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);
}
projects/hip-tests/catch/unit/multiThread/hipMultiThreadDevice.cc
+56 -62
View File
@@ -2,35 +2,35 @@
#include <hip_test_common.hh>
#ifdef _WIN32
#define MAX_BURST_SIZE 40
#define MAX_BURST_SIZE 40
#else
#define MAX_BURST_SIZE 100
#define MAX_BURST_SIZE 100
#endif
// Create a lot of streams and then destroy 'em.
void createThenDestroyStreams(int iterations, int burstSize) {
hipStream_t* streams = new hipStream_t[burstSize];
hipStream_t* streams = new hipStream_t[burstSize];
for (int i = 0; i < iterations; i++) {
for (int j = 0; j < burstSize; j++) {
HIPCHECK(hipStreamCreate(&streams[j]));
}
for (int j = 0; j < burstSize; j++) {
HIPCHECK(hipStreamDestroy(streams[j]));
}
for (int i = 0; i < iterations; i++) {
for (int j = 0; j < burstSize; j++) {
HIPCHECK(hipStreamCreate(&streams[j]));
}
for (int j = 0; j < burstSize; j++) {
HIPCHECK(hipStreamDestroy(streams[j]));
}
}
delete[] streams;
delete[] streams;
}
void waitStreams(int iterations) {
// Repeatedly sync and wait for all streams to complete.
// TO make this interesting, the test has other threads repeatedly adding and removing streams
// to the device.
for (int i = 0; i < iterations; i++) {
HIPCHECK(hipDeviceSynchronize());
}
// Repeatedly sync and wait for all streams to complete.
// TO make this interesting, the test has other threads repeatedly adding and removing streams
// to the device.
for (int i = 0; i < iterations; i++) {
HIPCHECK(hipDeviceSynchronize());
}
}
@@ -38,67 +38,61 @@ void waitStreams(int iterations) {
// Some create many queue, some not many.
//
void multiThread_pyramid(bool serialize, int iters) {
std::thread t1(createThenDestroyStreams, iters * 1, MAX_BURST_SIZE);
if (serialize) {
t1.join();
}
std::thread t1(createThenDestroyStreams, iters * 1, MAX_BURST_SIZE);
if (serialize) {
t1.join();
}
std::thread t2(createThenDestroyStreams, iters * 10, 10);
if (serialize) {
t2.join();
}
std::thread t2(createThenDestroyStreams, iters * 10, 10);
if (serialize) {
t2.join();
}
std::thread t3(createThenDestroyStreams, iters * 100, 1);
if (serialize) {
t3.join();
}
std::thread t3(createThenDestroyStreams, iters * 100, 1);
if (serialize) {
t3.join();
}
if (!serialize) {
t1.join();
t2.join();
t3.join();
}
if (!serialize) {
t1.join();
t2.join();
t3.join();
}
}
// Create 3 streams, all creating and destroying streams on the same device.
// Try to keep number of streams near zero, to cause problems.
void multiThread_nearzero(bool serialize, int iters) {
std::thread t1(createThenDestroyStreams, iters, 1);
if (serialize) {
t1.join();
}
std::thread t1(createThenDestroyStreams, iters, 1);
if (serialize) {
t1.join();
}
std::thread t2(createThenDestroyStreams, iters, 1);
if (serialize) {
t2.join();
}
std::thread t2(createThenDestroyStreams, iters, 1);
if (serialize) {
t2.join();
}
std::thread t3(waitStreams, iters * 50);
if (serialize) {
t3.join();
}
std::thread t3(waitStreams, iters * 50);
if (serialize) {
t3.join();
}
if (!serialize) {
t1.join();
t2.join();
t3.join();
}
if (!serialize) {
t1.join();
t2.join();
t3.join();
}
}
TEST_CASE("Unit_hipMultiThreadDevice_Streams") {
// Serial version, just call once:
createThenDestroyStreams(10, 10);
// Serial version, just call once:
createThenDestroyStreams(10, 10);
}
TEST_CASE("Unit_hipMultiThreadDevice_SerialPyramid") {
multiThread_pyramid(true, 3);
}
TEST_CASE("Unit_hipMultiThreadDevice_SerialPyramid") { multiThread_pyramid(true, 3); }
TEST_CASE("Unit_hipMultiThreadDevice_ParallelPyramid") {
multiThread_pyramid(false, 3);
}
TEST_CASE("Unit_hipMultiThreadDevice_ParallelPyramid") { multiThread_pyramid(false, 3); }
TEST_CASE("Unit_hipMultiThreadDevice_NearZero") {
multiThread_nearzero(false, 1000);
}
TEST_CASE("Unit_hipMultiThreadDevice_NearZero") { multiThread_nearzero(false, 1000); }
projects/hip-tests/catch/unit/multiThread/hipMultiThreadStreams1.cc
+72 -74
View File
@@ -35,106 +35,104 @@ unsigned threadsPerBlock = 256;
template <typename T = float, class P = HipTest::Unpinned, class C = HipTest::Memcpy>
void simpleVectorAdd(size_t numElements, int iters, hipStream_t stream) {
using HipTest::MemTraits;
size_t Nbytes = numElements * sizeof(T);
using HipTest::MemTraits;
size_t Nbytes = numElements * sizeof(T);
T *A_d, *B_d, *C_d;
T *A_h, *B_h, *C_h;
T *A_d, *B_d, *C_d;
T *A_h, *B_h, *C_h;
HipTest::initArrays(&A_d, &B_d, &C_d, &A_h, &B_h, &C_h, N, P::isPinned);
for (size_t i = 0; i < numElements; i++) {
A_h[i] = 1000.0f;
B_h[i] = 2000.0f;
C_h[i] = -1;
}
HipTest::initArrays(&A_d, &B_d, &C_d, &A_h, &B_h, &C_h, N, P::isPinned);
for (size_t i = 0; i < numElements; i++) {
A_h[i] = 1000.0f;
B_h[i] = 2000.0f;
C_h[i] = -1;
}
MemTraits<C>::Copy(B_d, B_h, Nbytes, hipMemcpyHostToDevice, stream);
MemTraits<C>::Copy(A_d, A_h, Nbytes, hipMemcpyHostToDevice, stream);
MemTraits<C>::Copy(C_d, C_h, Nbytes, hipMemcpyHostToDevice, stream);
HIPCHECK(hipDeviceSynchronize());
for (size_t i = 0; i < numElements; i++) {
A_h[i] = 1.0f;
B_h[i] = 2.0f;
C_h[i] = -1;
}
for (int i = 0; i < iters; i++) {
unsigned blocks = HipTest::setNumBlocks(blocksPerCU, threadsPerBlock, numElements);
MemTraits<C>::Copy(B_d, B_h, Nbytes, hipMemcpyHostToDevice, stream);
MemTraits<C>::Copy(A_d, A_h, Nbytes, hipMemcpyHostToDevice, stream);
MemTraits<C>::Copy(C_d, C_h, Nbytes, hipMemcpyHostToDevice, stream);
MemTraits<C>::Copy(B_d, B_h, Nbytes, hipMemcpyHostToDevice, stream);
hipLaunchKernelGGL(HipTest::vectorADDReverse, dim3(blocks), dim3(threadsPerBlock), 0, 0,
static_cast<const T*>(A_d), static_cast<const T*>(B_d), C_d, numElements);
HIP_CHECK(hipGetLastError());
MemTraits<C>::Copy(C_h, C_d, Nbytes, hipMemcpyDeviceToHost, stream);
HIPCHECK(hipDeviceSynchronize());
for (size_t i = 0; i < numElements; i++) {
A_h[i] = 1.0f;
B_h[i] = 2.0f;
C_h[i] = -1;
}
HipTest::checkVectorADD(A_h, B_h, C_h, numElements);
}
for (int i = 0; i < iters; i++) {
unsigned blocks = HipTest::setNumBlocks(blocksPerCU, threadsPerBlock, numElements);
MemTraits<C>::Copy(A_d, A_h, Nbytes, hipMemcpyHostToDevice, stream);
MemTraits<C>::Copy(B_d, B_h, Nbytes, hipMemcpyHostToDevice, stream);
hipLaunchKernelGGL(HipTest::vectorADDReverse, dim3(blocks), dim3(threadsPerBlock), 0, 0,
static_cast<const T*>(A_d), static_cast<const T*>(B_d), C_d, numElements);
HIP_CHECK(hipGetLastError());
MemTraits<C>::Copy(C_h, C_d, Nbytes, hipMemcpyDeviceToHost, stream);
HIPCHECK(hipDeviceSynchronize());
HipTest::checkVectorADD(A_h, B_h, C_h, numElements);
}
HipTest::freeArrays(A_d, B_d, C_d, A_h, B_h, C_h, P::isPinned);
HIPCHECK(hipDeviceSynchronize());
HipTest::freeArrays(A_d, B_d, C_d, A_h, B_h, C_h, P::isPinned);
HIPCHECK(hipDeviceSynchronize());
}
template <typename T, class C>
void test_multiThread_1(hipStream_t stream0, hipStream_t stream1, bool serialize) {
size_t numElements = N;
size_t numElements = N;
// Test 2 threads operating on same stream:
std::thread t1(simpleVectorAdd<T, HipTest::Pinned, C>, numElements, p_iters /*iters*/, stream0);
if (serialize) {
t1.join();
}
std::thread t2(simpleVectorAdd<T, HipTest::Pinned, C>, numElements, p_iters /*iters*/, stream1);
if (serialize) {
t2.join();
}
// Test 2 threads operating on same stream:
std::thread t1(simpleVectorAdd<T, HipTest::Pinned, C>, numElements, p_iters /*iters*/, stream0);
if (serialize) {
t1.join();
}
std::thread t2(simpleVectorAdd<T, HipTest::Pinned, C>, numElements, p_iters /*iters*/, stream1);
if (serialize) {
t2.join();
}
if (!serialize) {
t1.join();
t2.join();
}
if (!serialize) {
t1.join();
t2.join();
}
HIPCHECK(hipDeviceSynchronize());
HIPCHECK(hipDeviceSynchronize());
};
TEST_CASE("Unit_hipMultiThreadStreams1_AsyncSync") {
hipStream_t stream;
HIPCHECK(hipStreamCreate(&stream));
hipStream_t stream;
HIPCHECK(hipStreamCreate(&stream));
simpleVectorAdd<float, HipTest::Pinned, HipTest::MemcpyAsync>(N /*mb*/, 10 /*iters*/, stream);
simpleVectorAdd<float, HipTest::Pinned, HipTest::Memcpy>(N /*mb*/, 10 /*iters*/, stream);
simpleVectorAdd<float, HipTest::Pinned, HipTest::MemcpyAsync>(N /*mb*/, 10 /*iters*/, stream);
simpleVectorAdd<float, HipTest::Pinned, HipTest::Memcpy>(N /*mb*/, 10 /*iters*/, stream);
HIPCHECK(hipStreamDestroy(stream));
HIPCHECK(hipStreamDestroy(stream));
}
TEST_CASE("Unit_hipMultiThreadStreams1_AsyncAsync") {
hipStream_t stream0, stream1;
HIPCHECK(hipStreamCreate(&stream0));
HIPCHECK(hipStreamCreate(&stream1));
hipStream_t stream0, stream1;
HIPCHECK(hipStreamCreate(&stream0));
HIPCHECK(hipStreamCreate(&stream1));
// Easy tests to verify the test works - these don't allow overlap between the threads:
test_multiThread_1<float, HipTest::MemcpyAsync>(NULL, NULL, true);
test_multiThread_1<float, HipTest::MemcpyAsync>(stream0, stream1, true);
// Easy tests to verify the test works - these don't allow overlap between the threads:
test_multiThread_1<float, HipTest::MemcpyAsync>(NULL, NULL, true);
test_multiThread_1<float, HipTest::MemcpyAsync>(stream0, stream1, true);
HIPCHECK(hipStreamDestroy(stream0));
HIPCHECK(hipStreamDestroy(stream1));
HIPCHECK(hipStreamDestroy(stream0));
HIPCHECK(hipStreamDestroy(stream1));
}
TEST_CASE("Unit_hipMultiThreadStreams1_AsyncSame") {
hipStream_t stream;
HIPCHECK(hipStreamCreate(&stream));
hipStream_t stream;
HIPCHECK(hipStreamCreate(&stream));
// test_multiThread_1<float, HipTest::MemcpyAsync> ("Multithread with NULL stream", NULL,
// NULL, false); test_multiThread_1<float, HipTest::MemcpyAsync> ("Multithread with two
// streams", stream0, stream1, false);
test_multiThread_1<float, HipTest::MemcpyAsync>(stream, stream, false);
// test_multiThread_1<float, HipTest::MemcpyAsync> ("Multithread with NULL stream", NULL,
// NULL, false); test_multiThread_1<float, HipTest::MemcpyAsync> ("Multithread with two
// streams", stream0, stream1, false);
test_multiThread_1<float, HipTest::MemcpyAsync>(stream, stream, false);
HIPCHECK(hipStreamDestroy(stream));
HIPCHECK(hipStreamDestroy(stream));
}
projects/hip-tests/catch/unit/multiThread/hipMultiThreadStreams2.cc
+94 -95
View File
@@ -20,7 +20,7 @@ OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN
THE SOFTWARE.
*/
//This file is a port from hiprocclrtests (hipMultiThreadStreams2)
// This file is a port from hiprocclrtests (hipMultiThreadStreams2)
#include <iostream>
@@ -28,120 +28,119 @@ THE SOFTWARE.
#include <thread>
#define N 1000
template <typename T>
__global__ void Inc(T* Array) {
int tx = threadIdx.x + blockIdx.x * blockDim.x;
Array[tx] = Array[tx] + T(1);
template <typename T> __global__ void Inc(T* Array) {
int tx = threadIdx.x + blockIdx.x * blockDim.x;
Array[tx] = Array[tx] + T(1);
}
void run1(size_t size, hipStream_t stream) {
float *Ah, *Bh, *Cd, *Dd, *Eh;
float *snap = (float *) malloc(size);
float *Ah, *Bh, *Cd, *Dd, *Eh;
float* snap = (float*)malloc(size);
HIPCHECK(hipHostMalloc((void**)&Ah, size, hipHostMallocDefault));
HIPCHECK(hipHostMalloc((void**)&Bh, size, hipHostMallocDefault));
HIPCHECK(hipMalloc(&Cd, size));
HIPCHECK(hipMalloc(&Dd, size));
HIPCHECK(hipHostMalloc((void**)&Eh, size, hipHostMallocDefault));
HIPCHECK(hipHostMalloc((void**)&Ah, size, hipHostMallocDefault));
HIPCHECK(hipHostMalloc((void**)&Bh, size, hipHostMallocDefault));
HIPCHECK(hipMalloc(&Cd, size));
HIPCHECK(hipMalloc(&Dd, size));
HIPCHECK(hipHostMalloc((void**)&Eh, size, hipHostMallocDefault));
for (int i = 0; i < N; i++) {
Ah[i] = 1.0f;
}
for (int i = 0; i < N; i++) {
Ah[i] = 1.0f;
}
HIPCHECK(hipMemcpyAsync(Bh, Ah, size, hipMemcpyHostToHost, stream));
HIPCHECK(hipMemcpyAsync(Cd, Bh, size, hipMemcpyHostToDevice, stream));
hipLaunchKernelGGL(HIP_KERNEL_NAME(Inc), dim3(N / 500), dim3(500), 0, stream, Cd);
HIP_CHECK(hipGetLastError());
HIPCHECK(hipMemcpyAsync(Dd, Cd, size, hipMemcpyDeviceToDevice, stream));
HIPCHECK(hipMemcpyAsync(Eh, Dd, size, hipMemcpyDeviceToHost, stream));
HIPCHECK(hipDeviceSynchronize());
HIPCHECK(hipMemcpyAsync(Bh, Ah, size, hipMemcpyHostToHost, stream));
HIPCHECK(hipMemcpyAsync(Cd, Bh, size, hipMemcpyHostToDevice, stream));
hipLaunchKernelGGL(HIP_KERNEL_NAME(Inc), dim3(N / 500), dim3(500), 0, stream, Cd);
HIP_CHECK(hipGetLastError());
HIPCHECK(hipMemcpyAsync(Dd, Cd, size, hipMemcpyDeviceToDevice, stream));
HIPCHECK(hipMemcpyAsync(Eh, Dd, size, hipMemcpyDeviceToHost, stream));
HIPCHECK(hipDeviceSynchronize());
memcpy(snap, Eh, size);
for (int i = 0; i < N; i++) {
HIPASSERT(snap[i] == Ah[i] + 1.0f);
}
free(snap);
HIPCHECK(hipHostFree(Ah));
HIPCHECK(hipHostFree(Bh));
HIPCHECK(hipHostFree(Eh));
HIPCHECK(hipFree(Cd));
HIPCHECK(hipFree(Dd));
memcpy(snap, Eh, size);
for (int i = 0; i < N; i++) {
HIPASSERT(snap[i] == Ah[i] + 1.0f);
}
free(snap);
HIPCHECK(hipHostFree(Ah));
HIPCHECK(hipHostFree(Bh));
HIPCHECK(hipHostFree(Eh));
HIPCHECK(hipFree(Cd));
HIPCHECK(hipFree(Dd));
}
void run(size_t size, hipStream_t stream1, hipStream_t stream2) {
float *Ah, *Bh, *Cd, *Dd, *Eh;
float *Ahh, *Bhh, *Cdd, *Ddd, *Ehh;
float *snap, *snapp;
float *Ah, *Bh, *Cd, *Dd, *Eh;
float *Ahh, *Bhh, *Cdd, *Ddd, *Ehh;
float *snap, *snapp;
snap = (float *) malloc(size);
snapp = (float *) malloc(size);
snap = (float*)malloc(size);
snapp = (float*)malloc(size);
HIPCHECK(hipHostMalloc((void**)&Ah, size, hipHostMallocDefault));
HIPCHECK(hipHostMalloc((void**)&Bh, size, hipHostMallocDefault));
HIPCHECK(hipMalloc(&Cd, size));
HIPCHECK(hipMalloc(&Dd, size));
HIPCHECK(hipHostMalloc((void**)&Eh, size, hipHostMallocDefault));
HIPCHECK(hipHostMalloc((void**)&Ahh, size, hipHostMallocDefault));
HIPCHECK(hipHostMalloc((void**)&Bhh, size, hipHostMallocDefault));
HIPCHECK(hipMalloc(&Cdd, size));
HIPCHECK(hipMalloc(&Ddd, size));
HIPCHECK(hipHostMalloc((void**)&Ehh, size, hipHostMallocDefault));
HIPCHECK(hipHostMalloc((void**)&Ah, size, hipHostMallocDefault));
HIPCHECK(hipHostMalloc((void**)&Bh, size, hipHostMallocDefault));
HIPCHECK(hipMalloc(&Cd, size));
HIPCHECK(hipMalloc(&Dd, size));
HIPCHECK(hipHostMalloc((void**)&Eh, size, hipHostMallocDefault));
HIPCHECK(hipHostMalloc((void**)&Ahh, size, hipHostMallocDefault));
HIPCHECK(hipHostMalloc((void**)&Bhh, size, hipHostMallocDefault));
HIPCHECK(hipMalloc(&Cdd, size));
HIPCHECK(hipMalloc(&Ddd, size));
HIPCHECK(hipHostMalloc((void**)&Ehh, size, hipHostMallocDefault));
HIPCHECK(hipMemcpyAsync(Bh, Ah, size, hipMemcpyHostToHost, stream1));
HIPCHECK(hipMemcpyAsync(Bhh, Ahh, size, hipMemcpyHostToHost, stream2));
HIPCHECK(hipMemcpyAsync(Cd, Bh, size, hipMemcpyHostToDevice, stream1));
HIPCHECK(hipMemcpyAsync(Cdd, Bhh, size, hipMemcpyHostToDevice, stream2));
hipLaunchKernelGGL(HIP_KERNEL_NAME(Inc), dim3(N / 500), dim3(500), 0, stream1, Cd);
HIP_CHECK(hipGetLastError());
hipLaunchKernelGGL(HIP_KERNEL_NAME(Inc), dim3(N / 500), dim3(500), 0, stream2, Cdd);
HIP_CHECK(hipGetLastError());
HIPCHECK(hipMemcpyAsync(Dd, Cd, size, hipMemcpyDeviceToDevice, stream1));
HIPCHECK(hipMemcpyAsync(Ddd, Cdd, size, hipMemcpyDeviceToDevice, stream2));
HIPCHECK(hipMemcpyAsync(Eh, Dd, size, hipMemcpyDeviceToHost, stream1));
HIPCHECK(hipMemcpyAsync(Ehh, Ddd, size, hipMemcpyDeviceToHost, stream2));
HIPCHECK(hipDeviceSynchronize());
HIPCHECK(hipMemcpyAsync(Bh, Ah, size, hipMemcpyHostToHost, stream1));
HIPCHECK(hipMemcpyAsync(Bhh, Ahh, size, hipMemcpyHostToHost, stream2));
HIPCHECK(hipMemcpyAsync(Cd, Bh, size, hipMemcpyHostToDevice, stream1));
HIPCHECK(hipMemcpyAsync(Cdd, Bhh, size, hipMemcpyHostToDevice, stream2));
hipLaunchKernelGGL(HIP_KERNEL_NAME(Inc), dim3(N / 500), dim3(500), 0, stream1, Cd);
HIP_CHECK(hipGetLastError());
hipLaunchKernelGGL(HIP_KERNEL_NAME(Inc), dim3(N / 500), dim3(500), 0, stream2, Cdd);
HIP_CHECK(hipGetLastError());
HIPCHECK(hipMemcpyAsync(Dd, Cd, size, hipMemcpyDeviceToDevice, stream1));
HIPCHECK(hipMemcpyAsync(Ddd, Cdd, size, hipMemcpyDeviceToDevice, stream2));
HIPCHECK(hipMemcpyAsync(Eh, Dd, size, hipMemcpyDeviceToHost, stream1));
HIPCHECK(hipMemcpyAsync(Ehh, Ddd, size, hipMemcpyDeviceToHost, stream2));
HIPCHECK(hipDeviceSynchronize());
memcpy(snap, Eh, size);
memcpy(snapp, Ehh, size);
memcpy(snap, Eh, size);
memcpy(snapp, Ehh, size);
for (int i = 0; i < N; i++) {
HIPASSERT(snap[i] == Ah[i] + 1.0f);
HIPASSERT(snapp[i] == Ahh[i] + 1.0f);
}
free(snap);
free(snapp);
HIPCHECK(hipHostFree(Ah));
HIPCHECK(hipHostFree(Bh));
HIPCHECK(hipHostFree(Eh));
HIPCHECK(hipHostFree(Ahh));
HIPCHECK(hipHostFree(Bhh));
HIPCHECK(hipHostFree(Ehh));
HIPCHECK(hipFree(Cd));
HIPCHECK(hipFree(Dd));
HIPCHECK(hipFree(Cdd));
HIPCHECK(hipFree(Ddd));
for (int i = 0; i < N; i++) {
HIPASSERT(snap[i] == Ah[i] + 1.0f);
HIPASSERT(snapp[i] == Ahh[i] + 1.0f);
}
free(snap);
free(snapp);
HIPCHECK(hipHostFree(Ah));
HIPCHECK(hipHostFree(Bh));
HIPCHECK(hipHostFree(Eh));
HIPCHECK(hipHostFree(Ahh));
HIPCHECK(hipHostFree(Bhh));
HIPCHECK(hipHostFree(Ehh));
HIPCHECK(hipFree(Cd));
HIPCHECK(hipFree(Dd));
HIPCHECK(hipFree(Cdd));
HIPCHECK(hipFree(Ddd));
}
TEST_CASE("Unit_hipMultiThreadStreams2") {
int iterations = 100;
int iterations = 100;
hipStream_t stream[3];
for (int i = 0; i < 3; i++) {
HIPCHECK(hipStreamCreate(&stream[i]));
}
hipStream_t stream[3];
for (int i = 0; i < 3; i++) {
HIPCHECK(hipStreamCreate(&stream[i]));
}
const size_t size = N * sizeof(float);
for (int i = 0; i < iterations; i++) {
std::thread t1(run1, size, stream[0]);
std::thread t2(run1, size, stream[0]);
std::thread t3(run, size, stream[1], stream[2]);
const size_t size = N * sizeof(float);
for (int i = 0; i < iterations; i++) {
std::thread t1(run1, size, stream[0]);
std::thread t2(run1, size, stream[0]);
std::thread t3(run, size, stream[1], stream[2]);
t1.join();
t2.join();
t3.join();
}
t1.join();
t2.join();
t3.join();
}
for (int i = 0; i < 3; i++) {
HIPCHECK(hipStreamDestroy(stream[i]));
}
for (int i = 0; i < 3; i++) {
HIPCHECK(hipStreamDestroy(stream[i]));
}
}