From 87c311f8e55eb851e6bee2d040521ad84a5c6098 Mon Sep 17 00:00:00 2001
From: maximvassiliev <vassiliev.max@gmail.com>
Date: Mon, 8 Jul 2024 16:22:34 +0100
Subject: [PATCH 01/14] add sum without dims

---
 src/BandedMatrices.jl      |  1 +
 src/banded/BandedMatrix.jl | 21 +++++++++++++++++++++
 2 files changed, 22 insertions(+)

diff --git a/src/BandedMatrices.jl b/src/BandedMatrices.jl
index ff9b760a..0d89c893 100644
--- a/src/BandedMatrices.jl
+++ b/src/BandedMatrices.jl
@@ -99,4 +99,5 @@ end
 include("precompile.jl")
 
 
+
 end #module
diff --git a/src/banded/BandedMatrix.jl b/src/banded/BandedMatrix.jl
index f38834cf..065a4956 100644
--- a/src/banded/BandedMatrix.jl
+++ b/src/banded/BandedMatrix.jl
@@ -1007,3 +1007,24 @@ function resize(A::BandedSubBandedMatrix, n::Integer, m::Integer)
     l,u = bandwidths(A)
     _BandedMatrix(reshape(resize!(vec(copy(bandeddata(A))), (l+u+1)*m), l+u+1, m), n, l,u)
 end
+
+function Base.sum(A::BandedMatrix)
+    l,u = bandwidths(A)
+    height,width = size(A)
+    lower, upper = min(height-1,l), min(width-1,u)
+    data = zeros(1+lower+upper,min(height,width))
+    for i=-lower:upper
+        b = A[band(i)]
+        data[i+lower+1,1:length(b)] = b
+    end
+    sum(data)
+end
+
+function Base.sum(A::BandedMatrix; dims)
+    if(dims > 2)
+        A
+    elseif(dims == 2)
+        A
+    end
+end
+

From b6530b905464e28c4312679e41a610257d98dd11 Mon Sep 17 00:00:00 2001
From: maximvassiliev <vassiliev.max@gmail.com>
Date: Mon, 8 Jul 2024 16:35:09 +0100
Subject: [PATCH 02/14] add sum; dims=1

---
 src/banded/BandedMatrix.jl | 12 ++++++++++++
 1 file changed, 12 insertions(+)

diff --git a/src/banded/BandedMatrix.jl b/src/banded/BandedMatrix.jl
index 065a4956..4cf9a892 100644
--- a/src/banded/BandedMatrix.jl
+++ b/src/banded/BandedMatrix.jl
@@ -1011,6 +1011,7 @@ end
 function Base.sum(A::BandedMatrix)
     l,u = bandwidths(A)
     height,width = size(A)
+    #Only get nonempty bands
     lower, upper = min(height-1,l), min(width-1,u)
     data = zeros(1+lower+upper,min(height,width))
     for i=-lower:upper
@@ -1025,6 +1026,17 @@ function Base.sum(A::BandedMatrix; dims)
         A
     elseif(dims == 2)
         A
+    elseif(dims == 1)
+        l,u = bandwidths(A)
+        height,width = size(A)
+        #Only get nonempty bands
+        lower, upper = min(height-1,l), min(width-1,u)
+        data = zeros(1+lower+upper,min(height,width))
+        for i=-lower:upper
+            b = A[band(i)]
+            data[i+lower+1,(i <= 0 ? (1:length(b)) : (end-length(b)+1:end))] = b
+        end
+        sum(data; dims=1)
     end
 end
 

From bec277e9e79ee8e1effb6e1c01e8b83ac59e10af Mon Sep 17 00:00:00 2001
From: maximvassiliev <vassiliev.max@gmail.com>
Date: Mon, 8 Jul 2024 22:45:22 +0100
Subject: [PATCH 03/14] support for dims = 2 and error handling

---
 src/banded/BandedMatrix.jl | 19 +++++++++++++++----
 1 file changed, 15 insertions(+), 4 deletions(-)

diff --git a/src/banded/BandedMatrix.jl b/src/banded/BandedMatrix.jl
index 4cf9a892..6abd3013 100644
--- a/src/banded/BandedMatrix.jl
+++ b/src/banded/BandedMatrix.jl
@@ -1025,18 +1025,29 @@ function Base.sum(A::BandedMatrix; dims)
     if(dims > 2)
         A
     elseif(dims == 2)
-        A
+        l,u = bandwidths(A)
+        height,width = size(A)
+        #Only get nonempty bands
+        lower, upper = min(height-1,l), min(width-1,u)
+        data = zeros(height,1+lower+upper)
+        for i=-lower:upper
+            b = A[band(i)]
+            data[(i > 0 ? (1:length(b)) : (1-i:length(b)-i)),i+lower+1] = b
+        end
+        sum(data; dims=2)
     elseif(dims == 1)
         l,u = bandwidths(A)
         height,width = size(A)
         #Only get nonempty bands
         lower, upper = min(height-1,l), min(width-1,u)
-        data = zeros(1+lower+upper,min(height,width))
+        data = zeros(1+lower+upper,width)
         for i=-lower:upper
             b = A[band(i)]
-            data[i+lower+1,(i <= 0 ? (1:length(b)) : (end-length(b)+1:end))] = b
+            data[i+lower+1,(i <= 0 ? (1:length(b)) : (i+1:i+length(b)))] = b
         end
-        sum(data; dims=1)
+        sum(data;dims=1)
+    else
+        throw(ArgumentError("dimension must be ≥ 1, got $dims"))
     end
 end
 

From cd92d7f0caab3067c8ae89e616c86b89ed3a2b66 Mon Sep 17 00:00:00 2001
From: maximvassiliev <vassiliev.max@gmail.com>
Date: Mon, 8 Jul 2024 23:07:32 +0100
Subject: [PATCH 04/14] fix for empty matrices and added unit tests

---
 src/banded/BandedMatrix.jl | 14 ++++++++++++--
 test/test_sum.jl           | 11 +++++++++++
 2 files changed, 23 insertions(+), 2 deletions(-)
 create mode 100644 test/test_sum.jl

diff --git a/src/banded/BandedMatrix.jl b/src/banded/BandedMatrix.jl
index 6abd3013..924f0c99 100644
--- a/src/banded/BandedMatrix.jl
+++ b/src/banded/BandedMatrix.jl
@@ -1010,6 +1010,10 @@ end
 
 function Base.sum(A::BandedMatrix)
     l,u = bandwidths(A)
+    #For empty matrices
+    if(l + u < 0)
+        return convert(eltype(A),0)
+    end
     height,width = size(A)
     #Only get nonempty bands
     lower, upper = min(height-1,l), min(width-1,u)
@@ -1027,18 +1031,24 @@ function Base.sum(A::BandedMatrix; dims)
     elseif(dims == 2)
         l,u = bandwidths(A)
         height,width = size(A)
-        #Only get nonempty bands
+        if(l + u < 0)
+            return zeros(eltype(A),(height,1))
+        end
         lower, upper = min(height-1,l), min(width-1,u)
+        #dimension of interest preserved
         data = zeros(height,1+lower+upper)
         for i=-lower:upper
             b = A[band(i)]
+            #populates data array in such a way that everything along the same axis of interest is aligned
             data[(i > 0 ? (1:length(b)) : (1-i:length(b)-i)),i+lower+1] = b
         end
         sum(data; dims=2)
     elseif(dims == 1)
         l,u = bandwidths(A)
         height,width = size(A)
-        #Only get nonempty bands
+        if(l + u < 0)
+            return zeros(eltype(A),(1,width))
+        end
         lower, upper = min(height-1,l), min(width-1,u)
         data = zeros(1+lower+upper,width)
         for i=-lower:upper
diff --git a/test/test_sum.jl b/test/test_sum.jl
new file mode 100644
index 00000000..1acc84cc
--- /dev/null
+++ b/test/test_sum.jl
@@ -0,0 +1,11 @@
+using Test, BandedMatrices, Random
+
+r = brand(Float64,rand(1:10_000),rand(1:10_000),rand(-20:100),rand(-20:100))
+matr = Matrix(r)
+@testset "sum" begin
+    @test sum(r) ≈ sum(Matrix(r)) atol = 1e-8
+    @test sum(r; dims=2) ≈ sum(Matrix(r); dims=2) atol = 1e-8
+    @test sum(r; dims=1) ≈ sum(Matrix(r); dims=1) atol = 1e-8
+    @test sum(r; dims=3) == r
+    @test_throws ArgumentError sum(r; dims=0)
+end
\ No newline at end of file

From 0768d344f3a215a0ddff1fd870b7c8bd504b4838 Mon Sep 17 00:00:00 2001
From: maximvassiliev <vassiliev.max@gmail.com>
Date: Mon, 8 Jul 2024 23:46:08 +0100
Subject: [PATCH 05/14] style

---
 src/banded/BandedMatrix.jl | 26 +++++++++++++-------------
 1 file changed, 13 insertions(+), 13 deletions(-)

diff --git a/src/banded/BandedMatrix.jl b/src/banded/BandedMatrix.jl
index 924f0c99..9ca282d0 100644
--- a/src/banded/BandedMatrix.jl
+++ b/src/banded/BandedMatrix.jl
@@ -1029,33 +1029,33 @@ function Base.sum(A::BandedMatrix; dims)
     if(dims > 2)
         A
     elseif(dims == 2)
-        l,u = bandwidths(A)
-        height,width = size(A)
+        l, u = bandwidths(A)
+        height, width = size(A)
         if(l + u < 0)
-            return zeros(eltype(A),(height,1))
+            return zeros(eltype(A), (height, 1))
         end
-        lower, upper = min(height-1,l), min(width-1,u)
+        lower, upper = min(height-1, l), min(width-1, u)
         #dimension of interest preserved
-        data = zeros(height,1+lower+upper)
+        data = zeros(height, 1+lower+upper)
         for i=-lower:upper
             b = A[band(i)]
             #populates data array in such a way that everything along the same axis of interest is aligned
-            data[(i > 0 ? (1:length(b)) : (1-i:length(b)-i)),i+lower+1] = b
+            data[(i > 0 ? (1:length(b)) : (1-i:length(b)-i)), i+lower+1] = b
         end
         sum(data; dims=2)
     elseif(dims == 1)
-        l,u = bandwidths(A)
-        height,width = size(A)
+        l, u = bandwidths(A)
+        height, width = size(A)
         if(l + u < 0)
-            return zeros(eltype(A),(1,width))
+            return zeros(eltype(A), (1,width))
         end
-        lower, upper = min(height-1,l), min(width-1,u)
-        data = zeros(1+lower+upper,width)
+        lower, upper = min(height-1, l), min(width-1, u)
+        data = zeros(1+lower+upper, width)
         for i=-lower:upper
             b = A[band(i)]
-            data[i+lower+1,(i <= 0 ? (1:length(b)) : (i+1:i+length(b)))] = b
+            data[i+lower+1, (i <= 0 ? (1:length(b)) : (i+1:i+length(b)))] = b
         end
-        sum(data;dims=1)
+        sum(data; dims=1)
     else
         throw(ArgumentError("dimension must be ≥ 1, got $dims"))
     end

From cb197f5d06164f7469116c14873aafc71678f350 Mon Sep 17 00:00:00 2001
From: maximvassiliev <vassiliev.max@gmail.com>
Date: Tue, 9 Jul 2024 14:17:19 +0100
Subject: [PATCH 06/14] make improvements

---
 src/BandedMatrices.jl      |  2 +-
 src/banded/BandedMatrix.jl | 58 ++++++++++++++++++++------------------
 test/test_sum.jl           |  6 ++--
 3 files changed, 35 insertions(+), 31 deletions(-)

diff --git a/src/BandedMatrices.jl b/src/BandedMatrices.jl
index 0d89c893..c6584dc0 100644
--- a/src/BandedMatrices.jl
+++ b/src/BandedMatrices.jl
@@ -7,7 +7,7 @@ import Base: axes, axes1, getproperty, getindex, setindex!, *, +, -, ==, <, <=,
                >=, /, \, adjoint, transpose, showerror, convert, size, view,
                unsafe_indices, first, last, size, length, unsafe_length, step, to_indices,
                to_index, show, fill!, similar, copy, promote_rule, real, imag,
-               copyto!, Array
+               copyto!, Array, sum
 
 using Base.Broadcast: AbstractArrayStyle, DefaultArrayStyle, Broadcasted
 import Base.Broadcast: BroadcastStyle, broadcasted
diff --git a/src/banded/BandedMatrix.jl b/src/banded/BandedMatrix.jl
index 9ca282d0..15e55f3e 100644
--- a/src/banded/BandedMatrix.jl
+++ b/src/banded/BandedMatrix.jl
@@ -1008,54 +1008,58 @@ function resize(A::BandedSubBandedMatrix, n::Integer, m::Integer)
     _BandedMatrix(reshape(resize!(vec(copy(bandeddata(A))), (l+u+1)*m), l+u+1, m), n, l,u)
 end
 
-function Base.sum(A::BandedMatrix)
-    l,u = bandwidths(A)
-    #For empty matrices
+function sum(A::BandedMatrix)
+    l, u = bandwidths(A)
+    ret = zero(eltype(A))
     if(l + u < 0)
-        return convert(eltype(A),0)
+        return ret
     end
-    height,width = size(A)
+    n, m = size(A)
     #Only get nonempty bands
-    lower, upper = min(height-1,l), min(width-1,u)
-    data = zeros(1+lower+upper,min(height,width))
-    for i=-lower:upper
-        b = A[band(i)]
-        data[i+lower+1,1:length(b)] = b
+    lower, upper = min(n-1, l), min(m-1, u)
+    for i = -lower:upper
+        ret += sum(A[band(i)])
     end
-    sum(data)
+    ret
 end
 
-function Base.sum(A::BandedMatrix; dims)
+function sum(A::BandedMatrix; dims)
     if(dims > 2)
         A
     elseif(dims == 2)
         l, u = bandwidths(A)
-        height, width = size(A)
+        n, m = size(A)
+        ret = zeros(eltype(A), (n, 1))
         if(l + u < 0)
-            return zeros(eltype(A), (height, 1))
+            return ret
         end
-        lower, upper = min(height-1, l), min(width-1, u)
-        #dimension of interest preserved
-        data = zeros(height, 1+lower+upper)
-        for i=-lower:upper
+        lower, upper = min(n-1, l), min(m-1, u)
+        for i = -lower:upper
             b = A[band(i)]
-            #populates data array in such a way that everything along the same axis of interest is aligned
-            data[(i > 0 ? (1:length(b)) : (1-i:length(b)-i)), i+lower+1] = b
+            if(i <= 0)
+                ret[1-i:length(b)-i, 1] += b
+            else
+                ret[1:length(b), 1] += b
+            end
         end
-        sum(data; dims=2)
+        ret
     elseif(dims == 1)
         l, u = bandwidths(A)
-        height, width = size(A)
+        n, m = size(A)
+        ret = zeros(eltype(A), (1, m))
         if(l + u < 0)
-            return zeros(eltype(A), (1,width))
+            return ret
         end
-        lower, upper = min(height-1, l), min(width-1, u)
-        data = zeros(1+lower+upper, width)
+        lower, upper = min(n-1, l), min(m-1, u)
         for i=-lower:upper
             b = A[band(i)]
-            data[i+lower+1, (i <= 0 ? (1:length(b)) : (i+1:i+length(b)))] = b
+            if(i <= 0)
+                ret[1, 1:length(b)] += b
+            else
+                ret[1, i+1:i+length(b)] += b
+            end
         end
-        sum(data; dims=1)
+        ret
     else
         throw(ArgumentError("dimension must be ≥ 1, got $dims"))
     end
diff --git a/test/test_sum.jl b/test/test_sum.jl
index 1acc84cc..96c42da4 100644
--- a/test/test_sum.jl
+++ b/test/test_sum.jl
@@ -3,9 +3,9 @@ using Test, BandedMatrices, Random
 r = brand(Float64,rand(1:10_000),rand(1:10_000),rand(-20:100),rand(-20:100))
 matr = Matrix(r)
 @testset "sum" begin
-    @test sum(r) ≈ sum(Matrix(r)) atol = 1e-8
-    @test sum(r; dims=2) ≈ sum(Matrix(r); dims=2) atol = 1e-8
-    @test sum(r; dims=1) ≈ sum(Matrix(r); dims=1) atol = 1e-8
+    @test sum(r) ≈ sum(matr) atol = 1e-10
+    @test sum(r; dims=2) ≈ sum(matr; dims=2) atol = 1e-10
+    @test sum(r; dims=1) ≈ sum(matr; dims=1) atol = 1e-10
     @test sum(r; dims=3) == r
     @test_throws ArgumentError sum(r; dims=0)
 end
\ No newline at end of file

From d0109be4ab1f8f5fe2e3cb20fc7b0d03cc364063 Mon Sep 17 00:00:00 2001
From: maximvassiliev <vassiliev.max@gmail.com>
Date: Tue, 9 Jul 2024 14:42:53 +0100
Subject: [PATCH 07/14] add test_sum.jl to runtests.jl

---
 test/runtests.jl | 1 +
 test/test_sum.jl | 4 ++++
 2 files changed, 5 insertions(+)

diff --git a/test/runtests.jl b/test/runtests.jl
index 2e536839..7778453b 100644
--- a/test/runtests.jl
+++ b/test/runtests.jl
@@ -26,3 +26,4 @@ include("test_symbanded.jl")
 include("test_tribanded.jl")
 include("test_interface.jl")
 include("test_miscs.jl")
+include("test_sum.jl")
diff --git a/test/test_sum.jl b/test/test_sum.jl
index 96c42da4..ff37f95f 100644
--- a/test/test_sum.jl
+++ b/test/test_sum.jl
@@ -1,3 +1,5 @@
+module TestSum
+
 using Test, BandedMatrices, Random
 
 r = brand(Float64,rand(1:10_000),rand(1:10_000),rand(-20:100),rand(-20:100))
@@ -8,4 +10,6 @@ matr = Matrix(r)
     @test sum(r; dims=1) ≈ sum(matr; dims=1) atol = 1e-10
     @test sum(r; dims=3) == r
     @test_throws ArgumentError sum(r; dims=0)
+end
+
 end
\ No newline at end of file

From 8da60dfe142b0895c4e083c17d2593f87314816e Mon Sep 17 00:00:00 2001
From: maximvassiliev <vassiliev.max@gmail.com>
Date: Tue, 9 Jul 2024 15:26:37 +0100
Subject: [PATCH 08/14] fix method dispatch issue in a way that mimics Base.sum

---
 src/banded/BandedMatrix.jl | 32 +++++++++++++++-----------------
 1 file changed, 15 insertions(+), 17 deletions(-)

diff --git a/src/banded/BandedMatrix.jl b/src/banded/BandedMatrix.jl
index 15e55f3e..57cdf3e0 100644
--- a/src/banded/BandedMatrix.jl
+++ b/src/banded/BandedMatrix.jl
@@ -1008,23 +1008,21 @@ function resize(A::BandedSubBandedMatrix, n::Integer, m::Integer)
     _BandedMatrix(reshape(resize!(vec(copy(bandeddata(A))), (l+u+1)*m), l+u+1, m), n, l,u)
 end
 
-function sum(A::BandedMatrix)
-    l, u = bandwidths(A)
-    ret = zero(eltype(A))
-    if(l + u < 0)
-        return ret
-    end
-    n, m = size(A)
-    #Only get nonempty bands
-    lower, upper = min(n-1, l), min(m-1, u)
-    for i = -lower:upper
-        ret += sum(A[band(i)])
-    end
-    ret
-end
-
-function sum(A::BandedMatrix; dims)
-    if(dims > 2)
+function sum(A::BandedMatrix; dims=:)
+    if(dims == :)
+        l, u = bandwidths(A)
+        ret = zero(eltype(A))
+        if(l + u < 0)
+            return ret
+        end
+        n, m = size(A)
+        #Only get nonempty bands
+        lower, upper = min(n-1, l), min(m-1, u)
+        for i = -lower:upper
+            ret += sum(A[band(i)])
+        end
+        ret
+    elseif(dims > 2)
         A
     elseif(dims == 2)
         l, u = bandwidths(A)

From de15c533982793938ef91cf8848537c500f0ee7a Mon Sep 17 00:00:00 2001
From: maximvassiliev <vassiliev.max@gmail.com>
Date: Wed, 10 Jul 2024 14:03:53 +0100
Subject: [PATCH 09/14] update unit tests, reduce memory allocation, improve
 style

---
 src/banded/BandedMatrix.jl | 44 +++++++++++++-------------------------
 test/test_sum.jl           | 11 +++++++---
 2 files changed, 23 insertions(+), 32 deletions(-)

diff --git a/src/banded/BandedMatrix.jl b/src/banded/BandedMatrix.jl
index 57cdf3e0..70037d4a 100644
--- a/src/banded/BandedMatrix.jl
+++ b/src/banded/BandedMatrix.jl
@@ -1009,53 +1009,39 @@ function resize(A::BandedSubBandedMatrix, n::Integer, m::Integer)
 end
 
 function sum(A::BandedMatrix; dims=:)
-    if(dims == :)
+    if dims isa Colon
         l, u = bandwidths(A)
         ret = zero(eltype(A))
-        if(l + u < 0)
+        if l + u < 0
             return ret
         end
         n, m = size(A)
-        #Only get nonempty bands
-        lower, upper = min(n-1, l), min(m-1, u)
-        for i = -lower:upper
-            ret += sum(A[band(i)])
+        for i = 1:n, j = rowrange(A, i)
+            ret += A[i, j]
         end
         ret
-    elseif(dims > 2)
+    elseif dims > 2
         A
-    elseif(dims == 2)
+    elseif dims == 2
         l, u = bandwidths(A)
         n, m = size(A)
-        ret = zeros(eltype(A), (n, 1))
-        if(l + u < 0)
+        ret = zeros(eltype(A), n, 1)
+        if l + u < 0
             return ret
         end
-        lower, upper = min(n-1, l), min(m-1, u)
-        for i = -lower:upper
-            b = A[band(i)]
-            if(i <= 0)
-                ret[1-i:length(b)-i, 1] += b
-            else
-                ret[1:length(b), 1] += b
-            end
+        for i = 1:n, j = rowrange(A, i)
+            ret[i, 1] += A[i, j]
         end
         ret
-    elseif(dims == 1)
+    elseif dims == 1
         l, u = bandwidths(A)
         n, m = size(A)
-        ret = zeros(eltype(A), (1, m))
-        if(l + u < 0)
+        ret = zeros(eltype(A), 1, m)
+        if l + u < 0
             return ret
         end
-        lower, upper = min(n-1, l), min(m-1, u)
-        for i=-lower:upper
-            b = A[band(i)]
-            if(i <= 0)
-                ret[1, 1:length(b)] += b
-            else
-                ret[1, i+1:i+length(b)] += b
-            end
+        for i = 1:m, j = colrange(A, i)
+            ret[1, i] += A[j, i]
         end
         ret
     else
diff --git a/test/test_sum.jl b/test/test_sum.jl
index ff37f95f..54aee662 100644
--- a/test/test_sum.jl
+++ b/test/test_sum.jl
@@ -3,11 +3,16 @@ module TestSum
 using Test, BandedMatrices, Random
 
 r = brand(Float64,rand(1:10_000),rand(1:10_000),rand(-20:100),rand(-20:100))
+empty_r = brand(Float64,rand(1:1_000),rand(1:1_000),rand(1:100),rand(-200:-101))
+n,m = size(empty_r)
 matr = Matrix(r)
 @testset "sum" begin
-    @test sum(r) ≈ sum(matr) atol = 1e-10
-    @test sum(r; dims=2) ≈ sum(matr; dims=2) atol = 1e-10
-    @test sum(r; dims=1) ≈ sum(matr; dims=1) atol = 1e-10
+    @test sum(empty_r) ≈ 0
+    @test sum(empty_r; dims = 2) ≈ zeros(n,1)
+    @test sum(empty_r; dims = 1) ≈ zeros(1,m)
+    @test sum(r) ≈ sum(matr) rtol = 1e-10
+    @test sum(r; dims=2) ≈ sum(matr; dims=2) rtol = 1e-10
+    @test sum(r; dims=1) ≈ sum(matr; dims=1) rtol = 1e-10
     @test sum(r; dims=3) == r
     @test_throws ArgumentError sum(r; dims=0)
 end

From 0e0bc12ed40067f60d55b605ea4f81b594479163 Mon Sep 17 00:00:00 2001
From: maximvassiliev <vassiliev.max@gmail.com>
Date: Wed, 10 Jul 2024 16:15:19 +0100
Subject: [PATCH 10/14] update unit tests, add sum!, move to
 AbstractBandedMatrix.jl

---
 src/BandedMatrices.jl               |  2 +-
 src/banded/BandedMatrix.jl          | 41 -------------------
 src/generic/AbstractBandedMatrix.jl | 63 +++++++++++++++++++++++++++++
 test/test_sum.jl                    | 12 +++---
 4 files changed, 70 insertions(+), 48 deletions(-)

diff --git a/src/BandedMatrices.jl b/src/BandedMatrices.jl
index c6584dc0..5028f2ea 100644
--- a/src/BandedMatrices.jl
+++ b/src/BandedMatrices.jl
@@ -7,7 +7,7 @@ import Base: axes, axes1, getproperty, getindex, setindex!, *, +, -, ==, <, <=,
                >=, /, \, adjoint, transpose, showerror, convert, size, view,
                unsafe_indices, first, last, size, length, unsafe_length, step, to_indices,
                to_index, show, fill!, similar, copy, promote_rule, real, imag,
-               copyto!, Array, sum
+               copyto!, Array, sum, sum!
 
 using Base.Broadcast: AbstractArrayStyle, DefaultArrayStyle, Broadcasted
 import Base.Broadcast: BroadcastStyle, broadcasted
diff --git a/src/banded/BandedMatrix.jl b/src/banded/BandedMatrix.jl
index 70037d4a..4d1909fb 100644
--- a/src/banded/BandedMatrix.jl
+++ b/src/banded/BandedMatrix.jl
@@ -1008,44 +1008,3 @@ function resize(A::BandedSubBandedMatrix, n::Integer, m::Integer)
     _BandedMatrix(reshape(resize!(vec(copy(bandeddata(A))), (l+u+1)*m), l+u+1, m), n, l,u)
 end
 
-function sum(A::BandedMatrix; dims=:)
-    if dims isa Colon
-        l, u = bandwidths(A)
-        ret = zero(eltype(A))
-        if l + u < 0
-            return ret
-        end
-        n, m = size(A)
-        for i = 1:n, j = rowrange(A, i)
-            ret += A[i, j]
-        end
-        ret
-    elseif dims > 2
-        A
-    elseif dims == 2
-        l, u = bandwidths(A)
-        n, m = size(A)
-        ret = zeros(eltype(A), n, 1)
-        if l + u < 0
-            return ret
-        end
-        for i = 1:n, j = rowrange(A, i)
-            ret[i, 1] += A[i, j]
-        end
-        ret
-    elseif dims == 1
-        l, u = bandwidths(A)
-        n, m = size(A)
-        ret = zeros(eltype(A), 1, m)
-        if l + u < 0
-            return ret
-        end
-        for i = 1:m, j = colrange(A, i)
-            ret[1, i] += A[j, i]
-        end
-        ret
-    else
-        throw(ArgumentError("dimension must be ≥ 1, got $dims"))
-    end
-end
-
diff --git a/src/generic/AbstractBandedMatrix.jl b/src/generic/AbstractBandedMatrix.jl
index 55d1668a..0abeef70 100644
--- a/src/generic/AbstractBandedMatrix.jl
+++ b/src/generic/AbstractBandedMatrix.jl
@@ -336,3 +336,66 @@ if VERSION >= v"1.9"
     copy(A::Adjoint{T,<:AbstractBandedMatrix}) where T = copy(parent(A))'
     copy(A::Transpose{T,<:AbstractBandedMatrix}) where T = transpose(copy(parent(A)))
 end
+
+function sum!(ret::AbstractArray, A::AbstractBandedMatrix)
+    #Behaves similarly to Base.sum!
+    ret .= 0
+    n,m = size(A)
+    s = size(ret)
+    l = length(s)
+    #Check for singleton dimension and perform respective sum
+    if s[1] == 1 && (l == 1 || s[2]==1)
+        for j = 1:m, i = colrange(A, j)
+            ret .+= A[i, j]
+        end
+    elseif s[1] == n && (l == 1 || s[2]==1)
+        for i = 1:n, j = rowrange(A, i)
+            ret[i, 1] += A[i, j]
+        end
+    elseif s[1] == 1 && s[2] == m
+        for j = 1:m, i = colrange(A, j)
+            ret[1, j] += A[i, j]
+        end
+    elseif s[1] == n && s[2] == m
+        ret = A
+    else
+        throw(DimensionMismatch("reduction on matrix of size ($n, $m) with output size $s"))
+    end
+end
+
+function sum(A::AbstractBandedMatrix; dims=:)
+    if dims isa Colon
+        l, u = bandwidths(A)
+        ret = zero(eltype(A))
+        if l + u < 0
+            return ret
+        end
+        n, m = size(A)
+        for j = 1:m, i = colrange(A, j)
+            ret += A[i, j]
+        end
+        ret
+    elseif dims > 2
+        A
+    elseif dims == 2
+        l, u = bandwidths(A)
+        n, m = size(A)
+        ret = zeros(eltype(A), n, 1)
+        if l + u < 0
+            return ret
+        end
+        sum!(ret, A)
+        ret
+    elseif dims == 1
+        l, u = bandwidths(A)
+        n, m = size(A)
+        ret = zeros(eltype(A), 1, m)
+        if l + u < 0
+            return ret
+        end
+        sum!(ret, A)
+        ret
+    else
+        throw(ArgumentError("dimension must be ≥ 1, got $dims"))
+    end
+end
diff --git a/test/test_sum.jl b/test/test_sum.jl
index 54aee662..9a248e08 100644
--- a/test/test_sum.jl
+++ b/test/test_sum.jl
@@ -7,12 +7,12 @@ empty_r = brand(Float64,rand(1:1_000),rand(1:1_000),rand(1:100),rand(-200:-101))
 n,m = size(empty_r)
 matr = Matrix(r)
 @testset "sum" begin
-    @test sum(empty_r) ≈ 0
-    @test sum(empty_r; dims = 2) ≈ zeros(n,1)
-    @test sum(empty_r; dims = 1) ≈ zeros(1,m)
-    @test sum(r) ≈ sum(matr) rtol = 1e-10
-    @test sum(r; dims=2) ≈ sum(matr; dims=2) rtol = 1e-10
-    @test sum(r; dims=1) ≈ sum(matr; dims=1) rtol = 1e-10
+    @test sum(empty_r) == 0
+    @test sum(empty_r; dims = 2) == zeros(n,1)
+    @test sum(empty_r; dims = 1) == zeros(1,m)
+    @test sum(r) == sum(matr)
+    @test sum(r; dims=2) == sum(matr; dims=2)
+    @test sum(r; dims=1) == sum(matr; dims=1)
     @test sum(r; dims=3) == r
     @test_throws ArgumentError sum(r; dims=0)
 end

From cfc895e39ce5e5854d00a9b707b0ae5c8ccb6798 Mon Sep 17 00:00:00 2001
From: Maxim Vassiliev <vassiliev.max@gmail.com>
Date: Mon, 15 Jul 2024 15:38:15 +0100
Subject: [PATCH 11/14] revert tests to \approx

---
 test/test_sum.jl | 6 +++---
 1 file changed, 3 insertions(+), 3 deletions(-)

diff --git a/test/test_sum.jl b/test/test_sum.jl
index 9a248e08..a599dc09 100644
--- a/test/test_sum.jl
+++ b/test/test_sum.jl
@@ -10,9 +10,9 @@ matr = Matrix(r)
     @test sum(empty_r) == 0
     @test sum(empty_r; dims = 2) == zeros(n,1)
     @test sum(empty_r; dims = 1) == zeros(1,m)
-    @test sum(r) == sum(matr)
-    @test sum(r; dims=2) == sum(matr; dims=2)
-    @test sum(r; dims=1) == sum(matr; dims=1)
+    @test sum(r) ≈ sum(matr) rtol = 1e-10
+    @test sum(r; dims=2) ≈ sum(matr; dims=2) rtol = 1e-10
+    @test sum(r; dims=1) ≈ sum(matr; dims=1) rtol = 1e-10
     @test sum(r; dims=3) == r
     @test_throws ArgumentError sum(r; dims=0)
 end

From dc9a82d7c4e9465e2a871e6663ef89a7397d031f Mon Sep 17 00:00:00 2001
From: maximvassiliev <vassiliev.max@gmail.com>
Date: Tue, 16 Jul 2024 14:07:39 +0100
Subject: [PATCH 12/14] make improvements in AbstractBandedMatrix.jl

---
 src/generic/AbstractBandedMatrix.jl | 4 ++--
 1 file changed, 2 insertions(+), 2 deletions(-)

diff --git a/src/generic/AbstractBandedMatrix.jl b/src/generic/AbstractBandedMatrix.jl
index 0abeef70..8e714972 100644
--- a/src/generic/AbstractBandedMatrix.jl
+++ b/src/generic/AbstractBandedMatrix.jl
@@ -339,7 +339,7 @@ end
 
 function sum!(ret::AbstractArray, A::AbstractBandedMatrix)
     #Behaves similarly to Base.sum!
-    ret .= 0
+    fill!(ret, zero(eltype(ret)))
     n,m = size(A)
     s = size(ret)
     l = length(s)
@@ -357,7 +357,7 @@ function sum!(ret::AbstractArray, A::AbstractBandedMatrix)
             ret[1, j] += A[i, j]
         end
     elseif s[1] == n && s[2] == m
-        ret = A
+        copyto!(ret,A)
     else
         throw(DimensionMismatch("reduction on matrix of size ($n, $m) with output size $s"))
     end

From 393bf6218c0cfde36e86580c370c002cf3fed813 Mon Sep 17 00:00:00 2001
From: maximvassiliev <vassiliev.max@gmail.com>
Date: Tue, 16 Jul 2024 14:46:29 +0100
Subject: [PATCH 13/14] test special cases of sum!

---
 src/generic/AbstractBandedMatrix.jl |  2 ++
 test/test_sum.jl                    | 10 ++++++++++
 2 files changed, 12 insertions(+)

diff --git a/src/generic/AbstractBandedMatrix.jl b/src/generic/AbstractBandedMatrix.jl
index 8e714972..c5653c26 100644
--- a/src/generic/AbstractBandedMatrix.jl
+++ b/src/generic/AbstractBandedMatrix.jl
@@ -361,6 +361,8 @@ function sum!(ret::AbstractArray, A::AbstractBandedMatrix)
     else
         throw(DimensionMismatch("reduction on matrix of size ($n, $m) with output size $s"))
     end
+    #return the value to mimic Base.sum!
+    ret
 end
 
 function sum(A::AbstractBandedMatrix; dims=:)
diff --git a/test/test_sum.jl b/test/test_sum.jl
index a599dc09..5ca30fd0 100644
--- a/test/test_sum.jl
+++ b/test/test_sum.jl
@@ -2,6 +2,7 @@ module TestSum
 
 using Test, BandedMatrices, Random
 
+Random.seed!(0)
 r = brand(Float64,rand(1:10_000),rand(1:10_000),rand(-20:100),rand(-20:100))
 empty_r = brand(Float64,rand(1:1_000),rand(1:1_000),rand(1:100),rand(-200:-101))
 n,m = size(empty_r)
@@ -10,11 +11,20 @@ matr = Matrix(r)
     @test sum(empty_r) == 0
     @test sum(empty_r; dims = 2) == zeros(n,1)
     @test sum(empty_r; dims = 1) == zeros(1,m)
+
     @test sum(r) ≈ sum(matr) rtol = 1e-10
     @test sum(r; dims=2) ≈ sum(matr; dims=2) rtol = 1e-10
     @test sum(r; dims=1) ≈ sum(matr; dims=1) rtol = 1e-10
     @test sum(r; dims=3) == r
     @test_throws ArgumentError sum(r; dims=0)
+
+    v = [1.0]
+    sum!(v, r)
+    @test v == sum!(v, Matrix(r))
+    n2, m2 = size(r)
+    v = ones(Float64, n2)
+    @test sum!(v, r) == sum!(v, Matrix(r))
+    @test_throws DimensionMismatch sum!(zeros(Float64, n2 + 1, m2 + 1), r)
 end
 
 end
\ No newline at end of file

From 5ec4b3cdeb59847a87124ee9a189d792b13810ff Mon Sep 17 00:00:00 2001
From: Sheehan Olver <solver@mac.com>
Date: Tue, 16 Jul 2024 19:55:42 +0100
Subject: [PATCH 14/14] add some tests and avoid CI failure

---
 test/test_broadcasting.jl |  2 +-
 test/test_sum.jl          | 10 +++++++---
 2 files changed, 8 insertions(+), 4 deletions(-)

diff --git a/test/test_broadcasting.jl b/test/test_broadcasting.jl
index 720d2e4a..333e1f17 100644
--- a/test/test_broadcasting.jl
+++ b/test/test_broadcasting.jl
@@ -26,7 +26,7 @@ Random.seed!(0)
             @test identity.(A) isa BandedMatrix
             @test bandwidths(identity.(A)) == bandwidths(A)
 
-            @test (z -> exp(z)-1).(A) == (z -> exp(z)-1).(Matrix(A))
+            @test (z -> exp(z)-1).(A) ≈ (z -> exp(z)-1).(Matrix(A)) # for some reason == is breaking on Mac CI
             @test (z -> exp(z)-1).(A) isa BandedMatrix
             @test bandwidths((z -> exp(z)-1).(A)) == bandwidths(A)
 
diff --git a/test/test_sum.jl b/test/test_sum.jl
index 5ca30fd0..af08e767 100644
--- a/test/test_sum.jl
+++ b/test/test_sum.jl
@@ -3,8 +3,8 @@ module TestSum
 using Test, BandedMatrices, Random
 
 Random.seed!(0)
-r = brand(Float64,rand(1:10_000),rand(1:10_000),rand(-20:100),rand(-20:100))
-empty_r = brand(Float64,rand(1:1_000),rand(1:1_000),rand(1:100),rand(-200:-101))
+r = brand(rand(1:10_000),rand(1:10_000),rand(-20:100),rand(-20:100))
+empty_r = brand(rand(1:1_000),rand(1:1_000),rand(1:100),rand(-200:-101))
 n,m = size(empty_r)
 matr = Matrix(r)
 @testset "sum" begin
@@ -22,8 +22,12 @@ matr = Matrix(r)
     sum!(v, r)
     @test v == sum!(v, Matrix(r))
     n2, m2 = size(r)
-    v = ones(Float64, n2)
+    v = ones(n2)
     @test sum!(v, r) == sum!(v, Matrix(r))
+    V = zeros(1,m2)
+    @test sum!(V, r) === V ≈ sum!(zeros(1,m2), Matrix(r))
+    V = zeros(n2,m2)
+    @test sum!(V, r) === V == r
     @test_throws DimensionMismatch sum!(zeros(Float64, n2 + 1, m2 + 1), r)
 end