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

SWDEV-337452 - Changing Clock64 to WallClock64 in tests for gfx11. (#78)

Browse Source 
Change-Id: I484fe9ff7cd56c70a37a3ac5a4a55812f8557259
This commit is contained in:
ROCm CI Service Account
2023-01-07 04:35:21 +05:30
committed by GitHub
parent 70c61d55a4
commit 87fac87657
10 changed files with 339 additions and 50 deletions
Download Patch File Download Diff File Expand all files Collapse all files
catch/unit/stream/hipStreamACb_MultiThread.cc
+20 -1
View File
@@ -53,6 +53,24 @@ static __global__ void device_function(float* C_d, float* A_d, size_t Num) {
}
}
static __global__ void device_function_gfx11(float* C_d, float* A_d, size_t Num) {
#if HT_AMD
size_t gputhread = (blockIdx.x * blockDim.x + threadIdx.x);
size_t stride = blockDim.x * gridDim.x;
for (size_t i = gputhread; i < Num; i += stride) {
C_d[i] = A_d[i] * A_d[i];
}
// Delay thread 1 only in the GPU
if (gputhread == 1) {
uint64_t wait_t = 3200000000, start = wall_clock64(), cur;
do {
cur = wall_clock64() - start;
} while (cur < wait_t);
}
#endif
}
static void HIPRT_CB Thread1_Callback(hipStream_t stream, hipError_t status,
void* userData) {
@@ -128,7 +146,8 @@ TEST_CASE("Unit_hipStreamAddCallback_MultipleThreads") {
constexpr unsigned threadsPerBlock = 256;
constexpr unsigned blocks = (N + 255)/threadsPerBlock;
hipLaunchKernelGGL((device_function), dim3(blocks),
auto device_function_used = IsGfx11() ? device_function_gfx11 : device_function;
hipLaunchKernelGGL((device_function_used), dim3(blocks),
dim3(threadsPerBlock), 0,
mystream, C_d, A_d, N);
HIP_CHECK(hipGetLastError());
catch/unit/stream/hipStreamWaitEvent.cc
+19 -4
View File
@@ -94,6 +94,20 @@ __global__ void waitKernel(int clockRate, int seconds) {
}
}
__global__ void waitKernel_gfx11(int clockRate, int seconds) {
#if HT_AMD
auto start = wall_clock64();
auto ms = seconds * 1000;
long long waitTill = clockRate * (long long)ms;
while (1) {
auto end = wall_clock64();
if ((end - start) > waitTill) {
return;
}
}
#endif
}
TEST_CASE("Unit_hipStreamWaitEvent_Default") {
hipStream_t stream{nullptr};
hipEvent_t waitEvent{nullptr};
@@ -111,7 +125,8 @@ TEST_CASE("Unit_hipStreamWaitEvent_Default") {
HIP_CHECK(hipGetDeviceProperties(&prop, deviceId));
auto clockRate = prop.clockRate;
waitKernel<<<1, 1, 0, stream>>>(clockRate, 2); // Wait for 2 seconds
auto waitKernel_used = IsGfx11() ? waitKernel_gfx11 : waitKernel;
waitKernel_used<<<1, 1, 0, stream>>>(clockRate, 2); // Wait for 2 seconds
HIP_CHECK(hipEventRecord(waitEvent, stream));
@@ -145,8 +160,8 @@ TEST_CASE("Unit_hipStreamWaitEvent_DifferentStreams") {
hipDeviceProp_t prop{};
HIP_CHECK(hipGetDeviceProperties(&prop, deviceId));
auto clockRate = prop.clockRate;
waitKernel<<<1, 1, 0, blockedStreamA>>>(clockRate,
auto waitKernel_used = IsGfx11() ? waitKernel_gfx11 : waitKernel;
waitKernel_used<<<1, 1, 0, blockedStreamA>>>(clockRate,
3); // wait for 3 seconds
HIP_CHECK(hipEventRecord(waitEvent, blockedStreamA));
@@ -155,7 +170,7 @@ TEST_CASE("Unit_hipStreamWaitEvent_DifferentStreams") {
HIP_CHECK(hipStreamWaitEvent(streamBlockedOnStreamA, waitEvent, 0));
waitKernel<<<1, 1, 0, streamBlockedOnStreamA>>>(clockRate, 2); // Wait for 2 seconds
waitKernel_used<<<1, 1, 0, streamBlockedOnStreamA>>>(clockRate, 2); // Wait for 2 seconds
HIP_CHECK(hipStreamSynchronize(unblockingStream));