dv-flow-manager

分类: 内容与多媒体 | 上传者: dv-flow | 下载: 0 | 版本: v1.0（最新）
Create and modify DV Flow Manager (dfm) YAML-based build flows for silicon design and verification projects. Use when working with flow.yaml, flow.dv files, or dfm commands.
更新日志: Source: GitHub https://github.com/dv-flow/dv-flow-mgr
目录结构

当前层级: src/dv_flow/mgr/share/
📁 prompts/
- 📄 copilot_system.md 2.7 KB
- 📄 general_agent.md 4.2 KB
📁 references/
- 📄 advanced.md 8.4 KB
- 📄 cli.md 7.7 KB
- 📄 concepts.md 7.8 KB
- 📄 stdlib.md 8.1 KB
- 📄 task_development.md 23.0 KB
- 📄 tasks.md 9.1 KB
📄 dv.flow.schema.json 26.2 KB
📄 skill.md 12.8 KB
SKILL.md

---
name: dv-flow-manager
description: Create and modify DV Flow Manager (dfm) YAML-based build flows for silicon design and verification projects. Use when working with flow.yaml, flow.dv files, or dfm commands.
---

# DV Flow Manager (dfm)

DV Flow Manager is a YAML-based build system and execution engine designed for silicon design and verification projects. It orchestrates tasks through declarative workflows with dataflow-based dependency management.

## When to Use This Skill

Use this skill when:
- Creating or modifying `flow.yaml` or `flow.dv` files
- Writing task definitions for HDL compilation, simulation, or verification
- Configuring dataflow between tasks using `needs`, `consumes`, `produces`, `passthrough`
- Discovering tasks by their outputs using `produces` patterns
- Validating dataflow compatibility between producer and consumer tasks
- Setting up package parameters and configurations
- Running `dfm` commands (run, show, graph, validate)
- Debugging build flow issues
- Working with standard library tasks (std.FileSet, std.Message, etc.)
- Executing shell commands with `shell: bash` and `run:`
- **Executing dfm commands from within an LLM-driven Agent task**

## Quick Reference

### Minimal Flow Example

```yaml
package:
  name: my_project
  
  tasks:
    - name: rtl_files
      uses: std.FileSet
      with:
        type: systemVerilogSource
        include: "*.sv"
      produces:
        - type: std.FileSet
          filetype: systemVerilogSource
    
    - name: sim
      uses: hdlsim.vlt.SimImage
      needs: [rtl_files]
      consumes:
        - type: std.FileSet
          filetype: systemVerilogSource
      with:
        top: [my_top]
```

### Key Commands

```bash
# Run commands
dfm run [tasks...]          # Execute tasks
dfm run -j 4                # Run with 4 parallel jobs
dfm run --clean             # Clean rebuild
dfm run -c debug            # Use 'debug' configuration
dfm run -D param=value      # Override parameter

# Discovery commands (for humans and Agents)
dfm show packages           # List all available packages
dfm show packages --json    # JSON output for Agents
dfm show tasks              # List all visible tasks
dfm show tasks --search kw  # Search tasks by keyword
dfm show task std.FileSet   # Show detailed task info (includes produces)
dfm show types              # List data types and tags
dfm show project            # Show current project structure
dfm context --json          # Get full project context (for Agents)

# Task discovery by outputs (uses produces)
dfm show tasks --produces "type=std.FileSet,filetype=verilog"
dfm show tasks --produces "type=std.FileSet" --json

# Visualization
dfm graph task -o flow.dot  # Generate dependency graph

# Validation
dfm validate                # Validate flow configuration
dfm validate --json         # JSON output for programmatic use
```

### Expression Syntax

Use `${{ }}` for dynamic parameter evaluation:

```yaml
msg: "Building version ${{ version }}"
iff: ${{ debug_level > 0 }}
command: ${{ "make debug" if debug else "make release" }}
```

## LLM Call Interface (Running Inside Agent Tasks)

When running inside an LLM-driven `std.Agent` task, the `dfm` command automatically
connects to the parent DFM session via a Unix socket. This enables LLMs to:

- Execute tasks that share resources with the parent session
- Query project state and task information
- Validate configurations before execution

### Environment Detection

When `DFM_SERVER_SOCKET` environment variable is set, `dfm` runs in client mode:

```bash
# These commands work inside an Agent task:
dfm run task1 task2         # Execute tasks via parent session
dfm show tasks              # Query available tasks
dfm context --json          # Get project context
dfm validate                # Validate configuration
dfm ping                    # Health check
```

### Running Tasks from Within a Prompt

When an LLM needs to compile or simulate code it generated:

```bash
# 1. Create RTL files
cat > counter.sv << 'EOF'
module counter(input clk, rst_n, output logic [7:0] count);
  always_ff @(posedge clk or negedge rst_n)
    if (!rst_n) count <= 0;
    else count <= count + 1;
endmodule
EOF

# 2. Run compilation via parent DFM session
dfm run hdlsim.vlt.SimImage -D hdlsim.vlt.SimImage.top=counter

# 3. Check result (JSON output)
# Returns: {"status": 0, "outputs": [...], "markers": []}
```

### Querying Project State

```bash
# Get full project context
dfm context --json
# Returns:
# {
#   "project": {"name": "my_project", "root_dir": "/path/to/project"},
#   "tasks": [{"name": "my_project.build", "scope": "root", ...}],
#   "types": [...],
#   "skills": [...]
# }

# Get specific task details
dfm show task my_project.build --json
# Returns detailed task information including parameters and dependencies
```

### Benefits of Server Mode

1. **Resource Sharing**: Respects parent session's parallelism limits (`-j`)
2. **State Consistency**: Sees outputs from tasks already completed
3. **Cache Sharing**: Uses same memento cache for incremental builds
4. **Unified Logging**: All task output appears in parent session's logs

## Using Produces/Consumes for Task Discovery (AI Assistants)

When helping users build workflows, use produces/consumes to identify compatible tasks:

### Finding Tasks by Output Type

```bash
# Find all tasks that produce verilog files
dfm show tasks --produces "type=std.FileSet,filetype=verilog" --json

# Find tasks that produce any FileSet
dfm show tasks --produces "type=std.FileSet" --json

# Find tasks with specific output characteristics
dfm show tasks --produces "type=std.FileSet,filetype=verilog,stage=compiled"
```

### Understanding Task Relationships

When a task needs specific inputs, find compatible producers:

```bash
# 1. User wants to simulate - what tasks produce simulation inputs?
dfm show tasks --produces "type=std.FileSet,filetype=verilog" --json

# 2. Check what a specific task produces
dfm show task VerilogCompiler --json
# Look at "produces" field in output

# 3. Validate compatibility before suggesting workflow
dfm validate --json
# Check for warnings about produces/consumes mismatches
```

### Matching Logic

**OR Logic**: If ANY consume pattern matches ANY produce pattern → compatible

**Subset Matching**: Consumer can be less specific than producer

```yaml
# Producer (more specific)
produces:
  - type: std.FileSet
    filetype: verilog
    vendor: synopsys
    optimization: speed

# Consumer (less specific) - MATCHES!
consumes:
  - type: std.FileSet
    filetype: verilog
```

### Building Compatible Workflows

1. **Identify user's goal** - What output do they need?
2. **Find producers** - `dfm show tasks --produces "type=..."`
3. **Check consumer requirements** - `dfm show task ConsumerTask` → look at consumes
4. **Validate** - `dfm validate` to check compatibility
5. **Suggest workflow** - Connect producer → consumer via `needs`

### Example: Building a Compilation Pipeline

```bash
# User: "I need to compile and simulate my Verilog"

# 1. Find verilog compilers
dfm show tasks --produces "type=std.FileSet,filetype=verilog" --json
# Returns: [VerilogCompiler, PreProcessor, ...]

# 2. Find simulators that consume verilog
dfm show task Simulator --json
# Check consumes field: [{"type": "std.FileSet", "filetype": "verilog"}]

# 3. Suggest workflow
# - name: sim
#   uses: Simulator
#   needs: [VerilogCompiler]  # Compatible!
```

### Pattern Attribute Matching

Match patterns by any attributes defined in produces/consumes:

```yaml
# Common attributes for std.FileSet:
filetype: verilog|systemVerilog|vhdl|...
stage: preprocessed|compiled|optimized
vendor: synopsys|cadence|mentor|...
optimization: speed|area|power
format: json|xml|ucdb|...

# Custom types can have arbitrary attributes
produces:
  - type: custom.BuildArtifact
    language: python
    arch: x86_64
    debug: true
```

### Validation Warnings

If validation shows warnings, help fix them:

```bash
$ dfm validate
WARNING: Task 'Consumer' consumes [{'type': 'std.FileSet', 'filetype': 'vhdl'}]
         but 'Producer' produces [{'type': 'std.FileSet', 'filetype': 'verilog'}].
```

**Solutions:**
1. Change consumer to accept verilog: `filetype: verilog`
2. Find different producer that outputs vhdl
3. Add converter task in between
4. Check if parameter can adjust producer's output

## Detailed Documentation

For comprehensive documentation, see the following reference files:

- [Core Concepts](references/concepts.md) - Tasks, packages, dataflow, types
- [Task Reference](references/tasks.md) - Using and defining tasks
- [Standard Library](references/stdlib.md) - Built-in std.* tasks
- [CLI Reference](references/cli.md) - Command line interface
- [Advanced Patterns](references/advanced.md) - Complex workflows and optimization
- [Flow Schema](dv.flow.schema.json) - JSON Schema for flow.yaml validation

## Flow File Format

DV Flow uses YAML files (`flow.yaml` or `flow.dv`) to define workflows. The file structure is validated against a JSON Schema located at `dv.flow.schema.json`.

### Schema Validation

To validate your flow file or generate the schema:

```bash
# Get the JSON Schema
dfm util schema > flow.schema.json

# Use with YAML validators (e.g., VS Code YAML extension)
# Add to your flow.yaml:
# yaml-language-server: $schema=./dv.flow.schema.json
```

The schema defines two root types:
- **package** - Full project definition with tasks, types, configs, and imports
- **fragment** - Reusable partial definition for inclusion in packages

## Core Concepts Summary

### Tasks
Fundamental units of behavior. Tasks accept data from dependencies (`needs`) and produce outputs. Most tasks inherit from existing tasks using `uses`:

```yaml
- name: my_task
  uses: std.Message
  with:
    msg: "Hello!"
```

### Task Visibility
Control which tasks are entry points and which are API boundaries:

| Scope | Behavior |
|-------|----------|
| `root` | Entry point - shown in `dfm run` listing |
| `export` | Visible outside package for `needs` references |
| `local` | Only visible within its declaration fragment |
| (none) | Package-visible only (default) |

```yaml
# Entry point (shown in task listing)
- root: build
  desc: "Build project"
  run: make build

# Public API (other packages can use)
- export: compile
  run: ./compile.sh

# Both entry point AND public API
- name: main
  scope: [root, export]
  run: ./main.sh
```

**Best Practices:**
- Mark user-facing tasks as `root` so they appear in `dfm run` listing
- Mark tasks other packages should depend on as `export`
- Use `local` for helper tasks in compound task bodies
- Tasks without scope are only visible within the same package

### Packages
Parameterized namespaces that organize tasks. Defined in `flow.yaml` or `flow.dv`:

```yaml
package:
  name: my_package
  with:
    debug:
      type: bool
      value: false
  tasks:
    - name: task1
      uses: std.Message
```

### Dataflow & Produces/Consumes
Tasks communicate via typed data items, not global variables:
- `needs: [task1, task2]` - Specify dependencies
- `produces: [patterns]` - Declare what output datasets this task creates
- `consumes: all|none|[patterns]` - Control what inputs reach implementation
- `passthrough: all|none|unused|[patterns]` - Control what inputs forward to output

**Produces/Consumes enable:**
1. **Task Discovery** - Find tasks that produce specific outputs
2. **Dependency Validation** - Check dataflow compatibility
3. **Relationship Understanding** - Identify which tasks can work together

```yaml
# Producer declares outputs
- name: VerilogCompiler
  produces:
    - type: std.FileSet
      filetype: verilog
  run: compile_verilog.sh

# Consumer declares requirements  
- name: Simulator
  needs: [VerilogCompiler]
  consumes:
    - type: std.FileSet
      filetype: verilog
  run: simulate.sh
```

## File Structure

```
project/
├── flow.yaml           # Main package definition
├── rundir/             # Task execution workspace (created by dfm)
│   ├── cache/          # Task mementos and artifacts
│   └── log/            # Execution traces
└── packages/           # Optional sub-packages
```

## Installation

```bash
pip install dv-flow-mgr
pip install dv-flow-libhdlsim  # Optional: HDL simulator support
```

## Detailed Documentation

For comprehensive documentation, see the following reference files:

- [Core Concepts](references/concepts.md) - Tasks, packages, dataflow, types
- [Task Reference](references/tasks.md) - Using and defining tasks
- [Standard Library](references/stdlib.md) - Built-in std.* tasks
- [CLI Reference](references/cli.md) - Command line interface
- [Advanced Patterns](references/advanced.md) - Complex workflows and optimization
- [Task Development](references/task_development.md) - **Creating custom task implementations and plugin packages** (use when developing new tasks)

**For dataflow and produces/consumes:**
- See the User Guide: Dataflow & Produces (docs/userguide/dataflow.rst) for complete documentation
- See docs/produces.md for quick reference
登录后下载/点赞/收藏 ❤ 4 | ★ 0
dv-flow-manager

目录结构

SKILL.md

举报内容

提示
