Skip to content

Latest commit

 

History

History
462 lines (378 loc) · 24.2 KB

CHANGELOG.md

File metadata and controls

462 lines (378 loc) · 24.2 KB
Raw

Changelog

All notable changes to this project will be documented in this file.

The format is based on Keep a Changelog and this project adheres to Semantic Versioning.

[Unreleased]

Fixed

  • Avoid corner-case in which the sequencer issues the same instruction multiple times when two units become non-ready at the same time
  • The lane sequencer now calculates the correct number of elements to be requested by the MASKU operand requesters
  • When the instruction queue of the SLDU is not empty, read from it to update the commit counter, and not from the incoming request
  • Masked operations now target the MASKU as well for the issue checks in the main sequencer
  • The granularity of the issue checks/gold tickets is now at a PE level
  • Avoid losing hazard-related information in the pipeline between the main sequencer and the operand requesters
  • Fix anticipated grant bug from operand requester to LDU, SLDU, MASKU, because of the stream registers. Now, the three units wait for a final true grant before commiting
  • The mask unit does not require synchronized lanes anymore to commit an instruction
  • MASKU does not wait anymore for valid incoming data from inactive lanes
  • Fix corner-case comparison in MASKU to provide the expected behavior
  • Fix MaskB-queue vector length in the lane sequencer
  • VALU cleans up the partial result of a reduction when no more needed
  • VALU can step from normalOp->reduction and vice-versa without issues
  • VALU's counters can now count bits when operating on mask vectors
  • Fix the vector length for mask instructions that run on the VALU/VMFPU
  • Don't let indexed memory operations interfere with the reductions in MASKU
  • Enforce strict in-order execution of FPU operations in VMFPU
  • Fixed AXI-inval-filter policy for D$ lines invalidation upon vector stores that are misaligned w.r.t. the D$ line width
  • Fix lane sequencer checks for floating-point comparisons
  • Fix synthesis error occuring due to the continuous assignmnet in the always block of mask unit
  • Fix wrong variable in vmerge and vmv riscv-tests
  • Re-introduce WAIT_STATE to avoid hazards when changin LMUL
  • Fix the PEs-ready signals related conditions in the main sequencer
  • Fix misaligned memory operations with more than 255 beats (>= 256 beats)
  • Fix stripmining condition in dispatcher
  • CVA6 tracks writes to floating-point scalar registers by the accelerator
  • Fixed de-synch bug in vector-FPU
  • Fix masked VSLIDEUP. Use only the mask bits with index higher than the stride
  • Fix SLDU issue_counter modification upon new VSLIDEUP incoming instruction
  • Fix whole-register-move destination register re-encoding
  • vslide1up always writes the scalar element in vd
  • Don't trim vslide1up counters since it always writes the scalar element
  • Wait for the current reduction to be over to execute the next VALU instruction, also when reduction workload is unbalanced and some lanes are only a pass-through
  • Reshuffle the source registers vs when an in-lane operation operates on element with vsew != eew_q[vs]
  • Fix the data target for .spike app compilations
  • make -C apps clean performs a deeper clean of the temporary object files
  • Fix riscv-isa-sim patch and bump pointer to riscv-isa-sim submodule
  • Fix VALU issue_counter initialization for mask logical operations
  • vslideup instructions that have stride >= vl_q have no effect
  • benchmark.sh keeps into account the number of lanes during program compilation
  • Fix typo in performance.py related to fconv2d
  • Fix lmul checks on vs1 for conversion instructions
  • Fix eew_vs1 for widening instructions
  • Fix cleaning of the accumulator after partial reduction in valu
  • De-parametrize FFT on the data-type
  • New reductions should start only when the previous operation is over
  • Prevent VMFPU from acknowledging the opqueues when their issue is over
  • Operand requesters sanitize partial operands during a reduction
  • Fix load/store-complete signals to CVA6
  • Remove latches/repeated-signals from masku
  • Reshuffle all the registers of a register group that have eew_q != eew_d when LMUL > 1
  • VLXE and VSXE need to wait that the SlideAddrGenA opreq is free before being issued by the lane sequencer to the operand requester stage
  • Do not trap instructions with no operands in the main sequencer
  • Commit a reduction only after a grant from the VRF
  • Solve duplicate verify function in roi_align
  • Don't let the lanes sample more than once the same instruction when the workload is unbalanced and the instruction stalls in the main sequencer
  • Don't mute instructions on mask vectors in the lane sequencer when vl == 0 in that lane
  • Remove unintentional latches from valu, simd_mul, lane_sequencer
  • Fix vxsat CSR update in dispatcher
  • Fix parameter passing through the hierarchy for fixed point support
  • Decouple cmdBuffer and dataBuffer depths in opQueues
  • Avoid handshaking wrong results in VMFPU
  • Fix eew reshuffle for mask logical operations
  • Fix eew reshuffle for vmv.v.v operations
  • Remove timing loop between vmfpu and operand_requesters
  • Solve some of the Warnings from SpyGlass run
  • Fix masku handshake. All the lanes should handshake together in input
  • Start solving sldu counter widths warnings
  • Fix vslideup wrong counter trimming
  • Reset gating registers before the integer multipliers in vmfpu
  • Fix narrowing for vnclip and vnclipu

Added

  • The sequencer can issue instructions to non-full units even if the other units are full
  • Vector indexed unordered/ordered load (vluxei8, vluxei16, vluxei32, vluxei64, vloxei8, vloxei16, vloxei32, vloxei64)
  • Vector indexed unordered/ordered stores (vsuxei8, vsuxei16, vsuxei32, vsuxei64, vsoxei8, vsoxei16, vsoxei32, vsoxei64)
  • Vector integer reductions (vredsum, vredmaxu, vredmax, vredminu, vredmin, vredand, vredor, vredxor, vwredsumu, vwredsum)
  • Introduce the global hazard table in the main sequencer, to provide up-to-date information to the operand requesters about the status of the different dependant instructions
  • Integer and Floating-Point scalar move instructions (vmv.x.s, vmv.s.x, vmv.f.s, vmv.s.f)
  • Add support for apps simulation with spike
  • Support for vector single-width floating-point reduction instructions: vfredusum, vfredosum, vfredmin, vfredmax
  • Support for Vector widening floating-point reductions: vfwredusum, vfwredosum
  • fdotproduct benchmark to evaluate dot products with reductions
  • fredsum benchmark to evaluate fp-reductions
  • riscv-tests for vfredusum, vfredosum, vfredmin, vfredmax, vfwredusum, vfwredosum
  • Support for vector single-width fractional multiply with rounding and saturation instruction: vsmul
  • Support for vector single-width scaling shift instructions: vssra, vssrl
  • Support for vector narrowing fixed-point clip instructions: vnclip, vnclipu
  • Vector floating-point classify instruction (vfclass)
  • Vector floating-point divide instructions (vfdiv, vfrdiv)
  • Vector floating-point square-root instruction (vfsqrt)
  • Add simple test for source registers reshuffle, when VSEW != EEW_vs
  • Add support for ideal-dispatcher simulation
  • Add compile-time garbage-collection to strip unused functions out and decrease the memory footprint of the binary
  • Add benchmarking capability with the ideal-dispatcher system + performance plotting. The support is limited to simluation with QuestaSim only
  • Plot jacobi2d performance
  • Plot dropout performance
  • Add FFT benchmark and print its performance
  • Add DWT benchmark and print its performance
  • Add fp-exp, fp-cos, fp-log benchmarks from rivec bmark suite + print performance
  • Ideal Dispatcher tracer now supports strided memory operations
  • Add Softmax benchmark and print its performance
  • Add [f]dotproduct apps and benchmarks, and plot their performance
  • Add a specific script to handle [f]dotproduct performance plotting
  • Add a script to benchmark all the applications together locally
  • Add HW cycle counter to better measure runtime cycles with short vectors
  • Add Pathfinder app and benchmark, and plot its performance
  • Add Roi-Align app and benchmark, and plot its performance
  • Support for vector mask instructions: vmsbf, vmsof, vmsif, viota, vid
  • Add support for vector mask population count and find first set bit instructions: vcpop.m, vfirst.m
  • Add Spyglass linting script
  • Add parametrized support for Fixed-Point math
  • Add optimized ASM version of exp 64-bit
  • Add optimized ASM version of roi_align
  • Add support for cache warming before benchmarks
  • Add support to check the results of the ideal dispatcher runs
  • Add HW/SW environment for automatic VCD dumping
  • Support for vector floating-point reciprocal estimate instruction: vfrec7
  • Support for vector floating-point reciprocal square-root estimate instruction: vfrsqrt7
  • Support for vector narrowing floating-point convert instruction: vfncvt.rod.f.f
  • Parametrize vfrec7, vfrsqrt7 and vfncvt.rod.f.f support
  • Add guards in the testbench to successfully compile in post-layout simulations
  • Add VCD dumping features to imatmul
  • core_id_i added to the interface of the system
  • Clock-gate the system bank macros when not used (VRF, D$, I$)
  • Spill register on sldu input signals to better isolate the unit
  • Clock-gate the unusued SIMD-int multipliers to save power

Changed

  • Fixed-Point instructions were added (VSADDU, VSADD, VAADD, VAADDU, VSSUBU, VSSUB, VASUB, VASUBU)
  • The main sequencer issues instructions every time the target unit has a non-full instruction queue
  • The main sequencer stalls if the instructions target a lane, and its operand requesters are not ready
  • New instructions enter the main sequencer with a token that marks them as new, and the related counter is updated upon arrival
  • Update README with instructions on how to compile convolutions
  • Refactor benchmark app
  • Double the testbench memory size
  • Update the python-requirements list
  • Remove the assign keyword from the always block of masku.sv
  • Update LLVM to version 15.0.0 (RVV 1.0)
  • Update Spike to version 1.1.1-dev (RVV 1.0)
  • Update newlib from commit 84d068 to 5192d5
  • Ara's dispatcher goes to WAIT_STATE only when the new LMUL is lower than the old one
  • Updated target -march to rv64gcv_zfh_zvfh0p1 to enable half-floats support
  • Halve CVA6's L1 caches to ease backend timing closure
  • Remove CVA6's cache patch from hardware/patches (CVA6 is now updated)
  • Increase addrgen queue depth to four, to better hide memory latency
  • The RESHUFFLE state is now iterative and reshuffles all the vector registers that need this operation
  • Performance metrics are now calculated by an external performance.py script, instead of during the program simulation, which just prints out the cycle count
  • 7x7 kernel 2dconvs now support arbitrary vector lengths
  • Default con vlen in the config files is now NR_LANES*1024
  • Optimize jacobi2d in ASM, +align store address
  • Replace apps/common/script/datagen.sh with new input data source-of-truth (apps/common/default_arguments.mk) during app compilation
  • benchmark.sh can now also benchmark just one app at a time via an input argument
  • Adapt fdotproduct to dotproduct structure
  • Pre-calculate next-cycle aligned_start_address in addrgen for timing reasons
  • Add is_reduct signal to the operand queues, to gate the neutral value filling
  • The hw-counter can be enabled or disabled via software to work also with an environment with cache warming
  • The dropout vectors are now explicitly memory aligned
  • matmul generates data offline like all the other benchmarks
  • Refactor benchmark.sh script
  • Update LLVM to 16. This brings faster compilations and a prototype of auto-vectorization (disabled by default here!)
  • All the data-gen scripts generate now random data
  • Patch riscv-tests crt0.S also before compiling riscv-tests
  • Almost align all the vector/matrix sizes in benchmark.sh
  • Generate data for fmatmul at compile time
  • SIMD multipliers are now power gated
  • Roll-back to Verilator v4.214
  • Parametrize addrgen queue depth
  • SIMD-multipliers are now gated singularly depending on VSEW
  • Optimize pathfinder kernel
  • Optimize dwt kernel
  • Optimize dotproduct kernel
  • Optimize fft kernel
  • Simplify the reduction engine for both valu and vmfpu, to avoid spurious valid signals to the sldu
  • Fix commit for dtc installation (spike dependency)
  • Simplify the datapath of the slide unit. The sldu supports only powers of two, and cannot slide and reshuffle at the same time. Non-power-of-two slides are now handled with micro operations.
  • Bump Verilator to v5.012

2.2.0 - 2021-11-02

Fixed

  • Fix typo on the build instructions of the README
  • Fix Gnuplot installation on GitHub's CI
  • The number of elements requested by the Store Unit and the Element Requester now depends both on the requested eew and the past eew of the vector of the used register
  • When the VRF is written and EMUL > 1, the eew of all the interested registers is updated
  • Memory operations can change EMUL when EEW != VSEW
  • The LSU now correctly handles bursts with a saturated length of 256 beats
  • AXI transactions on an opposite channel w.r.t. the channel currently in use are started only after the completion of the previous transactions
  • Fix the number of elements to be requested for a vslidedown instruction
  • Generate data.S files before compiling the programs
  • Clean intermediate app object files with make clean
  • Add a fence before stopping the cycle counter, to let the last vector store complete

Added

  • benchmarks app to benchmark Ara
  • CI task to create roofline plots of imatmul and fmatmul, available as artifacts
  • Vector floating-point compare instructions (vmfeq, vmfne, vmflt, vmfle, vmfgt, vmfge)
  • Vector single-width floating-point/integer type-convert instructions (vfcvt.xu.f, vfcvt.x.f, vfcvt.rtz.xu.f, vfcvt.rtz.x.f, vfcvt.f.xu, vfcvt.f.x)
  • Vector widening floating-point/integer type-convert instructions (vfwcvt.xu.f, vfwcvt.x.f, vfwcvt.rtz.xu.f, vfwcvt.rtz.x.f, vfwcvt.f.xu, vfwcvt.f.x, vfwcvt.f.f)
  • Vector narrowing floating-point/integer type-convert instructions (vfncvt.xu.f, vfncvt.x.f, vfncvt.rtz.xu.f, vfncvt.rtz.x.f, vfncvt.f.xu, vfncvt.f.x, vfncvt.f.f)
  • Vector whole-register move instruction vmv<nr>
  • Vector whole-register load/store vl1r, vs1r
  • Vector load/store mask vle1, vse1
  • Whole-register instructions are executed also if vtype.vl == 0
  • Makefile option (trace=1) to generate waveform traces when running simulations with Verilator
  • Add fconv3d kernel, optimized for 7x7 filters
  • Optimize fconv2d and iconv2d kernels for 3x3 filters
  • Add convolutions to the benchmark app, and print the related roofline plots
  • Add corner case test to vslidedown instruction

Changed

  • Add spill register at the lane edge, to cut the timing-critical interface between the Mask unit and the VFUs
  • Increase latency of the 16-bit multiplier from 0 to 1 to cut an in-lane timing-critical path
  • Widen CVA6's cache lines
  • Implement back-to-back accelerator instruction issue mechanism on CVA6
  • Use https protocol when cloning DTC from main Makefile
  • Use https protocol for newlib-cygwin in .gitmodules
  • Cut a timing-critical path from Addrgen to Sequencer (1 cycle more to start an AXI transaction)
  • Cut a timing-critical path in the VSTU, relative to the calculation of the pointer to the VRF word received from the lanes
  • Create ara_system wrapper containing Ara, Ariane, and an AXI mux, instantiated from within Ara's SoC
  • Retime address calculation of the addrgen
  • Push MASKU operand muxing from the lanes to the Mask Unit
  • Reduce CVA6's default cache size
  • Update Verilator to v4.214
  • Update bender to v0.23.1

2.1.0 - 2021-07-16

Fixed

  • Fix calculation of vstu's vector length
  • Fix vslideup and vslidedown operand's vector length trimming
  • Mute mask requests on idle lanes
  • Mute instructions with vector length zero on the respective lane_sequencer and operand_requester
  • Fix simd_div's offset calculation
  • Delay acknowledgment of memory requests if the axi_inval_filter is busy

Added

  • Format source files in the apps folder with clang-format by running make format
  • Support for the 2_lanes, 8_lanes, and 16_lanes configurations, besides the default 4_lanes one

Changed

  • Compile Verilator and Ara's verilated model with LLVM, for a faster compile time.
  • Verilator updated to version v4.210.
  • Verilation is done with a hierarchical verilation flow
  • Replace ara_soc's LLC with a simple main memory
  • Reduce number of words on the main memory, for faster Verilation
  • Update common_cells to v1.22.1
  • Update axi to v0.29.1

2.0.0 - 2021-06-24

Added

  • Script to align all the elf sections to the AXI Data Width (the testbench requires it)
  • RISC-V V intrinsics can now be compiled
  • Add support for vsetivli, vmv<nr>r.v instructions
  • Add support for strided memory operations
  • Add support for stores misaligned w.r.t. the AXI Data Width

Changed

  • Alignment with lowRISC's coding guidelines
  • Update Ara support for RISC-V V extension to V 0.10, with the exception of the instructions that were already missing
  • Replace toolchain from GCC to LLVM when compiling for RISC-V V extension
  • Update toolchain and SPIKE support to RISC-V V 0.10
  • Patches for GCC and SPIKE are no longer required
  • Ara benchmarks are now compatible with RISC-V V 0.10

Fixed

  • Fix vrf_seq_byte definition in the Load Unit
  • Fix check to discriminate a valid byte in the VRF word, in the Load Unit
  • Fix axi_addrgen_d.len calculation in the Address Generation Unit
  • Correctly check whether the generated address corresponds to the vector load or the store unit
  • Typos on the ChangeLog's dates
  • Remove unwanted latches in the addrgen, simd_div, instr_queue, and decoder
  • Fix vl == 0 memory operations bug. Ara correctly tells Ariane that the memory operation is over

1.2.0 - 2021-04-12

Added

  • Hardware support for:
    • Vector slide instructions (vslideup, vslide1up, vfslide1up, vslidedown, vslide1down, vfslide1down)
  • Software implementation of a integer 2D convolution kernel
  • CI job to check the conv2d execution on Ara

Fixed

  • Removed dependency to a specific gcc g++ version in Makefile
  • Arithmetic and memory vector instructions with vl == 0 are considered as a NOP
  • Increment bit width of the vector length type (vlen_t), accounting for vectors whose length is VLMAX
  • Fix vector length calculation for the MaskB operand, which depends on vsew
  • Fix typo on the vrf_pnt updating logic at the Mask Unit
  • Update README to highlight dependency with Spike
  • Update Bender's link dependency to the public CVA6 repository
  • Retrigger the compile module if the ModelSim compilation did not succeed

Changed

  • The encoding.h in the common Ara runtime is now a copy from the encoding.h in the Spike submodule

1.1.1 - 2021-03-25

Added

  • Parametrization for FPU and FPU-specific formats support, through the FPUSupport ara_soc parameter

1.1.0 - 2021-03-18

Added

  • GitHub Actions-based CI
  • Hardware support for:
    • Vector single-width floating-point fused multiply-add instructions (vfnmacc, vfmsac, vfnmsac, vfnmadd, vfmsub, vfnmsub)
    • Vector floating-point sign-injection instructions (vfsgnj, vfsgnjn, vfsgnjx)
    • Vector widening floating-point add/subtract instructions (vfwadd, vfwsub, vfwadd.w, vfwsub.w)
    • Vector widening floating-point multiply instructions (vfwmul)
    • Vector widening floating-point fused multiply-add instructions (vfwmacc, vfwnmacc, vfwmsac, vfwnmsac)
    • Vector floating-point merge instruction (vfmerge)
    • Vector floating-point move instruction (vfmv)

Changed

  • Contributing guidelines updated to include commit message and C++ code style guidelines

1.0.0 - 2021-03-10

Added

  • Hardware support for:
    • Vector single-width floating-point add/subtract instructions (vfadd, vfsub, vfrsub)
    • Vector single-width floating-point multiply instructions (vfmul)
    • Vector single-width floating-point fused multiply-add instructions (vfmacc, vfmadd)
    • Vector single-width floating-point min/max instructions (vfmin, vfmax)
  • Software implementation of a floating-point matrix multiplication kernel

0.6.0 - 2021-02-24

Added

  • Support for a coherent mode between Ara and Ariane
    • Snoop AW channel from Ara to L2
    • Invalidate Ariane's L1 cache sets accordingly
    • Coherent mode can be toggled together with consistent mode using the LSB of CSR 0x702

Changed

  • Ariane's data cache is active by default
  • The matrix multiplication kernel achieves better performance
    • It reports the performance and the utilization for several matrix sizes

0.5.0 - 2021-02-14

Added

  • Hardware support for:
    • Vector single-width integer divide instructions (vdivu, vdiv, vremu, vrem)
    • Vector integer comparison instructions (vmseq, vmsne, vmsltu, vmslt, vmsleu, vmsle, vmsgtu, vmsgt)
    • Vector carry-out of add-with-carry and subtract-with-borrow instructions (vmadc, vmsbc)
  • Runtime measurement functions
  • Consistent mode which orders scalar and vector loads/stores.
    • Conservative ordering without address comparison
    • Consistent mode is enabled per default, can be disabled by clearing the LSB of CSR 0x702.

Fixed

  • Ariane's accelerator dispatcher module was rewritten, fixing a bug where instructions would get skipped.
  • The Vector Store unit takes the EEW of the source vector register into account to shuffle the elements before writing them to memory.

Changed

  • Vector mask instructions (vmand, vmnand, vmandnot, vmxor, vmor, vmnor, vmornot, vmxnor) no longer require the non-compliant constraint that the vector length is divisible by eight.

0.4.0 - 2021-02-04

Added

  • Hardware compilation with Verilator
  • Software implementation of a matrix multiplication kernel

Changed

  • The riscv_tests_simc Makefile target was deprecated. The riscv-tests are now run with the Verilated design, which can be called through the riscv_tests_simv Makefile target.
  • The operand queues now take as a parameter the type conversions they support (currently, SupportIntExt2, SupportIntExt4, and SupportIntExt8)
  • The Vector Multiplier unit now has independant pipelines for each element width.

0.3.0 - 2021-01-28

Added

  • Hardware support for:
    • Vector single-width integer multiply instructions (vmul, vmulh, vmulhu, vmulhsu)
    • Vector single-width integer multiply-add instructions (vmacc, vnmsac, vmadd, vnmsub)
    • Vector integer add-with-carry/subtract-with-borrow instructions (vadc, vsbc)
    • Vector widening integer multiply instructions (vwmul, vwmulu, vwmulsu)
    • Vector widening integer multiply-add instructions (vwmaccu, vwmacc, vwmaccsu, vwmaccus)

Changed

  • Explicit scan chain signals added to the lane's and Ara's interfaces

Fixed

  • Miscellaneous fixes for compatibility with Synopsys DC
  • Send the correct bits of the address to the Vector Register File's banks
  • Correctly calculate the initial address of each vector register in the VRF

0.2.0 - 2021-01-22

Added

  • Hardware support for:
    • Bit-shift instructions (vsll, vsrl, vsra)
    • Vector widening integer add/subtract (vwadd, vwaddu, vwsub, vwsubu)
    • Vector integer extension (vzext, vsext)
    • Vector integer merge and move instructions (vmerge, vmv)
    • Vector narrowing integer right shift instructions (vnsrl, vnsra)

Changed

  • Bender updated to version 0.21.0

Fixed

  • CVA6's forwarding mechanism of operand B for accelerator instructions

0.1.0 - 2021-01-06

Added

  • Hardware support for:

    • Vector configuration instructions (vsetvl/vsetvli)
    • Unit-strided vector loads and vector stores
    • Basic arithmetic and logic instructions (vand, vor, vxor, vadd, vsub, vrsub, vmin-u, vmax-u)
    • Predicated instructions through a mask unit
    • Vector mask instructions (vmand, vmnand, vmandnot, vmor, vmnor, vmornot, vmxor, vmxnor)
    • Length multipliers

  • Implementation of a synthesizable Ara SoC top-level

  • Software support for RISC-V Vector code

  • Continuous integration tests through riscv-tests executed both with Spike and on Ara