* Experimental support for BLAS for matrix multiplication.

Not enabled by default, yet.

math/linalg/__init__.py

@@ -19,15 +19,58 @@ import numpy.linalg
 from numpy import dot, trace
 from numpy.linalg import det, inv
 
+MATMUL_USE_BLAS = False
 
-def matmul(*Mats):
+def matmul(*Mats, **opts):
   """Do successive matrix product. For example,
      matmul(A,B,C,D)
   will evaluate a matrix multiplication ((A*B)*C)*D .
   The matrices must be of matching sizes."""
-  p = numpy.dot(Mats[0], Mats[1])
-  for M in Mats[2:]:
-    p = numpy.dot(p, M)
+  from numpy import asarray, dot, iscomplexobj
+  use_blas = opts.get('use_blas', MATMUL_USE_BLAS)
+  debug = opts.get('debug', True)
+  if debug:
+    def dbg(msg):
+      print msg,
+  else:
+    def dbg(msg):
+      pass
+  if use_blas:
+    try:
+      from scipy.linalg.blas import zgemm, dgemm
+    except:
+      # Older scipy (<= 0.10?)
+      from scipy.linalg.blas import fblas
+      zgemm = fblas.zgemm
+      dgemm = fblas.dgemm
+
+  if not use_blas:
+    p = dot(Mats[0], Mats[1])
+    for M in Mats[2:]:
+      p = dot(p, M)
+  else:
+    dbg("Using BLAS\n")
+    # FIXME: Right now only supporting double precision arithmetic.
+    M0 = asarray(Mats[0])
+    M1 = asarray(Mats[1])
+    if iscomplexobj(M0) or iscomplexobj(M1):
+      p = zgemm(alpha=1.0, a=M0, b=M1)
+      Cplx = True
+      dbg("- zgemm ")
+    else:
+      p = dgemm(alpha=1.0, a=M0, b=M1)
+      Cplx = False
+      dbg("- dgemm ")
+    for M in Mats[2:]:
+      M2 = asarray(M)
+      if Cplx or iscomplexobj(M2):
+        p = zgemm(alpha=1.0, a=p, b=M2)
+        Cplx = True
+        dbg(" zgemm")
+      else:
+        p = dgemm(alpha=1.0, a=p, b=M2)
+        dbg(" dgemm")
+    dbg("\n")
   return p