[UMD] Switching to new coord API (#17003)

### Ticket Related to #17002 ### Problem description One of the many changes leading towards tt-metal using new CoreCoords, both in soc_descriptor and in umd's interface. Due to complexity of all the maps present, the changes will be relatively small. ### What's changed - virtual_core_from_physical_core and get_virtual_coordinate_from_physical_coordinates rewritten to use translate_coord_to instead of convert_to_umd_coordinates and translate_to_noc_table_coords - removed worker_log_to_physical_routing and physical_routing_to_virtual_routing from metal_soc_descriptor - get_physical_tensix_core_from_logical and convert_to_umd_coordinates rewritten to use translate_coord_to - Rewritten some get_physical_core_from_logical_core to get_physical_tensix_core_from_logical - grayskull soc descriptor changed so that it is rowwise instead of columnwise. This matches the way it is written in UMD, and it matches other soc descriptors. - bump umd to pull in some new changes in UMD ### Checklist All runs on brosko/soc_new_api : - [x] All post-commit tests : https://github.com/tenstorrent/tt-metal/actions/runs/12986193862 - [x] Blackhole post-commit tests : https://github.com/tenstorrent/tt-metal/actions/runs/12986195849 - [x] (Single-card) Model perf tests : https://github.com/tenstorrent/tt-metal/actions/runs/12986198348 - [x] (Single-card) Device perf regressions : https://github.com/tenstorrent/tt-metal/actions/runs/12986201144 - [x] (T3K) T3000 unit tests : https://github.com/tenstorrent/tt-metal/actions/runs/12986203715 - [x] (T3K) T3000 demo tests : https://github.com/tenstorrent/tt-metal/actions/runs/12986205723 - [x] (TG) TG unit tests : https://github.com/tenstorrent/tt-metal/actions/runs/12986207280 - [x] (TG) TG demo tests : https://github.com/tenstorrent/tt-metal/actions/runs/12986209173 - [x] (TGG) TGG unit tests : https://github.com/tenstorrent/tt-metal/actions/runs/12986211545 - [x] (TGG) TGG demo tests : https://github.com/tenstorrent/tt-metal/actions/runs/12986213812
tenstorrent · Jan 30, 2025 · 3001628 · 3001628
1 parent 7deb7b2
commit 3001628
Show file tree

Hide file tree

Showing 9 changed files with 59 additions and 95 deletions.
diff --git a/tt_metal/api/tt-metalium/device_impl.hpp b/tt_metal/api/tt-metalium/device_impl.hpp
@@ -280,7 +280,7 @@ class Device : public IDevice {
     CoreCoord physical_worker_core_from_logical_core(const CoreCoord &logical_core) const;
     CoreCoord dram_core_from_dram_channel(uint32_t dram_channel) const;
     CoreType core_type_from_physical_core(const CoreCoord &physical_core) const;
-    CoreCoord virtual_core_from_physical_core(const CoreCoord &physical_coord, const CoreType& core_type) const;
+    CoreCoord virtual_core_from_physical_core(const CoreCoord& physical_coord) const;
 
     chip_id_t id_;
     uint32_t build_key_ = 0;

diff --git a/tt_metal/api/tt-metalium/metal_soc_descriptor.h b/tt_metal/api/tt-metalium/metal_soc_descriptor.h
@@ -29,12 +29,6 @@ struct metal_SocDescriptor : public tt_SocDescriptor {
     std::vector<tt_xy_pair> physical_harvested_workers;
     std::vector<tt_xy_pair> physical_ethernet_cores;
 
-    std::unordered_map<int, int> worker_log_to_physical_routing_x;
-    std::unordered_map<int, int> worker_log_to_physical_routing_y;
-    // Physical to virtual maps are only applicable for x and y of tensix workers
-    std::unordered_map<int, int> physical_routing_to_virtual_routing_x;
-    std::unordered_map<int, int> physical_routing_to_virtual_routing_y;
-
     std::map<CoreCoord, int> logical_eth_core_to_chan_map;
     std::map<int, CoreCoord> chan_to_logical_eth_core_map;
 

diff --git a/tt_metal/api/tt-metalium/tt_cluster.hpp b/tt_metal/api/tt-metalium/tt_cluster.hpp
@@ -70,7 +70,7 @@ class Cluster {
 
     const metal_SocDescriptor &get_soc_desc(chip_id_t chip) const;
     CoreCoord get_virtual_coordinate_from_logical_coordinates(chip_id_t chip_id, CoreCoord logical_coord, const CoreType& core_type) const;
-    CoreCoord get_virtual_coordinate_from_physical_coordinates(chip_id_t chip_id, CoreCoord physical_coord, const CoreType& core_type) const;
+    CoreCoord get_virtual_coordinate_from_physical_coordinates(chip_id_t chip_id, CoreCoord physical_coord) const;
     tt_cxy_pair get_virtual_coordinate_from_logical_coordinates(tt_cxy_pair logical_coordinate, const CoreType& core_type) const;
     const std::unordered_set<CoreCoord>& get_virtual_worker_cores(chip_id_t chip_id) const;
     const std::unordered_set<CoreCoord>& get_virtual_eth_cores(chip_id_t chip_id) const;

diff --git a/tt_metal/common/core_descriptor.cpp b/tt_metal/common/core_descriptor.cpp
@@ -182,10 +182,10 @@ const std::tuple<uint32_t, CoreRange>& get_physical_worker_grid_config(
         uint32_t tensix_num_worker_cores = tensix_num_worker_cols * tensix_num_worker_rows;
         const metal_SocDescriptor& soc_desc = tt::Cluster::instance().get_soc_desc(device_id);
         // Get physical compute grid range based on SOC Desc and Logical Coords
-        CoreCoord tensix_worker_start_phys = soc_desc.get_physical_core_from_logical_core(
-            CoreCoord(0, 0), CoreType::WORKER);  // Logical Worker Coords start at 0,0
-        CoreCoord tensix_worker_end_phys = soc_desc.get_physical_core_from_logical_core(
-            CoreCoord(tensix_num_worker_cols - 1, tensix_num_worker_rows - 1), CoreType::WORKER);
+        // Logical Worker Coords start at 0,0
+        CoreCoord tensix_worker_start_phys = soc_desc.get_physical_tensix_core_from_logical(CoreCoord(0, 0));
+        CoreCoord tensix_worker_end_phys = soc_desc.get_physical_tensix_core_from_logical(
+            CoreCoord(tensix_num_worker_cols - 1, tensix_num_worker_rows - 1));
         CoreRange tensix_worker_physical_grid = CoreRange(tensix_worker_start_phys, tensix_worker_end_phys);
         physical_grid_config_cache.insert(
             {config_hash, std::make_tuple(tensix_num_worker_cores, tensix_worker_physical_grid)});

diff --git a/tt_metal/common/metal_soc_descriptor.cpp b/tt_metal/common/metal_soc_descriptor.cpp
@@ -114,16 +114,9 @@ CoreCoord metal_SocDescriptor::get_logical_ethernet_core_from_physical(const Cor
 }
 
 CoreCoord metal_SocDescriptor::get_physical_tensix_core_from_logical(const CoreCoord& logical_coord) const {
-    TT_FATAL(
-        (logical_coord.x < this->worker_grid_size.x) and (logical_coord.y < this->worker_grid_size.y),
-        "Bounds-Error -- Logical_core={} is outside of logical_grid_size={}",
-        logical_coord.str(),
-        this->worker_grid_size.str());
-    CoreCoord physical_tensix_core({
-        static_cast<size_t>(this->worker_log_to_physical_routing_x.at(logical_coord.x)),
-        static_cast<size_t>(this->worker_log_to_physical_routing_y.at(logical_coord.y)),
-    });
-    return physical_tensix_core;
+    tt::umd::CoreCoord physical_coord =
+        translate_coord_to({logical_coord, CoreType::TENSIX, CoordSystem::LOGICAL}, CoordSystem::PHYSICAL);
+    return {physical_coord.x, physical_coord.y};
 }
 
 CoreCoord metal_SocDescriptor::get_physical_dram_core_from_logical(const CoreCoord& logical_coord) const {
@@ -205,33 +198,20 @@ void metal_SocDescriptor::load_dram_metadata_from_device_descriptor() {
 
 // UMD expects virtual NOC coordinates for worker cores
 tt_cxy_pair metal_SocDescriptor::convert_to_umd_coordinates(const tt_cxy_pair& physical_cxy) const {
-    CoreCoord physical_coord({physical_cxy.x, physical_cxy.y});
-    const CoreDescriptor& core_desc = this->physical_cores.at(physical_coord);
-    CoreCoord virtual_coord = physical_coord;
-    if (core_desc.type == CoreType::WORKER or core_desc.type == CoreType::HARVESTED) {
-        virtual_coord.x = static_cast<size_t>(this->physical_routing_to_virtual_routing_x.at(physical_cxy.x));
-        virtual_coord.y = static_cast<size_t>(this->physical_routing_to_virtual_routing_y.at(physical_cxy.y));
-    }
-    return tt_cxy_pair(physical_cxy.chip, virtual_coord);
+    CoordSystem target_system = (this->arch == tt::ARCH::GRAYSKULL) ? CoordSystem::PHYSICAL : CoordSystem::VIRTUAL;
+    tt::umd::CoreCoord virtual_coord =
+        translate_coord_to((tt_xy_pair)physical_cxy, CoordSystem::PHYSICAL, target_system);
+    return tt_cxy_pair(physical_cxy.chip, virtual_coord.x, virtual_coord.y);
 }
 
 void metal_SocDescriptor::generate_physical_descriptors_from_virtual(uint32_t harvesting_mask) {
     // No need to remap virtual descriptors to physical because Grayskull does not have translation tables enabled,
     // meaning UMD removes the physical harvested rows rather than using virtual coordinates
     if (harvesting_mask == 0 or (this->arch == tt::ARCH::GRAYSKULL)) {
-        this->worker_log_to_physical_routing_x = this->worker_log_to_routing_x;
-        this->worker_log_to_physical_routing_y = this->worker_log_to_routing_y;
         this->physical_cores = this->cores;
         this->physical_workers = this->workers;
         this->physical_harvested_workers = this->harvested_workers;
 
-        for (const auto& [virtual_noc_core, core_desc] : this->cores) {
-            if (core_desc.type == CoreType::WORKER or core_desc.type == CoreType::HARVESTED) {
-                this->physical_routing_to_virtual_routing_x.insert({virtual_noc_core.x, virtual_noc_core.x});
-                this->physical_routing_to_virtual_routing_y.insert({virtual_noc_core.y, virtual_noc_core.y});
-            }
-        }
-
         return;
     }
 
@@ -264,10 +244,8 @@ void metal_SocDescriptor::generate_physical_descriptors_from_virtual(uint32_t ha
     for (const auto& [virtual_noc_core, core_desc] : this->cores) {
         if (core_desc.type == CoreType::WORKER or core_desc.type == CoreType::HARVESTED) {
             virtual_y_coords.insert(virtual_noc_core.y);
-            this->physical_routing_to_virtual_routing_x.insert({virtual_noc_core.x, virtual_noc_core.x});
         }
     }
-    this->worker_log_to_physical_routing_x = this->worker_log_to_routing_x;
 
     std::unordered_map<int, int> virtual_routing_to_physical_routing_y;
     auto virtual_y_coord_it = virtual_y_coords.begin();
@@ -278,10 +256,12 @@ void metal_SocDescriptor::generate_physical_descriptors_from_virtual(uint32_t ha
         }
         int physical_y_coord = *virtual_y_coord_it;
         virtual_y_coord_it++;
-        this->worker_log_to_physical_routing_y.insert({logical_y_coord, physical_y_coord});
-        int virtual_y_coord = this->worker_log_to_routing_y.at(logical_y_coord);
-        this->physical_routing_to_virtual_routing_y.insert({physical_y_coord, virtual_y_coord});
-        virtual_routing_to_physical_routing_y.insert({virtual_y_coord, physical_y_coord});
+        // This branch will never be executed for Grayskull, but for completeness keeping it in here.
+        // This will go away in the next PR anyway.
+        CoordSystem target_system = (this->arch == tt::ARCH::GRAYSKULL) ? CoordSystem::PHYSICAL : CoordSystem::VIRTUAL;
+        tt::umd::CoreCoord virtual_coord =
+            translate_coord_to({0, logical_y_coord, CoreType::TENSIX, CoordSystem::LOGICAL}, target_system);
+        virtual_routing_to_physical_routing_y.insert({virtual_coord.y, physical_y_coord});
     }
 
     // map physical harvested rows to virtual harvested rows
@@ -290,7 +270,6 @@ void metal_SocDescriptor::generate_physical_descriptors_from_virtual(uint32_t ha
          v_it != virtual_harvested_rows.end() and p_it != row_coordinates_to_remove.end();
          ++v_it, ++p_it) {
         virtual_routing_to_physical_routing_y.insert({*v_it, *p_it});
-        this->physical_routing_to_virtual_routing_y.insert({*p_it, *v_it});
     }
 
     for (const auto& [virtual_noc_core, core_desc] : this->cores) {
@@ -309,11 +288,6 @@ void metal_SocDescriptor::generate_physical_descriptors_from_virtual(uint32_t ha
         }
         this->physical_cores.insert({physical_noc_core, phys_core_desc});
     }
-
-    TT_ASSERT(
-        this->physical_routing_to_virtual_routing_y.size() ==
-            this->worker_grid_size.y + row_coordinates_to_remove.size() and
-        this->physical_routing_to_virtual_routing_x.size() == this->worker_grid_size.x);
 }
 
 void metal_SocDescriptor::generate_logical_eth_coords_mapping() {

diff --git a/tt_metal/impl/device/device.cpp b/tt_metal/impl/device/device.cpp
@@ -287,7 +287,8 @@ std::unique_ptr<Allocator> Device::initialize_allocator(size_t l1_small_size, si
     }
     // Initialize core_type_from_noc_coord_table table
     for (const auto& core: soc_desc.physical_cores) {
-        config.core_type_from_noc_coord_table.insert({this->virtual_core_from_physical_core(core.first, core.second.type), AllocCoreType::Invalid});
+        config.core_type_from_noc_coord_table.insert(
+            {this->virtual_core_from_physical_core(core.first), AllocCoreType::Invalid});
     }
 
     for (const CoreCoord& core : tt::get_logical_compute_cores(id_, num_hw_cqs_, dispatch_core_config)) {
@@ -720,7 +721,7 @@ void Device::initialize_and_launch_firmware() {
         // Track Virtual Non Worker Cores (In this case only Eth) separately
         uint32_t virtual_non_worker_cores_idx = 0;
         for (const CoreCoord &core : eth_cores) {
-            auto virtual_core = this->virtual_core_from_physical_core(core, CoreType::ETH);
+            auto virtual_core = this->virtual_core_from_physical_core(core);
             core_info->virtual_non_worker_cores[virtual_non_worker_cores_idx++] = {virtual_core.x, virtual_core.y, AddressableCoreType::ETH};
         }
     }
@@ -1173,11 +1174,12 @@ CoreCoord Device::compute_with_storage_grid_size() const {
 }
 
 CoreType Device::core_type_from_physical_core(const CoreCoord &physical_coord) const {
-    const metal_SocDescriptor &soc_desc = tt::Cluster::instance().get_soc_desc(this->id_);
-    if (soc_desc.physical_cores.find(physical_coord) == soc_desc.physical_cores.end())
-        TT_THROW("Physical core {} doesn't exist in metal_SocDescriptor.", physical_coord);
-
-    return soc_desc.physical_cores.at(physical_coord).type;
+    const metal_SocDescriptor& soc_desc = tt::Cluster::instance().get_soc_desc(this->id_);
+    CoreType core_type = soc_desc.translate_coord_to(physical_coord, CoordSystem::PHYSICAL, CoordSystem::PHYSICAL).core_type;
+    if (core_type == CoreType::TENSIX) {
+        core_type = CoreType::WORKER;
+    }
+    return core_type;
 }
 
 CoreType Device::core_type_from_virtual_core(const CoreCoord &virtual_coord) const {
@@ -1196,7 +1198,7 @@ CoreCoord Device::virtual_noc0_coordinate(uint8_t noc_index, CoreCoord coord) co
     } else {
         const auto& grid_size = this->grid_size();
         // Coordinate in Physical NOC0 Space. Convert to Virtual.
-        coord = this->virtual_core_from_physical_core(coord, this->core_type_from_physical_core(coord));
+        coord = this->virtual_core_from_physical_core(coord);
         // Derive virtual coord in noc_index space.
         CoreCoord virtual_coord = {
             hal.noc_coordinate(noc_index, grid_size.x, coord.x),
@@ -1219,7 +1221,7 @@ CoreCoord Device::virtual_noc_coordinate(uint8_t noc_index, CoreCoord coord) con
             hal.noc_coordinate(noc_index, grid_size.x, coord.x),
             hal.noc_coordinate(noc_index, grid_size.y, coord.y)
         };
-        return this->virtual_core_from_physical_core(physical_coord, this->core_type_from_physical_core(physical_coord));
+        return this->virtual_core_from_physical_core(physical_coord);
     }
 }
 
@@ -1248,8 +1250,8 @@ CoreCoord Device::virtual_core_from_logical_core(const CoreCoord &logical_coord,
     return tt::Cluster::instance().get_virtual_coordinate_from_logical_coordinates(this->id_, logical_coord, core_type);
 }
 
-CoreCoord Device::virtual_core_from_physical_core(const CoreCoord &physical_coord, const CoreType& core_type) const {
-    return tt::Cluster::instance().get_virtual_coordinate_from_physical_coordinates(this->id_, physical_coord, core_type);
+CoreCoord Device::virtual_core_from_physical_core(const CoreCoord& physical_coord) const {
+    return tt::Cluster::instance().get_virtual_coordinate_from_physical_coordinates(this->id_, physical_coord);
 }
 
 CoreCoord Device::worker_core_from_logical_core(const CoreCoord &logical_core) const {

diff --git a/tt_metal/llrt/tt_cluster.cpp b/tt_metal/llrt/tt_cluster.cpp
@@ -345,7 +345,8 @@ void Cluster::generate_virtual_to_umd_coord_mapping() {
         this->virtual_eth_cores_[chip_id] = {};
         for (auto& core_desc : this->get_soc_desc(chip_id).physical_cores) {
             if (core_desc.second.type != CoreType::HARVESTED) {
-                CoreCoord virtual_coords = this->get_virtual_coordinate_from_physical_coordinates(chip_id, core_desc.first, core_desc.second.type);
+                CoreCoord virtual_coords =
+                    this->get_virtual_coordinate_from_physical_coordinates(chip_id, core_desc.first);
                 tt_cxy_pair virtual_core = tt_cxy_pair(chip_id, virtual_coords.x, virtual_coords.y);
                 tt_cxy_pair umd_core = this->get_soc_desc(chip_id).convert_to_umd_coordinates(tt_cxy_pair(chip_id, core_desc.first.x, core_desc.first.y));
                 this->virtual_to_umd_coord_mapping_[virtual_core] = umd_core;
@@ -370,18 +371,19 @@ void Cluster::generate_logical_to_virtual_coord_mapping() {
         this->worker_logical_to_virtual_y_.insert({board_type, {}});
         this->eth_logical_to_virtual_.insert({board_type, {}});
         for (auto x_coords : soc_desc.worker_log_to_routing_x) {
-            CoreCoord phys_core = soc_desc.get_physical_core_from_logical_core(CoreCoord(x_coords.first, 0), CoreType::WORKER);
-            CoreCoord virtual_coords = this->get_virtual_coordinate_from_physical_coordinates(chip_id, phys_core, CoreType::WORKER);
+            CoreCoord phys_core = soc_desc.get_physical_tensix_core_from_logical(CoreCoord(x_coords.first, 0));
+            CoreCoord virtual_coords = this->get_virtual_coordinate_from_physical_coordinates(chip_id, phys_core);
             this->worker_logical_to_virtual_x_.at(board_type).insert({x_coords.first, virtual_coords.x});
         }
         for (auto y_coords : soc_desc.worker_log_to_routing_y) {
-            CoreCoord phys_core = soc_desc.get_physical_core_from_logical_core(CoreCoord(0, y_coords.first), CoreType::WORKER);
-            CoreCoord virtual_coords = this->get_virtual_coordinate_from_physical_coordinates(chip_id, phys_core, CoreType::WORKER);
+            CoreCoord phys_core = soc_desc.get_physical_tensix_core_from_logical(CoreCoord(0, y_coords.first));
+            CoreCoord virtual_coords = this->get_virtual_coordinate_from_physical_coordinates(chip_id, phys_core);
             this->worker_logical_to_virtual_y_.at(board_type).insert({y_coords.first, virtual_coords.y});
         }
         for (std::size_t log_eth_core_y = 0; log_eth_core_y < soc_desc.physical_ethernet_cores.size(); log_eth_core_y++) {
             CoreCoord logical_eth_core = {0, log_eth_core_y};
-            CoreCoord virtual_coords = this->get_virtual_coordinate_from_physical_coordinates(chip_id, soc_desc.physical_ethernet_cores.at(log_eth_core_y), CoreType::ETH);
+            CoreCoord virtual_coords = this->get_virtual_coordinate_from_physical_coordinates(
+                chip_id, soc_desc.physical_ethernet_cores.at(log_eth_core_y));
             this->eth_logical_to_virtual_.at(board_type).insert({logical_eth_core, virtual_coords});
         }
     }
@@ -398,11 +400,10 @@ void Cluster::generate_virtual_to_profiler_flat_id_mapping() {
         this->virtual_routing_to_profiler_flat_id_.insert({board_type, {}});
         auto& soc_desc = this->get_soc_desc(chip_id);
         for (const auto& core_to_profiler_id : soc_desc.physical_routing_to_profiler_flat_id) {
-            if (std::find(soc_desc.physical_workers.begin(), soc_desc.physical_workers.end(), core_to_profiler_id.first) != soc_desc.physical_workers.end()) {
-                this->virtual_routing_to_profiler_flat_id_.at(board_type).insert({this->get_virtual_coordinate_from_physical_coordinates(chip_id, core_to_profiler_id.first, CoreType::WORKER), core_to_profiler_id.second});
-            } else {
-                this->virtual_routing_to_profiler_flat_id_.at(board_type).insert({this->get_virtual_coordinate_from_physical_coordinates(chip_id, core_to_profiler_id.first, CoreType::ETH), core_to_profiler_id.second});
-            }
+            this->virtual_routing_to_profiler_flat_id_.at(board_type)
+                .insert(
+                    {this->get_virtual_coordinate_from_physical_coordinates(chip_id, core_to_profiler_id.first),
+                     core_to_profiler_id.second});
         }
     }
 #endif
@@ -441,16 +442,11 @@ tt_cxy_pair Cluster::get_virtual_coordinate_from_logical_coordinates(tt_cxy_pair
     auto xy_virtual_coord = this->get_virtual_coordinate_from_logical_coordinates(logical_coordinate.chip, CoreCoord(logical_coordinate.x, logical_coordinate.y), core_type);
     return tt_cxy_pair(logical_coordinate.chip, xy_virtual_coord);
 }
-CoreCoord Cluster::get_virtual_coordinate_from_physical_coordinates(chip_id_t chip_id, CoreCoord physical_coord, const CoreType& core_type) const {
+CoreCoord Cluster::get_virtual_coordinate_from_physical_coordinates(chip_id_t chip_id, CoreCoord physical_coord) const {
     auto& soc_desc = this->get_soc_desc(chip_id);
-    if ((not (core_type == CoreType::WORKER or core_type == CoreType::ETH)) or this->target_type_ == TargetDevice::Simulator) {
-        return physical_coord;
-    }
-    tt_cxy_pair virtual_chip_coord = soc_desc.convert_to_umd_coordinates(tt_cxy_pair(chip_id, physical_coord.x, physical_coord.y));
-    std::size_t c = virtual_chip_coord.x;
-    std::size_t r = virtual_chip_coord.y;
-    this->driver_->translate_to_noc_table_coords(chip_id, r, c);
-    return CoreCoord{c, r};
+    tt::umd::CoreCoord translated_coord =
+        soc_desc.translate_coord_to(physical_coord, CoordSystem::PHYSICAL, CoordSystem::TRANSLATED);
+    return {translated_coord.x, translated_coord.y};
 }
 
 CoreCoord Cluster::get_logical_ethernet_core_from_virtual(chip_id_t chip, CoreCoord core) const {