Replace vshll_n_u32(v, 0) with vmovl_u32 in C4/C4S2 GEMM/IGEMM microkernels AArch32 GCC doesn't support vshll_n_u32 with shift count of 0. PiperOrigin-RevId: 418834748

commit: dc54e12aae2211d39f53d5f77069019aa7903d8d [log] [tgz]
author: Marat Dukhan <maratek@google.com> Wed Dec 29 12:56:08 2021 -0800
committer: XNNPACK Team <xnnpack-github-robot@google.com> Wed Dec 29 12:57:19 2021 -0800
tree: b3d268b9f5b8d06e270f459eceff87e1a1f240e1
parent: e1228e32037d88dcaa0cc710023a3b61e60774a7 [diff]
diff --git a/src/qc8-gemm/gen/1x8c4-minmax-fp32-neon-mlal-dup.c b/src/qc8-gemm/gen/1x8c4-minmax-fp32-neon-mlal-dup.c
index c1b728d..018992e 100644
--- a/src/qc8-gemm/gen/1x8c4-minmax-fp32-neon-mlal-dup.c
+++ b/src/qc8-gemm/gen/1x8c4-minmax-fp32-neon-mlal-dup.c

@@ -41,10 +41,10 @@
   int8_t* c0 = c;
 
   do {
-    int32x4_t vacc0x01 = vreinterpretq_s32_u64(vshll_n_u32(vld1_u32(w), 0)); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
-    int32x4_t vacc0x23 = vreinterpretq_s32_u64(vshll_n_u32(vld1_u32(w), 0)); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
-    int32x4_t vacc0x45 = vreinterpretq_s32_u64(vshll_n_u32(vld1_u32(w), 0)); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
-    int32x4_t vacc0x67 = vreinterpretq_s32_u64(vshll_n_u32(vld1_u32(w), 0)); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
+    int32x4_t vacc0x01 = vreinterpretq_s32_u64(vmovl_u32(vld1_u32(w))); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
+    int32x4_t vacc0x23 = vreinterpretq_s32_u64(vmovl_u32(vld1_u32(w))); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
+    int32x4_t vacc0x45 = vreinterpretq_s32_u64(vmovl_u32(vld1_u32(w))); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
+    int32x4_t vacc0x67 = vreinterpretq_s32_u64(vmovl_u32(vld1_u32(w))); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
 
     size_t k = kc;
 

diff --git a/src/qc8-gemm/gen/1x8c4-minmax-fp32-neon-mlal-ld1r.c b/src/qc8-gemm/gen/1x8c4-minmax-fp32-neon-mlal-ld1r.c
index bb0e943..4e7f7f1 100644
--- a/src/qc8-gemm/gen/1x8c4-minmax-fp32-neon-mlal-ld1r.c
+++ b/src/qc8-gemm/gen/1x8c4-minmax-fp32-neon-mlal-ld1r.c

@@ -41,10 +41,10 @@
   int8_t* c0 = c;
 
   do {
-    int32x4_t vacc0x01 = vreinterpretq_s32_u64(vshll_n_u32(vld1_u32(w), 0)); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
-    int32x4_t vacc0x23 = vreinterpretq_s32_u64(vshll_n_u32(vld1_u32(w), 0)); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
-    int32x4_t vacc0x45 = vreinterpretq_s32_u64(vshll_n_u32(vld1_u32(w), 0)); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
-    int32x4_t vacc0x67 = vreinterpretq_s32_u64(vshll_n_u32(vld1_u32(w), 0)); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
+    int32x4_t vacc0x01 = vreinterpretq_s32_u64(vmovl_u32(vld1_u32(w))); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
+    int32x4_t vacc0x23 = vreinterpretq_s32_u64(vmovl_u32(vld1_u32(w))); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
+    int32x4_t vacc0x45 = vreinterpretq_s32_u64(vmovl_u32(vld1_u32(w))); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
+    int32x4_t vacc0x67 = vreinterpretq_s32_u64(vmovl_u32(vld1_u32(w))); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
 
     size_t k = kc;
 

diff --git a/src/qc8-gemm/gen/1x8c4-minmax-fp32-neon-mlal-ld2r.c b/src/qc8-gemm/gen/1x8c4-minmax-fp32-neon-mlal-ld2r.c
index dd74420..8fafaa8 100644
--- a/src/qc8-gemm/gen/1x8c4-minmax-fp32-neon-mlal-ld2r.c
+++ b/src/qc8-gemm/gen/1x8c4-minmax-fp32-neon-mlal-ld2r.c

@@ -41,10 +41,10 @@
   int8_t* c0 = c;
 
   do {
-    int32x4_t vacc0x01 = vreinterpretq_s32_u64(vshll_n_u32(vld1_u32(w), 0)); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
-    int32x4_t vacc0x23 = vreinterpretq_s32_u64(vshll_n_u32(vld1_u32(w), 0)); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
-    int32x4_t vacc0x45 = vreinterpretq_s32_u64(vshll_n_u32(vld1_u32(w), 0)); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
-    int32x4_t vacc0x67 = vreinterpretq_s32_u64(vshll_n_u32(vld1_u32(w), 0)); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
+    int32x4_t vacc0x01 = vreinterpretq_s32_u64(vmovl_u32(vld1_u32(w))); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
+    int32x4_t vacc0x23 = vreinterpretq_s32_u64(vmovl_u32(vld1_u32(w))); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
+    int32x4_t vacc0x45 = vreinterpretq_s32_u64(vmovl_u32(vld1_u32(w))); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
+    int32x4_t vacc0x67 = vreinterpretq_s32_u64(vmovl_u32(vld1_u32(w))); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
 
     size_t k = kc;
 

diff --git a/src/qc8-gemm/gen/1x8c4-minmax-fp32-neonv8-mlal-dup.c b/src/qc8-gemm/gen/1x8c4-minmax-fp32-neonv8-mlal-dup.c
index 5117e4a..5b89a4c 100644
--- a/src/qc8-gemm/gen/1x8c4-minmax-fp32-neonv8-mlal-dup.c
+++ b/src/qc8-gemm/gen/1x8c4-minmax-fp32-neonv8-mlal-dup.c

@@ -42,10 +42,10 @@
   int8_t* c0 = c;
 
   do {
-    int32x4_t vacc0x01 = vreinterpretq_s32_u64(vshll_n_u32(vld1_u32(w), 0)); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
-    int32x4_t vacc0x23 = vreinterpretq_s32_u64(vshll_n_u32(vld1_u32(w), 0)); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
-    int32x4_t vacc0x45 = vreinterpretq_s32_u64(vshll_n_u32(vld1_u32(w), 0)); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
-    int32x4_t vacc0x67 = vreinterpretq_s32_u64(vshll_n_u32(vld1_u32(w), 0)); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
+    int32x4_t vacc0x01 = vreinterpretq_s32_u64(vmovl_u32(vld1_u32(w))); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
+    int32x4_t vacc0x23 = vreinterpretq_s32_u64(vmovl_u32(vld1_u32(w))); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
+    int32x4_t vacc0x45 = vreinterpretq_s32_u64(vmovl_u32(vld1_u32(w))); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
+    int32x4_t vacc0x67 = vreinterpretq_s32_u64(vmovl_u32(vld1_u32(w))); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
 
     size_t k = kc;
 

diff --git a/src/qc8-gemm/gen/1x8c4-minmax-fp32-neonv8-mlal-ld1r.c b/src/qc8-gemm/gen/1x8c4-minmax-fp32-neonv8-mlal-ld1r.c
index bff2e3a..20b09a4 100644
--- a/src/qc8-gemm/gen/1x8c4-minmax-fp32-neonv8-mlal-ld1r.c
+++ b/src/qc8-gemm/gen/1x8c4-minmax-fp32-neonv8-mlal-ld1r.c

@@ -42,10 +42,10 @@
   int8_t* c0 = c;
 
   do {
-    int32x4_t vacc0x01 = vreinterpretq_s32_u64(vshll_n_u32(vld1_u32(w), 0)); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
-    int32x4_t vacc0x23 = vreinterpretq_s32_u64(vshll_n_u32(vld1_u32(w), 0)); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
-    int32x4_t vacc0x45 = vreinterpretq_s32_u64(vshll_n_u32(vld1_u32(w), 0)); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
-    int32x4_t vacc0x67 = vreinterpretq_s32_u64(vshll_n_u32(vld1_u32(w), 0)); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
+    int32x4_t vacc0x01 = vreinterpretq_s32_u64(vmovl_u32(vld1_u32(w))); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
+    int32x4_t vacc0x23 = vreinterpretq_s32_u64(vmovl_u32(vld1_u32(w))); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
+    int32x4_t vacc0x45 = vreinterpretq_s32_u64(vmovl_u32(vld1_u32(w))); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
+    int32x4_t vacc0x67 = vreinterpretq_s32_u64(vmovl_u32(vld1_u32(w))); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
 
     size_t k = kc;
 

diff --git a/src/qc8-gemm/gen/1x8c4-minmax-fp32-neonv8-mlal-ld2r.c b/src/qc8-gemm/gen/1x8c4-minmax-fp32-neonv8-mlal-ld2r.c
index 2c09f1e..0a8ac86 100644
--- a/src/qc8-gemm/gen/1x8c4-minmax-fp32-neonv8-mlal-ld2r.c
+++ b/src/qc8-gemm/gen/1x8c4-minmax-fp32-neonv8-mlal-ld2r.c

@@ -42,10 +42,10 @@
   int8_t* c0 = c;
 
   do {
-    int32x4_t vacc0x01 = vreinterpretq_s32_u64(vshll_n_u32(vld1_u32(w), 0)); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
-    int32x4_t vacc0x23 = vreinterpretq_s32_u64(vshll_n_u32(vld1_u32(w), 0)); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
-    int32x4_t vacc0x45 = vreinterpretq_s32_u64(vshll_n_u32(vld1_u32(w), 0)); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
-    int32x4_t vacc0x67 = vreinterpretq_s32_u64(vshll_n_u32(vld1_u32(w), 0)); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
+    int32x4_t vacc0x01 = vreinterpretq_s32_u64(vmovl_u32(vld1_u32(w))); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
+    int32x4_t vacc0x23 = vreinterpretq_s32_u64(vmovl_u32(vld1_u32(w))); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
+    int32x4_t vacc0x45 = vreinterpretq_s32_u64(vmovl_u32(vld1_u32(w))); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
+    int32x4_t vacc0x67 = vreinterpretq_s32_u64(vmovl_u32(vld1_u32(w))); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
 
     size_t k = kc;
 

diff --git a/src/qc8-gemm/gen/1x8c4s2-minmax-fp32-neon-mlal.c b/src/qc8-gemm/gen/1x8c4s2-minmax-fp32-neon-mlal.c
index 4f357c7..9813ca7 100644
--- a/src/qc8-gemm/gen/1x8c4s2-minmax-fp32-neon-mlal.c
+++ b/src/qc8-gemm/gen/1x8c4s2-minmax-fp32-neon-mlal.c

@@ -40,10 +40,10 @@
   int8_t* c0 = c;
 
   do {
-    int32x4_t vacc0x01 = vreinterpretq_s32_u64(vshll_n_u32(vld1_u32(w), 0)); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
-    int32x4_t vacc0x23 = vreinterpretq_s32_u64(vshll_n_u32(vld1_u32(w), 0)); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
-    int32x4_t vacc0x45 = vreinterpretq_s32_u64(vshll_n_u32(vld1_u32(w), 0)); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
-    int32x4_t vacc0x67 = vreinterpretq_s32_u64(vshll_n_u32(vld1_u32(w), 0)); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
+    int32x4_t vacc0x01 = vreinterpretq_s32_u64(vmovl_u32(vld1_u32(w))); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
+    int32x4_t vacc0x23 = vreinterpretq_s32_u64(vmovl_u32(vld1_u32(w))); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
+    int32x4_t vacc0x45 = vreinterpretq_s32_u64(vmovl_u32(vld1_u32(w))); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
+    int32x4_t vacc0x67 = vreinterpretq_s32_u64(vmovl_u32(vld1_u32(w))); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
 
     size_t k = kc;
 

diff --git a/src/qc8-gemm/gen/1x8c4s2-minmax-fp32-neonv8-mlal.c b/src/qc8-gemm/gen/1x8c4s2-minmax-fp32-neonv8-mlal.c
index 4d86ed9..64dcea1 100644
--- a/src/qc8-gemm/gen/1x8c4s2-minmax-fp32-neonv8-mlal.c
+++ b/src/qc8-gemm/gen/1x8c4s2-minmax-fp32-neonv8-mlal.c

@@ -41,10 +41,10 @@
   int8_t* c0 = c;
 
   do {
-    int32x4_t vacc0x01 = vreinterpretq_s32_u64(vshll_n_u32(vld1_u32(w), 0)); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
-    int32x4_t vacc0x23 = vreinterpretq_s32_u64(vshll_n_u32(vld1_u32(w), 0)); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
-    int32x4_t vacc0x45 = vreinterpretq_s32_u64(vshll_n_u32(vld1_u32(w), 0)); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
-    int32x4_t vacc0x67 = vreinterpretq_s32_u64(vshll_n_u32(vld1_u32(w), 0)); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
+    int32x4_t vacc0x01 = vreinterpretq_s32_u64(vmovl_u32(vld1_u32(w))); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
+    int32x4_t vacc0x23 = vreinterpretq_s32_u64(vmovl_u32(vld1_u32(w))); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
+    int32x4_t vacc0x45 = vreinterpretq_s32_u64(vmovl_u32(vld1_u32(w))); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
+    int32x4_t vacc0x67 = vreinterpretq_s32_u64(vmovl_u32(vld1_u32(w))); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
 
     size_t k = kc;
 

diff --git a/src/qc8-gemm/gen/2x8c4-minmax-fp32-neon-mlal-dup.c b/src/qc8-gemm/gen/2x8c4-minmax-fp32-neon-mlal-dup.c
index a07161c..3b92e1a 100644
--- a/src/qc8-gemm/gen/2x8c4-minmax-fp32-neon-mlal-dup.c
+++ b/src/qc8-gemm/gen/2x8c4-minmax-fp32-neon-mlal-dup.c

@@ -47,10 +47,10 @@
   }
 
   do {
-    int32x4_t vacc0x01 = vreinterpretq_s32_u64(vshll_n_u32(vld1_u32(w), 0)); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
-    int32x4_t vacc0x23 = vreinterpretq_s32_u64(vshll_n_u32(vld1_u32(w), 0)); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
-    int32x4_t vacc0x45 = vreinterpretq_s32_u64(vshll_n_u32(vld1_u32(w), 0)); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
-    int32x4_t vacc0x67 = vreinterpretq_s32_u64(vshll_n_u32(vld1_u32(w), 0)); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
+    int32x4_t vacc0x01 = vreinterpretq_s32_u64(vmovl_u32(vld1_u32(w))); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
+    int32x4_t vacc0x23 = vreinterpretq_s32_u64(vmovl_u32(vld1_u32(w))); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
+    int32x4_t vacc0x45 = vreinterpretq_s32_u64(vmovl_u32(vld1_u32(w))); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
+    int32x4_t vacc0x67 = vreinterpretq_s32_u64(vmovl_u32(vld1_u32(w))); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
     int32x4_t vacc1x01 = vacc0x01;
     int32x4_t vacc1x23 = vacc0x23;
     int32x4_t vacc1x45 = vacc0x45;

diff --git a/src/qc8-gemm/gen/2x8c4-minmax-fp32-neon-mlal-ld1r.c b/src/qc8-gemm/gen/2x8c4-minmax-fp32-neon-mlal-ld1r.c
index 1247882..44d92d5 100644
--- a/src/qc8-gemm/gen/2x8c4-minmax-fp32-neon-mlal-ld1r.c
+++ b/src/qc8-gemm/gen/2x8c4-minmax-fp32-neon-mlal-ld1r.c

@@ -47,10 +47,10 @@
   }
 
   do {
-    int32x4_t vacc0x01 = vreinterpretq_s32_u64(vshll_n_u32(vld1_u32(w), 0)); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
-    int32x4_t vacc0x23 = vreinterpretq_s32_u64(vshll_n_u32(vld1_u32(w), 0)); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
-    int32x4_t vacc0x45 = vreinterpretq_s32_u64(vshll_n_u32(vld1_u32(w), 0)); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
-    int32x4_t vacc0x67 = vreinterpretq_s32_u64(vshll_n_u32(vld1_u32(w), 0)); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
+    int32x4_t vacc0x01 = vreinterpretq_s32_u64(vmovl_u32(vld1_u32(w))); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
+    int32x4_t vacc0x23 = vreinterpretq_s32_u64(vmovl_u32(vld1_u32(w))); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
+    int32x4_t vacc0x45 = vreinterpretq_s32_u64(vmovl_u32(vld1_u32(w))); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
+    int32x4_t vacc0x67 = vreinterpretq_s32_u64(vmovl_u32(vld1_u32(w))); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
     int32x4_t vacc1x01 = vacc0x01;
     int32x4_t vacc1x23 = vacc0x23;
     int32x4_t vacc1x45 = vacc0x45;

diff --git a/src/qc8-gemm/gen/2x8c4-minmax-fp32-neon-mlal-ld2r.c b/src/qc8-gemm/gen/2x8c4-minmax-fp32-neon-mlal-ld2r.c
index 3155667..11f8b2e 100644
--- a/src/qc8-gemm/gen/2x8c4-minmax-fp32-neon-mlal-ld2r.c
+++ b/src/qc8-gemm/gen/2x8c4-minmax-fp32-neon-mlal-ld2r.c

@@ -47,10 +47,10 @@
   }
 
   do {
-    int32x4_t vacc0x01 = vreinterpretq_s32_u64(vshll_n_u32(vld1_u32(w), 0)); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
-    int32x4_t vacc0x23 = vreinterpretq_s32_u64(vshll_n_u32(vld1_u32(w), 0)); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
-    int32x4_t vacc0x45 = vreinterpretq_s32_u64(vshll_n_u32(vld1_u32(w), 0)); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
-    int32x4_t vacc0x67 = vreinterpretq_s32_u64(vshll_n_u32(vld1_u32(w), 0)); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
+    int32x4_t vacc0x01 = vreinterpretq_s32_u64(vmovl_u32(vld1_u32(w))); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
+    int32x4_t vacc0x23 = vreinterpretq_s32_u64(vmovl_u32(vld1_u32(w))); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
+    int32x4_t vacc0x45 = vreinterpretq_s32_u64(vmovl_u32(vld1_u32(w))); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
+    int32x4_t vacc0x67 = vreinterpretq_s32_u64(vmovl_u32(vld1_u32(w))); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
     int32x4_t vacc1x01 = vacc0x01;
     int32x4_t vacc1x23 = vacc0x23;
     int32x4_t vacc1x45 = vacc0x45;

diff --git a/src/qc8-gemm/gen/2x8c4-minmax-fp32-neonv8-mlal-dup.c b/src/qc8-gemm/gen/2x8c4-minmax-fp32-neonv8-mlal-dup.c
index 71f73ae..b5ebf6c 100644
--- a/src/qc8-gemm/gen/2x8c4-minmax-fp32-neonv8-mlal-dup.c
+++ b/src/qc8-gemm/gen/2x8c4-minmax-fp32-neonv8-mlal-dup.c

@@ -48,10 +48,10 @@
   }
 
   do {
-    int32x4_t vacc0x01 = vreinterpretq_s32_u64(vshll_n_u32(vld1_u32(w), 0)); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
-    int32x4_t vacc0x23 = vreinterpretq_s32_u64(vshll_n_u32(vld1_u32(w), 0)); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
-    int32x4_t vacc0x45 = vreinterpretq_s32_u64(vshll_n_u32(vld1_u32(w), 0)); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
-    int32x4_t vacc0x67 = vreinterpretq_s32_u64(vshll_n_u32(vld1_u32(w), 0)); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
+    int32x4_t vacc0x01 = vreinterpretq_s32_u64(vmovl_u32(vld1_u32(w))); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
+    int32x4_t vacc0x23 = vreinterpretq_s32_u64(vmovl_u32(vld1_u32(w))); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
+    int32x4_t vacc0x45 = vreinterpretq_s32_u64(vmovl_u32(vld1_u32(w))); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
+    int32x4_t vacc0x67 = vreinterpretq_s32_u64(vmovl_u32(vld1_u32(w))); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
     int32x4_t vacc1x01 = vacc0x01;
     int32x4_t vacc1x23 = vacc0x23;
     int32x4_t vacc1x45 = vacc0x45;

diff --git a/src/qc8-gemm/gen/2x8c4-minmax-fp32-neonv8-mlal-ld1r.c b/src/qc8-gemm/gen/2x8c4-minmax-fp32-neonv8-mlal-ld1r.c
index 97ffacf..76432e4 100644
--- a/src/qc8-gemm/gen/2x8c4-minmax-fp32-neonv8-mlal-ld1r.c
+++ b/src/qc8-gemm/gen/2x8c4-minmax-fp32-neonv8-mlal-ld1r.c

@@ -48,10 +48,10 @@
   }
 
   do {
-    int32x4_t vacc0x01 = vreinterpretq_s32_u64(vshll_n_u32(vld1_u32(w), 0)); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
-    int32x4_t vacc0x23 = vreinterpretq_s32_u64(vshll_n_u32(vld1_u32(w), 0)); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
-    int32x4_t vacc0x45 = vreinterpretq_s32_u64(vshll_n_u32(vld1_u32(w), 0)); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
-    int32x4_t vacc0x67 = vreinterpretq_s32_u64(vshll_n_u32(vld1_u32(w), 0)); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
+    int32x4_t vacc0x01 = vreinterpretq_s32_u64(vmovl_u32(vld1_u32(w))); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
+    int32x4_t vacc0x23 = vreinterpretq_s32_u64(vmovl_u32(vld1_u32(w))); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
+    int32x4_t vacc0x45 = vreinterpretq_s32_u64(vmovl_u32(vld1_u32(w))); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
+    int32x4_t vacc0x67 = vreinterpretq_s32_u64(vmovl_u32(vld1_u32(w))); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
     int32x4_t vacc1x01 = vacc0x01;
     int32x4_t vacc1x23 = vacc0x23;
     int32x4_t vacc1x45 = vacc0x45;

diff --git a/src/qc8-gemm/gen/2x8c4-minmax-fp32-neonv8-mlal-ld2r.c b/src/qc8-gemm/gen/2x8c4-minmax-fp32-neonv8-mlal-ld2r.c
index 3eba002..1ae95a4 100644
--- a/src/qc8-gemm/gen/2x8c4-minmax-fp32-neonv8-mlal-ld2r.c
+++ b/src/qc8-gemm/gen/2x8c4-minmax-fp32-neonv8-mlal-ld2r.c

@@ -48,10 +48,10 @@
   }
 
   do {
-    int32x4_t vacc0x01 = vreinterpretq_s32_u64(vshll_n_u32(vld1_u32(w), 0)); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
-    int32x4_t vacc0x23 = vreinterpretq_s32_u64(vshll_n_u32(vld1_u32(w), 0)); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
-    int32x4_t vacc0x45 = vreinterpretq_s32_u64(vshll_n_u32(vld1_u32(w), 0)); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
-    int32x4_t vacc0x67 = vreinterpretq_s32_u64(vshll_n_u32(vld1_u32(w), 0)); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
+    int32x4_t vacc0x01 = vreinterpretq_s32_u64(vmovl_u32(vld1_u32(w))); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
+    int32x4_t vacc0x23 = vreinterpretq_s32_u64(vmovl_u32(vld1_u32(w))); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
+    int32x4_t vacc0x45 = vreinterpretq_s32_u64(vmovl_u32(vld1_u32(w))); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
+    int32x4_t vacc0x67 = vreinterpretq_s32_u64(vmovl_u32(vld1_u32(w))); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
     int32x4_t vacc1x01 = vacc0x01;
     int32x4_t vacc1x23 = vacc0x23;
     int32x4_t vacc1x45 = vacc0x45;

diff --git a/src/qc8-gemm/gen/2x8c4s2-minmax-fp32-neon-mlal.c b/src/qc8-gemm/gen/2x8c4s2-minmax-fp32-neon-mlal.c
index 84e415c..c3bc81f 100644
--- a/src/qc8-gemm/gen/2x8c4s2-minmax-fp32-neon-mlal.c
+++ b/src/qc8-gemm/gen/2x8c4s2-minmax-fp32-neon-mlal.c

@@ -46,10 +46,10 @@
   }
 
   do {
-    int32x4_t vacc0x01 = vreinterpretq_s32_u64(vshll_n_u32(vld1_u32(w), 0)); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
-    int32x4_t vacc0x23 = vreinterpretq_s32_u64(vshll_n_u32(vld1_u32(w), 0)); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
-    int32x4_t vacc0x45 = vreinterpretq_s32_u64(vshll_n_u32(vld1_u32(w), 0)); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
-    int32x4_t vacc0x67 = vreinterpretq_s32_u64(vshll_n_u32(vld1_u32(w), 0)); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
+    int32x4_t vacc0x01 = vreinterpretq_s32_u64(vmovl_u32(vld1_u32(w))); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
+    int32x4_t vacc0x23 = vreinterpretq_s32_u64(vmovl_u32(vld1_u32(w))); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
+    int32x4_t vacc0x45 = vreinterpretq_s32_u64(vmovl_u32(vld1_u32(w))); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
+    int32x4_t vacc0x67 = vreinterpretq_s32_u64(vmovl_u32(vld1_u32(w))); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
     int32x4_t vacc1x01 = vacc0x01;
     int32x4_t vacc1x23 = vacc0x23;
     int32x4_t vacc1x45 = vacc0x45;

diff --git a/src/qc8-gemm/gen/2x8c4s2-minmax-fp32-neonv8-mlal.c b/src/qc8-gemm/gen/2x8c4s2-minmax-fp32-neonv8-mlal.c
index d008bb1..1172f51 100644
--- a/src/qc8-gemm/gen/2x8c4s2-minmax-fp32-neonv8-mlal.c
+++ b/src/qc8-gemm/gen/2x8c4s2-minmax-fp32-neonv8-mlal.c

@@ -47,10 +47,10 @@
   }
 
   do {
-    int32x4_t vacc0x01 = vreinterpretq_s32_u64(vshll_n_u32(vld1_u32(w), 0)); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
-    int32x4_t vacc0x23 = vreinterpretq_s32_u64(vshll_n_u32(vld1_u32(w), 0)); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
-    int32x4_t vacc0x45 = vreinterpretq_s32_u64(vshll_n_u32(vld1_u32(w), 0)); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
-    int32x4_t vacc0x67 = vreinterpretq_s32_u64(vshll_n_u32(vld1_u32(w), 0)); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
+    int32x4_t vacc0x01 = vreinterpretq_s32_u64(vmovl_u32(vld1_u32(w))); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
+    int32x4_t vacc0x23 = vreinterpretq_s32_u64(vmovl_u32(vld1_u32(w))); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
+    int32x4_t vacc0x45 = vreinterpretq_s32_u64(vmovl_u32(vld1_u32(w))); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
+    int32x4_t vacc0x67 = vreinterpretq_s32_u64(vmovl_u32(vld1_u32(w))); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
     int32x4_t vacc1x01 = vacc0x01;
     int32x4_t vacc1x23 = vacc0x23;
     int32x4_t vacc1x45 = vacc0x45;

diff --git a/src/qc8-igemm/gen/1x8c4-minmax-fp32-neon-mlal-dup.c b/src/qc8-igemm/gen/1x8c4-minmax-fp32-neon-mlal-dup.c
index 3ff708b..51ede82 100644
--- a/src/qc8-igemm/gen/1x8c4-minmax-fp32-neon-mlal-dup.c
+++ b/src/qc8-igemm/gen/1x8c4-minmax-fp32-neon-mlal-dup.c

@@ -44,10 +44,10 @@
   int8_t* c0 = c;
 
   do {
-    int32x4_t vacc0x01 = vreinterpretq_s32_u64(vshll_n_u32(vld1_u32(w), 0)); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
-    int32x4_t vacc0x23 = vreinterpretq_s32_u64(vshll_n_u32(vld1_u32(w), 0)); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
-    int32x4_t vacc0x45 = vreinterpretq_s32_u64(vshll_n_u32(vld1_u32(w), 0)); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
-    int32x4_t vacc0x67 = vreinterpretq_s32_u64(vshll_n_u32(vld1_u32(w), 0)); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
+    int32x4_t vacc0x01 = vreinterpretq_s32_u64(vmovl_u32(vld1_u32(w))); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
+    int32x4_t vacc0x23 = vreinterpretq_s32_u64(vmovl_u32(vld1_u32(w))); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
+    int32x4_t vacc0x45 = vreinterpretq_s32_u64(vmovl_u32(vld1_u32(w))); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
+    int32x4_t vacc0x67 = vreinterpretq_s32_u64(vmovl_u32(vld1_u32(w))); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
 
     size_t p = ks;
     do {

diff --git a/src/qc8-igemm/gen/1x8c4-minmax-fp32-neon-mlal-ld1r.c b/src/qc8-igemm/gen/1x8c4-minmax-fp32-neon-mlal-ld1r.c
index 8990639..5022a96 100644
--- a/src/qc8-igemm/gen/1x8c4-minmax-fp32-neon-mlal-ld1r.c
+++ b/src/qc8-igemm/gen/1x8c4-minmax-fp32-neon-mlal-ld1r.c

@@ -44,10 +44,10 @@
   int8_t* c0 = c;
 
   do {
-    int32x4_t vacc0x01 = vreinterpretq_s32_u64(vshll_n_u32(vld1_u32(w), 0)); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
-    int32x4_t vacc0x23 = vreinterpretq_s32_u64(vshll_n_u32(vld1_u32(w), 0)); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
-    int32x4_t vacc0x45 = vreinterpretq_s32_u64(vshll_n_u32(vld1_u32(w), 0)); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
-    int32x4_t vacc0x67 = vreinterpretq_s32_u64(vshll_n_u32(vld1_u32(w), 0)); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
+    int32x4_t vacc0x01 = vreinterpretq_s32_u64(vmovl_u32(vld1_u32(w))); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
+    int32x4_t vacc0x23 = vreinterpretq_s32_u64(vmovl_u32(vld1_u32(w))); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
+    int32x4_t vacc0x45 = vreinterpretq_s32_u64(vmovl_u32(vld1_u32(w))); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
+    int32x4_t vacc0x67 = vreinterpretq_s32_u64(vmovl_u32(vld1_u32(w))); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
 
     size_t p = ks;
     do {

diff --git a/src/qc8-igemm/gen/1x8c4-minmax-fp32-neon-mlal-ld2r.c b/src/qc8-igemm/gen/1x8c4-minmax-fp32-neon-mlal-ld2r.c
index eb3b570..7dc398e 100644
--- a/src/qc8-igemm/gen/1x8c4-minmax-fp32-neon-mlal-ld2r.c
+++ b/src/qc8-igemm/gen/1x8c4-minmax-fp32-neon-mlal-ld2r.c

@@ -44,10 +44,10 @@
   int8_t* c0 = c;
 
   do {
-    int32x4_t vacc0x01 = vreinterpretq_s32_u64(vshll_n_u32(vld1_u32(w), 0)); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
-    int32x4_t vacc0x23 = vreinterpretq_s32_u64(vshll_n_u32(vld1_u32(w), 0)); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
-    int32x4_t vacc0x45 = vreinterpretq_s32_u64(vshll_n_u32(vld1_u32(w), 0)); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
-    int32x4_t vacc0x67 = vreinterpretq_s32_u64(vshll_n_u32(vld1_u32(w), 0)); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
+    int32x4_t vacc0x01 = vreinterpretq_s32_u64(vmovl_u32(vld1_u32(w))); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
+    int32x4_t vacc0x23 = vreinterpretq_s32_u64(vmovl_u32(vld1_u32(w))); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
+    int32x4_t vacc0x45 = vreinterpretq_s32_u64(vmovl_u32(vld1_u32(w))); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
+    int32x4_t vacc0x67 = vreinterpretq_s32_u64(vmovl_u32(vld1_u32(w))); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
 
     size_t p = ks;
     do {

diff --git a/src/qc8-igemm/gen/1x8c4-minmax-fp32-neonv8-mlal-dup.c b/src/qc8-igemm/gen/1x8c4-minmax-fp32-neonv8-mlal-dup.c
index 876903f..c4677c9 100644
--- a/src/qc8-igemm/gen/1x8c4-minmax-fp32-neonv8-mlal-dup.c
+++ b/src/qc8-igemm/gen/1x8c4-minmax-fp32-neonv8-mlal-dup.c

@@ -45,10 +45,10 @@
   int8_t* c0 = c;
 
   do {
-    int32x4_t vacc0x01 = vreinterpretq_s32_u64(vshll_n_u32(vld1_u32(w), 0)); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
-    int32x4_t vacc0x23 = vreinterpretq_s32_u64(vshll_n_u32(vld1_u32(w), 0)); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
-    int32x4_t vacc0x45 = vreinterpretq_s32_u64(vshll_n_u32(vld1_u32(w), 0)); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
-    int32x4_t vacc0x67 = vreinterpretq_s32_u64(vshll_n_u32(vld1_u32(w), 0)); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
+    int32x4_t vacc0x01 = vreinterpretq_s32_u64(vmovl_u32(vld1_u32(w))); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
+    int32x4_t vacc0x23 = vreinterpretq_s32_u64(vmovl_u32(vld1_u32(w))); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
+    int32x4_t vacc0x45 = vreinterpretq_s32_u64(vmovl_u32(vld1_u32(w))); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
+    int32x4_t vacc0x67 = vreinterpretq_s32_u64(vmovl_u32(vld1_u32(w))); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
 
     size_t p = ks;
     do {

diff --git a/src/qc8-igemm/gen/1x8c4-minmax-fp32-neonv8-mlal-ld1r.c b/src/qc8-igemm/gen/1x8c4-minmax-fp32-neonv8-mlal-ld1r.c
index bdb5e5c..ce5b676 100644
--- a/src/qc8-igemm/gen/1x8c4-minmax-fp32-neonv8-mlal-ld1r.c
+++ b/src/qc8-igemm/gen/1x8c4-minmax-fp32-neonv8-mlal-ld1r.c

@@ -45,10 +45,10 @@
   int8_t* c0 = c;
 
   do {
-    int32x4_t vacc0x01 = vreinterpretq_s32_u64(vshll_n_u32(vld1_u32(w), 0)); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
-    int32x4_t vacc0x23 = vreinterpretq_s32_u64(vshll_n_u32(vld1_u32(w), 0)); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
-    int32x4_t vacc0x45 = vreinterpretq_s32_u64(vshll_n_u32(vld1_u32(w), 0)); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
-    int32x4_t vacc0x67 = vreinterpretq_s32_u64(vshll_n_u32(vld1_u32(w), 0)); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
+    int32x4_t vacc0x01 = vreinterpretq_s32_u64(vmovl_u32(vld1_u32(w))); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
+    int32x4_t vacc0x23 = vreinterpretq_s32_u64(vmovl_u32(vld1_u32(w))); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
+    int32x4_t vacc0x45 = vreinterpretq_s32_u64(vmovl_u32(vld1_u32(w))); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
+    int32x4_t vacc0x67 = vreinterpretq_s32_u64(vmovl_u32(vld1_u32(w))); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
 
     size_t p = ks;
     do {

diff --git a/src/qc8-igemm/gen/1x8c4-minmax-fp32-neonv8-mlal-ld2r.c b/src/qc8-igemm/gen/1x8c4-minmax-fp32-neonv8-mlal-ld2r.c
index d3a3af8..6aad3d3 100644
--- a/src/qc8-igemm/gen/1x8c4-minmax-fp32-neonv8-mlal-ld2r.c
+++ b/src/qc8-igemm/gen/1x8c4-minmax-fp32-neonv8-mlal-ld2r.c

@@ -45,10 +45,10 @@
   int8_t* c0 = c;
 
   do {
-    int32x4_t vacc0x01 = vreinterpretq_s32_u64(vshll_n_u32(vld1_u32(w), 0)); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
-    int32x4_t vacc0x23 = vreinterpretq_s32_u64(vshll_n_u32(vld1_u32(w), 0)); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
-    int32x4_t vacc0x45 = vreinterpretq_s32_u64(vshll_n_u32(vld1_u32(w), 0)); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
-    int32x4_t vacc0x67 = vreinterpretq_s32_u64(vshll_n_u32(vld1_u32(w), 0)); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
+    int32x4_t vacc0x01 = vreinterpretq_s32_u64(vmovl_u32(vld1_u32(w))); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
+    int32x4_t vacc0x23 = vreinterpretq_s32_u64(vmovl_u32(vld1_u32(w))); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
+    int32x4_t vacc0x45 = vreinterpretq_s32_u64(vmovl_u32(vld1_u32(w))); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
+    int32x4_t vacc0x67 = vreinterpretq_s32_u64(vmovl_u32(vld1_u32(w))); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
 
     size_t p = ks;
     do {

diff --git a/src/qc8-igemm/gen/1x8c4s2-minmax-fp32-neon-mlal.c b/src/qc8-igemm/gen/1x8c4s2-minmax-fp32-neon-mlal.c
index b1d1432..c1675bb 100644
--- a/src/qc8-igemm/gen/1x8c4s2-minmax-fp32-neon-mlal.c
+++ b/src/qc8-igemm/gen/1x8c4s2-minmax-fp32-neon-mlal.c

@@ -43,10 +43,10 @@
   int8_t* c0 = c;
 
   do {
-    int32x4_t vacc0x01 = vreinterpretq_s32_u64(vshll_n_u32(vld1_u32(w), 0)); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
-    int32x4_t vacc0x23 = vreinterpretq_s32_u64(vshll_n_u32(vld1_u32(w), 0)); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
-    int32x4_t vacc0x45 = vreinterpretq_s32_u64(vshll_n_u32(vld1_u32(w), 0)); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
-    int32x4_t vacc0x67 = vreinterpretq_s32_u64(vshll_n_u32(vld1_u32(w), 0)); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
+    int32x4_t vacc0x01 = vreinterpretq_s32_u64(vmovl_u32(vld1_u32(w))); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
+    int32x4_t vacc0x23 = vreinterpretq_s32_u64(vmovl_u32(vld1_u32(w))); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
+    int32x4_t vacc0x45 = vreinterpretq_s32_u64(vmovl_u32(vld1_u32(w))); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
+    int32x4_t vacc0x67 = vreinterpretq_s32_u64(vmovl_u32(vld1_u32(w))); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
 
     size_t p = ks;
     do {

diff --git a/src/qc8-igemm/gen/1x8c4s2-minmax-fp32-neonv8-mlal.c b/src/qc8-igemm/gen/1x8c4s2-minmax-fp32-neonv8-mlal.c
index 716902c..183516d 100644
--- a/src/qc8-igemm/gen/1x8c4s2-minmax-fp32-neonv8-mlal.c
+++ b/src/qc8-igemm/gen/1x8c4s2-minmax-fp32-neonv8-mlal.c

@@ -44,10 +44,10 @@
   int8_t* c0 = c;
 
   do {
-    int32x4_t vacc0x01 = vreinterpretq_s32_u64(vshll_n_u32(vld1_u32(w), 0)); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
-    int32x4_t vacc0x23 = vreinterpretq_s32_u64(vshll_n_u32(vld1_u32(w), 0)); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
-    int32x4_t vacc0x45 = vreinterpretq_s32_u64(vshll_n_u32(vld1_u32(w), 0)); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
-    int32x4_t vacc0x67 = vreinterpretq_s32_u64(vshll_n_u32(vld1_u32(w), 0)); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
+    int32x4_t vacc0x01 = vreinterpretq_s32_u64(vmovl_u32(vld1_u32(w))); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
+    int32x4_t vacc0x23 = vreinterpretq_s32_u64(vmovl_u32(vld1_u32(w))); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
+    int32x4_t vacc0x45 = vreinterpretq_s32_u64(vmovl_u32(vld1_u32(w))); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
+    int32x4_t vacc0x67 = vreinterpretq_s32_u64(vmovl_u32(vld1_u32(w))); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
 
     size_t p = ks;
     do {

diff --git a/src/qc8-igemm/gen/2x8c4-minmax-fp32-neon-mlal-dup.c b/src/qc8-igemm/gen/2x8c4-minmax-fp32-neon-mlal-dup.c
index 4c6550f..b44172a 100644
--- a/src/qc8-igemm/gen/2x8c4-minmax-fp32-neon-mlal-dup.c
+++ b/src/qc8-igemm/gen/2x8c4-minmax-fp32-neon-mlal-dup.c

@@ -48,10 +48,10 @@
   }
 
   do {
-    int32x4_t vacc0x01 = vreinterpretq_s32_u64(vshll_n_u32(vld1_u32(w), 0)); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
-    int32x4_t vacc0x23 = vreinterpretq_s32_u64(vshll_n_u32(vld1_u32(w), 0)); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
-    int32x4_t vacc0x45 = vreinterpretq_s32_u64(vshll_n_u32(vld1_u32(w), 0)); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
-    int32x4_t vacc0x67 = vreinterpretq_s32_u64(vshll_n_u32(vld1_u32(w), 0)); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
+    int32x4_t vacc0x01 = vreinterpretq_s32_u64(vmovl_u32(vld1_u32(w))); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
+    int32x4_t vacc0x23 = vreinterpretq_s32_u64(vmovl_u32(vld1_u32(w))); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
+    int32x4_t vacc0x45 = vreinterpretq_s32_u64(vmovl_u32(vld1_u32(w))); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
+    int32x4_t vacc0x67 = vreinterpretq_s32_u64(vmovl_u32(vld1_u32(w))); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
     int32x4_t vacc1x01 = vacc0x01;
     int32x4_t vacc1x23 = vacc0x23;
     int32x4_t vacc1x45 = vacc0x45;

diff --git a/src/qc8-igemm/gen/2x8c4-minmax-fp32-neon-mlal-ld1r.c b/src/qc8-igemm/gen/2x8c4-minmax-fp32-neon-mlal-ld1r.c
index 1f3c60d..bdaae0c 100644
--- a/src/qc8-igemm/gen/2x8c4-minmax-fp32-neon-mlal-ld1r.c
+++ b/src/qc8-igemm/gen/2x8c4-minmax-fp32-neon-mlal-ld1r.c

@@ -48,10 +48,10 @@
   }
 
   do {
-    int32x4_t vacc0x01 = vreinterpretq_s32_u64(vshll_n_u32(vld1_u32(w), 0)); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
-    int32x4_t vacc0x23 = vreinterpretq_s32_u64(vshll_n_u32(vld1_u32(w), 0)); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
-    int32x4_t vacc0x45 = vreinterpretq_s32_u64(vshll_n_u32(vld1_u32(w), 0)); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
-    int32x4_t vacc0x67 = vreinterpretq_s32_u64(vshll_n_u32(vld1_u32(w), 0)); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
+    int32x4_t vacc0x01 = vreinterpretq_s32_u64(vmovl_u32(vld1_u32(w))); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
+    int32x4_t vacc0x23 = vreinterpretq_s32_u64(vmovl_u32(vld1_u32(w))); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
+    int32x4_t vacc0x45 = vreinterpretq_s32_u64(vmovl_u32(vld1_u32(w))); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
+    int32x4_t vacc0x67 = vreinterpretq_s32_u64(vmovl_u32(vld1_u32(w))); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
     int32x4_t vacc1x01 = vacc0x01;
     int32x4_t vacc1x23 = vacc0x23;
     int32x4_t vacc1x45 = vacc0x45;

diff --git a/src/qc8-igemm/gen/2x8c4-minmax-fp32-neon-mlal-ld2r.c b/src/qc8-igemm/gen/2x8c4-minmax-fp32-neon-mlal-ld2r.c
index f6586c8..1f22ed9 100644
--- a/src/qc8-igemm/gen/2x8c4-minmax-fp32-neon-mlal-ld2r.c
+++ b/src/qc8-igemm/gen/2x8c4-minmax-fp32-neon-mlal-ld2r.c

@@ -48,10 +48,10 @@
   }
 
   do {
-    int32x4_t vacc0x01 = vreinterpretq_s32_u64(vshll_n_u32(vld1_u32(w), 0)); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
-    int32x4_t vacc0x23 = vreinterpretq_s32_u64(vshll_n_u32(vld1_u32(w), 0)); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
-    int32x4_t vacc0x45 = vreinterpretq_s32_u64(vshll_n_u32(vld1_u32(w), 0)); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
-    int32x4_t vacc0x67 = vreinterpretq_s32_u64(vshll_n_u32(vld1_u32(w), 0)); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
+    int32x4_t vacc0x01 = vreinterpretq_s32_u64(vmovl_u32(vld1_u32(w))); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
+    int32x4_t vacc0x23 = vreinterpretq_s32_u64(vmovl_u32(vld1_u32(w))); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
+    int32x4_t vacc0x45 = vreinterpretq_s32_u64(vmovl_u32(vld1_u32(w))); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
+    int32x4_t vacc0x67 = vreinterpretq_s32_u64(vmovl_u32(vld1_u32(w))); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
     int32x4_t vacc1x01 = vacc0x01;
     int32x4_t vacc1x23 = vacc0x23;
     int32x4_t vacc1x45 = vacc0x45;

diff --git a/src/qc8-igemm/gen/2x8c4-minmax-fp32-neonv8-mlal-dup.c b/src/qc8-igemm/gen/2x8c4-minmax-fp32-neonv8-mlal-dup.c
index 4cc863a..63c6133 100644
--- a/src/qc8-igemm/gen/2x8c4-minmax-fp32-neonv8-mlal-dup.c
+++ b/src/qc8-igemm/gen/2x8c4-minmax-fp32-neonv8-mlal-dup.c

@@ -49,10 +49,10 @@
   }
 
   do {
-    int32x4_t vacc0x01 = vreinterpretq_s32_u64(vshll_n_u32(vld1_u32(w), 0)); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
-    int32x4_t vacc0x23 = vreinterpretq_s32_u64(vshll_n_u32(vld1_u32(w), 0)); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
-    int32x4_t vacc0x45 = vreinterpretq_s32_u64(vshll_n_u32(vld1_u32(w), 0)); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
-    int32x4_t vacc0x67 = vreinterpretq_s32_u64(vshll_n_u32(vld1_u32(w), 0)); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
+    int32x4_t vacc0x01 = vreinterpretq_s32_u64(vmovl_u32(vld1_u32(w))); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
+    int32x4_t vacc0x23 = vreinterpretq_s32_u64(vmovl_u32(vld1_u32(w))); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
+    int32x4_t vacc0x45 = vreinterpretq_s32_u64(vmovl_u32(vld1_u32(w))); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
+    int32x4_t vacc0x67 = vreinterpretq_s32_u64(vmovl_u32(vld1_u32(w))); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
     int32x4_t vacc1x01 = vacc0x01;
     int32x4_t vacc1x23 = vacc0x23;
     int32x4_t vacc1x45 = vacc0x45;

diff --git a/src/qc8-igemm/gen/2x8c4-minmax-fp32-neonv8-mlal-ld1r.c b/src/qc8-igemm/gen/2x8c4-minmax-fp32-neonv8-mlal-ld1r.c
index cc40522..37e73c0 100644
--- a/src/qc8-igemm/gen/2x8c4-minmax-fp32-neonv8-mlal-ld1r.c
+++ b/src/qc8-igemm/gen/2x8c4-minmax-fp32-neonv8-mlal-ld1r.c

@@ -49,10 +49,10 @@
   }
 
   do {
-    int32x4_t vacc0x01 = vreinterpretq_s32_u64(vshll_n_u32(vld1_u32(w), 0)); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
-    int32x4_t vacc0x23 = vreinterpretq_s32_u64(vshll_n_u32(vld1_u32(w), 0)); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
-    int32x4_t vacc0x45 = vreinterpretq_s32_u64(vshll_n_u32(vld1_u32(w), 0)); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
-    int32x4_t vacc0x67 = vreinterpretq_s32_u64(vshll_n_u32(vld1_u32(w), 0)); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
+    int32x4_t vacc0x01 = vreinterpretq_s32_u64(vmovl_u32(vld1_u32(w))); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
+    int32x4_t vacc0x23 = vreinterpretq_s32_u64(vmovl_u32(vld1_u32(w))); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
+    int32x4_t vacc0x45 = vreinterpretq_s32_u64(vmovl_u32(vld1_u32(w))); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
+    int32x4_t vacc0x67 = vreinterpretq_s32_u64(vmovl_u32(vld1_u32(w))); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
     int32x4_t vacc1x01 = vacc0x01;
     int32x4_t vacc1x23 = vacc0x23;
     int32x4_t vacc1x45 = vacc0x45;

diff --git a/src/qc8-igemm/gen/2x8c4-minmax-fp32-neonv8-mlal-ld2r.c b/src/qc8-igemm/gen/2x8c4-minmax-fp32-neonv8-mlal-ld2r.c
index 87ff201..668e848 100644
--- a/src/qc8-igemm/gen/2x8c4-minmax-fp32-neonv8-mlal-ld2r.c
+++ b/src/qc8-igemm/gen/2x8c4-minmax-fp32-neonv8-mlal-ld2r.c

@@ -49,10 +49,10 @@
   }
 
   do {
-    int32x4_t vacc0x01 = vreinterpretq_s32_u64(vshll_n_u32(vld1_u32(w), 0)); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
-    int32x4_t vacc0x23 = vreinterpretq_s32_u64(vshll_n_u32(vld1_u32(w), 0)); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
-    int32x4_t vacc0x45 = vreinterpretq_s32_u64(vshll_n_u32(vld1_u32(w), 0)); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
-    int32x4_t vacc0x67 = vreinterpretq_s32_u64(vshll_n_u32(vld1_u32(w), 0)); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
+    int32x4_t vacc0x01 = vreinterpretq_s32_u64(vmovl_u32(vld1_u32(w))); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
+    int32x4_t vacc0x23 = vreinterpretq_s32_u64(vmovl_u32(vld1_u32(w))); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
+    int32x4_t vacc0x45 = vreinterpretq_s32_u64(vmovl_u32(vld1_u32(w))); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
+    int32x4_t vacc0x67 = vreinterpretq_s32_u64(vmovl_u32(vld1_u32(w))); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
     int32x4_t vacc1x01 = vacc0x01;
     int32x4_t vacc1x23 = vacc0x23;
     int32x4_t vacc1x45 = vacc0x45;

diff --git a/src/qc8-igemm/gen/2x8c4s2-minmax-fp32-neon-mlal.c b/src/qc8-igemm/gen/2x8c4s2-minmax-fp32-neon-mlal.c
index da57e88..a871fba 100644
--- a/src/qc8-igemm/gen/2x8c4s2-minmax-fp32-neon-mlal.c
+++ b/src/qc8-igemm/gen/2x8c4s2-minmax-fp32-neon-mlal.c

@@ -47,10 +47,10 @@
   }
 
   do {
-    int32x4_t vacc0x01 = vreinterpretq_s32_u64(vshll_n_u32(vld1_u32(w), 0)); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
-    int32x4_t vacc0x23 = vreinterpretq_s32_u64(vshll_n_u32(vld1_u32(w), 0)); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
-    int32x4_t vacc0x45 = vreinterpretq_s32_u64(vshll_n_u32(vld1_u32(w), 0)); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
-    int32x4_t vacc0x67 = vreinterpretq_s32_u64(vshll_n_u32(vld1_u32(w), 0)); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
+    int32x4_t vacc0x01 = vreinterpretq_s32_u64(vmovl_u32(vld1_u32(w))); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
+    int32x4_t vacc0x23 = vreinterpretq_s32_u64(vmovl_u32(vld1_u32(w))); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
+    int32x4_t vacc0x45 = vreinterpretq_s32_u64(vmovl_u32(vld1_u32(w))); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
+    int32x4_t vacc0x67 = vreinterpretq_s32_u64(vmovl_u32(vld1_u32(w))); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
     int32x4_t vacc1x01 = vacc0x01;
     int32x4_t vacc1x23 = vacc0x23;
     int32x4_t vacc1x45 = vacc0x45;

diff --git a/src/qc8-igemm/gen/2x8c4s2-minmax-fp32-neonv8-mlal.c b/src/qc8-igemm/gen/2x8c4s2-minmax-fp32-neonv8-mlal.c
index 555daeb..abd5ed3 100644
--- a/src/qc8-igemm/gen/2x8c4s2-minmax-fp32-neonv8-mlal.c
+++ b/src/qc8-igemm/gen/2x8c4s2-minmax-fp32-neonv8-mlal.c

@@ -48,10 +48,10 @@
   }
 
   do {
-    int32x4_t vacc0x01 = vreinterpretq_s32_u64(vshll_n_u32(vld1_u32(w), 0)); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
-    int32x4_t vacc0x23 = vreinterpretq_s32_u64(vshll_n_u32(vld1_u32(w), 0)); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
-    int32x4_t vacc0x45 = vreinterpretq_s32_u64(vshll_n_u32(vld1_u32(w), 0)); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
-    int32x4_t vacc0x67 = vreinterpretq_s32_u64(vshll_n_u32(vld1_u32(w), 0)); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
+    int32x4_t vacc0x01 = vreinterpretq_s32_u64(vmovl_u32(vld1_u32(w))); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
+    int32x4_t vacc0x23 = vreinterpretq_s32_u64(vmovl_u32(vld1_u32(w))); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
+    int32x4_t vacc0x45 = vreinterpretq_s32_u64(vmovl_u32(vld1_u32(w))); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
+    int32x4_t vacc0x67 = vreinterpretq_s32_u64(vmovl_u32(vld1_u32(w))); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
     int32x4_t vacc1x01 = vacc0x01;
     int32x4_t vacc1x23 = vacc0x23;
     int32x4_t vacc1x45 = vacc0x45;

diff --git a/src/qs8-gemm/c4-neon-mull-dup.c.in b/src/qs8-gemm/c4-neon-mull-dup.c.in
index 0a0e77c..9a1d12d 100644
--- a/src/qs8-gemm/c4-neon-mull-dup.c.in
+++ b/src/qs8-gemm/c4-neon-mull-dup.c.in

@@ -68,7 +68,7 @@
 
   do {
     $for N in range(0, NR, 2):
-      int32x4_t vacc0x${ABC[N:N+2]} = vreinterpretq_s32_u64(vshll_n_u32(vld1_u32(w), 0)); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
+      int32x4_t vacc0x${ABC[N:N+2]} = vreinterpretq_s32_u64(vmovl_u32(vld1_u32(w))); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
     $for M in range(1, MR):
       $for N in range(0, NR, 2):
         int32x4_t vacc${M}x${ABC[N:N+2]} = vacc0x${ABC[N:N+2]};

diff --git a/src/qs8-gemm/c4-neon-mull-shuffle.c.in b/src/qs8-gemm/c4-neon-mull-shuffle.c.in
index 2fc2bc6..ad8c220 100644
--- a/src/qs8-gemm/c4-neon-mull-shuffle.c.in
+++ b/src/qs8-gemm/c4-neon-mull-shuffle.c.in

@@ -66,7 +66,7 @@
 
   do {
     $for N in range(0, NR, 2):
-      int32x4_t vacc0x${ABC[N:N+2]} = vreinterpretq_s32_u64(vshll_n_u32(vld1_u32(w), 0)); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
+      int32x4_t vacc0x${ABC[N:N+2]} = vreinterpretq_s32_u64(vmovl_u32(vld1_u32(w))); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
     $for M in range(1, MR):
       $for N in range(0, NR, 2):
         int32x4_t vacc${M}x${ABC[N:N+2]} = vacc0x${ABC[N:N+2]};

diff --git a/src/qs8-gemm/gen/1x16c4-minmax-rndnu-neon-mlal-dup.c b/src/qs8-gemm/gen/1x16c4-minmax-rndnu-neon-mlal-dup.c
index 69fc86a..e74e38f 100644
--- a/src/qs8-gemm/gen/1x16c4-minmax-rndnu-neon-mlal-dup.c
+++ b/src/qs8-gemm/gen/1x16c4-minmax-rndnu-neon-mlal-dup.c

@@ -41,14 +41,14 @@
   int8_t* c0 = c;
 
   do {
-    int32x4_t vacc0x01 = vreinterpretq_s32_u64(vshll_n_u32(vld1_u32(w), 0)); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
-    int32x4_t vacc0x23 = vreinterpretq_s32_u64(vshll_n_u32(vld1_u32(w), 0)); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
-    int32x4_t vacc0x45 = vreinterpretq_s32_u64(vshll_n_u32(vld1_u32(w), 0)); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
-    int32x4_t vacc0x67 = vreinterpretq_s32_u64(vshll_n_u32(vld1_u32(w), 0)); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
-    int32x4_t vacc0x89 = vreinterpretq_s32_u64(vshll_n_u32(vld1_u32(w), 0)); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
-    int32x4_t vacc0xAB = vreinterpretq_s32_u64(vshll_n_u32(vld1_u32(w), 0)); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
-    int32x4_t vacc0xCD = vreinterpretq_s32_u64(vshll_n_u32(vld1_u32(w), 0)); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
-    int32x4_t vacc0xEF = vreinterpretq_s32_u64(vshll_n_u32(vld1_u32(w), 0)); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
+    int32x4_t vacc0x01 = vreinterpretq_s32_u64(vmovl_u32(vld1_u32(w))); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
+    int32x4_t vacc0x23 = vreinterpretq_s32_u64(vmovl_u32(vld1_u32(w))); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
+    int32x4_t vacc0x45 = vreinterpretq_s32_u64(vmovl_u32(vld1_u32(w))); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
+    int32x4_t vacc0x67 = vreinterpretq_s32_u64(vmovl_u32(vld1_u32(w))); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
+    int32x4_t vacc0x89 = vreinterpretq_s32_u64(vmovl_u32(vld1_u32(w))); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
+    int32x4_t vacc0xAB = vreinterpretq_s32_u64(vmovl_u32(vld1_u32(w))); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
+    int32x4_t vacc0xCD = vreinterpretq_s32_u64(vmovl_u32(vld1_u32(w))); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
+    int32x4_t vacc0xEF = vreinterpretq_s32_u64(vmovl_u32(vld1_u32(w))); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
 
     size_t k = kc;
 

diff --git a/src/qs8-gemm/gen/1x16c4-minmax-rndnu-neon-mlal-ld1r.c b/src/qs8-gemm/gen/1x16c4-minmax-rndnu-neon-mlal-ld1r.c
index 92fcc62..3c04b4f 100644
--- a/src/qs8-gemm/gen/1x16c4-minmax-rndnu-neon-mlal-ld1r.c
+++ b/src/qs8-gemm/gen/1x16c4-minmax-rndnu-neon-mlal-ld1r.c

@@ -41,14 +41,14 @@
   int8_t* c0 = c;
 
   do {
-    int32x4_t vacc0x01 = vreinterpretq_s32_u64(vshll_n_u32(vld1_u32(w), 0)); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
-    int32x4_t vacc0x23 = vreinterpretq_s32_u64(vshll_n_u32(vld1_u32(w), 0)); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
-    int32x4_t vacc0x45 = vreinterpretq_s32_u64(vshll_n_u32(vld1_u32(w), 0)); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
-    int32x4_t vacc0x67 = vreinterpretq_s32_u64(vshll_n_u32(vld1_u32(w), 0)); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
-    int32x4_t vacc0x89 = vreinterpretq_s32_u64(vshll_n_u32(vld1_u32(w), 0)); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
-    int32x4_t vacc0xAB = vreinterpretq_s32_u64(vshll_n_u32(vld1_u32(w), 0)); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
-    int32x4_t vacc0xCD = vreinterpretq_s32_u64(vshll_n_u32(vld1_u32(w), 0)); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
-    int32x4_t vacc0xEF = vreinterpretq_s32_u64(vshll_n_u32(vld1_u32(w), 0)); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
+    int32x4_t vacc0x01 = vreinterpretq_s32_u64(vmovl_u32(vld1_u32(w))); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
+    int32x4_t vacc0x23 = vreinterpretq_s32_u64(vmovl_u32(vld1_u32(w))); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
+    int32x4_t vacc0x45 = vreinterpretq_s32_u64(vmovl_u32(vld1_u32(w))); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
+    int32x4_t vacc0x67 = vreinterpretq_s32_u64(vmovl_u32(vld1_u32(w))); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
+    int32x4_t vacc0x89 = vreinterpretq_s32_u64(vmovl_u32(vld1_u32(w))); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
+    int32x4_t vacc0xAB = vreinterpretq_s32_u64(vmovl_u32(vld1_u32(w))); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
+    int32x4_t vacc0xCD = vreinterpretq_s32_u64(vmovl_u32(vld1_u32(w))); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
+    int32x4_t vacc0xEF = vreinterpretq_s32_u64(vmovl_u32(vld1_u32(w))); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
 
     size_t k = kc;
 

diff --git a/src/qs8-gemm/gen/1x16c4-minmax-rndnu-neon-mlal-ld2r.c b/src/qs8-gemm/gen/1x16c4-minmax-rndnu-neon-mlal-ld2r.c
index ad51539..0ca1aef 100644
--- a/src/qs8-gemm/gen/1x16c4-minmax-rndnu-neon-mlal-ld2r.c
+++ b/src/qs8-gemm/gen/1x16c4-minmax-rndnu-neon-mlal-ld2r.c

@@ -41,14 +41,14 @@
   int8_t* c0 = c;
 
   do {
-    int32x4_t vacc0x01 = vreinterpretq_s32_u64(vshll_n_u32(vld1_u32(w), 0)); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
-    int32x4_t vacc0x23 = vreinterpretq_s32_u64(vshll_n_u32(vld1_u32(w), 0)); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
-    int32x4_t vacc0x45 = vreinterpretq_s32_u64(vshll_n_u32(vld1_u32(w), 0)); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
-    int32x4_t vacc0x67 = vreinterpretq_s32_u64(vshll_n_u32(vld1_u32(w), 0)); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
-    int32x4_t vacc0x89 = vreinterpretq_s32_u64(vshll_n_u32(vld1_u32(w), 0)); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
-    int32x4_t vacc0xAB = vreinterpretq_s32_u64(vshll_n_u32(vld1_u32(w), 0)); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
-    int32x4_t vacc0xCD = vreinterpretq_s32_u64(vshll_n_u32(vld1_u32(w), 0)); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
-    int32x4_t vacc0xEF = vreinterpretq_s32_u64(vshll_n_u32(vld1_u32(w), 0)); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
+    int32x4_t vacc0x01 = vreinterpretq_s32_u64(vmovl_u32(vld1_u32(w))); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
+    int32x4_t vacc0x23 = vreinterpretq_s32_u64(vmovl_u32(vld1_u32(w))); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
+    int32x4_t vacc0x45 = vreinterpretq_s32_u64(vmovl_u32(vld1_u32(w))); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
+    int32x4_t vacc0x67 = vreinterpretq_s32_u64(vmovl_u32(vld1_u32(w))); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
+    int32x4_t vacc0x89 = vreinterpretq_s32_u64(vmovl_u32(vld1_u32(w))); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
+    int32x4_t vacc0xAB = vreinterpretq_s32_u64(vmovl_u32(vld1_u32(w))); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
+    int32x4_t vacc0xCD = vreinterpretq_s32_u64(vmovl_u32(vld1_u32(w))); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
+    int32x4_t vacc0xEF = vreinterpretq_s32_u64(vmovl_u32(vld1_u32(w))); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
 
     size_t k = kc;
 

diff --git a/src/qs8-gemm/gen/1x16c4-minmax-rndnu-neon-mull-dup.c b/src/qs8-gemm/gen/1x16c4-minmax-rndnu-neon-mull-dup.c
index 2a52a2e..a5cdffb 100644
--- a/src/qs8-gemm/gen/1x16c4-minmax-rndnu-neon-mull-dup.c
+++ b/src/qs8-gemm/gen/1x16c4-minmax-rndnu-neon-mull-dup.c

@@ -41,14 +41,14 @@
   int8_t* c0 = c;
 
   do {
-    int32x4_t vacc0x01 = vreinterpretq_s32_u64(vshll_n_u32(vld1_u32(w), 0)); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
-    int32x4_t vacc0x23 = vreinterpretq_s32_u64(vshll_n_u32(vld1_u32(w), 0)); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
-    int32x4_t vacc0x45 = vreinterpretq_s32_u64(vshll_n_u32(vld1_u32(w), 0)); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
-    int32x4_t vacc0x67 = vreinterpretq_s32_u64(vshll_n_u32(vld1_u32(w), 0)); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
-    int32x4_t vacc0x89 = vreinterpretq_s32_u64(vshll_n_u32(vld1_u32(w), 0)); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
-    int32x4_t vacc0xAB = vreinterpretq_s32_u64(vshll_n_u32(vld1_u32(w), 0)); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
-    int32x4_t vacc0xCD = vreinterpretq_s32_u64(vshll_n_u32(vld1_u32(w), 0)); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
-    int32x4_t vacc0xEF = vreinterpretq_s32_u64(vshll_n_u32(vld1_u32(w), 0)); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
+    int32x4_t vacc0x01 = vreinterpretq_s32_u64(vmovl_u32(vld1_u32(w))); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
+    int32x4_t vacc0x23 = vreinterpretq_s32_u64(vmovl_u32(vld1_u32(w))); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
+    int32x4_t vacc0x45 = vreinterpretq_s32_u64(vmovl_u32(vld1_u32(w))); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
+    int32x4_t vacc0x67 = vreinterpretq_s32_u64(vmovl_u32(vld1_u32(w))); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
+    int32x4_t vacc0x89 = vreinterpretq_s32_u64(vmovl_u32(vld1_u32(w))); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
+    int32x4_t vacc0xAB = vreinterpretq_s32_u64(vmovl_u32(vld1_u32(w))); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
+    int32x4_t vacc0xCD = vreinterpretq_s32_u64(vmovl_u32(vld1_u32(w))); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
+    int32x4_t vacc0xEF = vreinterpretq_s32_u64(vmovl_u32(vld1_u32(w))); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
 
     size_t k = kc;
 

diff --git a/src/qs8-gemm/gen/1x16c4-minmax-rndnu-neon-mull-ld1r.c b/src/qs8-gemm/gen/1x16c4-minmax-rndnu-neon-mull-ld1r.c
index cde1b13..fb87b77 100644
--- a/src/qs8-gemm/gen/1x16c4-minmax-rndnu-neon-mull-ld1r.c
+++ b/src/qs8-gemm/gen/1x16c4-minmax-rndnu-neon-mull-ld1r.c

@@ -41,14 +41,14 @@
   int8_t* c0 = c;
 
   do {
-    int32x4_t vacc0x01 = vreinterpretq_s32_u64(vshll_n_u32(vld1_u32(w), 0)); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
-    int32x4_t vacc0x23 = vreinterpretq_s32_u64(vshll_n_u32(vld1_u32(w), 0)); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
-    int32x4_t vacc0x45 = vreinterpretq_s32_u64(vshll_n_u32(vld1_u32(w), 0)); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
-    int32x4_t vacc0x67 = vreinterpretq_s32_u64(vshll_n_u32(vld1_u32(w), 0)); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
-    int32x4_t vacc0x89 = vreinterpretq_s32_u64(vshll_n_u32(vld1_u32(w), 0)); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
-    int32x4_t vacc0xAB = vreinterpretq_s32_u64(vshll_n_u32(vld1_u32(w), 0)); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
-    int32x4_t vacc0xCD = vreinterpretq_s32_u64(vshll_n_u32(vld1_u32(w), 0)); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
-    int32x4_t vacc0xEF = vreinterpretq_s32_u64(vshll_n_u32(vld1_u32(w), 0)); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
+    int32x4_t vacc0x01 = vreinterpretq_s32_u64(vmovl_u32(vld1_u32(w))); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
+    int32x4_t vacc0x23 = vreinterpretq_s32_u64(vmovl_u32(vld1_u32(w))); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
+    int32x4_t vacc0x45 = vreinterpretq_s32_u64(vmovl_u32(vld1_u32(w))); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
+    int32x4_t vacc0x67 = vreinterpretq_s32_u64(vmovl_u32(vld1_u32(w))); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
+    int32x4_t vacc0x89 = vreinterpretq_s32_u64(vmovl_u32(vld1_u32(w))); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
+    int32x4_t vacc0xAB = vreinterpretq_s32_u64(vmovl_u32(vld1_u32(w))); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
+    int32x4_t vacc0xCD = vreinterpretq_s32_u64(vmovl_u32(vld1_u32(w))); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
+    int32x4_t vacc0xEF = vreinterpretq_s32_u64(vmovl_u32(vld1_u32(w))); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
 
     size_t k = kc;
 

diff --git a/src/qs8-gemm/gen/1x16c4-minmax-rndnu-neon-mull-ld2r.c b/src/qs8-gemm/gen/1x16c4-minmax-rndnu-neon-mull-ld2r.c
index 4534834..e146c6d 100644
--- a/src/qs8-gemm/gen/1x16c4-minmax-rndnu-neon-mull-ld2r.c
+++ b/src/qs8-gemm/gen/1x16c4-minmax-rndnu-neon-mull-ld2r.c

@@ -41,14 +41,14 @@
   int8_t* c0 = c;
 
   do {
-    int32x4_t vacc0x01 = vreinterpretq_s32_u64(vshll_n_u32(vld1_u32(w), 0)); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
-    int32x4_t vacc0x23 = vreinterpretq_s32_u64(vshll_n_u32(vld1_u32(w), 0)); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
-    int32x4_t vacc0x45 = vreinterpretq_s32_u64(vshll_n_u32(vld1_u32(w), 0)); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
-    int32x4_t vacc0x67 = vreinterpretq_s32_u64(vshll_n_u32(vld1_u32(w), 0)); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
-    int32x4_t vacc0x89 = vreinterpretq_s32_u64(vshll_n_u32(vld1_u32(w), 0)); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
-    int32x4_t vacc0xAB = vreinterpretq_s32_u64(vshll_n_u32(vld1_u32(w), 0)); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
-    int32x4_t vacc0xCD = vreinterpretq_s32_u64(vshll_n_u32(vld1_u32(w), 0)); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
-    int32x4_t vacc0xEF = vreinterpretq_s32_u64(vshll_n_u32(vld1_u32(w), 0)); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
+    int32x4_t vacc0x01 = vreinterpretq_s32_u64(vmovl_u32(vld1_u32(w))); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
+    int32x4_t vacc0x23 = vreinterpretq_s32_u64(vmovl_u32(vld1_u32(w))); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
+    int32x4_t vacc0x45 = vreinterpretq_s32_u64(vmovl_u32(vld1_u32(w))); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
+    int32x4_t vacc0x67 = vreinterpretq_s32_u64(vmovl_u32(vld1_u32(w))); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
+    int32x4_t vacc0x89 = vreinterpretq_s32_u64(vmovl_u32(vld1_u32(w))); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
+    int32x4_t vacc0xAB = vreinterpretq_s32_u64(vmovl_u32(vld1_u32(w))); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
+    int32x4_t vacc0xCD = vreinterpretq_s32_u64(vmovl_u32(vld1_u32(w))); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
+    int32x4_t vacc0xEF = vreinterpretq_s32_u64(vmovl_u32(vld1_u32(w))); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
 
     size_t k = kc;
 

diff --git a/src/qs8-gemm/gen/1x16c4s2-minmax-rndnu-neon-mlal.c b/src/qs8-gemm/gen/1x16c4s2-minmax-rndnu-neon-mlal.c
index 71b7cad..ec936c6 100644
--- a/src/qs8-gemm/gen/1x16c4s2-minmax-rndnu-neon-mlal.c
+++ b/src/qs8-gemm/gen/1x16c4s2-minmax-rndnu-neon-mlal.c

@@ -40,14 +40,14 @@
   int8_t* c0 = c;
 
   do {
-    int32x4_t vacc0x01 = vreinterpretq_s32_u64(vshll_n_u32(vld1_u32(w), 0)); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
-    int32x4_t vacc0x23 = vreinterpretq_s32_u64(vshll_n_u32(vld1_u32(w), 0)); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
-    int32x4_t vacc0x45 = vreinterpretq_s32_u64(vshll_n_u32(vld1_u32(w), 0)); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
-    int32x4_t vacc0x67 = vreinterpretq_s32_u64(vshll_n_u32(vld1_u32(w), 0)); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
-    int32x4_t vacc0x89 = vreinterpretq_s32_u64(vshll_n_u32(vld1_u32(w), 0)); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
-    int32x4_t vacc0xAB = vreinterpretq_s32_u64(vshll_n_u32(vld1_u32(w), 0)); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
-    int32x4_t vacc0xCD = vreinterpretq_s32_u64(vshll_n_u32(vld1_u32(w), 0)); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
-    int32x4_t vacc0xEF = vreinterpretq_s32_u64(vshll_n_u32(vld1_u32(w), 0)); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
+    int32x4_t vacc0x01 = vreinterpretq_s32_u64(vmovl_u32(vld1_u32(w))); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
+    int32x4_t vacc0x23 = vreinterpretq_s32_u64(vmovl_u32(vld1_u32(w))); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
+    int32x4_t vacc0x45 = vreinterpretq_s32_u64(vmovl_u32(vld1_u32(w))); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
+    int32x4_t vacc0x67 = vreinterpretq_s32_u64(vmovl_u32(vld1_u32(w))); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
+    int32x4_t vacc0x89 = vreinterpretq_s32_u64(vmovl_u32(vld1_u32(w))); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
+    int32x4_t vacc0xAB = vreinterpretq_s32_u64(vmovl_u32(vld1_u32(w))); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
+    int32x4_t vacc0xCD = vreinterpretq_s32_u64(vmovl_u32(vld1_u32(w))); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
+    int32x4_t vacc0xEF = vreinterpretq_s32_u64(vmovl_u32(vld1_u32(w))); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
 
     size_t k = kc;
 

diff --git a/src/qs8-gemm/gen/1x16c4s2-minmax-rndnu-neon-mull.c b/src/qs8-gemm/gen/1x16c4s2-minmax-rndnu-neon-mull.c
index 16e9396..293a29f 100644
--- a/src/qs8-gemm/gen/1x16c4s2-minmax-rndnu-neon-mull.c
+++ b/src/qs8-gemm/gen/1x16c4s2-minmax-rndnu-neon-mull.c

@@ -40,14 +40,14 @@
   int8_t* c0 = c;
 
   do {
-    int32x4_t vacc0x01 = vreinterpretq_s32_u64(vshll_n_u32(vld1_u32(w), 0)); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
-    int32x4_t vacc0x23 = vreinterpretq_s32_u64(vshll_n_u32(vld1_u32(w), 0)); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
-    int32x4_t vacc0x45 = vreinterpretq_s32_u64(vshll_n_u32(vld1_u32(w), 0)); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
-    int32x4_t vacc0x67 = vreinterpretq_s32_u64(vshll_n_u32(vld1_u32(w), 0)); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
-    int32x4_t vacc0x89 = vreinterpretq_s32_u64(vshll_n_u32(vld1_u32(w), 0)); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
-    int32x4_t vacc0xAB = vreinterpretq_s32_u64(vshll_n_u32(vld1_u32(w), 0)); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
-    int32x4_t vacc0xCD = vreinterpretq_s32_u64(vshll_n_u32(vld1_u32(w), 0)); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
-    int32x4_t vacc0xEF = vreinterpretq_s32_u64(vshll_n_u32(vld1_u32(w), 0)); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
+    int32x4_t vacc0x01 = vreinterpretq_s32_u64(vmovl_u32(vld1_u32(w))); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
+    int32x4_t vacc0x23 = vreinterpretq_s32_u64(vmovl_u32(vld1_u32(w))); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
+    int32x4_t vacc0x45 = vreinterpretq_s32_u64(vmovl_u32(vld1_u32(w))); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
+    int32x4_t vacc0x67 = vreinterpretq_s32_u64(vmovl_u32(vld1_u32(w))); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
+    int32x4_t vacc0x89 = vreinterpretq_s32_u64(vmovl_u32(vld1_u32(w))); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
+    int32x4_t vacc0xAB = vreinterpretq_s32_u64(vmovl_u32(vld1_u32(w))); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
+    int32x4_t vacc0xCD = vreinterpretq_s32_u64(vmovl_u32(vld1_u32(w))); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
+    int32x4_t vacc0xEF = vreinterpretq_s32_u64(vmovl_u32(vld1_u32(w))); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
 
     size_t k = kc;
 

diff --git a/src/qs8-gemm/gen/1x8c4-minmax-fp32-neon-mlal-dup.c b/src/qs8-gemm/gen/1x8c4-minmax-fp32-neon-mlal-dup.c
index ad1fa96..8bb7bba 100644
--- a/src/qs8-gemm/gen/1x8c4-minmax-fp32-neon-mlal-dup.c
+++ b/src/qs8-gemm/gen/1x8c4-minmax-fp32-neon-mlal-dup.c

@@ -41,10 +41,10 @@
   int8_t* c0 = c;
 
   do {
-    int32x4_t vacc0x01 = vreinterpretq_s32_u64(vshll_n_u32(vld1_u32(w), 0)); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
-    int32x4_t vacc0x23 = vreinterpretq_s32_u64(vshll_n_u32(vld1_u32(w), 0)); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
-    int32x4_t vacc0x45 = vreinterpretq_s32_u64(vshll_n_u32(vld1_u32(w), 0)); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
-    int32x4_t vacc0x67 = vreinterpretq_s32_u64(vshll_n_u32(vld1_u32(w), 0)); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
+    int32x4_t vacc0x01 = vreinterpretq_s32_u64(vmovl_u32(vld1_u32(w))); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
+    int32x4_t vacc0x23 = vreinterpretq_s32_u64(vmovl_u32(vld1_u32(w))); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
+    int32x4_t vacc0x45 = vreinterpretq_s32_u64(vmovl_u32(vld1_u32(w))); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
+    int32x4_t vacc0x67 = vreinterpretq_s32_u64(vmovl_u32(vld1_u32(w))); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
 
     size_t k = kc;
 

diff --git a/src/qs8-gemm/gen/1x8c4-minmax-fp32-neon-mlal-ld1r.c b/src/qs8-gemm/gen/1x8c4-minmax-fp32-neon-mlal-ld1r.c
index b992384..4964d6c 100644
--- a/src/qs8-gemm/gen/1x8c4-minmax-fp32-neon-mlal-ld1r.c
+++ b/src/qs8-gemm/gen/1x8c4-minmax-fp32-neon-mlal-ld1r.c

@@ -41,10 +41,10 @@
   int8_t* c0 = c;
 
   do {
-    int32x4_t vacc0x01 = vreinterpretq_s32_u64(vshll_n_u32(vld1_u32(w), 0)); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
-    int32x4_t vacc0x23 = vreinterpretq_s32_u64(vshll_n_u32(vld1_u32(w), 0)); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
-    int32x4_t vacc0x45 = vreinterpretq_s32_u64(vshll_n_u32(vld1_u32(w), 0)); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
-    int32x4_t vacc0x67 = vreinterpretq_s32_u64(vshll_n_u32(vld1_u32(w), 0)); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
+    int32x4_t vacc0x01 = vreinterpretq_s32_u64(vmovl_u32(vld1_u32(w))); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
+    int32x4_t vacc0x23 = vreinterpretq_s32_u64(vmovl_u32(vld1_u32(w))); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
+    int32x4_t vacc0x45 = vreinterpretq_s32_u64(vmovl_u32(vld1_u32(w))); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
+    int32x4_t vacc0x67 = vreinterpretq_s32_u64(vmovl_u32(vld1_u32(w))); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
 
     size_t k = kc;
 

diff --git a/src/qs8-gemm/gen/1x8c4-minmax-fp32-neon-mlal-ld2r.c b/src/qs8-gemm/gen/1x8c4-minmax-fp32-neon-mlal-ld2r.c
index 184fc02..ee9fe91 100644
--- a/src/qs8-gemm/gen/1x8c4-minmax-fp32-neon-mlal-ld2r.c
+++ b/src/qs8-gemm/gen/1x8c4-minmax-fp32-neon-mlal-ld2r.c

@@ -41,10 +41,10 @@
   int8_t* c0 = c;
 
   do {
-    int32x4_t vacc0x01 = vreinterpretq_s32_u64(vshll_n_u32(vld1_u32(w), 0)); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
-    int32x4_t vacc0x23 = vreinterpretq_s32_u64(vshll_n_u32(vld1_u32(w), 0)); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
-    int32x4_t vacc0x45 = vreinterpretq_s32_u64(vshll_n_u32(vld1_u32(w), 0)); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
-    int32x4_t vacc0x67 = vreinterpretq_s32_u64(vshll_n_u32(vld1_u32(w), 0)); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
+    int32x4_t vacc0x01 = vreinterpretq_s32_u64(vmovl_u32(vld1_u32(w))); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
+    int32x4_t vacc0x23 = vreinterpretq_s32_u64(vmovl_u32(vld1_u32(w))); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
+    int32x4_t vacc0x45 = vreinterpretq_s32_u64(vmovl_u32(vld1_u32(w))); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
+    int32x4_t vacc0x67 = vreinterpretq_s32_u64(vmovl_u32(vld1_u32(w))); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
 
     size_t k = kc;
 

diff --git a/src/qs8-gemm/gen/1x8c4-minmax-fp32-neonv8-mlal-dup.c b/src/qs8-gemm/gen/1x8c4-minmax-fp32-neonv8-mlal-dup.c
index 0005c8c..bdc4f30 100644
--- a/src/qs8-gemm/gen/1x8c4-minmax-fp32-neonv8-mlal-dup.c
+++ b/src/qs8-gemm/gen/1x8c4-minmax-fp32-neonv8-mlal-dup.c

@@ -42,10 +42,10 @@
   int8_t* c0 = c;
 
   do {
-    int32x4_t vacc0x01 = vreinterpretq_s32_u64(vshll_n_u32(vld1_u32(w), 0)); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
-    int32x4_t vacc0x23 = vreinterpretq_s32_u64(vshll_n_u32(vld1_u32(w), 0)); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
-    int32x4_t vacc0x45 = vreinterpretq_s32_u64(vshll_n_u32(vld1_u32(w), 0)); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
-    int32x4_t vacc0x67 = vreinterpretq_s32_u64(vshll_n_u32(vld1_u32(w), 0)); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
+    int32x4_t vacc0x01 = vreinterpretq_s32_u64(vmovl_u32(vld1_u32(w))); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
+    int32x4_t vacc0x23 = vreinterpretq_s32_u64(vmovl_u32(vld1_u32(w))); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
+    int32x4_t vacc0x45 = vreinterpretq_s32_u64(vmovl_u32(vld1_u32(w))); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
+    int32x4_t vacc0x67 = vreinterpretq_s32_u64(vmovl_u32(vld1_u32(w))); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
 
     size_t k = kc;
 

diff --git a/src/qs8-gemm/gen/1x8c4-minmax-fp32-neonv8-mlal-ld1r.c b/src/qs8-gemm/gen/1x8c4-minmax-fp32-neonv8-mlal-ld1r.c
index 8096463..18d76d6 100644
--- a/src/qs8-gemm/gen/1x8c4-minmax-fp32-neonv8-mlal-ld1r.c
+++ b/src/qs8-gemm/gen/1x8c4-minmax-fp32-neonv8-mlal-ld1r.c

@@ -42,10 +42,10 @@
   int8_t* c0 = c;
 
   do {
-    int32x4_t vacc0x01 = vreinterpretq_s32_u64(vshll_n_u32(vld1_u32(w), 0)); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
-    int32x4_t vacc0x23 = vreinterpretq_s32_u64(vshll_n_u32(vld1_u32(w), 0)); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
-    int32x4_t vacc0x45 = vreinterpretq_s32_u64(vshll_n_u32(vld1_u32(w), 0)); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
-    int32x4_t vacc0x67 = vreinterpretq_s32_u64(vshll_n_u32(vld1_u32(w), 0)); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
+    int32x4_t vacc0x01 = vreinterpretq_s32_u64(vmovl_u32(vld1_u32(w))); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
+    int32x4_t vacc0x23 = vreinterpretq_s32_u64(vmovl_u32(vld1_u32(w))); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
+    int32x4_t vacc0x45 = vreinterpretq_s32_u64(vmovl_u32(vld1_u32(w))); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
+    int32x4_t vacc0x67 = vreinterpretq_s32_u64(vmovl_u32(vld1_u32(w))); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
 
     size_t k = kc;
 

diff --git a/src/qs8-gemm/gen/1x8c4-minmax-fp32-neonv8-mlal-ld2r.c b/src/qs8-gemm/gen/1x8c4-minmax-fp32-neonv8-mlal-ld2r.c
index 2b71d83..ee5c327 100644
--- a/src/qs8-gemm/gen/1x8c4-minmax-fp32-neonv8-mlal-ld2r.c
+++ b/src/qs8-gemm/gen/1x8c4-minmax-fp32-neonv8-mlal-ld2r.c

@@ -42,10 +42,10 @@
   int8_t* c0 = c;
 
   do {
-    int32x4_t vacc0x01 = vreinterpretq_s32_u64(vshll_n_u32(vld1_u32(w), 0)); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
-    int32x4_t vacc0x23 = vreinterpretq_s32_u64(vshll_n_u32(vld1_u32(w), 0)); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
-    int32x4_t vacc0x45 = vreinterpretq_s32_u64(vshll_n_u32(vld1_u32(w), 0)); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
-    int32x4_t vacc0x67 = vreinterpretq_s32_u64(vshll_n_u32(vld1_u32(w), 0)); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
+    int32x4_t vacc0x01 = vreinterpretq_s32_u64(vmovl_u32(vld1_u32(w))); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
+    int32x4_t vacc0x23 = vreinterpretq_s32_u64(vmovl_u32(vld1_u32(w))); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
+    int32x4_t vacc0x45 = vreinterpretq_s32_u64(vmovl_u32(vld1_u32(w))); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
+    int32x4_t vacc0x67 = vreinterpretq_s32_u64(vmovl_u32(vld1_u32(w))); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
 
     size_t k = kc;
 

diff --git a/src/qs8-gemm/gen/1x8c4-minmax-rndnu-neon-mlal-dup.c b/src/qs8-gemm/gen/1x8c4-minmax-rndnu-neon-mlal-dup.c
index 33f9ac3..37c992a 100644
--- a/src/qs8-gemm/gen/1x8c4-minmax-rndnu-neon-mlal-dup.c
+++ b/src/qs8-gemm/gen/1x8c4-minmax-rndnu-neon-mlal-dup.c

@@ -41,10 +41,10 @@
   int8_t* c0 = c;
 
   do {
-    int32x4_t vacc0x01 = vreinterpretq_s32_u64(vshll_n_u32(vld1_u32(w), 0)); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
-    int32x4_t vacc0x23 = vreinterpretq_s32_u64(vshll_n_u32(vld1_u32(w), 0)); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
-    int32x4_t vacc0x45 = vreinterpretq_s32_u64(vshll_n_u32(vld1_u32(w), 0)); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
-    int32x4_t vacc0x67 = vreinterpretq_s32_u64(vshll_n_u32(vld1_u32(w), 0)); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
+    int32x4_t vacc0x01 = vreinterpretq_s32_u64(vmovl_u32(vld1_u32(w))); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
+    int32x4_t vacc0x23 = vreinterpretq_s32_u64(vmovl_u32(vld1_u32(w))); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
+    int32x4_t vacc0x45 = vreinterpretq_s32_u64(vmovl_u32(vld1_u32(w))); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
+    int32x4_t vacc0x67 = vreinterpretq_s32_u64(vmovl_u32(vld1_u32(w))); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
 
     size_t k = kc;
 

diff --git a/src/qs8-gemm/gen/1x8c4-minmax-rndnu-neon-mlal-ld1r.c b/src/qs8-gemm/gen/1x8c4-minmax-rndnu-neon-mlal-ld1r.c
index 1eaefc6..e1c4494 100644
--- a/src/qs8-gemm/gen/1x8c4-minmax-rndnu-neon-mlal-ld1r.c
+++ b/src/qs8-gemm/gen/1x8c4-minmax-rndnu-neon-mlal-ld1r.c

@@ -41,10 +41,10 @@
   int8_t* c0 = c;
 
   do {
-    int32x4_t vacc0x01 = vreinterpretq_s32_u64(vshll_n_u32(vld1_u32(w), 0)); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
-    int32x4_t vacc0x23 = vreinterpretq_s32_u64(vshll_n_u32(vld1_u32(w), 0)); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
-    int32x4_t vacc0x45 = vreinterpretq_s32_u64(vshll_n_u32(vld1_u32(w), 0)); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
-    int32x4_t vacc0x67 = vreinterpretq_s32_u64(vshll_n_u32(vld1_u32(w), 0)); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
+    int32x4_t vacc0x01 = vreinterpretq_s32_u64(vmovl_u32(vld1_u32(w))); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
+    int32x4_t vacc0x23 = vreinterpretq_s32_u64(vmovl_u32(vld1_u32(w))); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
+    int32x4_t vacc0x45 = vreinterpretq_s32_u64(vmovl_u32(vld1_u32(w))); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
+    int32x4_t vacc0x67 = vreinterpretq_s32_u64(vmovl_u32(vld1_u32(w))); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
 
     size_t k = kc;
 

diff --git a/src/qs8-gemm/gen/1x8c4-minmax-rndnu-neon-mlal-ld2r.c b/src/qs8-gemm/gen/1x8c4-minmax-rndnu-neon-mlal-ld2r.c
index 7f9f15c..502d4de 100644
--- a/src/qs8-gemm/gen/1x8c4-minmax-rndnu-neon-mlal-ld2r.c
+++ b/src/qs8-gemm/gen/1x8c4-minmax-rndnu-neon-mlal-ld2r.c

@@ -41,10 +41,10 @@
   int8_t* c0 = c;
 
   do {
-    int32x4_t vacc0x01 = vreinterpretq_s32_u64(vshll_n_u32(vld1_u32(w), 0)); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
-    int32x4_t vacc0x23 = vreinterpretq_s32_u64(vshll_n_u32(vld1_u32(w), 0)); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
-    int32x4_t vacc0x45 = vreinterpretq_s32_u64(vshll_n_u32(vld1_u32(w), 0)); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
-    int32x4_t vacc0x67 = vreinterpretq_s32_u64(vshll_n_u32(vld1_u32(w), 0)); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
+    int32x4_t vacc0x01 = vreinterpretq_s32_u64(vmovl_u32(vld1_u32(w))); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
+    int32x4_t vacc0x23 = vreinterpretq_s32_u64(vmovl_u32(vld1_u32(w))); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
+    int32x4_t vacc0x45 = vreinterpretq_s32_u64(vmovl_u32(vld1_u32(w))); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
+    int32x4_t vacc0x67 = vreinterpretq_s32_u64(vmovl_u32(vld1_u32(w))); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
 
     size_t k = kc;
 

diff --git a/src/qs8-gemm/gen/1x8c4-minmax-rndnu-neon-mull-dup.c b/src/qs8-gemm/gen/1x8c4-minmax-rndnu-neon-mull-dup.c
index 0d7617a..b94708c 100644
--- a/src/qs8-gemm/gen/1x8c4-minmax-rndnu-neon-mull-dup.c
+++ b/src/qs8-gemm/gen/1x8c4-minmax-rndnu-neon-mull-dup.c

@@ -41,10 +41,10 @@
   int8_t* c0 = c;
 
   do {
-    int32x4_t vacc0x01 = vreinterpretq_s32_u64(vshll_n_u32(vld1_u32(w), 0)); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
-    int32x4_t vacc0x23 = vreinterpretq_s32_u64(vshll_n_u32(vld1_u32(w), 0)); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
-    int32x4_t vacc0x45 = vreinterpretq_s32_u64(vshll_n_u32(vld1_u32(w), 0)); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
-    int32x4_t vacc0x67 = vreinterpretq_s32_u64(vshll_n_u32(vld1_u32(w), 0)); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
+    int32x4_t vacc0x01 = vreinterpretq_s32_u64(vmovl_u32(vld1_u32(w))); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
+    int32x4_t vacc0x23 = vreinterpretq_s32_u64(vmovl_u32(vld1_u32(w))); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
+    int32x4_t vacc0x45 = vreinterpretq_s32_u64(vmovl_u32(vld1_u32(w))); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
+    int32x4_t vacc0x67 = vreinterpretq_s32_u64(vmovl_u32(vld1_u32(w))); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
 
     size_t k = kc;
 

diff --git a/src/qs8-gemm/gen/1x8c4-minmax-rndnu-neon-mull-ld1r.c b/src/qs8-gemm/gen/1x8c4-minmax-rndnu-neon-mull-ld1r.c
index 679973f..999bf20 100644
--- a/src/qs8-gemm/gen/1x8c4-minmax-rndnu-neon-mull-ld1r.c
+++ b/src/qs8-gemm/gen/1x8c4-minmax-rndnu-neon-mull-ld1r.c

@@ -41,10 +41,10 @@
   int8_t* c0 = c;
 
   do {
-    int32x4_t vacc0x01 = vreinterpretq_s32_u64(vshll_n_u32(vld1_u32(w), 0)); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
-    int32x4_t vacc0x23 = vreinterpretq_s32_u64(vshll_n_u32(vld1_u32(w), 0)); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
-    int32x4_t vacc0x45 = vreinterpretq_s32_u64(vshll_n_u32(vld1_u32(w), 0)); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
-    int32x4_t vacc0x67 = vreinterpretq_s32_u64(vshll_n_u32(vld1_u32(w), 0)); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
+    int32x4_t vacc0x01 = vreinterpretq_s32_u64(vmovl_u32(vld1_u32(w))); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
+    int32x4_t vacc0x23 = vreinterpretq_s32_u64(vmovl_u32(vld1_u32(w))); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
+    int32x4_t vacc0x45 = vreinterpretq_s32_u64(vmovl_u32(vld1_u32(w))); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
+    int32x4_t vacc0x67 = vreinterpretq_s32_u64(vmovl_u32(vld1_u32(w))); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
 
     size_t k = kc;
 

diff --git a/src/qs8-gemm/gen/1x8c4-minmax-rndnu-neon-mull-ld2r.c b/src/qs8-gemm/gen/1x8c4-minmax-rndnu-neon-mull-ld2r.c
index 0ff83fc..e8cf945 100644
--- a/src/qs8-gemm/gen/1x8c4-minmax-rndnu-neon-mull-ld2r.c
+++ b/src/qs8-gemm/gen/1x8c4-minmax-rndnu-neon-mull-ld2r.c

@@ -41,10 +41,10 @@
   int8_t* c0 = c;
 
   do {
-    int32x4_t vacc0x01 = vreinterpretq_s32_u64(vshll_n_u32(vld1_u32(w), 0)); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
-    int32x4_t vacc0x23 = vreinterpretq_s32_u64(vshll_n_u32(vld1_u32(w), 0)); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
-    int32x4_t vacc0x45 = vreinterpretq_s32_u64(vshll_n_u32(vld1_u32(w), 0)); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
-    int32x4_t vacc0x67 = vreinterpretq_s32_u64(vshll_n_u32(vld1_u32(w), 0)); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
+    int32x4_t vacc0x01 = vreinterpretq_s32_u64(vmovl_u32(vld1_u32(w))); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
+    int32x4_t vacc0x23 = vreinterpretq_s32_u64(vmovl_u32(vld1_u32(w))); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
+    int32x4_t vacc0x45 = vreinterpretq_s32_u64(vmovl_u32(vld1_u32(w))); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
+    int32x4_t vacc0x67 = vreinterpretq_s32_u64(vmovl_u32(vld1_u32(w))); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
 
     size_t k = kc;
 

diff --git a/src/qs8-gemm/gen/1x8c4s2-minmax-fp32-neon-mlal.c b/src/qs8-gemm/gen/1x8c4s2-minmax-fp32-neon-mlal.c
index 56eac2d..3fa4271 100644
--- a/src/qs8-gemm/gen/1x8c4s2-minmax-fp32-neon-mlal.c
+++ b/src/qs8-gemm/gen/1x8c4s2-minmax-fp32-neon-mlal.c

@@ -40,10 +40,10 @@
   int8_t* c0 = c;
 
   do {
-    int32x4_t vacc0x01 = vreinterpretq_s32_u64(vshll_n_u32(vld1_u32(w), 0)); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
-    int32x4_t vacc0x23 = vreinterpretq_s32_u64(vshll_n_u32(vld1_u32(w), 0)); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
-    int32x4_t vacc0x45 = vreinterpretq_s32_u64(vshll_n_u32(vld1_u32(w), 0)); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
-    int32x4_t vacc0x67 = vreinterpretq_s32_u64(vshll_n_u32(vld1_u32(w), 0)); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
+    int32x4_t vacc0x01 = vreinterpretq_s32_u64(vmovl_u32(vld1_u32(w))); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
+    int32x4_t vacc0x23 = vreinterpretq_s32_u64(vmovl_u32(vld1_u32(w))); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
+    int32x4_t vacc0x45 = vreinterpretq_s32_u64(vmovl_u32(vld1_u32(w))); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
+    int32x4_t vacc0x67 = vreinterpretq_s32_u64(vmovl_u32(vld1_u32(w))); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
 
     size_t k = kc;
 

diff --git a/src/qs8-gemm/gen/1x8c4s2-minmax-fp32-neonv8-mlal.c b/src/qs8-gemm/gen/1x8c4s2-minmax-fp32-neonv8-mlal.c
index 9c75520..c46ee2d 100644
--- a/src/qs8-gemm/gen/1x8c4s2-minmax-fp32-neonv8-mlal.c
+++ b/src/qs8-gemm/gen/1x8c4s2-minmax-fp32-neonv8-mlal.c

@@ -41,10 +41,10 @@
   int8_t* c0 = c;
 
   do {
-    int32x4_t vacc0x01 = vreinterpretq_s32_u64(vshll_n_u32(vld1_u32(w), 0)); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
-    int32x4_t vacc0x23 = vreinterpretq_s32_u64(vshll_n_u32(vld1_u32(w), 0)); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
-    int32x4_t vacc0x45 = vreinterpretq_s32_u64(vshll_n_u32(vld1_u32(w), 0)); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
-    int32x4_t vacc0x67 = vreinterpretq_s32_u64(vshll_n_u32(vld1_u32(w), 0)); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
+    int32x4_t vacc0x01 = vreinterpretq_s32_u64(vmovl_u32(vld1_u32(w))); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
+    int32x4_t vacc0x23 = vreinterpretq_s32_u64(vmovl_u32(vld1_u32(w))); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
+    int32x4_t vacc0x45 = vreinterpretq_s32_u64(vmovl_u32(vld1_u32(w))); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
+    int32x4_t vacc0x67 = vreinterpretq_s32_u64(vmovl_u32(vld1_u32(w))); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
 
     size_t k = kc;
 

diff --git a/src/qs8-gemm/gen/1x8c4s2-minmax-rndnu-neon-mlal.c b/src/qs8-gemm/gen/1x8c4s2-minmax-rndnu-neon-mlal.c
index f59f511..a22e428 100644
--- a/src/qs8-gemm/gen/1x8c4s2-minmax-rndnu-neon-mlal.c
+++ b/src/qs8-gemm/gen/1x8c4s2-minmax-rndnu-neon-mlal.c

@@ -40,10 +40,10 @@
   int8_t* c0 = c;
 
   do {
-    int32x4_t vacc0x01 = vreinterpretq_s32_u64(vshll_n_u32(vld1_u32(w), 0)); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
-    int32x4_t vacc0x23 = vreinterpretq_s32_u64(vshll_n_u32(vld1_u32(w), 0)); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
-    int32x4_t vacc0x45 = vreinterpretq_s32_u64(vshll_n_u32(vld1_u32(w), 0)); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
-    int32x4_t vacc0x67 = vreinterpretq_s32_u64(vshll_n_u32(vld1_u32(w), 0)); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
+    int32x4_t vacc0x01 = vreinterpretq_s32_u64(vmovl_u32(vld1_u32(w))); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
+    int32x4_t vacc0x23 = vreinterpretq_s32_u64(vmovl_u32(vld1_u32(w))); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
+    int32x4_t vacc0x45 = vreinterpretq_s32_u64(vmovl_u32(vld1_u32(w))); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
+    int32x4_t vacc0x67 = vreinterpretq_s32_u64(vmovl_u32(vld1_u32(w))); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
 
     size_t k = kc;
 

diff --git a/src/qs8-gemm/gen/1x8c4s2-minmax-rndnu-neon-mull.c b/src/qs8-gemm/gen/1x8c4s2-minmax-rndnu-neon-mull.c
index 1c98a55..1b2eef9 100644
--- a/src/qs8-gemm/gen/1x8c4s2-minmax-rndnu-neon-mull.c
+++ b/src/qs8-gemm/gen/1x8c4s2-minmax-rndnu-neon-mull.c

@@ -40,10 +40,10 @@
   int8_t* c0 = c;
 
   do {
-    int32x4_t vacc0x01 = vreinterpretq_s32_u64(vshll_n_u32(vld1_u32(w), 0)); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
-    int32x4_t vacc0x23 = vreinterpretq_s32_u64(vshll_n_u32(vld1_u32(w), 0)); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
-    int32x4_t vacc0x45 = vreinterpretq_s32_u64(vshll_n_u32(vld1_u32(w), 0)); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
-    int32x4_t vacc0x67 = vreinterpretq_s32_u64(vshll_n_u32(vld1_u32(w), 0)); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
+    int32x4_t vacc0x01 = vreinterpretq_s32_u64(vmovl_u32(vld1_u32(w))); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
+    int32x4_t vacc0x23 = vreinterpretq_s32_u64(vmovl_u32(vld1_u32(w))); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
+    int32x4_t vacc0x45 = vreinterpretq_s32_u64(vmovl_u32(vld1_u32(w))); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
+    int32x4_t vacc0x67 = vreinterpretq_s32_u64(vmovl_u32(vld1_u32(w))); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
 
     size_t k = kc;
 

diff --git a/src/qs8-gemm/gen/2x16c4-minmax-rndnu-neon-mlal-dup.c b/src/qs8-gemm/gen/2x16c4-minmax-rndnu-neon-mlal-dup.c
index 29e9d80..79323fb 100644
--- a/src/qs8-gemm/gen/2x16c4-minmax-rndnu-neon-mlal-dup.c
+++ b/src/qs8-gemm/gen/2x16c4-minmax-rndnu-neon-mlal-dup.c

@@ -47,14 +47,14 @@
   }
 
   do {
-    int32x4_t vacc0x01 = vreinterpretq_s32_u64(vshll_n_u32(vld1_u32(w), 0)); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
-    int32x4_t vacc0x23 = vreinterpretq_s32_u64(vshll_n_u32(vld1_u32(w), 0)); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
-    int32x4_t vacc0x45 = vreinterpretq_s32_u64(vshll_n_u32(vld1_u32(w), 0)); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
-    int32x4_t vacc0x67 = vreinterpretq_s32_u64(vshll_n_u32(vld1_u32(w), 0)); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
-    int32x4_t vacc0x89 = vreinterpretq_s32_u64(vshll_n_u32(vld1_u32(w), 0)); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
-    int32x4_t vacc0xAB = vreinterpretq_s32_u64(vshll_n_u32(vld1_u32(w), 0)); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
-    int32x4_t vacc0xCD = vreinterpretq_s32_u64(vshll_n_u32(vld1_u32(w), 0)); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
-    int32x4_t vacc0xEF = vreinterpretq_s32_u64(vshll_n_u32(vld1_u32(w), 0)); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
+    int32x4_t vacc0x01 = vreinterpretq_s32_u64(vmovl_u32(vld1_u32(w))); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
+    int32x4_t vacc0x23 = vreinterpretq_s32_u64(vmovl_u32(vld1_u32(w))); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
+    int32x4_t vacc0x45 = vreinterpretq_s32_u64(vmovl_u32(vld1_u32(w))); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
+    int32x4_t vacc0x67 = vreinterpretq_s32_u64(vmovl_u32(vld1_u32(w))); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
+    int32x4_t vacc0x89 = vreinterpretq_s32_u64(vmovl_u32(vld1_u32(w))); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
+    int32x4_t vacc0xAB = vreinterpretq_s32_u64(vmovl_u32(vld1_u32(w))); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
+    int32x4_t vacc0xCD = vreinterpretq_s32_u64(vmovl_u32(vld1_u32(w))); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
+    int32x4_t vacc0xEF = vreinterpretq_s32_u64(vmovl_u32(vld1_u32(w))); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
     int32x4_t vacc1x01 = vacc0x01;
     int32x4_t vacc1x23 = vacc0x23;
     int32x4_t vacc1x45 = vacc0x45;

diff --git a/src/qs8-gemm/gen/2x16c4-minmax-rndnu-neon-mlal-ld1r.c b/src/qs8-gemm/gen/2x16c4-minmax-rndnu-neon-mlal-ld1r.c
index 5410db2..ced046f 100644
--- a/src/qs8-gemm/gen/2x16c4-minmax-rndnu-neon-mlal-ld1r.c
+++ b/src/qs8-gemm/gen/2x16c4-minmax-rndnu-neon-mlal-ld1r.c

@@ -47,14 +47,14 @@
   }
 
   do {
-    int32x4_t vacc0x01 = vreinterpretq_s32_u64(vshll_n_u32(vld1_u32(w), 0)); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
-    int32x4_t vacc0x23 = vreinterpretq_s32_u64(vshll_n_u32(vld1_u32(w), 0)); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
-    int32x4_t vacc0x45 = vreinterpretq_s32_u64(vshll_n_u32(vld1_u32(w), 0)); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
-    int32x4_t vacc0x67 = vreinterpretq_s32_u64(vshll_n_u32(vld1_u32(w), 0)); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
-    int32x4_t vacc0x89 = vreinterpretq_s32_u64(vshll_n_u32(vld1_u32(w), 0)); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
-    int32x4_t vacc0xAB = vreinterpretq_s32_u64(vshll_n_u32(vld1_u32(w), 0)); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
-    int32x4_t vacc0xCD = vreinterpretq_s32_u64(vshll_n_u32(vld1_u32(w), 0)); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
-    int32x4_t vacc0xEF = vreinterpretq_s32_u64(vshll_n_u32(vld1_u32(w), 0)); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
+    int32x4_t vacc0x01 = vreinterpretq_s32_u64(vmovl_u32(vld1_u32(w))); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
+    int32x4_t vacc0x23 = vreinterpretq_s32_u64(vmovl_u32(vld1_u32(w))); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
+    int32x4_t vacc0x45 = vreinterpretq_s32_u64(vmovl_u32(vld1_u32(w))); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
+    int32x4_t vacc0x67 = vreinterpretq_s32_u64(vmovl_u32(vld1_u32(w))); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
+    int32x4_t vacc0x89 = vreinterpretq_s32_u64(vmovl_u32(vld1_u32(w))); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
+    int32x4_t vacc0xAB = vreinterpretq_s32_u64(vmovl_u32(vld1_u32(w))); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
+    int32x4_t vacc0xCD = vreinterpretq_s32_u64(vmovl_u32(vld1_u32(w))); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
+    int32x4_t vacc0xEF = vreinterpretq_s32_u64(vmovl_u32(vld1_u32(w))); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
     int32x4_t vacc1x01 = vacc0x01;
     int32x4_t vacc1x23 = vacc0x23;
     int32x4_t vacc1x45 = vacc0x45;

diff --git a/src/qs8-gemm/gen/2x16c4-minmax-rndnu-neon-mlal-ld2r.c b/src/qs8-gemm/gen/2x16c4-minmax-rndnu-neon-mlal-ld2r.c
index a8e4a37..109e5e6 100644
--- a/src/qs8-gemm/gen/2x16c4-minmax-rndnu-neon-mlal-ld2r.c
+++ b/src/qs8-gemm/gen/2x16c4-minmax-rndnu-neon-mlal-ld2r.c

@@ -47,14 +47,14 @@
   }
 
   do {
-    int32x4_t vacc0x01 = vreinterpretq_s32_u64(vshll_n_u32(vld1_u32(w), 0)); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
-    int32x4_t vacc0x23 = vreinterpretq_s32_u64(vshll_n_u32(vld1_u32(w), 0)); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
-    int32x4_t vacc0x45 = vreinterpretq_s32_u64(vshll_n_u32(vld1_u32(w), 0)); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
-    int32x4_t vacc0x67 = vreinterpretq_s32_u64(vshll_n_u32(vld1_u32(w), 0)); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
-    int32x4_t vacc0x89 = vreinterpretq_s32_u64(vshll_n_u32(vld1_u32(w), 0)); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
-    int32x4_t vacc0xAB = vreinterpretq_s32_u64(vshll_n_u32(vld1_u32(w), 0)); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
-    int32x4_t vacc0xCD = vreinterpretq_s32_u64(vshll_n_u32(vld1_u32(w), 0)); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
-    int32x4_t vacc0xEF = vreinterpretq_s32_u64(vshll_n_u32(vld1_u32(w), 0)); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
+    int32x4_t vacc0x01 = vreinterpretq_s32_u64(vmovl_u32(vld1_u32(w))); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
+    int32x4_t vacc0x23 = vreinterpretq_s32_u64(vmovl_u32(vld1_u32(w))); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
+    int32x4_t vacc0x45 = vreinterpretq_s32_u64(vmovl_u32(vld1_u32(w))); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
+    int32x4_t vacc0x67 = vreinterpretq_s32_u64(vmovl_u32(vld1_u32(w))); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
+    int32x4_t vacc0x89 = vreinterpretq_s32_u64(vmovl_u32(vld1_u32(w))); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
+    int32x4_t vacc0xAB = vreinterpretq_s32_u64(vmovl_u32(vld1_u32(w))); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
+    int32x4_t vacc0xCD = vreinterpretq_s32_u64(vmovl_u32(vld1_u32(w))); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
+    int32x4_t vacc0xEF = vreinterpretq_s32_u64(vmovl_u32(vld1_u32(w))); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
     int32x4_t vacc1x01 = vacc0x01;
     int32x4_t vacc1x23 = vacc0x23;
     int32x4_t vacc1x45 = vacc0x45;

diff --git a/src/qs8-gemm/gen/2x16c4-minmax-rndnu-neon-mull-dup.c b/src/qs8-gemm/gen/2x16c4-minmax-rndnu-neon-mull-dup.c
index cbdbfea..8870218 100644
--- a/src/qs8-gemm/gen/2x16c4-minmax-rndnu-neon-mull-dup.c
+++ b/src/qs8-gemm/gen/2x16c4-minmax-rndnu-neon-mull-dup.c

@@ -47,14 +47,14 @@
   }
 
   do {
-    int32x4_t vacc0x01 = vreinterpretq_s32_u64(vshll_n_u32(vld1_u32(w), 0)); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
-    int32x4_t vacc0x23 = vreinterpretq_s32_u64(vshll_n_u32(vld1_u32(w), 0)); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
-    int32x4_t vacc0x45 = vreinterpretq_s32_u64(vshll_n_u32(vld1_u32(w), 0)); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
-    int32x4_t vacc0x67 = vreinterpretq_s32_u64(vshll_n_u32(vld1_u32(w), 0)); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
-    int32x4_t vacc0x89 = vreinterpretq_s32_u64(vshll_n_u32(vld1_u32(w), 0)); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
-    int32x4_t vacc0xAB = vreinterpretq_s32_u64(vshll_n_u32(vld1_u32(w), 0)); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
-    int32x4_t vacc0xCD = vreinterpretq_s32_u64(vshll_n_u32(vld1_u32(w), 0)); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
-    int32x4_t vacc0xEF = vreinterpretq_s32_u64(vshll_n_u32(vld1_u32(w), 0)); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
+    int32x4_t vacc0x01 = vreinterpretq_s32_u64(vmovl_u32(vld1_u32(w))); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
+    int32x4_t vacc0x23 = vreinterpretq_s32_u64(vmovl_u32(vld1_u32(w))); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
+    int32x4_t vacc0x45 = vreinterpretq_s32_u64(vmovl_u32(vld1_u32(w))); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
+    int32x4_t vacc0x67 = vreinterpretq_s32_u64(vmovl_u32(vld1_u32(w))); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
+    int32x4_t vacc0x89 = vreinterpretq_s32_u64(vmovl_u32(vld1_u32(w))); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
+    int32x4_t vacc0xAB = vreinterpretq_s32_u64(vmovl_u32(vld1_u32(w))); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
+    int32x4_t vacc0xCD = vreinterpretq_s32_u64(vmovl_u32(vld1_u32(w))); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
+    int32x4_t vacc0xEF = vreinterpretq_s32_u64(vmovl_u32(vld1_u32(w))); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
     int32x4_t vacc1x01 = vacc0x01;
     int32x4_t vacc1x23 = vacc0x23;
     int32x4_t vacc1x45 = vacc0x45;

diff --git a/src/qs8-gemm/gen/2x16c4-minmax-rndnu-neon-mull-ld1r.c b/src/qs8-gemm/gen/2x16c4-minmax-rndnu-neon-mull-ld1r.c
index 494885e..5b96e1e 100644
--- a/src/qs8-gemm/gen/2x16c4-minmax-rndnu-neon-mull-ld1r.c
+++ b/src/qs8-gemm/gen/2x16c4-minmax-rndnu-neon-mull-ld1r.c

@@ -47,14 +47,14 @@
   }
 
   do {
-    int32x4_t vacc0x01 = vreinterpretq_s32_u64(vshll_n_u32(vld1_u32(w), 0)); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
-    int32x4_t vacc0x23 = vreinterpretq_s32_u64(vshll_n_u32(vld1_u32(w), 0)); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
-    int32x4_t vacc0x45 = vreinterpretq_s32_u64(vshll_n_u32(vld1_u32(w), 0)); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
-    int32x4_t vacc0x67 = vreinterpretq_s32_u64(vshll_n_u32(vld1_u32(w), 0)); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
-    int32x4_t vacc0x89 = vreinterpretq_s32_u64(vshll_n_u32(vld1_u32(w), 0)); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
-    int32x4_t vacc0xAB = vreinterpretq_s32_u64(vshll_n_u32(vld1_u32(w), 0)); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
-    int32x4_t vacc0xCD = vreinterpretq_s32_u64(vshll_n_u32(vld1_u32(w), 0)); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
-    int32x4_t vacc0xEF = vreinterpretq_s32_u64(vshll_n_u32(vld1_u32(w), 0)); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
+    int32x4_t vacc0x01 = vreinterpretq_s32_u64(vmovl_u32(vld1_u32(w))); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
+    int32x4_t vacc0x23 = vreinterpretq_s32_u64(vmovl_u32(vld1_u32(w))); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
+    int32x4_t vacc0x45 = vreinterpretq_s32_u64(vmovl_u32(vld1_u32(w))); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
+    int32x4_t vacc0x67 = vreinterpretq_s32_u64(vmovl_u32(vld1_u32(w))); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
+    int32x4_t vacc0x89 = vreinterpretq_s32_u64(vmovl_u32(vld1_u32(w))); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
+    int32x4_t vacc0xAB = vreinterpretq_s32_u64(vmovl_u32(vld1_u32(w))); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
+    int32x4_t vacc0xCD = vreinterpretq_s32_u64(vmovl_u32(vld1_u32(w))); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
+    int32x4_t vacc0xEF = vreinterpretq_s32_u64(vmovl_u32(vld1_u32(w))); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
     int32x4_t vacc1x01 = vacc0x01;
     int32x4_t vacc1x23 = vacc0x23;
     int32x4_t vacc1x45 = vacc0x45;

diff --git a/src/qs8-gemm/gen/2x16c4-minmax-rndnu-neon-mull-ld2r.c b/src/qs8-gemm/gen/2x16c4-minmax-rndnu-neon-mull-ld2r.c
index 6d14b1b..40b65bf 100644
--- a/src/qs8-gemm/gen/2x16c4-minmax-rndnu-neon-mull-ld2r.c
+++ b/src/qs8-gemm/gen/2x16c4-minmax-rndnu-neon-mull-ld2r.c

@@ -47,14 +47,14 @@
   }
 
   do {
-    int32x4_t vacc0x01 = vreinterpretq_s32_u64(vshll_n_u32(vld1_u32(w), 0)); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
-    int32x4_t vacc0x23 = vreinterpretq_s32_u64(vshll_n_u32(vld1_u32(w), 0)); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
-    int32x4_t vacc0x45 = vreinterpretq_s32_u64(vshll_n_u32(vld1_u32(w), 0)); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
-    int32x4_t vacc0x67 = vreinterpretq_s32_u64(vshll_n_u32(vld1_u32(w), 0)); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
-    int32x4_t vacc0x89 = vreinterpretq_s32_u64(vshll_n_u32(vld1_u32(w), 0)); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
-    int32x4_t vacc0xAB = vreinterpretq_s32_u64(vshll_n_u32(vld1_u32(w), 0)); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
-    int32x4_t vacc0xCD = vreinterpretq_s32_u64(vshll_n_u32(vld1_u32(w), 0)); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
-    int32x4_t vacc0xEF = vreinterpretq_s32_u64(vshll_n_u32(vld1_u32(w), 0)); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
+    int32x4_t vacc0x01 = vreinterpretq_s32_u64(vmovl_u32(vld1_u32(w))); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
+    int32x4_t vacc0x23 = vreinterpretq_s32_u64(vmovl_u32(vld1_u32(w))); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
+    int32x4_t vacc0x45 = vreinterpretq_s32_u64(vmovl_u32(vld1_u32(w))); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
+    int32x4_t vacc0x67 = vreinterpretq_s32_u64(vmovl_u32(vld1_u32(w))); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
+    int32x4_t vacc0x89 = vreinterpretq_s32_u64(vmovl_u32(vld1_u32(w))); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
+    int32x4_t vacc0xAB = vreinterpretq_s32_u64(vmovl_u32(vld1_u32(w))); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
+    int32x4_t vacc0xCD = vreinterpretq_s32_u64(vmovl_u32(vld1_u32(w))); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
+    int32x4_t vacc0xEF = vreinterpretq_s32_u64(vmovl_u32(vld1_u32(w))); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
     int32x4_t vacc1x01 = vacc0x01;
     int32x4_t vacc1x23 = vacc0x23;
     int32x4_t vacc1x45 = vacc0x45;

diff --git a/src/qs8-gemm/gen/2x16c4s2-minmax-rndnu-neon-mlal.c b/src/qs8-gemm/gen/2x16c4s2-minmax-rndnu-neon-mlal.c
index 4fa8182..5abde66 100644
--- a/src/qs8-gemm/gen/2x16c4s2-minmax-rndnu-neon-mlal.c
+++ b/src/qs8-gemm/gen/2x16c4s2-minmax-rndnu-neon-mlal.c

@@ -46,14 +46,14 @@
   }
 
   do {
-    int32x4_t vacc0x01 = vreinterpretq_s32_u64(vshll_n_u32(vld1_u32(w), 0)); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
-    int32x4_t vacc0x23 = vreinterpretq_s32_u64(vshll_n_u32(vld1_u32(w), 0)); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
-    int32x4_t vacc0x45 = vreinterpretq_s32_u64(vshll_n_u32(vld1_u32(w), 0)); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
-    int32x4_t vacc0x67 = vreinterpretq_s32_u64(vshll_n_u32(vld1_u32(w), 0)); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
-    int32x4_t vacc0x89 = vreinterpretq_s32_u64(vshll_n_u32(vld1_u32(w), 0)); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
-    int32x4_t vacc0xAB = vreinterpretq_s32_u64(vshll_n_u32(vld1_u32(w), 0)); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
-    int32x4_t vacc0xCD = vreinterpretq_s32_u64(vshll_n_u32(vld1_u32(w), 0)); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
-    int32x4_t vacc0xEF = vreinterpretq_s32_u64(vshll_n_u32(vld1_u32(w), 0)); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
+    int32x4_t vacc0x01 = vreinterpretq_s32_u64(vmovl_u32(vld1_u32(w))); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
+    int32x4_t vacc0x23 = vreinterpretq_s32_u64(vmovl_u32(vld1_u32(w))); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
+    int32x4_t vacc0x45 = vreinterpretq_s32_u64(vmovl_u32(vld1_u32(w))); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
+    int32x4_t vacc0x67 = vreinterpretq_s32_u64(vmovl_u32(vld1_u32(w))); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
+    int32x4_t vacc0x89 = vreinterpretq_s32_u64(vmovl_u32(vld1_u32(w))); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
+    int32x4_t vacc0xAB = vreinterpretq_s32_u64(vmovl_u32(vld1_u32(w))); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
+    int32x4_t vacc0xCD = vreinterpretq_s32_u64(vmovl_u32(vld1_u32(w))); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
+    int32x4_t vacc0xEF = vreinterpretq_s32_u64(vmovl_u32(vld1_u32(w))); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
     int32x4_t vacc1x01 = vacc0x01;
     int32x4_t vacc1x23 = vacc0x23;
     int32x4_t vacc1x45 = vacc0x45;

diff --git a/src/qs8-gemm/gen/2x16c4s2-minmax-rndnu-neon-mull.c b/src/qs8-gemm/gen/2x16c4s2-minmax-rndnu-neon-mull.c
index 1c29818..4e5f3e3 100644
--- a/src/qs8-gemm/gen/2x16c4s2-minmax-rndnu-neon-mull.c
+++ b/src/qs8-gemm/gen/2x16c4s2-minmax-rndnu-neon-mull.c

@@ -46,14 +46,14 @@
   }
 
   do {
-    int32x4_t vacc0x01 = vreinterpretq_s32_u64(vshll_n_u32(vld1_u32(w), 0)); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
-    int32x4_t vacc0x23 = vreinterpretq_s32_u64(vshll_n_u32(vld1_u32(w), 0)); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
-    int32x4_t vacc0x45 = vreinterpretq_s32_u64(vshll_n_u32(vld1_u32(w), 0)); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
-    int32x4_t vacc0x67 = vreinterpretq_s32_u64(vshll_n_u32(vld1_u32(w), 0)); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
-    int32x4_t vacc0x89 = vreinterpretq_s32_u64(vshll_n_u32(vld1_u32(w), 0)); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
-    int32x4_t vacc0xAB = vreinterpretq_s32_u64(vshll_n_u32(vld1_u32(w), 0)); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
-    int32x4_t vacc0xCD = vreinterpretq_s32_u64(vshll_n_u32(vld1_u32(w), 0)); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
-    int32x4_t vacc0xEF = vreinterpretq_s32_u64(vshll_n_u32(vld1_u32(w), 0)); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
+    int32x4_t vacc0x01 = vreinterpretq_s32_u64(vmovl_u32(vld1_u32(w))); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
+    int32x4_t vacc0x23 = vreinterpretq_s32_u64(vmovl_u32(vld1_u32(w))); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
+    int32x4_t vacc0x45 = vreinterpretq_s32_u64(vmovl_u32(vld1_u32(w))); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
+    int32x4_t vacc0x67 = vreinterpretq_s32_u64(vmovl_u32(vld1_u32(w))); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
+    int32x4_t vacc0x89 = vreinterpretq_s32_u64(vmovl_u32(vld1_u32(w))); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
+    int32x4_t vacc0xAB = vreinterpretq_s32_u64(vmovl_u32(vld1_u32(w))); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
+    int32x4_t vacc0xCD = vreinterpretq_s32_u64(vmovl_u32(vld1_u32(w))); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
+    int32x4_t vacc0xEF = vreinterpretq_s32_u64(vmovl_u32(vld1_u32(w))); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
     int32x4_t vacc1x01 = vacc0x01;
     int32x4_t vacc1x23 = vacc0x23;
     int32x4_t vacc1x45 = vacc0x45;

diff --git a/src/qs8-gemm/gen/2x8c4-minmax-fp32-neon-mlal-dup.c b/src/qs8-gemm/gen/2x8c4-minmax-fp32-neon-mlal-dup.c
index 9928bbb..92e63eb 100644
--- a/src/qs8-gemm/gen/2x8c4-minmax-fp32-neon-mlal-dup.c
+++ b/src/qs8-gemm/gen/2x8c4-minmax-fp32-neon-mlal-dup.c

@@ -47,10 +47,10 @@
   }
 
   do {
-    int32x4_t vacc0x01 = vreinterpretq_s32_u64(vshll_n_u32(vld1_u32(w), 0)); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
-    int32x4_t vacc0x23 = vreinterpretq_s32_u64(vshll_n_u32(vld1_u32(w), 0)); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
-    int32x4_t vacc0x45 = vreinterpretq_s32_u64(vshll_n_u32(vld1_u32(w), 0)); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
-    int32x4_t vacc0x67 = vreinterpretq_s32_u64(vshll_n_u32(vld1_u32(w), 0)); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
+    int32x4_t vacc0x01 = vreinterpretq_s32_u64(vmovl_u32(vld1_u32(w))); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
+    int32x4_t vacc0x23 = vreinterpretq_s32_u64(vmovl_u32(vld1_u32(w))); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
+    int32x4_t vacc0x45 = vreinterpretq_s32_u64(vmovl_u32(vld1_u32(w))); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
+    int32x4_t vacc0x67 = vreinterpretq_s32_u64(vmovl_u32(vld1_u32(w))); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
     int32x4_t vacc1x01 = vacc0x01;
     int32x4_t vacc1x23 = vacc0x23;
     int32x4_t vacc1x45 = vacc0x45;

diff --git a/src/qs8-gemm/gen/2x8c4-minmax-fp32-neon-mlal-ld1r.c b/src/qs8-gemm/gen/2x8c4-minmax-fp32-neon-mlal-ld1r.c
index 4c0086b..8be9361 100644
--- a/src/qs8-gemm/gen/2x8c4-minmax-fp32-neon-mlal-ld1r.c
+++ b/src/qs8-gemm/gen/2x8c4-minmax-fp32-neon-mlal-ld1r.c

@@ -47,10 +47,10 @@
   }
 
   do {
-    int32x4_t vacc0x01 = vreinterpretq_s32_u64(vshll_n_u32(vld1_u32(w), 0)); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
-    int32x4_t vacc0x23 = vreinterpretq_s32_u64(vshll_n_u32(vld1_u32(w), 0)); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
-    int32x4_t vacc0x45 = vreinterpretq_s32_u64(vshll_n_u32(vld1_u32(w), 0)); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
-    int32x4_t vacc0x67 = vreinterpretq_s32_u64(vshll_n_u32(vld1_u32(w), 0)); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
+    int32x4_t vacc0x01 = vreinterpretq_s32_u64(vmovl_u32(vld1_u32(w))); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
+    int32x4_t vacc0x23 = vreinterpretq_s32_u64(vmovl_u32(vld1_u32(w))); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
+    int32x4_t vacc0x45 = vreinterpretq_s32_u64(vmovl_u32(vld1_u32(w))); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
+    int32x4_t vacc0x67 = vreinterpretq_s32_u64(vmovl_u32(vld1_u32(w))); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
     int32x4_t vacc1x01 = vacc0x01;
     int32x4_t vacc1x23 = vacc0x23;
     int32x4_t vacc1x45 = vacc0x45;

diff --git a/src/qs8-gemm/gen/2x8c4-minmax-fp32-neon-mlal-ld2r.c b/src/qs8-gemm/gen/2x8c4-minmax-fp32-neon-mlal-ld2r.c
index 7fe2531..53d700d 100644
--- a/src/qs8-gemm/gen/2x8c4-minmax-fp32-neon-mlal-ld2r.c
+++ b/src/qs8-gemm/gen/2x8c4-minmax-fp32-neon-mlal-ld2r.c

@@ -47,10 +47,10 @@
   }
 
   do {
-    int32x4_t vacc0x01 = vreinterpretq_s32_u64(vshll_n_u32(vld1_u32(w), 0)); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
-    int32x4_t vacc0x23 = vreinterpretq_s32_u64(vshll_n_u32(vld1_u32(w), 0)); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
-    int32x4_t vacc0x45 = vreinterpretq_s32_u64(vshll_n_u32(vld1_u32(w), 0)); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
-    int32x4_t vacc0x67 = vreinterpretq_s32_u64(vshll_n_u32(vld1_u32(w), 0)); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
+    int32x4_t vacc0x01 = vreinterpretq_s32_u64(vmovl_u32(vld1_u32(w))); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
+    int32x4_t vacc0x23 = vreinterpretq_s32_u64(vmovl_u32(vld1_u32(w))); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
+    int32x4_t vacc0x45 = vreinterpretq_s32_u64(vmovl_u32(vld1_u32(w))); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
+    int32x4_t vacc0x67 = vreinterpretq_s32_u64(vmovl_u32(vld1_u32(w))); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
     int32x4_t vacc1x01 = vacc0x01;
     int32x4_t vacc1x23 = vacc0x23;
     int32x4_t vacc1x45 = vacc0x45;

diff --git a/src/qs8-gemm/gen/2x8c4-minmax-fp32-neonv8-mlal-dup.c b/src/qs8-gemm/gen/2x8c4-minmax-fp32-neonv8-mlal-dup.c
index 5f45523..59ed7c3 100644
--- a/src/qs8-gemm/gen/2x8c4-minmax-fp32-neonv8-mlal-dup.c
+++ b/src/qs8-gemm/gen/2x8c4-minmax-fp32-neonv8-mlal-dup.c

@@ -48,10 +48,10 @@
   }
 
   do {
-    int32x4_t vacc0x01 = vreinterpretq_s32_u64(vshll_n_u32(vld1_u32(w), 0)); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
-    int32x4_t vacc0x23 = vreinterpretq_s32_u64(vshll_n_u32(vld1_u32(w), 0)); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
-    int32x4_t vacc0x45 = vreinterpretq_s32_u64(vshll_n_u32(vld1_u32(w), 0)); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
-    int32x4_t vacc0x67 = vreinterpretq_s32_u64(vshll_n_u32(vld1_u32(w), 0)); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
+    int32x4_t vacc0x01 = vreinterpretq_s32_u64(vmovl_u32(vld1_u32(w))); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
+    int32x4_t vacc0x23 = vreinterpretq_s32_u64(vmovl_u32(vld1_u32(w))); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
+    int32x4_t vacc0x45 = vreinterpretq_s32_u64(vmovl_u32(vld1_u32(w))); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
+    int32x4_t vacc0x67 = vreinterpretq_s32_u64(vmovl_u32(vld1_u32(w))); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
     int32x4_t vacc1x01 = vacc0x01;
     int32x4_t vacc1x23 = vacc0x23;
     int32x4_t vacc1x45 = vacc0x45;

diff --git a/src/qs8-gemm/gen/2x8c4-minmax-fp32-neonv8-mlal-ld1r.c b/src/qs8-gemm/gen/2x8c4-minmax-fp32-neonv8-mlal-ld1r.c
index e187d00..e66c2a9 100644
--- a/src/qs8-gemm/gen/2x8c4-minmax-fp32-neonv8-mlal-ld1r.c
+++ b/src/qs8-gemm/gen/2x8c4-minmax-fp32-neonv8-mlal-ld1r.c

@@ -48,10 +48,10 @@
   }
 
   do {
-    int32x4_t vacc0x01 = vreinterpretq_s32_u64(vshll_n_u32(vld1_u32(w), 0)); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
-    int32x4_t vacc0x23 = vreinterpretq_s32_u64(vshll_n_u32(vld1_u32(w), 0)); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
-    int32x4_t vacc0x45 = vreinterpretq_s32_u64(vshll_n_u32(vld1_u32(w), 0)); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
-    int32x4_t vacc0x67 = vreinterpretq_s32_u64(vshll_n_u32(vld1_u32(w), 0)); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
+    int32x4_t vacc0x01 = vreinterpretq_s32_u64(vmovl_u32(vld1_u32(w))); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
+    int32x4_t vacc0x23 = vreinterpretq_s32_u64(vmovl_u32(vld1_u32(w))); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
+    int32x4_t vacc0x45 = vreinterpretq_s32_u64(vmovl_u32(vld1_u32(w))); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
+    int32x4_t vacc0x67 = vreinterpretq_s32_u64(vmovl_u32(vld1_u32(w))); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
     int32x4_t vacc1x01 = vacc0x01;
     int32x4_t vacc1x23 = vacc0x23;
     int32x4_t vacc1x45 = vacc0x45;

diff --git a/src/qs8-gemm/gen/2x8c4-minmax-fp32-neonv8-mlal-ld2r.c b/src/qs8-gemm/gen/2x8c4-minmax-fp32-neonv8-mlal-ld2r.c
index fc915b0..4d8bd58 100644
--- a/src/qs8-gemm/gen/2x8c4-minmax-fp32-neonv8-mlal-ld2r.c
+++ b/src/qs8-gemm/gen/2x8c4-minmax-fp32-neonv8-mlal-ld2r.c

@@ -48,10 +48,10 @@
   }
 
   do {
-    int32x4_t vacc0x01 = vreinterpretq_s32_u64(vshll_n_u32(vld1_u32(w), 0)); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
-    int32x4_t vacc0x23 = vreinterpretq_s32_u64(vshll_n_u32(vld1_u32(w), 0)); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
-    int32x4_t vacc0x45 = vreinterpretq_s32_u64(vshll_n_u32(vld1_u32(w), 0)); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
-    int32x4_t vacc0x67 = vreinterpretq_s32_u64(vshll_n_u32(vld1_u32(w), 0)); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
+    int32x4_t vacc0x01 = vreinterpretq_s32_u64(vmovl_u32(vld1_u32(w))); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
+    int32x4_t vacc0x23 = vreinterpretq_s32_u64(vmovl_u32(vld1_u32(w))); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
+    int32x4_t vacc0x45 = vreinterpretq_s32_u64(vmovl_u32(vld1_u32(w))); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
+    int32x4_t vacc0x67 = vreinterpretq_s32_u64(vmovl_u32(vld1_u32(w))); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
     int32x4_t vacc1x01 = vacc0x01;
     int32x4_t vacc1x23 = vacc0x23;
     int32x4_t vacc1x45 = vacc0x45;

diff --git a/src/qs8-gemm/gen/2x8c4-minmax-rndnu-neon-mlal-dup.c b/src/qs8-gemm/gen/2x8c4-minmax-rndnu-neon-mlal-dup.c
index dcc53ce..d2d6bca 100644
--- a/src/qs8-gemm/gen/2x8c4-minmax-rndnu-neon-mlal-dup.c
+++ b/src/qs8-gemm/gen/2x8c4-minmax-rndnu-neon-mlal-dup.c

@@ -47,10 +47,10 @@
   }
 
   do {
-    int32x4_t vacc0x01 = vreinterpretq_s32_u64(vshll_n_u32(vld1_u32(w), 0)); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
-    int32x4_t vacc0x23 = vreinterpretq_s32_u64(vshll_n_u32(vld1_u32(w), 0)); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
-    int32x4_t vacc0x45 = vreinterpretq_s32_u64(vshll_n_u32(vld1_u32(w), 0)); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
-    int32x4_t vacc0x67 = vreinterpretq_s32_u64(vshll_n_u32(vld1_u32(w), 0)); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
+    int32x4_t vacc0x01 = vreinterpretq_s32_u64(vmovl_u32(vld1_u32(w))); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
+    int32x4_t vacc0x23 = vreinterpretq_s32_u64(vmovl_u32(vld1_u32(w))); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
+    int32x4_t vacc0x45 = vreinterpretq_s32_u64(vmovl_u32(vld1_u32(w))); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
+    int32x4_t vacc0x67 = vreinterpretq_s32_u64(vmovl_u32(vld1_u32(w))); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
     int32x4_t vacc1x01 = vacc0x01;
     int32x4_t vacc1x23 = vacc0x23;
     int32x4_t vacc1x45 = vacc0x45;

diff --git a/src/qs8-gemm/gen/2x8c4-minmax-rndnu-neon-mlal-ld1r.c b/src/qs8-gemm/gen/2x8c4-minmax-rndnu-neon-mlal-ld1r.c
index d8138c1..80456c9 100644
--- a/src/qs8-gemm/gen/2x8c4-minmax-rndnu-neon-mlal-ld1r.c
+++ b/src/qs8-gemm/gen/2x8c4-minmax-rndnu-neon-mlal-ld1r.c

@@ -47,10 +47,10 @@
   }
 
   do {
-    int32x4_t vacc0x01 = vreinterpretq_s32_u64(vshll_n_u32(vld1_u32(w), 0)); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
-    int32x4_t vacc0x23 = vreinterpretq_s32_u64(vshll_n_u32(vld1_u32(w), 0)); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
-    int32x4_t vacc0x45 = vreinterpretq_s32_u64(vshll_n_u32(vld1_u32(w), 0)); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
-    int32x4_t vacc0x67 = vreinterpretq_s32_u64(vshll_n_u32(vld1_u32(w), 0)); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
+    int32x4_t vacc0x01 = vreinterpretq_s32_u64(vmovl_u32(vld1_u32(w))); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
+    int32x4_t vacc0x23 = vreinterpretq_s32_u64(vmovl_u32(vld1_u32(w))); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
+    int32x4_t vacc0x45 = vreinterpretq_s32_u64(vmovl_u32(vld1_u32(w))); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
+    int32x4_t vacc0x67 = vreinterpretq_s32_u64(vmovl_u32(vld1_u32(w))); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
     int32x4_t vacc1x01 = vacc0x01;
     int32x4_t vacc1x23 = vacc0x23;
     int32x4_t vacc1x45 = vacc0x45;

diff --git a/src/qs8-gemm/gen/2x8c4-minmax-rndnu-neon-mlal-ld2r.c b/src/qs8-gemm/gen/2x8c4-minmax-rndnu-neon-mlal-ld2r.c
index ced5145..4bcc362 100644
--- a/src/qs8-gemm/gen/2x8c4-minmax-rndnu-neon-mlal-ld2r.c
+++ b/src/qs8-gemm/gen/2x8c4-minmax-rndnu-neon-mlal-ld2r.c

@@ -47,10 +47,10 @@
   }
 
   do {
-    int32x4_t vacc0x01 = vreinterpretq_s32_u64(vshll_n_u32(vld1_u32(w), 0)); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
-    int32x4_t vacc0x23 = vreinterpretq_s32_u64(vshll_n_u32(vld1_u32(w), 0)); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
-    int32x4_t vacc0x45 = vreinterpretq_s32_u64(vshll_n_u32(vld1_u32(w), 0)); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
-    int32x4_t vacc0x67 = vreinterpretq_s32_u64(vshll_n_u32(vld1_u32(w), 0)); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
+    int32x4_t vacc0x01 = vreinterpretq_s32_u64(vmovl_u32(vld1_u32(w))); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
+    int32x4_t vacc0x23 = vreinterpretq_s32_u64(vmovl_u32(vld1_u32(w))); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
+    int32x4_t vacc0x45 = vreinterpretq_s32_u64(vmovl_u32(vld1_u32(w))); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
+    int32x4_t vacc0x67 = vreinterpretq_s32_u64(vmovl_u32(vld1_u32(w))); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
     int32x4_t vacc1x01 = vacc0x01;
     int32x4_t vacc1x23 = vacc0x23;
     int32x4_t vacc1x45 = vacc0x45;

diff --git a/src/qs8-gemm/gen/2x8c4-minmax-rndnu-neon-mull-dup.c b/src/qs8-gemm/gen/2x8c4-minmax-rndnu-neon-mull-dup.c
index 359c8fb..722f85d 100644
--- a/src/qs8-gemm/gen/2x8c4-minmax-rndnu-neon-mull-dup.c
+++ b/src/qs8-gemm/gen/2x8c4-minmax-rndnu-neon-mull-dup.c

@@ -47,10 +47,10 @@
   }
 
   do {
-    int32x4_t vacc0x01 = vreinterpretq_s32_u64(vshll_n_u32(vld1_u32(w), 0)); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
-    int32x4_t vacc0x23 = vreinterpretq_s32_u64(vshll_n_u32(vld1_u32(w), 0)); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
-    int32x4_t vacc0x45 = vreinterpretq_s32_u64(vshll_n_u32(vld1_u32(w), 0)); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
-    int32x4_t vacc0x67 = vreinterpretq_s32_u64(vshll_n_u32(vld1_u32(w), 0)); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
+    int32x4_t vacc0x01 = vreinterpretq_s32_u64(vmovl_u32(vld1_u32(w))); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
+    int32x4_t vacc0x23 = vreinterpretq_s32_u64(vmovl_u32(vld1_u32(w))); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
+    int32x4_t vacc0x45 = vreinterpretq_s32_u64(vmovl_u32(vld1_u32(w))); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
+    int32x4_t vacc0x67 = vreinterpretq_s32_u64(vmovl_u32(vld1_u32(w))); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
     int32x4_t vacc1x01 = vacc0x01;
     int32x4_t vacc1x23 = vacc0x23;
     int32x4_t vacc1x45 = vacc0x45;

diff --git a/src/qs8-gemm/gen/2x8c4-minmax-rndnu-neon-mull-ld1r.c b/src/qs8-gemm/gen/2x8c4-minmax-rndnu-neon-mull-ld1r.c
index 35215b7..4d39823 100644
--- a/src/qs8-gemm/gen/2x8c4-minmax-rndnu-neon-mull-ld1r.c
+++ b/src/qs8-gemm/gen/2x8c4-minmax-rndnu-neon-mull-ld1r.c

@@ -47,10 +47,10 @@
   }
 
   do {
-    int32x4_t vacc0x01 = vreinterpretq_s32_u64(vshll_n_u32(vld1_u32(w), 0)); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
-    int32x4_t vacc0x23 = vreinterpretq_s32_u64(vshll_n_u32(vld1_u32(w), 0)); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
-    int32x4_t vacc0x45 = vreinterpretq_s32_u64(vshll_n_u32(vld1_u32(w), 0)); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
-    int32x4_t vacc0x67 = vreinterpretq_s32_u64(vshll_n_u32(vld1_u32(w), 0)); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
+    int32x4_t vacc0x01 = vreinterpretq_s32_u64(vmovl_u32(vld1_u32(w))); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
+    int32x4_t vacc0x23 = vreinterpretq_s32_u64(vmovl_u32(vld1_u32(w))); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
+    int32x4_t vacc0x45 = vreinterpretq_s32_u64(vmovl_u32(vld1_u32(w))); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
+    int32x4_t vacc0x67 = vreinterpretq_s32_u64(vmovl_u32(vld1_u32(w))); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
     int32x4_t vacc1x01 = vacc0x01;
     int32x4_t vacc1x23 = vacc0x23;
     int32x4_t vacc1x45 = vacc0x45;

diff --git a/src/qs8-gemm/gen/2x8c4-minmax-rndnu-neon-mull-ld2r.c b/src/qs8-gemm/gen/2x8c4-minmax-rndnu-neon-mull-ld2r.c
index 5440417..cd7ae3f 100644
--- a/src/qs8-gemm/gen/2x8c4-minmax-rndnu-neon-mull-ld2r.c
+++ b/src/qs8-gemm/gen/2x8c4-minmax-rndnu-neon-mull-ld2r.c

@@ -47,10 +47,10 @@
   }
 
   do {
-    int32x4_t vacc0x01 = vreinterpretq_s32_u64(vshll_n_u32(vld1_u32(w), 0)); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
-    int32x4_t vacc0x23 = vreinterpretq_s32_u64(vshll_n_u32(vld1_u32(w), 0)); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
-    int32x4_t vacc0x45 = vreinterpretq_s32_u64(vshll_n_u32(vld1_u32(w), 0)); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
-    int32x4_t vacc0x67 = vreinterpretq_s32_u64(vshll_n_u32(vld1_u32(w), 0)); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
+    int32x4_t vacc0x01 = vreinterpretq_s32_u64(vmovl_u32(vld1_u32(w))); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
+    int32x4_t vacc0x23 = vreinterpretq_s32_u64(vmovl_u32(vld1_u32(w))); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
+    int32x4_t vacc0x45 = vreinterpretq_s32_u64(vmovl_u32(vld1_u32(w))); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
+    int32x4_t vacc0x67 = vreinterpretq_s32_u64(vmovl_u32(vld1_u32(w))); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
     int32x4_t vacc1x01 = vacc0x01;
     int32x4_t vacc1x23 = vacc0x23;
     int32x4_t vacc1x45 = vacc0x45;

diff --git a/src/qs8-gemm/gen/2x8c4s2-minmax-fp32-neon-mlal.c b/src/qs8-gemm/gen/2x8c4s2-minmax-fp32-neon-mlal.c
index a7a0a04..e7a3399 100644
--- a/src/qs8-gemm/gen/2x8c4s2-minmax-fp32-neon-mlal.c
+++ b/src/qs8-gemm/gen/2x8c4s2-minmax-fp32-neon-mlal.c

@@ -46,10 +46,10 @@
   }
 
   do {
-    int32x4_t vacc0x01 = vreinterpretq_s32_u64(vshll_n_u32(vld1_u32(w), 0)); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
-    int32x4_t vacc0x23 = vreinterpretq_s32_u64(vshll_n_u32(vld1_u32(w), 0)); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
-    int32x4_t vacc0x45 = vreinterpretq_s32_u64(vshll_n_u32(vld1_u32(w), 0)); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
-    int32x4_t vacc0x67 = vreinterpretq_s32_u64(vshll_n_u32(vld1_u32(w), 0)); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
+    int32x4_t vacc0x01 = vreinterpretq_s32_u64(vmovl_u32(vld1_u32(w))); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
+    int32x4_t vacc0x23 = vreinterpretq_s32_u64(vmovl_u32(vld1_u32(w))); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
+    int32x4_t vacc0x45 = vreinterpretq_s32_u64(vmovl_u32(vld1_u32(w))); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
+    int32x4_t vacc0x67 = vreinterpretq_s32_u64(vmovl_u32(vld1_u32(w))); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
     int32x4_t vacc1x01 = vacc0x01;
     int32x4_t vacc1x23 = vacc0x23;
     int32x4_t vacc1x45 = vacc0x45;

diff --git a/src/qs8-gemm/gen/2x8c4s2-minmax-fp32-neonv8-mlal.c b/src/qs8-gemm/gen/2x8c4s2-minmax-fp32-neonv8-mlal.c
index b7da090..9a6bf8b 100644
--- a/src/qs8-gemm/gen/2x8c4s2-minmax-fp32-neonv8-mlal.c
+++ b/src/qs8-gemm/gen/2x8c4s2-minmax-fp32-neonv8-mlal.c

@@ -47,10 +47,10 @@
   }
 
   do {
-    int32x4_t vacc0x01 = vreinterpretq_s32_u64(vshll_n_u32(vld1_u32(w), 0)); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
-    int32x4_t vacc0x23 = vreinterpretq_s32_u64(vshll_n_u32(vld1_u32(w), 0)); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
-    int32x4_t vacc0x45 = vreinterpretq_s32_u64(vshll_n_u32(vld1_u32(w), 0)); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
-    int32x4_t vacc0x67 = vreinterpretq_s32_u64(vshll_n_u32(vld1_u32(w), 0)); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
+    int32x4_t vacc0x01 = vreinterpretq_s32_u64(vmovl_u32(vld1_u32(w))); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
+    int32x4_t vacc0x23 = vreinterpretq_s32_u64(vmovl_u32(vld1_u32(w))); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
+    int32x4_t vacc0x45 = vreinterpretq_s32_u64(vmovl_u32(vld1_u32(w))); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
+    int32x4_t vacc0x67 = vreinterpretq_s32_u64(vmovl_u32(vld1_u32(w))); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
     int32x4_t vacc1x01 = vacc0x01;
     int32x4_t vacc1x23 = vacc0x23;
     int32x4_t vacc1x45 = vacc0x45;

diff --git a/src/qs8-gemm/gen/2x8c4s2-minmax-rndnu-neon-mlal.c b/src/qs8-gemm/gen/2x8c4s2-minmax-rndnu-neon-mlal.c
index 103c779..28a2efc 100644
--- a/src/qs8-gemm/gen/2x8c4s2-minmax-rndnu-neon-mlal.c
+++ b/src/qs8-gemm/gen/2x8c4s2-minmax-rndnu-neon-mlal.c

@@ -46,10 +46,10 @@
   }
 
   do {
-    int32x4_t vacc0x01 = vreinterpretq_s32_u64(vshll_n_u32(vld1_u32(w), 0)); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
-    int32x4_t vacc0x23 = vreinterpretq_s32_u64(vshll_n_u32(vld1_u32(w), 0)); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
-    int32x4_t vacc0x45 = vreinterpretq_s32_u64(vshll_n_u32(vld1_u32(w), 0)); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
-    int32x4_t vacc0x67 = vreinterpretq_s32_u64(vshll_n_u32(vld1_u32(w), 0)); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
+    int32x4_t vacc0x01 = vreinterpretq_s32_u64(vmovl_u32(vld1_u32(w))); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
+    int32x4_t vacc0x23 = vreinterpretq_s32_u64(vmovl_u32(vld1_u32(w))); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
+    int32x4_t vacc0x45 = vreinterpretq_s32_u64(vmovl_u32(vld1_u32(w))); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
+    int32x4_t vacc0x67 = vreinterpretq_s32_u64(vmovl_u32(vld1_u32(w))); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
     int32x4_t vacc1x01 = vacc0x01;
     int32x4_t vacc1x23 = vacc0x23;
     int32x4_t vacc1x45 = vacc0x45;

diff --git a/src/qs8-gemm/gen/2x8c4s2-minmax-rndnu-neon-mull.c b/src/qs8-gemm/gen/2x8c4s2-minmax-rndnu-neon-mull.c
index a127c6f..db6c921 100644
--- a/src/qs8-gemm/gen/2x8c4s2-minmax-rndnu-neon-mull.c
+++ b/src/qs8-gemm/gen/2x8c4s2-minmax-rndnu-neon-mull.c

@@ -46,10 +46,10 @@
   }
 
   do {
-    int32x4_t vacc0x01 = vreinterpretq_s32_u64(vshll_n_u32(vld1_u32(w), 0)); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
-    int32x4_t vacc0x23 = vreinterpretq_s32_u64(vshll_n_u32(vld1_u32(w), 0)); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
-    int32x4_t vacc0x45 = vreinterpretq_s32_u64(vshll_n_u32(vld1_u32(w), 0)); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
-    int32x4_t vacc0x67 = vreinterpretq_s32_u64(vshll_n_u32(vld1_u32(w), 0)); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
+    int32x4_t vacc0x01 = vreinterpretq_s32_u64(vmovl_u32(vld1_u32(w))); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
+    int32x4_t vacc0x23 = vreinterpretq_s32_u64(vmovl_u32(vld1_u32(w))); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
+    int32x4_t vacc0x45 = vreinterpretq_s32_u64(vmovl_u32(vld1_u32(w))); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
+    int32x4_t vacc0x67 = vreinterpretq_s32_u64(vmovl_u32(vld1_u32(w))); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
     int32x4_t vacc1x01 = vacc0x01;
     int32x4_t vacc1x23 = vacc0x23;
     int32x4_t vacc1x45 = vacc0x45;

diff --git a/src/qs8-gemm/gen/3x16c4-minmax-rndnu-neon-mlal-dup.c b/src/qs8-gemm/gen/3x16c4-minmax-rndnu-neon-mlal-dup.c
index e449d85..cac5f50 100644
--- a/src/qs8-gemm/gen/3x16c4-minmax-rndnu-neon-mlal-dup.c
+++ b/src/qs8-gemm/gen/3x16c4-minmax-rndnu-neon-mlal-dup.c

@@ -53,14 +53,14 @@
   }
 
   do {
-    int32x4_t vacc0x01 = vreinterpretq_s32_u64(vshll_n_u32(vld1_u32(w), 0)); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
-    int32x4_t vacc0x23 = vreinterpretq_s32_u64(vshll_n_u32(vld1_u32(w), 0)); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
-    int32x4_t vacc0x45 = vreinterpretq_s32_u64(vshll_n_u32(vld1_u32(w), 0)); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
-    int32x4_t vacc0x67 = vreinterpretq_s32_u64(vshll_n_u32(vld1_u32(w), 0)); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
-    int32x4_t vacc0x89 = vreinterpretq_s32_u64(vshll_n_u32(vld1_u32(w), 0)); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
-    int32x4_t vacc0xAB = vreinterpretq_s32_u64(vshll_n_u32(vld1_u32(w), 0)); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
-    int32x4_t vacc0xCD = vreinterpretq_s32_u64(vshll_n_u32(vld1_u32(w), 0)); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
-    int32x4_t vacc0xEF = vreinterpretq_s32_u64(vshll_n_u32(vld1_u32(w), 0)); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
+    int32x4_t vacc0x01 = vreinterpretq_s32_u64(vmovl_u32(vld1_u32(w))); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
+    int32x4_t vacc0x23 = vreinterpretq_s32_u64(vmovl_u32(vld1_u32(w))); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
+    int32x4_t vacc0x45 = vreinterpretq_s32_u64(vmovl_u32(vld1_u32(w))); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
+    int32x4_t vacc0x67 = vreinterpretq_s32_u64(vmovl_u32(vld1_u32(w))); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
+    int32x4_t vacc0x89 = vreinterpretq_s32_u64(vmovl_u32(vld1_u32(w))); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
+    int32x4_t vacc0xAB = vreinterpretq_s32_u64(vmovl_u32(vld1_u32(w))); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
+    int32x4_t vacc0xCD = vreinterpretq_s32_u64(vmovl_u32(vld1_u32(w))); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
+    int32x4_t vacc0xEF = vreinterpretq_s32_u64(vmovl_u32(vld1_u32(w))); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
     int32x4_t vacc1x01 = vacc0x01;
     int32x4_t vacc1x23 = vacc0x23;
     int32x4_t vacc1x45 = vacc0x45;

diff --git a/src/qs8-gemm/gen/3x16c4-minmax-rndnu-neon-mlal-ld1r.c b/src/qs8-gemm/gen/3x16c4-minmax-rndnu-neon-mlal-ld1r.c
index 1b3be50..44f3be0 100644
--- a/src/qs8-gemm/gen/3x16c4-minmax-rndnu-neon-mlal-ld1r.c
+++ b/src/qs8-gemm/gen/3x16c4-minmax-rndnu-neon-mlal-ld1r.c

@@ -53,14 +53,14 @@
   }
 
   do {
-    int32x4_t vacc0x01 = vreinterpretq_s32_u64(vshll_n_u32(vld1_u32(w), 0)); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
-    int32x4_t vacc0x23 = vreinterpretq_s32_u64(vshll_n_u32(vld1_u32(w), 0)); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
-    int32x4_t vacc0x45 = vreinterpretq_s32_u64(vshll_n_u32(vld1_u32(w), 0)); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
-    int32x4_t vacc0x67 = vreinterpretq_s32_u64(vshll_n_u32(vld1_u32(w), 0)); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
-    int32x4_t vacc0x89 = vreinterpretq_s32_u64(vshll_n_u32(vld1_u32(w), 0)); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
-    int32x4_t vacc0xAB = vreinterpretq_s32_u64(vshll_n_u32(vld1_u32(w), 0)); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
-    int32x4_t vacc0xCD = vreinterpretq_s32_u64(vshll_n_u32(vld1_u32(w), 0)); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
-    int32x4_t vacc0xEF = vreinterpretq_s32_u64(vshll_n_u32(vld1_u32(w), 0)); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
+    int32x4_t vacc0x01 = vreinterpretq_s32_u64(vmovl_u32(vld1_u32(w))); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
+    int32x4_t vacc0x23 = vreinterpretq_s32_u64(vmovl_u32(vld1_u32(w))); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
+    int32x4_t vacc0x45 = vreinterpretq_s32_u64(vmovl_u32(vld1_u32(w))); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
+    int32x4_t vacc0x67 = vreinterpretq_s32_u64(vmovl_u32(vld1_u32(w))); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
+    int32x4_t vacc0x89 = vreinterpretq_s32_u64(vmovl_u32(vld1_u32(w))); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
+    int32x4_t vacc0xAB = vreinterpretq_s32_u64(vmovl_u32(vld1_u32(w))); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
+    int32x4_t vacc0xCD = vreinterpretq_s32_u64(vmovl_u32(vld1_u32(w))); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
+    int32x4_t vacc0xEF = vreinterpretq_s32_u64(vmovl_u32(vld1_u32(w))); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
     int32x4_t vacc1x01 = vacc0x01;
     int32x4_t vacc1x23 = vacc0x23;
     int32x4_t vacc1x45 = vacc0x45;

diff --git a/src/qs8-gemm/gen/3x16c4-minmax-rndnu-neon-mlal-ld2r.c b/src/qs8-gemm/gen/3x16c4-minmax-rndnu-neon-mlal-ld2r.c
index e80a181..9d4e0fa 100644
--- a/src/qs8-gemm/gen/3x16c4-minmax-rndnu-neon-mlal-ld2r.c
+++ b/src/qs8-gemm/gen/3x16c4-minmax-rndnu-neon-mlal-ld2r.c

@@ -53,14 +53,14 @@
   }
 
   do {
-    int32x4_t vacc0x01 = vreinterpretq_s32_u64(vshll_n_u32(vld1_u32(w), 0)); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
-    int32x4_t vacc0x23 = vreinterpretq_s32_u64(vshll_n_u32(vld1_u32(w), 0)); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
-    int32x4_t vacc0x45 = vreinterpretq_s32_u64(vshll_n_u32(vld1_u32(w), 0)); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
-    int32x4_t vacc0x67 = vreinterpretq_s32_u64(vshll_n_u32(vld1_u32(w), 0)); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
-    int32x4_t vacc0x89 = vreinterpretq_s32_u64(vshll_n_u32(vld1_u32(w), 0)); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
-    int32x4_t vacc0xAB = vreinterpretq_s32_u64(vshll_n_u32(vld1_u32(w), 0)); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
-    int32x4_t vacc0xCD = vreinterpretq_s32_u64(vshll_n_u32(vld1_u32(w), 0)); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
-    int32x4_t vacc0xEF = vreinterpretq_s32_u64(vshll_n_u32(vld1_u32(w), 0)); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
+    int32x4_t vacc0x01 = vreinterpretq_s32_u64(vmovl_u32(vld1_u32(w))); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
+    int32x4_t vacc0x23 = vreinterpretq_s32_u64(vmovl_u32(vld1_u32(w))); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
+    int32x4_t vacc0x45 = vreinterpretq_s32_u64(vmovl_u32(vld1_u32(w))); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
+    int32x4_t vacc0x67 = vreinterpretq_s32_u64(vmovl_u32(vld1_u32(w))); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
+    int32x4_t vacc0x89 = vreinterpretq_s32_u64(vmovl_u32(vld1_u32(w))); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
+    int32x4_t vacc0xAB = vreinterpretq_s32_u64(vmovl_u32(vld1_u32(w))); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
+    int32x4_t vacc0xCD = vreinterpretq_s32_u64(vmovl_u32(vld1_u32(w))); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
+    int32x4_t vacc0xEF = vreinterpretq_s32_u64(vmovl_u32(vld1_u32(w))); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
     int32x4_t vacc1x01 = vacc0x01;
     int32x4_t vacc1x23 = vacc0x23;
     int32x4_t vacc1x45 = vacc0x45;

diff --git a/src/qs8-gemm/gen/3x16c4-minmax-rndnu-neon-mull-dup.c b/src/qs8-gemm/gen/3x16c4-minmax-rndnu-neon-mull-dup.c
index 0095ee3..a420495 100644
--- a/src/qs8-gemm/gen/3x16c4-minmax-rndnu-neon-mull-dup.c
+++ b/src/qs8-gemm/gen/3x16c4-minmax-rndnu-neon-mull-dup.c

@@ -53,14 +53,14 @@
   }
 
   do {
-    int32x4_t vacc0x01 = vreinterpretq_s32_u64(vshll_n_u32(vld1_u32(w), 0)); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
-    int32x4_t vacc0x23 = vreinterpretq_s32_u64(vshll_n_u32(vld1_u32(w), 0)); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
-    int32x4_t vacc0x45 = vreinterpretq_s32_u64(vshll_n_u32(vld1_u32(w), 0)); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
-    int32x4_t vacc0x67 = vreinterpretq_s32_u64(vshll_n_u32(vld1_u32(w), 0)); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
-    int32x4_t vacc0x89 = vreinterpretq_s32_u64(vshll_n_u32(vld1_u32(w), 0)); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
-    int32x4_t vacc0xAB = vreinterpretq_s32_u64(vshll_n_u32(vld1_u32(w), 0)); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
-    int32x4_t vacc0xCD = vreinterpretq_s32_u64(vshll_n_u32(vld1_u32(w), 0)); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
-    int32x4_t vacc0xEF = vreinterpretq_s32_u64(vshll_n_u32(vld1_u32(w), 0)); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
+    int32x4_t vacc0x01 = vreinterpretq_s32_u64(vmovl_u32(vld1_u32(w))); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
+    int32x4_t vacc0x23 = vreinterpretq_s32_u64(vmovl_u32(vld1_u32(w))); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
+    int32x4_t vacc0x45 = vreinterpretq_s32_u64(vmovl_u32(vld1_u32(w))); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
+    int32x4_t vacc0x67 = vreinterpretq_s32_u64(vmovl_u32(vld1_u32(w))); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
+    int32x4_t vacc0x89 = vreinterpretq_s32_u64(vmovl_u32(vld1_u32(w))); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
+    int32x4_t vacc0xAB = vreinterpretq_s32_u64(vmovl_u32(vld1_u32(w))); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
+    int32x4_t vacc0xCD = vreinterpretq_s32_u64(vmovl_u32(vld1_u32(w))); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
+    int32x4_t vacc0xEF = vreinterpretq_s32_u64(vmovl_u32(vld1_u32(w))); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
     int32x4_t vacc1x01 = vacc0x01;
     int32x4_t vacc1x23 = vacc0x23;
     int32x4_t vacc1x45 = vacc0x45;

diff --git a/src/qs8-gemm/gen/3x16c4-minmax-rndnu-neon-mull-ld1r.c b/src/qs8-gemm/gen/3x16c4-minmax-rndnu-neon-mull-ld1r.c
index 57a95b4..a5bd2aa 100644
--- a/src/qs8-gemm/gen/3x16c4-minmax-rndnu-neon-mull-ld1r.c
+++ b/src/qs8-gemm/gen/3x16c4-minmax-rndnu-neon-mull-ld1r.c

@@ -53,14 +53,14 @@
   }
 
   do {
-    int32x4_t vacc0x01 = vreinterpretq_s32_u64(vshll_n_u32(vld1_u32(w), 0)); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
-    int32x4_t vacc0x23 = vreinterpretq_s32_u64(vshll_n_u32(vld1_u32(w), 0)); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
-    int32x4_t vacc0x45 = vreinterpretq_s32_u64(vshll_n_u32(vld1_u32(w), 0)); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
-    int32x4_t vacc0x67 = vreinterpretq_s32_u64(vshll_n_u32(vld1_u32(w), 0)); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
-    int32x4_t vacc0x89 = vreinterpretq_s32_u64(vshll_n_u32(vld1_u32(w), 0)); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
-    int32x4_t vacc0xAB = vreinterpretq_s32_u64(vshll_n_u32(vld1_u32(w), 0)); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
-    int32x4_t vacc0xCD = vreinterpretq_s32_u64(vshll_n_u32(vld1_u32(w), 0)); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
-    int32x4_t vacc0xEF = vreinterpretq_s32_u64(vshll_n_u32(vld1_u32(w), 0)); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
+    int32x4_t vacc0x01 = vreinterpretq_s32_u64(vmovl_u32(vld1_u32(w))); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
+    int32x4_t vacc0x23 = vreinterpretq_s32_u64(vmovl_u32(vld1_u32(w))); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
+    int32x4_t vacc0x45 = vreinterpretq_s32_u64(vmovl_u32(vld1_u32(w))); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
+    int32x4_t vacc0x67 = vreinterpretq_s32_u64(vmovl_u32(vld1_u32(w))); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
+    int32x4_t vacc0x89 = vreinterpretq_s32_u64(vmovl_u32(vld1_u32(w))); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
+    int32x4_t vacc0xAB = vreinterpretq_s32_u64(vmovl_u32(vld1_u32(w))); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
+    int32x4_t vacc0xCD = vreinterpretq_s32_u64(vmovl_u32(vld1_u32(w))); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
+    int32x4_t vacc0xEF = vreinterpretq_s32_u64(vmovl_u32(vld1_u32(w))); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
     int32x4_t vacc1x01 = vacc0x01;
     int32x4_t vacc1x23 = vacc0x23;
     int32x4_t vacc1x45 = vacc0x45;

diff --git a/src/qs8-gemm/gen/3x16c4-minmax-rndnu-neon-mull-ld2r.c b/src/qs8-gemm/gen/3x16c4-minmax-rndnu-neon-mull-ld2r.c
index d5c1c9f..89bd57f 100644
--- a/src/qs8-gemm/gen/3x16c4-minmax-rndnu-neon-mull-ld2r.c
+++ b/src/qs8-gemm/gen/3x16c4-minmax-rndnu-neon-mull-ld2r.c

@@ -53,14 +53,14 @@
   }
 
   do {
-    int32x4_t vacc0x01 = vreinterpretq_s32_u64(vshll_n_u32(vld1_u32(w), 0)); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
-    int32x4_t vacc0x23 = vreinterpretq_s32_u64(vshll_n_u32(vld1_u32(w), 0)); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
-    int32x4_t vacc0x45 = vreinterpretq_s32_u64(vshll_n_u32(vld1_u32(w), 0)); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
-    int32x4_t vacc0x67 = vreinterpretq_s32_u64(vshll_n_u32(vld1_u32(w), 0)); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
-    int32x4_t vacc0x89 = vreinterpretq_s32_u64(vshll_n_u32(vld1_u32(w), 0)); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
-    int32x4_t vacc0xAB = vreinterpretq_s32_u64(vshll_n_u32(vld1_u32(w), 0)); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
-    int32x4_t vacc0xCD = vreinterpretq_s32_u64(vshll_n_u32(vld1_u32(w), 0)); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
-    int32x4_t vacc0xEF = vreinterpretq_s32_u64(vshll_n_u32(vld1_u32(w), 0)); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
+    int32x4_t vacc0x01 = vreinterpretq_s32_u64(vmovl_u32(vld1_u32(w))); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
+    int32x4_t vacc0x23 = vreinterpretq_s32_u64(vmovl_u32(vld1_u32(w))); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
+    int32x4_t vacc0x45 = vreinterpretq_s32_u64(vmovl_u32(vld1_u32(w))); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
+    int32x4_t vacc0x67 = vreinterpretq_s32_u64(vmovl_u32(vld1_u32(w))); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
+    int32x4_t vacc0x89 = vreinterpretq_s32_u64(vmovl_u32(vld1_u32(w))); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
+    int32x4_t vacc0xAB = vreinterpretq_s32_u64(vmovl_u32(vld1_u32(w))); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
+    int32x4_t vacc0xCD = vreinterpretq_s32_u64(vmovl_u32(vld1_u32(w))); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
+    int32x4_t vacc0xEF = vreinterpretq_s32_u64(vmovl_u32(vld1_u32(w))); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
     int32x4_t vacc1x01 = vacc0x01;
     int32x4_t vacc1x23 = vacc0x23;
     int32x4_t vacc1x45 = vacc0x45;

diff --git a/src/qs8-gemm/gen/3x16c4s2-minmax-rndnu-neon-mlal.c b/src/qs8-gemm/gen/3x16c4s2-minmax-rndnu-neon-mlal.c
index f9a5b62..e87a583 100644
--- a/src/qs8-gemm/gen/3x16c4s2-minmax-rndnu-neon-mlal.c
+++ b/src/qs8-gemm/gen/3x16c4s2-minmax-rndnu-neon-mlal.c

@@ -52,14 +52,14 @@
   }
 
   do {
-    int32x4_t vacc0x01 = vreinterpretq_s32_u64(vshll_n_u32(vld1_u32(w), 0)); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
-    int32x4_t vacc0x23 = vreinterpretq_s32_u64(vshll_n_u32(vld1_u32(w), 0)); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
-    int32x4_t vacc0x45 = vreinterpretq_s32_u64(vshll_n_u32(vld1_u32(w), 0)); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
-    int32x4_t vacc0x67 = vreinterpretq_s32_u64(vshll_n_u32(vld1_u32(w), 0)); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
-    int32x4_t vacc0x89 = vreinterpretq_s32_u64(vshll_n_u32(vld1_u32(w), 0)); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
-    int32x4_t vacc0xAB = vreinterpretq_s32_u64(vshll_n_u32(vld1_u32(w), 0)); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
-    int32x4_t vacc0xCD = vreinterpretq_s32_u64(vshll_n_u32(vld1_u32(w), 0)); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
-    int32x4_t vacc0xEF = vreinterpretq_s32_u64(vshll_n_u32(vld1_u32(w), 0)); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
+    int32x4_t vacc0x01 = vreinterpretq_s32_u64(vmovl_u32(vld1_u32(w))); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
+    int32x4_t vacc0x23 = vreinterpretq_s32_u64(vmovl_u32(vld1_u32(w))); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
+    int32x4_t vacc0x45 = vreinterpretq_s32_u64(vmovl_u32(vld1_u32(w))); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
+    int32x4_t vacc0x67 = vreinterpretq_s32_u64(vmovl_u32(vld1_u32(w))); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
+    int32x4_t vacc0x89 = vreinterpretq_s32_u64(vmovl_u32(vld1_u32(w))); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
+    int32x4_t vacc0xAB = vreinterpretq_s32_u64(vmovl_u32(vld1_u32(w))); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
+    int32x4_t vacc0xCD = vreinterpretq_s32_u64(vmovl_u32(vld1_u32(w))); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
+    int32x4_t vacc0xEF = vreinterpretq_s32_u64(vmovl_u32(vld1_u32(w))); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
     int32x4_t vacc1x01 = vacc0x01;
     int32x4_t vacc1x23 = vacc0x23;
     int32x4_t vacc1x45 = vacc0x45;

diff --git a/src/qs8-gemm/gen/3x16c4s2-minmax-rndnu-neon-mull.c b/src/qs8-gemm/gen/3x16c4s2-minmax-rndnu-neon-mull.c
index c16794c..e432bb9 100644
--- a/src/qs8-gemm/gen/3x16c4s2-minmax-rndnu-neon-mull.c
+++ b/src/qs8-gemm/gen/3x16c4s2-minmax-rndnu-neon-mull.c

@@ -52,14 +52,14 @@
   }
 
   do {
-    int32x4_t vacc0x01 = vreinterpretq_s32_u64(vshll_n_u32(vld1_u32(w), 0)); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
-    int32x4_t vacc0x23 = vreinterpretq_s32_u64(vshll_n_u32(vld1_u32(w), 0)); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
-    int32x4_t vacc0x45 = vreinterpretq_s32_u64(vshll_n_u32(vld1_u32(w), 0)); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
-    int32x4_t vacc0x67 = vreinterpretq_s32_u64(vshll_n_u32(vld1_u32(w), 0)); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
-    int32x4_t vacc0x89 = vreinterpretq_s32_u64(vshll_n_u32(vld1_u32(w), 0)); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
-    int32x4_t vacc0xAB = vreinterpretq_s32_u64(vshll_n_u32(vld1_u32(w), 0)); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
-    int32x4_t vacc0xCD = vreinterpretq_s32_u64(vshll_n_u32(vld1_u32(w), 0)); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
-    int32x4_t vacc0xEF = vreinterpretq_s32_u64(vshll_n_u32(vld1_u32(w), 0)); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
+    int32x4_t vacc0x01 = vreinterpretq_s32_u64(vmovl_u32(vld1_u32(w))); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
+    int32x4_t vacc0x23 = vreinterpretq_s32_u64(vmovl_u32(vld1_u32(w))); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
+    int32x4_t vacc0x45 = vreinterpretq_s32_u64(vmovl_u32(vld1_u32(w))); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
+    int32x4_t vacc0x67 = vreinterpretq_s32_u64(vmovl_u32(vld1_u32(w))); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
+    int32x4_t vacc0x89 = vreinterpretq_s32_u64(vmovl_u32(vld1_u32(w))); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
+    int32x4_t vacc0xAB = vreinterpretq_s32_u64(vmovl_u32(vld1_u32(w))); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
+    int32x4_t vacc0xCD = vreinterpretq_s32_u64(vmovl_u32(vld1_u32(w))); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
+    int32x4_t vacc0xEF = vreinterpretq_s32_u64(vmovl_u32(vld1_u32(w))); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
     int32x4_t vacc1x01 = vacc0x01;
     int32x4_t vacc1x23 = vacc0x23;
     int32x4_t vacc1x45 = vacc0x45;

diff --git a/src/qs8-gemm/gen/3x8c4-minmax-rndnu-neon-mlal-dup.c b/src/qs8-gemm/gen/3x8c4-minmax-rndnu-neon-mlal-dup.c
index 9203975..c4d571d 100644
--- a/src/qs8-gemm/gen/3x8c4-minmax-rndnu-neon-mlal-dup.c
+++ b/src/qs8-gemm/gen/3x8c4-minmax-rndnu-neon-mlal-dup.c

@@ -53,10 +53,10 @@
   }
 
   do {
-    int32x4_t vacc0x01 = vreinterpretq_s32_u64(vshll_n_u32(vld1_u32(w), 0)); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
-    int32x4_t vacc0x23 = vreinterpretq_s32_u64(vshll_n_u32(vld1_u32(w), 0)); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
-    int32x4_t vacc0x45 = vreinterpretq_s32_u64(vshll_n_u32(vld1_u32(w), 0)); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
-    int32x4_t vacc0x67 = vreinterpretq_s32_u64(vshll_n_u32(vld1_u32(w), 0)); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
+    int32x4_t vacc0x01 = vreinterpretq_s32_u64(vmovl_u32(vld1_u32(w))); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
+    int32x4_t vacc0x23 = vreinterpretq_s32_u64(vmovl_u32(vld1_u32(w))); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
+    int32x4_t vacc0x45 = vreinterpretq_s32_u64(vmovl_u32(vld1_u32(w))); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
+    int32x4_t vacc0x67 = vreinterpretq_s32_u64(vmovl_u32(vld1_u32(w))); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
     int32x4_t vacc1x01 = vacc0x01;
     int32x4_t vacc1x23 = vacc0x23;
     int32x4_t vacc1x45 = vacc0x45;

diff --git a/src/qs8-gemm/gen/3x8c4-minmax-rndnu-neon-mlal-ld1r.c b/src/qs8-gemm/gen/3x8c4-minmax-rndnu-neon-mlal-ld1r.c
index 9ef8562..56eda78 100644
--- a/src/qs8-gemm/gen/3x8c4-minmax-rndnu-neon-mlal-ld1r.c
+++ b/src/qs8-gemm/gen/3x8c4-minmax-rndnu-neon-mlal-ld1r.c

@@ -53,10 +53,10 @@
   }
 
   do {
-    int32x4_t vacc0x01 = vreinterpretq_s32_u64(vshll_n_u32(vld1_u32(w), 0)); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
-    int32x4_t vacc0x23 = vreinterpretq_s32_u64(vshll_n_u32(vld1_u32(w), 0)); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
-    int32x4_t vacc0x45 = vreinterpretq_s32_u64(vshll_n_u32(vld1_u32(w), 0)); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
-    int32x4_t vacc0x67 = vreinterpretq_s32_u64(vshll_n_u32(vld1_u32(w), 0)); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
+    int32x4_t vacc0x01 = vreinterpretq_s32_u64(vmovl_u32(vld1_u32(w))); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
+    int32x4_t vacc0x23 = vreinterpretq_s32_u64(vmovl_u32(vld1_u32(w))); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
+    int32x4_t vacc0x45 = vreinterpretq_s32_u64(vmovl_u32(vld1_u32(w))); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
+    int32x4_t vacc0x67 = vreinterpretq_s32_u64(vmovl_u32(vld1_u32(w))); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
     int32x4_t vacc1x01 = vacc0x01;
     int32x4_t vacc1x23 = vacc0x23;
     int32x4_t vacc1x45 = vacc0x45;

diff --git a/src/qs8-gemm/gen/3x8c4-minmax-rndnu-neon-mlal-ld2r.c b/src/qs8-gemm/gen/3x8c4-minmax-rndnu-neon-mlal-ld2r.c
index 2ef6c97..7eb1257 100644
--- a/src/qs8-gemm/gen/3x8c4-minmax-rndnu-neon-mlal-ld2r.c
+++ b/src/qs8-gemm/gen/3x8c4-minmax-rndnu-neon-mlal-ld2r.c

@@ -53,10 +53,10 @@
   }
 
   do {
-    int32x4_t vacc0x01 = vreinterpretq_s32_u64(vshll_n_u32(vld1_u32(w), 0)); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
-    int32x4_t vacc0x23 = vreinterpretq_s32_u64(vshll_n_u32(vld1_u32(w), 0)); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
-    int32x4_t vacc0x45 = vreinterpretq_s32_u64(vshll_n_u32(vld1_u32(w), 0)); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
-    int32x4_t vacc0x67 = vreinterpretq_s32_u64(vshll_n_u32(vld1_u32(w), 0)); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
+    int32x4_t vacc0x01 = vreinterpretq_s32_u64(vmovl_u32(vld1_u32(w))); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
+    int32x4_t vacc0x23 = vreinterpretq_s32_u64(vmovl_u32(vld1_u32(w))); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
+    int32x4_t vacc0x45 = vreinterpretq_s32_u64(vmovl_u32(vld1_u32(w))); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
+    int32x4_t vacc0x67 = vreinterpretq_s32_u64(vmovl_u32(vld1_u32(w))); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
     int32x4_t vacc1x01 = vacc0x01;
     int32x4_t vacc1x23 = vacc0x23;
     int32x4_t vacc1x45 = vacc0x45;

diff --git a/src/qs8-gemm/gen/3x8c4-minmax-rndnu-neon-mull-dup.c b/src/qs8-gemm/gen/3x8c4-minmax-rndnu-neon-mull-dup.c
index 10b0197..0915fb8 100644
--- a/src/qs8-gemm/gen/3x8c4-minmax-rndnu-neon-mull-dup.c
+++ b/src/qs8-gemm/gen/3x8c4-minmax-rndnu-neon-mull-dup.c

@@ -53,10 +53,10 @@
   }
 
   do {
-    int32x4_t vacc0x01 = vreinterpretq_s32_u64(vshll_n_u32(vld1_u32(w), 0)); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
-    int32x4_t vacc0x23 = vreinterpretq_s32_u64(vshll_n_u32(vld1_u32(w), 0)); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
-    int32x4_t vacc0x45 = vreinterpretq_s32_u64(vshll_n_u32(vld1_u32(w), 0)); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
-    int32x4_t vacc0x67 = vreinterpretq_s32_u64(vshll_n_u32(vld1_u32(w), 0)); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
+    int32x4_t vacc0x01 = vreinterpretq_s32_u64(vmovl_u32(vld1_u32(w))); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
+    int32x4_t vacc0x23 = vreinterpretq_s32_u64(vmovl_u32(vld1_u32(w))); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
+    int32x4_t vacc0x45 = vreinterpretq_s32_u64(vmovl_u32(vld1_u32(w))); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
+    int32x4_t vacc0x67 = vreinterpretq_s32_u64(vmovl_u32(vld1_u32(w))); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
     int32x4_t vacc1x01 = vacc0x01;
     int32x4_t vacc1x23 = vacc0x23;
     int32x4_t vacc1x45 = vacc0x45;

diff --git a/src/qs8-gemm/gen/3x8c4-minmax-rndnu-neon-mull-ld1r.c b/src/qs8-gemm/gen/3x8c4-minmax-rndnu-neon-mull-ld1r.c
index c0a04e1..3ee3759 100644
--- a/src/qs8-gemm/gen/3x8c4-minmax-rndnu-neon-mull-ld1r.c
+++ b/src/qs8-gemm/gen/3x8c4-minmax-rndnu-neon-mull-ld1r.c

@@ -53,10 +53,10 @@
   }
 
   do {
-    int32x4_t vacc0x01 = vreinterpretq_s32_u64(vshll_n_u32(vld1_u32(w), 0)); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
-    int32x4_t vacc0x23 = vreinterpretq_s32_u64(vshll_n_u32(vld1_u32(w), 0)); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
-    int32x4_t vacc0x45 = vreinterpretq_s32_u64(vshll_n_u32(vld1_u32(w), 0)); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
-    int32x4_t vacc0x67 = vreinterpretq_s32_u64(vshll_n_u32(vld1_u32(w), 0)); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
+    int32x4_t vacc0x01 = vreinterpretq_s32_u64(vmovl_u32(vld1_u32(w))); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
+    int32x4_t vacc0x23 = vreinterpretq_s32_u64(vmovl_u32(vld1_u32(w))); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
+    int32x4_t vacc0x45 = vreinterpretq_s32_u64(vmovl_u32(vld1_u32(w))); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
+    int32x4_t vacc0x67 = vreinterpretq_s32_u64(vmovl_u32(vld1_u32(w))); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
     int32x4_t vacc1x01 = vacc0x01;
     int32x4_t vacc1x23 = vacc0x23;
     int32x4_t vacc1x45 = vacc0x45;

diff --git a/src/qs8-gemm/gen/3x8c4-minmax-rndnu-neon-mull-ld2r.c b/src/qs8-gemm/gen/3x8c4-minmax-rndnu-neon-mull-ld2r.c
index 02ac4d9..52a39d2 100644
--- a/src/qs8-gemm/gen/3x8c4-minmax-rndnu-neon-mull-ld2r.c
+++ b/src/qs8-gemm/gen/3x8c4-minmax-rndnu-neon-mull-ld2r.c

@@ -53,10 +53,10 @@
   }
 
   do {
-    int32x4_t vacc0x01 = vreinterpretq_s32_u64(vshll_n_u32(vld1_u32(w), 0)); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
-    int32x4_t vacc0x23 = vreinterpretq_s32_u64(vshll_n_u32(vld1_u32(w), 0)); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
-    int32x4_t vacc0x45 = vreinterpretq_s32_u64(vshll_n_u32(vld1_u32(w), 0)); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
-    int32x4_t vacc0x67 = vreinterpretq_s32_u64(vshll_n_u32(vld1_u32(w), 0)); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
+    int32x4_t vacc0x01 = vreinterpretq_s32_u64(vmovl_u32(vld1_u32(w))); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
+    int32x4_t vacc0x23 = vreinterpretq_s32_u64(vmovl_u32(vld1_u32(w))); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
+    int32x4_t vacc0x45 = vreinterpretq_s32_u64(vmovl_u32(vld1_u32(w))); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
+    int32x4_t vacc0x67 = vreinterpretq_s32_u64(vmovl_u32(vld1_u32(w))); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
     int32x4_t vacc1x01 = vacc0x01;
     int32x4_t vacc1x23 = vacc0x23;
     int32x4_t vacc1x45 = vacc0x45;

diff --git a/src/qs8-gemm/gen/3x8c4s2-minmax-rndnu-neon-mlal.c b/src/qs8-gemm/gen/3x8c4s2-minmax-rndnu-neon-mlal.c
index b42f7c8..4279c2d 100644
--- a/src/qs8-gemm/gen/3x8c4s2-minmax-rndnu-neon-mlal.c
+++ b/src/qs8-gemm/gen/3x8c4s2-minmax-rndnu-neon-mlal.c

@@ -52,10 +52,10 @@
   }
 
   do {
-    int32x4_t vacc0x01 = vreinterpretq_s32_u64(vshll_n_u32(vld1_u32(w), 0)); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
-    int32x4_t vacc0x23 = vreinterpretq_s32_u64(vshll_n_u32(vld1_u32(w), 0)); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
-    int32x4_t vacc0x45 = vreinterpretq_s32_u64(vshll_n_u32(vld1_u32(w), 0)); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
-    int32x4_t vacc0x67 = vreinterpretq_s32_u64(vshll_n_u32(vld1_u32(w), 0)); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
+    int32x4_t vacc0x01 = vreinterpretq_s32_u64(vmovl_u32(vld1_u32(w))); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
+    int32x4_t vacc0x23 = vreinterpretq_s32_u64(vmovl_u32(vld1_u32(w))); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
+    int32x4_t vacc0x45 = vreinterpretq_s32_u64(vmovl_u32(vld1_u32(w))); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
+    int32x4_t vacc0x67 = vreinterpretq_s32_u64(vmovl_u32(vld1_u32(w))); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
     int32x4_t vacc1x01 = vacc0x01;
     int32x4_t vacc1x23 = vacc0x23;
     int32x4_t vacc1x45 = vacc0x45;

diff --git a/src/qs8-gemm/gen/3x8c4s2-minmax-rndnu-neon-mull.c b/src/qs8-gemm/gen/3x8c4s2-minmax-rndnu-neon-mull.c
index 443d40b..f4100b7 100644
--- a/src/qs8-gemm/gen/3x8c4s2-minmax-rndnu-neon-mull.c
+++ b/src/qs8-gemm/gen/3x8c4s2-minmax-rndnu-neon-mull.c

@@ -52,10 +52,10 @@
   }
 
   do {
-    int32x4_t vacc0x01 = vreinterpretq_s32_u64(vshll_n_u32(vld1_u32(w), 0)); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
-    int32x4_t vacc0x23 = vreinterpretq_s32_u64(vshll_n_u32(vld1_u32(w), 0)); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
-    int32x4_t vacc0x45 = vreinterpretq_s32_u64(vshll_n_u32(vld1_u32(w), 0)); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
-    int32x4_t vacc0x67 = vreinterpretq_s32_u64(vshll_n_u32(vld1_u32(w), 0)); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
+    int32x4_t vacc0x01 = vreinterpretq_s32_u64(vmovl_u32(vld1_u32(w))); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
+    int32x4_t vacc0x23 = vreinterpretq_s32_u64(vmovl_u32(vld1_u32(w))); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
+    int32x4_t vacc0x45 = vreinterpretq_s32_u64(vmovl_u32(vld1_u32(w))); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
+    int32x4_t vacc0x67 = vreinterpretq_s32_u64(vmovl_u32(vld1_u32(w))); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
     int32x4_t vacc1x01 = vacc0x01;
     int32x4_t vacc1x23 = vacc0x23;
     int32x4_t vacc1x45 = vacc0x45;

diff --git a/src/qs8-gemm/gen/4x16c4-minmax-rndnu-neon-mlal-dup.c b/src/qs8-gemm/gen/4x16c4-minmax-rndnu-neon-mlal-dup.c
index f8beaf3..1ca6b28 100644
--- a/src/qs8-gemm/gen/4x16c4-minmax-rndnu-neon-mlal-dup.c
+++ b/src/qs8-gemm/gen/4x16c4-minmax-rndnu-neon-mlal-dup.c

@@ -59,14 +59,14 @@
   }
 
   do {
-    int32x4_t vacc0x01 = vreinterpretq_s32_u64(vshll_n_u32(vld1_u32(w), 0)); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
-    int32x4_t vacc0x23 = vreinterpretq_s32_u64(vshll_n_u32(vld1_u32(w), 0)); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
-    int32x4_t vacc0x45 = vreinterpretq_s32_u64(vshll_n_u32(vld1_u32(w), 0)); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
-    int32x4_t vacc0x67 = vreinterpretq_s32_u64(vshll_n_u32(vld1_u32(w), 0)); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
-    int32x4_t vacc0x89 = vreinterpretq_s32_u64(vshll_n_u32(vld1_u32(w), 0)); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
-    int32x4_t vacc0xAB = vreinterpretq_s32_u64(vshll_n_u32(vld1_u32(w), 0)); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
-    int32x4_t vacc0xCD = vreinterpretq_s32_u64(vshll_n_u32(vld1_u32(w), 0)); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
-    int32x4_t vacc0xEF = vreinterpretq_s32_u64(vshll_n_u32(vld1_u32(w), 0)); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
+    int32x4_t vacc0x01 = vreinterpretq_s32_u64(vmovl_u32(vld1_u32(w))); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
+    int32x4_t vacc0x23 = vreinterpretq_s32_u64(vmovl_u32(vld1_u32(w))); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
+    int32x4_t vacc0x45 = vreinterpretq_s32_u64(vmovl_u32(vld1_u32(w))); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
+    int32x4_t vacc0x67 = vreinterpretq_s32_u64(vmovl_u32(vld1_u32(w))); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
+    int32x4_t vacc0x89 = vreinterpretq_s32_u64(vmovl_u32(vld1_u32(w))); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
+    int32x4_t vacc0xAB = vreinterpretq_s32_u64(vmovl_u32(vld1_u32(w))); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
+    int32x4_t vacc0xCD = vreinterpretq_s32_u64(vmovl_u32(vld1_u32(w))); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
+    int32x4_t vacc0xEF = vreinterpretq_s32_u64(vmovl_u32(vld1_u32(w))); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
     int32x4_t vacc1x01 = vacc0x01;
     int32x4_t vacc1x23 = vacc0x23;
     int32x4_t vacc1x45 = vacc0x45;

diff --git a/src/qs8-gemm/gen/4x16c4-minmax-rndnu-neon-mlal-ld1r.c b/src/qs8-gemm/gen/4x16c4-minmax-rndnu-neon-mlal-ld1r.c
index 41e2f21..193d94b 100644
--- a/src/qs8-gemm/gen/4x16c4-minmax-rndnu-neon-mlal-ld1r.c
+++ b/src/qs8-gemm/gen/4x16c4-minmax-rndnu-neon-mlal-ld1r.c

@@ -59,14 +59,14 @@
   }
 
   do {
-    int32x4_t vacc0x01 = vreinterpretq_s32_u64(vshll_n_u32(vld1_u32(w), 0)); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
-    int32x4_t vacc0x23 = vreinterpretq_s32_u64(vshll_n_u32(vld1_u32(w), 0)); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
-    int32x4_t vacc0x45 = vreinterpretq_s32_u64(vshll_n_u32(vld1_u32(w), 0)); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
-    int32x4_t vacc0x67 = vreinterpretq_s32_u64(vshll_n_u32(vld1_u32(w), 0)); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
-    int32x4_t vacc0x89 = vreinterpretq_s32_u64(vshll_n_u32(vld1_u32(w), 0)); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
-    int32x4_t vacc0xAB = vreinterpretq_s32_u64(vshll_n_u32(vld1_u32(w), 0)); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
-    int32x4_t vacc0xCD = vreinterpretq_s32_u64(vshll_n_u32(vld1_u32(w), 0)); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
-    int32x4_t vacc0xEF = vreinterpretq_s32_u64(vshll_n_u32(vld1_u32(w), 0)); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
+    int32x4_t vacc0x01 = vreinterpretq_s32_u64(vmovl_u32(vld1_u32(w))); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
+    int32x4_t vacc0x23 = vreinterpretq_s32_u64(vmovl_u32(vld1_u32(w))); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
+    int32x4_t vacc0x45 = vreinterpretq_s32_u64(vmovl_u32(vld1_u32(w))); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
+    int32x4_t vacc0x67 = vreinterpretq_s32_u64(vmovl_u32(vld1_u32(w))); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
+    int32x4_t vacc0x89 = vreinterpretq_s32_u64(vmovl_u32(vld1_u32(w))); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
+    int32x4_t vacc0xAB = vreinterpretq_s32_u64(vmovl_u32(vld1_u32(w))); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
+    int32x4_t vacc0xCD = vreinterpretq_s32_u64(vmovl_u32(vld1_u32(w))); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
+    int32x4_t vacc0xEF = vreinterpretq_s32_u64(vmovl_u32(vld1_u32(w))); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
     int32x4_t vacc1x01 = vacc0x01;
     int32x4_t vacc1x23 = vacc0x23;
     int32x4_t vacc1x45 = vacc0x45;

diff --git a/src/qs8-gemm/gen/4x16c4-minmax-rndnu-neon-mlal-ld2r.c b/src/qs8-gemm/gen/4x16c4-minmax-rndnu-neon-mlal-ld2r.c
index 9297091..58c01da 100644
--- a/src/qs8-gemm/gen/4x16c4-minmax-rndnu-neon-mlal-ld2r.c
+++ b/src/qs8-gemm/gen/4x16c4-minmax-rndnu-neon-mlal-ld2r.c

@@ -59,14 +59,14 @@
   }
 
   do {
-    int32x4_t vacc0x01 = vreinterpretq_s32_u64(vshll_n_u32(vld1_u32(w), 0)); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
-    int32x4_t vacc0x23 = vreinterpretq_s32_u64(vshll_n_u32(vld1_u32(w), 0)); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
-    int32x4_t vacc0x45 = vreinterpretq_s32_u64(vshll_n_u32(vld1_u32(w), 0)); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
-    int32x4_t vacc0x67 = vreinterpretq_s32_u64(vshll_n_u32(vld1_u32(w), 0)); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
-    int32x4_t vacc0x89 = vreinterpretq_s32_u64(vshll_n_u32(vld1_u32(w), 0)); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
-    int32x4_t vacc0xAB = vreinterpretq_s32_u64(vshll_n_u32(vld1_u32(w), 0)); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
-    int32x4_t vacc0xCD = vreinterpretq_s32_u64(vshll_n_u32(vld1_u32(w), 0)); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
-    int32x4_t vacc0xEF = vreinterpretq_s32_u64(vshll_n_u32(vld1_u32(w), 0)); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
+    int32x4_t vacc0x01 = vreinterpretq_s32_u64(vmovl_u32(vld1_u32(w))); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
+    int32x4_t vacc0x23 = vreinterpretq_s32_u64(vmovl_u32(vld1_u32(w))); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
+    int32x4_t vacc0x45 = vreinterpretq_s32_u64(vmovl_u32(vld1_u32(w))); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
+    int32x4_t vacc0x67 = vreinterpretq_s32_u64(vmovl_u32(vld1_u32(w))); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
+    int32x4_t vacc0x89 = vreinterpretq_s32_u64(vmovl_u32(vld1_u32(w))); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
+    int32x4_t vacc0xAB = vreinterpretq_s32_u64(vmovl_u32(vld1_u32(w))); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
+    int32x4_t vacc0xCD = vreinterpretq_s32_u64(vmovl_u32(vld1_u32(w))); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
+    int32x4_t vacc0xEF = vreinterpretq_s32_u64(vmovl_u32(vld1_u32(w))); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
     int32x4_t vacc1x01 = vacc0x01;
     int32x4_t vacc1x23 = vacc0x23;
     int32x4_t vacc1x45 = vacc0x45;

diff --git a/src/qs8-gemm/gen/4x16c4-minmax-rndnu-neon-mull-dup.c b/src/qs8-gemm/gen/4x16c4-minmax-rndnu-neon-mull-dup.c
index 171f7cf..c87cbe6 100644
--- a/src/qs8-gemm/gen/4x16c4-minmax-rndnu-neon-mull-dup.c
+++ b/src/qs8-gemm/gen/4x16c4-minmax-rndnu-neon-mull-dup.c

@@ -59,14 +59,14 @@
   }
 
   do {
-    int32x4_t vacc0x01 = vreinterpretq_s32_u64(vshll_n_u32(vld1_u32(w), 0)); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
-    int32x4_t vacc0x23 = vreinterpretq_s32_u64(vshll_n_u32(vld1_u32(w), 0)); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
-    int32x4_t vacc0x45 = vreinterpretq_s32_u64(vshll_n_u32(vld1_u32(w), 0)); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
-    int32x4_t vacc0x67 = vreinterpretq_s32_u64(vshll_n_u32(vld1_u32(w), 0)); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
-    int32x4_t vacc0x89 = vreinterpretq_s32_u64(vshll_n_u32(vld1_u32(w), 0)); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
-    int32x4_t vacc0xAB = vreinterpretq_s32_u64(vshll_n_u32(vld1_u32(w), 0)); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
-    int32x4_t vacc0xCD = vreinterpretq_s32_u64(vshll_n_u32(vld1_u32(w), 0)); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
-    int32x4_t vacc0xEF = vreinterpretq_s32_u64(vshll_n_u32(vld1_u32(w), 0)); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
+    int32x4_t vacc0x01 = vreinterpretq_s32_u64(vmovl_u32(vld1_u32(w))); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
+    int32x4_t vacc0x23 = vreinterpretq_s32_u64(vmovl_u32(vld1_u32(w))); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
+    int32x4_t vacc0x45 = vreinterpretq_s32_u64(vmovl_u32(vld1_u32(w))); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
+    int32x4_t vacc0x67 = vreinterpretq_s32_u64(vmovl_u32(vld1_u32(w))); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
+    int32x4_t vacc0x89 = vreinterpretq_s32_u64(vmovl_u32(vld1_u32(w))); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
+    int32x4_t vacc0xAB = vreinterpretq_s32_u64(vmovl_u32(vld1_u32(w))); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
+    int32x4_t vacc0xCD = vreinterpretq_s32_u64(vmovl_u32(vld1_u32(w))); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
+    int32x4_t vacc0xEF = vreinterpretq_s32_u64(vmovl_u32(vld1_u32(w))); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
     int32x4_t vacc1x01 = vacc0x01;
     int32x4_t vacc1x23 = vacc0x23;
     int32x4_t vacc1x45 = vacc0x45;

diff --git a/src/qs8-gemm/gen/4x16c4-minmax-rndnu-neon-mull-ld1r.c b/src/qs8-gemm/gen/4x16c4-minmax-rndnu-neon-mull-ld1r.c
index f257fed..dbb62e5 100644
--- a/src/qs8-gemm/gen/4x16c4-minmax-rndnu-neon-mull-ld1r.c
+++ b/src/qs8-gemm/gen/4x16c4-minmax-rndnu-neon-mull-ld1r.c

@@ -59,14 +59,14 @@
   }
 
   do {
-    int32x4_t vacc0x01 = vreinterpretq_s32_u64(vshll_n_u32(vld1_u32(w), 0)); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
-    int32x4_t vacc0x23 = vreinterpretq_s32_u64(vshll_n_u32(vld1_u32(w), 0)); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
-    int32x4_t vacc0x45 = vreinterpretq_s32_u64(vshll_n_u32(vld1_u32(w), 0)); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
-    int32x4_t vacc0x67 = vreinterpretq_s32_u64(vshll_n_u32(vld1_u32(w), 0)); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
-    int32x4_t vacc0x89 = vreinterpretq_s32_u64(vshll_n_u32(vld1_u32(w), 0)); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
-    int32x4_t vacc0xAB = vreinterpretq_s32_u64(vshll_n_u32(vld1_u32(w), 0)); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
-    int32x4_t vacc0xCD = vreinterpretq_s32_u64(vshll_n_u32(vld1_u32(w), 0)); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
-    int32x4_t vacc0xEF = vreinterpretq_s32_u64(vshll_n_u32(vld1_u32(w), 0)); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
+    int32x4_t vacc0x01 = vreinterpretq_s32_u64(vmovl_u32(vld1_u32(w))); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
+    int32x4_t vacc0x23 = vreinterpretq_s32_u64(vmovl_u32(vld1_u32(w))); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
+    int32x4_t vacc0x45 = vreinterpretq_s32_u64(vmovl_u32(vld1_u32(w))); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
+    int32x4_t vacc0x67 = vreinterpretq_s32_u64(vmovl_u32(vld1_u32(w))); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
+    int32x4_t vacc0x89 = vreinterpretq_s32_u64(vmovl_u32(vld1_u32(w))); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
+    int32x4_t vacc0xAB = vreinterpretq_s32_u64(vmovl_u32(vld1_u32(w))); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
+    int32x4_t vacc0xCD = vreinterpretq_s32_u64(vmovl_u32(vld1_u32(w))); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
+    int32x4_t vacc0xEF = vreinterpretq_s32_u64(vmovl_u32(vld1_u32(w))); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
     int32x4_t vacc1x01 = vacc0x01;
     int32x4_t vacc1x23 = vacc0x23;
     int32x4_t vacc1x45 = vacc0x45;

diff --git a/src/qs8-gemm/gen/4x16c4-minmax-rndnu-neon-mull-ld2r.c b/src/qs8-gemm/gen/4x16c4-minmax-rndnu-neon-mull-ld2r.c
index 8f8d209..74eb882 100644
--- a/src/qs8-gemm/gen/4x16c4-minmax-rndnu-neon-mull-ld2r.c
+++ b/src/qs8-gemm/gen/4x16c4-minmax-rndnu-neon-mull-ld2r.c

@@ -59,14 +59,14 @@
   }
 
   do {
-    int32x4_t vacc0x01 = vreinterpretq_s32_u64(vshll_n_u32(vld1_u32(w), 0)); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
-    int32x4_t vacc0x23 = vreinterpretq_s32_u64(vshll_n_u32(vld1_u32(w), 0)); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
-    int32x4_t vacc0x45 = vreinterpretq_s32_u64(vshll_n_u32(vld1_u32(w), 0)); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
-    int32x4_t vacc0x67 = vreinterpretq_s32_u64(vshll_n_u32(vld1_u32(w), 0)); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
-    int32x4_t vacc0x89 = vreinterpretq_s32_u64(vshll_n_u32(vld1_u32(w), 0)); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
-    int32x4_t vacc0xAB = vreinterpretq_s32_u64(vshll_n_u32(vld1_u32(w), 0)); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
-    int32x4_t vacc0xCD = vreinterpretq_s32_u64(vshll_n_u32(vld1_u32(w), 0)); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
-    int32x4_t vacc0xEF = vreinterpretq_s32_u64(vshll_n_u32(vld1_u32(w), 0)); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
+    int32x4_t vacc0x01 = vreinterpretq_s32_u64(vmovl_u32(vld1_u32(w))); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
+    int32x4_t vacc0x23 = vreinterpretq_s32_u64(vmovl_u32(vld1_u32(w))); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
+    int32x4_t vacc0x45 = vreinterpretq_s32_u64(vmovl_u32(vld1_u32(w))); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
+    int32x4_t vacc0x67 = vreinterpretq_s32_u64(vmovl_u32(vld1_u32(w))); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
+    int32x4_t vacc0x89 = vreinterpretq_s32_u64(vmovl_u32(vld1_u32(w))); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
+    int32x4_t vacc0xAB = vreinterpretq_s32_u64(vmovl_u32(vld1_u32(w))); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
+    int32x4_t vacc0xCD = vreinterpretq_s32_u64(vmovl_u32(vld1_u32(w))); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
+    int32x4_t vacc0xEF = vreinterpretq_s32_u64(vmovl_u32(vld1_u32(w))); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
     int32x4_t vacc1x01 = vacc0x01;
     int32x4_t vacc1x23 = vacc0x23;
     int32x4_t vacc1x45 = vacc0x45;

diff --git a/src/qs8-gemm/gen/4x16c4s2-minmax-rndnu-neon-mlal.c b/src/qs8-gemm/gen/4x16c4s2-minmax-rndnu-neon-mlal.c
index 4b97dc2..0d48eb6 100644
--- a/src/qs8-gemm/gen/4x16c4s2-minmax-rndnu-neon-mlal.c
+++ b/src/qs8-gemm/gen/4x16c4s2-minmax-rndnu-neon-mlal.c

@@ -58,14 +58,14 @@
   }
 
   do {
-    int32x4_t vacc0x01 = vreinterpretq_s32_u64(vshll_n_u32(vld1_u32(w), 0)); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
-    int32x4_t vacc0x23 = vreinterpretq_s32_u64(vshll_n_u32(vld1_u32(w), 0)); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
-    int32x4_t vacc0x45 = vreinterpretq_s32_u64(vshll_n_u32(vld1_u32(w), 0)); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
-    int32x4_t vacc0x67 = vreinterpretq_s32_u64(vshll_n_u32(vld1_u32(w), 0)); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
-    int32x4_t vacc0x89 = vreinterpretq_s32_u64(vshll_n_u32(vld1_u32(w), 0)); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
-    int32x4_t vacc0xAB = vreinterpretq_s32_u64(vshll_n_u32(vld1_u32(w), 0)); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
-    int32x4_t vacc0xCD = vreinterpretq_s32_u64(vshll_n_u32(vld1_u32(w), 0)); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
-    int32x4_t vacc0xEF = vreinterpretq_s32_u64(vshll_n_u32(vld1_u32(w), 0)); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
+    int32x4_t vacc0x01 = vreinterpretq_s32_u64(vmovl_u32(vld1_u32(w))); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
+    int32x4_t vacc0x23 = vreinterpretq_s32_u64(vmovl_u32(vld1_u32(w))); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
+    int32x4_t vacc0x45 = vreinterpretq_s32_u64(vmovl_u32(vld1_u32(w))); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
+    int32x4_t vacc0x67 = vreinterpretq_s32_u64(vmovl_u32(vld1_u32(w))); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
+    int32x4_t vacc0x89 = vreinterpretq_s32_u64(vmovl_u32(vld1_u32(w))); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
+    int32x4_t vacc0xAB = vreinterpretq_s32_u64(vmovl_u32(vld1_u32(w))); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
+    int32x4_t vacc0xCD = vreinterpretq_s32_u64(vmovl_u32(vld1_u32(w))); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
+    int32x4_t vacc0xEF = vreinterpretq_s32_u64(vmovl_u32(vld1_u32(w))); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
     int32x4_t vacc1x01 = vacc0x01;
     int32x4_t vacc1x23 = vacc0x23;
     int32x4_t vacc1x45 = vacc0x45;

diff --git a/src/qs8-gemm/gen/4x16c4s2-minmax-rndnu-neon-mull.c b/src/qs8-gemm/gen/4x16c4s2-minmax-rndnu-neon-mull.c
index 1255627..541f8f5 100644
--- a/src/qs8-gemm/gen/4x16c4s2-minmax-rndnu-neon-mull.c
+++ b/src/qs8-gemm/gen/4x16c4s2-minmax-rndnu-neon-mull.c

@@ -58,14 +58,14 @@
   }
 
   do {
-    int32x4_t vacc0x01 = vreinterpretq_s32_u64(vshll_n_u32(vld1_u32(w), 0)); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
-    int32x4_t vacc0x23 = vreinterpretq_s32_u64(vshll_n_u32(vld1_u32(w), 0)); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
-    int32x4_t vacc0x45 = vreinterpretq_s32_u64(vshll_n_u32(vld1_u32(w), 0)); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
-    int32x4_t vacc0x67 = vreinterpretq_s32_u64(vshll_n_u32(vld1_u32(w), 0)); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
-    int32x4_t vacc0x89 = vreinterpretq_s32_u64(vshll_n_u32(vld1_u32(w), 0)); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
-    int32x4_t vacc0xAB = vreinterpretq_s32_u64(vshll_n_u32(vld1_u32(w), 0)); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
-    int32x4_t vacc0xCD = vreinterpretq_s32_u64(vshll_n_u32(vld1_u32(w), 0)); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
-    int32x4_t vacc0xEF = vreinterpretq_s32_u64(vshll_n_u32(vld1_u32(w), 0)); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
+    int32x4_t vacc0x01 = vreinterpretq_s32_u64(vmovl_u32(vld1_u32(w))); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
+    int32x4_t vacc0x23 = vreinterpretq_s32_u64(vmovl_u32(vld1_u32(w))); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
+    int32x4_t vacc0x45 = vreinterpretq_s32_u64(vmovl_u32(vld1_u32(w))); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
+    int32x4_t vacc0x67 = vreinterpretq_s32_u64(vmovl_u32(vld1_u32(w))); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
+    int32x4_t vacc0x89 = vreinterpretq_s32_u64(vmovl_u32(vld1_u32(w))); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
+    int32x4_t vacc0xAB = vreinterpretq_s32_u64(vmovl_u32(vld1_u32(w))); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
+    int32x4_t vacc0xCD = vreinterpretq_s32_u64(vmovl_u32(vld1_u32(w))); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
+    int32x4_t vacc0xEF = vreinterpretq_s32_u64(vmovl_u32(vld1_u32(w))); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
     int32x4_t vacc1x01 = vacc0x01;
     int32x4_t vacc1x23 = vacc0x23;
     int32x4_t vacc1x45 = vacc0x45;

diff --git a/src/qs8-gemm/gen/4x8c4-minmax-rndnu-neon-mlal-dup.c b/src/qs8-gemm/gen/4x8c4-minmax-rndnu-neon-mlal-dup.c
index ac85a24..b697b48 100644
--- a/src/qs8-gemm/gen/4x8c4-minmax-rndnu-neon-mlal-dup.c
+++ b/src/qs8-gemm/gen/4x8c4-minmax-rndnu-neon-mlal-dup.c

@@ -59,10 +59,10 @@
   }
 
   do {
-    int32x4_t vacc0x01 = vreinterpretq_s32_u64(vshll_n_u32(vld1_u32(w), 0)); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
-    int32x4_t vacc0x23 = vreinterpretq_s32_u64(vshll_n_u32(vld1_u32(w), 0)); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
-    int32x4_t vacc0x45 = vreinterpretq_s32_u64(vshll_n_u32(vld1_u32(w), 0)); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
-    int32x4_t vacc0x67 = vreinterpretq_s32_u64(vshll_n_u32(vld1_u32(w), 0)); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
+    int32x4_t vacc0x01 = vreinterpretq_s32_u64(vmovl_u32(vld1_u32(w))); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
+    int32x4_t vacc0x23 = vreinterpretq_s32_u64(vmovl_u32(vld1_u32(w))); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
+    int32x4_t vacc0x45 = vreinterpretq_s32_u64(vmovl_u32(vld1_u32(w))); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
+    int32x4_t vacc0x67 = vreinterpretq_s32_u64(vmovl_u32(vld1_u32(w))); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
     int32x4_t vacc1x01 = vacc0x01;
     int32x4_t vacc1x23 = vacc0x23;
     int32x4_t vacc1x45 = vacc0x45;

diff --git a/src/qs8-gemm/gen/4x8c4-minmax-rndnu-neon-mlal-ld1r.c b/src/qs8-gemm/gen/4x8c4-minmax-rndnu-neon-mlal-ld1r.c
index f12525b..afbc3e8 100644
--- a/src/qs8-gemm/gen/4x8c4-minmax-rndnu-neon-mlal-ld1r.c
+++ b/src/qs8-gemm/gen/4x8c4-minmax-rndnu-neon-mlal-ld1r.c

@@ -59,10 +59,10 @@
   }
 
   do {
-    int32x4_t vacc0x01 = vreinterpretq_s32_u64(vshll_n_u32(vld1_u32(w), 0)); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
-    int32x4_t vacc0x23 = vreinterpretq_s32_u64(vshll_n_u32(vld1_u32(w), 0)); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
-    int32x4_t vacc0x45 = vreinterpretq_s32_u64(vshll_n_u32(vld1_u32(w), 0)); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
-    int32x4_t vacc0x67 = vreinterpretq_s32_u64(vshll_n_u32(vld1_u32(w), 0)); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
+    int32x4_t vacc0x01 = vreinterpretq_s32_u64(vmovl_u32(vld1_u32(w))); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
+    int32x4_t vacc0x23 = vreinterpretq_s32_u64(vmovl_u32(vld1_u32(w))); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
+    int32x4_t vacc0x45 = vreinterpretq_s32_u64(vmovl_u32(vld1_u32(w))); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
+    int32x4_t vacc0x67 = vreinterpretq_s32_u64(vmovl_u32(vld1_u32(w))); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
     int32x4_t vacc1x01 = vacc0x01;
     int32x4_t vacc1x23 = vacc0x23;
     int32x4_t vacc1x45 = vacc0x45;

diff --git a/src/qs8-gemm/gen/4x8c4-minmax-rndnu-neon-mlal-ld2r.c b/src/qs8-gemm/gen/4x8c4-minmax-rndnu-neon-mlal-ld2r.c
index 0469fe7..ac3bdb7 100644
--- a/src/qs8-gemm/gen/4x8c4-minmax-rndnu-neon-mlal-ld2r.c
+++ b/src/qs8-gemm/gen/4x8c4-minmax-rndnu-neon-mlal-ld2r.c

@@ -59,10 +59,10 @@
   }
 
   do {
-    int32x4_t vacc0x01 = vreinterpretq_s32_u64(vshll_n_u32(vld1_u32(w), 0)); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
-    int32x4_t vacc0x23 = vreinterpretq_s32_u64(vshll_n_u32(vld1_u32(w), 0)); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
-    int32x4_t vacc0x45 = vreinterpretq_s32_u64(vshll_n_u32(vld1_u32(w), 0)); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
-    int32x4_t vacc0x67 = vreinterpretq_s32_u64(vshll_n_u32(vld1_u32(w), 0)); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
+    int32x4_t vacc0x01 = vreinterpretq_s32_u64(vmovl_u32(vld1_u32(w))); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
+    int32x4_t vacc0x23 = vreinterpretq_s32_u64(vmovl_u32(vld1_u32(w))); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
+    int32x4_t vacc0x45 = vreinterpretq_s32_u64(vmovl_u32(vld1_u32(w))); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
+    int32x4_t vacc0x67 = vreinterpretq_s32_u64(vmovl_u32(vld1_u32(w))); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
     int32x4_t vacc1x01 = vacc0x01;
     int32x4_t vacc1x23 = vacc0x23;
     int32x4_t vacc1x45 = vacc0x45;

diff --git a/src/qs8-gemm/gen/4x8c4-minmax-rndnu-neon-mull-dup.c b/src/qs8-gemm/gen/4x8c4-minmax-rndnu-neon-mull-dup.c
index bd2e9d0..29470db 100644
--- a/src/qs8-gemm/gen/4x8c4-minmax-rndnu-neon-mull-dup.c
+++ b/src/qs8-gemm/gen/4x8c4-minmax-rndnu-neon-mull-dup.c

@@ -59,10 +59,10 @@
   }
 
   do {
-    int32x4_t vacc0x01 = vreinterpretq_s32_u64(vshll_n_u32(vld1_u32(w), 0)); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
-    int32x4_t vacc0x23 = vreinterpretq_s32_u64(vshll_n_u32(vld1_u32(w), 0)); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
-    int32x4_t vacc0x45 = vreinterpretq_s32_u64(vshll_n_u32(vld1_u32(w), 0)); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
-    int32x4_t vacc0x67 = vreinterpretq_s32_u64(vshll_n_u32(vld1_u32(w), 0)); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
+    int32x4_t vacc0x01 = vreinterpretq_s32_u64(vmovl_u32(vld1_u32(w))); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
+    int32x4_t vacc0x23 = vreinterpretq_s32_u64(vmovl_u32(vld1_u32(w))); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
+    int32x4_t vacc0x45 = vreinterpretq_s32_u64(vmovl_u32(vld1_u32(w))); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
+    int32x4_t vacc0x67 = vreinterpretq_s32_u64(vmovl_u32(vld1_u32(w))); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
     int32x4_t vacc1x01 = vacc0x01;
     int32x4_t vacc1x23 = vacc0x23;
     int32x4_t vacc1x45 = vacc0x45;

diff --git a/src/qs8-gemm/gen/4x8c4-minmax-rndnu-neon-mull-ld1r.c b/src/qs8-gemm/gen/4x8c4-minmax-rndnu-neon-mull-ld1r.c
index 7d552d0..fda26e2 100644
--- a/src/qs8-gemm/gen/4x8c4-minmax-rndnu-neon-mull-ld1r.c
+++ b/src/qs8-gemm/gen/4x8c4-minmax-rndnu-neon-mull-ld1r.c

@@ -59,10 +59,10 @@
   }
 
   do {
-    int32x4_t vacc0x01 = vreinterpretq_s32_u64(vshll_n_u32(vld1_u32(w), 0)); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
-    int32x4_t vacc0x23 = vreinterpretq_s32_u64(vshll_n_u32(vld1_u32(w), 0)); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
-    int32x4_t vacc0x45 = vreinterpretq_s32_u64(vshll_n_u32(vld1_u32(w), 0)); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
-    int32x4_t vacc0x67 = vreinterpretq_s32_u64(vshll_n_u32(vld1_u32(w), 0)); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
+    int32x4_t vacc0x01 = vreinterpretq_s32_u64(vmovl_u32(vld1_u32(w))); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
+    int32x4_t vacc0x23 = vreinterpretq_s32_u64(vmovl_u32(vld1_u32(w))); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
+    int32x4_t vacc0x45 = vreinterpretq_s32_u64(vmovl_u32(vld1_u32(w))); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
+    int32x4_t vacc0x67 = vreinterpretq_s32_u64(vmovl_u32(vld1_u32(w))); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
     int32x4_t vacc1x01 = vacc0x01;
     int32x4_t vacc1x23 = vacc0x23;
     int32x4_t vacc1x45 = vacc0x45;

diff --git a/src/qs8-gemm/gen/4x8c4-minmax-rndnu-neon-mull-ld2r.c b/src/qs8-gemm/gen/4x8c4-minmax-rndnu-neon-mull-ld2r.c
index db17f0f..cb1b9e7 100644
--- a/src/qs8-gemm/gen/4x8c4-minmax-rndnu-neon-mull-ld2r.c
+++ b/src/qs8-gemm/gen/4x8c4-minmax-rndnu-neon-mull-ld2r.c

@@ -59,10 +59,10 @@
   }
 
   do {
-    int32x4_t vacc0x01 = vreinterpretq_s32_u64(vshll_n_u32(vld1_u32(w), 0)); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
-    int32x4_t vacc0x23 = vreinterpretq_s32_u64(vshll_n_u32(vld1_u32(w), 0)); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
-    int32x4_t vacc0x45 = vreinterpretq_s32_u64(vshll_n_u32(vld1_u32(w), 0)); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
-    int32x4_t vacc0x67 = vreinterpretq_s32_u64(vshll_n_u32(vld1_u32(w), 0)); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
+    int32x4_t vacc0x01 = vreinterpretq_s32_u64(vmovl_u32(vld1_u32(w))); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
+    int32x4_t vacc0x23 = vreinterpretq_s32_u64(vmovl_u32(vld1_u32(w))); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
+    int32x4_t vacc0x45 = vreinterpretq_s32_u64(vmovl_u32(vld1_u32(w))); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
+    int32x4_t vacc0x67 = vreinterpretq_s32_u64(vmovl_u32(vld1_u32(w))); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
     int32x4_t vacc1x01 = vacc0x01;
     int32x4_t vacc1x23 = vacc0x23;
     int32x4_t vacc1x45 = vacc0x45;

diff --git a/src/qs8-gemm/gen/4x8c4s2-minmax-rndnu-neon-mlal.c b/src/qs8-gemm/gen/4x8c4s2-minmax-rndnu-neon-mlal.c
index dc7c356..ec92556 100644
--- a/src/qs8-gemm/gen/4x8c4s2-minmax-rndnu-neon-mlal.c
+++ b/src/qs8-gemm/gen/4x8c4s2-minmax-rndnu-neon-mlal.c

@@ -58,10 +58,10 @@
   }
 
   do {
-    int32x4_t vacc0x01 = vreinterpretq_s32_u64(vshll_n_u32(vld1_u32(w), 0)); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
-    int32x4_t vacc0x23 = vreinterpretq_s32_u64(vshll_n_u32(vld1_u32(w), 0)); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
-    int32x4_t vacc0x45 = vreinterpretq_s32_u64(vshll_n_u32(vld1_u32(w), 0)); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
-    int32x4_t vacc0x67 = vreinterpretq_s32_u64(vshll_n_u32(vld1_u32(w), 0)); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
+    int32x4_t vacc0x01 = vreinterpretq_s32_u64(vmovl_u32(vld1_u32(w))); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
+    int32x4_t vacc0x23 = vreinterpretq_s32_u64(vmovl_u32(vld1_u32(w))); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
+    int32x4_t vacc0x45 = vreinterpretq_s32_u64(vmovl_u32(vld1_u32(w))); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
+    int32x4_t vacc0x67 = vreinterpretq_s32_u64(vmovl_u32(vld1_u32(w))); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
     int32x4_t vacc1x01 = vacc0x01;
     int32x4_t vacc1x23 = vacc0x23;
     int32x4_t vacc1x45 = vacc0x45;

diff --git a/src/qs8-gemm/gen/4x8c4s2-minmax-rndnu-neon-mull.c b/src/qs8-gemm/gen/4x8c4s2-minmax-rndnu-neon-mull.c
index 27b664d..b59e308 100644
--- a/src/qs8-gemm/gen/4x8c4s2-minmax-rndnu-neon-mull.c
+++ b/src/qs8-gemm/gen/4x8c4s2-minmax-rndnu-neon-mull.c

@@ -58,10 +58,10 @@
   }
 
   do {
-    int32x4_t vacc0x01 = vreinterpretq_s32_u64(vshll_n_u32(vld1_u32(w), 0)); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
-    int32x4_t vacc0x23 = vreinterpretq_s32_u64(vshll_n_u32(vld1_u32(w), 0)); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
-    int32x4_t vacc0x45 = vreinterpretq_s32_u64(vshll_n_u32(vld1_u32(w), 0)); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
-    int32x4_t vacc0x67 = vreinterpretq_s32_u64(vshll_n_u32(vld1_u32(w), 0)); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
+    int32x4_t vacc0x01 = vreinterpretq_s32_u64(vmovl_u32(vld1_u32(w))); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
+    int32x4_t vacc0x23 = vreinterpretq_s32_u64(vmovl_u32(vld1_u32(w))); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
+    int32x4_t vacc0x45 = vreinterpretq_s32_u64(vmovl_u32(vld1_u32(w))); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
+    int32x4_t vacc0x67 = vreinterpretq_s32_u64(vmovl_u32(vld1_u32(w))); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
     int32x4_t vacc1x01 = vacc0x01;
     int32x4_t vacc1x23 = vacc0x23;
     int32x4_t vacc1x45 = vacc0x45;

diff --git a/src/qs8-igemm/c4-neon-mull-dup.c.in b/src/qs8-igemm/c4-neon-mull-dup.c.in
index 54ab227..ffbf417 100644
--- a/src/qs8-igemm/c4-neon-mull-dup.c.in
+++ b/src/qs8-igemm/c4-neon-mull-dup.c.in

@@ -67,7 +67,7 @@
 
   do {
     $for N in range(0, NR, 2):
-      int32x4_t vacc0x${ABC[N:N+2]} = vreinterpretq_s32_u64(vshll_n_u32(vld1_u32(w), 0)); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
+      int32x4_t vacc0x${ABC[N:N+2]} = vreinterpretq_s32_u64(vmovl_u32(vld1_u32(w))); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
     $for M in range(1, MR):
       $for N in range(0, NR, 2):
         int32x4_t vacc${M}x${ABC[N:N+2]} = vacc0x${ABC[N:N+2]};

diff --git a/src/qs8-igemm/c4-neon-mull-shuffle.c.in b/src/qs8-igemm/c4-neon-mull-shuffle.c.in
index ee058da..0646505 100644
--- a/src/qs8-igemm/c4-neon-mull-shuffle.c.in
+++ b/src/qs8-igemm/c4-neon-mull-shuffle.c.in

@@ -65,7 +65,7 @@
 
   do {
     $for N in range(0, NR, 2):
-      int32x4_t vacc0x${ABC[N:N+2]} = vreinterpretq_s32_u64(vshll_n_u32(vld1_u32(w), 0)); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
+      int32x4_t vacc0x${ABC[N:N+2]} = vreinterpretq_s32_u64(vmovl_u32(vld1_u32(w))); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
     $for M in range(1, MR):
       $for N in range(0, NR, 2):
         int32x4_t vacc${M}x${ABC[N:N+2]} = vacc0x${ABC[N:N+2]};

diff --git a/src/qs8-igemm/gen/1x16c4-minmax-rndnu-neon-mlal-dup.c b/src/qs8-igemm/gen/1x16c4-minmax-rndnu-neon-mlal-dup.c
index 632abf2..b2bc78b 100644
--- a/src/qs8-igemm/gen/1x16c4-minmax-rndnu-neon-mlal-dup.c
+++ b/src/qs8-igemm/gen/1x16c4-minmax-rndnu-neon-mlal-dup.c

@@ -44,14 +44,14 @@
   int8_t* c0 = c;
 
   do {
-    int32x4_t vacc0x01 = vreinterpretq_s32_u64(vshll_n_u32(vld1_u32(w), 0)); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
-    int32x4_t vacc0x23 = vreinterpretq_s32_u64(vshll_n_u32(vld1_u32(w), 0)); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
-    int32x4_t vacc0x45 = vreinterpretq_s32_u64(vshll_n_u32(vld1_u32(w), 0)); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
-    int32x4_t vacc0x67 = vreinterpretq_s32_u64(vshll_n_u32(vld1_u32(w), 0)); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
-    int32x4_t vacc0x89 = vreinterpretq_s32_u64(vshll_n_u32(vld1_u32(w), 0)); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
-    int32x4_t vacc0xAB = vreinterpretq_s32_u64(vshll_n_u32(vld1_u32(w), 0)); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
-    int32x4_t vacc0xCD = vreinterpretq_s32_u64(vshll_n_u32(vld1_u32(w), 0)); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
-    int32x4_t vacc0xEF = vreinterpretq_s32_u64(vshll_n_u32(vld1_u32(w), 0)); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
+    int32x4_t vacc0x01 = vreinterpretq_s32_u64(vmovl_u32(vld1_u32(w))); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
+    int32x4_t vacc0x23 = vreinterpretq_s32_u64(vmovl_u32(vld1_u32(w))); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
+    int32x4_t vacc0x45 = vreinterpretq_s32_u64(vmovl_u32(vld1_u32(w))); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
+    int32x4_t vacc0x67 = vreinterpretq_s32_u64(vmovl_u32(vld1_u32(w))); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
+    int32x4_t vacc0x89 = vreinterpretq_s32_u64(vmovl_u32(vld1_u32(w))); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
+    int32x4_t vacc0xAB = vreinterpretq_s32_u64(vmovl_u32(vld1_u32(w))); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
+    int32x4_t vacc0xCD = vreinterpretq_s32_u64(vmovl_u32(vld1_u32(w))); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
+    int32x4_t vacc0xEF = vreinterpretq_s32_u64(vmovl_u32(vld1_u32(w))); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
 
     size_t p = ks;
     do {

diff --git a/src/qs8-igemm/gen/1x16c4-minmax-rndnu-neon-mlal-ld1r.c b/src/qs8-igemm/gen/1x16c4-minmax-rndnu-neon-mlal-ld1r.c
index b929b1c..89a17c1 100644
--- a/src/qs8-igemm/gen/1x16c4-minmax-rndnu-neon-mlal-ld1r.c
+++ b/src/qs8-igemm/gen/1x16c4-minmax-rndnu-neon-mlal-ld1r.c

@@ -44,14 +44,14 @@
   int8_t* c0 = c;
 
   do {
-    int32x4_t vacc0x01 = vreinterpretq_s32_u64(vshll_n_u32(vld1_u32(w), 0)); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
-    int32x4_t vacc0x23 = vreinterpretq_s32_u64(vshll_n_u32(vld1_u32(w), 0)); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
-    int32x4_t vacc0x45 = vreinterpretq_s32_u64(vshll_n_u32(vld1_u32(w), 0)); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
-    int32x4_t vacc0x67 = vreinterpretq_s32_u64(vshll_n_u32(vld1_u32(w), 0)); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
-    int32x4_t vacc0x89 = vreinterpretq_s32_u64(vshll_n_u32(vld1_u32(w), 0)); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
-    int32x4_t vacc0xAB = vreinterpretq_s32_u64(vshll_n_u32(vld1_u32(w), 0)); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
-    int32x4_t vacc0xCD = vreinterpretq_s32_u64(vshll_n_u32(vld1_u32(w), 0)); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
-    int32x4_t vacc0xEF = vreinterpretq_s32_u64(vshll_n_u32(vld1_u32(w), 0)); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
+    int32x4_t vacc0x01 = vreinterpretq_s32_u64(vmovl_u32(vld1_u32(w))); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
+    int32x4_t vacc0x23 = vreinterpretq_s32_u64(vmovl_u32(vld1_u32(w))); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
+    int32x4_t vacc0x45 = vreinterpretq_s32_u64(vmovl_u32(vld1_u32(w))); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
+    int32x4_t vacc0x67 = vreinterpretq_s32_u64(vmovl_u32(vld1_u32(w))); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
+    int32x4_t vacc0x89 = vreinterpretq_s32_u64(vmovl_u32(vld1_u32(w))); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
+    int32x4_t vacc0xAB = vreinterpretq_s32_u64(vmovl_u32(vld1_u32(w))); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
+    int32x4_t vacc0xCD = vreinterpretq_s32_u64(vmovl_u32(vld1_u32(w))); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
+    int32x4_t vacc0xEF = vreinterpretq_s32_u64(vmovl_u32(vld1_u32(w))); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
 
     size_t p = ks;
     do {

diff --git a/src/qs8-igemm/gen/1x16c4-minmax-rndnu-neon-mlal-ld2r.c b/src/qs8-igemm/gen/1x16c4-minmax-rndnu-neon-mlal-ld2r.c
index 064920f..1c425a4 100644
--- a/src/qs8-igemm/gen/1x16c4-minmax-rndnu-neon-mlal-ld2r.c
+++ b/src/qs8-igemm/gen/1x16c4-minmax-rndnu-neon-mlal-ld2r.c

@@ -44,14 +44,14 @@
   int8_t* c0 = c;
 
   do {
-    int32x4_t vacc0x01 = vreinterpretq_s32_u64(vshll_n_u32(vld1_u32(w), 0)); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
-    int32x4_t vacc0x23 = vreinterpretq_s32_u64(vshll_n_u32(vld1_u32(w), 0)); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
-    int32x4_t vacc0x45 = vreinterpretq_s32_u64(vshll_n_u32(vld1_u32(w), 0)); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
-    int32x4_t vacc0x67 = vreinterpretq_s32_u64(vshll_n_u32(vld1_u32(w), 0)); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
-    int32x4_t vacc0x89 = vreinterpretq_s32_u64(vshll_n_u32(vld1_u32(w), 0)); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
-    int32x4_t vacc0xAB = vreinterpretq_s32_u64(vshll_n_u32(vld1_u32(w), 0)); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
-    int32x4_t vacc0xCD = vreinterpretq_s32_u64(vshll_n_u32(vld1_u32(w), 0)); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
-    int32x4_t vacc0xEF = vreinterpretq_s32_u64(vshll_n_u32(vld1_u32(w), 0)); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
+    int32x4_t vacc0x01 = vreinterpretq_s32_u64(vmovl_u32(vld1_u32(w))); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
+    int32x4_t vacc0x23 = vreinterpretq_s32_u64(vmovl_u32(vld1_u32(w))); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
+    int32x4_t vacc0x45 = vreinterpretq_s32_u64(vmovl_u32(vld1_u32(w))); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
+    int32x4_t vacc0x67 = vreinterpretq_s32_u64(vmovl_u32(vld1_u32(w))); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
+    int32x4_t vacc0x89 = vreinterpretq_s32_u64(vmovl_u32(vld1_u32(w))); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
+    int32x4_t vacc0xAB = vreinterpretq_s32_u64(vmovl_u32(vld1_u32(w))); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
+    int32x4_t vacc0xCD = vreinterpretq_s32_u64(vmovl_u32(vld1_u32(w))); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
+    int32x4_t vacc0xEF = vreinterpretq_s32_u64(vmovl_u32(vld1_u32(w))); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
 
     size_t p = ks;
     do {

diff --git a/src/qs8-igemm/gen/1x16c4-minmax-rndnu-neon-mull-dup.c b/src/qs8-igemm/gen/1x16c4-minmax-rndnu-neon-mull-dup.c
index be0637a..6273d95 100644
--- a/src/qs8-igemm/gen/1x16c4-minmax-rndnu-neon-mull-dup.c
+++ b/src/qs8-igemm/gen/1x16c4-minmax-rndnu-neon-mull-dup.c

@@ -44,14 +44,14 @@
   int8_t* c0 = c;
 
   do {
-    int32x4_t vacc0x01 = vreinterpretq_s32_u64(vshll_n_u32(vld1_u32(w), 0)); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
-    int32x4_t vacc0x23 = vreinterpretq_s32_u64(vshll_n_u32(vld1_u32(w), 0)); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
-    int32x4_t vacc0x45 = vreinterpretq_s32_u64(vshll_n_u32(vld1_u32(w), 0)); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
-    int32x4_t vacc0x67 = vreinterpretq_s32_u64(vshll_n_u32(vld1_u32(w), 0)); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
-    int32x4_t vacc0x89 = vreinterpretq_s32_u64(vshll_n_u32(vld1_u32(w), 0)); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
-    int32x4_t vacc0xAB = vreinterpretq_s32_u64(vshll_n_u32(vld1_u32(w), 0)); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
-    int32x4_t vacc0xCD = vreinterpretq_s32_u64(vshll_n_u32(vld1_u32(w), 0)); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
-    int32x4_t vacc0xEF = vreinterpretq_s32_u64(vshll_n_u32(vld1_u32(w), 0)); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
+    int32x4_t vacc0x01 = vreinterpretq_s32_u64(vmovl_u32(vld1_u32(w))); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
+    int32x4_t vacc0x23 = vreinterpretq_s32_u64(vmovl_u32(vld1_u32(w))); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
+    int32x4_t vacc0x45 = vreinterpretq_s32_u64(vmovl_u32(vld1_u32(w))); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
+    int32x4_t vacc0x67 = vreinterpretq_s32_u64(vmovl_u32(vld1_u32(w))); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
+    int32x4_t vacc0x89 = vreinterpretq_s32_u64(vmovl_u32(vld1_u32(w))); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
+    int32x4_t vacc0xAB = vreinterpretq_s32_u64(vmovl_u32(vld1_u32(w))); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
+    int32x4_t vacc0xCD = vreinterpretq_s32_u64(vmovl_u32(vld1_u32(w))); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
+    int32x4_t vacc0xEF = vreinterpretq_s32_u64(vmovl_u32(vld1_u32(w))); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
 
     size_t p = ks;
     do {

diff --git a/src/qs8-igemm/gen/1x16c4-minmax-rndnu-neon-mull-ld1r.c b/src/qs8-igemm/gen/1x16c4-minmax-rndnu-neon-mull-ld1r.c
index ba07808..d0f215e 100644
--- a/src/qs8-igemm/gen/1x16c4-minmax-rndnu-neon-mull-ld1r.c
+++ b/src/qs8-igemm/gen/1x16c4-minmax-rndnu-neon-mull-ld1r.c

@@ -44,14 +44,14 @@
   int8_t* c0 = c;
 
   do {
-    int32x4_t vacc0x01 = vreinterpretq_s32_u64(vshll_n_u32(vld1_u32(w), 0)); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
-    int32x4_t vacc0x23 = vreinterpretq_s32_u64(vshll_n_u32(vld1_u32(w), 0)); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
-    int32x4_t vacc0x45 = vreinterpretq_s32_u64(vshll_n_u32(vld1_u32(w), 0)); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
-    int32x4_t vacc0x67 = vreinterpretq_s32_u64(vshll_n_u32(vld1_u32(w), 0)); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
-    int32x4_t vacc0x89 = vreinterpretq_s32_u64(vshll_n_u32(vld1_u32(w), 0)); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
-    int32x4_t vacc0xAB = vreinterpretq_s32_u64(vshll_n_u32(vld1_u32(w), 0)); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
-    int32x4_t vacc0xCD = vreinterpretq_s32_u64(vshll_n_u32(vld1_u32(w), 0)); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
-    int32x4_t vacc0xEF = vreinterpretq_s32_u64(vshll_n_u32(vld1_u32(w), 0)); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
+    int32x4_t vacc0x01 = vreinterpretq_s32_u64(vmovl_u32(vld1_u32(w))); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
+    int32x4_t vacc0x23 = vreinterpretq_s32_u64(vmovl_u32(vld1_u32(w))); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
+    int32x4_t vacc0x45 = vreinterpretq_s32_u64(vmovl_u32(vld1_u32(w))); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
+    int32x4_t vacc0x67 = vreinterpretq_s32_u64(vmovl_u32(vld1_u32(w))); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
+    int32x4_t vacc0x89 = vreinterpretq_s32_u64(vmovl_u32(vld1_u32(w))); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
+    int32x4_t vacc0xAB = vreinterpretq_s32_u64(vmovl_u32(vld1_u32(w))); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
+    int32x4_t vacc0xCD = vreinterpretq_s32_u64(vmovl_u32(vld1_u32(w))); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
+    int32x4_t vacc0xEF = vreinterpretq_s32_u64(vmovl_u32(vld1_u32(w))); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
 
     size_t p = ks;
     do {

diff --git a/src/qs8-igemm/gen/1x16c4-minmax-rndnu-neon-mull-ld2r.c b/src/qs8-igemm/gen/1x16c4-minmax-rndnu-neon-mull-ld2r.c
index 1b96f7f..09fdd31 100644
--- a/src/qs8-igemm/gen/1x16c4-minmax-rndnu-neon-mull-ld2r.c
+++ b/src/qs8-igemm/gen/1x16c4-minmax-rndnu-neon-mull-ld2r.c

@@ -44,14 +44,14 @@
   int8_t* c0 = c;
 
   do {
-    int32x4_t vacc0x01 = vreinterpretq_s32_u64(vshll_n_u32(vld1_u32(w), 0)); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
-    int32x4_t vacc0x23 = vreinterpretq_s32_u64(vshll_n_u32(vld1_u32(w), 0)); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
-    int32x4_t vacc0x45 = vreinterpretq_s32_u64(vshll_n_u32(vld1_u32(w), 0)); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
-    int32x4_t vacc0x67 = vreinterpretq_s32_u64(vshll_n_u32(vld1_u32(w), 0)); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
-    int32x4_t vacc0x89 = vreinterpretq_s32_u64(vshll_n_u32(vld1_u32(w), 0)); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
-    int32x4_t vacc0xAB = vreinterpretq_s32_u64(vshll_n_u32(vld1_u32(w), 0)); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
-    int32x4_t vacc0xCD = vreinterpretq_s32_u64(vshll_n_u32(vld1_u32(w), 0)); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
-    int32x4_t vacc0xEF = vreinterpretq_s32_u64(vshll_n_u32(vld1_u32(w), 0)); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
+    int32x4_t vacc0x01 = vreinterpretq_s32_u64(vmovl_u32(vld1_u32(w))); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
+    int32x4_t vacc0x23 = vreinterpretq_s32_u64(vmovl_u32(vld1_u32(w))); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
+    int32x4_t vacc0x45 = vreinterpretq_s32_u64(vmovl_u32(vld1_u32(w))); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
+    int32x4_t vacc0x67 = vreinterpretq_s32_u64(vmovl_u32(vld1_u32(w))); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
+    int32x4_t vacc0x89 = vreinterpretq_s32_u64(vmovl_u32(vld1_u32(w))); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
+    int32x4_t vacc0xAB = vreinterpretq_s32_u64(vmovl_u32(vld1_u32(w))); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
+    int32x4_t vacc0xCD = vreinterpretq_s32_u64(vmovl_u32(vld1_u32(w))); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
+    int32x4_t vacc0xEF = vreinterpretq_s32_u64(vmovl_u32(vld1_u32(w))); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
 
     size_t p = ks;
     do {

diff --git a/src/qs8-igemm/gen/1x16c4s2-minmax-rndnu-neon-mlal.c b/src/qs8-igemm/gen/1x16c4s2-minmax-rndnu-neon-mlal.c
index 3808403..bbc2c80 100644
--- a/src/qs8-igemm/gen/1x16c4s2-minmax-rndnu-neon-mlal.c
+++ b/src/qs8-igemm/gen/1x16c4s2-minmax-rndnu-neon-mlal.c

@@ -43,14 +43,14 @@
   int8_t* c0 = c;
 
   do {
-    int32x4_t vacc0x01 = vreinterpretq_s32_u64(vshll_n_u32(vld1_u32(w), 0)); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
-    int32x4_t vacc0x23 = vreinterpretq_s32_u64(vshll_n_u32(vld1_u32(w), 0)); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
-    int32x4_t vacc0x45 = vreinterpretq_s32_u64(vshll_n_u32(vld1_u32(w), 0)); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
-    int32x4_t vacc0x67 = vreinterpretq_s32_u64(vshll_n_u32(vld1_u32(w), 0)); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
-    int32x4_t vacc0x89 = vreinterpretq_s32_u64(vshll_n_u32(vld1_u32(w), 0)); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
-    int32x4_t vacc0xAB = vreinterpretq_s32_u64(vshll_n_u32(vld1_u32(w), 0)); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
-    int32x4_t vacc0xCD = vreinterpretq_s32_u64(vshll_n_u32(vld1_u32(w), 0)); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
-    int32x4_t vacc0xEF = vreinterpretq_s32_u64(vshll_n_u32(vld1_u32(w), 0)); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
+    int32x4_t vacc0x01 = vreinterpretq_s32_u64(vmovl_u32(vld1_u32(w))); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
+    int32x4_t vacc0x23 = vreinterpretq_s32_u64(vmovl_u32(vld1_u32(w))); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
+    int32x4_t vacc0x45 = vreinterpretq_s32_u64(vmovl_u32(vld1_u32(w))); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
+    int32x4_t vacc0x67 = vreinterpretq_s32_u64(vmovl_u32(vld1_u32(w))); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
+    int32x4_t vacc0x89 = vreinterpretq_s32_u64(vmovl_u32(vld1_u32(w))); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
+    int32x4_t vacc0xAB = vreinterpretq_s32_u64(vmovl_u32(vld1_u32(w))); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
+    int32x4_t vacc0xCD = vreinterpretq_s32_u64(vmovl_u32(vld1_u32(w))); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
+    int32x4_t vacc0xEF = vreinterpretq_s32_u64(vmovl_u32(vld1_u32(w))); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
 
     size_t p = ks;
     do {

diff --git a/src/qs8-igemm/gen/1x16c4s2-minmax-rndnu-neon-mull.c b/src/qs8-igemm/gen/1x16c4s2-minmax-rndnu-neon-mull.c
index 4f7e340..96998cd 100644
--- a/src/qs8-igemm/gen/1x16c4s2-minmax-rndnu-neon-mull.c
+++ b/src/qs8-igemm/gen/1x16c4s2-minmax-rndnu-neon-mull.c

@@ -43,14 +43,14 @@
   int8_t* c0 = c;
 
   do {
-    int32x4_t vacc0x01 = vreinterpretq_s32_u64(vshll_n_u32(vld1_u32(w), 0)); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
-    int32x4_t vacc0x23 = vreinterpretq_s32_u64(vshll_n_u32(vld1_u32(w), 0)); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
-    int32x4_t vacc0x45 = vreinterpretq_s32_u64(vshll_n_u32(vld1_u32(w), 0)); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
-    int32x4_t vacc0x67 = vreinterpretq_s32_u64(vshll_n_u32(vld1_u32(w), 0)); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
-    int32x4_t vacc0x89 = vreinterpretq_s32_u64(vshll_n_u32(vld1_u32(w), 0)); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
-    int32x4_t vacc0xAB = vreinterpretq_s32_u64(vshll_n_u32(vld1_u32(w), 0)); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
-    int32x4_t vacc0xCD = vreinterpretq_s32_u64(vshll_n_u32(vld1_u32(w), 0)); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
-    int32x4_t vacc0xEF = vreinterpretq_s32_u64(vshll_n_u32(vld1_u32(w), 0)); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
+    int32x4_t vacc0x01 = vreinterpretq_s32_u64(vmovl_u32(vld1_u32(w))); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
+    int32x4_t vacc0x23 = vreinterpretq_s32_u64(vmovl_u32(vld1_u32(w))); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
+    int32x4_t vacc0x45 = vreinterpretq_s32_u64(vmovl_u32(vld1_u32(w))); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
+    int32x4_t vacc0x67 = vreinterpretq_s32_u64(vmovl_u32(vld1_u32(w))); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
+    int32x4_t vacc0x89 = vreinterpretq_s32_u64(vmovl_u32(vld1_u32(w))); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
+    int32x4_t vacc0xAB = vreinterpretq_s32_u64(vmovl_u32(vld1_u32(w))); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
+    int32x4_t vacc0xCD = vreinterpretq_s32_u64(vmovl_u32(vld1_u32(w))); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
+    int32x4_t vacc0xEF = vreinterpretq_s32_u64(vmovl_u32(vld1_u32(w))); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
 
     size_t p = ks;
     do {

diff --git a/src/qs8-igemm/gen/1x8c4-minmax-fp32-neon-mlal-dup.c b/src/qs8-igemm/gen/1x8c4-minmax-fp32-neon-mlal-dup.c
index df491eb..055c7f8 100644
--- a/src/qs8-igemm/gen/1x8c4-minmax-fp32-neon-mlal-dup.c
+++ b/src/qs8-igemm/gen/1x8c4-minmax-fp32-neon-mlal-dup.c

@@ -44,10 +44,10 @@
   int8_t* c0 = c;
 
   do {
-    int32x4_t vacc0x01 = vreinterpretq_s32_u64(vshll_n_u32(vld1_u32(w), 0)); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
-    int32x4_t vacc0x23 = vreinterpretq_s32_u64(vshll_n_u32(vld1_u32(w), 0)); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
-    int32x4_t vacc0x45 = vreinterpretq_s32_u64(vshll_n_u32(vld1_u32(w), 0)); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
-    int32x4_t vacc0x67 = vreinterpretq_s32_u64(vshll_n_u32(vld1_u32(w), 0)); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
+    int32x4_t vacc0x01 = vreinterpretq_s32_u64(vmovl_u32(vld1_u32(w))); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
+    int32x4_t vacc0x23 = vreinterpretq_s32_u64(vmovl_u32(vld1_u32(w))); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
+    int32x4_t vacc0x45 = vreinterpretq_s32_u64(vmovl_u32(vld1_u32(w))); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
+    int32x4_t vacc0x67 = vreinterpretq_s32_u64(vmovl_u32(vld1_u32(w))); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
 
     size_t p = ks;
     do {

diff --git a/src/qs8-igemm/gen/1x8c4-minmax-fp32-neon-mlal-ld1r.c b/src/qs8-igemm/gen/1x8c4-minmax-fp32-neon-mlal-ld1r.c
index c1ca6f2..9d1efa4 100644
--- a/src/qs8-igemm/gen/1x8c4-minmax-fp32-neon-mlal-ld1r.c
+++ b/src/qs8-igemm/gen/1x8c4-minmax-fp32-neon-mlal-ld1r.c

@@ -44,10 +44,10 @@
   int8_t* c0 = c;
 
   do {
-    int32x4_t vacc0x01 = vreinterpretq_s32_u64(vshll_n_u32(vld1_u32(w), 0)); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
-    int32x4_t vacc0x23 = vreinterpretq_s32_u64(vshll_n_u32(vld1_u32(w), 0)); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
-    int32x4_t vacc0x45 = vreinterpretq_s32_u64(vshll_n_u32(vld1_u32(w), 0)); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
-    int32x4_t vacc0x67 = vreinterpretq_s32_u64(vshll_n_u32(vld1_u32(w), 0)); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
+    int32x4_t vacc0x01 = vreinterpretq_s32_u64(vmovl_u32(vld1_u32(w))); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
+    int32x4_t vacc0x23 = vreinterpretq_s32_u64(vmovl_u32(vld1_u32(w))); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
+    int32x4_t vacc0x45 = vreinterpretq_s32_u64(vmovl_u32(vld1_u32(w))); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
+    int32x4_t vacc0x67 = vreinterpretq_s32_u64(vmovl_u32(vld1_u32(w))); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
 
     size_t p = ks;
     do {

diff --git a/src/qs8-igemm/gen/1x8c4-minmax-fp32-neon-mlal-ld2r.c b/src/qs8-igemm/gen/1x8c4-minmax-fp32-neon-mlal-ld2r.c
index ce68a33..e2b3036 100644
--- a/src/qs8-igemm/gen/1x8c4-minmax-fp32-neon-mlal-ld2r.c
+++ b/src/qs8-igemm/gen/1x8c4-minmax-fp32-neon-mlal-ld2r.c

@@ -44,10 +44,10 @@
   int8_t* c0 = c;
 
   do {
-    int32x4_t vacc0x01 = vreinterpretq_s32_u64(vshll_n_u32(vld1_u32(w), 0)); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
-    int32x4_t vacc0x23 = vreinterpretq_s32_u64(vshll_n_u32(vld1_u32(w), 0)); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
-    int32x4_t vacc0x45 = vreinterpretq_s32_u64(vshll_n_u32(vld1_u32(w), 0)); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
-    int32x4_t vacc0x67 = vreinterpretq_s32_u64(vshll_n_u32(vld1_u32(w), 0)); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
+    int32x4_t vacc0x01 = vreinterpretq_s32_u64(vmovl_u32(vld1_u32(w))); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
+    int32x4_t vacc0x23 = vreinterpretq_s32_u64(vmovl_u32(vld1_u32(w))); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
+    int32x4_t vacc0x45 = vreinterpretq_s32_u64(vmovl_u32(vld1_u32(w))); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
+    int32x4_t vacc0x67 = vreinterpretq_s32_u64(vmovl_u32(vld1_u32(w))); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
 
     size_t p = ks;
     do {

diff --git a/src/qs8-igemm/gen/1x8c4-minmax-fp32-neonv8-mlal-dup.c b/src/qs8-igemm/gen/1x8c4-minmax-fp32-neonv8-mlal-dup.c
index 251856b..e7d491c 100644
--- a/src/qs8-igemm/gen/1x8c4-minmax-fp32-neonv8-mlal-dup.c
+++ b/src/qs8-igemm/gen/1x8c4-minmax-fp32-neonv8-mlal-dup.c

@@ -45,10 +45,10 @@
   int8_t* c0 = c;
 
   do {
-    int32x4_t vacc0x01 = vreinterpretq_s32_u64(vshll_n_u32(vld1_u32(w), 0)); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
-    int32x4_t vacc0x23 = vreinterpretq_s32_u64(vshll_n_u32(vld1_u32(w), 0)); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
-    int32x4_t vacc0x45 = vreinterpretq_s32_u64(vshll_n_u32(vld1_u32(w), 0)); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
-    int32x4_t vacc0x67 = vreinterpretq_s32_u64(vshll_n_u32(vld1_u32(w), 0)); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
+    int32x4_t vacc0x01 = vreinterpretq_s32_u64(vmovl_u32(vld1_u32(w))); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
+    int32x4_t vacc0x23 = vreinterpretq_s32_u64(vmovl_u32(vld1_u32(w))); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
+    int32x4_t vacc0x45 = vreinterpretq_s32_u64(vmovl_u32(vld1_u32(w))); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
+    int32x4_t vacc0x67 = vreinterpretq_s32_u64(vmovl_u32(vld1_u32(w))); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
 
     size_t p = ks;
     do {

diff --git a/src/qs8-igemm/gen/1x8c4-minmax-fp32-neonv8-mlal-ld1r.c b/src/qs8-igemm/gen/1x8c4-minmax-fp32-neonv8-mlal-ld1r.c
index e92ac99..9da338d 100644
--- a/src/qs8-igemm/gen/1x8c4-minmax-fp32-neonv8-mlal-ld1r.c
+++ b/src/qs8-igemm/gen/1x8c4-minmax-fp32-neonv8-mlal-ld1r.c

@@ -45,10 +45,10 @@
   int8_t* c0 = c;
 
   do {
-    int32x4_t vacc0x01 = vreinterpretq_s32_u64(vshll_n_u32(vld1_u32(w), 0)); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
-    int32x4_t vacc0x23 = vreinterpretq_s32_u64(vshll_n_u32(vld1_u32(w), 0)); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
-    int32x4_t vacc0x45 = vreinterpretq_s32_u64(vshll_n_u32(vld1_u32(w), 0)); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
-    int32x4_t vacc0x67 = vreinterpretq_s32_u64(vshll_n_u32(vld1_u32(w), 0)); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
+    int32x4_t vacc0x01 = vreinterpretq_s32_u64(vmovl_u32(vld1_u32(w))); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
+    int32x4_t vacc0x23 = vreinterpretq_s32_u64(vmovl_u32(vld1_u32(w))); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
+    int32x4_t vacc0x45 = vreinterpretq_s32_u64(vmovl_u32(vld1_u32(w))); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
+    int32x4_t vacc0x67 = vreinterpretq_s32_u64(vmovl_u32(vld1_u32(w))); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
 
     size_t p = ks;
     do {

diff --git a/src/qs8-igemm/gen/1x8c4-minmax-fp32-neonv8-mlal-ld2r.c b/src/qs8-igemm/gen/1x8c4-minmax-fp32-neonv8-mlal-ld2r.c
index a48ff28..bce8dc5 100644
--- a/src/qs8-igemm/gen/1x8c4-minmax-fp32-neonv8-mlal-ld2r.c
+++ b/src/qs8-igemm/gen/1x8c4-minmax-fp32-neonv8-mlal-ld2r.c

@@ -45,10 +45,10 @@
   int8_t* c0 = c;
 
   do {
-    int32x4_t vacc0x01 = vreinterpretq_s32_u64(vshll_n_u32(vld1_u32(w), 0)); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
-    int32x4_t vacc0x23 = vreinterpretq_s32_u64(vshll_n_u32(vld1_u32(w), 0)); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
-    int32x4_t vacc0x45 = vreinterpretq_s32_u64(vshll_n_u32(vld1_u32(w), 0)); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
-    int32x4_t vacc0x67 = vreinterpretq_s32_u64(vshll_n_u32(vld1_u32(w), 0)); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
+    int32x4_t vacc0x01 = vreinterpretq_s32_u64(vmovl_u32(vld1_u32(w))); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
+    int32x4_t vacc0x23 = vreinterpretq_s32_u64(vmovl_u32(vld1_u32(w))); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
+    int32x4_t vacc0x45 = vreinterpretq_s32_u64(vmovl_u32(vld1_u32(w))); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
+    int32x4_t vacc0x67 = vreinterpretq_s32_u64(vmovl_u32(vld1_u32(w))); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
 
     size_t p = ks;
     do {

diff --git a/src/qs8-igemm/gen/1x8c4-minmax-rndnu-neon-mlal-dup.c b/src/qs8-igemm/gen/1x8c4-minmax-rndnu-neon-mlal-dup.c
index 922acda..708fc34 100644
--- a/src/qs8-igemm/gen/1x8c4-minmax-rndnu-neon-mlal-dup.c
+++ b/src/qs8-igemm/gen/1x8c4-minmax-rndnu-neon-mlal-dup.c

@@ -44,10 +44,10 @@
   int8_t* c0 = c;
 
   do {
-    int32x4_t vacc0x01 = vreinterpretq_s32_u64(vshll_n_u32(vld1_u32(w), 0)); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
-    int32x4_t vacc0x23 = vreinterpretq_s32_u64(vshll_n_u32(vld1_u32(w), 0)); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
-    int32x4_t vacc0x45 = vreinterpretq_s32_u64(vshll_n_u32(vld1_u32(w), 0)); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
-    int32x4_t vacc0x67 = vreinterpretq_s32_u64(vshll_n_u32(vld1_u32(w), 0)); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
+    int32x4_t vacc0x01 = vreinterpretq_s32_u64(vmovl_u32(vld1_u32(w))); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
+    int32x4_t vacc0x23 = vreinterpretq_s32_u64(vmovl_u32(vld1_u32(w))); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
+    int32x4_t vacc0x45 = vreinterpretq_s32_u64(vmovl_u32(vld1_u32(w))); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
+    int32x4_t vacc0x67 = vreinterpretq_s32_u64(vmovl_u32(vld1_u32(w))); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
 
     size_t p = ks;
     do {

diff --git a/src/qs8-igemm/gen/1x8c4-minmax-rndnu-neon-mlal-ld1r.c b/src/qs8-igemm/gen/1x8c4-minmax-rndnu-neon-mlal-ld1r.c
index 8be89f3..d52b87f 100644
--- a/src/qs8-igemm/gen/1x8c4-minmax-rndnu-neon-mlal-ld1r.c
+++ b/src/qs8-igemm/gen/1x8c4-minmax-rndnu-neon-mlal-ld1r.c

@@ -44,10 +44,10 @@
   int8_t* c0 = c;
 
   do {
-    int32x4_t vacc0x01 = vreinterpretq_s32_u64(vshll_n_u32(vld1_u32(w), 0)); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
-    int32x4_t vacc0x23 = vreinterpretq_s32_u64(vshll_n_u32(vld1_u32(w), 0)); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
-    int32x4_t vacc0x45 = vreinterpretq_s32_u64(vshll_n_u32(vld1_u32(w), 0)); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
-    int32x4_t vacc0x67 = vreinterpretq_s32_u64(vshll_n_u32(vld1_u32(w), 0)); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
+    int32x4_t vacc0x01 = vreinterpretq_s32_u64(vmovl_u32(vld1_u32(w))); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
+    int32x4_t vacc0x23 = vreinterpretq_s32_u64(vmovl_u32(vld1_u32(w))); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
+    int32x4_t vacc0x45 = vreinterpretq_s32_u64(vmovl_u32(vld1_u32(w))); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
+    int32x4_t vacc0x67 = vreinterpretq_s32_u64(vmovl_u32(vld1_u32(w))); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
 
     size_t p = ks;
     do {

diff --git a/src/qs8-igemm/gen/1x8c4-minmax-rndnu-neon-mlal-ld2r.c b/src/qs8-igemm/gen/1x8c4-minmax-rndnu-neon-mlal-ld2r.c
index 02ae30e..dbd6b27 100644
--- a/src/qs8-igemm/gen/1x8c4-minmax-rndnu-neon-mlal-ld2r.c
+++ b/src/qs8-igemm/gen/1x8c4-minmax-rndnu-neon-mlal-ld2r.c

@@ -44,10 +44,10 @@
   int8_t* c0 = c;
 
   do {
-    int32x4_t vacc0x01 = vreinterpretq_s32_u64(vshll_n_u32(vld1_u32(w), 0)); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
-    int32x4_t vacc0x23 = vreinterpretq_s32_u64(vshll_n_u32(vld1_u32(w), 0)); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
-    int32x4_t vacc0x45 = vreinterpretq_s32_u64(vshll_n_u32(vld1_u32(w), 0)); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
-    int32x4_t vacc0x67 = vreinterpretq_s32_u64(vshll_n_u32(vld1_u32(w), 0)); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
+    int32x4_t vacc0x01 = vreinterpretq_s32_u64(vmovl_u32(vld1_u32(w))); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
+    int32x4_t vacc0x23 = vreinterpretq_s32_u64(vmovl_u32(vld1_u32(w))); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
+    int32x4_t vacc0x45 = vreinterpretq_s32_u64(vmovl_u32(vld1_u32(w))); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
+    int32x4_t vacc0x67 = vreinterpretq_s32_u64(vmovl_u32(vld1_u32(w))); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
 
     size_t p = ks;
     do {

diff --git a/src/qs8-igemm/gen/1x8c4-minmax-rndnu-neon-mull-dup.c b/src/qs8-igemm/gen/1x8c4-minmax-rndnu-neon-mull-dup.c
index 52f0414..97d5a7b 100644
--- a/src/qs8-igemm/gen/1x8c4-minmax-rndnu-neon-mull-dup.c
+++ b/src/qs8-igemm/gen/1x8c4-minmax-rndnu-neon-mull-dup.c

@@ -44,10 +44,10 @@
   int8_t* c0 = c;
 
   do {
-    int32x4_t vacc0x01 = vreinterpretq_s32_u64(vshll_n_u32(vld1_u32(w), 0)); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
-    int32x4_t vacc0x23 = vreinterpretq_s32_u64(vshll_n_u32(vld1_u32(w), 0)); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
-    int32x4_t vacc0x45 = vreinterpretq_s32_u64(vshll_n_u32(vld1_u32(w), 0)); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
-    int32x4_t vacc0x67 = vreinterpretq_s32_u64(vshll_n_u32(vld1_u32(w), 0)); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
+    int32x4_t vacc0x01 = vreinterpretq_s32_u64(vmovl_u32(vld1_u32(w))); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
+    int32x4_t vacc0x23 = vreinterpretq_s32_u64(vmovl_u32(vld1_u32(w))); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
+    int32x4_t vacc0x45 = vreinterpretq_s32_u64(vmovl_u32(vld1_u32(w))); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
+    int32x4_t vacc0x67 = vreinterpretq_s32_u64(vmovl_u32(vld1_u32(w))); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
 
     size_t p = ks;
     do {

diff --git a/src/qs8-igemm/gen/1x8c4-minmax-rndnu-neon-mull-ld1r.c b/src/qs8-igemm/gen/1x8c4-minmax-rndnu-neon-mull-ld1r.c
index 78fdf21..fe0261b 100644
--- a/src/qs8-igemm/gen/1x8c4-minmax-rndnu-neon-mull-ld1r.c
+++ b/src/qs8-igemm/gen/1x8c4-minmax-rndnu-neon-mull-ld1r.c

@@ -44,10 +44,10 @@
   int8_t* c0 = c;
 
   do {
-    int32x4_t vacc0x01 = vreinterpretq_s32_u64(vshll_n_u32(vld1_u32(w), 0)); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
-    int32x4_t vacc0x23 = vreinterpretq_s32_u64(vshll_n_u32(vld1_u32(w), 0)); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
-    int32x4_t vacc0x45 = vreinterpretq_s32_u64(vshll_n_u32(vld1_u32(w), 0)); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
-    int32x4_t vacc0x67 = vreinterpretq_s32_u64(vshll_n_u32(vld1_u32(w), 0)); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
+    int32x4_t vacc0x01 = vreinterpretq_s32_u64(vmovl_u32(vld1_u32(w))); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
+    int32x4_t vacc0x23 = vreinterpretq_s32_u64(vmovl_u32(vld1_u32(w))); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
+    int32x4_t vacc0x45 = vreinterpretq_s32_u64(vmovl_u32(vld1_u32(w))); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
+    int32x4_t vacc0x67 = vreinterpretq_s32_u64(vmovl_u32(vld1_u32(w))); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
 
     size_t p = ks;
     do {

diff --git a/src/qs8-igemm/gen/1x8c4-minmax-rndnu-neon-mull-ld2r.c b/src/qs8-igemm/gen/1x8c4-minmax-rndnu-neon-mull-ld2r.c
index c062468..d4e095c 100644
--- a/src/qs8-igemm/gen/1x8c4-minmax-rndnu-neon-mull-ld2r.c
+++ b/src/qs8-igemm/gen/1x8c4-minmax-rndnu-neon-mull-ld2r.c

@@ -44,10 +44,10 @@
   int8_t* c0 = c;
 
   do {
-    int32x4_t vacc0x01 = vreinterpretq_s32_u64(vshll_n_u32(vld1_u32(w), 0)); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
-    int32x4_t vacc0x23 = vreinterpretq_s32_u64(vshll_n_u32(vld1_u32(w), 0)); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
-    int32x4_t vacc0x45 = vreinterpretq_s32_u64(vshll_n_u32(vld1_u32(w), 0)); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
-    int32x4_t vacc0x67 = vreinterpretq_s32_u64(vshll_n_u32(vld1_u32(w), 0)); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
+    int32x4_t vacc0x01 = vreinterpretq_s32_u64(vmovl_u32(vld1_u32(w))); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
+    int32x4_t vacc0x23 = vreinterpretq_s32_u64(vmovl_u32(vld1_u32(w))); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
+    int32x4_t vacc0x45 = vreinterpretq_s32_u64(vmovl_u32(vld1_u32(w))); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
+    int32x4_t vacc0x67 = vreinterpretq_s32_u64(vmovl_u32(vld1_u32(w))); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
 
     size_t p = ks;
     do {

diff --git a/src/qs8-igemm/gen/1x8c4s2-minmax-fp32-neon-mlal.c b/src/qs8-igemm/gen/1x8c4s2-minmax-fp32-neon-mlal.c
index a21822b..2a4df35 100644
--- a/src/qs8-igemm/gen/1x8c4s2-minmax-fp32-neon-mlal.c
+++ b/src/qs8-igemm/gen/1x8c4s2-minmax-fp32-neon-mlal.c

@@ -43,10 +43,10 @@
   int8_t* c0 = c;
 
   do {
-    int32x4_t vacc0x01 = vreinterpretq_s32_u64(vshll_n_u32(vld1_u32(w), 0)); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
-    int32x4_t vacc0x23 = vreinterpretq_s32_u64(vshll_n_u32(vld1_u32(w), 0)); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
-    int32x4_t vacc0x45 = vreinterpretq_s32_u64(vshll_n_u32(vld1_u32(w), 0)); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
-    int32x4_t vacc0x67 = vreinterpretq_s32_u64(vshll_n_u32(vld1_u32(w), 0)); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
+    int32x4_t vacc0x01 = vreinterpretq_s32_u64(vmovl_u32(vld1_u32(w))); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
+    int32x4_t vacc0x23 = vreinterpretq_s32_u64(vmovl_u32(vld1_u32(w))); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
+    int32x4_t vacc0x45 = vreinterpretq_s32_u64(vmovl_u32(vld1_u32(w))); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
+    int32x4_t vacc0x67 = vreinterpretq_s32_u64(vmovl_u32(vld1_u32(w))); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
 
     size_t p = ks;
     do {

diff --git a/src/qs8-igemm/gen/1x8c4s2-minmax-fp32-neonv8-mlal.c b/src/qs8-igemm/gen/1x8c4s2-minmax-fp32-neonv8-mlal.c
index 71c5e2f..a20261b 100644
--- a/src/qs8-igemm/gen/1x8c4s2-minmax-fp32-neonv8-mlal.c
+++ b/src/qs8-igemm/gen/1x8c4s2-minmax-fp32-neonv8-mlal.c

@@ -44,10 +44,10 @@
   int8_t* c0 = c;
 
   do {
-    int32x4_t vacc0x01 = vreinterpretq_s32_u64(vshll_n_u32(vld1_u32(w), 0)); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
-    int32x4_t vacc0x23 = vreinterpretq_s32_u64(vshll_n_u32(vld1_u32(w), 0)); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
-    int32x4_t vacc0x45 = vreinterpretq_s32_u64(vshll_n_u32(vld1_u32(w), 0)); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
-    int32x4_t vacc0x67 = vreinterpretq_s32_u64(vshll_n_u32(vld1_u32(w), 0)); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
+    int32x4_t vacc0x01 = vreinterpretq_s32_u64(vmovl_u32(vld1_u32(w))); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
+    int32x4_t vacc0x23 = vreinterpretq_s32_u64(vmovl_u32(vld1_u32(w))); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
+    int32x4_t vacc0x45 = vreinterpretq_s32_u64(vmovl_u32(vld1_u32(w))); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
+    int32x4_t vacc0x67 = vreinterpretq_s32_u64(vmovl_u32(vld1_u32(w))); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
 
     size_t p = ks;
     do {

diff --git a/src/qs8-igemm/gen/1x8c4s2-minmax-rndnu-neon-mlal.c b/src/qs8-igemm/gen/1x8c4s2-minmax-rndnu-neon-mlal.c
index b03182f..889897d 100644
--- a/src/qs8-igemm/gen/1x8c4s2-minmax-rndnu-neon-mlal.c
+++ b/src/qs8-igemm/gen/1x8c4s2-minmax-rndnu-neon-mlal.c

@@ -43,10 +43,10 @@
   int8_t* c0 = c;
 
   do {
-    int32x4_t vacc0x01 = vreinterpretq_s32_u64(vshll_n_u32(vld1_u32(w), 0)); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
-    int32x4_t vacc0x23 = vreinterpretq_s32_u64(vshll_n_u32(vld1_u32(w), 0)); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
-    int32x4_t vacc0x45 = vreinterpretq_s32_u64(vshll_n_u32(vld1_u32(w), 0)); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
-    int32x4_t vacc0x67 = vreinterpretq_s32_u64(vshll_n_u32(vld1_u32(w), 0)); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
+    int32x4_t vacc0x01 = vreinterpretq_s32_u64(vmovl_u32(vld1_u32(w))); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
+    int32x4_t vacc0x23 = vreinterpretq_s32_u64(vmovl_u32(vld1_u32(w))); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
+    int32x4_t vacc0x45 = vreinterpretq_s32_u64(vmovl_u32(vld1_u32(w))); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
+    int32x4_t vacc0x67 = vreinterpretq_s32_u64(vmovl_u32(vld1_u32(w))); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
 
     size_t p = ks;
     do {

diff --git a/src/qs8-igemm/gen/1x8c4s2-minmax-rndnu-neon-mull.c b/src/qs8-igemm/gen/1x8c4s2-minmax-rndnu-neon-mull.c
index 5cd1d2b..abc493c 100644
--- a/src/qs8-igemm/gen/1x8c4s2-minmax-rndnu-neon-mull.c
+++ b/src/qs8-igemm/gen/1x8c4s2-minmax-rndnu-neon-mull.c

@@ -43,10 +43,10 @@
   int8_t* c0 = c;
 
   do {
-    int32x4_t vacc0x01 = vreinterpretq_s32_u64(vshll_n_u32(vld1_u32(w), 0)); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
-    int32x4_t vacc0x23 = vreinterpretq_s32_u64(vshll_n_u32(vld1_u32(w), 0)); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
-    int32x4_t vacc0x45 = vreinterpretq_s32_u64(vshll_n_u32(vld1_u32(w), 0)); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
-    int32x4_t vacc0x67 = vreinterpretq_s32_u64(vshll_n_u32(vld1_u32(w), 0)); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
+    int32x4_t vacc0x01 = vreinterpretq_s32_u64(vmovl_u32(vld1_u32(w))); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
+    int32x4_t vacc0x23 = vreinterpretq_s32_u64(vmovl_u32(vld1_u32(w))); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
+    int32x4_t vacc0x45 = vreinterpretq_s32_u64(vmovl_u32(vld1_u32(w))); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
+    int32x4_t vacc0x67 = vreinterpretq_s32_u64(vmovl_u32(vld1_u32(w))); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
 
     size_t p = ks;
     do {

diff --git a/src/qs8-igemm/gen/2x16c4-minmax-rndnu-neon-mlal-dup.c b/src/qs8-igemm/gen/2x16c4-minmax-rndnu-neon-mlal-dup.c
index 9295202..898e81f 100644
--- a/src/qs8-igemm/gen/2x16c4-minmax-rndnu-neon-mlal-dup.c
+++ b/src/qs8-igemm/gen/2x16c4-minmax-rndnu-neon-mlal-dup.c

@@ -48,14 +48,14 @@
   }
 
   do {
-    int32x4_t vacc0x01 = vreinterpretq_s32_u64(vshll_n_u32(vld1_u32(w), 0)); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
-    int32x4_t vacc0x23 = vreinterpretq_s32_u64(vshll_n_u32(vld1_u32(w), 0)); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
-    int32x4_t vacc0x45 = vreinterpretq_s32_u64(vshll_n_u32(vld1_u32(w), 0)); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
-    int32x4_t vacc0x67 = vreinterpretq_s32_u64(vshll_n_u32(vld1_u32(w), 0)); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
-    int32x4_t vacc0x89 = vreinterpretq_s32_u64(vshll_n_u32(vld1_u32(w), 0)); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
-    int32x4_t vacc0xAB = vreinterpretq_s32_u64(vshll_n_u32(vld1_u32(w), 0)); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
-    int32x4_t vacc0xCD = vreinterpretq_s32_u64(vshll_n_u32(vld1_u32(w), 0)); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
-    int32x4_t vacc0xEF = vreinterpretq_s32_u64(vshll_n_u32(vld1_u32(w), 0)); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
+    int32x4_t vacc0x01 = vreinterpretq_s32_u64(vmovl_u32(vld1_u32(w))); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
+    int32x4_t vacc0x23 = vreinterpretq_s32_u64(vmovl_u32(vld1_u32(w))); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
+    int32x4_t vacc0x45 = vreinterpretq_s32_u64(vmovl_u32(vld1_u32(w))); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
+    int32x4_t vacc0x67 = vreinterpretq_s32_u64(vmovl_u32(vld1_u32(w))); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
+    int32x4_t vacc0x89 = vreinterpretq_s32_u64(vmovl_u32(vld1_u32(w))); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
+    int32x4_t vacc0xAB = vreinterpretq_s32_u64(vmovl_u32(vld1_u32(w))); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
+    int32x4_t vacc0xCD = vreinterpretq_s32_u64(vmovl_u32(vld1_u32(w))); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
+    int32x4_t vacc0xEF = vreinterpretq_s32_u64(vmovl_u32(vld1_u32(w))); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
     int32x4_t vacc1x01 = vacc0x01;
     int32x4_t vacc1x23 = vacc0x23;
     int32x4_t vacc1x45 = vacc0x45;

diff --git a/src/qs8-igemm/gen/2x16c4-minmax-rndnu-neon-mlal-ld1r.c b/src/qs8-igemm/gen/2x16c4-minmax-rndnu-neon-mlal-ld1r.c
index 8f62977..742a8f9 100644
--- a/src/qs8-igemm/gen/2x16c4-minmax-rndnu-neon-mlal-ld1r.c
+++ b/src/qs8-igemm/gen/2x16c4-minmax-rndnu-neon-mlal-ld1r.c

@@ -48,14 +48,14 @@
   }
 
   do {
-    int32x4_t vacc0x01 = vreinterpretq_s32_u64(vshll_n_u32(vld1_u32(w), 0)); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
-    int32x4_t vacc0x23 = vreinterpretq_s32_u64(vshll_n_u32(vld1_u32(w), 0)); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
-    int32x4_t vacc0x45 = vreinterpretq_s32_u64(vshll_n_u32(vld1_u32(w), 0)); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
-    int32x4_t vacc0x67 = vreinterpretq_s32_u64(vshll_n_u32(vld1_u32(w), 0)); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
-    int32x4_t vacc0x89 = vreinterpretq_s32_u64(vshll_n_u32(vld1_u32(w), 0)); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
-    int32x4_t vacc0xAB = vreinterpretq_s32_u64(vshll_n_u32(vld1_u32(w), 0)); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
-    int32x4_t vacc0xCD = vreinterpretq_s32_u64(vshll_n_u32(vld1_u32(w), 0)); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
-    int32x4_t vacc0xEF = vreinterpretq_s32_u64(vshll_n_u32(vld1_u32(w), 0)); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
+    int32x4_t vacc0x01 = vreinterpretq_s32_u64(vmovl_u32(vld1_u32(w))); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
+    int32x4_t vacc0x23 = vreinterpretq_s32_u64(vmovl_u32(vld1_u32(w))); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
+    int32x4_t vacc0x45 = vreinterpretq_s32_u64(vmovl_u32(vld1_u32(w))); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
+    int32x4_t vacc0x67 = vreinterpretq_s32_u64(vmovl_u32(vld1_u32(w))); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
+    int32x4_t vacc0x89 = vreinterpretq_s32_u64(vmovl_u32(vld1_u32(w))); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
+    int32x4_t vacc0xAB = vreinterpretq_s32_u64(vmovl_u32(vld1_u32(w))); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
+    int32x4_t vacc0xCD = vreinterpretq_s32_u64(vmovl_u32(vld1_u32(w))); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
+    int32x4_t vacc0xEF = vreinterpretq_s32_u64(vmovl_u32(vld1_u32(w))); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
     int32x4_t vacc1x01 = vacc0x01;
     int32x4_t vacc1x23 = vacc0x23;
     int32x4_t vacc1x45 = vacc0x45;

diff --git a/src/qs8-igemm/gen/2x16c4-minmax-rndnu-neon-mlal-ld2r.c b/src/qs8-igemm/gen/2x16c4-minmax-rndnu-neon-mlal-ld2r.c
index 81e879c..3998485 100644
--- a/src/qs8-igemm/gen/2x16c4-minmax-rndnu-neon-mlal-ld2r.c
+++ b/src/qs8-igemm/gen/2x16c4-minmax-rndnu-neon-mlal-ld2r.c

@@ -48,14 +48,14 @@
   }
 
   do {
-    int32x4_t vacc0x01 = vreinterpretq_s32_u64(vshll_n_u32(vld1_u32(w), 0)); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
-    int32x4_t vacc0x23 = vreinterpretq_s32_u64(vshll_n_u32(vld1_u32(w), 0)); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
-    int32x4_t vacc0x45 = vreinterpretq_s32_u64(vshll_n_u32(vld1_u32(w), 0)); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
-    int32x4_t vacc0x67 = vreinterpretq_s32_u64(vshll_n_u32(vld1_u32(w), 0)); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
-    int32x4_t vacc0x89 = vreinterpretq_s32_u64(vshll_n_u32(vld1_u32(w), 0)); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
-    int32x4_t vacc0xAB = vreinterpretq_s32_u64(vshll_n_u32(vld1_u32(w), 0)); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
-    int32x4_t vacc0xCD = vreinterpretq_s32_u64(vshll_n_u32(vld1_u32(w), 0)); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
-    int32x4_t vacc0xEF = vreinterpretq_s32_u64(vshll_n_u32(vld1_u32(w), 0)); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
+    int32x4_t vacc0x01 = vreinterpretq_s32_u64(vmovl_u32(vld1_u32(w))); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
+    int32x4_t vacc0x23 = vreinterpretq_s32_u64(vmovl_u32(vld1_u32(w))); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
+    int32x4_t vacc0x45 = vreinterpretq_s32_u64(vmovl_u32(vld1_u32(w))); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
+    int32x4_t vacc0x67 = vreinterpretq_s32_u64(vmovl_u32(vld1_u32(w))); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
+    int32x4_t vacc0x89 = vreinterpretq_s32_u64(vmovl_u32(vld1_u32(w))); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
+    int32x4_t vacc0xAB = vreinterpretq_s32_u64(vmovl_u32(vld1_u32(w))); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
+    int32x4_t vacc0xCD = vreinterpretq_s32_u64(vmovl_u32(vld1_u32(w))); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
+    int32x4_t vacc0xEF = vreinterpretq_s32_u64(vmovl_u32(vld1_u32(w))); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
     int32x4_t vacc1x01 = vacc0x01;
     int32x4_t vacc1x23 = vacc0x23;
     int32x4_t vacc1x45 = vacc0x45;

diff --git a/src/qs8-igemm/gen/2x16c4-minmax-rndnu-neon-mull-dup.c b/src/qs8-igemm/gen/2x16c4-minmax-rndnu-neon-mull-dup.c
index 2f0b9ef..e2368ae 100644
--- a/src/qs8-igemm/gen/2x16c4-minmax-rndnu-neon-mull-dup.c
+++ b/src/qs8-igemm/gen/2x16c4-minmax-rndnu-neon-mull-dup.c

@@ -48,14 +48,14 @@
   }
 
   do {
-    int32x4_t vacc0x01 = vreinterpretq_s32_u64(vshll_n_u32(vld1_u32(w), 0)); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
-    int32x4_t vacc0x23 = vreinterpretq_s32_u64(vshll_n_u32(vld1_u32(w), 0)); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
-    int32x4_t vacc0x45 = vreinterpretq_s32_u64(vshll_n_u32(vld1_u32(w), 0)); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
-    int32x4_t vacc0x67 = vreinterpretq_s32_u64(vshll_n_u32(vld1_u32(w), 0)); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
-    int32x4_t vacc0x89 = vreinterpretq_s32_u64(vshll_n_u32(vld1_u32(w), 0)); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
-    int32x4_t vacc0xAB = vreinterpretq_s32_u64(vshll_n_u32(vld1_u32(w), 0)); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
-    int32x4_t vacc0xCD = vreinterpretq_s32_u64(vshll_n_u32(vld1_u32(w), 0)); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
-    int32x4_t vacc0xEF = vreinterpretq_s32_u64(vshll_n_u32(vld1_u32(w), 0)); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
+    int32x4_t vacc0x01 = vreinterpretq_s32_u64(vmovl_u32(vld1_u32(w))); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
+    int32x4_t vacc0x23 = vreinterpretq_s32_u64(vmovl_u32(vld1_u32(w))); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
+    int32x4_t vacc0x45 = vreinterpretq_s32_u64(vmovl_u32(vld1_u32(w))); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
+    int32x4_t vacc0x67 = vreinterpretq_s32_u64(vmovl_u32(vld1_u32(w))); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
+    int32x4_t vacc0x89 = vreinterpretq_s32_u64(vmovl_u32(vld1_u32(w))); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
+    int32x4_t vacc0xAB = vreinterpretq_s32_u64(vmovl_u32(vld1_u32(w))); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
+    int32x4_t vacc0xCD = vreinterpretq_s32_u64(vmovl_u32(vld1_u32(w))); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
+    int32x4_t vacc0xEF = vreinterpretq_s32_u64(vmovl_u32(vld1_u32(w))); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
     int32x4_t vacc1x01 = vacc0x01;
     int32x4_t vacc1x23 = vacc0x23;
     int32x4_t vacc1x45 = vacc0x45;

diff --git a/src/qs8-igemm/gen/2x16c4-minmax-rndnu-neon-mull-ld1r.c b/src/qs8-igemm/gen/2x16c4-minmax-rndnu-neon-mull-ld1r.c
index f73d7f0..816cf72 100644
--- a/src/qs8-igemm/gen/2x16c4-minmax-rndnu-neon-mull-ld1r.c
+++ b/src/qs8-igemm/gen/2x16c4-minmax-rndnu-neon-mull-ld1r.c

@@ -48,14 +48,14 @@
   }
 
   do {
-    int32x4_t vacc0x01 = vreinterpretq_s32_u64(vshll_n_u32(vld1_u32(w), 0)); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
-    int32x4_t vacc0x23 = vreinterpretq_s32_u64(vshll_n_u32(vld1_u32(w), 0)); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
-    int32x4_t vacc0x45 = vreinterpretq_s32_u64(vshll_n_u32(vld1_u32(w), 0)); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
-    int32x4_t vacc0x67 = vreinterpretq_s32_u64(vshll_n_u32(vld1_u32(w), 0)); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
-    int32x4_t vacc0x89 = vreinterpretq_s32_u64(vshll_n_u32(vld1_u32(w), 0)); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
-    int32x4_t vacc0xAB = vreinterpretq_s32_u64(vshll_n_u32(vld1_u32(w), 0)); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
-    int32x4_t vacc0xCD = vreinterpretq_s32_u64(vshll_n_u32(vld1_u32(w), 0)); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
-    int32x4_t vacc0xEF = vreinterpretq_s32_u64(vshll_n_u32(vld1_u32(w), 0)); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
+    int32x4_t vacc0x01 = vreinterpretq_s32_u64(vmovl_u32(vld1_u32(w))); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
+    int32x4_t vacc0x23 = vreinterpretq_s32_u64(vmovl_u32(vld1_u32(w))); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
+    int32x4_t vacc0x45 = vreinterpretq_s32_u64(vmovl_u32(vld1_u32(w))); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
+    int32x4_t vacc0x67 = vreinterpretq_s32_u64(vmovl_u32(vld1_u32(w))); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
+    int32x4_t vacc0x89 = vreinterpretq_s32_u64(vmovl_u32(vld1_u32(w))); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
+    int32x4_t vacc0xAB = vreinterpretq_s32_u64(vmovl_u32(vld1_u32(w))); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
+    int32x4_t vacc0xCD = vreinterpretq_s32_u64(vmovl_u32(vld1_u32(w))); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
+    int32x4_t vacc0xEF = vreinterpretq_s32_u64(vmovl_u32(vld1_u32(w))); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
     int32x4_t vacc1x01 = vacc0x01;
     int32x4_t vacc1x23 = vacc0x23;
     int32x4_t vacc1x45 = vacc0x45;

diff --git a/src/qs8-igemm/gen/2x16c4-minmax-rndnu-neon-mull-ld2r.c b/src/qs8-igemm/gen/2x16c4-minmax-rndnu-neon-mull-ld2r.c
index ad8b335..6afaa7c 100644
--- a/src/qs8-igemm/gen/2x16c4-minmax-rndnu-neon-mull-ld2r.c
+++ b/src/qs8-igemm/gen/2x16c4-minmax-rndnu-neon-mull-ld2r.c

@@ -48,14 +48,14 @@
   }
 
   do {
-    int32x4_t vacc0x01 = vreinterpretq_s32_u64(vshll_n_u32(vld1_u32(w), 0)); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
-    int32x4_t vacc0x23 = vreinterpretq_s32_u64(vshll_n_u32(vld1_u32(w), 0)); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
-    int32x4_t vacc0x45 = vreinterpretq_s32_u64(vshll_n_u32(vld1_u32(w), 0)); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
-    int32x4_t vacc0x67 = vreinterpretq_s32_u64(vshll_n_u32(vld1_u32(w), 0)); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
-    int32x4_t vacc0x89 = vreinterpretq_s32_u64(vshll_n_u32(vld1_u32(w), 0)); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
-    int32x4_t vacc0xAB = vreinterpretq_s32_u64(vshll_n_u32(vld1_u32(w), 0)); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
-    int32x4_t vacc0xCD = vreinterpretq_s32_u64(vshll_n_u32(vld1_u32(w), 0)); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
-    int32x4_t vacc0xEF = vreinterpretq_s32_u64(vshll_n_u32(vld1_u32(w), 0)); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
+    int32x4_t vacc0x01 = vreinterpretq_s32_u64(vmovl_u32(vld1_u32(w))); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
+    int32x4_t vacc0x23 = vreinterpretq_s32_u64(vmovl_u32(vld1_u32(w))); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
+    int32x4_t vacc0x45 = vreinterpretq_s32_u64(vmovl_u32(vld1_u32(w))); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
+    int32x4_t vacc0x67 = vreinterpretq_s32_u64(vmovl_u32(vld1_u32(w))); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
+    int32x4_t vacc0x89 = vreinterpretq_s32_u64(vmovl_u32(vld1_u32(w))); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
+    int32x4_t vacc0xAB = vreinterpretq_s32_u64(vmovl_u32(vld1_u32(w))); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
+    int32x4_t vacc0xCD = vreinterpretq_s32_u64(vmovl_u32(vld1_u32(w))); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
+    int32x4_t vacc0xEF = vreinterpretq_s32_u64(vmovl_u32(vld1_u32(w))); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
     int32x4_t vacc1x01 = vacc0x01;
     int32x4_t vacc1x23 = vacc0x23;
     int32x4_t vacc1x45 = vacc0x45;

diff --git a/src/qs8-igemm/gen/2x16c4s2-minmax-rndnu-neon-mlal.c b/src/qs8-igemm/gen/2x16c4s2-minmax-rndnu-neon-mlal.c
index 6a646ec..ca82e44 100644
--- a/src/qs8-igemm/gen/2x16c4s2-minmax-rndnu-neon-mlal.c
+++ b/src/qs8-igemm/gen/2x16c4s2-minmax-rndnu-neon-mlal.c

@@ -47,14 +47,14 @@
   }
 
   do {
-    int32x4_t vacc0x01 = vreinterpretq_s32_u64(vshll_n_u32(vld1_u32(w), 0)); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
-    int32x4_t vacc0x23 = vreinterpretq_s32_u64(vshll_n_u32(vld1_u32(w), 0)); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
-    int32x4_t vacc0x45 = vreinterpretq_s32_u64(vshll_n_u32(vld1_u32(w), 0)); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
-    int32x4_t vacc0x67 = vreinterpretq_s32_u64(vshll_n_u32(vld1_u32(w), 0)); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
-    int32x4_t vacc0x89 = vreinterpretq_s32_u64(vshll_n_u32(vld1_u32(w), 0)); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
-    int32x4_t vacc0xAB = vreinterpretq_s32_u64(vshll_n_u32(vld1_u32(w), 0)); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
-    int32x4_t vacc0xCD = vreinterpretq_s32_u64(vshll_n_u32(vld1_u32(w), 0)); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
-    int32x4_t vacc0xEF = vreinterpretq_s32_u64(vshll_n_u32(vld1_u32(w), 0)); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
+    int32x4_t vacc0x01 = vreinterpretq_s32_u64(vmovl_u32(vld1_u32(w))); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
+    int32x4_t vacc0x23 = vreinterpretq_s32_u64(vmovl_u32(vld1_u32(w))); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
+    int32x4_t vacc0x45 = vreinterpretq_s32_u64(vmovl_u32(vld1_u32(w))); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
+    int32x4_t vacc0x67 = vreinterpretq_s32_u64(vmovl_u32(vld1_u32(w))); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
+    int32x4_t vacc0x89 = vreinterpretq_s32_u64(vmovl_u32(vld1_u32(w))); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
+    int32x4_t vacc0xAB = vreinterpretq_s32_u64(vmovl_u32(vld1_u32(w))); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
+    int32x4_t vacc0xCD = vreinterpretq_s32_u64(vmovl_u32(vld1_u32(w))); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
+    int32x4_t vacc0xEF = vreinterpretq_s32_u64(vmovl_u32(vld1_u32(w))); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
     int32x4_t vacc1x01 = vacc0x01;
     int32x4_t vacc1x23 = vacc0x23;
     int32x4_t vacc1x45 = vacc0x45;

diff --git a/src/qs8-igemm/gen/2x16c4s2-minmax-rndnu-neon-mull.c b/src/qs8-igemm/gen/2x16c4s2-minmax-rndnu-neon-mull.c
index 0313a55..7e0884f 100644
--- a/src/qs8-igemm/gen/2x16c4s2-minmax-rndnu-neon-mull.c
+++ b/src/qs8-igemm/gen/2x16c4s2-minmax-rndnu-neon-mull.c

@@ -47,14 +47,14 @@
   }
 
   do {
-    int32x4_t vacc0x01 = vreinterpretq_s32_u64(vshll_n_u32(vld1_u32(w), 0)); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
-    int32x4_t vacc0x23 = vreinterpretq_s32_u64(vshll_n_u32(vld1_u32(w), 0)); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
-    int32x4_t vacc0x45 = vreinterpretq_s32_u64(vshll_n_u32(vld1_u32(w), 0)); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
-    int32x4_t vacc0x67 = vreinterpretq_s32_u64(vshll_n_u32(vld1_u32(w), 0)); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
-    int32x4_t vacc0x89 = vreinterpretq_s32_u64(vshll_n_u32(vld1_u32(w), 0)); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
-    int32x4_t vacc0xAB = vreinterpretq_s32_u64(vshll_n_u32(vld1_u32(w), 0)); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
-    int32x4_t vacc0xCD = vreinterpretq_s32_u64(vshll_n_u32(vld1_u32(w), 0)); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
-    int32x4_t vacc0xEF = vreinterpretq_s32_u64(vshll_n_u32(vld1_u32(w), 0)); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
+    int32x4_t vacc0x01 = vreinterpretq_s32_u64(vmovl_u32(vld1_u32(w))); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
+    int32x4_t vacc0x23 = vreinterpretq_s32_u64(vmovl_u32(vld1_u32(w))); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
+    int32x4_t vacc0x45 = vreinterpretq_s32_u64(vmovl_u32(vld1_u32(w))); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
+    int32x4_t vacc0x67 = vreinterpretq_s32_u64(vmovl_u32(vld1_u32(w))); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
+    int32x4_t vacc0x89 = vreinterpretq_s32_u64(vmovl_u32(vld1_u32(w))); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
+    int32x4_t vacc0xAB = vreinterpretq_s32_u64(vmovl_u32(vld1_u32(w))); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
+    int32x4_t vacc0xCD = vreinterpretq_s32_u64(vmovl_u32(vld1_u32(w))); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
+    int32x4_t vacc0xEF = vreinterpretq_s32_u64(vmovl_u32(vld1_u32(w))); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
     int32x4_t vacc1x01 = vacc0x01;
     int32x4_t vacc1x23 = vacc0x23;
     int32x4_t vacc1x45 = vacc0x45;

diff --git a/src/qs8-igemm/gen/2x8c4-minmax-fp32-neon-mlal-dup.c b/src/qs8-igemm/gen/2x8c4-minmax-fp32-neon-mlal-dup.c
index f79075c..91fa6d2 100644
--- a/src/qs8-igemm/gen/2x8c4-minmax-fp32-neon-mlal-dup.c
+++ b/src/qs8-igemm/gen/2x8c4-minmax-fp32-neon-mlal-dup.c

@@ -48,10 +48,10 @@
   }
 
   do {
-    int32x4_t vacc0x01 = vreinterpretq_s32_u64(vshll_n_u32(vld1_u32(w), 0)); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
-    int32x4_t vacc0x23 = vreinterpretq_s32_u64(vshll_n_u32(vld1_u32(w), 0)); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
-    int32x4_t vacc0x45 = vreinterpretq_s32_u64(vshll_n_u32(vld1_u32(w), 0)); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
-    int32x4_t vacc0x67 = vreinterpretq_s32_u64(vshll_n_u32(vld1_u32(w), 0)); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
+    int32x4_t vacc0x01 = vreinterpretq_s32_u64(vmovl_u32(vld1_u32(w))); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
+    int32x4_t vacc0x23 = vreinterpretq_s32_u64(vmovl_u32(vld1_u32(w))); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
+    int32x4_t vacc0x45 = vreinterpretq_s32_u64(vmovl_u32(vld1_u32(w))); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
+    int32x4_t vacc0x67 = vreinterpretq_s32_u64(vmovl_u32(vld1_u32(w))); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
     int32x4_t vacc1x01 = vacc0x01;
     int32x4_t vacc1x23 = vacc0x23;
     int32x4_t vacc1x45 = vacc0x45;

diff --git a/src/qs8-igemm/gen/2x8c4-minmax-fp32-neon-mlal-ld1r.c b/src/qs8-igemm/gen/2x8c4-minmax-fp32-neon-mlal-ld1r.c
index 0d26915..caf6c7a 100644
--- a/src/qs8-igemm/gen/2x8c4-minmax-fp32-neon-mlal-ld1r.c
+++ b/src/qs8-igemm/gen/2x8c4-minmax-fp32-neon-mlal-ld1r.c

@@ -48,10 +48,10 @@
   }
 
   do {
-    int32x4_t vacc0x01 = vreinterpretq_s32_u64(vshll_n_u32(vld1_u32(w), 0)); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
-    int32x4_t vacc0x23 = vreinterpretq_s32_u64(vshll_n_u32(vld1_u32(w), 0)); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
-    int32x4_t vacc0x45 = vreinterpretq_s32_u64(vshll_n_u32(vld1_u32(w), 0)); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
-    int32x4_t vacc0x67 = vreinterpretq_s32_u64(vshll_n_u32(vld1_u32(w), 0)); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
+    int32x4_t vacc0x01 = vreinterpretq_s32_u64(vmovl_u32(vld1_u32(w))); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
+    int32x4_t vacc0x23 = vreinterpretq_s32_u64(vmovl_u32(vld1_u32(w))); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
+    int32x4_t vacc0x45 = vreinterpretq_s32_u64(vmovl_u32(vld1_u32(w))); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
+    int32x4_t vacc0x67 = vreinterpretq_s32_u64(vmovl_u32(vld1_u32(w))); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
     int32x4_t vacc1x01 = vacc0x01;
     int32x4_t vacc1x23 = vacc0x23;
     int32x4_t vacc1x45 = vacc0x45;

diff --git a/src/qs8-igemm/gen/2x8c4-minmax-fp32-neon-mlal-ld2r.c b/src/qs8-igemm/gen/2x8c4-minmax-fp32-neon-mlal-ld2r.c
index e3d8a44..ffdc1b7 100644
--- a/src/qs8-igemm/gen/2x8c4-minmax-fp32-neon-mlal-ld2r.c
+++ b/src/qs8-igemm/gen/2x8c4-minmax-fp32-neon-mlal-ld2r.c

@@ -48,10 +48,10 @@
   }
 
   do {
-    int32x4_t vacc0x01 = vreinterpretq_s32_u64(vshll_n_u32(vld1_u32(w), 0)); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
-    int32x4_t vacc0x23 = vreinterpretq_s32_u64(vshll_n_u32(vld1_u32(w), 0)); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
-    int32x4_t vacc0x45 = vreinterpretq_s32_u64(vshll_n_u32(vld1_u32(w), 0)); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
-    int32x4_t vacc0x67 = vreinterpretq_s32_u64(vshll_n_u32(vld1_u32(w), 0)); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
+    int32x4_t vacc0x01 = vreinterpretq_s32_u64(vmovl_u32(vld1_u32(w))); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
+    int32x4_t vacc0x23 = vreinterpretq_s32_u64(vmovl_u32(vld1_u32(w))); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
+    int32x4_t vacc0x45 = vreinterpretq_s32_u64(vmovl_u32(vld1_u32(w))); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
+    int32x4_t vacc0x67 = vreinterpretq_s32_u64(vmovl_u32(vld1_u32(w))); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
     int32x4_t vacc1x01 = vacc0x01;
     int32x4_t vacc1x23 = vacc0x23;
     int32x4_t vacc1x45 = vacc0x45;

diff --git a/src/qs8-igemm/gen/2x8c4-minmax-fp32-neonv8-mlal-dup.c b/src/qs8-igemm/gen/2x8c4-minmax-fp32-neonv8-mlal-dup.c
index 1148d93..dc6737e 100644
--- a/src/qs8-igemm/gen/2x8c4-minmax-fp32-neonv8-mlal-dup.c
+++ b/src/qs8-igemm/gen/2x8c4-minmax-fp32-neonv8-mlal-dup.c

@@ -49,10 +49,10 @@
   }
 
   do {
-    int32x4_t vacc0x01 = vreinterpretq_s32_u64(vshll_n_u32(vld1_u32(w), 0)); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
-    int32x4_t vacc0x23 = vreinterpretq_s32_u64(vshll_n_u32(vld1_u32(w), 0)); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
-    int32x4_t vacc0x45 = vreinterpretq_s32_u64(vshll_n_u32(vld1_u32(w), 0)); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
-    int32x4_t vacc0x67 = vreinterpretq_s32_u64(vshll_n_u32(vld1_u32(w), 0)); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
+    int32x4_t vacc0x01 = vreinterpretq_s32_u64(vmovl_u32(vld1_u32(w))); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
+    int32x4_t vacc0x23 = vreinterpretq_s32_u64(vmovl_u32(vld1_u32(w))); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
+    int32x4_t vacc0x45 = vreinterpretq_s32_u64(vmovl_u32(vld1_u32(w))); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
+    int32x4_t vacc0x67 = vreinterpretq_s32_u64(vmovl_u32(vld1_u32(w))); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
     int32x4_t vacc1x01 = vacc0x01;
     int32x4_t vacc1x23 = vacc0x23;
     int32x4_t vacc1x45 = vacc0x45;

diff --git a/src/qs8-igemm/gen/2x8c4-minmax-fp32-neonv8-mlal-ld1r.c b/src/qs8-igemm/gen/2x8c4-minmax-fp32-neonv8-mlal-ld1r.c
index ff6e11d..3a4fff1 100644
--- a/src/qs8-igemm/gen/2x8c4-minmax-fp32-neonv8-mlal-ld1r.c
+++ b/src/qs8-igemm/gen/2x8c4-minmax-fp32-neonv8-mlal-ld1r.c

@@ -49,10 +49,10 @@
   }
 
   do {
-    int32x4_t vacc0x01 = vreinterpretq_s32_u64(vshll_n_u32(vld1_u32(w), 0)); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
-    int32x4_t vacc0x23 = vreinterpretq_s32_u64(vshll_n_u32(vld1_u32(w), 0)); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
-    int32x4_t vacc0x45 = vreinterpretq_s32_u64(vshll_n_u32(vld1_u32(w), 0)); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
-    int32x4_t vacc0x67 = vreinterpretq_s32_u64(vshll_n_u32(vld1_u32(w), 0)); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
+    int32x4_t vacc0x01 = vreinterpretq_s32_u64(vmovl_u32(vld1_u32(w))); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
+    int32x4_t vacc0x23 = vreinterpretq_s32_u64(vmovl_u32(vld1_u32(w))); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
+    int32x4_t vacc0x45 = vreinterpretq_s32_u64(vmovl_u32(vld1_u32(w))); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
+    int32x4_t vacc0x67 = vreinterpretq_s32_u64(vmovl_u32(vld1_u32(w))); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
     int32x4_t vacc1x01 = vacc0x01;
     int32x4_t vacc1x23 = vacc0x23;
     int32x4_t vacc1x45 = vacc0x45;

diff --git a/src/qs8-igemm/gen/2x8c4-minmax-fp32-neonv8-mlal-ld2r.c b/src/qs8-igemm/gen/2x8c4-minmax-fp32-neonv8-mlal-ld2r.c
index b4fce53..0a3b61c 100644
--- a/src/qs8-igemm/gen/2x8c4-minmax-fp32-neonv8-mlal-ld2r.c
+++ b/src/qs8-igemm/gen/2x8c4-minmax-fp32-neonv8-mlal-ld2r.c

@@ -49,10 +49,10 @@
   }
 
   do {
-    int32x4_t vacc0x01 = vreinterpretq_s32_u64(vshll_n_u32(vld1_u32(w), 0)); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
-    int32x4_t vacc0x23 = vreinterpretq_s32_u64(vshll_n_u32(vld1_u32(w), 0)); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
-    int32x4_t vacc0x45 = vreinterpretq_s32_u64(vshll_n_u32(vld1_u32(w), 0)); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
-    int32x4_t vacc0x67 = vreinterpretq_s32_u64(vshll_n_u32(vld1_u32(w), 0)); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
+    int32x4_t vacc0x01 = vreinterpretq_s32_u64(vmovl_u32(vld1_u32(w))); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
+    int32x4_t vacc0x23 = vreinterpretq_s32_u64(vmovl_u32(vld1_u32(w))); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
+    int32x4_t vacc0x45 = vreinterpretq_s32_u64(vmovl_u32(vld1_u32(w))); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
+    int32x4_t vacc0x67 = vreinterpretq_s32_u64(vmovl_u32(vld1_u32(w))); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
     int32x4_t vacc1x01 = vacc0x01;
     int32x4_t vacc1x23 = vacc0x23;
     int32x4_t vacc1x45 = vacc0x45;

diff --git a/src/qs8-igemm/gen/2x8c4-minmax-rndnu-neon-mlal-dup.c b/src/qs8-igemm/gen/2x8c4-minmax-rndnu-neon-mlal-dup.c
index 97f9dbb..65b8676 100644
--- a/src/qs8-igemm/gen/2x8c4-minmax-rndnu-neon-mlal-dup.c
+++ b/src/qs8-igemm/gen/2x8c4-minmax-rndnu-neon-mlal-dup.c

@@ -48,10 +48,10 @@
   }
 
   do {
-    int32x4_t vacc0x01 = vreinterpretq_s32_u64(vshll_n_u32(vld1_u32(w), 0)); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
-    int32x4_t vacc0x23 = vreinterpretq_s32_u64(vshll_n_u32(vld1_u32(w), 0)); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
-    int32x4_t vacc0x45 = vreinterpretq_s32_u64(vshll_n_u32(vld1_u32(w), 0)); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
-    int32x4_t vacc0x67 = vreinterpretq_s32_u64(vshll_n_u32(vld1_u32(w), 0)); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
+    int32x4_t vacc0x01 = vreinterpretq_s32_u64(vmovl_u32(vld1_u32(w))); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
+    int32x4_t vacc0x23 = vreinterpretq_s32_u64(vmovl_u32(vld1_u32(w))); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
+    int32x4_t vacc0x45 = vreinterpretq_s32_u64(vmovl_u32(vld1_u32(w))); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
+    int32x4_t vacc0x67 = vreinterpretq_s32_u64(vmovl_u32(vld1_u32(w))); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
     int32x4_t vacc1x01 = vacc0x01;
     int32x4_t vacc1x23 = vacc0x23;
     int32x4_t vacc1x45 = vacc0x45;

diff --git a/src/qs8-igemm/gen/2x8c4-minmax-rndnu-neon-mlal-ld1r.c b/src/qs8-igemm/gen/2x8c4-minmax-rndnu-neon-mlal-ld1r.c
index f600412..ea747d4 100644
--- a/src/qs8-igemm/gen/2x8c4-minmax-rndnu-neon-mlal-ld1r.c
+++ b/src/qs8-igemm/gen/2x8c4-minmax-rndnu-neon-mlal-ld1r.c

@@ -48,10 +48,10 @@
   }
 
   do {
-    int32x4_t vacc0x01 = vreinterpretq_s32_u64(vshll_n_u32(vld1_u32(w), 0)); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
-    int32x4_t vacc0x23 = vreinterpretq_s32_u64(vshll_n_u32(vld1_u32(w), 0)); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
-    int32x4_t vacc0x45 = vreinterpretq_s32_u64(vshll_n_u32(vld1_u32(w), 0)); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
-    int32x4_t vacc0x67 = vreinterpretq_s32_u64(vshll_n_u32(vld1_u32(w), 0)); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
+    int32x4_t vacc0x01 = vreinterpretq_s32_u64(vmovl_u32(vld1_u32(w))); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
+    int32x4_t vacc0x23 = vreinterpretq_s32_u64(vmovl_u32(vld1_u32(w))); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
+    int32x4_t vacc0x45 = vreinterpretq_s32_u64(vmovl_u32(vld1_u32(w))); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
+    int32x4_t vacc0x67 = vreinterpretq_s32_u64(vmovl_u32(vld1_u32(w))); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
     int32x4_t vacc1x01 = vacc0x01;
     int32x4_t vacc1x23 = vacc0x23;
     int32x4_t vacc1x45 = vacc0x45;

diff --git a/src/qs8-igemm/gen/2x8c4-minmax-rndnu-neon-mlal-ld2r.c b/src/qs8-igemm/gen/2x8c4-minmax-rndnu-neon-mlal-ld2r.c
index 66e4338..acc6e27 100644
--- a/src/qs8-igemm/gen/2x8c4-minmax-rndnu-neon-mlal-ld2r.c
+++ b/src/qs8-igemm/gen/2x8c4-minmax-rndnu-neon-mlal-ld2r.c

@@ -48,10 +48,10 @@
   }
 
   do {
-    int32x4_t vacc0x01 = vreinterpretq_s32_u64(vshll_n_u32(vld1_u32(w), 0)); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
-    int32x4_t vacc0x23 = vreinterpretq_s32_u64(vshll_n_u32(vld1_u32(w), 0)); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
-    int32x4_t vacc0x45 = vreinterpretq_s32_u64(vshll_n_u32(vld1_u32(w), 0)); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
-    int32x4_t vacc0x67 = vreinterpretq_s32_u64(vshll_n_u32(vld1_u32(w), 0)); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
+    int32x4_t vacc0x01 = vreinterpretq_s32_u64(vmovl_u32(vld1_u32(w))); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
+    int32x4_t vacc0x23 = vreinterpretq_s32_u64(vmovl_u32(vld1_u32(w))); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
+    int32x4_t vacc0x45 = vreinterpretq_s32_u64(vmovl_u32(vld1_u32(w))); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
+    int32x4_t vacc0x67 = vreinterpretq_s32_u64(vmovl_u32(vld1_u32(w))); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
     int32x4_t vacc1x01 = vacc0x01;
     int32x4_t vacc1x23 = vacc0x23;
     int32x4_t vacc1x45 = vacc0x45;

diff --git a/src/qs8-igemm/gen/2x8c4-minmax-rndnu-neon-mull-dup.c b/src/qs8-igemm/gen/2x8c4-minmax-rndnu-neon-mull-dup.c
index ed80c26..ff5cb74 100644
--- a/src/qs8-igemm/gen/2x8c4-minmax-rndnu-neon-mull-dup.c
+++ b/src/qs8-igemm/gen/2x8c4-minmax-rndnu-neon-mull-dup.c

@@ -48,10 +48,10 @@
   }
 
   do {
-    int32x4_t vacc0x01 = vreinterpretq_s32_u64(vshll_n_u32(vld1_u32(w), 0)); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
-    int32x4_t vacc0x23 = vreinterpretq_s32_u64(vshll_n_u32(vld1_u32(w), 0)); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
-    int32x4_t vacc0x45 = vreinterpretq_s32_u64(vshll_n_u32(vld1_u32(w), 0)); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
-    int32x4_t vacc0x67 = vreinterpretq_s32_u64(vshll_n_u32(vld1_u32(w), 0)); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
+    int32x4_t vacc0x01 = vreinterpretq_s32_u64(vmovl_u32(vld1_u32(w))); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
+    int32x4_t vacc0x23 = vreinterpretq_s32_u64(vmovl_u32(vld1_u32(w))); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
+    int32x4_t vacc0x45 = vreinterpretq_s32_u64(vmovl_u32(vld1_u32(w))); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
+    int32x4_t vacc0x67 = vreinterpretq_s32_u64(vmovl_u32(vld1_u32(w))); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
     int32x4_t vacc1x01 = vacc0x01;
     int32x4_t vacc1x23 = vacc0x23;
     int32x4_t vacc1x45 = vacc0x45;

diff --git a/src/qs8-igemm/gen/2x8c4-minmax-rndnu-neon-mull-ld1r.c b/src/qs8-igemm/gen/2x8c4-minmax-rndnu-neon-mull-ld1r.c
index 28db090..557fd26 100644
--- a/src/qs8-igemm/gen/2x8c4-minmax-rndnu-neon-mull-ld1r.c
+++ b/src/qs8-igemm/gen/2x8c4-minmax-rndnu-neon-mull-ld1r.c

@@ -48,10 +48,10 @@
   }
 
   do {
-    int32x4_t vacc0x01 = vreinterpretq_s32_u64(vshll_n_u32(vld1_u32(w), 0)); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
-    int32x4_t vacc0x23 = vreinterpretq_s32_u64(vshll_n_u32(vld1_u32(w), 0)); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
-    int32x4_t vacc0x45 = vreinterpretq_s32_u64(vshll_n_u32(vld1_u32(w), 0)); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
-    int32x4_t vacc0x67 = vreinterpretq_s32_u64(vshll_n_u32(vld1_u32(w), 0)); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
+    int32x4_t vacc0x01 = vreinterpretq_s32_u64(vmovl_u32(vld1_u32(w))); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
+    int32x4_t vacc0x23 = vreinterpretq_s32_u64(vmovl_u32(vld1_u32(w))); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
+    int32x4_t vacc0x45 = vreinterpretq_s32_u64(vmovl_u32(vld1_u32(w))); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
+    int32x4_t vacc0x67 = vreinterpretq_s32_u64(vmovl_u32(vld1_u32(w))); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
     int32x4_t vacc1x01 = vacc0x01;
     int32x4_t vacc1x23 = vacc0x23;
     int32x4_t vacc1x45 = vacc0x45;

diff --git a/src/qs8-igemm/gen/2x8c4-minmax-rndnu-neon-mull-ld2r.c b/src/qs8-igemm/gen/2x8c4-minmax-rndnu-neon-mull-ld2r.c
index e3d3ac9..b25f8dd 100644
--- a/src/qs8-igemm/gen/2x8c4-minmax-rndnu-neon-mull-ld2r.c
+++ b/src/qs8-igemm/gen/2x8c4-minmax-rndnu-neon-mull-ld2r.c

@@ -48,10 +48,10 @@
   }
 
   do {
-    int32x4_t vacc0x01 = vreinterpretq_s32_u64(vshll_n_u32(vld1_u32(w), 0)); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
-    int32x4_t vacc0x23 = vreinterpretq_s32_u64(vshll_n_u32(vld1_u32(w), 0)); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
-    int32x4_t vacc0x45 = vreinterpretq_s32_u64(vshll_n_u32(vld1_u32(w), 0)); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
-    int32x4_t vacc0x67 = vreinterpretq_s32_u64(vshll_n_u32(vld1_u32(w), 0)); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
+    int32x4_t vacc0x01 = vreinterpretq_s32_u64(vmovl_u32(vld1_u32(w))); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
+    int32x4_t vacc0x23 = vreinterpretq_s32_u64(vmovl_u32(vld1_u32(w))); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
+    int32x4_t vacc0x45 = vreinterpretq_s32_u64(vmovl_u32(vld1_u32(w))); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
+    int32x4_t vacc0x67 = vreinterpretq_s32_u64(vmovl_u32(vld1_u32(w))); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
     int32x4_t vacc1x01 = vacc0x01;
     int32x4_t vacc1x23 = vacc0x23;
     int32x4_t vacc1x45 = vacc0x45;

diff --git a/src/qs8-igemm/gen/2x8c4s2-minmax-fp32-neon-mlal.c b/src/qs8-igemm/gen/2x8c4s2-minmax-fp32-neon-mlal.c
index fc3890e..73419cc 100644
--- a/src/qs8-igemm/gen/2x8c4s2-minmax-fp32-neon-mlal.c
+++ b/src/qs8-igemm/gen/2x8c4s2-minmax-fp32-neon-mlal.c

@@ -47,10 +47,10 @@
   }
 
   do {
-    int32x4_t vacc0x01 = vreinterpretq_s32_u64(vshll_n_u32(vld1_u32(w), 0)); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
-    int32x4_t vacc0x23 = vreinterpretq_s32_u64(vshll_n_u32(vld1_u32(w), 0)); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
-    int32x4_t vacc0x45 = vreinterpretq_s32_u64(vshll_n_u32(vld1_u32(w), 0)); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
-    int32x4_t vacc0x67 = vreinterpretq_s32_u64(vshll_n_u32(vld1_u32(w), 0)); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
+    int32x4_t vacc0x01 = vreinterpretq_s32_u64(vmovl_u32(vld1_u32(w))); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
+    int32x4_t vacc0x23 = vreinterpretq_s32_u64(vmovl_u32(vld1_u32(w))); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
+    int32x4_t vacc0x45 = vreinterpretq_s32_u64(vmovl_u32(vld1_u32(w))); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
+    int32x4_t vacc0x67 = vreinterpretq_s32_u64(vmovl_u32(vld1_u32(w))); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
     int32x4_t vacc1x01 = vacc0x01;
     int32x4_t vacc1x23 = vacc0x23;
     int32x4_t vacc1x45 = vacc0x45;

diff --git a/src/qs8-igemm/gen/2x8c4s2-minmax-fp32-neonv8-mlal.c b/src/qs8-igemm/gen/2x8c4s2-minmax-fp32-neonv8-mlal.c
index e262fd8..2e3e476 100644
--- a/src/qs8-igemm/gen/2x8c4s2-minmax-fp32-neonv8-mlal.c
+++ b/src/qs8-igemm/gen/2x8c4s2-minmax-fp32-neonv8-mlal.c

@@ -48,10 +48,10 @@
   }
 
   do {
-    int32x4_t vacc0x01 = vreinterpretq_s32_u64(vshll_n_u32(vld1_u32(w), 0)); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
-    int32x4_t vacc0x23 = vreinterpretq_s32_u64(vshll_n_u32(vld1_u32(w), 0)); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
-    int32x4_t vacc0x45 = vreinterpretq_s32_u64(vshll_n_u32(vld1_u32(w), 0)); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
-    int32x4_t vacc0x67 = vreinterpretq_s32_u64(vshll_n_u32(vld1_u32(w), 0)); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
+    int32x4_t vacc0x01 = vreinterpretq_s32_u64(vmovl_u32(vld1_u32(w))); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
+    int32x4_t vacc0x23 = vreinterpretq_s32_u64(vmovl_u32(vld1_u32(w))); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
+    int32x4_t vacc0x45 = vreinterpretq_s32_u64(vmovl_u32(vld1_u32(w))); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
+    int32x4_t vacc0x67 = vreinterpretq_s32_u64(vmovl_u32(vld1_u32(w))); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
     int32x4_t vacc1x01 = vacc0x01;
     int32x4_t vacc1x23 = vacc0x23;
     int32x4_t vacc1x45 = vacc0x45;

diff --git a/src/qs8-igemm/gen/2x8c4s2-minmax-rndnu-neon-mlal.c b/src/qs8-igemm/gen/2x8c4s2-minmax-rndnu-neon-mlal.c
index 4bd5701..fd23a3a 100644
--- a/src/qs8-igemm/gen/2x8c4s2-minmax-rndnu-neon-mlal.c
+++ b/src/qs8-igemm/gen/2x8c4s2-minmax-rndnu-neon-mlal.c

@@ -47,10 +47,10 @@
   }
 
   do {
-    int32x4_t vacc0x01 = vreinterpretq_s32_u64(vshll_n_u32(vld1_u32(w), 0)); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
-    int32x4_t vacc0x23 = vreinterpretq_s32_u64(vshll_n_u32(vld1_u32(w), 0)); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
-    int32x4_t vacc0x45 = vreinterpretq_s32_u64(vshll_n_u32(vld1_u32(w), 0)); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
-    int32x4_t vacc0x67 = vreinterpretq_s32_u64(vshll_n_u32(vld1_u32(w), 0)); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
+    int32x4_t vacc0x01 = vreinterpretq_s32_u64(vmovl_u32(vld1_u32(w))); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
+    int32x4_t vacc0x23 = vreinterpretq_s32_u64(vmovl_u32(vld1_u32(w))); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
+    int32x4_t vacc0x45 = vreinterpretq_s32_u64(vmovl_u32(vld1_u32(w))); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
+    int32x4_t vacc0x67 = vreinterpretq_s32_u64(vmovl_u32(vld1_u32(w))); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
     int32x4_t vacc1x01 = vacc0x01;
     int32x4_t vacc1x23 = vacc0x23;
     int32x4_t vacc1x45 = vacc0x45;

diff --git a/src/qs8-igemm/gen/2x8c4s2-minmax-rndnu-neon-mull.c b/src/qs8-igemm/gen/2x8c4s2-minmax-rndnu-neon-mull.c
index faf70bb..5bed970 100644
--- a/src/qs8-igemm/gen/2x8c4s2-minmax-rndnu-neon-mull.c
+++ b/src/qs8-igemm/gen/2x8c4s2-minmax-rndnu-neon-mull.c

@@ -47,10 +47,10 @@
   }
 
   do {
-    int32x4_t vacc0x01 = vreinterpretq_s32_u64(vshll_n_u32(vld1_u32(w), 0)); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
-    int32x4_t vacc0x23 = vreinterpretq_s32_u64(vshll_n_u32(vld1_u32(w), 0)); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
-    int32x4_t vacc0x45 = vreinterpretq_s32_u64(vshll_n_u32(vld1_u32(w), 0)); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
-    int32x4_t vacc0x67 = vreinterpretq_s32_u64(vshll_n_u32(vld1_u32(w), 0)); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
+    int32x4_t vacc0x01 = vreinterpretq_s32_u64(vmovl_u32(vld1_u32(w))); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
+    int32x4_t vacc0x23 = vreinterpretq_s32_u64(vmovl_u32(vld1_u32(w))); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
+    int32x4_t vacc0x45 = vreinterpretq_s32_u64(vmovl_u32(vld1_u32(w))); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
+    int32x4_t vacc0x67 = vreinterpretq_s32_u64(vmovl_u32(vld1_u32(w))); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
     int32x4_t vacc1x01 = vacc0x01;
     int32x4_t vacc1x23 = vacc0x23;
     int32x4_t vacc1x45 = vacc0x45;

diff --git a/src/qs8-igemm/gen/3x16c4-minmax-rndnu-neon-mlal-dup.c b/src/qs8-igemm/gen/3x16c4-minmax-rndnu-neon-mlal-dup.c
index b783bae..4c99e05 100644
--- a/src/qs8-igemm/gen/3x16c4-minmax-rndnu-neon-mlal-dup.c
+++ b/src/qs8-igemm/gen/3x16c4-minmax-rndnu-neon-mlal-dup.c

@@ -52,14 +52,14 @@
   }
 
   do {
-    int32x4_t vacc0x01 = vreinterpretq_s32_u64(vshll_n_u32(vld1_u32(w), 0)); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
-    int32x4_t vacc0x23 = vreinterpretq_s32_u64(vshll_n_u32(vld1_u32(w), 0)); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
-    int32x4_t vacc0x45 = vreinterpretq_s32_u64(vshll_n_u32(vld1_u32(w), 0)); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
-    int32x4_t vacc0x67 = vreinterpretq_s32_u64(vshll_n_u32(vld1_u32(w), 0)); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
-    int32x4_t vacc0x89 = vreinterpretq_s32_u64(vshll_n_u32(vld1_u32(w), 0)); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
-    int32x4_t vacc0xAB = vreinterpretq_s32_u64(vshll_n_u32(vld1_u32(w), 0)); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
-    int32x4_t vacc0xCD = vreinterpretq_s32_u64(vshll_n_u32(vld1_u32(w), 0)); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
-    int32x4_t vacc0xEF = vreinterpretq_s32_u64(vshll_n_u32(vld1_u32(w), 0)); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
+    int32x4_t vacc0x01 = vreinterpretq_s32_u64(vmovl_u32(vld1_u32(w))); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
+    int32x4_t vacc0x23 = vreinterpretq_s32_u64(vmovl_u32(vld1_u32(w))); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
+    int32x4_t vacc0x45 = vreinterpretq_s32_u64(vmovl_u32(vld1_u32(w))); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
+    int32x4_t vacc0x67 = vreinterpretq_s32_u64(vmovl_u32(vld1_u32(w))); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
+    int32x4_t vacc0x89 = vreinterpretq_s32_u64(vmovl_u32(vld1_u32(w))); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
+    int32x4_t vacc0xAB = vreinterpretq_s32_u64(vmovl_u32(vld1_u32(w))); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
+    int32x4_t vacc0xCD = vreinterpretq_s32_u64(vmovl_u32(vld1_u32(w))); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
+    int32x4_t vacc0xEF = vreinterpretq_s32_u64(vmovl_u32(vld1_u32(w))); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
     int32x4_t vacc1x01 = vacc0x01;
     int32x4_t vacc1x23 = vacc0x23;
     int32x4_t vacc1x45 = vacc0x45;

diff --git a/src/qs8-igemm/gen/3x16c4-minmax-rndnu-neon-mlal-ld1r.c b/src/qs8-igemm/gen/3x16c4-minmax-rndnu-neon-mlal-ld1r.c
index 729b85e..879f829 100644
--- a/src/qs8-igemm/gen/3x16c4-minmax-rndnu-neon-mlal-ld1r.c
+++ b/src/qs8-igemm/gen/3x16c4-minmax-rndnu-neon-mlal-ld1r.c

@@ -52,14 +52,14 @@
   }
 
   do {
-    int32x4_t vacc0x01 = vreinterpretq_s32_u64(vshll_n_u32(vld1_u32(w), 0)); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
-    int32x4_t vacc0x23 = vreinterpretq_s32_u64(vshll_n_u32(vld1_u32(w), 0)); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
-    int32x4_t vacc0x45 = vreinterpretq_s32_u64(vshll_n_u32(vld1_u32(w), 0)); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
-    int32x4_t vacc0x67 = vreinterpretq_s32_u64(vshll_n_u32(vld1_u32(w), 0)); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
-    int32x4_t vacc0x89 = vreinterpretq_s32_u64(vshll_n_u32(vld1_u32(w), 0)); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
-    int32x4_t vacc0xAB = vreinterpretq_s32_u64(vshll_n_u32(vld1_u32(w), 0)); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
-    int32x4_t vacc0xCD = vreinterpretq_s32_u64(vshll_n_u32(vld1_u32(w), 0)); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
-    int32x4_t vacc0xEF = vreinterpretq_s32_u64(vshll_n_u32(vld1_u32(w), 0)); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
+    int32x4_t vacc0x01 = vreinterpretq_s32_u64(vmovl_u32(vld1_u32(w))); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
+    int32x4_t vacc0x23 = vreinterpretq_s32_u64(vmovl_u32(vld1_u32(w))); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
+    int32x4_t vacc0x45 = vreinterpretq_s32_u64(vmovl_u32(vld1_u32(w))); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
+    int32x4_t vacc0x67 = vreinterpretq_s32_u64(vmovl_u32(vld1_u32(w))); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
+    int32x4_t vacc0x89 = vreinterpretq_s32_u64(vmovl_u32(vld1_u32(w))); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
+    int32x4_t vacc0xAB = vreinterpretq_s32_u64(vmovl_u32(vld1_u32(w))); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
+    int32x4_t vacc0xCD = vreinterpretq_s32_u64(vmovl_u32(vld1_u32(w))); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
+    int32x4_t vacc0xEF = vreinterpretq_s32_u64(vmovl_u32(vld1_u32(w))); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
     int32x4_t vacc1x01 = vacc0x01;
     int32x4_t vacc1x23 = vacc0x23;
     int32x4_t vacc1x45 = vacc0x45;

diff --git a/src/qs8-igemm/gen/3x16c4-minmax-rndnu-neon-mlal-ld2r.c b/src/qs8-igemm/gen/3x16c4-minmax-rndnu-neon-mlal-ld2r.c
index 6fbe7da..9dd432d 100644
--- a/src/qs8-igemm/gen/3x16c4-minmax-rndnu-neon-mlal-ld2r.c
+++ b/src/qs8-igemm/gen/3x16c4-minmax-rndnu-neon-mlal-ld2r.c

@@ -52,14 +52,14 @@
   }
 
   do {
-    int32x4_t vacc0x01 = vreinterpretq_s32_u64(vshll_n_u32(vld1_u32(w), 0)); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
-    int32x4_t vacc0x23 = vreinterpretq_s32_u64(vshll_n_u32(vld1_u32(w), 0)); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
-    int32x4_t vacc0x45 = vreinterpretq_s32_u64(vshll_n_u32(vld1_u32(w), 0)); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
-    int32x4_t vacc0x67 = vreinterpretq_s32_u64(vshll_n_u32(vld1_u32(w), 0)); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
-    int32x4_t vacc0x89 = vreinterpretq_s32_u64(vshll_n_u32(vld1_u32(w), 0)); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
-    int32x4_t vacc0xAB = vreinterpretq_s32_u64(vshll_n_u32(vld1_u32(w), 0)); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
-    int32x4_t vacc0xCD = vreinterpretq_s32_u64(vshll_n_u32(vld1_u32(w), 0)); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
-    int32x4_t vacc0xEF = vreinterpretq_s32_u64(vshll_n_u32(vld1_u32(w), 0)); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
+    int32x4_t vacc0x01 = vreinterpretq_s32_u64(vmovl_u32(vld1_u32(w))); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
+    int32x4_t vacc0x23 = vreinterpretq_s32_u64(vmovl_u32(vld1_u32(w))); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
+    int32x4_t vacc0x45 = vreinterpretq_s32_u64(vmovl_u32(vld1_u32(w))); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
+    int32x4_t vacc0x67 = vreinterpretq_s32_u64(vmovl_u32(vld1_u32(w))); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
+    int32x4_t vacc0x89 = vreinterpretq_s32_u64(vmovl_u32(vld1_u32(w))); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
+    int32x4_t vacc0xAB = vreinterpretq_s32_u64(vmovl_u32(vld1_u32(w))); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
+    int32x4_t vacc0xCD = vreinterpretq_s32_u64(vmovl_u32(vld1_u32(w))); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
+    int32x4_t vacc0xEF = vreinterpretq_s32_u64(vmovl_u32(vld1_u32(w))); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
     int32x4_t vacc1x01 = vacc0x01;
     int32x4_t vacc1x23 = vacc0x23;
     int32x4_t vacc1x45 = vacc0x45;

diff --git a/src/qs8-igemm/gen/3x16c4-minmax-rndnu-neon-mull-dup.c b/src/qs8-igemm/gen/3x16c4-minmax-rndnu-neon-mull-dup.c
index 91b6427..294a797 100644
--- a/src/qs8-igemm/gen/3x16c4-minmax-rndnu-neon-mull-dup.c
+++ b/src/qs8-igemm/gen/3x16c4-minmax-rndnu-neon-mull-dup.c

@@ -52,14 +52,14 @@
   }
 
   do {
-    int32x4_t vacc0x01 = vreinterpretq_s32_u64(vshll_n_u32(vld1_u32(w), 0)); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
-    int32x4_t vacc0x23 = vreinterpretq_s32_u64(vshll_n_u32(vld1_u32(w), 0)); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
-    int32x4_t vacc0x45 = vreinterpretq_s32_u64(vshll_n_u32(vld1_u32(w), 0)); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
-    int32x4_t vacc0x67 = vreinterpretq_s32_u64(vshll_n_u32(vld1_u32(w), 0)); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
-    int32x4_t vacc0x89 = vreinterpretq_s32_u64(vshll_n_u32(vld1_u32(w), 0)); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
-    int32x4_t vacc0xAB = vreinterpretq_s32_u64(vshll_n_u32(vld1_u32(w), 0)); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
-    int32x4_t vacc0xCD = vreinterpretq_s32_u64(vshll_n_u32(vld1_u32(w), 0)); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
-    int32x4_t vacc0xEF = vreinterpretq_s32_u64(vshll_n_u32(vld1_u32(w), 0)); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
+    int32x4_t vacc0x01 = vreinterpretq_s32_u64(vmovl_u32(vld1_u32(w))); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
+    int32x4_t vacc0x23 = vreinterpretq_s32_u64(vmovl_u32(vld1_u32(w))); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
+    int32x4_t vacc0x45 = vreinterpretq_s32_u64(vmovl_u32(vld1_u32(w))); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
+    int32x4_t vacc0x67 = vreinterpretq_s32_u64(vmovl_u32(vld1_u32(w))); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
+    int32x4_t vacc0x89 = vreinterpretq_s32_u64(vmovl_u32(vld1_u32(w))); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
+    int32x4_t vacc0xAB = vreinterpretq_s32_u64(vmovl_u32(vld1_u32(w))); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
+    int32x4_t vacc0xCD = vreinterpretq_s32_u64(vmovl_u32(vld1_u32(w))); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
+    int32x4_t vacc0xEF = vreinterpretq_s32_u64(vmovl_u32(vld1_u32(w))); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
     int32x4_t vacc1x01 = vacc0x01;
     int32x4_t vacc1x23 = vacc0x23;
     int32x4_t vacc1x45 = vacc0x45;

diff --git a/src/qs8-igemm/gen/3x16c4-minmax-rndnu-neon-mull-ld1r.c b/src/qs8-igemm/gen/3x16c4-minmax-rndnu-neon-mull-ld1r.c
index 14ed5e9..1ef48fa 100644
--- a/src/qs8-igemm/gen/3x16c4-minmax-rndnu-neon-mull-ld1r.c
+++ b/src/qs8-igemm/gen/3x16c4-minmax-rndnu-neon-mull-ld1r.c

@@ -52,14 +52,14 @@
   }
 
   do {
-    int32x4_t vacc0x01 = vreinterpretq_s32_u64(vshll_n_u32(vld1_u32(w), 0)); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
-    int32x4_t vacc0x23 = vreinterpretq_s32_u64(vshll_n_u32(vld1_u32(w), 0)); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
-    int32x4_t vacc0x45 = vreinterpretq_s32_u64(vshll_n_u32(vld1_u32(w), 0)); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
-    int32x4_t vacc0x67 = vreinterpretq_s32_u64(vshll_n_u32(vld1_u32(w), 0)); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
-    int32x4_t vacc0x89 = vreinterpretq_s32_u64(vshll_n_u32(vld1_u32(w), 0)); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
-    int32x4_t vacc0xAB = vreinterpretq_s32_u64(vshll_n_u32(vld1_u32(w), 0)); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
-    int32x4_t vacc0xCD = vreinterpretq_s32_u64(vshll_n_u32(vld1_u32(w), 0)); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
-    int32x4_t vacc0xEF = vreinterpretq_s32_u64(vshll_n_u32(vld1_u32(w), 0)); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
+    int32x4_t vacc0x01 = vreinterpretq_s32_u64(vmovl_u32(vld1_u32(w))); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
+    int32x4_t vacc0x23 = vreinterpretq_s32_u64(vmovl_u32(vld1_u32(w))); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
+    int32x4_t vacc0x45 = vreinterpretq_s32_u64(vmovl_u32(vld1_u32(w))); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
+    int32x4_t vacc0x67 = vreinterpretq_s32_u64(vmovl_u32(vld1_u32(w))); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
+    int32x4_t vacc0x89 = vreinterpretq_s32_u64(vmovl_u32(vld1_u32(w))); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
+    int32x4_t vacc0xAB = vreinterpretq_s32_u64(vmovl_u32(vld1_u32(w))); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
+    int32x4_t vacc0xCD = vreinterpretq_s32_u64(vmovl_u32(vld1_u32(w))); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
+    int32x4_t vacc0xEF = vreinterpretq_s32_u64(vmovl_u32(vld1_u32(w))); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
     int32x4_t vacc1x01 = vacc0x01;
     int32x4_t vacc1x23 = vacc0x23;
     int32x4_t vacc1x45 = vacc0x45;

diff --git a/src/qs8-igemm/gen/3x16c4-minmax-rndnu-neon-mull-ld2r.c b/src/qs8-igemm/gen/3x16c4-minmax-rndnu-neon-mull-ld2r.c
index 28cb468..1130a20 100644
--- a/src/qs8-igemm/gen/3x16c4-minmax-rndnu-neon-mull-ld2r.c
+++ b/src/qs8-igemm/gen/3x16c4-minmax-rndnu-neon-mull-ld2r.c

@@ -52,14 +52,14 @@
   }
 
   do {
-    int32x4_t vacc0x01 = vreinterpretq_s32_u64(vshll_n_u32(vld1_u32(w), 0)); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
-    int32x4_t vacc0x23 = vreinterpretq_s32_u64(vshll_n_u32(vld1_u32(w), 0)); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
-    int32x4_t vacc0x45 = vreinterpretq_s32_u64(vshll_n_u32(vld1_u32(w), 0)); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
-    int32x4_t vacc0x67 = vreinterpretq_s32_u64(vshll_n_u32(vld1_u32(w), 0)); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
-    int32x4_t vacc0x89 = vreinterpretq_s32_u64(vshll_n_u32(vld1_u32(w), 0)); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
-    int32x4_t vacc0xAB = vreinterpretq_s32_u64(vshll_n_u32(vld1_u32(w), 0)); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
-    int32x4_t vacc0xCD = vreinterpretq_s32_u64(vshll_n_u32(vld1_u32(w), 0)); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
-    int32x4_t vacc0xEF = vreinterpretq_s32_u64(vshll_n_u32(vld1_u32(w), 0)); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
+    int32x4_t vacc0x01 = vreinterpretq_s32_u64(vmovl_u32(vld1_u32(w))); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
+    int32x4_t vacc0x23 = vreinterpretq_s32_u64(vmovl_u32(vld1_u32(w))); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
+    int32x4_t vacc0x45 = vreinterpretq_s32_u64(vmovl_u32(vld1_u32(w))); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
+    int32x4_t vacc0x67 = vreinterpretq_s32_u64(vmovl_u32(vld1_u32(w))); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
+    int32x4_t vacc0x89 = vreinterpretq_s32_u64(vmovl_u32(vld1_u32(w))); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
+    int32x4_t vacc0xAB = vreinterpretq_s32_u64(vmovl_u32(vld1_u32(w))); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
+    int32x4_t vacc0xCD = vreinterpretq_s32_u64(vmovl_u32(vld1_u32(w))); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
+    int32x4_t vacc0xEF = vreinterpretq_s32_u64(vmovl_u32(vld1_u32(w))); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
     int32x4_t vacc1x01 = vacc0x01;
     int32x4_t vacc1x23 = vacc0x23;
     int32x4_t vacc1x45 = vacc0x45;

diff --git a/src/qs8-igemm/gen/3x16c4s2-minmax-rndnu-neon-mlal.c b/src/qs8-igemm/gen/3x16c4s2-minmax-rndnu-neon-mlal.c
index 8cf171f..a82599c 100644
--- a/src/qs8-igemm/gen/3x16c4s2-minmax-rndnu-neon-mlal.c
+++ b/src/qs8-igemm/gen/3x16c4s2-minmax-rndnu-neon-mlal.c

@@ -51,14 +51,14 @@
   }
 
   do {
-    int32x4_t vacc0x01 = vreinterpretq_s32_u64(vshll_n_u32(vld1_u32(w), 0)); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
-    int32x4_t vacc0x23 = vreinterpretq_s32_u64(vshll_n_u32(vld1_u32(w), 0)); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
-    int32x4_t vacc0x45 = vreinterpretq_s32_u64(vshll_n_u32(vld1_u32(w), 0)); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
-    int32x4_t vacc0x67 = vreinterpretq_s32_u64(vshll_n_u32(vld1_u32(w), 0)); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
-    int32x4_t vacc0x89 = vreinterpretq_s32_u64(vshll_n_u32(vld1_u32(w), 0)); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
-    int32x4_t vacc0xAB = vreinterpretq_s32_u64(vshll_n_u32(vld1_u32(w), 0)); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
-    int32x4_t vacc0xCD = vreinterpretq_s32_u64(vshll_n_u32(vld1_u32(w), 0)); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
-    int32x4_t vacc0xEF = vreinterpretq_s32_u64(vshll_n_u32(vld1_u32(w), 0)); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
+    int32x4_t vacc0x01 = vreinterpretq_s32_u64(vmovl_u32(vld1_u32(w))); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
+    int32x4_t vacc0x23 = vreinterpretq_s32_u64(vmovl_u32(vld1_u32(w))); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
+    int32x4_t vacc0x45 = vreinterpretq_s32_u64(vmovl_u32(vld1_u32(w))); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
+    int32x4_t vacc0x67 = vreinterpretq_s32_u64(vmovl_u32(vld1_u32(w))); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
+    int32x4_t vacc0x89 = vreinterpretq_s32_u64(vmovl_u32(vld1_u32(w))); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
+    int32x4_t vacc0xAB = vreinterpretq_s32_u64(vmovl_u32(vld1_u32(w))); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
+    int32x4_t vacc0xCD = vreinterpretq_s32_u64(vmovl_u32(vld1_u32(w))); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
+    int32x4_t vacc0xEF = vreinterpretq_s32_u64(vmovl_u32(vld1_u32(w))); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
     int32x4_t vacc1x01 = vacc0x01;
     int32x4_t vacc1x23 = vacc0x23;
     int32x4_t vacc1x45 = vacc0x45;

diff --git a/src/qs8-igemm/gen/3x16c4s2-minmax-rndnu-neon-mull.c b/src/qs8-igemm/gen/3x16c4s2-minmax-rndnu-neon-mull.c
index 8a75ae1..3b468a7 100644
--- a/src/qs8-igemm/gen/3x16c4s2-minmax-rndnu-neon-mull.c
+++ b/src/qs8-igemm/gen/3x16c4s2-minmax-rndnu-neon-mull.c

@@ -51,14 +51,14 @@
   }
 
   do {
-    int32x4_t vacc0x01 = vreinterpretq_s32_u64(vshll_n_u32(vld1_u32(w), 0)); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
-    int32x4_t vacc0x23 = vreinterpretq_s32_u64(vshll_n_u32(vld1_u32(w), 0)); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
-    int32x4_t vacc0x45 = vreinterpretq_s32_u64(vshll_n_u32(vld1_u32(w), 0)); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
-    int32x4_t vacc0x67 = vreinterpretq_s32_u64(vshll_n_u32(vld1_u32(w), 0)); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
-    int32x4_t vacc0x89 = vreinterpretq_s32_u64(vshll_n_u32(vld1_u32(w), 0)); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
-    int32x4_t vacc0xAB = vreinterpretq_s32_u64(vshll_n_u32(vld1_u32(w), 0)); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
-    int32x4_t vacc0xCD = vreinterpretq_s32_u64(vshll_n_u32(vld1_u32(w), 0)); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
-    int32x4_t vacc0xEF = vreinterpretq_s32_u64(vshll_n_u32(vld1_u32(w), 0)); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
+    int32x4_t vacc0x01 = vreinterpretq_s32_u64(vmovl_u32(vld1_u32(w))); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
+    int32x4_t vacc0x23 = vreinterpretq_s32_u64(vmovl_u32(vld1_u32(w))); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
+    int32x4_t vacc0x45 = vreinterpretq_s32_u64(vmovl_u32(vld1_u32(w))); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
+    int32x4_t vacc0x67 = vreinterpretq_s32_u64(vmovl_u32(vld1_u32(w))); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
+    int32x4_t vacc0x89 = vreinterpretq_s32_u64(vmovl_u32(vld1_u32(w))); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
+    int32x4_t vacc0xAB = vreinterpretq_s32_u64(vmovl_u32(vld1_u32(w))); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
+    int32x4_t vacc0xCD = vreinterpretq_s32_u64(vmovl_u32(vld1_u32(w))); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
+    int32x4_t vacc0xEF = vreinterpretq_s32_u64(vmovl_u32(vld1_u32(w))); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
     int32x4_t vacc1x01 = vacc0x01;
     int32x4_t vacc1x23 = vacc0x23;
     int32x4_t vacc1x45 = vacc0x45;

diff --git a/src/qs8-igemm/gen/3x8c4-minmax-rndnu-neon-mlal-dup.c b/src/qs8-igemm/gen/3x8c4-minmax-rndnu-neon-mlal-dup.c
index 59e1dc5..5c9ccaf 100644
--- a/src/qs8-igemm/gen/3x8c4-minmax-rndnu-neon-mlal-dup.c
+++ b/src/qs8-igemm/gen/3x8c4-minmax-rndnu-neon-mlal-dup.c

@@ -52,10 +52,10 @@
   }
 
   do {
-    int32x4_t vacc0x01 = vreinterpretq_s32_u64(vshll_n_u32(vld1_u32(w), 0)); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
-    int32x4_t vacc0x23 = vreinterpretq_s32_u64(vshll_n_u32(vld1_u32(w), 0)); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
-    int32x4_t vacc0x45 = vreinterpretq_s32_u64(vshll_n_u32(vld1_u32(w), 0)); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
-    int32x4_t vacc0x67 = vreinterpretq_s32_u64(vshll_n_u32(vld1_u32(w), 0)); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
+    int32x4_t vacc0x01 = vreinterpretq_s32_u64(vmovl_u32(vld1_u32(w))); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
+    int32x4_t vacc0x23 = vreinterpretq_s32_u64(vmovl_u32(vld1_u32(w))); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
+    int32x4_t vacc0x45 = vreinterpretq_s32_u64(vmovl_u32(vld1_u32(w))); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
+    int32x4_t vacc0x67 = vreinterpretq_s32_u64(vmovl_u32(vld1_u32(w))); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
     int32x4_t vacc1x01 = vacc0x01;
     int32x4_t vacc1x23 = vacc0x23;
     int32x4_t vacc1x45 = vacc0x45;

diff --git a/src/qs8-igemm/gen/3x8c4-minmax-rndnu-neon-mlal-ld1r.c b/src/qs8-igemm/gen/3x8c4-minmax-rndnu-neon-mlal-ld1r.c
index ecb850c..a570533 100644
--- a/src/qs8-igemm/gen/3x8c4-minmax-rndnu-neon-mlal-ld1r.c
+++ b/src/qs8-igemm/gen/3x8c4-minmax-rndnu-neon-mlal-ld1r.c

@@ -52,10 +52,10 @@
   }
 
   do {
-    int32x4_t vacc0x01 = vreinterpretq_s32_u64(vshll_n_u32(vld1_u32(w), 0)); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
-    int32x4_t vacc0x23 = vreinterpretq_s32_u64(vshll_n_u32(vld1_u32(w), 0)); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
-    int32x4_t vacc0x45 = vreinterpretq_s32_u64(vshll_n_u32(vld1_u32(w), 0)); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
-    int32x4_t vacc0x67 = vreinterpretq_s32_u64(vshll_n_u32(vld1_u32(w), 0)); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
+    int32x4_t vacc0x01 = vreinterpretq_s32_u64(vmovl_u32(vld1_u32(w))); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
+    int32x4_t vacc0x23 = vreinterpretq_s32_u64(vmovl_u32(vld1_u32(w))); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
+    int32x4_t vacc0x45 = vreinterpretq_s32_u64(vmovl_u32(vld1_u32(w))); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
+    int32x4_t vacc0x67 = vreinterpretq_s32_u64(vmovl_u32(vld1_u32(w))); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
     int32x4_t vacc1x01 = vacc0x01;
     int32x4_t vacc1x23 = vacc0x23;
     int32x4_t vacc1x45 = vacc0x45;

diff --git a/src/qs8-igemm/gen/3x8c4-minmax-rndnu-neon-mlal-ld2r.c b/src/qs8-igemm/gen/3x8c4-minmax-rndnu-neon-mlal-ld2r.c
index 16a8426..6889e0d 100644
--- a/src/qs8-igemm/gen/3x8c4-minmax-rndnu-neon-mlal-ld2r.c
+++ b/src/qs8-igemm/gen/3x8c4-minmax-rndnu-neon-mlal-ld2r.c

@@ -52,10 +52,10 @@
   }
 
   do {
-    int32x4_t vacc0x01 = vreinterpretq_s32_u64(vshll_n_u32(vld1_u32(w), 0)); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
-    int32x4_t vacc0x23 = vreinterpretq_s32_u64(vshll_n_u32(vld1_u32(w), 0)); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
-    int32x4_t vacc0x45 = vreinterpretq_s32_u64(vshll_n_u32(vld1_u32(w), 0)); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
-    int32x4_t vacc0x67 = vreinterpretq_s32_u64(vshll_n_u32(vld1_u32(w), 0)); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
+    int32x4_t vacc0x01 = vreinterpretq_s32_u64(vmovl_u32(vld1_u32(w))); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
+    int32x4_t vacc0x23 = vreinterpretq_s32_u64(vmovl_u32(vld1_u32(w))); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
+    int32x4_t vacc0x45 = vreinterpretq_s32_u64(vmovl_u32(vld1_u32(w))); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
+    int32x4_t vacc0x67 = vreinterpretq_s32_u64(vmovl_u32(vld1_u32(w))); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
     int32x4_t vacc1x01 = vacc0x01;
     int32x4_t vacc1x23 = vacc0x23;
     int32x4_t vacc1x45 = vacc0x45;

diff --git a/src/qs8-igemm/gen/3x8c4-minmax-rndnu-neon-mull-dup.c b/src/qs8-igemm/gen/3x8c4-minmax-rndnu-neon-mull-dup.c
index f63f16f..442deb8 100644
--- a/src/qs8-igemm/gen/3x8c4-minmax-rndnu-neon-mull-dup.c
+++ b/src/qs8-igemm/gen/3x8c4-minmax-rndnu-neon-mull-dup.c

@@ -52,10 +52,10 @@
   }
 
   do {
-    int32x4_t vacc0x01 = vreinterpretq_s32_u64(vshll_n_u32(vld1_u32(w), 0)); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
-    int32x4_t vacc0x23 = vreinterpretq_s32_u64(vshll_n_u32(vld1_u32(w), 0)); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
-    int32x4_t vacc0x45 = vreinterpretq_s32_u64(vshll_n_u32(vld1_u32(w), 0)); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
-    int32x4_t vacc0x67 = vreinterpretq_s32_u64(vshll_n_u32(vld1_u32(w), 0)); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
+    int32x4_t vacc0x01 = vreinterpretq_s32_u64(vmovl_u32(vld1_u32(w))); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
+    int32x4_t vacc0x23 = vreinterpretq_s32_u64(vmovl_u32(vld1_u32(w))); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
+    int32x4_t vacc0x45 = vreinterpretq_s32_u64(vmovl_u32(vld1_u32(w))); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
+    int32x4_t vacc0x67 = vreinterpretq_s32_u64(vmovl_u32(vld1_u32(w))); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
     int32x4_t vacc1x01 = vacc0x01;
     int32x4_t vacc1x23 = vacc0x23;
     int32x4_t vacc1x45 = vacc0x45;

diff --git a/src/qs8-igemm/gen/3x8c4-minmax-rndnu-neon-mull-ld1r.c b/src/qs8-igemm/gen/3x8c4-minmax-rndnu-neon-mull-ld1r.c
index 4de4de5..b2abc9f 100644
--- a/src/qs8-igemm/gen/3x8c4-minmax-rndnu-neon-mull-ld1r.c
+++ b/src/qs8-igemm/gen/3x8c4-minmax-rndnu-neon-mull-ld1r.c

@@ -52,10 +52,10 @@
   }
 
   do {
-    int32x4_t vacc0x01 = vreinterpretq_s32_u64(vshll_n_u32(vld1_u32(w), 0)); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
-    int32x4_t vacc0x23 = vreinterpretq_s32_u64(vshll_n_u32(vld1_u32(w), 0)); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
-    int32x4_t vacc0x45 = vreinterpretq_s32_u64(vshll_n_u32(vld1_u32(w), 0)); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
-    int32x4_t vacc0x67 = vreinterpretq_s32_u64(vshll_n_u32(vld1_u32(w), 0)); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
+    int32x4_t vacc0x01 = vreinterpretq_s32_u64(vmovl_u32(vld1_u32(w))); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
+    int32x4_t vacc0x23 = vreinterpretq_s32_u64(vmovl_u32(vld1_u32(w))); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
+    int32x4_t vacc0x45 = vreinterpretq_s32_u64(vmovl_u32(vld1_u32(w))); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
+    int32x4_t vacc0x67 = vreinterpretq_s32_u64(vmovl_u32(vld1_u32(w))); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
     int32x4_t vacc1x01 = vacc0x01;
     int32x4_t vacc1x23 = vacc0x23;
     int32x4_t vacc1x45 = vacc0x45;

diff --git a/src/qs8-igemm/gen/3x8c4-minmax-rndnu-neon-mull-ld2r.c b/src/qs8-igemm/gen/3x8c4-minmax-rndnu-neon-mull-ld2r.c
index c1e1154..4476d13 100644
--- a/src/qs8-igemm/gen/3x8c4-minmax-rndnu-neon-mull-ld2r.c
+++ b/src/qs8-igemm/gen/3x8c4-minmax-rndnu-neon-mull-ld2r.c

@@ -52,10 +52,10 @@
   }
 
   do {
-    int32x4_t vacc0x01 = vreinterpretq_s32_u64(vshll_n_u32(vld1_u32(w), 0)); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
-    int32x4_t vacc0x23 = vreinterpretq_s32_u64(vshll_n_u32(vld1_u32(w), 0)); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
-    int32x4_t vacc0x45 = vreinterpretq_s32_u64(vshll_n_u32(vld1_u32(w), 0)); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
-    int32x4_t vacc0x67 = vreinterpretq_s32_u64(vshll_n_u32(vld1_u32(w), 0)); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
+    int32x4_t vacc0x01 = vreinterpretq_s32_u64(vmovl_u32(vld1_u32(w))); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
+    int32x4_t vacc0x23 = vreinterpretq_s32_u64(vmovl_u32(vld1_u32(w))); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
+    int32x4_t vacc0x45 = vreinterpretq_s32_u64(vmovl_u32(vld1_u32(w))); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
+    int32x4_t vacc0x67 = vreinterpretq_s32_u64(vmovl_u32(vld1_u32(w))); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
     int32x4_t vacc1x01 = vacc0x01;
     int32x4_t vacc1x23 = vacc0x23;
     int32x4_t vacc1x45 = vacc0x45;

diff --git a/src/qs8-igemm/gen/3x8c4s2-minmax-rndnu-neon-mlal.c b/src/qs8-igemm/gen/3x8c4s2-minmax-rndnu-neon-mlal.c
index b23ba78..fd86b80 100644
--- a/src/qs8-igemm/gen/3x8c4s2-minmax-rndnu-neon-mlal.c
+++ b/src/qs8-igemm/gen/3x8c4s2-minmax-rndnu-neon-mlal.c

@@ -51,10 +51,10 @@
   }
 
   do {
-    int32x4_t vacc0x01 = vreinterpretq_s32_u64(vshll_n_u32(vld1_u32(w), 0)); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
-    int32x4_t vacc0x23 = vreinterpretq_s32_u64(vshll_n_u32(vld1_u32(w), 0)); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
-    int32x4_t vacc0x45 = vreinterpretq_s32_u64(vshll_n_u32(vld1_u32(w), 0)); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
-    int32x4_t vacc0x67 = vreinterpretq_s32_u64(vshll_n_u32(vld1_u32(w), 0)); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
+    int32x4_t vacc0x01 = vreinterpretq_s32_u64(vmovl_u32(vld1_u32(w))); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
+    int32x4_t vacc0x23 = vreinterpretq_s32_u64(vmovl_u32(vld1_u32(w))); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
+    int32x4_t vacc0x45 = vreinterpretq_s32_u64(vmovl_u32(vld1_u32(w))); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
+    int32x4_t vacc0x67 = vreinterpretq_s32_u64(vmovl_u32(vld1_u32(w))); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
     int32x4_t vacc1x01 = vacc0x01;
     int32x4_t vacc1x23 = vacc0x23;
     int32x4_t vacc1x45 = vacc0x45;

diff --git a/src/qs8-igemm/gen/3x8c4s2-minmax-rndnu-neon-mull.c b/src/qs8-igemm/gen/3x8c4s2-minmax-rndnu-neon-mull.c
index b809c0a..f2bc198 100644
--- a/src/qs8-igemm/gen/3x8c4s2-minmax-rndnu-neon-mull.c
+++ b/src/qs8-igemm/gen/3x8c4s2-minmax-rndnu-neon-mull.c

@@ -51,10 +51,10 @@
   }
 
   do {
-    int32x4_t vacc0x01 = vreinterpretq_s32_u64(vshll_n_u32(vld1_u32(w), 0)); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
-    int32x4_t vacc0x23 = vreinterpretq_s32_u64(vshll_n_u32(vld1_u32(w), 0)); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
-    int32x4_t vacc0x45 = vreinterpretq_s32_u64(vshll_n_u32(vld1_u32(w), 0)); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
-    int32x4_t vacc0x67 = vreinterpretq_s32_u64(vshll_n_u32(vld1_u32(w), 0)); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
+    int32x4_t vacc0x01 = vreinterpretq_s32_u64(vmovl_u32(vld1_u32(w))); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
+    int32x4_t vacc0x23 = vreinterpretq_s32_u64(vmovl_u32(vld1_u32(w))); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
+    int32x4_t vacc0x45 = vreinterpretq_s32_u64(vmovl_u32(vld1_u32(w))); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
+    int32x4_t vacc0x67 = vreinterpretq_s32_u64(vmovl_u32(vld1_u32(w))); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
     int32x4_t vacc1x01 = vacc0x01;
     int32x4_t vacc1x23 = vacc0x23;
     int32x4_t vacc1x45 = vacc0x45;

diff --git a/src/qs8-igemm/gen/4x16c4-minmax-rndnu-neon-mlal-dup.c b/src/qs8-igemm/gen/4x16c4-minmax-rndnu-neon-mlal-dup.c
index 0167061..22db6a4 100644
--- a/src/qs8-igemm/gen/4x16c4-minmax-rndnu-neon-mlal-dup.c
+++ b/src/qs8-igemm/gen/4x16c4-minmax-rndnu-neon-mlal-dup.c

@@ -56,14 +56,14 @@
   }
 
   do {
-    int32x4_t vacc0x01 = vreinterpretq_s32_u64(vshll_n_u32(vld1_u32(w), 0)); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
-    int32x4_t vacc0x23 = vreinterpretq_s32_u64(vshll_n_u32(vld1_u32(w), 0)); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
-    int32x4_t vacc0x45 = vreinterpretq_s32_u64(vshll_n_u32(vld1_u32(w), 0)); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
-    int32x4_t vacc0x67 = vreinterpretq_s32_u64(vshll_n_u32(vld1_u32(w), 0)); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
-    int32x4_t vacc0x89 = vreinterpretq_s32_u64(vshll_n_u32(vld1_u32(w), 0)); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
-    int32x4_t vacc0xAB = vreinterpretq_s32_u64(vshll_n_u32(vld1_u32(w), 0)); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
-    int32x4_t vacc0xCD = vreinterpretq_s32_u64(vshll_n_u32(vld1_u32(w), 0)); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
-    int32x4_t vacc0xEF = vreinterpretq_s32_u64(vshll_n_u32(vld1_u32(w), 0)); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
+    int32x4_t vacc0x01 = vreinterpretq_s32_u64(vmovl_u32(vld1_u32(w))); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
+    int32x4_t vacc0x23 = vreinterpretq_s32_u64(vmovl_u32(vld1_u32(w))); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
+    int32x4_t vacc0x45 = vreinterpretq_s32_u64(vmovl_u32(vld1_u32(w))); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
+    int32x4_t vacc0x67 = vreinterpretq_s32_u64(vmovl_u32(vld1_u32(w))); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
+    int32x4_t vacc0x89 = vreinterpretq_s32_u64(vmovl_u32(vld1_u32(w))); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
+    int32x4_t vacc0xAB = vreinterpretq_s32_u64(vmovl_u32(vld1_u32(w))); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
+    int32x4_t vacc0xCD = vreinterpretq_s32_u64(vmovl_u32(vld1_u32(w))); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
+    int32x4_t vacc0xEF = vreinterpretq_s32_u64(vmovl_u32(vld1_u32(w))); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
     int32x4_t vacc1x01 = vacc0x01;
     int32x4_t vacc1x23 = vacc0x23;
     int32x4_t vacc1x45 = vacc0x45;

diff --git a/src/qs8-igemm/gen/4x16c4-minmax-rndnu-neon-mlal-ld1r.c b/src/qs8-igemm/gen/4x16c4-minmax-rndnu-neon-mlal-ld1r.c
index d7ed9f8..a35e318 100644
--- a/src/qs8-igemm/gen/4x16c4-minmax-rndnu-neon-mlal-ld1r.c
+++ b/src/qs8-igemm/gen/4x16c4-minmax-rndnu-neon-mlal-ld1r.c

@@ -56,14 +56,14 @@
   }
 
   do {
-    int32x4_t vacc0x01 = vreinterpretq_s32_u64(vshll_n_u32(vld1_u32(w), 0)); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
-    int32x4_t vacc0x23 = vreinterpretq_s32_u64(vshll_n_u32(vld1_u32(w), 0)); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
-    int32x4_t vacc0x45 = vreinterpretq_s32_u64(vshll_n_u32(vld1_u32(w), 0)); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
-    int32x4_t vacc0x67 = vreinterpretq_s32_u64(vshll_n_u32(vld1_u32(w), 0)); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
-    int32x4_t vacc0x89 = vreinterpretq_s32_u64(vshll_n_u32(vld1_u32(w), 0)); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
-    int32x4_t vacc0xAB = vreinterpretq_s32_u64(vshll_n_u32(vld1_u32(w), 0)); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
-    int32x4_t vacc0xCD = vreinterpretq_s32_u64(vshll_n_u32(vld1_u32(w), 0)); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
-    int32x4_t vacc0xEF = vreinterpretq_s32_u64(vshll_n_u32(vld1_u32(w), 0)); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
+    int32x4_t vacc0x01 = vreinterpretq_s32_u64(vmovl_u32(vld1_u32(w))); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
+    int32x4_t vacc0x23 = vreinterpretq_s32_u64(vmovl_u32(vld1_u32(w))); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
+    int32x4_t vacc0x45 = vreinterpretq_s32_u64(vmovl_u32(vld1_u32(w))); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
+    int32x4_t vacc0x67 = vreinterpretq_s32_u64(vmovl_u32(vld1_u32(w))); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
+    int32x4_t vacc0x89 = vreinterpretq_s32_u64(vmovl_u32(vld1_u32(w))); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
+    int32x4_t vacc0xAB = vreinterpretq_s32_u64(vmovl_u32(vld1_u32(w))); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
+    int32x4_t vacc0xCD = vreinterpretq_s32_u64(vmovl_u32(vld1_u32(w))); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
+    int32x4_t vacc0xEF = vreinterpretq_s32_u64(vmovl_u32(vld1_u32(w))); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
     int32x4_t vacc1x01 = vacc0x01;
     int32x4_t vacc1x23 = vacc0x23;
     int32x4_t vacc1x45 = vacc0x45;

diff --git a/src/qs8-igemm/gen/4x16c4-minmax-rndnu-neon-mlal-ld2r.c b/src/qs8-igemm/gen/4x16c4-minmax-rndnu-neon-mlal-ld2r.c
index 385e433..c9e0ed8 100644
--- a/src/qs8-igemm/gen/4x16c4-minmax-rndnu-neon-mlal-ld2r.c
+++ b/src/qs8-igemm/gen/4x16c4-minmax-rndnu-neon-mlal-ld2r.c

@@ -56,14 +56,14 @@
   }
 
   do {
-    int32x4_t vacc0x01 = vreinterpretq_s32_u64(vshll_n_u32(vld1_u32(w), 0)); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
-    int32x4_t vacc0x23 = vreinterpretq_s32_u64(vshll_n_u32(vld1_u32(w), 0)); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
-    int32x4_t vacc0x45 = vreinterpretq_s32_u64(vshll_n_u32(vld1_u32(w), 0)); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
-    int32x4_t vacc0x67 = vreinterpretq_s32_u64(vshll_n_u32(vld1_u32(w), 0)); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
-    int32x4_t vacc0x89 = vreinterpretq_s32_u64(vshll_n_u32(vld1_u32(w), 0)); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
-    int32x4_t vacc0xAB = vreinterpretq_s32_u64(vshll_n_u32(vld1_u32(w), 0)); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
-    int32x4_t vacc0xCD = vreinterpretq_s32_u64(vshll_n_u32(vld1_u32(w), 0)); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
-    int32x4_t vacc0xEF = vreinterpretq_s32_u64(vshll_n_u32(vld1_u32(w), 0)); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
+    int32x4_t vacc0x01 = vreinterpretq_s32_u64(vmovl_u32(vld1_u32(w))); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
+    int32x4_t vacc0x23 = vreinterpretq_s32_u64(vmovl_u32(vld1_u32(w))); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
+    int32x4_t vacc0x45 = vreinterpretq_s32_u64(vmovl_u32(vld1_u32(w))); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
+    int32x4_t vacc0x67 = vreinterpretq_s32_u64(vmovl_u32(vld1_u32(w))); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
+    int32x4_t vacc0x89 = vreinterpretq_s32_u64(vmovl_u32(vld1_u32(w))); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
+    int32x4_t vacc0xAB = vreinterpretq_s32_u64(vmovl_u32(vld1_u32(w))); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
+    int32x4_t vacc0xCD = vreinterpretq_s32_u64(vmovl_u32(vld1_u32(w))); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
+    int32x4_t vacc0xEF = vreinterpretq_s32_u64(vmovl_u32(vld1_u32(w))); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
     int32x4_t vacc1x01 = vacc0x01;
     int32x4_t vacc1x23 = vacc0x23;
     int32x4_t vacc1x45 = vacc0x45;

diff --git a/src/qs8-igemm/gen/4x16c4-minmax-rndnu-neon-mull-dup.c b/src/qs8-igemm/gen/4x16c4-minmax-rndnu-neon-mull-dup.c
index 6cbdfa2..18761e2 100644
--- a/src/qs8-igemm/gen/4x16c4-minmax-rndnu-neon-mull-dup.c
+++ b/src/qs8-igemm/gen/4x16c4-minmax-rndnu-neon-mull-dup.c

@@ -56,14 +56,14 @@
   }
 
   do {
-    int32x4_t vacc0x01 = vreinterpretq_s32_u64(vshll_n_u32(vld1_u32(w), 0)); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
-    int32x4_t vacc0x23 = vreinterpretq_s32_u64(vshll_n_u32(vld1_u32(w), 0)); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
-    int32x4_t vacc0x45 = vreinterpretq_s32_u64(vshll_n_u32(vld1_u32(w), 0)); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
-    int32x4_t vacc0x67 = vreinterpretq_s32_u64(vshll_n_u32(vld1_u32(w), 0)); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
-    int32x4_t vacc0x89 = vreinterpretq_s32_u64(vshll_n_u32(vld1_u32(w), 0)); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
-    int32x4_t vacc0xAB = vreinterpretq_s32_u64(vshll_n_u32(vld1_u32(w), 0)); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
-    int32x4_t vacc0xCD = vreinterpretq_s32_u64(vshll_n_u32(vld1_u32(w), 0)); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
-    int32x4_t vacc0xEF = vreinterpretq_s32_u64(vshll_n_u32(vld1_u32(w), 0)); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
+    int32x4_t vacc0x01 = vreinterpretq_s32_u64(vmovl_u32(vld1_u32(w))); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
+    int32x4_t vacc0x23 = vreinterpretq_s32_u64(vmovl_u32(vld1_u32(w))); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
+    int32x4_t vacc0x45 = vreinterpretq_s32_u64(vmovl_u32(vld1_u32(w))); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
+    int32x4_t vacc0x67 = vreinterpretq_s32_u64(vmovl_u32(vld1_u32(w))); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
+    int32x4_t vacc0x89 = vreinterpretq_s32_u64(vmovl_u32(vld1_u32(w))); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
+    int32x4_t vacc0xAB = vreinterpretq_s32_u64(vmovl_u32(vld1_u32(w))); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
+    int32x4_t vacc0xCD = vreinterpretq_s32_u64(vmovl_u32(vld1_u32(w))); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
+    int32x4_t vacc0xEF = vreinterpretq_s32_u64(vmovl_u32(vld1_u32(w))); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
     int32x4_t vacc1x01 = vacc0x01;
     int32x4_t vacc1x23 = vacc0x23;
     int32x4_t vacc1x45 = vacc0x45;

diff --git a/src/qs8-igemm/gen/4x16c4-minmax-rndnu-neon-mull-ld1r.c b/src/qs8-igemm/gen/4x16c4-minmax-rndnu-neon-mull-ld1r.c
index 5139a26..17141b8 100644
--- a/src/qs8-igemm/gen/4x16c4-minmax-rndnu-neon-mull-ld1r.c
+++ b/src/qs8-igemm/gen/4x16c4-minmax-rndnu-neon-mull-ld1r.c

@@ -56,14 +56,14 @@
   }
 
   do {
-    int32x4_t vacc0x01 = vreinterpretq_s32_u64(vshll_n_u32(vld1_u32(w), 0)); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
-    int32x4_t vacc0x23 = vreinterpretq_s32_u64(vshll_n_u32(vld1_u32(w), 0)); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
-    int32x4_t vacc0x45 = vreinterpretq_s32_u64(vshll_n_u32(vld1_u32(w), 0)); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
-    int32x4_t vacc0x67 = vreinterpretq_s32_u64(vshll_n_u32(vld1_u32(w), 0)); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
-    int32x4_t vacc0x89 = vreinterpretq_s32_u64(vshll_n_u32(vld1_u32(w), 0)); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
-    int32x4_t vacc0xAB = vreinterpretq_s32_u64(vshll_n_u32(vld1_u32(w), 0)); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
-    int32x4_t vacc0xCD = vreinterpretq_s32_u64(vshll_n_u32(vld1_u32(w), 0)); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
-    int32x4_t vacc0xEF = vreinterpretq_s32_u64(vshll_n_u32(vld1_u32(w), 0)); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
+    int32x4_t vacc0x01 = vreinterpretq_s32_u64(vmovl_u32(vld1_u32(w))); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
+    int32x4_t vacc0x23 = vreinterpretq_s32_u64(vmovl_u32(vld1_u32(w))); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
+    int32x4_t vacc0x45 = vreinterpretq_s32_u64(vmovl_u32(vld1_u32(w))); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
+    int32x4_t vacc0x67 = vreinterpretq_s32_u64(vmovl_u32(vld1_u32(w))); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
+    int32x4_t vacc0x89 = vreinterpretq_s32_u64(vmovl_u32(vld1_u32(w))); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
+    int32x4_t vacc0xAB = vreinterpretq_s32_u64(vmovl_u32(vld1_u32(w))); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
+    int32x4_t vacc0xCD = vreinterpretq_s32_u64(vmovl_u32(vld1_u32(w))); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
+    int32x4_t vacc0xEF = vreinterpretq_s32_u64(vmovl_u32(vld1_u32(w))); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
     int32x4_t vacc1x01 = vacc0x01;
     int32x4_t vacc1x23 = vacc0x23;
     int32x4_t vacc1x45 = vacc0x45;

diff --git a/src/qs8-igemm/gen/4x16c4-minmax-rndnu-neon-mull-ld2r.c b/src/qs8-igemm/gen/4x16c4-minmax-rndnu-neon-mull-ld2r.c
index 3bf269f..fb9fd01 100644
--- a/src/qs8-igemm/gen/4x16c4-minmax-rndnu-neon-mull-ld2r.c
+++ b/src/qs8-igemm/gen/4x16c4-minmax-rndnu-neon-mull-ld2r.c

@@ -56,14 +56,14 @@
   }
 
   do {
-    int32x4_t vacc0x01 = vreinterpretq_s32_u64(vshll_n_u32(vld1_u32(w), 0)); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
-    int32x4_t vacc0x23 = vreinterpretq_s32_u64(vshll_n_u32(vld1_u32(w), 0)); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
-    int32x4_t vacc0x45 = vreinterpretq_s32_u64(vshll_n_u32(vld1_u32(w), 0)); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
-    int32x4_t vacc0x67 = vreinterpretq_s32_u64(vshll_n_u32(vld1_u32(w), 0)); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
-    int32x4_t vacc0x89 = vreinterpretq_s32_u64(vshll_n_u32(vld1_u32(w), 0)); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
-    int32x4_t vacc0xAB = vreinterpretq_s32_u64(vshll_n_u32(vld1_u32(w), 0)); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
-    int32x4_t vacc0xCD = vreinterpretq_s32_u64(vshll_n_u32(vld1_u32(w), 0)); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
-    int32x4_t vacc0xEF = vreinterpretq_s32_u64(vshll_n_u32(vld1_u32(w), 0)); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
+    int32x4_t vacc0x01 = vreinterpretq_s32_u64(vmovl_u32(vld1_u32(w))); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
+    int32x4_t vacc0x23 = vreinterpretq_s32_u64(vmovl_u32(vld1_u32(w))); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
+    int32x4_t vacc0x45 = vreinterpretq_s32_u64(vmovl_u32(vld1_u32(w))); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
+    int32x4_t vacc0x67 = vreinterpretq_s32_u64(vmovl_u32(vld1_u32(w))); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
+    int32x4_t vacc0x89 = vreinterpretq_s32_u64(vmovl_u32(vld1_u32(w))); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
+    int32x4_t vacc0xAB = vreinterpretq_s32_u64(vmovl_u32(vld1_u32(w))); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
+    int32x4_t vacc0xCD = vreinterpretq_s32_u64(vmovl_u32(vld1_u32(w))); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
+    int32x4_t vacc0xEF = vreinterpretq_s32_u64(vmovl_u32(vld1_u32(w))); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
     int32x4_t vacc1x01 = vacc0x01;
     int32x4_t vacc1x23 = vacc0x23;
     int32x4_t vacc1x45 = vacc0x45;

diff --git a/src/qs8-igemm/gen/4x16c4s2-minmax-rndnu-neon-mlal.c b/src/qs8-igemm/gen/4x16c4s2-minmax-rndnu-neon-mlal.c
index 3f7462e..2e65871 100644
--- a/src/qs8-igemm/gen/4x16c4s2-minmax-rndnu-neon-mlal.c
+++ b/src/qs8-igemm/gen/4x16c4s2-minmax-rndnu-neon-mlal.c

@@ -55,14 +55,14 @@
   }
 
   do {
-    int32x4_t vacc0x01 = vreinterpretq_s32_u64(vshll_n_u32(vld1_u32(w), 0)); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
-    int32x4_t vacc0x23 = vreinterpretq_s32_u64(vshll_n_u32(vld1_u32(w), 0)); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
-    int32x4_t vacc0x45 = vreinterpretq_s32_u64(vshll_n_u32(vld1_u32(w), 0)); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
-    int32x4_t vacc0x67 = vreinterpretq_s32_u64(vshll_n_u32(vld1_u32(w), 0)); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
-    int32x4_t vacc0x89 = vreinterpretq_s32_u64(vshll_n_u32(vld1_u32(w), 0)); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
-    int32x4_t vacc0xAB = vreinterpretq_s32_u64(vshll_n_u32(vld1_u32(w), 0)); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
-    int32x4_t vacc0xCD = vreinterpretq_s32_u64(vshll_n_u32(vld1_u32(w), 0)); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
-    int32x4_t vacc0xEF = vreinterpretq_s32_u64(vshll_n_u32(vld1_u32(w), 0)); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
+    int32x4_t vacc0x01 = vreinterpretq_s32_u64(vmovl_u32(vld1_u32(w))); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
+    int32x4_t vacc0x23 = vreinterpretq_s32_u64(vmovl_u32(vld1_u32(w))); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
+    int32x4_t vacc0x45 = vreinterpretq_s32_u64(vmovl_u32(vld1_u32(w))); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
+    int32x4_t vacc0x67 = vreinterpretq_s32_u64(vmovl_u32(vld1_u32(w))); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
+    int32x4_t vacc0x89 = vreinterpretq_s32_u64(vmovl_u32(vld1_u32(w))); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
+    int32x4_t vacc0xAB = vreinterpretq_s32_u64(vmovl_u32(vld1_u32(w))); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
+    int32x4_t vacc0xCD = vreinterpretq_s32_u64(vmovl_u32(vld1_u32(w))); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
+    int32x4_t vacc0xEF = vreinterpretq_s32_u64(vmovl_u32(vld1_u32(w))); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
     int32x4_t vacc1x01 = vacc0x01;
     int32x4_t vacc1x23 = vacc0x23;
     int32x4_t vacc1x45 = vacc0x45;

diff --git a/src/qs8-igemm/gen/4x16c4s2-minmax-rndnu-neon-mull.c b/src/qs8-igemm/gen/4x16c4s2-minmax-rndnu-neon-mull.c
index 9b4fc8d..8d8904c 100644
--- a/src/qs8-igemm/gen/4x16c4s2-minmax-rndnu-neon-mull.c
+++ b/src/qs8-igemm/gen/4x16c4s2-minmax-rndnu-neon-mull.c

@@ -55,14 +55,14 @@
   }
 
   do {
-    int32x4_t vacc0x01 = vreinterpretq_s32_u64(vshll_n_u32(vld1_u32(w), 0)); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
-    int32x4_t vacc0x23 = vreinterpretq_s32_u64(vshll_n_u32(vld1_u32(w), 0)); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
-    int32x4_t vacc0x45 = vreinterpretq_s32_u64(vshll_n_u32(vld1_u32(w), 0)); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
-    int32x4_t vacc0x67 = vreinterpretq_s32_u64(vshll_n_u32(vld1_u32(w), 0)); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
-    int32x4_t vacc0x89 = vreinterpretq_s32_u64(vshll_n_u32(vld1_u32(w), 0)); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
-    int32x4_t vacc0xAB = vreinterpretq_s32_u64(vshll_n_u32(vld1_u32(w), 0)); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
-    int32x4_t vacc0xCD = vreinterpretq_s32_u64(vshll_n_u32(vld1_u32(w), 0)); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
-    int32x4_t vacc0xEF = vreinterpretq_s32_u64(vshll_n_u32(vld1_u32(w), 0)); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
+    int32x4_t vacc0x01 = vreinterpretq_s32_u64(vmovl_u32(vld1_u32(w))); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
+    int32x4_t vacc0x23 = vreinterpretq_s32_u64(vmovl_u32(vld1_u32(w))); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
+    int32x4_t vacc0x45 = vreinterpretq_s32_u64(vmovl_u32(vld1_u32(w))); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
+    int32x4_t vacc0x67 = vreinterpretq_s32_u64(vmovl_u32(vld1_u32(w))); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
+    int32x4_t vacc0x89 = vreinterpretq_s32_u64(vmovl_u32(vld1_u32(w))); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
+    int32x4_t vacc0xAB = vreinterpretq_s32_u64(vmovl_u32(vld1_u32(w))); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
+    int32x4_t vacc0xCD = vreinterpretq_s32_u64(vmovl_u32(vld1_u32(w))); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
+    int32x4_t vacc0xEF = vreinterpretq_s32_u64(vmovl_u32(vld1_u32(w))); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
     int32x4_t vacc1x01 = vacc0x01;
     int32x4_t vacc1x23 = vacc0x23;
     int32x4_t vacc1x45 = vacc0x45;

diff --git a/src/qs8-igemm/gen/4x8c4-minmax-rndnu-neon-mlal-dup.c b/src/qs8-igemm/gen/4x8c4-minmax-rndnu-neon-mlal-dup.c
index beb2ad6..1dca65b 100644
--- a/src/qs8-igemm/gen/4x8c4-minmax-rndnu-neon-mlal-dup.c
+++ b/src/qs8-igemm/gen/4x8c4-minmax-rndnu-neon-mlal-dup.c

@@ -56,10 +56,10 @@
   }
 
   do {
-    int32x4_t vacc0x01 = vreinterpretq_s32_u64(vshll_n_u32(vld1_u32(w), 0)); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
-    int32x4_t vacc0x23 = vreinterpretq_s32_u64(vshll_n_u32(vld1_u32(w), 0)); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
-    int32x4_t vacc0x45 = vreinterpretq_s32_u64(vshll_n_u32(vld1_u32(w), 0)); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
-    int32x4_t vacc0x67 = vreinterpretq_s32_u64(vshll_n_u32(vld1_u32(w), 0)); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
+    int32x4_t vacc0x01 = vreinterpretq_s32_u64(vmovl_u32(vld1_u32(w))); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
+    int32x4_t vacc0x23 = vreinterpretq_s32_u64(vmovl_u32(vld1_u32(w))); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
+    int32x4_t vacc0x45 = vreinterpretq_s32_u64(vmovl_u32(vld1_u32(w))); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
+    int32x4_t vacc0x67 = vreinterpretq_s32_u64(vmovl_u32(vld1_u32(w))); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
     int32x4_t vacc1x01 = vacc0x01;
     int32x4_t vacc1x23 = vacc0x23;
     int32x4_t vacc1x45 = vacc0x45;

diff --git a/src/qs8-igemm/gen/4x8c4-minmax-rndnu-neon-mlal-ld1r.c b/src/qs8-igemm/gen/4x8c4-minmax-rndnu-neon-mlal-ld1r.c
index 76864e2..8a4e740 100644
--- a/src/qs8-igemm/gen/4x8c4-minmax-rndnu-neon-mlal-ld1r.c
+++ b/src/qs8-igemm/gen/4x8c4-minmax-rndnu-neon-mlal-ld1r.c

@@ -56,10 +56,10 @@
   }
 
   do {
-    int32x4_t vacc0x01 = vreinterpretq_s32_u64(vshll_n_u32(vld1_u32(w), 0)); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
-    int32x4_t vacc0x23 = vreinterpretq_s32_u64(vshll_n_u32(vld1_u32(w), 0)); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
-    int32x4_t vacc0x45 = vreinterpretq_s32_u64(vshll_n_u32(vld1_u32(w), 0)); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
-    int32x4_t vacc0x67 = vreinterpretq_s32_u64(vshll_n_u32(vld1_u32(w), 0)); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
+    int32x4_t vacc0x01 = vreinterpretq_s32_u64(vmovl_u32(vld1_u32(w))); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
+    int32x4_t vacc0x23 = vreinterpretq_s32_u64(vmovl_u32(vld1_u32(w))); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
+    int32x4_t vacc0x45 = vreinterpretq_s32_u64(vmovl_u32(vld1_u32(w))); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
+    int32x4_t vacc0x67 = vreinterpretq_s32_u64(vmovl_u32(vld1_u32(w))); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
     int32x4_t vacc1x01 = vacc0x01;
     int32x4_t vacc1x23 = vacc0x23;
     int32x4_t vacc1x45 = vacc0x45;

diff --git a/src/qs8-igemm/gen/4x8c4-minmax-rndnu-neon-mlal-ld2r.c b/src/qs8-igemm/gen/4x8c4-minmax-rndnu-neon-mlal-ld2r.c
index eb171db..71a31bd 100644
--- a/src/qs8-igemm/gen/4x8c4-minmax-rndnu-neon-mlal-ld2r.c
+++ b/src/qs8-igemm/gen/4x8c4-minmax-rndnu-neon-mlal-ld2r.c

@@ -56,10 +56,10 @@
   }
 
   do {
-    int32x4_t vacc0x01 = vreinterpretq_s32_u64(vshll_n_u32(vld1_u32(w), 0)); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
-    int32x4_t vacc0x23 = vreinterpretq_s32_u64(vshll_n_u32(vld1_u32(w), 0)); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
-    int32x4_t vacc0x45 = vreinterpretq_s32_u64(vshll_n_u32(vld1_u32(w), 0)); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
-    int32x4_t vacc0x67 = vreinterpretq_s32_u64(vshll_n_u32(vld1_u32(w), 0)); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
+    int32x4_t vacc0x01 = vreinterpretq_s32_u64(vmovl_u32(vld1_u32(w))); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
+    int32x4_t vacc0x23 = vreinterpretq_s32_u64(vmovl_u32(vld1_u32(w))); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
+    int32x4_t vacc0x45 = vreinterpretq_s32_u64(vmovl_u32(vld1_u32(w))); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
+    int32x4_t vacc0x67 = vreinterpretq_s32_u64(vmovl_u32(vld1_u32(w))); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
     int32x4_t vacc1x01 = vacc0x01;
     int32x4_t vacc1x23 = vacc0x23;
     int32x4_t vacc1x45 = vacc0x45;

diff --git a/src/qs8-igemm/gen/4x8c4-minmax-rndnu-neon-mull-dup.c b/src/qs8-igemm/gen/4x8c4-minmax-rndnu-neon-mull-dup.c
index 574a962..b9d2fa2 100644
--- a/src/qs8-igemm/gen/4x8c4-minmax-rndnu-neon-mull-dup.c
+++ b/src/qs8-igemm/gen/4x8c4-minmax-rndnu-neon-mull-dup.c

@@ -56,10 +56,10 @@
   }
 
   do {
-    int32x4_t vacc0x01 = vreinterpretq_s32_u64(vshll_n_u32(vld1_u32(w), 0)); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
-    int32x4_t vacc0x23 = vreinterpretq_s32_u64(vshll_n_u32(vld1_u32(w), 0)); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
-    int32x4_t vacc0x45 = vreinterpretq_s32_u64(vshll_n_u32(vld1_u32(w), 0)); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
-    int32x4_t vacc0x67 = vreinterpretq_s32_u64(vshll_n_u32(vld1_u32(w), 0)); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
+    int32x4_t vacc0x01 = vreinterpretq_s32_u64(vmovl_u32(vld1_u32(w))); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
+    int32x4_t vacc0x23 = vreinterpretq_s32_u64(vmovl_u32(vld1_u32(w))); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
+    int32x4_t vacc0x45 = vreinterpretq_s32_u64(vmovl_u32(vld1_u32(w))); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
+    int32x4_t vacc0x67 = vreinterpretq_s32_u64(vmovl_u32(vld1_u32(w))); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
     int32x4_t vacc1x01 = vacc0x01;
     int32x4_t vacc1x23 = vacc0x23;
     int32x4_t vacc1x45 = vacc0x45;

diff --git a/src/qs8-igemm/gen/4x8c4-minmax-rndnu-neon-mull-ld1r.c b/src/qs8-igemm/gen/4x8c4-minmax-rndnu-neon-mull-ld1r.c
index 3085dd2..a1f72f1 100644
--- a/src/qs8-igemm/gen/4x8c4-minmax-rndnu-neon-mull-ld1r.c
+++ b/src/qs8-igemm/gen/4x8c4-minmax-rndnu-neon-mull-ld1r.c

@@ -56,10 +56,10 @@
   }
 
   do {
-    int32x4_t vacc0x01 = vreinterpretq_s32_u64(vshll_n_u32(vld1_u32(w), 0)); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
-    int32x4_t vacc0x23 = vreinterpretq_s32_u64(vshll_n_u32(vld1_u32(w), 0)); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
-    int32x4_t vacc0x45 = vreinterpretq_s32_u64(vshll_n_u32(vld1_u32(w), 0)); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
-    int32x4_t vacc0x67 = vreinterpretq_s32_u64(vshll_n_u32(vld1_u32(w), 0)); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
+    int32x4_t vacc0x01 = vreinterpretq_s32_u64(vmovl_u32(vld1_u32(w))); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
+    int32x4_t vacc0x23 = vreinterpretq_s32_u64(vmovl_u32(vld1_u32(w))); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
+    int32x4_t vacc0x45 = vreinterpretq_s32_u64(vmovl_u32(vld1_u32(w))); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
+    int32x4_t vacc0x67 = vreinterpretq_s32_u64(vmovl_u32(vld1_u32(w))); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
     int32x4_t vacc1x01 = vacc0x01;
     int32x4_t vacc1x23 = vacc0x23;
     int32x4_t vacc1x45 = vacc0x45;

diff --git a/src/qs8-igemm/gen/4x8c4-minmax-rndnu-neon-mull-ld2r.c b/src/qs8-igemm/gen/4x8c4-minmax-rndnu-neon-mull-ld2r.c
index fe12cf2..fdf9cc5 100644
--- a/src/qs8-igemm/gen/4x8c4-minmax-rndnu-neon-mull-ld2r.c
+++ b/src/qs8-igemm/gen/4x8c4-minmax-rndnu-neon-mull-ld2r.c

@@ -56,10 +56,10 @@
   }
 
   do {
-    int32x4_t vacc0x01 = vreinterpretq_s32_u64(vshll_n_u32(vld1_u32(w), 0)); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
-    int32x4_t vacc0x23 = vreinterpretq_s32_u64(vshll_n_u32(vld1_u32(w), 0)); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
-    int32x4_t vacc0x45 = vreinterpretq_s32_u64(vshll_n_u32(vld1_u32(w), 0)); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
-    int32x4_t vacc0x67 = vreinterpretq_s32_u64(vshll_n_u32(vld1_u32(w), 0)); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
+    int32x4_t vacc0x01 = vreinterpretq_s32_u64(vmovl_u32(vld1_u32(w))); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
+    int32x4_t vacc0x23 = vreinterpretq_s32_u64(vmovl_u32(vld1_u32(w))); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
+    int32x4_t vacc0x45 = vreinterpretq_s32_u64(vmovl_u32(vld1_u32(w))); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
+    int32x4_t vacc0x67 = vreinterpretq_s32_u64(vmovl_u32(vld1_u32(w))); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
     int32x4_t vacc1x01 = vacc0x01;
     int32x4_t vacc1x23 = vacc0x23;
     int32x4_t vacc1x45 = vacc0x45;

diff --git a/src/qs8-igemm/gen/4x8c4s2-minmax-rndnu-neon-mlal.c b/src/qs8-igemm/gen/4x8c4s2-minmax-rndnu-neon-mlal.c
index fbb3f9a..bb6a629 100644
--- a/src/qs8-igemm/gen/4x8c4s2-minmax-rndnu-neon-mlal.c
+++ b/src/qs8-igemm/gen/4x8c4s2-minmax-rndnu-neon-mlal.c

@@ -55,10 +55,10 @@
   }
 
   do {
-    int32x4_t vacc0x01 = vreinterpretq_s32_u64(vshll_n_u32(vld1_u32(w), 0)); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
-    int32x4_t vacc0x23 = vreinterpretq_s32_u64(vshll_n_u32(vld1_u32(w), 0)); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
-    int32x4_t vacc0x45 = vreinterpretq_s32_u64(vshll_n_u32(vld1_u32(w), 0)); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
-    int32x4_t vacc0x67 = vreinterpretq_s32_u64(vshll_n_u32(vld1_u32(w), 0)); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
+    int32x4_t vacc0x01 = vreinterpretq_s32_u64(vmovl_u32(vld1_u32(w))); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
+    int32x4_t vacc0x23 = vreinterpretq_s32_u64(vmovl_u32(vld1_u32(w))); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
+    int32x4_t vacc0x45 = vreinterpretq_s32_u64(vmovl_u32(vld1_u32(w))); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
+    int32x4_t vacc0x67 = vreinterpretq_s32_u64(vmovl_u32(vld1_u32(w))); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
     int32x4_t vacc1x01 = vacc0x01;
     int32x4_t vacc1x23 = vacc0x23;
     int32x4_t vacc1x45 = vacc0x45;

diff --git a/src/qs8-igemm/gen/4x8c4s2-minmax-rndnu-neon-mull.c b/src/qs8-igemm/gen/4x8c4s2-minmax-rndnu-neon-mull.c
index 50a17b8..4571858 100644
--- a/src/qs8-igemm/gen/4x8c4s2-minmax-rndnu-neon-mull.c
+++ b/src/qs8-igemm/gen/4x8c4s2-minmax-rndnu-neon-mull.c

@@ -55,10 +55,10 @@
   }
 
   do {
-    int32x4_t vacc0x01 = vreinterpretq_s32_u64(vshll_n_u32(vld1_u32(w), 0)); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
-    int32x4_t vacc0x23 = vreinterpretq_s32_u64(vshll_n_u32(vld1_u32(w), 0)); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
-    int32x4_t vacc0x45 = vreinterpretq_s32_u64(vshll_n_u32(vld1_u32(w), 0)); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
-    int32x4_t vacc0x67 = vreinterpretq_s32_u64(vshll_n_u32(vld1_u32(w), 0)); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
+    int32x4_t vacc0x01 = vreinterpretq_s32_u64(vmovl_u32(vld1_u32(w))); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
+    int32x4_t vacc0x23 = vreinterpretq_s32_u64(vmovl_u32(vld1_u32(w))); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
+    int32x4_t vacc0x45 = vreinterpretq_s32_u64(vmovl_u32(vld1_u32(w))); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
+    int32x4_t vacc0x67 = vreinterpretq_s32_u64(vmovl_u32(vld1_u32(w))); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
     int32x4_t vacc1x01 = vacc0x01;
     int32x4_t vacc1x23 = vacc0x23;
     int32x4_t vacc1x45 = vacc0x45;
commit	dc54e12aae2211d39f53d5f77069019aa7903d8d	[log] [tgz]
author	Marat Dukhan <maratek@google.com>	Wed Dec 29 12:56:08 2021 -0800
committer	XNNPACK Team <xnnpack-github-robot@google.com>	Wed Dec 29 12:57:19 2021 -0800
tree	b3d268b9f5b8d06e270f459eceff87e1a1f240e1
parent	e1228e32037d88dcaa0cc710023a3b61e60774a7 [diff]