victordion / cusp-library Goto Github PK

Due to the lack of implicit conversion of int to cusp::complex the code does 
not work.
The attached patch add the explicit conversion.

What steps will reproduce the problem?
1. Consider a simple test case cusp_test.cpp:

#include <cusp/blas.h>

int main()
{
  return 0;
}

2. compile with

g++ -I /path/to/cusp -I /path/to/thrust -I /usr/local/cuda/include cusp_test.cpp

What is the expected output? What do you see instead?

compiler error:
In file included from /path/to/cusp/cusp/blas.h:199,
                 from cusp_test.cpp:1:
/path/to/cusp/cusp/detail/blas.inl: In function 'typename 
thrust::iterator_value<Iterator>::type cusp::blas::nrm2(InputIterator, 
InputIterator)':
/path/to/cusp/cusp/detail/blas.inl:395: error: 'sqrt' is not a member of 'std'

What version of the product are you using? On what operating system?

Tested on:

thrust 1150:9f5c19852f16
cusp 259:9fdf9bde9f6d

Ubuntu 8.04 2.6.24-24-server #1 SMP Wed Apr 15 15:41:09 UTC 2009 x86_64 
GNU/Linux
Ubuntu 9.10 2.6.31-22-generic #61-Ubuntu SMP Wed Jul 28 02:02:56 UTC 2010 
x86_64 GNU/Linux

g++-4.4 (Ubuntu 4.4.1-4ubuntu9) 4.4.1
g++-4.3 (Ubuntu 4.3.4-5ubuntu1) 4.3.4
g++-4.2 (GCC) 4.2.4 (Ubuntu 4.2.4-5ubuntu1)

Based on the comments by Florian in this thread [1] on cusp-users it seems we 
should allow polymorphic usage of Cusp monitors and linear_operators.

[1] 
http://groups.google.com/group/cusp-users/browse_thread/thread/73dbba85ee83efb9

Here's a quick patch:

Index: cusp/detail/host/conversion.h
===================================================================
--- cusp/detail/host/conversion.h   (revision 76)
+++ cusp/detail/host/conversion.h   (working copy)
@@ -328,7 +328,7 @@
         for(IndexType n = 0; n < num_entries_per_row; n++)
         {
             const IndexType j = src.column_indices(i,n);
-            const IndexType v = src.values(i,n);
+            const ValueType v = src.values(i,n);
             if(j != invalid_index)
             {
                 dst.column_indices[num_entries] = j;

Would store an array of integer indices encoding a permutation of [0,num_rows). 
 Specializing transpose(permutation_matrix,...) and 
multiply(permutation_matrix, ...) for matrix-vector and matrix-matrix 
multiplication would have a noticeable performance advantage.

Questions:
Can permutation_matrix be implemented more concisely as a coo_matrix_view?

At minimum everything should be tested for IndexType=int and long.

The current fix for 'volatile complex' does not appear to work on MSVC.

On WinXP, MSVC 2005, nvcc 3.2.
$ cd testing; scons
.
.
.
complex.cu(333) : error C2593: 'operator +' is ambiguous
        c:/documents and settings/mgarland/my documents/work/cusp-library\cusp/c
omplex.h(547): could be 'cusp::complex<float> cusp::operator +<float>(volatile c
usp::complex<float> &,volatile cusp::complex<float> &)' [found using argument-de
pendent lookup]
        c:/documents and settings/mgarland/my documents/work/cusp-library\cusp/c
omplex.h(533): or       'cusp::complex<float> cusp::operator +<float>(const cusp
::complex<float> &,const cusp::complex<float> &)' [found using argument-dependen
t lookup]
        while trying to match the argument list '(cusp::complex<float>, cusp::co
mplex<float>)'
        complex.cu(456) : see reference to function template instantiation 'cusp
::complex<float> test_complex_compilation_entry<ValueType>(void)' being compiled

        with
        [
            ValueType=float
        ]
        complex.cu(469) : see reference to function template instantiation 'void
 TestComplex<float>(void)' being compiled
complex.cu(333) : error C2593: 'operator +' is ambiguous
        c:/documents and settings/mgarland/my documents/work/cusp-library\cusp/c
omplex.h(547): could be 'cusp::complex<double> cusp::operator +<double>(volatile
 cusp::complex<double> &,volatile cusp::complex<double> &)' [found using argumen
t-dependent lookup]
        c:/documents and settings/mgarland/my documents/work/cusp-library\cusp/c
omplex.h(533): or       'cusp::complex<double> cusp::operator +<double>(const cu
sp::complex<double> &,const cusp::complex<double> &)' [found using argument-depe
ndent lookup]
        while trying to match the argument list '(cusp::complex<double>, cusp::c
omplex<double>)'
        complex.cu(456) : see reference to function template instantiation 'cusp
::complex<double> test_complex_compilation_entry<ValueType>(void)' being compile
d
        with
        [
            ValueType=double
        ]
        complex.cu(469) : see reference to function template instantiation 'void
 TestComplex<double>(void)' being compiled
scons: *** [complex.obj] Error 2
scons: building terminated because of errors.

Specifically, array2d and array2d_view should not assume that the stride is 
equal to the minor dimension.

What steps will reproduce the problem?
1. Apply the smoothed_aggregation preconditioner to a big enough matrix.

What is the expected output? What do you see instead?
The solver fails with bad::alloc because the GPU memory is exhausted by 
cusp::detail::device::spmm_coo.


What version of the product are you using? On what operating system?
cusp v0.1.1
thrust v1.3.0
nvidia cuda 3.0
windows xp professional x64 edition

Please provide any additional information below.
I think, the problem is general, but if you need, I can upload my particular 
matrix. It only has 8M of nonzeros and fails on a GTX285 card which has 1GB RAM.

I wonder if the spmm algorithm could be optimized or avoided in 
smoothed_aggregation by keeping the original matrices, let's say A and B, and 
computing y=A*(B*x) instead of y=(A*B)*x?

Thanks!

Is there an easy way to implement an efficient parallel sparse triangular
matrix backsubstitution step using cusp/thrust in the library? This would
enable the utilization of more elaborate preconditioners like SSOR or
custom incomplete Cholesky / LU to be easily integrated into the krylov pcg
solver.

Hello,

At line 368 in testing/complex.cu, you test if the architecture supports the 
double precision, hence is higher than 1.3, but the test seems wrong if you 
have a 2.0 for example.

Thanks a lot for your great job!
Luc.

currently the host code assumes the row entries are unsorted and first
computes a histogram, etc.

In light of issue #27 it seems we'll need to make convergence_monitor [1] 
handle multiple calls to .finished() per iteration.  Possible remedies are

1) Record all residual norms and use iteration_count() to compute the 
convergence rates.
2) Record only one residual norm per iteration
3) Record both the iteration number and norm for all calls to .finished()

For #2 we could keep the first, last or minimum norm by default and support 
general BinaryOperators so any behavior can be handled.

My initial preference is for #2.  The problem with #1 is that the 
immediate_rate() computation is no longer straightforward.  The last option 
keeps the most information but is probably not what most people want.  Option 
#2 also has the nice property that (for reasonable BinaryOperators) any 
redundant calls to finished() won't affect the results.

So, I like #2, but I haven't given much thought to how monitors will work with 
multiple RHS or handle methods like GMRES.

Thoughts?

[1] http://code.google.com/p/cusp-library/source/browse/cusp/monitor.h#256

Example MatrixFormats/ell.cu doesn't currently compile:

./cusp/ell_matrix.h(206): error: namespace "cusp" has no member "minimum_space"

Adding #include <cusp/memory.h> to cusp/ell_matrix.h seems to fix the problem, 
but I don't know if there's something deeper going on.

We'll also need a unit test on some simple matrices with anisotropic stencils.

The most general form might allow users to specify whether to filter out 
explicit zeros (defaulting to true).

In particular
  warning: comparison between signed and unsigned integer expressions

It may be best to use size_t for all the matrix dimensions.

following the implementation of std::complex<T>

Specializations for float and double should use float2 and double2 for 
coalescing.

thrust::gather is now deprecated. The latest CUSP should be compileable with 
the latest Thrust, shouldn't it?

Currently only COO * COO is supported

Currently only coo_matrix, csr_matrix, and array2d are supported.

We should choose THREADS_PER_VECTOR=2,4,8,16,32 based on num_entries/num_rows

http://code.google.com/p/cusp-library/source/browse/trunk/cusp/detail/device/spm
v/csr_vector.h

The CSR (vector) kernel does not consider alignment when accessing the
matrix arrays.  For matrices with large row sizes are likely to see a
significant benefit.

Cusp's bicgstab can't solve the system Ix = b.  Instead of converging in one 
step, it yields NaN's.

For example, take "examples/Solvers/bicgstab.cu"

    http://code.google.com/p/cusp-library/source/browse/examples/Solvers/bicgstab.cu

Change line 32,

-    cusp::krylov::bicgstab(A, x, b, monitor, M);
+    cusp::krylov::bicgstab(M, x, b, monitor, M);

The solver fails to converge.

Provide an OpenMP backend like Thrust's

Appears to be a problem with std::getline()

http://lists.apple.com/archives/cocoa-dev/2009/Sep/msg01199.html
http://stackoverflow.com/questions/2234557/c-using-getline-prints-pointer-being-
freed-was-not-allocated-in-xcode

Reported in cusp-user thread:
http://groups.google.com/group/cusp-users/browse_thread/thread/6aa484f775da2070?
hl=en

$ cd testing
$ scons        # arch=sm_10 is the default
$ ./tester     # Tesla C2050, CUDA 3.1, Linux & Windows

================================================================
FAILURE: TestComplexDouble
[complex.cu:452] values are not approximately equal: 1.41894 5.38788e-315 [type=
'double']
================================================================

This test should probably just be disabled when arch=sm_10, since sm_13 is 
required for double support.  The code appears to be trying to do that, but it 
is unsuccessful.

cusp/detail/device/arch.h is missing after revision 57fd568deacf which requires 
it.

At the moment the absolute routine in blas.inl just works for real numers.
The attached one line patch makes it work for complex number also.

I get this compiling PETSc with the latest Thrust and Cusp:

knepley@Matthew-Knepleys-MacBook-Air:/PETSc3/petsc/petsc-dev/src/mat/impls/aij/s
eq/seqcuda$ make PETSC_DIR=/PETSc3/petsc/petsc-dev PETSC_ARCH=darwin-cuda
make PETSC_DIR=/PETSc3/petsc/petsc-dev PETSC_ARCH=darwin-cuda
nvcc -m64  -c --compiler-options="-PIC -Wall -Wwrite-strings 
-Wno-strict-aliasing -Wno-unknown-pragmas -g3   
-I/PETSc3/petsc/petsc-dev/darwin-cuda/include -I/PETSc3/petsc/petsc-dev/include 
-I/usr/local/cuda/include -I/usr/local/include -I/PETSc3/multicore/cusp/ 
-I/PETSc3/multicore/thrust/ -I/PETSc3/petsc/petsc-dev/include/mpiuni      
-D__INSDIR__=src/mat/impls/aij/seq/seqcuda/" aijcuda.cu
/PETSc3/multicore/thrust/thrust/detail/device/cuda/detail/b40c/vector_types.h(37
): error: expected an identifier

/PETSc3/multicore/thrust/thrust/detail/device/cuda/detail/b40c/vector_types.h(45
): error: expected an identifier

/PETSc3/multicore/thrust/thrust/detail/device/cuda/detail/b40c/vector_types.h(51
): error: expected an identifier

/PETSc3/multicore/thrust/thrust/detail/device/cuda/detail/b40c/vector_types.h(54
): error: expected an identifier

/PETSc3/multicore/thrust/thrust/detail/device/cuda/detail/b40c/vector_types.h(54
): error: expected an identifier

/PETSc3/multicore/thrust/thrust/detail/device/cuda/detail/b40c/vector_types.h(58
): error: expected an identifier

/PETSc3/multicore/thrust/thrust/detail/device/cuda/detail/b40c/vector_types.h(63
): error: expected an identifier

/PETSc3/multicore/thrust/thrust/detail/device/cuda/detail/b40c/vector_types.h(63
): error: expected an identifier

/PETSc3/multicore/thrust/thrust/detail/device/cuda/detail/b40c/vector_types.h(75
): error: expected either a definition or a tag name

/PETSc3/multicore/thrust/thrust/detail/device/cuda/detail/b40c/vector_types.h(75
): error: expected either a definition or a tag name

/PETSc3/multicore/thrust/thrust/detail/device/cuda/detail/b40c/vector_types.h(75
): error: expected either a definition or a tag name

/PETSc3/multicore/thrust/thrust/detail/device/cuda/detail/b40c/vector_types.h(76
): error: expected either a definition or a tag name

/PETSc3/multicore/thrust/thrust/detail/device/cuda/detail/b40c/vector_types.h(76
): error: expected either a definition or a tag name

/PETSc3/multicore/thrust/thrust/detail/device/cuda/detail/b40c/vector_types.h(76
): error: expected either a definition or a tag name

/PETSc3/multicore/thrust/thrust/detail/device/cuda/detail/b40c/vector_types.h(77
): error: expected either a definition or a tag name

/PETSc3/multicore/thrust/thrust/detail/device/cuda/detail/b40c/vector_types.h(77
): error: expected either a definition or a tag name

/PETSc3/multicore/thrust/thrust/detail/device/cuda/detail/b40c/vector_types.h(77
): error: expected either a definition or a tag name

/PETSc3/multicore/thrust/thrust/detail/device/cuda/detail/b40c/vector_types.h(78
): error: expected either a definition or a tag name

/PETSc3/multicore/thrust/thrust/detail/device/cuda/detail/b40c/vector_types.h(78
): error: expected either a definition or a tag name

/PETSc3/multicore/thrust/thrust/detail/device/cuda/detail/b40c/vector_types.h(78
): error: expected either a definition or a tag name

/PETSc3/multicore/thrust/thrust/detail/device/cuda/detail/b40c/vector_types.h(79
): error: expected either a definition or a tag name

/PETSc3/multicore/thrust/thrust/detail/device/cuda/detail/b40c/vector_types.h(79
): error: expected either a definition or a tag name

/PETSc3/multicore/thrust/thrust/detail/device/cuda/detail/b40c/vector_types.h(79
): error: expected either a definition or a tag name

/PETSc3/multicore/thrust/thrust/detail/device/cuda/detail/b40c/vector_types.h(80
): error: expected either a definition or a tag name

/PETSc3/multicore/thrust/thrust/detail/device/cuda/detail/b40c/vector_types.h(80
): error: expected either a definition or a tag name

/PETSc3/multicore/thrust/thrust/detail/device/cuda/detail/b40c/vector_types.h(80
): error: expected either a definition or a tag name

/PETSc3/multicore/thrust/thrust/detail/device/cuda/detail/b40c/vector_types.h(81
): error: expected either a definition or a tag name

/PETSc3/multicore/thrust/thrust/detail/device/cuda/detail/b40c/vector_types.h(81
): error: expected either a definition or a tag name

/PETSc3/multicore/thrust/thrust/detail/device/cuda/detail/b40c/vector_types.h(81
): error: expected either a definition or a tag name

/PETSc3/multicore/thrust/thrust/detail/device/cuda/detail/b40c/vector_types.h(82
): error: expected either a definition or a tag name

/PETSc3/multicore/thrust/thrust/detail/device/cuda/detail/b40c/vector_types.h(82
): error: expected either a definition or a tag name

/PETSc3/multicore/thrust/thrust/detail/device/cuda/detail/b40c/vector_types.h(82
): error: expected either a definition or a tag name

/PETSc3/multicore/thrust/thrust/detail/device/cuda/detail/b40c/vector_types.h(83
): error: expected either a definition or a tag name

/PETSc3/multicore/thrust/thrust/detail/device/cuda/detail/b40c/vector_types.h(83
): error: expected either a definition or a tag name

/PETSc3/multicore/thrust/thrust/detail/device/cuda/detail/b40c/vector_types.h(83
): error: expected either a definition or a tag name

/PETSc3/multicore/thrust/thrust/detail/device/cuda/detail/b40c/vector_types.h(84
): error: expected either a definition or a tag name

/PETSc3/multicore/thrust/thrust/detail/device/cuda/detail/b40c/vector_types.h(84
): error: expected either a definition or a tag name

/PETSc3/multicore/thrust/thrust/detail/device/cuda/detail/b40c/vector_types.h(84
): error: expected either a definition or a tag name

/PETSc3/multicore/thrust/thrust/detail/device/cuda/detail/b40c/vector_types.h(85
): error: expected either a definition or a tag name

/PETSc3/multicore/thrust/thrust/detail/device/cuda/detail/b40c/vector_types.h(85
): error: expected either a definition or a tag name

/PETSc3/multicore/thrust/thrust/detail/device/cuda/detail/b40c/vector_types.h(85
): error: expected either a definition or a tag name

/PETSc3/multicore/thrust/thrust/detail/device/cuda/detail/b40c/vector_types.h(86
): error: expected either a definition or a tag name

/PETSc3/multicore/thrust/thrust/detail/device/cuda/detail/b40c/vector_types.h(86
): error: expected either a definition or a tag name

/PETSc3/multicore/thrust/thrust/detail/device/cuda/detail/b40c/vector_types.h(86
): error: expected either a definition or a tag name

/PETSc3/multicore/thrust/thrust/detail/device/cuda/detail/b40c/radixsort_scansca
tter_kernel.h(124): error: "DecodeDigits" is not a function or static data 
member

/PETSc3/multicore/thrust/thrust/detail/device/cuda/detail/b40c/radixsort_scansca
tter_kernel.h(125): error: expected a ";"

/PETSc3/multicore/thrust/thrust/detail/device/cuda/detail/b40c/radixsort_scansca
tter_kernel.h(130): warning: this pragma must immediately precede a statement

/PETSc3/multicore/thrust/thrust/detail/device/cuda/detail/b40c/radixsort_scansca
tter_kernel.h(184): warning: this pragma must immediately precede a statement

/PETSc3/multicore/thrust/thrust/detail/device/cuda/detail/b40c/radixsort_scansca
tter_kernel.h(197): warning: parsing restarts here after previous syntax error

/PETSc3/multicore/thrust/thrust/detail/device/cuda/detail/b40c/radixsort_scansca
tter_kernel.h(198): error: expected a declaration

/PETSc3/multicore/thrust/thrust/detail/device/cuda/detail/b40c/radixsort_scansca
tter_kernel.h(199): error: expected a declaration

/PETSc3/multicore/thrust/thrust/detail/device/cuda/detail/b40c/radixsort_scansca
tter_kernel.h(200): error: expected a declaration

/PETSc3/multicore/thrust/thrust/detail/device/cuda/detail/b40c/radixsort_scansca
tter_kernel.h(401): error: expected an identifier

/PETSc3/multicore/thrust/thrust/detail/device/cuda/detail/b40c/radixsort_scansca
tter_kernel.h(401): error: expected a type specifier

/PETSc3/multicore/thrust/thrust/detail/device/cuda/detail/b40c/radixsort_scansca
tter_kernel.h(455): error: expected an identifier

/PETSc3/multicore/thrust/thrust/detail/device/cuda/detail/b40c/radixsort_scansca
tter_kernel.h(455): error: expected a type specifier

/PETSc3/multicore/thrust/thrust/detail/device/cuda/detail/b40c/radixsort_scansca
tter_kernel.h(501): error: expected an identifier

/PETSc3/multicore/thrust/thrust/detail/device/cuda/detail/b40c/radixsort_scansca
tter_kernel.h(501): error: expected a type specifier

/PETSc3/multicore/thrust/thrust/detail/device/cuda/detail/b40c/radixsort_scansca
tter_kernel.h(518): error: expected an identifier

/PETSc3/multicore/thrust/thrust/detail/device/cuda/detail/b40c/radixsort_scansca
tter_kernel.h(518): error: expected a type specifier

/PETSc3/multicore/thrust/thrust/detail/device/cuda/detail/b40c/radixsort_scansca
tter_kernel.h(549): error: "SwapAndScatterSm13" is not a function or static 
data member

/PETSc3/multicore/thrust/thrust/detail/device/cuda/detail/b40c/radixsort_scansca
tter_kernel.h(550): error: expected a ";"

/PETSc3/multicore/thrust/thrust/detail/device/cuda/detail/b40c/radixsort_scansca
tter_kernel.h(565): warning: this pragma must immediately precede a statement

/PETSc3/multicore/thrust/thrust/detail/device/cuda/detail/b40c/radixsort_scansca
tter_kernel.h(568): warning: this pragma must immediately precede a statement

/PETSc3/multicore/thrust/thrust/detail/device/cuda/detail/b40c/radixsort_scansca
tter_kernel.h(577): warning: this pragma must immediately precede a statement

/PETSc3/multicore/thrust/thrust/detail/device/cuda/detail/b40c/radixsort_scansca
tter_kernel.h(579): warning: parsing restarts here after previous syntax error

/PETSc3/multicore/thrust/thrust/detail/device/cuda/detail/b40c/radixsort_scansca
tter_kernel.h(580): error: identifier "K" is undefined

/PETSc3/multicore/thrust/thrust/detail/device/cuda/detail/b40c/radixsort_scansca
tter_kernel.h(580): error: identifier "UNGUARDED_IO" is undefined

/PETSc3/multicore/thrust/thrust/detail/device/cuda/detail/b40c/radixsort_scansca
tter_kernel.h(580): error: identifier "PASSES_PER_CYCLE" is undefined

/PETSc3/multicore/thrust/thrust/detail/device/cuda/detail/b40c/radixsort_scansca
tter_kernel.h(580): error: identifier "SETS_PER_PASS" is undefined

/PETSc3/multicore/thrust/thrust/detail/device/cuda/detail/b40c/radixsort_scansca
tter_kernel.h(580): error: identifier "PostprocessFunctor" is undefined

/PETSc3/multicore/thrust/thrust/detail/device/cuda/detail/b40c/radixsort_scansca
tter_kernel.h(580): error: expected a declaration

/PETSc3/multicore/thrust/thrust/detail/device/cuda/detail/b40c/radixsort_scansca
tter_kernel.h(583): error: expected a declaration

/PETSc3/multicore/thrust/thrust/detail/device/cuda/detail/b40c/radixsort_scansca
tter_kernel.h(600): warning: this pragma must immediately precede a statement

/PETSc3/multicore/thrust/thrust/detail/device/cuda/detail/b40c/radixsort_scansca
tter_kernel.h(605): warning: parsing restarts here after previous syntax error

/PETSc3/multicore/thrust/thrust/detail/device/cuda/detail/b40c/radixsort_scansca
tter_kernel.h(666): error: "SwapAndScatterPairs" is not a function or static 
data member

/PETSc3/multicore/thrust/thrust/detail/device/cuda/detail/b40c/radixsort_scansca
tter_kernel.h(667): error: expected a ";"

/PETSc3/multicore/thrust/thrust/detail/device/cuda/detail/b40c/radixsort_scansca
tter_kernel.h(814): warning: parsing restarts here after previous syntax error

/PETSc3/multicore/thrust/thrust/detail/device/cuda/detail/b40c/radixsort_scansca
tter_kernel.h(815): error: expected a declaration

/PETSc3/multicore/thrust/thrust/detail/device/cuda/detail/b40c/radixsort_scansca
tter_kernel.h(822): error: expected a declaration

/PETSc3/multicore/thrust/thrust/detail/device/cuda/detail/b40c/radixsort_scansca
tter_kernel.h(847): warning: this pragma must immediately precede a statement

/PETSc3/multicore/thrust/thrust/detail/device/cuda/detail/b40c/radixsort_scansca
tter_kernel.h(863): warning: this pragma must immediately precede a statement

/PETSc3/multicore/thrust/thrust/detail/device/cuda/detail/b40c/radixsort_scansca
tter_kernel.h(866): warning: this pragma must immediately precede a statement

/PETSc3/multicore/thrust/thrust/detail/device/cuda/detail/b40c/radixsort_scansca
tter_kernel.h(875): warning: parsing restarts here after previous syntax error

/PETSc3/multicore/thrust/thrust/detail/device/cuda/detail/b40c/radixsort_scansca
tter_kernel.h(878): error: identifier "WarpScan" is undefined

/PETSc3/multicore/thrust/thrust/detail/device/cuda/detail/b40c/radixsort_scansca
tter_kernel.h(878): error: identifier "RADIX_DIGITS" is undefined

/PETSc3/multicore/thrust/thrust/detail/device/cuda/detail/b40c/radixsort_scansca
tter_kernel.h(878): error: expected an identifier

/PETSc3/multicore/thrust/thrust/detail/device/cuda/detail/b40c/radixsort_scansca
tter_kernel.h(879): error: identifier "digit_scan" is undefined

/PETSc3/multicore/thrust/thrust/detail/device/cuda/detail/b40c/radixsort_scansca
tter_kernel.h(880): error: identifier "inclusive_total" is undefined

/PETSc3/multicore/thrust/thrust/detail/device/cuda/detail/b40c/radixsort_scansca
tter_kernel.h(881): error: expected a type specifier

/PETSc3/multicore/thrust/thrust/detail/device/cuda/detail/b40c/radixsort_scansca
tter_kernel.h(884): error: this operator is not allowed in an integral constant 
expression

/PETSc3/multicore/thrust/thrust/detail/device/cuda/detail/b40c/radixsort_scansca
tter_kernel.h(884): error: identifier "my_carry" is undefined

/PETSc3/multicore/thrust/thrust/detail/device/cuda/detail/b40c/radixsort_scansca
tter_kernel.h(887): error: expected a declaration

/PETSc3/multicore/thrust/thrust/detail/device/cuda/detail/b40c/radixsort_scansca
tter_kernel.h(889): warning: this pragma must immediately precede a statement

/PETSc3/multicore/thrust/thrust/detail/device/cuda/detail/b40c/radixsort_scansca
tter_kernel.h(894): warning: parsing restarts here after previous syntax error

/PETSc3/multicore/thrust/thrust/detail/device/cuda/detail/b40c/radixsort_scansca
tter_kernel.h(896): error: explicit type is missing ("int" assumed)

/PETSc3/multicore/thrust/thrust/detail/device/cuda/detail/b40c/radixsort_scansca
tter_kernel.h(896): error: cannot overload functions distinguished by return 
type alone

/PETSc3/multicore/thrust/thrust/detail/device/cuda/detail/b40c/radixsort_scansca
tter_kernel.h(902): error: UpdateRanks is not a template

/PETSc3/multicore/thrust/thrust/detail/device/cuda/detail/b40c/radixsort_scansca
tter_kernel.h(902): error: identifier "RADIX_DIGITS" is undefined

/PETSc3/multicore/thrust/thrust/detail/device/cuda/detail/b40c/radixsort_scansca
tter_kernel.h(902): error: identifier "PASSES_PER_CYCLE" is undefined

/PETSc3/multicore/thrust/thrust/detail/device/cuda/detail/b40c/radixsort_scansca
tter_kernel.h(902): error: identifier "SETS_PER_PASS" is undefined

/PETSc3/multicore/thrust/thrust/detail/device/cuda/detail/b40c/radixsort_scansca
tter_kernel.h(902): error: expected a ")"

/PETSc3/multicore/thrust/thrust/detail/device/cuda/detail/b40c/radixsort_scansca
tter_kernel.h(902): error: expected a ";"

/PETSc3/multicore/thrust/thrust/detail/device/cuda/detail/b40c/radixsort_scansca
tter_kernel.h(911): error: SwapAndScatterSm10 is not a template

/PETSc3/multicore/thrust/thrust/detail/device/cuda/detail/b40c/radixsort_scansca
tter_kernel.h(911): error: identifier "K" is undefined

/PETSc3/multicore/thrust/thrust/detail/device/cuda/detail/b40c/radixsort_scansca
tter_kernel.h(911): error: identifier "V" is undefined

/PETSc3/multicore/thrust/thrust/detail/device/cuda/detail/b40c/radixsort_scansca
tter_kernel.h(911): error: identifier "UNGUARDED_IO" is undefined

/PETSc3/multicore/thrust/thrust/detail/device/cuda/detail/b40c/radixsort_scansca
tter_kernel.h(911): error: identifier "PostprocessFunctor" is undefined

/PETSc3/multicore/thrust/thrust/detail/device/cuda/detail/b40c/radixsort_scansca
tter_kernel.h(912): error: expected a ")"

/PETSc3/multicore/thrust/thrust/detail/device/cuda/detail/b40c/radixsort_scansca
tter_kernel.h(913): error: variable "ranks" has already been defined

/PETSc3/multicore/thrust/thrust/detail/device/cuda/detail/b40c/radixsort_scansca
tter_kernel.h(920): error: expected a ";"

/PETSc3/multicore/thrust/thrust/detail/device/cuda/detail/b40c/radixsort_scansca
tter_kernel.h(936): error: explicit type is missing ("int" assumed)

/PETSc3/multicore/thrust/thrust/detail/device/cuda/detail/b40c/radixsort_scansca
tter_kernel.h(936): error: cannot overload functions distinguished by return 
type alone

/PETSc3/multicore/thrust/thrust/detail/device/cuda/detail/b40c/radixsort_scansca
tter_kernel.h(938): error: expected a declaration

/PETSc3/multicore/thrust/thrust/detail/device/cuda/detail/b40c/radixsort_scansca
tter_kernel.h(978): warning: parsing restarts here after previous syntax error

/PETSc3/multicore/thrust/thrust/detail/device/cuda/detail/b40c/radixsort_scansca
tter_kernel.h(980): error: identifier "LOG_RAKING_THREADS_PER_PASS" is undefined

/PETSc3/multicore/thrust/thrust/detail/device/cuda/detail/b40c/radixsort_scansca
tter_kernel.h(980): error: identifier "LOG_SCAN_LANES_PER_PASS" is undefined

$ cd testing
$ scons
.
.
complex.cu(451): error: identifier "is_same_type" is undefined
          detected during instantiation of "void TestComplex<ValueType>() [with
ValueType=float]"
(460): here

complex.cu(451): error: type name is not allowed
          detected during instantiation of "void TestComplex<ValueType>() [with
ValueType=float]"
(460): here

complex.cu(451): error: type name is not allowed
          detected during instantiation of "void TestComplex<ValueType>() [with
ValueType=float]"
(460): here

complex.cu(451): error: the global scope has no "result"
          detected during instantiation of "void TestComplex<ValueType>() [with
ValueType=float]"
(460): here

complex.cu(451): error: identifier "is_same_type" is undefined
          detected during instantiation of "void TestComplex<ValueType>() [with
ValueType=double]"
(460): here

complex.cu(451): error: type name is not allowed
          detected during instantiation of "void TestComplex<ValueType>() [with
ValueType=double]"
(460): here

complex.cu(451): error: type name is not allowed
          detected during instantiation of "void TestComplex<ValueType>() [with
ValueType=double]"
(460): here

complex.cu(451): error: the global scope has no "result"
          detected during instantiation of "void TestComplex<ValueType>() [with
ValueType=double]"
(460): here

Support allocation and deallocation of Cusp matrix containers in CPP files.  

Sizes of arrays passed to cusp::blas:: functions should be checked for
consistency.  An exception should be raised if sizes are inconsistent.

Our spmv kernels require coo_matrix row indices to be in sorted order. 
Whenever a coo_matrix is created (e.g. read from disk or converted from
another format) it should satisfy this expectation.

Currently it returns the number of iterations of Luby's algorithm.  If we 
provided a Luby-specific method it would be sensible to return the number of 
iterations also.

Proposed function signature:

template <typename Matrix1,
          typename Array1,
          typename Matrix2>
size_t induced_subgraph(const Matrix1& G,
                        const Array1& stencil,
                              Matrix2& S);

Assumptions:
 G is an NxN matrix
 stencil is an array with length N whose value_type is convertible to bool
 S can be .resize()-ed to hold the induced subgraph

Semantics:
 Returns K, where K is the number of true values in stencil
 S will be a KxK matrix
 The valid vertices of G will be mapped to [0,K) in order
 For each edge (i,j) in G where (stencil[i] && stencil[j]) is true the edge (M[i],M[j]) will exist in S, where M is the mapping above


Possible extensions:

template <typename Matrix1,
          typename Array1,
          typename Matrix2,
          typename Array2>
size_t induced_subgraph(const Matrix1& G,
                        const Array1& stencil,
                              Matrix2& S,
                              Array2& map);

Here output array map stores the mapping from S vertices to G vertices.  Note 
this is not the mapping mentioned in the Semantics section above.

The header is always "real" regardless of the matrix.
The attached patch checks the matrix type and changes the header appropriately.

Problem:
Trying to use the smoothed_aggregation preconditioner with double-precision 
vectors/matrix doesn't work because there are scalar arguments in _solve 
hardcoded to be float (e.g, 1.0f) which are ruining accuracy.  This can be 
solved by casting to the templated "ValueType" as has been done elsewhere.


What version of the product are you using? On what operating system?
The latest mercurial pull on both Max OS X and Ubuntu

For example

nathan@droog:~/NV/cusp/performance/spmv$ ./spmv
~/NV/research/datasets/matrices/cell_paper/matrix_market/dense2.mtx

Computing SpMV with 'float' values.

There are 2 devices supporting CUDA

Device 0: "Tesla C1060"
  Major revision number:                         1
  Minor revision number:                         3
  Total amount of global memory:                 4294705152 bytes

Device 1: "GeForce 8800 GT"
  Major revision number:                         1
  Minor revision number:                         1
  Total amount of global memory:                 536150016 bytes

Running on Device 0

Read matrix
(/home/nathan/NV/research/datasets/matrices/cell_paper/matrix_market/dense2.mtx)
with shape (2000,2000) and 4000000 entries

    coo_flat            :   1.3561 ms (  5.90 GFLOP/s  47.2 GB/s) [L2 error
0.000000]
    coo_flat_tex        :   1.3534 ms (  5.91 GFLOP/s  47.3 GB/s) [L2 error
0.000000]
    coo_flat_k          :   1.3874 ms (  5.77 GFLOP/s  46.1 GB/s) [L2 error
0.000000]
    coo_flat_k_tex      :   1.2837 ms (  6.23 GFLOP/s  49.9 GB/s) [L2 error
0.000000]
    csr_scalar          :  10.8679 ms (  0.74 GFLOP/s   4.4 GB/s) [L2 error
0.000000]
    csr_scalar_tex      :  10.4112 ms (  0.77 GFLOP/s   4.6 GB/s) [L2 error
0.000000]
    csr_vector          :   0.7620 ms ( 10.50 GFLOP/s  63.0 GB/s) [L2 error
0.000000]
    csr_vector_tex      :   0.5633 ms ( 14.20 GFLOP/s  85.3 GB/s) [L2 error
0.000000]
    dia                 :   2.3368 ms (  3.42 GFLOP/s  13.7 GB/s) [L2 error
1.000000]
    dia_tex             :   2.2275 ms (  3.59 GFLOP/s  14.4 GB/s) [L2 error
1.000000]
    ell                 :   1.9842 ms (  4.03 GFLOP/s  24.2 GB/s) [L2 error
0.000000]
    ell_tex             :   1.9803 ms (  4.04 GFLOP/s  24.2 GB/s) [L2 error
0.000000]
    hyb                 :   1.2869 ms (  6.22 GFLOP/s  49.8 GB/s) [L2 error
0.000000]
    hyb_tex             :   1.2705 ms (  6.30 GFLOP/s  50.4 GB/s) [L2 error
0.000000]

build-env.py copies LD_LIBRARY_PATH in posix systems, but macosx uses 
DYLD_LIBRARY_PATH.
The attached patch checks for this and changes the assignment.

Sparse matrices are quite cheap to transpose, and it can be done in place.
It would be great if the library could provide such a function. 

Specifically in the DIA format, the access to the elements changes,
however, if only the indices of the diagonals are inverted. Thus, the
matrix should "know" whether it is transposed, making this an "intrusive"
change. 

I wrote transposed access before your major restructuring and it worked
without loss of speed, but I'd rather see this in the "main" branch...

In particular, the sparse matrix-vector multiplication routines need to support 
cusp::complex [1].

[1] 
http://groups.google.com/group/cusp-users/browse_thread/thread/6cabe7d6078ff33c

   (Appendix G.2)  Addition and multiplication are often combined
   into a single multiply-add instruction (FMAD), which truncates 
   (i.e. without rounding) the intermediate mantissa of the multiplication;
   Source:
http://developer.download.nvidia.com/compute/cuda/3_0/toolkit/docs/NVIDIA_CUDA_P
rogrammingGuide.pdf

Provide a general transform_elementwise function which implements arbitrary 
binary transformations between sparse matrices.

For example, cusp::blas::dotc needs to conjugate values.

What steps will reproduce the problem?
1. Compile and run test.cu (in attachments)

What is the expected output? What do you see instead?
Sparse matrix multiply is working on the host (for CSR), but not on the device.

What version of the product are you using? On what operating system?
0.1, Fedora 10, kernel : 2.6.27.5-117.fc10.i686.PAE

Please provide any additional information below.
I may be doing something wrong, but I could only get this point from the
documentation available. You can contact me @ pyalama (at) g (dot) clemson
(dot) edu

Hi,

I am using the Conjugate Gradient Solver cg.cu to solve A*x=b.
I changed it a little bit so that it could initialize the vector b from reading 
a matrix market file as well. It works fine with the matrix A in the given 
example. However, when I changed the matrix to my own sparse matrix, it always 
tends to further way (fails to converge). On the other hand, I can get the 
solution when I use Matlab to solve it. Thus, I am wondering if there is any 
specific requirement to the matrix by using Conjugate Gradient method?

What version of the product are you using? On what operating system?

I am using CUDA3.0, cusp0.1 and thrust1.2. The operating system is Windows7.

Thank you

A cusp::print function could either replace cusp::print_matrix just call 
print_matrix using default arguments.

If cusp::print_matrix survives it needs to offer printing options, such as 
dense vs. sparse output.

cusp::print should work for at least all the cusp containers and views.

Possible function signatures:
  template <typename Matrix>
  cusp::print(const Matrix& A);

  template <typename Matrix, typename Stream>
  cusp::print(const Matrix& A, Stream& ostream);

Reported by Jordan on cusp-users

http://groups.google.com/group/cusp-users/browse_thread/thread/f92f73ee66439de7

Hi,

I get the following issues when I try to compile 

1. compile the version.cu returns wrong thrust version: v1.1 instead of v1.2. I 
downloaded thrust from  http://thrust.googlecode.com/files/thrust-v1.2.zip

2. compile of any example results in the following error output which I think 
is related to the wrong thrust version:

:nvcc cg.cu -o cg -I /home/fdelpin/GPU/
/usr/local/cuda/bin/../include/thrust/sorting/detail/device/cuda/stable_merge_so
rt.inl(94): error: expression must have a constant value

/home/fdelpin/GPU/thrust/iterator/permutation_iterator.h(152): error: namespace 
"thrust::detail" has no member "enable_if_convertible"

/home/fdelpin/GPU/thrust/iterator/permutation_iterator.h(152): error: expected 
a ")"

/home/fdelpin/GPU/thrust/iterator/detail/permutation_iterator.inl(31): error: 
namespace "thrust::detail::device" has no member "dereference_result"

/home/fdelpin/GPU/thrust/iterator/detail/permutation_iterator.inl(31): error: 
expected an identifier

/home/fdelpin/GPU/thrust/iterator/detail/permutation_iterator.inl(46): error: a 
template argument list is not allowed in a declaration of a primary template

/home/fdelpin/GPU/thrust/iterator/detail/permutation_iterator.inl(54): error: 
dereference_result is not a template

/home/fdelpin/GPU/thrust/iterator/detail/permutation_iterator.inl(63): error: 
dereference_result is not a template

8 errors detected in the compilation of 
"/tmp/tmpxft_000011ea_00000000-4_cg.cpp1.ii".

I'm using CUSP V0.1 on linux SuSe 11

Thanks!