clamp unroll extrema and add div to cost table (copying fdiv for now...)

JuliaSIMD · Apr 26, 2021 · 96aef96 · 96aef96
1 parent 8b5f640
commit 96aef96
Show file tree

Hide file tree

Showing 3 changed files with 8 additions and 7 deletions.
diff --git a/src/codegen/operation_evaluation_order.jl b/src/codegen/operation_evaluation_order.jl
@@ -105,7 +105,7 @@ function fillorder!(ls::LoopSet, order::Vector{Symbol}, u₁loop::Symbol, u₂lo
         ro[_n] = loopsym = order[n]
         #loopsym = order[n]
         for op ∈ ops
-            addoptoorder!( ls, included_vars, place_after_loop, op, loopsym, _n, u₁loop, u₂loop, vectorized, u₂max )
+          addoptoorder!( ls, included_vars, place_after_loop, op, loopsym, _n, u₁loop, u₂loop, vectorized, u₂max )
         end
     end
 end

diff --git a/src/modeling/costs.jl b/src/modeling/costs.jl
@@ -140,6 +140,7 @@ const COST = Dict{Symbol,InstructionCost}(
     :rem_fast => InstructionCost(13,4.0,-2.0), # FIXME
     :div_fast => InstructionCost(13,4.0,-2.0),
     :vdiv_fast => InstructionCost(20,4.0,-2.0), # FIXME
+    :÷ => InstructionCost(13,4.0,-2.0),
     # :evadd => InstructionCost(4,0.5),
     # :evsub => InstructionCost(4,0.5),
     # :evmul => InstructionCost(4,0.5),
@@ -494,6 +495,8 @@ const FUNCTIONSYMBOLS = IdDict{Type{<:Function},Instruction}(
     # typeof(VectorizationBase.vfdiv) => :(/),
     # typeof(VectorizationBase.vfdiv!) => :(/),
     typeof(VectorizationBase.vdiv) => :(/),
+    typeof(÷) => :(÷),
+    typeof(Base.FastMath.div_fast) => :div_fast,
     typeof(Base.FastMath.div_fast) => :div_fast,
     typeof(Base.FastMath.rem_fast) => :rem_fast,
     typeof(==) => :(==),

diff --git a/src/modeling/determinestrategy.jl b/src/modeling/determinestrategy.jl
@@ -469,19 +469,17 @@ function solve_unroll_lagrange(X, R, u₁L, u₂L, u₁step::Int, u₂step::Int,
     u₂low = max(u₂step, floor(Int, 0.8u₂float)) # must be at least 1
     u₁high = solve_unroll_constT(R, u₂low) + u₁step
     u₂high = solve_unroll_constU(R, u₁low) + u₂step
-    # @show u₁low, u₁high, u₂low, u₂high
     if u₁low ≥ u₁high
         u₁low = solve_unroll_constT(R, u₂high)
     end
     if u₂low ≥ u₂high
         u₂low = solve_unroll_constU(R, u₁high)
     end
     maxunroll = atleast32registers ? (((X₂ > 0) & (X₃ > 0)) ? 10 : 8) : 6
-    u₁low = (min(u₁low, maxunroll) ÷ u₁step) * u₁step
-    u₂low = (min(u₂low, maxunroll) ÷ u₂step) * u₂step
-    u₁high = min(u₁high, maxunroll)
-    u₂high = min(u₂high, maxunroll)
-    # @show u₁low, u₁high, u₂low, u₂high, u₁float, u₂float
+    u₁low = (clamp(u₁low, 1, maxunroll) ÷ u₁step) * u₁step
+    u₂low = (clamp(u₂low, 1, maxunroll) ÷ u₂step) * u₂step
+    u₁high = clamp(u₁high, 1, maxunroll)
+    u₂high = clamp(u₂high, 1, maxunroll)
     solve_unroll_iter(X, R, u₁L, u₂L, reverse(u₁low:u₁step:u₁high), reverse(u₂low:u₂step:u₂high))
 end