Two new performance tests for #1073: stream fusion and bounds-checking

test/perf2/bounds.jl

@@ -0,0 +1,25 @@
+function laplacian2(A::Matrix, B::Matrix)
+    d1 = 1
+    d2 = size(A,1)
+    for j = 2:size(B,2)-1
+        offset = (j-1)*size(A,1)
+        for i = 2:size(B,1)-1
+            ii = offset+i
+            B[ii] = A[ii+d1] + A[ii-d1] + A[ii+d2] + A[ii-d2] - 4*A[ii]
+        end
+    end
+end
+
+function time_laplacian(A::Matrix, niter::Int)
+    B = similar(A)
+    print("Laplacian of a matrix: ")
+    @time begin
+        for n = 1:niter
+            laplacian2(A, B)
+        end
+    end
+end
+
+time_laplacian(randn(1000,1000), 100)
+
+# Note: run time_laplacian with bounds-checking on and off (turn off by commenting out the middle two lines in cgutils.cpp's "emit_bounds_check"

test/perf2/fusion.jl

@@ -0,0 +1,29 @@
+function arith_vectorized(b,c,d)
+    a = b.*c + d + 1.0
+end
+
+function arith_loop(b,c,d)
+    a = similar(b)
+    for i = 1:length(b)
+        a[i] = b[i]*c[i] + d[i] + 1.0
+    end
+end
+
+function compare_arith(len::Int, niter::Int)
+    b = randn(len)
+    c = randn(len)
+    d = randn(len)
+    print("With vectorized: ")
+    @time begin
+        for n in 1:niter
+            a = arith_vectorized(b,c,d)
+        end
+    end
+    print("With loop: ")
+    @time begin
+        for n in 1:niter
+            a = arith_loop(b,c,d)
+        end
+    end
+end
+compare_arith(1_000_000, 10)