use SIMD.jl directly instead of LV.jl for fast_findmin()

Author: Moelf

Project.toml

@@ -9,31 +9,31 @@ CodecZlib = "944b1d66-785c-5afd-91f1-9de20f533193"
 EnumX = "4e289a0a-7415-4d19-859d-a7e5c4648b56"
 JSON = "682c06a0-de6a-54ab-a142-c8b1cf79cde6"
 Logging = "56ddb016-857b-54e1-b83d-db4d58db5568"
-LoopVectorization = "bdcacae8-1622-11e9-2a5c-532679323890"
 LorentzVectorHEP = "f612022c-142a-473f-8cfd-a09cf3793c6c"
 LorentzVectors = "3f54b04b-17fc-5cd4-9758-90c048d965e3"
 MuladdMacro = "46d2c3a1-f734-5fdb-9937-b9b9aeba4221"
+SIMD = "fdea26ae-647d-5447-a871-4b548cad5224"
 StructArrays = "09ab397b-f2b6-538f-b94a-2f83cf4a842a"
 
 [weakdeps]
-Makie = "ee78f7c6-11fb-53f2-987a-cfe4a2b5a57a"
 EDM4hep = "eb32b910-dde9-4347-8fce-cd6be3498f0c"
+Makie = "ee78f7c6-11fb-53f2-987a-cfe4a2b5a57a"
 
 [extensions]
-JetVisualisation = "Makie"
 EDM4hepJets = "EDM4hep"
+JetVisualisation = "Makie"
 
 [compat]
 Accessors = "0.1.36"
 CodecZlib = "0.7.4"
 EDM4hep = "0.4.0"
 EnumX = "1.0.4"
 JSON = "0.21.4"
-LoopVectorization = "0.12.170"
 LorentzVectorHEP = "0.1.6"
 LorentzVectors = "0.4.3"
 Makie = "0.20, 0.21"
 MuladdMacro = "0.2.4"
+SIMD = "3.6"
 StructArrays = "0.6.18"
 julia = "1.9"
 

src/JetReconstruction.jl

@@ -18,6 +18,7 @@ module JetReconstruction
 using LorentzVectorHEP
 using MuladdMacro
 using StructArrays
+using SIMD
 
 # Import from LorentzVectorHEP methods for those 4-vector types
 pt2(p::LorentzVector) = LorentzVectorHEP.pt2(p)

src/PlainAlgo.jl

@@ -1,5 +1,3 @@
-using LoopVectorization
-
 """
     dist(i, j, rapidity_array, phi_array)
 

src/TiledAlgoLL.jl

@@ -5,7 +5,6 @@
 
 using Logging
 using Accessors
-using LoopVectorization
 
 # Include struct definitions and basic operations
 include("TiledAlgoLLStructs.jl")

src/Utils.jl

@@ -123,7 +123,7 @@ end
     fast_findmin(dij, n)
 
 Find the minimum value and its index in the first `n` elements of the `dij`
-array. The use of `@turbo` macro gives a significiant performance boost.
+array.
 
 # Arguments
 - `dij`: An array of values.
@@ -133,14 +133,35 @@ array. The use of `@turbo` macro gives a significiant performance boost.
 - `dij_min`: The minimum value in the first `n` elements of the `dij` array.
 - `best`: The index of the minimum value in the `dij` array.
 """
-fast_findmin(dij, n) = begin
-    # findmin(@inbounds @view dij[1:n])
-    best = 1
-    @inbounds dij_min = dij[1]
-    @turbo for here in 2:n
-        newmin = dij[here] < dij_min
-        best = newmin ? here : best
-        dij_min = newmin ? dij[here] : dij_min
+function fast_findmin(x, n)
+    laneIndices = SIMD.Vec{8, Int64}((1, 2, 3, 4, 5, 6, 7, 8))
+    minvals = SIMD.Vec{8, Float64}(Inf)
+    min_indices = SIMD.Vec{8, Int64}(0)
+
+    n_batches, remainder = divrem(n, 8)
+    lane = VecRange{8}(0)
+    i = 1
+    @inbounds @fastmath for _ in 1:n_batches
+        predicate = x[lane + i] < minvals
+        minvals = vifelse(predicate, x[lane + i], minvals)
+        min_indices = vifelse(predicate, laneIndices, min_indices)
+
+        i += 8
+        laneIndices += 8
     end
-    dij_min, best
+
+    min_value = SIMD.minimum(minvals)
+    min_index = min_value == minvals[1] ? min_indices[1] : min_value == minvals[2] ? min_indices[2] :
+                min_value == minvals[3] ? min_indices[3] : min_value == minvals[4] ? min_indices[4] :
+                min_value == minvals[5] ? min_indices[5] : min_value == minvals[6] ? min_indices[6] :
+                min_value == minvals[7] ? min_indices[7] : min_indices[8]
+
+    @inbounds @fastmath for _ in 1:remainder
+        xi = x[i]
+        pred = x[i] < min_value
+        min_value = ifelse(pred, xi, min_value)
+        min_index = ifelse(pred, i, min_index)
+        i += 1
+    end
+    return min_value, min_index
 end