Add option for manually specifying unroll or tile, tweaked unrolling in absence of reductions.

Author: chriselrod

src/constructors.jl

@@ -98,6 +98,20 @@ macro avx(q)
     end
     esc(q2)
 end
-
+macro avx(arg, q)
+    @assert q.head === :for
+    @assert arg.head === :(=)
+    local U::Int, T::Int
+    if arg.args[1] === :unroll
+        U = arg.args[2]
+        T = -1
+    elseif arg.args[1] === :tile
+        tup = arg.args[2]
+        @assert tup.head === :tuple
+        U = tup.args[1]
+        T = tup.args[2]
+    end
+    esc(lower(LoopSet(q), U, T))
+end
 
 

src/determinestrategy.jl

@@ -131,8 +131,10 @@ function unroll_no_reductions(ls, order, vectorized, Wshift, size_T)
         end
     end
     # heuristic guess
-    # @show compute_rt, load_rt
-    roundpow2(min(4, round(Int, (compute_rt + load_rt + 1) / compute_rt)))
+    @show compute_rt, load_rt
+    # roundpow2(min(4, round(Int, (compute_rt + load_rt + 1) / compute_rt)))
+    rt = max(compute_rt, load_rt)
+    roundpow2( min( 4, round(Int, 16 / rt) ) )
 end
 function determine_unroll_factor(
     ls::LoopSet, order::Vector{Symbol}, unrolled::Symbol, vectorized::Symbol = first(order)
@@ -151,7 +153,7 @@ function determine_unroll_factor(
     if iszero(num_reductions)
         # if only 1 loop, no need to unroll
         # if more than 1 loop, there is some cost. Picking 2 here as a heuristic.
-        return length(order) == 1 ? 1 : unroll_no_reductions(ls, order, vectorized, Wshift, size_T)
+        return unroll_no_reductions(ls, order, vectorized, Wshift, size_T)
     end
     # So if num_reductions > 0, we set the unroll factor to be high enough so that the CPU can be kept busy
     # if there are, U = max(1, round(Int, max(latency) * throughput / num_reductions)) = max(1, round(Int, latency / (recip_throughput * num_reductions)))

src/lowering.jl

@@ -888,6 +888,13 @@ function lower(ls::LoopSet)
     fillorder!(ls, order, istiled)
     istiled ? lower_tiled(ls, vectorized, U, T) : lower_unrolled(ls, vectorized, U)
 end
+function lower(ls::LoopSet, U, T = -1)
+    num_loops(ls) == 1 && @assert T == -1
+    order, vectorized, _U, _T = choose_order(ls)
+    istiled = T != -1
+    fillorder!(ls, order, istiled)
+    istiled ? lower_tiled(ls, vectorized, U, T) : lower_unrolled(ls, vectorized, U)
+end
 
 Base.convert(::Type{Expr}, ls::LoopSet) = lower(ls)
 Base.show(io::IO, ls::LoopSet) = println(io, lower(ls))