ROCmPort-AI / backend /agents /translator.py
tazwarrrr's picture
fixing bugs
a5be23e
Raw History Blame Contribute Delete
3.65 kB
# pylint: disable=broad-exception-caught
from ..models import TranslatorResult, AnalyzerResult
from ..tools.llm_client import LLMClient
from ..tools.hipify_wrapper import HipifyWrapper
from ..tools.json_utils import safe_json_loads
llm_client = LLMClient()
hipify_wrapper = HipifyWrapper()
def chat_complete(messages: list, temperature: float = 0.7, max_tokens: int = 4000) -> str:
"""Wrapper for LLM client chat completion"""
return llm_client.chat_completion(messages, temperature=temperature, max_tokens=max_tokens)
def run_hipify(cuda_code: str) -> str:
"""Wrapper for hipify wrapper"""
return hipify_wrapper.hipify_code(cuda_code)
SYSTEM_PROMPT = """You are an expert AMD ROCm/HIP engineer. You receive CUDA code that has already gone through hipify (basic syntax replacement) and you fix what hipify missed.
Your specific jobs:
1. Fix warp size assumptions: any code assuming warpSize=32 must be updated for AMD wavefront size of 64
- Hardcoded 32 in reductions -> use 64 explicitly or warpSize
- __ballot_sync(0xffffffff, ...) -> __ballot(...)
- __shfl_sync -> __shfl (HIP equivalent)
2. Fix kernel launch syntax if broken
3. Fix any CUDA intrinsics with no direct HIP equivalent
4. Ensure #include uses hip/hip_runtime.h not cuda_runtime.h
Return ONLY this JSON, no markdown:
{
"fixed_code": "the complete fixed HIP code here",
"llm_changes": [
{
"description": "Fixed warp size assumption: changed hardcoded 32 to 64 for AMD wavefront",
"confidence": "high"
}
]
}
If nothing needs fixing beyond what hipify did, return the code unchanged with empty llm_changes array."""
def run(cuda_code: str, analyzer_result: AnalyzerResult) -> TranslatorResult:
# Pass 1: hipify (mechanical replacements)
hip_code_pass1, hipify_changes = run_hipify(cuda_code)
# Pass 2: LLM fixes what hipify missed
context = f"""
The following code has already been through hipify (basic CUDA->HIP syntax replacement).
Analyzer findings:
- Warp size issue detected: {analyzer_result.warp_size_issue}
- Warp size detail: {analyzer_result.warp_size_detail or 'none'}
- Workload type: {analyzer_result.workload_type}
- CUDA APIs found: {', '.join(analyzer_result.cuda_apis)}
Fix what hipify missed, especially warp size issues.
Code after hipify:
```
{hip_code_pass1}
```
"""
try:
raw = chat_complete(
messages=[
{"role": "system", "content": SYSTEM_PROMPT},
{"role": "user", "content": context}
],
temperature=0.1,
max_tokens=4096,
)
data = safe_json_loads(raw)
except Exception:
# Fallback to hipify output if LLM fails
data = {
"fixed_code": hip_code_pass1,
"llm_changes": []
}
final_code = data.get("fixed_code", hip_code_pass1)
llm_changes = data.get("llm_changes", [])
diff_lines = _build_diff(cuda_code, final_code)
return TranslatorResult(
hip_code=final_code,
total_changes=len(hipify_changes) + len(llm_changes),
hipify_changes=len(hipify_changes),
llm_changes=len(llm_changes),
diff_lines=diff_lines,
)
def _build_diff(original: str, converted: str) -> list[dict]:
orig_lines = original.splitlines()
conv_lines = converted.splitlines()
diff = []
max_len = max(len(orig_lines), len(conv_lines))
for i in range(max_len):
o = orig_lines[i] if i < len(orig_lines) else ""
c = conv_lines[i] if i < len(conv_lines) else ""
if o != c:
diff.append({"line": i + 1, "old": o, "new": c})
return diff