Merge pull request #1429 from kokkos/half-precision

Minor changes for half precision paper
kokkos · Jun 9, 2022 · 1369cdd · 1369cdd
2 parents 09dfb65 + bae78fb
commit 1369cdd
Show file tree

Hide file tree

Showing 3 changed files with 31 additions and 3 deletions.
diff --git a/perf_test/blas/blas3/KokkosBatched_BatchedGemm_benchmark.sh b/perf_test/blas/blas3/KokkosBatched_BatchedGemm_benchmark.sh
@@ -50,7 +50,7 @@ export KOKKOS_SRC_DIR=$(realpath $KOKKOS_SRC_DIR)
 export KOKKOS_SHA=${KOKKOS_SHA:-"tags/3.6.00"}
 export KOKKOSKERNELS_SRC_DIR=${KOKKOSKERNELS_SRC_DIR:-"$HOME/KOKKOS.base/kokkos-kernels"}
 export KOKKOSKERNELS_SRC_DIR=$(realpath $KOKKOSKERNELS_SRC_DIR)
-export KOKKOSKERNELS_SHA=${KOKKOSKERNELS_SHA:-"half_examples"}
+export KOKKOSKERNELS_SHA=${KOKKOSKERNELS_SHA:-"tags/papers/us-rse-escience-2022"}
 envprint KOKKOS_SRC_DIR KOKKOS_SHA KOKKOSKERNELS_SRC_DIR KOKKOSKERNELS_SHA
 
 dry_run="off"
@@ -208,7 +208,7 @@ echo "cd $benchmark_dir" >> $KOKKOSKERNELS_BUILD_DIR/bench.sh
 echo "$KOKKOSKERNELS_BUILD_DIR/perf_test/blas/blas3/KokkosBlas3_perf_test \
       --test=batched_heuristic --routines=gemm --loop_type=parallel --batch_size_last_dim=0 \
       --matrix_size_start=2x2,2x2,2x2 --matrix_size_stop=64x64,64x64,64x64 \
-      --matrix_size_step=2 --batch_size=1024 \
+      --matrix_size_step=2 --batch_size=$((32*1024)) \
       --warm_up_loop=10 --iter=20 --verify=1 \
       ${use_simd} \
       --csv=${benchmark_dir}/${precision}_bench.csv" \

diff --git a/perf_test/blas/blas3/KokkosBlas3_gemm_perf_test.hpp b/perf_test/blas/blas3/KokkosBlas3_gemm_perf_test.hpp
@@ -263,9 +263,11 @@ static std::string gemm_csv_header_str =
 // Flop count formula from lapack working note 41:
 // http://www.icl.utk.edu/~mgates3/docs/lawn41.pdf
 static inline double __gemm_flop_count(double a_m, double a_n, double b_n) {
+  // TODO: if not Kokkos::complex.
   if (std::is_same<double, default_scalar>::value ||
       std::is_same<float, default_scalar>::value ||
-      std::is_same<Kokkos::Experimental::half_t, default_scalar>::value)
+      std::is_same<Kokkos::Experimental::half_t, default_scalar>::value ||
+      std::is_same<Kokkos::Experimental::bhalf_t, default_scalar>::value)
     return 2 * a_m * b_n * a_n;
   else
     // For complex, we need to count 2 flops for each add and 6 flops for each

diff --git a/perf_test/blas/blas3/papers/kokkos-half-t-rse-escience-2022/reproducer.md b/perf_test/blas/blas3/papers/kokkos-half-t-rse-escience-2022/reproducer.md
@@ -0,0 +1,26 @@
+## To reproduce the half precision results for batched-GEMM:
+```bash
+git clone https://github.com/kokkos/kokkos.git
+git clone https://github.com/kokkos/kokkos-kernels.git
+cd kokkos-kernels
+git checkout tags/papers/us-rse-escience-2022
+cd perf_test/blas/blas3
+export KOKKOS_SRC_DIR=/path/to/kokkos
+export KOKKOSKERNELS_SRC_DIR=/path/to/kokkos-kernels
+```
+
+### On V100
+```bash
+./KokkosBatched_BatchedGemm_benchmark.sh double SNB VOLTA70
+./KokkosBatched_BatchedGemm_benchmark.sh float SNB VOLTA70
+./KokkosBatched_BatchedGemm_benchmark.sh half SNB VOLTA70
+./KokkosBatched_BatchedGemm_benchmark.sh bhalf SNB VOLTA70
+```
+
+### On A100
+```bash
+./KokkosBatched_BatchedGemm_benchmark.sh double DEFAULT AMPERE80
+./KokkosBatched_BatchedGemm_benchmark.sh float DEFAULT AMPERE80
+./KokkosBatched_BatchedGemm_benchmark.sh half DEFAULT AMPERE80
+./KokkosBatched_BatchedGemm_benchmark.sh bhalf DEFAULT AMPERE80
+```