alap toposort, improve tests

src/kyupy/logic_sim.py

@@ -105,7 +105,7 @@ class LogicSim(sim.SimOps):
                     elif op == sim.OAI211:self.c[o0] = ~((self.c[i0] | self.c[i1]) & self.c[i2] & self.c[i3])
                     elif op == sim.MUX21: self.c[o0] = (self.c[i0] & ~self.c[i2]) | (self.c[i1] & self.c[i2])
                     else: print(f'unknown op {op}')
-                    inject_cb(o0, self.s[o0])
+                    inject_cb(o0, self.c[o0])
         elif self.m == 4:
             for op, o0, i0, i1, i2, i3 in self.ops[:,:6]:
                 o0, i0, i1, i2, i3 = [self.c_locs[x] for x in (o0, i0, i1, i2, i3)]
@@ -181,6 +181,7 @@ class LogicSim(sim.SimOps):
                     logic.bp4v_and(self.c[t1], self.c[i1], self.c[i2])
                     logic.bp4v_or(self.c[o0], self.c[t0], self.c[t1])
                 else: print(f'unknown op {op}')
+                if inject_cb is not None: inject_cb(o0, self.c[o0])
         else:
             for op, o0, i0, i1, i2, i3 in self.ops[:,:6]:
                 o0, i0, i1, i2, i3 = [self.c_locs[x] for x in (o0, i0, i1, i2, i3)]
@@ -256,7 +257,7 @@ class LogicSim(sim.SimOps):
                     logic.bp8v_and(self.c[t1], self.c[i1], self.c[i2])
                     logic.bp8v_or(self.c[o0], self.c[t0], self.c[t1])
                 else: print(f'unknown op {op}')
-                if inject_cb is not None: inject_cb(o0, self.s[o0])
+                if inject_cb is not None: inject_cb(o0, self.c[o0])
 
     def c_to_s(self):
         """Copies (captures) the results of the combinational portion to ``s[1]``.

src/kyupy/sim.py

@@ -9,6 +9,9 @@ from .circuit import Circuit
 BUF1 = np.uint16(0b1010_1010_1010_1010)
 INV1 = ~BUF1
 
+__const0__ = BUF1
+__const1__ = INV1
+
 AND2 = np.uint16(0b1000_1000_1000_1000)
 AND3 = np.uint16(0b1000_0000_1000_0000)
 AND4 = np.uint16(0b1000_0000_0000_0000)
@@ -41,7 +44,10 @@ AOI211, OAI211 = ~AO211, ~OA211
 
 MUX21 = np.uint16(0b1100_1010_1100_1010)  # z = i1 if i2 else i0 (i2 is select)
 
-names = dict([(v, k) for k, v in globals().items() if isinstance(v, np.uint16)])
+names = dict([(v, k) for k, v in globals().items() if isinstance(v, np.uint16) and '__' not in k])
+
+prim2name = dict([(v, k) for k, v in globals().items() if isinstance(v, np.uint16) and '__' not in k])
+name2prim = dict([(k, v) for k, v in globals().items() if isinstance(v, np.uint16)])
 
 kind_prefixes = {
     'nand': (NAND4, NAND3, NAND2),
@@ -177,84 +183,75 @@ class SimOps:
         self.ppo_offset = self.ppi_offset + self.s_len
         self.c_locs_len = self.ppo_offset + self.s_len
 
-        # translate circuit structure into self.ops
+        # ALAP-toposort the circuit into self.ops
-        ops = []
+        levels = []
-        interface_dict = dict((n, i) for i, n in enumerate(circuit.s_nodes))
+
-        for n in circuit.topological_order():
+        ppio2idx = dict((n, i) for i, n in enumerate(circuit.s_nodes))
-            if n in interface_dict:
+        pis = set([n for n in circuit.s_nodes if len(n.ins) == 0])
-                inp_idx = self.ppi_offset + interface_dict[n]
+        ppos = set([n for n in circuit.s_nodes if len(n.ins) > 0])
-                if len(n.outs) > 0 and n.outs[0] is not None:  # first output of a PI/PPI
+        readers = np.array([1 if l.reader in ppos else len(l.reader.outs) for l in circuit.lines], dtype=np.int32)  # for ref-counting forks
-                    ops.append((BUF1, n.outs[0].index, inp_idx, self.zero_idx, self.zero_idx, self.zero_idx, *a_ctrl[n.outs[0]]))
+
-                if 'dff' in n.kind.lower():  # second output of DFF is inverted
+        level_lines = [n.ins[0] for n in ppos]  # start from PPOs
-                    if len(n.outs) > 1 and n.outs[1] is not None:
+        # FIXME: Should probably instanciate buffers for PPOs and attach DFF clocks
-                        ops.append((INV1, n.outs[1].index, inp_idx, self.zero_idx, self.zero_idx, self.zero_idx, *a_ctrl[n.outs[1]]))
+
-                else:  # if not DFF, no output is inverted.
+        while len(level_lines) > 0:  # traverse the circuit level-wise back towards (P)PIs
-                    for o_line in n.outs[1:]:
+            level_ops = []
-                        if o_line is not None:
+            prev_level_lines = []
-                            ops.append((BUF1, o_line.index, inp_idx, self.zero_idx, self.zero_idx, self.zero_idx, *a_ctrl[o_line]))
+
-                continue
+            for l in level_lines:
-            # regular node, not PI/PPI or PO/PPO
+                n = l.driver
-            o0_idx = n.outs[0].index if len(n.outs) > 0 and n.outs[0] is not None else self.tmp_idx
+                in_idxs = [n.ins[x].index if len(n.ins) > x and n.ins[x] is not None else self.zero_idx for x in [0,1,2,3]]
-            i0_idx = n.ins[0].index if len(n.ins) > 0 and n.ins[0] is not None else self.zero_idx
+                if n in ppio2idx:
-            i1_idx = n.ins[1].index if len(n.ins) > 1 and n.ins[1] is not None else self.zero_idx
+                    in_idxs[0] = self.ppi_offset + ppio2idx[n]
-            i2_idx = n.ins[2].index if len(n.ins) > 2 and n.ins[2] is not None else self.zero_idx
+                    if l.driver_pin == 1 and 'dff' in n.kind.lower():  # second output of DFF is inverted
-            i3_idx = n.ins[3].index if len(n.ins) > 3 and n.ins[3] is not None else self.zero_idx
+                        level_ops.append((INV1, l.index, *in_idxs, *a_ctrl[l]))
-            kind = n.kind.lower()
+                    else:
-            if kind == '__fork__':
+                        level_ops.append((BUF1, l.index, *in_idxs, *a_ctrl[l]))
-                if not strip_forks:
+                elif n.kind == '__fork__':
-                    for o_line in n.outs:
+                    readers[n.ins[0]] -= 1
-                        if o_line is not None:
+                    if readers[n.ins[0]] == 0: prev_level_lines.append(n.ins[0])
-                            ops.append((BUF1, o_line.index, i0_idx, i1_idx, i2_idx, i3_idx, *a_ctrl[o_line]))
+                    if not strip_forks: level_ops.append((BUF1, l.index, *in_idxs, *a_ctrl[l]))
-                continue
+                else:
-            sp = None
+                    prev_level_lines += n.ins
-            for prefix, prims in kind_prefixes.items():
+                    sp = None
-                if kind.startswith(prefix):
+                    kind = n.kind.lower()
-                    sp = prims[0]
+                    for prefix, prims in kind_prefixes.items():
-                    if i3_idx == self.zero_idx:
+                        if kind.startswith(prefix):
-                        sp = prims[1]
+                            sp = prims[0]
-                        if i2_idx == self.zero_idx:
+                            if in_idxs[3] == self.zero_idx:
-                            sp = prims[2]
+                                sp = prims[1]
-                    break
+                                if in_idxs[2] == self.zero_idx:
-            if sp is None:
+                                    sp = prims[2]
-                print('unknown cell type', kind)
+                            break
-            else:
+                    if sp is None:
-                ops.append((sp, o0_idx, i0_idx, i1_idx, i2_idx, i3_idx, *a_ctrl[o0_idx]))
+                        print('unknown cell type', kind)
-
+                    else:
-        self.ops = np.asarray(ops, dtype='int32')
+                        level_ops.append((sp, l.index, *in_idxs, *a_ctrl[l]))
+
+            if len(level_ops) > 0: levels.append(level_ops)
+            level_lines = prev_level_lines
+
+        self.levels = [np.asarray(lv, dtype=np.int32) for lv in levels[::-1]]
+        level_sums = np.cumsum([0]+[len(lv) for lv in self.levels], dtype=np.int32)
+        self.level_starts, self.level_stops = level_sums[:-1], level_sums[1:]
+        self.ops = np.vstack(self.levels)
 
         # create a map from fanout lines to stem lines for fork stripping
-        stems = np.zeros(self.c_locs_len, dtype='int32') - 1  # default to -1: 'no fanout line'
+        stems = np.full(self.c_locs_len, -1, dtype=np.int32)  # default to -1: 'no fanout line'
         if strip_forks:
             for f in circuit.forks.values():
                 prev_line = f.ins[0]
                 while prev_line.driver.kind == '__fork__':
                     prev_line = prev_line.driver.ins[0]
-                stem_idx = prev_line.index
                 for ol in f.outs:
                     if ol is not None:
-                        stems[ol] = stem_idx
+                        stems[ol] = prev_line.index
-
+
-        # calculate level (distance from PI/PPI) and reference count for each line
+        ref_count = np.zeros(self.c_locs_len, dtype=np.int32)
-        levels = np.zeros(self.c_locs_len, dtype='int32')
+
-        ref_count = np.zeros(self.c_locs_len, dtype='int32')
+        for op in self.ops:
-        level_starts = [0]
+            for x in [2, 3, 4, 5]:
-        current_level = 1
+                ref_count[stems[op[x]] if stems[op[x]] >= 0 else op[x]] += 1
-        for i, op in enumerate(self.ops):
-            # if we fork-strip, always take the stems for determining fan-in level
-            i0_idx = stems[op[2]] if stems[op[2]] >= 0 else op[2]
-            i1_idx = stems[op[3]] if stems[op[3]] >= 0 else op[3]
-            i2_idx = stems[op[4]] if stems[op[4]] >= 0 else op[4]
-            i3_idx = stems[op[5]] if stems[op[5]] >= 0 else op[5]
-            if levels[i0_idx] >= current_level or levels[i1_idx] >= current_level or levels[i2_idx] >= current_level or levels[i3_idx] >= current_level:
-                current_level += 1
-                level_starts.append(i)
-            levels[op[1]] = current_level  # set level of the output line
-            ref_count[i0_idx] += 1
-            ref_count[i1_idx] += 1
-            ref_count[i2_idx] += 1
-            ref_count[i3_idx] += 1
-        self.level_starts = np.asarray(level_starts, dtype='int32')
-        self.level_stops = np.asarray(level_starts[1:] + [len(self.ops)], dtype='int32')
 
         # combinational signal allocation table. maps line and interface indices to self.c memory locations
         self.c_locs = np.full((self.c_locs_len,), -1, dtype=np.int32)
@@ -280,9 +277,9 @@ class SimOps:
                 ref_count[i0_idx] += 1
 
         # allocate memory for the rest of the circuit
-        for op_start, op_stop in zip(self.level_starts, self.level_stops):
+        for ops in self.levels:
             free_set = set()
-            for op in self.ops[op_start:op_stop]:
+            for op in ops:
                 # if we fork-strip, always take the stems
                 i0_idx = stems[op[2]] if stems[op[2]] >= 0 else op[2]
                 i1_idx = stems[op[3]] if stems[op[3]] >= 0 else op[3]
@@ -301,7 +298,8 @@ class SimOps:
                 self.c_locs[o_idx], self.c_caps[o_idx] = h.alloc(cap), cap
             if c_reuse:
                 for loc in free_set:
-                    h.free(loc)
+                    if loc >= 0:  # DFF clocks are not allocated. Ignore for now.
+                        h.free(loc)
 
         # copy memory location and capacity from stems to fanout lines
         for lidx, stem in enumerate(stems):

src/kyupy/wave_sim.py

@@ -377,9 +377,11 @@ class WaveSimCuda(WaveSim):
 
     def _grid_dim(self, x, y): return cdiv(x, self._block_dim[0]), cdiv(y, self._block_dim[1])
 
-    def c_prop(self, sims=None, seed=1):
+    def c_prop(self, sims=None, seed=1, op_from=0, op_to=None):
         sims = min(sims or self.sims, self.sims)
         for op_start, op_stop in zip(self.level_starts, self.level_stops):
+            if op_from > op_start: continue
+            if op_to is not None and op_to <= op_start: break
             grid_dim = self._grid_dim(sims, op_stop - op_start)
             wave_eval_gpu[grid_dim, self._block_dim](self.ops, op_start, op_stop, self.c, self.c_locs, self.c_caps, self.abuf, int(0),
                 sims, self.delays, self.simctl_int, seed)

tests/conftest.py

@@ -13,6 +13,13 @@ def b15_2ig_circuit(mydir):
     from kyupy.techlib import SAED32
     return verilog.load(mydir / 'b15_2ig.v.gz', branchforks=True, tlib=SAED32)
 
+@pytest.fixture(scope='session')
+def b15_2ig_circuit_resolved(b15_2ig_circuit):
+    from kyupy.techlib import SAED32
+    cr = b15_2ig_circuit.copy()
+    cr.resolve_tlib_cells(SAED32)
+    return cr
+
 @pytest.fixture(scope='session')
 def b15_2ig_delays(mydir, b15_2ig_circuit):
     from kyupy import sdf

tests/test_logic_sim.py

@@ -173,3 +173,50 @@ def test_b01(mydir):
     s.c_prop()
     s.c_to_s()
     bp_to_mv(s.s[1])
+
+
+def sim_and_compare(v_file, stil_file, m=8):
+    from kyupy import verilog, stil
+    from kyupy.techlib import SAED32
+    c = verilog.load(v_file, branchforks=True, tlib=SAED32)
+    c.resolve_tlib_cells(SAED32)
+    s = stil.load(stil_file)
+    tests = s.tests(c)[:,1:]
+    resp = s.responses(c)[:,1:]
+    lsim = LogicSim(c, m=m, sims=tests.shape[1])
+    lsim.s[0] = logic.mv_to_bp(tests)
+    lsim.s_to_c()
+    lsim.c_prop()
+    lsim.c_to_s()
+    resp_sim = logic.bp_to_mv(lsim.s[1])[:,:tests.shape[1]]
+    idxs, pats = np.nonzero(((resp == logic.ONE) & (resp_sim != logic.ONE)) | ((resp == logic.ZERO) & (resp_sim != logic.ZERO)))
+    for i, (idx, pat) in enumerate(zip(idxs, pats)):
+        if i >= 10:
+            print(f'...')
+            break
+        print(f'mismatch pattern:{pat} ppio:{idx} exp:{logic.mv_str(resp[idx,pat])} act:{logic.mv_str(resp_sim[idx,pat])}')
+    assert len(idxs) == 0
+
+
+def test_b15_2ig_sa_2v(mydir):
+    sim_and_compare(mydir / 'b15_2ig.v.gz', mydir / 'b15_2ig.sa_nf.stil.gz', m=2)
+
+
+def test_b15_2ig_sa_4v(mydir):
+    sim_and_compare(mydir / 'b15_2ig.v.gz', mydir / 'b15_2ig.sa_nf.stil.gz', m=4)
+
+
+def test_b15_2ig_sa_8v(mydir):
+    sim_and_compare(mydir / 'b15_2ig.v.gz', mydir / 'b15_2ig.sa_nf.stil.gz', m=8)
+
+
+def test_b15_4ig_sa_2v(mydir):
+    sim_and_compare(mydir / 'b15_4ig.v.gz', mydir / 'b15_4ig.sa_rf.stil.gz', m=2)
+
+
+def test_b15_4ig_sa_4v(mydir):
+    sim_and_compare(mydir / 'b15_4ig.v.gz', mydir / 'b15_4ig.sa_rf.stil.gz', m=4)
+
+
+def test_b15_4ig_sa_8v(mydir):
+    sim_and_compare(mydir / 'b15_4ig.v.gz', mydir / 'b15_4ig.sa_rf.stil.gz', m=8)

tests/test_wave_sim.py

@@ -177,7 +177,7 @@ def compare_to_logic_sim(wsim: WaveSim):
     lsim.s_to_c()
     lsim.c_prop()
     lsim.c_to_s()
-    exp = logic.bp_to_mv(lsim.s[1])
+    exp = logic.bp_to_mv(lsim.s[1])[:,:tests.shape[-1]]
 
     resp[resp == logic.PPULSE] = logic.ZERO
     resp[resp == logic.NPULSE] = logic.ONE
@@ -188,13 +188,13 @@ def compare_to_logic_sim(wsim: WaveSim):
     np.testing.assert_allclose(resp, exp)
 
 
-def test_b15(b15_2ig_circuit, b15_2ig_delays):
+def test_b15(b15_2ig_circuit_resolved, b15_2ig_delays):
-    compare_to_logic_sim(WaveSim(b15_2ig_circuit, b15_2ig_delays, 8))
+    compare_to_logic_sim(WaveSim(b15_2ig_circuit_resolved, b15_2ig_delays, 8))
 
 
-def test_b15_strip_forks(b15_2ig_circuit, b15_2ig_delays):
+def test_b15_strip_forks(b15_2ig_circuit_resolved, b15_2ig_delays):
-    compare_to_logic_sim(WaveSim(b15_2ig_circuit, b15_2ig_delays, 8, strip_forks=True))
+    compare_to_logic_sim(WaveSim(b15_2ig_circuit_resolved, b15_2ig_delays, 8, strip_forks=True))
 
 
-def test_b15_cuda(b15_2ig_circuit, b15_2ig_delays):
+def test_b15_cuda(b15_2ig_circuit_resolved, b15_2ig_delays):
-    compare_to_logic_sim(WaveSimCuda(b15_2ig_circuit, b15_2ig_delays, 8, strip_forks=True))
+    compare_to_logic_sim(WaveSimCuda(b15_2ig_circuit_resolved, b15_2ig_delays, 8, strip_forks=True))