hw: CSR-based hardware cluster barrier

hw/snitch/src/riscv_instr.sv

@@ -1133,6 +1133,7 @@ package riscv_instr;
   localparam logic [11:0] CSR_MHARTID = 12'hf14;
   localparam logic [11:0] CSR_SSR = 12'h7c0;
   localparam logic [11:0] CSR_FPMODE = 12'h7c1;
+  localparam logic [11:0] CSR_BARRIER = 12'h7c2;
   localparam logic [11:0] CSR_HTIMEDELTAH = 12'h615;
   localparam logic [11:0] CSR_CYCLEH = 12'hc80;
   localparam logic [11:0] CSR_TIMEH = 12'hc81;

hw/snitch/src/snitch.sv

@@ -102,7 +102,10 @@ module snitch import snitch_pkg::*; import riscv_instr::*; #(
   output fpnew_pkg::fmt_mode_t      fpu_fmt_mode_o,
   input  fpnew_pkg::status_t        fpu_status_i,
   // Core events for performance counters
-  output snitch_pkg::core_events_t  core_events_o
+  output snitch_pkg::core_events_t  core_events_o,
+  // Cluster HW barrier
+  output logic          barrier_o,
+  input  logic          barrier_i
 );
   // Debug module's base address
   localparam logic [31:0] DmBaseAddress = 0;
@@ -232,6 +235,7 @@ module snitch import snitch_pkg::*; import riscv_instr::*; #(
   // -----
   logic [31:0] csr_rvalue;
   logic csr_en;
+  logic csr_stall_d, csr_stall_q;
 
   localparam logic M = 0;
   localparam logic S = 1;
@@ -293,6 +297,8 @@ module snitch import snitch_pkg::*; import riscv_instr::*; #(
   `FFNR(dscratch_q, dscratch_d, clk_i)
   `FFAR(debug_q, debug_d, '0, clk_i, rst_i) // Debug mode
 
+  `FFAR(csr_stall_q, csr_stall_d, '0, clk_i, rst_i)
+
   typedef struct packed {
     fpnew_pkg::fmt_mode_t  fmode;
     fpnew_pkg::roundmode_e frm;
@@ -398,7 +404,7 @@ module snitch import snitch_pkg::*; import riscv_instr::*; #(
     | (~{(PPNSize){trans_active}} & {{{AddrWidth-32}{1'b0}}, pc_q[31:PageShift]});
   assign inst_addr_o[PageShift-1:0] = pc_q[PageShift-1:0];
   assign inst_cacheable_o = snitch_pma_pkg::is_inside_cacheable_regions(SnitchPMACfg, inst_addr_o);
-  assign inst_valid_o = ~wfi_q;
+  assign inst_valid_o = ~wfi_q && ~csr_stall_q;
 
   // --------------------
   // Control
@@ -454,7 +460,7 @@ module snitch import snitch_pkg::*; import riscv_instr::*; #(
     pc_d = pc_q;
     npc = pc_q; // the next PC if we wouldn't be in debug mode
     // if we got a valid instruction word increment the PC unless we are waiting for an event
-    if (!stall && !wfi_q) begin
+    if (!stall && !wfi_q && !csr_stall_q) begin
       casez (next_pc)
         Consec: npc = consec_pc;
         Alu: npc = alu_result & {{31{1'b1}}, ~zero_lsb};
@@ -2273,6 +2279,11 @@ module snitch import snitch_pkg::*; import riscv_instr::*; #(
     dpc_d = dpc_q;
     dscratch_d = dscratch_q;
 
+    csr_stall_d = csr_stall_q;
+
+    if (barrier_i) csr_stall_d = 1'b0;
+    barrier_o = 1'b0;
+
     // DPC and DCSR update logic
     if (!debug_q) begin
       if (valid_instr && inst_data_i == EBREAK) begin
@@ -2468,6 +2479,11 @@ module snitch import snitch_pkg::*; import riscv_instr::*; #(
               if (!exception) fcsr_d = fcsr_t'(alu_result[9:0]);
             end else illegal_csr = 1'b1;
           end
+          // HW cluster barrier
+          CSR_BARRIER: begin
+            barrier_o = 1'b1;
+            csr_stall_d = 1'b1;
+          end
           default: csr_rvalue = '0;
         endcase
       end else illegal_csr = 1'b1;

hw/snitch_cluster/src/snitch_barrier.sv

@@ -1,80 +1,37 @@
-// Copyright 2020 ETH Zurich and University of Bologna.
+// Copyright 2023 ETH Zurich and University of Bologna.
 // Solderpad Hardware License, Version 0.51, see LICENSE for details.
 // SPDX-License-Identifier: SHL-0.51
-
-// Author: Florian Zaruba <[email protected]>
-// Author: Fabian Schuiki <[email protected]>
+//
+// Author: Luca Colagrande <[email protected]>
 
 `include "common_cells/registers.svh"
 
 /// Hardware barrier to synchronize all cores in a cluster.
-module snitch_barrier
-  import snitch_pkg::*;
-  import snitch_cluster_peripheral_reg_pkg::*;
-#(
-  parameter int unsigned AddrWidth = 0,
-  parameter int  NrPorts = 0,
-  parameter type dreq_t = logic,
-  parameter type drsp_t = logic,
-  /// Derived parameter *Do not override*
-  parameter type addr_t = logic [AddrWidth-1:0]
+module snitch_barrier #(
+  parameter int NrCores = 0
 ) (
-  input  logic clk_i,
-  input  logic rst_ni,
-  input  dreq_t [NrPorts-1:0] in_req_i,
-  output drsp_t [NrPorts-1:0] in_rsp_o,
-
-  output dreq_t [NrPorts-1:0] out_req_o,
-  input  drsp_t [NrPorts-1:0] out_rsp_i,
-
-  input  addr_t              cluster_periph_start_address_i
+  input  logic               clk_i,
+  input  logic               rst_ni,
+  input  logic [NrCores-1:0] barrier_i,
+  output logic               barrier_o
 );
 
-  typedef enum logic [1:0] {
-    Idle,
-    Wait,
-    Take
-  } barrier_state_e;
-  barrier_state_e [NrPorts-1:0] state_d, state_q;
-  logic [NrPorts-1:0] is_barrier;
-  logic take_barrier;
+  logic [NrCores-1:0] arrival_d, arrival_q;
+
+  generate
+    for (genvar i = 0; i < NrCores; i++) begin : gen_arrival_bit
 
-  assign take_barrier = &is_barrier;
+      `FF(arrival_q[i], arrival_d[i], 1'b0, clk_i, rst_ni)
 
-  always_comb begin
-    state_d     = state_q;
-    is_barrier  = '0;
-    out_req_o = in_req_i;
-    in_rsp_o = out_rsp_i;
+      always_comb begin
+        if (barrier_o) arrival_d[i] = 1'b0;
+        else if (barrier_i[i]) arrival_d[i] = 1'b1;
+        else arrival_d[i] = arrival_q[i];
+      end
 
-    for (int i = 0; i < NrPorts; i++) begin
-      case (state_q[i])
-        Idle: begin
-          if (in_req_i[i].q_valid &&
-            (in_req_i[i].q.addr ==
-                cluster_periph_start_address_i +
-                SNITCH_CLUSTER_PERIPHERAL_HW_BARRIER_OFFSET)) begin
-            state_d[i] = Wait;
-            out_req_o[i].q_valid = 0;
-            in_rsp_o[i].q_ready  = 0;
-          end
-        end
-        Wait: begin
-          is_barrier[i]  = 1;
-          out_req_o[i].q_valid = 0;
-          in_rsp_o[i].q_ready  = 0;
-          if (take_barrier) state_d[i] = Take;
-        end
-        Take: begin
-          if (out_req_o[i].q_valid && in_rsp_o[i].q_ready) state_d[i] = Idle;
-        end
-        default: state_d[i] = Idle;
-      endcase
     end
-  end
+  endgenerate
 
-  for (genvar i = 0; i < NrPorts; i++) begin : gen_ff
-    `FFARN(state_q[i], state_d[i], Idle, clk_i, rst_ni)
-  end
+  assign barrier_o = &arrival_q;
 
 endmodule

hw/snitch_cluster/src/snitch_cc.sv

@@ -126,7 +126,10 @@ module snitch_cc #(
   output dma_events_t                axi_dma_events_o,
   // Core event strobes
   output snitch_pkg::core_events_t   core_events_o,
-  input  addr_t                      tcdm_addr_base_i
+  input  addr_t                      tcdm_addr_base_i,
+  // Cluster HW barrier
+  output logic                       barrier_o,
+  input  logic                       barrier_i
 );
 
   // FMA architecture is "merged" -> mulexp and macexp instructions are supported
@@ -252,7 +255,9 @@ module snitch_cc #(
     .fpu_rnd_mode_o ( fpu_rnd_mode ),
     .fpu_fmt_mode_o ( fpu_fmt_mode ),
     .fpu_status_i ( fpu_status ),
-    .core_events_o ( snitch_events)
+    .core_events_o ( snitch_events),
+    .barrier_o ( barrier_o ),
+    .barrier_i ( barrier_i )
   );
 
   reqrsp_iso #(

hw/snitch_cluster/src/snitch_cluster.sv

@@ -490,8 +490,8 @@ module snitch_cluster
   snitch_icache_pkg::icache_events_t [NrCores-1:0] icache_events;
 
   // 4. Memory Subsystem (Core side).
-  reqrsp_req_t [NrCores-1:0] core_req, filtered_core_req;
-  reqrsp_rsp_t [NrCores-1:0] core_rsp, filtered_core_rsp;
+  reqrsp_req_t [NrCores-1:0] core_req;
+  reqrsp_rsp_t [NrCores-1:0] core_rsp;
   reqrsp_req_t [NrHives-1:0] ptw_req;
   reqrsp_rsp_t [NrHives-1:0] ptw_rsp;
 
@@ -502,6 +502,8 @@ module snitch_cluster
   // 5. Misc. Wires.
   logic icache_prefetch_enable;
   logic [NrCores-1:0] cl_interrupt;
+  logic [NrCores-1:0] barrier_in;
+  logic barrier_out;
 
   // -------------
   // DMA Subsystem
@@ -889,7 +891,9 @@ module snitch_cluster
         .axi_dma_perf_o (),
         .axi_dma_events_o (dma_core_events),
         .core_events_o (core_events[i]),
-        .tcdm_addr_base_i (tcdm_start_address)
+        .tcdm_addr_base_i (tcdm_start_address),
+        .barrier_o (barrier_in[i]),
+        .barrier_i (barrier_out)
       );
       for (genvar j = 0; j < TcdmPorts; j++) begin : gen_tcdm_user
         always_comb begin
@@ -1000,19 +1004,14 @@ module snitch_cluster
   // --------
   // Coes SoC
   // --------
+
   snitch_barrier #(
-    .AddrWidth (PhysicalAddrWidth),
-    .NrPorts (NrCores),
-    .dreq_t  (reqrsp_req_t),
-    .drsp_t  (reqrsp_rsp_t)
+    .NrCores(NrCores)
   ) i_snitch_barrier (
     .clk_i,
     .rst_ni,
-    .in_req_i (core_req),
-    .in_rsp_o (core_rsp),
-    .out_req_o (filtered_core_req),
-    .out_rsp_i (filtered_core_rsp),
-    .cluster_periph_start_address_i (cluster_periph_start_address)
+    .barrier_i(barrier_in),
+    .barrier_o(barrier_out)
   );
 
   reqrsp_req_t core_to_axi_req;
@@ -1032,8 +1031,8 @@ module snitch_cluster
   ) i_reqrsp_mux_core (
     .clk_i,
     .rst_ni,
-    .slv_req_i (filtered_core_req),
-    .slv_rsp_o (filtered_core_rsp),
+    .slv_req_i (core_req),
+    .slv_rsp_o (core_rsp),
     .mst_req_o (core_to_axi_req),
     .mst_rsp_i (core_to_axi_rsp),
     .idx_o (/*unused*/)

sw/snRuntime/src/start.c

@@ -13,13 +13,6 @@ static inline uint32_t snrt_cls_base_addr() {
     return l1_end_addr - cdata_size - cbss_size;
 }
 
-static inline void snrt_crt0_cluster_hw_barrier() {
-    uint32_t register r;
-    uint32_t hw_barrier =
-        SNRT_CLUSTER_HW_BARRIER_ADDR + snrt_cluster_idx() * SNRT_CLUSTER_OFFSET;
-    asm volatile("lw %0, 0(%1)" : "=r"(r) : "r"(hw_barrier) : "memory");
-}
-
 static inline void snrt_init_tls() {
     extern volatile uint32_t __tdata_start, __tdata_end;
     extern volatile uint32_t __tbss_start, __tbss_end;
@@ -130,7 +123,7 @@ void snrt_main() {
 #endif
 
 #ifdef SNRT_CRT0_PRE_BARRIER
-    snrt_crt0_cluster_hw_barrier();
+    snrt_cluster_hw_barrier();
 #endif
 
 #ifdef SNRT_CRT0_CALLBACK5
@@ -147,7 +140,7 @@ void snrt_main() {
 #endif
 
 #ifdef SNRT_CRT0_POST_BARRIER
-    snrt_crt0_cluster_hw_barrier();
+    snrt_cluster_hw_barrier();
 #endif
 
 #ifdef SNRT_CRT0_CALLBACK7

sw/snRuntime/src/sync.h

@@ -56,12 +56,7 @@ inline void snrt_mutex_release(volatile uint32_t *pmtx) {
 
 /// Synchronize cores in a cluster with a hardware barrier
 inline void snrt_cluster_hw_barrier() {
-    uint32_t register r;
-
-    asm volatile("lw %0, 0(%1)"
-                 : "=r"(r)
-                 : "r"((uint32_t)snrt_cluster_hw_barrier_addr())
-                 : "memory");
+    asm volatile("csrr x0, 0x7C2" ::: "memory");
 }
 
 /// Synchronize clusters globally with a global software barrier

util/generate-opcodes.sh

@@ -3,7 +3,7 @@
 # Solderpad Hardware License, Version 0.51, see LICENSE for details.
 # SPDX-License-Identifier: SHL-0.51
 
-# Generate the opcodes for the Snith system.
+# Generate the opcodes for the Snitch system.
 set -e
 ROOT=$(cd "$(dirname "${BASH_SOURCE[0]}")/.." && pwd)