leavelet/cutlass
1
0
Fork 0
Code Issues Pull Requests Actions 28 Packages Projects Releases Wiki Activity

CUTLASS 3.1 (#915)

Browse Source 
Co-authored-by: Aniket Shivam <ashivam@nvidia.com>
This commit is contained in:
ANIKET SHIVAM
2023-04-14 23:19:34 -04:00
committed by GitHub
co-authored by Aniket Shivam
parent 9b8166e3f0
commit d572cc1aab
482 changed files with 37175 additions and 16410 deletions
Download Patch File Download Diff File Expand all files Collapse all files
include/cutlass/barrier.h
+6 -6
View File
@@ -81,7 +81,7 @@ protected:
CUTLASS_DEVICE
static void red_release(int *ptr, int val)
{
#if defined(__NVCC__) || (defined(__clang__) && defined(__CUDA__)) || defined(__CUDACC_RTC__)
#if !defined(CUTLASS_PYTHON_HOST_CC)
#if (__CUDA_ARCH__ >= 700)
/// SM70 and newer use memory consistency qualifiers
@@ -104,7 +104,7 @@ public:
CUTLASS_DEVICE
static void wait_lt(void *lock_ptr, int thread_idx, int flag_idx, int count)
{
#if defined(__NVCC__) || (defined(__clang__) && defined(__CUDA__)) || defined(__CUDACC_RTC__)
#if !defined(CUTLASS_PYTHON_HOST_CC)
T *flag_ptr = reinterpret_cast<T*>(lock_ptr) + flag_idx;
if (thread_idx == 0)
@@ -122,7 +122,7 @@ public:
CUTLASS_DEVICE
static void wait_eq(void *lock_ptr, int thread_idx, int flag_idx, T val = 1)
{
#if defined(__NVCC__) || (defined(__clang__) && defined(__CUDA__)) || defined(__CUDACC_RTC__)
#if !defined(CUTLASS_PYTHON_HOST_CC)
T *flag_ptr = reinterpret_cast<T*>(lock_ptr) + flag_idx;
if (thread_idx == 0)
@@ -138,7 +138,7 @@ public:
/// Uses thread[0] to wait for the specified count of signals on the given flag counter
CUTLASS_DEVICE
static void wait_eq_reset(void *lock_ptr, int thread_idx, int flag_idx, T val = 1) {
#if defined(__NVCC__) || (defined(__clang__) && defined(__CUDA__)) || defined(__CUDACC_RTC__)
#if !defined(CUTLASS_PYTHON_HOST_CC)
T *flag_ptr = reinterpret_cast<T*>(lock_ptr) + flag_idx;
if (thread_idx == 0)
@@ -156,7 +156,7 @@ public:
CUTLASS_DEVICE
static void arrive_inc(void *lock_ptr, int thread_idx, int flag_idx)
{
#if defined(__NVCC__) || (defined(__clang__) && defined(__CUDA__)) || defined(__CUDACC_RTC__)
#if !defined(CUTLASS_PYTHON_HOST_CC)
T* flag_ptr = reinterpret_cast<T*>(lock_ptr) + flag_idx;
__syncthreads();
@@ -173,7 +173,7 @@ public:
CUTLASS_DEVICE
static void arrive_range_inc(void *lock_ptr, int thread_idx, int first_flag_idx, int count = 1)
{
#if defined(__NVCC__) || (defined(__clang__) && defined(__CUDA__)) || defined(__CUDACC_RTC__)
#if !defined(CUTLASS_PYTHON_HOST_CC)
int flag_idx = first_flag_idx + thread_idx;
T* flag_ptr = reinterpret_cast<T*>(lock_ptr) + flag_idx;