AI/rocm-systems
2
0
Fork 0
Code Issues Pull Requests Actions Packages Projects Releases Wiki Activity

*_wait* routines changed parameter from ptr to ivars to match OpenSHMEM

Browse Source
This commit is contained in:
Yiltan Hassan Temucin
2024-10-10 06:28:43 -07:00
parent e419a8b963
commit bcf3fdff10
16 changed files with 268 additions and 267 deletions
Download Patch File Download Diff File Expand all files Collapse all files
src/context_tmpl_device.hpp
+19 -19
View File
@@ -199,9 +199,9 @@ __device__ void Context::broadcast(T *dest, const T *source, int nelems,
}
template <typename T>
__device__ __forceinline__ void Context::wait_until(T *ptr, roc_shmem_cmps cmp,
__device__ __forceinline__ void Context::wait_until(T *ivars, roc_shmem_cmps cmp,
T val) {
while (!test(ptr, cmp, val)) {
while (!test(ivars, cmp, val)) {
}
}
@@ -219,7 +219,7 @@ __device__ __forceinline__ size_t status_entry(size_t nelems,
template <typename T>
__device__ __forceinline__
size_t Context::wait_until_any(T* ptr, size_t nelems,
size_t Context::wait_until_any(T *ivars, size_t nelems,
const int *status,
roc_shmem_cmps cmp, T val) {
// zero nelems error condition
@@ -240,7 +240,7 @@ size_t Context::wait_until_any(T* ptr, size_t nelems,
if (status[i]) {
continue;
}
if (test(ptr + i, cmp, val)) {
if (test(ivars + i, cmp, val)) {
return i;
}
}
@@ -249,7 +249,7 @@ size_t Context::wait_until_any(T* ptr, size_t nelems,
template <typename T>
__device__ __forceinline__
void Context::wait_until_all(T* ptr, size_t nelems,
void Context::wait_until_all(T *ivars, size_t nelems,
const int *status,
roc_shmem_cmps cmp, T val) {
// zero nelems error condition
@@ -268,14 +268,14 @@ void Context::wait_until_all(T* ptr, size_t nelems,
if (status[i]) {
continue;
}
while (!test(ptr + i, cmp, val)) {
while (!test(ivars + i, cmp, val)) {
}
}
}
template <typename T>
__device__ __forceinline__
size_t Context::wait_until_some(T* ptr, size_t nelems,
size_t Context::wait_until_some(T *ivars, size_t nelems,
size_t* indices,
const int *status,
roc_shmem_cmps cmp, T val) {
@@ -299,7 +299,7 @@ size_t Context::wait_until_some(T* ptr, size_t nelems,
if (status[i]) {
continue;
}
if (test(ptr + i, cmp, val)) {
if (test(ivars + i, cmp, val)) {
done = true;
indices[ncompleted] = i;
ncompleted++;
@@ -311,7 +311,7 @@ size_t Context::wait_until_some(T* ptr, size_t nelems,
template <typename T>
__device__ __forceinline__
void Context::wait_until_all_vector(T* ptr, size_t nelems,
void Context::wait_until_all_vector(T *ivars, size_t nelems,
const int *status,
roc_shmem_cmps cmp, T* vals) {
;
@@ -319,7 +319,7 @@ void Context::wait_until_all_vector(T* ptr, size_t nelems,
template <typename T>
__device__ __forceinline__
size_t Context::wait_until_any_vector(T* ptr, size_t nelems,
size_t Context::wait_until_any_vector(T *ivars, size_t nelems,
const int *status,
roc_shmem_cmps cmp, T* vals) {
return 0;
@@ -327,7 +327,7 @@ size_t Context::wait_until_any_vector(T* ptr, size_t nelems,
template <typename T>
__device__ __forceinline__
size_t Context::wait_until_some_vector(T* ptr, size_t nelems,
size_t Context::wait_until_some_vector(T *ivars, size_t nelems,
size_t* indices,
const int *status,
roc_shmem_cmps cmp, T* vals) {
@@ -335,38 +335,38 @@ size_t Context::wait_until_some_vector(T* ptr, size_t nelems,
}
template <typename T>
__device__ __forceinline__ int Context::test(T *ptr, roc_shmem_cmps cmp,
__device__ __forceinline__ int Context::test(T *ivars, roc_shmem_cmps cmp,
T val) {
int ret = 0;
volatile T *vol_ptr = reinterpret_cast<T *>(ptr);
volatile T *vol_ivars = reinterpret_cast<T *>(ivars);
switch (cmp) {
case ROC_SHMEM_CMP_EQ:
if (uncached_load(vol_ptr) == val) {
if (uncached_load(vol_ivars) == val) {
ret = 1;
}
break;
case ROC_SHMEM_CMP_NE:
if (uncached_load(vol_ptr) != val) {
if (uncached_load(vol_ivars) != val) {
ret = 1;
}
break;
case ROC_SHMEM_CMP_GT:
if (uncached_load(vol_ptr) > val) {
if (uncached_load(vol_ivars) > val) {
ret = 1;
}
break;
case ROC_SHMEM_CMP_GE:
if (uncached_load(vol_ptr) >= val) {
if (uncached_load(vol_ivars) >= val) {
ret = 1;
}
break;
case ROC_SHMEM_CMP_LT:
if (uncached_load(vol_ptr) < val) {
if (uncached_load(vol_ivars) < val) {
ret = 1;
}
break;
case ROC_SHMEM_CMP_LE:
if (uncached_load(vol_ptr) <= val) {
if (uncached_load(vol_ivars) <= val) {
ret = 1;
}
break;