add conditional synchronize

Author: vchuravy

src/KernelAbstractions.jl

@@ -115,13 +115,34 @@ end
 
 """
    @synchronize()
+
+After a `@synchronize` statement all read and writes to global and local memory
+from each thread in the workgroup are visible in from all other threads in the
+workgroup.
 """
 macro synchronize()
     quote
         $__synchronize()
     end
 end
 
+"""
+   @synchronize(cond)
+
+After a `@synchronize` statement all read and writes to global and local memory
+from each thread in the workgroup are visible in from all other threads in the
+workgroup. `cond` is not allowed to have any visible sideffects.
+
+# Platform differences
+  - `GPU`: This synchronization will only occur if the `cond` evaluates.
+  - `CPU`: This synchronization will always occur.
+"""
+macro synchronize(cond)
+    quote
+        $(esc(cond)) && $__synchronize()
+    end
+end
+
 """
    @index
 

src/macros.jl

@@ -103,7 +103,13 @@ struct WorkgroupLoop
     private :: Vector{Any}
 end
 
-is_sync(expr) = @capture(expr, @synchronize)
+is_sync(expr) = @capture(expr, @synchronize() | @synchronize(a_))
+
+function is_scope_construct(expr::Expr)
+    expr.head === :block # ||
+    # expr.head === :let
+end
+
 function find_sync(stmt)
     result = false
     postwalk(stmt) do expr
@@ -131,7 +137,7 @@ function split(stmts,
             push!(new_stmts, emit(loop))
             allocations = Any[]
             current     = Any[]
-            @capture(stmt, @synchronize) && continue
+            is_sync(stmt) && continue
 
             # Recurse into scope constructs
             # TODO: This currently implements hard scoping
@@ -140,7 +146,7 @@ function split(stmts,
             recurse(x) = x
             function recurse(expr::Expr)
                 expr = unblock(expr)
-                if any(is_sync, expr.args)
+                if is_scope_construct(expr) && any(is_sync, expr.args)
                     new_args = unblock(split(expr.args, deepcopy(indicies), deepcopy(private)))
                     return Expr(expr.head, new_args...)
                 else 

test/unroll.jl

@@ -9,13 +9,16 @@ using StaticArrays
 end
 
 @kernel function kernel_unroll!(a, ::Val{N}) where N
-  @unroll for i in 1:N
-    @inbounds a[i] = i
+  let M = N+5
+    @unroll for i in 6:M
+      @inbounds a[i-5] = i
+    end
+    @synchronize
   end
 end
 
 # Check that nested `@unroll` doesn't throw a syntax error
-@kernel function kernel_unroll!(a, ::Val{N}) where N
+@kernel function kernel_unroll2!(A)
   @uniform begin
     a = MVector{3, Float64}(1, 2, 3)
     b = MVector{3, Float64}(3, 2, 1)
@@ -28,16 +31,16 @@ end
         c[1, j] = m * a[1] * b[j]
       end
     end
-    a[I] = c[1, 1]
-    m % 2 == 0 && @synchronize
+    A[I] = c[1, 1]
+    @synchronize(m % 2 == 0)
   end
 end
 
 let
   a = zeros(5)
   kernel! = kernel_unroll!(CPU(), 1, 1)
-  event = kernel!(a)
-  wait(event)
-  event = kernel!(a, Val(5))
-  wait(event)
+  wait(kernel!(a))
+  wait(kernel!(a, Val(5)))
+  kernel2! = kernel_unroll2!(CPU(), 1, 1)
+  wait(kernel2!(a))
 end