diff --git a/src/Octavian.jl b/src/Octavian.jl
index 1866e95..67bcd54 100644
--- a/src/Octavian.jl
+++ b/src/Octavian.jl
@@ -12,8 +12,10 @@ include("macros.jl")
 include("types.jl")
 
 include("block_sizes.jl")
-include("macrokernel.jl")
+include("macrokernels.jl")
 include("matmul.jl")
+include("memory_buffer.jl")
+include("pointer_matrix.jl")
 include("utils.jl")
 
 const BCACHE = UInt8[]
diff --git a/src/block_sizes.jl b/src/block_sizes.jl
index e923af3..9f2a424 100644
--- a/src/block_sizes.jl
+++ b/src/block_sizes.jl
@@ -53,11 +53,7 @@ function block_sizes(::Type{T}) where {T}
     st = VectorizationBase.static_sizeof(T)
 
     L2 = (StaticInt{_L2}() - StaticInt{_L1}()) ÷ st
-    if 2_L2 * L2c > _L3 * L3c
-        L3 = StaticInt{_L3}() ÷ st
-    else
-        L3 = (StaticInt{_L3}() - StaticInt{_L2}()) ÷ st
-    end
+    L3 = _calculate_L3(StaticInt{_L2}(), StaticInt{_L3}(), st, VectorizationBase.CACHE_INCLUSIVITY[3])
 
     W = VectorizationBase.pick_vector_width_val(T)
     Mr = StaticInt{LoopVectorization.mᵣ}()
@@ -69,3 +65,11 @@ function block_sizes(::Type{T}) where {T}
 
     Mc, Kc, Nc
 end
+
+@inline function _calculate_L3(_L2, _L3, st, cache_inclusivity_3::Bool)
+    if cache_inclusivity_3
+        return (_L3 - _L2) ÷ st
+    else
+        return _L3 ÷ st
+    end
+end
diff --git a/src/macrokernel.jl b/src/macrokernels.jl
similarity index 100%
rename from src/macrokernel.jl
rename to src/macrokernels.jl
diff --git a/src/memory_buffer.jl b/src/memory_buffer.jl
new file mode 100644
index 0000000..b804c82
--- /dev/null
+++ b/src/memory_buffer.jl
@@ -0,0 +1,5 @@
+Base.unsafe_convert(::Type{Ptr{T}}, m::MemoryBuffer) where {T} = Base.unsafe_convert(Ptr{T}, Base.pointer_from_objref(m))
+@inline MemoryBuffer(::StaticInt{L}, ::Type{T}) where {L,T} = MemoryBuffer{L,T}(undef)
+@inline function L2Buffer(::Type{T}) where {T}
+    MemoryBuffer(StaticInt{VectorizationBase.CACHE_SIZE[2]}() ÷ VectorizationBase.static_sizeof(T), T)
+end
diff --git a/src/pointer_matrix.jl b/src/pointer_matrix.jl
new file mode 100644
index 0000000..fcc7c33
--- /dev/null
+++ b/src/pointer_matrix.jl
@@ -0,0 +1,33 @@
+PointerMatrix(p::P, s::S) where {T,P<:VectorizationBase.AbstractStridedPointer{T},S} = PointerMatrix{T,P,S}(p, s)
+Base.size(A::PointerMatrix) = map(Int, A.s)
+VectorizationBase.stridedpointer(A::PointerMatrix) = A.p
+Base.unsafe_convert(::Type{Ptr{T}}, A::PointerMatrix{T}) where {T} = pointer(A.p)
+@inline function Base.getindex(A::PointerMatrix, i::Integer, j::Integer)
+    @boundscheck checkbounds(A, i, j)
+    VectorizationBase.vload(VectorizationBase.stridedpointer(A), (i,j))
+end
+@inline function Base.getindex(A::PointerMatrix, i::Integer)
+    @boundscheck checkbounds(A, i)
+    VectorizationBase.vload(VectorizationBase.stridedpointer(A), (i-1,))
+end
+@inline function Base.setindex!(A::PointerMatrix{T}, v, i::Integer, j::Integer) where {T}
+    @boundscheck checkbounds(A, i, j)
+    VectorizationBase.vstore!(VectorizationBase.stridedpointer(A), convert(T, v), (i,j))
+    v
+end
+@inline function Base.setindex!(A::PointerMatrix{T}, v, i::Integer) where {T}
+    @boundscheck checkbounds(A, i)
+    VectorizationBase.vstore!(VectorizationBase.stridedpointer(A), convert(T, v), (i-1,))
+    v
+end
+
+function PointerMatrix(Bptr::Ptr{T}, (M,N), padcols::Bool = false) where {T}
+    st = VectorizationBase.static_sizeof(T)
+    _M = padcols ? VectorizationBase.align(M, T) : M
+    # Should maybe add a more convenient column major constructor
+    Bsptr = VectorizationBase.stridedpointer(
+        Bptr, VectorizationBase.ArrayInterface.Contiguous{1}(), VectorizationBase.ArrayInterface.ContiguousBatch{0}(),
+        VectorizationBase.ArrayInterface.StrideRank{(1,2)}(), (st, _M*st), (StaticInt{1}(),StaticInt{1}())
+    )
+    PointerMatrix(Bsptr, (M,N))
+end
diff --git a/src/types.jl b/src/types.jl
index c3b47d4..72c8ba4 100644
--- a/src/types.jl
+++ b/src/types.jl
@@ -3,48 +3,7 @@ struct PointerMatrix{T,P<:VectorizationBase.AbstractStridedPointer,S<:Tuple{Vara
     s::S
 end
 
-PointerMatrix(p::P, s::S) where {T,P<:VectorizationBase.AbstractStridedPointer{T},S} = PointerMatrix{T,P,S}(p, s)
-Base.size(A::PointerMatrix) = map(Int, A.s)
-VectorizationBase.stridedpointer(A::PointerMatrix) = A.p
-Base.unsafe_convert(::Type{Ptr{T}}, A::PointerMatrix{T}) where {T} = pointer(A.p)
-@inline function Base.getindex(A::PointerMatrix, i::Integer, j::Integer)
-    @boundscheck checkbounds(A, i, j)
-    VectorizationBase.vload(VectorizationBase.stridedpointer(A), (i,j))
-end
-@inline function Base.getindex(A::PointerMatrix, i::Integer)
-    @boundscheck checkbounds(A, i)
-    VectorizationBase.vload(VectorizationBase.stridedpointer(A), (i-1,))
-end
-@inline function Base.setindex!(A::PointerMatrix{T}, v, i::Integer, j::Integer) where {T}
-    @boundscheck checkbounds(A, i, j)
-    VectorizationBase.vstore!(VectorizationBase.stridedpointer(A), convert(T, v), (i,j))
-    v
-end
-@inline function Base.setindex!(A::PointerMatrix{T}, v, i::Integer) where {T}
-    @boundscheck checkbounds(A, i)
-    VectorizationBase.vstore!(VectorizationBase.stridedpointer(A), convert(T, v), (i-1,))
-    v
-end
-
-function PointerMatrix(Bptr::Ptr{T}, (M,N), padcols::Bool = false) where {T}
-    st = VectorizationBase.static_sizeof(T)
-    _M = padcols ? VectorizationBase.align(M, T) : M
-    # Should maybe add a more convenient column major constructor
-    Bsptr = VectorizationBase.stridedpointer(
-        Bptr, VectorizationBase.ArrayInterface.Contiguous{1}(), VectorizationBase.ArrayInterface.ContiguousBatch{0}(),
-        VectorizationBase.ArrayInterface.StrideRank{(1,2)}(), (st, _M*st), (StaticInt{1}(),StaticInt{1}())
-    )
-    PointerMatrix(Bsptr, (M,N))
-end
-
-
 mutable struct MemoryBuffer{L,T}
     data::NTuple{L,T}
     MemoryBuffer{L,T}(::UndefInitializer) where {L,T} = new{L,T}()
 end
-
-Base.unsafe_convert(::Type{Ptr{T}}, m::MemoryBuffer) where {T} = Base.unsafe_convert(Ptr{T}, Base.pointer_from_objref(m))
-@inline MemoryBuffer(::StaticInt{L}, ::Type{T}) where {L,T} = MemoryBuffer{L,T}(undef)
-@inline function L2Buffer(::Type{T}) where {T}
-    MemoryBuffer(StaticInt{VectorizationBase.CACHE_SIZE[2]}() ÷ VectorizationBase.static_sizeof(T), T)
-end
diff --git a/src/utils.jl b/src/utils.jl
index f1a0a62..a2842cc 100644
--- a/src/utils.jl
+++ b/src/utils.jl
@@ -1,5 +1,5 @@
 check_sizes(::StaticInt{M}, ::StaticInt{M}) where {M} = StaticInt{M}()
-check_sizes(::StaticInt{M}, ::StaticInt{N}) where {M,N} = throw("$M ≠ $N")
+check_sizes(::StaticInt{M}, ::StaticInt{N}) where {M,N} = throw(ErrorException("$M ≠ $N"))
 check_sizes(::StaticInt{M}, m) where {M} = (@assert M == m; StaticInt{M}())
 check_sizes(m, ::StaticInt{M}) where {M} = (@assert M == m; StaticInt{M}())
 check_sizes(m, n) = (@assert m == n; m)
diff --git a/test/block_sizes.jl b/test/block_sizes.jl
new file mode 100644
index 0000000..6a385ba
--- /dev/null
+++ b/test/block_sizes.jl
@@ -0,0 +1,4 @@
+@time @testset "block_sizes" begin
+    @test Octavian._calculate_L3(1, 1, 1, true) == 0
+    @test Octavian._calculate_L3(1, 1, 1, false) == 1
+end
diff --git a/test/macrokernels.jl b/test/macrokernels.jl
new file mode 100644
index 0000000..9bf9ee6
--- /dev/null
+++ b/test/macrokernels.jl
@@ -0,0 +1,16 @@
+@time @testset "Macrokernels" begin
+    m = 20
+    n = 30
+    k = 40
+    A1 = rand(Float64, m, k)
+    B1 = rand(Float64, k, n)
+    C1 = rand(Float64, m, n)
+    A2 = deepcopy(A1)
+    B2 = deepcopy(B1)
+    C2 = deepcopy(C1)
+    α = Float64(2.0)
+    β = Float64(2.0)
+    Octavian.macrokernel!(C1, A1, B1, α, β)
+    C2 = α*A2*B2 + β*C2
+    @test C1 ≈ C2
+end
diff --git a/test/matmul-coverage.jl b/test/matmul_coverage.jl
similarity index 100%
rename from test/matmul-coverage.jl
rename to test/matmul_coverage.jl
diff --git a/test/pointer_matrix.jl b/test/pointer_matrix.jl
new file mode 100644
index 0000000..ff00a1b
--- /dev/null
+++ b/test/pointer_matrix.jl
@@ -0,0 +1,12 @@
+@time Test.@testset "PointerMatrix" begin
+    mem = Octavian.L2Buffer(Float64);
+    ptr = Base.unsafe_convert(Ptr{Float64}, mem)
+    block = Octavian.PointerMatrix(ptr, (10, 20))
+    Test.@test Base.unsafe_convert(Ptr{Float64}, block) == pointer(block.p)
+    GC.@preserve mem begin
+        block[1] = 2.3
+        Test.@test block[1] == 2.3
+        block[4, 5] = 67.89
+        Test.@test block[4, 5] == 67.89
+    end
+end
diff --git a/test/runtests.jl b/test/runtests.jl
index 45cbecf..4f92efe 100644
--- a/test/runtests.jl
+++ b/test/runtests.jl
@@ -5,16 +5,17 @@ import InteractiveUtils
 import LinearAlgebra
 import Test
 
-using Test: @test, @testset
+using Test: @testset, @test, @test_throws
 
-include("test-suite-preamble.jl")
+include("test_suite_preamble.jl")
 
+include("block_sizes.jl")
+include("macrokernels.jl")
 include("macros.jl")
+include("matmul_coverage.jl")
+include("pointer_matrix.jl")
+include("utils.jl")
 
-if (run_all_tests) || (coverage)
-    include("matmul-coverage.jl")
-end
-
-if (run_all_tests) || (!coverage)
+if !coverage
     include("matmul.jl")
 end
diff --git a/test/test-suite-preamble.jl b/test/test_suite_preamble.jl
similarity index 67%
rename from test/test-suite-preamble.jl
rename to test/test_suite_preamble.jl
index 2465a35..1dca2e4 100644
--- a/test/test-suite-preamble.jl
+++ b/test/test_suite_preamble.jl
@@ -11,9 +11,3 @@ end
 const coverage = is_coverage()
 
 @info("Code coverage is $(coverage ? "enabled" : "disabled")")
-
-const run_all_tests = get(ENV, "RUN_ALL_TESTS", "false") == "true"
-
-if run_all_tests
-    @info("RUN_ALL_TESTS is $(run_all_tests)")
-end
diff --git a/test/utils.jl b/test/utils.jl
new file mode 100644
index 0000000..6062834
--- /dev/null
+++ b/test/utils.jl
@@ -0,0 +1,12 @@
+@time @testset "utils" begin
+    @testset "check_sizes" begin
+        a = Octavian.StaticInt{1}()
+        b = Octavian.StaticInt{2}()
+        @test Octavian.check_sizes(a, a) == a
+        @test_throws ErrorException Octavian.check_sizes(a, b)
+        @test Octavian.check_sizes(a, 1) == a
+        @test_throws AssertionError Octavian.check_sizes(a, 100)
+        @test Octavian.check_sizes(2, b) == b
+        @test_throws AssertionError Octavian.check_sizes(200, b)
+    end
+end