use deplane_slow for unaligned arrays

tlnagy · Jan 12, 2024 · 4e3bb5d · 4e3bb5d
1 parent ea0995b
commit 4e3bb5d
Show file tree

Hide file tree

Showing 2 changed files with 29 additions and 12 deletions.
diff --git a/src/ifds.jl b/src/ifds.jl
@@ -320,10 +320,7 @@ function Base.read!(target::AbstractArray{T, N}, tf::TiffFile{O, S}, ifd::IFD{O}
 
     if isplanar(ifd)
         samplesv = vec(samples)
-        temp = deplane(samplesv, spp)
-        GC.@preserve samplesv temp target begin
-            memcpy(pointer(samplesv), pointer(temp), sizeof(target))
-        end
+        deplane!(samplesv, spp)
     end
 end
 
@@ -428,19 +425,37 @@ function reverse_prediction!(ifd::IFD, arr::AbstractArray{T,N}) where {T, N}
                     end
                 end
                 vw = view(reinterpret(UInt8, arr), start+1:start+columns)
-                deplane_simd!(buffer, vw, Val(sizeof(T)))
-                vw .= buffer
+                deplane!(buffer, vw, sizeof(T))
             end
 
             arr .= bswap.(arr)
         end
     end
 end
 
-function deplane(arr::AbstractVector{T}, n::Integer) where T
+# {AAA...BBB...CCC...} => {ABCABCABC...}
+function deplane!(arr::AbstractVector{T}, n::Integer) where T
     out = Vector{T}(undef, length(arr))
-    deplane_simd!(out, arr, Val(n))
-    out
+    deplane!(out, arr, n)
+end
+
+# {AAA...BBB...CCC...} => {ABCABCABC...}
+function deplane!(buffer::AbstractVector{T}, arr::AbstractVector{T}, n::Integer) where T
+    @assert length(buffer) == length(arr)
+    @assert length(arr) % n == 0
+
+    GC.@preserve arr buffer begin
+        if Int(pointer(arr)) & 0x3f > 0 || length(arr) < 64
+            # small or not 64-byte aligned
+            temp = deplane_slow(arr, n)
+            GC.@preserve temp begin
+                memcpy(pointer(arr), pointer(temp), sizeof(temp))
+            end
+        else
+            deplane_simd!(buffer, arr, Val(n))
+            memcpy(pointer(arr), pointer(buffer), sizeof(buffer))
+        end
+    end
 end
 
 # {AAA...BBB...CCC...} => {ABCABCABC...}

diff --git a/test/runtests.jl b/test/runtests.jl
@@ -234,7 +234,7 @@ end
 
     for typ in [Int8,UInt16,Float32]
         for planes in 1:33
-            for size in 100:164
+            for size in 64:164
                 out = Vector{typ}(undef, size * planes)
                 a=reduce(vcat,[fill(typ(x),size) for x in 1:planes])
                 TiffImages.deplane_simd!(out, a, Val(planes))
@@ -245,9 +245,11 @@ end
 
     for typ in [Int8,UInt16,Float32]
         for planes in 1:33
-            for size in 100:164
+            for size in 1:164
                 a=reduce(vcat,[fill(typ(x),size) for x in 1:planes])
-                @test TiffImages.deplane(a, planes) == TiffImages.deplane_slow(a, planes)
+                b=copy(a)
+                TiffImages.deplane!(a, planes)
+                @test a == TiffImages.deplane_slow(b, planes)
             end
         end
     end