From 1500107c4e3b6e405f02d25ea5c40c191df48359 Mon Sep 17 00:00:00 2001
From: Yaman Umuroglu <yamanu@xilinx.com>
Date: Wed, 13 Oct 2021 19:47:17 +0200
Subject: [PATCH] [CustomOp] sketch CustomOp for lookup
---
src/finn/custom_op/fpgadataflow/lookup.py | 294 ++++++++++++++++++++++
1 file changed, 294 insertions(+)
create mode 100644 src/finn/custom_op/fpgadataflow/lookup.py
diff --git a/src/finn/custom_op/fpgadataflow/lookup.py b/src/finn/custom_op/fpgadataflow/lookup.py
new file mode 100644
index 000000000..e7a957169
--- /dev/null
+++ b/src/finn/custom_op/fpgadataflow/lookup.py
@@ -0,0 +1,294 @@
+# Copyright (c) 2020, Xilinx
+# All rights reserved.
+#
+# Redistribution and use in source and binary forms, with or without
+# modification, are permitted provided that the following conditions are met:
+#
+# * Redistributions of source code must retain the above copyright notice, this
+# list of conditions and the following disclaimer.
+#
+# * Redistributions in binary form must reproduce the above copyright notice,
+# this list of conditions and the following disclaimer in the documentation
+# and/or other materials provided with the distribution.
+#
+# * Neither the name of FINN nor the names of its
+# contributors may be used to endorse or promote products derived from
+# this software without specific prior written permission.
+#
+# THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS "AS IS"
+# AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
+# IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE ARE
+# DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT HOLDER OR CONTRIBUTORS BE LIABLE
+# FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL
+# DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR
+# SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER
+# CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY,
+# OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE
+# OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
+
+import numpy as np
+import os
+import warnings
+
+from finn.core.datatype import DataType
+from finn.custom_op.fpgadataflow.hlscustomop import HLSCustomOp
+from finn.util.data_packing import npy_to_rtlsim_input, rtlsim_output_to_npy
+
+
+class Lookup(HLSCustomOp):
+ "Streaming elementwise HLS lookup, mapping indices to values."
+
+ def __init__(self, onnx_node):
+ super().__init__(onnx_node)
+
+ def get_nodeattr_types(self):
+ my_attrs = {
+ # Number of embeddings ("memory depth")
+ "NumEmbeddings": ("i", True, 0),
+ # Dimensionality of each embedding (part of "memory width")
+ "EmbeddingDim": ("i", True, 0),
+ # Datatype for embeddings (part of "memory width")
+ "EmbeddingType": ("s", True, ""),
+ # Datatype for inputs
+ "InputType": ("s", True, ""),
+ # Input shape
+ "InputShape": ("ints", False, [1]),
+ }
+ my_attrs.update(super().get_nodeattr_types())
+ return my_attrs
+
+ def get_exp_cycles(self):
+ n_inputs = np.prod(self.get_nodeattr("InputShape"))
+ exp_cycles = int(n_inputs)
+ return exp_cycles
+
+ def get_normal_input_shape(self):
+ return self.get_nodeattr("InputShape")
+
+ def get_normal_output_shape(self):
+ ishape = self.get_normal_input_shape()
+ oshape = list(ishape) + [self.get_nodeattr("EmbeddingDim")]
+ return tuple(oshape)
+
+ def get_folded_input_shape(self):
+ ishape = self.get_normal_input_shape()
+ folded_ishape = list(ishape) + [1]
+ return tuple(folded_ishape)
+
+ def get_folded_output_shape(self):
+ return self.get_normal_output_shape()
+
+ def make_shape_compatible_op(self, model):
+ exp_ishape = self.get_normal_input_shape()
+ oshape = self.get_normal_output_shape()
+ ishape = tuple(model.get_tensor_shape(self.onnx_node.input[0]))
+ assert ishape == exp_ishape, "Unexpect input shape for Lookup: %s vs %s" % (
+ str(exp_ishape),
+ str(ishape),
+ )
+ return super().make_const_shape_op(oshape)
+
+ def infer_node_datatype(self, model):
+ node = self.onnx_node
+ idt = model.get_tensor_datatype(node.input[0])
+ if idt != self.get_input_datatype():
+ warn_str = "InputType changing for %s: %s -> %s " % (
+ node.name,
+ str(self.get_input_datatype()),
+ str(idt),
+ )
+ warnings.warn(warn_str)
+ self.set_nodeattr("InputType", idt.name)
+ odt = DataType[self.get_nodeattr("EmbeddingType")]
+ model.set_tensor_datatype(node.output[0], odt)
+
+ def verify_node(self):
+ pass
+
+ def get_input_datatype(self):
+ ret = DataType[self.get_nodeattr("InputType")]
+ return ret
+
+ def get_output_datatype(self):
+ ret = DataType[self.get_nodeattr("EmbeddingType")]
+ return ret
+
+ def get_instream_width(self):
+ ibits = self.get_input_datatype().bitwidth()
+ return ibits
+
+ def get_outstream_width(self):
+ obits = self.get_output_datatype().bitwidth()
+ ofm_ch = self.get_nodeattr("EmbeddingDim")
+ return obits * ofm_ch
+
+ def get_number_output_values(self):
+ folded_oshape = self.get_folded_output_shape()
+ return np.prod(folded_oshape[:-1])
+
+ def global_includes(self):
+ self.code_gen_dict["$GLOBALS$"] = ['#include "lookup.hpp"']
+
+ def defines(self, var):
+ n_inputs = np.prod(self.get_folded_input_shape()[:-1])
+ dtype = self.get_input_datatype()
+ elem_hls_type = dtype.get_hls_datatype_str()
+ emb_type = DataType[self.get_nodeattr("EmbeddingType")]
+ emb_hls_type = emb_type.get_hls_datatype_str()
+ my_defines = []
+ my_defines.append(
+ "#define NumEmbeddings %d" % self.get_nodeattr("NumEmbeddings")
+ )
+ my_defines.append("#define EmbeddingDim %d" % self.get_nodeattr("EmbeddingDim"))
+ my_defines.append("#define NumInputs %d" % n_inputs)
+ my_defines.append("#define InputType %s" % elem_hls_type)
+ my_defines.append("#define EmbeddingType %s" % emb_hls_type)
+ self.code_gen_dict["$DEFINES$"] = my_defines
+
+ def read_npy_data(self):
+ code_gen_dir = self.get_nodeattr("code_gen_dir_cppsim")
+ dtype = self.get_input_datatype()
+ if dtype == DataType["BIPOLAR"]:
+ # use binary for bipolar storage
+ dtype = DataType["BINARY"]
+ elem_bits = dtype.bitwidth()
+ packed_bits = self.get_instream_width()
+ packed_hls_type = "ap_uint<%d>" % packed_bits
+ elem_hls_type = dtype.get_hls_datatype_str()
+ npy_type = "float"
+ npy_in = "%s/input_0.npy" % code_gen_dir
+ self.code_gen_dict["$READNPYDATA$"] = []
+ self.code_gen_dict["$READNPYDATA$"].append(
+ 'npy2apintstream<%s, %s, %d, %s>("%s", in0);'
+ % (packed_hls_type, elem_hls_type, elem_bits, npy_type, npy_in)
+ )
+
+ def dataoutstrm(self):
+ code_gen_dir = self.get_nodeattr("code_gen_dir_cppsim")
+ dtype = self.get_output_datatype()
+ if dtype == DataType["BIPOLAR"]:
+ # use binary for bipolar storage
+ dtype = DataType["BINARY"]
+ elem_bits = dtype.bitwidth()
+ packed_bits = self.get_outstream_width()
+ packed_hls_type = "ap_uint<%d>" % packed_bits
+ elem_hls_type = dtype.get_hls_datatype_str()
+ npy_type = "float"
+ npy_out = "%s/output.npy" % code_gen_dir
+ oshape = self.get_folded_output_shape()
+ oshape_cpp_str = str(oshape).replace("(", "{").replace(")", "}")
+
+ self.code_gen_dict["$DATAOUTSTREAM$"] = [
+ 'apintstream2npy<%s, %s, %d, %s>(out, %s, "%s");'
+ % (
+ packed_hls_type,
+ elem_hls_type,
+ elem_bits,
+ npy_type,
+ oshape_cpp_str,
+ npy_out,
+ )
+ ]
+
+ def save_as_npy(self):
+ self.code_gen_dict["$SAVEASCNPY$"] = []
+
+ def strm_decl(self):
+ self.code_gen_dict["$STREAMDECLARATIONS$"] = []
+ self.code_gen_dict["$STREAMDECLARATIONS$"].append(
+ 'hls::stream<ap_uint<{}>> in0 ("in0");'.format(self.get_instream_width())
+ )
+ self.code_gen_dict["$STREAMDECLARATIONS$"].append(
+ 'hls::stream<ap_uint<{}>> out ("out");'.format(self.get_outstream_width())
+ )
+
+ def docompute(self):
+ self.code_gen_dict["$DOCOMPUTE$"] = [
+ """StreamingLookup<NumEmbeddings, EmbeddingDim, NumInputs,
+ InputType, EmbeddingType >(in0, out, embeddings);"""
+ ]
+
+ def blackboxfunction(self):
+ ibits = self.get_instream_width()
+ packed_input_hls_type = "ap_uint<%d>" % ibits
+ obits = self.get_outstream_width()
+ packed_output_hls_type = "ap_uint<%d>" % obits
+ self.code_gen_dict["$BLACKBOXFUNCTION$"] = [
+ "void %s(hls::stream<%s > &in0, hls::stream<%s > &out)"
+ % (self.onnx_node.name, packed_input_hls_type, packed_output_hls_type)
+ ]
+
+ def pragmas(self):
+ my_pragmas = ["#pragma HLS INTERFACE axis port=in0"]
+ my_pragmas.append("#pragma HLS INTERFACE axis port=out")
+ my_pragmas.append("#pragma HLS INTERFACE ap_ctrl_none port=return")
+ self.code_gen_dict["$PRAGMAS$"] = my_pragmas
+
+ def execute_node(self, context, graph):
+ mode = self.get_nodeattr("exec_mode")
+ node = self.onnx_node
+ exp_ishape = self.get_normal_input_shape()
+ exp_oshape = self.get_normal_output_shape()
+ folded_ishape = self.get_folded_input_shape()
+ folded_oshape = self.get_folded_output_shape()
+
+ if mode == "cppsim":
+ code_gen_dir = self.get_nodeattr("code_gen_dir_cppsim")
+ elif mode == "rtlsim":
+ code_gen_dir = self.get_nodeattr("code_gen_dir_ipgen")
+ else:
+ raise Exception(
+ """Invalid value for attribute exec_mode! Is currently set to: {}
+ has to be set to one of the following value ("cppsim", "rtlsim")""".format(
+ mode
+ )
+ )
+
+ inp = context[node.input[0]]
+ assert str(inp.dtype) == "float32", "Input datatype is not float32"
+ assert inp.shape == exp_ishape, """Input shape doesn't match expected shape."""
+ 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."""
--
GitLab