loading from file crash at thrust reduce_by_key

[georgezhlw]

Hi Anton, alenka crashed when loading, do you see what should be done to solve this? This is on Mac OS X. Test Case: Alenka]$ cat testtable.tbl 1|a 2|b 3|c Alenka]$ cat load_testtable.sql T := LOAD 'testtable.tbl' USING ('|') AS (id{1}:int, value{2}:varchar(20)); STORE T INTO 'testtable' BINARY; Alenka]$./alenka -l 1 load_testtable.sql Process count = 1 LOAD: T testtable.tbl 2 | STORE: T testtable LOADING testtable.tbl | libc++abi.dylib: terminate called throwing an exception Abort trap: 6

Analysis: (gdb) s pfor_delta_compress (source=0x500240000, source_len=8, file_name=0x7fff5fbfe9b8 "testtable.1.0", host=@0x102d22c00, tp=false, sz=0) at compress.cu:287

389 thrust::device_ptr fin_seq = thrust::device_malloc(cnt); (gdb) 391 thrust::reduce_by_key(add_seq, add_seq+recCount,s_copy1,thrust::make_discard_iterator(), (gdb) libc++abi.dylib: terminate called throwing an exception

Program received signal SIGABRT, Aborted. 0x00007fff956b8212 in __pthread_kill ()

thanks, George

[georgezhlw]

BTW, i got some warning when compile, not sure if it's related to the crash:

Alenka]$ make /usr/local/cuda/bin/nvcc -O3 -arch=sm_20 -m64 -I/usr/local/boost_1_53_0 -lcudpp -c merge.cu merge.cu(364): error: space required between adjacent ">" delimiters of nested template argument lists (">>" is the right shift operator)

/usr/local/cuda/bin/../include/thrust/detail/allocator/allocator_traits.h(124): error: name followed by "::" must be a class or namespace name detected during: instantiation of class "thrust::detail::allocator_traits [with Alloc=]" /usr/local/cuda/bin/../include/thrust/detail/contiguous_storage.h(39): here instantiation of class "thrust::detail::contiguous_storage<T, Alloc> [with T=float_type, Alloc=]" /usr/local/cuda/bin/../include/thrust/detail/vector_base.h(46): here instantiation of class "thrust::detail::vector_base<T, Alloc> [with T=float_type, Alloc=]" /usr/local/cuda/bin/../include/thrust/host_vector.h(53): here instantiation of class "thrust::host_vector<T, Alloc> [with T=float_type, Alloc=]"

modify source file, change non-shift ">>" to "> >" in merge.cu, cm.cu near "locate push_back("

make: /usr/local/cuda/bin/nvcc -O3 -arch=sm_20 -m64 -I/usr/local/boost_1_53_0 -lcudpp -c merge.cu /usr/local/cuda/bin/../include/thrust/detail/allocator/allocator_traits.inl(136): warning: calling a host function from a host device function is not allowed detected during: instantiation of "thrust::detail::enable_if<thrust::detail::allocator_traits_detail::has_member_destroy<Alloc, T>::value, void>::type thrust::detail::allocator_traits_detail::destroy(Alloc &, T ) [with Alloc=uninitialized_host_allocator, T=int_type]" (265): here instantiation of "void thrust::detail::allocator_traits::destroy(thrust::detail::allocator_traits::allocator_type &, T ) [with Alloc=uninitialized_host_allocator, T=int_type]" /usr/local/cuda/bin/../include/thrust/detail/allocator/destroy_range.inl(103): here instantiation of "void thrust::detail::allocator_traits_detail::destroy_via_allocator::operator()(T &) [with Allocator=uninitialized_host_allocator, T=int_type]" /usr/local/cuda/bin/../include/thrust/detail/function.h(47): here

[antonmks]

Hi George The warnings are ok, just complaining about uninitialized host allocators. I run your script and it works fine. I assume you compiled any CUDA programs on your machine before and it run ok ? See if you can compile and run some example Thrust programs. Also, if you use varchars, use lengths from 1 to 10,most of the other lengths are not yet defined in strings.cu file ( you can add it there if you want to)

[georgezhlw]

Anton, CUDA and basic Thrust program works. Since alenka fails at reduce_by_key, i tested thrust::reduce_by_key, it fails with same error, so this should be a thrust 1.7 problem on my machine.

thrust]$ cat reducebykey.cu

include < thrust/reduce.h>

include < thrust/unique.h>

include < thrust/iterator/discard_iterator.h>

include < thrust/host_vector.h>

include < thrust/device_vector.h>

int main(void) { thrust::device_vector keys; thrust::device_vector values; typename thrust::pair<typename thrust::device_vector::iterator, typename thrust::device_vector::iterator> new_last; keys.resize(9); keys[0] = 11; keys[1] = 11; keys[2] = 21; keys[3] = 20; keys[4] = 21; keys[5] = 21; keys[6] = 21; keys[7] = 37; keys[8] = 37; values.resize(9); values[0] = 0; values[1] = 1; values[2] = 2; values[3] = 3; values[4] = 4; values[5] = 5; values[6] = 6; values[7] = 7; values[8] = 8; thrust::device_vector output_keys(keys.size()); thrust::device_vector output_values(values.size()); new_last = thrust::reduce_by_key(keys.begin(), keys.end(), values.begin(), output_keys.begin(), output_values.begin()); return 0; } thrust]$ nvcc -arch=sm_20 -m64 reducebykey.cu -o reducebykey thrust]$ ./reducebykey libc++abi.dylib: terminate called throwing an exception Abort trap: 6 regards, George

[antonmks]

Which version of CUDA do you have on your machine ? nvcc --version

[georgezhlw]

$ nvcc --version nvcc: NVIDIA (R) Cuda compiler driver Copyright (c) 2005-2012 NVIDIA Corporation Built on Fri_Sep_28_16:10:16_PDT_2012 Cuda compilation tools, release 5.0, V0.2.1221

[antonmks]

Ok, that should work. Which GPU do you use ?

[georgezhlw]

GPU is GeForce GT 650M; global mem: 536543232B; compute v3.0; clock: 774500 kHz

I found that the failure is because i use -arch=sm_20 (or sm_30) together with -m64. reduce_by_key will success if i use only one of the option.

For compiling Alenka, if I don't use -arch, if fails at: /usr/local/cuda/bin/nvcc -O3 -m64 -I/usr/local/boost_1_53_0 -c strings.cu ptxas /tmp/tmpxft_0000079f_00000000-5_strings.ptx, line 0; fatal : (C9999) max reg limit too low

if i don't use -m64, it will compile into 32 bit, and by default cudpp compiled to 64 bit on my system. I really don't want to compile the whole thing to 32bit. So i am stuck (:

[antonmks]

You definitely should compile 64 bit version. I haven't seen this error (reg limit) before but it might be related to the limited hardware resources of your GPU. Because it fails to compile strings.cu I can suggest the following thing : comment out from strings.cu all processing of strings longer than, say, 10 characters and see if that compiles.