Merge pull request #2476 from devitocodes/tweak-cire-cm

compiler: Improve CIRE's cost model

devito/passes/clusters/aliases.py

@@ -958,15 +958,58 @@ def pick_best(variants):
 
         flops = flops0 + flops1
 
-        # Data movement in the two sweeps
-        indexeds0 = search([sa.pivot for sa in i.schedule], Indexed)
+        # Estimate the data movement in the two sweeps
+
+        # With cross-loop blocking, a Function appearing in both sweeps is
+        # much more likely to be in cache during the second sweep, hence
+        # we count it only once
+        functions0 = set()
+        functions1 = set()
+        for sa in i.schedule:
+            indexeds0 = search(sa.pivot, Indexed)
+
+            if any(d.is_Block for d in sa.ispace.itdims):
+                functions1.update({i.function for i in indexeds0})
+            else:
+                functions0.update({i.function for i in indexeds0})
+
         indexeds1 = search(i.exprs, Indexed)
+        functions1.update({i.function for i in indexeds1})
+
+        nfunctions0 = len(functions0)
+        nfunctions1 = len(functions1)
+
+        # All temporaries impact data movement, but some kind of temporaries
+        # are more likely to be in cache than others, so they are given a
+        # lighter weight
+        for ii in indexeds1:
+            grid = ii.function.grid
+            if grid is None:
+                continue
 
-        ntemps = len(i.schedule)
-        nfunctions0 = len({i.function for i in indexeds0})
-        nfunctions1 = len({i.function for i in indexeds1})
+            ntemps = 0
+            for sa in i.schedule:
+                if len(sa.writeto) < grid.dim:
+                    # Tiny temporary, extremely likely to be in cache, hardly
+                    # impacting data movement in a significant way
+                    ntemps += 0.1
+                elif any(d.is_Block for d in sa.writeto.itdims):
+                    # Cross-loop blocking temporary, likely to be in some level
+                    # of cache (but unlikely to be in the fastest level)
+                    ntemps += 1
+                else:
+                    # Grid-size temporary, likely _not_ to be in cache, and
+                    # therefore requiring at least two costly accesses per
+                    # grid point
+                    ntemps += 2
+
+            ntemps = int(ntemps)
+
+            break
+        else:
+            ntemps = len(i.schedule)
 
-        ws = ntemps*2 + nfunctions0 + nfunctions1
+        ws = ntemps + nfunctions0 + nfunctions1
 
         if best is None:
             best, best_flops, best_ws = i, flops, ws

tests/test_dse.py

@@ -1915,6 +1915,70 @@ def test_tti_adjoint_akin_v2(self):
         # all redundancies have been detected correctly
         assert summary1[('section1', None)].ops == 75
 
+    @switchconfig(profiling='advanced')
+    def test_tti_adjoint_akin_v3(self):
+        so = 8
+        fd_order = 2
+
+        grid = Grid(shape=(20, 20, 20))
+        x, y, z = grid.dimensions
+
+        vx = TimeFunction(name="vx", grid=grid, space_order=so)
+        vy = TimeFunction(name="vy", grid=grid, space_order=so)
+        vz = TimeFunction(name="vz", grid=grid, space_order=so)
+        txy = TimeFunction(name="txy", grid=grid, space_order=so)
+        txz = TimeFunction(name="txz", grid=grid, space_order=so)
+        theta = Function(name='theta', grid=grid, space_order=so)
+        phi = Function(name='phi', grid=grid, space_order=so)
+
+        r00 = cos(theta)*cos(phi)
+        r01 = cos(theta)*sin(phi)
+        r02 = -sin(theta)
+        r10 = -sin(phi)
+        r11 = cos(phi)
+        r12 = cos(theta)
+        r20 = sin(theta)*cos(phi)
+        r21 = sin(theta)*sin(phi)
+        r22 = cos(theta)
+
+        def foo0(field):
+            return ((r00 * field).dx(x0=x+x.spacing/2) +
+                    Derivative(r01 * field, x, deriv_order=0, fd_order=fd_order,
+                               x0=x+x.spacing/2).dy(x0=y) +
+                    Derivative(r02 * field, x, deriv_order=0, fd_order=fd_order,
+                               x0=x+x.spacing/2).dz(x0=z))
+
+        def foo1(field):
+            return (Derivative(r10 * field, y, deriv_order=0, fd_order=fd_order,
+                               x0=y+y.spacing/2).dx(x0=x) +
+                    (r11 * field).dy(x0=y+y.spacing/2) +
+                    Derivative(r12 * field, y, deriv_order=0, fd_order=fd_order,
+                               x0=y+y.spacing/2).dz(x0=z))
+
+        def foo2(field):
+            return (Derivative(r20 * field, z, deriv_order=0, fd_order=fd_order,
+                               x0=z+z.spacing/2).dx(x0=x) +
+                    Derivative(r21 * field, z, deriv_order=0, fd_order=fd_order,
+                               x0=z+z.spacing/2).dy(x0=y) +
+                    (r22 * field).dz(x0=z+z.spacing/2))
+
+        eqns = [Eq(txz.forward, txz + foo0(vz.forward) + foo2(vx.forward)),
+                Eq(txy.forward, txy + foo0(vy.forward) + foo1(vx.forward))]
+
+        op = Operator(eqns, subs=grid.spacing_map,
+                      opt=('advanced', {'openmp': True,
+                                        'cire-rotate': True}))
+
+        # Check code generation
+        bns, _ = assert_blocking(op, {'x0_blk0'})
+        xs, ys, zs = get_params(op, 'x0_blk0_size', 'y0_blk0_size', 'z_size')
+        arrays = [i for i in FindSymbols().visit(bns['x0_blk0']) if i.is_Array]
+        assert len(arrays) == 10
+        assert len([i for i in arrays if i.shape == (zs,)]) == 2
+        assert len([i for i in arrays if i.shape == (9, zs)]) == 2
+
+        assert op._profiler._sections['section1'].sops == 184
+
     @pytest.mark.parametrize('rotate', [False, True])
     @switchconfig(profiling='advanced')
     def test_nested_first_derivatives(self, rotate):
@@ -2153,7 +2217,7 @@ def test_multiple_rotating_dims(self):
         # to jit-compile `op1`. However, we also check numerical correctness
         op1.apply(time_m=0, time_M=nt-2, dt=dt, u=u1, vx=vx1, vy=vy1)
 
-        assert np.allclose(u.data, u1.data, rtol=1e-5)
+        assert np.allclose(u.data, u1.data, rtol=1e-3)
 
     def test_maxpar_option_v2(self):
         """