[Harvesting] noc_translation_enabled added to tt_SocDescriptor (#443)

### Issue
Related to #439 

### Description
The last missing feature of Coordinate Manager is noc_translation
enabled feature. If this is disabled, translated coords will be same as
physical coords.

### List of the changes
- Add bool noc_translation_enabled to CoordinateManager constructor
- Added default versions of translate-physical mapping, which do
identity mapping.
- Added a test to test if this flag works
- The translation flag is now required, so added it explicitly to tests
where needed.

### Testing
Added another test, existing CI tests.

### API Changes
There are no API changes in this PR.

device/api/umd/device/blackhole_coordinate_manager.h

@@ -12,6 +12,7 @@
 class BlackholeCoordinateManager : public CoordinateManager {
 public:
     BlackholeCoordinateManager(
+        const bool noc_translation_enabled,
         const tt_xy_pair& tensix_grid_size,
         const std::vector<tt_xy_pair>& tensix_cores,
         const size_t tensix_harvesting_mask,
@@ -29,14 +30,15 @@ class BlackholeCoordinateManager : public CoordinateManager {
 protected:
     void assert_coordinate_manager_constructor() override;
 
-    void translate_dram_coords() override;
     void translate_tensix_coords() override;
+    void translate_dram_coords() override;
     void translate_eth_coords() override;
 
     void fill_tensix_physical_translated_mapping() override;
+    void fill_dram_physical_translated_mapping() override;
     void fill_eth_physical_translated_mapping() override;
     void fill_pcie_physical_translated_mapping() override;
-    void fill_dram_physical_translated_mapping() override;
+    void fill_arc_physical_translated_mapping() override;
 
     std::vector<tt::umd::CoreCoord> get_tensix_cores() const override;
     std::vector<tt::umd::CoreCoord> get_harvested_tensix_cores() const override;

device/api/umd/device/cluster.h

@@ -868,8 +868,16 @@ class Cluster : public tt_device {
 
     // This functions has to be called for local chip, and then it will wait for all connected remote chips to flush.
     void wait_for_connected_non_mmio_flush(chip_id_t chip_id);
+
+    // Helper functions for constructing the chips from the cluster descriptor.
     std::unique_ptr<Chip> construct_chip_from_cluster(
         chip_id_t chip_id, tt_ClusterDescriptor* cluster_desc, tt_SocDescriptor& soc_desc);
+    std::unique_ptr<Chip> construct_chip_from_cluster(
+        const std::string& soc_desc_path,
+        chip_id_t chip_id,
+        tt_ClusterDescriptor* cluster_desc,
+        bool perform_harvesting,
+        std::unordered_map<chip_id_t, uint32_t>& simulated_harvesting_masks);
     std::unique_ptr<Chip> construct_chip_from_cluster(
         chip_id_t logical_device_id,
         tt_ClusterDescriptor* cluster_desc,
@@ -887,6 +895,8 @@ class Cluster : public tt_device {
         const bool clean_system_resources,
         bool perform_harvesting,
         std::unordered_map<chip_id_t, uint32_t> simulated_harvesting_masks);
+
+    // Helper function for translating chip coordinates.
     tt::umd::CoreCoord translate_chip_coord(
         const chip_id_t chip, const tt::umd::CoreCoord core_coord, const CoordSystem coord_system) const;
 

device/api/umd/device/coordinate_manager.h

@@ -24,6 +24,7 @@ class CoordinateManager {
      */
     static std::shared_ptr<CoordinateManager> create_coordinate_manager(
         tt::ARCH arch,
+        const bool noc_translation_enabled,
         const tt_xy_pair& tensix_grid_size,
         const std::vector<tt_xy_pair>& tensix_cores,
         const size_t tensix_harvesting_mask,
@@ -40,6 +41,7 @@ class CoordinateManager {
 
     static std::shared_ptr<CoordinateManager> create_coordinate_manager(
         tt::ARCH arch,
+        const bool noc_translation_enabled,
         const size_t tensix_harvesting_mask = 0,
         const size_t dram_harvesting_mask = 0,
         const size_t eth_harvesting_mask = 0,
@@ -88,6 +90,7 @@ class CoordinateManager {
      * layout of the tensix cores.
      */
     CoordinateManager(
+        const bool noc_translation_enabled,
         const tt_xy_pair& tensix_grid_size,
         const std::vector<tt_xy_pair>& tensix_cores,
         const size_t tensix_harvesting_mask,
@@ -129,13 +132,23 @@ class CoordinateManager {
     virtual tt_xy_pair get_harvested_eth_grid_size() const;
 
     /*
-     * Fills the logical to translated mapping for the tensix cores.
+     * By default, translated coordinates are the same as physical coordinates.
+     * This will be true for all architectures if noc_translation_enabled is false.
+     */
+    void fill_tensix_default_physical_translated_mapping();
+    void fill_eth_default_physical_translated_mapping();
+    void fill_dram_default_physical_translated_mapping();
+    void fill_pcie_default_physical_translated_mapping();
+    void fill_arc_default_physical_translated_mapping();
+
+    /*
+     * Fills the physical to translated mapping for the tensix cores.
      * By default, translated coordinates are the same as physical coordinates.
      * Derived coordinate managers that need to implement different mapping
      * should override this method. Wormhole and Blackhole coordinate managers
      * override this method to implement different mapping.
      */
-    virtual void fill_tensix_physical_translated_mapping();
+    virtual void fill_tensix_physical_translated_mapping() = 0;
 
     /*
      * Fills the physical to translated mapping for the ethernet cores.
@@ -144,7 +157,7 @@ class CoordinateManager {
      * should override this method. Wormhole and Blackhole coordinate managers
      * override this method to implement different mapping.
      */
-    virtual void fill_eth_physical_translated_mapping();
+    virtual void fill_eth_physical_translated_mapping() = 0;
 
     /*
      * Fills the physical to translated mapping for the DRAM cores.
@@ -153,7 +166,7 @@ class CoordinateManager {
      * should override this method. Blackhole coordinate manager overrides
      * this method to implement different mapping.
      */
-    virtual void fill_dram_physical_translated_mapping();
+    virtual void fill_dram_physical_translated_mapping() = 0;
 
     /*
      * Fills the physical to translated mapping for the PCIE cores.
@@ -162,19 +175,25 @@ class CoordinateManager {
      * should override this method. Blackhole coordinate manager overrides
      * this method to implement different mapping.
      */
-    virtual void fill_pcie_physical_translated_mapping();
+    virtual void fill_pcie_physical_translated_mapping() = 0;
 
     /*
      * Fills the physical to translated mapping for the ARC cores.
      * By default, translated coordinates are the same as physical coordinates.
      * Derived coordinate managers that need to implement different mapping
      * should override this method.
      */
-    virtual void fill_arc_physical_translated_mapping();
+    virtual void fill_arc_physical_translated_mapping() = 0;
 
     std::map<tt::umd::CoreCoord, tt_xy_pair> to_physical_map;
     std::map<std::pair<tt_xy_pair, CoordSystem>, tt::umd::CoreCoord> from_physical_map;
 
+    // Whether NOC translation is enabled on chip.
+    // This flag affects how Translated coords are calculated. If translation is enabled on the chip, than we can
+    // interface it with a coordinate system which abstracts away harvested cores. If it is not enabled, then we need to
+    // interface it with noc0 coordinates.
+    bool noc_translation_enabled;
+
     tt_xy_pair tensix_grid_size;
     const std::vector<tt_xy_pair> tensix_cores;
     size_t tensix_harvesting_mask;

device/api/umd/device/grayskull_coordinate_manager.h

@@ -27,5 +27,9 @@ class GrayskullCoordinateManager : public CoordinateManager {
         const std::vector<tt_xy_pair>& pcie_cores);
 
 protected:
+    void fill_tensix_physical_translated_mapping() override;
     void fill_eth_physical_translated_mapping() override;
+    void fill_dram_physical_translated_mapping() override;
+    void fill_pcie_physical_translated_mapping() override;
+    void fill_arc_physical_translated_mapping() override;
 };

device/api/umd/device/tt_soc_descriptor.h

@@ -52,6 +52,7 @@ class tt_SocDescriptor {
     // Constructor used to build object from device descriptor file.
     tt_SocDescriptor(
         std::string device_descriptor_path,
+        const bool noc_translation_enabled,
         const size_t tensix_harvesting_mask = 0,
         const size_t dram_harvesting_mask = 0,
         const size_t eth_harvesting_mask = 0);
@@ -114,7 +115,10 @@ class tt_SocDescriptor {
 
 private:
     void create_coordinate_manager(
-        const size_t tensix_harvesting_mask, const size_t dram_harvesting_mask, const size_t eth_harvesting_mask);
+        const bool noc_translation_enabled,
+        const size_t tensix_harvesting_mask,
+        const size_t dram_harvesting_mask,
+        const size_t eth_harvesting_mask);
     void load_core_descriptors_from_device_descriptor(YAML::Node &device_descriptor_yaml);
     void load_soc_features_from_device_descriptor(YAML::Node &device_descriptor_yaml);
     void get_cores_and_grid_size_from_coordinate_manager();

device/api/umd/device/wormhole_coordinate_manager.h

@@ -12,6 +12,7 @@
 class WormholeCoordinateManager : public CoordinateManager {
 public:
     WormholeCoordinateManager(
+        const bool noc_translation_enabled,
         const tt_xy_pair& tensix_grid_size,
         const std::vector<tt_xy_pair>& tensix_cores,
         const size_t tensix_harvesting_mask,
@@ -28,5 +29,8 @@ class WormholeCoordinateManager : public CoordinateManager {
 
 protected:
     void fill_tensix_physical_translated_mapping() override;
+    void fill_dram_physical_translated_mapping() override;
     void fill_eth_physical_translated_mapping() override;
+    void fill_pcie_physical_translated_mapping() override;
+    void fill_arc_physical_translated_mapping() override;
 };

device/blackhole/blackhole_coordinate_manager.cpp

@@ -10,6 +10,7 @@
 using namespace tt::umd;
 
 BlackholeCoordinateManager::BlackholeCoordinateManager(
+    const bool noc_translation_enabled,
     const tt_xy_pair& tensix_grid_size,
     const std::vector<tt_xy_pair>& tensix_cores,
     const size_t tensix_harvesting_mask,
@@ -24,6 +25,7 @@ BlackholeCoordinateManager::BlackholeCoordinateManager(
     const tt_xy_pair& pcie_grid_size,
     const std::vector<tt_xy_pair>& pcie_cores) :
     CoordinateManager(
+        noc_translation_enabled,
         tensix_grid_size,
         tensix_cores,
         tensix_harvesting_mask,
@@ -89,7 +91,11 @@ void BlackholeCoordinateManager::translate_tensix_coords() {
         }
     }
 
-    fill_tensix_physical_translated_mapping();
+    if (noc_translation_enabled) {
+        fill_tensix_physical_translated_mapping();
+    } else {
+        fill_tensix_default_physical_translated_mapping();
+    }
 }
 
 void BlackholeCoordinateManager::fill_tensix_physical_translated_mapping() {
@@ -146,7 +152,11 @@ void BlackholeCoordinateManager::translate_dram_coords() {
         }
     }
 
-    fill_dram_physical_translated_mapping();
+    if (noc_translation_enabled) {
+        fill_dram_physical_translated_mapping();
+    } else {
+        fill_dram_default_physical_translated_mapping();
+    }
 }
 
 void BlackholeCoordinateManager::translate_eth_coords() {
@@ -184,7 +194,11 @@ void BlackholeCoordinateManager::translate_eth_coords() {
         }
     }
 
-    fill_eth_physical_translated_mapping();
+    if (noc_translation_enabled) {
+        fill_eth_physical_translated_mapping();
+    } else {
+        fill_eth_default_physical_translated_mapping();
+    }
 }
 
 void BlackholeCoordinateManager::fill_eth_physical_translated_mapping() {
@@ -231,6 +245,11 @@ void BlackholeCoordinateManager::fill_pcie_physical_translated_mapping() {
     add_core_translation(translated_coord, physical_pair);
 }
 
+void BlackholeCoordinateManager::fill_arc_physical_translated_mapping() {
+    // ARC cores are not translated in Blackhole.
+    fill_arc_default_physical_translated_mapping();
+}
+
 void BlackholeCoordinateManager::map_column_of_dram_banks(
     const size_t start_bank, const size_t end_bank, const size_t x_coord) {
     size_t translated_y = blackhole::dram_translated_coordinate_start_y;

device/cluster.cpp

@@ -436,19 +436,30 @@ std::unique_ptr<Chip> Cluster::construct_chip_from_cluster(
 }
 
 std::unique_ptr<Chip> Cluster::construct_chip_from_cluster(
+    const std::string& soc_desc_path,
     chip_id_t chip_id,
     tt_ClusterDescriptor* cluster_desc,
     bool perform_harvesting,
     std::unordered_map<chip_id_t, uint32_t>& simulated_harvesting_masks) {
-    tt::ARCH arch = cluster_desc->get_arch(chip_id);
-    const BoardType chip_board_type = cluster_desc->get_board_type(chip_id);
-    std::string soc_desc_path = tt_SocDescriptor::get_soc_descriptor_path(arch, chip_board_type);
     uint32_t tensix_harvesting_mask =
         get_tensix_harvesting_mask(chip_id, cluster_desc, perform_harvesting, simulated_harvesting_masks);
-    tt_SocDescriptor soc_desc = tt_SocDescriptor(soc_desc_path, tensix_harvesting_mask);
+    tt_SocDescriptor soc_desc = tt_SocDescriptor(
+        soc_desc_path, cluster_desc->get_noc_translation_table_en().at(chip_id), tensix_harvesting_mask);
     return construct_chip_from_cluster(chip_id, cluster_desc, soc_desc);
 }
 
+std::unique_ptr<Chip> Cluster::construct_chip_from_cluster(
+    chip_id_t chip_id,
+    tt_ClusterDescriptor* cluster_desc,
+    bool perform_harvesting,
+    std::unordered_map<chip_id_t, uint32_t>& simulated_harvesting_masks) {
+    tt::ARCH arch = cluster_desc->get_arch(chip_id);
+    const BoardType chip_board_type = cluster_desc->get_board_type(chip_id);
+    std::string soc_desc_path = tt_SocDescriptor::get_soc_descriptor_path(arch, chip_board_type);
+    return construct_chip_from_cluster(
+        soc_desc_path, chip_id, cluster_desc, perform_harvesting, simulated_harvesting_masks);
+}
+
 void Cluster::add_chip(chip_id_t chip_id, std::unique_ptr<Chip> chip) {
     log_assert(
         chips_.find(chip_id) == chips_.end(),
@@ -567,16 +578,16 @@ Cluster::Cluster(
             cluster_desc->get_all_chips().find(chip_id) != cluster_desc->get_all_chips().end(),
             "Target device {} not present in current cluster!",
             chip_id);
-        size_t tensix_harvesting_mask =
-            get_tensix_harvesting_mask(chip_id, cluster_desc.get(), perform_harvesting, simulated_harvesting_masks);
-        tt_SocDescriptor soc_desc = tt_SocDescriptor(sdesc_path, tensix_harvesting_mask);
+        add_chip(
+            chip_id,
+            construct_chip_from_cluster(
+                sdesc_path, chip_id, cluster_desc.get(), perform_harvesting, simulated_harvesting_masks));
         log_assert(
-            cluster_desc->get_arch(chip_id) == soc_desc.arch,
+            cluster_desc->get_arch(chip_id) == chips_.at(chip_id)->get_soc_descriptor().arch,
             "Passed soc descriptor has {} arch, but for chip id {} has arch {}",
-            arch_to_str(soc_desc.arch),
+            arch_to_str(chips_.at(chip_id)->get_soc_descriptor().arch),
             chip_id,
             arch_to_str(cluster_desc->get_arch(chip_id)));
-        add_chip(chip_id, construct_chip_from_cluster(chip_id, cluster_desc.get(), soc_desc));
     }
 
     // TODO: work on removing this member altogether. Currently assumes all have the same arch.
@@ -623,7 +634,8 @@ Cluster::Cluster(
     // rather than ClusterDescriptor.
     tt::ARCH arch = tt::ARCH::GRAYSKULL;
     chip_id_t mock_chip_id = 0;
-    tt_SocDescriptor soc_desc = tt_SocDescriptor(tt_SocDescriptor::get_soc_descriptor_path(arch, BoardType::UNKNOWN));
+    tt_SocDescriptor soc_desc =
+        tt_SocDescriptor(tt_SocDescriptor::get_soc_descriptor_path(arch, BoardType::UNKNOWN), false);
     std::unique_ptr<Chip> chip = std::make_unique<MockChip>(soc_desc);
 
     std::unordered_map<chip_id_t, std::unique_ptr<Chip>> chips;