IP support update for 2.2.0 release

This update contains:

- Warp width detection for G78AE.
- Execution engine count detection for Midgard GPUs.
- Workaround for kernel interface vs. version mismatch on some devices with Mali-G715 GPUs.
- Improved counter naming and counter definitions for many recent CSF GPUs.

Author: markbarnett-arm

CMakeLists.txt

@@ -6,7 +6,7 @@
 
 cmake_minimum_required(VERSION 3.13.5)
 
-project(hwcpipe VERSION 2.1)
+project(hwcpipe VERSION 2.2)
 
 set(CMAKE_CXX_STANDARD 14)
 set(CMAKE_CXX_STANDARD_REQUIRED ON)

backend/device/CMakeLists.txt

@@ -1,5 +1,5 @@
 #
-# Copyright (c) 2021-2022 ARM Limited.
+# Copyright (c) 2021-2023 ARM Limited.
 #
 # SPDX-License-Identifier: MIT
 #
@@ -29,6 +29,8 @@ add_library(
     src/device/hwcnt/reader.cpp
     src/device/hwcnt/sampler/detail/backend.cpp
     src/device/hwcnt/sampler/kinstr_prfcnt/block_index_remap.cpp
+    src/device/hwcnt/sampler/kinstr_prfcnt/enum_info_parser.cpp
+    src/device/hwcnt/sampler/kinstr_prfcnt/metadata_parser.cpp
     src/device/instance.cpp
     src/device/num_exec_engines.cpp
 )

backend/device/include/device/hwcnt/block_extents.hpp

@@ -60,16 +60,16 @@ class block_extents {
 
     /** Number of block types. */
     static const constexpr num_block_types_type num_block_types =
-        static_cast<num_block_types_type>(block_type::num_block_types);
+        static_cast<num_block_types_type>(block_type::last) + 1;
 
-    using num_blocks_of_type_type = std::array<uint16_t, num_block_types>;
+    using num_blocks_of_type_type = std::array<uint8_t, num_block_types>;
 
     /**
      * Construct block extents.
      *
-     * @param num_blocks_of_type_v[in] Array of number of blocks on per type basis.
-     * @param counters_per_block[in] Number of counters per block.
-     * @param values_type[in]        Hardware counters values type.
+     * @param[in] num_blocks_of_type_v Array of number of blocks on per type basis.
+     * @param[in] counters_per_block   Number of counters per block.
+     * @param[in] values_type          Hardware counters values type.
      */
     block_extents(num_blocks_of_type_type num_blocks_of_type_v, uint16_t counters_per_block,
                   sample_values_type values_type)
@@ -86,16 +86,18 @@ class block_extents {
 
     /** @return Number of hardware counters blocks. */
     uint8_t num_blocks() const {
-        return num_blocks_of_type(block_type::fe)       //
-               + num_blocks_of_type(block_type::tiler)  //
-               + num_blocks_of_type(block_type::memory) //
-               + num_blocks_of_type(block_type::core);
+        uint8_t result{};
+
+        for (const auto num_blocks : num_blocks_of_type_)
+            result = static_cast<uint8_t>(result + num_blocks);
+
+        return result;
     }
 
     /**
      * Number of blocks of a given type.
      *
-     * @param type[in]    Block type.
+     * @param[in] type Block type.
      * @return Number of blocks of type @p type.
      */
     uint8_t num_blocks_of_type(block_type type) const { return num_blocks_of_type_[static_cast<size_t>(type)]; }

backend/device/include/device/hwcnt/block_metadata.hpp

@@ -43,11 +43,18 @@ class reader;
 
 /** Hardware counters block type. */
 enum class block_type : uint8_t {
-    fe,              /**< Front End. */
-    tiler,           /**< Tiler. */
-    memory,          /**< Memory System. */
-    core,            /**< Shader Core. */
-    num_block_types, /**< Sentinel. */
+    /** Front End. */
+    fe,
+    /** Tiler. */
+    tiler,
+    /** Memory System. */
+    memory,
+    /** Shader Core. */
+    core,
+    /** First block type. */
+    first = fe,
+    /** Last block type. */
+    last = core,
 };
 
 /**

backend/device/include/device/hwcnt/reader.hpp

@@ -264,9 +264,9 @@ class HWCPIPE_DEVICE_API reader {
     /**
      * Constructor.
      *
-     * @param fd[in]            Counters file descriptor.
-     * @param features_v[in]    Reader features.
-     * @param extents[in]       Hardware counters block extents.
+     * @param[in] fd            Counters file descriptor.
+     * @param[in] features_v    Reader features.
+     * @param[in] extents       Hardware counters block extents.
      */
     reader(int fd, features features_v, block_extents extents)
         : fd_(fd)

backend/device/include/device/hwcnt/sampler/detail/backend.hpp

@@ -55,7 +55,7 @@ class reader;
  *  - @ref configuration to configure which counters to sample for a given
  *    block type.
  *
- * Also contains the low level @ref detail::backend interface,
+ * Also contains the low level @c detail::backend interface,
  * which is used for samplers implementation. The interface is implemented
  * for each hardware counters ioctl API version.
  */

backend/device/include/device/hwcnt/sampler/manual.hpp

@@ -45,6 +45,11 @@ namespace sampler {
 /** Manual hardware counters sampler. */
 class manual {
   public:
+    using backend_type = detail::backend;
+
+    explicit manual(std::unique_ptr<backend_type> backend = nullptr)
+        : backend_(std::move(backend)) {}
+
     /**
      * Constructor.
      *
@@ -66,6 +71,12 @@ class manual {
     manual(const instance &inst, std::initializer_list<configuration> lst)
         : manual(inst, lst.begin(), lst.size()) {}
 
+    /** Default move construct. */
+    manual(manual &&) = default;
+
+    /** Default move assign. */
+    manual &operator=(manual &&) = default;
+
     /**
      * Check if the sampler's back-end was initialized successfully.
      *
@@ -159,7 +170,6 @@ class manual {
     }
 
   private:
-    using backend_type = detail::backend;
     std::unique_ptr<backend_type> backend_;
 };
 

backend/device/include/device/hwcnt/sampler/periodic.hpp

@@ -45,6 +45,11 @@ namespace sampler {
 /** Periodic hardware counters sampler. */
 class periodic {
   public:
+    using backend_type = detail::backend;
+
+    explicit periodic(std::unique_ptr<backend_type> backend = nullptr)
+        : backend_(std::move(backend)) {}
+
     /**
      * Constructor.
      *
@@ -74,6 +79,12 @@ class periodic {
     periodic(const instance &inst, uint64_t period_ns, std::initializer_list<const configuration> lst)
         : periodic(inst, period_ns, lst.begin(), lst.size()) {}
 
+    /** Default move construct. */
+    periodic(periodic &&) = default;
+
+    /** Default move assign. */
+    periodic &operator=(periodic &&) = default;
+
     /**
      * Check if the sampler's back-end was initialized successfully.
      *
@@ -146,7 +157,6 @@ class periodic {
     }
 
   private:
-    using backend_type = detail::backend;
     std::unique_ptr<backend_type> backend_;
 };
 

backend/device/include/device/product_id.hpp

@@ -81,20 +81,33 @@ class product_id {
     /**
      * Construct product id from a GPU id.
      *
-     * @param gpu_id[in] GPU id to parse.
+     * @param[in] raw_product_id Raw unmasked product id.
      */
-    explicit constexpr product_id(uint64_t gpu_id)
-        : product_id_(from_gpu_id(gpu_id)) {}
+    explicit constexpr product_id(uint32_t raw_product_id)
+        : product_id_(from_raw_product_id(raw_product_id)) {}
 
     /**
      * Construct product id using product/arch major versions pair.
      *
-     * @param arch_major[in]       Arch major version.
-     * @param product_major[in]    Product major version.
+     * @param[in] arch_major       Arch major version.
+     * @param[in] product_major    Product major version.
      */
     explicit constexpr product_id(uint32_t arch_major, uint32_t product_major)
         : product_id_(from_versions(arch_major, product_major)) {}
 
+    /**
+     * Construct product id from the raw GPU id value.
+     *
+     * @param gpu_id Raw gpu id.
+     * @return Product id.
+     */
+    static constexpr product_id from_raw_gpu_id(uint64_t gpu_id) {
+        constexpr uint64_t product_id_shift{16};
+        constexpr uint64_t product_id_mask{0xFFFF};
+
+        return product_id{static_cast<uint32_t>((gpu_id >> product_id_shift) & product_id_mask)};
+    }
+
     /** @return Version style of the product ID. */
     constexpr version_style get_version_style() const { return get_version_style(product_id_); }
 
@@ -156,40 +169,40 @@ class product_id {
     /**
      * Get version style of a gpu_id or product_id value.
      *
-     * @param value[in]    Value to get the style from.
+     * @param[in] value    Value to get the style from.
      * @return Version style of @p value.
      */
     constexpr static version_style get_version_style(uint32_t value) {
         if (value == product_id_t60x)
             return version_style::legacy_t60x;
-        else if (value < 0x1000)
+        else if (value < 0x1000U)
             return version_style::legacy_txxx;
         else
             return version_style::arch_product_major;
     }
 
     /**
-     * Get product id from a GPU id.
+     * Get product id from a raw unmasked value.
      *
-     * @param gpu_id[in] GPU id to parse.
+     * @param[in] raw_product_id Raw unmasked product id.
      * @return Product id value.
      */
-    static constexpr uint32_t from_gpu_id(uint64_t gpu_id) {
-        switch (get_version_style(gpu_id)) {
+    static constexpr uint32_t from_raw_product_id(uint32_t raw_product_id) {
+        switch (get_version_style(raw_product_id)) {
         case version_style::legacy_t60x:
         case version_style::legacy_txxx:
-            return gpu_id & product_id_mask_legacy;
+            return raw_product_id & product_id_mask_legacy;
         case version_style::arch_product_major:
         default:
-            return gpu_id & product_id_mask_modern;
+            return raw_product_id & product_id_mask_modern;
         }
     }
 
     /**
      * Get product id using product/arch major versions pair.
      *
-     * @param arch_major[in]       Arch major version.
-     * @param product_major[in]    Product major version.
+     * @param[in] arch_major       Arch major version.
+     * @param[in] product_major    Product major version.
      * @return Product id value.
      */
     static constexpr uint32_t from_versions(uint32_t arch_major, uint32_t product_major) {

backend/device/src/device/detail/enum_operators.hpp

@@ -22,7 +22,7 @@
  * SOFTWARE.
  */
 
-/** @file define_enum_operators.hpp */
+/** @file enum_operators.hpp */
 
 #pragma once