diff --git a/docker/finn_entrypoint.sh b/docker/finn_entrypoint.sh
index 0074cce02f7de57dc778e0b671c484233df72a8a..80e5261c9b6568e0f340ae0add86d269c036ff16 100644
--- a/docker/finn_entrypoint.sh
+++ b/docker/finn_entrypoint.sh
@@ -15,7 +15,7 @@ gecho () {
# the repos themselves are cloned in the Dockerfile
BREVITAS_COMMIT=026a509186b7e7b0b65d46a2f905043d41069306
CNPY_COMMIT=4e8810b1a8637695171ed346ce68f6984e585ef4
-HLSLIB_COMMIT=13e9b0772a27a3a1efc40c878d8e78ed09efb716
+HLSLIB_COMMIT=8aed899c278c36c977a249558d71795086cf852c
PYVERILATOR_COMMIT=c97a5ba41bbc7c419d6f25c74cdf3bdc3393174f
PYNQSHELL_COMMIT=0c82a61b0ec1a07fa275a14146233824ded7a13d
OMX_COMMIT=1bae737669901e762f581af73348332b5c4b2ada
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/pool_batch.py b/src/finn/custom_op/fpgadataflow/pool_batch.py
new file mode 100644
index 0000000000000000000000000000000000000000..1aedf7056b1381a8904ac63b3a64f86b1243cb36
--- /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/SIMD, K, K, SIMD)
+
+ 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 = model.get_tensor_datatype(node.input[0])
+ 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/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 2dae826cf9712bef17d0053a0878c41ef51fec36..0060e5d400f30055d532671c8cf1680f0668442a 100644
--- a/src/finn/custom_op/registry.py
+++ b/src/finn/custom_op/registry.py
@@ -44,7 +44,8 @@ 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.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
@@ -66,6 +67,7 @@ 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
diff --git a/src/finn/transformation/fpgadataflow/convert_to_hls_layers.py b/src/finn/transformation/fpgadataflow/convert_to_hls_layers.py
index d421a5f3ef8ca980b399087de1482b2ae913da1b..b70b126680d650547cf376dd601c048c73a1cfd4 100644
--- a/src/finn/transformation/fpgadataflow/convert_to_hls_layers.py
+++ b/src/finn/transformation/fpgadataflow/convert_to_hls_layers.py
@@ -34,6 +34,7 @@ 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.basic import get_by_name
class InferConvInpGen(Transformation):
@@ -56,6 +57,7 @@ class InferConvInpGen(Transformation):
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 +69,11 @@ 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
@@ -112,6 +118,7 @@ class InferConvInpGen(Transformation):
Stride=stride,
inputDataType=dt.name,
outputDataType=dt.name,
+ depthwise=depthwise,
)
graph.node.insert(ConvInpGen_node_idx, ConvInpGen_node)
# remove old nodes
@@ -169,6 +176,137 @@ 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
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/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_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())