Removed two probably unreachable add_constant methods (constants in a loop-expr may be of type Expr/Symbol/Number), adjusted tests.

Author: chriselrod

src/add_compute.jl

@@ -94,9 +94,9 @@ function isreductzero(op::Operation, ls::LoopSet, reduct_zero::Symbol)
     isconstant(op) || return false
     reduct_zero === op.instruction.mod && return true
     if reduct_zero === :zero
-        identifier(op) ∈ ls.preamble_zeros && return true
+        iszero(ls, op) && return true
     elseif reduct_zero === :one
-        identifier(op) ∈ ls.preamble_ones && return true
+        isone(ls, op) && return true
     end
     false
 end

src/add_constants.jl

@@ -3,11 +3,11 @@ function add_constant!(ls::LoopSet, var::Symbol, elementbytes::Int)
     pushpreamble!(ls, op, var)
     pushop!(ls, op, var)
 end
-function add_constant!(ls::LoopSet, var, elementbytes::Int = 8)
-    sym = gensym(:loopconstant)
-    pushpreamble!(ls, Expr(:(=), sym, var))
-    add_constant!(ls, sym, elementbytes)
-end
+# function add_constant!(ls::LoopSet, var, elementbytes::Int = 8)
+#     sym = gensym(:loopconstant)
+#     pushpreamble!(ls, Expr(:(=), sym, var))
+#     add_constant!(ls, sym, elementbytes)
+# end
 function add_constant!(ls::LoopSet, var::Number, elementbytes::Int = 8)
     op = Operation(length(operations(ls)), gensym(:loopconstnumber), elementbytes, LOOPCONSTANT, constant, NODEPENDENCY, Symbol[], NOPARENTS)
     ops = operations(ls)
@@ -52,13 +52,13 @@ function add_constant!(
     op = Operation(length(operations(ls)), assignedsym, elementbytes, Instruction(f, value), constant, deps, NODEPENDENCY, NOPARENTS)
     pushop!(ls, op, assignedsym)
 end
-function add_constant!(
-    ls::LoopSet, value, deps::Vector{Symbol}, assignedsym::Symbol, elementbytes::Int, f::Symbol = Symbol("")
-)
-    intermediary = gensym(:intermediate) # hack, passing meta info here
-    pushpreamble!(ls, Expr(:(=), intermediary, value))
-    add_constant!(ls, intermediary, deps, assignedsym, f, elementbytes)
-end
+# function add_constant!(
+#     ls::LoopSet, value, deps::Vector{Symbol}, assignedsym::Symbol, elementbytes::Int, f::Symbol = Symbol("")
+# )
+#     intermediary = gensym(:intermediate) # hack, passing meta info here
+#     pushpreamble!(ls, Expr(:(=), intermediary, value))
+#     add_constant!(ls, intermediary, deps, assignedsym, f, elementbytes)
+# end
 function add_constant!(
     ls::LoopSet, value::Number, deps::Vector{Symbol}, assignedsym::Symbol, elementbytes::Int
 )

src/costs.jl

@@ -228,11 +228,11 @@ function reduction_to_single_vector(x::Float64)
     x == 1.0 ? :evadd : x == 2.0 ? :evmul : x == 5.0 ? :max : x == 6.0 ? :min : throw("Reduction not found.")
 end
 reduction_to_single_vector(x) = reduction_to_single_vector(reduction_instruction_class(x))
-function reduction_to_scalar(x::Float64)
-    # x == 1.0 ? :vsum : x == 2.0 ? :vprod : x == 3.0 ? :vany : x == 4.0 ? :vall : x == 5.0 ? :maximum : x == 6.0 ? :minimum : throw("Reduction not found.")
-    x == 1.0 ? :vsum : x == 2.0 ? :vprod : x == 5.0 ? :maximum : x == 6.0 ? :minimum : throw("Reduction not found.")
-end
-reduction_to_scalar(x) = reduction_to_scalar(reduction_instruction_class(x))
+# function reduction_to_scalar(x::Float64)
+#     # x == 1.0 ? :vsum : x == 2.0 ? :vprod : x == 3.0 ? :vany : x == 4.0 ? :vall : x == 5.0 ? :maximum : x == 6.0 ? :minimum : throw("Reduction not found.")
+#     x == 1.0 ? :vsum : x == 2.0 ? :vprod : x == 5.0 ? :maximum : x == 6.0 ? :minimum : throw("Reduction not found.")
+# end
+# reduction_to_scalar(x) = reduction_to_scalar(reduction_instruction_class(x))
 function reduction_scalar_combine(x::Float64)
     # x == 1.0 ? :reduced_add : x == 2.0 ? :reduced_prod : x == 3.0 ? :reduced_any : x == 4.0 ? :reduced_all : x == 5.0 ? :reduced_max : x == 6.0 ? :reduced_min : throw("Reduction not found.")
     x == 1.0 ? :reduced_add : x == 2.0 ? :reduced_prod : x == 5.0 ? :reduced_max : x == 6.0 ? :reduced_min : throw("Reduction not found.")

src/graphs.jl

@@ -234,7 +234,20 @@ function zerotype(ls::LoopSet, op::Operation)
     end
     INVALID
 end
-
+function Base.iszero(ls::LoopSet, op::Operation)
+    opid = identifier(op)
+    for (id,_) ∈ ls.preamble_zeros
+        opid == id && return true
+    end
+    false
+end
+function Base.isone(ls::LoopSet, op::Operation)
+    opid = identifier(op)
+    for (id,_) ∈ ls.preamble_ones
+        opid == id && return true
+    end
+    false
+end
 
 
 includesarray(ls::LoopSet, array::Symbol) = array ∈ ls.includedarrays

src/lower_compute.jl

@@ -1,17 +1,5 @@
 # A compute op needs to know the unrolling and tiling status of each of its parents.
-#
-function lower_compute_scalar!(
-    q::Expr, op::Operation, vectorized::Symbol, W::Symbol, unrolled::Symbol, tiled::Symbol, U::Int,
-    suffix::Union{Nothing,Int}, mask::Union{Nothing,Symbol,Unsigned} = nothing
-)
-    lower_compute!(q, op, vectorized, W, unrolled, tiled, U, suffix, mask, false)
-end
-function lower_compute_unrolled!(
-    q::Expr, op::Operation, vectorized::Symbol, W::Symbol, unrolled::Symbol, tiled::Symbol, U::Int,
-    suffix::Union{Nothing,Int}, mask::Union{Nothing,Symbol,Unsigned} = nothing
-)
-    lower_compute!(q, op, vectorized, W, unrolled, tiled, U, suffix, mask, true)
-end
+
 struct FalseCollection end
 Base.getindex(::FalseCollection, i...) = false
 function lower_compute!(

test/runtests.jl

@@ -995,33 +995,23 @@ end
             softmax3_core_avx4!(lse, qq, xx, tmpmax, maxk, nk)
         end
 
-        function mysumavx(x)
+        function sumprodavx(x)
             s = zero(eltype(x))
+            p = one(eltype(x))
             @avx for i ∈ eachindex(x)
                 s += x[i]
+                p *=x[i]
             end
-            s
+            s, p
         end
-        function mysum_avx(x)
+        function sumprod_avx(x)
             s = zero(eltype(x))
-            @_avx for i ∈ eachindex(x)
-                s += x[i]
-            end
-            s
-        end
-        function myprodavx(x)
-            p = one(eltype(x))
-            @avx for i ∈ eachindex(x)
-                p *= x[i]
-            end
-            p
-        end
-        function myprod_avx(x)
             p = one(eltype(x))
             @_avx for i ∈ eachindex(x)
-                p *= x[i]
+                s += x[i]
+                p *=x[i]
             end
-            p
+            s, p
         end
 
         function test_bit_shift(counter)
@@ -1155,19 +1145,21 @@ end
             @test sum(q2; dims=3) ≈ ones(T,ni,nj)
 
             x .+= 0.545;
-            s = sum(x)
-            @test s ≈ mysumavx(x)
-            @test s ≈ mysum_avx(x)
-            p = prod(x)
-            @test p ≈ myprodavx(x)
-            @test p ≈ myprod_avx(x)
-            r = T == Float32 ? (Int32(-10):Int32(234)) : -10:234
-            s = sum(r)
-            @test s ≈ mysumavx(r)
-            @test s ≈ mysum_avx(r)
-            p = prod(r)
-            @test p ≈ myprodavx(r)
-            @test p ≈ myprod_avx(r)
+            s = sum(x); p = prod(x)
+            s1, p1 = sumprodavx(x)
+            @test s ≈ s1
+            @test p ≈ p1
+            s1, p1 = sumprod_avx(x)
+            @test s ≈ s1
+            @test p ≈ p1
+            r = T == Float32 ? (Int32(-10):Int32(107)) : (Int64(-10):Int64(107))
+            s = sum(r); p = prod(r)
+            s1, p1 = sumprodavx(r)
+            @test s ≈ s1
+            @test p ≈ p1
+            s1, p1 = sumprod_avx(r)
+            @test s ≈ s1
+            @test p ≈ p1
 
             @test test_bit_shift(r) == test_bit_shiftavx(r)
             @test test_bit_shift(r) == test_bit_shift_avx(r)
@@ -1178,9 +1170,13 @@ end
             else
                 sum(identity, r)
             end
-            @test s ≈ mysumavx(r)
-            @test s ≈ mysum_avx(r)
-            
+            p = prod(r);
+            s1, p1 = sumprodavx(r)
+            @test s ≈ s1
+            @test p ≈ p1
+            s1, p1 = sumprod_avx(r)
+            @test s ≈ s1
+            @test p ≈ p1
         end
 end
 
@@ -1541,34 +1537,34 @@ end
             C[m,n] > 0 && (C[m,n] = Cₘₙ)
         end
     end
-    function condstore!(y, x)
-        @inbounds for i ∈ eachindex(y, x)
-            x1 = x[i]
+    function condstore!(x)
+        @inbounds for i ∈ eachindex(x)
+            x1 = 2*x[i]-100
             x2 = x1*x1
             x3 = x2 + x1
-            y[i] = x1
-            (x1 < 30) && (y[i] = x2)
-            (x1 < 80) || (y[i] = x3)
+            x[i] = x1
+            (x1 < -50) && (x[i] = x2)
+            (x1 < 60) || (x[i] = x3)
         end
     end
-    function condstoreavx!(y, x)
-        @avx for i ∈ eachindex(y, x)
-            x1 = x[i]
+    function condstoreavx!(x)
+        @avx for i ∈ eachindex(x)
+            x1 = 2*x[i]-100
             x2 = x1*x1
             x3 = x2 + x1
-            y[i] = x1
-            (x1 < 30) && (y[i] = x2)
-            (x1 < 80) || (y[i] = x3)
+            x[i] = x1
+            (x1 < -50) && (x[i] = x2)
+            (x1 < 60) || (x[i] = x3)
         end
     end
-    function condstore_avx!(y, x)
-        @_avx for i ∈ eachindex(y, x)
-            x1 = x[i]
+    function condstore_avx!(x)
+        @_avx for i ∈ eachindex(x)
+            x1 = 2*x[i]-100
             x2 = x1*x1
             x3 = x2 + x1
-            y[i] = x1
-            (x1 < 30) && (y[i] = x2)
-            (x1 < 80) || (y[i] = x3)
+            x[i] = x1
+            (x1 < -50) && (x[i] = x2)
+            (x1 < 60) || (x[i] = x3)
         end
     end
 
@@ -1603,12 +1599,13 @@ end
         if T <: Union{Float32,Float64}
             a .*= 100;
         end
-        b2 = similar(b);
-        condstore!(b, a)
-        condstoreavx!(b2, a)
-        @test b == b2
-        fill!(b2, -999999); condstore_avx!(b2, a)
-        @test b == b2
+        b1 = copy(a);
+        b2 = copy(a);
+        condstore!(b1)
+        condstoreavx!(b2)
+        @test b1 == b2
+        copyto!(b2, a); condstore_avx!(b2)
+        @test b1 == b2
 
         M, K, N = 83, 85, 79;
         if T <: Integer