def run(self, desired_result, input, limit):
"""
Compile and run a given configuration then
return performance
"""
cfg = desired_result.configuration.data
run_cmd = (
"CHECK=0"
" LIBXSMM_TGEMM_M=" + str(self.granularity * cfg["M"]) +
" LIBXSMM_TGEMM_N=" + str(self.granularity * cfg["N"]) +
" LIBXSMM_TGEMM_K=" + str(self.granularity * cfg["K"]) +
" ./xgemm.sh")
dimset = libxsmm_utilities.load_mnklist(self.args.mnk, 0, -1)
geoperf = 0 # geometric mean
compensation = 0 # see Kahan
for dims in dimset:
run_result = self.call_program(
run_cmd + " " + " ".join(map(str, dims)))
assert(run_result["returncode"] == 0)
match = re.search(
"\\s*LIBXSMM:\\s+([0-9]+(\\.[0-9]*)*)",
run_result["stdout"])
assert(match is not None)
gflops = float(match.group(1))
assert(0 < gflops)
kha = math.log(gflops) - compensation
khb = geoperf + kha
compensation = (khb - geoperf) - kha
geoperf = khb
geoperf = math.exp(geoperf / len(dimset))
geotime = 1000000.0 / geoperf
mnk = (self.granularity**3) * cfg["M"] * cfg["N"] * cfg["K"]
return Result(time=geotime, accuracy=geoperf, size=mnk)
def run(self, desired_result, input, limit):
"""
Compile and run a given configuration then
return performance
"""
cfg = desired_result.configuration.data
run_cmd = ("CHECK=0"
" LIBXSMM_TGEMM_M=" + str(self.granularity * cfg["M"]) +
" LIBXSMM_TGEMM_N=" + str(self.granularity * cfg["N"]) +
" LIBXSMM_TGEMM_K=" + str(self.granularity * cfg["K"]) +
" ./xgemm.sh")
dimset = libxsmm_utilities.load_mnklist(self.args.mnk, 0, -1)
geoperf = 0 # geometric mean
compensation = 0 # see Kahan
for dims in dimset:
run_result = self.call_program(run_cmd + " " +
" ".join(map(str, dims)))
assert (run_result["returncode"] == 0)
match = re.search("\\s*LIBXSMM:\\s+([0-9]+(\\.[0-9]*)*)",
run_result["stdout"])
assert (match is not None)
gflops = float(match.group(1))
assert (0 < gflops)
kha = math.log(gflops) - compensation
khb = geoperf + kha
compensation = (khb - geoperf) - kha
geoperf = khb
geoperf = math.exp(geoperf / len(dimset))
geotime = 1000000.0 / geoperf
mnk = (self.granularity**3) * cfg["M"] * cfg["N"] * cfg["K"]
return Result(time=geotime, accuracy=geoperf, size=mnk)
def save_final_config(self, configuration):
"""called at the end of tuning"""
dimset = libxsmm_utilities.load_mnklist(self.args.mnk, 0, -1)
matrices = ( # collects requested matrix shapes into string
"-".join(map(str, map(lambda mnk: "x".join(map(str, mnk)),
dimset))))
filename = "xgemm-" + matrices + time.strftime(
"-%Y%m%d-%H%M%S") + ".json"
print("Optimal block size written to " + filename + ": ",
configuration.data)
self.manipulator().save_to_file(configuration.data, filename)
def save_final_config(self, configuration):
"""called at the end of tuning"""
dimset = libxsmm_utilities.load_mnklist(self.args.mnk, 0, -1)
matrices = ( # collects requested matrix shapes into string
"-".join(map(str, map(lambda mnk: "x".join(
map(str, mnk)), dimset))))
filename = "xgemm-" + matrices + time.strftime(
"-%Y%m%d-%H%M%S") + ".json"
print("Optimal block size written to " + filename +
": ", configuration.data)
self.manipulator().save_to_file(configuration.data, filename)
def manipulator(self):
"""
Define the search space by creating a
ConfigurationManipulator
"""
self.dimset = libxsmm_utilities.load_mnklist(self.args.mnk, 0, -1)
self.granularity = 1
assert (0 < self.granularity)
m_max = (64 + self.granularity - 1) / self.granularity
n_max = (256 + self.granularity - 1) / self.granularity
k_max = (256 + self.granularity - 1) / self.granularity
m_param = IntegerParameter("M", self.granularity, m_max)
n_param = IntegerParameter("N", self.granularity, n_max)
k_param = IntegerParameter("K", self.granularity, k_max)
manipulator = ConfigurationManipulator()
manipulator.add_parameter(m_param)
manipulator.add_parameter(n_param)
manipulator.add_parameter(k_param)
return manipulator
if (1 < argc):
# required argument(s)
filename = sys.argv[1]
# default configuration if no arguments are given
precision = 0 # all
prefetch = -1 # auto
mnklist = list()
# optional argument(s)
if (2 < argc):
precision = int(sys.argv[2])
if (3 < argc):
prefetch = int(sys.argv[3])
if (4 < argc):
mnklist = sorted(libxsmm_utilities.load_mnklist(sys.argv[4:], 0))
template = Template(open(filename, "r").read())
if (fnmatch.fnmatch(filename, "*.h*")):
optional = [", ...", ""][0 <= prefetch]
substitute = {"MNK_INTERFACE_LIST": ""}
for mnk in mnklist:
mnkstr = "_".join(map(str, mnk))
if (2 != precision):
pfsig = [optional + ");", ",\n "
"const float* pa, "
"const float* pb, "
"const float* pc);"][0 < prefetch]
substitute["MNK_INTERFACE_LIST"] += (
"\nLIBXSMM_API void libxsmm_smm_" + mnkstr +
"(const float* a, const float* b, float* c" +
else: alignment = 64
if (6 < argc): prefetch = int(sys.argv[6])
else: prefetch = 0
if (7 < argc): threshold = int(sys.argv[7])
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)
if 8 < argc:
sync = int(sys.argv[8])
if 9 < argc:
jit = int(sys.argv[9])
if 10 < argc:
flags = int(sys.argv[10])
if 11 < argc:
alpha = int(sys.argv[11])
if 12 < argc:
beta = int(sys.argv[12])
if 13 < argc:
wrap = int(sys.argv[13])
if 14 < argc:
malloc = int(sys.argv[14])
if 15 < argc:
mnklist = sorted(libxsmm_utilities.load_mnklist(sys.argv[15:], 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)
# PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY THEORY OF #
# LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT (INCLUDING #
# NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE OF THIS #
# SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE. #
###############################################################################
# Hans Pabst (Intel Corp.)
###############################################################################
import libxsmm_utilities
import sys
if __name__ == "__main__":
argc = len(sys.argv)
if (3 < argc):
precision = int(sys.argv[1])
threshold = int(sys.argv[2])
mnklist = libxsmm_utilities.load_mnklist(sys.argv[3:], 0)
print("libxsmm_gemm_descriptor desc;")
print("libxsmm_xmmfunction func;")
print("unsigned int hash, indx;")
print("#if defined(_MSC_VER)")
print("# pragma warning(push)")
print("# pragma warning(disable: 4127)")
print("#endif")
for mnk in mnklist:
mstr, nstr, kstr, mnkstr = \
str(mnk[0]), str(mnk[1]), str(mnk[2]), "_".join(map(str, mnk))
mnksig = mstr + ", " + nstr + ", " + kstr
ldxsig = mstr + ", " + kstr + ", " + mstr
# prefer registering double-precision kernels
# when approaching an exhausted registry
## NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE OF THIS ##
## SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE. ##
###############################################################################
## Hans Pabst (Intel Corp.)
###############################################################################
import libxsmm_utilities
import sys
import os
if __name__ == "__main__":
argc = len(sys.argv)
if (3 < argc):
precision = int(sys.argv[1])
threshold = int(sys.argv[2])
mnklist = libxsmm_utilities.load_mnklist(sys.argv[3:], threshold)
print("libxsmm_gemm_descriptor desc;")
print("libxsmm_xmmfunction func;")
print("unsigned int hash, indx;")
for mnk in mnklist:
mstr, nstr, kstr, mnkstr = str(mnk[0]), str(mnk[1]), str(mnk[2]), "_".join(map(str, mnk))
mnksig = "LIBXSMM_LD(" + mstr + ", " + nstr + "), LIBXSMM_LD(" + nstr + ", " + mstr + "), " + kstr
ldxsig = "LIBXSMM_LD(" + mstr + ", " + nstr + "), " + kstr + ", LIBXSMM_LD(" + mstr + ", " + nstr + ")"
# prefer registering double-precision kernels when approaching an exhausted registry
if (1 != precision): # only double-precision
print("LIBXSMM_GEMM_DESCRIPTOR(desc, LIBXSMM_ALIGNMENT, LIBXSMM_FLAGS,")
print(" " + mnksig + ", " + ldxsig + ",")
print(" LIBXSMM_ALPHA, LIBXSMM_BETA, INTERNAL_PREFETCH);")
print("LIBXSMM_HASH_FUNCTION_CALL(hash, indx, desc);")
print("func.dmm = (libxsmm_dmmfunction)libxsmm_dmm_" + mnkstr + ";")
arg1_filename = [sys.argv[1], ""]["0" == sys.argv[1]]
arg1_isfile = os.path.isfile(arg1_filename)
base = 1
if (arg1_isfile):
print("#if !defined(_WIN32)")
print("{ static const char *const build_state =")
print("# include \"" + sys.argv[1] + "\"")
print(" ;")
print(" internal_build_state = build_state;")
print("}")
print("#endif")
base = 2
if ((base + 2) < argc):
precision = int(sys.argv[base+0])
threshold = int(sys.argv[base+1])
mnklist = libxsmm_utilities.load_mnklist(sys.argv[base+2:], 0)
print("/* omit registering code if JIT is enabled"
" and if an ISA extension is found")
print(" * which is beyond the static code"
" path used to compile the library")
print(" */")
print("#if (0 != LIBXSMM_JIT) && !defined(__MIC__)")
print("/* check if target arch. permits execution"
" (arch. may be overridden) */")
print("if (LIBXSMM_STATIC_TARGET_ARCH"
" <= libxsmm_target_archid &&")
print(" (LIBXSMM_X86_SSE3 > libxsmm_target_archid "
"/* JIT code gen. is not available */")
print(" /* condition allows to avoid JIT "
"(if static code is good enough) */")
print(" || LIBXSMM_STATIC_TARGET_ARCH"
if (1 < argc):
# required argument(s)
filename = sys.argv[1]
# optional argument(s)
precision = int(sys.argv[2]) if (2 < argc) else 0
ilp64 = int(sys.argv[3]) if (3 < argc) else 0
alignment = libxsmm_utilities.sanitize_alignment(int(sys.argv[4])) if (4 < argc) else 64
row_major = int(sys.argv[5]) if (5 < argc) else 0
prefetch = int(sys.argv[6]) if (6 < argc) else 0
threshold = int(sys.argv[7]) if (7 < argc) else 0
jit = int(sys.argv[8]) if (8 < argc) else 0
flags = int(sys.argv[9]) if (9 < argc) else 0
alpha = int(sys.argv[10]) if (10 < argc) else 1
beta = int(sys.argv[11]) if (11 < argc) else 1
mnklist = libxsmm_utilities.load_mnklist(sys.argv[12:], threshold) if (12 < argc) else 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)
version, branch = libxsmm_utilities.version_branch()
if __name__ == "__main__":
argc = len(sys.argv)
if (1 < argc):
# required argument(s)
filename = sys.argv[1]
# optional argument(s)
row_major = int(sys.argv[2]) if (2 < argc) else 0
alignment = libxsmm_utilities.sanitize_alignment(int(sys.argv[3])) if (3 < argc) else 64
aligned_stores = libxsmm_utilities.sanitize_alignment(int(sys.argv[4])) if (4 < argc) else 1
aligned_loads = libxsmm_utilities.sanitize_alignment(int(sys.argv[5])) if (5 < argc) else 1
prefetch = int(sys.argv[6]) if (6 < argc) else 0
jit = int(sys.argv[7]) if (7 < argc) else 0
threshold = int(sys.argv[8]) if (8 < argc) else 0
beta = int(sys.argv[9]) if (9 < argc) else 1
mnklist = libxsmm_utilities.load_mnklist(sys.argv[10:], 0, threshold) if (10 < argc) else 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(map(lambda mnk: mnk[0], mnklist), avgdim, False)
avgn = libxsmm_utilities.median(map(lambda mnk: mnk[1], mnklist), avgdim, False)
avgk = libxsmm_utilities.median(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 = { \
## NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE OF THIS ##
## SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE. ##
###############################################################################
## Hans Pabst (Intel Corp.)
###############################################################################
import libxsmm_utilities
import sys
import os
if __name__ == "__main__":
argc = len(sys.argv)
if (3 < argc):
precision = int(sys.argv[1])
threshold = int(sys.argv[2])
mnklist = libxsmm_utilities.load_mnklist(sys.argv[3:], threshold)
print("libxsmm_gemm_descriptor desc;")
print("unsigned int indx;")
for mnk in mnklist:
mstr, nstr, kstr, mnkstr = str(mnk[0]), str(mnk[1]), str(mnk[2]), "_".join(map(str, mnk))
mnksig = "LIBXSMM_LD(" + mstr + ", " + nstr + "), LIBXSMM_LD(" + nstr + ", " + mstr + "), " + kstr
ldxsig = "LIBXSMM_LD(" + mstr + ", " + nstr + "), " + kstr + ", LIBXSMM_LD(" + mstr + ", " + nstr + ")"
if (2 != precision): # only double-precision
print("LIBXSMM_GEMM_DESCRIPTOR(desc, LIBXSMM_ALIGNMENT, LIBXSMM_FLAGS | LIBXSMM_GEMM_FLAG_F32PREC,")
print(" " + mnksig + ", " + ldxsig + ",")
print(" LIBXSMM_ALPHA, LIBXSMM_BETA, LIBXSMM_PREFETCH);")
print("indx = libxsmm_crc32(&desc, LIBXSMM_GEMM_DESCRIPTOR_SIZE, LIBXSMM_HASH_SEED) % (LIBXSMM_CACHESIZE);")
print("if (0 == result[indx].code.xmm) { /* no further effort to handle collision */")
print(" result[indx].code.smm = (libxsmm_smmfunction)libxsmm_smm_" + mnkstr + ";")
print(" result[indx].code_size = 0; /* statically generated code */")
else: alignment = 64
if (6 < argc): prefetch = int(sys.argv[6])
else: prefetch = 0
if (7 < argc): threshold = int(sys.argv[7])
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:], threshold))
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)
else: alignment = 64
if (5 < argc): prefetch = int(sys.argv[5])
else: prefetch = 0
if (6 < argc): threshold = int(sys.argv[6])
else: threshold = 0
if (7 < argc): sync = int(sys.argv[7])
else: sync = 0
if (8 < argc): jit = int(sys.argv[8])
else: jit = 0
if (9 < argc): flags = int(sys.argv[9])
else: flags = 0
if (10 < argc): alpha = int(sys.argv[10])
else: alpha = 1
if (11 < argc): beta = int(sys.argv[11])
else: beta = 1
if (12 < argc): mnklist = sorted(libxsmm_utilities.load_mnklist(sys.argv[12:], 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)