Merge pull request #3244 from AlexandreSinger/feature-ap-full-none-3d

[APPack][3D] 3D APPack Support

Author: amin1377

vpr/src/analytical_place/global_placer.cpp

@@ -411,7 +411,15 @@ PartialPlacement SimPLGlobalPlacer::place() {
 
         // Exit condition: If the upper-bound and lower-bound HPWLs are
         // sufficiently close together then stop.
-        double hpwl_relative_gap = (ub_hpwl - lb_hpwl) / ub_hpwl;
+        double hpwl_gap = ub_hpwl - lb_hpwl;
+        double hpwl_relative_gap;
+        if (ub_hpwl != 0.0)
+            hpwl_relative_gap = hpwl_gap / ub_hpwl;
+        else if (lb_hpwl != 0.0)
+            hpwl_relative_gap = hpwl_gap / lb_hpwl;
+        else
+            hpwl_relative_gap = 0.0;
+
         if (hpwl_relative_gap < target_hpwl_relative_gap_)
             break;
     }

vpr/src/base/flat_placement_utils.h

@@ -41,10 +41,14 @@ inline float get_manhattan_distance_to_tile(const t_flat_pl_loc& src_flat_loc,
     // the src_flat_loc. To do this, we project the point in L1 space.
     float proj_x = std::clamp(src_flat_loc.x, tile_xmin, tile_xmax);
     float proj_y = std::clamp(src_flat_loc.y, tile_ymin, tile_ymax);
+    // Note: We assume that tiles do not cross layers, so the projected layer
+    //       is just the layer that contains the tile.
+    float proj_layer = tile_loc.layer_num;
 
     // Then compute the L1 distance from the src_flat_loc to the projected
     // position. This will be the minimum distance this point needs to move.
     float dx = std::abs(proj_x - src_flat_loc.x);
     float dy = std::abs(proj_y - src_flat_loc.y);
-    return dx + dy;
+    float dlayer = std::abs(proj_layer - src_flat_loc.layer);
+    return dx + dy + dlayer;
 }

vpr/src/pack/greedy_candidate_selector.cpp

@@ -153,23 +153,23 @@ void GreedyCandidateSelector::initialize_unrelated_clustering_data(const t_molec
             max_loc.layer = std::max(max_loc.layer, mol_pos.layer);
         }
 
-        VTR_ASSERT_MSG(max_loc.layer == 0,
-                       "APPack unrelated clustering does not support 3D "
-                       "FPGAs yet");
-
         // Initialize the data structure with empty arrays with enough space
         // for each molecule.
+        size_t flat_grid_num_layers = max_loc.layer + 1;
         size_t flat_grid_width = max_loc.x + 1;
         size_t flat_grid_height = max_loc.y + 1;
         appack_unrelated_clustering_data_ =
-            vtr::NdMatrix<std::vector<std::vector<PackMoleculeId>>, 2>({flat_grid_width,
+            vtr::NdMatrix<std::vector<std::vector<PackMoleculeId>>, 3>({flat_grid_num_layers,
+                                                                        flat_grid_width,
                                                                         flat_grid_height});
-        for (size_t x = 0; x < flat_grid_width; x++) {
-            for (size_t y = 0; y < flat_grid_height; y++) {
-                // Resize to the maximum number of used external pins. This is
-                // to ensure that every molecule below can be inserted into a
-                // valid list based on their number of external pins.
-                appack_unrelated_clustering_data_[x][y].resize(max_molecule_stats.num_used_ext_pins + 1);
+        for (size_t layer_num = 0; layer_num < flat_grid_num_layers; layer_num++) {
+            for (size_t x = 0; x < flat_grid_width; x++) {
+                for (size_t y = 0; y < flat_grid_height; y++) {
+                    // Resize to the maximum number of used external pins. This is
+                    // to ensure that every molecule below can be inserted into a
+                    // valid list based on their number of external pins.
+                    appack_unrelated_clustering_data_[layer_num][x][y].resize(max_molecule_stats.num_used_ext_pins + 1);
+                }
             }
         }
 
@@ -185,7 +185,7 @@ void GreedyCandidateSelector::initialize_unrelated_clustering_data(const t_molec
             int ext_inps = molecule_stats.num_used_ext_inputs;
 
             //Insert the molecule into the unclustered lists by number of external inputs
-            auto& tile_uc_data = appack_unrelated_clustering_data_[mol_pos.x][mol_pos.y];
+            auto& tile_uc_data = appack_unrelated_clustering_data_[mol_pos.layer][mol_pos.x][mol_pos.y];
             tile_uc_data[ext_inps].push_back(mol_id);
         }
     } else {
@@ -1258,21 +1258,33 @@ PackMoleculeId GreedyCandidateSelector::get_unrelated_candidate_for_cluster_appa
     // to the max number of inputs a molecule could have.
     size_t inputs_avail = cluster_legalizer.get_num_cluster_inputs_available(cluster_id);
     VTR_ASSERT_SAFE(!appack_unrelated_clustering_data_.empty());
-    size_t max_molecule_inputs_avail = appack_unrelated_clustering_data_[0][0].size() - 1;
+    size_t max_molecule_inputs_avail = appack_unrelated_clustering_data_[0][0][0].size() - 1;
+    size_t flat_grid_num_layers = appack_unrelated_clustering_data_.dim_size(0);
+    size_t flat_grid_width = appack_unrelated_clustering_data_.dim_size(1);
+    size_t flat_grid_height = appack_unrelated_clustering_data_.dim_size(2);
     if (inputs_avail >= max_molecule_inputs_avail) {
         inputs_avail = max_molecule_inputs_avail;
     }
 
     // Create a queue of locations to search and a map of visited grid locations.
     std::queue<t_physical_tile_loc> search_queue;
-    vtr::NdMatrix<bool, 2> visited({appack_unrelated_clustering_data_.dim_size(0),
-                                    appack_unrelated_clustering_data_.dim_size(1)},
-                                   false);
-    // Push the position of the cluster to the queue.
+    vtr::NdMatrix<bool, 3> visited({flat_grid_num_layers,
+                                    flat_grid_width,
+                                    flat_grid_height},
+                                    false);
+
     t_physical_tile_loc cluster_tile_loc(cluster_gain_stats.flat_cluster_position.x,
                                          cluster_gain_stats.flat_cluster_position.y,
                                          cluster_gain_stats.flat_cluster_position.layer);
-    search_queue.push(cluster_tile_loc);
+
+    // Push the position of the cluster to the queue. We push this position on
+    // each layer such that each layer is searched independently.
+    for (size_t layer_num = 0; layer_num < flat_grid_num_layers; layer_num++) {
+        t_physical_tile_loc tile_loc(cluster_tile_loc.x,
+                                     cluster_tile_loc.y,
+                                     layer_num);
+        search_queue.push(tile_loc);
+    }
 
     // Get the max unrelated tile distance for the block type of this cluster.
     t_logical_block_type_ptr cluster_type = cluster_legalizer.get_cluster_type(cluster_id);
@@ -1288,10 +1300,12 @@ PackMoleculeId GreedyCandidateSelector::get_unrelated_candidate_for_cluster_appa
     while (!search_queue.empty()) {
         // Pop a position to search from the queue.
         const t_physical_tile_loc& node_loc = search_queue.front();
-        VTR_ASSERT_SAFE(node_loc.layer_num == 0);
 
         // Get the distance from the cluster to the current tile in tiles.
-        float dist = std::abs(node_loc.x - cluster_tile_loc.x) + std::abs(node_loc.y - cluster_tile_loc.y);
+        float node_dx = std::abs(node_loc.x - cluster_tile_loc.x);
+        float node_dy = std::abs(node_loc.y - cluster_tile_loc.y);
+        float node_dlayer = std::abs(node_loc.layer_num - cluster_tile_loc.layer_num);
+        float dist = node_dx + node_dy + node_dlayer;
 
         // If this position is too far from the source, skip it.
         if (dist > max_dist) {
@@ -1309,18 +1323,18 @@ PackMoleculeId GreedyCandidateSelector::get_unrelated_candidate_for_cluster_appa
         }
 
         // If this position has been visited, skip it.
-        if (visited[node_loc.x][node_loc.y]) {
+        if (visited[node_loc.layer_num][node_loc.x][node_loc.y]) {
             search_queue.pop();
             continue;
         }
-        visited[node_loc.x][node_loc.y] = true;
+        visited[node_loc.layer_num][node_loc.x][node_loc.y] = true;
 
         // Explore this position from highest number of inputs available to lowest.
         // Here, we are trying to find the closest compatible molecule, where we
         // break ties based on whoever has more external inputs.
         PackMoleculeId best_candidate = PackMoleculeId::INVALID();
         float best_candidate_distance = std::numeric_limits<float>::max();
-        const auto& uc_data = appack_unrelated_clustering_data_[node_loc.x][node_loc.y];
+        const auto& uc_data = appack_unrelated_clustering_data_[node_loc.layer_num][node_loc.x][node_loc.y];
         VTR_ASSERT_SAFE(inputs_avail < uc_data.size());
         for (int ext_inps = inputs_avail; ext_inps >= 0; ext_inps--) {
             // Get the molecule by the number of external inputs.

vpr/src/pack/greedy_candidate_selector.h

@@ -596,14 +596,14 @@ class GreedyCandidateSelector {
     /// @brief Data pre-computed to help select unrelated molecules when APPack
     ///        is being used. This is the same data as unrelated_clustering_data_,
     ///        but it is spatially distributed over the device.
-    /// For each grid location on the device (x, y), this provides a list of
+    /// For each grid location on the device (layer, x, y), this provides a list of
     /// molecules sorted by their gain, where the first dimension is the number
     /// of external outputs of the molecule.
     /// When APPack is not used, this will be uninitialized.
-    ///     [0..flat_grid_width][0..flat_grid_height][0..max_num_used_ext_pins]
+    ///     [0..flat_grid_num_layers][0..flat_grid_width][0..flat_grid_height][0..max_num_used_ext_pins]
     /// Here, flat_grid width/height is the maximum x and y positions given in
     /// the flat placement.
-    vtr::NdMatrix<std::vector<std::vector<PackMoleculeId>>, 2> appack_unrelated_clustering_data_;
+    vtr::NdMatrix<std::vector<std::vector<PackMoleculeId>>, 3> appack_unrelated_clustering_data_;
 
     /// @brief The APPack state which contains the options used to configure
     ///        APPack and the flat placement.

vpr/src/place/initial_placement.cpp

@@ -738,35 +738,39 @@ static inline t_pl_loc find_nearest_compatible_loc(const t_flat_pl_loc& src_flat
     const auto& compressed_block_grid = g_vpr_ctx.placement().compressed_block_grids[block_type->index];
     const DeviceGrid& device_grid = g_vpr_ctx.device().grid;
     const int num_layers = device_grid.get_num_layers();
-    // This method does not support 3D FPGAs yet. The search performed will only
-    // traverse the same layer as the src_loc.
-    VTR_ASSERT(num_layers == 1);
-    constexpr int layer = 0;
-
-    // Get the closest (approximately) compressed location to the src location.
-    // This does not need to be perfect (in fact I do not think it is), but the
-    // closer it is, the faster the BFS will find the best solution.
-    t_physical_tile_loc src_grid_loc(src_flat_loc.x, src_flat_loc.y, src_flat_loc.layer);
-    const t_physical_tile_loc compressed_src_loc = compressed_block_grid.grid_loc_to_compressed_loc_approx(src_grid_loc);
 
     // Weighted-BFS search the compressed grid for an empty compatible subtile.
-    size_t num_rows = compressed_block_grid.get_num_rows(layer);
-    size_t num_cols = compressed_block_grid.get_num_columns(layer);
-    vtr::NdMatrix<bool, 2> visited({num_cols, num_rows}, false);
+    std::vector<vtr::NdMatrix<bool, 2>> per_layer_visited(num_layers);
+    for (int layer = 0; layer < num_layers; layer++) {
+        size_t num_rows = compressed_block_grid.get_num_rows(layer);
+        size_t num_cols = compressed_block_grid.get_num_columns(layer);
+        per_layer_visited[layer].resize({num_cols, num_rows}, false);
+    }
     float best_dist = std::numeric_limits<float>::max();
     t_pl_loc best_loc(OPEN, OPEN, OPEN, OPEN);
 
+    // Get the closest (approximately) compressed location to the src location
+    // on each layer and enqueue them. We only want to enqueue locations onto
+    // layers that can feasibly implement this block.
+    // This does not need to be perfect (in fact I do not think it is), but the
+    // closer it is, the faster the BFS will find the best solution.
     std::queue<t_physical_tile_loc> loc_queue;
-    loc_queue.push(compressed_src_loc);
+    for (int layer_num : compressed_block_grid.get_layer_nums()) {
+        t_physical_tile_loc src_grid_loc(src_flat_loc.x, src_flat_loc.y, layer_num);
+        const t_physical_tile_loc compressed_src_loc = compressed_block_grid.grid_loc_to_compressed_loc_approx(src_grid_loc);
+        if (compressed_src_loc.x != OPEN && compressed_src_loc.y != OPEN)
+            loc_queue.push(compressed_src_loc);
+    }
+
     while (!loc_queue.empty()) {
         // Pop the top element off the queue.
         t_physical_tile_loc loc = loc_queue.front();
         loc_queue.pop();
 
         // If this location has already been visited, skip it.
-        if (visited[loc.x][loc.y])
+        if (per_layer_visited[loc.layer_num][loc.x][loc.y])
             continue;
-        visited[loc.x][loc.y] = true;
+        per_layer_visited[loc.layer_num][loc.x][loc.y] = true;
 
         // Get the minimum distance the cluster would need to move (relative to
         // its global placement solution) to be within the tile at the given
@@ -795,7 +799,7 @@ static inline t_pl_loc find_nearest_compatible_loc(const t_flat_pl_loc& src_flat
         // (i.e. no tile exists there). This is fine, we just need to check for
         // them to ensure we never try to put a cluster there.
         bool is_valid_compressed_loc = false;
-        const auto& compressed_col_blk_map = compressed_block_grid.get_column_block_map(loc.x, layer);
+        const auto& compressed_col_blk_map = compressed_block_grid.get_column_block_map(loc.x, loc.layer_num);
         if (compressed_col_blk_map.count(loc.y) != 0)
             is_valid_compressed_loc = true;
 
@@ -837,6 +841,8 @@ static inline t_pl_loc find_nearest_compatible_loc(const t_flat_pl_loc& src_flat
         // been visited. The code above checks for these cases to prevent extra
         // work and invalid lookups. This must be done this way to ensure that
         // the closest location can be found efficiently.
+        size_t num_rows = compressed_block_grid.get_num_rows(loc.layer_num);
+        size_t num_cols = compressed_block_grid.get_num_columns(loc.layer_num);
         if (loc.x > 0) {
             t_physical_tile_loc new_comp_loc = t_physical_tile_loc(loc.x - 1,
                                                                    loc.y,

vtr_flow/tasks/regression_tests/vtr_reg_strong/strong_ap/basic_3d_ap/config/config.txt

@@ -0,0 +1,45 @@
+##############################################
+# Configuration file for running experiments
+##############################################
+
+# Path to directory of circuits to use
+circuits_dir=benchmarks/blif/4
+
+# Path to directory of architectures to use
+archs_dir=arch/multi_die/simple_arch
+
+# Add architectures to list to sweep
+arch_list_add=3d_k4_N4_90nm.xml
+
+# Add circuits to list to sweep
+# This is a sweep of blif files which pack to a density between 50% and 90% of
+# the max density on this device.
+circuit_list_add=s820.blif
+circuit_list_add=s838.1.blif
+circuit_list_add=bw.blif
+circuit_list_add=rd84.blif
+circuit_list_add=s832.blif
+circuit_list_add=mm9a.blif
+circuit_list_add=alu2.blif
+circuit_list_add=x1.blif
+circuit_list_add=t481.blif
+circuit_list_add=mm9b.blif
+circuit_list_add=styr.blif
+circuit_list_add=s953.blif
+
+# Parse info and how to parse
+parse_file=vpr_fixed_chan_width.txt
+
+# How to parse QoR info
+qor_parse_file=qor_ap_fixed_chan_width.txt
+
+# Pass requirements
+pass_requirements_file=pass_requirements_ap_fixed_chan_width.txt
+
+script_params_common=-starting_stage vpr -track_memory_usage --analytical_place --route --device FPGA3D --route_chan_width 100
+
+# Test only the packer and the initial placer of the AP flow.
+script_params_list_add=--ap_analytical_solver identity --ap_partial_legalizer none
+# Force unrelated clustering on.
+script_params_list_add=--ap_analytical_solver identity --ap_partial_legalizer none --allow_unrelated_clustering on
+