PyCUDA December 2012

pycuda@tiker.net

6 participants
13 discussions

by Freddie Witherden

Hello, While attempting to compile PyCUDA under Python 3: $ virtualenv -p python3.2 --system-site-packages myenv $ cd myenv $ source bin/activate $ git clone https://github.com/inducer/pycuda.git $ cd pycuda $ git submodule init $ git submodule update $ python setup.py install I received: x86_64-pc-linux-gnu-g++ -pthread -fPIC -I/usr/lib/python3.2/site-packages/numpy/core/include -I/usr/lib/python3.2/site-packages/numpy/core/include -I/usr/include/python3.2 -c src/wrapper/_pvt_struct_v3.cpp -o build/temp.linux-x86_64-3.2/src/wrapper/_pvt_struct_v3.o src/wrapper/_pvt_struct_v3.cpp: In function ‘int s_init(PyObject*, PyObject*, PyObject*)’: src/wrapper/_pvt_struct_v3.cpp:1045:41: warning: deprecated conversion from string constant to ‘char*’ [-Wwrite-strings] src/wrapper/_pvt_struct_v3.cpp:1047:5: error: ‘PyStructType’ was not declared in this scope src/wrapper/_pvt_struct_v3.cpp: In function ‘PyObject* s_unpack(PyObject*, PyObject*)’: src/wrapper/_pvt_struct_v3.cpp:1138:5: error: ‘PyStructType’ was not declared in this scope src/wrapper/_pvt_struct_v3.cpp: In function ‘PyObject* s_unpack_from(PyObject*, PyObject*, PyObject*)’: src/wrapper/_pvt_struct_v3.cpp:1164:51: warning: deprecated conversion from string constant to ‘char*’ [-Wwrite-strings] src/wrapper/_pvt_struct_v3.cpp:1164:51: warning: deprecated conversion from string constant to ‘char*’ [-Wwrite-strings] src/wrapper/_pvt_struct_v3.cpp:1172:5: error: ‘PyStructType’ was not declared in this scope src/wrapper/_pvt_struct_v3.cpp: In function ‘PyObject* s_pack(PyObject*, PyObject*)’: src/wrapper/_pvt_struct_v3.cpp:1296:5: error: ‘PyStructType’ was not declared in this scope src/wrapper/_pvt_struct_v3.cpp: In function ‘PyObject* s_pack_into(PyObject*, PyObject*)’: src/wrapper/_pvt_struct_v3.cpp:1336:5: error: ‘PyStructType’ was not declared in this scope src/wrapper/_pvt_struct_v3.cpp: At global scope: src/wrapper/_pvt_struct_v3.cpp:1414:1: warning: deprecated conversion from string constant to ‘char*’ [-Wwrite-strings] src/wrapper/_pvt_struct_v3.cpp:1414:1: warning: deprecated conversion from string constant to ‘char*’ [-Wwrite-strings] src/wrapper/_pvt_struct_v3.cpp:1414:1: warning: deprecated conversion from string constant to ‘char*’ [-Wwrite-strings] src/wrapper/_pvt_struct_v3.cpp:1414:1: warning: deprecated conversion from string constant to ‘char*’ [-Wwrite-strings] error: command 'x86_64-pc-linux-gnu-g++' failed with exit status 1 which is very similar to https://github.com/inducer/pycuda/issues/11 in that it can also be fixed by passing the -DNDEBUG flag. Would it be possible for this fix to be ported to _pvt_struct_v3? (Or just ensure that -DNDEBUG is always passed.) Also, are there any other potential issues with PyCUDA and Python 3.x that I should be aware of? Regards, Freddie.

7 years, 10 months

cuda-gdm with "set cuda memcheck on"

by Youssef Barhomi

Hi All, I am not sure if this is the right place to post this (nvidia forum would be better I guess?), but since I am running my cuda kernels inside python (using pycuda) I am going to give this list a try. My problem is that if I run: cuda-memcheck python -m pycuda.debug <my_pycuda_code.py> I end up with few global memory access violations that look like this ========= CUDA-MEMCHECK *** compiler output in /tmp/tmpwmbHa1 Traceback (most recent call last): File "/usr/lib/python2.7/runpy.py", line 162, in _run_module_as_main "__main__", fname, loader, pkg_name) File "/usr/lib/python2.7/runpy.py", line 72, in _run_code exec code in run_globals File "/usr/local/lib/python2.7/dist-packages/pycuda-2012.1-py2.7-linux-x86_64.egg/pycuda/debug.py", line 25, in <module> execfile(mainpyfile) File "./layers.py", line 239, in <module> main() File "./layers.py", line 233, in main filt, s1, c1, s2, c2 = extract_all_layers(img, params, index_gpu) File "./layers.py", line 213, in extract_all_layers test= True) File "./layers.py", line 117, in extract_s2_c2 s2_w += [s2_d_temp.get()] File "/usr/local/lib/python2.7/dist-packages/pycuda-2012.1-py2.7-linux-x86_64.egg/pycuda/gpuarray.py", line 254, in get drv.memcpy_dtoh(ary, self.gpudata) pycuda._driver.LaunchError: cuMemcpyDtoH failed: launch failed PyCUDA WARNING: a clean-up operation failed (dead context maybe?) cuMemFree failed: launch failed PyCUDA WARNING: a clean-up operation failed (dead context maybe?) cuMemFree failed: launch failed PyCUDA WARNING: a clean-up operation failed (dead context maybe?) cuModuleUnload failed: launch failed ========= Invalid __global__ read of size 4 ========= at 0x00000970 in /tmp/tmpwmbHa1/kernel.cu:135 :extract_s2_matlab_investigation ========= by thread (31,20,0) in block (0,0,0) ========= Address 0xb00350d6c is out of bounds ========= ========= Invalid __global__ read of size 4 ========= at 0x00000970 in /tmp/tmpwmbHa1/kernel.cu:135 :extract_s2_matlab_investigation ========= by thread (30,20,0) in block (0,0,0) ========= Address 0xb00350d68 is out of bounds ..... ========= Invalid __global__ read of size 4 ========= at 0x00000970 in /tmp/tmpwmbHa1/kernel.cu:135 :extract_s2_matlab_investigation ========= by thread (0,0,0) in block (1,1,0) ========= Address 0xb00351730 is out of bounds ========= ========= Program hit error 700 on CUDA API call to cuMemcpyDtoH_v2 ========= ========= Program hit error 700 on CUDA API call to cuMemFree_v2 ========= ========= Program hit error 700 on CUDA API call to cuMemFree_v2 ========= ========= Program hit error 700 on CUDA API call to cuModuleUnload ========= ========= ERROR SUMMARY: 1096 errors Now, if I run my code (./<my_python_cuda.py>), everything works fine (no crash), even if I run it thousands of times. The output that I am getting also makes sense: I work in computer vision and I compare my pycuda code with a scipy implementation of my algorithm and they both output the same arrays. Now, I used cuda-gdb on the pycuda code and put breakpoints where cuda-memcheck tells me that global memory access violation occurs. I checked all my variables and everything made sense (I run my kernel line by line and I didn't "notice" any out of bounds access). Also, I tried to "set cuda memcheck on" before starting to run my code on cuda-gdb and it didn't stop where global memory access violation seems to occur according to cuda-memcheck. Any ideas? PS: I am new to debugging with cuda-gdb so please let me know if I am missing something crucial related to cuda-memcheck Thank you so much!! Youssef -- Youssef Barhomi, MSc, MEng. Research Software Engineer at the CLPS department Brown University T: +1 (617) 797 9929 | GMT -5:00

8 years, 1 month

Problem with demo.py and cuda-gdb

by Tomi Pieviläinen

I tried to start debugging within cuda, but can't even get the demo.py to work as shown at http://wiki.tiker.net/PyCuda/FrequentlyAskedQuestions#Is_it_possible_to_use…. After setting the breakpoint and starting the problem, it just finishes without ever breaking when calling the doublify function. Am I missing something? I'm using the cuda 5.0. -- Tomi Pieviläinen, +358 400 487 504 A: Because it disrupts the natural way of thinking. Q: Why is top posting frowned upon?

8 years, 1 month

Even synchronising when using aligned memory

by Zhe Weng

Hi all, I'm doing some experiments on paged-locked memory, but got some errors that I failed to address when I use aligned memory to invoke the kernel function. My code is like below: 1 import pycuda.driver as drv 2 import autoinit 3 4 temp = drv.aligned_empty(my_numpy_array.shape, dtype=my_numpy_array.dtype, order='C') 5 temp = drv.register_host_memory( temp, flags=drv.mem_host_register_flags.DEVICEMAP) 6 my_plptr = numpy.intp(temp.base.get_device_pointer()) 7 8 start_event.record() 9 kernel_function( my_plptr, rest_parameters, block, grid) 10 end_event.record() 11 end_event.synchronize() But when executing, it reports error saying Traceback (most recent call last): end.synchronize() pycuda._driver.LaunchError: cuEventSynchronize failed: launch failed PyCUDA WARNING: a clean-up operation failed (dead context maybe?) cuMemFree failed: launch failed This code https://gist.github.com/4036292 works good on the same machine. Thanks. John -------------------------------- M: (+61) 415786645 Canberra, Australia Sent with Sparrow (http://www.sparrowmailapp.com/?sig)

8 years, 1 month

Re: [PyCUDA] Passing pagelocked memory to a kernel

by Andreas Kloeckner

Daisuke Nishino <niboshi000(a)gmail.com> writes: > Andreas, > Thank you for your reply. > > When I add mem_flags=pycuda.driver.host_alloc_flags.DEVICEMAP in > pagelocked_empty(), > it results in error: > pycuda._driver.LogicError: cuMemHostAlloc failed: invalid value > > Does this mean my card doesn't support this functionality? > I'm not sure if it is relevant, but CAN_MAP_HOST_MEMORY entry in > pycuda.driver.Context.get_driver().get_attributes() is 1. http://documen.tician.de/pycuda/driver.html#pycuda.driver.ctx_flags.MAP_HOST Andreas

8 years, 1 month

Re: [PyCUDA] Passing pagelocked memory to a kernel

by Daisuke Nishino

Andreas, I found that the previous error is the problem of initialization. Instead of pycuda.autoinit, I made the context by myself with MAP_HOST flag. It seems working fine now. Thank you so much! My code looks like below now. ##CODE START###################### import pycuda.driver as drv drv.init() dev = drv.Device(0) ctx = dev.make_context(drv.ctx_flags.SCHED_AUTO | drv.ctx_flags.MAP_HOST) k = pycuda.compiler.SourceModule(""" __global__ void krnl(float* a) { int i = blockIdx.x * blockDim.x + threadIdx.x; a[i] = i; } """).get_function("krnl") a = drv.pagelocked_empty((10, 10), numpy.float32, mem_flags=drv.host_alloc_flags.DEVICEMAP) aa = numpy.intp(a.base.get_device_pointer()) k(aa, grid=(100,1), block=(1,1,1)) ctx.pop() ##CODE END###################### On 5 December 2012 17:14, Daisuke Nishino <niboshi000(a)gmail.com> wrote: > Andreas, > Thank you for your reply. > > When I add mem_flags=pycuda.driver.host_alloc_flags.DEVICEMAP in > pagelocked_empty(), > it results in error: > pycuda._driver.LogicError: cuMemHostAlloc failed: invalid value > > Does this mean my card doesn't support this functionality? > I'm not sure if it is relevant, but CAN_MAP_HOST_MEMORY entry in > pycuda.driver.Context.get_driver().get_attributes() is 1. > > Thanks, > Daisuke > > > > On 5 December 2012 16:47, Andreas Kloeckner <lists(a)informa.tiker.net>wrote: > >> Daisuke Nishino <niboshi000(a)gmail.com> writes: >> >> > Hi, all*. >> > >> > *I have a problem using pagelocked memory. >> > I allocated one with pagelocked_xxx or PageLockedMemoryPool, but how >> can I >> > pass it into a kernel? >> > I put a simple code below. >> > What should "aa" be? >> > >> > aa = pycuda.driver.Out(a) works just fine, but I guess it involves a >> copy. >> >> >> http://documen.tician.de/pycuda/driver.html#pycuda.driver.host_alloc_flags.… >> >> >> http://documen.tician.de/pycuda/driver.html#pycuda.driver.HostPointer.get_d… >> >> Andreas >> >> > > > -- > Daisuke Nishino > -- Daisuke Nishino

8 years, 1 month

Re: [PyCUDA] Passing pagelocked memory to a kernel

by Andreas Kloeckner

Daisuke Nishino <niboshi000(a)gmail.com> writes: > Hi, all*. > > *I have a problem using pagelocked memory. > I allocated one with pagelocked_xxx or PageLockedMemoryPool, but how can I > pass it into a kernel? > I put a simple code below. > What should "aa" be? > > aa = pycuda.driver.Out(a) works just fine, but I guess it involves a copy. http://documen.tician.de/pycuda/driver.html#pycuda.driver.host_alloc_flags.… http://documen.tician.de/pycuda/driver.html#pycuda.driver.HostPointer.get_d… Andreas

8 years, 1 month

Passing pagelocked memory to a kernel

by Daisuke Nishino

Hi, all*. *I have a problem using pagelocked memory. I allocated one with pagelocked_xxx or PageLockedMemoryPool, but how can I pass it into a kernel? I put a simple code below. What should "aa" be? aa = pycuda.driver.Out(a) works just fine, but I guess it involves a copy. Thanks, Daisuke ##CODE START###################### k = pycuda.compiler.SourceModule(""" __global__ void krnl(float* a) { int i = blockIdx.x * blockDim.x + threadIdx.x; a[i] = i; } """).get_function("krnl") a = pycuda.driver.pagelocked_empty((10, 10), numpy.float32) aa = ?? k(aa, grid=(100,1), block=(1,1,1)) ##CODE END###################### -- Daisuke Nishino

8 years, 1 month

Using Kernels from .cu file

by Zhe Weng

Hi all, Is there any way I can edit the kernel function in .h or .cu extension file, which is syntax highlight, instead of editing it as a string? For example, I have a kernel called swap(float *a, float *b, int length) stored in swap.h and implemented in swap.cu, and in the Python file where the module is like below SourceModule(""" #include"swap.h" """) But currently it just reports fail to find the swap.h file. Thanks! Cheers, John -------------------------------- M: (+61) 415786645 Canberra, Australia Sent with Sparrow (http://www.sparrowmailapp.com/?sig)

8 years, 1 month

Re: [PyCUDA] Forcing kernel launchs to be serialized

by Leandro Demarco Vedelago

After some reading and profiling I realized that as long as they ain't launched in different streams they'll be serialized. Sorry for the inconvenience On Tue, Dec 4, 2012 at 11:51 AM, Leandro Demarco Vedelago <leandrodemarco(a)gmail.com> wrote: > Hi there. I have 2 kernels, one which must be launched only after the > first one has finished it's computations as it uses the results > computed by the first. > > My code looks something like: > > def main(): > ----- Something ------ > src = SourceModule(cudaCode) > f1 = src.get_function("f1") > f2 = src.get_function("f2") > f1(args1, res1, block=..., grid=...) > f2(args2, res2, block=..., grid=...) > ---- Something else ----- > > where res1 is contained in args2 > > So, I was wondering if in "normal" conditions (I'm using the same > context and just one stream for both launches) f2 is launched only > after f1 has ended or it's possible (maybe because of the Cuda > scheduler) that it's launched before f1 is done, in which case I > should seek a way to prevent this. > > Thanks in advance, Leandro Demarco.

8 years, 1 month

2021

2020

2019

2018

2017

2016

2015

2014

2013

2012

2011

2010

2009

2008

PyCUDA December 2012