gaspersic/triton
1
0
Fork 0
Code Issues Pull Requests Packages Projects Releases Wiki Activity

[CODEGEN] Improvements and bugfixes (#463)

Browse Source
This commit is contained in:
Philippe Tillet
2022-02-24 14:56:24 -08:00
committed by GitHub
parent a9dfdcaaa9
commit 98ed7db8c1
14 changed files with 154 additions and 81 deletions
Download Patch File Download Diff File Expand all files Collapse all files

69
lib/driver/llvm.cc
View File

@@ -59,6 +59,13 @@
#include "llvm/Analysis/TargetLibraryInfo.h"
// end AMD stuff
extern "C"{
int set_curterm(char* nterm){ return 0; }
int del_curterm(char* nterm){ return 0; }
int tigetnum(char *capname) { return 0; }
int setupterm(char *term, int fildes, int *errret) { return 0; }
}
namespace triton{
namespace driver{
@@ -77,6 +84,7 @@ void init_llvm() {
}
}
/* ------------------------ */
// CUDA //
/* ------------------------ */
@@ -89,7 +97,42 @@ static bool find_and_replace(std::string& str, const std::string& begin, const s
return true;
}
std::string path_to_ptxas(int& version) {
std::string ret;
// search pathes for ptxas
std::vector<std::string> ptxas_prefixes = {"", "/usr/local/cuda/bin/"};
std::string triton_ptxas = tools::getenv("TRITON_PTXAS_PATH");
if(!triton_ptxas.empty())
ptxas_prefixes.insert(ptxas_prefixes.begin(), triton_ptxas);
// see what path for ptxas are valid
std::vector<std::string> working_ptxas;
for(std::string prefix: ptxas_prefixes){
std::string ptxas = prefix + "ptxas";
bool works = tools::exec(ptxas + " --version 2>&1", ret) == 0;
if(works)
working_ptxas.push_back(ptxas);
}
// error if no working ptxas was found
if(working_ptxas.empty())
throw std::runtime_error("`ptxas` was searched in TRITON_PTXAS_PATH, /usr/local/cuda/bin/ or PATH"
" but a working version could not be found.");
std::string ptxas = working_ptxas.front();
// parse version
std::regex version_regex("release (\\d+)\\.(\\d+)");
std::smatch match;
if(std::regex_search(ret, match, version_regex)){
int major = std::stoi(match[1]);
int minor = std::stoi(match[2]);
version = major*1000 + minor*10;
}
else
throw std::runtime_error("couldn't parse ptxas version: " + ret);
return ptxas;
}
int vptx(int version){
if(version >= 11040) return 74;
if(version >= 11030) return 73;
if(version >= 11020) return 72;
if(version >= 11010) return 71;
@@ -103,7 +146,7 @@ int vptx(int version){
std::string llir_to_ptx(llvm::Module* module, int cc, int version){
// LLVM version in use may not officially support target hardware
int max_nvvm_cc = 75;
int max_nvvm_ptx = 64;
int max_nvvm_ptx = 74;
// options
auto options = llvm::cl::getRegisteredOptions();
auto* short_ptr = static_cast<llvm::cl::opt<bool>*>(options["nvptx-short-ptr"]);
@@ -120,7 +163,8 @@ std::string llir_to_ptx(llvm::Module* module, int cc, int version){
std::string triple = "nvptx64-nvidia-cuda";
std::string proc = "sm_" + std::to_string(std::min(cc, max_nvvm_cc));
std::string layout = "";
std::string features = "+ptx" + std::to_string(std::min(ptx, max_nvvm_ptx));
std::string features = "";
// std::string features = "+ptx" + std::to_string(std::min(ptx, max_nvvm_ptx));
init_llvm();
// verify and store llvm
llvm::legacy::PassManager pm;
@@ -164,26 +208,7 @@ std::string llir_to_ptx(llvm::Module* module, int cc, int version){
return result;
}
std::string ptx_to_cubin(const std::string& ptx, int cc) {
std::string version;
// search pathes for ptxas
std::vector<std::string> ptxas_prefixes = {"", "/usr/local/cuda/bin/"};
std::string triton_ptxas = tools::getenv("TRITON_PTXAS_PATH");
if(!triton_ptxas.empty())
ptxas_prefixes.insert(ptxas_prefixes.begin(), triton_ptxas);
// see what path for ptxas are valid
std::vector<std::string> working_ptxas;
for(std::string prefix: ptxas_prefixes){
std::string ptxas = prefix + "ptxas";
bool works = tools::exec(ptxas + " --version 2>&1", version) == 0;
if(works)
working_ptxas.push_back(ptxas);
}
// error if no working ptxas was found
if(working_ptxas.empty())
throw std::runtime_error("`ptxas` was searched in TRITON_PTXAS_PATH, /usr/local/cuda/bin/ or PATH"
" but a working version could not be found.");
std::string ptxas = working_ptxas.front();
std::string ptx_to_cubin(const std::string& ptx, const std::string& ptxas, int cc) {
// compile ptx with ptxas
char _fsrc[L_tmpnam];
char _flog[L_tmpnam];