path: root/python/openvino/demo/ip/intel_ai_ip/verilog/dla_aux_depthwise_top.sv

// Copyright 2020-2023 Intel Corporation.
//
// This software and the related documents are Intel copyrighted materials,
// and your use of them is governed by the express license under which they
// were provided to you ("License"). Unless the License provides otherwise,
// you may not use, modify, copy, publish, distribute, disclose or transmit
// this software or the related documents without Intel's prior written
// permission.
//
// This software and the related documents are provided as is, with no express
// or implied warranties, other than those that are expressly stated in the
// License.

/*
 * Module `dla_aux_depthwise_top`
 *
 * Top level of the aux block.
 *
 * WARNING!  DO NOT EDIT THIS FILE FOR FUTURE-COMPATIBILITY
 *
 * This module is a template for all auxiliary blocks and the content should
 * not be modified for future compatibility.
 *
 * Do not modify any of fields other than the module name and included package
 * name. In case of a limitation or bug please contact owner of the Example
 * Aux block.
 *
 * See README.md of the Example Aux block for more details.
 */

`undefineall
`resetall
`default_nettype none

module dla_aux_depthwise_top
  import dla_aux_depthwise_pkg::*;
#(
  parameter aux_data_pack_params_t AUX_DATA_PACK_PARAMS,
  parameter stream_params_t        CONFIG_STREAM_PARAMS,
  parameter debug_axi_params_t     DEBUG_AXI_PARAMS,
  parameter aux_generic_params_t   AUX_GENERIC_PARAMS,
  parameter aux_special_params_t   AUX_SPECIAL_PARAMS,
  parameter vector_dot_arch_t      AUX_DEPTHWISE_VECTOR_ARCH,
  parameter vector_dot_arch_info_t DEPTHWISE_VECTOR_ARCH_INFO,
  //
  localparam stream_params_t       DATA_STREAM_PARAMS  = '{ // Data stream parameterization
    DATA_WIDTH : aux_params_to_bus_width(AUX_DATA_PACK_PARAMS)},
  localparam int GROUP_DELAY  = AUX_DATA_PACK_PARAMS.GROUP_DELAY,
  localparam int GROUP_NUM    = AUX_DATA_PACK_PARAMS.GROUP_NUM,
  localparam int GROUP_SIZE   = AUX_DATA_PACK_PARAMS.GROUP_SIZE,
  localparam int VECTOR_SIZE  = AUX_DATA_PACK_PARAMS.VECTOR_SIZE,
  localparam int ELEMENT_BITS = AUX_DATA_PACK_PARAMS.ELEMENT_BITS
) (
  input  wire                                             clk                , // Clock
  input  wire                                             i_aresetn          , // Active-low async reset
  //
  input var logic [GROUP_NUM   -1:0]
                  [GROUP_SIZE  -1:0]
                  [VECTOR_SIZE -1:0]
                  [ELEMENT_BITS-1:0]                      i_data             , // Data input stream port
  input  var logic                                        i_data_valid       , // Data input stream port valid
  output generic_response_t                               o_data             , // Data input stream port response
  //
  input  generic_response_t                               i_result           , // Result output stream port response
  output logic [GROUP_NUM   -1:0]
               [GROUP_SIZE  -1:0]
               [VECTOR_SIZE -1:0]
               [ELEMENT_BITS-1:0]                         o_result           , // Result output stream port
  output logic                                            o_result_valid     , // Result output stream port valid
  //
  input  var logic  [CONFIG_STREAM_PARAMS.DATA_WIDTH-1:0] i_config           , // Config stream port
  input  var logic                                        i_config_valid     , // Config stream port valid
  output generic_response_t                               o_config           , // Config stream port response
  input  var logic                                        i_config_filter_bias_valid, // Config (actual data) for cache
  input  var logic  [CONFIG_STREAM_PARAMS.DATA_WIDTH-1:0] i_config_filter_bias_data,
  output var logic                                        o_config_filter_bias_ready,
  //
  input var logic       [DEBUG_AXI_PARAMS.ADDR_WIDTH-1:0] i_debug_raddr      , // Debug AXI read-address port
  input var logic                                         i_debug_raddr_valid, // Debug AXI read-address port valid
  output generic_response_t                               o_debug_raddr      , // Debug AXI read-address port response
  input  generic_response_t                               i_debug_rdata      , // Debug AXI read-data port response
  output logic          [DEBUG_AXI_PARAMS.DATA_WIDTH-1:0] o_debug_rdata      , // Debug AXI read-data port
  output logic                                            o_debug_rdata_valid  // Debug AXI read-data port valid
);

  localparam int GROUP_SHIFT = (GROUP_NUM-1)*GROUP_DELAY+1;

  // Parameter checking
  initial /* synthesis enable_verilog_initial_construct */
  begin
    // check AUX_DATA_PACK_PARAMS
    if (AUX_DATA_PACK_PARAMS.ELEMENT_BITS > 32 || AUX_DATA_PACK_PARAMS.ELEMENT_BITS <= 0) begin
      $fatal(1, {"Illegal parameterization (dla_aux_depthwise_top.sv), ELEMENT_BITS = %0d, ",
        "expected 0 < ELEMENT_BITS <= 32"}, AUX_DATA_PACK_PARAMS.ELEMENT_BITS);
    end
    if (AUX_DATA_PACK_PARAMS.VECTOR_SIZE <= 0) begin
      $fatal(1, {"Illegal parameterization (dla_aux_depthwise_top.sv), VECTOR_SIZE = %0d, ",
        "expected VECTOR_SIZE > 0"}, AUX_DATA_PACK_PARAMS.VECTOR_SIZE);
    end
    if (AUX_DATA_PACK_PARAMS.NATIVE_VECTOR_SIZE <= 0) begin
      $fatal(1, {"Illegal parameterization (dla_aux_depthwise_top.sv), NATIVE_VECTOR_SIZE = %0d, ",
        "expected NATIVE_VECTOR_SIZE > 0"}, AUX_DATA_PACK_PARAMS.NATIVE_VECTOR_SIZE);
    end
    if (AUX_DATA_PACK_PARAMS.GROUP_SIZE <= 0) begin
      $fatal(1, {"Illegal parameterization (dla_aux_depthwise_top.sv), GROUP_SIZE = %0d, ",
        "expected GROUP_SIZE > 0"}, AUX_DATA_PACK_PARAMS.GROUP_SIZE);
    end
    if (AUX_DATA_PACK_PARAMS.GROUP_NUM <= 0) begin
      $fatal(1, {"Illegal parameterization (dla_aux_depthwise_top.sv), GROUP_NUM = %0d, ",
        "expected GROUP_NUM > 0"}, AUX_DATA_PACK_PARAMS.GROUP_NUM);
    end
    if (AUX_DATA_PACK_PARAMS.GROUP_DELAY < 0) begin
      $fatal(1, {"Illegal parameterization (dla_aux_depthwise_top.sv), GROUP_DELAY = %0d, ",
        "expected GROUP_DELAY >= 0"}, AUX_DATA_PACK_PARAMS.GROUP_DELAY);
    end

    // check CONFIG_STREAM_PARAMS
    if (CONFIG_STREAM_PARAMS.DATA_WIDTH <= 0) begin
      $fatal(1, {"Illegal parameterization (dla_aux_depthwise_top.sv), CONFIG_STREAM_PARAMS.DATA_WIDTH ",
        "= %0d, expected CONFIG_STREAM_PARAMS.DATA_WIDTH > 0"}, CONFIG_STREAM_PARAMS.DATA_WIDTH);
    end

    // check DEBUG_AXI_PARAMS
    if (DEBUG_AXI_PARAMS.DATA_WIDTH <= 0) begin
      $fatal(1, {"Illegal parameterization (dla_aux_depthwise_top.sv), DEBUG_AXI_PARAMS.DATA_WIDTH = ",
        "%0d, expected DEBUG_AXI_PARAMS.DATA_WIDTH > 0"}, DEBUG_AXI_PARAMS.DATA_WIDTH);
    end
    if (DEBUG_AXI_PARAMS.ADDR_WIDTH <= 0) begin
      $fatal(1, {"Illegal parameterization (dla_aux_depthwise_top.sv), DEBUG_AXI_PARAMS.ADDR_WIDTH = ",
        "%0d, expected DEBUG_AXI_PARAMS.ADDR_WIDTH > 0"}, DEBUG_AXI_PARAMS.ADDR_WIDTH);
    end

    // check AUX_GENERIC_PARAMS
    if (AUX_GENERIC_PARAMS.INPUT_BUFFER_REG_STAGES < 0) begin
      $fatal(1, {"Illegal parameterization (dla_aux_depthwise_top.sv), AUX_GENERIC_PARAMS.",
        "INPUT_BUFFER_REG_STAGES = %0d, expected AUX_GENERIC_PARAMS.INPUT_BUFFER_REG_STAGES >= 0"},
        AUX_GENERIC_PARAMS.INPUT_BUFFER_REG_STAGES);
    end
    if (AUX_GENERIC_PARAMS.COMMAND_BUFFER_DEPTH < 0) begin
      $fatal(1, {"Illegal parameterization (dla_aux_depthwise_top.sv), AUX_GENERIC_PARAMS.",
        "COMMAND_BUFFER_DEPTH = %0d, expected AUX_GENERIC_PARAMS.COMMAND_BUFFER_DEPTH >= 0"},
        AUX_GENERIC_PARAMS.COMMAND_BUFFER_DEPTH);
    end
    if (AUX_GENERIC_PARAMS.PER_GROUP_CONTROL != 0 &&
      AUX_GENERIC_PARAMS.PER_GROUP_CONTROL != 1) begin
      $fatal(1, {"Illegal parameterization (dla_aux_depthwise_top.sv), AUX_GENERIC_PARAMS.",
        "PER_GROUP_CONTROL = %0d, expected AUX_GENERIC_PARAMS.PER_GROUP_CONTROL = 0 or 1"},
        AUX_GENERIC_PARAMS.PER_GROUP_CONTROL);
    end
    if (AUX_GENERIC_PARAMS.DEBUG_LEVEL < 0) begin
      $fatal(1, {"Illegal parameterization (dla_aux_depthwise_top.sv), AUX_GENERIC_PARAMS.DEBUG_LEVEL ",
        "= %0d, expected AUX_GENERIC_PARAMS.DEBUG_LEVEL >= 0"}, AUX_GENERIC_PARAMS.DEBUG_LEVEL);
    end
    if (AUX_GENERIC_PARAMS.DEBUG_ID < 0) begin
      $fatal(1, {"Illegal parameterization (dla_aux_depthwise_top.sv), AUX_GENERIC_PARAMS.DEBUG_ID = ",
        "%0d, expected AUX_GENERIC_PARAMS.DEBUG_ID >= 0"}, AUX_GENERIC_PARAMS.DEBUG_ID);
    end
    if (AUX_GENERIC_PARAMS.DEBUG_EVENT_DEPTH < 0) begin
      $fatal(1, {"Illegal parameterization (dla_aux_depthwise_top.sv), AUX_GENERIC_PARAMS.",
        "DEBUG_EVENT_DEPTH = %0d, expected AUX_GENERIC_PARAMS.DEBUG_EVENT_DEPTH >= 0"},
        AUX_GENERIC_PARAMS.DEBUG_EVENT_DEPTH);
    end

    // Note: AUX_SPECIAL_PARAMS must be checked in relevant files (config_decoder, control and core)
  end

  // Arch assignment
  localparam aux_depthwise_arch_params_t ARCH = '{
    AUX_DATA_PACK_PARAMS : AUX_DATA_PACK_PARAMS,
    CONFIG_STREAM_PARAMS : CONFIG_STREAM_PARAMS,
    DEBUG_AXI_PARAMS     : DEBUG_AXI_PARAMS,
    AUX_GENERIC_PARAMS   : AUX_GENERIC_PARAMS,
    AUX_SPECIAL_PARAMS   : AUX_SPECIAL_PARAMS
  };

  // Reset module constants
  localparam RST_USE_SYNCHRONIZER = 1; // yes clock domain crossings
  localparam RST_PIPE_DEPTH       = 3; // stages of synchronization registers
  localparam RST_NUM_COPIES       = 3; // number of reset signals

  // synchronized reset
  logic [RST_NUM_COPIES-1:0] sclrn;

  /*------------------------------------------------------------------------------
  --  Reset handler
  ------------------------------------------------------------------------------*/
  dla_reset_handler_simple #(
    .USE_SYNCHRONIZER(RST_USE_SYNCHRONIZER),
    .PIPE_DEPTH      (RST_PIPE_DEPTH      ),
    .NUM_COPIES      (RST_NUM_COPIES      )
  ) dla_reset_handler_simple_inst (
    .clk     (clk      ),
    .i_resetn(i_aresetn), // active-low asyncronous reset input
    .o_sclrn (sclrn    )  // one or more copies of synchronized reset, 'dont_merge' constraints applied
  );

  // Group 0 only signals
  control_to_config_t control_to_config;

  // Systolic Group signals
  // config-control interface and related signals
  depthwise_config_to_control_if #(
    .special_params  (AUX_SPECIAL_PARAMS  ),
    .data_pack_params(AUX_DATA_PACK_PARAMS)
  ) config_to_control[GROUP_SHIFT]();

  // control-lane interface
  depthwise_control_to_lane_if #(
    .special_params  (AUX_SPECIAL_PARAMS  ),
    .data_pack_params(AUX_DATA_PACK_PARAMS)
  ) control_to_lane[GROUP_SHIFT]();

  // data and result handshaking signals
  logic stream_data_valid[GROUP_SHIFT-1:0];
  generic_response_t stream_result_request [GROUP_SHIFT-1:0];

  // synchronized reset
  logic group_sreset_n [GROUP_SHIFT-1:0];

  logic [ARCH.AUX_DATA_PACK_PARAMS.VECTOR_SIZE-1:0][ARCH.AUX_SPECIAL_PARAMS.MAX_WINDOW_HEIGHT*ARCH.AUX_SPECIAL_PARAMS.MAX_WINDOW_WIDTH-1:0][ARCH.AUX_DATA_PACK_PARAMS.ELEMENT_BITS-1:0] filter_data_buf;
  logic [ARCH.AUX_DATA_PACK_PARAMS.VECTOR_SIZE-1:0][AUX_DEPTHWISE_VECTOR_ARCH.BIAS_WIDTH-1:0] bias_data_buf; // fp32 bias
  // debug connections
  debug_config_t debug_from_config;
  debug_group_t  debug_from_group ;

  // Assign group delay input signals
  assign group_sreset_n       [0]       = sclrn[0];
  assign stream_data_valid    [0]       = (i_data_valid & o_data.ready);
  assign stream_result_request[0].ready = i_result.ready;

  //
  // Group-delay chains of all group delayed signals
  //
  if (GROUP_DELAY > 0) begin : gen_non_zero_delay
    for(genvar i = 1; i < GROUP_SHIFT; i++) begin : gen_group_delay
      always_ff @(posedge clk) begin : proc_group_delay
        group_sreset_n       [i]            <= group_sreset_n       [i-1];
        stream_result_request[i]            <= stream_result_request[i-1];
        stream_data_valid    [i]            <= stream_data_valid    [i-1];
        control_to_lane      [i].data[0][0] <= control_to_lane      [i-1].data[0][0];
        config_to_control    [i].data[0][0] <= config_to_control    [i-1].data[0][0];
      end : proc_group_delay
    end : gen_group_delay
  end : gen_non_zero_delay

  //
  // Config decoder
  //
  dla_aux_depthwise_config_decoder #(
    .ARCH(ARCH)
  ) dla_aux_depthwise_config_decoder_inst (
    .clk                 ( clk                        ),
    .i_resetn            ( sclrn   [RST_NUM_COPIES-2] ),
    .i_config            ( i_config                   ),
    .i_config_valid      ( i_config_valid             ),
    .o_config            ( o_config                   ),
    .i_control_to_config ( control_to_config          ),
    .o_config_to_control ( config_to_control[0]       ),
    .o_debug             ( debug_from_config          )
  );

  //
  // Multiple groups (phases) are generated
  //
  // First group is responsible for interacting with Config Decoder and generating debug information
  dla_aux_depthwise_group #(
    .ID  (0   ),
    .ARCH(ARCH),
    .AUX_DEPTHWISE_VECTOR_ARCH(AUX_DEPTHWISE_VECTOR_ARCH),
    .DEPTHWISE_VECTOR_ARCH_INFO(DEPTHWISE_VECTOR_ARCH_INFO)
  ) dla_aux_depthwise_group_inst (
    .clk                        ( clk                              ),
    .i_resetn                   ( group_sreset_n               [0] ),
    .i_data                     ( i_data                       [0] ),
    .i_data_valid               ( stream_data_valid            [0] ),
    .o_data                     ( {>>{o_data}}                     ),
    .i_result                   ( stream_result_request        [0] ),
    .o_result                   ( o_result                     [0] ),
    .o_result_valid             ( o_result_valid                   ),
    .i_filter                   ( filter_data_buf                  ),
    .i_bias                     ( bias_data_buf                    ),
    .i_config_to_control        ( config_to_control            [0] ),
    .o_control_to_config        ( control_to_config                ),
    .i_control_to_lane          ( control_to_lane              [0] ),
    .o_control_to_lane_next     ( control_to_lane              [0] ),
    .i_config_filter_bias_valid ( i_config_filter_bias_valid       ),
    .i_config_filter_bias_data  ( i_config_filter_bias_data        ),
    .o_config_filter_bias_ready ( o_config_filter_bias_ready       ),
    .o_debug                    ( debug_from_group                 )
  );
  // Other groups only consume decoded config
  for (genvar i = 1; i < GROUP_NUM; i++) begin : gen_groups
    // Must pass an interface to o_control_lane_next, so create a placeholder interface
    depthwise_control_to_lane_if #(.special_params(AUX_SPECIAL_PARAMS), .data_pack_params(AUX_DATA_PACK_PARAMS)) output_control_to_lane();

    dla_aux_depthwise_group #(
      .ID  (i   ),
      .ARCH(ARCH),
      .AUX_DEPTHWISE_VECTOR_ARCH(AUX_DEPTHWISE_VECTOR_ARCH),
      .DEPTHWISE_VECTOR_ARCH_INFO(DEPTHWISE_VECTOR_ARCH_INFO)
    ) dla_aux_depthwise_group_inst (
      .clk                        ( clk                                    ),
      .i_resetn                   ( group_sreset_n       [i * GROUP_DELAY] ),
      .i_data                     ( i_data               [i              ] ),
      .i_data_valid               ( stream_data_valid    [i * GROUP_DELAY] ),
      .o_data                     (                                        ),
      .i_result                   ( stream_result_request[i * GROUP_DELAY] ),
      .o_result                   ( o_result             [i              ] ),
      .o_result_valid             (                                        ),
      .i_filter                   ( filter_data_buf                        ),
      .i_bias                     ( bias_data_buf                          ),
      .i_config_to_control        ( config_to_control    [i * GROUP_DELAY] ),
      .o_control_to_config        (                                        ),
      .i_control_to_lane          ( control_to_lane      [i * GROUP_DELAY] ),
      .o_control_to_lane_next     ( output_control_to_lane                 ),
      .i_config_filter_bias_valid ( i_config_filter_bias_valid             ),
      .i_config_filter_bias_data  ( i_config_filter_bias_data              ),
      .o_config_filter_bias_ready (                                        ),
      .o_debug                    (                                        )
    );
  end : gen_groups

  //
  // Filter bias cache
  //
  logic filter_cache_ready;
  logic configured_and_ready;
  initial begin
    if (o_config_filter_bias_ready != (filter_cache_ready && configured_and_ready)) begin
      $display("Double plumbed filters not matching ready %d %d %d", o_config_filter_bias_ready, filter_cache_ready, configured_and_ready);
    end
  end
  //assign o_config_filter_bias_ready = filter_cache_ready && configured_and_ready;
  assign configured_and_ready = control_to_lane[0].data[0][0].configured_starting !=
                                control_to_lane[0].data[0][0].configured_ending;

  dla_aux_depthwise_filter_bias_cache #(
    .MAX_WINDOW_HEIGHT      (ARCH.AUX_SPECIAL_PARAMS.MAX_WINDOW_HEIGHT             ),
    .MAX_WINDOW_WIDTH       (ARCH.AUX_SPECIAL_PARAMS.MAX_WINDOW_WIDTH              ),
    .ELEMENT_BITS           (ARCH.AUX_DATA_PACK_PARAMS.ELEMENT_BITS                ),
    .CONFIG_BIT_WIDTH       (32                                                    ),
    .VECTOR_SIZE            (ARCH.AUX_DATA_PACK_PARAMS.VECTOR_SIZE                 ),
    .WINDOW_BITS_VERTICAL   ($clog2(ARCH.AUX_SPECIAL_PARAMS.MAX_WINDOW_HEIGHT + 1) ),
    .WINDOW_BITS_HORIZONTAL ($clog2(ARCH.AUX_SPECIAL_PARAMS.MAX_WINDOW_WIDTH + 1)  ),
    .BIAS_WIDTH             (AUX_DEPTHWISE_VECTOR_ARCH.BIAS_WIDTH                  )
  ) filter_bias_cache (
    .clk                        ( clk                                            ),
    .i_resetn                   ( group_sreset_n      [0]                        ),
    .i_data_valid               ( i_config_filter_bias_valid                     ),
    .i_data                     ( i_config_filter_bias_data                      ),
    .o_config_filter_bias_ready ( filter_cache_ready                             ),
    .i_configured_and_ready     ( configured_and_ready                           ),
    .i_done                     ( control_to_lane[GROUP_SHIFT-1].data[0][0].done ), // use the last lane's data
    .i_window_width             ( control_to_lane[0].data[0][0].window_width     ),
    .i_window_height            ( control_to_lane[0].data[0][0].window_height    ),
    .o_filter_valid             (                                                ),
    .o_filter                   ( filter_data_buf                                ),
    .o_bias                     ( bias_data_buf                                  )
  );

  //
  // Debug
  //
  dla_aux_depthwise_debug #(
    .ARCH(ARCH)
  ) dla_aux_depthwise_debug_inst (
    .clk     (clk                     ),
    .i_resetn(sclrn[RST_NUM_COPIES-1] ),
    //
    .i_config(debug_from_config       ),
    .i_group (debug_from_group        ),
    //
    .i_raddr (i_debug_raddr           ),
    .i_raddr_valid(i_debug_raddr_valid),
    .o_raddr (o_debug_raddr           ),
    .i_rdata (i_debug_rdata           ),
    .o_rdata (o_debug_rdata           ),
    .o_rdata_valid(o_debug_rdata_valid)
  );

endmodule