synth: add SYNTH_MOCK_LARGE_MEMORIES

docs/user/FlowVariables.md

@@ -243,8 +243,9 @@ configuration file.
 | <a name="SYNTH_HIERARCHICAL"></a>SYNTH_HIERARCHICAL| Enable to Synthesis hierarchically, otherwise considered flat synthesis.| 0|
 | <a name="SYNTH_HIER_SEPARATOR"></a>SYNTH_HIER_SEPARATOR| Separator used for the synthesis flatten stage.| .|
 | <a name="SYNTH_KEEP_MODULES"></a>SYNTH_KEEP_MODULES| Mark modules to keep from getting removed in flattening.| |
-| <a name="SYNTH_MEMORY_MAX_BITS"></a>SYNTH_MEMORY_MAX_BITS| Maximum number of bits for memory synthesis.| 4096|
+| <a name="SYNTH_MEMORY_MAX_BITS"></a>SYNTH_MEMORY_MAX_BITS| Maximum number of bits for memory synthesis. Ideally, real RAM or realistic fakeram should be used for RAMs much larger than 1024 bits. To temporarily ignore the RAM concerns and investigate other aspects of the design, consider setting `SYNTH_MOCK_LARGE_MEMORIES=1`, or adjusting `SYNTH_MEMORY_MAX_BITS`.| 4096|
 | <a name="SYNTH_MINIMUM_KEEP_SIZE"></a>SYNTH_MINIMUM_KEEP_SIZE| For hierarchical synthesis, we keep modules of larger area than given by this variable and flatten smaller modules. The area unit used is the size of a basic nand2 gate from the platform's standard cell library. The default value is platform specific.| 0|
+| <a name="SYNTH_MOCK_LARGE_MEMORIES"></a>SYNTH_MOCK_LARGE_MEMORIES| Reduce memories larger than SYNTH_MEMORY_MAX_BITS to 1 row. This is useful and convenient to separate the concern of instantiating and placing memories from investigating other issues with a design, though it comes at the expense of the increased accuracy that using realistic fakemem would provide. Memories with a single 1 row will of course have unrealistically good timing and area characteristics, but timing will still correctly terminate in a register. Large port memories, typically register files, will still have the retain a lot of the port logic that can be useful to investigate issues. Consider using SYNTH_KEEP_MODULES to keep the modules of the mocked memories so that code outside the mocked memories is not optimized as a consequence of mocking a memory, yielding better insight into issues running the rest of the design through the ORFS flow.| 0|
 | <a name="SYNTH_NETLIST_FILES"></a>SYNTH_NETLIST_FILES| Skips synthesis and uses the supplied netlist files. If the netlist files contains duplicate modules, which can happen when using hierarchical synthesis on indvidual netlist files and combining here, subsequent modules are silently ignored and only the first module is used.| |
 | <a name="SYNTH_OPT_HIER"></a>SYNTH_OPT_HIER| Optimize constants across hierarchical boundaries.| |
 | <a name="SYNTH_RETIME_MODULES"></a>SYNTH_RETIME_MODULES| *This is an experimental option and may cause adverse effects.* *No effort has been made to check if the retimed RTL is logically equivalent to the non-retimed RTL.* List of modules to apply automatic retiming to. These modules must not get dissolved and as such they should either be the top module or be included in SYNTH_KEEP_MODULES. The main use case is to quickly identify if performance can be improved by manually retiming the input RTL. Retiming will treat module ports like register endpoints/startpoints. The objective function of retiming isn't informed by SDC, even the clock period is ignored. As such, retiming will optimize for best delay at potentially high register number cost. Automatic retiming can produce suboptimal results as its timing model is crude and it doesn't find the optimal distribution of registers on long pipelines. See OR discussion #8080.| |
@@ -284,6 +285,7 @@ configuration file.
 - [SYNTH_KEEP_MODULES](#SYNTH_KEEP_MODULES)
 - [SYNTH_MEMORY_MAX_BITS](#SYNTH_MEMORY_MAX_BITS)
 - [SYNTH_MINIMUM_KEEP_SIZE](#SYNTH_MINIMUM_KEEP_SIZE)
+- [SYNTH_MOCK_LARGE_MEMORIES](#SYNTH_MOCK_LARGE_MEMORIES)
 - [SYNTH_NETLIST_FILES](#SYNTH_NETLIST_FILES)
 - [SYNTH_OPT_HIER](#SYNTH_OPT_HIER)
 - [SYNTH_RETIME_MODULES](#SYNTH_RETIME_MODULES)

flow/designs/sky130hd/microwatt/README.md

@@ -0,0 +1,178 @@
+# Mocking vs fake memories
+
+Configuring fake memories for this design would speed up the ORFS flow and increase accuracy of results, but some effort is required. Two methods for learning something about a design without setting up SRAM are explained here.
+
+## Synthesis with large flip flop memories
+
+By default `SYNTH_MEMORY_MAX_BITS=42000` since no fake memories have been configured for this example design. This is simple, but results in slow builds and unrealistically large amount of flip flops and singificantly slower timing than fakeram or real RAMs.
+
+![RAMs as flops histogram](histogram-sram-as-flops.png)
+
+## Results with `SYNTH_MOCK_LARGE_MEMORIES=1`
+
+To ensure a quick synthesis run and to better understand the design without being slowed down by large memory blocks, we set a `SYNTH_MEMORY_MAX_BITS=1024`. This helps us bypass potential memory-related issues and focus on other ORFS flow issues of the design.
+
+During synthesis, certain modules are reported in error messages when `SYNTH_MEMORY_MAX_BITS=1024` and `SYNTH_MOCK_LARGE_MEMORIES=0`. By explicitly listing these modules in `SYNTH_KEEP_MODULES`, we avoid further optimizations outside of the mocked memories that could obscure the behavior of the rest of the design.
+
+The goal of these settings is to enable a rapid exploration of the flow, providing insights into the design while minimizing complications from large memory structures.
+
+![Mocked memory histogram](histogram-mock-memory.png)
+
+## Other ways to speed up synthesis
+
+There is a small advantage in synthesis time for `SYNTH_MOCK_LARGE_MEMORIES=1`, it shaves off ca. 1 minute on a 3 minute build on a test on a laptop. However, larger designs can have synthesis run into hours if memories are not managed with a bit of care.
+
+For synthesis, yosys-abc actually takes most of the time and SRAMs don't generally change in a design, even if other RTL development continues. It is possible to keep the synthesized netlist for SRAMs list them in `SYNTH_BLACKBOXES` and simply concatenate the already built netlists onto the `1_synth.v` files before continuing the flow.
+
+If large modules that change rarely are kept in a large design and only a small part of the design changes during RTL development, then it is possible to set up a build flow that completes in minutes instead of hours.
+
+## A/B run times
+
+The difference in run-times for mocking and simply instantiating larger flip flop based RAMs is not large on this design, but on designs with bigger SRAMs, the difference can be substantial.
+
+    make DESIGN_CONFIG=designs/sky130hd/microwatt/config.mk SYNTH_MOCK_LARGE_MEMORIES=1 FLOW_VARIANT=mock
+    make DESIGN_CONFIG=designs/sky130hd/microwatt/config.mk
+
+| Step                      | Mock RAM/s | Default/s |
+|---------------------------|------------|-----------|
+| 1_1_yosys_canonicalize    | 4          | 4         |
+| 1_2_yosys                 | 161        | 182       |
+| 1_3_synth                 |            | 1         |
+| 2_1_floorplan             | 72         | 83        |
+| 2_2_floorplan_macro       | 16         | 16        |
+| 2_3_floorplan_tapcell     | 1          | 0         |
+| 2_4_floorplan_pdn         | 7          | 9         |
+| 3_1_place_gp_skip_io      | 43         | 45        |
+| 3_2_place_iop             | 1          | 1         |
+| 3_3_place_gp              | 331        | 327       |
+| 3_4_place_resized         | 68         | 65        |
+| 3_5_place_dp              | 79         | 74        |
+| 4_1_cts                   | 152        | 180       |
+| 5_1_grt                   | 385        | 404       |
+| 5_2_route                 | 3827       | 3960      |
+
+## `SYNTH_MOCK_LARGE_MEMORIES=1` worst ext_clk path
+
+```
+Startpoint: soc0/processor/icache_0/rams:1.way/cache_ram_0
+            (rising edge-triggered flip-flop clocked by ext_clk)
+Endpoint: soc0/processor/icache_0/_163_[147]$_DFFE_PP_
+          (rising edge-triggered flip-flop clocked by ext_clk)
+Path Group: ext_clk
+Path Type: max
+
+  Delay    Time   Description
+---------------------------------------------------------
+   0.00    0.00   clock ext_clk (rise edge)
+   4.07    4.07   clock network delay (propagated)
+   0.00    4.07 ^ soc0/processor/icache_0/rams:1.way/cache_ram_0/CLK (RAM32_1RW1R)
+  11.44   15.51 v soc0/processor/icache_0/rams:1.way/cache_ram_0/Do1[31] (RAM32_1RW1R)
+   0.62   16.14 v soc0/processor/icache_0/rams:1.way/_43_/X (sky130_fd_sc_hd__mux2_4)
+   0.42   16.56 v soc0/processor/icache_0/_2550_/X (sky130_fd_sc_hd__mux2_4)
+   0.19   16.75 v place24125/X (sky130_fd_sc_hd__buf_12)
+   0.15   16.90 v soc0/processor/decode1_0/_2318_/Y (sky130_fd_sc_hd__nand2b_4)
+   0.38   17.27 v soc0/processor/decode1_0/_2375_/X (sky130_fd_sc_hd__or3_4)
+   0.14   17.41 ^ soc0/processor/decode1_0/_2737_/Y (sky130_fd_sc_hd__nor2_4)
+   0.07   17.48 v soc0/processor/decode1_0/_2738_/Y (sky130_fd_sc_hd__inv_2)
+   0.19   17.67 ^ soc0/processor/decode1_0/_2740_/Y (sky130_fd_sc_hd__a21oi_4)
+   0.24   17.91 v soc0/processor/decode1_0/_3744_/Y (sky130_fd_sc_hd__nand4b_1)
+   0.16   18.06 ^ soc0/processor/decode1_0/_4248_/Y (sky130_fd_sc_hd__nand2b_1)
+   0.16   18.22 ^ soc0/processor/_318_/X (sky130_fd_sc_hd__or2_4)
+   0.05   18.27 v soc0/processor/icache_0/_2130_/Y (sky130_fd_sc_hd__nor2_4)
+   0.09   18.36 ^ soc0/processor/icache_0/_2131_/Y (sky130_fd_sc_hd__nand2_4)
+   0.05   18.41 v soc0/processor/icache_0/_2132_/Y (sky130_fd_sc_hd__a211oi_4)
+   0.19   18.60 v rebuffer29423/X (sky130_fd_sc_hd__buf_12)
+   0.11   18.72 ^ soc0/processor/icache_0/_2133_/Y (sky130_fd_sc_hd__nand2_8)
+   0.06   18.78 v soc0/processor/icache_0/_2155_/Y (sky130_fd_sc_hd__inv_12)
+   0.14   18.92 v place19388/X (sky130_fd_sc_hd__buf_12)
+   0.14   19.06 v place19392/X (sky130_fd_sc_hd__buf_12)
+   0.14   19.19 v place19394/X (sky130_fd_sc_hd__buf_12)
+   0.16   19.35 v soc0/processor/icache_0/_2494_/X (sky130_fd_sc_hd__and2_4)
+   0.00   19.35 v soc0/processor/icache_0/_163_[147]$_DFFE_PP_/D (sky130_fd_sc_hd__edfxtp_1)
+          19.35   data arrival time
+
+  15.00   15.00   clock ext_clk (rise edge)
+   3.42   18.42   clock network delay (propagated)
+  -0.25   18.17   clock uncertainty
+   0.14   18.31   clock reconvergence pessimism
+          18.31 ^ soc0/processor/icache_0/_163_[147]$_DFFE_PP_/CLK (sky130_fd_sc_hd__edfxtp_1)
+  -0.24   18.07   library setup time
+          18.07   data required time
+---------------------------------------------------------
+          18.07   data required time
+         -19.35   data arrival time
+---------------------------------------------------------
+          -1.27   slack (VIOLATED)
+```
+
+## `SYNTH_MOCK_LARGE_MEMORIES=0` worst ext_clk path
+
+As can be seen, there's no significant difference in the worst negative slack path for ext_clk.
+
+```
+Startpoint: soc0/processor/icache_0/rams:1.way/cache_ram_0
+            (rising edge-triggered flip-flop clocked by ext_clk)
+Endpoint: soc0/processor/icache_0/_163_[14]$_SDFFE_PP0P_
+          (rising edge-triggered flip-flop clocked by ext_clk)
+Path Group: ext_clk
+Path Type: max
+
+  Delay    Time   Description
+---------------------------------------------------------
+   0.00    0.00   clock ext_clk (rise edge)
+   4.04    4.04   clock network delay (propagated)
+   0.00    4.04 ^ soc0/processor/icache_0/rams:1.way/cache_ram_0/CLK (RAM32_1RW1R)
+  11.44   15.48 v soc0/processor/icache_0/rams:1.way/cache_ram_0/Do1[59] (RAM32_1RW1R)
+   0.64   16.12 v soc0/processor/icache_0/rams:1.way/_76_/X (sky130_fd_sc_hd__mux2_4)
+   0.36   16.48 v soc0/processor/icache_0/_2544_/X (sky130_fd_sc_hd__mux2_4)
+   0.15   16.64 v place27067/X (sky130_fd_sc_hd__buf_6)
+   0.06   16.70 ^ soc0/processor/decode1_0/_3560_/Y (sky130_fd_sc_hd__inv_4)
+   0.16   16.85 ^ soc0/processor/decode1_0/_6875_/COUT (sky130_fd_sc_hd__ha_4)
+   0.07   16.92 v soc0/processor/decode1_0/_3695_/Y (sky130_fd_sc_hd__nand2b_4)
+   0.39   17.31 ^ soc0/processor/decode1_0/_3696_/Y (sky130_fd_sc_hd__nor3_4)
+   0.20   17.51 ^ place24130/X (sky130_fd_sc_hd__buf_6)
+   0.06   17.57 v soc0/processor/decode1_0/_5317_/Y (sky130_fd_sc_hd__nand2_4)
+   0.30   17.87 ^ soc0/processor/decode1_0/_5318_/Y (sky130_fd_sc_hd__a21oi_4)
+   0.19   18.06 ^ place23148/X (sky130_fd_sc_hd__buf_6)
+   0.22   18.28 v soc0/processor/decode1_0/_6350_/Y (sky130_fd_sc_hd__nand4b_1)
+   0.29   18.57 v place22875/X (sky130_fd_sc_hd__buf_6)
+   0.10   18.67 ^ soc0/processor/decode1_0/_6854_/Y (sky130_fd_sc_hd__nand2b_4)
+   0.15   18.82 ^ soc0/processor/_318_/X (sky130_fd_sc_hd__or2_4)
+   0.12   18.94 ^ place22433/X (sky130_fd_sc_hd__buf_12)
+   0.05   18.99 v soc0/processor/icache_0/_2130_/Y (sky130_fd_sc_hd__nor2_4)
+   0.16   19.15 v place22148/X (sky130_fd_sc_hd__buf_6)
+   0.08   19.23 ^ soc0/processor/icache_0/_2131_/Y (sky130_fd_sc_hd__nand2_4)
+   0.07   19.30 v soc0/processor/icache_0/_2132_/Y (sky130_fd_sc_hd__a211oi_4)
+   0.18   19.48 v place21617/X (sky130_fd_sc_hd__buf_12)
+   0.11   19.59 ^ soc0/processor/icache_0/_2133_/Y (sky130_fd_sc_hd__nand2_8)
+   0.06   19.65 v soc0/processor/icache_0/_2155_/Y (sky130_fd_sc_hd__inv_8)
+   0.15   19.80 v place21391/X (sky130_fd_sc_hd__buf_12)
+   0.13   19.93 v place21403/X (sky130_fd_sc_hd__buf_12)
+   0.14   20.07 v rebuffer32771/X (sky130_fd_sc_hd__buf_4)
+   0.19   20.26 ^ soc0/processor/icache_0/_2285_/Y (sky130_fd_sc_hd__mux2i_1)
+   0.15   20.40 ^ place21348/X (sky130_fd_sc_hd__buf_4)
+   0.04   20.44 v soc0/processor/icache_0/_2286_/Y (sky130_fd_sc_hd__nor2_1)
+   0.00   20.44 v soc0/processor/icache_0/_163_[14]$_SDFFE_PP0P_/D (sky130_fd_sc_hd__dfxtp_1)
+          20.44   data arrival time
+
+  15.00   15.00   clock ext_clk (rise edge)
+   3.53   18.53   clock network delay (propagated)
+  -0.25   18.28   clock uncertainty
+   0.14   18.42   clock reconvergence pessimism
+          18.42 ^ soc0/processor/icache_0/_163_[14]$_SDFFE_PP0P_/CLK (sky130_fd_sc_hd__dfxtp_1)
+  -0.11   18.31   library setup time
+          18.31   data required time
+---------------------------------------------------------
+          18.31   data required time
+         -20.44   data arrival time
+---------------------------------------------------------
+          -2.13   slack (VIOLATED)
+```
+
+## Conclusion
+
+Above there's no visible difference in the Endpoint Slack histogram for the two approaches. In other words, the design doesn't appear to be terribly sensitive to how RAMs are mocked, other factors dominate and merit further investigation.
+
+ORFS is built on `make`, which shines for simple, fast flows. For larger, complicated, designs and with flows that take a long time to run, it is worth looking beyond `make` to [bazel-orfs](https://github.com/The-OpenROAD-Project/bazel-orfs)
+
+

flow/designs/sky130hd/microwatt/config.mk

@@ -36,7 +36,23 @@ export SETUP_SLACK_MARGIN = 0.2
 # GRT non-default config
 export FASTROUTE_TCL = $(DESIGN_HOME)/$(PLATFORM)/$(DESIGN_NICKNAME)/fastroute.tcl
 
-# This is high, some SRAMs should probably be converted
-# to real SRAMs and not instantiated as flops
-export SYNTH_MEMORY_MAX_BITS = 42000
-
+ifeq ($(SYNTH_MOCK_LARGE_MEMORIES),1)
+    # ca. 3 minutes to run make synth
+    #
+    # These module names comes from the error report when setting SYNTH_MEMORY_MAX_BITS=2048
+    # and SYNTH_MOCK_LARGE_MEMORIES=0
+    #
+    # Keeping them avoids mocking them away, which would lead to further optimizations
+    # that would obscure what is going on in the rest of the design.
+    export SYNTH_KEEP_MODULES=decode1_0_bf8b4530d8d246dd74ac53a13471bba17941dff7 \
+      decode1_0_bf8b4530d8d246dd74ac53a13471bba17941dff7 \
+      fpu \
+      decode1_0_bf8b4530d8d246dd74ac53a13471bba17941dff7 \
+      decode1_0_bf8b4530d8d246dd74ac53a13471bba17941dff7 \
+      decode1_0_bf8b4530d8d246dd74ac53a13471bba17941dff7
+    # The goal is to run through the flow quickly to learn what we can
+    # about the design without getting bogged down in memory issues.
+    export SYNTH_MEMORY_MAX_BITS ?= 1024
+else
+    export SYNTH_MEMORY_MAX_BITS ?= 42000
+endif

flow/scripts/synth.tcl

@@ -55,6 +55,26 @@ if { !$::env(SYNTH_HIERARCHICAL) } {
   synth -flatten -run coarse:fine {*}$synth_full_args
 }
 
+
+if { $::env(SYNTH_MOCK_LARGE_MEMORIES) } {
+  memory_collect
+  set select [tee -q -s result.string select -list t:\$mem_v2]
+  foreach path [split [string trim $select] "\n"] {
+    set index [string first "/" $path]
+    set module [string range $path 0 [expr { $index - 1 }]]
+    set instance [string range $path [expr { $index + 1 }] end]
+
+    set width [rtlil::get_param -uint $module $instance WIDTH]
+    set size [rtlil::get_param -uint $module $instance SIZE]
+    set nbits [expr $width * $size]
+    puts "Memory $path has dimensions $size x $width = $nbits"
+    if { $nbits > $::env(SYNTH_MEMORY_MAX_BITS) } {
+      rtlil::set_param -uint $module $instance SIZE 1
+      puts "Shrunk memory $path from $size rows to 1"
+    }
+  }
+}
+
 json -o $::env(RESULTS_DIR)/mem.json
 # Run report and check here so as to fail early if this synthesis run is doomed
 exec -- $::env(PYTHON_EXE) $::env(SCRIPTS_DIR)/mem_dump.py \

flow/scripts/variables.yaml

@@ -174,9 +174,39 @@ SYNTH_HIERARCHICAL:
 SYNTH_MEMORY_MAX_BITS:
   description: >
     Maximum number of bits for memory synthesis.
+
+    Ideally, real RAM or realistic fakeram should be used for RAMs
+    much larger than 1024 bits.
+
+    To temporarily ignore the RAM concerns and investigate other
+    aspects of the design, consider setting `SYNTH_MOCK_LARGE_MEMORIES=1`,
+    or adjusting `SYNTH_MEMORY_MAX_BITS`.
   default: 4096
   stages:
     - synth
+SYNTH_MOCK_LARGE_MEMORIES:
+  description: >
+    Reduce memories larger than SYNTH_MEMORY_MAX_BITS to 1 row.
+
+    This is useful and convenient to separate the concern of instantiating
+    and placing memories from investigating other issues with a design,
+    though it comes at the expense of the increased accuracy that using
+    realistic fakemem would provide.
+
+    Memories with a single 1 row will of course have unrealistically good
+    timing and area characteristics, but timing will still correctly terminate
+    in a register.
+
+    Large port memories, typically register files, will still have the
+    retain a lot of the port logic that can be useful to investigate issues.
+
+    Consider using SYNTH_KEEP_MODULES to keep the modules of the mocked
+    memories so that code outside the mocked memories is not
+    optimized as a consequence of mocking a memory, yielding better insight
+    into issues running the rest of the design through the ORFS flow.
+  default: 0
+  stages:
+    - synth
 SYNTH_HDL_FRONTEND:
   description: >
     Select an alternative language frontend to ingest the design. Available option