The bench2 from silver-ymz

bench2's Introduction

benchmark on Neoverse-V1 (Graviton 3)

CFLAGS=-Wall -O3 -march=armv8-a+sve:

❯ ./main 300 10000
v_sparse_dot ------ Value: 3.028991, Time: 907.99ns/iter
v_sparse_dot_omp -- Value: 3.028991, Time: 1215.47ns/iter
v_sparse_dot_sve -- Value: 3.028991, Time: 1110.44ns/iter

CFLAGS=-Wall -O3 -mcpu=neoverse-v1:

❯ ./main 300 10000
v_sparse_dot ------ Value: 4.321673, Time: 1696.15ns/iter
v_sparse_dot_omp -- Value: 4.321673, Time: 1218.40ns/iter
v_sparse_dot_sve -- Value: 4.321673, Time: 1085.07ns/iter

bench2's People

Contributors

Watchers

bench2's Issues

The makefile is missing the -fopenmp C flag and thus the related pragmas are ignored by the compiler.

I played with the code here since these was a concern with -mcpu=neoverse-v1 versus -march=armv8-a+sve flags.
With GCC-11 two of the tests are faster than with clang with either cpu/arch flag:

v_sparse_dot ------ Value: 2.551899, Time: 494.75ns/iter
v_sparse_dot_omp -- Value: 2.551900, Time: 2318.14ns/iter
v_sparse_dot_sve -- Value: 2.551899, Time: 951.34ns/iter

It appears that for the function v_sparse_dot_omp the #pragma omp directives are ignored when using the Makefile as-is.
When -fopenmp is given to enable the omp options the build fails:

clang-15 -Wall -O3 -mcpu=neoverse-v1 -fopenmp -c -o lib.o lib.c
lib.c:35:13: error: expected an OpenMP directive
#pragma omp reduction(| : m1) reduction(| : m2)
            ^
lib.c:33:3: error: statement after '#pragma omp parallel for' must be a for loop
  while (lhs_pos < lhs_loop_len && rhs_pos < rhs_loop_len) {
  ^
2 errors generated.
make: *** [Makefile:11: lib.o] Error 1

The same is true with -march=armv8-a+sve.

Recommend Projects