Set kernel names directly

Author: thomasfaingnaert

test/perf/matmul_kernels/wmma/cutlass-mma-turing.sh

@@ -16,7 +16,7 @@ for i in {7..14}; do
     N=$((2**i))
 
     # runtime in ns
-    runtime=$(nv-nsight-cu-cli -f --summary per-kernel --csv --units base -k Kernel ${CUTLASS_BUILD_PATH}/tools/profiler/cutlass_profiler --op_class=tensorop --A=f16:col --B=f16:col --C=f32 --accum=f32 --m=$N --n=$N --k=$N --inst_m=16 --inst_n=8 --inst_k=8 --warmup-iterations=1 --profiling-iterations=10 --verification-enabled=false 2>/dev/null | grep 'gpu__time_duration' | tail -1 | awk -F',' '{print $NF}' | sed 's/"//g')
+    runtime=$(nv-nsight-cu-cli -f --summary per-kernel --csv --units base -k Kernel ${CUTLASS_BUILD_PATH}/tools/profiler/cutlass_profiler --m=$N --n=$N --k=$N --warmup-iterations=1 --profiling-iterations=10 --verification-enabled=false --kernels=cutlass_tensorop_s1688gemm_f16_128x128_nn 2>/dev/null | grep 'gpu__time_duration' | tail -1 | awk -F',' '{print $NF}' | sed 's/"//g')
 
     printf "$N,$runtime\n" >>cutlass-mma-turing.csv
 done

test/perf/matmul_kernels/wmma/cutlass-mma.sh

@@ -16,7 +16,7 @@ for i in {7..14}; do
     N=$((2**i))
 
     # runtime in ns
-    runtime=$(nv-nsight-cu-cli -f --summary per-kernel --csv --units base -k Kernel ${CUTLASS_BUILD_PATH}/tools/profiler/cutlass_profiler --op_class=tensorop --A=f16:col --B=f16:col --C=f32 --accum=f32 --m=$N --n=$N --k=$N --inst_m=8 --inst_n=8 --inst_k=4 --warmup-iterations=1 --profiling-iterations=10 --verification-enabled=false 2>/dev/null | grep 'gpu__time_duration' | tail -1 | awk -F',' '{print $NF}' | sed 's/"//g')
+    runtime=$(nv-nsight-cu-cli -f --summary per-kernel --csv --units base -k Kernel ${CUTLASS_BUILD_PATH}/tools/profiler/cutlass_profiler --m=$N --n=$N --k=$N --warmup-iterations=1 --profiling-iterations=10 --verification-enabled=false --kernels=cutlass_tensorop_s884gemm_f16_128x128_nn 2>/dev/null | grep 'gpu__time_duration' | tail -1 | awk -F',' '{print $NF}' | sed 's/"//g')
 
     printf "$N,$runtime\n" >>cutlass-mma.csv
 done

test/perf/matmul_kernels/wmma/cutlass-wmma.sh

@@ -16,7 +16,7 @@ for i in {7..14}; do
     N=$((2**i))
 
     # runtime in ns
-    runtime=$(nv-nsight-cu-cli -f --summary per-kernel --csv --units base -k Kernel ${CUTLASS_BUILD_PATH}/tools/profiler/cutlass_profiler --op_class=wmmatensorop --A=f16:col --B=f16:col --C=f32 --accum=f32 --m=$N --n=$N --k=$N --warmup-iterations=1 --profiling-iterations=10 --verification-enabled=false 2>/dev/null | grep 'gpu__time_duration' | tail -1 | awk -F',' '{print $NF}' | sed 's/"//g')
+    runtime=$(nv-nsight-cu-cli -f --summary per-kernel --csv --units base -k Kernel ${CUTLASS_BUILD_PATH}/tools/profiler/cutlass_profiler --m=$N --n=$N --k=$N --warmup-iterations=1 --profiling-iterations=10 --verification-enabled=false --kernels=cutlass_wmma_tensorop_s161616gemm_f16_128x128_nn 2>/dev/null | grep 'gpu__time_duration' | tail -1 | awk -F',' '{print $NF}' | sed 's/"//g')
 
     printf "$N,$runtime\n" >>cutlass-wmma.csv
 done