_mm_load_ps caused segment fault - c++

Question

I have a code snippet. The snippet just loads 2 arrays and calculates dot product between them using SSE.
Code here:
using namespace std;
long long size = 3200000;
float* _random()
{
unsigned int seed = 123;
// float *t = malloc(size*sizeof(float));
float *t = new float[size];
int i;
float num = 0.0;
for(i=0; i < size; i++) {
num = rand()/(RAND_MAX+1.0);
t[i] = num;
}
return t;
}
float _dotProductVectorSSE(float *s1, float *s2)
{
float prod;
int i;
__m128 X, Y, Z;
for(i=0; i<size; i+=4)
{
X = _mm_load_ps(&s1[i]);
Y = _mm_load_ps(&s2[i]);
X = _mm_mul_ps(X, Y);
Z = _mm_add_ps(X, Z);
}
float *v = new float[4];
_mm_store_ps(v,Z);
for(i=0; i<4; i++)
{
// prod += Z[i];
std::cout << v[i] << endl;
}
return prod;
}
int main(int argc, char *argv[])
{
QCoreApplication a(argc, argv);
time_t start, stop;
double avg_time = 0;
double cur_time;
float* s1 = NULL;
float* s2 = NULL;
for(int i = 0; i < 100; i++)
{
s1 = _random();
s2 = _random();
start = clock();
float sse_product = _dotProductVectorSSE(s1, s2);
stop = clock();
cur_time = ((double) stop-start) / CLOCKS_PER_SEC;
avg_time += cur_time;
}
std::cout << "Averagely used " << avg_time/100 << " seconds." << endl;
return a.exec();
}
When I run, I got segment fault. Here is the backtrace:
(gdb) bt
0 0x0804965f in _mm_load_ps (__P=0xb6b56008) at /usr/lib/gcc/i586-suse-linux/4.6/include/xmmintrin.h:899
1 _dotProductVectorSSE (s1=0xb6b56008, s2=0xb5f20008) at ../simd/simd.cpp:37
2 0x0804987f in main (argc=1, argv=0xbfffee84) at ../simd/simd.cpp:80
Diassembler:
0x8049b30 push %ebp
0x8049b31 <+0x0001> push %edi
0x8049b32 <+0x0002> push %esi
0x8049b33 <+0x0003> push %ebx
0x8049b34 <+0x0004> sub $0x2c,%esp
0x8049b37 <+0x0007> mov 0x804c0a4,%esi
0x8049b3d <+0x000d> mov 0x40(%esp),%edx
0x8049b41 <+0x0011> mov 0x44(%esp),%ecx
0x8049b45 <+0x0015> mov 0x804c0a0,%ebx
0x8049b4b <+0x001b> cmp $0x0,%esi
0x8049b4e <+0x001e> jl 0x8049b7a <_Z20_dotProductVectorSSEPfS_+74>
0x8049b50 <+0x0020> jle 0x8049c10 <_Z20_dotProductVectorSSEPfS_+224>
0x8049b56 <+0x0026> add $0xffffffff,%ebx
0x8049b59 <+0x0029> adc $0xffffffff,%esi
0x8049b5c <+0x002c> xor %eax,%eax
0x8049b5e <+0x002e> shrd $0x2,%esi,%ebx
0x8049b62 <+0x0032> add $0x1,%ebx
0x8049b65 <+0x0035> shl $0x2,%ebx
**0x8049b68 <+0x0038> movaps (%edx,%eax,4),%xmm0**
0x8049b6c <+0x003c> mulps (%ecx,%eax,4),%xmm0
0x8049b70 <+0x0040> add $0x4,%eax
0x8049b73 <+0x0043> cmp %ebx,%eax
0x8049b75 <+0x0045> addps %xmm0,%xmm1
0x8049b78 <+0x0048> jne 0x8049b68 <_Z20_dotProductVectorSSEPfS_+56>
0x8049b7a <+0x004a> movaps %xmm1,0x10(%esp)
0x8049b7f <+0x004f> xor %ebx,%ebx
I am using QtCreator and defined in .pro file:
QMAKE_CXXFLAGS += -msse -msse2
DEFINES += __SSE__
DEFINES += __SSE2__
DEFINES += __MMX__
Please tell me how to fix that problem !

Answer 1

You are not ensuring that your data is 16 byte aligned (malloc/new are not sufficient in general) - you will either need to use _mm_loadu_ps instead of _mm_load_ps to deal with your potentially misaligned data, or preferably use a suitable method to allocate aligned memory (e.g. posix_memalign on Linux).
Note that you should _mm_load_ps and 16 byte aligned memory if you possibly can, otherwise use _mm_loadu_ps but note that this may reduce performance signficantly on some (older) CPUs.

Answer 2

Try the link below.
http://flyeater.wordpress.com/2010/11/29/memory-allocation-and-data-alignment-custom-mallocfree/
You basically allocate a bit more memory than you need, then calculate the address which is modulo 16 and use memory beginning from that address to load/store data.
Take care of pointer arithmetic.
Most of the code here ideone.com/fXKQhR is taken from the above link, sample usage.

Answer 3

I think, the _mm_malloc maybe helpful with you.