hw: Add Reduction Feature

Bender.lock

@@ -51,6 +51,21 @@ packages:
     source:
       Path: pd
     dependencies: []
+  fpnew:
+    revision: e5aa6a01b5bbe1675c3aa8872e1203413ded83d1
+    version: null
+    source:
+      Git: https://github.com/pulp-platform/cvfpu.git
+    dependencies:
+    - common_cells
+    - fpu_div_sqrt_mvp
+  fpu_div_sqrt_mvp:
+    revision: 86e1f558b3c95e91577c41b2fc452c86b04e85ac
+    version: 1.0.4
+    source:
+      Git: https://github.com/pulp-platform/fpu_div_sqrt_mvp.git
+    dependencies:
+    - common_cells
   idma:
     revision: 28a36e5e07705549e59fc33db96ab681bc1ca88e
     version: 0.6.5

Bender.yml

@@ -17,8 +17,9 @@ dependencies:
   axi_riscv_atomics: 0.8.2
   common_cells: 1.39.0
   common_verification: 0.2.3
-  idma: {version: 0.6.2, upstream_name: iDMA}
-  floo_noc_pd: {path: ./pd, target: "floo_synth"}
+  idma: { version: 0.6.2, upstream_name: iDMA }
+  FPnew: { git: "https://github.com/pulp-platform/cvfpu.git", rev: pulp-v0.2.3 }
+  floo_noc_pd: { path: ./pd, target: "floo_synth" }
 
 export_include_dirs:
   - hw/include
@@ -39,15 +40,18 @@ sources:
   - hw/floo_rob_wrapper.sv
   - hw/floo_reduction_sync.sv
   - hw/floo_route_xymask.sv
+  - hw/floo_alu.sv
   # Level 2
   - hw/floo_route_select.sv
   - hw/floo_route_comp.sv
   - hw/floo_meta_buffer.sv
   - hw/floo_reduction_arbiter.sv
+  - hw/floo_reduction_unit.sv
   # Level 3
   - hw/floo_output_arbiter.sv
   # Level 4
   - hw/floo_nw_join.sv
+    # TODO(lleone): Put back and fis axi_chimney
   - hw/floo_axi_chimney.sv
   - hw/floo_nw_chimney.sv
   - hw/floo_router.sv
@@ -70,13 +74,15 @@ sources:
       - hw/test/axi_reorder_remap_compare.sv
       - hw/test/axi_bw_monitor.sv
       - hw/test/floo_hbm_model.sv
+#      - hw/test/floo_reduction_offloads.sv
       - hw/test/axi_aw_w_sync.sv
       # Level 2
       - hw/tb/tb_floo_axi_chimney.sv
       - hw/tb/tb_floo_nw_chimney.sv
       - hw/tb/tb_floo_router.sv
       - hw/tb/tb_floo_rob.sv
       - hw/tb/tb_floo_rob_multicast.sv
+      #- hw/tb/tb_floo_rob_multicast.sv
 
   - target: all(any(floo_test, floo_synth), axi_mesh)
     files:
@@ -109,6 +115,7 @@ sources:
     - hw/synth/floo_synth_nw_chimney.sv
     - hw/synth/floo_synth_axi_router.sv
     - hw/synth/floo_synth_nw_router.sv
+    - hw/synth/floo_synth_nw_2tiles.sv
 
   - target: floo_deprecated_hw
     include_dirs:

floogen/examples/collective.yml

@@ -0,0 +1,99 @@
+# Copyright 2025 ETH Zurich and University of Bologna.
+# Licensed under the Apache License, Version 2.0, see LICENSE for details.
+# SPDX-License-Identifier: Apache-2.0
+#
+# Example: full collective support
+#   - narrow + wide multicast
+#   - barrier (LSB-AND)
+#   - narrow reduction: all integer ops (Add, Mul, MinS, MinU, MaxS, MaxU)
+#   - wide reduction: only Add and Max (selective ops example)
+
+name: nw_mesh_collective
+description: "4x4 NW mesh with full collective operation support"
+network_type: "narrow-wide"
+
+routing:
+  route_algo: "XY"
+  use_id_table: true
+  collective:
+    en_narrow_multicast: true
+    en_wide_multicast: true
+    en_barrier: true
+    en_narrow_reduction:
+      rd_pipeline_depth: 1   # all integer ops, custom hw config
+    en_wide_reduction:       # selective FP ops, default hw config
+      ops: [Add, Max]
+      rd_pipeline_depth: 5
+      cut_offload_intf: true
+  # decouple_rw: Phys
+  vc_impl: preempt
+
+protocols:
+  - name: "narrow_in"
+    type: "narrow"
+    protocol: "AXI4"
+    data_width: 64
+    addr_width: 48
+    id_width: 5
+    user_width:
+      collective_mask: 48
+      collective_op: 4
+      user: 5
+  - name: "narrow_out"
+    type: "narrow"
+    protocol: "AXI4"
+    data_width: 64
+    addr_width: 48
+    id_width: 2
+    user_width:
+      collective_mask: 48
+      collective_op: 4
+      user: 5
+  - name: "wide_in"
+    type: "wide"
+    protocol: "AXI4"
+    data_width: 512
+    addr_width: 48
+    id_width: 3
+    user_width:
+      collective_mask: 48
+      collective_op: 4
+  - name: "wide_out"
+    type: "wide"
+    protocol: "AXI4"
+    data_width: 512
+    addr_width: 48
+    id_width: 1
+    user_width:
+      collective_mask: 48
+      collective_op: 4
+
+endpoints:
+  - name: "cluster"
+    array: [4, 4]
+    addr_range:
+      base: 0x2000_0000
+      size: 0x0004_0000
+      en_collective: true
+    mgr_port_protocol:
+      - "narrow_in"
+      - "wide_in"
+    sbr_port_protocol:
+      - "narrow_out"
+      - "wide_out"
+
+routers:
+  - name: "router"
+    array: [4, 4]
+    degree: 5
+
+connections:
+  - src: "cluster"
+    dst: "router"
+    src_range:
+    - [0, 3]
+    - [0, 3]
+    dst_range:
+    - [0, 3]
+    - [0, 3]
+    dst_dir: "Eject"

floogen/examples/mcast.yml

@@ -0,0 +1,86 @@
+# Copyright 2025 ETH Zurich and University of Bologna.
+# Licensed under the Apache License, Version 2.0, see LICENSE for details.
+# SPDX-License-Identifier: Apache-2.0
+#
+# Example: narrow + wide multicast only (no reduction, no barrier)
+
+name: nw_mesh_mcast
+description: "4x4 NW mesh with narrow and wide multicast support"
+network_type: "narrow-wide"
+
+routing:
+  route_algo: "XY"
+  use_id_table: true
+  collective:
+    en_narrow_multicast: true
+    en_wide_multicast: true
+
+protocols:
+  - name: "narrow_in"
+    type: "narrow"
+    protocol: "AXI4"
+    data_width: 64
+    addr_width: 48
+    id_width: 5
+    user_width:
+      collective_mask: 48
+      collective_op: 4
+      user: 5
+  - name: "narrow_out"
+    type: "narrow"
+    protocol: "AXI4"
+    data_width: 64
+    addr_width: 48
+    id_width: 2
+    user_width:
+      collective_mask: 48
+      collective_op: 4
+      user: 5
+  - name: "wide_in"
+    type: "wide"
+    protocol: "AXI4"
+    data_width: 512
+    addr_width: 48
+    id_width: 3
+    user_width:
+      collective_mask: 48
+      collective_op: 4
+  - name: "wide_out"
+    type: "wide"
+    protocol: "AXI4"
+    data_width: 512
+    addr_width: 48
+    id_width: 1
+    user_width:
+      collective_mask: 48
+      collective_op: 4
+
+endpoints:
+  - name: "cluster"
+    array: [4, 4]
+    addr_range:
+      base: 0x2000_0000
+      size: 0x0004_0000
+      en_collective: true
+    mgr_port_protocol:
+      - "narrow_in"
+      - "wide_in"
+    sbr_port_protocol:
+      - "narrow_out"
+      - "wide_out"
+
+routers:
+  - name: "router"
+    array: [4, 4]
+    degree: 5
+
+connections:
+  - src: "cluster"
+    dst: "router"
+    src_range:
+    - [0, 3]
+    - [0, 3]
+    dst_range:
+    - [0, 3]
+    - [0, 3]
+    dst_dir: "Eject"

floogen/model/endpoint.py

@@ -90,9 +90,9 @@ def is_only_mgr(self) -> bool:
         """Return true if the endpoint is only a manager."""
         return self.is_mgr() and not self.is_sbr()
 
-    def is_multicast_ep(self) -> bool:
-        """Return true if the endpoint supports multicast."""
-        return any([b for b in self.addr_range if b.en_multicast])
+    def is_collective_ep(self) -> bool:
+        """Return true if the endpoint supports collective operations."""
+        return any([b for b in self.addr_range if b.en_collective])
 
     def get_ni_name(self, name: str) -> str:
         """Return the name of the NI."""
@@ -118,9 +118,9 @@ def from_desc(cls, desc: EndpointDesc,
     def render_ports(self, pkg_name=""):
         """Render the ports of the endpoint."""
         ports = []
-        mcast_prefix = "mcast" if self.is_multicast_ep() else ""
+        collective_prefix = "collective" if self.is_collective_ep() else ""
         for port in self.mgr_ports:
-            ports += port.render_port(pkg_name, prefix=mcast_prefix)
+            ports += port.render_port(pkg_name, prefix=collective_prefix)
         for port in self.sbr_ports:
-            ports += port.render_port(pkg_name, prefix=mcast_prefix)
+            ports += port.render_port(pkg_name, prefix=collective_prefix)
         return ports

floogen/model/link.py

@@ -94,11 +94,14 @@ def wide_name(self, is_reversed=False):
         return f"{self.source}_to_{self.dest}_wide"
 
     @classmethod
-    def render_typedefs(cls, axi_narrow, axi_wide, cfg_n, cfg_w):
+    def render_typedefs(cls, axi_narrow, axi_wide, cfg_n, cfg_w, vc_num=None, phy_num=None):
         """Render the typedefs of the links."""
         string = f"`FLOO_TYPEDEF_NW_CHAN_ALL(axi, req, rsp, wide, \
             {axi_narrow}, {axi_wide}, {cfg_n}, {cfg_w}, hdr_t)\n\n"
-        string += "`FLOO_TYPEDEF_NW_LINK_ALL(req, rsp, wide, req, rsp, wide)\n"
+        if vc_num is not None and phy_num is not None:
+            string += f"`FLOO_TYPEDEF_NW_VIRT_CHAN_LINK_ALL(req, rsp, wide, req, rsp, wide, {vc_num}, {phy_num})\n"
+        else:
+            string += "`FLOO_TYPEDEF_NW_LINK_ALL(req, rsp, wide, req, rsp, wide)\n"
         return string
 
     def declare(self):

floogen/model/network.py

@@ -13,7 +13,7 @@
 import matplotlib.pyplot as plt
 from pydantic import BaseModel, ConfigDict, StringConstraints, field_validator, model_validator
 
-from floogen.model.routing import Routing, RouteAlgo, RouteMapRule, RouteRule, RouteMap, RouteTable, RouteMapRuleMcast
+from floogen.model.routing import Routing, RouteAlgo, RouteMapRule, RouteRule, RouteMap, RouteTable, RouteMapRuleCollective, WideRwDecouple
 from floogen.model.routing import Coord, SimpleId, AddrRange, XYDirections
 from floogen.model.graph import Graph
 from floogen.model.endpoint import EndpointDesc, Endpoint
@@ -92,7 +92,7 @@ def validate_protocols(self):
             if len(set(prot.data_width for prot in self.protocols if prot.type == "wide")) != 1:
                 raise ValueError("All `wide` protocols must have the same data width")
             # Check that `narrow` and `wide` protocols have the same user width
-            if not self.routing.en_multicast:
+            if not self.routing.en_collective:
                 if len(set(prot.user_width for prot in self.protocols if prot.type == "narrow")) != 1:
                     raise ValueError("All `narrow` protocols must have the same user width")
                 if len(set(prot.user_width for prot in self.protocols if prot.type == "wide")) != 1:
@@ -575,8 +575,8 @@ def gen_routing_info(self):
                     f"Routing algorithm {self.routing.route_algo} is not supported yet"
                 )
         self.routing.sam = self.gen_sam()
-        if self.routing.en_multicast:
-            self.routing.multicast_sam = self.gen_multicast_sam()
+        if self.routing.en_collective:
+            self.routing.collective_sam = self.gen_collective_sam()
         # Provide the routing info to the network interfaces
         for ni in self.graph.get_ni_nodes():
             ni.routing = self.routing
@@ -660,26 +660,26 @@ def gen_sam(self):
                 rule_name = ni.endpoint.name
                 if addr_range.desc is not None:
                     rule_name += f"_{addr_range.desc}"
-                addr_rule = RouteMapRule(dest=dest, addr_range=addr_range, en_multicast=addr_range.en_multicast, desc=rule_name)
+                addr_rule = RouteMapRule(dest=dest, addr_range=addr_range, en_collective=addr_range.en_collective, desc=rule_name)
                 addr_table.append(addr_rule)
         return RouteMap(name="sam", rules=addr_table)
 
-    def gen_multicast_sam(self):
-        """Generate the multicast system address map, which contains additional mask
-        information."""
+    def gen_collective_sam(self):
+        """Generate the collective system address map, which contains additional mask
+        information for collective endpoints."""
         addr_table = []
         for addr_rule in self.routing.sam.rules:
             mask_fields = {}
-            if addr_rule.addr_range.en_multicast:
+            if addr_rule.addr_range.en_collective:
                 mask_offset_x = clog2(addr_rule.addr_range.size)
                 mask_fields = {
                     "mask_len": (self.routing.num_x_bits, self.routing.num_y_bits),
                     "mask_offset": (mask_offset_x + self.routing.num_y_bits, mask_offset_x),
                     "base_id": (addr_rule.dest.x - addr_rule.addr_range.arr_idx[0],
                                 addr_rule.dest.y - addr_rule.addr_range.arr_idx[1]),
                 }
-            addr_table.append(RouteMapRuleMcast(**addr_rule.model_dump(), **mask_fields))
-        return RouteMap(name="mcast_sam", rules=addr_table)
+            addr_table.append(RouteMapRuleCollective(**addr_rule.model_dump(), **mask_fields))
+        return RouteMap(name="collective_sam", rules=addr_table)
 
     def render_ports(self, pkg_name=""):
         """Render the ports in the generated code."""
@@ -707,8 +707,13 @@ def render_link_typedefs(self):
             string += AXI4.render_cfg("AxiCfgN", narrow_in_prot, narrow_out_prot)
             string += AXI4.render_cfg("AxiCfgW", wide_in_prot, wide_out_prot)
 
+            vc_num, phy_num = {
+                WideRwDecouple.PHYS: (2, 2),
+                WideRwDecouple.VC:   (2, 1),
+            }.get(self.routing.decouple_rw, (None, None))
             string += NarrowWideLink.render_typedefs(
-                narrow_in_prot.type_name(), wide_in_prot.type_name(), "AxiCfgN", "AxiCfgW"
+                narrow_in_prot.type_name(), wide_in_prot.type_name(), "AxiCfgN", "AxiCfgW",
+                vc_num, phy_num
             )
         else:
             in_prot = next((prot for prot in self.protocols if prot.direction == "input"), None)

floogen/model/network_interface.py

@@ -47,9 +47,9 @@ def is_only_mgr(self) -> bool:
         """Return true if the network interface is only a manager."""
         return self.endpoint.is_mgr() and not self.endpoint.is_sbr()
 
-    def is_multicast_ni(self) -> bool:
-        """Return true if the network interface supports multicast."""
-        return any([b for b in self.addr_range if b.en_multicast])
+    def is_collective_ni(self) -> bool:
+        """Return true if the network interface supports collective operations."""
+        return any([b for b in self.addr_range if b.en_collective])
 
     def render_enum_name(self) -> str:
         """Render the enum name."""