Now direct connection builder in tileable rr_graph supports any subtile index as the from_pin (#3318)

* [test] add a strong test where clb, dsp and bram in a same tile, to validate the direct connections across subtiles

* [test] deploy new test to strong reg test

* [test] add new test

* [test] add new arch from openfpga

* [test] add new blif benchmark for tileable direct connection strong tests

* [test] add new example arch to validate tileable direct connections

* [test] reorg tasks

* [test] deploy new tests to validate supertile with z_offset in direct connection for tileable rr_graph

* [test] debug

* [test] debug

* [core] debug

* [core] now search all the sub tiles which contains the ports defined in the direct

* [core] syntax

* [test] debug

* [test] update golden

* [test] add missing golden results

* [core] add doxygen-style comments to vpr utility function

* [core] simplify swap in if condition

* [test] add arch file to validate direct connections across subtiles for regular rr_graph generator

* [test] add new testcases for strong reg tests about direct connections between subtiles

* [core] fixed the bug where IPIN coordinate does not consider pin offset at tile-level (pin location)

* [core] should sit with root location for IPIN

* [core] debug

* [core] debug

* [core] debug

* [test] update golden

* [doc] add examples

* [doc] typo

* [core] typo

* [doc] typo

Author: tangxifan

doc/src/arch/example_subtile_direct_connection.png

@@ -2365,6 +2365,8 @@ Direct Inter-block Connections
 The content within the ``<directlist>`` tag consists of a group of ``<direct>`` tags.
 The ``<direct>`` tag and its contents are described below.
 
+.. note:: ``from_pin`` and ``to_pin`` only support big endian! For example, ``clb.out[8:0]``
+
 .. arch:tag:: <direct name="string" from_pin="string" to_pin="string" x_offset="int" y_offset="int" z_offset="int" switch_name="string" from_side="{left|right|top|bottom}" to_side="{left|right|top|bottom}"/>
 
     :req_param name: is a unique alphanumeric string to name the connection.
@@ -2384,12 +2386,89 @@ The ``<direct>`` tag and its contents are described below.
     The ``from_side`` and ``to_side`` options can usually be left unspecified.
     However they can be used to explicitly control how direct connections to physically equivalent pins (which may appear on multiple sides) are handled.
 
-    **Example:**
-    Consider a carry chain where the ``cout`` of each CLB drives the ``cin`` of the CLB immediately below it, using the delay-less switch one would enter the following:
+**Example: Inter-tile connection**
+Consider a carry chain where the ``cout`` of each CLB drives the ``cin`` of the CLB immediately below it, using the delay-less switch one would enter the following:
 
-    .. code-block:: xml
+.. code-block:: xml
+
+    <direct name="adder_carry" from_pin="clb.cout" to_pin="clb.cin" x_offset="0" y_offset="-1" z_offset="0"/>
+
+**Example: Inner-tile feedback**
+
+Consider a feedback connection where the ``out`` of each CLB drives the ``in`` of the CLB in the same location, using the connection block switch one would enter the following:
+
+.. code-block:: xml
+
+    <direct name="feedback" from_pin="clb.out" to_pin="clb.in" x_offset="0" y_offset="0" z_offset="0" switch_name="cb_mux"/>
+
+**Example: Cross-sub-tile connection**
+
+In this example, a tile ``cim8_1k`` is defined, under which there are two types of sub-tiles:
+
+- ``mult_8``: the first sub-tile
+- ``memory``: the second, and the third sub-tile
+
+.. code-block:: xml
+
+    <tile name="cim8_1k" height="2" area="396000">
+      <sub_tile name="mult_8" capacity="1">
+        <equivalent_sites>
+          <site pb_type="mult_8" pin_mapping="direct"/>
+        </equivalent_sites>
+        <input name="a" num_pins="8"/>
+        <input name="b" num_pins="8"/>
+        <output name="out" num_pins="16"/>
+        <fc in_type="frac" in_val="0.15" out_type="frac" out_val="0.10">
+          <fc_override port_name="out" fc_type="frac" fc_val="0"/>
+        </fc>
+        <pinlocations pattern="custom">
+          <loc side="left"/>
+          <loc side="top"/>
+          <loc side="right" yoffset="0">mult_8.a[0:2] mult_8.b[0:2] mult_8.out[0:5]</loc>
+          <loc side="right" yoffset="1">mult_8.a[3:5] mult_8.b[3:5] mult_8.out[6:10]</loc>
+          <loc side="bottom">mult_8.a[6:7] mult_8.b[6:7] mult_8.out[11:15]</loc>
+        </pinlocations>
+      </sub_tile>
+      <sub_tile name="memory" capacity="2">
+        <equivalent_sites>
+          <site pb_type="memory"/>
+        </equivalent_sites>
+        <input name="waddr" num_pins="7"/>
+        <input name="raddr" num_pins="7"/>
+        <input name="data_in" num_pins="8"/>
+        <input name="wen" num_pins="1"/>
+        <input name="ren" num_pins="1"/>
+        <output name="data_out" num_pins="8"/>
+        <clock name="clk" num_pins="1"/>
+        <fc in_type="frac" in_val="0.15" out_type="frac" out_val="0.10">
+          <fc_override port_name="clk" fc_type="frac" fc_val="0"/>
+          <fc_override port_name="data_in" fc_type="frac" fc_val="0"/>
+        </fc>
+        <pinlocations pattern="custom">
+          <loc side="left" yoffset="0">memory.clk memory.waddr[0:0] memory.raddr[0:0] memory.data_in[0:0] memory.data_out[0:0]</loc>
+          <loc side="left" yoffset="1">memory.waddr[1:1] memory.raddr[1:1] memory.data_in[1:1] memory.data_out[1:1]</loc>
+          <loc side="top" yoffset="1">memory.waddr[2:2] memory.raddr[2:2] memory.data_in[2:2] memory.data_out[2:2] memory.waddr[3:3] memory.raddr[3:3] memory.data_in[3:3] memory.data_out[3:3]</loc>
+          <loc side="right" yoffset="0">memory.waddr[4:4] memory.raddr[4:4] memory.data_in[4:4] memory.data_out[4:4]</loc>
+          <loc side="right" yoffset="1">memory.waddr[5:5] memory.raddr[5:5] memory.data_in[5:5] memory.data_out[5:5]</loc>
+          <loc side="bottom" yoffset="0">memory.wen memory.waddr[6:6] memory.raddr[6:6] memory.data_in[6:6] memory.data_out[6:6] memory.ren memory.data_in[7:7] memory.data_out[7:7]</loc>
+        </pinlocations>
+      </sub_tile>
+    </tile>
+
+As shown in :numref:`fig_example_subtile_direct_connection`, consider a connection where the ``out`` of a sub tile ``mult_8`` of tile ``cim8_1k`` drives the ``data_in`` of the sub tile ``memory`` of tile ``cim8_1k`` with an offset, using the delayless switch one would enter the following:
+
+.. code-block:: xml
+
+    <direct name="cim_direct0" from_pin="cim8_1k.out[7:0]" to_pin="cim8_1k.data_in[7:0]" x_offset="0" y_offset="0" z_offset="1"/>
+    <direct name="cim_direct1" from_pin="cim8_1k.out[15:8]" to_pin="cim8_1k.data_in[7:0]" x_offset="0" y_offset="0" z_offset="2"/>
+
+.. _fig_example_subtile_direct_connection:
+
+.. figure:: ./example_subtile_direct_connection.png
+   :width: 60%
+   :alt: Example of direct connections across sub-tiles
 
-        <direct name="adder_carry" from_pin="clb.cout" to_pin="clb.cin" x_offset="0" y_offset="-1" z_offset="0"/>
+   Example of direct connections across sub-tiles
 
 .. _custom_switch_blocks:
 

doc/src/arch/reference.rst

@@ -33,6 +33,11 @@ std::vector<t_clb_to_clb_directs> alloc_and_load_clb_to_clb_directs(const std::v
         clb_to_clb_directs[i].from_clb_type = physical_tile;
 
         t_physical_tile_port tile_port = find_tile_port_by_name(physical_tile, port_name);
+        /* Find the sub tile indices */
+        clb_to_clb_directs[i].from_sub_tiles = find_sub_tile_indices_by_port_name(physical_tile, port_name); 
+        if (clb_to_clb_directs[i].from_sub_tiles.empty()) {
+            VPR_THROW(VPR_ERROR_ARCH, "Unable to find sub tile under tile '%s' which contains the port %s.\n", tile_name.c_str(), port_name.data());
+        }
 
         if (start_pin_index == UNDEFINED) {
             VTR_ASSERT(start_pin_index == end_pin_index);

vpr/src/route/rr_graph_generation/clb2clb_directs.cpp

@@ -13,6 +13,7 @@
  */
 struct t_clb_to_clb_directs {
     t_physical_tile_type_ptr from_clb_type;
+    std::vector<int> from_sub_tiles;
     int from_clb_pin_start_index;
     int from_clb_pin_end_index;
     t_physical_tile_type_ptr to_clb_type;

vpr/src/route/rr_graph_generation/clb2clb_directs.h

@@ -2939,7 +2939,7 @@ static int get_opin_direct_connections(RRGraphBuilder& rr_graph_builder,
     auto [z, relative_opin] = get_capacity_location_from_physical_pin(curr_type, opin);
     VTR_ASSERT(z >= 0 && z < curr_type->capacity);
     const int num_directs = directs.size();
-
+    
     // Iterate through all direct connections
     for (int i = 0; i < num_directs; i++) {
         // Find matching direct clb-to-clb connections with the same type as current grid location
@@ -3012,15 +3012,21 @@ static int get_opin_direct_connections(RRGraphBuilder& rr_graph_builder,
                         // Add new ipin edge to list of edges
                         std::vector<RRNodeId> inodes;
 
+                        int target_width_offset = device_ctx.grid.get_width_offset({x + directs[i].x_offset, y + directs[i].y_offset, layer});
+                        int target_height_offset = device_ctx.grid.get_height_offset({x + directs[i].x_offset, y + directs[i].y_offset, layer});
+                        int final_ipin_x = x + directs[i].x_offset - target_width_offset + target_type->pin_width_offset[ipin];
+                        int final_ipin_y = y + directs[i].y_offset - target_height_offset + target_type->pin_height_offset[ipin];
+
                         if (directs[i].to_side != NUM_2D_SIDES) {
                             //Explicit side specified, only create if pin exists on that side
-                            RRNodeId inode = rr_graph_builder.node_lookup().find_node(layer, x + directs[i].x_offset, y + directs[i].y_offset, e_rr_type::IPIN, ipin, directs[i].to_side);
+                            RRNodeId inode = rr_graph_builder.node_lookup().find_node(layer, final_ipin_x, final_ipin_y,
+                                                                                      e_rr_type::IPIN, ipin, directs[i].to_side);
                             if (inode) {
                                 inodes.push_back(inode);
                             }
                         } else {
                             //No side specified, get all candidates
-                            inodes = rr_graph_builder.node_lookup().find_nodes_at_all_sides(layer, x + directs[i].x_offset, y + directs[i].y_offset, e_rr_type::IPIN, ipin);
+                            inodes = rr_graph_builder.node_lookup().find_nodes_at_all_sides(layer, final_ipin_x, final_ipin_y, e_rr_type::IPIN, ipin);
                         }
 
                         if (inodes.size() > 0) {

vpr/src/route/rr_graph_generation/rr_graph.cpp

@@ -1722,74 +1722,81 @@ void build_direct_connections_for_one_gsb(const RRGraphView& rr_graph,
         }
 
         /* get every opin in the range */
-        for (int opin = min_index; opin <= max_index; ++opin) {
-            int offset = opin - min_index;
+        for (int relative_opin = min_index; relative_opin <= max_index; ++relative_opin) {
+            int offset = relative_opin - min_index;
             //Capacity location determined by pin number relative to pins per capacity instance
-            auto [z, relative_opin] = get_capacity_location_from_physical_pin(grid_type, opin);
-            VTR_ASSERT(z >= 0 && z < grid_type->capacity);
-
-            if ((to_grid_coordinate.x() < grids.width() - 1)
-                && (to_grid_coordinate.y() < grids.height() - 1)) {
-                int relative_ipin = UNDEFINED;
-                if (clb_to_clb_directs[i].to_clb_pin_start_index
-                    > clb_to_clb_directs[i].to_clb_pin_end_index) {
-                    if (true == swap) {
-                        relative_ipin = clb_to_clb_directs[i].to_clb_pin_end_index + offset;
+            for (int z : clb_to_clb_directs[i].from_sub_tiles) {
+                int opin = get_physical_pin_from_capacity_location(grid_type, relative_opin, z);
+                VTR_ASSERT(z >= 0 && z < grid_type->capacity);
+
+                if ((to_grid_coordinate.x() < grids.width() - 1)
+                    && (to_grid_coordinate.y() < grids.height() - 1)) {
+                    int relative_ipin = UNDEFINED;
+                    if (clb_to_clb_directs[i].to_clb_pin_start_index
+                        > clb_to_clb_directs[i].to_clb_pin_end_index) {
+                        if (swap) {
+                            relative_ipin = clb_to_clb_directs[i].to_clb_pin_end_index + offset;
+                        } else {
+                            relative_ipin = clb_to_clb_directs[i].to_clb_pin_start_index - offset;
+                        }
                     } else {
-                        relative_ipin = clb_to_clb_directs[i].to_clb_pin_start_index - offset;
+                        if (swap) {
+                            relative_ipin = clb_to_clb_directs[i].to_clb_pin_end_index - offset;
+                        } else {
+                            relative_ipin = clb_to_clb_directs[i].to_clb_pin_start_index + offset;
+                        }
                     }
-                } else {
-                    if (true == swap) {
-                        relative_ipin = clb_to_clb_directs[i].to_clb_pin_end_index - offset;
-                    } else {
-                        relative_ipin = clb_to_clb_directs[i].to_clb_pin_start_index + offset;
+
+                    /* Get the pin index in the rr_graph */
+                    t_physical_tile_loc from_tile_loc(from_grid_coordinate.x(), from_grid_coordinate.y(), layer);
+                    t_physical_tile_loc to_tile_loc(to_grid_coordinate.x(), to_grid_coordinate.y(), layer);
+
+                    /* Find the side of grid pins, the pin location should be unique!
+                     * Pin location is required by searching a node in rr_graph
+                     */
+                    std::vector<e_side> opin_grid_side = find_grid_pin_sides(grids, layer, from_grid_coordinate.x() + grid_type->pin_width_offset[opin], from_grid_coordinate.y() + grid_type->pin_height_offset[opin], opin);
+                    if (1 != opin_grid_side.size()) {
+                      VPR_FATAL_ERROR(VPR_ERROR_ARCH, "[Arch LINE %d] From pin (index=%d) of direct connection '%s' does not exist on any side of the programmable block '%s'.\n", directs[i].line, opin, directs[i].from_pin.c_str());
                     }
-                }
 
-                /* Get the pin index in the rr_graph */
-                t_physical_tile_loc from_tile_loc(from_grid_coordinate.x(), from_grid_coordinate.y(), layer);
-                t_physical_tile_loc to_tile_loc(to_grid_coordinate.x(), to_grid_coordinate.y(), layer);
+                    /* directs[i].sub_tile_offset is added to from_capacity(z) to get the target_capacity */
+                    int to_subtile_cap = z + directs[i].sub_tile_offset;
+                    /* Iterate over all sub_tiles to get the sub_tile which the target_cap belongs to. */
+                    const t_sub_tile* to_sub_tile = nullptr;
+                    for (const t_sub_tile& sub_tile : to_grid_type->sub_tiles) {
+                        if (sub_tile.capacity.is_in_range(to_subtile_cap)) {
+                            to_sub_tile = &sub_tile;
+                            break;
+                        }
+                    }
+                    VTR_ASSERT(to_sub_tile != nullptr);
+                    if (relative_ipin >= to_sub_tile->num_phy_pins) continue;
+                    // If this block has capacity > 1 then the pins of z position > 0 are offset
+                    // by the number of pins per capacity instance
+                    int ipin = get_physical_pin_from_capacity_location(to_grid_type, relative_ipin, to_subtile_cap);
+                    std::vector<e_side> ipin_grid_side = find_grid_pin_sides(grids, layer, to_grid_coordinate.x() + to_grid_type->pin_width_offset[ipin], to_grid_coordinate.y() + to_grid_type->pin_height_offset[ipin], ipin);
+                    if (1 != ipin_grid_side.size()) {
+                      VPR_FATAL_ERROR(VPR_ERROR_ARCH, "[Arch LINE %d] To pin (index=%d) of direct connection '%s' does not exist on any side of the programmable block '%s'.\n", directs[i].line, relative_ipin, directs[i].to_pin.c_str());
+                    }
 
-                /* Find the side of grid pins, the pin location should be unique!
-                 * Pin location is required by searching a node in rr_graph
-                 */
-                std::vector<e_side> opin_grid_side = find_grid_pin_sides(grids, layer, from_grid_coordinate.x() + grid_type->pin_width_offset[opin], from_grid_coordinate.y() + grid_type->pin_height_offset[opin], opin);
-                if (1 != opin_grid_side.size()) {
-                    VPR_FATAL_ERROR(VPR_ERROR_ARCH, "[Arch LINE %d] From pin (index=%d) of direct connection '%s' does not exist on any side of the programmable block '%s'.\n", directs[i].line, opin, directs[i].from_pin.c_str());
-                }
-
-                /* directs[i].sub_tile_offset is added to from_capacity(z) to get the target_capacity */
-                int to_subtile_cap = z + directs[i].sub_tile_offset;
-                /* Iterate over all sub_tiles to get the sub_tile which the target_cap belongs to. */
-                const t_sub_tile* to_sub_tile = nullptr;
-                for (const t_sub_tile& sub_tile : to_grid_type->sub_tiles) {
-                    if (sub_tile.capacity.is_in_range(to_subtile_cap)) {
-                        to_sub_tile = &sub_tile;
-                        break;
+                    RRNodeId opin_node_id = rr_graph.node_lookup().find_node(layer,
+                                                                             from_grid_coordinate.x() + grid_type->pin_width_offset[opin],
+                                                                             from_grid_coordinate.y() + grid_type->pin_height_offset[opin],
+                                                                             e_rr_type::OPIN, opin, opin_grid_side[0]);
+                    RRNodeId ipin_node_id = rr_graph.node_lookup().find_node(layer,
+                                                                             to_grid_coordinate.x() + to_grid_type->pin_width_offset[ipin],
+                                                                             to_grid_coordinate.y() + to_grid_type->pin_height_offset[ipin],
+                                                                             e_rr_type::IPIN, ipin, ipin_grid_side[0]);
+
+                    /* add edges to the opin_node */
+                    if (!opin_node_id) {
+                      VTR_ASSERT(opin_node_id);
+                    }
+                    if (!ipin_node_id) {
+                      VTR_ASSERT(opin_node_id);
                     }
+                    rr_graph_builder.create_edge_in_cache(opin_node_id, ipin_node_id, RRSwitchId(clb_to_clb_directs[i].switch_index), false);
                 }
-                VTR_ASSERT(to_sub_tile != nullptr);
-                if (relative_ipin >= to_sub_tile->num_phy_pins) continue;
-                // If this block has capacity > 1 then the pins of z position > 0 are offset
-                // by the number of pins per capacity instance
-                int ipin = get_physical_pin_from_capacity_location(to_grid_type, relative_ipin, to_subtile_cap);
-                std::vector<e_side> ipin_grid_side = find_grid_pin_sides(grids, layer, to_grid_coordinate.x() + to_grid_type->pin_width_offset[ipin], to_grid_coordinate.y() + to_grid_type->pin_height_offset[ipin], ipin);
-                if (1 != ipin_grid_side.size()) {
-                    VPR_FATAL_ERROR(VPR_ERROR_ARCH, "[Arch LINE %d] To pin (index=%d) of direct connection '%s' does not exist on any side of the programmable block '%s'.\n", directs[i].line, relative_ipin, directs[i].to_pin.c_str());
-                }
-
-                RRNodeId opin_node_id = rr_graph.node_lookup().find_node(layer,
-                                                                         from_grid_coordinate.x() + grid_type->pin_width_offset[opin],
-                                                                         from_grid_coordinate.y() + grid_type->pin_height_offset[opin],
-                                                                         e_rr_type::OPIN, opin, opin_grid_side[0]);
-                RRNodeId ipin_node_id = rr_graph.node_lookup().find_node(layer,
-                                                                         to_grid_coordinate.x() + to_grid_type->pin_width_offset[ipin],
-                                                                         to_grid_coordinate.y() + to_grid_type->pin_height_offset[ipin],
-                                                                         e_rr_type::IPIN, ipin, ipin_grid_side[0]);
-
-                /* add edges to the opin_node */
-                VTR_ASSERT(opin_node_id && ipin_node_id);
-                rr_graph_builder.create_edge_in_cache(opin_node_id, ipin_node_id, RRSwitchId(clb_to_clb_directs[i].switch_index), false);
             }
         }
     }