Update dclapack to do automatic size-of-array'ing

3 files changed, 24 insertions, 16 deletions

diff --git a/redist/dclapack.h b/redist/dclapack.h
index 1dc7b77..af8869c 100644
--- a/redist/dclapack.h
+++ b/redist/dclapack.h
@@ -9,10 +9,14 @@
 
 #define _ABS(a)  ( (a)<=0 ? 0-(a) : (a) )
 
+//Tricky: If you want to use this with pointers, instead of 2D arrays, you will
+//need to #define DYNAMIC_INDEX, as well as, for all arrays, suffix their name
+//with 'c'
+
 #ifdef DYNAMIC_INDEX
-	#define _(A,O,P) A[O*A##c+P]
+	#define _(AM,O,P) AM[O*AM##c+P]
 #else
-	#define _(A,O,P) A[O][P]
+	#define _(AM,O,P) AM[O][P]
 #endif
 
 
diff --git a/redist/dclhelpers.c b/redist/dclhelpers.c
index 9fa9266..c3725f8 100644
--- a/redist/dclhelpers.c
+++ b/redist/dclhelpers.c
@@ -69,34 +69,40 @@ void dcldgemm(
 	int k,
 	DCL_FLOAT alpha,
 	const DCL_FLOAT* A,
-	int lda, //must be n
+	int Ac, //must be n
 	const DCL_FLOAT* B,
-	int ldb, //must be m
+	int Bc, //must be m
 	DCL_FLOAT beta,
-	const DCL_FLOAT * C,
-	int ldc //must be n
+	DCL_FLOAT * C,
+	int Cc //must be n
 	 )
 {
 	const DCL_FLOAT * ta;
 	const DCL_FLOAT * tb;
+	int tac = Ac;
+	int tbc = Bc;
 	if( transA )
 	{
-		ta = alloca( sizeof( DCL_FLOAT ) * n * m );
-		//TRANSP(	ta, A, n, m );
+		DCL_FLOAT * la = alloca( sizeof( DCL_FLOAT ) * n * m );
+		const int lac = m;
+		TRANSP(	la, A, n, m );
+		ta = la;
+		tac = lac;
 	}
 	else
 		ta = A;
 
 	if( transB )
 	{
-		tb = alloca( sizeof( DCL_FLOAT ) * n * m );
-		//TRANSP(	tb, B, n, m );
+		DCL_FLOAT * lb = alloca( sizeof( DCL_FLOAT ) * n * m );
+		const int lbc = m;
+		TRANSP( lb, B, n, m );
+		tb = lb;
+		tbc = lbc;
 	}
 	else
 		tb = B;
 
-
-	//XXX Tricky: In here, A is mxn
-
+	GMULADD(C,ta,tb,C,alpha,beta,n,m,k);
 }
 
diff --git a/redist/dclhelpers.h b/redist/dclhelpers.h
index 5010e02..8779e1a 100644
--- a/redist/dclhelpers.h
+++ b/redist/dclhelpers.h
@@ -3,8 +3,6 @@
 
 #define DCL_FLOAT FLT
 
-//XXX XXX XXX WARNING XXX XXX XXX The argument order may be changing!!!
-
 /* Prints matrix A of size[n][m] */
 void dclPrint( const DCL_FLOAT * A, int Ac, int n, int m );
 
@@ -75,7 +73,7 @@ void dcldgemm(
 	const DCL_FLOAT* B,
 	int Bc,
 	DCL_FLOAT beta,
-	const DCL_FLOAT * C,
+	DCL_FLOAT * C,
 	int Cc
 	 );