intel/ds: allow user to select metric set at start time

Rather than using always the same metric set, let the user choose when
starting the producer with :

  INTEL_PERFETTO_METRIC_SET=RasterizerAndPixelBackend ./build/src/tool/pps/pps-producer

Signed-off-by: Lionel Landwerlin <lionel.g.landwerlin@intel.com>
Reviewed-by: Rohan Garg <rohan.garg@intel.com>
Acked-by: Antonio Caggiano <antonio.caggiano@collabora.com>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/13996>

5 files changed, 94 insertions, 98 deletions

diff --git a/docs/perfetto.rst b/docs/perfetto.rst
index 17379e6f770..689031d73e5 100644
--- a/docs/perfetto.rst
+++ b/docs/perfetto.rst
@@ -154,6 +154,13 @@ Another option to enable access wide data without root permissions would be runn
 
 Alternatively using the ``CAP_PERFMON`` permission on the binary should work too.
 
+A particular metric set can also be selected to capture a different
+set of HW counters :
+
+.. code-block:: console
+
+   INTEL_PERFETTO_METRIC_SET=RasterizerAndPixelBackend ./build/src/tool/pps/pps-producer
+
 Panfrost
 ^^^^^^^^
 
diff --git a/src/intel/ds/intel_pps_driver.cc b/src/intel/ds/intel_pps_driver.cc
index 2cd399a12b6..56f08dd64fb 100644
--- a/src/intel/ds/intel_pps_driver.cc
+++ b/src/intel/ds/intel_pps_driver.cc
@@ -58,38 +58,17 @@ IntelDriver::~IntelDriver()
 void IntelDriver::enable_counter(uint32_t counter_id)
 {
    auto &counter = counters[counter_id];
-   auto &group = groups[counter.group];
-   if (perf->query) {
-      if (perf->query->symbol_name != group.name) {
-         PPS_LOG_ERROR(
-            "Unable to enable metrics from different sets: %u "
-            "belongs to %s but %s is currently in use.",
-            counter_id,
-            perf->query->symbol_name,
-            group.name.c_str());
-         return;
-      }
-   }
 
    enabled_counters.emplace_back(counter);
-   if (!perf->query) {
-      perf->query = perf->find_query_by_name(group.name);
-   }
 }
 
 void IntelDriver::enable_all_counters()
 {
-   // We can only enable one metric set at a time so at least enable one.
-   for (auto &group : groups) {
-      if (group.name == "RenderBasic") {
-         for (uint32_t counter_id : group.counters) {
-            auto &counter = counters[counter_id];
-            enabled_counters.emplace_back(counter);
-         }
-
-         perf->query = perf->find_query_by_name(group.name);
-         break;
-      }
+   // We should only have one group
+   assert(groups.size() == 1);
+   for (uint32_t counter_id : groups[0].counters) {
+      auto &counter = counters[counter_id];
+      enabled_counters.emplace_back(counter);
    }
 }
 
@@ -99,49 +78,76 @@ bool IntelDriver::init_perfcnt()
 
    perf = std::make_unique<IntelPerf>(drm_device.fd);
 
+   const char *metric_set_name = getenv("INTEL_PERFETTO_METRIC_SET");
+
+   struct intel_perf_query_info *default_query = nullptr;
+   selected_query = nullptr;
    for (auto &query : perf->get_queries()) {
-      // Create group
-      CounterGroup group = {};
-      group.id = groups.size();
-      group.name = query->symbol_name;
-
-      for (int i = 0; i < query->n_counters; ++i) {
-         intel_perf_query_counter &counter = query->counters[i];
-
-         // Create counter
-         Counter counter_desc = {};
-         counter_desc.id = counters.size();
-         counter_desc.name = counter.symbol_name;
-         counter_desc.group = group.id;
-         counter_desc.getter = [counter, query, this](
-                                  const Counter &c, const Driver &dri) -> Counter::Value {
-            switch (counter.data_type) {
-            case INTEL_PERF_COUNTER_DATA_TYPE_UINT64:
-            case INTEL_PERF_COUNTER_DATA_TYPE_UINT32:
-            case INTEL_PERF_COUNTER_DATA_TYPE_BOOL32:
-               return (int64_t)counter.oa_counter_read_uint64(perf->cfg, query, &perf->result);
-               break;
-            case INTEL_PERF_COUNTER_DATA_TYPE_DOUBLE:
-            case INTEL_PERF_COUNTER_DATA_TYPE_FLOAT:
-               return counter.oa_counter_read_float(perf->cfg, query, &perf->result);
-               break;
-            }
-
-            return {};
-         };
-
-         // Add counter id to the group
-         group.counters.emplace_back(counter_desc.id);
-
-         // Store counter
-         counters.emplace_back(std::move(counter_desc));
+      if (!strcmp(query->symbol_name, "RenderBasic"))
+         default_query = query;
+      if (metric_set_name && !strcmp(query->symbol_name, metric_set_name))
+         selected_query = query;
+   }
+
+   assert(default_query);
+
+   if (!selected_query) {
+      if (metric_set_name) {
+         PPS_LOG_ERROR("Available metric sets:");
+         for (auto &query : perf->get_queries())
+            PPS_LOG_ERROR("   %s", query->symbol_name);
+         PPS_LOG_FATAL("Metric set '%s' not available.", metric_set_name);
       }
+      selected_query = default_query;
+   }
+
+   PPS_LOG("Using metric set '%s': %s",
+           selected_query->symbol_name, selected_query->name);
+
+   // Create group
+   CounterGroup group = {};
+   group.id = groups.size();
+   group.name = selected_query->symbol_name;
+
+   for (int i = 0; i < selected_query->n_counters; ++i) {
+      intel_perf_query_counter &counter = selected_query->counters[i];
+
+      // Create counter
+      Counter counter_desc = {};
+      counter_desc.id = counters.size();
+      counter_desc.name = counter.symbol_name;
+      counter_desc.group = group.id;
+      counter_desc.getter = [counter, this](
+         const Counter &c, const Driver &dri) -> Counter::Value {
+         switch (counter.data_type) {
+         case INTEL_PERF_COUNTER_DATA_TYPE_UINT64:
+         case INTEL_PERF_COUNTER_DATA_TYPE_UINT32:
+         case INTEL_PERF_COUNTER_DATA_TYPE_BOOL32:
+            return (int64_t)counter.oa_counter_read_uint64(perf->cfg,
+                                                           selected_query,
+                                                           &perf->result);
+            break;
+         case INTEL_PERF_COUNTER_DATA_TYPE_DOUBLE:
+         case INTEL_PERF_COUNTER_DATA_TYPE_FLOAT:
+            return counter.oa_counter_read_float(perf->cfg,
+                                                 selected_query,
+                                                 &perf->result);
+            break;
+         }
+
+         return {};
+      };
 
-      // Store group
-      groups.emplace_back(std::move(group));
+      // Add counter id to the group
+      group.counters.emplace_back(counter_desc.id);
+
+      // Store counter
+      counters.emplace_back(std::move(counter_desc));
    }
 
-   assert(groups.size() && "Failed to query groups");
+   // Store group
+   groups.emplace_back(std::move(group));
+
    assert(counters.size() && "Failed to query counters");
 
    // Clear accumulations
@@ -154,7 +160,7 @@ void IntelDriver::enable_perfcnt(uint64_t sampling_period_ns)
 {
    this->sampling_period_ns = sampling_period_ns;
 
-   if (!perf->open(sampling_period_ns)) {
+   if (!perf->open(sampling_period_ns, selected_query)) {
       PPS_LOG_FATAL("Failed to open intel perf");
    }
 }
@@ -197,7 +203,7 @@ std::vector<PerfRecord> IntelDriver::parse_perf_records(const std::vector<uint8_
          // Report is next to the header
          const uint32_t *report = reinterpret_cast<const uint32_t *>(header + 1);
          uint64_t gpu_timestamp_ldw =
-            intel_perf_report_timestamp(&perf->query.value(), report);
+            intel_perf_report_timestamp(selected_query, report);
 
          /* Our HW only provides us with the lower 32 bits of the 36bits
           * timestamp counter value. If we haven't captured the top bits yet,
@@ -292,11 +298,11 @@ uint64_t IntelDriver::gpu_next()
    auto record_b = reinterpret_cast<const drm_i915_perf_record_header *>(records[1].data.data());
 
    intel_perf_query_result_accumulate_fields(&perf->result,
-      &perf->query.value(),
-      &perf->devinfo,
-      record_a + 1,
-      record_b + 1,
-      false /* no_oa_accumulate */);
+                                             selected_query,
+                                             &perf->devinfo,
+                                             record_a + 1,
+                                             record_b + 1,
+                                             false /* no_oa_accumulate */);
 
    // Get last timestamp
    auto gpu_timestamp = records[1].timestamp;
diff --git a/src/intel/ds/intel_pps_driver.h b/src/intel/ds/intel_pps_driver.h
index 404fc651588..fb02327c39f 100644
--- a/src/intel/ds/intel_pps_driver.h
+++ b/src/intel/ds/intel_pps_driver.h
@@ -9,6 +9,10 @@
 
 #include <pps/pps_driver.h>
 
+extern "C" {
+struct intel_perf_query_info;
+};
+
 namespace pps
 {
 
@@ -82,6 +86,9 @@ class IntelDriver : public Driver
 
    // Gpu clock ID used to correlate GPU/CPU timestamps
    uint32_t clock_id = 0;
+
+   // Selected query
+   intel_perf_query_info *selected_query = nullptr;
 };
 
 } // namespace pps
diff --git a/src/intel/ds/intel_pps_perf.cc b/src/intel/ds/intel_pps_perf.cc
index 441d06cd584..f72b80a9c6b 100644
--- a/src/intel/ds/intel_pps_perf.cc
+++ b/src/intel/ds/intel_pps_perf.cc
@@ -36,13 +36,6 @@ IntelPerf::IntelPerf(const int drm_fd)
       false, // no pipeline statistics
       false  // no register snapshots
    );
-
-   // Enable RenderBasic counters
-   auto query_name = "RenderBasic";
-   query = find_query_by_name(query_name);
-   if (!query) {
-      PPS_LOG_FATAL("Failed to find %s query", query_name);
-   }
 }
 
 IntelPerf::~IntelPerf()
@@ -58,20 +51,6 @@ IntelPerf::~IntelPerf()
    }
 }
 
-/// @return A query info, which is something like a group of counters
-std::optional<struct intel_perf_query_info> IntelPerf::find_query_by_name(
-   const std::string &name) const
-{
-   for (int i = 0; i < cfg->n_queries; ++i) {
-      struct intel_perf_query_info query = cfg->queries[i];
-      if (name == query.symbol_name) {
-         return query;
-      }
-   }
-
-   return std::nullopt;
-}
-
 std::vector<struct intel_perf_query_info *> IntelPerf::get_queries() const
 {
    assert(cfg && "Intel perf config should be valid");
@@ -98,7 +77,8 @@ static uint32_t get_oa_exponent(const intel_device_info *devinfo, const uint64_t
    return static_cast<uint32_t>(log2(sampling_period_ns * devinfo->timestamp_frequency / 1000000000ull)) - 1;
 }
 
-bool IntelPerf::open(const uint64_t sampling_period_ns)
+bool IntelPerf::open(const uint64_t sampling_period_ns,
+                     struct intel_perf_query_info *query)
 {
    assert(!ctx && "Perf context should not be initialized at this point");
 
diff --git a/src/intel/ds/intel_pps_perf.h b/src/intel/ds/intel_pps_perf.h
index 0c2831577a5..2db622ae466 100644
--- a/src/intel/ds/intel_pps_perf.h
+++ b/src/intel/ds/intel_pps_perf.h
@@ -23,11 +23,9 @@ class IntelPerf
    IntelPerf(int drm_fd);
    ~IntelPerf();
 
-   std::optional<struct intel_perf_query_info> find_query_by_name(const std::string &name) const;
-
    std::vector<struct intel_perf_query_info*> get_queries() const;
 
-   bool open(uint64_t sampling_period_ns);
+   bool open(uint64_t sampling_period_ns, struct intel_perf_query_info *query);
    void close();
 
    bool oa_stream_ready() const;
@@ -45,8 +43,6 @@ class IntelPerf
    struct intel_perf_query_result result = {};
 
    struct intel_device_info devinfo = {};
-
-   std::optional<struct intel_perf_query_info> query = std::nullopt;
 };
 
 } // namespace pps