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

Change-Id: I484fe9ff7cd56c70a37a3ac5a4a55812f8557259

catch/include/hip_test_common.hh

@@ -140,6 +140,31 @@ static void initHipCtx(hipCtx_t* pcontext) {
 #define HIP_ARRAY hipArray*
 #endif
 
+static inline bool IsGfx11() {
+#if defined(HT_NVIDIA)
+  return false;
+#elif defined(HT_AMD)
+  int device = -1;
+  hipDeviceProp_t props{};
+  HIP_CHECK(hipGetDevice(&device));
+  HIP_CHECK(hipGetDeviceProperties(&props, device));
+
+   // Get GCN Arch Name and compare to check if it is gfx11
+  std::string arch = std::string(props.gcnArchName);
+  auto pos = arch.find(":");
+  if (pos != std::string::npos)
+    arch = arch.substr(0, pos);
+
+  if(arch.size() >= 5)
+    arch = arch.substr(0,5);
+
+  return (arch == std::string("gfx11")) ? true : false;
+#else
+  std::cout<<"Have to be either Nvidia or AMD platform, asserting"<<std::endl;
+  assert(false);
+#endif
+}
+
 
 // Utility Functions
 namespace HipTest {
@@ -335,6 +360,14 @@ static __global__ void waitKernel(clock_t offset) {
   }
 }
 
+static __global__ void waitKernel_gfx11(clock_t offset) {
+#if HT_AMD
+  auto start = wall_clock64();
+  while ((wall_clock64() - start) < offset) {
+  }
+#endif
+}
+
 // helper function used to set the device frequency variable
 // estimates the number of clock ticks in 1 second
 static size_t findTicksPerSecond() {
@@ -350,9 +383,9 @@ static size_t findTicksPerSecond() {
   hipEvent_t start, stop;
   HIP_CHECK(hipEventCreate(&start));
   HIP_CHECK(hipEventCreate(&stop));
-
+  auto waitKernel_used = IsGfx11() ? waitKernel_gfx11 : waitKernel;
   // Warmup
-  hipLaunchKernelGGL(waitKernel, dim3(1), dim3(1), 0, 0, clockTicksPerSecond);
+  hipLaunchKernelGGL(waitKernel_used, dim3(1), dim3(1), 0, 0, clockTicksPerSecond);
   HIP_CHECK(hipGetLastError());
   HIP_CHECK(hipDeviceSynchronize());
 
@@ -360,7 +393,7 @@ static size_t findTicksPerSecond() {
   // after 10 attempts the result is likely good enough so just accept it
   for (int attempts = 10; attempts > 0; --attempts) {
     HIP_CHECK(hipEventRecord(start));
-    hipLaunchKernelGGL(waitKernel, dim3(1), dim3(1), 0, 0, clockTicksPerSecond);
+    hipLaunchKernelGGL(waitKernel_used, dim3(1), dim3(1), 0, 0, clockTicksPerSecond);
     HIP_CHECK(hipEventRecord(stop));
     HIP_CHECK(hipGetLastError());
     HIP_CHECK(hipEventSynchronize(stop));
@@ -396,7 +429,8 @@ static inline void runKernelForDuration(std::chrono::milliseconds duration,
   // precision so that's acceptable.
   static size_t ticksPerSecond = findTicksPerSecond();
   const auto millis = duration.count();
-  hipLaunchKernelGGL(waitKernel, dim3(1), dim3(1), 0, stream, ticksPerSecond * millis / 1000);
+  auto waitKernel_used = IsGfx11() ? waitKernel_gfx11 : waitKernel;
+  hipLaunchKernelGGL(waitKernel_used, dim3(1), dim3(1), 0, stream, ticksPerSecond * millis / 1000);
 }
 
 }  // namespace HipTest

catch/include/utils.hh

@@ -128,7 +128,11 @@ __global__ void Iota(T* const out, size_t pitch, size_t w, size_t h, size_t d) {
 inline void LaunchDelayKernel(const std::chrono::milliseconds interval, const hipStream_t stream) {
   int ticks_per_ms = 0;
   // Clock rate is in kHz => number of clock ticks in a millisecond
-  HIP_CHECK(hipDeviceGetAttribute(&ticks_per_ms, hipDeviceAttributeClockRate, 0));
+  if (IsGfx11()) {
+    HIPCHECK(hipDeviceGetAttribute(&ticks_per_ms, hipDeviceAttributeWallClockRate, 0));
+  } else {
+    HIPCHECK(hipDeviceGetAttribute(&ticks_per_ms, hipDeviceAttributeClockRate, 0));
+  }
   Delay<<<1, 1, 0, stream>>>(interval.count(), ticks_per_ms);
   HIP_CHECK(hipGetLastError());
 }