Vector256<ulong> accumulator0 = Vector256.Create(Prime32.Prime2, Prime64.Prime0, Prime64.Prime1, Prime64.Prime2);
        Vector256<ulong> accumulator1 = Vector256.Create(Prime64.Prime3, Prime32.Prime1, Prime64.Prime4, Prime32.Prime0);

        [MethodImpl(MethodImplOptions.AggressiveInlining)]
        static ValueTuple<Vector256<ulong>, Vector256<ulong>> Pass(byte* data, byte* secret, in Vector256<ulong> accumulator0, in Vector256<ulong> accumulator1, uint offset) {
            PrefetchNonTemporalNext(data + (offset * StripeLength) + 0x040);
            PrefetchNonTemporalNext(data + (offset * StripeLength) + 0x080);

            byte* lData = data + (offset * StripeLength);
            byte* lSecret = secret + (offset * 8u);

            var dataVec0 = *(Vector256<ulong>*)(lData + 0x00u);
            var dataVec1 = *(Vector256<ulong>*)(lData + 0x20u);
            var keyVec0 = *(Vector256<ulong>*)(lSecret + 0x00u);
            var keyVec1 = *(Vector256<ulong>*)(lSecret + 0x20u);

            var dataKey0 = Avx2.Xor(dataVec0, keyVec0);
            var dataKey1 = Avx2.Xor(dataVec1, keyVec1);
            var dataKeyLo0 = Avx2.Shuffle(dataKey0.AsUInt32(), ShuffleDataKey);
            var dataKeyLo1 = Avx2.Shuffle(dataKey1.AsUInt32(), ShuffleDataKey);
            var product0 = Avx2.Multiply(dataKey0.AsUInt32(), dataKeyLo0);
            var product1 = Avx2.Multiply(dataKey1.AsUInt32(), dataKeyLo1);

            var dataSwap0 = Avx2.Shuffle(dataVec0.AsUInt32(), ShuffleDataSwap);
            var dataSwap1 = Avx2.Shuffle(dataVec1.AsUInt32(), ShuffleDataSwap);
            var addend0 = accumulator0;
            var addend1 = accumulator1;
            var sum0 = Avx2.Add(addend0, dataSwap0.AsUInt64());
            var sum1 = Avx2.Add(addend1, dataSwap1.AsUInt64());

            var result0 = Avx2.Add(product0, sum0);
            var result1 = Avx2.Add(product1, sum1);

            addend0 = result0;
            addend1 = result1;

            var dataVec2 = *(Vector256<ulong>*)(lData + 0x40u);
            var dataVec3 = *(Vector256<ulong>*)(lData + 0x60u);
            var keyVec2 = *(Vector256<ulong>*)(lSecret + 0x08u);
            var keyVec3 = *(Vector256<ulong>*)(lSecret + 0x28u);

            var dataKey2 = Avx2.Xor(dataVec2, keyVec2);
            var dataKey3 = Avx2.Xor(dataVec3, keyVec3);
            var dataKeyLo2 = Avx2.Shuffle(dataKey2.AsUInt32(), ShuffleDataKey);
            var dataKeyLo3 = Avx2.Shuffle(dataKey3.AsUInt32(), ShuffleDataKey);
            var product2 = Avx2.Multiply(dataKey2.AsUInt32(), dataKeyLo2);
            var product3 = Avx2.Multiply(dataKey3.AsUInt32(), dataKeyLo3);

            var dataSwap2 = Avx2.Shuffle(dataVec2.AsUInt32(), ShuffleDataSwap);
            var dataSwap3 = Avx2.Shuffle(dataVec3.AsUInt32(), ShuffleDataSwap);
            var sum2 = Avx2.Add(addend0, dataSwap2.AsUInt64());
            var sum3 = Avx2.Add(addend1, dataSwap3.AsUInt64());

            var result2 = Avx2.Add(product2, sum2);
            var result3 = Avx2.Add(product3, sum3);

            return (
                result2,
                result3
            );
        }

        {
            // 0
            (accumulator0, accumulator1) = Pass(data, secret, in accumulator0, in accumulator1, 0);
            (accumulator0, accumulator1) = Pass(data, secret, in accumulator0, in accumulator1, 2);
            (accumulator0, accumulator1) = Pass(data, secret, in accumulator0, in accumulator1, 4);
            (accumulator0, accumulator1) = Pass(data, secret, in accumulator0, in accumulator1, 6);
            (accumulator0, accumulator1) = Pass(data, secret, in accumulator0, in accumulator1, 8);
            (accumulator0, accumulator1) = Pass(data, secret, in accumulator0, in accumulator1, 10);
            (accumulator0, accumulator1) = Pass(data, secret, in accumulator0, in accumulator1, 12);