Skip to content

THU-DSP-LAB/ventus-gpgpu-isa-simulator

Folders and files

NameName
Last commit message
Last commit date

Latest commit

 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 

Repository files navigation

Ventus RISC-V GPGPU ISA Simulator

About

This is Ventus RISC-V GPGPU isa simulator based on Spike.

We recommend using ventus-llvm to build the complete toolchain for Ventus GPGPU.

This repository can also be built independently. In this situation, you can use a custom driver in gpgpu-testcase/driver to simulate, see following "compile ventus GPGPU program with clang and run in spike".

This simulator is developed by 杨轲翔, 张泰然 and 郑名宸.

This simulator adds the following features into origin spike simulator:

  • support custom instruction for vector branch control and warp control
    • vector branch control: vbeq, vbne, vblt, vbge, vbltu, vbgeu
    • warp control: barrier, endprg
  • support custom csr: numw, numt, numwg, tid, wid, pds, lds, knl
  • enhenced interactive mode commands to facilitate debugging of Ventus GPGPU programs

For more information of the original Spike, please refer to README-spike-origin.md

installation and usage

build step

Build step is identical to the origin build step of Spike. Assume that the SPIKE_TARGET_DIR enironment variable is set to spike install path.

$ apt-get install device-tree-compiler
$ mkdir build
$ cd build
$ ../configure --prefix=$SPIKE_TARGET_DIR --enable-commitlog
$ make
$ [sudo] make install

(update) compile ventus GPGPU program with clang and run in spike

Now you can test spike-device with spike-driver in a baremetal mode.

About Using clang and llvm to compile kernel.cl into kernel.elf, see ventus-LLVM.

Then modify elf-filename, kernel argument and buffer data in gpgpu-testcase/driver/test.cpp .After that, run:

$ cd gpgpu-testcase/driver
$ export SPIKE_SRC_DIR=<path-to-spike-dir>
$ export SPIKE_TARGET_DIR=<path-to-spike-target-path>
$ mkdir build && cd build
$ cmake ..
$ make
$ ./spike_test

Spike running logfile elf-filename.log will be generated at current path.

if spike_main.so cannot be found, add <path-to-spike-target-path>/lib into your LD_LIBRARY_PATH

(deprecated) compiling and running Ventus GPGPU program in Spike

prerequisites

We execute Ventus GPGPU program with Spike in machine mode. To produce executable file, we utilize libgloss-htif and the modifed version of riscv-gnu-toolchain.

# install riscv64-unknown-elf toolchain
$ git clone https://github.com/THU-DSP-LAB/riscv-gnu-toolchain.git
$ cd riscv-gnu-toolchain
$
$ git submodule init riscv-binutils
$ git submodule update riscv-binutils
$ cd riscv-binutils
$ git checkout main
$ cd ..
$
$ mkdir build && cd build
$ ../configure --prefix=$RISCV --with-isa=rv64gv --with-abi=lp64d --with-cmodel=medany
$ make

# install libgloss-htif
$ cd ..
$ git clone https://github.com/ucb-bar/libgloss-htif.git
$ cd ligbloss-htif
$ mkdir build && cd build
$ ../configure --prefix=${RISCV}/riscv64-unknown-elf --host=riscv64-unknown-elf
$ make
$ make install

For futher reference, pleace refer to the corresponding instructions provided by libgloss-htif and riscv64-gnu-toolchain

compiling and running

Assuming assembly code test.s, we use following commands.

$ riscv64-unknown-elf-gcc -fno-common -fno-builtin-printf -specs=htif_nano.specs -Wl,--defsym=__main=main -c test.s 
$ riscv64-unknown-elf-gcc -static -specs=htif_nano.specs -Wl,--defsym=__main=main test.o -o test.riscv

And use Spike to execute test.riscv.

$ spike -d -p4 --isa=rv64gv --varch=vlen:256,elen:32 --gpgpuarch numw:4,numt:8,numwg:1 test.riscv

We add gpgpuarch option to configure uArch parameters for Ventus GPGPU (numw to set total warp number and numt to set thread number per warp). Note that there are some constraints about the parameters:

  1. numw * numwg must be equal to processor count (-p)
  2. numt must be equal to vlen divided by elen
  3. more option parameters could be added after numw, numt and numwg, here is an example:numw:4,numt:8,numwg:2,kernelx:1,kernely:2,kernelz:1,ldssize:0x80000000,ldsbase:0x00001000,pdssize:0x90000000,pdsbase:0x00001000,knlbase:0x90000000, you can offer workgroup dimensions with kernel_x/y/z, and ldssize(sharedmemory size) & pdssize(privatememory size) for each workgroup.

For more spike options, use spike -h

interactive mode command enhancement

GPGPU execution is different from conventional CPU with vector extension in two aspect:

  1. use simt stack to control divergence and loop
  2. multiwarp managment

We implement software models of simt stack and warp scheduling to support above features, and add enhanced interactive mode commands to display simt stack and warp scheduling information.

command usage function
simt-stack simt-stack <core> Display simt stack info given hartid
warp-barrier warp-barrier Display global warp barrier info(numw, numt and barrier bit vector)
mem-region mem-region <hex addr> <hex size> Show contents of physical memory given base memory and size

For more interactive mode commands, use help in interactive mode.

assembly programming introduction

In this section, we will introduce how to program programs with Ventus GPGPU extension that can run in Spike. Since the software stack is not mature, this section may be changed a lot in the future.

branch and loop assembly code

First, we introduce how to program branch and loop with custom instructions.

Psuedo code for branch is shown as below.

# if(condition(for example: vs1 == vs2))
#   ture_branch_code;
# else
#   false_branch_code;

BRANCH:
    vbne        vs1, vs2, FALSE_BRANCH
    true_branch_assembly_code
    join        v0, v0, BRANCH_JOIN
FALSE_BRANCH:
    false_branch_assenbly_code
    join        v0, v0, BRANCH_JOIN
BRANCH_JOIN:
    end_of_the_branch

Psuedo code for loop is shown as below.

    vblt    v0, v1, LOOP
    join    v0, v0, LOOP_END
LOOP:
    loop_body
    join    v0, v0, LOOP_COND_EVAL
LOOP_COND_EVAL:
    vblt    v0, v1, LOOP
    join    v0, v0, LOOP_END

extra codes to produce executable file

Since we use libgloss-htif to produce machine mode executable, extra codes are needed.

main:
    addi    sp,sp,-16
    sd      s0,8(sp)
    addi    s0,sp,16
    j       start_of_your_program

main_end:
    li      a5,0
    mv      a0,a5
    ld      s0,8(sp)
    addi    sp,sp,16
    ret

start_of_your_program:
    ...
    j       main_end

In future version, we may develop custom linker scripts to replace libgloss-htif and the extra codes will be unnecessary.

Now we develop scripts to add extra codes automatically. See here.

other conventions

Current version needs the assembly program to obey following conventions to be executed sucessfully in Spike. Some of them may be changed in future version.

  1. In Ventus GPGPU, vl is set in hardware. But in this simulator, extra vsetvli instrucions are needed to configure vl and other related parameters.
  2. We use .data directive to store data. Unlike in Ventus GPGPU hardware, we don't use gds/lds to get the global/local memory address. Instead we use la instruction to load start address of global/local memory into the register.

testcases

We provided testcased in path gpgpu-testcase. To run the testcase, change directory to gpgpu-testcase and execute the makefile script.

  1. compile the testcase $ make TEST=[testcasename]

  2. invoke spike to run the testcase $ make spike-sim TEST=[testcasename]

PS:

  • make spike-sim will alse compile the asssembly
  • testcasename is the name of testcase, e.g. loop
  • spike-sim have other parameters that can be configured via commandline
    • P: processor count, defalut value is 4
    • VARCH: rvv uArch parameter, default value is vlen:256,elen:32
    • DEBUG: enable interactive mode, default value is -d
    • ISA:supported isa, default value is rv64gv

We provide the following testcases

name function
branch-basic test for simple branch
branch-nested test for nested branch
loop test for loop
barrier test for inter warp barrier
gaussian gaussian testcase
saxpy saxpy testcase implemented with extended gpgpu instructions
saxpy2 saxpy testcase implemented with conventional rvv instructions