Cosmetics

1 files changed, 94 insertions, 73 deletions

diff --git a/iqk_mul_mat.cpp b/iqk_mul_mat.cpp
index 02878372..c10ca438 100644
--- a/iqk_mul_mat.cpp
+++ b/iqk_mul_mat.cpp
@@ -1947,39 +1947,73 @@ struct DequantizerIQ4XS final : public BaseDequantizer<block_iq4_xs> {
     float d;
 };
 
-static const int8_t keven_signs_q2xs[1024] = {
-     1,  1,  1,  1,  1,  1,  1,  1, -1,  1,  1,  1,  1,  1,  1, -1,  1, -1,  1,  1,  1,  1,  1, -1, -1, -1,  1,  1,  1,  1,  1,  1,
-     1,  1, -1,  1,  1,  1,  1, -1, -1,  1, -1,  1,  1,  1,  1,  1,  1, -1, -1,  1,  1,  1,  1,  1, -1, -1, -1,  1,  1,  1,  1, -1,
-     1,  1,  1, -1,  1,  1,  1, -1, -1,  1,  1, -1,  1,  1,  1,  1,  1, -1,  1, -1,  1,  1,  1,  1, -1, -1,  1, -1,  1,  1,  1, -1,
-     1,  1, -1, -1,  1,  1,  1,  1, -1,  1, -1, -1,  1,  1,  1, -1,  1, -1, -1, -1,  1,  1,  1, -1, -1, -1, -1, -1,  1,  1,  1,  1,
-     1,  1,  1,  1, -1,  1,  1, -1, -1,  1,  1,  1, -1,  1,  1,  1,  1, -1,  1,  1, -1,  1,  1,  1, -1, -1,  1,  1, -1,  1,  1, -1,
-     1,  1, -1,  1, -1,  1,  1,  1, -1,  1, -1,  1, -1,  1,  1, -1,  1, -1, -1,  1, -1,  1,  1, -1, -1, -1, -1,  1, -1,  1,  1,  1,
-     1,  1,  1, -1, -1,  1,  1,  1, -1,  1,  1, -1, -1,  1,  1, -1,  1, -1,  1, -1, -1,  1,  1, -1, -1, -1,  1, -1, -1,  1,  1,  1,
-     1,  1, -1, -1, -1,  1,  1, -1, -1,  1, -1, -1, -1,  1,  1,  1,  1, -1, -1, -1, -1,  1,  1,  1, -1, -1, -1, -1, -1,  1,  1, -1,
-     1,  1,  1,  1,  1, -1,  1, -1, -1,  1,  1,  1,  1, -1,  1,  1,  1, -1,  1,  1,  1, -1,  1,  1, -1, -1,  1,  1,  1, -1,  1, -1,
-     1,  1, -1,  1,  1, -1,  1,  1, -1,  1, -1,  1,  1, -1,  1, -1,  1, -1, -1,  1,  1, -1,  1, -1, -1, -1, -1,  1,  1, -1,  1,  1,
-     1,  1,  1, -1,  1, -1,  1,  1, -1,  1,  1, -1,  1, -1,  1, -1,  1, -1,  1, -1,  1, -1,  1, -1, -1, -1,  1, -1,  1, -1,  1,  1,
-     1,  1, -1, -1,  1, -1,  1, -1, -1,  1, -1, -1,  1, -1,  1,  1,  1, -1, -1, -1,  1, -1,  1,  1, -1, -1, -1, -1,  1, -1,  1, -1,
-     1,  1,  1,  1, -1, -1,  1,  1, -1,  1,  1,  1, -1, -1,  1, -1,  1, -1,  1,  1, -1, -1,  1, -1, -1, -1,  1,  1, -1, -1,  1,  1,
-     1,  1, -1,  1, -1, -1,  1, -1, -1,  1, -1,  1, -1, -1,  1,  1,  1, -1, -1,  1, -1, -1,  1,  1, -1, -1, -1,  1, -1, -1,  1, -1,
-     1,  1,  1, -1, -1, -1,  1, -1, -1,  1,  1, -1, -1, -1,  1,  1,  1, -1,  1, -1, -1, -1,  1,  1, -1, -1,  1, -1, -1, -1,  1, -1,
-     1,  1, -1, -1, -1, -1,  1,  1, -1,  1, -1, -1, -1, -1,  1, -1,  1, -1, -1, -1, -1, -1,  1, -1, -1, -1, -1, -1, -1, -1,  1,  1,
-     1,  1,  1,  1,  1,  1, -1, -1, -1,  1,  1,  1,  1,  1, -1,  1,  1, -1,  1,  1,  1,  1, -1,  1, -1, -1,  1,  1,  1,  1, -1, -1,
-     1,  1, -1,  1,  1,  1, -1,  1, -1,  1, -1,  1,  1,  1, -1, -1,  1, -1, -1,  1,  1,  1, -1, -1, -1, -1, -1,  1,  1,  1, -1,  1,
-     1,  1,  1, -1,  1,  1, -1,  1, -1,  1,  1, -1,  1,  1, -1, -1,  1, -1,  1, -1,  1,  1, -1, -1, -1, -1,  1, -1,  1,  1, -1,  1,
-     1,  1, -1, -1,  1,  1, -1, -1, -1,  1, -1, -1,  1,  1, -1,  1,  1, -1, -1, -1,  1,  1, -1,  1, -1, -1, -1, -1,  1,  1, -1, -1,
-     1,  1,  1,  1, -1,  1, -1,  1, -1,  1,  1,  1, -1,  1, -1, -1,  1, -1,  1,  1, -1,  1, -1, -1, -1, -1,  1,  1, -1,  1, -1,  1,
-     1,  1, -1,  1, -1,  1, -1, -1, -1,  1, -1,  1, -1,  1, -1,  1,  1, -1, -1,  1, -1,  1, -1,  1, -1, -1, -1,  1, -1,  1, -1, -1,
-     1,  1,  1, -1, -1,  1, -1, -1, -1,  1,  1, -1, -1,  1, -1,  1,  1, -1,  1, -1, -1,  1, -1,  1, -1, -1,  1, -1, -1,  1, -1, -1,
-     1,  1, -1, -1, -1,  1, -1,  1, -1,  1, -1, -1, -1,  1, -1, -1,  1, -1, -1, -1, -1,  1, -1, -1, -1, -1, -1, -1, -1,  1, -1,  1,
-     1,  1,  1,  1,  1, -1, -1,  1, -1,  1,  1,  1,  1, -1, -1, -1,  1, -1,  1,  1,  1, -1, -1, -1, -1, -1,  1,  1,  1, -1, -1,  1,
-     1,  1, -1,  1,  1, -1, -1, -1, -1,  1, -1,  1,  1, -1, -1,  1,  1, -1, -1,  1,  1, -1, -1,  1, -1, -1, -1,  1,  1, -1, -1, -1,
-     1,  1,  1, -1,  1, -1, -1, -1, -1,  1,  1, -1,  1, -1, -1,  1,  1, -1,  1, -1,  1, -1, -1,  1, -1, -1,  1, -1,  1, -1, -1, -1,
-     1,  1, -1, -1,  1, -1, -1,  1, -1,  1, -1, -1,  1, -1, -1, -1,  1, -1, -1, -1,  1, -1, -1, -1, -1, -1, -1, -1,  1, -1, -1,  1,
-     1,  1,  1,  1, -1, -1, -1, -1, -1,  1,  1,  1, -1, -1, -1,  1,  1, -1,  1,  1, -1, -1, -1,  1, -1, -1,  1,  1, -1, -1, -1, -1,
-     1,  1, -1,  1, -1, -1, -1,  1, -1,  1, -1,  1, -1, -1, -1, -1,  1, -1, -1,  1, -1, -1, -1, -1, -1, -1, -1,  1, -1, -1, -1,  1,
-     1,  1,  1, -1, -1, -1, -1,  1, -1,  1,  1, -1, -1, -1, -1, -1,  1, -1,  1, -1, -1, -1, -1, -1, -1, -1,  1, -1, -1, -1, -1,  1,
-     1,  1, -1, -1, -1, -1, -1, -1, -1,  1, -1, -1, -1, -1, -1,  1,  1, -1, -1, -1, -1, -1, -1,  1, -1, -1, -1, -1, -1, -1, -1, -1,
+//static const int8_t keven_signs_q2xs[1024] = {
+//     1,  1,  1,  1,  1,  1,  1,  1, -1,  1,  1,  1,  1,  1,  1, -1,  1, -1,  1,  1,  1,  1,  1, -1, -1, -1,  1,  1,  1,  1,  1,  1,
+//     1,  1, -1,  1,  1,  1,  1, -1, -1,  1, -1,  1,  1,  1,  1,  1,  1, -1, -1,  1,  1,  1,  1,  1, -1, -1, -1,  1,  1,  1,  1, -1,
+//     1,  1,  1, -1,  1,  1,  1, -1, -1,  1,  1, -1,  1,  1,  1,  1,  1, -1,  1, -1,  1,  1,  1,  1, -1, -1,  1, -1,  1,  1,  1, -1,
+//     1,  1, -1, -1,  1,  1,  1,  1, -1,  1, -1, -1,  1,  1,  1, -1,  1, -1, -1, -1,  1,  1,  1, -1, -1, -1, -1, -1,  1,  1,  1,  1,
+//     1,  1,  1,  1, -1,  1,  1, -1, -1,  1,  1,  1, -1,  1,  1,  1,  1, -1,  1,  1, -1,  1,  1,  1, -1, -1,  1,  1, -1,  1,  1, -1,
+//     1,  1, -1,  1, -1,  1,  1,  1, -1,  1, -1,  1, -1,  1,  1, -1,  1, -1, -1,  1, -1,  1,  1, -1, -1, -1, -1,  1, -1,  1,  1,  1,
+//     1,  1,  1, -1, -1,  1,  1,  1, -1,  1,  1, -1, -1,  1,  1, -1,  1, -1,  1, -1, -1,  1,  1, -1, -1, -1,  1, -1, -1,  1,  1,  1,
+//     1,  1, -1, -1, -1,  1,  1, -1, -1,  1, -1, -1, -1,  1,  1,  1,  1, -1, -1, -1, -1,  1,  1,  1, -1, -1, -1, -1, -1,  1,  1, -1,
+//     1,  1,  1,  1,  1, -1,  1, -1, -1,  1,  1,  1,  1, -1,  1,  1,  1, -1,  1,  1,  1, -1,  1,  1, -1, -1,  1,  1,  1, -1,  1, -1,
+//     1,  1, -1,  1,  1, -1,  1,  1, -1,  1, -1,  1,  1, -1,  1, -1,  1, -1, -1,  1,  1, -1,  1, -1, -1, -1, -1,  1,  1, -1,  1,  1,
+//     1,  1,  1, -1,  1, -1,  1,  1, -1,  1,  1, -1,  1, -1,  1, -1,  1, -1,  1, -1,  1, -1,  1, -1, -1, -1,  1, -1,  1, -1,  1,  1,
+//     1,  1, -1, -1,  1, -1,  1, -1, -1,  1, -1, -1,  1, -1,  1,  1,  1, -1, -1, -1,  1, -1,  1,  1, -1, -1, -1, -1,  1, -1,  1, -1,
+//     1,  1,  1,  1, -1, -1,  1,  1, -1,  1,  1,  1, -1, -1,  1, -1,  1, -1,  1,  1, -1, -1,  1, -1, -1, -1,  1,  1, -1, -1,  1,  1,
+//     1,  1, -1,  1, -1, -1,  1, -1, -1,  1, -1,  1, -1, -1,  1,  1,  1, -1, -1,  1, -1, -1,  1,  1, -1, -1, -1,  1, -1, -1,  1, -1,
+//     1,  1,  1, -1, -1, -1,  1, -1, -1,  1,  1, -1, -1, -1,  1,  1,  1, -1,  1, -1, -1, -1,  1,  1, -1, -1,  1, -1, -1, -1,  1, -1,
+//     1,  1, -1, -1, -1, -1,  1,  1, -1,  1, -1, -1, -1, -1,  1, -1,  1, -1, -1, -1, -1, -1,  1, -1, -1, -1, -1, -1, -1, -1,  1,  1,
+//     1,  1,  1,  1,  1,  1, -1, -1, -1,  1,  1,  1,  1,  1, -1,  1,  1, -1,  1,  1,  1,  1, -1,  1, -1, -1,  1,  1,  1,  1, -1, -1,
+//     1,  1, -1,  1,  1,  1, -1,  1, -1,  1, -1,  1,  1,  1, -1, -1,  1, -1, -1,  1,  1,  1, -1, -1, -1, -1, -1,  1,  1,  1, -1,  1,
+//     1,  1,  1, -1,  1,  1, -1,  1, -1,  1,  1, -1,  1,  1, -1, -1,  1, -1,  1, -1,  1,  1, -1, -1, -1, -1,  1, -1,  1,  1, -1,  1,
+//     1,  1, -1, -1,  1,  1, -1, -1, -1,  1, -1, -1,  1,  1, -1,  1,  1, -1, -1, -1,  1,  1, -1,  1, -1, -1, -1, -1,  1,  1, -1, -1,
+//     1,  1,  1,  1, -1,  1, -1,  1, -1,  1,  1,  1, -1,  1, -1, -1,  1, -1,  1,  1, -1,  1, -1, -1, -1, -1,  1,  1, -1,  1, -1,  1,
+//     1,  1, -1,  1, -1,  1, -1, -1, -1,  1, -1,  1, -1,  1, -1,  1,  1, -1, -1,  1, -1,  1, -1,  1, -1, -1, -1,  1, -1,  1, -1, -1,
+//     1,  1,  1, -1, -1,  1, -1, -1, -1,  1,  1, -1, -1,  1, -1,  1,  1, -1,  1, -1, -1,  1, -1,  1, -1, -1,  1, -1, -1,  1, -1, -1,
+//     1,  1, -1, -1, -1,  1, -1,  1, -1,  1, -1, -1, -1,  1, -1, -1,  1, -1, -1, -1, -1,  1, -1, -1, -1, -1, -1, -1, -1,  1, -1,  1,
+//     1,  1,  1,  1,  1, -1, -1,  1, -1,  1,  1,  1,  1, -1, -1, -1,  1, -1,  1,  1,  1, -1, -1, -1, -1, -1,  1,  1,  1, -1, -1,  1,
+//     1,  1, -1,  1,  1, -1, -1, -1, -1,  1, -1,  1,  1, -1, -1,  1,  1, -1, -1,  1,  1, -1, -1,  1, -1, -1, -1,  1,  1, -1, -1, -1,
+//     1,  1,  1, -1,  1, -1, -1, -1, -1,  1,  1, -1,  1, -1, -1,  1,  1, -1,  1, -1,  1, -1, -1,  1, -1, -1,  1, -1,  1, -1, -1, -1,
+//     1,  1, -1, -1,  1, -1, -1,  1, -1,  1, -1, -1,  1, -1, -1, -1,  1, -1, -1, -1,  1, -1, -1, -1, -1, -1, -1, -1,  1, -1, -1,  1,
+//     1,  1,  1,  1, -1, -1, -1, -1, -1,  1,  1,  1, -1, -1, -1,  1,  1, -1,  1,  1, -1, -1, -1,  1, -1, -1,  1,  1, -1, -1, -1, -1,
+//     1,  1, -1,  1, -1, -1, -1,  1, -1,  1, -1,  1, -1, -1, -1, -1,  1, -1, -1,  1, -1, -1, -1, -1, -1, -1, -1,  1, -1, -1, -1,  1,
+//     1,  1,  1, -1, -1, -1, -1,  1, -1,  1,  1, -1, -1, -1, -1, -1,  1, -1,  1, -1, -1, -1, -1, -1, -1, -1,  1, -1, -1, -1, -1,  1,
+//     1,  1, -1, -1, -1, -1, -1, -1, -1,  1, -1, -1, -1, -1, -1,  1,  1, -1, -1, -1, -1, -1, -1,  1, -1, -1, -1, -1, -1, -1, -1, -1,
+//};
+const uint64_t keven_signs[128] = {
+    0x0101010101010101, 0xff010101010101ff, 0xff0101010101ff01, 0x010101010101ffff,
+    0xff01010101ff0101, 0x0101010101ff01ff, 0x0101010101ffff01, 0xff01010101ffffff,
+    0xff010101ff010101, 0x01010101ff0101ff, 0x01010101ff01ff01, 0xff010101ff01ffff,
+    0x01010101ffff0101, 0xff010101ffff01ff, 0xff010101ffffff01, 0x01010101ffffffff,
+    0xff0101ff01010101, 0x010101ff010101ff, 0x010101ff0101ff01, 0xff0101ff0101ffff,
+    0x010101ff01ff0101, 0xff0101ff01ff01ff, 0xff0101ff01ffff01, 0x010101ff01ffffff,
+    0x010101ffff010101, 0xff0101ffff0101ff, 0xff0101ffff01ff01, 0x010101ffff01ffff,
+    0xff0101ffffff0101, 0x010101ffffff01ff, 0x010101ffffffff01, 0xff0101ffffffffff,
+    0xff01ff0101010101, 0x0101ff01010101ff, 0x0101ff010101ff01, 0xff01ff010101ffff,
+    0x0101ff0101ff0101, 0xff01ff0101ff01ff, 0xff01ff0101ffff01, 0x0101ff0101ffffff,
+    0x0101ff01ff010101, 0xff01ff01ff0101ff, 0xff01ff01ff01ff01, 0x0101ff01ff01ffff,
+    0xff01ff01ffff0101, 0x0101ff01ffff01ff, 0x0101ff01ffffff01, 0xff01ff01ffffffff,
+    0x0101ffff01010101, 0xff01ffff010101ff, 0xff01ffff0101ff01, 0x0101ffff0101ffff,
+    0xff01ffff01ff0101, 0x0101ffff01ff01ff, 0x0101ffff01ffff01, 0xff01ffff01ffffff,
+    0xff01ffffff010101, 0x0101ffffff0101ff, 0x0101ffffff01ff01, 0xff01ffffff01ffff,
+    0x0101ffffffff0101, 0xff01ffffffff01ff, 0xff01ffffffffff01, 0x0101ffffffffffff,
+    0xffff010101010101, 0x01ff0101010101ff, 0x01ff01010101ff01, 0xffff01010101ffff,
+    0x01ff010101ff0101, 0xffff010101ff01ff, 0xffff010101ffff01, 0x01ff010101ffffff,
+    0x01ff0101ff010101, 0xffff0101ff0101ff, 0xffff0101ff01ff01, 0x01ff0101ff01ffff,
+    0xffff0101ffff0101, 0x01ff0101ffff01ff, 0x01ff0101ffffff01, 0xffff0101ffffffff,
+    0x01ff01ff01010101, 0xffff01ff010101ff, 0xffff01ff0101ff01, 0x01ff01ff0101ffff,
+    0xffff01ff01ff0101, 0x01ff01ff01ff01ff, 0x01ff01ff01ffff01, 0xffff01ff01ffffff,
+    0xffff01ffff010101, 0x01ff01ffff0101ff, 0x01ff01ffff01ff01, 0xffff01ffff01ffff,
+    0x01ff01ffffff0101, 0xffff01ffffff01ff, 0xffff01ffffffff01, 0x01ff01ffffffffff,
+    0x01ffff0101010101, 0xffffff01010101ff, 0xffffff010101ff01, 0x01ffff010101ffff,
+    0xffffff0101ff0101, 0x01ffff0101ff01ff, 0x01ffff0101ffff01, 0xffffff0101ffffff,
+    0xffffff01ff010101, 0x01ffff01ff0101ff, 0x01ffff01ff01ff01, 0xffffff01ff01ffff,
+    0x01ffff01ffff0101, 0xffffff01ffff01ff, 0xffffff01ffffff01, 0x01ffff01ffffffff,
+    0xffffffff01010101, 0x01ffffff010101ff, 0x01ffffff0101ff01, 0xffffffff0101ffff,
+    0x01ffffff01ff0101, 0xffffffff01ff01ff, 0xffffffff01ffff01, 0x01ffffff01ffffff,
+    0x01ffffffff010101, 0xffffffffff0101ff, 0xffffffffff01ff01, 0x01ffffffff01ffff,
+    0xffffffffffff0101, 0x01ffffffffff01ff, 0x01ffffffffffff01, 0xffffffffffffffff,
 };
 
 struct DequantizerIQ2XXS final : public BaseDequantizer<block_iq2_xxs> {
@@ -1991,55 +2025,42 @@ struct DequantizerIQ2XXS final : public BaseDequantizer<block_iq2_xxs> {
     template <typename Q8>
     inline int32x4x2_t new_block(int i, const Q8& /*q8*/, float32x4_t * /*acc*/) {
         d = 0.125f * GGML_FP16_TO_FP32(x[i].d);
+
         auto tmp = vld1q_u32_x4((const uint32_t *)x[i].qs);
         data.val[0] = vuzp1q_u32(tmp.val[0], tmp.val[1]);  // codebook indices for blocks 0...3
-        data.val[1] = vuzp1q_u32(tmp.val[2], tmp.val[3]);  // codebook indices for blocks 4...7
-        data.val[2] = vuzp2q_u32(tmp.val[0], tmp.val[1]);  // scales and signs for blocks 0...3
+        data.val[1] = vuzp2q_u32(tmp.val[0], tmp.val[1]);  // scales and signs for blocks 0...3
+        data.val[2] = vuzp1q_u32(tmp.val[2], tmp.val[3]);  // codebook indices for blocks 4...7
         data.val[3] = vuzp2q_u32(tmp.val[2], tmp.val[3]);  // scales and signs for blocks 4...7
+
         int32x4x2_t scales;
-        scales.val[0] = vreinterpretq_s32_u32(vorrq_u32(vshlq_n_u32(vshrq_n_u32(data.val[2], 28), 1), vdupq_n_u32(1)));
+        scales.val[0] = vreinterpretq_s32_u32(vorrq_u32(vshlq_n_u32(vshrq_n_u32(data.val[1], 28), 1), vdupq_n_u32(1)));
         scales.val[1] = vreinterpretq_s32_u32(vorrq_u32(vshlq_n_u32(vshrq_n_u32(data.val[3], 28), 1), vdupq_n_u32(1)));
         return scales;
     }
 
     inline void prepare(int /*i*/, int j) {
-        const uint8_t * idx = (const uint8_t *)(data.val + j);
-        bits.b1.val[0] = vcombine_s8(vld1_s8((const int8_t *)(iq2xxs_grid + idx[ 0])), vld1_s8((const int8_t *)(iq2xxs_grid + idx[ 1])));
-        bits.b1.val[1] = vcombine_s8(vld1_s8((const int8_t *)(iq2xxs_grid + idx[ 2])), vld1_s8((const int8_t *)(iq2xxs_grid + idx[ 3])));
-        bits.b1.val[2] = vcombine_s8(vld1_s8((const int8_t *)(iq2xxs_grid + idx[ 4])), vld1_s8((const int8_t *)(iq2xxs_grid + idx[ 5])));
-        bits.b1.val[3] = vcombine_s8(vld1_s8((const int8_t *)(iq2xxs_grid + idx[ 6])), vld1_s8((const int8_t *)(iq2xxs_grid + idx[ 7])));
-        bits.b2.val[0] = vcombine_s8(vld1_s8((const int8_t *)(iq2xxs_grid + idx[ 8])), vld1_s8((const int8_t *)(iq2xxs_grid + idx[ 9])));
-        bits.b2.val[1] = vcombine_s8(vld1_s8((const int8_t *)(iq2xxs_grid + idx[10])), vld1_s8((const int8_t *)(iq2xxs_grid + idx[11])));
-        bits.b2.val[2] = vcombine_s8(vld1_s8((const int8_t *)(iq2xxs_grid + idx[12])), vld1_s8((const int8_t *)(iq2xxs_grid + idx[13])));
-        bits.b2.val[3] = vcombine_s8(vld1_s8((const int8_t *)(iq2xxs_grid + idx[14])), vld1_s8((const int8_t *)(iq2xxs_grid + idx[15])));
-        const uint64_t * signs64 = (const uint64_t *)keven_signs_q2xs;
-        const uint32_t * sidx = (const uint32_t *)(data.val + 2 + j);
-        bits.b1.val[0] = vmulq_s8(bits.b1.val[0], vcombine_s8(vld1_s8((const int8_t *)(signs64 + ((sidx[0] >>  0) & 127))), vld1_s8((const int8_t *)(signs64 + ((sidx[0] >>  7) & 127)))));
-        bits.b1.val[1] = vmulq_s8(bits.b1.val[1], vcombine_s8(vld1_s8((const int8_t *)(signs64 + ((sidx[0] >> 14) & 127))), vld1_s8((const int8_t *)(signs64 + ((sidx[0] >> 21) & 127)))));
-        bits.b1.val[2] = vmulq_s8(bits.b1.val[2], vcombine_s8(vld1_s8((const int8_t *)(signs64 + ((sidx[1] >>  0) & 127))), vld1_s8((const int8_t *)(signs64 + ((sidx[1] >>  7) & 127)))));
-        bits.b1.val[3] = vmulq_s8(bits.b1.val[3], vcombine_s8(vld1_s8((const int8_t *)(signs64 + ((sidx[1] >> 14) & 127))), vld1_s8((const int8_t *)(signs64 + ((sidx[1] >> 21) & 127)))));
-        bits.b2.val[0] = vmulq_s8(bits.b2.val[0], vcombine_s8(vld1_s8((const int8_t *)(signs64 + ((sidx[2] >>  0) & 127))), vld1_s8((const int8_t *)(signs64 + ((sidx[2] >>  7) & 127)))));
-        bits.b2.val[1] = vmulq_s8(bits.b2.val[1], vcombine_s8(vld1_s8((const int8_t *)(signs64 + ((sidx[2] >> 14) & 127))), vld1_s8((const int8_t *)(signs64 + ((sidx[2] >> 21) & 127)))));
-        bits.b2.val[2] = vmulq_s8(bits.b2.val[2], vcombine_s8(vld1_s8((const int8_t *)(signs64 + ((sidx[3] >>  0) & 127))), vld1_s8((const int8_t *)(signs64 + ((sidx[3] >>  7) & 127)))));
-        bits.b2.val[3] = vmulq_s8(bits.b2.val[3], vcombine_s8(vld1_s8((const int8_t *)(signs64 + ((sidx[3] >> 14) & 127))), vld1_s8((const int8_t *)(signs64 + ((sidx[3] >> 21) & 127)))));
-        //auto mask = vdupq_n_u32(127);
-        //uint32x4_t sindex;
-        //sindex = vandq_u32(data.val[2+j], mask);
-        //mask = vshlq_n_u32(mask, 7);
-        //sindex = vorrq_u32(sindex, vshlq_n_u32(vandq_u32(data.val[2+j], mask), 1));
-        //mask = vshlq_n_u32(mask, 7);
-        //sindex = vorrq_u32(sindex, vshlq_n_u32(vandq_u32(data.val[2+j], mask), 2));
-        //mask = vshlq_n_u32(mask, 7);
-        //sindex = vorrq_u32(sindex, vshlq_n_u32(vandq_u32(data.val[2+j], mask), 3));
-        //const uint8_t * sidx = (const uint8_t *)&sindex;
-        //bits.b1.val[0] = vmulq_s8(bits.b1.val[0], vcombine_s8(vld1_s8((const int8_t *)(signs64 + sidx[ 0])), vld1_s8((const int8_t *)(signs64 + sidx[ 1]))));
-        //bits.b1.val[1] = vmulq_s8(bits.b1.val[1], vcombine_s8(vld1_s8((const int8_t *)(signs64 + sidx[ 2])), vld1_s8((const int8_t *)(signs64 + sidx[ 3]))));
-        //bits.b1.val[2] = vmulq_s8(bits.b1.val[2], vcombine_s8(vld1_s8((const int8_t *)(signs64 + sidx[ 4])), vld1_s8((const int8_t *)(signs64 + sidx[ 5]))));
-        //bits.b1.val[3] = vmulq_s8(bits.b1.val[3], vcombine_s8(vld1_s8((const int8_t *)(signs64 + sidx[ 6])), vld1_s8((const int8_t *)(signs64 + sidx[ 7]))));
-        //bits.b2.val[0] = vmulq_s8(bits.b2.val[0], vcombine_s8(vld1_s8((const int8_t *)(signs64 + sidx[ 8])), vld1_s8((const int8_t *)(signs64 + sidx[ 9]))));
-        //bits.b2.val[1] = vmulq_s8(bits.b2.val[1], vcombine_s8(vld1_s8((const int8_t *)(signs64 + sidx[10])), vld1_s8((const int8_t *)(signs64 + sidx[11]))));
-        //bits.b2.val[2] = vmulq_s8(bits.b2.val[2], vcombine_s8(vld1_s8((const int8_t *)(signs64 + sidx[12])), vld1_s8((const int8_t *)(signs64 + sidx[13]))));
-        //bits.b2.val[3] = vmulq_s8(bits.b2.val[3], vcombine_s8(vld1_s8((const int8_t *)(signs64 + sidx[14])), vld1_s8((const int8_t *)(signs64 + sidx[15]))));
+        const uint8_t * idx = (const uint8_t *)(data.val + 2*j);
+        bits.b1.val[0] = vreinterpretq_u8_u64(uint64x2_t{iq2xxs_grid[idx[ 0]], iq2xxs_grid[idx[ 1]]});
+        bits.b1.val[1] = vreinterpretq_u8_u64(uint64x2_t{iq2xxs_grid[idx[ 2]], iq2xxs_grid[idx[ 3]]});
+        bits.b1.val[2] = vreinterpretq_u8_u64(uint64x2_t{iq2xxs_grid[idx[ 4]], iq2xxs_grid[idx[ 5]]});
+        bits.b1.val[3] = vreinterpretq_u8_u64(uint64x2_t{iq2xxs_grid[idx[ 6]], iq2xxs_grid[idx[ 7]]});
+        bits.b2.val[0] = vreinterpretq_u8_u64(uint64x2_t{iq2xxs_grid[idx[ 8]], iq2xxs_grid[idx[ 9]]});
+        bits.b2.val[1] = vreinterpretq_u8_u64(uint64x2_t{iq2xxs_grid[idx[10]], iq2xxs_grid[idx[11]]});
+        bits.b2.val[2] = vreinterpretq_u8_u64(uint64x2_t{iq2xxs_grid[idx[12]], iq2xxs_grid[idx[13]]});
+        bits.b2.val[3] = vreinterpretq_u8_u64(uint64x2_t{iq2xxs_grid[idx[14]], iq2xxs_grid[idx[15]]});
+
+        //const uint64_t * signs64 = (const uint64_t *)keven_signs_q2xs;
+        const uint64_t * signs = keven_signs;
+        const uint32_t * sidx = (const uint32_t *)(data.val + 2*j+1);
+
+        bits.b1.val[0] = vmulq_s8(bits.b1.val[0], vcombine_s8(vld1_s8((const int8_t *)(signs + ((sidx[0] >>  0) & 127))), vld1_s8((const int8_t *)(signs + ((sidx[0] >>  7) & 127)))));
+        bits.b1.val[1] = vmulq_s8(bits.b1.val[1], vcombine_s8(vld1_s8((const int8_t *)(signs + ((sidx[0] >> 14) & 127))), vld1_s8((const int8_t *)(signs + ((sidx[0] >> 21) & 127)))));
+        bits.b1.val[2] = vmulq_s8(bits.b1.val[2], vcombine_s8(vld1_s8((const int8_t *)(signs + ((sidx[1] >>  0) & 127))), vld1_s8((const int8_t *)(signs + ((sidx[1] >>  7) & 127)))));
+        bits.b1.val[3] = vmulq_s8(bits.b1.val[3], vcombine_s8(vld1_s8((const int8_t *)(signs + ((sidx[1] >> 14) & 127))), vld1_s8((const int8_t *)(signs + ((sidx[1] >> 21) & 127)))));
+        bits.b2.val[0] = vmulq_s8(bits.b2.val[0], vcombine_s8(vld1_s8((const int8_t *)(signs + ((sidx[2] >>  0) & 127))), vld1_s8((const int8_t *)(signs + ((sidx[2] >>  7) & 127)))));
+        bits.b2.val[1] = vmulq_s8(bits.b2.val[1], vcombine_s8(vld1_s8((const int8_t *)(signs + ((sidx[2] >> 14) & 127))), vld1_s8((const int8_t *)(signs + ((sidx[2] >> 21) & 127)))));
+        bits.b2.val[2] = vmulq_s8(bits.b2.val[2], vcombine_s8(vld1_s8((const int8_t *)(signs + ((sidx[3] >>  0) & 127))), vld1_s8((const int8_t *)(signs + ((sidx[3] >>  7) & 127)))));
+        bits.b2.val[3] = vmulq_s8(bits.b2.val[3], vcombine_s8(vld1_s8((const int8_t *)(signs + ((sidx[3] >> 14) & 127))), vld1_s8((const int8_t *)(signs + ((sidx[3] >> 21) & 127)))));
     }
 
     uint32x4x4_t data;