nand.py

from huggingface_hub import InferenceClient
from langchain_community.embeddings.sentence_transformer import SentenceTransformerEmbeddings
from langchain_community.vectorstores import Chroma
from transformers import pipeline
from sentence_transformers.cross_encoder import CrossEncoder
import re

def setupDB(domain, hasLLM):
    history = []
    history.append("")
    history.append("")
    crossmodel = CrossEncoder("cross-encoder/stsb-distilroberta-base")
    models,allState = nandState()
    support_db = nandGetChroma(domain) 
    
    insts_db = nandGetChroma("insts")


    pdf_dbs = []
    if domain == 'en':
        pdfs = [] #"pdf_0em", "pdf_1em", "pdf_2em", "pdf_3em","pdf_4em"]
        for onepdf in pdfs:
            pdfdb =  nandGetChroma(onepdf)
            pdf_dbs.append(pdfdb)
    para = {}
    para['history'] = history
    para['disnum'] = 10
    para['domain'] = domain
    para['crossmodel'] = crossmodel
    para['insts_db'] = insts_db
    para['support_db'] = support_db
    para['pdf_dbs'] = pdf_dbs
    para['hasLLM'] = hasLLM
    return para
def remapScore(domain, inscore):
    if domain == 'ch':
        xin = 1 - inscore
        a = -0.2
        b = 1.2
        y = a * xin * xin + b * xin
        return int(y * 100)
    else:
        xin = 1 - inscore
        a = -1.2
        b = 2.2
        y = a * xin * xin + b * xin
        return int(y * 100)
       
def process_query(iniquery, para):
    query = re.sub("<br>", "", iniquery)
    ch2en, query = toEn(query)
    if ch2en:
        print(f"Received from connected users : {query}")
    else:
        print(f"Received from connected users : {query}", end='')
    disnum = para['disnum']
    domain = para['domain']
    history = para['history']
    crossmodel = para['crossmodel']
    insts_db = para['insts_db']
    support_db = para['support_db']
    pdf_dbs = para['pdf_dbs']
    hasLLM = para['hasLLM']
    ret = ""

    needScriptScores = crossmodel.predict([["write a perl ECO script", query]])
    print(f"THE QUERY SCORE for creating eco script: score={needScriptScores[0]}") 
    allapis = []
    threshold = 0.45
    if needScriptScores[0] > threshold:
        print(f"THE QUERY REQUIRES CREATING AN ECO SCRIPT score={needScriptScores[0]} > {threshold}") 
        retinsts = insts_db.similarity_search_with_score(query, k=10)
        accu = 0
        for inst in retinsts:
            instdoc = inst[0]
            instscore = inst[1]
            instname = instdoc.metadata['source']
            otherfile = re.sub("^insts", "src_en", instname)
            otherfile = re.sub("\.\d+", "", otherfile)
            if not otherfile in allapis:
                allapis.append(otherfile)
                apisize = os.path.getsize(otherfile)
                accu += apisize
                print(f"INST: {instname} SCORE: {instscore} API-size: {apisize} Accu: {accu}")
    
    results = []
    docs = support_db.similarity_search_with_score(query, k=8)
    for doc in docs:
        results.append([doc[0], doc[1]])
    for onepdfdb in pdf_dbs:
        pdocs = onepdfdb.similarity_search_with_score(query, k=8)
        for doc in pdocs:
            results.append([doc[0], doc[1]+0.2])
    results.sort(key=lambda x: x[1])
    docnum = len(results)
    index = 1
    for ii in range(docnum):
        doc = results[ii][0]
        source = doc.metadata['source']
        path = source #source.replace("\\", "/")
        #print(f"path={path}")
        if path in allapis:
            print(f"dont use path={path}, it's in instruction list")
            continue
        prefix = "Help:"
        if re.search("api\.", source):
            prefix = "API:"
        elif re.search("man\.", source):
            prefix = "Manual:"
        elif re.search("\.pdf$", source):
            prefix = "PDF:";
        score = remapScore(domain, results[ii][1])
        retcont = doc.page_content
        if re.search("\.pdf$", source):
            page = doc.metadata['page'] + 1
            subpage = doc.metadata['subpage']
            retcont += f"\n<a target='_blank' href='/AI/{path}#page={page}'>PDF{page} {subpage}</a>\n"
        ret += f"Return {index} ({score}) {prefix} {retcont}\n"
        if len(ret) > 6000:
            break
        index += 1
        if index > disnum:
            break
    if hasLLM:
        context = "Context information is below\n---------------------\n"
        if len(allapis):
            context += scriptExamples()
            for oneapi in allapis:
                cont = GetContent(oneapi)
                cont = re.sub("</h3>", " API Detail:", cont)
                cont = re.sub('<.*?>', '', cont)
                cont = re.sub('Examples:.*', '', cont, flags=re.DOTALL)
                context += cont
        context += ret
        prompt = f"{context}\n"
        prompt += "------------------------------------------\n"
        if len(allapis):
            prompt += "Given the context information and not prior knowledge, creat a Perl ECO script by following the format and sequence in the script examples provided above.\n"
            #prompt += "1. Following the format in the script examples provided above.\n"
            #prompt += "2. Following the API sequence in the script examples above, for instance, APIs get_spare_cells and map_spare_cells should be after fix_design.\n"
        else:
            prompt += "Given the context information and not prior knowledge, answer the query.\n"
        prompt += f"Query: {query}\n"
        
        llmout = llmGenerate(prompt)
        history[0] = query
        history[1] = llmout
        #return llmout
        outlen = len(llmout)
        prolen = len(prompt)
        print(f"Prompt len: {prolen} LLMOUT len: {outlen}")
        allret = "LLM_OUTPUT_START:"+llmout+"\nEND OF LLM OUTPUT\n"+prompt
        return allret
    return ret

def toEn(intxt):
    pattern = re.compile(r'[\u4e00-\u9fff]+')
    if pattern.search(intxt):
        translator = pipeline(task="translation", model="Helsinki-NLP/opus-mt-zh-en")
        ini_text = translator(intxt, max_length=500)[0]['translation_text']
        out_text = re.sub("ECO foot", "ECO Script", ini_text)
        out_text = re.sub("web-based", "netlist", out_text)
        out_text = re.sub(r"\bweb\b", "netlist", out_text)
        out_text = re.sub(r"\bwebsheet\b", "netlist", out_text)
        out_text = re.sub(r"\bweblists?\b", "netlist", out_text)
        print(f"AFTER RESULT: {out_text}")
        return 1, out_text
    return 0, intxt
    


def nandGetChroma(domain):
    models,allState = nandState()
    chdb = allState[domain]['chroma']
    print(f"domain: {domain} has chroma dir {chdb}")
    model_ind = allState[domain]['model']
    model_name = models[model_ind]
    embedding_function = SentenceTransformerEmbeddings(model_name=model_name)
    chroma_db = Chroma(persist_directory=chdb, embedding_function=embedding_function)
    return chroma_db
def nandState():
    models = {'em': "all-MiniLM-L6-v2",
              'en': "all-mpnet-base-v2",
              'ch': "shibing624/text2vec-base-chinese-sentence"}
    # chunk is to cut the big PDF page to smaller, 1000byte chunks, and chinese page into smaller chunks
    allState = {'insts':{'cstate':{},'pstate':{},'dir':'insts','json':'filestatus.insts.json','chroma':'chroma_db_insts','model':'en','chunk':0},
                'en':{'cstate':{},'pstate':{},'dir':'src_en','json':'filestatus.english.json','chroma':'chroma_db_en','model':'en','chunk':0},
                'ch':{'cstate':{},'pstate':{},'dir':'src_ch','json':'filestatus.chinese.json','chroma':'chroma_db_ch','model':'ch','chunk':1}
                }

    for ind in range(12):
        name = f"pdf_{ind}em"
        allState[name] = {'cstate':{},'pstate':{},'dir':f"pdf_sub{ind}",'json':f"filestatus.{name}.json",'chroma':f"chroma_db_{name}",'model':'em','chunk':1}
    return models, allState
def formatPrompt(message, history):
    if history[0]:
        prompt = "Create a new query based on previous query/answer paire and current query:\n"
        prompt += f"Previous query: {history[0]}"
        prompt += f"Previous answer: {histroy[1]}"
        prompt += f"Current query: {message}"
        prompt += "New query:"
        return prompt
    return message

def llmNewQuery(prompt, history):
    newpend = formatPrompt(prompt, history)
    newquery = llmGenerate(newpend)
    return newquery

def llmGenerate(prompt, temperature=0.001, max_new_tokens=2048, top_p=0.95, repetition_penalty=1.0):
    #temperature = float(temperature)
    #if temperature < 1e-2:
    #    temperature = 1e-2
    top_p = float(top_p)

    generate_kwargs = dict(
        temperature=temperature,
        max_new_tokens=max_new_tokens,
        top_p=top_p,
        repetition_penalty=repetition_penalty,
        do_sample=True,
        seed=42,
    )
    llmclient = InferenceClient("mistralai/Mistral-7B-Instruct-v0.2")
   
    stream = llmclient.text_generation(prompt, **generate_kwargs, stream=True, details=True, return_full_text=False)
    output = ""

    for response in stream:
        output += response.token.text
        #yield output
    return output

   
def thoseRemove():
    those = ["www.synopsys.com", "sy ?nopsys", "cadence", "mentor", "solvnetplus", "solvnet"]
    return those

def GetContent(file):
    fcont = ""
    with open(file) as f:
        fcont = f.read()
    return fcont

def scriptExamples():
    exp = """
#The first ECO scipt example for manual ECO:
use strict;
setup_eco("eco_example");
read_library("tsmc.5nm.lib");
read_design("-imp", "implementation.gv");
set_top("topmod");
change_pin("u_abc/state_reg_0_/D", "INVX1", "", "-");
change_pin("u_abc/state_reg_1_/D", "INVX1", "", "-");
change_pin("u_abc/state_reg_2_/D", "INVX1", "", "-");
report_eco(); # ECO report
check_design();
write_verilog("eco_verilog.v");# Write out ECO result in Verilog
#End of the manual ECO script example

#The second ECO script example for automatic ECO:
use strict;
setup_eco("eco_example");# Setup ECO name
read_library("tsmc.5nm.lib");# Read in standard library
# SVF files are optional, best to be used when the design involves multibit flops
#read_svf("-ref", "reference.svf.txt");       
#read_svf("-imp", "implementation.svf.txt"); 
read_design("-ref", "reference.gv");
read_design("-imp", "implementation.gv");
set_top("topmod");# Set the top module
# Preserve DFT Test Logic
set_ignore_output("scan_out*");
set_pin_constant("scan_enable", 0);
set_pin_constant("scan_mode", 0);
fix_design();
report_eco(); # ECO report
check_design();
write_verilog("eco_verilog.v");# Write out ECO result in Verilog
run_lec(); # Run GOF LEC to generate Formality help files
#End of automatic ECO script example


#The third ECO script example is for automatic metal only ECO:
use strict;
setup_eco("eco_example");# Setup ECO name
read_library("tsmc.5nm.lib");# Read in standard library
# SVF files are optional, best to be used when the design involves multibit flops
#read_svf("-ref", "reference.svf.txt");     
#read_svf("-imp", "implementation.svf.txt"); 
read_design("-ref", "reference.gv");# Read in Reference Netlist
read_design("-imp", "implementation.gv");
set_top("topmod");# Set the top module
set_ignore_output("scan_out*");
set_pin_constant("scan_enable", 0);
set_pin_constant("scan_mode", 0);
read_lef("tsmc.lef"); # Read LEF
read_def("topmod.def"); # Read Design Exchange Format file
fix_design(); # Must run before get_spare_cells and map_spare_cells
get_spare_cells("*/*_SPARE*");
map_spare_cells();
report_eco(); # ECO report
check_design();# Check if the ECO causes any issue, like floating
write_verilog("eco_verilog.v");# Write out ECO result in Verilog
write_perl("eco_result.pl");# Write out result in Perl script
run_lec(); # Run GOF LEC to generate Formality help files
#End of automatic ECO script example

#The four ECO script example is the same as the third ECO script, except fix_design
# list_file option to load in the ECO points list file converted from RTL-to-RTL LEC result
fix_design("-list_file", "the_eco_points.txt");    

#The 5th ECO script example is the same as the 3rd ECO script, except fix_design
# Enable flatten mode ECO. The default mode is hierarchical. The flatten mode is for small fix but the changes go across
# module boundaries
fix_design("-flatten");

#The 6th ECO script is similar to the third ECO script, but it dumps formality help file after LEC 
run_lec(); # Run GOF LEC to generate Formality help files
write_compare_points("compare_points.report");
write_formality_help_files("fm_dir/formality_help"); # formality_help files are generated in fm_dir folder    

#The 7th ECO script is similar to the third ECO script, but it uses gate array spare cells
fix_design(); # Must run before get_spare_cells and map_spare_cells
# Enable Gate Array Spare Cells Metal Only ECO Flow, map_spare_cells will map to Gate Array Cells only
get_spare_cells("-gate_array", "G*", "-gate_array_filler", "GFILL*|GDCAP*");
map_spare_cells();

#The 8th ECO script is similar to the third ECO script, but it uses only deleted gates or freed up gates in ECO as spare cells
fix_design(); # Must run before get_spare_cells and map_spare_cells
get_spare_cells("-addfreed");
map_spare_cells();

#The 9th ECO script is manual ECO, find all memory hierarchically and tie the pin TEST_SHIFT of memory to net "TEST_EN"
use strict;
setup_eco("eco_example");
read_library("tsmc.3nm.lib");
read_design("-imp", "from_backend.gv");
set_top("topmod");
# Get all memories hierarchically, instance naming, "U_HMEM*"
my @mems = get_cells("-hier", "U_HMEM*");
foreach my $mem (@mems){
    change_pin("$mem/TEST_SHIFT", "TEST_EN");
}
report_eco(); # ECO report
check_design();
write_verilog("mem_eco.v"); 
    
    """
    return exp