Ara integration in Cheshire (1)

Bender.lock

@@ -14,6 +14,17 @@ packages:
     dependencies:
     - apb
     - register_interface
+  ara:
+    revision: 3dad93de70c6bb20c4c0b20780d96d4243f94136
+    version: null
+    source:
+      Git: https://github.com/pulp-platform/ara.git
+    dependencies:
+    - apb
+    - axi
+    - common_cells
+    - cva6
+    - tech_cells_generic
   axi:
     revision: ac5deb3ff086aa34b168f392c051e92603d6c0e2
     version: 0.39.2
@@ -92,7 +103,7 @@ packages:
       Git: https://github.com/pulp-platform/common_verification.git
     dependencies: []
   cva6:
-    revision: 9338c2ca7cf1a47aef54322f89ce867825c3c8d5
+    revision: 99ae53bde1a94b90c1d9bbbe7fe272a9336200a6
     version: null
     source:
       Git: https://github.com/pulp-platform/cva6.git

Bender.yml

@@ -22,7 +22,8 @@ dependencies:
   clint:                    { git: "https://github.com/pulp-platform/clint.git",                  version: 0.2.0  }
   common_cells:             { git: "https://github.com/pulp-platform/common_cells.git",           version: 1.33.0 }
   common_verification:      { git: "https://github.com/pulp-platform/common_verification.git",    version: 0.2.0  }
-  cva6:                     { git: "https://github.com/pulp-platform/cva6.git",                   rev: pulp-v1.0.0 }
+  cva6:                     { git: "https://github.com/pulp-platform/cva6.git",                   rev: pulp-v1    }
+  ara:                      { git: "https://github.com/pulp-platform/ara.git",                    rev: mp/cva6-pulpv1/rebase }
   iDMA:                     { git: "https://github.com/pulp-platform/iDMA.git",                   version: 0.5.1  }
   irq_router:               { git: "https://github.com/pulp-platform/irq_router.git",             version: 0.0.1-beta.1 }
   opentitan_peripherals:    { git: "https://github.com/pulp-platform/opentitan_peripherals.git",  version: 0.4.0  }

cheshire.mk

@@ -58,7 +58,7 @@ chs-clean-deps:
 ######################
 
 CHS_NONFREE_REMOTE ?= [email protected]:pulp-restricted/cheshire-nonfree.git
-CHS_NONFREE_COMMIT ?= f731b17
+CHS_NONFREE_COMMIT ?= dc0a4e4c08d185829e7e7934fb458ed429cd10ec
 
 CHS_PHONY += chs-nonfree-init
 chs-nonfree-init:

hw/cheshire_pkg.sv

@@ -138,6 +138,7 @@ package cheshire_pkg;
     bit     Clic;
     bit     IrqRouter;
     bit     BusErr;
+    bit     Ara;
     // Parameters for Debug Module
     jtag_idcode_t DbgIdCode;
     dw_bt   DbgMaxReqs;
@@ -196,6 +197,9 @@ package cheshire_pkg;
     aw_bt   AxiRtNumAddrRegions;
     bit     AxiRtCutPaths;
     bit     AxiRtEnableChecks;
+    // Parameters for Ara
+    byte_bt AraNrLanes;
+    word_bt AraVLEN;
   } cheshire_cfg_t;
 
   //////////////////
@@ -295,6 +299,7 @@ package cheshire_pkg;
   typedef struct packed {
     aw_bt [2**MaxCoresWidth-1:0] cores;
     aw_bt dbg;
+    aw_bt ara;
     aw_bt dma;
     aw_bt slink;
     aw_bt vga;
@@ -308,6 +313,7 @@ package cheshire_pkg;
     int unsigned i = 0;
     for (int j = 0; j < cfg.NumCores; j++) begin ret.cores[i] = i; i++; end
     ret.dbg = i;
+    if (cfg.Ara)        begin i++; ret.ara   = i; end
     if (cfg.Dma)        begin i++; ret.dma   = i; end
     if (cfg.SerialLink) begin i++; ret.slink = i; end
     if (cfg.Vga)        begin i++; ret.vga   = i; end
@@ -499,9 +505,9 @@ package cheshire_pkg;
       XF8ALT                : 0,
       RVA                   : 1,
       RVB                   : 0,
-      RVV                   : 0,
+      RVV                   : cfg.Ara,
       RVC                   : 1,
-      RVH                   : 1,
+      RVH                   : ~cfg.Ara,
       RVZCB                 : 1,
       XFVec                 : 0,
       CvxifEn               : 0,
@@ -612,6 +618,7 @@ package cheshire_pkg;
     Clic              : 0,
     IrqRouter         : 0,
     BusErr            : 1,
+    Ara               : 0,
     // Debug
     DbgIdCode         : CheshireIdCode,
     DbgMaxReqs        : 4,
@@ -669,6 +676,9 @@ package cheshire_pkg;
     AxiRtWBufferDepth   : 16,
     AxiRtNumAddrRegions : 2,
     AxiRtCutPaths       : 1,
+    // Ara
+    AraNrLanes          : 2,
+    AraVLEN             : 2048,
     // All non-set values should be zero
     default: '0
   };

hw/cheshire_soc.sv

@@ -556,6 +556,26 @@ module cheshire_soc import cheshire_pkg::*; #(
 
   // TODO: Implement X interface support
 
+  // Accelerator ports
+  acc_pkg::accelerator_req_t  acc_req;
+  acc_pkg::accelerator_resp_t acc_resp;
+
+  // CVA6-Ara memory consistency
+  logic                     acc_cons_en;
+  logic [Cfg.AddrWidth-1:0] inval_addr;
+  logic                     inval_valid;
+  logic                     inval_ready;
+
+  // Pack invalidation interface into acc interface
+  acc_pkg::accelerator_resp_t acc_resp_pack;
+  always_comb begin : pack_inval
+    acc_resp_pack             = acc_resp;
+    acc_resp_pack.inval_valid = inval_valid;
+    acc_resp_pack.inval_addr  = inval_addr;
+    inval_ready               = acc_req.inval_ready;
+    acc_cons_en               = acc_req.acc_cons_en;
+  end
+
   `CHESHIRE_TYPEDEF_AXI_CT(axi_cva6, addr_t, cva6_id_t, axi_data_t, axi_strb_t, axi_user_t)
 
   localparam config_pkg::cva6_cfg_t Cva6Cfg = gen_cva6_cfg(Cfg);
@@ -606,6 +626,8 @@ module cheshire_soc import cheshire_pkg::*; #(
       .axi_w_chan_t   ( axi_cva6_w_chan_t  ),
       .b_chan_t       ( axi_cva6_b_chan_t  ),
       .r_chan_t       ( axi_cva6_r_chan_t  ),
+      .cvxif_req_t    ( acc_pkg::accelerator_req_t  ),
+      .cvxif_resp_t   ( acc_pkg::accelerator_resp_t ),
       .noc_req_t      ( axi_cva6_req_t ),
       .noc_resp_t     ( axi_cva6_rsp_t )
     ) i_core_cva6 (
@@ -626,8 +648,8 @@ module cheshire_soc import cheshire_pkg::*; #(
       .clic_kill_req_i  ( clic_irq_kill_req ),
       .clic_kill_ack_o  ( clic_irq_kill_ack ),
       .rvfi_probes_o    ( ),
-      .cvxif_req_o      ( ),
-      .cvxif_resp_i     ( '0 ),
+      .cvxif_req_o      ( acc_req       ),
+      .cvxif_resp_i     ( acc_resp_pack ),
       .noc_req_o        ( core_out_req ),
       .noc_resp_i       ( core_out_rsp )
     );
@@ -747,6 +769,105 @@ module cheshire_soc import cheshire_pkg::*; #(
       .mst_req_o  ( axi_in_req[AxiIn.cores[i]] ),
       .mst_resp_i ( axi_in_rsp[AxiIn.cores[i]] )
     );
+
+    // Generate Ara RVV vector processor if enabled
+    if (Cfg.Ara) begin : gen_ara
+      // Configure Ara with the right AXI id width
+      typedef logic [Cfg.AxiMstIdWidth-1:0] ara_id_t;
+      // Default Ara AXI data width
+      localparam int unsigned AraDataWideWidth = 32 * Cfg.AraNrLanes;
+      typedef logic [AraDataWideWidth   -1 : 0] axi_ara_wide_data_t;
+      typedef logic [AraDataWideWidth/8 -1 : 0] axi_ara_wide_strb_t;
+      `AXI_TYPEDEF_ALL(
+        axi_ara_wide, addr_t, ara_id_t, axi_ara_wide_data_t, axi_ara_wide_strb_t, axi_user_t)
+      axi_ara_wide_req_t  axi_ara_wide_req_inval, axi_ara_wide_req;
+      axi_ara_wide_resp_t axi_ara_wide_resp_inval, axi_ara_wide_resp;
+
+      axi_mst_req_t axi_ara_narrow_req;
+      axi_mst_rsp_t axi_ara_narrow_resp;
+
+      ara #(
+        .NrLanes      ( Cfg.AraNrLanes         ),
+        .VLEN         ( Cfg.AraVLEN            ),
+        .AxiDataWidth ( AraDataWideWidth       ),
+        .AxiAddrWidth ( Cfg.AddrWidth          ),
+        .axi_ar_t     ( axi_ara_wide_ar_chan_t ),
+        .axi_r_t      ( axi_ara_wide_r_chan_t  ),
+        .axi_aw_t     ( axi_ara_wide_aw_chan_t ),
+        .axi_w_t      ( axi_ara_wide_w_chan_t  ),
+        .axi_b_t      ( axi_ara_wide_b_chan_t  ),
+        .axi_req_t    ( axi_ara_wide_req_t     ),
+        .axi_resp_t   ( axi_ara_wide_resp_t    )
+      ) i_ara (
+        .clk_i           ( clk_i             ),
+        .rst_ni          ( rst_ni            ),
+        .scan_enable_i   ( 1'b0              ),
+        .scan_data_i     ( 1'b0              ),
+        .scan_data_o     ( /* Unused */      ),
+        .acc_req_i       ( acc_req           ),
+        .acc_resp_o      ( acc_resp          ),
+        .axi_req_o       ( axi_ara_wide_req  ),
+        .axi_resp_i      ( axi_ara_wide_resp )
+      );
+
+      // Issue invalidations to CVA6 L1D$
+      axi_inval_filter #(
+        .MaxTxns    ( 4                               ),
+        .AddrWidth  ( Cfg.AddrWidth                   ),
+        .L1LineWidth( ariane_pkg::DCACHE_LINE_WIDTH/8 ),
+        .aw_chan_t  ( axi_ara_wide_aw_chan_t          ),
+        .req_t      ( axi_ara_wide_req_t              ),
+        .resp_t     ( axi_ara_wide_resp_t             )
+      ) i_ara_axi_inval_filter (
+        .clk_i        ( clk_i                   ),
+        .rst_ni       ( rst_ni                  ),
+        .en_i         ( acc_cons_en             ),
+        .slv_req_i    ( axi_ara_wide_req        ),
+        .slv_resp_o   ( axi_ara_wide_resp       ),
+        .mst_req_o    ( axi_ara_wide_req_inval  ),
+        .mst_resp_i   ( axi_ara_wide_resp_inval ),
+        .inval_addr_o ( inval_addr              ),
+        .inval_valid_o( inval_valid             ),
+        .inval_ready_i( inval_ready             )
+      );
+
+      // Convert from AraDataWideWidth (axi_ara_wide) to Cfg.AxiDataWidth (axi_ara_narrow)
+      axi_dw_converter #(
+        .AxiSlvPortDataWidth ( AraDataWideWidth       ),
+        .AxiMstPortDataWidth ( Cfg.AxiDataWidth       ),
+        .AxiMaxReads         ( 4                      ), // TODO: Tune this w.r.t. ARA_NR_LANES
+        .AxiAddrWidth        ( Cfg.AddrWidth          ),
+        .AxiIdWidth          ( Cfg.AxiMstIdWidth      ),
+        .aw_chan_t           ( axi_ara_wide_aw_chan_t ),
+        .mst_w_chan_t        ( axi_mst_w_chan_t       ),
+        .slv_w_chan_t        ( axi_ara_wide_w_chan_t  ),
+        .b_chan_t            ( axi_ara_wide_b_chan_t  ),
+        .ar_chan_t           ( axi_ara_wide_ar_chan_t ),
+        .mst_r_chan_t        ( axi_mst_r_chan_t       ),
+        .slv_r_chan_t        ( axi_ara_wide_r_chan_t  ),
+        .axi_mst_req_t       ( axi_mst_req_t          ),
+        .axi_mst_resp_t      ( axi_mst_rsp_t          ),
+        .axi_slv_req_t       ( axi_ara_wide_req_t     ),
+        .axi_slv_resp_t      ( axi_ara_wide_resp_t    )
+      ) i_ara_axi_dw_converter (
+        .clk_i      ( clk_i                   ),
+        .rst_ni     ( rst_ni                  ),
+        .slv_req_i  ( axi_ara_wide_req_inval  ),
+        .slv_resp_o ( axi_ara_wide_resp_inval ),
+        .mst_req_o  ( axi_ara_narrow_req      ),
+        .mst_resp_i ( axi_ara_narrow_resp     )
+      );
+
+      // Assign to crossbar input/master
+      assign axi_in_req[AxiIn.ara] = axi_ara_narrow_req;
+      assign axi_ara_narrow_resp   = axi_in_rsp[AxiIn.ara];
+
+    end else begin : gen_no_ara
+      // Tie-to-safe the Ara-related signals
+      assign acc_resp              = '0;
+      assign inval_valid           = '0;
+      assign inval_addr            = '0;
+    end
   end
 
   /////////////////////////
@@ -1729,4 +1850,7 @@ module cheshire_soc import cheshire_pkg::*; #(
   // TODO: many other things I most likely forgot
   // TODO: check that LLC only exists if its output is connected (the reverse is allowed)
 
+  if (Cfg.Ara && (NumIntHarts > 1))
+    $error("Ara is only compatible with a single-core architecture.");
+
 endmodule

target/sim/src/tb_cheshire_pkg.sv

@@ -16,11 +16,21 @@ package tb_cheshire_pkg;
       return ret;
     endfunction
 
+    // A dedicated Ara config
+    function automatic cheshire_cfg_t gen_cheshire_ara_cfg();
+      cheshire_cfg_t ret = DefaultCfg;
+      ret.Ara        = 1;
+      ret.AraNrLanes = 2;
+      ret.AraVLEN    = 2048;
+      return ret;
+    endfunction
+
     // Number of Cheshire configurations
-    localparam int unsigned NumCheshireConfigs = 32'd2;
+    localparam int unsigned NumCheshireConfigs = 32'd3;
 
     // Assemble a configuration array indexed by a numeric parameter
     localparam cheshire_cfg_t [NumCheshireConfigs-1:0] TbCheshireConfigs = {
+        gen_cheshire_ara_cfg(),  // 2: Ara-enabled configuration
         gen_cheshire_rt_cfg(),   // 1: RT-enabled configuration
         DefaultCfg               // 0: Default configuration
     };