more auto-tune, fix reproducibility issue in test

lib/gpt/core/auto_tune.py

@@ -46,18 +46,26 @@ def wrapper(self, *args):
         if self.at_tuned_params is not None:
             return method(self, self.at_tuned_params["params"], *args)
 
+        # create a snapshot of parameters to restore
+        args = list(args)
+        args_snapshot = g.copy(args)
+
         # do experiments
         dt_warmup = -g.time()
         g.message(f"Auto-tune {self.at_tag} warmup")
         method(self, self.at_default_param, *args)
         dt_warmup += g.time()
 
+        g.copy(args, args_snapshot)
+
         dts = []
         for p in self.at_params:
             dt = -g.time()
             g.message(f"Auto-tune {self.at_tag} with {p}")
             method(self, p, *args)
             dt += g.time()
+
+            g.copy(args, args_snapshot)
             dts.append(dt)
 
         g.message(f"Tuning result for {self.at_tag}:")

lib/gpt/core/local_stencil/matrix.py

@@ -17,6 +17,12 @@
 #    51 Franklin Street, Fifth Floor, Boston, MA 02110-1301 USA.
 #
 import cgpt
+from gpt.core import auto_tuned_class, auto_tuned_method
+import hashlib
+
+
+def hash_code(code):
+    return str(len(code)) + "-" + str(hashlib.sha256(str(code).encode("utf-8")).hexdigest())
 
 
 def parse(c):
@@ -26,7 +32,7 @@ def parse(c):
     return c
 
 
-class matrix:
+class matrix(auto_tuned_class):
     def __init__(self, lat, points, code, code_parallel_block_size=None, local=1):
         self.points = points
         self.code = [parse(c) for c in code]
@@ -36,10 +42,14 @@ def __init__(self, lat, points, code, code_parallel_block_size=None, local=1):
         self.obj = cgpt.stencil_matrix_create(
             lat.v_obj[0], lat.grid.obj, points, self.code, code_parallel_block_size, local
         )
-        self.fast_osites = 0
 
-    def __call__(self, *fields):
-        cgpt.stencil_matrix_execute(self.obj, list(fields), self.fast_osites)
+        # auto tuner
+        tag = f"local_matrix({lat.otype.__name__}, {lat.grid.describe()}, {str(points)}, {code_parallel_block_size}, {hash_code(code)}, {local})"
+        super().__init__(tag, [0, 1], 0)
+
+    @auto_tuned_method
+    def __call__(self, fast_osites, *fields):
+        cgpt.stencil_matrix_execute(self.obj, list(fields), fast_osites)
 
     def __del__(self):
         cgpt.stencil_matrix_delete(self.obj)

lib/gpt/core/object_type/su_n.py

@@ -182,6 +182,9 @@ def _convert(dst, src):
         self.ctab = {f"ot_matrix_su_n_fundamental_group({Nc})": _convert}
         self.CA = Nc
 
+    def project(self, A, method):
+        A @= gpt.qcd.gauge.project.traceless_hermitian(A)
+
     def generators(self, dt):
         r = []
 

lib/gpt/qcd/scalar/action/mass_term.py

@@ -154,7 +154,7 @@ def draw(self, pi, rng):
 
         pi_prime = self.inv_sqrt_M(pi)
         for mu in range(n):
-            pi[mu] @= pi_prime[mu]
+            pi[mu] @= g.project(pi_prime[mu], "defect")
 
         return value
 
@@ -172,7 +172,7 @@ def gradient(self, pi, dpi):
 
             ret = pi_prime[mu]
 
-            ret = g(g.qcd.gauge.project.traceless_hermitian(ret))
+            ret = g(g.project(ret, "defect"))
             dS.append(ret)
 
         return dS

tests/qcd/scalar.py

@@ -119,7 +119,7 @@
 A1 = g.qcd.scalar.action.fourier_mass_term(sqrt_mass)
 A1.assert_gradient_error(rng, U_mom, U_mom, 1e-3, 1e-8)
 r = A1.draw(U_mom, rng)
-assert abs(A1(U_mom)/r - 1) < 1e-10
+assert abs(A1(U_mom) / r - 1) < 1e-10
 
 # group defect would be triggered if sqrt_mass does not have sqrt_mass[k] = sqrt_mass[-k]
 eps = g.group.defect(U_mom[0])
@@ -165,6 +165,7 @@ def __slap(dst, src):
 
 tmp = g.copy(U_mom)
 tmp2 = g.copy(U_mom)
+
 lap(tmp, U_mom)
 __slap(tmp2, U_mom)
 
@@ -174,7 +175,7 @@ def __slap(dst, src):
     assert eps < 1e-10
 
 cg = g.algorithms.inverter.block_cg({"eps": 1e-12, "maxiter": 100})
-slap = g.matrix_operator(mat=lap, inv_mat=cg(lap), accept_list=True)
+slap = g.matrix_operator(mat=lap, inv_mat=cg(lap), accept_list=True, accept_guess=(False, True))
 slap2 = slap * slap
 
 A2 = g.qcd.scalar.action.general_mass_term(M=slap2, sqrt_M=slap)