ChatHLS-HLSTuner is a specialized Large Language Model fine-tuned for High-Level Synthesis (HLS) C/C++ code optimization. It acts as the optimization analysis agent within the ChatHLS framework.

Quickstart

Prerequisites

pip install transformers torch accelerate

Example

Here is a simple example of how to format your prompt and generate optimization analysis for HLS code:

python

from transformers import AutoModelForCausalLM, AutoTokenizer
import torch

model_name = "XXXiong/ChatHLS-HLSTuner"
tokenizer = AutoTokenizer.from_pretrained(model_name)
model = AutoModelForCausalLM.from_pretrained(
    model_name,
    torch_dtype=torch.bfloat16,
    device_map="auto"
)

hls_code = """
#include "ap_fixed.h"
#include "hls_math.h"

typedef ap_fixed<32,16> t_ap_fixed;

void gemm(  
		 t_ap_fixed alpha,
		 t_ap_fixed beta,
		 t_ap_fixed C[ 20 + 0][25 + 0],
		 t_ap_fixed A[ 20 + 0][30 + 0],
		 t_ap_fixed B[ 30 + 0][25 + 0])
{
  #pragma HLS top name=gemm

    const int ni = 20;
    const int nj = 25;
    const int nk = 30;

  int i, j, k;
L1:  for (i = 0; i < ni; i++) {
L2:    for (j = 0; j < nj; j++)
	C[i][j] *= beta;
L3:    for (k = 0; k < nk; k++) {
L4:       for (j = 0; j < nj; j++)
	  C[i][j] += alpha * A[i][k] * B[k][j];
    }
  }
}
"""

hls_code_ppa = """
latency: 15661 cycles
util_dsp: 10.5%
util_ff: 0.4%
util_lut: 2.1%
"""

loop_array_info = """
kernel_gemm
L1,loop,20
L2,loop,25
L3,loop,30
L4,loop,25
C,array,20/25
A,array,20/30
B,array,30/25
"""

prompt = f"""Your task is to optimize the following HLS kernel by inserting `#pragma` directives 
to enhance loops and arrays for lower latency and relatively low resource utilization. 
You may use the following pragma types:

- Loop Pipeline
- Loop Unroll
- Array Partition

Provided code:
{hls_code}

Current latency and hardware resource utilization:
{hls_code_ppa}

Loop and array information:
{loop_array_info}

Please suggest pragma modifications, specifying where to insert them and the reasoning behind each. 
Provide only the suggestions without additional explanations or formatting.

**Expected Output Format**:

Detailed Reasoning:
```text
[Your step-by-step analysis here]
```

Optimization Strategies:
```json
{{
    "Optimization Strategies": [
        {{
            "Code Segment Before Optimization (Loop/Array)": "[Code line before optimization]",
            "Code Segment After Optimization (Loop/Array)": "[Code line after inserting the pragma]",
            "Reason": "[Explanation of why this pragma was inserted]",
            "Result Analysis": {{
                "Impact of Inserting This Pragma on PPA": "[Description of the impact on Performance, Power, and Area]",
                "Reason": "[Explanation of why this impact occurred based on the performance analysis]"
            }}
        }}
    ]
}}
```
"""

messages = [
    {"role": "system", "content": """You are an expert proficient in HLS algorithm optimization,
                improving algorithm PPA through inserting pragmas into arrays and loops in HLS code."""},
    {"role": "user", "content": prompt}
]

text = tokenizer.apply_chat_template(
    messages,
    tokenize=False,
    add_generation_prompt=True
)

model_inputs = tokenizer([text], return_tensors="pt").to(model.device)

generated_ids = model.generate(
    **model_inputs,
    max_new_tokens=2048,
    temperature=0.7
)
generated_ids = [
    output_ids[len(input_ids):] for input_ids, output_ids in zip(model_inputs.input_ids, generated_ids)
]

response = tokenizer.batch_decode(generated_ids, skip_special_tokens=True)[0]
print(response)