Merge pull request #2669 from verilog-to-routing/temp_place_ref

Remove accesses to global placement state during placement stage

Author: vaughnbetz

libs/EXTERNAL/libtatum/libtatum/tatum/TimingReporter.cpp

@@ -606,7 +606,7 @@ Time TimingReporter::report_timing_data_arrival_subpath(std::ostream& os,
 
     {
         //Input constraint
-        TATUM_ASSERT(subpath.elements().size() > 0);
+        TATUM_ASSERT(!subpath.elements().empty());
         const TimingPathElem& path_elem = *(subpath.elements().begin());
 
         Time input_constraint;
@@ -712,7 +712,7 @@ bool TimingReporter::nearly_equal(const Time& lhs, const Time& rhs) const {
 
 size_t TimingReporter::estimate_point_print_width(const TimingPath& path) const {
     size_t width = 60; //default
-    for(auto subpath : {path.clock_launch_path(), path.data_arrival_path(), path.clock_capture_path()}) {
+    for(const auto& subpath : {path.clock_launch_path(), path.data_arrival_path(), path.clock_capture_path()}) {
         for(auto elem : subpath.elements()) {
             //Take the longest typical point name
             std::string point = name_resolver_.node_name(elem.node()) + " (" + name_resolver_.node_type_name(elem.node()) + ")";

utils/fasm/src/fasm.cpp

@@ -37,9 +37,9 @@ void FasmWriterVisitor::visit_top_impl(const char* top_level_name) {
 }
 
 void FasmWriterVisitor::visit_clb_impl(ClusterBlockId blk_id, const t_pb* clb) {
-    auto& place_ctx = g_vpr_ctx.placement();
     auto& device_ctx = g_vpr_ctx.device();
     auto& cluster_ctx = g_vpr_ctx.clustering();
+    auto& block_locs = g_vpr_ctx.placement().block_locs();
 
     current_blk_id_ = blk_id;
 
@@ -48,10 +48,10 @@ void FasmWriterVisitor::visit_clb_impl(ClusterBlockId blk_id, const t_pb* clb) {
 
     root_clb_ = clb->pb_graph_node;
 
-    int x = place_ctx.block_locs[blk_id].loc.x;
-    int y = place_ctx.block_locs[blk_id].loc.y;
-    int layer_num = place_ctx.block_locs[blk_id].loc.layer;
-    int sub_tile = place_ctx.block_locs[blk_id].loc.sub_tile;
+    int x = block_locs[blk_id].loc.x;
+    int y = block_locs[blk_id].loc.y;
+    int layer_num = block_locs[blk_id].loc.layer;
+    int sub_tile = block_locs[blk_id].loc.sub_tile;
     physical_tile_ = device_ctx.grid.get_physical_type({x, y, layer_num});
     logical_block_ = cluster_ctx.clb_nlist.block_type(blk_id);
     const auto& grid_meta = device_ctx.grid.get_metadata({x, y, layer_num});

utils/fasm/test/test_fasm.cpp

@@ -569,7 +569,7 @@ TEST_CASE("fasm_integration_test", "[fasm]") {
 
     // Verify occupied grid LOCs
     const auto & place_ctx = g_vpr_ctx.placement();
-    for (const auto& loc: place_ctx.block_locs) {
+    for (const auto& loc: place_ctx.block_locs()) {
 
         // Do not consider "IOB" tiles. They do not have fasm features
         // defined in the arch.

vpr/src/analysis/timing_reports.cpp

@@ -12,13 +12,18 @@
 
 #include "VprTimingGraphResolver.h"
 
-void generate_setup_timing_stats(const std::string& prefix, const SetupTimingInfo& timing_info, const AnalysisDelayCalculator& delay_calc, const t_analysis_opts& analysis_opts, bool is_flat) {
+void generate_setup_timing_stats(const std::string& prefix,
+                                 const SetupTimingInfo& timing_info,
+                                 const AnalysisDelayCalculator& delay_calc,
+                                 const t_analysis_opts& analysis_opts,
+                                 bool is_flat,
+                                 const BlkLocRegistry& blk_loc_registry) {
     auto& timing_ctx = g_vpr_ctx.timing();
     auto& atom_ctx = g_vpr_ctx.atom();
 
     print_setup_timing_summary(*timing_ctx.constraints, *timing_info.setup_analyzer(), "Final ", analysis_opts.write_timing_summary);
 
-    VprTimingGraphResolver resolver(atom_ctx.nlist, atom_ctx.lookup, *timing_ctx.graph, delay_calc, is_flat);
+    VprTimingGraphResolver resolver(atom_ctx.nlist, atom_ctx.lookup, *timing_ctx.graph, delay_calc, is_flat, blk_loc_registry);
     resolver.set_detail_level(analysis_opts.timing_report_detail);
 
     tatum::TimingReporter timing_reporter(resolver, *timing_ctx.graph, *timing_ctx.constraints);
@@ -32,13 +37,18 @@ void generate_setup_timing_stats(const std::string& prefix, const SetupTimingInf
     timing_reporter.report_unconstrained_setup(prefix + "report_unconstrained_timing.setup.rpt", *timing_info.setup_analyzer());
 }
 
-void generate_hold_timing_stats(const std::string& prefix, const HoldTimingInfo& timing_info, const AnalysisDelayCalculator& delay_calc, const t_analysis_opts& analysis_opts, bool is_flat) {
+void generate_hold_timing_stats(const std::string& prefix,
+                                const HoldTimingInfo& timing_info,
+                                const AnalysisDelayCalculator& delay_calc,
+                                const t_analysis_opts& analysis_opts,
+                                bool is_flat,
+                                const BlkLocRegistry& blk_loc_registry) {
     auto& timing_ctx = g_vpr_ctx.timing();
     auto& atom_ctx = g_vpr_ctx.atom();
 
     print_hold_timing_summary(*timing_ctx.constraints, *timing_info.hold_analyzer(), "Final ");
 
-    VprTimingGraphResolver resolver(atom_ctx.nlist, atom_ctx.lookup, *timing_ctx.graph, delay_calc, is_flat);
+    VprTimingGraphResolver resolver(atom_ctx.nlist, atom_ctx.lookup, *timing_ctx.graph, delay_calc, is_flat, blk_loc_registry);
     resolver.set_detail_level(analysis_opts.timing_report_detail);
 
     tatum::TimingReporter timing_reporter(resolver, *timing_ctx.graph, *timing_ctx.constraints);

vpr/src/analysis/timing_reports.h

@@ -5,7 +5,20 @@
 #include "AnalysisDelayCalculator.h"
 #include "vpr_types.h"
 
-void generate_setup_timing_stats(const std::string& prefix, const SetupTimingInfo& timing_info, const AnalysisDelayCalculator& delay_calc, const t_analysis_opts& report_detail, bool is_flat);
-void generate_hold_timing_stats(const std::string& prefix, const HoldTimingInfo& timing_info, const AnalysisDelayCalculator& delay_calc, const t_analysis_opts& report_detail, bool is_flat);
+class BlkLocRegistry;
+
+void generate_setup_timing_stats(const std::string& prefix,
+                                 const SetupTimingInfo& timing_info,
+                                 const AnalysisDelayCalculator& delay_calc,
+                                 const t_analysis_opts& report_detail,
+                                 bool is_flat,
+                                 const BlkLocRegistry& blk_loc_registry);
+
+void generate_hold_timing_stats(const std::string& prefix,
+                                const HoldTimingInfo& timing_info,
+                                const AnalysisDelayCalculator& delay_calc,
+                                const t_analysis_opts& report_detail,
+                                bool is_flat,
+                                const BlkLocRegistry& blk_loc_registry);
 
 #endif

vpr/src/base/ShowSetup.cpp

@@ -527,10 +527,10 @@ static void ShowPlacerOpts(const t_placer_opts& PlacerOpts,
 
         VTR_LOG("PlacerOpts.pad_loc_type: ");
         switch (PlacerOpts.pad_loc_type) {
-            case FREE:
+            case e_pad_loc_type::FREE:
                 VTR_LOG("FREE\n");
                 break;
-            case RANDOM:
+            case e_pad_loc_type::RANDOM:
                 VTR_LOG("RANDOM\n");
                 break;
             default:

vpr/src/base/blk_loc_registry.cpp

@@ -0,0 +1,114 @@
+
+#include "blk_loc_registry.h"
+#include "globals.h"
+
+const vtr::vector_map<ClusterBlockId, t_block_loc>& BlkLocRegistry::block_locs() const {
+    return  block_locs_;
+}
+
+vtr::vector_map<ClusterBlockId, t_block_loc>& BlkLocRegistry::mutable_block_locs() {
+    return  block_locs_;
+}
+
+const GridBlock& BlkLocRegistry::grid_blocks() const {
+    return grid_blocks_;
+}
+
+GridBlock& BlkLocRegistry::mutable_grid_blocks() {
+    return grid_blocks_;
+}
+
+const vtr::vector_map<ClusterPinId, int>& BlkLocRegistry::physical_pins() const {
+    return physical_pins_;
+}
+
+vtr::vector_map<ClusterPinId, int>& BlkLocRegistry::mutable_physical_pins() {
+    return physical_pins_;
+}
+
+int BlkLocRegistry::tile_pin_index(const ClusterPinId pin) const {
+    return physical_pins_[pin];
+}
+
+int BlkLocRegistry::net_pin_to_tile_pin_index(const ClusterNetId net_id, int net_pin_index) const {
+    auto& cluster_ctx = g_vpr_ctx.clustering();
+
+    // Get the logical pin index of pin within its logical block type
+    ClusterPinId pin_id = cluster_ctx.clb_nlist.net_pin(net_id, net_pin_index);
+
+    return this->tile_pin_index(pin_id);
+}
+
+void BlkLocRegistry::set_block_location(ClusterBlockId blk_id, const t_pl_loc& location) {
+    auto& device_ctx = g_vpr_ctx.device();
+    auto& cluster_ctx = g_vpr_ctx.clustering();
+
+    const std::string& block_name = cluster_ctx.clb_nlist.block_name(blk_id);
+
+    //Check if block location is out of range of grid dimensions
+    if (location.x < 0 || location.x > int(device_ctx.grid.width() - 1)
+        || location.y < 0 || location.y > int(device_ctx.grid.height() - 1)) {
+        VPR_THROW(VPR_ERROR_PLACE, "Block %s with ID %d is out of range at location (%d, %d). \n",
+                  block_name.c_str(), blk_id, location.x, location.y);
+    }
+
+    //Set the location of the block
+    block_locs_[blk_id].loc = location;
+
+    //Check if block is at an illegal location
+    auto physical_tile = device_ctx.grid.get_physical_type({location.x, location.y, location.layer});
+    auto logical_block = cluster_ctx.clb_nlist.block_type(blk_id);
+
+    if (location.sub_tile >= physical_tile->capacity || location.sub_tile < 0) {
+        VPR_THROW(VPR_ERROR_PLACE, "Block %s subtile number (%d) is out of range. \n", block_name.c_str(), location.sub_tile);
+    }
+
+    if (!is_sub_tile_compatible(physical_tile, logical_block, block_locs_[blk_id].loc.sub_tile)) {
+        VPR_THROW(VPR_ERROR_PLACE, "Attempt to place block %s with ID %d at illegal location (%d,%d,%d). \n",
+                  block_name.c_str(),
+                  blk_id,
+                  location.x,
+                  location.y,
+                  location.layer);
+    }
+
+    //Mark the grid location and usage of the block
+    grid_blocks_.set_block_at_location(location, blk_id);
+    grid_blocks_.increment_usage({location.x, location.y, location.layer});
+
+    place_sync_external_block_connections(blk_id);
+}
+
+void BlkLocRegistry::place_sync_external_block_connections(ClusterBlockId iblk) {
+    const auto& cluster_ctx = g_vpr_ctx.clustering();
+    const auto& clb_nlist = cluster_ctx.clb_nlist;
+
+    t_pl_loc block_loc = block_locs_[iblk].loc;
+
+    auto physical_tile = physical_tile_type(block_loc);
+    auto logical_block = clb_nlist.block_type(iblk);
+
+    int sub_tile_index = get_sub_tile_index(iblk, block_locs_);
+    auto sub_tile = physical_tile->sub_tiles[sub_tile_index];
+
+    VTR_ASSERT(sub_tile.num_phy_pins % sub_tile.capacity.total() == 0);
+
+    int max_num_block_pins = sub_tile.num_phy_pins / sub_tile.capacity.total();
+    /* Logical location and physical location is offset by z * max_num_block_pins */
+
+    int rel_capacity = block_loc.sub_tile - sub_tile.capacity.low;
+
+    for (ClusterPinId pin : clb_nlist.block_pins(iblk)) {
+        int logical_pin_index = clb_nlist.pin_logical_index(pin);
+        int sub_tile_pin_index = get_sub_tile_physical_pin(sub_tile_index, physical_tile, logical_block, logical_pin_index);
+
+        int new_physical_pin_index = sub_tile.sub_tile_to_tile_pin_indices[sub_tile_pin_index + rel_capacity * max_num_block_pins];
+
+        auto result = physical_pins_.find(pin);
+        if (result != physical_pins_.end()) {
+            physical_pins_[pin] = new_physical_pin_index;
+        } else {
+            physical_pins_.insert(pin, new_physical_pin_index);
+        }
+    }
+}

vpr/src/base/blk_loc_registry.h

@@ -0,0 +1,85 @@
+#ifndef VTR_BLK_LOC_REGISTRY_H
+#define VTR_BLK_LOC_REGISTRY_H
+
+#include "clustered_netlist_fwd.h"
+#include "vtr_vector_map.h"
+#include "vpr_types.h"
+#include "grid_block.h"
+
+struct t_block_loc;
+
+/**
+ * @class BlkLocRegistry contains information about the placement of clustered blocks.
+ * More specifically:
+ *      1) block_locs stores the location where each clustered blocks is placed at.
+ *      2) grid_blocks stores which blocks (if any) are placed at a given location.
+ *      3) physical_pins stores the mapping between the pins of a clustered block and
+ *      the pins of the physical tile where the clustered blocks is placed.
+ *
+ */
+class BlkLocRegistry {
+  public:
+    BlkLocRegistry() = default;
+    ~BlkLocRegistry() = default;
+    BlkLocRegistry(const BlkLocRegistry&) = delete;
+    BlkLocRegistry& operator=(const BlkLocRegistry&) = default;
+    BlkLocRegistry(BlkLocRegistry&&) = delete;
+    BlkLocRegistry& operator=(BlkLocRegistry&&) = delete;
+
+  private:
+    ///@brief Clustered block placement locations
+    vtr::vector_map<ClusterBlockId, t_block_loc> block_locs_;
+
+    ///@brief Clustered block associated with each grid location (i.e. inverse of block_locs)
+    GridBlock grid_blocks_;
+
+    ///@brief Clustered pin placement mapping with physical pin
+    vtr::vector_map<ClusterPinId, int> physical_pins_;
+
+  public:
+    const vtr::vector_map<ClusterBlockId, t_block_loc>& block_locs() const;
+    vtr::vector_map<ClusterBlockId, t_block_loc>& mutable_block_locs();
+
+    const GridBlock& grid_blocks() const;
+    GridBlock& mutable_grid_blocks();
+
+    const vtr::vector_map<ClusterPinId, int>& physical_pins() const;
+    vtr::vector_map<ClusterPinId, int>& mutable_physical_pins();
+
+    ///@brief Returns the physical pin of the tile, related to the given ClusterPinId
+    int tile_pin_index(const ClusterPinId pin) const;
+
+    ///@brief Returns the physical pin of the tile, related to the given ClusterNedId, and the net pin index.
+    int net_pin_to_tile_pin_index(const ClusterNetId net_id, int net_pin_index) const;
+
+    /**
+     * @brief Performs error checking to see if location is legal for block type,
+     * and sets the location and grid usage of the block if it is legal.
+     * @param blk_id The unique ID of the clustered block whose location is to set.
+     * @param location The location where the clustered block should placed at.
+     */
+    void set_block_location(ClusterBlockId blk_id, const t_pl_loc& location);
+
+    /**
+     * @brief Syncs the logical block pins corresponding to the input iblk with the corresponding chosen physical tile
+     * @param iblk cluster block ID to sync within the assigned physical tile
+     *
+     * This routine updates the physical pins vector of the place context after the placement step
+     * to synchronize the pins related to the logical block with the actual connection interface of
+     * the belonging physical tile with the RR graph.
+     *
+     * This step is required as the logical block can be placed at any compatible sub tile locations
+     * within a physical tile.
+     * Given that it is possible to have equivalent logical blocks within a specific sub tile, with
+     * a different subset of IO pins, the various pins offsets must be correctly computed and assigned
+     * to the physical pins vector, so that, when the net RR terminals are computed, the correct physical
+     * tile IO pins are selected.
+     *
+     * This routine uses the x,y and sub_tile coordinates of the clb netlist, and expects those to place each netlist block
+     * at a legal location that can accommodate it.
+     * It does not check for overuse of locations, therefore it can be used with placements that have resource overuse.
+     */
+    void place_sync_external_block_connections(ClusterBlockId iblk);
+};
+
+#endif //VTR_BLK_LOC_REGISTRY_H

vpr/src/base/grid_block.cpp

@@ -0,0 +1,59 @@
+
+#include "grid_block.h"
+
+#include "globals.h"
+
+void GridBlock::zero_initialize() {
+    auto& device_ctx = g_vpr_ctx.device();
+
+    /* Initialize all occupancy to zero. */
+    for (int layer_num = 0; layer_num < (int)device_ctx.grid.get_num_layers(); layer_num++) {
+        for (int i = 0; i < (int)device_ctx.grid.width(); i++) {
+            for (int j = 0; j < (int)device_ctx.grid.height(); j++) {
+                set_usage({i, j, layer_num}, 0);
+                auto tile = device_ctx.grid.get_physical_type({i, j, layer_num});
+
+                for (const auto& sub_tile : tile->sub_tiles) {
+                    auto capacity = sub_tile.capacity;
+
+                    for (int k = 0; k < capacity.total(); k++) {
+                        set_block_at_location({i, j, k + capacity.low, layer_num}, ClusterBlockId::INVALID());
+                    }
+                }
+            }
+        }
+    }
+}
+
+void GridBlock::load_from_block_locs(const vtr::vector_map<ClusterBlockId, t_block_loc>& block_locs) {
+    auto& cluster_ctx = g_vpr_ctx.clustering();
+    auto& device_ctx = g_vpr_ctx.device();
+
+    zero_initialize();
+
+    for (ClusterBlockId blk_id : cluster_ctx.clb_nlist.blocks()) {
+        t_pl_loc location = block_locs[blk_id].loc;
+
+        VTR_ASSERT(location.x < (int)device_ctx.grid.width());
+        VTR_ASSERT(location.y < (int)device_ctx.grid.height());
+
+        set_block_at_location(location, blk_id);
+        increment_usage({location.x, location.y, location.layer});
+    }
+}
+
+int GridBlock::increment_usage(const t_physical_tile_loc& loc) {
+    int curr_usage = get_usage(loc);
+    int updated_usage = set_usage(loc, curr_usage + 1);
+
+    return updated_usage;
+}
+
+int GridBlock::decrement_usage(const t_physical_tile_loc& loc) {
+    int curr_usage = get_usage(loc);
+    int updated_usage = set_usage(loc, curr_usage - 1);
+
+    return updated_usage;
+}
+
+