More metrics for RF, RS and ROB (#632)

coreblocks/cache/icache.py

@@ -115,7 +115,7 @@ def __init__(self, layouts: ICacheLayouts, params: ICacheParameters, refiller: C
         self.perf_misses = HwCounter("frontend.icache.misses")
         self.perf_errors = HwCounter("frontend.icache.fetch_errors")
         self.perf_flushes = HwCounter("frontend.icache.flushes")
-        self.req_latency = LatencyMeasurer(
+        self.req_latency = FIFOLatencyMeasurer(
             "frontend.icache.req_latency", "Latencies of cache requests", slots_number=2, max_latency=500
         )
 

coreblocks/core_structs/rf.py

@@ -1,7 +1,9 @@
 from amaranth import *
-from transactron import Method, def_method, TModule
+from transactron import Method, Transaction, def_method, TModule
 from coreblocks.interface.layouts import RFLayouts
 from coreblocks.params import GenParams
+from transactron.lib.metrics import HwExpHistogram, TaggedLatencyMeasurer
+from transactron.utils.amaranth_ext.functions import popcount
 from transactron.utils.transactron_helpers import make_layout
 
 __all__ = ["RegisterFile"]
@@ -20,9 +22,24 @@ def __init__(self, *, gen_params: GenParams):
         self.write = Method(i=layouts.rf_write)
         self.free = Method(i=layouts.rf_free)
 
+        self.perf_rf_valid_time = TaggedLatencyMeasurer(
+            "struct.rf.valid_time",
+            description="Distribution of time registers are valid in RF",
+            slots_number=2**gen_params.phys_regs_bits,
+            max_latency=1000,
+        )
+        self.perf_num_valid = HwExpHistogram(
+            "struct.rf.num_valid",
+            description="Number of valid registers in RF",
+            bucket_count=gen_params.phys_regs_bits + 1,
+            sample_width=gen_params.phys_regs_bits + 1,
+        )
+
     def elaborate(self, platform):
         m = TModule()
 
+        m.submodules += [self.perf_rf_valid_time, self.perf_num_valid]
+
         being_written = Signal(self.gen_params.phys_regs_bits)
         written_value = Signal(self.gen_params.isa.xlen)
 
@@ -56,10 +73,20 @@ def _(reg_id: Value, reg_val: Value):
             with m.If(~(zero_reg)):
                 m.d.sync += self.entries[reg_id].reg_val.eq(reg_val)
                 m.d.sync += self.entries[reg_id].valid.eq(1)
+                self.perf_rf_valid_time.start(m, slot=reg_id)
 
         @def_method(m, self.free)
         def _(reg_id: Value):
             with m.If(reg_id != 0):
                 m.d.sync += self.entries[reg_id].valid.eq(0)
+                self.perf_rf_valid_time.stop(m, slot=reg_id)
+
+        if self.perf_num_valid.metrics_enabled():
+            num_valid = Signal(self.gen_params.phys_regs_bits + 1)
+            m.d.comb += num_valid.eq(
+                popcount(Cat(self.entries[reg_id].valid for reg_id in range(2**self.gen_params.phys_regs_bits)))
+            )
+            with Transaction(name="perf").body(m):
+                self.perf_num_valid.add(m, num_valid)
 
         return m

coreblocks/core_structs/rob.py

@@ -1,5 +1,5 @@
 from amaranth import *
-from transactron import Method, def_method, TModule
+from transactron import Method, Transaction, def_method, TModule
 from transactron.lib.metrics import *
 from coreblocks.interface.layouts import ROBLayouts
 from coreblocks.params import GenParams
@@ -18,17 +18,23 @@ def __init__(self, gen_params: GenParams) -> None:
         self.data = Array(Signal(layouts.internal_layout) for _ in range(2**gen_params.rob_entries_bits))
         self.get_indices = Method(o=layouts.get_indices, nonexclusive=True)
 
-        self.perf_rob_wait_time = LatencyMeasurer(
+        self.perf_rob_wait_time = FIFOLatencyMeasurer(
             "backend.rob.wait_time",
             description="Distribution of time instructions spend in ROB",
             slots_number=(2**gen_params.rob_entries_bits + 1),
             max_latency=1000,
         )
+        self.perf_rob_size = HwExpHistogram(
+            "backend.rob.size",
+            description="Number of instructions in ROB",
+            bucket_count=gen_params.rob_entries_bits + 1,
+            sample_width=gen_params.rob_entries_bits,
+        )
 
     def elaborate(self, platform):
         m = TModule()
 
-        m.submodules += [self.perf_rob_wait_time]
+        m.submodules += [self.perf_rob_wait_time, self.perf_rob_size]
 
         start_idx = Signal(self.params.rob_entries_bits)
         end_idx = Signal(self.params.rob_entries_bits)
@@ -70,4 +76,10 @@ def _(rob_id: Value, exception):
         def _():
             return {"start": start_idx, "end": end_idx}
 
+        if self.perf_rob_size.metrics_enabled():
+            rob_size = Signal(self.params.rob_entries_bits)
+            m.d.comb += rob_size.eq((end_idx - start_idx)[0 : self.params.rob_entries_bits])
+            with Transaction(name="perf").body(m):
+                self.perf_rob_size.add(m, rob_size)
+
         return m

coreblocks/func_blocks/csr/csr.py

@@ -236,6 +236,7 @@ def _(rob_id: Value, side_fx: Value):
         return m
 
 
+@dataclass(frozen=True)
 class CSRBlockComponent(BlockComponentParams):
     def get_module(self, gen_params: GenParams) -> FuncBlock:
         connections = gen_params.get(DependencyManager)

coreblocks/func_blocks/fu/common/rs.py

@@ -2,19 +2,25 @@
 from typing import Optional
 from amaranth import *
 from amaranth.lib.coding import PriorityEncoder
-from transactron import Method, def_method, TModule
+from transactron import Method, Transaction, def_method, TModule
 from coreblocks.params import GenParams
 from coreblocks.frontend.decoder import OpType
 from coreblocks.interface.layouts import RSLayouts
+from transactron.lib.metrics import HwExpHistogram, TaggedLatencyMeasurer
 from transactron.utils import RecordDict
+from transactron.utils.amaranth_ext.functions import popcount
 from transactron.utils.transactron_helpers import make_layout
 
 __all__ = ["RS"]
 
 
 class RS(Elaboratable):
     def __init__(
-        self, gen_params: GenParams, rs_entries: int, ready_for: Optional[Iterable[Iterable[OpType]]] = None
+        self,
+        gen_params: GenParams,
+        rs_entries: int,
+        rs_number: int,
+        ready_for: Optional[Iterable[Iterable[OpType]]] = None,
     ) -> None:
         ready_for = ready_for or ((op for op in OpType),)
         self.gen_params = gen_params
@@ -38,10 +44,24 @@ def __init__(
         self.data = Array(Signal(self.internal_layout) for _ in range(self.rs_entries))
         self.data_ready = Signal(self.rs_entries)
 
+        self.perf_rs_wait_time = TaggedLatencyMeasurer(
+            f"fu.block_{rs_number}.rs.valid_time",
+            description=f"Distribution of time instructions wait in RS {rs_number}",
+            slots_number=2**self.rs_entries_bits,
+            max_latency=1000,
+        )
+        self.perf_num_full = HwExpHistogram(
+            f"fu.block_{rs_number}.rs.num_full",
+            description=f"Number of full entries in RS {rs_number}",
+            bucket_count=self.rs_entries_bits + 1,
+            sample_width=self.rs_entries_bits + 1,
+        )
+
     def elaborate(self, platform):
         m = TModule()
 
         m.submodules.enc_select = PriorityEncoder(width=self.rs_entries)
+        m.submodules += [self.perf_rs_wait_time, self.perf_num_full]
 
         for i, record in enumerate(self.data):
             m.d.comb += self.data_ready[i].eq(
@@ -71,6 +91,7 @@ def _(rs_entry_id: Value, rs_data: Value) -> None:
             m.d.sync += self.data[rs_entry_id].rs_data.eq(rs_data)
             m.d.sync += self.data[rs_entry_id].rec_full.eq(1)
             m.d.sync += self.data[rs_entry_id].rec_reserved.eq(1)
+            self.perf_rs_wait_time.start(m, slot=rs_entry_id)
 
         @def_method(m, self.update)
         def _(reg_id: Value, reg_val: Value) -> None:
@@ -89,6 +110,7 @@ def _(rs_entry_id: Value) -> RecordDict:
             record = self.data[rs_entry_id]
             m.d.sync += record.rec_reserved.eq(0)
             m.d.sync += record.rec_full.eq(0)
+            self.perf_rs_wait_time.stop(m, slot=rs_entry_id)
             return {
                 "s1_val": record.rs_data.s1_val,
                 "s2_val": record.rs_data.s2_val,
@@ -105,4 +127,10 @@ def _(rs_entry_id: Value) -> RecordDict:
             def _() -> RecordDict:
                 return {"ready_list": ready_list}
 
+        if self.perf_num_full.metrics_enabled():
+            num_full = Signal(self.rs_entries_bits + 1)
+            m.d.comb += num_full.eq(popcount(Cat(self.data[entry_id].rec_full for entry_id in range(self.rs_entries))))
+            with Transaction(name="perf").body(m):
+                self.perf_num_full.add(m, num_full)
+
         return m

coreblocks/func_blocks/fu/common/rs_func_block.py

@@ -31,7 +31,9 @@ class RSFuncBlock(FuncBlock, Elaboratable):
         layout described by `FuncUnitLayouts`.
     """
 
-    def __init__(self, gen_params: GenParams, func_units: Iterable[tuple[FuncUnit, set[OpType]]], rs_entries: int):
+    def __init__(
+        self, gen_params: GenParams, func_units: Iterable[tuple[FuncUnit, set[OpType]]], rs_entries: int, rs_number: int
+    ):
         """
         Parameters
         ----------
@@ -41,10 +43,13 @@ def __init__(self, gen_params: GenParams, func_units: Iterable[tuple[FuncUnit, s
             Functional units to be used by this module.
         rs_entries: int
             Number of entries in RS.
+        rs_number: int
+            The number of this RS block. Used for debugging.
         """
         self.gen_params = gen_params
         self.rs_entries = rs_entries
         self.rs_entries_bits = (rs_entries - 1).bit_length()
+        self.rs_number = rs_number
         self.rs_layouts = gen_params.get(RSLayouts, rs_entries_bits=self.rs_entries_bits)
         self.fu_layouts = gen_params.get(FuncUnitLayouts)
         self.func_units = list(func_units)
@@ -60,6 +65,7 @@ def elaborate(self, platform):
         m.submodules.rs = self.rs = RS(
             gen_params=self.gen_params,
             rs_entries=self.rs_entries,
+            rs_number=self.rs_number,
             ready_for=(optypes for _, optypes in self.func_units),
         )
 
@@ -87,10 +93,13 @@ def elaborate(self, platform):
 class RSBlockComponent(BlockComponentParams):
     func_units: Collection[FunctionalComponentParams]
     rs_entries: int
+    rs_number: int = -1  # overwritten by CoreConfiguration
 
     def get_module(self, gen_params: GenParams) -> FuncBlock:
         modules = list((u.get_module(gen_params), u.get_optypes()) for u in self.func_units)
-        rs_unit = RSFuncBlock(gen_params=gen_params, func_units=modules, rs_entries=self.rs_entries)
+        rs_unit = RSFuncBlock(
+            gen_params=gen_params, func_units=modules, rs_entries=self.rs_entries, rs_number=self.rs_number
+        )
         return rs_unit
 
     def get_optypes(self) -> set[OpType]:

coreblocks/func_blocks/lsu/dummyLsu.py

@@ -1,3 +1,4 @@
+from dataclasses import dataclass
 from amaranth import *
 from amaranth.lib.data import View
 
@@ -320,6 +321,7 @@ def _(rob_id: Value, side_fx: Value):
         return m
 
 
+@dataclass(frozen=True)
 class LSUBlockComponent(BlockComponentParams):
     def get_module(self, gen_params: GenParams) -> FuncBlock:
         connections = gen_params.get(DependencyManager)

coreblocks/params/configurations.py

@@ -74,6 +74,12 @@ class CoreConfiguration:
         Definitions of PMAs per contiguous segments of memory.
     """
 
+    def __post_init__(self):
+        self.func_units_config = [
+            dataclasses.replace(conf, rs_number=k) if hasattr(conf, "rs_number") else conf
+            for k, conf in enumerate(self.func_units_config)
+        ]
+
     xlen: int = 32
     func_units_config: Collection[BlockComponentParams] = basic_configuration
 

coreblocks/params/fu_params.py

@@ -1,4 +1,5 @@
 from abc import abstractmethod, ABC
+from dataclasses import dataclass
 from collections.abc import Collection, Iterable
 
 from coreblocks.func_blocks.interface.func_protocols import FuncBlock, FuncUnit
@@ -20,6 +21,7 @@
 ]
 
 
+@dataclass(frozen=True)
 class BlockComponentParams(ABC):
     @abstractmethod
     def get_module(self, gen_params: "GenParams") -> FuncBlock:

test/regression/cocotb/benchmark.Makefile

@@ -14,7 +14,7 @@ SIM_BUILD = build/benchmark
 
 # Yosys/Amaranth borkedness workaround
 ifeq ($(SIM),verilator)
-  EXTRA_ARGS += -Wno-CASEINCOMPLETE -Wno-CASEOVERLAP -Wno-WIDTHEXPAND -Wno-WIDTHTRUNC
+  EXTRA_ARGS += -Wno-CASEINCOMPLETE -Wno-CASEOVERLAP -Wno-WIDTHEXPAND -Wno-WIDTHTRUNC -Wno-UNSIGNED
   BUILD_ARGS += -j`nproc`
 endif
 

test/regression/cocotb/signature.Makefile

@@ -14,7 +14,7 @@ SIM_BUILD = build/signature
 
 # Yosys/Amaranth borkedness workaround
 ifeq ($(SIM),verilator)
-  EXTRA_ARGS += -Wno-CASEINCOMPLETE -Wno-CASEOVERLAP -Wno-WIDTHEXPAND -Wno-WIDTHTRUNC
+  EXTRA_ARGS += -Wno-CASEINCOMPLETE -Wno-CASEOVERLAP -Wno-WIDTHEXPAND -Wno-WIDTHTRUNC -Wno-UNSIGNED
   BUILD_ARGS += -j`nproc`
 endif
 

test/regression/cocotb/test.Makefile

@@ -14,7 +14,7 @@ SIM_BUILD = build/test
 
 # Yosys/Amaranth borkedness workaround
 ifeq ($(SIM),verilator)
-  EXTRA_ARGS += -Wno-CASEINCOMPLETE -Wno-CASEOVERLAP -Wno-WIDTHEXPAND -Wno-WIDTHTRUNC
+  EXTRA_ARGS += -Wno-CASEINCOMPLETE -Wno-CASEOVERLAP -Wno-WIDTHEXPAND -Wno-WIDTHTRUNC -Wno-UNSIGNED
   BUILD_ARGS += -j`nproc`
 endif
 

test/scheduler/test_scheduler.py

@@ -127,7 +127,7 @@ def setUp(self):
         self.rs_count = len(self.optype_sets)
         self.gen_params = GenParams(
             test_core_config.replace(
-                func_units_config=tuple(RSBlockComponent([], rs_entries=4) for _ in range(self.rs_count))
+                func_units_config=tuple(RSBlockComponent([], rs_entries=4, rs_number=k) for k in range(self.rs_count))
             )
         )
         self.expected_rename_queue = deque()

test/scheduler/test_wakeup_select.py

@@ -43,7 +43,9 @@ def elaborate(self, platform):
 class TestWakeupSelect(TestCaseWithSimulator):
     def setUp(self):
         self.gen_params = GenParams(
-            test_core_config.replace(func_units_config=tuple(RSBlockComponent([], rs_entries=16) for _ in range(2)))
+            test_core_config.replace(
+                func_units_config=tuple(RSBlockComponent([], rs_entries=16, rs_number=k) for k in range(2))
+            )
         )
         self.m = WakeupTestCircuit(self.gen_params)
         self.cycles = 50

test/structs_common/test_rs.py

@@ -24,7 +24,7 @@ class TestRSMethodInsert(TestCaseWithSimulator):
     def test_insert(self):
         self.gen_params = GenParams(test_core_config)
         self.rs_entries_bits = self.gen_params.max_rs_entries_bits
-        self.m = SimpleTestCircuit(RS(self.gen_params, 2**self.rs_entries_bits, None))
+        self.m = SimpleTestCircuit(RS(self.gen_params, 2**self.rs_entries_bits, 0, None))
         self.insert_list = [
             {
                 "rs_entry_id": id,
@@ -69,7 +69,7 @@ class TestRSMethodSelect(TestCaseWithSimulator):
     def test_select(self):
         self.gen_params = GenParams(test_core_config)
         self.rs_entries_bits = self.gen_params.max_rs_entries_bits
-        self.m = SimpleTestCircuit(RS(self.gen_params, 2**self.rs_entries_bits, None))
+        self.m = SimpleTestCircuit(RS(self.gen_params, 2**self.rs_entries_bits, 0, None))
         self.insert_list = [
             {
                 "rs_entry_id": id,
@@ -132,7 +132,7 @@ class TestRSMethodUpdate(TestCaseWithSimulator):
     def test_update(self):
         self.gen_params = GenParams(test_core_config)
         self.rs_entries_bits = self.gen_params.max_rs_entries_bits
-        self.m = SimpleTestCircuit(RS(self.gen_params, 2**self.rs_entries_bits, None))
+        self.m = SimpleTestCircuit(RS(self.gen_params, 2**self.rs_entries_bits, 0, None))
         self.insert_list = [
             {
                 "rs_entry_id": id,
@@ -223,7 +223,7 @@ class TestRSMethodTake(TestCaseWithSimulator):
     def test_take(self):
         self.gen_params = GenParams(test_core_config)
         self.rs_entries_bits = self.gen_params.max_rs_entries_bits
-        self.m = SimpleTestCircuit(RS(self.gen_params, 2**self.rs_entries_bits, None))
+        self.m = SimpleTestCircuit(RS(self.gen_params, 2**self.rs_entries_bits, 0, None))
         self.insert_list = [
             {
                 "rs_entry_id": id,
@@ -322,7 +322,7 @@ class TestRSMethodGetReadyList(TestCaseWithSimulator):
     def test_get_ready_list(self):
         self.gen_params = GenParams(test_core_config)
         self.rs_entries_bits = self.gen_params.max_rs_entries_bits
-        self.m = SimpleTestCircuit(RS(self.gen_params, 2**self.rs_entries_bits, None))
+        self.m = SimpleTestCircuit(RS(self.gen_params, 2**self.rs_entries_bits, 0, None))
         self.insert_list = [
             {
                 "rs_entry_id": id,
@@ -378,7 +378,7 @@ def test_two_get_ready_lists(self):
         self.rs_entries = self.gen_params.max_rs_entries
         self.rs_entries_bits = self.gen_params.max_rs_entries_bits
         self.m = SimpleTestCircuit(
-            RS(self.gen_params, 2**self.rs_entries_bits, [[OpType(1), OpType(2)], [OpType(3), OpType(4)]])
+            RS(self.gen_params, 2**self.rs_entries_bits, 0, [[OpType(1), OpType(2)], [OpType(3), OpType(4)]])
         )
         self.insert_list = [
             {