Hi, all,
I am a new CNN learning and trying run cuda-convnet on gpgpu-sim_distribution simulator.
Cuda-convnet can run well on my server, but when I run cuda-convnet with the gpgpusim simulator, an error happenes. The error is "ImportError: /usr/local/cuda/lib64/libcublas.so.4: undefined symbol: cudaMemsetAsync".
The details are like this:

Importing _ConvNet C++ module
Traceback (most recent call last):
File "convnet.py", line 203, in
model = ConvNet(op, load_dic)
File "convnet.py", line 43, in init
IGPUModel.init(self, "ConvNet", op, load_dic, filename_options, dp_params=dp_params)
File "/root/convnet/trunk/gpumodel.py", line 88, in init
self.import_model()
File "convnet.py", line 49, in import_model
self.libmodel = import(lib_name)
ImportError: /usr/local/cuda/lib64/libcublas.so.4: undefined symbol: cudaMemsetAsync

The wired thing is Cuda-convnet can run well on my server. After I change my $PATH and $LD_LIBRARY_PATH to the original value, the error disappears.
I setup my gpgpusim simulator following the tutorial and copy the config files into the convnet's work path.

I build the gpgpusim with gcc-5.4.1. The GPU config chose GTX480.
My cuda version is 4.0.
Here is the output of my libcublas.so
nm -D /usr/local/cuda/lib64/libcublas.so |grep ‘cuda’

         U __cudaRegisterFatBinary
         U __cudaRegisterFunction
         U __cudaRegisterTexture
         U __cudaUnregisterFatBinary
         U cudaBindTexture
         U cudaConfigureCall
         U cudaCreateChannelDesc
         U cudaEventCreateWithFlags
         U cudaEventDestroy
         U cudaEventQuery
         U cudaEventRecord
         U cudaEventSynchronize
         U cudaFree
         U cudaFuncGetAttributes
         U cudaGetDevice
         U cudaGetDeviceProperties
         U cudaGetExportTable
         U cudaGetLastError
         U cudaLaunch
         U cudaMalloc
         U cudaMemcpy
         U cudaMemcpy2D
         U cudaMemcpy2DAsync
         U cudaMemcpyAsync
         U cudaMemsetAsync
         U cudaSetupArgument
         U cudaThreadSynchronize
         U cudaUnbindTexture

ll /usr/local/cuda/lib64/libcublas.so

lrwxrwxrwx 1 root root 14 Mar 17 16:10 /usr/local/cuda/lib64/libcublas.so -> libcublas.so.4*

Anyone know the reason for this error and how to fix it? Could anyone help me out with this issue? Thanks a lot in advance!

For the master branch, it is possible to compile the ispass workloads. However, with the dev branch, there is no common.mk and hence the workloads cannot be built. Is there a new set for the dev branch?

Issue: Syntax and Parse errors in the libcudnn ptx files.

Set up:
GPGPUSIM version 4.0.0
GCC: 4.9.4
CUDA: 9.1
CUDNN: 7.1
PYTORCH_BIN: /usr/local/cuda-9.1/lib64/libcudnn.so

Program: saxpy.cu link: https://devblogs.nvidia.com/easy-introduction-cuda-c-and-c/

Description: Standalone GPGPUSIM works fine with the above set-up and the example program run successfully.

After installing the PyTorch_GPGPUSIM and setting the PYTORCH_BIN = /usr/.../CUDA-9.1/.../libcudnn.so, the above mentioned programs fail to execute. The simulator extracts a large number of PTX files from libcudnn and the simulator generates syntax and parse errors.

Simulation log file attached here.
out_pytorch_sm70.txt

Thank you.

I have recently downloaded gpgpu-sim_distribution. I am using ubuntu 14.04 (64bit), gcc 4.4.7 and cudatoolkit 4.0, bison 2.4.1 and flex 2.5.35 with all other dependencies as mentioned in the README docs. But when I am building gpgpu-sim , I am getting the follwing error:

ptx_parser.cc:30: fatal error: ptx.tab.h: No such file or directory
compilation terminated.
make[1]: *** [/home/ssb/gpgpu-sim_distribution/build/gcc-4.4.7/cuda-4000/release/cuda-sim/ptx_parser.o] Error 1
make[1]: Leaving directory `/home/ssb/gpgpu-sim_distribution/src/cuda-sim'
make: *** [cuda-sim] Error 2

Earlier with the same versions of s/w, I was not getting any error and able to run gpgpusim and the ispass2009 benchmarks.
Kindly help what I have to do to get the successful build.

Thanks

For a cuda written code, I get this error from the simulator

ptx_ir.h:116: bool type_info_key_compare::operator()(const type_info_key&, const type_info_key&) const: Assertion `a.m_init && b.m_init' failed.

The call stack shows that in the following function

type_info *symbol_table::get_array_type( type_info *base_type, unsigned array_dim ) 
{
   type_info_key t = base_type->get_key();
   t.set_array_dim(array_dim);
   type_info *pt;
   pt = m_types[t] = new type_info(this,t);       // HERE
   return pt;
}

The following assertion fails

struct type_info_key_compare {
   bool operator()( const type_info_key &a, const type_info_key &b ) const
   {
      assert( a.m_init && b.m_init );
      if ( a.m_space_spec < b.m_space_spec ) return true;
      if ( a.m_scalar_type_spec < b.m_scalar_type_spec ) return true;
      if ( a.m_vector_spec < b.m_vector_spec ) return true;
      if ( a.m_alignment_spec < b.m_alignment_spec ) return true;
      if ( a.m_extern_spec < b.m_extern_spec ) return true;
      if ( a.m_array_dim < b.m_array_dim ) return true;
      if ( a.m_is_function < b.m_is_function ) return true;

      return false;
   }
};

I can not track what happens here although I attached GDB. Can someone shed a light? What is the purpose of that assertion?

It seems that the ispass benchmarks are not compatible with all boost version. Also, it has not beed stated which boost version is used by the developers. May I know that?

I was trying to run the gpgpu-sim_simulation.

When I was trying to run the benmark ibackprop, the following error pops:

terminate called after throwing an instance of 'std::ios_base::failure[abi:cxx11]'
  what():  basic_ios::clear: iostream error
Aborted (core dumped)

The command I use is following:

/root/gpgpusim-dev/gpgpu-sim_distribution/gpgpu-sim_simulations/util/job_launching/../../benchmarks/bin/4.2/release/backprop-rodinia-2.0-ft 4096 ./data/result-4096.txt

The following are my work environment:

echo $GPGPUSIM_CONFIG

> gcc-4.5.3/cuda-4020/release

echo $GPGPUSIM_ROOT

> /root/gpgpusim-dev/gpgpu-sim_distribution

My cuda version is cudnn 4.2.

nvcc -V
nvcc: NVIDIA (R) Cuda compiler driver
Copyright (c) 2005-2012 NVIDIA Corporation
Built on Thu_Apr__5_00:24:31_PDT_2012
Cuda compilation tools, release 4.2, V0.2.1221

Is the error due to my cuda version? Is there anyone know how to fix this issue?

When i run GPGPU-SIM based on CUDA it operates well. All things are operate well.
But when i run Pytorch-GPGPU-SIM it is not operate well.

I don't know exactly what is wrong

Here is my setting and my code . i use docker

Ubuntu : 16.04
gcc : 4.8.4
cuda : 8.0
cudnn : 6.0

Finally i get this error and stop.
My pytorch code is very simple for test.

import torch
from torch.autograd import Variable

a= torch.ones(2,2)
b= torch.ones(2,2)
print(a)

a=Variable(a, requires_grad=True).cuda()
b=Variable(a, requires_grad=True).cuda()
b= a+2
print(b)

It operates well without gpgpu-sim but in gpgpu-sim when b=a+2 the error is occur.

In addition to my code, in /pytorch-gpgpu-sim/test/ many sample codes are all generate error like this. What is problem? please help me...

Tagging @mkhairy as he seems to be working mostly on the dev branch.

tesnor -> tensor ?

I was trying to run the gpgpu-sim_simulation.

When I was trying to run any benmarks, They all appera the same mistake

virtual void shader_core_config::init(): Assertion `toks' failed.
12260 Aborted (core dumped)

I'm trying to run some simulations on gpgpu-sim. I'm using the dev branch that supports some of the newer CUDA versions. The build succeeds but when I try to run a program, the CUDA calls go to the system installation of CUDA instead of GPGPU-Sim. No errors or warnings are thrown. I have copied config files to the folder where I run the program from. No modifications were made to the config files from the tested_configs folder.

I'm using nvidia-docker (nvidia/cuda:9.1-devel) and have installed all the dependencies. Environment setup and build are successful.

My $PATH is /root/gpgpu-sim_distribution/bin/:usr/local/sbin:/usr/local/bin:/usr/sbin:/usr/bin:/sbin:/bin
My $CUDA_INSTALL_PATH is /usr/local/cuda

Any help would be highly appreciated

I'm trying to run tensor core applications with dev branch.
When I try to use certain cuda api such as, cudaFuncSetAttributes and cudaMallocManaged, it says it's not in library.

Do you have more recent cuda library or should I have to find another way to use tensor core without those higher version CUDA libraries?

How to solve this error?
dev branch, above sm_10,

Error: No PTX sections found with sm capability that is lower than current forced maximum capability
minimum ptx capability found = 0, maximum forced ptx capability = 20
User might want to change either the forced maximum capability from gpgpusim configuration or update the compilation to generate the required PTX version

Hi,

I am trying to simulate a behavior in which I need to disable the performance for only one thread in a wrap. Is it possible to do?

hi everyone
in gpu-cache.h in class mshr_table there is this code:

bool m_current_response_ready;
std::list<new_addr_type> m_current_response;

i dont see the m_current_response_ready in anywhere of the source code!

is there anything wrong ?

There is no relevant information in the Debugging Errors in Performance Simulation section.
I am relatively new to GPGPU sim and would like to know how should determine whether my application is faulty or the config file is?

In the Dev branch it is stated that after building gpgpusim, it is possible to run cuda based applications by chnaging the cuda library paths to gpgusim library folder. Also, the GPU config file should be in the working directory.

Well, I did that but the application says no cuda device found! I also ran deviceQuery and it also said there is no cuda.

mahmood@u1604:~/gpgpu-sim_distribution$ echo $LD_LIBRARY_PATH

mahmood@u1604:~/gpgpu-sim_distribution$ source setup_environment release
GPGPU-Sim version 3.2.2 (build ) WARNING ** GPGPU-Sim version 3.2.2 not fully tested with CUDA version 7.5 (please see README)
configured with GPUWattch.
setup_environment succeeded
mahmood@u1604:~/gpgpu-sim_distribution$ echo $LD_LIBRARY_PATH
/home/mahmood/gpgpu-sim_distribution/lib/gcc-4.8.5/cuda-7050/release:
mahmood@u1604:~/gpgpu-sim_distribution$ ls /home/mahmood/gpgpu-sim_distribution/lib/gcc-4.8.5/cuda-7050/release
libcudart.so    libcudart.so.4    libcudart.so.6.0  libcudart.so.8.0
libcudart.so.2  libcudart.so.5.0  libcudart.so.6.5
libcudart.so.3  libcudart.so.5.5  libcudart.so.7.5
mahmood@u1604:~/gunrock/build/bin$ ls
bc   chesapeake.mtx            hits   shared_lib_bc   shared_lib_example  sssp
bfs  config_quadro_islip.icnt  pr     shared_lib_bfs  shared_lib_pr       topk
cc   gpgpusim.config           salsa  shared_lib_cc   shared_lib_sssp     wtf
mahmood@u1604:~/gunrock/build/bin$ ./bfs market chesapeake.mtx --src=0 --unidirected
Loading Matrix-market coordinate-formatted graph ...
Reading from chesapeake.mtx:
  Parsing MARKET COO format (39 nodes, 340 directed edges)... Done parsing (0s).
  Converting 39 vertices, 340 directed edges (unordered tuples) to CSR format...
Done converting (0s).

Degree Histogram (39 vertices, 340 edges):
    Degree   0: 0 (0.00%)
    Degree 2^0: 0 (0.00%)
    Degree 2^1: 1 (2.56%)
    Degree 2^2: 22 (56.41%)
    Degree 2^3: 13 (33.33%)
    Degree 2^4: 2 (5.13%)
    Degree 2^5: 1 (2.56%)

  Converting 39 vertices, 340 directed edges (unordered tuples) to CSR format...
Done converting (0s).
Source vertex: 0
Using 1 GPU: [ 0 ].
[/home/mahmood/gunrock/gunrock/util/info.cuh, 470 @ gpu 32764] cudaGetDevice failed (CUDA error 38: no CUDA-capable device is detected)
[/home/mahmood/gunrock/gunrock/util/test_utils.cu, 61 @ gpu 0] cudaSetDevice failed. (CUDA error 38: no CUDA-capable device is detected)
[/home/mahmood/gunrock/gunrock/util/info.cuh, 484 @ gpu 0] cudaStreamCreate failed. (CUDA error 38: no CUDA-capable device is detected)
CODE REQUESTED INVALID CUDA DEVICE -2050039464
mahmood@u1604:~/NVIDIA_CUDA-7.5_Samples/1_Utilities/deviceQuery$ make
"/usr/local/cuda-7.5"/bin/nvcc -ccbin g++ -I../../common/inc  -m64    -gencode arch=compute_20,code=sm_20 -gencode arch=compute_30,code=sm_30 -gencode arch=compute_35,code=sm_35 -gencode arch=compute_37,code=sm_37 -gencode arch=compute_50,code=sm_50 -gencode arch=compute_52,code=sm_52 -gencode arch=compute_52,code=compute_52 -o deviceQuery.o -c deviceQuery.cpp
"/usr/local/cuda-7.5"/bin/nvcc -ccbin g++   -m64      -gencode arch=compute_20,code=sm_20 -gencode arch=compute_30,code=sm_30 -gencode arch=compute_35,code=sm_35 -gencode arch=compute_37,code=sm_37 -gencode arch=compute_50,code=sm_50 -gencode arch=compute_52,code=sm_52 -gencode arch=compute_52,code=compute_52 -o deviceQuery deviceQuery.o 
mkdir -p ../../bin/x86_64/linux/release
cp deviceQuery ../../bin/x86_64/linux/release
mahmood@u1604:~/NVIDIA_CUDA-7.5_Samples/1_Utilities/deviceQuery$ ../../bin/x86_64/linux/release/deviceQuery 
../../bin/x86_64/linux/release/deviceQuery Starting...

 CUDA Device Query (Runtime API) version (CUDART static linking)

cudaGetDeviceCount returned 38
-> no CUDA-capable device is detected
Result = FAIL

So, what should I do now?

We have a couple of problems running with cuda 7.5 according to an email discussion I had with a GPGPU-Sim user.
He was able to get his benchmark(s) running by doing some modifications which are available on his/her fork https://github.com/andpic/gpgpu-sim_distribution

Reported by GPGPU-Sim user:
OpenCL 1.1 C++ API compatibility. I added to gpgpu-sim's source code a couple of API functions that are necessary to be able to compile code with the C++ interface. In particular, the "clRetain" functions and "clCreateKernelsInProgram".
PTX parsing. The problem was due to a constant argument passed to the kernel. In other words, gpgpu doesn't like when kernel arguments are not passed with buffers. I fixed the problem by just changing that .const to blank space during the parsing phase.
ptxas output parsing. Ptxas outputs an extra line with the gmem keyword, which is not recognised by gpgpu-sim's parser. I added a bash trick to exclude that line when the file is read.

The simulator compiles and runs well with the latest CUDA 7.5.

Hi, i am using the given VM, but when i am going to make the SDK, an error occurs "/usr/bin/ld: cannot find -lrendercheckgl_x86_64-lGL". Do you know how to fix it? Thank you

Hello, I am trying to simulate pytorch-gpgpu-sim with gpgpu-sim.dev.

on README of pytorch_gpgpusim, it says "export PYTORCH_BIN=/path/to/libcudnn.so", which I think it is doing dynamic linking with cudnn.
Is it true that pytorch use that "PYTORCH_BIN" path?
Is it right to dynamic link the cudnn when using PyTorch?

Thank you.

Hi,

I am planning to start using gpgpu-sim for my GPU architecture related research. I stumbled upon this branch yesterday and saw that there are quite a lot of updates to the previous gpgpu-sim that is on the master branch. Is there a plan to do a full documented release of this new version ? I am just curious as to what are the new features that will be supported here. Thank you.

Used Environments

• Ubuntu 16.04.5
• CUDA 8.0
• cuDNN 7.1.4
• gcc 5.4.0
• g++ 5.4.0
• python 2.7.12

Problem Situation

Hello, I am trying to run pytorch-gpgpu-sim with gpgpu-sim_distribution dev branch.

I've built gpgpu-sim_distribution and pytorch-gpgpu-sim successfully.

However, I could not run my PyTorch model.

When I'm trying to run it, the PyTorch is terminated with following message.

THCudaCheck FAIL file=/home/redcarrottt/pytorch-gpgpu-sim/aten/src/THC/THCGeneral.cpp line=74 error=30 : unknown error
Traceback (most recent call last):
  File "./infer.py", line 130, in <module>
    infer(net, args)
  File "./infer.py", line 47, in infer
    net_gpu = net.cuda()
  File "/home/redcarrottt/.local/lib/python2.7/site-packages/torch/nn/modules/module.py", line 258, in cuda
    return self._apply(lambda t: t.cuda(device))
  File "/home/redcarrottt/.local/lib/python2.7/site-packages/torch/nn/modules/module.py", line 185, in _apply
    module._apply(fn)
  File "/home/redcarrottt/.local/lib/python2.7/site-packages/torch/nn/modules/module.py", line 191, in _apply
    param.data = fn(param.data)
  File "/home/redcarrottt/.local/lib/python2.7/site-packages/torch/nn/modules/module.py", line 258, in <lambda>
    return self._apply(lambda t: t.cuda(device))
RuntimeError: cuda runtime error (30) : unknown error at /home/redcarrottt/pytorch-gpgpu-sim/aten/src/THC/THCGeneral.cpp:74

As I know, the CUDA function which makes the error is cudaGetDeviceCount().
(called at aten/src/THC/THCGeneral.cpp:74)

I guess that my PyTorch's linking configuration is wrong.

Anyone knows what is the exact reason?
Thank you.

When we launched a kernel with like gridDim(0,0,1), which is exactly an empty kernel, the kernel will launch successfully, and the gpu->active() will always be true because the variable "pending" is true, and stream->op will have nothing to do because there are no more CTAs left in that kernel to be executed.
The loop {}while(gpu->active) will never stop: src/gpgpusim_entrypoint.cc:102

I tried to run pytorch with gpgpusim-dev, but i found that gpgpu-sim are not able to register cuda kernels in pytorch(like kernelPointwiseApply1 in aten/src/THC/THCApply.cuh).
Does anyone know that it is intended result of author or it is just my fault!

Thank you

Used env:
Ubuntu 16.04.6
gcc&g++ 5.4.0
CUDA 8.0
cuDNN 7.1.4

while sourcing setup_environment was successful, I get an error that can not find uda-sim/Makefile.makedepend

`mahmood@vb:~/gpgpu-sim_distribution$ make

    Building GPGPU-Sim version 3.2.2 (build ) with CUDA version 4.2

if [ ! -d lib/gcc-4.6.3/cuda-4020/release ]; then mkdir -p lib/gcc-4.6.3/cuda-4020/release; fi;
if [ ! -d /home/mahmood/gpgpu-sim_distribution/build/gcc-4.6.3/cuda-4020/release/libcuda ]; then mkdir -p /home/mahmood/gpgpu-sim_distribution/build/gcc-4.6.3/cuda-4020/release/libcuda; fi;
if [ ! -d /home/mahmood/gpgpu-sim_distribution/build/gcc-4.6.3/cuda-4020/release/cuda-sim ]; then mkdir -p /home/mahmood/gpgpu-sim_distribution/build/gcc-4.6.3/cuda-4020/release/cuda-sim; fi;
if [ ! -d /home/mahmood/gpgpu-sim_distribution/build/gcc-4.6.3/cuda-4020/release/cuda-sim/decuda_pred_table ]; then mkdir -p /home/mahmood/gpgpu-sim_distribution/build/gcc-4.6.3/cuda-4020/release/cuda-sim/decuda_pred_table; fi;
if [ ! -d /home/mahmood/gpgpu-sim_distribution/build/gcc-4.6.3/cuda-4020/release/gpgpu-sim ]; then mkdir -p /home/mahmood/gpgpu-sim_distribution/build/gcc-4.6.3/cuda-4020/release/gpgpu-sim; fi;
if [ ! -d /home/mahmood/gpgpu-sim_distribution/build/gcc-4.6.3/cuda-4020/release/libopencl ]; then mkdir -p /home/mahmood/gpgpu-sim_distribution/build/gcc-4.6.3/cuda-4020/release/libopencl; fi;
if [ ! -d /home/mahmood/gpgpu-sim_distribution/build/gcc-4.6.3/cuda-4020/release/libopencl/bin ]; then mkdir -p /home/mahmood/gpgpu-sim_distribution/build/gcc-4.6.3/cuda-4020/release/libopencl/bin; fi;
if [ ! -d /home/mahmood/gpgpu-sim_distribution/build/gcc-4.6.3/cuda-4020/release/intersim2 ]; then mkdir -p /home/mahmood/gpgpu-sim_distribution/build/gcc-4.6.3/cuda-4020/release/intersim2; fi;
if [ ! -d /home/mahmood/gpgpu-sim_distribution/build/gcc-4.6.3/cuda-4020/release/cuobjdump_to_ptxplus ]; then mkdir -p /home/mahmood/gpgpu-sim_distribution/build/gcc-4.6.3/cuda-4020/release/cuobjdump_to_ptxplus; fi;
if [ ! -d /home/mahmood/gpgpu-sim_distribution/build/gcc-4.6.3/cuda-4020/release/gpuwattch ]; then mkdir -p /home/mahmood/gpgpu-sim_distribution/build/gcc-4.6.3/cuda-4020/release/gpuwattch; fi;
if [ ! -d /home/mahmood/gpgpu-sim_distribution/build/gcc-4.6.3/cuda-4020/release/gpuwattch/cacti ]; then mkdir -p /home/mahmood/gpgpu-sim_distribution/build/gcc-4.6.3/cuda-4020/release/gpuwattch/cacti; fi;
make -C ./src/cuda-sim/ depend
make[1]: Entering directory /home/mahmood/gpgpu-sim_distribution/src/cuda-sim' Makefile:149: /home/mahmood/gpgpu-sim_distribution/build/gcc-4.6.3/cuda-4020/release/cuda-sim/Makefile.makedepend: No such file or directory touch /home/mahmood/gpgpu-sim_distribution/build/gcc-4.6.3/cuda-4020/release/cuda-sim/Makefile.makedepend makedepend -f/home/mahmood/gpgpu-sim_distribution/build/gcc-4.6.3/cuda-4020/release/cuda-sim/Makefile.makedepend -p/home/mahmood/gpgpu-sim_distribution/build/gcc-4.6.3/cuda-4020/release/cuda-sim/ cuda_device_printf.cc cuda-sim.cc instructions.cc memory.cc ptx_ir.cc ptx_loader.cc ptx_parser.cc ptx_sim.cc ptx-stats.cc 2> /dev/null make[1]: *** [depend] Error 127 make[1]: Leaving directory /home/mahmood/gpgpu-sim_distribution/src/cuda-sim'
make: *** [cuda-sim] Error 2
`

Since they have moved to here:

https://github.com/booksim/booksim2

I have not found a good way to integrate this booksim2 repository with our repository. I have modifications to their booksim2 source code. Git submodule may not work under this case.

I have successfully built dev branch on the following system

    ubuntu 16.04
    CUDA 8 GA1
    gcc 5.4

The following packages are needed

For CUDA:

    sudo apt-get install nvidia-384-dev libglu1-mesa-dev libxi-dev libxmu-dev
    bash cuda_8.0.44_linux.run

I didn't allow the run file to install the driver since I did that via apt-get.

For GPGPUSIM:

sudo apt-get install flex bison xutils-dev zlib1g-dev

Changes in the code:
1- Replace isnan() with std::isnan() in instructions.cc and cuda-math.h
2- Replace std::cout << "Failed to execute: " << cmd << std::endl; with std::cout << "Failed to execute: " << cmd.str() << std::endl; in cuda_runtime_api.cc
3- Replace std::cout << "Using command: " << cmd << std::endl; with std::cout << "Using command: " << cmd.str() << std::endl; in cuda_runtime_api.cc
4- Replace std::cout << "Trying to parse " << libcodfn << std::endl; with std::cout << "Trying to parse " << libcodfn.str() << std::endl; in cuda_runtime_api.cc

P.S: I wasn't able to create a new pull request.

I built the gpgpu-sim 4.0.0 version and used it with CUDA tookkit 9.1 and CUDA SDK.

I successfully executed matrix multiplication in CUDA SDK.
However, I ran into error with CUDA compute capability when I tried to run "cudaTensorCoreGemm" in CUDA SDK sample application.

It all starts to execute nicely at first, but after loading all ptx configuration releated lines, error comes out like this.

I copied configuration files from SM7-Titan V, which should have compute capability 7.0 and it is shown in very first output of execution.

Do I have to change this compute capability number some how by looking inside gpgpu-sim simulator code? Or should I have to find some other way to link this 7.0 capability written in configuration file to simulator?

Thank you for any advice

Hello，I want to make some changes to the LRU algorithm,but I can't find where is it in gpgpu-sim.Could you help me ,it help me a lot .Thanks a lot!

I'm trying to run a very large scale application which consumes over 4GB GPU memory on GPGPU-sim dev branch with GTX1080Ti config. (CUDA7.5, gcc-4.4.7, g++-4.4.7, Ubuntu14.04)

However, it doesn't work since GPU memory corruption (maybe overflow). So, I inspected source codes of GPGPU-sim and found "address_type" variable type at src/abstract_hardware_model.h.

src/abstract_hardware_model.h: line 70, typedef unsigned address_type;

I thought it may cause the memory overflow, I modified "unsigned" to "unsigned long long" for 64bit addressing, but it causes new problems with my GPU kernel. (My CUDA application works fine with native GPU GTX1080Ti)

Is it correct current version of GPGPU-sim has <4GB limitation? If so, how can I overcome it?

Hi ,
I export the environment variables (export GPGPUSIM_DEBUG=1).
When I start to executing the NVIDIA Benchmark , there are some commands for users to use .
But I don't have any idea about the commands ?
Here are the commands and my problem :

"""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""
1 . b : - set breakpoint
I don't know what's the file mean (is it a benckmark ?) and how to calculate the thread uid ?
2 . w - set watchpoint
The global address range ?
3 . del - delete breakpoint
What is the n mean ?
4 . dp - display pipeline contents on SM
There are shader cores in the CTAs will be initial state while starting executing the program.
If I set the 1 to n , it means shader core 1 in every CTA(maybe 0~5) or shader core 1 in CTA 0 ?

Regards.,

With the following specs
ubuntu 16.04
cuda 7.5
gcc/g++ 4.8

there are some warnings in ptx_parser.cc such as this

ptx_parser.cc:438:41: warning: format ‘%x’ expects argument of type ‘unsigned int’, but argument 2 has type ‘new_addr_type {aka long long unsigned int}’ [-Wformat=]
               addr+addr_pad + num_bits/8);

That line of code is

printf("from 0x%x to 0x%lx (global memory space) %u\n",   
                    addr+addr_pad,    
                    addr+addr_pad + num_bits/8,
                    g_const_alloc++);

The data types are

   size_t num_bits;
   unsigned addr_pad;
   new_addr_type addr;

and the last one is actually unsigned long long. So, I think the correct format is %llx for all of the statements that use addr.

gpgpu sim 3.x version can disable L2 cache.

but, the latest version can't do it.

Is there anyone who knows this problem

Hi.

i've an error very strange that i do not know how to resolve.

The problem is in the end of the compilation when the linker ld is executed. This is the problem:

/usr/bin/ld: no se puede encontrar -lGL //(here is saying that it is not possible to find -lGL option)
collect2: error: ld returned 1 exit status
Makefile:143: fallo en las instrucciones para el objetivo 'lib/gcc-5.4.0/cuda-4000/release/libcudart.so

I am using ubuntu 16.04 and the version of ld came with the operative system. By other hand i've been searching for this option and i am not able to find nothing.

Somebody can help me?

Thanks you very much.

Wrong question. Please delete it.

Using dev branch commit 551da3e, CUDA 8.0, and gcc-5.4.0

I haven't checked other commits or CUDA versions.

Compiling the following kernel (full code here)

__global__ void kernel(float *a, int offset)
{
  int i = offset + threadIdx.x + blockIdx.x*blockDim.x;
  float x = (float)i;
  float s = sinf(x);  // Generates problematic instructions
  float c = cosf(x); // Generates problematic instructions
  a[i] = a[i] + sqrtf(s*s+c*c); // Does not cause a problem
}

Will generate code causing a ptx runtime error:

./ptx_ir.h:1039: const operand_info& ptx_instruction::operand_lookup(unsigned int) const: Assertion `n < m_operands.size()' failed.

This is caused by the madc line, below. gpgpu-sim expects madc to have a fifth operand specifying the carry.

...
BB0_4:
.pragma "nounroll";
ld.const.u32 %r82, [%rd28];
	{
mad.lo.cc.u32 %r80, %r82, %r3, %r175;
madc.hi.u32 %r175, %r82, %r3, 0;
}
...

The program runs to completion if Taylor expansion versions of sinf and cosf are used instead of the cuda intrinsics.

While testing the dev branch with
ubuntu 16.04
gcc/g++ 4.8
cuda 7.5

In case of error: ‘isnan’ was not declared in this scope, just use std::isnan or add using namespace std; at the top of the problematic file

I’ve tried to run cudnn_samples_v7 with GPGPU-Sim, but its cuDNN kernels does not run on GPGPU-Sim.
It makes following message when "g_debug_execution = 3".

GPGPU-Sim PTX: CUDA API function "cudaError_t cudaMemcpy(void*, const void*, size_t, cudaMemcpyKind)" has been called.
GPGPU-Sim PTX: cudaMemcpy(): devPtr = 0xc01a5300
GPGPU-Sim API: Stream Manager State
GPGPU-Sim API:    stream 0 has 1 operations
GPGPU-Sim API:       0 :  stream operation memcpy host-to-device
GPGPU-Sim: ** START simulation thread (detected work) **
GPGPU-Sim API: Stream Manager State
GPGPU-Sim API:    stream 0 has 1 operations
GPGPU-Sim API:       0 :  stream operation memcpy host-to-device
GPGPU-Sim API: stream 0 performing memcpy host-to-device
GPGPU-Sim PTX: copying 3136 bytes from CPU[0x7fffdc43aa00] to GPU[0xc01a5300] ...  done.
GPGPU-Sim: ** STOP simulation thread (no work) **
GPGPU-Sim: *** simulation thread starting and spinning waiting for work ***
Testing cudnnGetConvolutionForwardAlgorithm ...
Fastest algorithm is Algo 1
Testing cudnnFindConvolutionForwardAlgorithm ...
^^^^ CUDNN_STATUS_SUCCESS for Algo 0: 0.020256 time requiring 0 memory
^^^^ CUDNN_STATUS_SUCCESS for Algo 1: 0.029696 time requiring 3464 memory
^^^^ CUDNN_STATUS_SUCCESS for Algo 2: 0.037888 time requiring 57600 memory
^^^^ CUDNN_STATUS_SUCCESS for Algo 4: 0.070240 time requiring 207360 memory
^^^^ CUDNN_STATUS_SUCCESS for Algo 7: 0.072352 time requiring 2057744 memory

GPGPU-Sim PTX: CUDA API function "cudaError_t cudaMalloc(void**, size_t)" has been called.
GPGPU-Sim PTX: allocating 46080 bytes on GPU starting at address 0xc01a6000
GPGPU-Sim PTX: cudaMallocing 46080 bytes starting at 0xc01a6000..

cudaLaunchKernel function should be called after “Testing cudnnFindConvolutionForwardAlgorithm”, but it has never been called.

On the other hand, if I try to run CUDA sample(such as vectorAdd), it does work well.

I guess that my cuDNN library does not execute cudaLaunchKernel function in GPGPU-Sim ‘libcudart.so'.
It seems to call cudaLaunchKernel in original 'libcudart.so'.

I used g++ to compile an OpenCL project. Then execute the code with "run_gpgpu-sim.sh".. However, it only outputed:

GPGPU-Sim finished running "/home/gpgpu-sim/a.out"
Used rundir=/home/gpgpu-sim/GTX480_rundir

How can I get the detailed report as running a C programme?
I am using the fully setup virtual machine provided by you.

In the GTX480 configs folder, the dram_latency is defined as 100 core cycles. Can anyone help me understand how did they arrive at this number?

Thank you.

I found some of PTX ISA are not implemented while I was parsing pytorch libraries.
(found in 151th PTX file in libcaffe2_gpu.so)

After examining ptx.y and ptx.l files, following statements are seemed to not be implemented.
.callprototype,
.calltargets,
.branchtargets,
(above are critical in control flow)
.byte,
.maxnreg,
.reqntid,
.section,
.struct,
.union

wish that there's any progress

in gpu-cache.h:677,in tag(new_addr_type addr) function,
return addr& ~(m_line_sz-1). while addr is unsigned long long, and m_line_sz-1 is unsigned int, this code will loose the higher 32 bit infomation of addr.

for example:
unsigned long long a=0xFFFFFFFFFFFFFFFF;
unsigned int mask= 16;
unsigned long long mask_2=16;

unsigned long long result1=a & ~(mask-1);
unsigned long long result2=a &~(mask_2-1);
printf("%llx\n",result1);
printf("%llx\n",result2);
return 0;

the output is
fffffff0
fffffffffffffff0

To be more specific, in GPUTrafficManager::_GeneratePacket, I want to be able to look at the actual contents of the flit.

There's a line that seemed promising, f->data = data that seemed promising but as far as I can tell it's just a mem_fetch object which doesn't contain the actual data.

Is it possible to simulate DNN workloads with the dev branch? Has anyone tried that?

Environment
Ubuntu 16.04.4 LTS
gcc/g++ 5.4.0 20160609
Python 2.7
CUDA 8.0
cuDNN 7.1.4
gpgpu-sim_distribution dev branch
pytorch-gpgpu-sim (removed the git dependencies of nervanagpu @ d4eefd5, since the repo is no longer there)

PYTORCH_BIN /usr/lib/x86_64-linux-gnu/libcudnn.so

The configurations used are from the configs folder of gpgpu-sim dev branch

MNIST
I use the MNIST sample from here. (Following deval281shah 's suggestions in another discussion. )
https://github.com/gpgpu-sim/gpgpu-sim_simulations

Config
I use TITANV config.
I also tried TITANX config, but a deadlock happened with that configuration.

The simulation runs for 39 minutes, and I checked it has stimulated a number of kernels, and it reported some related information such as IPC.

In the beginning, it generated a large amount of .. it cannot find all device function required.

Warning: cannot find deviceFun maxwell_zgemmBatched_32x32_raggedMn_ct
Warning: cannot find deviceFun maxwell_zgemmBatched_64x32_raggedMn_ct
Warning: cannot find deviceFun maxwell_zgemmBatched_32x32_raggedMn_cn
Warning: cannot find deviceFun maxwell_zgemmBatched_64x32_raggedMn_cn
Warning: cannot find deviceFun maxwell_zgemmBatched_32x32_raggedMn_tc
Warning: cannot find deviceFun maxwell_zgemmBatched_64x32_raggedMn_tc
Warning: cannot find deviceFun maxwell_zgemmBatched_32x32_raggedMn_tt
Warning: cannot find deviceFun maxwell_zgemmBatched_64x32_raggedMn_tt
Warning: cannot find deviceFun maxwell_zgemmBatched_32x32_raggedMn_tn
Warning: cannot find deviceFun maxwell_zgemmBatched_64x32_raggedMn_tn
Warning: cannot find deviceFun maxwell_zgemmBatched_32x32_raggedMn_nc

Therefore, cudaLaunchKernel fails to find the device function.

However, it ends up with
...

GPGPU-Sim PTX: Setting up arguments for 4 bytes starting at 0x7ffc0b96aa68..

GPGPU-Sim PTX: cudaLaunch for 0x0x4321f0 (mode=performance simulation) on stream 0
GPGPU-Sim PTX: ERROR launching kernel -- no PTX implementation found for 0x4321f0

Does anyone encounter this problem before? Any suggestions?

Thank you so much for your help.

Hello I am using 'gpgpusim-dev ver.' and running 'cudnn_samples_v7/mnistCUDNN' I get this error message.

_"CUDNN failure
Error: CUDNN_STATUS_NOT_INITIALIZED"_

My environment settings are like below.
(Virtualbox)
Ubuntu 16.04.5 LTS
gcc 5.4.0 g++ 5.4.0
python 3.6.7
cuda 9.1
cudnn 7.0.5

Does anyone know how to solve that error message?

In GTX480 gpgpusim.config, L1D block size is set to 128B. However, in gpuwattch_gtx480.xml, L1D block size is set to 32B. Is there any reason for setting value like that? or is this just an error?

Used Environments

• Ubuntu 16.04.5
• CUDA 8.0
• cuDNN 7.1.4
• gcc 5.4.0
• g++ 5.4.0
• python 2.7.12
• GPU: GTX 1080 Ti

Problem Situation

Hello, I'm trying to run pytorch-gpgpu-sim but I'm facing a problem.

After it extract all the ptx files from libcudnn.so, it fails to get pruned PTX section list as shown at following message.

Extracting PTX file and ptxas options  194: libcudnn.194.sm_50.ptx -arch=sm_50
Extracting specific PTX file named libcudnn.195.sm_61.ptx
Extracting PTX file and ptxas options  195: libcudnn.195.sm_61.ptx -arch=sm_61
Extracting specific PTX file named libcudnn.196.sm_35.ptx
Extracting PTX file and ptxas options  196: libcudnn.196.sm_35.ptx -arch=sm_35
Error: No PTX sections found with sm capability that is lower than current forced maximum capability
 minimum ptx capability found = 0, maximum forced ptx capability = 61
 User might want to change either the forced maximum capability from gpgpusim configuration or update the compilation to generate the required PTX version

I guess that I have problem in setting sm capability, but I don't know where I have to modify. 

I have successfully built gpgpusim dev branch, but when I try to simulate my own kernel(which does not require cuBLAS), I face a runtime error indicating:

symbol lookup error: /ROOT/gpgpu-sim_distribution/lib/gcc-5.4.0/cuda-9000/release/libcudart.so.9.0: undefined symbol: yyparse

Here is the configuration of my setup:

• Ubuntu 16.04
• CUDA 9.0
• GCC 5.4.0
• GPGPU-Sim 4.0.0

Would really appreciate your help.
Thanks