Skip to content
New issue

Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.

Sign up for GitHub

By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.

Already on GitHub? Sign in to your account

Jump to bottom

Performance improvements #67

Open
wants to merge 4 commits into
base: dev
Choose a base branch
from
Open

Performance improvements #67

Show file tree
Hide file tree
Changes from 3 commits
Commits
Show all changes
4 commits
Select commit Hold shift + click to select a range
296be30
performance inprovements
JRPan Feb 21, 2024
ccc6a9b
use node_id before incremented
JRPan Feb 21, 2024
229edda
Cleanup iSLIP
JRPan Feb 28, 2024
8e05bd9
run set_dram_power_stats only when power model enabled
JRPan Jul 12, 2024
File filter

Filter by extension

Filter by extension
Conversations
Failed to load comments.
Loading
Jump to
Jump to file
Failed to load files.
Loading
Diff view
Diff view
8 changes: 6 additions & 2 deletions src/gpgpu-sim/gpu-sim.cc
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -1935,8 +1935,10 @@ void gpgpu_sim::cycle() {
if (mf) partiton_reqs_in_parallel_per_cycle++;
}
m_memory_sub_partition[i]->cache_cycle(gpu_sim_cycle + gpu_tot_sim_cycle);
m_memory_sub_partition[i]->accumulate_L2cache_stats(
m_power_stats->pwr_mem_stat->l2_cache_stats[CURRENT_STAT_IDX]);
if (m_config.g_power_simulation_enabled) {
m_memory_sub_partition[i]->accumulate_L2cache_stats(
m_power_stats->pwr_mem_stat->l2_cache_stats[CURRENT_STAT_IDX]);
}
}
}
partiton_reqs_in_parallel += partiton_reqs_in_parallel_per_cycle;
Expand All @@ -1958,11 +1960,13 @@ void gpgpu_sim::cycle() {
*active_sms += m_cluster[i]->get_n_active_sms();
}
// Update core icnt/cache stats for AccelWattch
if (m_config.g_power_simulation_enabled) {
m_cluster[i]->get_icnt_stats(
m_power_stats->pwr_mem_stat->n_simt_to_mem[CURRENT_STAT_IDX][i],
m_power_stats->pwr_mem_stat->n_mem_to_simt[CURRENT_STAT_IDX][i]);
m_cluster[i]->get_cache_stats(
m_power_stats->pwr_mem_stat->core_cache_stats[CURRENT_STAT_IDX]);
}
m_cluster[i]->get_current_occupancy(
gpu_occupancy.aggregate_warp_slot_filled,
gpu_occupancy.aggregate_theoretical_warp_slots);
Expand Down
52 changes: 28 additions & 24 deletions src/gpgpu-sim/local_interconnect.cc
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -177,25 +177,23 @@ void xbar_router::RR_Advance() {
// IEEE/ACM transactions on networking 2 (1999): 188-201.
// https://www.cs.rutgers.edu/~sn624/552-F18/papers/islip.pdf
void xbar_router::iSLIP_Advance() {
vector<unsigned> node_tmp;
bool active = false;

unsigned conflict_sub = 0;
unsigned reqs = 0;

// calcaulte how many conflicts are there for stats
std::set<unsigned> input_nodes;
std::set<unsigned> destination_set;
for (unsigned i = 0; i < total_nodes; ++i) {
if (!in_buffers[i].empty()) {
Packet _packet_tmp = in_buffers[i].front();
if (!node_tmp.empty()) {
if (std::find(node_tmp.begin(), node_tmp.end(),
_packet_tmp.output_deviceID) != node_tmp.end()) {
conflict_sub++;
} else
node_tmp.push_back(_packet_tmp.output_deviceID);
} else {
node_tmp.push_back(_packet_tmp.output_deviceID);
input_nodes.insert(i);
unsigned out_node = in_buffers[i].front().output_deviceID;

if(destination_set.find(out_node) != destination_set.end()) {
conflict_sub++;
}
destination_set.insert(out_node);
active = true;
}
}
Expand All @@ -206,41 +204,47 @@ void xbar_router::iSLIP_Advance() {
cycles_util++;
}
// do iSLIP
for (unsigned i = 0; i < total_nodes; ++i) {
if (Has_Buffer_Out(i, 1)) {
for (unsigned j = 0; j < total_nodes; ++j) {
unsigned node_id = (j + next_node[i]) % total_nodes;

if (!in_buffers[node_id].empty()) {
for (auto dest : destination_set) {
if (Has_Buffer_Out(dest, 1)) {
unsigned start_node = next_node[dest];
auto it = std::upper_bound(input_nodes.begin(), input_nodes.end(),
start_node);
for (unsigned j = 0; j < input_nodes.size(); j++, it++) {
if (it == input_nodes.end()) {
it = input_nodes.begin();
}
unsigned node_id = *it;
Copy link

@FJShen FJShen Jul 10, 2024 •
edited
Loading

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

I have a concern whether this piece of code is equivalent to the original code.

I understand that std::set<unsigned> destination_set is a subset of the range from 0 to total_nodes. The outer loop iterates over it. So, i is equivalent to dest.

In the original code, the first value which node_id can be assigned is (0 + next_node[i]) % total_nodes.

In your PR, the first value that node_id can get is (1 + next_node[dest]) % total_nodes. This offset of 1 comes from std::upper_bound which yields the first iterator iter that can satisfy bool(comp(start_node, *iter))==true. (citing cppreference.com). comp(start_node, start_node) yields false.

Which one (original vs PR) is correct? Or does this matter?

Copy link
Author

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

The point of this is round-robin. So starting from 0 or 1 should be fine either way. The difference should be minimal.

Copy link

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

Is this specified in the ISLIP paper or not? If it was, better do it right.

Copy link

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

if it wasn't specified in the paper, I'm fine with the PR

assert(!in_buffers[node_id].empty());
Packet _packet = in_buffers[node_id].front();
if (_packet.output_deviceID == i) {
if (_packet.output_deviceID == dest) {
out_buffers[_packet.output_deviceID].push(_packet);
in_buffers[node_id].pop();
input_nodes.erase(node_id); //can only be used once
if (verbose)
printf("%d : cycle %llu : send req from %d to %d\n", m_id, cycles,
node_id, i - _n_shader);
node_id, dest - _n_shader);
if (grant_cycles_count == 1)
next_node[i] = (++node_id % total_nodes);
next_node[dest] = (++node_id % total_nodes);
if (verbose) {
for (unsigned k = j + 1; k < total_nodes; ++k) {
unsigned node_id2 = (k + next_node[i]) % total_nodes;
unsigned node_id2 = (k + next_node[dest]) % total_nodes;
if (!in_buffers[node_id2].empty()) {
Packet _packet2 = in_buffers[node_id2].front();

if (_packet2.output_deviceID == i)
if (_packet2.output_deviceID == dest)
printf("%d : cycle %llu : cannot send req from %d to %d\n",
m_id, cycles, node_id2, i - _n_shader);
m_id, cycles, node_id2, dest - _n_shader);
}
}
}

reqs++;
break;
}
}
}
} else
} else {
out_buffer_full++;
}
}

if (active) {
Expand Down