القائمة
Comprehensive guidelines for writing production-quality RTL code in Verilog and SystemVerilog for hardware design
التثبيت باستخدام Codex أو Claude انسخ هذا Prompt والصقه في Codex أو Claude أو مساعد آخر ليراجع صفحة Skill ويثبّتها لك.
استنادا إلى تصنيف SOC المهني
| name | rtl-coding |
| description | Comprehensive guidelines for writing production-quality RTL code in Verilog and SystemVerilog for hardware design |
This skill provides comprehensive guidelines for writing production-quality RTL code in SystemVerilog.
Purpose: Comprehensive guidelines for writing SystemVerilog RTL code based on production-quality implementations from high-performance graphics processing unit design.
Audience: Hardware engineers, RTL designers, and AI agents performing RTL design and implementation.
Scope: Applicable to all digital design projects including processors, accelerators, peripherals, and SoC components.
Note for AI Agents: This document provides structured patterns and examples for generating RTL code. Each section includes rationale, implementation patterns, and best practices. Use these patterns as templates, adapting signal names and parameters to match the specific design context.
Use this skill when you need to:
A comprehensive file header serves multiple critical purposes in professional RTL development. It provides legal protection, version control integration, and essential metadata for project management. The header establishes ownership, tracks changes over time, and helps developers understand the module's purpose and history.
The standardized header format ensures consistency across all project files and integrates seamlessly with version control systems like Perforce or Git. The RCS (Revision Control System) tags are automatically updated by the version control system, providing real-time tracking of file modifications.
//------------------------------------------------------------------------------
// COMPANY Proprietary
// Copyright (c) YEAR, COMPANY Incorporated. All rights reserved.
//
// All data and information contained in or disclosed by this document are
// confidential and proprietary information of COMPANY Incorporated, and
// all rights therein are expressly reserved. By accepting this material,
// the recipient agrees that this material and the information contained
// therein are held in confidence and in trust and will not be used,
// copied, reproduced in whole or in part, nor its contents revealed in
// any manner to others without the express written permission of COMPANY
// Incorporated.
//
// This technology was exported from the United States in accordance with
// the Export Administration Regulations. Diversion contrary to U.S. law
// prohibited.
// ------------------------------------------------------------------------------
// RCS File : $Source: /path/to/file $
// Revision : $Revision: x.xx.x.xx $
// Id : $Id: filename.sv,v x.xx.x.xx date time author Exp $
// ------------------------------------------------------------------------------
// Block : module_name
// Description : Brief description of module functionality
//
//R- Author : Name (email@domain.com)
//R- Created : Date
// ------------------------------------------------------------------------------
//R- Revision Log
//R- Who When What
//R==============================================================================
//-------------------------------------------------------------------------------
Consistent file naming is crucial for large-scale projects where hundreds or thousands of RTL files must be organized and maintained. The naming convention should immediately convey the file's purpose, hierarchical position, and functional area within the design.
The hierarchical naming approach helps developers quickly locate files and understand the design structure. Using lowercase with underscores ensures compatibility across different operating systems and tools, while avoiding potential issues with case-sensitive file systems.
Naming Rules:
gfx_<blk>_cluster_wrap.svgfx_<blk>p_prefetch_dbg.sv.sv for SystemVerilog files_core, _wrap, _ctrl, _dbgModule naming and declaration style form the foundation of readable and maintainable RTL code. A well-structured module declaration immediately communicates the module's purpose, dependencies, and interface to other developers. The naming convention should reflect the design hierarchy and functional relationships between modules.
The systematic approach to module naming enables automated tools to understand design structure and facilitates script-based operations across the design database. Consistent naming also helps with debugging, as signal names in waveforms and synthesis reports directly correspond to the hierarchical structure.
The hierarchical naming convention serves multiple purposes: it indicates the module's position in the design hierarchy, its functional area, and its specific role within that area. This approach scales effectively from small designs to complex SoCs with thousands of modules.
Naming Guidelines:
gfx_<blk>_cluster_wrapgfx_<blk>p_, gfx_<blk>__wrap, _core, _ctrl, _dbgThe module declaration format establishes a clear structure that separates concerns: package imports define the type system, parameters configure the module, and ports define the interface. This separation makes the module easier to understand and modify.
Package imports should be placed immediately after the module name to establish the type context before parameters and ports are declared. This ordering ensures that all custom types are available for use in parameter and port declarations.
module gfx_<blk>_cluster_wrap
import gfx_common_datatype_pkg::*;
import gfx_<blk>_datatype_pkg::*;
import gfx_usp_datatype_pkg::*;
import gfx_<blk>_internal_pkg::*;
#(
parameter BLK_CL_ID = 0,
parameter DV_RAND_DLY = 0
)
(
// Clock & Reset
input ares,
input clk,
// Interface Groups (organized by function)
gfx_rbbm_cgc_p2s_if.rx rbbm_<blk>_cgc,
gfx_slice_info_if.slice slice_id,
// Status outputs
output logic [NUM_CB_PIPES-1:0] <blk>_cluster_busy,
output logic [NUM_CB_PIPES-1:0] <blk>_cluster_active
);
Parameters and localparams are fundamental to creating scalable, configurable RTL designs. They enable design reuse across different configurations and facilitate maintenance by centralizing configuration values. Proper parameter usage distinguishes between externally configurable values (parameters) and internally calculated constants (localparams).
The distinction between parameters and localparams is crucial for design intent and tool optimization. Parameters can be overridden during instantiation, making modules configurable, while localparams are computed internally and cannot be modified externally. This separation ensures design integrity while providing necessary flexibility.
Parameters should be used sparingly and only for values that genuinely need external configuration. Each parameter should have a clear purpose and sensible default value. Localparams handle all derived calculations and internal constants, keeping the parameter interface clean and focused.
The parameter declaration style should clearly separate external configuration from internal calculations. This organization makes it immediately obvious which values can be modified during instantiation and which are computed internally.
// Module parameters (configurable from outside)
parameter BLK_CL_ID = 0,
parameter DV_RAND_DLY = 0
// Local parameters (internal calculations)
localparam RBUF_DEPTH = BLK_BUF_DEPTH/2;
localparam CGC_ID_BLK_CLUSTER = get_cgc_id(CGC_P2S_BLK_<BLK>, (BLK_CL_ID + 1));
localparam NUM_BLK_CLUSTER = NUM_BLK_CLUSTERS;
Consistent naming conventions for parameters and localparams improve code readability and reduce errors. The naming should immediately convey the parameter's purpose and data type. Using ALL_CAPS distinguishes parameters from variables and signals.
NUM_BLK_CLUSTERS not N_BLKADDR_W, DATA_W_W for width, _DEPTH for depthOrganizing localparams into logical categories improves maintainability and helps developers understand the design structure. Each category should group related constants that serve similar purposes within the design.
// RAM/Memory configuration
localparam BLK_GL_MASTER_RAM_FLOP_SEL = 0;
localparam BLK_IDX_FIFO_MASTER_RAM_FLOP_SEL = 0;
localparam BLK_RBUF_MASTER_RAM_FLOP_SEL = 0;
// Pipeline configuration
localparam PIPELINE_EN = gfx_common_datatype_pkg::HPGPU_2_5_GHZ_EN;
localparam PIPELINE_MODE__NO_STAGE = 0;
localparam PIPELINE_MODE__STAGE = 1;
localparam PIPELINE_MODE__DCPL = 2;
// Width and depth calculations
localparam FETCH_SRAM_WIDTH = 49;
localparam DEST_SRAM_WIDTH = 12;
localparam IAB_DEPTH = 36;
localparam IAB_ADDR_BIT = $clog2(IAB_DEPTH);
// Performance and timing parameters
localparam MAX_OUTSTANDING_REQUESTS = 16;
localparam TIMEOUT_CYCLES = 1000;
localparam RETRY_LIMIT = 3;
Package organization is critical for large-scale designs with complex type systems. Packages should be organized hierarchically with clear dependencies and consistent naming. The USP datatype implementation demonstrates sophisticated package organization with conditional compilation and architecture-specific parameters.
Packages should be organized by functionality and dependency hierarchy. Common types should be in base packages, while block-specific types should be in dedicated packages. Architecture-specific parameters should be handled through conditional compilation or parameter overrides.
package gfx_<blk>_datatype_pkg;
// Include architecture-specific defines
`include "<blk>_top_defines.vh"
// Import dependencies in order
import gfx_common_datatype_pkg::*;
import gfx_hlsq_datatype_pkg::*;
// Architecture-specific parameters with conditional logic
localparam int NUM_CLUSTER_PER_SP = 2;
localparam int GFX_<BLK>_VS_IF_QUAD = (GFX_ARCH == 2) ? 1 : ((GFX_ARCH == 4) ? 1 : 2);
localparam int GFX_<BLK>_FS_IF_QUAD = (GFX_ARCH == 2) ? 2 : ((GFX_ARCH == 4) ? 4 : 8);
localparam int GFX_<BLK>_WAVE_SIZE = (GFX_ARCH == 2) ? 32 : ((GFX_ARCH == 4) ? 64 : 128);
// Shared constants with descriptive names
localparam int STCHE_CLI_WIDTH = 6;
localparam int STCHE_CLI_NULL = 6'd32;
localparam int STCHE_CLI_OOB = 6'd33;
localparam int STCHE_CLI_2D = 6'd34;
// Enumeration definitions with comprehensive coverage
typedef enum logic [5:0] {
UCOORD = 0,
VCOORD = 1,
RCOORD = 2,
QCOORD = 3,
OFFSET_X = 4,
OFFSET_Y = 5,
OFFSET_Z = 6,
MISC = 7,
LIGHTZ = 8,
LODCLAMP = 9,
DUX = 10,
DUY = 11,
DVX = 12,
DVY = 13,
DRX = 14,
DRY = 15,
RAY_BVH_NODE = 16,
RAY_ORIGIN_X = 17,
RAY_ORIGIN_Y = 18,
RAY_ORIGIN_Z = 19,
PCMN_A_SCALE = 20,
PCMN_F_SCALE = 21,
SAD_A_WX_WY = 22,
SAD_R_WX_WY = 23,
SRC_FBID = 24,
FB_DATA = 25,
WTEX_U = 26,
WTEX_V = 27,
WTEX_F = 28,
WTEX_B = 29,
TENSOR_X = 30,
TENSOR_Y = 31,
TENSOR_F = 32,
TENSOR_B = 33,
PACKED_SAMPLE_C0 = 34,
PACKED_SAMPLE_C1 = 35
} gfx_<blk>_attr_id_e;
localparam NUM_ATTR = 36; // Document total count for validation
// Complex union structures with opaque overlay
typedef union packed {
logic [3:0][31:0] isamLOD;
logic [3:0][31:0] isammSampleID;
logic [3:0][31:0] sambLODbias;
logic [3:0][31:0] samlLOD;
logic [3:0][31:0] convmSampleID;
logic [3:0][31:0] getsizeMipLevel;
logic [3:0][31:0] getPosSampleID;
logic [127:0] opaque; // Always include opaque for full structure access
} gfx_<blk>_attr_misc_u;
// Nested structure definitions
typedef struct packed {
logic [127:32] rsvd;
logic [31:0] value;
} gfx_<blk>_attr_derivitive_s;
typedef struct packed {
gfx_<blk>_attr_derivitive_s quad;
} gfx_<blk>_attr_du_dv_dr_s;
// Complex parameterized unions
typedef union packed {
gfx_<blk>_attr_coord_s [GFX_<BLK>_NUM_QUAD-1:0] coord_U;
gfx_<blk>_attr_coord_s [GFX_<BLK>_NUM_QUAD-1:0] coord_V;
gfx_<blk>_attr_coord_s [GFX_<BLK>_NUM_QUAD-1:0] coord_R;
gfx_<blk>_attr_coord_s [GFX_<BLK>_NUM_QUAD-1:0] coord_Q;
gfx_<blk>_attr_quad_data_u [GFX_<BLK>_NUM_QUAD-1:0] offset_X;
gfx_<blk>_attr_quad_data_u [GFX_<BLK>_NUM_QUAD-1:0] offset_Y;
gfx_<blk>_attr_quad_data_u [GFX_<BLK>_NUM_QUAD-1:0] offset_Z;
gfx_<blk>_attr_misc_u [GFX_<BLK>_NUM_QUAD-1:0] misc;
logic [(GFX_<BLK>_NUM_QUAD*128)-1:0] opaque;
} gfx_<blk>_attr_u;
// Utility functions for type conversion and validation
function automatic logic is_<blk>_tag;
input gfx_<blk>_tag_e tag;
is_<blk>_tag = tag inside { [TAG_SP0:TAG_SP2] };
endfunction
function automatic gfx_<blk>_tag_e convert_to_<blk>_tag;
input int index;
convert_to_<blk>_tag = gfx_<blk>_tag_e'( index + TAG_BASE );
endfunction
endpackage: gfx_<blk>_datatype_pkg
For complex, repetitive type definitions, use macros to ensure consistency and reduce maintenance overhead. This is particularly useful for interface definitions and parameterized structures.
// Macro for consistent interface definitions
`define TYPEDEF_ATTR_T \
typedef union packed { \
coord_s [NUM_QUAD-1:0] coord_U; \
coord_s [NUM_QUAD-1:0] coord_V; \
coord_s [NUM_QUAD-1:0] coord_R; \
coord_s [NUM_QUAD-1:0] coord_Q; \
quad_data_u [NUM_QUAD-1:0] offset_X; \
quad_data_u [NUM_QUAD-1:0] offset_Y; \
quad_data_u [NUM_QUAD-1:0] offset_Z; \
misc_u [NUM_QUAD-1:0] misc; \
logic [(NUM_QUAD*128)-1:0] opaque; \
} gfx_<blk>_attr_t;
// Usage in interfaces
`include "example_defns.sv"
interface example_if #(NUM_QUAD=4) (input clk);
`TYPEDEF_ATTR_T
logic srdy;
logic rrdy;
gfx_<blk>_attr_t attr_bus;
modport sp (output srdy, attr_bus, input rrdy);
modport tp (input srdy, attr_bus, output rrdy);
endinterface
Use conditional compilation to support multiple architectures while maintaining code clarity. Document architecture-specific behavior clearly.
// Architecture-specific configuration
`ifdef GLYMUR_STAR_HC_MEM
localparam HC_MEM = 1;
`else
localparam HC_MEM = 0;
`endif
`ifdef GLYMUR_REG_ARRAY_CGC_DIS
localparam REG_ARRAY_CGC_EN = 0;
`else
localparam REG_ARRAY_CGC_EN = 1;
`endif
// Library-specific parameters
`ifdef GFX_<BLK>_LIB_7FF_KONA
`ifdef RF_MEM_DP
localparam int <BLK>_GPR_S_ACC_WIDTH = 10;
`else
localparam int <BLK>_GPR_S_ACC_WIDTH = 10;
`endif
`elsif GFX_<BLK>_LIB_14LPE
localparam int <BLK>_GPR_S_ACC_WIDTH = 8;
`else
localparam int <BLK>_GPR_S_ACC_WIDTH = 14;
`endif
Package imports should follow a strict hierarchy to avoid circular dependencies. Use makefile dependencies to enforce proper compilation order.
Import Order:
Makefile Dependencies:
# makefile.dep - Document package dependencies
gfx_common_datatype
gfx_hlsq_datatype
gfx_tp_ram_pkg
gfx_tp_datatype
Package Import Best Practices:
Data type definitions are the foundation of SystemVerilog's type system and enable creation of complex, self-documenting designs. Proper type organization improves code readability, reduces errors, and facilitates design reuse. The type system should be hierarchical, with common types in shared packages and specialized types in block-specific packages.
Structures should be packed for synthesis and clearly document bit allocation. Field ordering should follow a logical pattern, typically with control fields first, followed by data fields, and status fields last. Always include reserved fields for future expansion in critical structures.
typedef struct packed {
// Control fields (MSB)
logic valid;
logic last;
logic [2:0] cmd_type;
logic [1:0] priority;
// Address and data fields
logic [ADDR_WIDTH-1:0] addr;
logic [DATA_WIDTH-1:0] data;
logic [DATA_WIDTH/8-1:0] byte_enable;
// Identification fields
logic [TXN_ID_WIDTH-1:0] transaction_id;
logic [SRC_ID_WIDTH-1:0] source_id;
// Status and control (LSB)
logic error;
logic timeout;
logic [1:0] reserved; // For future expansion
} memory_request_s;
// Complex nested structure example
typedef struct packed {
logic [7:0] reg_id;
logic [3:0] write_mask;
logic [VFD_ATTR_DATA_W-1:0] data;
logic shared;
logic attr_valid;
logic last_attr;
logic no_more_iab;
} decoder_packed_attr_s;
Unions provide multiple views of the same data and are essential for protocol conversion and data interpretation. Always include an opaque member for full-width access and debugging. Document the purpose of each union member clearly.
typedef union packed {
// Structured access for different instruction types
struct packed {
logic [31:24] opcode;
logic [23:16] dest_reg;
logic [15:8] src1_reg;
logic [7:0] src2_reg;
} r_type;
struct packed {
logic [31:24] opcode;
logic [23:16] dest_reg;
logic [15:8] src_reg;
logic [7:0] immediate;
} i_type;
struct packed {
logic [31:24] opcode;
logic [23:0] jump_addr;
} j_type;
// Raw access for debugging and protocol conversion
logic [31:0] raw_instruction;
logic [3:0][7:0] byte_array;
// Always include opaque for full structure access
logic [31:0] opaque;
} instruction_format_u;
// Complex union with parameterized arrays (from USP analysis)
typedef union packed {
coordinate_s [NUM_QUAD-1:0] coord_U;
coordinate_s [NUM_QUAD-1:0] coord_V;
coordinate_s [NUM_QUAD-1:0] coord_R;
coordinate_s [NUM_QUAD-1:0] coord_Q;
quad_data_u [NUM_QUAD-1:0] offset_X;
quad_data_u [NUM_QUAD-1:0] offset_Y;
quad_data_u [NUM_QUAD-1:0] offset_Z;
misc_data_u [NUM_QUAD-1:0] misc;
logic [(NUM_QUAD*128)-1:0] opaque; // Full-width access
} graphics_attribute_u;
Enumerations should use explicit values for hardware interfaces and include comprehensive coverage of all valid states. Use appropriate bit widths and consider one-hot encoding for performance-critical state machines.
// Standard binary encoding for simple state machines
typedef enum logic [2:0] {
IDLE = 3'b000,
FETCH = 3'b001,
DECODE = 3'b010,
EXECUTE = 3'b011,
WRITEBACK = 3'b100,
ERROR_RECOVERY = 3'b101
// Reserve 3'b110, 3'b111 for future states
} processor_state_e;
// One-hot encoding for performance-critical paths
typedef enum logic [7:0] {
RESET_STATE = 8'b00000001,
INIT_STATE = 8'b00000010,
ACTIVE_STATE = 8'b00000100,
WAIT_STATE = 8'b00001000,
FLUSH_STATE = 8'b00010000,
ERROR_STATE = 8'b00100000,
RECOVERY_STATE = 8'b01000000,
SHUTDOWN_STATE = 8'b10000000
} cache_controller_state_e;
// Hardware interface enumeration with explicit values
typedef enum logic [5:0] {
CMD_NOP = 6'h00,
CMD_READ = 6'h01,
CMD_WRITE = 6'h02,
CMD_READ_MODIFY = 6'h03,
CMD_FLUSH = 6'h10,
CMD_INVALIDATE = 6'h11,
CMD_PREFETCH = 6'h20,
CMD_BARRIER = 6'h30,
CMD_ATOMIC_ADD = 6'h31,
CMD_ATOMIC_CAS = 6'h32,
CMD_DEBUG_READ = 6'h3E,
CMD_DEBUG_WRITE = 6'h3F
} memory_command_e;
// Document total count for validation and array sizing
localparam NUM_MEMORY_COMMANDS = 12;
For complex designs, use advanced SystemVerilog features like parameterized types, type definitions with functions, and hierarchical type organization.
// Parameterized structure for reusable components
typedef struct packed {
logic valid;
logic ready;
logic [ADDR_W-1:0] address;
logic [DATA_W-1:0] data;
logic [DATA_W/8-1:0] strobe;
logic [ID_W-1:0] id;
} generic_bus_transaction_s #(
parameter ADDR_W = 32,
parameter DATA_W = 64,
parameter ID_W = 8
);
// Type definitions with utility functions
typedef logic [COORD_WIDTH-1:0] coordinate_t;
function automatic coordinate_t add_coordinates(
input coordinate_t a,
input coordinate_t b
);
add_coordinates = a + b;
endfunction
function automatic logic coordinates_equal(
input coordinate_t a,
input coordinate_t b
);
coordinates_equal = (a == b);
endfunction
// Hierarchical type organization
typedef struct packed {
logic [15:0] x;
logic [15:0] y;
logic [15:0] z;
logic valid;
} vertex_3d_s;
typedef struct packed {
vertex_3d_s [2:0] vertices; // Triangle vertices
logic [7:0] material_id;
logic [3:0] flags;
logic visible;
} triangle_s;
typedef struct packed {
triangle_s [MAX_TRIANGLES-1:0] triangles;
logic [15:0] triangle_count;
logic [31:0] mesh_id;
} mesh_s;
Consistent naming conventions are essential for maintainable code. Follow these rules strictly across all type definitions:
Naming Rules:
_s suffix (e.g., memory_request_s)_u suffix (e.g., instruction_format_u)_e suffix (e.g., processor_state_e)_t suffix (e.g., coordinate_t)_if suffix (e.g., memory_bus_if)Organization Guidelines:
Include compile-time and runtime validation for critical types to catch errors early and ensure design integrity.
// Compile-time validation using static assertions
typedef struct packed {
logic [ADDR_WIDTH-1:0] addr;
logic [DATA_WIDTH-1:0] data;
logic [ID_WIDTH-1:0] id;
} validated_transaction_s;
// Static assertions for type validation
initial begin
assert (ADDR_WIDTH >= 16) else $fatal("Address width too small");
assert (DATA_WIDTH % 8 == 0) else $fatal("Data width must be byte-aligned");
assert ($bits(validated_transaction_s) <= MAX_TRANSACTION_BITS)
else $fatal("Transaction structure too large");
end
// Runtime validation functions
function automatic logic is_valid_command(memory_command_e cmd);
case (cmd)
CMD_NOP, CMD_READ, CMD_WRITE, CMD_READ_MODIFY,
CMD_FLUSH, CMD_INVALIDATE, CMD_PREFETCH, CMD_BARRIER,
CMD_ATOMIC_ADD, CMD_ATOMIC_CAS, CMD_DEBUG_READ, CMD_DEBUG_WRITE:
is_valid_command = 1'b1;
default:
is_valid_command = 1'b0;
endcase
endfunction
// Use SystemVerilog interfaces for complex protocols
gfx_rl_if.target cp_vfdp_rl[NUM_CB_PIPES],
gfx_rbbm_cgc_p2s_if.rx rbbm_cgc_p2s_slice,
gfx_pc_s_vfd_fiber_if.vfd pc_vfdp_fiber[NUM_GPC_SLICE_INTF],
// Use modports to specify direction
gfx_vfd_cche_req_if.vfd vfd_cche_req[NUM_VFD],
gfx_cche_vfd_rdata_if.vfd cche_vfd_rdata[NUM_VFD],
// Use consistent array notation
input logic [NUM_CB_PIPES-1:0] vfd_busy,
output logic [NUM_SP_PER_SLICE-1:0] vfd_hlsq_br_workload_busy,
// For interface arrays, specify range clearly
gfx_vfd_hlsq_info_if.vfd vfd_hlsq_info[NUM_VFD_CLUSTERS * NUM_CB_PIPES],
vfd_cluster_busyvfd_cluster_cgw_wake_up_srdy, _rrdy for ready signals_valid for valid signals_en for enable signals_sel for select signals_cnt for counters_ptr for pointers_busy, _active for status_rt for retimed signals// Group related signals with consistent naming
logic [NUM_CB_PIPES-1:0] vfd_busy;
logic [NUM_CB_PIPES-1:0] vfd_active;
logic [NUM_CB_PIPES-1:0] vfd_error;
// Use arrays for multi-dimensional signals
logic [NUM_CB_PIPES-1:0][NUM_VFD_CLUSTERS-1:0] vfd_cluster_cb_busy;
logic [NUM_VFD_CLUSTERS-1:0][NUM_CB_PIPES-1:0] vfd_cluster_busy;
// Use descriptive suffixes for internal signals
logic [NUM_CB_PIPES-1:0] pc_vfd_pass1_wave_done_rt; // retimed
logic [NUM_CB_PIPES-1:0] vfd_ctxt_done_decr_t; // temporary
logic [NUM_CB_PIPES-1:0] vfd_ctxt_done_decr_o; // output
logic [NUM_CB_PIPES-1:0] vfd_cluster_ctxt_done_int; // internal
generate
for (genvar g0=0; g0<NUM_VFD_CLUSTERS; g0++) begin: VFD_CLUSTERS
gfx_vfd_cluster_wrap #(
.VFD_CL_ID ( g0 ),
.DV_RAND_DLY ( DV_RAND_DLY )
) u_vfd_cluster (
.clk ( clk ),
.ares ( vfd_ares[g0] ),
.vfd_cluster_busy ( vfd_cluster_busy[g0] ),
.vfd_cluster_active( vfd_cluster_active[g0] )
);
end
endgenerate
g0, h0, i0 (not i, j, k)VFD_CLUSTERS, PER_CB, PER_COREgenerate
if (NUM_CB_PIPES == 2) begin: CB_BUSY
assign vfd_hlsq_bv_workload_busy[h0] = vfd_hlsq_workload_busy[1][h0];
end
else begin: BV_TIE
assign vfd_hlsq_bv_workload_busy[h0] = 1'b0;
end
endgenerate
Clock and reset design is fundamental to reliable digital systems. Proper clock distribution, reset synchronization, and power management directly impact system performance, power consumption, and reliability. Clock gating is essential for power optimization in modern designs, while reset strategies ensure predictable system initialization and recovery.
Clock and reset signals should follow consistent naming conventions and be properly documented. Multiple clock domains require careful consideration of clock domain crossing (CDC) issues. Reset hierarchy should match the design hierarchy to enable selective reset of subsystems.
// Standard clock and reset naming
input clk, // Primary system clock
input ares, // Asynchronous reset (active high)
input <blk>p_ares, // Module-specific async reset
input <blk>_ares[NUM_CLUSTERS], // Per-cluster reset array
// Clock domain specific signals
input clk_core, // Core logic clock (gated)
input clk_mem, // Memory interface clock
input clk_debug, // Debug interface clock (always-on)
// Reset synchronizers for different domains
logic core_reset_sync;
logic mem_reset_sync;
Clock gating is critical for power reduction in high-performance designs. The clock gating controller should provide configurable wake-up conditions, hysteresis to prevent excessive gating/ungating, and debug visibility. Proper clock gating can reduce dynamic power by 20-50% in typical designs.
// Clock gating controller with comprehensive configuration
gfx_cgctrl_leaf #(
.NUM_WAKE_UP ( CGC_SIG_NUM_<BLK> ), // Number of wake-up signals
.CLK_ALIVE_BIT ( NUM_CLK_ALIVE_<BLK> ), // Always-on clock bits
.CGC_ID ( CGC_ID_<BLK> ), // Unique CGC identifier
.CGC_MODE ( CGC_MODE_<BLK> ), // Gating mode (aggressive/conservative)
.CGC_DELAY ( CGC_DELAY_<BLK> ), // Gating delay cycles
.CGC_HYST ( CGC_HYST_<BLK> ), // Hysteresis cycles
.CGC_FORCE_ON ( CGC_FORCE_ON_<BLK> ) // Force-on for debug
) u_cgctrl (
.ares ( ares ),
.clk ( clk ),
.clk_core ( clk_core ), // Gated output clock
.clk_alive ( clk_alive ), // Always-on clock
.cgw_cgctrl_wake_up ( cgw_cgctrl_wake_up ), // Wake-up condition signals
.cgw_cgctrl_keep_on ( cgw_cgctrl_keep_on ), // Keep-on condition signals
.cgc_status ( cgc_status ), // Gating status for debug
.cgc_override ( cgc_debug_override ) // Debug override control
);
// Wake-up signal generation
assign cgw_cgctrl_wake_up = {
fifo_not_empty, // Data available for processing
interface_request_pending, // External request pending
state_machine_active, // State machine not idle
error_condition, // Error requires attention
debug_access_active // Debug access in progress
};
// Keep-on signal generation
assign cgw_cgctrl_keep_on = {
critical_operation_active, // Critical operation in progress
memory_transaction_pending,// Memory access pending
reset_sequence_active, // Reset/initialization sequence
performance_monitor_active // Performance monitoring enabled
};
Reset synchronization prevents metastability issues while ensuring proper reset release timing. Reset distribution should be hierarchical and match the clock distribution to minimize skew and ensure reliable operation.
// Reset synchronizer for core domain
always_ff @(posedge clk_core or posedge ares) begin
if (ares) begin
core_reset_sync_ff1 <= 1'b1;
core_reset_sync_ff2 <= 1'b1;
core_reset_sync <= 1'b1;
end else begin
core_reset_sync_ff1 <= 1'b0;
core_reset_sync_ff2 <= core_reset_sync_ff1;
core_reset_sync <= core_reset_sync_ff2;
end
end
// Reset distribution tree with proper buffering
generate
for (genvar i = 0; i < NUM_CLUSTERS; i++) begin: RESET_DISTRIBUTION
// Local reset buffer for each cluster
logic cluster_reset_local;
always_ff @(posedge clk or posedge ares) begin
if (ares) begin
cluster_reset_local <= 1'b1;
end else begin
cluster_reset_local <= core_reset_sync | cluster_specific_reset[i];
end
end
assign <blk>_ares[i] = cluster_reset_local;
end
endgenerate
Power management should be integrated with clock gating to provide comprehensive power optimization. This includes voltage scaling coordination, retention register control, and power domain isolation.
// Power management controller interface
typedef struct packed {
logic power_down_req; // Power down request
logic retention_enable; // Enable retention registers
logic isolation_enable; // Enable power domain isolation
logic [3:0] voltage_level; // Target voltage level
logic dvfs_transition; // Dynamic voltage/frequency scaling
} power_mgmt_ctrl_s;
// Power state machine
typedef enum logic [2:0] {
PWR_ACTIVE = 3'b000,
PWR_IDLE = 3'b001,
PWR_DROWSY = 3'b010,
PWR_RETENTION = 3'b011,
PWR_OFF = 3'b100
} power_state_e;
power_state_e current_power_state, next_power_state;
// Power state transitions with proper sequencing
always_ff @(posedge clk or posedge ares) begin
if (ares) begin
current_power_state <= PWR_ACTIVE;
power_transition_counter <= '0;
end else begin
current_power_state <= next_power_state;
if (current_power_state != next_power_state) begin
power_transition_counter <= power_transition_counter + 1;
end
end
end
// Comprehensive sequential logic template with error handling
always_ff @ (posedge clk or posedge ares) begin
if (ares) begin
// Reset all registers to known states
signal <= '0;
counter <= '0;
state_machine <= IDLE;
error_flags <= '0;
transaction_id <= '0;
end else begin
// Default assignments to prevent latches
error_flags.timeout <= 1'b0;
error_flags.overflow <= 1'b0;
if (enable) begin
signal <= next_signal;
counter <= counter + 1;
// Timeout detection
if (counter > TIMEOUT_THRESHOLD) begin
error_flags.timeout <= 1'b1;
counter <= '0;
end
// Overflow protection
if (counter == MAX_COUNT) begin
error_flags.overflow <= 1'b1;
counter <= '0;
end
end
// State machine with error handling
case (state_machine)
IDLE: begin
if (start_request) begin
state_machine <= ACTIVE;
transaction_id <= transaction_id + 1;
end
end
ACTIVE: begin
if (completion_detected) begin
state_machine <= IDLE;
end else if (error_detected) begin
state_machine <= ERROR_RECOVERY;
end
end
ERROR_RECOVERY: begin
if (recovery_complete) begin
state_machine <= IDLE;
error_flags <= '0;
end
end
default: begin
state_machine <= IDLE; // Safe fallback
error_flags.invalid_state <= 1'b1;
end
endcase
end
end
// Use always_comb for combinational logic
always_comb begin
pc_vfdp_fiber_srdy[g0] = pc_vfdp_fiber[g0].srdy;
pc_vfd_pass1_wave_done_srdy[g0] = pc_vfd_pass1_wave_done[g0].srdy;
vfd_pc_us_dealloc_srdy[g0] = (g0==0) ? '0 : vfd_pc_us_dealloc.srdy;
end
// Use assign for simple assignments
assign vfd_error[g0] = '0;
assign vfd_cluster_error = '0;
always_comb begin
vfd_attr_miss_count = '0;
for(int i=0; i<NUM_VFD_CLUSTERS; i++)
vfd_attr_miss_count = (vfd_attr_miss_count + ({'0, vfd_attr_dealloc_count[i]}));
end
<=) in clocked blocks=) in combinational blocks'0 for zero initialization of any widthDebug infrastructure and performance monitoring are critical for silicon bring-up, performance optimization, and field debugging. A well-designed debug system provides visibility into internal operations without significantly impacting area, power, or timing. The debug infrastructure should be planned early in the design phase and integrated systematically throughout the design hierarchy.
Performance counters enable quantitative analysis of design behavior, bottleneck identification, and performance optimization. They should capture key metrics that directly relate to system performance and provide actionable insights for both hardware and software teams.
The debug bus provides real-time visibility into internal signals and state machines. Signal selection should focus on critical control paths, state transitions, and interface handshakes. The debug bus width is typically constrained, so signal prioritization is essential.
Debug Signal Selection Criteria:
// Debug bus signal naming convention
logic [63:0] GFX_DBGBUS_misr_data_<blk>_fetch;
logic [63:0] GFX_DBGBUS_misr_data_<blk>_decode;
logic [3:0] GFX_DBGBUS_misr_side_<blk>_fetch;
logic [3:0] GFX_DBGBUS_misr_side_<blk>_decode;
// Debug bus assignments with detailed bit allocation
// DEBBUS_DEFINE_BEGIN_BV top_block=<blk>; my_block=fetch_dbg; radix=10;
assign GFX_DBGBUS_misr_data_<blk>_fetch[0] = fetch_state_machine.current_state[0];
assign GFX_DBGBUS_misr_data_<blk>_fetch[3:1] = fetch_state_machine.current_state[3:1];
assign GFX_DBGBUS_misr_data_<blk>_fetch[4] = fetch_fifo_full;
assign GFX_DBGBUS_misr_data_<blk>_fetch[5] = fetch_fifo_empty;
assign GFX_DBGBUS_misr_data_<blk>_fetch[13:6] = fetch_fifo_occupancy[7:0];
assign GFX_DBGBUS_misr_data_<blk>_fetch[14] = interface_srdy;
assign GFX_DBGBUS_misr_data_<blk>_fetch[15] = interface_rrdy;
assign GFX_DBGBUS_misr_data_<blk>_fetch[16] = error_detected;
assign GFX_DBGBUS_misr_data_<blk>_fetch[17] = timeout_occurred;
assign GFX_DBGBUS_misr_data_<blk>_fetch[25:18] = transaction_id[7:0];
assign GFX_DBGBUS_misr_data_<blk>_fetch[26] = clock_gated;
assign GFX_DBGBUS_misr_data_<blk>_fetch[27] = power_down_req;
assign GFX_DBGBUS_misr_data_<blk>_fetch[63:28] = debug_timestamp[35:0];
// Side channel for transaction tracking
assign GFX_DBGBUS_misr_side_<blk>_fetch[0] = interface_srdy & interface_rrdy; // Valid transaction
assign GFX_DBGBUS_misr_side_<blk>_fetch[1] = interface_srdy; // Request pending
assign GFX_DBGBUS_misr_side_<blk>_fetch[2] = error_detected; // Error condition
assign GFX_DBGBUS_misr_side_<blk>_fetch[3] = state_transition; // State change
// DEBBUS_DEFINE_END_BV
Performance counters should be organized hierarchically to provide both high-level system metrics and detailed subsystem analysis. Counter width should be sufficient to avoid overflow during typical measurement periods, typically 32 or 64 bits.
Performance Counter Categories:
typedef struct packed {
// Utilization metrics
logic [31:0] pfc_busy_cycles;
logic [31:0] pfc_active_cycles;
logic [31:0] pfc_idle_cycles;
// Stall analysis
logic [31:0] pfc_stall_cycles_input_fifo_full;
logic [31:0] pfc_stall_cycles_output_fifo_full;
logic [31:0] pfc_stall_cycles_memory_conflict;
logic [31:0] pfc_stall_cycles_resource_unavail;
// Throughput measurements
logic [31:0] pfc_transactions_completed;
logic [31:0] pfc_bytes_processed;
logic [31:0] pfc_operations_per_cycle;
// Latency tracking
logic [31:0] pfc_avg_latency_cycles;
logic [31:0] pfc_max_latency_cycles;
logic [31:0] pfc_min_latency_cycles;
// Error monitoring
logic [15:0] pfc_error_count;
logic [15:0] pfc_timeout_count;
logic [15:0] pfc_retry_count;
// Power and clock gating
logic [31:0] pfc_clock_gated_cycles;
logic [31:0] pfc_power_down_cycles;
logic [7:0] pfc_cgc_efficiency_percent;
} block_performance_counters_s;
// Performance counter implementation with enable control
always_ff @(posedge clk or posedge ares) begin
if (ares) begin
perf_counters <= '0;
end else if (perf_counter_enable) begin
// Increment counters based on conditions
if (block_busy)
perf_counters.pfc_busy_cycles <= perf_counters.pfc_busy_cycles + 1;
if (input_stall)
perf_counters.pfc_stall_cycles_input_fifo_full <= perf_counters.pfc_stall_cycles_input_fifo_full + 1;
if (transaction_complete)
perf_counters.pfc_transactions_completed <= perf_counters.pfc_transactions_completed + 1;
if (error_detected)
perf_counters.pfc_error_count <= perf_counters.pfc_error_count + 1;
end
end
Trigger and Capture Logic:
// Debug trigger conditions
logic debug_trigger_condition;
logic [7:0] debug_trigger_count;
always_comb begin
debug_trigger_condition = (state_machine == ERROR_STATE) |
(fifo_overflow_detected) |
(timeout_counter > TIMEOUT_THRESHOLD) |
(performance_degradation_detected);
end
// Circular buffer for debug data capture
logic [DEBUG_BUFFER_DEPTH-1:0][DEBUG_DATA_WIDTH-1:0] debug_capture_buffer;
logic [DEBUG_ADDR_WIDTH-1:0] debug_write_ptr;
always_ff @(posedge clk or posedge ares) begin
if (ares) begin
debug_write_ptr <= '0;
debug_capture_buffer <= '0;
end else if (debug_capture_enable) begin
debug_capture_buffer[debug_write_ptr] <= {
current_timestamp,
state_machine_state,
interface_status,
error_flags,
performance_indicators
};
debug_write_ptr <= debug_write_ptr + 1;
end
end
//==============================================================================
// Clock Gating Controller
//==============================================================================
//============================================================================
// VFD CLUSTERS
//============================================================================
//==============================================================================
// Import Packages
//==============================================================================
// TODO: Use NUM_MICRO_SP_PER_SP ??
if (NUM_SPTP_PER_SP == 2) begin : PC_PASS1_WAVE_DONE
// Handle dual SPTP case
assign pc_vfd_pass1_wave_done_rt[g0].rrdy = pc_vfd_pass1_wave_done_rt[g0].payload.sptp_id[1] ?
pc_dqsync_pass1_done[g0+NUM_CB_PIPES].rrdy :
pc_dqsync_pass1_done[g0].rrdy;
end
// Retiming buffer for performance
gfx_retiming_fifo #(
.DSIZE ( $bits(pc_vfd_pass1_wave_done_s) ),
.NUM_INST ( GFX_GPC_VFD_PC_PASS1_BUF_NUM[g0] )
) u_pc_vfd_pass_1_done_retiming_buf (
// connections...
);
// Function: get_cgc_id
// Purpose: Calculate clock gating controller ID based on block and instance
// Parameters:
// - blk_id: Block identifier
// - inst_id: Instance identifier within block
// Returns: Calculated CGC ID
//synopsys translate_off
// Simulation-only code
logic [NUM_VFD_PER_SP*NUM_VFD_CLUSTERS -1 :0] vfd_uche_req_state_id_lsb;
logic [NUM_VFD_PER_SP*NUM_VFD_CLUSTERS -1 :0][VFD_CTXTID_W-1:0] vfd_uche_req_ctxt_id;
//synopsys translate_on
`ifdef VFD_DV_RAND_DLY_FIFO
localparam DV_RAND_DLY = 1;
`else
localparam DV_RAND_DLY = 0;
`endif
`ifdef GFX_VFD_WBOX_COV
`include "gfx_vfd_top_wb.sv"
`endif
//synopsys translate_off
`ifdef WBOX_COV_EN
`include "gfx_vfd_cluster_wb.sv"
`endif
//synopsys translate_on
FIFOs are fundamental building blocks in digital systems, providing data buffering, clock domain crossing, and flow control. Proper FIFO design ensures data integrity, optimal performance, and reliable operation across varying conditions.
FIFO depth should be determined based on system requirements, including burst sizes, latency tolerance, and throughput requirements. Shallow FIFOs (2-8 entries) are suitable for simple buffering, while deep FIFOs (64+ entries) handle large bursts and provide substantial buffering.
// Parameterized FIFO with comprehensive features
module gfx_<blk>_fifo #(
parameter DEPTH = 16,
parameter DATA_WIDTH = 64,
parameter ADDR_WIDTH = $clog2(DEPTH),
parameter ALMOST_FULL_THRESH = DEPTH - 4,
parameter ALMOST_EMPTY_THRESH = 4,
parameter SYNC_STAGES = 2,
parameter RAM_STYLE = "auto" // "auto", "distributed", "block"
)(
// Write interface
input wr_clk,
input wr_ares,
input wr_en,
input [DATA_WIDTH-1:0] wr_data,
output wr_full,
output wr_almost_full,
output [ADDR_WIDTH:0] wr_count,
// Read interface
input rd_clk,
input rd_ares,
input rd_en,
output [DATA_WIDTH-1:0] rd_data,
output rd_empty,
output rd_almost_empty,
output [ADDR_WIDTH:0] rd_count,
// Status and debug
output fifo_error,
output [ADDR_WIDTH:0] max_occupancy
);
Retiming FIFOs break critical timing paths while maintaining throughput. They should be strategically placed at pipeline boundaries and interface crossings to optimize timing closure.
// Retiming FIFO for timing optimization
module gfx_retiming_fifo #(
parameter DSIZE = 64,
parameter NUM_INST = 2, // Number of pipeline stages
parameter BYPASS_EN = 0 // Enable bypass for zero latency
)(
input clk,
input ares,
// Input interface
input data_in_srdy,
output data_in_rrdy,
input [DSIZE-1:0] data_in,
// Output interface
output data_out_srdy,
input data_out_rrdy,
output [DSIZE-1:0] data_out,
// Status
output busy
);
// Implementation with configurable stages
generate
if (NUM_INST == 0 || BYPASS_EN) begin: BYPASS_PATH
assign data_out_srdy = data_in_srdy;
assign data_in_rrdy = data_out_rrdy;
assign data_out = data_in;
assign busy = 1'b0;
end else begin: PIPELINE_STAGES
logic [NUM_INST:0] stage_srdy;
logic [NUM_INST:0] stage_rrdy;
logic [NUM_INST-1:0][DSIZE-1:0] stage_data;
assign stage_srdy[0] = data_in_srdy;
assign data_in_rrdy = stage_rrdy[0];
assign data_out_srdy = stage_srdy[NUM_INST];
assign stage_rrdy[NUM_INST] = data_out_rrdy;
assign data_out = stage_data[NUM_INST-1];
// Pipeline stage implementation
for (genvar i = 0; i < NUM_INST; i++) begin: STAGE
always_ff @(posedge clk or posedge ares) begin
if (ares) begin
stage_srdy[i+1] <= 1'b0;
stage_data[i] <= '0;
end else begin
if (stage_rrdy[i] || ~stage_srdy[i+1]) begin
stage_srdy[i+1] <= stage_srdy[i];
if (stage_srdy[i] && stage_rrdy[i]) begin
stage_data[i] <= (i == 0) ? data_in : stage_data[i-1];
end
end
end
end
// Ready propagation with lookahead
assign stage_rrdy[i] = ~stage_srdy[i+1] || stage_rrdy[i+1];
end
// Busy when any stage contains valid data
assign busy = |stage_srdy[NUM_INST:1];
end
endgenerate
endmodule
### Asynchronous FIFO Implementation
Asynchronous FIFOs enable reliable data transfer between different clock domains. The key challenges include metastability prevention, proper gray code pointer synchronization, and accurate full/empty detection.
```systemverilog
// Asynchronous FIFO with gray code pointers
module gfx_async_fifo #(
parameter DEPTH = 16,
parameter DATA_WIDTH = 64,
parameter ADDR_WIDTH = $clog2(DEPTH),
parameter SYNC_STAGES = 2
)(
// Write domain
input wr_clk,
input wr_ares,
input wr_en,
input [DATA_WIDTH-1:0] wr_data,
output wr_full,
// Read domain
input rd_clk,
input rd_ares,
input rd_en,
output [DATA_WIDTH-1:0] rd_data,
output rd_empty,
// Status
output overflow_error,
output underflow_error
);
// Internal registers
logic [ADDR_WIDTH:0] wr_ptr_bin;
logic [ADDR_WIDTH:0] rd_ptr_bin;
logic [ADDR_WIDTH:0] wr_ptr_gray;
logic [ADDR_WIDTH:0] rd_ptr_gray;
logic [ADDR_WIDTH:0] wr_ptr_gray_sync;
logic [ADDR_WIDTH:0] rd_ptr_gray_sync;
// Memory array
logic [DATA_WIDTH-1:0] fifo_mem [0:DEPTH-1];
// Binary to Gray code conversion
function automatic [ADDR_WIDTH:0] bin_to_gray(input [ADDR_WIDTH:0] bin);
bin_to_gray = bin ^ (bin >> 1);
endfunction
// Gray to Binary conversion
function automatic [ADDR_WIDTH:0] gray_to_bin(input [ADDR_WIDTH:0] gray);
logic [ADDR_WIDTH:0] bin;
bin = gray;
for (int i = 1; i <= ADDR_WIDTH; i = i << 1)
bin = bin ^ (bin >> i);
return bin;
endfunction
// Write pointer logic
always_ff @(posedge wr_clk or posedge wr_ares) begin
if (wr_ares) begin
wr_ptr_bin <= '0;
wr_ptr_gray <= '0;
end else if (wr_en && !wr_full) begin
wr_ptr_bin <= wr_ptr_bin + 1'b1;
wr_ptr_gray <= bin_to_gray(wr_ptr_bin + 1'b1);
end
end
// Read pointer logic
always_ff @(posedge rd_clk or posedge rd_ares) begin
if (rd_ares) begin
rd_ptr_bin <= '0;
rd_ptr_gray <= '0;
end else if (rd_en && !rd_empty) begin
rd_ptr_bin <= rd_ptr_bin + 1'b1;
rd_ptr_gray <= bin_to_gray(rd_ptr_bin + 1'b1);
end
end
// Synchronize read pointer to write clock domain
gfx_synchronizer #(
.WIDTH(ADDR_WIDTH+1),
.STAGES(SYNC_STAGES)
) rd_ptr_synchronizer (
.clk(wr_clk),
.ares(wr_ares),
.data_in(rd_ptr_gray),
.data_out(rd_ptr_gray_sync)
);
// Synchronize write pointer to read clock domain
gfx_synchronizer #(
.WIDTH(ADDR_WIDTH+1),
.STAGES(SYNC_STAGES)
) wr_ptr_synchronizer (
.clk(rd_clk),
.ares(rd_ares),
.data_in(wr_ptr_gray),
.data_out(wr_ptr_gray_sync)
);
// Full and empty generation
assign wr_full = (wr_ptr_gray[ADDR_WIDTH] != rd_ptr_gray_sync[ADDR_WIDTH]) &&
(wr_ptr_gray[ADDR_WIDTH-1] != rd_ptr_gray_sync[ADDR_WIDTH-1]) &&
(wr_ptr_gray[ADDR_WIDTH-2:0] == rd_ptr_gray_sync[ADDR_WIDTH-2:0]);
assign rd_empty = (rd_ptr_gray == wr_ptr_gray_sync);
// Memory write operation
always_ff @(posedge wr_clk) begin
if (wr_en && !wr_full) begin
fifo_mem[wr_ptr_bin[ADDR_WIDTH-1:0]] <= wr_data;
end
end
// Memory read operation
assign rd_data = fifo_mem[rd_ptr_bin[ADDR_WIDTH-1:0]];
// Error detection
logic wr_overflow, rd_underflow;
always_ff @(posedge wr_clk or posedge wr_ares) begin
if (wr_ares) begin
wr_overflow <= 1'b0;
end else begin
wr_overflow <= wr_overflow | (wr_en && wr_full);
end
end
always_ff @(posedge rd_clk or posedge rd_ares) begin
if (rd_ares) begin
rd_underflow <= 1'b0;
end else begin
rd_underflow <= rd_underflow | (rd_en && rd_empty);
end
end
assign overflow_error = wr_overflow;
assign underflow_error = rd_underflow;
endmodule
Ready/Valid (or srdy/rrdy) protocol is commonly used for flow control in high-performance designs. This FIFO implementation supports this protocol with configurable options for performance and area optimization.
// FIFO with ready/valid protocol
module gfx_srdy_rrdy_fifo #(
parameter DEPTH = 8,
parameter DATA_WIDTH = 64,
parameter ADDR_WIDTH = $clog2(DEPTH),
parameter REGISTER_OUTPUT = 1, // Register output for timing
parameter LOOKAHEAD = 1 // Enable lookahead for back-to-back transfers
)(
input clk,
input ares,
// Input interface
input in_srdy,
output in_rrdy,
input [DATA_WIDTH-1:0] in_data,
// Output interface
output out_srdy,
input out_rrdy,
output [DATA_WIDTH-1:0] out_data,
// Status
output [ADDR_WIDTH:0] occupancy,
output full,
output empty
);
// Internal registers
logic [ADDR_WIDTH:0] wr_ptr;
logic [ADDR_WIDTH:0] rd_ptr;
logic [DATA_WIDTH-1:0] fifo_mem [0:DEPTH-1];
logic output_valid;
logic [DATA_WIDTH-1:0] output_reg;
// Control signals
logic fifo_write;
logic fifo_read;
logic internal_empty;
// Write and read pointers
always_ff @(posedge clk or posedge ares) begin
if (ares) begin
wr_ptr <= '0;
rd_ptr <= '0;
end else begin
if (fifo_write)
wr_ptr <= wr_ptr + 1'b1;
if (fifo_read)
rd_ptr <= rd_ptr + 1'b1;
end
end
// Memory write operation
always_ff @(posedge clk) begin
if (fifo_write) begin
fifo_mem[wr_ptr[ADDR_WIDTH-1:0]] <= in_data;
end
end
// Occupancy calculation
assign occupancy = wr_ptr - rd_ptr;
assign full = occupancy == DEPTH;
assign internal_empty = (wr_ptr == rd_ptr);
// Input ready generation
assign in_rrdy = ~full;
assign fifo_write = in_srdy && in_rrdy;
// Output path with optional registration
generate
if (REGISTER_OUTPUT) begin: REG_OUTPUT
// Registered output path
always_ff @(posedge clk or posedge ares) begin
if (ares) begin
output_valid <= 1'b0;
output_reg <= '0;
end else begin
if (fifo_read) begin
output_valid <= 1'b1;
output_reg <= fifo_mem[rd_ptr[ADDR_WIDTH-1:0]];
end else if (out_rrdy) begin
output_valid <= 1'b0;
end
end
end
assign out_srdy = output_valid;
assign out_data = output_reg;
assign empty = internal_empty && ~output_valid;
// Read logic with lookahead
if (LOOKAHEAD) begin: LOOKAHEAD_ENABLED
assign fifo_read = ~internal_empty && (~output_valid || out_rrdy);
end else begin: STANDARD_READ
assign fifo_read = ~internal_empty && ~output_valid;
end
end else begin: DIRECT_OUTPUT
// Direct output path
assign out_srdy = ~internal_empty;
assign out_data = fifo_mem[rd_ptr[ADDR_WIDTH-1:0]];
assign empty = internal_empty;
assign fifo_read = out_srdy && out_rrdy;
end
endgenerate
endmodule
Memory interfaces are critical components in digital systems, providing access to various memory types including SRAM, register files, and external memories. Proper memory interface design ensures optimal performance, reliability, and power efficiency.
SRAM interfaces should be designed for maximum throughput while maintaining timing margins. Key considerations include address setup time, write enable timing, and read data latency. For high-performance designs, consider pipelining the interface to improve timing closure.
// SRAM interface module with configurable parameters
module gfx_sram_interface #(
parameter ADDR_WIDTH = 10,
parameter DATA_WIDTH = 64,
parameter WRITE_ENABLE_WIDTH = DATA_WIDTH/8,
parameter READ_LATENCY = 1,
parameter REGISTER_OUTPUTS = 1
)(
input clk,
input ares,
// Client interface
input req_valid,
output req_ready,
input req_write,
input [ADDR_WIDTH-1:0] req_addr,
input [DATA_WIDTH-1:0] req_wdata,
input [WRITE_ENABLE_WIDTH-1:0] req_wstrb,
output resp_valid,
output [DATA_WIDTH-1:0] resp_rdata,
// SRAM interface
output sram_cs,
output sram_we,
output [ADDR_WIDTH-1:0] sram_addr,
output [DATA_WIDTH-1:0] sram_wdata,
output [WRITE_ENABLE_WIDTH-1:0] sram_wstrb,
input [DATA_WIDTH-1:0] sram_rdata
);
// Request handling
logic req_accepted;
logic read_pending;
logic [READ_LATENCY-1:0] read_valid_shift;
// Accept requests when not busy
assign req_ready = ~read_pending || (READ_LATENCY == 1 && resp_valid && ~req_write);
assign req_accepted = req_valid && req_ready;
// SRAM control signals
assign sram_cs = req_accepted;
assign sram_we = req_accepted && req_write;
assign sram_addr = req_addr;
assign sram_wdata = req_wdata;
assign sram_wstrb = req_write ? req_wstrb : '0;
// Read response tracking
always_ff @(posedge clk or posedge ares) begin
if (ares) begin
read_valid_shift <= '0;
read_pending <= 1'b0;
end else begin
// Shift register to track read latency
read_valid_shift <= {read_valid_shift[READ_LATENCY-2:0],
req_accepted && ~req_write};
// Track pending reads
if (req_accepted && ~req_write) begin
read_pending <= 1'b1;
end else if (resp_valid) begin
read_pending <= 1'b0;
end
end
end
// Read data path
generate
if (REGISTER_OUTPUTS) begin: REG_OUTPUTS
// Registered output for timing improvement
logic [DATA_WIDTH-1:0] rdata_reg;
always_ff @(posedge clk) begin
if (read_valid_shift[READ_LATENCY-2]) begin
rdata_reg <= sram_rdata;
end
end
assign resp_valid = read_valid_shift[READ_LATENCY-1];
assign resp_rdata = rdata_reg;
end else begin: DIRECT_OUTPUTS
// Direct connection to SRAM outputs
assign resp_valid = read_valid_shift[READ_LATENCY-1];
assign resp_rdata = sram_rdata;
end
endgenerate
endmodule
Register files require multi-port access with minimal latency. The design should optimize for area efficiency while maintaining high performance. For high-performance designs, consider banking the register file to support multiple simultaneous accesses.
// Multi-port register file with configurable parameters
module gfx_register_file #(
parameter ADDR_WIDTH = 5,
parameter DATA_WIDTH = 32,
parameter NUM_READ_PORTS = 2,
parameter NUM_WRITE_PORTS = 1,
parameter DEPTH = 2**ADDR_WIDTH,
parameter RESET_REGS = 0, // Set to 1 to reset all registers
parameter FORWARDING_ENABLE = 1 // Enable write-to-read forwarding
)(
input clk,
input ares,
// Read ports
input [NUM_READ_PORTS-1:0] rd_en,
input [NUM_READ_PORTS-1:0][ADDR_WIDTH-1:0] rd_addr,
output [NUM_READ_PORTS-1:0][DATA_WIDTH-1:0] rd_data,
// Write ports
input [NUM_WRITE_PORTS-1:0] wr_en,
input [NUM_WRITE_PORTS-1:0][ADDR_WIDTH-1:0] wr_addr,
input [NUM_WRITE_PORTS-1:0][DATA_WIDTH-1:0] wr_data
);
// Register file storage
logic [DEPTH-1:0][DATA_WIDTH-1:0] registers;
// Write logic with priority resolution
always_ff @(posedge clk or posedge ares) begin
if (ares && RESET_REGS) begin
registers <= '0;
end else begin
// Process write ports in priority order (lower index has higher priority)
for (int w = 0; w < NUM_WRITE_PORTS; w++) begin
if (wr_en[w]) begin
registers[wr_addr[w]] <= wr_data[w];
end
end
end
end
// Read logic with write forwarding
generate
for (genvar r = 0; r < NUM_READ_PORTS; r++) begin: READ_PORT
if (FORWARDING_ENABLE) begin: WITH_FORWARDING
// Forwarding logic to handle read-after-write hazards
logic [DATA_WIDTH-1:0] forwarded_data;
logic forwarding_active;
always_comb begin
forwarded_data = registers[rd_addr[r]];
forwarding_active = 1'b0;
// Check all write ports for forwarding (priority to lower indices)
for (int w = NUM_WRITE_PORTS-1; w >= 0; w--) begin
if (wr_en[w] && (wr_addr[w] == rd_addr[r])) begin
forwarded_data = wr_data[w];
forwarding_active = 1'b1;
end
end
end
assign rd_data[r] = rd_en[r] ? forwarded_data : '0;
end else begin: NO_FORWARDING
// Simple read without forwarding
assign rd_data[r] = rd_en[r] ? registers[rd_addr[r]] : '0;
end
end
endgenerate
endmodule
Memory controllers manage access to external memories like DDR SDRAM. They handle complex timing requirements, reordering for efficiency, and refresh operations. The design should optimize for bandwidth utilization while maintaining latency targets.
// Memory controller with request reordering and timing management
module gfx_memory_controller #(
parameter ADDR_WIDTH = 32,
parameter DATA_WIDTH = 512,
parameter ID_WIDTH = 8,
parameter MAX_OUTSTANDING = 16,
parameter REORDER_ENABLE = 1,
parameter BANK_INTERLEAVE = 1
)(
input clk,
input ares,
// Client request interface
input req_valid,
output req_ready,
input req_write,
input [ADDR_WIDTH-1:0] req_addr,
input [DATA_WIDTH-1:0] req_wdata,
input [DATA_WIDTH/8-1:0] req_wstrb,
input [ID_WIDTH-1:0] req_id,
// Client response interface
output resp_valid,
output [DATA_WIDTH-1:0] resp_rdata,
output [ID_WIDTH-1:0] resp_id,
output resp_error,
// Memory PHY interface
output mem_cmd_valid,
input mem_cmd_ready,
output [2:0] mem_cmd_opcode,
output [ADDR_WIDTH-1:0] mem_cmd_addr,
output [7:0] mem_cmd_len,
output mem_wr_valid,
input mem_wr_ready,
output [DATA_WIDTH-1:0] mem_wr_data,
output [DATA_WIDTH/8-1:0] mem_wr_strb,
output mem_wr_last,
input mem_rd_valid,
output mem_rd_ready,
input [DATA_WIDTH-1:0] mem_rd_data,
input mem_rd_last,
input mem_rd_error
);
// Request queue and tracking
typedef struct packed {
logic valid;
logic write;
logic [ADDR_WIDTH-1:0] addr;
logic [DATA_WIDTH-1:0] wdata;
logic [DATA_WIDTH/8-1:0] wstrb;
logic [ID_WIDTH-1:0] id;
logic [3:0] bank_id;
logic scheduled;
} mem_request_s;
mem_request_s [MAX_OUTSTANDING-1:0] req_queue;
logic [$clog2(MAX_OUTSTANDING)-1:0] enq_ptr, deq_ptr;
logic [$clog2(MAX_OUTSTANDING):0] queue_count;
// Bank status tracking for reordering
logic [15:0] bank_active;
logic [15:0] bank_timer;
// Queue management
assign req_ready = (queue_count < MAX_OUTSTANDING);
// Enqueue new requests
always_ff @(posedge clk or posedge ares) begin
if (ares) begin
enq_ptr <= '0;
for (int i = 0; i < MAX_OUTSTANDING; i++) begin
req_queue[i].valid <= 1'b0;
end
end else if (req_valid && req_ready) begin
req_queue[enq_ptr].valid <= 1'b1;
req_queue[enq_ptr].write <= req_write;
req_queue[enq_ptr].addr <= req_addr;
req_queue[enq_ptr].wdata <= req_wdata;
req_queue[enq_ptr].wstrb <= req_wstrb;
req_queue[enq_ptr].id <= req_id;
req_queue[enq_ptr].bank_id <= req_addr[7:4]; // Example bank mapping
req_queue[enq_ptr].scheduled <= 1'b0;
enq_ptr <= enq_ptr + 1'b1;
end
end
// Queue count tracking
always_ff @(posedge clk or posedge ares) begin
if (ares) begin
queue_count <= '0;
end else begin
case ({req_valid && req_ready, mem_cmd_valid && mem_cmd_ready})
2'b10: queue_count <= queue_count + 1'b1;
2'b01: queue_count <= queue_count - 1'b1;
default: queue_count <= queue_count;
endcase
end
end
// Request scheduling logic with bank interleaving
logic [$clog2(MAX_OUTSTANDING)-1:0] schedule_ptr;
logic schedule_valid;
generate
if (REORDER_ENABLE && BANK_INTERLEAVE) begin: REORDER_LOGIC
// Reordering logic to optimize bank access patterns
always_comb begin
schedule_valid = 1'b0;
schedule_ptr = deq_ptr;
// First pass: look for requests to idle banks
for (int i = 0; i < MAX_OUTSTANDING; i++) begin
logic [$clog2(MAX_OUTSTANDING)-1:0] check_ptr;
check_ptr = (deq_ptr + i) % MAX_OUTSTANDING;
if (req_queue[check_ptr].valid && !req_queue[check_ptr].scheduled &&
!bank_active[req_queue[check_ptr].bank_id]) begin
schedule_valid = 1'b1;
schedule_ptr = check_ptr;
break;
end
end
// Second pass: if no idle banks, take oldest request
if (!schedule_valid) begin
for (int i = 0; i < MAX_OUTSTANDING; i++) begin
logic [$clog2(MAX_OUTSTANDING)-1:0] check_ptr;
check_ptr = (deq_ptr + i) % MAX_OUTSTANDING;
if (req_queue[check_ptr].valid && !req_queue[check_ptr].scheduled) begin
schedule_valid = 1'b1;
schedule_ptr = check_ptr;
break;
end
end
end
end
end else begin: IN_ORDER_LOGIC
// Simple in-order scheduling
assign schedule_valid = req_queue[deq_ptr].valid && !req_queue[deq_ptr].scheduled;
assign schedule_ptr = deq_ptr;
end
endgenerate
// Memory command generation
assign mem_cmd_valid = schedule_valid && !req_queue[schedule_ptr].scheduled;
assign mem_cmd_opcode = req_queue[schedule_ptr].write ? 3'b001 : 3'b000; // Write/Read
assign mem_cmd_addr = req_queue[schedule_ptr].addr;
assign mem_cmd_len = 8'd0; // Single transfer
// Mark request as scheduled
always_ff @(posedge clk) begin
if (mem_cmd_valid && mem_cmd_ready) begin
req_queue[schedule_ptr].scheduled <= 1'b1;
bank_active[req_queue[schedule_ptr].bank_id] <= 1'b1;
bank_timer[req_queue[schedule_ptr].bank_id] <= 4'd10; // Example timing
end
end
// Bank timing management
always_ff @(posedge clk or posedge ares) begin
if (ares) begin
bank_active <= '0;
bank_timer <= '0;
end else begin
for (int b = 0; b < 16; b++) begin
if (bank_active[b]) begin
if (bank_timer[b] > 0) begin
bank_timer[b] <= bank_timer[b] - 1'b1;
end else begin
bank_active[b] <= 1'b0;
end
end
end
end
end
// Write data channel
assign mem_wr_valid = req_queue[deq_ptr].valid && req_queue[deq_ptr].scheduled &&
req_queue[deq_ptr].write;
assign mem_wr_data = req_queue[deq_ptr].wdata;
assign mem_wr_strb = req_queue[deq_ptr].wstrb;
assign mem_wr_last = 1'b1; // Single transfer
// Read data handling
assign mem_rd_ready = 1'b1; // Always ready to receive read data
// Response generation
assign resp_valid = (mem_rd_valid && !req_queue[deq_ptr].write) ||
(mem_wr_valid && mem_wr_ready && req_queue[deq_ptr].write);
assign resp_rdata = mem_rd_data;
assign resp_id = req_queue[deq_ptr].id;
assign resp_error = mem_rd_error;
// Dequeue completed requests
always_ff @(posedge clk or posedge ares) begin
if (ares) begin
deq_ptr <= '0;
end else if (resp_valid) begin
req_queue[deq_ptr].valid <= 1'b0;
deq_ptr <= deq_ptr + 1'b1;
end
end
endmodule
Communication protocols enable reliable data exchange between modules, chips, and systems. Proper protocol implementation ensures interoperability, error detection, and efficient data transfer. This section covers common on-chip and chip-to-chip protocols used in digital systems.
AXI (Advanced eXtensible Interface) is a widely used on-chip communication protocol that supports high-performance, flexible data transfers. The implementation should adhere to the ARM AMBA specification while optimizing for the specific design requirements.
// AXI4 Master Interface
module gfx_axi_master #(
parameter ADDR_WIDTH = 32,
parameter DATA_WIDTH = 64,
parameter ID_WIDTH = 4,
parameter USER_WIDTH = 1,
parameter MAX_BURST_LEN = 16
)(
input clk,
input ares,
// Client interface
input req_valid,
output req_ready,
input req_write,
input [ADDR_WIDTH-1:0] req_addr,
input [7:0] req_len,
input [2:0] req_size,
input [1:0] req_burst,
input [ID_WIDTH-1:0] req_id,
input [DATA_WIDTH-1:0] req_wdata,
input [DATA_WIDTH/8-1:0] req_wstrb,
output resp_valid,
output [DATA_WIDTH-1:0] resp_rdata,
output [ID_WIDTH-1:0] resp_id,
output [1:0] resp_resp,
// AXI interface
// Write address channel
output axi_awvalid,
input axi_awready,
output [ADDR_WIDTH-1:0] axi_awaddr,
output [7:0] axi_awlen,
output [2:0] axi_awsize,
output [1:0] axi_awburst,
output [ID_WIDTH-1:0] axi_awid,
output [USER_WIDTH-1:0] axi_awuser,
// Write data channel
output axi_wvalid,
input axi_wready,
output [DATA_WIDTH-1:0] axi_wdata,
output [DATA_WIDTH/8-1:0] axi_wstrb,
output axi_wlast,
// Write response channel
input axi_bvalid,
output axi_bready,
input [ID_WIDTH-1:0] axi_bid,
input [1:0] axi_bresp,
// Read address channel
output axi_arvalid,
input axi_arready,
output [ADDR_WIDTH-1:0] axi_araddr,
output [7:0] axi_arlen,
output [2:0] axi_arsize,
output [1:0] axi_arburst,
output [ID_WIDTH-1:0] axi_arid,
output [USER_WIDTH-1:0] axi_aruser,
// Read data channel
input axi_rvalid,
output axi_rready,
input [DATA_WIDTH-1:0] axi_rdata,