Tweaks to get BLIS compiling again on clarksville.

Details:
- Updated header files and make_defs.mk in config/clarksville.
- Fixes to bl2_mem.c (now that SMEM_M, SMEM_N are gone).
- Moved definition of blksz_t from bl2_cntl.h to bl2_type_defs.h.
- Shuffled include statements in blis2.h.

config/clarksville/bl2_config.h

@@ -46,6 +46,10 @@
   #define BLIS_MEMORY_ALIGNMENT_BOUNDARY 16
 #endif
 
+// Static memory pool size.
+#define BLIS_STATIC_MEM_POOL_SIZE (256 * 5000 * sizeof(double))
+
+
 // Enable mixed domain operations?
 //#define BLIS_ENABLE_MIXED_DOMAIN_SUPPORT
 

config/clarksville/bl2_kernel.h

@@ -57,6 +57,24 @@
 
 
 
+// -- LEVEL-1M KERNEL DEFINITIONS ----------------------------------------------
+
+// -- packm --
+
+#define PACKM_2XK_KERNEL     packm_ref_2xk
+#define PACKM_4XK_KERNEL     packm_ref_4xk
+#define PACKM_6XK_KERNEL     packm_ref_6xk
+#define PACKM_8XK_KERNEL     packm_ref_8xk
+
+// -- unpackm --
+
+#define UNPACKM_2XK_KERNEL   unpackm_ref_2xk
+#define UNPACKM_4XK_KERNEL   unpackm_ref_4xk
+#define UNPACKM_6XK_KERNEL   unpackm_ref_6xk
+#define UNPACKM_8XK_KERNEL   unpackm_ref_8xk
+
+
+
 // -- LEVEL-1F KERNEL DEFINITIONS ----------------------------------------------
 
 #include "bl2_axpy2v_opt_var1.h"

config/clarksville/make_defs.mk

@@ -33,8 +33,8 @@
 #
 
 # Only include this block of code once.
-ifndef CONFIG_MK_INCLUDED
-CONFIG_MK_INCLUDED := yes
+ifndef MAKE_DEFS_MK_INCLUDED
+MAKE_DEFS_MK_INCLUDED := yes
 
 
 
@@ -42,19 +42,10 @@ CONFIG_MK_INCLUDED := yes
 # --- Build definitions --------------------------------------------------------
 #
 
-# This variable will identify the machine architecture (ie: i686, x86_64, etc.)
-# It will be used to construct a string that is appended to the BLIS library
-# name.
-ARCH                            := x86_64
-
-# The install prefix tell us where to install the libraries and header file
-# directory. Notice that we support the use of DESTDIR so that advanced users
-# may install to a temporary location.
-PREFIX                          := $(HOME)/flame
-INSTALL_PREFIX                  := $(DESTDIR)$(PREFIX)
-
 # Variables corresponding to other configure-time options.
 BLIS_ENABLE_VERBOSE_MAKE_OUTPUT := no
+BLIS_ENABLE_STATIC_BUILD        := yes
+BLIS_ENABLE_DYNAMIC_BUILD       := no
 
 
 
@@ -73,6 +64,10 @@ XARGS      := xargs
 RANLIB     := ranlib
 INSTALL    := install -c
 
+# Used to refresh CHANGELOG.
+GIT        := git
+GIT_LOG    := $(GIT) log --decorate
+
 
 
 #
@@ -81,11 +76,11 @@ INSTALL    := install -c
 
 # --- Determine the C compiler and related flags ---
 CC           := gcc
-CPPROCFLAGS  := 
+CPPROCFLAGS  :=
 CMISCFLAGS   := -std=c99 # -fopenmp -pg
-CDBGFLAGS    := # -g
-CWARNFLAGS   := -Wall -Wno-comment
-COPTFLAGS    := -O2
+CDBGFLAGS    := -g
+CWARNFLAGS   := -Wall
+COPTFLAGS    := -O2 #-malign-double
 CVECFLAGS    := -msse3 -march=nocona -mfpmath=sse
 
 # Aggregate all of the flags into two groups: one for optimizable code, and
@@ -103,5 +98,5 @@ LDFLAGS      :=
 
 
 
-# end of ifndef CONFIG_MK_INCLUDED conditional block
+# end of ifndef MAKE_DEFS_MK_INCLUDED conditional block
 endif

frame/base/bl2_mem.c

@@ -34,8 +34,9 @@
 
 #include "blis2.h"
 
+#define N_ELEM_SMEM ( BLIS_STATIC_MEM_POOL_SIZE / sizeof( double ) )
 
-double  smem[ BLIS_STATIC_MEM_POOL_SIZE / sizeof( double ) ];
+double  smem[ N_ELEM_SMEM ];
 
 double* mc      = smem;
 int     counter = 0;
@@ -109,7 +110,7 @@ void* bl2_malloc_s( siz_t buf_size )
 	rmem = ( void* )mc;
 	mc += ( buf_size / sizeof( double ) );
 
-	if ( mc >= smem + ( SMEM_M * SMEM_N ) )
+	if ( mc >= smem + ( N_ELEM_SMEM ) )
 		bl2_abort();
 
 	++counter;
@@ -127,7 +128,7 @@ void bl2_free_s( void* p )
 
 void bl2_mm_clear_smem( void )
 {
-	dim_t n = SMEM_M * SMEM_N;
+	dim_t n = N_ELEM_SMEM;
 	dim_t i;
 
 	for ( i = 0; i < n; ++i )

frame/cntl/bl2_cntl.h

@@ -55,11 +55,6 @@ typedef enum
 	BLIS_VARIANT9,
 } varnum_t;
 
-typedef struct blksz_s
-{
-	dim_t v[BLIS_NUM_FP_TYPES];
-} blksz_t;
-
 
 void bl2_cntl_obj_free( void* cntl );
 

frame/include/bl2_type_defs.h

@@ -71,6 +71,13 @@ typedef struct mem_s
 	dim_t  n;
 } mem_t;
 
+// Blocksize object type
+
+typedef struct blksz_s
+{
+	dim_t v[BLIS_NUM_FP_TYPES];
+} blksz_t;
+
 // Atom type
 // Note: atom types are used to hold "bufferless" scalar object values. Note
 // that it needs to be as large as the largest possible scalar value we might

frame/include/blis2.h

@@ -53,7 +53,6 @@ extern "C" {
 
 #include "bl2_config.h"
 #include "bl2_arch.h"
-#include "bl2_kernel.h"
 
 
 // -- System headers --
@@ -72,11 +71,12 @@ extern "C" {
 #include "bl2_type_defs.h"
 #include "bl2_macro_defs.h"
 #include "bl2_const_defs.h"
-#include "bl2_cntl.h"
-
 #include "bl2_extern_defs.h"
 
-#include "bl2_f2c.h"
+
+// -- BLIS architecture / kernel definitions --
+
+#include "bl2_kernel.h"
 
 
 // -- Base operation prototypes --
@@ -92,8 +92,12 @@ extern "C" {
 #include "bl2_clock.h"
 #include "bl2_check.h"
 #include "bl2_error.h"
+#include "bl2_f2c.h"
 #include "bl2_machval.h"
 
+// Control tree definitions.
+#include "bl2_cntl.h"
+
 
 // -- Level-0 operations --
 

version

@@ -1 +1 @@
-0.0.0-5
+0.0.0-6