Merge pull request verilog-to-routing#1841 from verilog-to-routing/change_place_flow_in_init_place

Change place flow in init place

Author: vaughnbetz

vpr/src/base/echo_files.cpp

@@ -122,6 +122,7 @@ void alloc_and_load_echo_file_info() {
     setEchoFileName(E_ECHO_ENDPOINT_TIMING, "endpoint_timing.echo.json");
     setEchoFileName(E_ECHO_LOOKAHEAD_MAP, "lookahead_map.echo");
     setEchoFileName(E_ECHO_RR_GRAPH_INDEXED_DATA, "rr_indexed_data.echo");
+    setEchoFileName(E_ECHO_COMPRESSED_GRIDS, "compressed_grids.echo");
 }
 
 void free_echo_file_info() {

vpr/src/base/echo_files.h

@@ -52,6 +52,7 @@ enum e_echo_files {
     E_ECHO_ENDPOINT_TIMING,
     E_ECHO_LOOKAHEAD_MAP,
     E_ECHO_RR_GRAPH_INDEXED_DATA,
+    E_ECHO_COMPRESSED_GRIDS,
 
     //Timing Graphs
     E_ECHO_PRE_PACKING_TIMING_GRAPH,

vpr/src/place/compressed_grid.cpp

@@ -142,3 +142,40 @@ int grid_to_compressed_approx(const std::vector<int>& coords, int point) {
         return std::distance(coords.begin(), itr - 1);
     return std::distance(coords.begin(), itr);
 }
+
+/*Print the contents of the compressed grids to an echo file*/
+void echo_compressed_grids(char* filename, const std::vector<t_compressed_block_grid>& comp_grids) {
+    FILE* fp;
+    fp = vtr::fopen(filename, "w");
+
+    auto& device_ctx = g_vpr_ctx.device();
+
+    fprintf(fp, "--------------------------------------------------------------\n");
+    fprintf(fp, "Compressed Grids: \n");
+    fprintf(fp, "--------------------------------------------------------------\n");
+    fprintf(fp, "\n");
+
+    for (int i = 0; i < (int)comp_grids.size(); i++) {
+        fprintf(fp, "\n\nGrid type: %s \n", device_ctx.logical_block_types[i].name);
+
+        fprintf(fp, "X coordinates: \n");
+        for (int j = 0; j < (int)comp_grids[i].compressed_to_grid_x.size(); j++) {
+            fprintf(fp, "%d ", comp_grids[i].compressed_to_grid_x[j]);
+        }
+        fprintf(fp, "\n");
+
+        fprintf(fp, "Y coordinates: \n");
+        for (int k = 0; k < (int)comp_grids[i].compressed_to_grid_y.size(); k++) {
+            fprintf(fp, "%d ", comp_grids[i].compressed_to_grid_y[k]);
+        }
+        fprintf(fp, "\n");
+
+        fprintf(fp, "Subtiles: \n");
+        for (int s = 0; s < (int)comp_grids[i].compatible_sub_tiles_for_tile.size(); s++) {
+            fprintf(fp, "%d ", comp_grids[i].compressed_to_grid_y[s]);
+        }
+        fprintf(fp, "\n");
+    }
+
+    fclose(fp);
+}

vpr/src/place/compressed_grid.h

@@ -64,4 +64,13 @@ int grid_to_compressed(const std::vector<int>& coords, int point);
  */
 int grid_to_compressed_approx(const std::vector<int>& coords, int point);
 
+/**
+ * @brief  print the contents of the compressed grids to an echo file
+ *
+ *   @param filename the name of the file to print to
+ *   @param comp_grids the compressed grids that are used during placement
+ *
+ */
+void echo_compressed_grids(char* filename, const std::vector<t_compressed_block_grid>& comp_grids);
+
 #endif

vpr/src/place/initial_placement.cpp

@@ -36,13 +36,19 @@ e_block_move_result record_block_move(t_pl_blocks_to_be_moved& blocks_affected,
 //Moves the blocks in blocks_affected to their new locations
 void apply_move_blocks(const t_pl_blocks_to_be_moved& blocks_affected) {
     auto& place_ctx = g_vpr_ctx.mutable_placement();
+    auto& device_ctx = g_vpr_ctx.device();
 
     //Swap the blocks, but don't swap the nets or update place_ctx.grid_blocks
     //yet since we don't know whether the swap will be accepted
     for (int iblk = 0; iblk < blocks_affected.num_moved_blocks; ++iblk) {
         ClusterBlockId blk = blocks_affected.moved_blocks[iblk].block_num;
 
         place_ctx.block_locs[blk].loc = blocks_affected.moved_blocks[iblk].new_loc;
+
+        //if physical tile type of old location does not equal physical tile type of new location, sync the new physical pins
+        if (device_ctx.grid[blocks_affected.moved_blocks[iblk].old_loc.x][blocks_affected.moved_blocks[iblk].old_loc.y].type != device_ctx.grid[blocks_affected.moved_blocks[iblk].new_loc.x][blocks_affected.moved_blocks[iblk].new_loc.y].type) {
+            place_sync_external_block_connections(blk);
+        }
     }
 }
 
@@ -78,6 +84,7 @@ void commit_move_blocks(const t_pl_blocks_to_be_moved& blocks_affected) {
 //Moves the blocks in blocks_affected to their old locations
 void revert_move_blocks(t_pl_blocks_to_be_moved& blocks_affected) {
     auto& place_ctx = g_vpr_ctx.mutable_placement();
+    auto& device_ctx = g_vpr_ctx.device();
 
     // Swap the blocks back, nets not yet swapped they don't need to be changed
     for (int iblk = 0; iblk < blocks_affected.num_moved_blocks; ++iblk) {
@@ -87,6 +94,11 @@ void revert_move_blocks(t_pl_blocks_to_be_moved& blocks_affected) {
 
         place_ctx.block_locs[blk].loc = old;
 
+        //if physical tile type of old location does not equal physical tile type of new location, sync the new physical pins
+        if (device_ctx.grid[blocks_affected.moved_blocks[iblk].old_loc.x][blocks_affected.moved_blocks[iblk].old_loc.y].type != device_ctx.grid[blocks_affected.moved_blocks[iblk].new_loc.x][blocks_affected.moved_blocks[iblk].new_loc.y].type) {
+            place_sync_external_block_connections(blk);
+        }
+
         VTR_ASSERT_SAFE_MSG(place_ctx.grid_blocks[old.x][old.y].blocks[old.sub_tile] == blk, "Grid blocks should only have been updated if swap commited (not reverted)");
     }
 }

vpr/src/place/move_transactions.cpp

@@ -510,8 +510,7 @@ void try_place(const t_placer_opts& placer_opts,
 
     vtr::ScopedStartFinishTimer timer("Placement");
 
-    initial_placement(placer_opts.pad_loc_type,
-                      placer_opts.constraints_file.c_str());
+    initial_placement(placer_opts.pad_loc_type, placer_opts.constraints_file.c_str());
 
 #ifdef ENABLE_ANALYTIC_PLACE
     /*
@@ -912,6 +911,11 @@ void try_place(const t_placer_opts& placer_opts,
     }
 #endif
 
+    // Update physical pin values
+    for (auto block_id : cluster_ctx.clb_nlist.blocks()) {
+        place_sync_external_block_connections(block_id);
+    }
+
     check_place(costs, place_delay_model.get(), placer_criticalities.get(),
                 placer_opts.place_algorithm);
 
@@ -1088,10 +1092,10 @@ static void placement_inner_loop(const t_annealing_state* state,
         }
 #ifdef VERBOSE
         VTR_LOG("t = %g  cost = %g   bb_cost = %g timing_cost = %g move = %d\n",
-                t, costs->cost, costs->bb_cost, costs->timing_cost, inner_iter);
+                state->t, costs->cost, costs->bb_cost, costs->timing_cost, inner_iter);
         if (fabs((costs->bb_cost) - comp_bb_cost(CHECK)) > (costs->bb_cost) * ERROR_TOL)
-            VPR_ERROR(VPR_ERROR_PLACE,
-                      "fabs((*bb_cost) - comp_bb_cost(CHECK)) > (*bb_cost) * ERROR_TOL");
+            VPR_ERROR(VPR_ERROR_PLACE, "bb_cost is %g, comp_bb_cost is %g\n", costs->bb_cost, comp_bb_cost(CHECK));
+            //"fabs((*bb_cost) - comp_bb_cost(CHECK)) > (*bb_cost) * ERROR_TOL");
 #endif
 
         /* Lines below prevent too much round-off error from accumulating
@@ -1101,8 +1105,10 @@ static void placement_inner_loop(const t_annealing_state* state,
          */
         ++(*moves_since_cost_recompute);
         if (*moves_since_cost_recompute > MAX_MOVES_BEFORE_RECOMPUTE) {
+            //VTR_LOG("recomputing costs from scratch, old bb_cost is %g\n", costs->bb_cost);
             recompute_costs_from_scratch(placer_opts, delay_model,
                                          criticalities, costs);
+            //VTR_LOG("new_bb_cost is %g\n", costs->bb_cost);
             *moves_since_cost_recompute = 0;
         }
 
@@ -1551,7 +1557,7 @@ static e_move_result try_swap(const t_annealing_state* state,
     //VTR_ASSERT(check_macro_placement_consistency() == 0);
 #if 0
     //Check that each accepted swap yields a valid placement
-    check_place(*costs, delay_model, place_algorithm);
+    check_place(*costs, delay_model, criticalities, place_algorithm);
 #endif
 
     return move_outcome;
@@ -2005,9 +2011,9 @@ static double comp_bb_cost(e_cost_methods method) {
     }
 
     if (method == CHECK) {
-        VTR_LOG("\n");
-        VTR_LOG("BB estimate of min-dist (placement) wire length: %.0f\n",
-                expected_wirelength);
+        /*VTR_LOG("\n");
+         * VTR_LOG("BB estimate of min-dist (placement) wire length: %.0f\n",
+         * expected_wirelength);*/
     }
     return cost;
 }
@@ -2313,6 +2319,7 @@ static void get_non_updateable_bb(ClusterNetId net_id, t_bb* bb_coord_new) {
 
     ClusterBlockId bnum = cluster_ctx.clb_nlist.net_driver_block(net_id);
     pnum = net_pin_to_tile_pin_index(net_id, 0);
+
     x = place_ctx.block_locs[bnum].loc.x
         + physical_tile_type(bnum)->pin_width_offset[pnum];
     y = place_ctx.block_locs[bnum].loc.y

vpr/src/place/place.cpp

@@ -7,11 +7,11 @@ routed_wirelength;RangeAbs(0.60,1.50,5)
 #Area metrics
 logic_block_area_total;Range(0.8,1.3)
 logic_block_area_used;Range(0.8,1.3)
-min_chan_width_routing_area_total;Range(0.8,1.3)
-min_chan_width_routing_area_per_tile;Range(0.8,1.3)
+min_chan_width_routing_area_total;Range(0.7,1.3)
+min_chan_width_routing_area_per_tile;Range(0.7,1.3)
 
 #Run-time metrics
-min_chan_width_route_time;RangeAbs(0.10,10.0,2)
+min_chan_width_route_time;RangeAbs(0.10,14.0,2)
 
 #Peak memory
 #We set a 100MiB minimum threshold since the memory 

vtr_flow/parse/pass_requirements/common/pass_requirements.vpr_route_min_chan_width.txt

@@ -5,8 +5,8 @@
 crit_path_routed_wirelength;Range(0.60,1.50)
 
 #Area Metrics
-crit_path_routing_area_total;Range(0.8,1.3)
-crit_path_routing_area_per_tile;Range(0.8,1.3)
+crit_path_routing_area_total;Range(0.7,1.3)
+crit_path_routing_area_per_tile;Range(0.7,1.3)
 
 #Run-time Metrics
 crit_path_route_time;RangeAbs(0.10,10.0,2)