Make vector times matrix faster

src/flint/fmpz_mat.jl

@@ -1784,16 +1784,6 @@
 addmul!(z::ZZMatrixOrPtr, a::ZZMatrixOrPtr, b::Integer) = addmul!(z, a, flintify(b))
 addmul!(z::ZZMatrixOrPtr, a::IntegerUnionOrPtr, b::ZZMatrixOrPtr) = addmul!(z, b, a)
 
-function mul!(z::Vector{ZZRingElem}, a::ZZMatrixOrPtr, b::Vector{ZZRingElem})
-  @ccall libflint.fmpz_mat_mul_fmpz_vec_ptr(z::Ptr{Ref{ZZRingElem}}, a::Ref{ZZMatrix}, b::Ptr{Ref{ZZRingElem}}, length(b)::Int)::Nothing
-  return z
-end
-
-function mul!(z::Vector{ZZRingElem}, a::Vector{ZZRingElem}, b::ZZMatrixOrPtr)
-  @ccall libflint.fmpz_mat_fmpz_vec_mul_ptr(z::Ptr{Ref{ZZRingElem}}, a::Ptr{Ref{ZZRingElem}}, length(a)::Int, b::Ref{ZZMatrix})::Nothing
-  return z
-end
-
 function Generic.add_one!(a::ZZMatrix, i::Int, j::Int)
   @boundscheck _checkbounds(a, i, j)
   GC.@preserve a begin
@@ -1819,6 +1809,96 @@
   return g
 end
 
+################################################################################
+#
+#  Vector * Matrix and Matrix * Vector
+#
+################################################################################
+
+# Vector{fmpz} * fmpz_mat can be performed using
+# - fmpz_mat_fmpz_vec_mul_ptr
+# - or conversion + fmpz_mat_mul
+#
+# The fmpz_mat_fmpz_vec_mul_ptr variants are not optimized.
+# Thus, if the conversion is negliable, we convert and call fmpz_mat.
+# The conversion is done on the julia side, trying to reduce the number of
+# allocations and objects tracked by the GC.
+
+function _very_unsafe_convert(::Type{ZZMatrix}, a::Vector{ZZRingElem}, row = true)
+  # a must be GC.@preserved
+  # row = true -> make it a row
+  # row = false -> make it a column
+  M = Nemo.@new_struct(ZZMatrix)
+  Me = zeros(Int, length(a))
+  M.entries = reinterpret(Ptr{ZZRingElem}, pointer(Me))
+  if row
+    Mep = [pointer(Me)]
+    M.rows = reinterpret(Ptr{Ptr{ZZRingElem}}, pointer(Mep))
+    M.r = 1
+    M.c = length(a)
+  else
+    M.r = length(a)
+    M.c = 1
+    Mep = [pointer(Me) + 8*(i - 1) for i in 1:length(a)]
+    M.rows = reinterpret(Ptr{Ptr{ZZRingElem}}, pointer(Mep))
+  end
+  for i in 1:length(a)
+    Me[i] = a[i].d
+  end
+  return M, Me, Mep
+end
+
+function mul!_flint(z::Vector{ZZRingElem}, a::ZZMatrixOrPtr, b::Vector{ZZRingElem})
+  ccall((:fmpz_mat_mul_fmpz_vec_ptr, libflint), Nothing,
+        (Ptr{Ref{ZZRingElem}}, Ref{ZZMatrix}, Ptr{Ref{ZZRingElem}}, Int),
+        z, a, b, length(b))
+  return z
+end
+
+function mul!_flint(z::Vector{ZZRingElem}, a::Vector{ZZRingElem}, b::ZZMatrixOrPtr)
+  ccall((:fmpz_mat_fmpz_vec_mul_ptr, libflint), Nothing,
+        (Ptr{Ref{ZZRingElem}}, Ptr{Ref{ZZRingElem}}, Int, Ref{ZZMatrix}),
+        z, a, length(a), b)
+  return z
+end
+
+function mul!(z::Vector{ZZRingElem}, a::ZZMatrixOrPtr, b::Vector{ZZRingElem})
+  # cutoff for the flint method
+  if nrows(a) < 50 && maximum(nbits, a) < 10
+    return mul!_flint(z, a, b)
+  end
+
+  GC.@preserve z b begin
+    bb, dk1, dk2 = _very_unsafe_convert(ZZMatrix, b, false)
+    zz, dk3, dk4 = _very_unsafe_convert(ZZMatrix, z, false)
+    GC.@preserve dk1 dk2 dk3 dk4 begin
+      mul!(zz, a, bb)
+      for i in 1:length(z)
+        z[i].d = unsafe_load(zz.entries, i).d
+      end
+    end
+  end
+  return z
+end
+
+function mul!(z::Vector{ZZRingElem}, a::Vector{ZZRingElem}, b::ZZMatrixOrPtr)
+  # cutoff for the flint method
+  if nrows(b) < 50 && maximum(nbits, b) < 10
+    return mul!_flint(z, a, b)
+  end
+  GC.@preserve z a begin
+    aa, dk1, dk2 = _very_unsafe_convert(ZZMatrix, a)
+    zz, dk3, dk4 = _very_unsafe_convert(ZZMatrix, z)
+    GC.@preserve dk1 dk2 dk3 dk4 begin
+      mul!(zz, aa, b)
+      for i in 1:length(z)
+        z[i].d = unsafe_load(zz.entries, i).d
+      end
+    end
+  end
+  return z
+end
+
 ###############################################################################
 #
 #   Parent object call overloads