def create_dispatch_direct_function(typeflag, mnklist):
print "LIBXSMM_EXTERN_C LIBXSMM_RETARGETABLE libxsmm_" + typeflag + "mm_function libxsmm_" + typeflag + "mm_dispatch(int m, int n, int k)"
print "{"
maxm = libxsmm_utilities.max_mnk(mnklist, 0, 0)
maxn = libxsmm_utilities.max_mnk(mnklist, 0, 1)
maxk = libxsmm_utilities.max_mnk(mnklist, 0, 2)
d, h = maxk + 1, maxn + 1
print " static /*const*/ libxsmm_" + typeflag + "mm_function functions[/*" + str(d * h * (maxm + 1)) + "*/] = {"
sys.stdout.write(" ")
begin, m, n, r = 0, 0, 0, 8
s = r * 6
for mnk in mnklist:
end = calc_direct_index(mnk, d, h)
for i in range(begin, end):
m = m + 1; n = n + 1
if (0 == (m % s)):
sys.stdout.write(",\n ")
elif (1 < n):
sys.stdout.write(", ")
sys.stdout.write("0")
begin, m, n = end + 1, m + r, n + 1
if (r > (m % s)):
sys.stdout.write(",\n ")
m = m + m % s
elif (1 < n):
sys.stdout.write(", ")
sys.stdout.write("libxsmm_" + typeflag + "mm_" + "_".join(map(str, mnk)))
print
print " };"
print " return (" + str(maxm) + " >= m && " + str(maxn) + " >= n && " + str(maxk) + " >= k) " + \
"? functions[" + str(d) + "*(" + str(h) + "*m+n)+k] " + \
": 0;"
print "}"
else: threshold = 0
if (8 < argc): sync = int(sys.argv[8])
else: sync = 0
if (9 < argc): jit = int(sys.argv[9])
else: jit = 0
if (10 < argc): flags = int(sys.argv[10])
else: flags = 0
if (11 < argc): alpha = int(sys.argv[11])
else: alpha = 1
if (12 < argc): beta = int(sys.argv[12])
else: beta = 1
if (13 < argc): mnklist = sorted(libxsmm_utilities.load_mnklist(sys.argv[13:], 0))
else: mnklist = list()
template = Template(open(filename, "r").read())
maxmnk = libxsmm_utilities.max_mnk(mnklist, threshold)
maxdim = int(maxmnk ** (1.0 / 3.0) + 0.5)
avgdim = int(0.5 * maxdim + 0.5)
avgm = libxsmm_utilities.median(list(map(lambda mnk: mnk[0], mnklist)), avgdim, False)
avgn = libxsmm_utilities.median(list(map(lambda mnk: mnk[1], mnklist)), avgdim, False)
avgk = libxsmm_utilities.median(list(map(lambda mnk: mnk[2], mnklist)), avgdim, False)
maxm = libxsmm_utilities.max_mnk(mnklist, avgdim, 0)
maxn = libxsmm_utilities.max_mnk(mnklist, avgdim, 1)
maxk = libxsmm_utilities.max_mnk(mnklist, avgdim, 2)
substitute = { \
"LIBXSMM_OFFLOAD_BUILD": ["", "\n#define LIBXSMM_OFFLOAD_BUILD"][0!=offload], \
"ALIGNMENT": alignment, \
"PREFETCH": [-1, prefetch][0<=prefetch], \
alpha = int(sys.argv[11])
if (12 < argc):
beta = int(sys.argv[12])
if (13 < argc):
wrap = int(sys.argv[13])
if (14 < argc):
mnklist = sorted(libxsmm_utilities.load_mnklist(sys.argv[14:], 0))
version, branch, realversion = \
libxsmm_utilities.version_branch()
major, minor, update, patch = \
libxsmm_utilities.version_numbers(version)
if (0 == threshold):
threshold = 64 * 64 * 64
maxmnk = libxsmm_utilities.max_mnk(mnklist, threshold)
maxdim = int(maxmnk ** (1.0 / 3.0) + 0.5)
avgdim = int(0.5 * maxdim + 0.5)
avgm = libxsmm_utilities.median(
list(map(lambda mnk: mnk[0], mnklist)), avgdim, False)
avgn = libxsmm_utilities.median(
list(map(lambda mnk: mnk[1], mnklist)), avgdim, False)
avgk = libxsmm_utilities.median(
list(map(lambda mnk: mnk[2], mnklist)), avgdim, False)
maxm = libxsmm_utilities.max_mnk(mnklist, avgdim, 0)
maxn = libxsmm_utilities.max_mnk(mnklist, avgdim, 1)
maxk = libxsmm_utilities.max_mnk(mnklist, avgdim, 2)
substitute = {
