diff --git a/test/3m4m/Makefile b/test/3m4m/Makefile index bfe55935e..ecc6a84ef 100644 --- a/test/3m4m/Makefile +++ b/test/3m4m/Makefile @@ -221,9 +221,6 @@ STR_NAT := -DSTR=\"asm\" STR_OBL := -DSTR=\"openblas\" STR_MKL := -DSTR=\"mkl\" STR_ACML := -DSTR=\"acml\" -STR_AMK := -DSTR=\"amk\" -STR_AKN := -DSTR=\"akn\" -STR_AMN := -DSTR=\"amn\" # Single or multithreaded string STR_ST := -DTHR_STR=\"st\" @@ -238,10 +235,6 @@ PDEF_MT := -DP_BEGIN=400 \ -DP_END=8000 \ -DP_INC=400 -ADJ_MK := -DADJ_MK -ADJ_KN := -DADJ_KN -ADJ_MN := -DADJ_MN - # @@ -281,14 +274,7 @@ blis: \ test_cgemm_4m1a_blis_st.x \ test_zgemm_4m1a_blis_st.x \ test_cgemm_asm_blis_st.x \ - test_zgemm_asm_blis_st.x \ - \ - test_cgemm_amk_blis_st.x \ - test_zgemm_amk_blis_st.x \ - test_cgemm_akn_blis_st.x \ - test_zgemm_akn_blis_st.x \ - test_cgemm_amn_blis_st.x \ - test_zgemm_amn_blis_st.x + test_zgemm_asm_blis_st.x blis-mt: \ test_sgemm_asm_blis_mt.x \ @@ -312,6 +298,8 @@ blis-mt: \ test_zgemm_asm_blis_mt.x openblas: \ + test_sgemm_openblas_st.x \ + test_dgemm_openblas_st.x \ test_cgemm_openblas_st.x \ test_zgemm_openblas_st.x @@ -320,6 +308,8 @@ openblas-mt: \ test_zgemm_openblas_mt.x mkl: \ + test_sgemm_mkl_st.x \ + test_dgemm_mkl_st.x \ test_cgemm_mkl_st.x \ test_zgemm_mkl_st.x @@ -328,6 +318,8 @@ mkl-mt: \ test_zgemm_mkl_mt.x acml: \ + test_sgemm_acml_st.x \ + test_dgemm_acml_st.x \ test_cgemm_acml_st.x \ test_zgemm_acml_st.x @@ -458,26 +450,13 @@ test_z%_asm_blis_mt.o: test_%.c test_c%_asm_blis_mt.o: test_%.c $(CC) $(CFLAGS) $(PDEF_MT) $(DT_C) $(BLI_DEF) $(DNAT) $(STR_NAT) $(STR_MT) -c $< -o $@ -# blis asm (adjusted) -test_z%_amk_blis_st.o: test_%.c - $(CC) $(CFLAGS) $(PDEF_ST) $(DT_Z) $(BLI_DEF) $(DNAT) $(STR_AMK) $(STR_ST) $(ADJ_MK) -c $< -o $@ - -test_c%_amk_blis_st.o: test_%.c - $(CC) $(CFLAGS) $(PDEF_ST) $(DT_C) $(BLI_DEF) $(DNAT) $(STR_AMK) $(STR_ST) $(ADJ_MK) -c $< -o $@ - -test_z%_akn_blis_st.o: test_%.c - $(CC) $(CFLAGS) $(PDEF_ST) $(DT_Z) $(BLI_DEF) $(DNAT) $(STR_AKN) $(STR_ST) $(ADJ_KN) -c $< -o $@ - -test_c%_akn_blis_st.o: test_%.c - $(CC) $(CFLAGS) $(PDEF_ST) $(DT_C) $(BLI_DEF) $(DNAT) $(STR_AKN) $(STR_ST) $(ADJ_KN) -c $< -o $@ - -test_z%_amn_blis_st.o: test_%.c - $(CC) $(CFLAGS) $(PDEF_ST) $(DT_Z) $(BLI_DEF) $(DNAT) $(STR_AMN) $(STR_ST) $(ADJ_MN) -c $< -o $@ - -test_c%_amn_blis_st.o: test_%.c - $(CC) $(CFLAGS) $(PDEF_ST) $(DT_C) $(BLI_DEF) $(DNAT) $(STR_AMN) $(STR_ST) $(ADJ_MN) -c $< -o $@ - # openblas +test_d%_openblas_st.o: test_%.c + $(CC) $(CFLAGS) $(PDEF_ST) $(DT_D) $(BLA_DEF) $(DNAT) $(STR_OBL) $(STR_ST) -c $< -o $@ + +test_s%_openblas_st.o: test_%.c + $(CC) $(CFLAGS) $(PDEF_ST) $(DT_S) $(BLA_DEF) $(DNAT) $(STR_OBL) $(STR_ST) -c $< -o $@ + test_z%_openblas_st.o: test_%.c $(CC) $(CFLAGS) $(PDEF_ST) $(DT_Z) $(BLA_DEF) $(DNAT) $(STR_OBL) $(STR_ST) -c $< -o $@ @@ -491,6 +470,12 @@ test_c%_openblas_mt.o: test_%.c $(CC) $(CFLAGS) $(PDEF_MT) $(DT_C) $(BLA_DEF) $(DNAT) $(STR_OBL) $(STR_MT) -c $< -o $@ # mkl +test_d%_mkl_st.o: test_%.c + $(CC) $(CFLAGS) $(PDEF_ST) $(DT_D) $(BLA_DEF) $(DNAT) $(STR_MKL) $(STR_ST) -c $< -o $@ + +test_s%_mkl_st.o: test_%.c + $(CC) $(CFLAGS) $(PDEF_ST) $(DT_S) $(BLA_DEF) $(DNAT) $(STR_MKL) $(STR_ST) -c $< -o $@ + test_z%_mkl_st.o: test_%.c $(CC) $(CFLAGS) $(PDEF_ST) $(DT_Z) $(BLA_DEF) $(DNAT) $(STR_MKL) $(STR_ST) -c $< -o $@ @@ -504,6 +489,12 @@ test_c%_mkl_mt.o: test_%.c $(CC) $(CFLAGS) $(PDEF_MT) $(DT_C) $(BLA_DEF) $(DNAT) $(STR_MKL) $(STR_MT) -c $< -o $@ # acml +test_d%_acml_st.o: test_%.c + $(CC) $(CFLAGS) $(PDEF_ST) $(DT_D) $(BLA_DEF) $(DNAT) $(STR_ACML) $(STR_ST) -c $< -o $@ + +test_s%_acml_st.o: test_%.c + $(CC) $(CFLAGS) $(PDEF_ST) $(DT_S) $(BLA_DEF) $(DNAT) $(STR_ACML) $(STR_ST) -c $< -o $@ + test_z%_acml_st.o: test_%.c $(CC) $(CFLAGS) $(PDEF_ST) $(DT_Z) $(BLA_DEF) $(DNAT) $(STR_ACML) $(STR_ST) -c $< -o $@ diff --git a/test/3m4m/runme.sh b/test/3m4m/runme.sh index 2333b0367..9bd11f02a 100755 --- a/test/3m4m/runme.sh +++ b/test/3m4m/runme.sh @@ -4,8 +4,8 @@ exec_root="test" out_root="output" -sys="stampede" -#sys="wahlberg" +#sys="stampede" +sys="wahlberg" # Bind threads to processors. #export OMP_PROC_BIND=true diff --git a/test/3m4m/test_gemm.c b/test/3m4m/test_gemm.c index afc2209bf..0a45266d9 100644 --- a/test/3m4m/test_gemm.c +++ b/test/3m4m/test_gemm.c @@ -49,7 +49,7 @@ int main( int argc, char** argv ) dim_t p; dim_t p_begin, p_end, p_inc; int m_input, n_input, k_input; - num_t dt, dt_real; + num_t dt; char dt_ch; int r, n_repeats; trans_t transa; @@ -61,7 +61,6 @@ int main( int argc, char** argv ) double dtime_save; double gflops; - extern blksz_t* gemm_kc; bli_init(); @@ -70,7 +69,6 @@ int main( int argc, char** argv ) n_repeats = 3; dt = DT; - dt_real = bli_datatype_proj_to_real( DT ); p_begin = P_BEGIN; p_end = P_END; @@ -80,6 +78,11 @@ int main( int argc, char** argv ) n_input = -1; k_input = -1; +#if 0 + extern blksz_t* gemm_kc; + + num_t dt_real = bli_datatype_proj_to_real( DT ); + // Extract the kc blocksize for the requested datatype and its // real analogue. dim_t kc = bli_blksz_get_def( dt, gemm_kc ); @@ -92,14 +95,6 @@ int main( int argc, char** argv ) else if ( IND == BLIS_3M1 ) k_input = kc_real / 3; else if ( IND == BLIS_4M1A ) k_input = kc_real / 2; else k_input = kc_real; - - // Adjust the relative dimensions, if requested. -#if (defined ADJ_MK) - m_input = -2; k_input = -2; n_input = -1; -#elif (defined ADJ_KN) - k_input = -2; n_input = -2; m_input = -1; -#elif (defined ADJ_MN) - m_input = -2; n_input = -2; k_input = -1; #endif // Choose the char corresponding to the requested datatype. @@ -131,6 +126,7 @@ int main( int argc, char** argv ) for ( p = p_begin; p <= p_end; p += p_inc ) { + if ( m_input < 0 ) m = p / ( dim_t )abs(m_input); else m = ( dim_t ) m_input; if ( n_input < 0 ) n = p / ( dim_t )abs(n_input);