Add new level-0 macro layer.

.circleci/config.yml

@@ -30,6 +30,7 @@ workflows:
       #     OOT: 0
       #     TEST: FAST
       #     SDE: 0
+      #     LEVEL0: 0
       #     THR: none
       #     CONF: auto
       #     BLD: ''
@@ -41,20 +42,18 @@ workflows:
       - build:
           OOT: 1
           TEST: ALL
-          SDE: 0
           CONF: x86_64
 
       # SDE testing for x86_64
       - build:
           # linuxvm must be used because it provides 8G RAM and SDE fails with 4G RAM
           os: linuxvm
-          OOT: 0
-          TEST: FAST
           SDE: 1
           CONF: x86_64
 
-      # openmp build
+      # openmp build + LEVEL0
       - build:
+          LEVEL0: 1
           THR: openmp
 
       # pthreads build
@@ -153,6 +152,9 @@ jobs:
       SDE:
         type: integer
         default: 0
+      LEVEL0:
+        type: integer
+        default: 0
       THR:
         type: string
         default: none
@@ -197,6 +199,7 @@ jobs:
             export BLD="<< parameters.BLD >>"
             export LDFLAGS="<< parameters.LDFLAGS >>"
             export SDE="<< parameters.SDE >>"
+            export LEVEL0="<< parameters.LEVEL0 >>"
             export THR="<< parameters.THR >>"
             export TESTSUITE_WRAPPER="<< parameters.TESTSUITE_WRAPPER >>"
 
@@ -232,6 +235,7 @@ jobs:
             echo "TEST              = $TEST"
             echo "BLD               = $BLD"
             echo "SDE               = $SDE"
+            echo "LEVEL0            = $LEVEL0"
             echo "DIST_PATH         = $DIST_PATH"
             echo "LDFLAGS           = $LDFLAGS"
             echo "TESTSUITE_WRAPPER = $TESTSUITE_WRAPPER"
@@ -251,3 +255,4 @@ jobs:
             if [ "$CONF" = "armsve" ]; then sed -i 's/.*\<gemmt\>.*/0/' $DIST_PATH/testsuite/input.operations.fast; fi
             if [ "$TEST" != "0" ]; then $DIST_PATH/ci/do_testsuite.sh; fi
             if [ "$SDE" = "1" ]; then $DIST_PATH/ci/do_sde.sh; fi
+            if [ "$LEVEL0" = "1" ]; then $DIST_PATH/ci/do_level0.sh; fi

blastest/src/dblat1.c

@@ -1044,7 +1044,7 @@ static real c_b81 = 0.f;
 
     /* Local variables */
     real sd;
-    extern real s_epsilon_();
+    extern real s_epsilon_(real *);
 
     /* Fortran I/O blocks */
     static cilist io___125 = { 0, 6, 0, fmt_99999, 0 };

build/libblis-symbols.def

@@ -121,7 +121,7 @@ bli_cger_ex
 bli_cgetijm
 bli_cgetijv
 bli_cgetsc
-bli_cgtesc
+bli_cgesc
 bli_cgtsc
 bli_check_error_code_helper
 bli_chemm
@@ -149,7 +149,7 @@ bli_cinvscalv
 bli_cinvscalv_ex
 bli_clock
 bli_clock_min_diff
-bli_cltesc
+bli_clesc
 bli_cltsc
 bli_cmachval
 bli_cmkherm
@@ -351,7 +351,7 @@ bli_dger_ex
 bli_dgetijm
 bli_dgetijv
 bli_dgetsc
-bli_dgtesc
+bli_dgesc
 bli_dgtsc
 bli_dhemm
 bli_dhemm_ex
@@ -377,7 +377,7 @@ bli_dinvscalm_ex
 bli_dinvscalv
 bli_dinvscalv_ex
 bli_divsc
-bli_dltesc
+bli_dlesc
 bli_dltsc
 bli_dmachval
 bli_dmkherm
@@ -532,7 +532,7 @@ bli_gks_l3_ukr_impl_type
 bli_gks_query_cntx
 bli_gks_query_ind_cntx
 bli_gks_query_nat_cntx
-bli_gtesc
+bli_gesc
 bli_gtsc
 bli_hemm
 bli_hemm_ex
@@ -630,7 +630,7 @@ bli_iprintm
 bli_iprintv
 bli_isetsc
 bli_l3_thrinfo_create
-bli_ltesc
+bli_lesc
 bli_ltsc
 bli_machval
 bli_malloc_user
@@ -824,7 +824,7 @@ bli_sger_ex
 bli_sgetijm
 bli_sgetijv
 bli_sgetsc
-bli_sgtesc
+bli_sgesc
 bli_sgtsc
 bli_shemm
 bli_shemm_ex
@@ -852,7 +852,7 @@ bli_sinvscalm_ex
 bli_sinvscalv
 bli_sinvscalv_ex
 bli_sleep
-bli_sltesc
+bli_slesc
 bli_sltsc
 bli_smachval
 bli_smkherm
@@ -1097,7 +1097,7 @@ bli_zger_ex
 bli_zgetijm
 bli_zgetijv
 bli_zgetsc
-bli_zgtesc
+bli_zgesc
 bli_zgtsc
 bli_zhemm
 bli_zhemm_ex
@@ -1123,7 +1123,7 @@ bli_zinvscalm_ex
 bli_zinvscalv
 bli_zinvscalv_ex
 bli_zipsc
-bli_zltesc
+bli_zlesc
 bli_zltsc
 bli_zmachval
 bli_zmkherm

build/plugin/my_kernel_1_ref.c

@@ -48,7 +48,7 @@ void PASTEMAC(ch,opname,arch,suf) \
 \
 	for ( dim_t i = 0; i < n; ++i ) \
 	{ \
-		PASTEMAC(ch,copys)( *a, x[ i ] ); \
+		bli_tcopys( ch,ch, *a, x[ i ] ); \
 	} \
 }
 

build/plugin/my_kernel_2_ref.c

@@ -56,7 +56,7 @@ void PASTEMAC(ch,opname,arch,suf) \
 		{ \
 			for ( dim_t i = 0; i < m; ++i ) \
 			{ \
-				PASTEMAC(ch,seti0s)( a[ i*n + j ] ); \
+				bli_tseti0s( ch, a[ i*n + j ] ); \
 			} \
 		} \
 	} \
@@ -66,7 +66,7 @@ void PASTEMAC(ch,opname,arch,suf) \
 		{ \
 			for ( dim_t j = 0; j < n; ++j ) \
 			{ \
-				PASTEMAC(ch,seti0s)( a[ i + j*m ] ); \
+				bli_tseti0s( ch, a[ i + j*m ] ); \
 			} \
 		} \
 	} \

ci/do_level0.sh

@@ -0,0 +1,16 @@
+#!/bin/bash
+
+set -e
+set -x
+
+: ${SRCDIR:=../..}
+
+if ! [ -d test/level0 ]; then
+    mkdir -p test/level0
+    ln -s $SRCDIR/test/level0/* test/level0/
+fi
+
+cd test/level0
+make -j2
+
+./test_l0.x

common.mk

@@ -63,7 +63,6 @@ $(eval $(call store-var-for,CC,         $(1)))
 $(eval $(call store-var-for,CC_VENDOR,  $(1)))
 $(eval $(call store-var-for,CPPROCFLAGS,$(1)))
 $(eval $(call store-var-for,CLANGFLAGS, $(1)))
-$(eval $(call store-var-for,CXXLANGFLAGS,$(1)))
 $(eval $(call store-var-for,CMISCFLAGS, $(1)))
 $(eval $(call store-var-for,CPICFLAGS,  $(1)))
 $(eval $(call store-var-for,CWARNFLAGS, $(1)))
@@ -109,8 +108,8 @@ get-noopt-cxxflags-for   = $(strip $(CXXFLAGS_PRESET) \
                                    $(call load-var-for,CWARNFLAGS,$(1)) \
                                    $(call load-var-for,CPICFLAGS,$(1)) \
                                    $(call load-var-for,CMISCFLAGS,$(1)) \
-                                   $(call load-var-for,CXXLANGFLAGS,$(1)) \
                                    $(call load-var-for,CPPROCFLAGS,$(1)) \
+                                   $(CXXLANGFLAGS) \
                                    $(CTHREADFLAGS) \
                                    $(CXXTHREADFLAGS) \
                                    $(CINCFLAGS) \
@@ -776,6 +775,11 @@ endif
 
 CWARNFLAGS :=
 
+# Do not allow functions with implicit definitions to be called
+ifneq ($(CC_VENDOR),ibm)
+CWARNFLAGS += -Werror=implicit-function-declaration
+endif
+
 # Disable unused function warnings and stop compiling on first error for
 # all compilers that accept such options: gcc, clang, and icc.
 ifneq ($(CC_VENDOR),ibm)
@@ -914,7 +918,6 @@ endif
 else
 CXXLANGFLAGS :=
 endif
-$(foreach c, $(CONFIG_LIST_FAM), $(eval $(call append-var-for,CXXLANGFLAGS,$(c))))
 
 # --- C Preprocessor flags ---
 

config/template/kernels/1/bli_axpyv_template_noopt_var1.c

@@ -117,7 +117,7 @@ void bli_zaxpyv_template_noopt
 
 	if ( bli_zero_dim1( n ) ) return;
 
-	if ( bli_zeq0( *alpha ) ) return;
+	if ( bli_teq0s( z, *alpha ) ) return;
 
 
 	// If there is anything that would interfere with our use of aligned
@@ -179,7 +179,7 @@ void bli_zaxpyv_template_noopt
 		// Compute front edge cases if x and y were unaligned.
 		for ( i = 0; i < n_pre; ++i )
 		{
-			bli_zaxpys( *alpha, *xp, *yp );
+			bli_taxpys( z,z,z,z, *alpha, *xp, *yp );
 
 			xp += 1; yp += 1;
 		}
@@ -188,7 +188,7 @@ void bli_zaxpyv_template_noopt
 		// yp are guaranteed to be aligned to BLIS_SIMD_ALIGN_SIZE.
 		for ( i = 0; i < n_iter; ++i )
 		{
-			bli_zaxpys( *alpha, *xp, *yp );
+			bli_taxpys( z,z,z,z, *alpha, *xp, *yp );
 
 			xp += n_elem_per_iter;
 			yp += n_elem_per_iter;
@@ -197,7 +197,7 @@ void bli_zaxpyv_template_noopt
 		// Compute tail edge cases, if applicable.
 		for ( i = 0; i < n_left; ++i )
 		{
-			bli_zaxpys( *alpha, *xp, *yp );
+			bli_taxpys( z,z,z,z, *alpha, *xp, *yp );
 
 			xp += 1; yp += 1;
 		}
@@ -207,7 +207,7 @@ void bli_zaxpyv_template_noopt
 		// Compute front edge cases if x and y were unaligned.
 		for ( i = 0; i < n_pre; ++i )
 		{
-			bli_zaxpyjs( *alpha, *xp, *yp );
+			bli_taxpyjs( z,z,z,z, *alpha, *xp, *yp );
 
 			xp += 1; yp += 1;
 		}
@@ -216,7 +216,7 @@ void bli_zaxpyv_template_noopt
 		// yp are guaranteed to be aligned to BLIS_SIMD_ALIGN_SIZE.
 		for ( i = 0; i < n_iter; ++i )
 		{
-			bli_zaxpyjs( *alpha, *xp, *yp );
+			bli_taxpyjs( z,z,z,z, *alpha, *xp, *yp );
 
 			xp += n_elem_per_iter;
 			yp += n_elem_per_iter;
@@ -225,7 +225,7 @@ void bli_zaxpyv_template_noopt
 		// Compute tail edge cases, if applicable.
 		for ( i = 0; i < n_left; ++i )
 		{
-			bli_zaxpyjs( *alpha, *xp, *yp );
+			bli_taxpyjs( z,z,z,z, *alpha, *xp, *yp );
 
 			xp += 1; yp += 1;
 		}

config/template/kernels/1/bli_dotv_template_noopt_var1.c

@@ -127,7 +127,7 @@ void bli_zdotv_template_noopt
 	// If the vector lengths are zero, set rho to zero and return.
 	if ( bli_zero_dim1( n ) )
 	{
-		bli_zset0s( *rho );
+		bli_tset0s( z, *rho );
 		return;
 	}
 
@@ -185,7 +185,7 @@ void bli_zdotv_template_noopt
 
 
 	// Initialize accumulator to zero.
-	bli_zset0s( dotxy );
+	bli_tset0s( z, dotxy );
 
 
 	conjx_use = conjx;
@@ -204,7 +204,7 @@ void bli_zdotv_template_noopt
 		// Compute front edge cases if x and y were unaligned.
 		for ( i = 0; i < n_pre; ++i )
 		{
-			bli_zdots( *xp, *yp, dotxy );
+			bli_tdots( z,z,z,z, *xp, *yp, dotxy );
 
 			xp += 1; yp += 1;
 		}
@@ -213,7 +213,7 @@ void bli_zdotv_template_noopt
 		// yp are guaranteed to be aligned to BLIS_SIMD_ALIGN_SIZE.
 		for ( i = 0; i < n_iter; ++i )
 		{
-			bli_zdots( *xp, *yp, dotxy );
+			bli_tdots( z,z,z,z, *xp, *yp, dotxy );
 
 			xp += n_elem_per_iter;
 			yp += n_elem_per_iter;
@@ -222,7 +222,7 @@ void bli_zdotv_template_noopt
 		// Compute tail edge cases, if applicable.
 		for ( i = 0; i < n_left; ++i )
 		{
-			bli_zdots( *xp, *yp, dotxy );
+			bli_tdots( z,z,z,z, *xp, *yp, dotxy );
 
 			xp += 1; yp += 1;
 		}
@@ -232,7 +232,7 @@ void bli_zdotv_template_noopt
 		// Compute front edge cases if x and y were unaligned.
 		for ( i = 0; i < n_pre; ++i )
 		{
-			bli_zdotjs( *xp, *yp, dotxy );
+			bli_tdotjs( z,z,z,z, *xp, *yp, dotxy );
 
 			xp += 1; yp += 1;
 		}
@@ -241,7 +241,7 @@ void bli_zdotv_template_noopt
 		// yp are guaranteed to be aligned to BLIS_SIMD_ALIGN_SIZE.
 		for ( i = 0; i < n_iter; ++i )
 		{
-			bli_zdotjs( *xp, *yp, dotxy );
+			bli_tdotjs( z,z,z,z, *xp, *yp, dotxy );
 
 			xp += n_elem_per_iter;
 			yp += n_elem_per_iter;
@@ -250,7 +250,7 @@ void bli_zdotv_template_noopt
 		// Compute tail edge cases, if applicable.
 		for ( i = 0; i < n_left; ++i )
 		{
-			bli_zdotjs( *xp, *yp, dotxy );
+			bli_tdotjs( z,z,z,z, *xp, *yp, dotxy );
 
 			xp += 1; yp += 1;
 		}
@@ -259,8 +259,8 @@ void bli_zdotv_template_noopt
 	// If conjugation on y was requested, we induce it by conjugating
 	// the contents of dotxy.
 	if ( bli_is_conj( conjy ) )
-		bli_zconjs( dotxy );
+		bli_tconjs( z, dotxy );
 
-	bli_zcopys( dotxy, *rho );
+	bli_tcopys( z,z, dotxy, *rho );
 }