sync

SNU-ARC · Feb 7, 2024 · 23d4d89 · 23d4d89
1 parent 31397d9
commit 23d4d89
Show file tree

Hide file tree

Showing 3 changed files with 13 additions and 7 deletions.
diff --git a/build/examples/15_ampere_sparse_tensorop_gemm/ncu_run.sh b/build/examples/15_ampere_sparse_tensorop_gemm/ncu_run.sh
@@ -4,4 +4,4 @@ if [ $# != 1 ]; then
   echo "e.g. ./ncu_run.sh output_file_name "
   exit 100
 fi
-sudo /usr/local/cuda/bin/ncu --call-stack --nvtx -o $1 --set full ./15_ampere_sparse_tensorop_gemm
+sudo /usr/local/cuda/bin/ncu --call-stack --nvtx -o $1 --set full ./15_ampere_sparse_tensorop_gemm 
diff --git a/build/examples/15_ampere_sparse_tensorop_gemm/nsys_run.sh b/build/examples/15_ampere_sparse_tensorop_gemm/nsys_run.sh
@@ -4,14 +4,18 @@ if [ $# != 1 ]; then
   echo "e.g. ./nsys_run.sh output_file_name "
   exit 100
 fi
+#for i in 0 32 64 96 128 224 320 416 512 1024 2048 4096 8192
 #for i in 0 32 64 96 128 224 320 416 512
 #for i in 0 32 64 96 128 224 256
 #for i in 0 32 64 96 128
 #for i in 0 32 64 
-for i in 0 32 
+#for i in 0 32 
+for i in 3264
+#for i in 0 
 do
   nsys profile -t cuda,osrt,nvtx,cudnn,cublas -o $1_$i.qdstrm --stats=true -w true ./15_ampere_sparse_tensorop_gemm $1 $i | grep -e SparseGemm -e Gemm -e vecAddOpt -e M: >>$1.log
   echo ////////////////////////////////////////////////////////// >>$1.log
   echo  >>$1.log
 done
+cat $1.log
 #nsys profile -t cuda,osrt,nvtx,cudnn,cublas -o $1.qdstrm --stats=true -w true ./15_ampere_sparse_tensorop_gemm 1024 0 |grep -e SparseGemm -e Gemm -e vecAddOpt -e M:
diff --git a/examples/15_ampere_sparse_tensorop_gemm/ampere_sparse_tensorop_gemm.cu b/examples/15_ampere_sparse_tensorop_gemm/ampere_sparse_tensorop_gemm.cu
@@ -59,16 +59,18 @@ efficiently.
 ///////////////////////////////////////////////
 ///// TEST CONFIGURATION
 ///////////////////////////////////////////////
-#define DENSE_GEMM_EN   1 // 0: disable, 1: enable
+#define DENSE_GEMM_EN   0 // 0: disable, 1: enable
 #define VEC_ADD_EN      1 // 0: disable, 1: enable
 #define REF_EN          2 // 0: disable, 1: host, 2: cutlass
 #define DBG_LOG_EN      0 // 0: disable, 1: enable
 
-#define LIST_ENTRY_NUM  2048 
+#define LIST_ENTRY_NUM  8192
 
-#define M_SIZE          512
-#define K_SIZE          20480
-#define N_SIZE          5120
+#define M_SIZE          8192
+#define K_SIZE          16384
+#define N_SIZE          4096
+//#define K_SIZE          20480
+//#define N_SIZE          5120
 #define M_EXTRA_SIZE    0