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

[CODEGEN] Major performance improvements on A100 (#70)

Browse Source 
Improved handling of asynchronous copy, scheduling and synchronization for A100. Now achieving CUTLASS-like performance on large square dense matrix multiplication tasks
This commit is contained in:
Philippe Tillet
2021-02-21 15:19:39 -08:00
committed by Philippe Tillet
parent 045ab5d62a
commit 5b83259592
31 changed files with 1331 additions and 1115 deletions
Download Patch File Download Diff File Expand all files Collapse all files

16
lib/driver/kernel.cc
View File

@@ -70,7 +70,21 @@ host_kernel::host_kernel(driver::module* program, const char *name): kernel(prog
cu_kernel::cu_kernel(driver::module *program, const char * name) : kernel(program, CUfunction(), true) {
dispatch::cuModuleGetFunction(&*cu_, *program->cu(), name);
// dispatch::cuFuncSetCacheConfig(*cu_, CU_FUNC_CACHE_PREFER_SHARED);
dispatch::cuFuncSetCacheConfig(*cu_, CU_FUNC_CACHE_PREFER_SHARED);
// properties
int shared_total, shared_optin, shared_static;
int n_spills, n_reg;
CUdevice dev;
dispatch::cuCtxGetDevice(&dev);
dispatch::cuDeviceGetAttribute(&shared_total, CU_DEVICE_ATTRIBUTE_MAX_SHARED_MEMORY_PER_MULTIPROCESSOR, dev);
dispatch::cuDeviceGetAttribute(&shared_optin, CU_DEVICE_ATTRIBUTE_MAX_SHARED_MEMORY_PER_BLOCK_OPTIN, dev);
dispatch::cuFuncGetAttribute(&shared_static, CU_FUNC_ATTRIBUTE_SHARED_SIZE_BYTES, *cu_);
dispatch::cuFuncGetAttribute(&n_spills, CU_FUNC_ATTRIBUTE_LOCAL_SIZE_BYTES, *cu_);
dispatch::cuFuncGetAttribute(&n_reg, CU_FUNC_ATTRIBUTE_NUM_REGS, *cu_);
if (shared_optin > 49152){
// std::cout << "dynamic shared memory " << shared_optin << " " << shared_static << std::endl;
dispatch::cuFuncSetAttribute(*cu_, CU_FUNC_ATTRIBUTE_MAX_DYNAMIC_SHARED_SIZE_BYTES, shared_optin - shared_static);
}
}
}

1
lib/driver/module.cc
View File

@@ -282,6 +282,7 @@ std::string cu_module::compile_llvm_module(std::unique_ptr<llvm::Module> module,
void cu_module::init_from_ptx(const std::string& ptx) {
// JIT compile source-code
// std::cout << ptx << std::endl;
try{
// // compile ptx with ptxas

6
lib/driver/stream.cc
View File

@@ -76,7 +76,7 @@ void host_stream::synchronize() {
hst_->args.clear();
}
void host_stream::enqueue(driver::kernel* kernel, std::array<size_t, 3> grid, std::array<size_t, 3> block, void* args, size_t args_size) {
void host_stream::enqueue(driver::kernel* kernel, std::array<size_t, 3> grid, std::array<size_t, 3> block, void* args, size_t args_size, size_t) {
auto hst = kernel->module()->hst();
hst_->futures->reserve(hst_->futures->size() + grid[0]*grid[1]*grid[2]);
char* params = new char[args_size];
@@ -113,13 +113,13 @@ void cu_stream::synchronize() {
dispatch::cuStreamSynchronize(*cu_);
}
void cu_stream::enqueue(driver::kernel* kernel, std::array<size_t, 3> grid, std::array<size_t, 3> block, void* args, size_t args_size) {
void cu_stream::enqueue(driver::kernel* kernel, std::array<size_t, 3> grid, std::array<size_t, 3> block, void* args, size_t args_size, size_t shared_mem) {
void *config[] = {
CU_LAUNCH_PARAM_BUFFER_POINTER, args,
CU_LAUNCH_PARAM_BUFFER_SIZE, &args_size,
CU_LAUNCH_PARAM_END
};
dispatch::cuLaunchKernel(*kernel->cu(), grid[0], grid[1], grid[2], block[0], block[1], block[2], 0, *cu_, nullptr, config);
dispatch::cuLaunchKernel(*kernel->cu(), grid[0], grid[1], grid[2], block[0], block[1], block[2], shared_mem, *cu_, nullptr, config);
}
void cu_stream::write(driver::buffer* buffer, bool blocking, std::size_t offset, std::size_t size, void const* ptr) {