Skip to content

synthesis_analysis.md

Latest commit

 

History

History
556 lines (397 loc) · 21.6 KB

synthesis_analysis.md

File metadata and controls

556 lines (397 loc) · 21.6 KB
Raw

Synthesis and Power Analysis

This guide explains how to run synthesis, timing analysis, area analysis, and power analysis for Vortex across all supported back-ends: Xilinx (Vivado), Altera (Quartus), Yosys (open-source), and Synopsys Design Compiler.

Table of Contents

Design Configuration

All synthesis flows accept a CONFIGS variable to customize the hardware design at build time. CONFIGS is a string of preprocessor macro definitions (-D flags) that control core count, cache hierarchy, extensions, and other parameters.

Common configuration flags (all parameters live in the VX_CFG_* namespace; see VX_config.toml at the repo root for the full list):

Flag Description
-DVX_CFG_NUM_CLUSTERS=N Number of clusters
-DVX_CFG_NUM_CORES=N Number of cores per cluster
-DVX_CFG_NUM_WARPS=N Number of warps per core
-DVX_CFG_NUM_THREADS=N Number of threads per warp
-DVX_CFG_L2_ENABLE Enable shared L2 cache
-DVX_CFG_L3_ENABLE Enable shared L3 cache
-DVX_CFG_EXT_TCU_ENABLE Enable Tensor Core Unit
-DVX_CFG_EXT_DXA_ENABLE Enable DXA extension
-DVX_CFG_DCACHE_SIZE=N Set data cache size in bytes

Example:

CONFIGS="-DVX_CFG_NUM_CLUSTERS=1 -DVX_CFG_NUM_CORES=4 -DVX_CFG_L2_ENABLE -DVX_CFG_EXT_TCU_ENABLE"

Overriding top-module parameters (-G)

CONFIGS may also carry Verilog parameter overrides as -G<NAME>=<value> alongside the -D macros. gen_sources.sh forwards these to repl_params.py, which rewrites the parameter's default in the top module's per-build copy (the source tree is untouched, so concurrent builds with different overrides stay isolated). gen_config ignores -G tokens, so no separate Makefile variable is needed.

This is mainly for the DUT unittest wrappers, whose knobs are Verilog parameters rather than VX_CFG_* macros — e.g. the cache wrapper's AMO_ENABLE and IS_LLC:

# LLC cache DUT, AMO disabled vs enabled (NT=NW=32)
CONFIGS="-DVX_CFG_NUM_THREADS=32 -DVX_CFG_NUM_WARPS=32 -GAMO_ENABLE=0 -GIS_LLC=1" PREFIX=amo0 make cache
CONFIGS="-DVX_CFG_NUM_THREADS=32 -DVX_CFG_NUM_WARPS=32 -GAMO_ENABLE=1 -GIS_LLC=1" PREFIX=amo1 make cache

All flows also support the NUM_CORES Makefile shorthand which auto-selects a pre-defined cluster/core/L2 configuration:

NUM_CORES=4   # equivalent to -DVX_CFG_NUM_CLUSTERS=1 -DVX_CFG_NUM_CORES=4 -DVX_CFG_L2_ENABLE
NUM_CORES=16  # equivalent to -DVX_CFG_NUM_CLUSTERS=1 -DVX_CFG_NUM_CORES=16 -DVX_CFG_L2_ENABLE
NUM_CORES=32  # equivalent to -DVX_CFG_NUM_CLUSTERS=2 -DVX_CFG_NUM_CORES=16 -DVX_CFG_L2_ENABLE

Using PREFIX for Isolated Builds

Use PREFIX=<unique_build_dir> to keep builds separate. Each flow creates a build directory derived from PREFIX so that multiple configurations can coexist without overwriting each other:

# Xilinx XRT: creates build_4c_<platform>_<target>/
PREFIX=build_4c NUM_CORES=4 make -C hw/syn/xilinx/xrt

# Synopsys: creates my_test_Vortex/
PREFIX=my_test make -C hw/syn/synopsys synthesis

Generating SAIF Files

SAIF (Switching Activity Interchange Format) files capture signal toggle rates during simulation and are used to produce accurate power estimates. Vortex supports SAIF generation through its RTL simulators: rtlsim, opaesim, and xrtsim.

Method 1: Build the Simulator Directly

Build the simulator with SAIF tracing enabled, then run a workload:

# Build rtlsim with SAIF support
make -C sim/rtlsim SAIF=1

# Run a test application
make -C tests/regression/sgemm run-rtlsim

The SAIF file is written to trace.saif in the application directory.

Method 2: Use the Blackbox Test Driver

The ci/blackbox.sh script provides a convenient wrapper:

./ci/blackbox.sh --driver=rtlsim --app=sgemm --cores=4 --l2cache --saif

When --saif is passed, blackbox.sh:

  1. Builds the simulator with SAIF=1
  2. Runs the application
  3. Copies the resulting trace.saif to the current directory

Available drivers for SAIF generation:

Driver Simulator Use Case
rtlsim Verilator RTL sim General-purpose RTL power analysis
opaesim OPAE AFU simulator Intel/Altera platform-specific analysis
xrtsim XRT simulator Xilinx platform-specific analysis

Specifying SAIF_INST

When reading a SAIF file, the tool must strip the testbench hierarchy prefix from signal names so they align with the synthesized netlist. SAIF_INST specifies this prefix.

Typical values:

Flow SAIF_INST
Xilinx DUT TOP.rtlsim_shim.vortex
Xilinx XRT TOP.vortex_afu_shim.vortex_afu
Synopsys / Yosys Instance path matching your simulation hierarchy

The path does not have to be absolute. A module instance name works if the tool can resolve it unambiguously (find-first semantics).

If the SAIF root scope already matches the top module, leave SAIF_INST empty.

DUT Sub-Component Evaluation

Both Xilinx and Altera provide DUT (Device Under Test) flows for synthesizing and analyzing sub-components in isolation, without the full platform wrapper. This is useful for evaluating individual units such as the TCU, FPU, cache, or a single core.

Xilinx DUT Targets

Located in hw/syn/xilinx/dut/. Available sub-component targets:

Target Module Description
unittest Unit tests Basic block tests
scope Scope analyzer Debug scope
mem_unit Memory unit Memory subsystem
lmem Local memory Local/shared memory
cache Cache Cache subsystem
fpu FPU Floating-point unit
tcu TCU Tensor Core Unit
dxa DXA DXA extension
core Core Single core
issue Issue unit Instruction issue
vortex Vortex Full processor (no AFU wrapper)
top Top Full design with AFU 
cd hw/syn/xilinx/dut

# Synthesize the TCU in isolation
CONFIGS="-DVX_CFG_EXT_TCU_ENABLE" make tcu

# Synthesize a 4-core Vortex without the platform wrapper
CONFIGS="-DVX_CFG_NUM_CORES=4 -DVX_CFG_L2_ENABLE" make vortex

# Run power analysis on an existing tcu
make tcu-power SAIF_FILE=/path/to/trace.saif SAIF_INST=*.tensor_unit

Each target creates its build under <target>/<BUILD_DIR>/ (e.g., tcu/build/).

Altera DUT Targets

Located in hw/syn/altera/dut/. Same set of sub-component targets as Xilinx. Requires the DEVICE_FAMILY variable and IP cache generation:

cd hw/syn/altera/dut

# Generate IP cache first (required for fpu, vortex, top)
make ip-gen

# Synthesize TCU for Arria 10
DEVICE_FAMILY=arria10 CONFIGS="-DVX_CFG_EXT_TCU_ENABLE" make tcu

# Synthesize a single core for Stratix 10
DEVICE_FAMILY=stratix10 make core

Build directories include the device family: <target>/build_<device_family>/.

Xilinx (Vivado)

XRT Full-Platform Flow

Located in hw/syn/xilinx/xrt/. Builds a complete Vitis xclbin for deployment on Xilinx FPGAs.

Supported platforms: Alveo U50, U55C, U200, U250, U280, Versal VCK5000.

Running Synthesis

cd hw/syn/xilinx/xrt

# Build a 4-core design for U280
PREFIX=build_4c NUM_CORES=4 TARGET=hw \
  PLATFORM=xilinx_u280_gen3x16_xdma_1_202310_1 \
  CONFIGS="-DVX_CFG_L2_ENABLE -DVX_CFG_DCACHE_SIZE=8192" \
  make > build.log 2>&1 &

Key variables:

Variable Default Description
PREFIX build$(XLEN) Build directory prefix
TARGET hw hw for hardware, hw_emu for emulation
PLATFORM (required) Xilinx platform identifier
NUM_CORES - Shorthand for core configuration
CONFIGS - Additional design macros
MAX_JOBS 8 Parallel Vivado jobs

Power Analysis

make power SAIF_FILE=/path/to/trace.saif SAIF_INST=TOP.vortex_afu_shim.vortex_afu BUILD_DIR=<build_dir>

The script (hw/scripts/xilinx_power_analysis.tcl) resolves the post-implementation checkpoint from BUILD_DIR automatically.

Where to Find Reports

XRT flow (under <BUILD_DIR>/):

Report Location Content
Utilization <BUILD_DIR>/bin/utilization.rpt LUTs, FFs, BRAM, DSP
Timing <BUILD_DIR>/bin/timing.rpt Worst setup paths
Power (vectorless) power_vectorless.rpt Baseline power estimate
Power (SAIF) power_saif.rpt Activity-annotated power

DUT flow (under <target>/<BUILD_DIR>/):

Report Location Content
Post-synth utilization post_synth_util.rpt Hierarchical resource usage
Post-impl utilization post_impl_util.rpt Hierarchical resource usage after P&R
Timing timing.rpt 100 worst setup paths
Methodology methodology.rpt Design rule checks
Clock utilization clock_utilization.rpt Clock tree and register usage
RAM utilization ram_utilization.rpt Detailed RAM/BRAM usage
Power (vectorless) power_vectorless.rpt Baseline power
Power (VCD) power_vcd.rpt VCD-annotated power (if VCD_FILE set)
Power (SAIF) power_saif.rpt SAIF-annotated power (via make power)
DRC drc.rpt Design rule violations
High fanout nets high_fanout_nets.rpt Nets with >100 fanout

Finding Key Metrics

  • Fmax: Look in timing.rpt for the worst negative slack (WNS). Fmax = 1 / (clock_period - WNS).
  • Total LUTs: In post_impl_util.rpt, find the row for CLB LUTs or Slice LUTs.
  • Total DSPs: In post_impl_util.rpt, find the row for DSPs or DSP48E2.
  • Total BRAM: In post_impl_util.rpt, find the row for Block RAM Tile or RAMB36/RAMB18.

Altera (Quartus)

OPAE Full-Platform Flow

Located in hw/syn/altera/opae/. Builds AFU images for Intel OPAE platforms (Arria 10, Stratix 10).

Running Synthesis

cd hw/syn/altera/opae

# Full build: IP generation, setup, and synthesis
DEVICE_FAMILY=arria10 PREFIX=build_4c NUM_CORES=4 TARGET=fpga make

# For ASE simulation build
DEVICE_FAMILY=stratix10 TARGET=asesim make

Key variables:

Variable Default Description
DEVICE_FAMILY arria10 arria10 or stratix10
PREFIX build$(XLEN) Build directory prefix
TARGET fpga fpga, asesim, or ase
NUM_CORES - Shorthand for core configuration
CONFIGS - Additional design macros

Build directory: <PREFIX>_<device_family>_<target>_<num_cores>c/.

Power Analysis

Altera power analysis uses Quartus PowerPlay with VCD-based toggle annotation:

# Located in hw/syn/altera/power_play.sh
quartus_pow --input_vcd=trace.vcd \
  --vcd_filter_glitches=on \
  --default_input_io_toggle_rate=10000transitions/s \
  $ProjectName

Where to Find Reports

Reports are generated in the synthesis build directory by the Quartus report scripts.

Area reports (from report_area.tcl):

Report Content
*.syn.area.resource_summary.csv Synthesis resource summary
*.syn.area.resource_breakdown.csv Resource breakdown by entity
*.syn.area.ram_summary.csv Synthesis RAM summary
*.syn.area.stats.csv Post-synthesis netlist statistics
*.fit.area.resource_summary.csv Fitter resource summary (post-P&R)
*.fit.area.resource_breakdown.csv Fitter resource breakdown by entity
*.fit.area.ram_summary.csv Fitter RAM summary
*.fit.area.routing_summary.csv Routing utilization
*.fit.area.routing_global.csv Global signal routing
*.fit.area.routing_high_fanout.csv High fanout signal routing

Timing reports (from analyze_timing.tcl):

Report Content
*.fit.timing.summary.txt Summary with Fmax, setup/hold, clock summary
*.fit.timing.setup.html Top 200 setup violation paths (with routing)
*.fit.timing.hold.html Top 200 hold violation paths
*.fit.timing.recovery.html Recovery timing paths
*.fit.timing.removal.html Removal timing paths
*.fit.timing.check_errors.html Timing DRC (no clock, multiple clock, loops)
*.fit.timing.check_metastability.html Metastability report
*.fit.timing_histogram.*.setup.html Per-clock setup slack histograms
*.fit.timing.setup.bottlenecks.txt Bottleneck analysis (TNS, fanout, fanin)
*.fit.timing.summary.fmax.csv Fmax summary (CSV)
*.fit.timing.summary.setup.csv Setup summary (CSV)
*.fit.timing.summary.hold.csv Hold summary (CSV)
*.fit.timing.summary.multicorner.csv Multi-corner timing summary

Finding Key Metrics

  • Fmax: Open *.fit.timing.summary.txt or *.fit.timing.summary.fmax.csv. The Fmax summary reports the restricted Fmax for each clock domain.
  • Total ALMs/LUTs: In *.fit.area.resource_summary.csv, look for ALMs needed (Stratix 10) or Logic utilization (Arria 10).
  • Total DSPs: In *.fit.area.resource_summary.csv, look for the DSP row.
  • Total BRAM (M20K/M10K): In *.fit.area.resource_summary.csv, look for M20K blocks or M10K blocks. Also see *.fit.area.ram_summary.csv for detailed RAM usage by entity.

Yosys (Open-Source)

Located in hw/syn/yosys/. Uses Yosys for synthesis, optional ABC for technology mapping, and OpenSTA for static timing analysis.

Running Synthesis

cd hw/syn/yosys

# Synthesis only (generic gates)
PREFIX=test NUM_CORES=1 make synthesis

# Synthesis + technology mapping
PREFIX=test NUM_CORES=1 make techmap

# Full flow: synthesis + mapping + STA + power
PREFIX=test NUM_CORES=1 SAIF_FILE=/path/to/trace.saif SAIF_INST=<inst> make timing

Key variables:

Variable Default Description
PREFIX build Build directory prefix
TOP_LEVEL_ENTITY Vortex Top module name
NUM_CORES - Shorthand for core configuration
CONFIGS - Additional design macros
CLOCK_FREQ 800 Target clock frequency in MHz
DELAY_UNC 0.02 Clock uncertainty (fraction of period)
DELAY_IO 0.05 I/O delay (fraction of period)
LIB_TGT NanGate 15nm OCL Liberty file for technology mapping
SAIF_FILE - SAIF file for power annotation
SAIF_INST - Instance path prefix to strip

Build directory: <PREFIX>_<TOP_LEVEL_ENTITY>/ (e.g., test_Vortex/).

The flow uses sv2v to convert SystemVerilog sources to Verilog before feeding them to Yosys.

SRAM Area Estimation

Yosys uses blackbox modules (VX_dp_ram_asic, VX_sp_ram_asic) for SRAM. The sram_cost.py script estimates SRAM area from the Yosys JSON netlist by inferring width and depth from port connectivity:

Area = (width x depth x SRAM_BIT_AREA) + SRAM_OVERHEAD

Defaults: SRAM_BIT_AREA=0.1 um^2/bit, SRAM_OVERHEAD=100.0 um^2. These can be overridden via environment variables.

Where to Find Reports

All reports are under <BUILD_DIR>/reports/:

Report Content
yosys.log Full Yosys synthesis log
stat_lib.rpt Cell count and area (post-mapping, by liberty cell type)
sram_area.rpt Estimated SRAM area breakdown
sta.log OpenSTA timing log
power.rpt Power estimate (vectorless or SAIF-annotated)
power_hier.rpt Hierarchical power breakdown
saif_unannotated.rpt Signals not covered by SAIF

Netlists are written to <BUILD_DIR>/out/:

File Content
<TOP>_syn.v Post-synthesis generic netlist
<TOP>_mapped.v Post-mapping technology netlist
<TOP>.json Yosys JSON netlist (used by sram_cost.py)

Finding Key Metrics

  • Total area: In stat_lib.rpt, look for Chip area for top module. Add the estimated SRAM area from sram_area.rpt for the total.
  • Cell count: In stat_lib.rpt, the per-cell-type breakdown shows gate counts.
  • Timing (WNS/TNS): In sta.log, look for the report_wns and report_tns outputs. Fmax = 1 / (target_period - WNS).
  • Power: In power.rpt, look for total power. power_hier.rpt breaks it down by hierarchy.

Synopsys Design Compiler

Located in hw/syn/synopsys/. Uses Synopsys DC for ASIC synthesis with support for multiple technology libraries.

Supported Libraries

LIB_TYPE Technology Path
DEFAULT NanGate 15nm OCL Bundled in hw/syn/libs/
ASAP7 ASAP7 7nm /mnt/nas0/eda.libs/asap7/asap7sc7p5t_28/LIB/NLDM
SAED14 SAED 14nm SLVT /mnt/nas0/eda.libs/saed14/EDK_03_2025

Running Synthesis

cd hw/syn/synopsys

# Default library, 1 core
PREFIX=test make synthesis

# ASAP7 library, 4 cores, with SAIF power
PREFIX=test NUM_CORES=4 LIB_TYPE=ASAP7 \
  SAIF_FILE=/path/to/trace.saif SAIF_INST=<inst> \
  make synthesis

# Synthesis without SRAM macros (blackbox)
PREFIX=test make synthesis-nosram

# Synthesis with estimated SRAM area
PREFIX=test make synthesis-estsram

Key variables:

Variable Default Description
PREFIX build Build directory prefix
TOP_LEVEL_ENTITY Vortex Top module name
NUM_CORES - Shorthand for core configuration
CONFIGS - Additional design macros
CLOCK_FREQ 800 Target frequency in MHz
DELAY_UNC 0.02 Clock uncertainty (fraction of period)
DELAY_IO 0.05 I/O delay (fraction of period)
LIB_TYPE DEFAULT Technology library selection
SAIF_FILE - SAIF file for power annotation
SAIF_INST - Instance path prefix in SAIF

Build directory: <PREFIX>_<TOP_LEVEL_ENTITY>/ (e.g., test_Vortex/).

Synthesis Variants

  • synthesis: Full synthesis with generated SRAM wrappers from the technology library's SRAM .db files. Requires the library to provide SRAM models.
  • synthesis-nosram: Synthesis without any SRAM logic. RAM modules are inferred by DC.
  • synthesis-estsram: Blackboxes VX_dp_ram_asic and VX_sp_ram_asic, then estimates their area from port dimensions (same approach as Yosys).

Where to Find Reports

All reports are under <BUILD_DIR>/reports/:

Report Content
area.rpt Hierarchical area breakdown
qor.rpt Quality of Results summary (area, timing, utilization)
timing_max.rpt Setup timing (50 worst paths, with nets/transitions/capacitance)
timing_min.rpt Hold timing (50 worst paths)
clock_skew.rpt Clock skew analysis
constraints_violators.rpt All constraint violations
check_design.rpt Pre-synthesis design checks
power_active.rpt SAIF-annotated hierarchical power (if SAIF_FILE provided)
power_vectorless.rpt Vectorless power estimate (if no SAIF_FILE)
saif_annotation_coverage.rpt SAIF annotation coverage statistics

Outputs are under <BUILD_DIR>/out/:

File Content
<TOP>.mapped.ddc Synopsys binary netlist
<TOP>.mapped.v Mapped gate-level Verilog
<TOP>.mapped.sdf Standard Delay Format for back-annotation
<TOP>.post_compile.sdc Post-compile timing constraints

Finding Key Metrics

  • Total area: In area.rpt, look for the top-level Total cell area. The SRAM estimated area (if using synthesis-estsram) is printed in the build log as Total Estimated SRAM Area.
  • Timing / Fmax: In timing_max.rpt, the slack of the first path gives the worst negative slack (WNS). Fmax = 1 / (target_period - WNS). Also check qor.rpt for a summary.
  • Power: In power_active.rpt (with SAIF) or power_vectorless.rpt (without), the hierarchical breakdown shows internal, switching, and leakage power per module.
  • Gate count: In qor.rpt, look for Design Area and Number of cells.

Understanding Power Reports

Power reports across all flows break down total power into similar categories. Understanding these helps identify optimization targets.

Power Components

Component Description
Dynamic power Power consumed by signal switching activity
  Internal Short-circuit current during output transitions within cells
  Switching Charging/discharging of interconnect and load capacitances
Static (leakage) power Power consumed even when signals are not switching; due to sub-threshold and gate leakage currents

Total Power = Dynamic (Internal + Switching) + Static (Leakage)

Vectorless vs. Activity-Annotated

  • Vectorless: The tool assumes a default toggle rate (typically 12.5%) and static probability (0.5) for all signals. Provides a rough baseline but can significantly over- or under-estimate actual power.
  • SAIF/VCD-annotated: Uses real switching activity captured during simulation. Much more accurate for the specific workload simulated. Signals not covered by the SAIF/VCD fall back to the default toggle rate.

Always compare the vectorless and annotated reports to understand which modules differ most from the default assumption.

Tips

  • Run a representative workload when generating SAIF files. Short or trivial tests will underestimate steady-state power.
  • Check SAIF annotation coverage reports (saif_annotation_coverage.rpt in Synopsys, saif_unannotated.rpt in Yosys) to ensure good signal coverage.
  • For Xilinx, the power report includes device-specific contributions (clocking, I/O, BRAM, DSP power) that are not present in ASIC flows.
  • For hierarchical analysis, look at per-module power breakdowns to identify the most power-hungry blocks (e.g., caches, FPU, TCU).