From 3ecdfda2b19a6b86f8c8d5a306dd4ce3befb9c6a Mon Sep 17 00:00:00 2001
From: "Nemallapudi, Jaikrishna" <nemallapudi.jaikrishna@intel.com>
Date: Thu, 7 May 2026 09:35:31 +0000
Subject: [PATCH] intel/pps: classify OA counters into Perfetto counter groups

Implement Driver::classify_counter_groups() for the Intel PPS driver so
that Perfetto's GpuCounterDescriptor exposes a non-empty 'groups' field
for every Intel OA counter.

The classifier matches Intel OA counter names against well-known
prefixes / suffixes / substrings to assign each counter to one or more
Perfetto buckets:

  VERTICES     starts with VS, VERTEX; contains _VERTEX; ends with _VS
  FRAGMENTS    starts with PS, PIXEL, POSTPS, RASTERIZER, SHADER_PIX,
               SAMPLES; ends with _PS
  COMPUTE      starts with CS, ASYNCCS, ASYNC_GPGPU, GPGPU
  MEMORY       starts with GTI, L3, SLM, SHADERMEMORY, LOAD_STORE,
               COLOR_L3, GPU_MEMORY, SAMPLER_
  PRIMITIVES   starts with CLIPPER, SO, IA_, STREAMOUT, STRIPSFAN;
               contains PRIMITIVE
  RAY_TRACING  starts with RT_; contains RAYTRACE or _RT_

A counter may belong to several buckets at once -- for example
CLIPPER_INPUT_VERTEX is both PRIMITIVES and VERTICES -- which is
allowed by the repeated GpuCounterSpec.groups field.

This fixes the Android CTS test
  CtsGpuProfilingDataTest#testProfilingDataProducersAvailable
which previously failed on Intel platforms with:
  Required counter groups missing.
  Found:    []
  Required: [COMPUTE, FRAGMENTS, MEMORY, PRIMITIVES, VERTICES, RAY_TRACING]

Signed-off-by: Nemallapudi, Jaikrishna <nemallapudi.jaikrishna@intel.com>
---
 src/intel/ds/intel_pps_driver.cc | 57 ++++++++++++++++++++++++++++++++
 src/intel/ds/intel_pps_driver.h  |  4 +++
 2 files changed, 61 insertions(+)

diff --git a/src/intel/ds/intel_pps_driver.cc b/src/intel/ds/intel_pps_driver.cc
index b13377d161b..ee81200d98b 100644
--- a/src/intel/ds/intel_pps_driver.cc
+++ b/src/intel/ds/intel_pps_driver.cc
@@ -8,6 +8,8 @@
 
 #include "intel_pps_driver.h"
 
+#include <cctype>
+#include <cstring>
 #include <dirent.h>
 #include <fcntl.h>
 #include <math.h>
@@ -384,4 +386,59 @@ bool IntelDriver::cpu_gpu_timestamp(uint64_t &cpu_timestamp,
    return true;
 }
 
+void IntelDriver::classify_counter_groups(
+   const std::string &cname_in,
+   const std::function<void(CounterGroupType)> &add_group) const
+{
+   // Normalise to upper-case so the heuristics work for both Mixed-case
+   // (e.g. "VsThreads", "PostPsDepthTestFails") and UPPER_SNAKE_CASE
+   // (e.g. "CLIPPER_INPUT_VERTEX") counter names used across Intel OA XMLs.
+   std::string cname;
+   cname.reserve(cname_in.size());
+   for (char c : cname_in)
+      cname.push_back(static_cast<char>(std::toupper(static_cast<unsigned char>(c))));
+
+   auto starts_with = [](const std::string &s, const char *prefix) {
+      return s.compare(0, strlen(prefix), prefix) == 0;
+   };
+   auto ends_with = [](const std::string &s, const char *suffix) {
+      size_t slen = strlen(suffix);
+      return s.size() >= slen && s.compare(s.size() - slen, slen, suffix) == 0;
+   };
+   auto contains = [](const std::string &s, const char *needle) {
+      return s.find(needle) != std::string::npos;
+   };
+
+   if (starts_with(cname, "VS") || starts_with(cname, "VERTEX") ||
+       contains(cname, "_VERTEX") || ends_with(cname, "_VS")) {
+      add_group(CounterGroupType::VERTICES);
+   }
+   if (starts_with(cname, "PS") || starts_with(cname, "PIXEL") ||
+       starts_with(cname, "POSTPS") || starts_with(cname, "RASTERIZER") ||
+       starts_with(cname, "SHADERPIX") || starts_with(cname, "SHADER_PIX") ||
+       starts_with(cname, "SAMPLES") || ends_with(cname, "_PS")) {
+      add_group(CounterGroupType::FRAGMENTS);
+   }
+   if (starts_with(cname, "CS") || starts_with(cname, "ASYNCCS") ||
+       starts_with(cname, "ASYNC_GPGPU") || starts_with(cname, "GPGPU")) {
+      add_group(CounterGroupType::COMPUTE);
+   }
+   if (starts_with(cname, "GTI") || starts_with(cname, "L3") ||
+       starts_with(cname, "SLM") || starts_with(cname, "SHADERMEMORY") ||
+       starts_with(cname, "LOAD_STORE") || starts_with(cname, "LOADSTORE") ||
+       starts_with(cname, "COLOR_L3") || starts_with(cname, "GPU_MEMORY") ||
+       starts_with(cname, "SAMPLER_")) {
+      add_group(CounterGroupType::MEMORY);
+   }
+   if (starts_with(cname, "CLIPPER") || starts_with(cname, "SO") ||
+       starts_with(cname, "IA_") || starts_with(cname, "STREAMOUT") ||
+       starts_with(cname, "STRIPSFAN") || contains(cname, "PRIMITIVE")) {
+      add_group(CounterGroupType::PRIMITIVES);
+   }
+   if (starts_with(cname, "RT_") || contains(cname, "RAYTRACE") ||
+       contains(cname, "_RT_")) {
+      add_group(CounterGroupType::RAY_TRACING);
+   }
+}
+
 } // namespace pps
diff --git a/src/intel/ds/intel_pps_driver.h b/src/intel/ds/intel_pps_driver.h
index 0daabd41682..2ed57ebe79e 100644
--- a/src/intel/ds/intel_pps_driver.h
+++ b/src/intel/ds/intel_pps_driver.h
@@ -52,6 +52,10 @@ class IntelDriver : public Driver
    bool cpu_gpu_timestamp(uint64_t &cpu_timestamp,
                           uint64_t &gpu_timestamp) const override;
 
+   void classify_counter_groups(
+      const std::string &counter_name,
+      const std::function<void(CounterGroupType)> &add_group) const override;
+
    private:
    /// @brief Requests the next perf sample
    /// @return The sample GPU timestamp