From ae97e38e9304e017cdd90ee99fd5dd01169ab90e Mon Sep 17 00:00:00 2001
From: mmrahorovic <mmrahorovic@hotmail.com>
Date: Thu, 28 Mar 2024 14:31:27 +0000
Subject: [PATCH] [hls vvau]: added lut and dsp estimation methods

---
 .../hls/vectorvectoractivation_hls.py         | 79 +++++++++++++++++++
 1 file changed, 79 insertions(+)

diff --git a/src/finn/custom_op/fpgadataflow/hls/vectorvectoractivation_hls.py b/src/finn/custom_op/fpgadataflow/hls/vectorvectoractivation_hls.py
index fbae9eb9b8..3e10b640c5 100644
--- a/src/finn/custom_op/fpgadataflow/hls/vectorvectoractivation_hls.py
+++ b/src/finn/custom_op/fpgadataflow/hls/vectorvectoractivation_hls.py
@@ -26,6 +26,7 @@
 # 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 math
 import numpy as np
 import os
 from qonnx.core.datatype import DataType
@@ -47,6 +48,84 @@ def get_nodeattr_types(self):
         my_attrs.update(HLSBackend.get_nodeattr_types(self))
         return my_attrs
 
+    def lut_estimation(self):
+        """Calculates resource estimations for LUTs based on:
+        - FINN-R: An End-to-End Deep-Learning Framework for Fast
+        Exploration of Quantized Neural Networks
+        - M. Blott, T. B. Preusser, N. J. Fraser, G. Gambardella, K. O'Brien,
+        Y. Umuroglu, M. Leeser and K. Vissers
+        - 12. Sep 2018
+        """
+        # TODO add in/out FIFO contributions
+        P = self.get_nodeattr("PE")
+        Q = self.get_nodeattr("SIMD")
+        wdt = self.get_weight_datatype()
+        W = wdt.bitwidth()
+        # determine tdt with input and weight data types
+        idt = self.get_input_datatype()
+        A = idt.bitwidth()
+        # parameters from experiments in paper mentioned above
+        c0 = 300
+        c1 = 1.1
+        c2 = 0
+        mmode = self.get_nodeattr("mem_mode")
+        mstyle = self.get_nodeattr("ram_style")
+        if (mmode == "internal_decoupled" and mstyle == "distributed") or (
+            mmode == "internal_embedded" and self.calc_wmem() <= 128
+        ):
+            c2 = (P * Q * W) * math.ceil(self.calc_wmem() / 64)
+
+        # multiplication
+        res_type = self.get_nodeattr("resType")
+        if res_type == "dsp":
+            mult_luts = 0
+        else:
+            mult_luts = Q * (2 * math.ceil((W + A) / 6) - 1) * (W + A)
+        # adder tree
+        addertree_luts = (W + A) * (2 * Q - 1)
+        # accumulator
+        acc_datatype = self.get_accumulator_datatype()
+        acc_bits = acc_datatype.bitwidth()
+        k_h, k_w = self.get_nodeattr("Kernel")
+        # if accDataType is not set, then it will default to INT32, which would
+        # be a large overestimate in most (if not all) cases. In this scenario,
+        # we would use the minimum accumulator as determined by the data types
+        # bound, derived in https://arxiv.org/abs/2301.13376
+        alpha = math.log(k_h * k_w, 2) + W + A - 1 - int(idt.signed())
+        acc_bits = min(
+            acc_datatype.bitwidth(),
+            np.ceil(alpha + math.log(1 + pow(2, -alpha), 2) + 1),
+        )
+        acc_luts = acc_bits
+        # thresholds and threshold comparators
+        thr_luts = 0
+        comp_luts = 0
+        noact = self.get_nodeattr("noActivation")
+        # TODO - add 'ram_style_threshold' node attribute
+        if noact == 0:
+            odt = self.get_output_datatype()
+            B = odt.bitwidth()
+            thr_luts = (2**B - 1) * acc_bits * self.calc_tmem() / 64
+            comp_luts = (2**B - 1) * acc_bits
+
+        return int(
+            c0 + c1 * (P * (mult_luts + addertree_luts + acc_luts + thr_luts + comp_luts)) + c2
+        )
+
+    def dsp_estimation(self):
+        # multiplication
+        P = self.get_nodeattr("PE")
+        res_type = self.get_nodeattr("resType")
+        wdt = self.get_weight_datatype()
+        W = wdt.bitwidth()
+        idt = self.get_input_datatype()
+        A = idt.bitwidth()
+        if res_type == "dsp":
+            mult_dsp = P * np.ceil((W + A) / 48)  # TODO: more accurate modelling
+        else:
+            mult_dsp = 0
+        return int(mult_dsp)
+
     def execute_node(self, context, graph):
         mode = self.get_nodeattr("exec_mode")
         mem_mode = self.get_nodeattr("mem_mode")