hw(output_arb): Remove gen_path_mask and clean up output_arbiter

Author: fischeti

Bender.yml

@@ -34,7 +34,6 @@ sources:
   - hw/floo_rob_wrapper.sv
   - hw/floo_reduction_sync.sv
   - hw/floo_route_xymask.sv
-  - hw/floo_gen_path_mask.sv
   # Level 2
   - hw/floo_route_select.sv
   - hw/floo_route_comp.sv

hw/floo_gen_path_mask.sv

@@ -6,7 +6,9 @@
 
 module floo_output_arbiter import floo_pkg::*;
 #(
+  /// Number of input ports
   parameter int unsigned NumRoutes  = 1,
+  /// Type definitions
   parameter type         flit_t     = logic,
   parameter type         payload_t  = logic,
   parameter payload_t    NarrowRspMask = '0,
@@ -15,86 +17,75 @@ module floo_output_arbiter import floo_pkg::*;
 ) (
   input  logic                   clk_i,
   input  logic                   rst_ni,
-  /// Ports towards the input routes
+  /// Input ports
   input  logic  [NumRoutes-1:0]  valid_i,
   output logic  [NumRoutes-1:0]  ready_o,
   input  flit_t [NumRoutes-1:0]  data_i,
-  input  id_t                    node_id_i,
-  /// Ports towards the output route
+  input  id_t                    xy_id_i,
+  /// Output port
   output logic                   valid_o,
   input  logic                   ready_i,
   output flit_t                  data_o
 );
 
-  flit_t[NumRoutes-1:0]  in_reducing_data, in_normal_data;
-  logic [NumRoutes-1:0]  in_reducing_valid, in_reducing_ready;
-  logic [NumRoutes-1:0]  in_normal_valid, in_normal_ready;
+  flit_t[NumRoutes-1:0]  reduce_data_in, unicast_data_in;
+  logic [NumRoutes-1:0]  reduce_valid_in, reduce_ready_in;
+  logic [NumRoutes-1:0]  unicast_valid_in, unicast_ready_in;
 
-  // 1 for every bit means that the corresponding flit is normal, 0 means that it needs to be reduced
-  logic [NumRoutes-1:0]  path_mask;
+  flit_t reduced_data_out, unicast_data_out;
+  logic reduced_valid_out, unicast_valid_out;
 
-  flit_t out_reduced_data, out_normal_data;
-  logic out_reduced_valid, out_reduced_ready;
-  logic out_normal_valid, out_normal_ready;
+  logic [NumRoutes-1:0]  reduce_mask;
 
-  assign out_reduced_ready = ready_i;
-  assign out_normal_ready = ready_i;
+  // Determine which input ports are to be reduced
+  for (genvar i = 0; i < NumRoutes; i++) begin : gen_reduce_mask
+    assign reduce_mask[i] = (data_i[i].hdr.commtype == CollectB);
+  end
 
-  floo_gen_path_mask #(
-    .NumRoutes ( NumRoutes ),
-    .flit_t    ( flit_t    )
-  ) i_gen_path_mask (
-    .clk_i,
-    .rst_ni,
-    .data_i ( data_i    ),
-    .mask_o ( path_mask )
-  );
-
-  // the normal arbitrater side
-  assign in_normal_valid = valid_i & path_mask;
-  assign in_normal_data  = data_i ;
+  // Arbitrate unicasts
+  assign unicast_valid_in = valid_i & ~reduce_mask;
+  assign unicast_data_in  = data_i;
 
   floo_wormhole_arbiter #(
     .NumRoutes  ( NumRoutes ),
-    .flit_t     ( flit_t          )
+    .flit_t     ( flit_t    )
   ) i_wormhole_arbiter (
     .clk_i,
     .rst_ni,
-
-    .valid_i ( in_normal_valid ),
-    .ready_o ( in_normal_ready ),
-    .data_i  ( in_normal_data  ),
-
-    .valid_o ( out_normal_valid ),
-    .ready_i ( out_reduced_ready ),
-    .data_o  ( out_normal_data )
+    .valid_i ( unicast_valid_in  ),
+    .ready_o ( unicast_ready_in  ),
+    .data_i  ( unicast_data_in   ),
+    .valid_o ( unicast_valid_out ),
+    .ready_i ( ready_i           ),
+    .data_o  ( unicast_data_out  )
   );
 
-  // the reduction side
-  assign in_reducing_valid = valid_i & ~path_mask;
-  assign in_reducing_data  = data_i ;
+  // Arbitrate reductions
+  assign reduce_valid_in = valid_i & reduce_mask;
+  assign reduce_data_in  = data_i;
 
   floo_reduction_arbiter #(
-    .NumRoutes  ( NumRoutes ),
-    .flit_t     ( flit_t          ),
-    .payload_t  ( payload_t       ),
-    .NarrowRspMask ( NarrowRspMask ),
-    .WideRspMask   ( WideRspMask   ),
-    .id_t       ( id_t            )
+    .NumRoutes      ( NumRoutes     ),
+    .flit_t         ( flit_t        ),
+    .payload_t      ( payload_t     ),
+    .id_t           ( id_t          ),
+    .NarrowRspMask  ( NarrowRspMask ),
+    .WideRspMask    ( WideRspMask   )
   ) i_reduction_arbiter (
-    .valid_i ( in_reducing_valid ),
-    .ready_o ( in_reducing_ready ),
-    .data_i  ( in_reducing_data  ),
-    .node_id_i ( node_id_i ),
-
-    .valid_o ( out_reduced_valid ),
-    .ready_i ( out_normal_ready ),
-    .data_o  ( out_reduced_data )
+    .valid_i   ( reduce_valid_in   ),
+    .ready_o   ( reduce_ready_in   ),
+    .data_i    ( reduce_data_in    ),
+    .node_id_i ( xy_id_i           ),
+    .valid_o   ( reduced_valid_out ),
+    .ready_i   ( ready_i           ),
+    .data_o    ( reduced_data_out  )
   );
 
-  // choose which side to output, prefer the reduced side
-  assign valid_o = (out_reduced_valid & |in_reducing_valid) ? out_reduced_valid : out_normal_valid;
-  assign data_o  = (out_reduced_valid & |in_reducing_valid) ? out_reduced_data  : out_normal_data;
-  assign ready_o = (out_reduced_valid & |in_reducing_valid) ? in_reducing_ready : in_normal_ready;
+  // Arbitrate between wormhole and reduction arbiter
+  // TODO(fischeti): Discuss with Chen if the handshaking is correctly handled
+  // I believe that the `ready_i` of the two arbiters should be masked.
+  assign valid_o = (reduced_valid_out & |reduce_valid_in) ? reduced_valid_out : unicast_valid_out;
+  assign data_o  = (reduced_valid_out & |reduce_valid_in) ? reduced_data_out  : unicast_data_out;
+  assign ready_o = (reduced_valid_out & |reduce_valid_in) ? reduce_ready_in : unicast_ready_in;
 
 endmodule

hw/floo_output_arbiter.sv

@@ -210,7 +210,7 @@ module floo_router
           .valid_i  ( masked_valid[out][v] ),
           .ready_o  ( masked_ready[out][v] ),
           .data_i   ( masked_data [out][v] ),
-          .node_id_i( xy_id_i              ),
+          .xy_id_i  ( xy_id_i              ),
 
           .valid_o ( out_valid[out][v] ),
           .ready_i ( out_ready[out][v] ),