tune

lehner · Sep 26, 2024 · 833cb2c · 833cb2c
1 parent 4a473d5
commit 833cb2c
Show file tree

Hide file tree

Showing 3 changed files with 26 additions and 26 deletions.
diff --git a/lib/gpt/core/foundation/lattice/__init__.py b/lib/gpt/core/foundation/lattice/__init__.py
@@ -29,7 +29,22 @@ def rank_inner_product(a, b, use_accelerator):
 
 
 def inner_product(a, b, use_accelerator):
-    return a[0].grid.globalsum(rank_inner_product(a, b, use_accelerator))
+    t = gpt.timer("inner_product")
+    gpt.barrier()
+    t("rank inner product")
+    rip = rank_inner_product(a, b, use_accelerator)
+    t("barrier")
+    gpt.barrier()
+    t("global sum")
+    rip2 = numpy.array(rip)
+    gs = a[0].grid.globalsum(rip)
+    t("barrier")
+    gpt.barrier()
+    t("reduce")
+    gs2 = a[0].grid.reduce(rip2, lambda a, b: a.__iadd__(b))
+    t()
+    gpt.message(t, gs, gs2)
+    return gs
 
 
 def norm2(l):

diff --git a/lib/gpt/core/grid.py b/lib/gpt/core/grid.py
@@ -42,9 +42,10 @@ def global_sum_default(grid, x):
         if isinstance(x, np.ndarray):
             return grid.reduce(x, lambda a, b: a.__iadd__(b))
         elif gpt.util.is_num(x):
+            t = type(x)
             x = np.array([x], dtype=np.complex128)
             global_sum_default(grid, x)
-            return gpt.util.to_num(x[0])
+            return t(x[0])
         else:
             raise Exception(f"Unknown data type in global sum: {type(x)}")
 

diff --git a/lib/gpt/qcd/fermion/register.py b/lib/gpt/qcd/fermion/register.py
@@ -11,40 +11,24 @@ def register(reg, op):
     reg.Mdiag = lambda dst, src: op.apply_unary_operator(2009, dst, src)
     reg.Dminus = lambda dst, src: op.apply_unary_operator(2010, dst, src)
     reg.DminusDag = lambda dst, src: op.apply_unary_operator(2011, dst, src)
-    reg.ImportPhysicalFermionSource = lambda dst, src: op.apply_unary_operator(
-        2012, dst, src
-    )
-    reg.ImportUnphysicalFermion = lambda dst, src: op.apply_unary_operator(
-        2013, dst, src
-    )
-    reg.ExportPhysicalFermionSolution = lambda dst, src: op.apply_unary_operator(
-        2014, dst, src
-    )
-    reg.ExportPhysicalFermionSource = lambda dst, src: op.apply_unary_operator(
-        2015, dst, src
-    )
+    reg.ImportPhysicalFermionSource = lambda dst, src: op.apply_unary_operator(2012, dst, src)
+    reg.ImportUnphysicalFermion = lambda dst, src: op.apply_unary_operator(2013, dst, src)
+    reg.ExportPhysicalFermionSolution = lambda dst, src: op.apply_unary_operator(2014, dst, src)
+    reg.ExportPhysicalFermionSource = lambda dst, src: op.apply_unary_operator(2015, dst, src)
     reg.Dhop = lambda dst, src: op.apply_unary_operator(3001, dst, src)
     reg.DhopDag = lambda dst, src: op.apply_unary_operator(4001, dst, src)
     reg.DhopEO = lambda dst, src: op.apply_unary_operator(3002, dst, src)
     reg.DhopEODag = lambda dst, src: op.apply_unary_operator(4002, dst, src)
-    reg.Mdir = lambda dst, src, dir, disp: op.apply_dirdisp_operator(
-        5001, dst, src, dir, disp
-    )
+    reg.Mdir = lambda dst, src, dir, disp: op.apply_dirdisp_operator(5001, dst, src, dir, disp)
     reg.MDeriv = lambda mat, dst, src: op.apply_deriv_operator(6001, mat, dst, src)
     reg.MDerivDag = lambda mat, dst, src: op.apply_deriv_operator(7001, mat, dst, src)
     reg.MoeDeriv = lambda mat, dst, src: op.apply_deriv_operator(6002, mat, dst, src)
     reg.MoeDerivDag = lambda mat, dst, src: op.apply_deriv_operator(7002, mat, dst, src)
     reg.MeoDeriv = lambda mat, dst, src: op.apply_deriv_operator(6003, mat, dst, src)
     reg.MeoDerivDag = lambda mat, dst, src: op.apply_deriv_operator(7003, mat, dst, src)
     reg.DhopDeriv = lambda mat, dst, src: op.apply_deriv_operator(6004, mat, dst, src)
-    reg.DhopDerivDag = lambda mat, dst, src: op.apply_deriv_operator(
-        7004, mat, dst, src
-    )
+    reg.DhopDerivDag = lambda mat, dst, src: op.apply_deriv_operator(7004, mat, dst, src)
     reg.DhopDerivEO = lambda mat, dst, src: op.apply_deriv_operator(6005, mat, dst, src)
-    reg.DhopDerivEODag = lambda mat, dst, src: op.apply_deriv_operator(
-        7005, mat, dst, src
-    )
+    reg.DhopDerivEODag = lambda mat, dst, src: op.apply_deriv_operator(7005, mat, dst, src)
     reg.DhopDerivOE = lambda mat, dst, src: op.apply_deriv_operator(6006, mat, dst, src)
-    reg.DhopDerivOEDag = lambda mat, dst, src: op.apply_deriv_operator(
-        7006, mat, dst, src
-    )
+    reg.DhopDerivOEDag = lambda mat, dst, src: op.apply_deriv_operator(7006, mat, dst, src)