Update FixedEffectSolverCPU.jl (#53)

* Update FixedEffectSolverCPU.jl

Correct GLFixedEffectModels

* Update Project.toml

* Update FixedEffectSolverCPU.jl

* Update solve.jl

* Update FixedEffectSolverGPU.jl

* Update Project.toml

Project.toml

@@ -1,6 +1,6 @@
 name = "FixedEffects"
 uuid = "c8885935-8500-56a7-9867-7708b20db0eb"
-version = "2.0.8"
+version = "2.1.0"
 
 [deps]
 LinearAlgebra = "37e2e46d-f89d-539d-b4ee-838fcccc9c8e"

src/FixedEffectSolvers/FixedEffectSolverCPU.jl

@@ -132,6 +132,7 @@ end
 
 
 function scale!(scale::AbstractVector, refs::AbstractVector, interaction::AbstractVector, weights::AbstractVector)
+        fill!(scale, 0)
 	@inbounds @simd for i in eachindex(refs)
 		scale[refs[i]] += abs2(interaction[i]) * weights[i]
 	end
@@ -153,14 +154,13 @@ function solve_residuals!(r::AbstractVector, feM::FixedEffectSolverCPU{T}; tol::
 		feM.r .*=  sqrt.(feM.weights)
 	end
 	copyto!(feM.b, feM.r)
-	if length(feM.x.x) == 1
-		mul!(feM.x, feM.m', feM.b, 1, 0)
-		iter, converged = 1, true
-	else
-		mul!(feM.x, feM.m', feM.b, 1, 0)
-		x, ch = lsmr!(feM.x, feM.m, feM.b, feM.v, feM.h, feM.hbar; atol = tol, btol = tol, maxiter = maxiter)
-		iter, converged = ch.mvps + 1, ch.isconverged
-	end
+	mul!(feM.x, feM.m', feM.b, 1, 0)
+        if length(feM.x.x) > 1
+            x, ch = lsmr!(feM.x, feM.m, feM.b, feM.v, feM.h, feM.hbar; atol = tol, btol = tol, maxiter = maxiter)
+	    iter, converged = ch.mvps + 1, ch.isconverged
+        else
+            iter, converged = 1, true
+        end
 	mul!(feM.r, feM.m, feM.x, -1, 1)
 	if !(feM.weights isa UnitWeights)
 		feM.r ./=  sqrt.(feM.weights)

src/FixedEffectSolvers/FixedEffectSolverGPU.jl

@@ -146,6 +146,7 @@ end
 
 function scale!(scale::CuVector, refs::CuVector, interaction::CuVector, weights::CuVector, nthreads::Integer)
 	nblocks = cld(length(refs), nthreads) 
+        fill!(scale, 0)
 	@cuda threads=nthreads blocks=nblocks scale_kernel!(scale, refs, interaction, weights)
 	@cuda threads=nthreads blocks=nblocks inv_kernel!(scale)
 end

test/solve.jl

@@ -22,15 +22,6 @@ r_ols =  [-0.2015993617092453,  0.2015993617092464, -0.2015993617092463,  0.2015
 (c, iter, conv) = solve_residuals!([x x], fes)
 
 
-# test update_weights
-weights = ones(10)
-feM = FixedEffects.AbstractFixedEffectSolver{Float64}(fes, Weights(weights), Val{:cpu})
-weights = Weights([1, 2, 3, 4, 5, 6, 7, 8, 9, 10])
-FixedEffects.update_weights!(feM, weights) 
-solve_residuals!(deepcopy(x), feM)[1] ≈ solve_residuals!(deepcopy(x), fes, weights)[1]
-
-
-
 method_s = [:cpu]
 if FixedEffects.has_CUDA()
 	push!(method_s, :gpu)
@@ -45,3 +36,15 @@ end
 fe = FixedEffect([1, 2])
 @test_throws "FixedEffects must have the same length as y" ỹ = solve_residuals!(ones(100), [fe])
 
+
+# test update_weights
+weights = ones(10)
+fes = [FixedEffect(p1)]
+feM = FixedEffects.AbstractFixedEffectSolver{Float64}(fes, Weights(weights), Val{:cpu})
+weights = Weights([1, 2, 3, 4, 5, 6, 7, 8, 9, 10])
+FixedEffects.update_weights!(feM, weights) 
+solve_residuals!(deepcopy(x), feM)[1] ≈ solve_residuals!(deepcopy(x), fes, weights)[1]
+
+weights = Weights(reverse([1, 2, 3, 4, 5, 6, 7, 8, 9, 10]))
+FixedEffects.update_weights!(feM, weights) 
+solve_residuals!(deepcopy(x), feM)[1] ≈ solve_residuals!(deepcopy(x), fes, weights)[1]