From 987a23f84f8705c0566a47aeda2132fa33a82618 Mon Sep 17 00:00:00 2001 From: JRPan <25518778+JRPan@users.noreply.github.com> Date: Wed, 3 Sep 2025 16:01:54 -0400 Subject: Fix correlation on A100 CUTLASS GEMM Kernels (#127) * 1. Add a queue to promote fairness CTA scheduling 2. Add a Perfect xbar interconnect option * remove merge queue * Should not change this * rename pending cta to allocated_ctas * run formatter --- src/gpgpu-sim/local_interconnect.h | 7 ++++++- 1 file changed, 6 insertions(+), 1 deletion(-) (limited to 'src/gpgpu-sim/local_interconnect.h') diff --git a/src/gpgpu-sim/local_interconnect.h b/src/gpgpu-sim/local_interconnect.h index dd10a06..2db7f3c 100644 --- a/src/gpgpu-sim/local_interconnect.h +++ b/src/gpgpu-sim/local_interconnect.h @@ -37,7 +37,11 @@ using namespace std; enum Interconnect_type { REQ_NET = 0, REPLY_NET = 1 }; -enum Arbiteration_type { NAIVE_RR = 0, iSLIP = 1 }; +enum Arbiteration_type { + NAIVE_RR = 0, // + iSLIP, + PERFECT +}; struct inct_config { // config for local interconnect @@ -80,6 +84,7 @@ class xbar_router { private: void iSLIP_Advance(); void RR_Advance(); + void Perfect_Advance(); struct Packet { Packet(void* m_data, unsigned m_output_deviceID) { -- cgit v1.3