summaryrefslogtreecommitdiff
path: root/src/abstract_hardware_model.cc
diff options
context:
space:
mode:
authorJRPan <[email protected]>2025-09-03 16:01:54 -0400
committerGitHub <[email protected]>2025-09-03 13:01:54 -0700
commit987a23f84f8705c0566a47aeda2132fa33a82618 (patch)
tree67cc731de018ab6573a1dd1926cc005b3e5980db /src/abstract_hardware_model.cc
parent21b18be3832b80b00d4890f041818bac431866e7 (diff)
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
Diffstat (limited to 'src/abstract_hardware_model.cc')
-rw-r--r--src/abstract_hardware_model.cc2
1 files changed, 2 insertions, 0 deletions
diff --git a/src/abstract_hardware_model.cc b/src/abstract_hardware_model.cc
index e8ddf95..7123c24 100644
--- a/src/abstract_hardware_model.cc
+++ b/src/abstract_hardware_model.cc
@@ -782,6 +782,7 @@ kernel_info_t::kernel_info_t(dim3 gridDim, dim3 blockDim,
num_blocks() * entry->gpgpu_ctx->device_runtime->g_TB_launch_latency;
cache_config_set = false;
+ allocated_ctas = 0;
}
/*A snapshot of the texture mappings needs to be stored in the kernel's info as
@@ -815,6 +816,7 @@ kernel_info_t::kernel_info_t(
cache_config_set = false;
m_NameToCudaArray = nameToCudaArray;
m_NameToTextureInfo = nameToTextureInfo;
+ allocated_ctas = 0;
}
kernel_info_t::~kernel_info_t() {