Multilayer models

.github/workflows/verilator-linting.yml

@@ -0,0 +1,39 @@
+name: Verilog Code Linting with Verilator
+
+on:
+  push:
+    branches: [ "main" ]
+  pull_request:
+    branches: [ "main" ]
+  workflow_dispatch:
+    inputs:
+      logLevel:
+        description: 'Log level'
+        required: true
+        default: 'warning'
+        type: choice
+        options:
+          - info
+          - warning
+          - debug
+
+jobs:
+  verilog-lint:
+    runs-on: ubuntu-latest
+    container:
+      image: verilator/verilator:latest  # Use your Docker image
+
+    steps:
+      - name: Checkout code
+        uses: actions/checkout@v3
+
+      - name: Run Verilator Linter
+        run: |
+          mkdir -p lint_reports
+          find . -name "*.sv" -or -name "*.v" | xargs -I {} sh -c 'verilator --lint-only  {} > lint_reports/$(basename {}).lint || true'
+
+      - name: Upload lint reports
+        uses: actions/upload-artifact@v4
+        with:
+          name: verilator-lint-reports
+          path: lint_reports/

.github/workflows/verilog_linting.yml

@@ -0,0 +1,47 @@
+name: Verilog Code Linting with Verible
+
+on:
+  push:
+    branches: [ "main" ]
+  pull_request:
+    branches: [ "main" ]
+  workflow_dispatch:
+    inputs:
+      logLevel:
+        description: 'Log level'
+        required: true
+        default: 'warning'
+        type: choice
+        options:
+          - info
+          - warning
+          - debug
+
+jobs:
+  verilog-lint:
+    runs-on: ubuntu-latest
+    container:
+      image: alexanderlwhite/ample-image:latest  # Use your Docker image with Verilator and Verible
+
+    steps:
+      - name: Checkout code
+        uses: actions/checkout@v3
+
+      - name: Cache Verible Lint Configuration
+        uses: actions/cache@v3
+        with:
+          path: .verible_lint.rules
+          key: ${{ runner.os }}-verible-lint-config-${{ hashFiles('.verible_lint.rules') }}
+          restore-keys: |
+            ${{ runner.os }}-verible-lint-config-
+
+      - name: Run Verible Linter
+        run: |
+          mkdir -p lint_reports
+          find . -name "*.sv" -or -name "*.v" | xargs -I {} sh -c 'verible-verilog-lint --rules_config .verible_lint.rules {} > lint_reports/$(basename {}).lint || true'
+
+      - name: Upload lint reports
+        uses: actions/upload-artifact@v4
+        with:
+          name: verible-lint-reports
+          path: lint_reports/

.gitignore

@@ -9,6 +9,9 @@ vivado_*
 *.runs
 *.sim
 *.tmp
+*.pt
+*.pth
+*.csv
 
 *.log
 *.jou
@@ -18,9 +21,13 @@ vivado_*
 *.vdb
 *.sdb
 
+
+
+
 # Simulation
 xsim.dir/
 .Xil/
+.nfs*
 
 vagrant/
 
@@ -42,6 +49,7 @@ graph_dump.txt
 nodeslot_programming.json
 layer_config.json
 *.mem
+Cora/
 
 hw/build/
 hw/sim/xsim.ini
@@ -53,6 +61,7 @@ hw/update_regbanks.tcl
 # Modelsim simulation
 sim_build/
 modelsim_lib/
+verilator_build/
 
 # Cocotb
 results.xml
@@ -61,4 +70,19 @@ transcript
 *.wlf
 compile.do
 simulate.do
-opt.do
+opt.do
+
+
+.vscode/
+
+
+hw/sim/-debugDB
+hw/sim/sim_cycles.txt
+hw/sim/vsim.dbg
+hw/sim/sim_time.txt
+imports/verilog-axi
+imports/nocrouter
+imports/
+
+hw/tb/module_tests/lib/test/*
+

Dockerfile

@@ -0,0 +1,18 @@
+FROM ubuntu:latest
+
+# Install dependencies
+RUN apt-get update && apt-get install -y \
+    wget \
+    tar \
+    xz-utils \
+    build-essential
+
+# Install Verible
+RUN apt-get update && apt-get upgrade -y \
+    && wget https://github.com/chipsalliance/verible/releases/download/v0.0-3724-gdec56671/verible-v0.0-3724-gdec56671-linux-static-x86_64.tar.gz \
+    &&   tar -xzf verible-v0.0-3724-gdec56671-linux-static-x86_64.tar.gz \
+    && mv verible-v0.0-3724-gdec56671/bin/* /usr/local/bin/ 
+
+CMD ["bash"]
+
+

README.md

@@ -117,7 +117,7 @@ source $WORKAREA/scripts/build.sh
 6. Generate the simulation payloads. For example, for the KarateClub dataset:
 
 ```bash
-$WORKAREA/scripts/initialize.py --karate --gcn --payloads --random
+python3 $WORKAREA/scripts/initialize.py --karate --gcn --payloads --random
 ```
 
 7. Build the testbench.
@@ -131,6 +131,19 @@ make build
 make sim GUI=1
 ```
 
+
+8. Benchmark AMPLE against a CPU
+```bash
+make build 
+python3 $WORKAREA/scripts/initialize.py --karate --mlp --payloads --random --layers 10 --sim --cpu --tb_log_level INFO
+```
+
+
+
+Note: Defining ```SIMULATION``` in the Modelsim makefile  will simualte without Xillinx IP and ```SIMULATION_QUICK``` will simulate without FP units, activaiton, bias or aggregation
+
+
+
 <p align="right">(<a href="#readme-top">back to top</a>)</p>
 
 

hw/ip/aggregation_engine/include/age_pkg.sv

@@ -21,6 +21,11 @@ typedef struct packed {
 
     logic [MAX_AGC_PER_NODE-1:0] [$clog2(noc_pkg::MAX_MESH_COLS)-1:0] coords_x;
     logic [MAX_AGC_PER_NODE-1:0] [$clog2(noc_pkg::MAX_MESH_ROWS)-1:0] coords_y;
+
+    logic [MAX_AGC_PER_NODE-1:0] [$clog2(top_pkg::MAX_FEATURE_COUNT)-1:0] num_features;
+
+
+
 } AGE_AGM_REQ_t;
 
 typedef struct packed {

hw/ip/aggregation_engine/rtl/aggregation_core.sv

@@ -35,15 +35,17 @@ module aggregation_core #(
 );
 
 parameter ALLOCATION_PKT_AGGR_FUNC_OFFSET = $clog2(top_pkg::MAX_NODESLOT_COUNT);
+parameter ALLOCATION_PKT_NUM_FEATURES_OFFSET = ALLOCATION_PKT_AGGR_FUNC_OFFSET + $bits(top_pkg::AGGREGATION_FUNCTION_e) ;
 
 parameter EXPECTED_FLITS_PER_PACKET = 1;
 
-typedef enum logic [3:0] {
+typedef enum logic [4:0] {
     AGC_FSM_IDLE,
     AGC_FSM_NODESLOT_ALLOCATION,
     AGC_FSM_WAIT_FEATURE_HEAD,
     AGC_FSM_WAIT_FEATURE_BODY,
     AGC_FSM_UPDATE_ACCS,
+    AGC_FSM_WAIT_UPDATE_ACCS,
     AGC_FSM_WAIT_BUFFER_REQ,
     AGC_FSM_SEND_BUFF_MAN,
     AGC_FSM_WAIT_DRAIN
@@ -119,6 +121,11 @@ logic                                           noc_router_waiting;
 
 logic [SCALE_FACTOR_QUEUE_READ_WIDTH-1:0]       scale_factor_q;
 
+
+logic [$clog2(top_pkg::MAX_FEATURE_COUNT)-1:0] num_features;
+logic [MESH_NODE_ID_WIDTH - 1 : 0] agc_loc;
+
+
 // ==================================================================================================================================================
 // Instantiations
 // ==================================================================================================================================================
@@ -202,16 +209,23 @@ always_comb begin
     end
 
     AGC_FSM_UPDATE_ACCS: begin
+        agc_state_n = feature_aggregator_in_feature_ready ? AGC_FSM_WAIT_UPDATE_ACCS
+                    : AGC_FSM_UPDATE_ACCS;
+    end
+
+
+    AGC_FSM_WAIT_UPDATE_ACCS: begin
         agc_state_n = 
                     // Updating last feature accumulator and last packet flag has already been received
                     (received_flits == EXPECTED_FLITS_PER_PACKET[3:0]) && &feature_updated && aggregation_manager_packet_last_q ? AGC_FSM_WAIT_BUFFER_REQ // updating final features
 
                     // Updating last feature accumulator but packets still pending
                     : (received_flits == EXPECTED_FLITS_PER_PACKET[3:0]) && &feature_updated ? AGC_FSM_WAIT_FEATURE_HEAD
 
-                    : AGC_FSM_UPDATE_ACCS;
+                    : AGC_FSM_WAIT_UPDATE_ACCS;
     end
 
+
     AGC_FSM_WAIT_BUFFER_REQ: begin
         agc_state_n = router_aggregation_core_valid && aggregation_manager_pkt && received_buffer_req_head && tail_packet && correct_pkt_dest ? AGC_FSM_SEND_BUFF_MAN
                     : AGC_FSM_WAIT_BUFFER_REQ;
@@ -240,6 +254,7 @@ always_comb begin
     head_packet = (router_aggregation_core_data.flit_label == noc_pkg::HEAD);
     tail_packet = (router_aggregation_core_data.flit_label == noc_pkg::TAIL);
 
+
     packet_source = router_aggregation_core_data.data.head_data.head_pl[noc_pkg::HEAD_PAYLOAD_SIZE-1 : noc_pkg::HEAD_PAYLOAD_SIZE-MESH_NODE_ID_WIDTH];    
     packet_source_col = packet_source[MESH_NODE_ID_WIDTH - 1 : MESH_NODE_ID_WIDTH - $clog2(noc_pkg::MAX_MESH_COLS)];
     packet_source_row = packet_source[$clog2(noc_pkg::MAX_MESH_ROWS)-1:0];
@@ -351,10 +366,23 @@ always_comb begin
                                             : sent_flits_counter == EXPECTED_FLITS_PER_PACKET[$clog2(FEATURE_COUNT/2)-1:0] ? noc_pkg::TAIL
                                             : noc_pkg::BODY;
 
-    aggregation_core_router_data.data.bt_pl = {buffer_manager_pkt_dest_col, buffer_manager_pkt_dest_row,
-                                                X_COORD[$clog2(MAX_MESH_COLS)-1:0], Y_COORD[$clog2(MAX_MESH_ROWS)-1:0], // source node coordinates
-                                                bm_chosen_data };
-end
+    if (sent_flits_counter == '0) begin
+
+        aggregation_core_router_data.data.head_data.x_dest = {buffer_manager_pkt_dest_col};
+        aggregation_core_router_data.data.head_data.y_dest = {buffer_manager_pkt_dest_row};
+
+        agc_loc[MESH_NODE_ID_WIDTH - 1 : MESH_NODE_ID_WIDTH - $clog2(MAX_MESH_COLS)] = X_COORD[$clog2(MAX_MESH_COLS)-1:0]; 
+        agc_loc[$clog2(MAX_MESH_ROWS)-1:0] = Y_COORD[$clog2(MAX_MESH_ROWS)-1:0];
+
+        aggregation_core_router_data.data.head_data.head_pl[noc_pkg::HEAD_PAYLOAD_SIZE-1 : noc_pkg::HEAD_PAYLOAD_SIZE-MESH_NODE_ID_WIDTH] = agc_loc;
+
+        aggregation_core_router_data.data.head_data.head_pl[noc_pkg::HEAD_PAYLOAD_SIZE-MESH_NODE_ID_WIDTH- 1 : noc_pkg::HEAD_PAYLOAD_SIZE-MESH_NODE_ID_WIDTH - $bits(num_features)] = {num_features};
+
+
+    end else begin
+        aggregation_core_router_data.data.bt_pl = {bm_chosen_data};
+    end
+end 
 
 // Nodeslot allocation
 // --------------------------------------------
@@ -385,7 +413,10 @@ always_ff @(posedge core_clk or negedge resetn) begin
                                                 : router_aggregation_core_data.data.bt_pl [ALLOCATION_PKT_AGGR_FUNC_OFFSET + $bits(top_pkg::AGGREGATION_FUNCTION_e) - 1 : ALLOCATION_PKT_AGGR_FUNC_OFFSET] == 2'd2 ? top_pkg::WEIGHTED_SUM
                                                 : AGGR_FUNC_RESERVED;
 
+        num_features <= router_aggregation_core_data.data.bt_pl [ALLOCATION_PKT_NUM_FEATURES_OFFSET + $bits(top_pkg::MAX_FEATURE_COUNT) - 1 : ALLOCATION_PKT_NUM_FEATURES_OFFSET];  
+
     end
+
 end
 
 

hw/ip/aggregation_engine/rtl/aggregation_core_allocator.sv

@@ -74,18 +74,24 @@ if (ALLOCATION_MODE == age_pkg::AGC_ALLOCATION_MODE_STATIC) begin
         agm_req.nsb_req = allocation_req;
         agm_req.required_agcs = layer_config_in_features_count[9:4] + (|layer_config_in_features_count[3:0] ? 1'b1 : '0);
 
-        // Mask of allocated cores, size NUM_CORES = (AGGREGATION_ROWS * AGGREGATION_COLUMNS)
-        // Used later for deallocating AGCs when AGM is finished so they become available for next AGM
-        // Not needed here since allocation is static
         agm_req.allocated_cores = '0;
     end
 
     for (genvar allocation_slot = 0; allocation_slot < age_pkg::MAX_AGC_PER_NODE; allocation_slot++) begin
         always_comb begin
             agm_req.coords_x [allocation_slot] = 1'b1 + allocation_slot;
             agm_req.coords_y [allocation_slot] = (allocation_req.nodeslot % NUM_MANAGERS);
+
+            //Final AGC may take less than 16 features
+            if (allocation_slot == (agm_req.required_agcs-1) && |layer_config_in_features_count[3:0]) begin
+                agm_req.num_features [allocation_slot] <= layer_config_in_features_count[3:0];
+            end
+            else begin
+                agm_req.num_features [allocation_slot] <= 16;
+            end
         end
     end
+
 
 end
 

hw/ip/aggregation_engine/rtl/aggregation_core_allocator_sequential_rr.sv

@@ -1,7 +1,7 @@
 import age_pkg::*;
 import noc_pkg::*;
 
-module aggregation_core_allocator #(
+module aggregation_core_allocator_sequential_rr #(
     parameter NUM_CORES            = top_pkg::TRANSFORMATION_CHANNELS * top_pkg::AGGREGATION_CHANNELS,
     parameter NUM_MANAGERS         = top_pkg::TRANSFORMATION_CHANNELS,
     parameter AGGREGATION_COLUMNS  = top_pkg::AGGREGATION_CHANNELS
@@ -66,7 +66,6 @@ always_comb begin
     // Static AGM req payloads
     agm_req.nsb_req = allocation_req_q;
     agm_req.required_agcs = layer_config_in_features_count[9:4] + (|layer_config_in_features_count[3:0] ? 1'b1 : '0);
-
     allocation_req_ready = !busy;
     done = (agc_counter == agm_req.required_agcs);
 end
@@ -102,10 +101,17 @@ for (genvar allocation_slot = 0; allocation_slot < age_pkg::MAX_AGC_PER_NODE; al
         if (!resetn) begin
             agm_req.coords_x [allocation_slot] <= '0;
             agm_req.coords_y [allocation_slot] <= '0;
-
+            agm_req.num_features [allocation_slot] <= 0;
         end else if (busy && !done && (agc_counter == allocation_slot)) begin
             agm_req.coords_x [allocation_slot] <= 1'b1 + (allocated_core_bin % AGGREGATION_COLUMNS); // TO DO: replace with AGGREGATION_COLS when merging with generalized aggregation mesh
             agm_req.coords_y [allocation_slot] <= allocated_core_bin / AGGREGATION_COLUMNS;
+
+            agm_req.num_features [allocation_slot] <= 16;
+
+            //Final AGC may take less than 16 features
+            if (allocation_slot == (agm_req.required_agcs-1) && |layer_config_in_features_count[3:0]) begin
+                agm_req.num_features [allocation_slot] <= layer_config_in_features_count[3:0];
+            end
         end
     end
 end