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

SWDEV-470698 - fix formatting, add format check workflow (#657)

Browse Source
This commit is contained in:
Danylo Lytovchenko
2025-08-20 16:28:06 +02:00
committed by GitHub
parent 5840940caa
commit f7338717ae
1574 changed files with 162972 additions and 199346 deletions
Download Patch File Download Diff File Expand all files Collapse all files
projects/hip-tests/catch/unit/multiThread/hipMemsetAsyncMultiThread.cc
+34 -45
View File
@@ -33,26 +33,24 @@ enum TestType { SameStream = 0, DifferentStreams };
// Func -> Memset/MemsetD[8/16/32], T - type of data to be worked on
// basically each thread memsets value present in input one by one on its
// allocated buffer
template <typename Func, typename T>
void threadCall(Func f, hipStream_t stream) {
template <typename Func, typename T> void threadCall(Func f, hipStream_t stream) {
// Should match hipMemsetAsync or hipMemsetD*Async arguments
static_assert(
(std::is_same<Func, hipError_t (*)(void *, int, size_t, hipStream_t)>::
value || // hipMemsetAsync
(std::is_same<Func,
hipError_t (*)(void*, int, size_t, hipStream_t)>::value || // hipMemsetAsync
std::is_same<Func,
hipError_t (*)(hipDeviceptr_t, int, size_t,
hipStream_t)>::value || // hipMemsetD32Async
hipStream_t)>::value || // hipMemsetD32Async
std::is_same<Func,
hipError_t (*)(hipDeviceptr_t, unsigned short, size_t,
hipStream_t)>::value || // hipMemsetD16Async
hipStream_t)>::value || // hipMemsetD16Async
std::is_same<Func,
hipError_t (*)(hipDeviceptr_t, unsigned char, size_t,
hipStream_t)>::value) && // hipMemsetD8Async
hipStream_t)>::value) && // hipMemsetD8Async
"Func f should be hipMemsetAsync or hipMemsetD*Async");
constexpr bool cast_2_void =
std::is_same<Func,
hipError_t (*)(void *, int, size_t, hipStream_t)>::value;
std::is_same<Func, hipError_t (*)(void*, int, size_t, hipStream_t)>::value;
// Use the unsiged type, since memset concerns with set bit values over a mem
// address
@@ -63,10 +61,10 @@ void threadCall(Func f, hipStream_t stream) {
std::mt19937_64 engine(std::random_device{}());
auto distribution = std::uniform_int_distribution<unsigned int>(
min, max); // this needs to be unsigned because windows does not treats
// char as numeric types
min, max); // this needs to be unsigned because windows does not treats
// char as numeric types
T *ptr{nullptr};
T* ptr{nullptr};
constexpr size_t size = 1024;
constexpr size_t iter = 1024;
HIP_CHECK_THREAD(hipMalloc(&ptr, sizeof(T) * size));
@@ -80,19 +78,17 @@ void threadCall(Func f, hipStream_t stream) {
std::vector<T> dst(size, 0);
for (size_t i = 0; i < iter; i++) {
overlay_val.u_val = static_cast<unsigned_t>(
distribution(engine)); // generate an unsigned int number
overlay_val.u_val =
static_cast<unsigned_t>(distribution(engine)); // generate an unsigned int number
if constexpr (cast_2_void) {
HIP_CHECK_THREAD(f((void *)ptr, overlay_val.t_val, size, stream));
HIP_CHECK_THREAD(f((void*)ptr, overlay_val.t_val, size, stream));
} else {
HIP_CHECK_THREAD(
f(*(hipDeviceptr_t *)&ptr, overlay_val.t_val, size, stream));
HIP_CHECK_THREAD(f(*(hipDeviceptr_t*)&ptr, overlay_val.t_val, size, stream));
}
HIP_CHECK_THREAD(hipMemcpyAsync(dst.data(), ptr, size * sizeof(T),
hipMemcpyDeviceToHost, stream));
HIP_CHECK_THREAD(hipEventRecord(event, stream));
HIP_CHECK_THREAD(
hipStreamWaitEvent(stream, event, 0)); // wait till memcpy is done
hipMemcpyAsync(dst.data(), ptr, size * sizeof(T), hipMemcpyDeviceToHost, stream));
HIP_CHECK_THREAD(hipEventRecord(event, stream));
HIP_CHECK_THREAD(hipStreamWaitEvent(stream, event, 0)); // wait till memcpy is done
REQUIRE_THREAD(std::all_of(dst.begin(), dst.end(), [&](T v) {
// If this test ever fails, add prints here on mismatch
return v == overlay_val.t_val;
@@ -109,25 +105,24 @@ template <typename Func, typename T> void launchThreads(Func f, TestType type) {
// Should match hipMemsetAsync or hipMemsetD*Async arguments
static_assert(
(std::is_same<Func, hipError_t (*)(void *, int, size_t, hipStream_t)>::
value || // hipMemsetAsync
(std::is_same<Func,
hipError_t (*)(void*, int, size_t, hipStream_t)>::value || // hipMemsetAsync
std::is_same<Func,
hipError_t (*)(hipDeviceptr_t, int, size_t,
hipStream_t)>::value || // hipMemsetD32Async
hipStream_t)>::value || // hipMemsetD32Async
std::is_same<Func,
hipError_t (*)(hipDeviceptr_t, unsigned short, size_t,
hipStream_t)>::value || // hipMemsetD16Async
hipStream_t)>::value || // hipMemsetD16Async
std::is_same<Func,
hipError_t (*)(hipDeviceptr_t, unsigned char, size_t,
hipStream_t)>::value) && // hipMemsetD8Async
hipStream_t)>::value) && // hipMemsetD8Async
"Func f should be hipMemsetAsync or hipMemsetD*Async");
const size_t num_threads =
(std::thread::hardware_concurrency() > 8)
? (((std::thread::hardware_concurrency() / 4) >= 127)
? 127
: (std::thread::hardware_concurrency() / 4))
: 2; // thread count between 2 - 127
const size_t num_threads = (std::thread::hardware_concurrency() > 8)
? (((std::thread::hardware_concurrency() / 4) >= 127)
? 127
: (std::thread::hardware_concurrency() / 4))
: 2; // thread count between 2 - 127
const size_t num_streams = (type == SameStream) ? 1 : num_threads;
std::vector<hipStream_t> streams(num_streams, nullptr);
@@ -149,7 +144,7 @@ template <typename Func, typename T> void launchThreads(Func f, TestType type) {
thread_pool[i].join();
}
HIP_CHECK_THREAD_FINALIZE(); // Make sure all thread have exited properly
HIP_CHECK_THREAD_FINALIZE(); // Make sure all thread have exited properly
for (size_t i = 0; i < num_streams; i++) {
HIP_CHECK(hipStreamDestroy(streams[i]));
@@ -157,26 +152,20 @@ template <typename Func, typename T> void launchThreads(Func f, TestType type) {
}
TEST_CASE("Unit_hipMemsetAsync_QueueJobsMultithreaded") {
using hipMemsetAsync_t =
hipError_t (*)(void *, int, const size_t, hipStream_t);
using hipMemsetAsync_t = hipError_t (*)(void*, int, const size_t, hipStream_t);
using hipMemsetAsyncD8_t =
hipError_t (*)(hipDeviceptr_t, unsigned char, const size_t, hipStream_t);
using hipMemsetAsyncD16_t =
hipError_t (*)(hipDeviceptr_t, unsigned short, const size_t, hipStream_t);
using hipMemsetAsyncD32_t =
hipError_t (*)(hipDeviceptr_t, int, const size_t, hipStream_t);
using hipMemsetAsyncD32_t = hipError_t (*)(hipDeviceptr_t, int, const size_t, hipStream_t);
launchThreads<hipMemsetAsync_t, char>(hipMemsetAsync, SameStream);
launchThreads<hipMemsetAsync_t, char>(hipMemsetAsync, DifferentStreams);
launchThreads<hipMemsetAsyncD8_t, unsigned char>(hipMemsetD8Async,
SameStream);
launchThreads<hipMemsetAsyncD8_t, unsigned char>(hipMemsetD8Async,
DifferentStreams);
launchThreads<hipMemsetAsyncD16_t, unsigned short>(hipMemsetD16Async,
SameStream);
launchThreads<hipMemsetAsyncD16_t, unsigned short>(hipMemsetD16Async,
DifferentStreams);
launchThreads<hipMemsetAsyncD8_t, unsigned char>(hipMemsetD8Async, SameStream);
launchThreads<hipMemsetAsyncD8_t, unsigned char>(hipMemsetD8Async, DifferentStreams);
launchThreads<hipMemsetAsyncD16_t, unsigned short>(hipMemsetD16Async, SameStream);
launchThreads<hipMemsetAsyncD16_t, unsigned short>(hipMemsetD16Async, DifferentStreams);
launchThreads<hipMemsetAsyncD32_t, int>(hipMemsetD32Async, SameStream);
launchThreads<hipMemsetAsyncD32_t, int>(hipMemsetD32Async, DifferentStreams);
}