Fix profiling timestamps after resume from suspend

On gfx1151 (Strix Halo), the GPU timestamp counter used by CP, SDMA and
s_memrealtime restarts from 0 on resume from s2idle, while the KFD clock
counter (SMUIO golden TSC) keeps counting. Since 6e6e616744 (#5093), ROCr
rejects all dispatch and copy timestamps that are older than agent
creation, so hipEventElapsedTime returns 0 or negative values and MIOpen
Find mode fails.

Estimate the epoch offset, like it is done on Windows (a54bd1c7b9) and
proposed for the opposite direction in
https://github.com/ROCm/rocm-systems/pull/12577 and
https://github.com/ROCm/rocm-systems/issues/9018.
--- a/runtime/hsa-runtime/core/inc/amd_gpu_agent.h
+++ b/runtime/hsa-runtime/core/inc/amd_gpu_agent.h
@@ -618,6 +618,9 @@ class GpuAgent : public GpuAgentInt {
   // @brief Update ::t1_ tick count.
   void SyncClocks();
 
+  // @brief Validates raw timestamps of a signal and updates gpu_clock_offset_.
+  bool CheckRawTs(uint64_t start, uint64_t end);
+
   // @brief Binds the second-level trap handler to this node.
   void BindTrapHandler();
 
@@ -721,7 +724,8 @@ class GpuAgent : public GpuAgentInt {
   // @brief Offset (in GPU ticks) between AQL dispatch timestamp clock and
   // D3DKMTQueryClockCalibration GPU clock.  Non-zero on Windows where the two
   // GPU clock domains share frequency but differ in epoch.  Computed on first
-  // TranslateTime call; zero on Linux (clocks match).
+  // TranslateTime call.  On Linux, non-zero after the GPU timestamp counter
+  // restarted on resume from suspend.
   int64_t gpu_clock_offset_;
 
   // @brief s_memrealtime nominal clock frequency
--- a/runtime/hsa-runtime/core/runtime/amd_gpu_agent.cpp
+++ b/runtime/hsa-runtime/core/runtime/amd_gpu_agent.cpp
@@ -3197,7 +3197,7 @@ void GpuAgent::TranslateTime(core::Signal* signal, hsa_amd_profiling_dispatch_ti
   uint64_t start, end;
   signal->GetRawTs(false, start, end);
 
-  if ((start == 0) || (end == 0) || (start < t0_.GPUClockCounter) || (end < t0_.GPUClockCounter)) {
+  if (!CheckRawTs(start, end)) {
     debug_print("Signal %p time stamps may be invalid (start=%" PRIu64 ", end=%" PRIu64 ", t0=%" PRIu64 ").\n",
                 &signal->signal_, start, end, t0_.GPUClockCounter);
     time.start = 0;
@@ -3215,7 +3215,7 @@ void GpuAgent::TranslateTime(core::Signal* signal, hsa_amd_profiling_async_copy_
   uint64_t start, end;
   signal->GetRawTs(true, start, end);
 
-  if ((start == 0) || (end == 0) || (start < t0_.GPUClockCounter) || (end < t0_.GPUClockCounter)) {
+  if (!CheckRawTs(start, end)) {
     debug_print("Signal %p async copy time stamps may be invalid (start=%" PRIu64 ", end=%" PRIu64 ", t0=%" PRIu64 ").\n",
                 &signal->signal_, start, end, t0_.GPUClockCounter);
     time.start = 0;
@@ -3229,6 +3229,24 @@ void GpuAgent::TranslateTime(core::Signal* signal, hsa_amd_profiling_async_copy_
   time.start = TranslateTime(start);
 }
 
+bool GpuAgent::CheckRawTs(uint64_t start, uint64_t end) {
+  if ((start == 0) || (end == 0)) return false;
+
+  std::lock_guard<std::mutex> lock(t1_lock_);
+#ifndef _WIN32
+  // The GPU timestamp counter (CP, SDMA, s_memrealtime) can restart from 0 on resume from
+  // suspend (e.g. s2idle on gfx1151), while the KFD clock counter keeps counting.  Ticks
+  // written after agent creation can't predate t0_, so estimate the new epoch offset,
+  // assuming that the packet has just completed.
+  if ((end >= start) && (end - gpu_clock_offset_ < t0_.GPUClockCounter)) {
+    SyncClocks();
+    gpu_clock_offset_ = int64_t(end - t1_.GPUClockCounter);
+  }
+#endif
+  return (start - gpu_clock_offset_ >= t0_.GPUClockCounter) &&
+      (end - gpu_clock_offset_ >= t0_.GPUClockCounter);
+}
+
 /*
 Times during program execution are interpolated to adjust for relative clock drift.
 Interval timing may appear as ticks well before process start, leading to large errors due to
@@ -3244,13 +3262,11 @@ uint64_t GpuAgent::TranslateTime(uint64_t tick) {
 
   std::lock_guard<std::mutex> lock(t1_lock_);
 
-#ifdef _WIN32
-  // On Windows, AQL dispatch timestamps may have a fixed epoch offset from
-  // D3DKMTQueryClockCalibration's GPUClockCounter (same clock domain, different
-  // base).  Subtract the offset before interpolation (0 until first detection).
+  // AQL dispatch timestamps may have an epoch offset from the clock counters (fixed
+  // on Windows, after resume from suspend on Linux).  Subtract the offset before
+  // interpolation (0 until first detection).
   // gpu_clock_offset_ is read and written under t1_lock_ to avoid data races.
   tick -= gpu_clock_offset_;
-#endif
   // Limit errors due to correlated pair certainty to ~0.5us.
   // extrapolated time < (0.5us / half clock read certainty) * delay between clock measures
   // clock read certainty is <4us.