pps: allow drivers to report timestamps in their own time domain

For this each driver must :

  - report its clock_id (if no particular clock just default to cpu
    boottime one)

  - be able to sample its clock (gpu_timestamp())

The PPSDataSource will then emit timestamp correlation events in the
trace ensuring perfetto is able to display GPU & CPU events
appropriately on its timeline.

Signed-off-by: Lionel Landwerlin <lionel.g.landwerlin@intel.com>
Reviewed-by: Antonio Caggiano <antonio.caggiano@collabora.com>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/13831>

9 files changed, 109 insertions, 12 deletions

diff --git a/src/freedreno/ds/fd_pps_driver.cc b/src/freedreno/ds/fd_pps_driver.cc
index 88697bf6111..97305c911ae 100644
--- a/src/freedreno/ds/fd_pps_driver.cc
+++ b/src/freedreno/ds/fd_pps_driver.cc
@@ -389,4 +389,16 @@ FreedrenoDriver::counter(std::string name, Counter::Units units,
    return counter;
 }
 
+uint32_t
+FreedrenoDriver::gpu_clock_id() const
+{
+   return perfetto::protos::pbzero::BUILTIN_CLOCK_BOOTTIME;
+}
+
+uint64_t
+FreedrenoDriver::gpu_timestamp() const
+{
+   return perfetto::base::GetBootTimeNs().count();
+}
+
 } // namespace pps
diff --git a/src/freedreno/ds/fd_pps_driver.h b/src/freedreno/ds/fd_pps_driver.h
index 8a738e26dbd..29f738b313e 100644
--- a/src/freedreno/ds/fd_pps_driver.h
+++ b/src/freedreno/ds/fd_pps_driver.h
@@ -28,6 +28,8 @@ public:
    void disable_perfcnt() override;
    bool dump_perfcnt() override;
    uint64_t next() override;
+   uint32_t gpu_clock_id() const override;
+   uint64_t gpu_timestamp() const override;
 
 private:
    struct fd_device *dev;
diff --git a/src/intel/ds/intel_pps_driver.cc b/src/intel/ds/intel_pps_driver.cc
index 6c79acd0ad0..6966db8fdc1 100644
--- a/src/intel/ds/intel_pps_driver.cc
+++ b/src/intel/ds/intel_pps_driver.cc
@@ -400,4 +400,14 @@ uint64_t IntelDriver::next()
    return cpu_next();
 }
 
+uint32_t IntelDriver::gpu_clock_id() const
+{
+   return perfetto::protos::pbzero::BUILTIN_CLOCK_BOOTTIME;
+}
+
+uint64_t IntelDriver::gpu_timestamp() const
+{
+   return perfetto::base::GetBootTimeNs().count();
+}
+
 } // namespace pps
diff --git a/src/intel/ds/intel_pps_driver.h b/src/intel/ds/intel_pps_driver.h
index 5df48e2f4ec..cac029a973e 100644
--- a/src/intel/ds/intel_pps_driver.h
+++ b/src/intel/ds/intel_pps_driver.h
@@ -51,7 +51,10 @@ class IntelDriver : public Driver
    void disable_perfcnt() override;
    bool dump_perfcnt() override;
    uint64_t next() override;
+   uint32_t gpu_clock_id() const override;
+   uint64_t gpu_timestamp() const override;
 
+   private:
    /// @brief Requests the next perf sample
    /// @return The sample GPU timestamp
    uint32_t gpu_next();
diff --git a/src/panfrost/ds/pan_pps_driver.cc b/src/panfrost/ds/pan_pps_driver.cc
index fbe9263748d..0453475b46d 100644
--- a/src/panfrost/ds/pan_pps_driver.cc
+++ b/src/panfrost/ds/pan_pps_driver.cc
@@ -157,4 +157,14 @@ void PanfrostDriver::disable_perfcnt()
    enabled_counters.clear();
 }
 
+uint32_t PanfrostDriver::gpu_clock_id() const
+{
+   return perfetto::protos::pbzero::BUILTIN_CLOCK_BOOTTIME;
+}
+
+uint64_t PanfrostDriver::gpu_timestamp() const
+{
+   return perfetto::base::GetBootTimeNs().count();
+}
+
 } // namespace pps
diff --git a/src/panfrost/ds/pan_pps_driver.h b/src/panfrost/ds/pan_pps_driver.h
index de764b18aae..f6476d9dee3 100644
--- a/src/panfrost/ds/pan_pps_driver.h
+++ b/src/panfrost/ds/pan_pps_driver.h
@@ -41,6 +41,8 @@ class PanfrostDriver : public Driver
    void disable_perfcnt() override;
    bool dump_perfcnt() override;
    uint64_t next() override;
+   uint32_t gpu_clock_id() const override;
+   uint64_t gpu_timestamp() const override;
 
    uint64_t last_dump_ts = 0;
 
diff --git a/src/tool/pps/pps_datasource.cc b/src/tool/pps/pps_datasource.cc
index 3b5a083b80f..9eac41b9ade 100644
--- a/src/tool/pps/pps_datasource.cc
+++ b/src/tool/pps/pps_datasource.cc
@@ -18,6 +18,8 @@
 // Minimum supported sampling period in nanoseconds
 #define MIN_SAMPLING_PERIOD_NS 50000
 
+#define CORRELATION_TIMESTAMP_PERIOD (1000000000ull)
+
 namespace pps
 {
 static std::string driver_name;
@@ -232,11 +234,37 @@ void add_samples(perfetto::protos::pbzero::GpuCounterEvent &event, const Driver
    }
 }
 
+void add_timestamp(perfetto::protos::pbzero::ClockSnapshot *event, const Driver *driver)
+{
+   uint32_t gpu_clock_id = driver->gpu_clock_id();
+   if (perfetto::protos::pbzero::BUILTIN_CLOCK_BOOTTIME == gpu_clock_id)
+      return;
+
+   // Send a correlation event between GPU & CPU timestamps
+   uint64_t cpu_ts = perfetto::base::GetBootTimeNs().count();
+   uint64_t gpu_ts = driver->gpu_timestamp();
+
+   {
+      auto clock = event->add_clocks();
+
+      clock->set_clock_id(perfetto::protos::pbzero::BUILTIN_CLOCK_BOOTTIME);
+      clock->set_timestamp(cpu_ts);
+   }
+
+   {
+      auto clock = event->add_clocks();
+
+      clock->set_clock_id(gpu_clock_id);
+      clock->set_timestamp(gpu_ts);
+   }
+}
+
 void GpuDataSource::trace(TraceContext &ctx)
 {
    using namespace perfetto::protos::pbzero;
 
-   if (auto state = ctx.GetIncrementalState(); state->was_cleared) {
+   auto state = ctx.GetIncrementalState();
+   if (state->was_cleared) {
       // Mark any incremental state before this point invalid
       {
          auto packet = ctx.NewTracePacket();
@@ -248,14 +276,23 @@ void GpuDataSource::trace(TraceContext &ctx)
       descriptor_timestamp = perfetto::base::GetBootTimeNs().count();
       packet->set_timestamp(descriptor_timestamp);
 
-      auto event = packet->set_gpu_counter_event();
-      event->set_gpu_id(driver->drm_device.gpu_num);
+      {
+         auto event = packet->set_gpu_counter_event();
+         event->set_gpu_id(driver->drm_device.gpu_num);
 
-      auto &groups = driver->groups;
-      auto &counters = driver->enabled_counters;
-      PPS_LOG("Sending counter descriptors");
-      add_descriptors(event, groups, counters, *driver);
+         auto &groups = driver->groups;
+         auto &counters = driver->enabled_counters;
+         PPS_LOG("Sending counter descriptors");
+         add_descriptors(event, groups, counters, *driver);
+      }
 
+      {
+         last_correlation_timestamp = perfetto::base::GetBootTimeNs().count();
+         auto event = packet->set_clock_snapshot();
+         add_timestamp(event, driver);
+      }
+
+      descriptor_gpu_timestamp = driver->gpu_timestamp();
       state->was_cleared = false;
    }
 
@@ -272,15 +309,16 @@ void GpuDataSource::trace(TraceContext &ctx)
    sched_setscheduler(0, sched_policy, &priority_param);
 
    if (driver->dump_perfcnt()) {
-      while (auto timestamp = driver->next()) {
-         if (timestamp <= descriptor_timestamp) {
+      while (auto gpu_timestamp = driver->next()) {
+         if (gpu_timestamp <= descriptor_gpu_timestamp) {
             // Do not send counter values before counter descriptors
             PPS_LOG_ERROR("Skipping counter values coming before descriptors");
             continue;
          }
 
          auto packet = ctx.NewTracePacket();
-         packet->set_timestamp(timestamp);
+         packet->set_timestamp_clock_id(driver->gpu_clock_id());
+         packet->set_timestamp(gpu_timestamp);
 
          auto event = packet->set_gpu_counter_event();
          event->set_gpu_id(driver->drm_device.gpu_num);
@@ -289,6 +327,14 @@ void GpuDataSource::trace(TraceContext &ctx)
       }
    }
 
+   uint64_t cpu_ts = perfetto::base::GetBootTimeNs().count();
+   if ((cpu_ts - last_correlation_timestamp) > CORRELATION_TIMESTAMP_PERIOD) {
+      auto packet = ctx.NewTracePacket();
+      auto event = packet->set_clock_snapshot();
+      add_timestamp(event, driver);
+      last_correlation_timestamp = cpu_ts;
+   }
+
    // Reset normal scheduler
    sched_setscheduler(0, prev_sched_policy, &prev_priority_param);
 }
diff --git a/src/tool/pps/pps_datasource.h b/src/tool/pps/pps_datasource.h
index 96a83b5ae97..f2946cbd6a5 100644
--- a/src/tool/pps/pps_datasource.h
+++ b/src/tool/pps/pps_datasource.h
@@ -53,12 +53,18 @@ class GpuDataSource : public perfetto::DataSource<GpuDataSource, GpuDataSourceTr
    /// Used to check whether the datasource is quick enough
    std::chrono::nanoseconds time_to_trace;
 
+   /// Last CPU timestamp at which we correlated CPU/GPU timestamps
+   uint64_t last_correlation_timestamp = 0;
+
    /// A data source supports one driver at a time, but if you need more
    /// than one gpu datasource you can just run another producer
    Driver *driver = nullptr;
 
-   /// Timestamp of packet sent with counter descriptors
+   /// CPU timestamp of packet sent with counter descriptors
    uint64_t descriptor_timestamp = 0;
+
+   /// GPU timestamp of packet sent with counter descriptors
+   uint64_t descriptor_gpu_timestamp = 0;
 };
 
 } // namespace pps
diff --git a/src/tool/pps/pps_driver.h b/src/tool/pps/pps_driver.h
index 55849d07ada..8d1c59631ac 100644
--- a/src/tool/pps/pps_driver.h
+++ b/src/tool/pps/pps_driver.h
@@ -72,9 +72,15 @@ class Driver
 
    /// @brief After dumping performance counters, with this function you can iterate
    /// through the samples collected.
-   /// @return The CPU timestamp associated to current sample, or 0 if there are no more samples
+   /// @return The GPU timestamp associated to current sample, or 0 if there are no more samples
    virtual uint64_t next() = 0;
 
+   /// Clock ID in which the values returned by gpu_timestamp() belong
+   virtual uint32_t gpu_clock_id() const = 0;
+
+   /// Sample a timestamp from the GPU
+   virtual uint64_t gpu_timestamp() const = 0;
+
    DrmDevice drm_device;
 
    /// List of counter groups