dgpnorm for CPU vs. GPU

[kibaekkim]

I think the CPU and GPU codes are not using the same norm.

ExaTron.jl/src/CPU/dgpnorm.jl

Lines 1 to 20 in 3ebc6bc

	function dgpnorm(n::Int, x::Array{Float64}, xl::Array{Float64},
	xu::Array{Float64}, g::Array{Float64})
	inf_norm = 0.0
	for i=1:n
	if xl[i] != xu[i]
	if x[i] == xl[i]
	v = (min(g[i], 0.0))^2
	elseif x[i] == xu[i]
	v = (max(g[i], 0.0))^2
	else
	v = g[i]^2
	end
	v = sqrt(v)
	inf_norm = (inf_norm > v) ? inf_norm : v
	end
	end
	return inf_norm
	end

ExaTron.jl/src/CUDA/dgpnorm.jl

Lines 1 to 32 in 3ebc6bc

	@inline function ExaTron.dgpnorm(n::Int, x::CuDeviceArray{Float64,1}, xl::CuDeviceArray{Float64,1},
	xu::CuDeviceArray{Float64,1}, g::CuDeviceArray{Float64,1})
	tx = threadIdx().x
	v = 0.0
	if tx <= n
	@inbounds begin
	if xl[tx] != xu[tx]
	if x[tx] == xl[tx]
	v = min(g[tx], 0.0)
	elseif x[tx] == xu[tx]
	v = max(g[tx], 0.0)
	else
	v = g[tx]
	end
	v = abs(v)
	end
	end
	end
	# shfl_down_sync() will automatically sync threads in a warp.
	offset = 16
	while offset > 0
	v = max(v, CUDA.shfl_down_sync(0xffffffff, v, offset))
	offset >>= 1
	end
	v = CUDA.shfl_sync(0xffffffff, v, 1)
	return v
	end

ExaTron.jl/src/KA/dgpnorm.jl

Lines 1 to 39 in 3ebc6bc

	@inline function ExaTron.dgpnorm(n::Int, x, xl,
	xu, g,
	tx)
	@synchronize
	res = 0.0
	inf_norm = @localmem Float64 (1,)
	v = 0.0
	if tx == 1
	inf_norm[1] = 0.0
	for i in 1:n
	@inbounds begin
	if xl[i] != xu[i]
	if x[i] == xl[i]
	v = min(g[i], 0.0)
	v = v*v
	elseif x[i] == xu[i]
	v = max(g[i], 0.0)
	v = v*v
	else
	v = g[i]*g[i]
	end
	v = sqrt(v)
	if inf_norm[1] > v
	inf_norm[1] = inf_norm[1]
	else
	inf_norm[1] = v
	end
	end
	end
	end
	end
	@synchronize
	res = inf_norm[1]
	return res
	end

ExaTron.jl/src/driver.jl

Lines 42 to 67 in 3ebc6bc

	function gpnorm(n, x, x_l, x_u, g)
	two_norm = 0.0
	inf_norm = 0.0
	for i=1:n
	if x_l[i] != x_u[i]
	if x[i] == x_l[i]
	val = (min(g[i], 0.0))^2
	elseif x[i] == x_u[i]
	val = (max(g[i], 0.0))^2
	else
	val = g[i]^2
	end
	two_norm += val
	val = sqrt(val)
	if inf_norm < val
	inf_norm = val
	end
	end
	end
	two_norm = sqrt(two_norm)
	return two_norm, inf_norm
	end

I am not sure if this is intended. We need @youngdae to confirm this.

[kibaekkim]

Never mind.. they are same.. but computing in different ways.