SWDEV-444257 - Less # of CUs i.e. 2 tends to ..

- make blockdim to become 0 and hence need validation.
- run cooperative kernel with param lesser # of threads
  than actual # of elements and hence need correction in
  validation part.

Change-Id: Ie48a038fe6f8cf25d7f6460703648a360f511f35


[ROCm/hip-tests commit: ad602b0011]
Этот коммит содержится в:
Jaydeep Patel
2024-02-22 14:36:50 +00:00
коммит произвёл Rakesh Roy
родитель 19a19e773f
Коммит 6f750d2361
2 изменённых файлов: 4 добавлений и 1 удалений
+3 -1
Просмотреть файл
@@ -70,7 +70,9 @@ template <class T> bool CheckDimensions(unsigned int device, T kernel, dim3 bloc
HIP_CHECK(hipGetDeviceProperties(&props, device));
num_sm = props.multiProcessorCount;
if ((blocks.x * blocks.y * blocks.z) > max_blocks_per_sm * num_sm) {
if ((blocks.x * blocks.y * blocks.z) > max_blocks_per_sm * num_sm ||
blocks.x <= 0 || blocks.y <= 0 || blocks.z <= 0 ||
threads.x <= 0 || threads.y <= 0 || threads.z <= 0) {
return false;
}
+1
Просмотреть файл
@@ -505,6 +505,7 @@ TEST_CASE("Unit_hipStreamPerThread_CoopLaunch") {
warp_size, 0));
int max_active_blocks = max_blocks_per_sm * num_sms;
int *Ad = nullptr, *Ah = nullptr, *DNumElms = nullptr, NumElms = 4096;
NumElms = max_active_blocks * warp_size > NumElms ? NumElms : max_active_blocks * warp_size;
int Const = 123;
Ah = new int[NumElms];
for (int i = 0; i < NumElms; ++i) {