loop-unrolling and memory access performance

actually my inner loop has large enough burst read. In the profiler, I see cache hit rate is almost 1 and memory access efficiency is 100%. In another level, I want to have multiple parallel access to improve the bandwidth.

1- First let me ask, if there are multiple accesses in different parts of code to a same global memory variable, like

line 100: X = GlMem;

line 200: y = glmem[j];

does it lead to port replication? or access are done serially through a single port to glmem variable? according to your statement, ports are replicated, right?

2- if so, is it logical to manually unroll the loop, taking care about port replication points? unroll those part i need, and leave the rest rolled? for example, instead of :

# pragma unroll 4

for (unsigned i = 0;i < 4000000; i++)

{

// i,j, acc, s, e are local, rest are global.)

acc = 0.0;

s = start_index;

e = end_index;

for(unsigned j = s;j < e;++j)

acc += value[j]; // target memory access

value_next = acc ;

}

I do this:

for (unsigned i = 0;i < 4000000; i=i+4 ){

// kept rolled

for(unsigned j = 0; j < 4; j++){

acc[j] = 0.0;

s[j] = start_index[i+j];

e[j] = end_index[i+j];

}

// unrolled, I want to improve performance of reading value[j] variable.

for(unsigned j = s[0];j < e[0];++j) // a large burst access

acc[0] += value[j];

for(unsigned j = s[1];j < e[1];++j) // a large burst access

acc[1] += value[j];

for(unsigned j = s[2];j < e[2];++j) // a large burst access

acc[2] += value[j];

for(unsigned j = s[3];j < e[3];++j) // a large burst access

acc[3] += value[j];

// kept rolled

for(unsigned j = 0; j < 4; j++){

value_next[i+j] = acc[j];

}

}

Thanks a lot for your help :)