Add support for derived metrics (#65)

Adds support for derived metrics, which combine multiple hardware counters
into a single value.

Author: bengaineyarm

examples/api_example.cpp

@@ -29,6 +29,17 @@ const char *get_product_family_name(hwcpipe::device::product_id::gpu_family f) {
     }
 }
 
+void print_sample_value(const hwcpipe::counter_sample &sample) {
+    switch (sample.type) {
+    case hwcpipe::counter_sample::type::uint64:
+        std::cout << sample.value.uint64;
+        return;
+    case hwcpipe::counter_sample::type::float64:
+        std::cout << sample.value.float64;
+        return;
+    }
+}
+
 int main() {
     // Detect all GPUs & print some info
     for (const auto &gpu : hwcpipe::find_gpus()) {
@@ -84,6 +95,11 @@ int main() {
         std::cout << "Fragment Active Cycles counter not supported by this GPU." << std::endl;
         return -1;
     }
+    ec = config.add_counter(MaliGeomSampleCullRate);
+    if (ec) {
+        std::cout << "Geometry Sample Cull Rate counter not supported by this GPU." << std::endl;
+        return -1;
+    }
 
     auto sampler = hwcpipe::sampler<>(config);
 
@@ -109,19 +125,32 @@ int main() {
             std::cout << ec.message() << std::endl;
             continue;
         }
-        uint64_t active_cycles = sample.value;
+
+        std::cout << "GPU Active cycles [";
+        print_sample_value(sample);
+        std::cout << "] ; ";
 
         // Fetch the Fragment Active Cycles counter
         ec = sampler.get_counter_value(MaliFragActiveCy, sample);
         if (ec) {
             std::cout << ec.message() << std::endl;
             continue;
         }
-        uint64_t fragment_cycles = sample.value;
 
-        std::cout << "    " << std::setw(3) << i << ": "
-                  << "GPU Active Cycles [" << active_cycles << "] ; "
-                  << "Fragment Active Cycles [" << fragment_cycles << "]" << std::endl;
+        std::cout << "Fragment Active Cycles [";
+        print_sample_value(sample);
+        std::cout << "] ; ";
+
+        // Fetch the Geometry Sample Cull Rate counter
+        ec = sampler.get_counter_value(MaliGeomSampleCullRate, sample);
+        if (ec) {
+            std::cout << ec.message() << std::endl;
+            continue;
+        }
+
+        std::cout << "Geometry Sample Cull Rate [";
+        print_sample_value(sample);
+        std::cout << "]" << std::endl;
     }
 
     ec = sampler.stop_sampling();

hwcpipe/CMakeLists.txt

@@ -11,6 +11,7 @@ add_library(hwcpipe
     src/hwcpipe/detail/counter_database.cpp
     src/hwcpipe/all_gpu_counters.cpp
     src/hwcpipe/counter_metadata.cpp
+    src/hwcpipe/derived_functions.cpp
 )
 
 if(HWCPIPE_PIC)

hwcpipe/include/hwcpipe/detail/counter_database.hpp

@@ -15,7 +15,7 @@
 namespace hwcpipe {
 namespace detail {
 
-using backing_map_type = std::unordered_map<hwcpipe_counter, block_offset>;
+using backing_map_type = std::unordered_map<hwcpipe_counter, detail::counter_definition>;
 
 /**
  * @brief An type that provides an enumerable view over the counters for a
@@ -138,7 +138,7 @@ class counter_database {
      * @brief Queries the database to find the block/offset address of a counter
      * for the specified GPU.
      */
-    HWCP_NODISCARD block_offset get_counter_address(gpu_id_type id, hwcpipe_counter counter, std::error_code &ec);
+    HWCP_NODISCARD counter_definition get_counter_def(gpu_id_type id, hwcpipe_counter counter, std::error_code &ec);
 };
 
 } // namespace detail

hwcpipe/include/hwcpipe/detail/internal_types.hpp

@@ -6,6 +6,7 @@
 
 #pragma once
 
+#include "hwcpipe/hwcpipe_counter.h"
 #include "hwcpipe/types.hpp"
 
 #include <device/handle.hpp>
@@ -14,24 +15,179 @@
 #include <device/hwcnt/sampler/manual.hpp>
 #include <device/instance.hpp>
 
+#include <cstdint>
+#include <initializer_list>
+#include <system_error>
 namespace hwcpipe {
 namespace detail {
 
+namespace expression {
 /**
- * Structure representing the block type/offset address of a counter within
- * a particular GPU's PMU data.
+ * The expression evaluation context provides an abstraction over some block of
+ * hardware counter storage. It breaks an otherwise cyclic dependency between
+ * the sampler class & the expression evaluator function type.
+ */
+class context {
+  public:
+    /**
+     * @brief Returns the value a hardware counter from the sampler to be used
+     * in the expression evaluator
+     *
+     * @param counter Hardware counter to get the value for.
+     * @return the counter value
+     */
+    HWCP_NODISCARD virtual double get_counter_value(hwcpipe_counter counter) const = 0;
+
+    /**
+     * @brief Returns the AXI bus width in bytes
+     */
+    HWCP_NODISCARD virtual double get_mali_config_ext_bus_byte_size() const = 0;
+
+    /**
+     * @brief Returns the number of shader core (constant) to be used
+     * in evaluator function
+     */
+    HWCP_NODISCARD virtual double get_mali_config_shader_core_count() const = 0;
+
+    /**
+     * @brief
+     */
+    HWCP_NODISCARD virtual double get_mali_config_l2_cache_count() const = 0;
+};
+
+// Signature for generated evaluation functions.
+using evaluator = double (*)(const context &);
+/**
+ * Holds information about the expression that the sampler will need when
+ * registering the counters and evaluating.
+ */
+struct expression_definition {
+    /**
+     * Pointer to the function that will evaluate the expression and return the
+     * calculated result.
+     */
+    evaluator eval;
+    /**
+     * The list of counters that this expression depends on. These will need to
+     * be implicitly registered with the sampler so that they can be collected
+     * when it is polled.
+     */
+    const std::initializer_list<hwcpipe_counter> dependencies;
+};
+
+} // namespace expression
+/**
+ * Structure representing the block type/offset address and shift
+ * scaling of a counter within a particular GPU's PMU data.
  */
 struct block_offset {
     uint32_t offset;
+    uint32_t shift;
     hwcpipe::device::hwcnt::block_type block_type;
 };
 
+/**
+ * Counters can either be raw hardware counters, or they are expressions based
+ * on some combination of hardware counters, constants or other expressions.
+ *
+ * For hardware counters, we need to know their block/offset addresses so that
+ * we can read them as we iterate over the block types. For expression counters
+ * we need to call the function that will evaluate the formula.
+ */
+struct counter_definition {
+    enum class type { invalid, hardware, expression };
+    union u {
+        block_offset address{};
+        expression::expression_definition expression;
+        explicit u(expression::expression_definition expression)
+            : expression(expression) {}
+        explicit u(block_offset address)
+            : address(address) {}
+    } u;
+    type tag;
+
+    counter_definition()
+        : tag(type::invalid)
+        , u(block_offset{0, 0, device::hwcnt::block_type::fe}) {}
+    explicit counter_definition(expression::expression_definition expression)
+        : tag(type::expression)
+        , u(expression) {}
+    explicit counter_definition(block_offset address)
+        : tag(type::hardware)
+        , u(address) {}
+};
+
 struct hwcpipe_backend_policy {
     using handle_type = device::handle;
     using instance_type = device::instance;
     using sampler_type = device::hwcnt::sampler::manual;
     using sample_type = device::hwcnt::sample;
 };
 
+/**
+ * This class should **only** be used in the derived expressions (derived_functions.cpp).
+ *
+ * Used to override the result of 0.0 / 0.0 to return 0.0. Operators for other types included
+ * to resolve ambiguity.
+ */
+class hwcpipe_double {
+  public:
+    double value;
+
+    hwcpipe_double(double value)
+        : value(value) {}
+
+    hwcpipe_double operator+(const hwcpipe_double &other) const { return hwcpipe_double(value + other.value); }
+
+    hwcpipe_double operator-(const hwcpipe_double &other) const { return hwcpipe_double(value - other.value); }
+
+    hwcpipe_double operator*(const hwcpipe_double &other) const { return hwcpipe_double(value * other.value); }
+
+    hwcpipe_double operator/(const hwcpipe_double &other) const {
+        if (value == 0.0 && other.value == 0.0) {
+            return hwcpipe_double(0.0);
+        }
+        return hwcpipe_double(value / other.value);
+    }
+
+    hwcpipe_double operator+(const int &other) const { return hwcpipe_double(value + other); }
+
+    hwcpipe_double operator-(const int &other) const { return hwcpipe_double(value - other); }
+
+    hwcpipe_double operator*(const int &other) const { return hwcpipe_double(value * other); }
+
+    hwcpipe_double operator/(const int &other) const {
+        if (value == 0.0 && other == 0) {
+            return hwcpipe_double(0.0);
+        }
+        return hwcpipe_double(value / other);
+    }
+
+    hwcpipe_double operator+(const double &other) const { return hwcpipe_double(value + other); }
+
+    hwcpipe_double operator-(const double &other) const { return hwcpipe_double(value - other); }
+
+    hwcpipe_double operator*(const double &other) const { return hwcpipe_double(value * other); }
+
+    hwcpipe_double operator/(const double &other) const {
+        if (value == 0.0 && other == 0.0) {
+            return hwcpipe_double(0.0);
+        }
+        return hwcpipe_double(value / other);
+    }
+
+    bool operator<(const hwcpipe_double &other) const { return value < other.value; }
+
+    bool operator<=(const hwcpipe_double &other) const { return value <= other.value; }
+
+    bool operator>(const hwcpipe_double &other) const { return value > other.value; }
+
+    bool operator>=(const hwcpipe_double &other) const { return value >= other.value; }
+
+    bool operator==(const hwcpipe_double &other) const { return value == other.value; }
+
+    operator double() const { return value; }
+};
+
 } // namespace detail
 } // namespace hwcpipe