Merge pull request #3 from makslevental/max_op

Max op

Author: makslevental

bragghls/compiler/compile.py

@@ -203,7 +203,7 @@ def compile(
             vals,
             csts,
             pe_idxs,
-            for_testbench=do_testbench
+            for_testbench=do_testbench,
         )
         verilog_file = verilog_file.replace("%", "p_")
         with open(f"{artifacts_dir}/{name}.sv", "w") as f:
@@ -224,9 +224,12 @@ def compile(
     logger.info(f"RTL top-level {name}")
 
     for ip_core_sv in [
-        f"flopoco_fadd_{width_exponent}_{width_fraction}.sv",
-        f"flopoco_fmul_{width_exponent}_{width_fraction}.sv",
-        f"flopoco_fdiv_{width_exponent}_{width_fraction}.sv",
+        f"fadd/flopoco_fadd_{width_exponent}_{width_fraction}.sv",
+        f"fsub/flopoco_fsub_{width_exponent}_{width_fraction}.sv",
+        f"fmul/flopoco_fmul_{width_exponent}_{width_fraction}.sv",
+        f"fdiv/flopoco_fdiv_{width_exponent}_{width_fraction}.sv",
+        f"fcmplt/flopoco_fcmplt_{width_exponent}_{width_fraction}.sv",
+        f"flopoco_max.sv",
         f"flopoco_neg.sv",
         f"flopoco_relu.sv",
         f"alveo-u280-xdc.xdc",
@@ -251,7 +254,6 @@ def compile(
             },
             width_exponent=width_exponent,
             width_fraction=width_fraction,
-            ip_cores_path=os.path.dirname(ip_cores.__file__),
             n_test_vectors=n_test_vectors,
             threshold=threshold,
         )

bragghls/compiler/state.py

@@ -3,6 +3,7 @@
 import networkx as nx
 
 from bragghls.config import VAL_PREFIX, DTYPE, DEBUG, INCLUDE_AUX_DEPS
+from bragghls.util import extend_idx
 
 logging.basicConfig(encoding="utf-8", level=logging.INFO)
 logger = logging.getLogger(__name__)
@@ -98,7 +99,8 @@ def update_current_pe_idx(self, *, pe_idx=None, val=None):
             src = self.get_arg_src(val)
             if isinstance(src, str):
                 assert src in {INPUT_ARG, MEMREF_ARG, GLOBAL_MEMREF_ARG, CONSTANT}
-                return
+                if src in {MEMREF_ARG, GLOBAL_MEMREF_ARG}:
+                    self.pe_idx = extend_idx(tuple(map(int, val.id.split("_"))))
             else:
                 self.pe_idx = src.pe_idx
         else:

bragghls/config.py

@@ -33,38 +33,43 @@
     pipeline_depth_re = re.compile(r"Pipeline depth: (\d+) cycles")
     with open(
         Path(ip_cores.__file__).parent
+        / "fmul"
         / f"flopoco_fmul_{WIDTH_EXPONENT}_{WIDTH_FRACTION}.vhdl"
     ) as f:
         depths = pipeline_depth_re.findall(f.read())
         MUL_PIPELINE_DEPTH = int(depths[-1])
     with open(
         Path(ip_cores.__file__).parent
+        / "fadd"
         / f"flopoco_fadd_{WIDTH_EXPONENT}_{WIDTH_FRACTION}.vhdl"
     ) as f:
         depths = pipeline_depth_re.findall(f.read())
         ADD_PIPELINE_DEPTH = int(depths[-1])
     with open(
         Path(ip_cores.__file__).parent
+        / "fsub"
+        / f"flopoco_fsub_{WIDTH_EXPONENT}_{WIDTH_FRACTION}.vhdl"
+    ) as f:
+        depths = pipeline_depth_re.findall(f.read())
+        SUB_PIPELINE_DEPTH = int(depths[-1])
+    with open(
+        Path(ip_cores.__file__).parent
+        / "fdiv"
         / f"flopoco_fdiv_{WIDTH_EXPONENT}_{WIDTH_FRACTION}.vhdl"
     ) as f:
         depths = pipeline_depth_re.findall(f.read())
         DIV_PIPELINE_DEPTH = int(depths[-1])
-else:
-    MUL_PIPELINE_DEPTH = int(
-        os.getenv("MUL_PIPELINE_DEPTH") or config["ip"].get("MulPipelineDepth")
-    )
-    DIV_PIPELINE_DEPTH = int(
-        os.getenv("DIV_PIPELINE_DEPTH") or config["ip"].get("DivPipelineDepth")
-    )
-    # DIV_RADIX = int(os.getenv("DIV_RADIX") or config["ip"].get("DivRadix"))
-    ADD_PIPELINE_DEPTH = int(
-        os.getenv("ADD_PIPELINE_DEPTH") or config["ip"].get("AddPipelineDepth")
-    )
 
 MUL_LATENCY = MUL_PIPELINE_DEPTH + 1
 DIV_LATENCY = DIV_PIPELINE_DEPTH + 1
 ADD_LATENCY = ADD_PIPELINE_DEPTH + 1
+SUB_LATENCY = SUB_PIPELINE_DEPTH + 1
 
+# comb ops with registered outputs
+MAX_LATENCY = 2
+GT_LATENCY = 2
+NEG_LATENCY = 2
+RELU_LATENCY = 2
 
 _nameToLevel = {
     "CRITICAL": logging.CRITICAL,

bragghls/flopoco/convert_flopoco.py

@@ -12,7 +12,9 @@
 
 def convert_flopoco_binary_str_to_float(s, width_exp, width_frac):
     assert len(s) == width_exp + width_frac + 2 + 1
-    return float(flopoco_converter.bin2fp(width_exp, width_frac, s))
+    return float(
+        flopoco_converter.bin2fp(width_exp, width_frac, s).strip().replace("\x01", "")
+    )
 
 
 if __name__ == "__main__":

bragghls/testbench/filter_process.py bragghls/flopoco/filter_process.pybragghls/testbench/filter_process.py renamed to bragghls/flopoco/filter_process.py

@@ -1,14 +1,14 @@
 #!/usr/bin/env python3
 import sys
 
-from bragghls.flopoco.convert_flopoco import convert_flopoco_binary_str_to_float
+from convert_flopoco import convert_flopoco_binary_str_to_float
 
 
 def main():
     fh_in = sys.stdin
     fh_out = sys.stdout
 
-    width_exponent = width_fraction = 8
+    width_exponent = width_fraction = 4
     while True:
         # incoming values have newline
         l = fh_in.readline()
@@ -27,8 +27,8 @@ def main():
             )
             fh_out.write("%s\n" % s)
         except Exception as e:
-            print(e, file=sys.stderr)
-            fh_out.write("%s" % l)
+            # print(e, file=sys.stderr)
+            fh_out.write("%s\n" % l)
         fh_out.flush()
 
 

bragghls/flopoco/ops.py

@@ -1,35 +1,46 @@
+import operator
 from collections import namedtuple
 from dataclasses import dataclass
 from functools import reduce
 
 import numpy as np
 
-from bragghls.compiler import state
-from bragghls.config import WIDTH_EXPONENT, WIDTH_FRACTION
-from bragghls.util import idx_to_str, chunks
-
 try:
     from . import flopoco_converter
 except:
     import flopoco_converter
 
+from bragghls.compiler import state
+from bragghls.config import WIDTH_EXPONENT, WIDTH_FRACTION
+from bragghls.util import idx_to_str, chunks
+
 FPNUMBER = namedtuple("FPNUMBER", "pe_idx")(None)
 
 
-def reducer(accum, val):
+def reducer(accum, val, reduce_op):
     if len(val) > 1:
-        return accum + [val[0] + val[1]]
+        res = reduce_op(val[0], val[1])
+        return accum + [res]
     else:
         return accum + val
 
 
 def ReduceAdd(vals):
     pairs = list(chunks(list(vals), 2))
     while len(pairs) > 1:
-        pairs = list(chunks(reduce(reducer, pairs, []), 2))
+        pairs = list(
+            chunks(reduce(lambda x, y: reducer(x, y, operator.add), pairs, []), 2)
+        )
     return pairs[0][0] + pairs[0][1]
 
 
+def ReduceMax(vals):
+    pairs = list(chunks(list(vals), 2))
+    while len(pairs) > 1:
+        pairs = list(chunks(reduce(lambda x, y: reducer(x, y, max), pairs, []), 2))
+    return max(pairs[0][0], pairs[0][1])
+
+
 def check_make_val(v, width_exponent, width_fraction):
     if not isinstance(v, Val):
         assert isinstance(v, (float, int)), v
@@ -70,6 +81,12 @@ def __eq__(self, other):
         other = check_make_val(other, self.width_exponent, self.width_fraction)
         return self.fp == other.fp
 
+    def __lt__(self, other):
+        other = check_make_val(other, self.width_exponent, self.width_fraction)
+        # print(self.fp, other.fp, self.fp - other.fp, (self.fp - other.fp).sign())
+        # print((self.fp - other.fp).sign())
+        return (self.fp - other.fp).sign() == 1
+
     def __add__(self, other):
         other = check_make_val(other, self.width_exponent, self.width_fraction)
         v = add(self, other)
@@ -97,6 +114,10 @@ def __repr__(self):
             f"<IEEE {self.ieee:.5e}> {self.fp} {self.width_exponent} {self.width_fraction}"
         )
 
+    @property
+    def fp_float(self):
+        return float(f"{str(self.fp).split(':')[0].split(' ')[1]}")
+
 
 def mul(x: Val, y: Val):
     assert x.width_exponent == y.width_exponent
@@ -171,6 +192,9 @@ def numel(self):
     def reduce_add(self):
         return ReduceAdd(self.registers.flatten())
 
+    def reduce_max(self):
+        return ReduceMax(list(self.registers.flatten()))
+
     @property
     def val_names_map(self):
         assert self.input or self.output
@@ -250,6 +274,9 @@ def __getitem__(self, index):
     def numel(self):
         return np.prod(self.shape)
 
+    def reduce_max(self):
+        return ReduceMax(list(self.vals.flatten()))
+
     @staticmethod
     def from_global_memref(memref, width_exponent, width_fraction):
         return GlobalMemRef(
@@ -302,8 +329,19 @@ def Div(cst, val):
 
 
 def main():
-    five = Val(4.0, 4, 4)
-    print(Div(1.0, five))
+    a = Val(2, 4, 4)
+    b = Val(1, 4, 4)
+    print(a - a)
+    print(a, b)
+    print(a - b)
+    a = flopoco_converter.FPNumber(2, 4, 4)
+    b = flopoco_converter.FPNumber(1, 4, 4)
+    print(a + a)
+    print(a - a)
+    print(a - b)
+    print(b - b)
+    print(a - b)
+    print(a + b)
 
 
 if __name__ == "__main__":