # if both is True, also insert marker on input
+ if self.both:
+ graph_in_name = model.graph.input[0].name
+ first_node = model.find_consumer(graph_in_name)
+ if first_node.op_type != "TLastMarker" and not (
+ first_node.op_type == "IODMA"
+ and get_by_name(first_node.attribute, "direction").s.decode("UTF-8")
+ == "in"
+ ):
+
+ custom_op = getCustomOp(first_node)
+ num_iters = np.prod(custom_op.get_folded_input_shape()[1:-1])
+ stream_width = int(custom_op.get_instream_width())
+ in_shape = model.get_tensor_shape(graph_in_name)
+ in_dtype = model.get_tensor_datatype(graph_in_name)
+ elem_width = in_dtype.bitwidth()
+ # 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 first_node_in_name
+ first_node.input[0] = first_node_in.name
+ tlast_node = oh.make_node(
+ "TLastMarker",
+ [graph_in_name],
+ [first_node_in.name],
+ NumIters=num_iters,
+ StreamWidth=stream_width,
+ ElemWidth=elem_width,
+ DynIters=(1 if self.dyniters else 0),
+ Direction="in",
+ Protocol=("external" if self.external else "internal"),
+ domain="finn",
+ backend="fpgadataflow",
+ )
+ model.graph.node.insert(0, tlast_node)
+ graph_modified = True
+ return (model, graph_modified)
diff --git a/src/finn/transformation/fpgadataflow/prepare_cppsim.py b/src/finn/transformation/fpgadataflow/prepare_cppsim.py
index a1524322ec03a4e96ef41f999144e3eed349c5af..4f050be8540ddf5ef48699d1658b571852ff4510 100644
--- a/src/finn/transformation/fpgadataflow/prepare_cppsim.py
+++ b/src/finn/transformation/fpgadataflow/prepare_cppsim.py
@@ -29,9 +29,12 @@
import os
import finn.custom_op.registry as registry
-from finn.transformation import Transformation
from finn.util.basic import make_build_dir
from finn.util.fpgadataflow import is_fpgadataflow_node
+from finn.transformation import Transformation
+from finn.util.basic import get_num_default_workers
+import multiprocessing as mp
+import copy
def _codegen_single_node(node, model):
@@ -66,8 +69,40 @@ class PrepareCppSim(Transformation):
that contains generated C++ code that can be used to simulate node using cppsim.
The subsequent transformation is CompileCppSim"""
+ def __init__(self, num_workers=None):
+ super().__init__()
+ if num_workers is None:
+ self._num_workers = get_num_default_workers()
+ else:
+ self._num_workers = num_workers
+ assert self._num_workers >= 0, "Number of workers must be nonnegative."
+ if self._num_workers == 0:
+ 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)
+
def apply(self, model):
- for node in model.graph.node:
- if is_fpgadataflow_node(node) is True:
- _codegen_single_node(node, model)
- return (model, False)
+ # Remove old nodes from the current model
+ self.model = copy.deepcopy(model)
+ old_nodes = []
+ for i in range(len(model.graph.node)):
+ old_nodes.append(model.graph.node.pop())
+
+ # Execute transformation in parallel
+ with mp.Pool(self._num_workers) as p:
+ new_nodes_and_bool = p.map(self.prepareCppSim_node, old_nodes, chunksize=1)
+
+ # extract nodes and check if the transformation needs to run again
+ # Note: .pop() had initially reversed the node order
+ run_again = False
+ for node, run in reversed(new_nodes_and_bool):
+ # Reattach new nodes to old model
+ model.graph.node.append(node)
+ if run is True:
+ run_again = True
+
+ return (model, run_again)
diff --git a/src/finn/transformation/infer_datatypes.py b/src/finn/transformation/infer_datatypes.py
index 1acd4e3abe2d77248810cf15c15475e806a3bd32..39b7a787be8c725e7b6d474757dd96fc4848dfe0 100644
--- a/src/finn/transformation/infer_datatypes.py
+++ b/src/finn/transformation/infer_datatypes.py
@@ -71,7 +71,13 @@ def _infer_node_datatype(model, node):
else:
# unknown, assume node produces float32 outputs
for o in node.output:
- model.set_tensor_datatype(o, DataType.FLOAT32)
+ # check if output datatype is already set to a value != FLOAT32
+ odtype = model.get_tensor_datatype(o)
+ if odtype is not None and odtype != DataType.FLOAT32:
+ # don't change data type
+ model.set_tensor_datatype(o, odtype)
+ else:
+ model.set_tensor_datatype(o, DataType.FLOAT32)
# compare old and new output dtypes to see if anything changed
new_odtypes = list(map(lambda x: model.get_tensor_datatype(x), node.output))
graph_modified = new_odtypes != odtypes
diff --git a/src/finn/transformation/lower_convs_to_matmul.py b/src/finn/transformation/lower_convs_to_matmul.py
index 3da785d8dd21b2c6701bffc8ce3869fb14b237a9..aa231a43a3865a161a501b4997ff2f538800554f 100644
--- a/src/finn/transformation/lower_convs_to_matmul.py
+++ b/src/finn/transformation/lower_convs_to_matmul.py
@@ -80,14 +80,19 @@ class LowerConvsToMatMul(Transformation):
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)
+ need_im2col = True
+ if k == 1 and pad == 0 and stride == 1:
+ need_im2col = False
+
+ if need_im2col:
+ 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)
matmul_out = helper.make_tensor_value_info(
model.make_new_valueinfo_name(),
@@ -104,19 +109,23 @@ class LowerConvsToMatMul(Transformation):
"Transpose", [cnv_input], [inp_trans_out], perm=[0, 2, 3, 1]
)
# lower input tensor
- im2col_node = helper.make_node(
- "Im2Col",
- [inp_trans_out],
- [im2col_out],
- domain="finn",
- stride=stride,
- kernel_size=k,
- pad_amount=pad,
- input_shape="(1,{},{},{})".format(ifm_dim, ifm_dim, ifm_ch),
- )
+ matmul_input = inp_trans_out
+ if need_im2col:
+ matmul_input = im2col_out
+ im2col_node = helper.make_node(
+ "Im2Col",
+ [inp_trans_out],
+ [im2col_out],
+ domain="finn",
+ stride=stride,
+ kernel_size=k,
+ pad_amount=pad,
+ input_shape="(1,{},{},{})".format(ifm_dim, ifm_dim, ifm_ch),
+ )
+
# do matmul
matmul_node = helper.make_node(
- "MatMul", [im2col_out, weight_name], [matmul_out]
+ "MatMul", [matmul_input, weight_name], [matmul_out]
)
# NHWC -> NCHW
out_trans_node = helper.make_node(
@@ -124,9 +133,13 @@ class LowerConvsToMatMul(Transformation):
)
# 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, matmul_node)
- graph.node.insert(node_ind + 3, out_trans_node)
+ if need_im2col:
+ graph.node.insert(node_ind + 1, im2col_node)
+ graph.node.insert(node_ind + 2, matmul_node)
+ graph.node.insert(node_ind + 3, out_trans_node)
+ else:
+ graph.node.insert(node_ind + 1, matmul_node)
+ graph.node.insert(node_ind + 2, out_trans_node)
# remove old nodes
graph.node.remove(n)
model = model.transform(InferShapes())
diff --git a/src/finn/transformation/streamline/collapse_repeated.py b/src/finn/transformation/streamline/collapse_repeated.py
index 67824ad4f633983b93e3178d03118927a1ddd85b..769bed841ce07c1c9c62f762de4b2c0937a6d68f 100644
--- a/src/finn/transformation/streamline/collapse_repeated.py
+++ b/src/finn/transformation/streamline/collapse_repeated.py
@@ -30,6 +30,7 @@ from onnx import helper as oh
from finn.transformation import Transformation
from finn.transformation.infer_shapes import InferShapes
+from finn.core.datatype import DataType
class CollapseRepeatedOp(Transformation):
@@ -83,6 +84,9 @@ class CollapseRepeatedOp(Transformation):
graph.node.insert(node_ind, new_node)
# replace parameter value
model.set_initializer(new_node_param_name, new_param)
+ # be conservative with param/output DataTypes
+ model.set_tensor_datatype(new_node_param_name, DataType.FLOAT32)
+ model.set_tensor_datatype(end_name, DataType.FLOAT32)
# remove old nodes
graph.node.remove(n)
graph.node.remove(consumer)
diff --git a/src/finn/transformation/streamline/remove.py b/src/finn/transformation/streamline/remove.py
new file mode 100644
index 0000000000000000000000000000000000000000..ddc4233ddafbc70c4d20d316ea72ea6bba1b82a8
--- /dev/null
+++ b/src/finn/transformation/streamline/remove.py
@@ -0,0 +1,69 @@
+# 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
+from finn.transformation.infer_shapes import InferShapes
+import numpy as np
+
+class RemoveIdentityOps(Transformation):
+ """Remove identity ops like Add/Sub with zero or Mul/Div with one"""
+
+ 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 ["Add", "Sub"]
+ and not model.is_fork_node(n)
+ and not model.is_join_node(n)
+ ):
+ A = model.get_initializer(n.input[1])
+ if A is not None and (A == np.zeros_like(A)).all():
+ producer = model.find_producer(n.input[0])
+ # remove node and wire output tensor to
+ # output of producer node
+ producer.output[0] = n.output[0]
+ graph.node.remove(n)
+
+ elif (
+ n.op_type in ["Mul", "Div"]
+ and not model.is_fork_node(n)
+ and not model.is_join_node(n)
+ ):
+ A = model.get_initializer(n.input[1])
+ if A is not None and (A == np.ones_like(A)).all():
+ producer = model.find_producer(n.input[0])
+ # remove node and wire output tensor to
+ # output of producer node
+ producer.output[0] = n.output[0]
+ graph.node.remove(n)
+ model = model.transform(InferShapes())
+ return (model, graph_modified)
diff --git a/src/finn/transformation/streamline/reorder.py b/src/finn/transformation/streamline/reorder.py
index 0b6259a61d3eb67b7b38d4c6939019ce2893a875..b46b82c77a3f1b70a3b05d87cd3c48fc1d94fd45 100644
--- a/src/finn/transformation/streamline/reorder.py
+++ b/src/finn/transformation/streamline/reorder.py
@@ -27,12 +27,14 @@
# OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
import numpy as np
+import warnings
from onnx import helper as oh
from finn.transformation import Transformation
from finn.transformation.infer_shapes import InferShapes
from finn.core.onnx_exec import execute_node
from finn.util.basic import get_by_name
+from finn.custom_op.registry import getCustomOp
class MoveAddPastMul(Transformation):
@@ -531,3 +533,67 @@ class MoveMulPastFork(MoveOpPastFork):
class MoveLinearPastFork(MoveOpPastFork):
def __init__(self):
super().__init__(["Add", "Mul"])
+
+
+class MoveMaxPoolPastMultiThreshold(Transformation):
+ """Move MaxPool nodes past MultiThreshold nodes on linear segments of the graph."""
+
+ def apply(self, model):
+ graph = model.graph
+ node_ind = 0
+ graph_modified = False
+ nodes = [n for n in graph.node]
+ for n in nodes:
+ node_ind += 1
+ if n.op_type == "MaxPool" and not model.is_fork_node(n):
+ consumer = model.find_consumer(n.output[0])
+ pads = get_by_name(n.attribute, "pads")
+ has_padding = False
+ if pads is not None:
+ pads = list(pads.ints)
+ has_padding = np.prod(pads) != 0
+ if consumer is not None and consumer.op_type == "MultiThreshold":
+ mt_out = consumer.output[0]
+ mt_odt = model.get_tensor_datatype(mt_out)
+ if mt_odt.signed() and has_padding:
+ warnings.warn(
+ "Skipping padded MaxPool + signed-output MultiThreshold"
+ )
+ continue
+ # check for non-decreasing thresholds and nonnegative
+ # scale factor in MultiThreshold
+ # otherwise we cannot do the reordering
+ T = model.get_initializer(consumer.input[1])
+ T_sorted = np.sort(T, axis=1)
+ assert (
+ T == T_sorted
+ ).all(), "MultiThreshold must have non-decreasing thresholds"
+ mt_inst = getCustomOp(consumer)
+ if mt_inst.get_nodeattr("out_scale") < 0:
+ warnings.warn("Skipping MultiThreshold with negative out_scale")
+ continue
+
+ # remove old nodes
+ graph.node.remove(n)
+ graph.node.remove(consumer)
+
+ # swap conections
+ group_in = n.input[0]
+ # new tensor because dims change
+ group_middle = model.make_new_valueinfo_name()
+ group_out = consumer.output[0]
+
+ consumer.input[0] = group_in
+ consumer.output[0] = group_middle
+
+ n.input[0] = group_middle
+ n.output[0] = group_out
+
+ # insert them back in
+ graph.node.insert(node_ind - 1, consumer)
+ graph.node.insert(node_ind, n)
+
+ graph_modified = True
+
+ model = model.transform(InferShapes())
+ return (model, graph_modified)
diff --git a/src/finn/util/basic.py b/src/finn/util/basic.py
index eb3d46bcd66e3dc307a679e6b8dfbb9913398d36..4a8277e08d3fc21e0b20668edf2ecad947b36647 100644
--- a/src/finn/util/basic.py
+++ b/src/finn/util/basic.py
@@ -31,6 +31,7 @@ import random
import string
import subprocess
import tempfile
+import warnings
import numpy as np
@@ -105,6 +106,25 @@ def get_finn_root():
)
+def get_execution_error_thresh():
+ "Return the max error that is allowed for rounding in FINN execution."
+ try:
+ return float(os.environ["ERROR_THRESH"])
+ except KeyError:
+ return 1e-2
+
+
+def get_sanitize_quant_tensors():
+ """Return whether tensors with quantization annotations should be sanitized.
+ Enabled by default, disabling will yield faster ONNX execution but may give
+ incorrect results. Use with caution."""
+ try:
+ return int(os.environ["SANITIZE_QUANT_TENSORS"])
+ except KeyError:
+ # enabled by default
+ return 1
+
+
def make_build_dir(prefix=""):
"""Creates a temporary folder with given prefix to be used as a build dir.
Use this function instead of tempfile.mkdtemp to ensure any generated files
@@ -264,6 +284,69 @@ def calculate_signed_dot_prod_range(dt_a, dt_b, len):
return (min_prod, max_prod)
+def sanitize_quant_values(model, node_tensors, execution_context, check_values=False):
+ """ Sanitize given list of tensors in execution_context by rounding values
+ that are supposed to be integers (as indicated by their quantization
+ annotation). Will raise an assertion if the amount of rounding is too large.
+ Returns the sanitized execution context.
+
+ If check_values is specified, an extra DataType.allowed() check will be
+ performed on any rounded tensors.
+
+ Background:
+ FINN uses floating point tensors as a carrier data type to represent
+ integers. Floating point arithmetic can introduce rounding errors, e.g.
+ (int_num * float_scale) / float_scale is not always equal to int_num.
+ We use this function to ensure that the values that are supposed to be
+ integers are indeed integers.
+ """
+
+ for tensor in node_tensors:
+ dtype = model.get_tensor_datatype(tensor)
+ # floats don't need sanitization, skip to next
+ # introduces less quicker runtime
+ if dtype == DataType.FLOAT32:
+ continue
+ current_values = execution_context[tensor]
+ updated_values = current_values
+ has_to_be_rounded = False
+ # TODO: vectorize with numpy
+ for value in np.nditer(current_values):
+ if not dtype.allowed(value):
+ has_to_be_rounded = True
+ break
+ if has_to_be_rounded:
+ updated_values = np.round(current_values)
+ warnings.warn(
+ "The values of tensor {} can't be represented "
+ "with the set FINN datatype ({}), they will be rounded to match the "
+ "FINN datatype.".format(tensor, dtype)
+ )
+ # check if rounded values are not too far from original values
+ max_error = max(np.abs(current_values - updated_values).flatten())
+ if max_error <= get_execution_error_thresh():
+ if check_values is True:
+ # check again if values can now be represented with set finn datatype
+ # TODO: vectorize with numpy
+ for value in np.nditer(updated_values):
+ if not dtype.allowed(value):
+ raise Exception(
+ """Values can't be represented with set
+ finn datatype ({}) for input {}""".format(
+ dtype, tensor
+ )
+ )
+ execution_context[tensor] = updated_values
+ else:
+ raise Exception(
+ """Rounding error is too high to match set FINN
+ datatype ({}) for input {}""".format(
+ dtype, tensor
+ )
+ )
+ return execution_context
+
+
class CppBuilder:
"""Builds the g++ compiler command to produces the executable of the c++ code
in code_gen_dir which is passed to the function build() of this class."""
diff --git a/src/finn/util/vivado.py b/src/finn/util/vivado.py
index 0f82c52cb2c1fc5ee4ed5a1927f46e222e0ab9b5..6b6df3940cfeeed292345382471719c49f725de6 100644
--- a/src/finn/util/vivado.py
+++ b/src/finn/util/vivado.py
@@ -28,6 +28,7 @@
import os
import subprocess
+import stat
from finn.util.basic import get_remote_vivado
@@ -91,6 +92,7 @@ def out_of_context_synth(
vivado_proj_folder = "%s/results_%s" % (verilog_dir, top_name)
res_counts_path = vivado_proj_folder + "/res.txt"
if remote_server is not None:
+ print("Using remote Vivado OOC synth, remote server %s" % remote_server)
run_synth = """
#!/bin/bash
which vivado;
@@ -105,14 +107,17 @@ cat %s
)
with open(vivado_proj_folder + "/run.sh", "w") as f:
f.write(run_synth)
+ st = os.stat(vivado_proj_folder + "/run.sh")
+ os.chmod(vivado_proj_folder + "/run.sh", st.st_mode | stat.S_IEXEC)
# note that this assumes the same temp folder can be created on the
# remote server
- remote_server_uri = remote_server + ":" + verilog_dir
- copy_files = "rsync -avz %s %s" % (verilog_dir + "/", remote_server_uri + "/")
+ # note we set target path as / due to use of -R (relative)
+ remote_server_uri = remote_server + ":/"
+ copy_files = "rsync -avzR %s %s" % (verilog_dir + "/", remote_server_uri)
copy_files = copy_files.split()
proc = subprocess.Popen(copy_files, cwd=verilog_dir, env=os.environ)
proc.communicate()
- vivado_cmd = "bash %s/run.sh" % vivado_proj_folder
+ vivado_cmd = "bash -ic %s/run.sh" % vivado_proj_folder
run_vivado = ["ssh", "-t", remote_server, vivado_cmd]
proc = subprocess.Popen(run_vivado, cwd=verilog_dir, env=os.environ)
proc.communicate()
diff --git a/tests/brevitas/test_brevitas_avg_pool_export.py b/tests/brevitas/test_brevitas_avg_pool_export.py
new file mode 100644
index 0000000000000000000000000000000000000000..24854a2153df9af78feb8352ca119e831a9ac9eb
--- /dev/null
+++ b/tests/brevitas/test_brevitas_avg_pool_export.py
@@ -0,0 +1,103 @@
+import os
+
+import onnx # noqa
+import torch
+import numpy as np
+import brevitas.onnx as bo
+from brevitas.nn import QuantAvgPool2d
+from brevitas.quant_tensor import pack_quant_tensor
+from brevitas.core.quant import QuantType
+from finn.core.modelwrapper import ModelWrapper
+from finn.core.datatype import DataType
+from finn.transformation.infer_shapes import InferShapes
+from finn.transformation.infer_datatypes import InferDataTypes
+from finn.util.basic import gen_finn_dt_tensor
+import finn.core.onnx_exec as oxe
+
+import pytest
+
+export_onnx_path = "test_avg_pool.onnx"
+
+
+@pytest.mark.parametrize("kernel_size", [2, 3])
+@pytest.mark.parametrize("stride", [1, 2])
+@pytest.mark.parametrize("signed", [False, True])
+@pytest.mark.parametrize("bit_width", [2, 4])
+@pytest.mark.parametrize("input_bit_width", [4, 8, 32])
+@pytest.mark.parametrize("channels", [2, 4])
+@pytest.mark.parametrize("idim", [7, 8])
+def test_brevitas_avg_pool_export(
+ kernel_size, stride, signed, bit_width, input_bit_width, channels, idim
+):
+ ishape = (1, channels, idim, idim)
+ ibw_tensor = torch.Tensor([input_bit_width])
+
+ b_avgpool = QuantAvgPool2d(
+ kernel_size=kernel_size,
+ stride=stride,
+ signed=signed,
+ min_overall_bit_width=bit_width,
+ max_overall_bit_width=bit_width,
+ quant_type=QuantType.INT,
+ )
+ # call forward pass manually once to cache scale factor and bitwidth
+ input_tensor = torch.from_numpy(np.zeros(ishape)).float()
+ scale = np.ones((1, channels, 1, 1))
+ output_scale = torch.from_numpy(scale).float()
+ input_quant_tensor = pack_quant_tensor(
+ tensor=input_tensor, scale=output_scale, bit_width=ibw_tensor
+ )
+ bo.export_finn_onnx(b_avgpool, ishape, export_onnx_path, input_t=input_quant_tensor)
+ model = ModelWrapper(export_onnx_path)
+
+ # determine input FINN datatype
+ if signed is True:
+ prefix = "INT"
+ else:
+ prefix = "UINT"
+ dt_name = prefix + str(input_bit_width // 2)
+ dtype = DataType[dt_name]
+ model = model.transform(InferShapes())
+ model = model.transform(InferDataTypes())
+
+ # execution with input tensor using integers and scale = 1
+ # calculate golden output
+ inp = gen_finn_dt_tensor(dtype, ishape)
+ input_tensor = torch.from_numpy(inp).float()
+ input_quant_tensor = pack_quant_tensor(
+ tensor=input_tensor, scale=output_scale, bit_width=ibw_tensor
+ )
+ b_avgpool.eval()
+ expected = b_avgpool.forward(input_quant_tensor).tensor.detach().numpy()
+
+ # finn execution
+ idict = {model.graph.input[0].name: inp}
+ odict = oxe.execute_onnx(model, idict, True)
+ produced = odict[model.graph.output[0].name]
+ assert (expected == produced).all()
+
+ # execution with input tensor using float and scale != 1
+ scale = np.random.uniform(low=0, high=1, size=(1, channels, 1, 1)).astype(
+ np.float32
+ )
+ inp_tensor = inp * scale
+ input_tensor = torch.from_numpy(inp_tensor).float()
+ input_scale = torch.from_numpy(scale).float()
+ input_quant_tensor = pack_quant_tensor(
+ tensor=input_tensor, scale=input_scale, bit_width=ibw_tensor
+ )
+ # export again to set the scale values correctly
+ bo.export_finn_onnx(b_avgpool, ishape, export_onnx_path, input_t=input_quant_tensor)
+ model = ModelWrapper(export_onnx_path)
+ model = model.transform(InferShapes())
+ model = model.transform(InferDataTypes())
+ b_avgpool.eval()
+ expected = b_avgpool.forward(input_quant_tensor).tensor.detach().numpy()
+ # finn execution
+ idict = {model.graph.input[0].name: inp_tensor}
+ odict = oxe.execute_onnx(model, idict, True)
+ produced = odict[model.graph.output[0].name]
+
+ assert np.isclose(expected, produced).all()
+
+ os.remove(export_onnx_path)
diff --git a/tests/brevitas/test_brevitas_relu_act_export.py b/tests/brevitas/test_brevitas_relu_act_export.py
index c9d8f2d812bc7bea1a2fd2598a7711099ad421e6..c5ddad12ca3e8d353682fbb20449d44358485f69 100644
--- a/tests/brevitas/test_brevitas_relu_act_export.py
+++ b/tests/brevitas/test_brevitas_relu_act_export.py
@@ -23,6 +23,7 @@ export_onnx_path = "test_act.onnx"
def test_brevitas_act_export_relu(abits, max_val, scaling_impl_type):
min_val = -1.0
ishape = (1, 15)
+
b_act = QuantReLU(
bit_width=abits,
max_val=max_val,
@@ -67,3 +68,60 @@ scaling_impl.learned_value": torch.tensor(
assert np.isclose(produced, expected, atol=1e-3).all()
os.remove(export_onnx_path)
+
+
+@pytest.mark.parametrize("abits", [1, 2, 4, 8])
+@pytest.mark.parametrize("max_val", [1.0, 1.5, 1 - 2 ** (-7)])
+@pytest.mark.parametrize("scaling_per_channel", [True, False])
+def test_brevitas_act_export_relu_imagenet(abits, max_val, scaling_per_channel):
+ out_channels = 32
+ ishape = (1, out_channels, 1, 1)
+ min_val = -1.0
+ b_act = QuantReLU(
+ bit_width=abits,
+ quant_type=QuantType.INT,
+ scaling_impl_type=ScalingImplType.PARAMETER,
+ scaling_per_channel=scaling_per_channel,
+ restrict_scaling_type=RestrictValueType.LOG_FP,
+ scaling_min_val=2e-16,
+ max_val=6.0,
+ return_quant_tensor=True,
+ per_channel_broadcastable_shape=(1, out_channels, 1, 1),
+ )
+ if scaling_per_channel is True:
+ rand_tensor = (2) * torch.rand((1, out_channels, 1, 1))
+ else:
+ rand_tensor = torch.tensor(1.2398)
+ checkpoint = {
+ "act_quant_proxy.fused_activation_quant_proxy.tensor_quant.\
+scaling_impl.learned_value": rand_tensor.type(
+ torch.FloatTensor
+ )
+ }
+ b_act.load_state_dict(checkpoint)
+ bo.export_finn_onnx(b_act, ishape, export_onnx_path)
+ model = ModelWrapper(export_onnx_path)
+ model = model.transform(InferShapes())
+ inp_tensor = np.random.uniform(low=min_val, high=max_val, size=ishape).astype(
+ np.float32
+ )
+ idict = {model.graph.input[0].name: inp_tensor}
+ odict = oxe.execute_onnx(model, idict, True)
+ produced = odict[model.graph.output[0].name]
+ inp_tensor = torch.from_numpy(inp_tensor).float()
+ b_act.eval()
+ expected = b_act.forward(inp_tensor).tensor.detach().numpy()
+ if not np.isclose(produced, expected, atol=1e-3).all():
+ print(abits, max_val)
+ print("scale: ", b_act.quant_act_scale().type(torch.FloatTensor).detach())
+ if abits < 5:
+ print(
+ "thres:",
+ ", ".join(["{:8.4f}".format(x) for x in b_act.export_thres[0]]),
+ )
+ print("input:", ", ".join(["{:8.4f}".format(x) for x in inp_tensor[0]]))
+ print("prod :", ", ".join(["{:8.4f}".format(x) for x in produced[0]]))
+ print("expec:", ", ".join(["{:8.4f}".format(x) for x in expected[0]]))
+
+ assert np.isclose(produced, expected, atol=1e-3).all()
+ os.remove(export_onnx_path)
diff --git a/tests/core/test_basic_onnx_exec.py b/tests/core/test_basic_onnx_exec.py
index a7b6da9965aa5912870812a8c1f8d6da2ee0d181..7b0412432cc6360cb9c42d66417bd187ed142563 100644
--- a/tests/core/test_basic_onnx_exec.py
+++ b/tests/core/test_basic_onnx_exec.py
@@ -35,6 +35,8 @@ import onnx.numpy_helper as np_helper
import finn.core.onnx_exec as oxe
from finn.core.modelwrapper import ModelWrapper
from finn.transformation.infer_shapes import InferShapes
+from finn.core.datatype import DataType
+from finn.util.basic import gen_finn_dt_tensor
def test_mnist_onnx_download_extract_run():
@@ -53,3 +55,30 @@ def test_mnist_onnx_download_extract_run():
assert np.isclose(
np_helper.to_array(output_tensor), output_dict["Plus214_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"],)
+ graph = onnx.helper.make_graph(
+ nodes=[mul_node], name="mul_graph", inputs=[inp0, inp1], outputs=[outp]
+ )
+
+ model = onnx.helper.make_model(graph, producer_name="mul-model")
+ model = ModelWrapper(model)
+ idt = DataType.INT2
+ model.set_tensor_datatype("inp0", idt)
+ model.set_tensor_datatype("inp1", idt)
+ model.transform(InferShapes())
+
+ mul_value = np.asarray([-1], dtype=np.float32)
+ inp_int = gen_finn_dt_tensor(idt, [2, 2])
+ scale = np.random.uniform(low=0, high=1, size=(2, 2)).astype(np.float32)
+ inp_rounded = (inp_int * scale) / (scale + 1e-7)
+ input_dict = {"inp0": inp_rounded, "inp1": mul_value}
+ output_dict = oxe.execute_onnx(model, input_dict)
+ produced = output_dict["outp"]
+ expected = np.multiply(inp_int, mul_value)
+ assert (produced == expected).all()
diff --git a/tests/fpgadataflow/test_fpgadataflow_fmpadding.py b/tests/fpgadataflow/test_fpgadataflow_fmpadding.py
index 9d6390b2673e5d2c0e72748183ac04ed222d078e..5ff3da87228a2a32a41226bb46e0b16b1a44df50 100644
--- a/tests/fpgadataflow/test_fpgadataflow_fmpadding.py
+++ b/tests/fpgadataflow/test_fpgadataflow_fmpadding.py
@@ -23,7 +23,7 @@ test_fpga_part = pynq_part_map[test_pynq_board]
target_clk_ns = 10
-def make_single_fmpadding_modelwrapper(idim, padding, num_ch, idt, pad_style):
+def make_single_fmpadding_modelwrapper(idim, padding, num_ch, simd, idt, pad_style):
assert pad_style == 2, "only pad_style == 2 supported in hlslib"
assert padding > 0, "Output dim should be greater than input dim"
odim = idim + padding
@@ -47,6 +47,7 @@ def make_single_fmpadding_modelwrapper(idim, padding, num_ch, idt, pad_style):
inputDataType=str(idt.name),
PaddingStyle=pad_style,
numInputVectors=1,
+ SIMD=simd,
)
graph = helper.make_graph(
@@ -63,11 +64,13 @@ def make_single_fmpadding_modelwrapper(idim, padding, num_ch, idt, pad_style):
# input image dimension
-@pytest.mark.parametrize("idim", [8, 16])
+@pytest.mark.parametrize("idim", [8])
# number of rows and number of cols to add
@pytest.mark.parametrize("pad", [2, 3])
# number of channels
-@pytest.mark.parametrize("num_ch", [1, 2])
+@pytest.mark.parametrize("num_ch", [2, 4])
+# Input parallelism
+@pytest.mark.parametrize("simd", [1, 2])
# PaddingStyle: selects behavior when (odim-idim)%2 != 0
@pytest.mark.parametrize("pad_style", [2])
# FINN input datatype
@@ -76,14 +79,15 @@ def make_single_fmpadding_modelwrapper(idim, padding, num_ch, idt, pad_style):
@pytest.mark.parametrize("mode", ["cppsim", "rtlsim"])
@pytest.mark.slow
@pytest.mark.vivado
-def test_fpgadataflow_fmpadding(idim, pad, num_ch, pad_style, idt, mode):
-
+def test_fpgadataflow_fmpadding(idim, pad, num_ch, simd, pad_style, idt, mode):
+ if num_ch % simd != 0:
+ pytest.skip(" num_ch % simd != 0, skipping")
# generate input data
x = gen_finn_dt_tensor(idt, [1, idim, idim, num_ch])
input_dict = {"inp": x}
odim = idim + pad
- model = make_single_fmpadding_modelwrapper(idim, pad, num_ch, idt, pad_style)
+ model = make_single_fmpadding_modelwrapper(idim, pad, num_ch, simd, idt, pad_style)
model = model.transform(InferShapes())
model = model.transform(SetExecMode(mode))
model = model.transform(GiveUniqueNodeNames())
diff --git a/tests/fpgadataflow/test_fpgadataflow_ip_stitch.py b/tests/fpgadataflow/test_fpgadataflow_ipstitch.py
similarity index 99%
rename from tests/fpgadataflow/test_fpgadataflow_ip_stitch.py
rename to tests/fpgadataflow/test_fpgadataflow_ipstitch.py
index 61dd81b728aafcd8ccc812cf0cb4c27eff00f471..b830693c32afe629dd6fc70868d0bddacac4c887 100644
--- a/tests/fpgadataflow/test_fpgadataflow_ip_stitch.py
+++ b/tests/fpgadataflow/test_fpgadataflow_ipstitch.py
@@ -300,7 +300,7 @@ def test_fpgadataflow_ipstitch_synth_ooc():
assert ret["FF"] > 0
assert ret["DSP"] == 0
assert ret["BRAM"] == 0
- assert ret["fmax_mz"] > 100
+ assert ret["fmax_mhz"] > 100
@pytest.mark.vivado
diff --git a/tests/transformation/test_conv_lowering.py b/tests/transformation/test_conv_lowering.py
index 2cbc8e558940517168678b05c3bb46af8170abce..73891ded1b9691c7c48a2075ad6ca4668fcf6bfe 100644
--- a/tests/transformation/test_conv_lowering.py
+++ b/tests/transformation/test_conv_lowering.py
@@ -26,12 +26,13 @@
# 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 onnx.helper as oh
+from onnx import TensorProto
import os
import pkg_resources as pk
import brevitas.onnx as bo
import numpy as np
-
from finn.core.modelwrapper import ModelWrapper
from finn.transformation.fold_constants import FoldConstants
from finn.transformation.infer_shapes import InferShapes
@@ -65,3 +66,51 @@ def test_conv_lowering_cnv_w1a1():
assert np.isclose(produced, expected).all()
assert np.argmax(produced) == 3
os.remove(export_onnx_path)
+
+
+def test_conv_lowering_conv_1x1():
+ np.random.seed(0)
+
+ in_feature_dim = 7
+ in_chn = 3
+ kernel_size = 1
+ out_feature_dim = in_feature_dim
+
+ input_shape = [1, in_chn, in_feature_dim, in_feature_dim]
+ output_shape = [1, in_chn, out_feature_dim, out_feature_dim]
+
+ conv_param_shape = [in_chn, in_chn, kernel_size, kernel_size]
+
+ conv_config = {}
+ conv_config["dilations"] = [1, 1]
+ conv_config["group"] = 1
+ conv_config["kernel_shape"] = [kernel_size, kernel_size]
+ conv_config["pads"] = [0, 0, 0, 0]
+ conv_config["strides"] = [1, 1]
+
+ top_in = oh.make_tensor_value_info("top_in", TensorProto.FLOAT, input_shape)
+ top_out = oh.make_tensor_value_info("top_out", TensorProto.FLOAT, output_shape)
+
+ value_info = [oh.make_tensor_value_info("p1", TensorProto.FLOAT, conv_param_shape)]
+
+ modelproto = oh.make_model(
+ oh.make_graph(
+ name="test",
+ inputs=[top_in],
+ outputs=[top_out],
+ value_info=value_info,
+ nodes=[oh.make_node("Conv", ["top_in", "p1"], ["top_out"], **conv_config)],
+ )
+ )
+ model = ModelWrapper(modelproto)
+ model = model.transform(InferShapes())
+ model.set_initializer("p1", np.random.rand(*conv_param_shape).astype(np.float32))
+
+ new_model = model.transform(LowerConvsToMatMul())
+ inp_dict = {"top_in": np.random.rand(*input_shape).astype(np.float32)}
+
+ assert oxe.compare_execution(model, new_model, inp_dict)
+ assert new_model.graph.node[0].op_type == "Transpose"
+ assert new_model.graph.node[1].op_type == "MatMul"
+ assert new_model.graph.node[2].op_type == "Transpose"
+ assert len(new_model.graph.node) == 3
diff --git a/tests/transformation/test_move_maxpool_past_multithreshold.py b/tests/transformation/test_move_maxpool_past_multithreshold.py
new file mode 100644
index 0000000000000000000000000000000000000000..2fc19debf8d6fc89d15e3d731f1e54daa491c321
--- /dev/null
+++ b/tests/transformation/test_move_maxpool_past_multithreshold.py
@@ -0,0 +1,100 @@
+from onnx import TensorProto, helper
+import numpy as np
+
+import finn.core.onnx_exec as oxe
+from finn.core.modelwrapper import ModelWrapper
+from finn.transformation.streamline.reorder import MoveMaxPoolPastMultiThreshold
+from finn.transformation.infer_shapes import InferShapes
+from finn.transformation.infer_datatypes import InferDataTypes
+
+
+def get_multithreshold_rand_params(channels, num_of_thres, seed=None):
+ if seed is not None:
+ np.random.seed(seed)
+ steps = np.random.rand(channels, 1) * 2
+ bias = np.random.rand(channels, 1) * 10
+ thres = [np.arange(num_of_thres) for chn in range(channels)]
+ thres = ((thres - bias) * steps).astype(np.float32)
+ return thres
+
+
+def test_move_maxpool_past_multithreshold():
+ # generate test vectors of correct shape
+ ch = 64
+ ifmdim = 16
+ ofmdim = 16 // 4
+ input_shape = (1, ch, ifmdim, ifmdim)
+ output_shape = (1, ch, ofmdim, ofmdim)
+
+ 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)
+
+ maxpool_config = {}
+ maxpool_config["pads"] = [1, 1, 1, 1]
+ maxpool_config["kernel_shape"] = [3, 3]
+ maxpool_config["strides"] = [2, 2]
+
+ value_info = []
+ thres1_shape = [1, 1]
+ value_info += [
+ helper.make_tensor_value_info("thres1", TensorProto.FLOAT, thres1_shape)
+ ]
+
+ thres2_shape = [ch, 14]
+ value_info += [
+ helper.make_tensor_value_info("thres2", TensorProto.FLOAT, thres2_shape)
+ ]
+
+ nodes = []
+ nodes += [helper.make_node("MaxPool", ["top_in"], ["t1"], **maxpool_config)]
+ nodes += [
+ helper.make_node(
+ "MultiThreshold",
+ ["t1", "thres1"],
+ ["t2"],
+ domain="finn",
+ out_dtype="BIPOLAR",
+ out_bias=-1.0,
+ out_scale=1.0,
+ )
+ ]
+ nodes += [helper.make_node("MaxPool", ["t2"], ["t3"], **maxpool_config)]
+ nodes += [
+ helper.make_node(
+ "MultiThreshold",
+ ["t3", "thres2"],
+ ["top_out"],
+ domain="finn",
+ out_dtype="UINT4",
+ )
+ ]
+
+ modelproto = helper.make_model(
+ helper.make_graph(
+ name="test",
+ inputs=[top_in],
+ outputs=[top_out],
+ value_info=value_info,
+ nodes=nodes,
+ )
+ )
+ model = ModelWrapper(modelproto)
+ model = model.transform(InferShapes())
+ model = model.transform(InferDataTypes())
+
+ model.set_initializer("thres1", np.array([[0]]))
+ model.set_initializer(
+ "thres2", get_multithreshold_rand_params(*thres2_shape, seed=0)
+ )
+
+ # Transform
+ new_model = model.transform(MoveMaxPoolPastMultiThreshold())
+ inp_dict = {"top_in": np.random.rand(*input_shape).astype(np.float32)}
+
+ # Test
+ assert oxe.compare_execution(model, new_model, inp_dict)
+ assert new_model.graph.node[0].op_type == "MaxPool"
+ assert new_model.graph.node[1].op_type == "MultiThreshold"
+ assert new_model.graph.node[2].op_type == "MultiThreshold"
+ assert new_model.graph.node[3].op_type == "MaxPool"
+ assert len(new_model.graph.node) == 4
diff --git a/tests/transformation/test_remove_identity_ops.py b/tests/transformation/test_remove_identity_ops.py
new file mode 100644
index 0000000000000000000000000000000000000000..536c1ab0b48fa44388da23f45b528da3c5f3b2f2
--- /dev/null
+++ b/tests/transformation/test_remove_identity_ops.py
@@ -0,0 +1,81 @@
+import pytest
+
+import numpy as np
+from onnx import helper, TensorProto
+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.transformation.streamline.remove import RemoveIdentityOps
+from finn.util.basic import gen_finn_dt_tensor
+
+
+def insert_identity_op(model, op):
+ if op in ["Add", "Sub"]:
+ val = np.asarray([0.0], dtype=np.float32)
+ elif op in ["Mul", "Div"]:
+ val = np.asarray([1.0], dtype=np.float32)
+ else:
+ return
+
+ identity_node = helper.make_node(op, ["div_out", "value"], ["ident_out"])
+ graph = model.graph
+ graph.node.insert(3, identity_node)
+ graph.node[-1].input[0] = "ident_out"
+ model.set_initializer("value", val)
+
+ return model
+
+
+# identity operations to be inserted
+@pytest.mark.parametrize("op", ["Add", "Sub", "Mul", "Div"])
+def test_remove_identity_ops(op):
+
+ # set up onnx model
+ inp = helper.make_tensor_value_info("inp", TensorProto.FLOAT, [1, 4, 1, 1])
+ mul = helper.make_tensor_value_info("mul", TensorProto.FLOAT, [])
+ shape = helper.make_tensor_value_info("shape", TensorProto.FLOAT, [2])
+ div = helper.make_tensor_value_info("div", TensorProto.FLOAT, [])
+ matmul = helper.make_tensor_value_info("matmul", TensorProto.FLOAT, [4, 2])
+ outp = helper.make_tensor_value_info("outp", TensorProto.FLOAT, [1, 2])
+
+ mul_node = helper.make_node("Mul", ["inp", "mul"], ["mul_out"])
+ reshape_node = helper.make_node("Reshape", ["mul_out", "shape"], ["reshape_out"])
+ div_node = helper.make_node("Div", ["reshape_out", "div"], ["div_out"])
+ matmul_node = helper.make_node("MatMul", ["div_out", "matmul"], ["outp"])
+
+ graph = helper.make_graph(
+ nodes=[mul_node, reshape_node, div_node, matmul_node],
+ name="identity-graph",
+ inputs=[inp],
+ outputs=[outp],
+ value_info=[mul, shape, div, matmul],
+ )
+
+ model = helper.make_model(graph, producer_name="mulpastconv-model")
+ model = ModelWrapper(model)
+ inp_values = gen_finn_dt_tensor(DataType.INT2, [1, 4, 1, 1])
+ mul_values = np.random.uniform(low=0.1, high=0.99, size=(1)).astype(np.float32)
+ shape_values = np.asarray([1, -1], dtype=np.int64)
+ div_values = np.random.uniform(low=0.1, high=0.99, size=(1)).astype(np.float32)
+ matmul_values = gen_finn_dt_tensor(DataType.INT2, [4, 2])
+ model.set_initializer("mul", mul_values)
+ model.set_initializer("shape", shape_values)
+ model.set_initializer("div", div_values)
+ model.set_initializer("matmul", matmul_values)
+ insert_identity_op(model, op)
+ model = model.transform(InferShapes())
+ model = model.transform(InferDataTypes())
+ idict = {"inp": inp_values}
+ odict = oxe.execute_onnx(model, idict)
+ out_before = odict["outp"]
+ num_of_nodes_before = len(model.graph.node)
+
+ model = model.transform(RemoveIdentityOps())
+ num_of_nodes_after = len(model.graph.node)
+ assert num_of_nodes_before - 1 == num_of_nodes_after
+
+ odict = oxe.execute_onnx(model, idict)
+ out_after = odict["outp"]
+ assert (out_before == out_after).all()