Unroll KC loop to do MULL and then MLAL to 16 bit before lengthening to 32 bit. PiperOrigin-RevId: 356886292

commit: c8532ae6ac9550ee44b126052ac39cc02ff58052 [log] [tgz]
author: Frank Barchard <fbarchard@google.com> Wed Feb 10 20:01:39 2021 -0800
committer: XNNPACK Team <xnnpack-github-robot@google.com> Wed Feb 10 20:02:26 2021 -0800
tree: 11fec338f8f304e36fac5fdcc2f16be34637d9d9
parent: 4c032f34f870ee35a10ba4c5e203de6cc2565865 [diff]
diff --git a/src/qs8-gemm/c8-neon-mull-padal.c.in b/src/qs8-gemm/c8-neon-mull-padal.c.in
index b7785f9..d3f4756 100644
--- a/src/qs8-gemm/c8-neon-mull-padal.c.in
+++ b/src/qs8-gemm/c8-neon-mull-padal.c.in

@@ -64,15 +64,34 @@
         int32x4_t vacc${M}x${N} = vacc0x${N};
 
     size_t k = kc;
-    while (k >= 8 * sizeof(int8_t)) {
+    while (k >= 16 * sizeof(int8_t)) {
+      $for M in range(MR):
+        const int8x8_t va${M}x0 = vld1_s8(a${M}); a${M} += 8;
+        const int8x8_t va${M}x1 = vld1_s8(a${M}); a${M} += 8;
+
+      $for N in range(0, NR):
+        const int8x8_t vb${N}x0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
+
+      $for N in range(0, NR):
+        const int8x8_t vb${N}x1 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
+        $for M in range(MR):
+          int16x8_t vprod${M}x${N} = vmull_s8(vb${N}x0, va${M}x0);
+        $for M in range(MR):
+          vprod${M}x${N} = vmlal_s8(vprod${M}x${N}, vb${N}x1, va${M}x1);
+        $for M in range(MR):
+          vacc${M}x${N} = vpadalq_s16(vacc${M}x${N}, vprod${M}x${N});
+
+      k -= 16 * sizeof(int8_t);
+    }
+    if (k >= 8 * sizeof(int8_t)) {
       $for M in range(MR):
         const int8x8_t va${M} = vld1_s8(a${M}); a${M} += 8;
 
       $for N in range(0, NR):
         const int8x8_t vb${N} = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
-
         $for M in range(MR):
           const int16x8_t vprod${M}x${N} = vmull_s8(vb${N}, va${M});
+        $for M in range(MR):
           vacc${M}x${N} = vpadalq_s16(vacc${M}x${N}, vprod${M}x${N});
 
       k -= 8 * sizeof(int8_t);
@@ -83,9 +102,9 @@
 
       $for N in range(0, NR):
         const int8x8_t vb${N} = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
-
         $for M in range(MR):
           const int16x8_t vprod${M}x${N} = vmull_s8(vb${N}, va${M});
+        $for M in range(MR):
           vacc${M}x${N} = vpadalq_s16(vacc${M}x${N}, vprod${M}x${N});
     }
 

diff --git a/src/qs8-gemm/gen/1x16c8-minmax-neon-mull-padal.c b/src/qs8-gemm/gen/1x16c8-minmax-neon-mull-padal.c
index 08ba26f..306d130 100644
--- a/src/qs8-gemm/gen/1x16c8-minmax-neon-mull-padal.c
+++ b/src/qs8-gemm/gen/1x16c8-minmax-neon-mull-padal.c

@@ -58,71 +58,143 @@
     int32x4_t vacc0x15 = vld1q_lane_s32(w, vmovq_n_s32(0), 0); w = (const void*) ((uintptr_t) w + sizeof(int32_t));
 
     size_t k = kc;
-    while (k >= 8 * sizeof(int8_t)) {
+    while (k >= 16 * sizeof(int8_t)) {
+      const int8x8_t va0x0 = vld1_s8(a0); a0 += 8;
+      const int8x8_t va0x1 = vld1_s8(a0); a0 += 8;
+
+      const int8x8_t vb0x0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
+      const int8x8_t vb1x0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
+      const int8x8_t vb2x0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
+      const int8x8_t vb3x0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
+      const int8x8_t vb4x0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
+      const int8x8_t vb5x0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
+      const int8x8_t vb6x0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
+      const int8x8_t vb7x0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
+      const int8x8_t vb8x0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
+      const int8x8_t vb9x0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
+      const int8x8_t vb10x0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
+      const int8x8_t vb11x0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
+      const int8x8_t vb12x0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
+      const int8x8_t vb13x0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
+      const int8x8_t vb14x0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
+      const int8x8_t vb15x0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
+
+      const int8x8_t vb0x1 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
+      int16x8_t vprod0x0 = vmull_s8(vb0x0, va0x0);
+      vprod0x0 = vmlal_s8(vprod0x0, vb0x1, va0x1);
+      vacc0x0 = vpadalq_s16(vacc0x0, vprod0x0);
+      const int8x8_t vb1x1 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
+      int16x8_t vprod0x1 = vmull_s8(vb1x0, va0x0);
+      vprod0x1 = vmlal_s8(vprod0x1, vb1x1, va0x1);
+      vacc0x1 = vpadalq_s16(vacc0x1, vprod0x1);
+      const int8x8_t vb2x1 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
+      int16x8_t vprod0x2 = vmull_s8(vb2x0, va0x0);
+      vprod0x2 = vmlal_s8(vprod0x2, vb2x1, va0x1);
+      vacc0x2 = vpadalq_s16(vacc0x2, vprod0x2);
+      const int8x8_t vb3x1 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
+      int16x8_t vprod0x3 = vmull_s8(vb3x0, va0x0);
+      vprod0x3 = vmlal_s8(vprod0x3, vb3x1, va0x1);
+      vacc0x3 = vpadalq_s16(vacc0x3, vprod0x3);
+      const int8x8_t vb4x1 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
+      int16x8_t vprod0x4 = vmull_s8(vb4x0, va0x0);
+      vprod0x4 = vmlal_s8(vprod0x4, vb4x1, va0x1);
+      vacc0x4 = vpadalq_s16(vacc0x4, vprod0x4);
+      const int8x8_t vb5x1 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
+      int16x8_t vprod0x5 = vmull_s8(vb5x0, va0x0);
+      vprod0x5 = vmlal_s8(vprod0x5, vb5x1, va0x1);
+      vacc0x5 = vpadalq_s16(vacc0x5, vprod0x5);
+      const int8x8_t vb6x1 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
+      int16x8_t vprod0x6 = vmull_s8(vb6x0, va0x0);
+      vprod0x6 = vmlal_s8(vprod0x6, vb6x1, va0x1);
+      vacc0x6 = vpadalq_s16(vacc0x6, vprod0x6);
+      const int8x8_t vb7x1 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
+      int16x8_t vprod0x7 = vmull_s8(vb7x0, va0x0);
+      vprod0x7 = vmlal_s8(vprod0x7, vb7x1, va0x1);
+      vacc0x7 = vpadalq_s16(vacc0x7, vprod0x7);
+      const int8x8_t vb8x1 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
+      int16x8_t vprod0x8 = vmull_s8(vb8x0, va0x0);
+      vprod0x8 = vmlal_s8(vprod0x8, vb8x1, va0x1);
+      vacc0x8 = vpadalq_s16(vacc0x8, vprod0x8);
+      const int8x8_t vb9x1 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
+      int16x8_t vprod0x9 = vmull_s8(vb9x0, va0x0);
+      vprod0x9 = vmlal_s8(vprod0x9, vb9x1, va0x1);
+      vacc0x9 = vpadalq_s16(vacc0x9, vprod0x9);
+      const int8x8_t vb10x1 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
+      int16x8_t vprod0x10 = vmull_s8(vb10x0, va0x0);
+      vprod0x10 = vmlal_s8(vprod0x10, vb10x1, va0x1);
+      vacc0x10 = vpadalq_s16(vacc0x10, vprod0x10);
+      const int8x8_t vb11x1 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
+      int16x8_t vprod0x11 = vmull_s8(vb11x0, va0x0);
+      vprod0x11 = vmlal_s8(vprod0x11, vb11x1, va0x1);
+      vacc0x11 = vpadalq_s16(vacc0x11, vprod0x11);
+      const int8x8_t vb12x1 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
+      int16x8_t vprod0x12 = vmull_s8(vb12x0, va0x0);
+      vprod0x12 = vmlal_s8(vprod0x12, vb12x1, va0x1);
+      vacc0x12 = vpadalq_s16(vacc0x12, vprod0x12);
+      const int8x8_t vb13x1 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
+      int16x8_t vprod0x13 = vmull_s8(vb13x0, va0x0);
+      vprod0x13 = vmlal_s8(vprod0x13, vb13x1, va0x1);
+      vacc0x13 = vpadalq_s16(vacc0x13, vprod0x13);
+      const int8x8_t vb14x1 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
+      int16x8_t vprod0x14 = vmull_s8(vb14x0, va0x0);
+      vprod0x14 = vmlal_s8(vprod0x14, vb14x1, va0x1);
+      vacc0x14 = vpadalq_s16(vacc0x14, vprod0x14);
+      const int8x8_t vb15x1 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
+      int16x8_t vprod0x15 = vmull_s8(vb15x0, va0x0);
+      vprod0x15 = vmlal_s8(vprod0x15, vb15x1, va0x1);
+      vacc0x15 = vpadalq_s16(vacc0x15, vprod0x15);
+
+      k -= 16 * sizeof(int8_t);
+    }
+    if (k >= 8 * sizeof(int8_t)) {
       const int8x8_t va0 = vld1_s8(a0); a0 += 8;
 
       const int8x8_t vb0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
-
       const int16x8_t vprod0x0 = vmull_s8(vb0, va0);
       vacc0x0 = vpadalq_s16(vacc0x0, vprod0x0);
       const int8x8_t vb1 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
-
       const int16x8_t vprod0x1 = vmull_s8(vb1, va0);
       vacc0x1 = vpadalq_s16(vacc0x1, vprod0x1);
       const int8x8_t vb2 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
-
       const int16x8_t vprod0x2 = vmull_s8(vb2, va0);
       vacc0x2 = vpadalq_s16(vacc0x2, vprod0x2);
       const int8x8_t vb3 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
-
       const int16x8_t vprod0x3 = vmull_s8(vb3, va0);
       vacc0x3 = vpadalq_s16(vacc0x3, vprod0x3);
       const int8x8_t vb4 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
-
       const int16x8_t vprod0x4 = vmull_s8(vb4, va0);
       vacc0x4 = vpadalq_s16(vacc0x4, vprod0x4);
       const int8x8_t vb5 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
-
       const int16x8_t vprod0x5 = vmull_s8(vb5, va0);
       vacc0x5 = vpadalq_s16(vacc0x5, vprod0x5);
       const int8x8_t vb6 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
-
       const int16x8_t vprod0x6 = vmull_s8(vb6, va0);
       vacc0x6 = vpadalq_s16(vacc0x6, vprod0x6);
       const int8x8_t vb7 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
-
       const int16x8_t vprod0x7 = vmull_s8(vb7, va0);
       vacc0x7 = vpadalq_s16(vacc0x7, vprod0x7);
       const int8x8_t vb8 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
-
       const int16x8_t vprod0x8 = vmull_s8(vb8, va0);
       vacc0x8 = vpadalq_s16(vacc0x8, vprod0x8);
       const int8x8_t vb9 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
-
       const int16x8_t vprod0x9 = vmull_s8(vb9, va0);
       vacc0x9 = vpadalq_s16(vacc0x9, vprod0x9);
       const int8x8_t vb10 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
-
       const int16x8_t vprod0x10 = vmull_s8(vb10, va0);
       vacc0x10 = vpadalq_s16(vacc0x10, vprod0x10);
       const int8x8_t vb11 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
-
       const int16x8_t vprod0x11 = vmull_s8(vb11, va0);
       vacc0x11 = vpadalq_s16(vacc0x11, vprod0x11);
       const int8x8_t vb12 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
-
       const int16x8_t vprod0x12 = vmull_s8(vb12, va0);
       vacc0x12 = vpadalq_s16(vacc0x12, vprod0x12);
       const int8x8_t vb13 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
-
       const int16x8_t vprod0x13 = vmull_s8(vb13, va0);
       vacc0x13 = vpadalq_s16(vacc0x13, vprod0x13);
       const int8x8_t vb14 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
-
       const int16x8_t vprod0x14 = vmull_s8(vb14, va0);
       vacc0x14 = vpadalq_s16(vacc0x14, vprod0x14);
       const int8x8_t vb15 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
-
       const int16x8_t vprod0x15 = vmull_s8(vb15, va0);
       vacc0x15 = vpadalq_s16(vacc0x15, vprod0x15);
 
@@ -132,67 +204,51 @@
       const int8x8_t va0 = vld1_s8(a0); a0 += k;
 
       const int8x8_t vb0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
-
       const int16x8_t vprod0x0 = vmull_s8(vb0, va0);
       vacc0x0 = vpadalq_s16(vacc0x0, vprod0x0);
       const int8x8_t vb1 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
-
       const int16x8_t vprod0x1 = vmull_s8(vb1, va0);
       vacc0x1 = vpadalq_s16(vacc0x1, vprod0x1);
       const int8x8_t vb2 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
-
       const int16x8_t vprod0x2 = vmull_s8(vb2, va0);
       vacc0x2 = vpadalq_s16(vacc0x2, vprod0x2);
       const int8x8_t vb3 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
-
       const int16x8_t vprod0x3 = vmull_s8(vb3, va0);
       vacc0x3 = vpadalq_s16(vacc0x3, vprod0x3);
       const int8x8_t vb4 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
-
       const int16x8_t vprod0x4 = vmull_s8(vb4, va0);
       vacc0x4 = vpadalq_s16(vacc0x4, vprod0x4);
       const int8x8_t vb5 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
-
       const int16x8_t vprod0x5 = vmull_s8(vb5, va0);
       vacc0x5 = vpadalq_s16(vacc0x5, vprod0x5);
       const int8x8_t vb6 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
-
       const int16x8_t vprod0x6 = vmull_s8(vb6, va0);
       vacc0x6 = vpadalq_s16(vacc0x6, vprod0x6);
       const int8x8_t vb7 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
-
       const int16x8_t vprod0x7 = vmull_s8(vb7, va0);
       vacc0x7 = vpadalq_s16(vacc0x7, vprod0x7);
       const int8x8_t vb8 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
-
       const int16x8_t vprod0x8 = vmull_s8(vb8, va0);
       vacc0x8 = vpadalq_s16(vacc0x8, vprod0x8);
       const int8x8_t vb9 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
-
       const int16x8_t vprod0x9 = vmull_s8(vb9, va0);
       vacc0x9 = vpadalq_s16(vacc0x9, vprod0x9);
       const int8x8_t vb10 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
-
       const int16x8_t vprod0x10 = vmull_s8(vb10, va0);
       vacc0x10 = vpadalq_s16(vacc0x10, vprod0x10);
       const int8x8_t vb11 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
-
       const int16x8_t vprod0x11 = vmull_s8(vb11, va0);
       vacc0x11 = vpadalq_s16(vacc0x11, vprod0x11);
       const int8x8_t vb12 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
-
       const int16x8_t vprod0x12 = vmull_s8(vb12, va0);
       vacc0x12 = vpadalq_s16(vacc0x12, vprod0x12);
       const int8x8_t vb13 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
-
       const int16x8_t vprod0x13 = vmull_s8(vb13, va0);
       vacc0x13 = vpadalq_s16(vacc0x13, vprod0x13);
       const int8x8_t vb14 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
-
       const int16x8_t vprod0x14 = vmull_s8(vb14, va0);
       vacc0x14 = vpadalq_s16(vacc0x14, vprod0x14);
       const int8x8_t vb15 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
-
       const int16x8_t vprod0x15 = vmull_s8(vb15, va0);
       vacc0x15 = vpadalq_s16(vacc0x15, vprod0x15);
     }

diff --git a/src/qs8-gemm/gen/1x8c8-minmax-neon-mull-padal.c b/src/qs8-gemm/gen/1x8c8-minmax-neon-mull-padal.c
index 5c2e845..c4c2a58 100644
--- a/src/qs8-gemm/gen/1x8c8-minmax-neon-mull-padal.c
+++ b/src/qs8-gemm/gen/1x8c8-minmax-neon-mull-padal.c

@@ -50,39 +50,79 @@
     int32x4_t vacc0x7 = vld1q_lane_s32(w, vmovq_n_s32(0), 0); w = (const void*) ((uintptr_t) w + sizeof(int32_t));
 
     size_t k = kc;
-    while (k >= 8 * sizeof(int8_t)) {
+    while (k >= 16 * sizeof(int8_t)) {
+      const int8x8_t va0x0 = vld1_s8(a0); a0 += 8;
+      const int8x8_t va0x1 = vld1_s8(a0); a0 += 8;
+
+      const int8x8_t vb0x0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
+      const int8x8_t vb1x0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
+      const int8x8_t vb2x0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
+      const int8x8_t vb3x0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
+      const int8x8_t vb4x0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
+      const int8x8_t vb5x0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
+      const int8x8_t vb6x0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
+      const int8x8_t vb7x0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
+
+      const int8x8_t vb0x1 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
+      int16x8_t vprod0x0 = vmull_s8(vb0x0, va0x0);
+      vprod0x0 = vmlal_s8(vprod0x0, vb0x1, va0x1);
+      vacc0x0 = vpadalq_s16(vacc0x0, vprod0x0);
+      const int8x8_t vb1x1 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
+      int16x8_t vprod0x1 = vmull_s8(vb1x0, va0x0);
+      vprod0x1 = vmlal_s8(vprod0x1, vb1x1, va0x1);
+      vacc0x1 = vpadalq_s16(vacc0x1, vprod0x1);
+      const int8x8_t vb2x1 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
+      int16x8_t vprod0x2 = vmull_s8(vb2x0, va0x0);
+      vprod0x2 = vmlal_s8(vprod0x2, vb2x1, va0x1);
+      vacc0x2 = vpadalq_s16(vacc0x2, vprod0x2);
+      const int8x8_t vb3x1 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
+      int16x8_t vprod0x3 = vmull_s8(vb3x0, va0x0);
+      vprod0x3 = vmlal_s8(vprod0x3, vb3x1, va0x1);
+      vacc0x3 = vpadalq_s16(vacc0x3, vprod0x3);
+      const int8x8_t vb4x1 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
+      int16x8_t vprod0x4 = vmull_s8(vb4x0, va0x0);
+      vprod0x4 = vmlal_s8(vprod0x4, vb4x1, va0x1);
+      vacc0x4 = vpadalq_s16(vacc0x4, vprod0x4);
+      const int8x8_t vb5x1 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
+      int16x8_t vprod0x5 = vmull_s8(vb5x0, va0x0);
+      vprod0x5 = vmlal_s8(vprod0x5, vb5x1, va0x1);
+      vacc0x5 = vpadalq_s16(vacc0x5, vprod0x5);
+      const int8x8_t vb6x1 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
+      int16x8_t vprod0x6 = vmull_s8(vb6x0, va0x0);
+      vprod0x6 = vmlal_s8(vprod0x6, vb6x1, va0x1);
+      vacc0x6 = vpadalq_s16(vacc0x6, vprod0x6);
+      const int8x8_t vb7x1 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
+      int16x8_t vprod0x7 = vmull_s8(vb7x0, va0x0);
+      vprod0x7 = vmlal_s8(vprod0x7, vb7x1, va0x1);
+      vacc0x7 = vpadalq_s16(vacc0x7, vprod0x7);
+
+      k -= 16 * sizeof(int8_t);
+    }
+    if (k >= 8 * sizeof(int8_t)) {
       const int8x8_t va0 = vld1_s8(a0); a0 += 8;
 
       const int8x8_t vb0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
-
       const int16x8_t vprod0x0 = vmull_s8(vb0, va0);
       vacc0x0 = vpadalq_s16(vacc0x0, vprod0x0);
       const int8x8_t vb1 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
-
       const int16x8_t vprod0x1 = vmull_s8(vb1, va0);
       vacc0x1 = vpadalq_s16(vacc0x1, vprod0x1);
       const int8x8_t vb2 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
-
       const int16x8_t vprod0x2 = vmull_s8(vb2, va0);
       vacc0x2 = vpadalq_s16(vacc0x2, vprod0x2);
       const int8x8_t vb3 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
-
       const int16x8_t vprod0x3 = vmull_s8(vb3, va0);
       vacc0x3 = vpadalq_s16(vacc0x3, vprod0x3);
       const int8x8_t vb4 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
-
       const int16x8_t vprod0x4 = vmull_s8(vb4, va0);
       vacc0x4 = vpadalq_s16(vacc0x4, vprod0x4);
       const int8x8_t vb5 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
-
       const int16x8_t vprod0x5 = vmull_s8(vb5, va0);
       vacc0x5 = vpadalq_s16(vacc0x5, vprod0x5);
       const int8x8_t vb6 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
-
       const int16x8_t vprod0x6 = vmull_s8(vb6, va0);
       vacc0x6 = vpadalq_s16(vacc0x6, vprod0x6);
       const int8x8_t vb7 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
-
       const int16x8_t vprod0x7 = vmull_s8(vb7, va0);
       vacc0x7 = vpadalq_s16(vacc0x7, vprod0x7);
 
@@ -92,35 +132,27 @@
       const int8x8_t va0 = vld1_s8(a0); a0 += k;
 
       const int8x8_t vb0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
-
       const int16x8_t vprod0x0 = vmull_s8(vb0, va0);
       vacc0x0 = vpadalq_s16(vacc0x0, vprod0x0);
       const int8x8_t vb1 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
-
       const int16x8_t vprod0x1 = vmull_s8(vb1, va0);
       vacc0x1 = vpadalq_s16(vacc0x1, vprod0x1);
       const int8x8_t vb2 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
-
       const int16x8_t vprod0x2 = vmull_s8(vb2, va0);
       vacc0x2 = vpadalq_s16(vacc0x2, vprod0x2);
       const int8x8_t vb3 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
-
       const int16x8_t vprod0x3 = vmull_s8(vb3, va0);
       vacc0x3 = vpadalq_s16(vacc0x3, vprod0x3);
       const int8x8_t vb4 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
-
       const int16x8_t vprod0x4 = vmull_s8(vb4, va0);
       vacc0x4 = vpadalq_s16(vacc0x4, vprod0x4);
       const int8x8_t vb5 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
-
       const int16x8_t vprod0x5 = vmull_s8(vb5, va0);
       vacc0x5 = vpadalq_s16(vacc0x5, vprod0x5);
       const int8x8_t vb6 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
-
       const int16x8_t vprod0x6 = vmull_s8(vb6, va0);
       vacc0x6 = vpadalq_s16(vacc0x6, vprod0x6);
       const int8x8_t vb7 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
-
       const int16x8_t vprod0x7 = vmull_s8(vb7, va0);
       vacc0x7 = vpadalq_s16(vacc0x7, vprod0x7);
     }

diff --git a/src/qs8-gemm/gen/2x16c8-minmax-neon-mull-padal.c b/src/qs8-gemm/gen/2x16c8-minmax-neon-mull-padal.c
index 647c83e..76f1f81 100644
--- a/src/qs8-gemm/gen/2x16c8-minmax-neon-mull-padal.c
+++ b/src/qs8-gemm/gen/2x16c8-minmax-neon-mull-padal.c

@@ -80,105 +80,227 @@
     int32x4_t vacc1x15 = vacc0x15;
 
     size_t k = kc;
-    while (k >= 8 * sizeof(int8_t)) {
+    while (k >= 16 * sizeof(int8_t)) {
+      const int8x8_t va0x0 = vld1_s8(a0); a0 += 8;
+      const int8x8_t va0x1 = vld1_s8(a0); a0 += 8;
+      const int8x8_t va1x0 = vld1_s8(a1); a1 += 8;
+      const int8x8_t va1x1 = vld1_s8(a1); a1 += 8;
+
+      const int8x8_t vb0x0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
+      const int8x8_t vb1x0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
+      const int8x8_t vb2x0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
+      const int8x8_t vb3x0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
+      const int8x8_t vb4x0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
+      const int8x8_t vb5x0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
+      const int8x8_t vb6x0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
+      const int8x8_t vb7x0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
+      const int8x8_t vb8x0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
+      const int8x8_t vb9x0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
+      const int8x8_t vb10x0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
+      const int8x8_t vb11x0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
+      const int8x8_t vb12x0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
+      const int8x8_t vb13x0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
+      const int8x8_t vb14x0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
+      const int8x8_t vb15x0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
+
+      const int8x8_t vb0x1 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
+      int16x8_t vprod0x0 = vmull_s8(vb0x0, va0x0);
+      int16x8_t vprod1x0 = vmull_s8(vb0x0, va1x0);
+      vprod0x0 = vmlal_s8(vprod0x0, vb0x1, va0x1);
+      vprod1x0 = vmlal_s8(vprod1x0, vb0x1, va1x1);
+      vacc0x0 = vpadalq_s16(vacc0x0, vprod0x0);
+      vacc1x0 = vpadalq_s16(vacc1x0, vprod1x0);
+      const int8x8_t vb1x1 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
+      int16x8_t vprod0x1 = vmull_s8(vb1x0, va0x0);
+      int16x8_t vprod1x1 = vmull_s8(vb1x0, va1x0);
+      vprod0x1 = vmlal_s8(vprod0x1, vb1x1, va0x1);
+      vprod1x1 = vmlal_s8(vprod1x1, vb1x1, va1x1);
+      vacc0x1 = vpadalq_s16(vacc0x1, vprod0x1);
+      vacc1x1 = vpadalq_s16(vacc1x1, vprod1x1);
+      const int8x8_t vb2x1 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
+      int16x8_t vprod0x2 = vmull_s8(vb2x0, va0x0);
+      int16x8_t vprod1x2 = vmull_s8(vb2x0, va1x0);
+      vprod0x2 = vmlal_s8(vprod0x2, vb2x1, va0x1);
+      vprod1x2 = vmlal_s8(vprod1x2, vb2x1, va1x1);
+      vacc0x2 = vpadalq_s16(vacc0x2, vprod0x2);
+      vacc1x2 = vpadalq_s16(vacc1x2, vprod1x2);
+      const int8x8_t vb3x1 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
+      int16x8_t vprod0x3 = vmull_s8(vb3x0, va0x0);
+      int16x8_t vprod1x3 = vmull_s8(vb3x0, va1x0);
+      vprod0x3 = vmlal_s8(vprod0x3, vb3x1, va0x1);
+      vprod1x3 = vmlal_s8(vprod1x3, vb3x1, va1x1);
+      vacc0x3 = vpadalq_s16(vacc0x3, vprod0x3);
+      vacc1x3 = vpadalq_s16(vacc1x3, vprod1x3);
+      const int8x8_t vb4x1 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
+      int16x8_t vprod0x4 = vmull_s8(vb4x0, va0x0);
+      int16x8_t vprod1x4 = vmull_s8(vb4x0, va1x0);
+      vprod0x4 = vmlal_s8(vprod0x4, vb4x1, va0x1);
+      vprod1x4 = vmlal_s8(vprod1x4, vb4x1, va1x1);
+      vacc0x4 = vpadalq_s16(vacc0x4, vprod0x4);
+      vacc1x4 = vpadalq_s16(vacc1x4, vprod1x4);
+      const int8x8_t vb5x1 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
+      int16x8_t vprod0x5 = vmull_s8(vb5x0, va0x0);
+      int16x8_t vprod1x5 = vmull_s8(vb5x0, va1x0);
+      vprod0x5 = vmlal_s8(vprod0x5, vb5x1, va0x1);
+      vprod1x5 = vmlal_s8(vprod1x5, vb5x1, va1x1);
+      vacc0x5 = vpadalq_s16(vacc0x5, vprod0x5);
+      vacc1x5 = vpadalq_s16(vacc1x5, vprod1x5);
+      const int8x8_t vb6x1 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
+      int16x8_t vprod0x6 = vmull_s8(vb6x0, va0x0);
+      int16x8_t vprod1x6 = vmull_s8(vb6x0, va1x0);
+      vprod0x6 = vmlal_s8(vprod0x6, vb6x1, va0x1);
+      vprod1x6 = vmlal_s8(vprod1x6, vb6x1, va1x1);
+      vacc0x6 = vpadalq_s16(vacc0x6, vprod0x6);
+      vacc1x6 = vpadalq_s16(vacc1x6, vprod1x6);
+      const int8x8_t vb7x1 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
+      int16x8_t vprod0x7 = vmull_s8(vb7x0, va0x0);
+      int16x8_t vprod1x7 = vmull_s8(vb7x0, va1x0);
+      vprod0x7 = vmlal_s8(vprod0x7, vb7x1, va0x1);
+      vprod1x7 = vmlal_s8(vprod1x7, vb7x1, va1x1);
+      vacc0x7 = vpadalq_s16(vacc0x7, vprod0x7);
+      vacc1x7 = vpadalq_s16(vacc1x7, vprod1x7);
+      const int8x8_t vb8x1 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
+      int16x8_t vprod0x8 = vmull_s8(vb8x0, va0x0);
+      int16x8_t vprod1x8 = vmull_s8(vb8x0, va1x0);
+      vprod0x8 = vmlal_s8(vprod0x8, vb8x1, va0x1);
+      vprod1x8 = vmlal_s8(vprod1x8, vb8x1, va1x1);
+      vacc0x8 = vpadalq_s16(vacc0x8, vprod0x8);
+      vacc1x8 = vpadalq_s16(vacc1x8, vprod1x8);
+      const int8x8_t vb9x1 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
+      int16x8_t vprod0x9 = vmull_s8(vb9x0, va0x0);
+      int16x8_t vprod1x9 = vmull_s8(vb9x0, va1x0);
+      vprod0x9 = vmlal_s8(vprod0x9, vb9x1, va0x1);
+      vprod1x9 = vmlal_s8(vprod1x9, vb9x1, va1x1);
+      vacc0x9 = vpadalq_s16(vacc0x9, vprod0x9);
+      vacc1x9 = vpadalq_s16(vacc1x9, vprod1x9);
+      const int8x8_t vb10x1 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
+      int16x8_t vprod0x10 = vmull_s8(vb10x0, va0x0);
+      int16x8_t vprod1x10 = vmull_s8(vb10x0, va1x0);
+      vprod0x10 = vmlal_s8(vprod0x10, vb10x1, va0x1);
+      vprod1x10 = vmlal_s8(vprod1x10, vb10x1, va1x1);
+      vacc0x10 = vpadalq_s16(vacc0x10, vprod0x10);
+      vacc1x10 = vpadalq_s16(vacc1x10, vprod1x10);
+      const int8x8_t vb11x1 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
+      int16x8_t vprod0x11 = vmull_s8(vb11x0, va0x0);
+      int16x8_t vprod1x11 = vmull_s8(vb11x0, va1x0);
+      vprod0x11 = vmlal_s8(vprod0x11, vb11x1, va0x1);
+      vprod1x11 = vmlal_s8(vprod1x11, vb11x1, va1x1);
+      vacc0x11 = vpadalq_s16(vacc0x11, vprod0x11);
+      vacc1x11 = vpadalq_s16(vacc1x11, vprod1x11);
+      const int8x8_t vb12x1 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
+      int16x8_t vprod0x12 = vmull_s8(vb12x0, va0x0);
+      int16x8_t vprod1x12 = vmull_s8(vb12x0, va1x0);
+      vprod0x12 = vmlal_s8(vprod0x12, vb12x1, va0x1);
+      vprod1x12 = vmlal_s8(vprod1x12, vb12x1, va1x1);
+      vacc0x12 = vpadalq_s16(vacc0x12, vprod0x12);
+      vacc1x12 = vpadalq_s16(vacc1x12, vprod1x12);
+      const int8x8_t vb13x1 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
+      int16x8_t vprod0x13 = vmull_s8(vb13x0, va0x0);
+      int16x8_t vprod1x13 = vmull_s8(vb13x0, va1x0);
+      vprod0x13 = vmlal_s8(vprod0x13, vb13x1, va0x1);
+      vprod1x13 = vmlal_s8(vprod1x13, vb13x1, va1x1);
+      vacc0x13 = vpadalq_s16(vacc0x13, vprod0x13);
+      vacc1x13 = vpadalq_s16(vacc1x13, vprod1x13);
+      const int8x8_t vb14x1 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
+      int16x8_t vprod0x14 = vmull_s8(vb14x0, va0x0);
+      int16x8_t vprod1x14 = vmull_s8(vb14x0, va1x0);
+      vprod0x14 = vmlal_s8(vprod0x14, vb14x1, va0x1);
+      vprod1x14 = vmlal_s8(vprod1x14, vb14x1, va1x1);
+      vacc0x14 = vpadalq_s16(vacc0x14, vprod0x14);
+      vacc1x14 = vpadalq_s16(vacc1x14, vprod1x14);
+      const int8x8_t vb15x1 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
+      int16x8_t vprod0x15 = vmull_s8(vb15x0, va0x0);
+      int16x8_t vprod1x15 = vmull_s8(vb15x0, va1x0);
+      vprod0x15 = vmlal_s8(vprod0x15, vb15x1, va0x1);
+      vprod1x15 = vmlal_s8(vprod1x15, vb15x1, va1x1);
+      vacc0x15 = vpadalq_s16(vacc0x15, vprod0x15);
+      vacc1x15 = vpadalq_s16(vacc1x15, vprod1x15);
+
+      k -= 16 * sizeof(int8_t);
+    }
+    if (k >= 8 * sizeof(int8_t)) {
       const int8x8_t va0 = vld1_s8(a0); a0 += 8;
       const int8x8_t va1 = vld1_s8(a1); a1 += 8;
 
       const int8x8_t vb0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
-
       const int16x8_t vprod0x0 = vmull_s8(vb0, va0);
-      vacc0x0 = vpadalq_s16(vacc0x0, vprod0x0);
       const int16x8_t vprod1x0 = vmull_s8(vb0, va1);
+      vacc0x0 = vpadalq_s16(vacc0x0, vprod0x0);
       vacc1x0 = vpadalq_s16(vacc1x0, vprod1x0);
       const int8x8_t vb1 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
-
       const int16x8_t vprod0x1 = vmull_s8(vb1, va0);
-      vacc0x1 = vpadalq_s16(vacc0x1, vprod0x1);
       const int16x8_t vprod1x1 = vmull_s8(vb1, va1);
+      vacc0x1 = vpadalq_s16(vacc0x1, vprod0x1);
       vacc1x1 = vpadalq_s16(vacc1x1, vprod1x1);
       const int8x8_t vb2 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
-
       const int16x8_t vprod0x2 = vmull_s8(vb2, va0);
-      vacc0x2 = vpadalq_s16(vacc0x2, vprod0x2);
       const int16x8_t vprod1x2 = vmull_s8(vb2, va1);
+      vacc0x2 = vpadalq_s16(vacc0x2, vprod0x2);
       vacc1x2 = vpadalq_s16(vacc1x2, vprod1x2);
       const int8x8_t vb3 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
-
       const int16x8_t vprod0x3 = vmull_s8(vb3, va0);
-      vacc0x3 = vpadalq_s16(vacc0x3, vprod0x3);
       const int16x8_t vprod1x3 = vmull_s8(vb3, va1);
+      vacc0x3 = vpadalq_s16(vacc0x3, vprod0x3);
       vacc1x3 = vpadalq_s16(vacc1x3, vprod1x3);
       const int8x8_t vb4 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
-
       const int16x8_t vprod0x4 = vmull_s8(vb4, va0);
-      vacc0x4 = vpadalq_s16(vacc0x4, vprod0x4);
       const int16x8_t vprod1x4 = vmull_s8(vb4, va1);
+      vacc0x4 = vpadalq_s16(vacc0x4, vprod0x4);
       vacc1x4 = vpadalq_s16(vacc1x4, vprod1x4);
       const int8x8_t vb5 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
-
       const int16x8_t vprod0x5 = vmull_s8(vb5, va0);
-      vacc0x5 = vpadalq_s16(vacc0x5, vprod0x5);
       const int16x8_t vprod1x5 = vmull_s8(vb5, va1);
+      vacc0x5 = vpadalq_s16(vacc0x5, vprod0x5);
       vacc1x5 = vpadalq_s16(vacc1x5, vprod1x5);
       const int8x8_t vb6 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
-
       const int16x8_t vprod0x6 = vmull_s8(vb6, va0);
-      vacc0x6 = vpadalq_s16(vacc0x6, vprod0x6);
       const int16x8_t vprod1x6 = vmull_s8(vb6, va1);
+      vacc0x6 = vpadalq_s16(vacc0x6, vprod0x6);
       vacc1x6 = vpadalq_s16(vacc1x6, vprod1x6);
       const int8x8_t vb7 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
-
       const int16x8_t vprod0x7 = vmull_s8(vb7, va0);
-      vacc0x7 = vpadalq_s16(vacc0x7, vprod0x7);
       const int16x8_t vprod1x7 = vmull_s8(vb7, va1);
+      vacc0x7 = vpadalq_s16(vacc0x7, vprod0x7);
       vacc1x7 = vpadalq_s16(vacc1x7, vprod1x7);
       const int8x8_t vb8 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
-
       const int16x8_t vprod0x8 = vmull_s8(vb8, va0);
-      vacc0x8 = vpadalq_s16(vacc0x8, vprod0x8);
       const int16x8_t vprod1x8 = vmull_s8(vb8, va1);
+      vacc0x8 = vpadalq_s16(vacc0x8, vprod0x8);
       vacc1x8 = vpadalq_s16(vacc1x8, vprod1x8);
       const int8x8_t vb9 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
-
       const int16x8_t vprod0x9 = vmull_s8(vb9, va0);
-      vacc0x9 = vpadalq_s16(vacc0x9, vprod0x9);
       const int16x8_t vprod1x9 = vmull_s8(vb9, va1);
+      vacc0x9 = vpadalq_s16(vacc0x9, vprod0x9);
       vacc1x9 = vpadalq_s16(vacc1x9, vprod1x9);
       const int8x8_t vb10 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
-
       const int16x8_t vprod0x10 = vmull_s8(vb10, va0);
-      vacc0x10 = vpadalq_s16(vacc0x10, vprod0x10);
       const int16x8_t vprod1x10 = vmull_s8(vb10, va1);
+      vacc0x10 = vpadalq_s16(vacc0x10, vprod0x10);
       vacc1x10 = vpadalq_s16(vacc1x10, vprod1x10);
       const int8x8_t vb11 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
-
       const int16x8_t vprod0x11 = vmull_s8(vb11, va0);
-      vacc0x11 = vpadalq_s16(vacc0x11, vprod0x11);
       const int16x8_t vprod1x11 = vmull_s8(vb11, va1);
+      vacc0x11 = vpadalq_s16(vacc0x11, vprod0x11);
       vacc1x11 = vpadalq_s16(vacc1x11, vprod1x11);
       const int8x8_t vb12 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
-
       const int16x8_t vprod0x12 = vmull_s8(vb12, va0);
-      vacc0x12 = vpadalq_s16(vacc0x12, vprod0x12);
       const int16x8_t vprod1x12 = vmull_s8(vb12, va1);
+      vacc0x12 = vpadalq_s16(vacc0x12, vprod0x12);
       vacc1x12 = vpadalq_s16(vacc1x12, vprod1x12);
       const int8x8_t vb13 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
-
       const int16x8_t vprod0x13 = vmull_s8(vb13, va0);
-      vacc0x13 = vpadalq_s16(vacc0x13, vprod0x13);
       const int16x8_t vprod1x13 = vmull_s8(vb13, va1);
+      vacc0x13 = vpadalq_s16(vacc0x13, vprod0x13);
       vacc1x13 = vpadalq_s16(vacc1x13, vprod1x13);
       const int8x8_t vb14 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
-
       const int16x8_t vprod0x14 = vmull_s8(vb14, va0);
-      vacc0x14 = vpadalq_s16(vacc0x14, vprod0x14);
       const int16x8_t vprod1x14 = vmull_s8(vb14, va1);
+      vacc0x14 = vpadalq_s16(vacc0x14, vprod0x14);
       vacc1x14 = vpadalq_s16(vacc1x14, vprod1x14);
       const int8x8_t vb15 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
-
       const int16x8_t vprod0x15 = vmull_s8(vb15, va0);
-      vacc0x15 = vpadalq_s16(vacc0x15, vprod0x15);
       const int16x8_t vprod1x15 = vmull_s8(vb15, va1);
+      vacc0x15 = vpadalq_s16(vacc0x15, vprod0x15);
       vacc1x15 = vpadalq_s16(vacc1x15, vprod1x15);
 
       k -= 8 * sizeof(int8_t);
@@ -188,100 +310,84 @@
       const int8x8_t va1 = vld1_s8(a1); a1 += k;
 
       const int8x8_t vb0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
-
       const int16x8_t vprod0x0 = vmull_s8(vb0, va0);
-      vacc0x0 = vpadalq_s16(vacc0x0, vprod0x0);
       const int16x8_t vprod1x0 = vmull_s8(vb0, va1);
+      vacc0x0 = vpadalq_s16(vacc0x0, vprod0x0);
       vacc1x0 = vpadalq_s16(vacc1x0, vprod1x0);
       const int8x8_t vb1 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
-
       const int16x8_t vprod0x1 = vmull_s8(vb1, va0);
-      vacc0x1 = vpadalq_s16(vacc0x1, vprod0x1);
       const int16x8_t vprod1x1 = vmull_s8(vb1, va1);
+      vacc0x1 = vpadalq_s16(vacc0x1, vprod0x1);
       vacc1x1 = vpadalq_s16(vacc1x1, vprod1x1);
       const int8x8_t vb2 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
-
       const int16x8_t vprod0x2 = vmull_s8(vb2, va0);
-      vacc0x2 = vpadalq_s16(vacc0x2, vprod0x2);
       const int16x8_t vprod1x2 = vmull_s8(vb2, va1);
+      vacc0x2 = vpadalq_s16(vacc0x2, vprod0x2);
       vacc1x2 = vpadalq_s16(vacc1x2, vprod1x2);
       const int8x8_t vb3 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
-
       const int16x8_t vprod0x3 = vmull_s8(vb3, va0);
-      vacc0x3 = vpadalq_s16(vacc0x3, vprod0x3);
       const int16x8_t vprod1x3 = vmull_s8(vb3, va1);
+      vacc0x3 = vpadalq_s16(vacc0x3, vprod0x3);
       vacc1x3 = vpadalq_s16(vacc1x3, vprod1x3);
       const int8x8_t vb4 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
-
       const int16x8_t vprod0x4 = vmull_s8(vb4, va0);
-      vacc0x4 = vpadalq_s16(vacc0x4, vprod0x4);
       const int16x8_t vprod1x4 = vmull_s8(vb4, va1);
+      vacc0x4 = vpadalq_s16(vacc0x4, vprod0x4);
       vacc1x4 = vpadalq_s16(vacc1x4, vprod1x4);
       const int8x8_t vb5 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
-
       const int16x8_t vprod0x5 = vmull_s8(vb5, va0);
-      vacc0x5 = vpadalq_s16(vacc0x5, vprod0x5);
       const int16x8_t vprod1x5 = vmull_s8(vb5, va1);
+      vacc0x5 = vpadalq_s16(vacc0x5, vprod0x5);
       vacc1x5 = vpadalq_s16(vacc1x5, vprod1x5);
       const int8x8_t vb6 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
-
       const int16x8_t vprod0x6 = vmull_s8(vb6, va0);
-      vacc0x6 = vpadalq_s16(vacc0x6, vprod0x6);
       const int16x8_t vprod1x6 = vmull_s8(vb6, va1);
+      vacc0x6 = vpadalq_s16(vacc0x6, vprod0x6);
       vacc1x6 = vpadalq_s16(vacc1x6, vprod1x6);
       const int8x8_t vb7 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
-
       const int16x8_t vprod0x7 = vmull_s8(vb7, va0);
-      vacc0x7 = vpadalq_s16(vacc0x7, vprod0x7);
       const int16x8_t vprod1x7 = vmull_s8(vb7, va1);
+      vacc0x7 = vpadalq_s16(vacc0x7, vprod0x7);
       vacc1x7 = vpadalq_s16(vacc1x7, vprod1x7);
       const int8x8_t vb8 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
-
       const int16x8_t vprod0x8 = vmull_s8(vb8, va0);
-      vacc0x8 = vpadalq_s16(vacc0x8, vprod0x8);
       const int16x8_t vprod1x8 = vmull_s8(vb8, va1);
+      vacc0x8 = vpadalq_s16(vacc0x8, vprod0x8);
       vacc1x8 = vpadalq_s16(vacc1x8, vprod1x8);
       const int8x8_t vb9 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
-
       const int16x8_t vprod0x9 = vmull_s8(vb9, va0);
-      vacc0x9 = vpadalq_s16(vacc0x9, vprod0x9);
       const int16x8_t vprod1x9 = vmull_s8(vb9, va1);
+      vacc0x9 = vpadalq_s16(vacc0x9, vprod0x9);
       vacc1x9 = vpadalq_s16(vacc1x9, vprod1x9);
       const int8x8_t vb10 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
-
       const int16x8_t vprod0x10 = vmull_s8(vb10, va0);
-      vacc0x10 = vpadalq_s16(vacc0x10, vprod0x10);
       const int16x8_t vprod1x10 = vmull_s8(vb10, va1);
+      vacc0x10 = vpadalq_s16(vacc0x10, vprod0x10);
       vacc1x10 = vpadalq_s16(vacc1x10, vprod1x10);
       const int8x8_t vb11 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
-
       const int16x8_t vprod0x11 = vmull_s8(vb11, va0);
-      vacc0x11 = vpadalq_s16(vacc0x11, vprod0x11);
       const int16x8_t vprod1x11 = vmull_s8(vb11, va1);
+      vacc0x11 = vpadalq_s16(vacc0x11, vprod0x11);
       vacc1x11 = vpadalq_s16(vacc1x11, vprod1x11);
       const int8x8_t vb12 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
-
       const int16x8_t vprod0x12 = vmull_s8(vb12, va0);
-      vacc0x12 = vpadalq_s16(vacc0x12, vprod0x12);
       const int16x8_t vprod1x12 = vmull_s8(vb12, va1);
+      vacc0x12 = vpadalq_s16(vacc0x12, vprod0x12);
       vacc1x12 = vpadalq_s16(vacc1x12, vprod1x12);
       const int8x8_t vb13 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
-
       const int16x8_t vprod0x13 = vmull_s8(vb13, va0);
-      vacc0x13 = vpadalq_s16(vacc0x13, vprod0x13);
       const int16x8_t vprod1x13 = vmull_s8(vb13, va1);
+      vacc0x13 = vpadalq_s16(vacc0x13, vprod0x13);
       vacc1x13 = vpadalq_s16(vacc1x13, vprod1x13);
       const int8x8_t vb14 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
-
       const int16x8_t vprod0x14 = vmull_s8(vb14, va0);
-      vacc0x14 = vpadalq_s16(vacc0x14, vprod0x14);
       const int16x8_t vprod1x14 = vmull_s8(vb14, va1);
+      vacc0x14 = vpadalq_s16(vacc0x14, vprod0x14);
       vacc1x14 = vpadalq_s16(vacc1x14, vprod1x14);
       const int8x8_t vb15 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
-
       const int16x8_t vprod0x15 = vmull_s8(vb15, va0);
-      vacc0x15 = vpadalq_s16(vacc0x15, vprod0x15);
       const int16x8_t vprod1x15 = vmull_s8(vb15, va1);
+      vacc0x15 = vpadalq_s16(vacc0x15, vprod0x15);
       vacc1x15 = vpadalq_s16(vacc1x15, vprod1x15);
     }
 

diff --git a/src/qs8-gemm/gen/2x8c8-minmax-neon-mull-padal.c b/src/qs8-gemm/gen/2x8c8-minmax-neon-mull-padal.c
index b49cf81..034dceb 100644
--- a/src/qs8-gemm/gen/2x8c8-minmax-neon-mull-padal.c
+++ b/src/qs8-gemm/gen/2x8c8-minmax-neon-mull-padal.c

@@ -64,57 +64,123 @@
     int32x4_t vacc1x7 = vacc0x7;
 
     size_t k = kc;
-    while (k >= 8 * sizeof(int8_t)) {
+    while (k >= 16 * sizeof(int8_t)) {
+      const int8x8_t va0x0 = vld1_s8(a0); a0 += 8;
+      const int8x8_t va0x1 = vld1_s8(a0); a0 += 8;
+      const int8x8_t va1x0 = vld1_s8(a1); a1 += 8;
+      const int8x8_t va1x1 = vld1_s8(a1); a1 += 8;
+
+      const int8x8_t vb0x0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
+      const int8x8_t vb1x0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
+      const int8x8_t vb2x0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
+      const int8x8_t vb3x0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
+      const int8x8_t vb4x0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
+      const int8x8_t vb5x0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
+      const int8x8_t vb6x0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
+      const int8x8_t vb7x0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
+
+      const int8x8_t vb0x1 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
+      int16x8_t vprod0x0 = vmull_s8(vb0x0, va0x0);
+      int16x8_t vprod1x0 = vmull_s8(vb0x0, va1x0);
+      vprod0x0 = vmlal_s8(vprod0x0, vb0x1, va0x1);
+      vprod1x0 = vmlal_s8(vprod1x0, vb0x1, va1x1);
+      vacc0x0 = vpadalq_s16(vacc0x0, vprod0x0);
+      vacc1x0 = vpadalq_s16(vacc1x0, vprod1x0);
+      const int8x8_t vb1x1 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
+      int16x8_t vprod0x1 = vmull_s8(vb1x0, va0x0);
+      int16x8_t vprod1x1 = vmull_s8(vb1x0, va1x0);
+      vprod0x1 = vmlal_s8(vprod0x1, vb1x1, va0x1);
+      vprod1x1 = vmlal_s8(vprod1x1, vb1x1, va1x1);
+      vacc0x1 = vpadalq_s16(vacc0x1, vprod0x1);
+      vacc1x1 = vpadalq_s16(vacc1x1, vprod1x1);
+      const int8x8_t vb2x1 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
+      int16x8_t vprod0x2 = vmull_s8(vb2x0, va0x0);
+      int16x8_t vprod1x2 = vmull_s8(vb2x0, va1x0);
+      vprod0x2 = vmlal_s8(vprod0x2, vb2x1, va0x1);
+      vprod1x2 = vmlal_s8(vprod1x2, vb2x1, va1x1);
+      vacc0x2 = vpadalq_s16(vacc0x2, vprod0x2);
+      vacc1x2 = vpadalq_s16(vacc1x2, vprod1x2);
+      const int8x8_t vb3x1 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
+      int16x8_t vprod0x3 = vmull_s8(vb3x0, va0x0);
+      int16x8_t vprod1x3 = vmull_s8(vb3x0, va1x0);
+      vprod0x3 = vmlal_s8(vprod0x3, vb3x1, va0x1);
+      vprod1x3 = vmlal_s8(vprod1x3, vb3x1, va1x1);
+      vacc0x3 = vpadalq_s16(vacc0x3, vprod0x3);
+      vacc1x3 = vpadalq_s16(vacc1x3, vprod1x3);
+      const int8x8_t vb4x1 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
+      int16x8_t vprod0x4 = vmull_s8(vb4x0, va0x0);
+      int16x8_t vprod1x4 = vmull_s8(vb4x0, va1x0);
+      vprod0x4 = vmlal_s8(vprod0x4, vb4x1, va0x1);
+      vprod1x4 = vmlal_s8(vprod1x4, vb4x1, va1x1);
+      vacc0x4 = vpadalq_s16(vacc0x4, vprod0x4);
+      vacc1x4 = vpadalq_s16(vacc1x4, vprod1x4);
+      const int8x8_t vb5x1 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
+      int16x8_t vprod0x5 = vmull_s8(vb5x0, va0x0);
+      int16x8_t vprod1x5 = vmull_s8(vb5x0, va1x0);
+      vprod0x5 = vmlal_s8(vprod0x5, vb5x1, va0x1);
+      vprod1x5 = vmlal_s8(vprod1x5, vb5x1, va1x1);
+      vacc0x5 = vpadalq_s16(vacc0x5, vprod0x5);
+      vacc1x5 = vpadalq_s16(vacc1x5, vprod1x5);
+      const int8x8_t vb6x1 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
+      int16x8_t vprod0x6 = vmull_s8(vb6x0, va0x0);
+      int16x8_t vprod1x6 = vmull_s8(vb6x0, va1x0);
+      vprod0x6 = vmlal_s8(vprod0x6, vb6x1, va0x1);
+      vprod1x6 = vmlal_s8(vprod1x6, vb6x1, va1x1);
+      vacc0x6 = vpadalq_s16(vacc0x6, vprod0x6);
+      vacc1x6 = vpadalq_s16(vacc1x6, vprod1x6);
+      const int8x8_t vb7x1 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
+      int16x8_t vprod0x7 = vmull_s8(vb7x0, va0x0);
+      int16x8_t vprod1x7 = vmull_s8(vb7x0, va1x0);
+      vprod0x7 = vmlal_s8(vprod0x7, vb7x1, va0x1);
+      vprod1x7 = vmlal_s8(vprod1x7, vb7x1, va1x1);
+      vacc0x7 = vpadalq_s16(vacc0x7, vprod0x7);
+      vacc1x7 = vpadalq_s16(vacc1x7, vprod1x7);
+
+      k -= 16 * sizeof(int8_t);
+    }
+    if (k >= 8 * sizeof(int8_t)) {
       const int8x8_t va0 = vld1_s8(a0); a0 += 8;
       const int8x8_t va1 = vld1_s8(a1); a1 += 8;
 
       const int8x8_t vb0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
-
       const int16x8_t vprod0x0 = vmull_s8(vb0, va0);
-      vacc0x0 = vpadalq_s16(vacc0x0, vprod0x0);
       const int16x8_t vprod1x0 = vmull_s8(vb0, va1);
+      vacc0x0 = vpadalq_s16(vacc0x0, vprod0x0);
       vacc1x0 = vpadalq_s16(vacc1x0, vprod1x0);
       const int8x8_t vb1 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
-
       const int16x8_t vprod0x1 = vmull_s8(vb1, va0);
-      vacc0x1 = vpadalq_s16(vacc0x1, vprod0x1);
       const int16x8_t vprod1x1 = vmull_s8(vb1, va1);
+      vacc0x1 = vpadalq_s16(vacc0x1, vprod0x1);
       vacc1x1 = vpadalq_s16(vacc1x1, vprod1x1);
       const int8x8_t vb2 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
-
       const int16x8_t vprod0x2 = vmull_s8(vb2, va0);
-      vacc0x2 = vpadalq_s16(vacc0x2, vprod0x2);
       const int16x8_t vprod1x2 = vmull_s8(vb2, va1);
+      vacc0x2 = vpadalq_s16(vacc0x2, vprod0x2);
       vacc1x2 = vpadalq_s16(vacc1x2, vprod1x2);
       const int8x8_t vb3 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
-
       const int16x8_t vprod0x3 = vmull_s8(vb3, va0);
-      vacc0x3 = vpadalq_s16(vacc0x3, vprod0x3);
       const int16x8_t vprod1x3 = vmull_s8(vb3, va1);
+      vacc0x3 = vpadalq_s16(vacc0x3, vprod0x3);
       vacc1x3 = vpadalq_s16(vacc1x3, vprod1x3);
       const int8x8_t vb4 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
-
       const int16x8_t vprod0x4 = vmull_s8(vb4, va0);
-      vacc0x4 = vpadalq_s16(vacc0x4, vprod0x4);
       const int16x8_t vprod1x4 = vmull_s8(vb4, va1);
+      vacc0x4 = vpadalq_s16(vacc0x4, vprod0x4);
       vacc1x4 = vpadalq_s16(vacc1x4, vprod1x4);
       const int8x8_t vb5 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
-
       const int16x8_t vprod0x5 = vmull_s8(vb5, va0);
-      vacc0x5 = vpadalq_s16(vacc0x5, vprod0x5);
       const int16x8_t vprod1x5 = vmull_s8(vb5, va1);
+      vacc0x5 = vpadalq_s16(vacc0x5, vprod0x5);
       vacc1x5 = vpadalq_s16(vacc1x5, vprod1x5);
       const int8x8_t vb6 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
-
       const int16x8_t vprod0x6 = vmull_s8(vb6, va0);
-      vacc0x6 = vpadalq_s16(vacc0x6, vprod0x6);
       const int16x8_t vprod1x6 = vmull_s8(vb6, va1);
+      vacc0x6 = vpadalq_s16(vacc0x6, vprod0x6);
       vacc1x6 = vpadalq_s16(vacc1x6, vprod1x6);
       const int8x8_t vb7 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
-
       const int16x8_t vprod0x7 = vmull_s8(vb7, va0);
-      vacc0x7 = vpadalq_s16(vacc0x7, vprod0x7);
       const int16x8_t vprod1x7 = vmull_s8(vb7, va1);
+      vacc0x7 = vpadalq_s16(vacc0x7, vprod0x7);
       vacc1x7 = vpadalq_s16(vacc1x7, vprod1x7);
 
       k -= 8 * sizeof(int8_t);
@@ -124,52 +190,44 @@
       const int8x8_t va1 = vld1_s8(a1); a1 += k;
 
       const int8x8_t vb0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
-
       const int16x8_t vprod0x0 = vmull_s8(vb0, va0);
-      vacc0x0 = vpadalq_s16(vacc0x0, vprod0x0);
       const int16x8_t vprod1x0 = vmull_s8(vb0, va1);
+      vacc0x0 = vpadalq_s16(vacc0x0, vprod0x0);
       vacc1x0 = vpadalq_s16(vacc1x0, vprod1x0);
       const int8x8_t vb1 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
-
       const int16x8_t vprod0x1 = vmull_s8(vb1, va0);
-      vacc0x1 = vpadalq_s16(vacc0x1, vprod0x1);
       const int16x8_t vprod1x1 = vmull_s8(vb1, va1);
+      vacc0x1 = vpadalq_s16(vacc0x1, vprod0x1);
       vacc1x1 = vpadalq_s16(vacc1x1, vprod1x1);
       const int8x8_t vb2 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
-
       const int16x8_t vprod0x2 = vmull_s8(vb2, va0);
-      vacc0x2 = vpadalq_s16(vacc0x2, vprod0x2);
       const int16x8_t vprod1x2 = vmull_s8(vb2, va1);
+      vacc0x2 = vpadalq_s16(vacc0x2, vprod0x2);
       vacc1x2 = vpadalq_s16(vacc1x2, vprod1x2);
       const int8x8_t vb3 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
-
       const int16x8_t vprod0x3 = vmull_s8(vb3, va0);
-      vacc0x3 = vpadalq_s16(vacc0x3, vprod0x3);
       const int16x8_t vprod1x3 = vmull_s8(vb3, va1);
+      vacc0x3 = vpadalq_s16(vacc0x3, vprod0x3);
       vacc1x3 = vpadalq_s16(vacc1x3, vprod1x3);
       const int8x8_t vb4 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
-
       const int16x8_t vprod0x4 = vmull_s8(vb4, va0);
-      vacc0x4 = vpadalq_s16(vacc0x4, vprod0x4);
       const int16x8_t vprod1x4 = vmull_s8(vb4, va1);
+      vacc0x4 = vpadalq_s16(vacc0x4, vprod0x4);
       vacc1x4 = vpadalq_s16(vacc1x4, vprod1x4);
       const int8x8_t vb5 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
-
       const int16x8_t vprod0x5 = vmull_s8(vb5, va0);
-      vacc0x5 = vpadalq_s16(vacc0x5, vprod0x5);
       const int16x8_t vprod1x5 = vmull_s8(vb5, va1);
+      vacc0x5 = vpadalq_s16(vacc0x5, vprod0x5);
       vacc1x5 = vpadalq_s16(vacc1x5, vprod1x5);
       const int8x8_t vb6 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
-
       const int16x8_t vprod0x6 = vmull_s8(vb6, va0);
-      vacc0x6 = vpadalq_s16(vacc0x6, vprod0x6);
       const int16x8_t vprod1x6 = vmull_s8(vb6, va1);
+      vacc0x6 = vpadalq_s16(vacc0x6, vprod0x6);
       vacc1x6 = vpadalq_s16(vacc1x6, vprod1x6);
       const int8x8_t vb7 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
-
       const int16x8_t vprod0x7 = vmull_s8(vb7, va0);
-      vacc0x7 = vpadalq_s16(vacc0x7, vprod0x7);
       const int16x8_t vprod1x7 = vmull_s8(vb7, va1);
+      vacc0x7 = vpadalq_s16(vacc0x7, vprod0x7);
       vacc1x7 = vpadalq_s16(vacc1x7, vprod1x7);
     }
 

diff --git a/src/qs8-gemm/gen/3x16c8-minmax-neon-mull-padal.c b/src/qs8-gemm/gen/3x16c8-minmax-neon-mull-padal.c
index f7c9e38..9c3fca7 100644
--- a/src/qs8-gemm/gen/3x16c8-minmax-neon-mull-padal.c
+++ b/src/qs8-gemm/gen/3x16c8-minmax-neon-mull-padal.c

@@ -102,138 +102,310 @@
     int32x4_t vacc2x15 = vacc0x15;
 
     size_t k = kc;
-    while (k >= 8 * sizeof(int8_t)) {
+    while (k >= 16 * sizeof(int8_t)) {
+      const int8x8_t va0x0 = vld1_s8(a0); a0 += 8;
+      const int8x8_t va0x1 = vld1_s8(a0); a0 += 8;
+      const int8x8_t va1x0 = vld1_s8(a1); a1 += 8;
+      const int8x8_t va1x1 = vld1_s8(a1); a1 += 8;
+      const int8x8_t va2x0 = vld1_s8(a2); a2 += 8;
+      const int8x8_t va2x1 = vld1_s8(a2); a2 += 8;
+
+      const int8x8_t vb0x0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
+      const int8x8_t vb1x0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
+      const int8x8_t vb2x0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
+      const int8x8_t vb3x0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
+      const int8x8_t vb4x0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
+      const int8x8_t vb5x0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
+      const int8x8_t vb6x0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
+      const int8x8_t vb7x0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
+      const int8x8_t vb8x0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
+      const int8x8_t vb9x0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
+      const int8x8_t vb10x0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
+      const int8x8_t vb11x0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
+      const int8x8_t vb12x0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
+      const int8x8_t vb13x0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
+      const int8x8_t vb14x0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
+      const int8x8_t vb15x0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
+
+      const int8x8_t vb0x1 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
+      int16x8_t vprod0x0 = vmull_s8(vb0x0, va0x0);
+      int16x8_t vprod1x0 = vmull_s8(vb0x0, va1x0);
+      int16x8_t vprod2x0 = vmull_s8(vb0x0, va2x0);
+      vprod0x0 = vmlal_s8(vprod0x0, vb0x1, va0x1);
+      vprod1x0 = vmlal_s8(vprod1x0, vb0x1, va1x1);
+      vprod2x0 = vmlal_s8(vprod2x0, vb0x1, va2x1);
+      vacc0x0 = vpadalq_s16(vacc0x0, vprod0x0);
+      vacc1x0 = vpadalq_s16(vacc1x0, vprod1x0);
+      vacc2x0 = vpadalq_s16(vacc2x0, vprod2x0);
+      const int8x8_t vb1x1 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
+      int16x8_t vprod0x1 = vmull_s8(vb1x0, va0x0);
+      int16x8_t vprod1x1 = vmull_s8(vb1x0, va1x0);
+      int16x8_t vprod2x1 = vmull_s8(vb1x0, va2x0);
+      vprod0x1 = vmlal_s8(vprod0x1, vb1x1, va0x1);
+      vprod1x1 = vmlal_s8(vprod1x1, vb1x1, va1x1);
+      vprod2x1 = vmlal_s8(vprod2x1, vb1x1, va2x1);
+      vacc0x1 = vpadalq_s16(vacc0x1, vprod0x1);
+      vacc1x1 = vpadalq_s16(vacc1x1, vprod1x1);
+      vacc2x1 = vpadalq_s16(vacc2x1, vprod2x1);
+      const int8x8_t vb2x1 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
+      int16x8_t vprod0x2 = vmull_s8(vb2x0, va0x0);
+      int16x8_t vprod1x2 = vmull_s8(vb2x0, va1x0);
+      int16x8_t vprod2x2 = vmull_s8(vb2x0, va2x0);
+      vprod0x2 = vmlal_s8(vprod0x2, vb2x1, va0x1);
+      vprod1x2 = vmlal_s8(vprod1x2, vb2x1, va1x1);
+      vprod2x2 = vmlal_s8(vprod2x2, vb2x1, va2x1);
+      vacc0x2 = vpadalq_s16(vacc0x2, vprod0x2);
+      vacc1x2 = vpadalq_s16(vacc1x2, vprod1x2);
+      vacc2x2 = vpadalq_s16(vacc2x2, vprod2x2);
+      const int8x8_t vb3x1 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
+      int16x8_t vprod0x3 = vmull_s8(vb3x0, va0x0);
+      int16x8_t vprod1x3 = vmull_s8(vb3x0, va1x0);
+      int16x8_t vprod2x3 = vmull_s8(vb3x0, va2x0);
+      vprod0x3 = vmlal_s8(vprod0x3, vb3x1, va0x1);
+      vprod1x3 = vmlal_s8(vprod1x3, vb3x1, va1x1);
+      vprod2x3 = vmlal_s8(vprod2x3, vb3x1, va2x1);
+      vacc0x3 = vpadalq_s16(vacc0x3, vprod0x3);
+      vacc1x3 = vpadalq_s16(vacc1x3, vprod1x3);
+      vacc2x3 = vpadalq_s16(vacc2x3, vprod2x3);
+      const int8x8_t vb4x1 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
+      int16x8_t vprod0x4 = vmull_s8(vb4x0, va0x0);
+      int16x8_t vprod1x4 = vmull_s8(vb4x0, va1x0);
+      int16x8_t vprod2x4 = vmull_s8(vb4x0, va2x0);
+      vprod0x4 = vmlal_s8(vprod0x4, vb4x1, va0x1);
+      vprod1x4 = vmlal_s8(vprod1x4, vb4x1, va1x1);
+      vprod2x4 = vmlal_s8(vprod2x4, vb4x1, va2x1);
+      vacc0x4 = vpadalq_s16(vacc0x4, vprod0x4);
+      vacc1x4 = vpadalq_s16(vacc1x4, vprod1x4);
+      vacc2x4 = vpadalq_s16(vacc2x4, vprod2x4);
+      const int8x8_t vb5x1 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
+      int16x8_t vprod0x5 = vmull_s8(vb5x0, va0x0);
+      int16x8_t vprod1x5 = vmull_s8(vb5x0, va1x0);
+      int16x8_t vprod2x5 = vmull_s8(vb5x0, va2x0);
+      vprod0x5 = vmlal_s8(vprod0x5, vb5x1, va0x1);
+      vprod1x5 = vmlal_s8(vprod1x5, vb5x1, va1x1);
+      vprod2x5 = vmlal_s8(vprod2x5, vb5x1, va2x1);
+      vacc0x5 = vpadalq_s16(vacc0x5, vprod0x5);
+      vacc1x5 = vpadalq_s16(vacc1x5, vprod1x5);
+      vacc2x5 = vpadalq_s16(vacc2x5, vprod2x5);
+      const int8x8_t vb6x1 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
+      int16x8_t vprod0x6 = vmull_s8(vb6x0, va0x0);
+      int16x8_t vprod1x6 = vmull_s8(vb6x0, va1x0);
+      int16x8_t vprod2x6 = vmull_s8(vb6x0, va2x0);
+      vprod0x6 = vmlal_s8(vprod0x6, vb6x1, va0x1);
+      vprod1x6 = vmlal_s8(vprod1x6, vb6x1, va1x1);
+      vprod2x6 = vmlal_s8(vprod2x6, vb6x1, va2x1);
+      vacc0x6 = vpadalq_s16(vacc0x6, vprod0x6);
+      vacc1x6 = vpadalq_s16(vacc1x6, vprod1x6);
+      vacc2x6 = vpadalq_s16(vacc2x6, vprod2x6);
+      const int8x8_t vb7x1 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
+      int16x8_t vprod0x7 = vmull_s8(vb7x0, va0x0);
+      int16x8_t vprod1x7 = vmull_s8(vb7x0, va1x0);
+      int16x8_t vprod2x7 = vmull_s8(vb7x0, va2x0);
+      vprod0x7 = vmlal_s8(vprod0x7, vb7x1, va0x1);
+      vprod1x7 = vmlal_s8(vprod1x7, vb7x1, va1x1);
+      vprod2x7 = vmlal_s8(vprod2x7, vb7x1, va2x1);
+      vacc0x7 = vpadalq_s16(vacc0x7, vprod0x7);
+      vacc1x7 = vpadalq_s16(vacc1x7, vprod1x7);
+      vacc2x7 = vpadalq_s16(vacc2x7, vprod2x7);
+      const int8x8_t vb8x1 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
+      int16x8_t vprod0x8 = vmull_s8(vb8x0, va0x0);
+      int16x8_t vprod1x8 = vmull_s8(vb8x0, va1x0);
+      int16x8_t vprod2x8 = vmull_s8(vb8x0, va2x0);
+      vprod0x8 = vmlal_s8(vprod0x8, vb8x1, va0x1);
+      vprod1x8 = vmlal_s8(vprod1x8, vb8x1, va1x1);
+      vprod2x8 = vmlal_s8(vprod2x8, vb8x1, va2x1);
+      vacc0x8 = vpadalq_s16(vacc0x8, vprod0x8);
+      vacc1x8 = vpadalq_s16(vacc1x8, vprod1x8);
+      vacc2x8 = vpadalq_s16(vacc2x8, vprod2x8);
+      const int8x8_t vb9x1 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
+      int16x8_t vprod0x9 = vmull_s8(vb9x0, va0x0);
+      int16x8_t vprod1x9 = vmull_s8(vb9x0, va1x0);
+      int16x8_t vprod2x9 = vmull_s8(vb9x0, va2x0);
+      vprod0x9 = vmlal_s8(vprod0x9, vb9x1, va0x1);
+      vprod1x9 = vmlal_s8(vprod1x9, vb9x1, va1x1);
+      vprod2x9 = vmlal_s8(vprod2x9, vb9x1, va2x1);
+      vacc0x9 = vpadalq_s16(vacc0x9, vprod0x9);
+      vacc1x9 = vpadalq_s16(vacc1x9, vprod1x9);
+      vacc2x9 = vpadalq_s16(vacc2x9, vprod2x9);
+      const int8x8_t vb10x1 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
+      int16x8_t vprod0x10 = vmull_s8(vb10x0, va0x0);
+      int16x8_t vprod1x10 = vmull_s8(vb10x0, va1x0);
+      int16x8_t vprod2x10 = vmull_s8(vb10x0, va2x0);
+      vprod0x10 = vmlal_s8(vprod0x10, vb10x1, va0x1);
+      vprod1x10 = vmlal_s8(vprod1x10, vb10x1, va1x1);
+      vprod2x10 = vmlal_s8(vprod2x10, vb10x1, va2x1);
+      vacc0x10 = vpadalq_s16(vacc0x10, vprod0x10);
+      vacc1x10 = vpadalq_s16(vacc1x10, vprod1x10);
+      vacc2x10 = vpadalq_s16(vacc2x10, vprod2x10);
+      const int8x8_t vb11x1 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
+      int16x8_t vprod0x11 = vmull_s8(vb11x0, va0x0);
+      int16x8_t vprod1x11 = vmull_s8(vb11x0, va1x0);
+      int16x8_t vprod2x11 = vmull_s8(vb11x0, va2x0);
+      vprod0x11 = vmlal_s8(vprod0x11, vb11x1, va0x1);
+      vprod1x11 = vmlal_s8(vprod1x11, vb11x1, va1x1);
+      vprod2x11 = vmlal_s8(vprod2x11, vb11x1, va2x1);
+      vacc0x11 = vpadalq_s16(vacc0x11, vprod0x11);
+      vacc1x11 = vpadalq_s16(vacc1x11, vprod1x11);
+      vacc2x11 = vpadalq_s16(vacc2x11, vprod2x11);
+      const int8x8_t vb12x1 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
+      int16x8_t vprod0x12 = vmull_s8(vb12x0, va0x0);
+      int16x8_t vprod1x12 = vmull_s8(vb12x0, va1x0);
+      int16x8_t vprod2x12 = vmull_s8(vb12x0, va2x0);
+      vprod0x12 = vmlal_s8(vprod0x12, vb12x1, va0x1);
+      vprod1x12 = vmlal_s8(vprod1x12, vb12x1, va1x1);
+      vprod2x12 = vmlal_s8(vprod2x12, vb12x1, va2x1);
+      vacc0x12 = vpadalq_s16(vacc0x12, vprod0x12);
+      vacc1x12 = vpadalq_s16(vacc1x12, vprod1x12);
+      vacc2x12 = vpadalq_s16(vacc2x12, vprod2x12);
+      const int8x8_t vb13x1 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
+      int16x8_t vprod0x13 = vmull_s8(vb13x0, va0x0);
+      int16x8_t vprod1x13 = vmull_s8(vb13x0, va1x0);
+      int16x8_t vprod2x13 = vmull_s8(vb13x0, va2x0);
+      vprod0x13 = vmlal_s8(vprod0x13, vb13x1, va0x1);
+      vprod1x13 = vmlal_s8(vprod1x13, vb13x1, va1x1);
+      vprod2x13 = vmlal_s8(vprod2x13, vb13x1, va2x1);
+      vacc0x13 = vpadalq_s16(vacc0x13, vprod0x13);
+      vacc1x13 = vpadalq_s16(vacc1x13, vprod1x13);
+      vacc2x13 = vpadalq_s16(vacc2x13, vprod2x13);
+      const int8x8_t vb14x1 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
+      int16x8_t vprod0x14 = vmull_s8(vb14x0, va0x0);
+      int16x8_t vprod1x14 = vmull_s8(vb14x0, va1x0);
+      int16x8_t vprod2x14 = vmull_s8(vb14x0, va2x0);
+      vprod0x14 = vmlal_s8(vprod0x14, vb14x1, va0x1);
+      vprod1x14 = vmlal_s8(vprod1x14, vb14x1, va1x1);
+      vprod2x14 = vmlal_s8(vprod2x14, vb14x1, va2x1);
+      vacc0x14 = vpadalq_s16(vacc0x14, vprod0x14);
+      vacc1x14 = vpadalq_s16(vacc1x14, vprod1x14);
+      vacc2x14 = vpadalq_s16(vacc2x14, vprod2x14);
+      const int8x8_t vb15x1 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
+      int16x8_t vprod0x15 = vmull_s8(vb15x0, va0x0);
+      int16x8_t vprod1x15 = vmull_s8(vb15x0, va1x0);
+      int16x8_t vprod2x15 = vmull_s8(vb15x0, va2x0);
+      vprod0x15 = vmlal_s8(vprod0x15, vb15x1, va0x1);
+      vprod1x15 = vmlal_s8(vprod1x15, vb15x1, va1x1);
+      vprod2x15 = vmlal_s8(vprod2x15, vb15x1, va2x1);
+      vacc0x15 = vpadalq_s16(vacc0x15, vprod0x15);
+      vacc1x15 = vpadalq_s16(vacc1x15, vprod1x15);
+      vacc2x15 = vpadalq_s16(vacc2x15, vprod2x15);
+
+      k -= 16 * sizeof(int8_t);
+    }
+    if (k >= 8 * sizeof(int8_t)) {
       const int8x8_t va0 = vld1_s8(a0); a0 += 8;
       const int8x8_t va1 = vld1_s8(a1); a1 += 8;
       const int8x8_t va2 = vld1_s8(a2); a2 += 8;
 
       const int8x8_t vb0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
-
       const int16x8_t vprod0x0 = vmull_s8(vb0, va0);
-      vacc0x0 = vpadalq_s16(vacc0x0, vprod0x0);
       const int16x8_t vprod1x0 = vmull_s8(vb0, va1);
-      vacc1x0 = vpadalq_s16(vacc1x0, vprod1x0);
       const int16x8_t vprod2x0 = vmull_s8(vb0, va2);
+      vacc0x0 = vpadalq_s16(vacc0x0, vprod0x0);
+      vacc1x0 = vpadalq_s16(vacc1x0, vprod1x0);
       vacc2x0 = vpadalq_s16(vacc2x0, vprod2x0);
       const int8x8_t vb1 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
-
       const int16x8_t vprod0x1 = vmull_s8(vb1, va0);
-      vacc0x1 = vpadalq_s16(vacc0x1, vprod0x1);
       const int16x8_t vprod1x1 = vmull_s8(vb1, va1);
-      vacc1x1 = vpadalq_s16(vacc1x1, vprod1x1);
       const int16x8_t vprod2x1 = vmull_s8(vb1, va2);
+      vacc0x1 = vpadalq_s16(vacc0x1, vprod0x1);
+      vacc1x1 = vpadalq_s16(vacc1x1, vprod1x1);
       vacc2x1 = vpadalq_s16(vacc2x1, vprod2x1);
       const int8x8_t vb2 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
-
       const int16x8_t vprod0x2 = vmull_s8(vb2, va0);
-      vacc0x2 = vpadalq_s16(vacc0x2, vprod0x2);
       const int16x8_t vprod1x2 = vmull_s8(vb2, va1);
-      vacc1x2 = vpadalq_s16(vacc1x2, vprod1x2);
       const int16x8_t vprod2x2 = vmull_s8(vb2, va2);
+      vacc0x2 = vpadalq_s16(vacc0x2, vprod0x2);
+      vacc1x2 = vpadalq_s16(vacc1x2, vprod1x2);
       vacc2x2 = vpadalq_s16(vacc2x2, vprod2x2);
       const int8x8_t vb3 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
-
       const int16x8_t vprod0x3 = vmull_s8(vb3, va0);
-      vacc0x3 = vpadalq_s16(vacc0x3, vprod0x3);
       const int16x8_t vprod1x3 = vmull_s8(vb3, va1);
-      vacc1x3 = vpadalq_s16(vacc1x3, vprod1x3);
       const int16x8_t vprod2x3 = vmull_s8(vb3, va2);
+      vacc0x3 = vpadalq_s16(vacc0x3, vprod0x3);
+      vacc1x3 = vpadalq_s16(vacc1x3, vprod1x3);
       vacc2x3 = vpadalq_s16(vacc2x3, vprod2x3);
       const int8x8_t vb4 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
-
       const int16x8_t vprod0x4 = vmull_s8(vb4, va0);
-      vacc0x4 = vpadalq_s16(vacc0x4, vprod0x4);
       const int16x8_t vprod1x4 = vmull_s8(vb4, va1);
-      vacc1x4 = vpadalq_s16(vacc1x4, vprod1x4);
       const int16x8_t vprod2x4 = vmull_s8(vb4, va2);
+      vacc0x4 = vpadalq_s16(vacc0x4, vprod0x4);
+      vacc1x4 = vpadalq_s16(vacc1x4, vprod1x4);
       vacc2x4 = vpadalq_s16(vacc2x4, vprod2x4);
       const int8x8_t vb5 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
-
       const int16x8_t vprod0x5 = vmull_s8(vb5, va0);
-      vacc0x5 = vpadalq_s16(vacc0x5, vprod0x5);
       const int16x8_t vprod1x5 = vmull_s8(vb5, va1);
-      vacc1x5 = vpadalq_s16(vacc1x5, vprod1x5);
       const int16x8_t vprod2x5 = vmull_s8(vb5, va2);
+      vacc0x5 = vpadalq_s16(vacc0x5, vprod0x5);
+      vacc1x5 = vpadalq_s16(vacc1x5, vprod1x5);
       vacc2x5 = vpadalq_s16(vacc2x5, vprod2x5);
       const int8x8_t vb6 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
-
       const int16x8_t vprod0x6 = vmull_s8(vb6, va0);
-      vacc0x6 = vpadalq_s16(vacc0x6, vprod0x6);
       const int16x8_t vprod1x6 = vmull_s8(vb6, va1);
-      vacc1x6 = vpadalq_s16(vacc1x6, vprod1x6);
       const int16x8_t vprod2x6 = vmull_s8(vb6, va2);
+      vacc0x6 = vpadalq_s16(vacc0x6, vprod0x6);
+      vacc1x6 = vpadalq_s16(vacc1x6, vprod1x6);
       vacc2x6 = vpadalq_s16(vacc2x6, vprod2x6);
       const int8x8_t vb7 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
-
       const int16x8_t vprod0x7 = vmull_s8(vb7, va0);
-      vacc0x7 = vpadalq_s16(vacc0x7, vprod0x7);
       const int16x8_t vprod1x7 = vmull_s8(vb7, va1);
-      vacc1x7 = vpadalq_s16(vacc1x7, vprod1x7);
       const int16x8_t vprod2x7 = vmull_s8(vb7, va2);
+      vacc0x7 = vpadalq_s16(vacc0x7, vprod0x7);
+      vacc1x7 = vpadalq_s16(vacc1x7, vprod1x7);
       vacc2x7 = vpadalq_s16(vacc2x7, vprod2x7);
       const int8x8_t vb8 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
-
       const int16x8_t vprod0x8 = vmull_s8(vb8, va0);
-      vacc0x8 = vpadalq_s16(vacc0x8, vprod0x8);
       const int16x8_t vprod1x8 = vmull_s8(vb8, va1);
-      vacc1x8 = vpadalq_s16(vacc1x8, vprod1x8);
       const int16x8_t vprod2x8 = vmull_s8(vb8, va2);
+      vacc0x8 = vpadalq_s16(vacc0x8, vprod0x8);
+      vacc1x8 = vpadalq_s16(vacc1x8, vprod1x8);
       vacc2x8 = vpadalq_s16(vacc2x8, vprod2x8);
       const int8x8_t vb9 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
-
       const int16x8_t vprod0x9 = vmull_s8(vb9, va0);
-      vacc0x9 = vpadalq_s16(vacc0x9, vprod0x9);
       const int16x8_t vprod1x9 = vmull_s8(vb9, va1);
-      vacc1x9 = vpadalq_s16(vacc1x9, vprod1x9);
       const int16x8_t vprod2x9 = vmull_s8(vb9, va2);
+      vacc0x9 = vpadalq_s16(vacc0x9, vprod0x9);
+      vacc1x9 = vpadalq_s16(vacc1x9, vprod1x9);
       vacc2x9 = vpadalq_s16(vacc2x9, vprod2x9);
       const int8x8_t vb10 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
-
       const int16x8_t vprod0x10 = vmull_s8(vb10, va0);
-      vacc0x10 = vpadalq_s16(vacc0x10, vprod0x10);
       const int16x8_t vprod1x10 = vmull_s8(vb10, va1);
-      vacc1x10 = vpadalq_s16(vacc1x10, vprod1x10);
       const int16x8_t vprod2x10 = vmull_s8(vb10, va2);
+      vacc0x10 = vpadalq_s16(vacc0x10, vprod0x10);
+      vacc1x10 = vpadalq_s16(vacc1x10, vprod1x10);
       vacc2x10 = vpadalq_s16(vacc2x10, vprod2x10);
       const int8x8_t vb11 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
-
       const int16x8_t vprod0x11 = vmull_s8(vb11, va0);
-      vacc0x11 = vpadalq_s16(vacc0x11, vprod0x11);
       const int16x8_t vprod1x11 = vmull_s8(vb11, va1);
-      vacc1x11 = vpadalq_s16(vacc1x11, vprod1x11);
       const int16x8_t vprod2x11 = vmull_s8(vb11, va2);
+      vacc0x11 = vpadalq_s16(vacc0x11, vprod0x11);
+      vacc1x11 = vpadalq_s16(vacc1x11, vprod1x11);
       vacc2x11 = vpadalq_s16(vacc2x11, vprod2x11);
       const int8x8_t vb12 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
-
       const int16x8_t vprod0x12 = vmull_s8(vb12, va0);
-      vacc0x12 = vpadalq_s16(vacc0x12, vprod0x12);
       const int16x8_t vprod1x12 = vmull_s8(vb12, va1);
-      vacc1x12 = vpadalq_s16(vacc1x12, vprod1x12);
       const int16x8_t vprod2x12 = vmull_s8(vb12, va2);
+      vacc0x12 = vpadalq_s16(vacc0x12, vprod0x12);
+      vacc1x12 = vpadalq_s16(vacc1x12, vprod1x12);
       vacc2x12 = vpadalq_s16(vacc2x12, vprod2x12);
       const int8x8_t vb13 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
-
       const int16x8_t vprod0x13 = vmull_s8(vb13, va0);
-      vacc0x13 = vpadalq_s16(vacc0x13, vprod0x13);
       const int16x8_t vprod1x13 = vmull_s8(vb13, va1);
-      vacc1x13 = vpadalq_s16(vacc1x13, vprod1x13);
       const int16x8_t vprod2x13 = vmull_s8(vb13, va2);
+      vacc0x13 = vpadalq_s16(vacc0x13, vprod0x13);
+      vacc1x13 = vpadalq_s16(vacc1x13, vprod1x13);
       vacc2x13 = vpadalq_s16(vacc2x13, vprod2x13);
       const int8x8_t vb14 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
-
       const int16x8_t vprod0x14 = vmull_s8(vb14, va0);
-      vacc0x14 = vpadalq_s16(vacc0x14, vprod0x14);
       const int16x8_t vprod1x14 = vmull_s8(vb14, va1);
-      vacc1x14 = vpadalq_s16(vacc1x14, vprod1x14);
       const int16x8_t vprod2x14 = vmull_s8(vb14, va2);
+      vacc0x14 = vpadalq_s16(vacc0x14, vprod0x14);
+      vacc1x14 = vpadalq_s16(vacc1x14, vprod1x14);
       vacc2x14 = vpadalq_s16(vacc2x14, vprod2x14);
       const int8x8_t vb15 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
-
       const int16x8_t vprod0x15 = vmull_s8(vb15, va0);
-      vacc0x15 = vpadalq_s16(vacc0x15, vprod0x15);
       const int16x8_t vprod1x15 = vmull_s8(vb15, va1);
-      vacc1x15 = vpadalq_s16(vacc1x15, vprod1x15);
       const int16x8_t vprod2x15 = vmull_s8(vb15, va2);
+      vacc0x15 = vpadalq_s16(vacc0x15, vprod0x15);
+      vacc1x15 = vpadalq_s16(vacc1x15, vprod1x15);
       vacc2x15 = vpadalq_s16(vacc2x15, vprod2x15);
 
       k -= 8 * sizeof(int8_t);
@@ -244,132 +416,116 @@
       const int8x8_t va2 = vld1_s8(a2); a2 += k;
 
       const int8x8_t vb0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
-
       const int16x8_t vprod0x0 = vmull_s8(vb0, va0);
-      vacc0x0 = vpadalq_s16(vacc0x0, vprod0x0);
       const int16x8_t vprod1x0 = vmull_s8(vb0, va1);
-      vacc1x0 = vpadalq_s16(vacc1x0, vprod1x0);
       const int16x8_t vprod2x0 = vmull_s8(vb0, va2);
+      vacc0x0 = vpadalq_s16(vacc0x0, vprod0x0);
+      vacc1x0 = vpadalq_s16(vacc1x0, vprod1x0);
       vacc2x0 = vpadalq_s16(vacc2x0, vprod2x0);
       const int8x8_t vb1 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
-
       const int16x8_t vprod0x1 = vmull_s8(vb1, va0);
-      vacc0x1 = vpadalq_s16(vacc0x1, vprod0x1);
       const int16x8_t vprod1x1 = vmull_s8(vb1, va1);
-      vacc1x1 = vpadalq_s16(vacc1x1, vprod1x1);
       const int16x8_t vprod2x1 = vmull_s8(vb1, va2);
+      vacc0x1 = vpadalq_s16(vacc0x1, vprod0x1);
+      vacc1x1 = vpadalq_s16(vacc1x1, vprod1x1);
       vacc2x1 = vpadalq_s16(vacc2x1, vprod2x1);
       const int8x8_t vb2 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
-
       const int16x8_t vprod0x2 = vmull_s8(vb2, va0);
-      vacc0x2 = vpadalq_s16(vacc0x2, vprod0x2);
       const int16x8_t vprod1x2 = vmull_s8(vb2, va1);
-      vacc1x2 = vpadalq_s16(vacc1x2, vprod1x2);
       const int16x8_t vprod2x2 = vmull_s8(vb2, va2);
+      vacc0x2 = vpadalq_s16(vacc0x2, vprod0x2);
+      vacc1x2 = vpadalq_s16(vacc1x2, vprod1x2);
       vacc2x2 = vpadalq_s16(vacc2x2, vprod2x2);
       const int8x8_t vb3 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
-
       const int16x8_t vprod0x3 = vmull_s8(vb3, va0);
-      vacc0x3 = vpadalq_s16(vacc0x3, vprod0x3);
       const int16x8_t vprod1x3 = vmull_s8(vb3, va1);
-      vacc1x3 = vpadalq_s16(vacc1x3, vprod1x3);
       const int16x8_t vprod2x3 = vmull_s8(vb3, va2);
+      vacc0x3 = vpadalq_s16(vacc0x3, vprod0x3);
+      vacc1x3 = vpadalq_s16(vacc1x3, vprod1x3);
       vacc2x3 = vpadalq_s16(vacc2x3, vprod2x3);
       const int8x8_t vb4 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
-
       const int16x8_t vprod0x4 = vmull_s8(vb4, va0);
-      vacc0x4 = vpadalq_s16(vacc0x4, vprod0x4);
       const int16x8_t vprod1x4 = vmull_s8(vb4, va1);
-      vacc1x4 = vpadalq_s16(vacc1x4, vprod1x4);
       const int16x8_t vprod2x4 = vmull_s8(vb4, va2);
+      vacc0x4 = vpadalq_s16(vacc0x4, vprod0x4);
+      vacc1x4 = vpadalq_s16(vacc1x4, vprod1x4);
       vacc2x4 = vpadalq_s16(vacc2x4, vprod2x4);
       const int8x8_t vb5 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
-
       const int16x8_t vprod0x5 = vmull_s8(vb5, va0);
-      vacc0x5 = vpadalq_s16(vacc0x5, vprod0x5);
       const int16x8_t vprod1x5 = vmull_s8(vb5, va1);
-      vacc1x5 = vpadalq_s16(vacc1x5, vprod1x5);
       const int16x8_t vprod2x5 = vmull_s8(vb5, va2);
+      vacc0x5 = vpadalq_s16(vacc0x5, vprod0x5);
+      vacc1x5 = vpadalq_s16(vacc1x5, vprod1x5);
       vacc2x5 = vpadalq_s16(vacc2x5, vprod2x5);
       const int8x8_t vb6 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
-
       const int16x8_t vprod0x6 = vmull_s8(vb6, va0);
-      vacc0x6 = vpadalq_s16(vacc0x6, vprod0x6);
       const int16x8_t vprod1x6 = vmull_s8(vb6, va1);
-      vacc1x6 = vpadalq_s16(vacc1x6, vprod1x6);
       const int16x8_t vprod2x6 = vmull_s8(vb6, va2);
+      vacc0x6 = vpadalq_s16(vacc0x6, vprod0x6);
+      vacc1x6 = vpadalq_s16(vacc1x6, vprod1x6);
       vacc2x6 = vpadalq_s16(vacc2x6, vprod2x6);
       const int8x8_t vb7 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
-
       const int16x8_t vprod0x7 = vmull_s8(vb7, va0);
-      vacc0x7 = vpadalq_s16(vacc0x7, vprod0x7);
       const int16x8_t vprod1x7 = vmull_s8(vb7, va1);
-      vacc1x7 = vpadalq_s16(vacc1x7, vprod1x7);
       const int16x8_t vprod2x7 = vmull_s8(vb7, va2);
+      vacc0x7 = vpadalq_s16(vacc0x7, vprod0x7);
+      vacc1x7 = vpadalq_s16(vacc1x7, vprod1x7);
       vacc2x7 = vpadalq_s16(vacc2x7, vprod2x7);
       const int8x8_t vb8 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
-
       const int16x8_t vprod0x8 = vmull_s8(vb8, va0);
-      vacc0x8 = vpadalq_s16(vacc0x8, vprod0x8);
       const int16x8_t vprod1x8 = vmull_s8(vb8, va1);
-      vacc1x8 = vpadalq_s16(vacc1x8, vprod1x8);
       const int16x8_t vprod2x8 = vmull_s8(vb8, va2);
+      vacc0x8 = vpadalq_s16(vacc0x8, vprod0x8);
+      vacc1x8 = vpadalq_s16(vacc1x8, vprod1x8);
       vacc2x8 = vpadalq_s16(vacc2x8, vprod2x8);
       const int8x8_t vb9 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
-
       const int16x8_t vprod0x9 = vmull_s8(vb9, va0);
-      vacc0x9 = vpadalq_s16(vacc0x9, vprod0x9);
       const int16x8_t vprod1x9 = vmull_s8(vb9, va1);
-      vacc1x9 = vpadalq_s16(vacc1x9, vprod1x9);
       const int16x8_t vprod2x9 = vmull_s8(vb9, va2);
+      vacc0x9 = vpadalq_s16(vacc0x9, vprod0x9);
+      vacc1x9 = vpadalq_s16(vacc1x9, vprod1x9);
       vacc2x9 = vpadalq_s16(vacc2x9, vprod2x9);
       const int8x8_t vb10 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
-
       const int16x8_t vprod0x10 = vmull_s8(vb10, va0);
-      vacc0x10 = vpadalq_s16(vacc0x10, vprod0x10);
       const int16x8_t vprod1x10 = vmull_s8(vb10, va1);
-      vacc1x10 = vpadalq_s16(vacc1x10, vprod1x10);
       const int16x8_t vprod2x10 = vmull_s8(vb10, va2);
+      vacc0x10 = vpadalq_s16(vacc0x10, vprod0x10);
+      vacc1x10 = vpadalq_s16(vacc1x10, vprod1x10);
       vacc2x10 = vpadalq_s16(vacc2x10, vprod2x10);
       const int8x8_t vb11 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
-
       const int16x8_t vprod0x11 = vmull_s8(vb11, va0);
-      vacc0x11 = vpadalq_s16(vacc0x11, vprod0x11);
       const int16x8_t vprod1x11 = vmull_s8(vb11, va1);
-      vacc1x11 = vpadalq_s16(vacc1x11, vprod1x11);
       const int16x8_t vprod2x11 = vmull_s8(vb11, va2);
+      vacc0x11 = vpadalq_s16(vacc0x11, vprod0x11);
+      vacc1x11 = vpadalq_s16(vacc1x11, vprod1x11);
       vacc2x11 = vpadalq_s16(vacc2x11, vprod2x11);
       const int8x8_t vb12 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
-
       const int16x8_t vprod0x12 = vmull_s8(vb12, va0);
-      vacc0x12 = vpadalq_s16(vacc0x12, vprod0x12);
       const int16x8_t vprod1x12 = vmull_s8(vb12, va1);
-      vacc1x12 = vpadalq_s16(vacc1x12, vprod1x12);
       const int16x8_t vprod2x12 = vmull_s8(vb12, va2);
+      vacc0x12 = vpadalq_s16(vacc0x12, vprod0x12);
+      vacc1x12 = vpadalq_s16(vacc1x12, vprod1x12);
       vacc2x12 = vpadalq_s16(vacc2x12, vprod2x12);
       const int8x8_t vb13 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
-
       const int16x8_t vprod0x13 = vmull_s8(vb13, va0);
-      vacc0x13 = vpadalq_s16(vacc0x13, vprod0x13);
       const int16x8_t vprod1x13 = vmull_s8(vb13, va1);
-      vacc1x13 = vpadalq_s16(vacc1x13, vprod1x13);
       const int16x8_t vprod2x13 = vmull_s8(vb13, va2);
+      vacc0x13 = vpadalq_s16(vacc0x13, vprod0x13);
+      vacc1x13 = vpadalq_s16(vacc1x13, vprod1x13);
       vacc2x13 = vpadalq_s16(vacc2x13, vprod2x13);
       const int8x8_t vb14 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
-
       const int16x8_t vprod0x14 = vmull_s8(vb14, va0);
-      vacc0x14 = vpadalq_s16(vacc0x14, vprod0x14);
       const int16x8_t vprod1x14 = vmull_s8(vb14, va1);
-      vacc1x14 = vpadalq_s16(vacc1x14, vprod1x14);
       const int16x8_t vprod2x14 = vmull_s8(vb14, va2);
+      vacc0x14 = vpadalq_s16(vacc0x14, vprod0x14);
+      vacc1x14 = vpadalq_s16(vacc1x14, vprod1x14);
       vacc2x14 = vpadalq_s16(vacc2x14, vprod2x14);
       const int8x8_t vb15 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
-
       const int16x8_t vprod0x15 = vmull_s8(vb15, va0);
-      vacc0x15 = vpadalq_s16(vacc0x15, vprod0x15);
       const int16x8_t vprod1x15 = vmull_s8(vb15, va1);
-      vacc1x15 = vpadalq_s16(vacc1x15, vprod1x15);
       const int16x8_t vprod2x15 = vmull_s8(vb15, va2);
+      vacc0x15 = vpadalq_s16(vacc0x15, vprod0x15);
+      vacc1x15 = vpadalq_s16(vacc1x15, vprod1x15);
       vacc2x15 = vpadalq_s16(vacc2x15, vprod2x15);
     }
 

diff --git a/src/qs8-gemm/gen/3x8c8-minmax-neon-mull-padal.c b/src/qs8-gemm/gen/3x8c8-minmax-neon-mull-padal.c
index 0f5d1f7..29ec316 100644
--- a/src/qs8-gemm/gen/3x8c8-minmax-neon-mull-padal.c
+++ b/src/qs8-gemm/gen/3x8c8-minmax-neon-mull-padal.c

@@ -78,74 +78,166 @@
     int32x4_t vacc2x7 = vacc0x7;
 
     size_t k = kc;
-    while (k >= 8 * sizeof(int8_t)) {
+    while (k >= 16 * sizeof(int8_t)) {
+      const int8x8_t va0x0 = vld1_s8(a0); a0 += 8;
+      const int8x8_t va0x1 = vld1_s8(a0); a0 += 8;
+      const int8x8_t va1x0 = vld1_s8(a1); a1 += 8;
+      const int8x8_t va1x1 = vld1_s8(a1); a1 += 8;
+      const int8x8_t va2x0 = vld1_s8(a2); a2 += 8;
+      const int8x8_t va2x1 = vld1_s8(a2); a2 += 8;
+
+      const int8x8_t vb0x0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
+      const int8x8_t vb1x0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
+      const int8x8_t vb2x0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
+      const int8x8_t vb3x0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
+      const int8x8_t vb4x0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
+      const int8x8_t vb5x0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
+      const int8x8_t vb6x0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
+      const int8x8_t vb7x0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
+
+      const int8x8_t vb0x1 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
+      int16x8_t vprod0x0 = vmull_s8(vb0x0, va0x0);
+      int16x8_t vprod1x0 = vmull_s8(vb0x0, va1x0);
+      int16x8_t vprod2x0 = vmull_s8(vb0x0, va2x0);
+      vprod0x0 = vmlal_s8(vprod0x0, vb0x1, va0x1);
+      vprod1x0 = vmlal_s8(vprod1x0, vb0x1, va1x1);
+      vprod2x0 = vmlal_s8(vprod2x0, vb0x1, va2x1);
+      vacc0x0 = vpadalq_s16(vacc0x0, vprod0x0);
+      vacc1x0 = vpadalq_s16(vacc1x0, vprod1x0);
+      vacc2x0 = vpadalq_s16(vacc2x0, vprod2x0);
+      const int8x8_t vb1x1 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
+      int16x8_t vprod0x1 = vmull_s8(vb1x0, va0x0);
+      int16x8_t vprod1x1 = vmull_s8(vb1x0, va1x0);
+      int16x8_t vprod2x1 = vmull_s8(vb1x0, va2x0);
+      vprod0x1 = vmlal_s8(vprod0x1, vb1x1, va0x1);
+      vprod1x1 = vmlal_s8(vprod1x1, vb1x1, va1x1);
+      vprod2x1 = vmlal_s8(vprod2x1, vb1x1, va2x1);
+      vacc0x1 = vpadalq_s16(vacc0x1, vprod0x1);
+      vacc1x1 = vpadalq_s16(vacc1x1, vprod1x1);
+      vacc2x1 = vpadalq_s16(vacc2x1, vprod2x1);
+      const int8x8_t vb2x1 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
+      int16x8_t vprod0x2 = vmull_s8(vb2x0, va0x0);
+      int16x8_t vprod1x2 = vmull_s8(vb2x0, va1x0);
+      int16x8_t vprod2x2 = vmull_s8(vb2x0, va2x0);
+      vprod0x2 = vmlal_s8(vprod0x2, vb2x1, va0x1);
+      vprod1x2 = vmlal_s8(vprod1x2, vb2x1, va1x1);
+      vprod2x2 = vmlal_s8(vprod2x2, vb2x1, va2x1);
+      vacc0x2 = vpadalq_s16(vacc0x2, vprod0x2);
+      vacc1x2 = vpadalq_s16(vacc1x2, vprod1x2);
+      vacc2x2 = vpadalq_s16(vacc2x2, vprod2x2);
+      const int8x8_t vb3x1 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
+      int16x8_t vprod0x3 = vmull_s8(vb3x0, va0x0);
+      int16x8_t vprod1x3 = vmull_s8(vb3x0, va1x0);
+      int16x8_t vprod2x3 = vmull_s8(vb3x0, va2x0);
+      vprod0x3 = vmlal_s8(vprod0x3, vb3x1, va0x1);
+      vprod1x3 = vmlal_s8(vprod1x3, vb3x1, va1x1);
+      vprod2x3 = vmlal_s8(vprod2x3, vb3x1, va2x1);
+      vacc0x3 = vpadalq_s16(vacc0x3, vprod0x3);
+      vacc1x3 = vpadalq_s16(vacc1x3, vprod1x3);
+      vacc2x3 = vpadalq_s16(vacc2x3, vprod2x3);
+      const int8x8_t vb4x1 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
+      int16x8_t vprod0x4 = vmull_s8(vb4x0, va0x0);
+      int16x8_t vprod1x4 = vmull_s8(vb4x0, va1x0);
+      int16x8_t vprod2x4 = vmull_s8(vb4x0, va2x0);
+      vprod0x4 = vmlal_s8(vprod0x4, vb4x1, va0x1);
+      vprod1x4 = vmlal_s8(vprod1x4, vb4x1, va1x1);
+      vprod2x4 = vmlal_s8(vprod2x4, vb4x1, va2x1);
+      vacc0x4 = vpadalq_s16(vacc0x4, vprod0x4);
+      vacc1x4 = vpadalq_s16(vacc1x4, vprod1x4);
+      vacc2x4 = vpadalq_s16(vacc2x4, vprod2x4);
+      const int8x8_t vb5x1 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
+      int16x8_t vprod0x5 = vmull_s8(vb5x0, va0x0);
+      int16x8_t vprod1x5 = vmull_s8(vb5x0, va1x0);
+      int16x8_t vprod2x5 = vmull_s8(vb5x0, va2x0);
+      vprod0x5 = vmlal_s8(vprod0x5, vb5x1, va0x1);
+      vprod1x5 = vmlal_s8(vprod1x5, vb5x1, va1x1);
+      vprod2x5 = vmlal_s8(vprod2x5, vb5x1, va2x1);
+      vacc0x5 = vpadalq_s16(vacc0x5, vprod0x5);
+      vacc1x5 = vpadalq_s16(vacc1x5, vprod1x5);
+      vacc2x5 = vpadalq_s16(vacc2x5, vprod2x5);
+      const int8x8_t vb6x1 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
+      int16x8_t vprod0x6 = vmull_s8(vb6x0, va0x0);
+      int16x8_t vprod1x6 = vmull_s8(vb6x0, va1x0);
+      int16x8_t vprod2x6 = vmull_s8(vb6x0, va2x0);
+      vprod0x6 = vmlal_s8(vprod0x6, vb6x1, va0x1);
+      vprod1x6 = vmlal_s8(vprod1x6, vb6x1, va1x1);
+      vprod2x6 = vmlal_s8(vprod2x6, vb6x1, va2x1);
+      vacc0x6 = vpadalq_s16(vacc0x6, vprod0x6);
+      vacc1x6 = vpadalq_s16(vacc1x6, vprod1x6);
+      vacc2x6 = vpadalq_s16(vacc2x6, vprod2x6);
+      const int8x8_t vb7x1 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
+      int16x8_t vprod0x7 = vmull_s8(vb7x0, va0x0);
+      int16x8_t vprod1x7 = vmull_s8(vb7x0, va1x0);
+      int16x8_t vprod2x7 = vmull_s8(vb7x0, va2x0);
+      vprod0x7 = vmlal_s8(vprod0x7, vb7x1, va0x1);
+      vprod1x7 = vmlal_s8(vprod1x7, vb7x1, va1x1);
+      vprod2x7 = vmlal_s8(vprod2x7, vb7x1, va2x1);
+      vacc0x7 = vpadalq_s16(vacc0x7, vprod0x7);
+      vacc1x7 = vpadalq_s16(vacc1x7, vprod1x7);
+      vacc2x7 = vpadalq_s16(vacc2x7, vprod2x7);
+
+      k -= 16 * sizeof(int8_t);
+    }
+    if (k >= 8 * sizeof(int8_t)) {
       const int8x8_t va0 = vld1_s8(a0); a0 += 8;
       const int8x8_t va1 = vld1_s8(a1); a1 += 8;
       const int8x8_t va2 = vld1_s8(a2); a2 += 8;
 
       const int8x8_t vb0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
-
       const int16x8_t vprod0x0 = vmull_s8(vb0, va0);
-      vacc0x0 = vpadalq_s16(vacc0x0, vprod0x0);
       const int16x8_t vprod1x0 = vmull_s8(vb0, va1);
-      vacc1x0 = vpadalq_s16(vacc1x0, vprod1x0);
       const int16x8_t vprod2x0 = vmull_s8(vb0, va2);
+      vacc0x0 = vpadalq_s16(vacc0x0, vprod0x0);
+      vacc1x0 = vpadalq_s16(vacc1x0, vprod1x0);
       vacc2x0 = vpadalq_s16(vacc2x0, vprod2x0);
       const int8x8_t vb1 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
-
       const int16x8_t vprod0x1 = vmull_s8(vb1, va0);
-      vacc0x1 = vpadalq_s16(vacc0x1, vprod0x1);
       const int16x8_t vprod1x1 = vmull_s8(vb1, va1);
-      vacc1x1 = vpadalq_s16(vacc1x1, vprod1x1);
       const int16x8_t vprod2x1 = vmull_s8(vb1, va2);
+      vacc0x1 = vpadalq_s16(vacc0x1, vprod0x1);
+      vacc1x1 = vpadalq_s16(vacc1x1, vprod1x1);
       vacc2x1 = vpadalq_s16(vacc2x1, vprod2x1);
       const int8x8_t vb2 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
-
       const int16x8_t vprod0x2 = vmull_s8(vb2, va0);
-      vacc0x2 = vpadalq_s16(vacc0x2, vprod0x2);
       const int16x8_t vprod1x2 = vmull_s8(vb2, va1);
-      vacc1x2 = vpadalq_s16(vacc1x2, vprod1x2);
       const int16x8_t vprod2x2 = vmull_s8(vb2, va2);
+      vacc0x2 = vpadalq_s16(vacc0x2, vprod0x2);
+      vacc1x2 = vpadalq_s16(vacc1x2, vprod1x2);
       vacc2x2 = vpadalq_s16(vacc2x2, vprod2x2);
       const int8x8_t vb3 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
-
       const int16x8_t vprod0x3 = vmull_s8(vb3, va0);
-      vacc0x3 = vpadalq_s16(vacc0x3, vprod0x3);
       const int16x8_t vprod1x3 = vmull_s8(vb3, va1);
-      vacc1x3 = vpadalq_s16(vacc1x3, vprod1x3);
       const int16x8_t vprod2x3 = vmull_s8(vb3, va2);
+      vacc0x3 = vpadalq_s16(vacc0x3, vprod0x3);
+      vacc1x3 = vpadalq_s16(vacc1x3, vprod1x3);
       vacc2x3 = vpadalq_s16(vacc2x3, vprod2x3);
       const int8x8_t vb4 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
-
       const int16x8_t vprod0x4 = vmull_s8(vb4, va0);
-      vacc0x4 = vpadalq_s16(vacc0x4, vprod0x4);
       const int16x8_t vprod1x4 = vmull_s8(vb4, va1);
-      vacc1x4 = vpadalq_s16(vacc1x4, vprod1x4);
       const int16x8_t vprod2x4 = vmull_s8(vb4, va2);
+      vacc0x4 = vpadalq_s16(vacc0x4, vprod0x4);
+      vacc1x4 = vpadalq_s16(vacc1x4, vprod1x4);
       vacc2x4 = vpadalq_s16(vacc2x4, vprod2x4);
       const int8x8_t vb5 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
-
       const int16x8_t vprod0x5 = vmull_s8(vb5, va0);
-      vacc0x5 = vpadalq_s16(vacc0x5, vprod0x5);
       const int16x8_t vprod1x5 = vmull_s8(vb5, va1);
-      vacc1x5 = vpadalq_s16(vacc1x5, vprod1x5);
       const int16x8_t vprod2x5 = vmull_s8(vb5, va2);
+      vacc0x5 = vpadalq_s16(vacc0x5, vprod0x5);
+      vacc1x5 = vpadalq_s16(vacc1x5, vprod1x5);
       vacc2x5 = vpadalq_s16(vacc2x5, vprod2x5);
       const int8x8_t vb6 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
-
       const int16x8_t vprod0x6 = vmull_s8(vb6, va0);
-      vacc0x6 = vpadalq_s16(vacc0x6, vprod0x6);
       const int16x8_t vprod1x6 = vmull_s8(vb6, va1);
-      vacc1x6 = vpadalq_s16(vacc1x6, vprod1x6);
       const int16x8_t vprod2x6 = vmull_s8(vb6, va2);
+      vacc0x6 = vpadalq_s16(vacc0x6, vprod0x6);
+      vacc1x6 = vpadalq_s16(vacc1x6, vprod1x6);
       vacc2x6 = vpadalq_s16(vacc2x6, vprod2x6);
       const int8x8_t vb7 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
-
       const int16x8_t vprod0x7 = vmull_s8(vb7, va0);
-      vacc0x7 = vpadalq_s16(vacc0x7, vprod0x7);
       const int16x8_t vprod1x7 = vmull_s8(vb7, va1);
-      vacc1x7 = vpadalq_s16(vacc1x7, vprod1x7);
       const int16x8_t vprod2x7 = vmull_s8(vb7, va2);
+      vacc0x7 = vpadalq_s16(vacc0x7, vprod0x7);
+      vacc1x7 = vpadalq_s16(vacc1x7, vprod1x7);
       vacc2x7 = vpadalq_s16(vacc2x7, vprod2x7);
 
       k -= 8 * sizeof(int8_t);
@@ -156,68 +248,60 @@
       const int8x8_t va2 = vld1_s8(a2); a2 += k;
 
       const int8x8_t vb0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
-
       const int16x8_t vprod0x0 = vmull_s8(vb0, va0);
-      vacc0x0 = vpadalq_s16(vacc0x0, vprod0x0);
       const int16x8_t vprod1x0 = vmull_s8(vb0, va1);
-      vacc1x0 = vpadalq_s16(vacc1x0, vprod1x0);
       const int16x8_t vprod2x0 = vmull_s8(vb0, va2);
+      vacc0x0 = vpadalq_s16(vacc0x0, vprod0x0);
+      vacc1x0 = vpadalq_s16(vacc1x0, vprod1x0);
       vacc2x0 = vpadalq_s16(vacc2x0, vprod2x0);
       const int8x8_t vb1 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
-
       const int16x8_t vprod0x1 = vmull_s8(vb1, va0);
-      vacc0x1 = vpadalq_s16(vacc0x1, vprod0x1);
       const int16x8_t vprod1x1 = vmull_s8(vb1, va1);
-      vacc1x1 = vpadalq_s16(vacc1x1, vprod1x1);
       const int16x8_t vprod2x1 = vmull_s8(vb1, va2);
+      vacc0x1 = vpadalq_s16(vacc0x1, vprod0x1);
+      vacc1x1 = vpadalq_s16(vacc1x1, vprod1x1);
       vacc2x1 = vpadalq_s16(vacc2x1, vprod2x1);
       const int8x8_t vb2 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
-
       const int16x8_t vprod0x2 = vmull_s8(vb2, va0);
-      vacc0x2 = vpadalq_s16(vacc0x2, vprod0x2);
       const int16x8_t vprod1x2 = vmull_s8(vb2, va1);
-      vacc1x2 = vpadalq_s16(vacc1x2, vprod1x2);
       const int16x8_t vprod2x2 = vmull_s8(vb2, va2);
+      vacc0x2 = vpadalq_s16(vacc0x2, vprod0x2);
+      vacc1x2 = vpadalq_s16(vacc1x2, vprod1x2);
       vacc2x2 = vpadalq_s16(vacc2x2, vprod2x2);
       const int8x8_t vb3 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
-
       const int16x8_t vprod0x3 = vmull_s8(vb3, va0);
-      vacc0x3 = vpadalq_s16(vacc0x3, vprod0x3);
       const int16x8_t vprod1x3 = vmull_s8(vb3, va1);
-      vacc1x3 = vpadalq_s16(vacc1x3, vprod1x3);
       const int16x8_t vprod2x3 = vmull_s8(vb3, va2);
+      vacc0x3 = vpadalq_s16(vacc0x3, vprod0x3);
+      vacc1x3 = vpadalq_s16(vacc1x3, vprod1x3);
       vacc2x3 = vpadalq_s16(vacc2x3, vprod2x3);
       const int8x8_t vb4 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
-
       const int16x8_t vprod0x4 = vmull_s8(vb4, va0);
-      vacc0x4 = vpadalq_s16(vacc0x4, vprod0x4);
       const int16x8_t vprod1x4 = vmull_s8(vb4, va1);
-      vacc1x4 = vpadalq_s16(vacc1x4, vprod1x4);
       const int16x8_t vprod2x4 = vmull_s8(vb4, va2);
+      vacc0x4 = vpadalq_s16(vacc0x4, vprod0x4);
+      vacc1x4 = vpadalq_s16(vacc1x4, vprod1x4);
       vacc2x4 = vpadalq_s16(vacc2x4, vprod2x4);
       const int8x8_t vb5 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
-
       const int16x8_t vprod0x5 = vmull_s8(vb5, va0);
-      vacc0x5 = vpadalq_s16(vacc0x5, vprod0x5);
       const int16x8_t vprod1x5 = vmull_s8(vb5, va1);
-      vacc1x5 = vpadalq_s16(vacc1x5, vprod1x5);
       const int16x8_t vprod2x5 = vmull_s8(vb5, va2);
+      vacc0x5 = vpadalq_s16(vacc0x5, vprod0x5);
+      vacc1x5 = vpadalq_s16(vacc1x5, vprod1x5);
       vacc2x5 = vpadalq_s16(vacc2x5, vprod2x5);
       const int8x8_t vb6 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
-
       const int16x8_t vprod0x6 = vmull_s8(vb6, va0);
-      vacc0x6 = vpadalq_s16(vacc0x6, vprod0x6);
       const int16x8_t vprod1x6 = vmull_s8(vb6, va1);
-      vacc1x6 = vpadalq_s16(vacc1x6, vprod1x6);
       const int16x8_t vprod2x6 = vmull_s8(vb6, va2);
+      vacc0x6 = vpadalq_s16(vacc0x6, vprod0x6);
+      vacc1x6 = vpadalq_s16(vacc1x6, vprod1x6);
       vacc2x6 = vpadalq_s16(vacc2x6, vprod2x6);
       const int8x8_t vb7 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
-
       const int16x8_t vprod0x7 = vmull_s8(vb7, va0);
-      vacc0x7 = vpadalq_s16(vacc0x7, vprod0x7);
       const int16x8_t vprod1x7 = vmull_s8(vb7, va1);
-      vacc1x7 = vpadalq_s16(vacc1x7, vprod1x7);
       const int16x8_t vprod2x7 = vmull_s8(vb7, va2);
+      vacc0x7 = vpadalq_s16(vacc0x7, vprod0x7);
+      vacc1x7 = vpadalq_s16(vacc1x7, vprod1x7);
       vacc2x7 = vpadalq_s16(vacc2x7, vprod2x7);
     }
 

diff --git a/src/qs8-gemm/gen/4x16c8-minmax-neon-mull-padal.c b/src/qs8-gemm/gen/4x16c8-minmax-neon-mull-padal.c
index 0ee11e6..a1169bb 100644
--- a/src/qs8-gemm/gen/4x16c8-minmax-neon-mull-padal.c
+++ b/src/qs8-gemm/gen/4x16c8-minmax-neon-mull-padal.c

@@ -124,171 +124,393 @@
     int32x4_t vacc3x15 = vacc0x15;
 
     size_t k = kc;
-    while (k >= 8 * sizeof(int8_t)) {
+    while (k >= 16 * sizeof(int8_t)) {
+      const int8x8_t va0x0 = vld1_s8(a0); a0 += 8;
+      const int8x8_t va0x1 = vld1_s8(a0); a0 += 8;
+      const int8x8_t va1x0 = vld1_s8(a1); a1 += 8;
+      const int8x8_t va1x1 = vld1_s8(a1); a1 += 8;
+      const int8x8_t va2x0 = vld1_s8(a2); a2 += 8;
+      const int8x8_t va2x1 = vld1_s8(a2); a2 += 8;
+      const int8x8_t va3x0 = vld1_s8(a3); a3 += 8;
+      const int8x8_t va3x1 = vld1_s8(a3); a3 += 8;
+
+      const int8x8_t vb0x0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
+      const int8x8_t vb1x0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
+      const int8x8_t vb2x0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
+      const int8x8_t vb3x0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
+      const int8x8_t vb4x0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
+      const int8x8_t vb5x0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
+      const int8x8_t vb6x0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
+      const int8x8_t vb7x0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
+      const int8x8_t vb8x0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
+      const int8x8_t vb9x0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
+      const int8x8_t vb10x0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
+      const int8x8_t vb11x0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
+      const int8x8_t vb12x0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
+      const int8x8_t vb13x0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
+      const int8x8_t vb14x0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
+      const int8x8_t vb15x0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
+
+      const int8x8_t vb0x1 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
+      int16x8_t vprod0x0 = vmull_s8(vb0x0, va0x0);
+      int16x8_t vprod1x0 = vmull_s8(vb0x0, va1x0);
+      int16x8_t vprod2x0 = vmull_s8(vb0x0, va2x0);
+      int16x8_t vprod3x0 = vmull_s8(vb0x0, va3x0);
+      vprod0x0 = vmlal_s8(vprod0x0, vb0x1, va0x1);
+      vprod1x0 = vmlal_s8(vprod1x0, vb0x1, va1x1);
+      vprod2x0 = vmlal_s8(vprod2x0, vb0x1, va2x1);
+      vprod3x0 = vmlal_s8(vprod3x0, vb0x1, va3x1);
+      vacc0x0 = vpadalq_s16(vacc0x0, vprod0x0);
+      vacc1x0 = vpadalq_s16(vacc1x0, vprod1x0);
+      vacc2x0 = vpadalq_s16(vacc2x0, vprod2x0);
+      vacc3x0 = vpadalq_s16(vacc3x0, vprod3x0);
+      const int8x8_t vb1x1 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
+      int16x8_t vprod0x1 = vmull_s8(vb1x0, va0x0);
+      int16x8_t vprod1x1 = vmull_s8(vb1x0, va1x0);
+      int16x8_t vprod2x1 = vmull_s8(vb1x0, va2x0);
+      int16x8_t vprod3x1 = vmull_s8(vb1x0, va3x0);
+      vprod0x1 = vmlal_s8(vprod0x1, vb1x1, va0x1);
+      vprod1x1 = vmlal_s8(vprod1x1, vb1x1, va1x1);
+      vprod2x1 = vmlal_s8(vprod2x1, vb1x1, va2x1);
+      vprod3x1 = vmlal_s8(vprod3x1, vb1x1, va3x1);
+      vacc0x1 = vpadalq_s16(vacc0x1, vprod0x1);
+      vacc1x1 = vpadalq_s16(vacc1x1, vprod1x1);
+      vacc2x1 = vpadalq_s16(vacc2x1, vprod2x1);
+      vacc3x1 = vpadalq_s16(vacc3x1, vprod3x1);
+      const int8x8_t vb2x1 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
+      int16x8_t vprod0x2 = vmull_s8(vb2x0, va0x0);
+      int16x8_t vprod1x2 = vmull_s8(vb2x0, va1x0);
+      int16x8_t vprod2x2 = vmull_s8(vb2x0, va2x0);
+      int16x8_t vprod3x2 = vmull_s8(vb2x0, va3x0);
+      vprod0x2 = vmlal_s8(vprod0x2, vb2x1, va0x1);
+      vprod1x2 = vmlal_s8(vprod1x2, vb2x1, va1x1);
+      vprod2x2 = vmlal_s8(vprod2x2, vb2x1, va2x1);
+      vprod3x2 = vmlal_s8(vprod3x2, vb2x1, va3x1);
+      vacc0x2 = vpadalq_s16(vacc0x2, vprod0x2);
+      vacc1x2 = vpadalq_s16(vacc1x2, vprod1x2);
+      vacc2x2 = vpadalq_s16(vacc2x2, vprod2x2);
+      vacc3x2 = vpadalq_s16(vacc3x2, vprod3x2);
+      const int8x8_t vb3x1 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
+      int16x8_t vprod0x3 = vmull_s8(vb3x0, va0x0);
+      int16x8_t vprod1x3 = vmull_s8(vb3x0, va1x0);
+      int16x8_t vprod2x3 = vmull_s8(vb3x0, va2x0);
+      int16x8_t vprod3x3 = vmull_s8(vb3x0, va3x0);
+      vprod0x3 = vmlal_s8(vprod0x3, vb3x1, va0x1);
+      vprod1x3 = vmlal_s8(vprod1x3, vb3x1, va1x1);
+      vprod2x3 = vmlal_s8(vprod2x3, vb3x1, va2x1);
+      vprod3x3 = vmlal_s8(vprod3x3, vb3x1, va3x1);
+      vacc0x3 = vpadalq_s16(vacc0x3, vprod0x3);
+      vacc1x3 = vpadalq_s16(vacc1x3, vprod1x3);
+      vacc2x3 = vpadalq_s16(vacc2x3, vprod2x3);
+      vacc3x3 = vpadalq_s16(vacc3x3, vprod3x3);
+      const int8x8_t vb4x1 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
+      int16x8_t vprod0x4 = vmull_s8(vb4x0, va0x0);
+      int16x8_t vprod1x4 = vmull_s8(vb4x0, va1x0);
+      int16x8_t vprod2x4 = vmull_s8(vb4x0, va2x0);
+      int16x8_t vprod3x4 = vmull_s8(vb4x0, va3x0);
+      vprod0x4 = vmlal_s8(vprod0x4, vb4x1, va0x1);
+      vprod1x4 = vmlal_s8(vprod1x4, vb4x1, va1x1);
+      vprod2x4 = vmlal_s8(vprod2x4, vb4x1, va2x1);
+      vprod3x4 = vmlal_s8(vprod3x4, vb4x1, va3x1);
+      vacc0x4 = vpadalq_s16(vacc0x4, vprod0x4);
+      vacc1x4 = vpadalq_s16(vacc1x4, vprod1x4);
+      vacc2x4 = vpadalq_s16(vacc2x4, vprod2x4);
+      vacc3x4 = vpadalq_s16(vacc3x4, vprod3x4);
+      const int8x8_t vb5x1 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
+      int16x8_t vprod0x5 = vmull_s8(vb5x0, va0x0);
+      int16x8_t vprod1x5 = vmull_s8(vb5x0, va1x0);
+      int16x8_t vprod2x5 = vmull_s8(vb5x0, va2x0);
+      int16x8_t vprod3x5 = vmull_s8(vb5x0, va3x0);
+      vprod0x5 = vmlal_s8(vprod0x5, vb5x1, va0x1);
+      vprod1x5 = vmlal_s8(vprod1x5, vb5x1, va1x1);
+      vprod2x5 = vmlal_s8(vprod2x5, vb5x1, va2x1);
+      vprod3x5 = vmlal_s8(vprod3x5, vb5x1, va3x1);
+      vacc0x5 = vpadalq_s16(vacc0x5, vprod0x5);
+      vacc1x5 = vpadalq_s16(vacc1x5, vprod1x5);
+      vacc2x5 = vpadalq_s16(vacc2x5, vprod2x5);
+      vacc3x5 = vpadalq_s16(vacc3x5, vprod3x5);
+      const int8x8_t vb6x1 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
+      int16x8_t vprod0x6 = vmull_s8(vb6x0, va0x0);
+      int16x8_t vprod1x6 = vmull_s8(vb6x0, va1x0);
+      int16x8_t vprod2x6 = vmull_s8(vb6x0, va2x0);
+      int16x8_t vprod3x6 = vmull_s8(vb6x0, va3x0);
+      vprod0x6 = vmlal_s8(vprod0x6, vb6x1, va0x1);
+      vprod1x6 = vmlal_s8(vprod1x6, vb6x1, va1x1);
+      vprod2x6 = vmlal_s8(vprod2x6, vb6x1, va2x1);
+      vprod3x6 = vmlal_s8(vprod3x6, vb6x1, va3x1);
+      vacc0x6 = vpadalq_s16(vacc0x6, vprod0x6);
+      vacc1x6 = vpadalq_s16(vacc1x6, vprod1x6);
+      vacc2x6 = vpadalq_s16(vacc2x6, vprod2x6);
+      vacc3x6 = vpadalq_s16(vacc3x6, vprod3x6);
+      const int8x8_t vb7x1 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
+      int16x8_t vprod0x7 = vmull_s8(vb7x0, va0x0);
+      int16x8_t vprod1x7 = vmull_s8(vb7x0, va1x0);
+      int16x8_t vprod2x7 = vmull_s8(vb7x0, va2x0);
+      int16x8_t vprod3x7 = vmull_s8(vb7x0, va3x0);
+      vprod0x7 = vmlal_s8(vprod0x7, vb7x1, va0x1);
+      vprod1x7 = vmlal_s8(vprod1x7, vb7x1, va1x1);
+      vprod2x7 = vmlal_s8(vprod2x7, vb7x1, va2x1);
+      vprod3x7 = vmlal_s8(vprod3x7, vb7x1, va3x1);
+      vacc0x7 = vpadalq_s16(vacc0x7, vprod0x7);
+      vacc1x7 = vpadalq_s16(vacc1x7, vprod1x7);
+      vacc2x7 = vpadalq_s16(vacc2x7, vprod2x7);
+      vacc3x7 = vpadalq_s16(vacc3x7, vprod3x7);
+      const int8x8_t vb8x1 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
+      int16x8_t vprod0x8 = vmull_s8(vb8x0, va0x0);
+      int16x8_t vprod1x8 = vmull_s8(vb8x0, va1x0);
+      int16x8_t vprod2x8 = vmull_s8(vb8x0, va2x0);
+      int16x8_t vprod3x8 = vmull_s8(vb8x0, va3x0);
+      vprod0x8 = vmlal_s8(vprod0x8, vb8x1, va0x1);
+      vprod1x8 = vmlal_s8(vprod1x8, vb8x1, va1x1);
+      vprod2x8 = vmlal_s8(vprod2x8, vb8x1, va2x1);
+      vprod3x8 = vmlal_s8(vprod3x8, vb8x1, va3x1);
+      vacc0x8 = vpadalq_s16(vacc0x8, vprod0x8);
+      vacc1x8 = vpadalq_s16(vacc1x8, vprod1x8);
+      vacc2x8 = vpadalq_s16(vacc2x8, vprod2x8);
+      vacc3x8 = vpadalq_s16(vacc3x8, vprod3x8);
+      const int8x8_t vb9x1 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
+      int16x8_t vprod0x9 = vmull_s8(vb9x0, va0x0);
+      int16x8_t vprod1x9 = vmull_s8(vb9x0, va1x0);
+      int16x8_t vprod2x9 = vmull_s8(vb9x0, va2x0);
+      int16x8_t vprod3x9 = vmull_s8(vb9x0, va3x0);
+      vprod0x9 = vmlal_s8(vprod0x9, vb9x1, va0x1);
+      vprod1x9 = vmlal_s8(vprod1x9, vb9x1, va1x1);
+      vprod2x9 = vmlal_s8(vprod2x9, vb9x1, va2x1);
+      vprod3x9 = vmlal_s8(vprod3x9, vb9x1, va3x1);
+      vacc0x9 = vpadalq_s16(vacc0x9, vprod0x9);
+      vacc1x9 = vpadalq_s16(vacc1x9, vprod1x9);
+      vacc2x9 = vpadalq_s16(vacc2x9, vprod2x9);
+      vacc3x9 = vpadalq_s16(vacc3x9, vprod3x9);
+      const int8x8_t vb10x1 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
+      int16x8_t vprod0x10 = vmull_s8(vb10x0, va0x0);
+      int16x8_t vprod1x10 = vmull_s8(vb10x0, va1x0);
+      int16x8_t vprod2x10 = vmull_s8(vb10x0, va2x0);
+      int16x8_t vprod3x10 = vmull_s8(vb10x0, va3x0);
+      vprod0x10 = vmlal_s8(vprod0x10, vb10x1, va0x1);
+      vprod1x10 = vmlal_s8(vprod1x10, vb10x1, va1x1);
+      vprod2x10 = vmlal_s8(vprod2x10, vb10x1, va2x1);
+      vprod3x10 = vmlal_s8(vprod3x10, vb10x1, va3x1);
+      vacc0x10 = vpadalq_s16(vacc0x10, vprod0x10);
+      vacc1x10 = vpadalq_s16(vacc1x10, vprod1x10);
+      vacc2x10 = vpadalq_s16(vacc2x10, vprod2x10);
+      vacc3x10 = vpadalq_s16(vacc3x10, vprod3x10);
+      const int8x8_t vb11x1 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
+      int16x8_t vprod0x11 = vmull_s8(vb11x0, va0x0);
+      int16x8_t vprod1x11 = vmull_s8(vb11x0, va1x0);
+      int16x8_t vprod2x11 = vmull_s8(vb11x0, va2x0);
+      int16x8_t vprod3x11 = vmull_s8(vb11x0, va3x0);
+      vprod0x11 = vmlal_s8(vprod0x11, vb11x1, va0x1);
+      vprod1x11 = vmlal_s8(vprod1x11, vb11x1, va1x1);
+      vprod2x11 = vmlal_s8(vprod2x11, vb11x1, va2x1);
+      vprod3x11 = vmlal_s8(vprod3x11, vb11x1, va3x1);
+      vacc0x11 = vpadalq_s16(vacc0x11, vprod0x11);
+      vacc1x11 = vpadalq_s16(vacc1x11, vprod1x11);
+      vacc2x11 = vpadalq_s16(vacc2x11, vprod2x11);
+      vacc3x11 = vpadalq_s16(vacc3x11, vprod3x11);
+      const int8x8_t vb12x1 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
+      int16x8_t vprod0x12 = vmull_s8(vb12x0, va0x0);
+      int16x8_t vprod1x12 = vmull_s8(vb12x0, va1x0);
+      int16x8_t vprod2x12 = vmull_s8(vb12x0, va2x0);
+      int16x8_t vprod3x12 = vmull_s8(vb12x0, va3x0);
+      vprod0x12 = vmlal_s8(vprod0x12, vb12x1, va0x1);
+      vprod1x12 = vmlal_s8(vprod1x12, vb12x1, va1x1);
+      vprod2x12 = vmlal_s8(vprod2x12, vb12x1, va2x1);
+      vprod3x12 = vmlal_s8(vprod3x12, vb12x1, va3x1);
+      vacc0x12 = vpadalq_s16(vacc0x12, vprod0x12);
+      vacc1x12 = vpadalq_s16(vacc1x12, vprod1x12);
+      vacc2x12 = vpadalq_s16(vacc2x12, vprod2x12);
+      vacc3x12 = vpadalq_s16(vacc3x12, vprod3x12);
+      const int8x8_t vb13x1 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
+      int16x8_t vprod0x13 = vmull_s8(vb13x0, va0x0);
+      int16x8_t vprod1x13 = vmull_s8(vb13x0, va1x0);
+      int16x8_t vprod2x13 = vmull_s8(vb13x0, va2x0);
+      int16x8_t vprod3x13 = vmull_s8(vb13x0, va3x0);
+      vprod0x13 = vmlal_s8(vprod0x13, vb13x1, va0x1);
+      vprod1x13 = vmlal_s8(vprod1x13, vb13x1, va1x1);
+      vprod2x13 = vmlal_s8(vprod2x13, vb13x1, va2x1);
+      vprod3x13 = vmlal_s8(vprod3x13, vb13x1, va3x1);
+      vacc0x13 = vpadalq_s16(vacc0x13, vprod0x13);
+      vacc1x13 = vpadalq_s16(vacc1x13, vprod1x13);
+      vacc2x13 = vpadalq_s16(vacc2x13, vprod2x13);
+      vacc3x13 = vpadalq_s16(vacc3x13, vprod3x13);
+      const int8x8_t vb14x1 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
+      int16x8_t vprod0x14 = vmull_s8(vb14x0, va0x0);
+      int16x8_t vprod1x14 = vmull_s8(vb14x0, va1x0);
+      int16x8_t vprod2x14 = vmull_s8(vb14x0, va2x0);
+      int16x8_t vprod3x14 = vmull_s8(vb14x0, va3x0);
+      vprod0x14 = vmlal_s8(vprod0x14, vb14x1, va0x1);
+      vprod1x14 = vmlal_s8(vprod1x14, vb14x1, va1x1);
+      vprod2x14 = vmlal_s8(vprod2x14, vb14x1, va2x1);
+      vprod3x14 = vmlal_s8(vprod3x14, vb14x1, va3x1);
+      vacc0x14 = vpadalq_s16(vacc0x14, vprod0x14);
+      vacc1x14 = vpadalq_s16(vacc1x14, vprod1x14);
+      vacc2x14 = vpadalq_s16(vacc2x14, vprod2x14);
+      vacc3x14 = vpadalq_s16(vacc3x14, vprod3x14);
+      const int8x8_t vb15x1 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
+      int16x8_t vprod0x15 = vmull_s8(vb15x0, va0x0);
+      int16x8_t vprod1x15 = vmull_s8(vb15x0, va1x0);
+      int16x8_t vprod2x15 = vmull_s8(vb15x0, va2x0);
+      int16x8_t vprod3x15 = vmull_s8(vb15x0, va3x0);
+      vprod0x15 = vmlal_s8(vprod0x15, vb15x1, va0x1);
+      vprod1x15 = vmlal_s8(vprod1x15, vb15x1, va1x1);
+      vprod2x15 = vmlal_s8(vprod2x15, vb15x1, va2x1);
+      vprod3x15 = vmlal_s8(vprod3x15, vb15x1, va3x1);
+      vacc0x15 = vpadalq_s16(vacc0x15, vprod0x15);
+      vacc1x15 = vpadalq_s16(vacc1x15, vprod1x15);
+      vacc2x15 = vpadalq_s16(vacc2x15, vprod2x15);
+      vacc3x15 = vpadalq_s16(vacc3x15, vprod3x15);
+
+      k -= 16 * sizeof(int8_t);
+    }
+    if (k >= 8 * sizeof(int8_t)) {
       const int8x8_t va0 = vld1_s8(a0); a0 += 8;
       const int8x8_t va1 = vld1_s8(a1); a1 += 8;
       const int8x8_t va2 = vld1_s8(a2); a2 += 8;
       const int8x8_t va3 = vld1_s8(a3); a3 += 8;
 
       const int8x8_t vb0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
-
       const int16x8_t vprod0x0 = vmull_s8(vb0, va0);
-      vacc0x0 = vpadalq_s16(vacc0x0, vprod0x0);
       const int16x8_t vprod1x0 = vmull_s8(vb0, va1);
-      vacc1x0 = vpadalq_s16(vacc1x0, vprod1x0);
       const int16x8_t vprod2x0 = vmull_s8(vb0, va2);
-      vacc2x0 = vpadalq_s16(vacc2x0, vprod2x0);
       const int16x8_t vprod3x0 = vmull_s8(vb0, va3);
+      vacc0x0 = vpadalq_s16(vacc0x0, vprod0x0);
+      vacc1x0 = vpadalq_s16(vacc1x0, vprod1x0);
+      vacc2x0 = vpadalq_s16(vacc2x0, vprod2x0);
       vacc3x0 = vpadalq_s16(vacc3x0, vprod3x0);
       const int8x8_t vb1 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
-
       const int16x8_t vprod0x1 = vmull_s8(vb1, va0);
-      vacc0x1 = vpadalq_s16(vacc0x1, vprod0x1);
       const int16x8_t vprod1x1 = vmull_s8(vb1, va1);
-      vacc1x1 = vpadalq_s16(vacc1x1, vprod1x1);
       const int16x8_t vprod2x1 = vmull_s8(vb1, va2);
-      vacc2x1 = vpadalq_s16(vacc2x1, vprod2x1);
       const int16x8_t vprod3x1 = vmull_s8(vb1, va3);
+      vacc0x1 = vpadalq_s16(vacc0x1, vprod0x1);
+      vacc1x1 = vpadalq_s16(vacc1x1, vprod1x1);
+      vacc2x1 = vpadalq_s16(vacc2x1, vprod2x1);
       vacc3x1 = vpadalq_s16(vacc3x1, vprod3x1);
       const int8x8_t vb2 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
-
       const int16x8_t vprod0x2 = vmull_s8(vb2, va0);
-      vacc0x2 = vpadalq_s16(vacc0x2, vprod0x2);
       const int16x8_t vprod1x2 = vmull_s8(vb2, va1);
-      vacc1x2 = vpadalq_s16(vacc1x2, vprod1x2);
       const int16x8_t vprod2x2 = vmull_s8(vb2, va2);
-      vacc2x2 = vpadalq_s16(vacc2x2, vprod2x2);
       const int16x8_t vprod3x2 = vmull_s8(vb2, va3);
+      vacc0x2 = vpadalq_s16(vacc0x2, vprod0x2);
+      vacc1x2 = vpadalq_s16(vacc1x2, vprod1x2);
+      vacc2x2 = vpadalq_s16(vacc2x2, vprod2x2);
       vacc3x2 = vpadalq_s16(vacc3x2, vprod3x2);
       const int8x8_t vb3 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
-
       const int16x8_t vprod0x3 = vmull_s8(vb3, va0);
-      vacc0x3 = vpadalq_s16(vacc0x3, vprod0x3);
       const int16x8_t vprod1x3 = vmull_s8(vb3, va1);
-      vacc1x3 = vpadalq_s16(vacc1x3, vprod1x3);
       const int16x8_t vprod2x3 = vmull_s8(vb3, va2);
-      vacc2x3 = vpadalq_s16(vacc2x3, vprod2x3);
       const int16x8_t vprod3x3 = vmull_s8(vb3, va3);
+      vacc0x3 = vpadalq_s16(vacc0x3, vprod0x3);
+      vacc1x3 = vpadalq_s16(vacc1x3, vprod1x3);
+      vacc2x3 = vpadalq_s16(vacc2x3, vprod2x3);
       vacc3x3 = vpadalq_s16(vacc3x3, vprod3x3);
       const int8x8_t vb4 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
-
       const int16x8_t vprod0x4 = vmull_s8(vb4, va0);
-      vacc0x4 = vpadalq_s16(vacc0x4, vprod0x4);
       const int16x8_t vprod1x4 = vmull_s8(vb4, va1);
-      vacc1x4 = vpadalq_s16(vacc1x4, vprod1x4);
       const int16x8_t vprod2x4 = vmull_s8(vb4, va2);
-      vacc2x4 = vpadalq_s16(vacc2x4, vprod2x4);
       const int16x8_t vprod3x4 = vmull_s8(vb4, va3);
+      vacc0x4 = vpadalq_s16(vacc0x4, vprod0x4);
+      vacc1x4 = vpadalq_s16(vacc1x4, vprod1x4);
+      vacc2x4 = vpadalq_s16(vacc2x4, vprod2x4);
       vacc3x4 = vpadalq_s16(vacc3x4, vprod3x4);
       const int8x8_t vb5 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
-
       const int16x8_t vprod0x5 = vmull_s8(vb5, va0);
-      vacc0x5 = vpadalq_s16(vacc0x5, vprod0x5);
       const int16x8_t vprod1x5 = vmull_s8(vb5, va1);
-      vacc1x5 = vpadalq_s16(vacc1x5, vprod1x5);
       const int16x8_t vprod2x5 = vmull_s8(vb5, va2);
-      vacc2x5 = vpadalq_s16(vacc2x5, vprod2x5);
       const int16x8_t vprod3x5 = vmull_s8(vb5, va3);
+      vacc0x5 = vpadalq_s16(vacc0x5, vprod0x5);
+      vacc1x5 = vpadalq_s16(vacc1x5, vprod1x5);
+      vacc2x5 = vpadalq_s16(vacc2x5, vprod2x5);
       vacc3x5 = vpadalq_s16(vacc3x5, vprod3x5);
       const int8x8_t vb6 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
-
       const int16x8_t vprod0x6 = vmull_s8(vb6, va0);
-      vacc0x6 = vpadalq_s16(vacc0x6, vprod0x6);
       const int16x8_t vprod1x6 = vmull_s8(vb6, va1);
-      vacc1x6 = vpadalq_s16(vacc1x6, vprod1x6);
       const int16x8_t vprod2x6 = vmull_s8(vb6, va2);
-      vacc2x6 = vpadalq_s16(vacc2x6, vprod2x6);
       const int16x8_t vprod3x6 = vmull_s8(vb6, va3);
+      vacc0x6 = vpadalq_s16(vacc0x6, vprod0x6);
+      vacc1x6 = vpadalq_s16(vacc1x6, vprod1x6);
+      vacc2x6 = vpadalq_s16(vacc2x6, vprod2x6);
       vacc3x6 = vpadalq_s16(vacc3x6, vprod3x6);
       const int8x8_t vb7 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
-
       const int16x8_t vprod0x7 = vmull_s8(vb7, va0);
-      vacc0x7 = vpadalq_s16(vacc0x7, vprod0x7);
       const int16x8_t vprod1x7 = vmull_s8(vb7, va1);
-      vacc1x7 = vpadalq_s16(vacc1x7, vprod1x7);
       const int16x8_t vprod2x7 = vmull_s8(vb7, va2);
-      vacc2x7 = vpadalq_s16(vacc2x7, vprod2x7);
       const int16x8_t vprod3x7 = vmull_s8(vb7, va3);
+      vacc0x7 = vpadalq_s16(vacc0x7, vprod0x7);
+      vacc1x7 = vpadalq_s16(vacc1x7, vprod1x7);
+      vacc2x7 = vpadalq_s16(vacc2x7, vprod2x7);
       vacc3x7 = vpadalq_s16(vacc3x7, vprod3x7);
       const int8x8_t vb8 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
-
       const int16x8_t vprod0x8 = vmull_s8(vb8, va0);
-      vacc0x8 = vpadalq_s16(vacc0x8, vprod0x8);
       const int16x8_t vprod1x8 = vmull_s8(vb8, va1);
-      vacc1x8 = vpadalq_s16(vacc1x8, vprod1x8);
       const int16x8_t vprod2x8 = vmull_s8(vb8, va2);
-      vacc2x8 = vpadalq_s16(vacc2x8, vprod2x8);
       const int16x8_t vprod3x8 = vmull_s8(vb8, va3);
+      vacc0x8 = vpadalq_s16(vacc0x8, vprod0x8);
+      vacc1x8 = vpadalq_s16(vacc1x8, vprod1x8);
+      vacc2x8 = vpadalq_s16(vacc2x8, vprod2x8);
       vacc3x8 = vpadalq_s16(vacc3x8, vprod3x8);
       const int8x8_t vb9 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
-
       const int16x8_t vprod0x9 = vmull_s8(vb9, va0);
-      vacc0x9 = vpadalq_s16(vacc0x9, vprod0x9);
       const int16x8_t vprod1x9 = vmull_s8(vb9, va1);
-      vacc1x9 = vpadalq_s16(vacc1x9, vprod1x9);
       const int16x8_t vprod2x9 = vmull_s8(vb9, va2);
-      vacc2x9 = vpadalq_s16(vacc2x9, vprod2x9);
       const int16x8_t vprod3x9 = vmull_s8(vb9, va3);
+      vacc0x9 = vpadalq_s16(vacc0x9, vprod0x9);
+      vacc1x9 = vpadalq_s16(vacc1x9, vprod1x9);
+      vacc2x9 = vpadalq_s16(vacc2x9, vprod2x9);
       vacc3x9 = vpadalq_s16(vacc3x9, vprod3x9);
       const int8x8_t vb10 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
-
       const int16x8_t vprod0x10 = vmull_s8(vb10, va0);
-      vacc0x10 = vpadalq_s16(vacc0x10, vprod0x10);
       const int16x8_t vprod1x10 = vmull_s8(vb10, va1);
-      vacc1x10 = vpadalq_s16(vacc1x10, vprod1x10);
       const int16x8_t vprod2x10 = vmull_s8(vb10, va2);
-      vacc2x10 = vpadalq_s16(vacc2x10, vprod2x10);
       const int16x8_t vprod3x10 = vmull_s8(vb10, va3);
+      vacc0x10 = vpadalq_s16(vacc0x10, vprod0x10);
+      vacc1x10 = vpadalq_s16(vacc1x10, vprod1x10);
+      vacc2x10 = vpadalq_s16(vacc2x10, vprod2x10);
       vacc3x10 = vpadalq_s16(vacc3x10, vprod3x10);
       const int8x8_t vb11 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
-
       const int16x8_t vprod0x11 = vmull_s8(vb11, va0);
-      vacc0x11 = vpadalq_s16(vacc0x11, vprod0x11);
       const int16x8_t vprod1x11 = vmull_s8(vb11, va1);
-      vacc1x11 = vpadalq_s16(vacc1x11, vprod1x11);
       const int16x8_t vprod2x11 = vmull_s8(vb11, va2);
-      vacc2x11 = vpadalq_s16(vacc2x11, vprod2x11);
       const int16x8_t vprod3x11 = vmull_s8(vb11, va3);
+      vacc0x11 = vpadalq_s16(vacc0x11, vprod0x11);
+      vacc1x11 = vpadalq_s16(vacc1x11, vprod1x11);
+      vacc2x11 = vpadalq_s16(vacc2x11, vprod2x11);
       vacc3x11 = vpadalq_s16(vacc3x11, vprod3x11);
       const int8x8_t vb12 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
-
       const int16x8_t vprod0x12 = vmull_s8(vb12, va0);
-      vacc0x12 = vpadalq_s16(vacc0x12, vprod0x12);
       const int16x8_t vprod1x12 = vmull_s8(vb12, va1);
-      vacc1x12 = vpadalq_s16(vacc1x12, vprod1x12);
       const int16x8_t vprod2x12 = vmull_s8(vb12, va2);
-      vacc2x12 = vpadalq_s16(vacc2x12, vprod2x12);
       const int16x8_t vprod3x12 = vmull_s8(vb12, va3);
+      vacc0x12 = vpadalq_s16(vacc0x12, vprod0x12);
+      vacc1x12 = vpadalq_s16(vacc1x12, vprod1x12);
+      vacc2x12 = vpadalq_s16(vacc2x12, vprod2x12);
       vacc3x12 = vpadalq_s16(vacc3x12, vprod3x12);
       const int8x8_t vb13 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
-
       const int16x8_t vprod0x13 = vmull_s8(vb13, va0);
-      vacc0x13 = vpadalq_s16(vacc0x13, vprod0x13);
       const int16x8_t vprod1x13 = vmull_s8(vb13, va1);
-      vacc1x13 = vpadalq_s16(vacc1x13, vprod1x13);
       const int16x8_t vprod2x13 = vmull_s8(vb13, va2);
-      vacc2x13 = vpadalq_s16(vacc2x13, vprod2x13);
       const int16x8_t vprod3x13 = vmull_s8(vb13, va3);
+      vacc0x13 = vpadalq_s16(vacc0x13, vprod0x13);
+      vacc1x13 = vpadalq_s16(vacc1x13, vprod1x13);
+      vacc2x13 = vpadalq_s16(vacc2x13, vprod2x13);
       vacc3x13 = vpadalq_s16(vacc3x13, vprod3x13);
       const int8x8_t vb14 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
-
       const int16x8_t vprod0x14 = vmull_s8(vb14, va0);
-      vacc0x14 = vpadalq_s16(vacc0x14, vprod0x14);
       const int16x8_t vprod1x14 = vmull_s8(vb14, va1);
-      vacc1x14 = vpadalq_s16(vacc1x14, vprod1x14);
       const int16x8_t vprod2x14 = vmull_s8(vb14, va2);
-      vacc2x14 = vpadalq_s16(vacc2x14, vprod2x14);
       const int16x8_t vprod3x14 = vmull_s8(vb14, va3);
+      vacc0x14 = vpadalq_s16(vacc0x14, vprod0x14);
+      vacc1x14 = vpadalq_s16(vacc1x14, vprod1x14);
+      vacc2x14 = vpadalq_s16(vacc2x14, vprod2x14);
       vacc3x14 = vpadalq_s16(vacc3x14, vprod3x14);
       const int8x8_t vb15 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
-
       const int16x8_t vprod0x15 = vmull_s8(vb15, va0);
-      vacc0x15 = vpadalq_s16(vacc0x15, vprod0x15);
       const int16x8_t vprod1x15 = vmull_s8(vb15, va1);
-      vacc1x15 = vpadalq_s16(vacc1x15, vprod1x15);
       const int16x8_t vprod2x15 = vmull_s8(vb15, va2);
-      vacc2x15 = vpadalq_s16(vacc2x15, vprod2x15);
       const int16x8_t vprod3x15 = vmull_s8(vb15, va3);
+      vacc0x15 = vpadalq_s16(vacc0x15, vprod0x15);
+      vacc1x15 = vpadalq_s16(vacc1x15, vprod1x15);
+      vacc2x15 = vpadalq_s16(vacc2x15, vprod2x15);
       vacc3x15 = vpadalq_s16(vacc3x15, vprod3x15);
 
       k -= 8 * sizeof(int8_t);
@@ -300,164 +522,148 @@
       const int8x8_t va3 = vld1_s8(a3); a3 += k;
 
       const int8x8_t vb0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
-
       const int16x8_t vprod0x0 = vmull_s8(vb0, va0);
-      vacc0x0 = vpadalq_s16(vacc0x0, vprod0x0);
       const int16x8_t vprod1x0 = vmull_s8(vb0, va1);
-      vacc1x0 = vpadalq_s16(vacc1x0, vprod1x0);
       const int16x8_t vprod2x0 = vmull_s8(vb0, va2);
-      vacc2x0 = vpadalq_s16(vacc2x0, vprod2x0);
       const int16x8_t vprod3x0 = vmull_s8(vb0, va3);
+      vacc0x0 = vpadalq_s16(vacc0x0, vprod0x0);
+      vacc1x0 = vpadalq_s16(vacc1x0, vprod1x0);
+      vacc2x0 = vpadalq_s16(vacc2x0, vprod2x0);
       vacc3x0 = vpadalq_s16(vacc3x0, vprod3x0);
       const int8x8_t vb1 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
-
       const int16x8_t vprod0x1 = vmull_s8(vb1, va0);
-      vacc0x1 = vpadalq_s16(vacc0x1, vprod0x1);
       const int16x8_t vprod1x1 = vmull_s8(vb1, va1);
-      vacc1x1 = vpadalq_s16(vacc1x1, vprod1x1);
       const int16x8_t vprod2x1 = vmull_s8(vb1, va2);
-      vacc2x1 = vpadalq_s16(vacc2x1, vprod2x1);
       const int16x8_t vprod3x1 = vmull_s8(vb1, va3);
+      vacc0x1 = vpadalq_s16(vacc0x1, vprod0x1);
+      vacc1x1 = vpadalq_s16(vacc1x1, vprod1x1);
+      vacc2x1 = vpadalq_s16(vacc2x1, vprod2x1);
       vacc3x1 = vpadalq_s16(vacc3x1, vprod3x1);
       const int8x8_t vb2 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
-
       const int16x8_t vprod0x2 = vmull_s8(vb2, va0);
-      vacc0x2 = vpadalq_s16(vacc0x2, vprod0x2);
       const int16x8_t vprod1x2 = vmull_s8(vb2, va1);
-      vacc1x2 = vpadalq_s16(vacc1x2, vprod1x2);
       const int16x8_t vprod2x2 = vmull_s8(vb2, va2);
-      vacc2x2 = vpadalq_s16(vacc2x2, vprod2x2);
       const int16x8_t vprod3x2 = vmull_s8(vb2, va3);
+      vacc0x2 = vpadalq_s16(vacc0x2, vprod0x2);
+      vacc1x2 = vpadalq_s16(vacc1x2, vprod1x2);
+      vacc2x2 = vpadalq_s16(vacc2x2, vprod2x2);
       vacc3x2 = vpadalq_s16(vacc3x2, vprod3x2);
       const int8x8_t vb3 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
-
       const int16x8_t vprod0x3 = vmull_s8(vb3, va0);
-      vacc0x3 = vpadalq_s16(vacc0x3, vprod0x3);
       const int16x8_t vprod1x3 = vmull_s8(vb3, va1);
-      vacc1x3 = vpadalq_s16(vacc1x3, vprod1x3);
       const int16x8_t vprod2x3 = vmull_s8(vb3, va2);
-      vacc2x3 = vpadalq_s16(vacc2x3, vprod2x3);
       const int16x8_t vprod3x3 = vmull_s8(vb3, va3);
+      vacc0x3 = vpadalq_s16(vacc0x3, vprod0x3);
+      vacc1x3 = vpadalq_s16(vacc1x3, vprod1x3);
+      vacc2x3 = vpadalq_s16(vacc2x3, vprod2x3);
       vacc3x3 = vpadalq_s16(vacc3x3, vprod3x3);
       const int8x8_t vb4 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
-
       const int16x8_t vprod0x4 = vmull_s8(vb4, va0);
-      vacc0x4 = vpadalq_s16(vacc0x4, vprod0x4);
       const int16x8_t vprod1x4 = vmull_s8(vb4, va1);
-      vacc1x4 = vpadalq_s16(vacc1x4, vprod1x4);
       const int16x8_t vprod2x4 = vmull_s8(vb4, va2);
-      vacc2x4 = vpadalq_s16(vacc2x4, vprod2x4);
       const int16x8_t vprod3x4 = vmull_s8(vb4, va3);
+      vacc0x4 = vpadalq_s16(vacc0x4, vprod0x4);
+      vacc1x4 = vpadalq_s16(vacc1x4, vprod1x4);
+      vacc2x4 = vpadalq_s16(vacc2x4, vprod2x4);
       vacc3x4 = vpadalq_s16(vacc3x4, vprod3x4);
       const int8x8_t vb5 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
-
       const int16x8_t vprod0x5 = vmull_s8(vb5, va0);
-      vacc0x5 = vpadalq_s16(vacc0x5, vprod0x5);
       const int16x8_t vprod1x5 = vmull_s8(vb5, va1);
-      vacc1x5 = vpadalq_s16(vacc1x5, vprod1x5);
       const int16x8_t vprod2x5 = vmull_s8(vb5, va2);
-      vacc2x5 = vpadalq_s16(vacc2x5, vprod2x5);
       const int16x8_t vprod3x5 = vmull_s8(vb5, va3);
+      vacc0x5 = vpadalq_s16(vacc0x5, vprod0x5);
+      vacc1x5 = vpadalq_s16(vacc1x5, vprod1x5);
+      vacc2x5 = vpadalq_s16(vacc2x5, vprod2x5);
       vacc3x5 = vpadalq_s16(vacc3x5, vprod3x5);
       const int8x8_t vb6 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
-
       const int16x8_t vprod0x6 = vmull_s8(vb6, va0);
-      vacc0x6 = vpadalq_s16(vacc0x6, vprod0x6);
       const int16x8_t vprod1x6 = vmull_s8(vb6, va1);
-      vacc1x6 = vpadalq_s16(vacc1x6, vprod1x6);
       const int16x8_t vprod2x6 = vmull_s8(vb6, va2);
-      vacc2x6 = vpadalq_s16(vacc2x6, vprod2x6);
       const int16x8_t vprod3x6 = vmull_s8(vb6, va3);
+      vacc0x6 = vpadalq_s16(vacc0x6, vprod0x6);
+      vacc1x6 = vpadalq_s16(vacc1x6, vprod1x6);
+      vacc2x6 = vpadalq_s16(vacc2x6, vprod2x6);
       vacc3x6 = vpadalq_s16(vacc3x6, vprod3x6);
       const int8x8_t vb7 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
-
       const int16x8_t vprod0x7 = vmull_s8(vb7, va0);
-      vacc0x7 = vpadalq_s16(vacc0x7, vprod0x7);
       const int16x8_t vprod1x7 = vmull_s8(vb7, va1);
-      vacc1x7 = vpadalq_s16(vacc1x7, vprod1x7);
       const int16x8_t vprod2x7 = vmull_s8(vb7, va2);
-      vacc2x7 = vpadalq_s16(vacc2x7, vprod2x7);
       const int16x8_t vprod3x7 = vmull_s8(vb7, va3);
+      vacc0x7 = vpadalq_s16(vacc0x7, vprod0x7);
+      vacc1x7 = vpadalq_s16(vacc1x7, vprod1x7);
+      vacc2x7 = vpadalq_s16(vacc2x7, vprod2x7);
       vacc3x7 = vpadalq_s16(vacc3x7, vprod3x7);
       const int8x8_t vb8 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
-
       const int16x8_t vprod0x8 = vmull_s8(vb8, va0);
-      vacc0x8 = vpadalq_s16(vacc0x8, vprod0x8);
       const int16x8_t vprod1x8 = vmull_s8(vb8, va1);
-      vacc1x8 = vpadalq_s16(vacc1x8, vprod1x8);
       const int16x8_t vprod2x8 = vmull_s8(vb8, va2);
-      vacc2x8 = vpadalq_s16(vacc2x8, vprod2x8);
       const int16x8_t vprod3x8 = vmull_s8(vb8, va3);
+      vacc0x8 = vpadalq_s16(vacc0x8, vprod0x8);
+      vacc1x8 = vpadalq_s16(vacc1x8, vprod1x8);
+      vacc2x8 = vpadalq_s16(vacc2x8, vprod2x8);
       vacc3x8 = vpadalq_s16(vacc3x8, vprod3x8);
       const int8x8_t vb9 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
-
       const int16x8_t vprod0x9 = vmull_s8(vb9, va0);
-      vacc0x9 = vpadalq_s16(vacc0x9, vprod0x9);
       const int16x8_t vprod1x9 = vmull_s8(vb9, va1);
-      vacc1x9 = vpadalq_s16(vacc1x9, vprod1x9);
       const int16x8_t vprod2x9 = vmull_s8(vb9, va2);
-      vacc2x9 = vpadalq_s16(vacc2x9, vprod2x9);
       const int16x8_t vprod3x9 = vmull_s8(vb9, va3);
+      vacc0x9 = vpadalq_s16(vacc0x9, vprod0x9);
+      vacc1x9 = vpadalq_s16(vacc1x9, vprod1x9);
+      vacc2x9 = vpadalq_s16(vacc2x9, vprod2x9);
       vacc3x9 = vpadalq_s16(vacc3x9, vprod3x9);
       const int8x8_t vb10 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
-
       const int16x8_t vprod0x10 = vmull_s8(vb10, va0);
-      vacc0x10 = vpadalq_s16(vacc0x10, vprod0x10);
       const int16x8_t vprod1x10 = vmull_s8(vb10, va1);
-      vacc1x10 = vpadalq_s16(vacc1x10, vprod1x10);
       const int16x8_t vprod2x10 = vmull_s8(vb10, va2);
-      vacc2x10 = vpadalq_s16(vacc2x10, vprod2x10);
       const int16x8_t vprod3x10 = vmull_s8(vb10, va3);
+      vacc0x10 = vpadalq_s16(vacc0x10, vprod0x10);
+      vacc1x10 = vpadalq_s16(vacc1x10, vprod1x10);
+      vacc2x10 = vpadalq_s16(vacc2x10, vprod2x10);
       vacc3x10 = vpadalq_s16(vacc3x10, vprod3x10);
       const int8x8_t vb11 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
-
       const int16x8_t vprod0x11 = vmull_s8(vb11, va0);
-      vacc0x11 = vpadalq_s16(vacc0x11, vprod0x11);
       const int16x8_t vprod1x11 = vmull_s8(vb11, va1);
-      vacc1x11 = vpadalq_s16(vacc1x11, vprod1x11);
       const int16x8_t vprod2x11 = vmull_s8(vb11, va2);
-      vacc2x11 = vpadalq_s16(vacc2x11, vprod2x11);
       const int16x8_t vprod3x11 = vmull_s8(vb11, va3);
+      vacc0x11 = vpadalq_s16(vacc0x11, vprod0x11);
+      vacc1x11 = vpadalq_s16(vacc1x11, vprod1x11);
+      vacc2x11 = vpadalq_s16(vacc2x11, vprod2x11);
       vacc3x11 = vpadalq_s16(vacc3x11, vprod3x11);
       const int8x8_t vb12 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
-
       const int16x8_t vprod0x12 = vmull_s8(vb12, va0);
-      vacc0x12 = vpadalq_s16(vacc0x12, vprod0x12);
       const int16x8_t vprod1x12 = vmull_s8(vb12, va1);
-      vacc1x12 = vpadalq_s16(vacc1x12, vprod1x12);
       const int16x8_t vprod2x12 = vmull_s8(vb12, va2);
-      vacc2x12 = vpadalq_s16(vacc2x12, vprod2x12);
       const int16x8_t vprod3x12 = vmull_s8(vb12, va3);
+      vacc0x12 = vpadalq_s16(vacc0x12, vprod0x12);
+      vacc1x12 = vpadalq_s16(vacc1x12, vprod1x12);
+      vacc2x12 = vpadalq_s16(vacc2x12, vprod2x12);
       vacc3x12 = vpadalq_s16(vacc3x12, vprod3x12);
       const int8x8_t vb13 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
-
       const int16x8_t vprod0x13 = vmull_s8(vb13, va0);
-      vacc0x13 = vpadalq_s16(vacc0x13, vprod0x13);
       const int16x8_t vprod1x13 = vmull_s8(vb13, va1);
-      vacc1x13 = vpadalq_s16(vacc1x13, vprod1x13);
       const int16x8_t vprod2x13 = vmull_s8(vb13, va2);
-      vacc2x13 = vpadalq_s16(vacc2x13, vprod2x13);
       const int16x8_t vprod3x13 = vmull_s8(vb13, va3);
+      vacc0x13 = vpadalq_s16(vacc0x13, vprod0x13);
+      vacc1x13 = vpadalq_s16(vacc1x13, vprod1x13);
+      vacc2x13 = vpadalq_s16(vacc2x13, vprod2x13);
       vacc3x13 = vpadalq_s16(vacc3x13, vprod3x13);
       const int8x8_t vb14 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
-
       const int16x8_t vprod0x14 = vmull_s8(vb14, va0);
-      vacc0x14 = vpadalq_s16(vacc0x14, vprod0x14);
       const int16x8_t vprod1x14 = vmull_s8(vb14, va1);
-      vacc1x14 = vpadalq_s16(vacc1x14, vprod1x14);
       const int16x8_t vprod2x14 = vmull_s8(vb14, va2);
-      vacc2x14 = vpadalq_s16(vacc2x14, vprod2x14);
       const int16x8_t vprod3x14 = vmull_s8(vb14, va3);
+      vacc0x14 = vpadalq_s16(vacc0x14, vprod0x14);
+      vacc1x14 = vpadalq_s16(vacc1x14, vprod1x14);
+      vacc2x14 = vpadalq_s16(vacc2x14, vprod2x14);
       vacc3x14 = vpadalq_s16(vacc3x14, vprod3x14);
       const int8x8_t vb15 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
-
       const int16x8_t vprod0x15 = vmull_s8(vb15, va0);
-      vacc0x15 = vpadalq_s16(vacc0x15, vprod0x15);
       const int16x8_t vprod1x15 = vmull_s8(vb15, va1);
-      vacc1x15 = vpadalq_s16(vacc1x15, vprod1x15);
       const int16x8_t vprod2x15 = vmull_s8(vb15, va2);
-      vacc2x15 = vpadalq_s16(vacc2x15, vprod2x15);
       const int16x8_t vprod3x15 = vmull_s8(vb15, va3);
+      vacc0x15 = vpadalq_s16(vacc0x15, vprod0x15);
+      vacc1x15 = vpadalq_s16(vacc1x15, vprod1x15);
+      vacc2x15 = vpadalq_s16(vacc2x15, vprod2x15);
       vacc3x15 = vpadalq_s16(vacc3x15, vprod3x15);
     }
 

diff --git a/src/qs8-gemm/gen/4x8c8-minmax-neon-mull-padal.c b/src/qs8-gemm/gen/4x8c8-minmax-neon-mull-padal.c
index 03fdbc7..53bcd2c 100644
--- a/src/qs8-gemm/gen/4x8c8-minmax-neon-mull-padal.c
+++ b/src/qs8-gemm/gen/4x8c8-minmax-neon-mull-padal.c

@@ -92,91 +92,209 @@
     int32x4_t vacc3x7 = vacc0x7;
 
     size_t k = kc;
-    while (k >= 8 * sizeof(int8_t)) {
+    while (k >= 16 * sizeof(int8_t)) {
+      const int8x8_t va0x0 = vld1_s8(a0); a0 += 8;
+      const int8x8_t va0x1 = vld1_s8(a0); a0 += 8;
+      const int8x8_t va1x0 = vld1_s8(a1); a1 += 8;
+      const int8x8_t va1x1 = vld1_s8(a1); a1 += 8;
+      const int8x8_t va2x0 = vld1_s8(a2); a2 += 8;
+      const int8x8_t va2x1 = vld1_s8(a2); a2 += 8;
+      const int8x8_t va3x0 = vld1_s8(a3); a3 += 8;
+      const int8x8_t va3x1 = vld1_s8(a3); a3 += 8;
+
+      const int8x8_t vb0x0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
+      const int8x8_t vb1x0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
+      const int8x8_t vb2x0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
+      const int8x8_t vb3x0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
+      const int8x8_t vb4x0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
+      const int8x8_t vb5x0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
+      const int8x8_t vb6x0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
+      const int8x8_t vb7x0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
+
+      const int8x8_t vb0x1 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
+      int16x8_t vprod0x0 = vmull_s8(vb0x0, va0x0);
+      int16x8_t vprod1x0 = vmull_s8(vb0x0, va1x0);
+      int16x8_t vprod2x0 = vmull_s8(vb0x0, va2x0);
+      int16x8_t vprod3x0 = vmull_s8(vb0x0, va3x0);
+      vprod0x0 = vmlal_s8(vprod0x0, vb0x1, va0x1);
+      vprod1x0 = vmlal_s8(vprod1x0, vb0x1, va1x1);
+      vprod2x0 = vmlal_s8(vprod2x0, vb0x1, va2x1);
+      vprod3x0 = vmlal_s8(vprod3x0, vb0x1, va3x1);
+      vacc0x0 = vpadalq_s16(vacc0x0, vprod0x0);
+      vacc1x0 = vpadalq_s16(vacc1x0, vprod1x0);
+      vacc2x0 = vpadalq_s16(vacc2x0, vprod2x0);
+      vacc3x0 = vpadalq_s16(vacc3x0, vprod3x0);
+      const int8x8_t vb1x1 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
+      int16x8_t vprod0x1 = vmull_s8(vb1x0, va0x0);
+      int16x8_t vprod1x1 = vmull_s8(vb1x0, va1x0);
+      int16x8_t vprod2x1 = vmull_s8(vb1x0, va2x0);
+      int16x8_t vprod3x1 = vmull_s8(vb1x0, va3x0);
+      vprod0x1 = vmlal_s8(vprod0x1, vb1x1, va0x1);
+      vprod1x1 = vmlal_s8(vprod1x1, vb1x1, va1x1);
+      vprod2x1 = vmlal_s8(vprod2x1, vb1x1, va2x1);
+      vprod3x1 = vmlal_s8(vprod3x1, vb1x1, va3x1);
+      vacc0x1 = vpadalq_s16(vacc0x1, vprod0x1);
+      vacc1x1 = vpadalq_s16(vacc1x1, vprod1x1);
+      vacc2x1 = vpadalq_s16(vacc2x1, vprod2x1);
+      vacc3x1 = vpadalq_s16(vacc3x1, vprod3x1);
+      const int8x8_t vb2x1 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
+      int16x8_t vprod0x2 = vmull_s8(vb2x0, va0x0);
+      int16x8_t vprod1x2 = vmull_s8(vb2x0, va1x0);
+      int16x8_t vprod2x2 = vmull_s8(vb2x0, va2x0);
+      int16x8_t vprod3x2 = vmull_s8(vb2x0, va3x0);
+      vprod0x2 = vmlal_s8(vprod0x2, vb2x1, va0x1);
+      vprod1x2 = vmlal_s8(vprod1x2, vb2x1, va1x1);
+      vprod2x2 = vmlal_s8(vprod2x2, vb2x1, va2x1);
+      vprod3x2 = vmlal_s8(vprod3x2, vb2x1, va3x1);
+      vacc0x2 = vpadalq_s16(vacc0x2, vprod0x2);
+      vacc1x2 = vpadalq_s16(vacc1x2, vprod1x2);
+      vacc2x2 = vpadalq_s16(vacc2x2, vprod2x2);
+      vacc3x2 = vpadalq_s16(vacc3x2, vprod3x2);
+      const int8x8_t vb3x1 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
+      int16x8_t vprod0x3 = vmull_s8(vb3x0, va0x0);
+      int16x8_t vprod1x3 = vmull_s8(vb3x0, va1x0);
+      int16x8_t vprod2x3 = vmull_s8(vb3x0, va2x0);
+      int16x8_t vprod3x3 = vmull_s8(vb3x0, va3x0);
+      vprod0x3 = vmlal_s8(vprod0x3, vb3x1, va0x1);
+      vprod1x3 = vmlal_s8(vprod1x3, vb3x1, va1x1);
+      vprod2x3 = vmlal_s8(vprod2x3, vb3x1, va2x1);
+      vprod3x3 = vmlal_s8(vprod3x3, vb3x1, va3x1);
+      vacc0x3 = vpadalq_s16(vacc0x3, vprod0x3);
+      vacc1x3 = vpadalq_s16(vacc1x3, vprod1x3);
+      vacc2x3 = vpadalq_s16(vacc2x3, vprod2x3);
+      vacc3x3 = vpadalq_s16(vacc3x3, vprod3x3);
+      const int8x8_t vb4x1 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
+      int16x8_t vprod0x4 = vmull_s8(vb4x0, va0x0);
+      int16x8_t vprod1x4 = vmull_s8(vb4x0, va1x0);
+      int16x8_t vprod2x4 = vmull_s8(vb4x0, va2x0);
+      int16x8_t vprod3x4 = vmull_s8(vb4x0, va3x0);
+      vprod0x4 = vmlal_s8(vprod0x4, vb4x1, va0x1);
+      vprod1x4 = vmlal_s8(vprod1x4, vb4x1, va1x1);
+      vprod2x4 = vmlal_s8(vprod2x4, vb4x1, va2x1);
+      vprod3x4 = vmlal_s8(vprod3x4, vb4x1, va3x1);
+      vacc0x4 = vpadalq_s16(vacc0x4, vprod0x4);
+      vacc1x4 = vpadalq_s16(vacc1x4, vprod1x4);
+      vacc2x4 = vpadalq_s16(vacc2x4, vprod2x4);
+      vacc3x4 = vpadalq_s16(vacc3x4, vprod3x4);
+      const int8x8_t vb5x1 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
+      int16x8_t vprod0x5 = vmull_s8(vb5x0, va0x0);
+      int16x8_t vprod1x5 = vmull_s8(vb5x0, va1x0);
+      int16x8_t vprod2x5 = vmull_s8(vb5x0, va2x0);
+      int16x8_t vprod3x5 = vmull_s8(vb5x0, va3x0);
+      vprod0x5 = vmlal_s8(vprod0x5, vb5x1, va0x1);
+      vprod1x5 = vmlal_s8(vprod1x5, vb5x1, va1x1);
+      vprod2x5 = vmlal_s8(vprod2x5, vb5x1, va2x1);
+      vprod3x5 = vmlal_s8(vprod3x5, vb5x1, va3x1);
+      vacc0x5 = vpadalq_s16(vacc0x5, vprod0x5);
+      vacc1x5 = vpadalq_s16(vacc1x5, vprod1x5);
+      vacc2x5 = vpadalq_s16(vacc2x5, vprod2x5);
+      vacc3x5 = vpadalq_s16(vacc3x5, vprod3x5);
+      const int8x8_t vb6x1 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
+      int16x8_t vprod0x6 = vmull_s8(vb6x0, va0x0);
+      int16x8_t vprod1x6 = vmull_s8(vb6x0, va1x0);
+      int16x8_t vprod2x6 = vmull_s8(vb6x0, va2x0);
+      int16x8_t vprod3x6 = vmull_s8(vb6x0, va3x0);
+      vprod0x6 = vmlal_s8(vprod0x6, vb6x1, va0x1);
+      vprod1x6 = vmlal_s8(vprod1x6, vb6x1, va1x1);
+      vprod2x6 = vmlal_s8(vprod2x6, vb6x1, va2x1);
+      vprod3x6 = vmlal_s8(vprod3x6, vb6x1, va3x1);
+      vacc0x6 = vpadalq_s16(vacc0x6, vprod0x6);
+      vacc1x6 = vpadalq_s16(vacc1x6, vprod1x6);
+      vacc2x6 = vpadalq_s16(vacc2x6, vprod2x6);
+      vacc3x6 = vpadalq_s16(vacc3x6, vprod3x6);
+      const int8x8_t vb7x1 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
+      int16x8_t vprod0x7 = vmull_s8(vb7x0, va0x0);
+      int16x8_t vprod1x7 = vmull_s8(vb7x0, va1x0);
+      int16x8_t vprod2x7 = vmull_s8(vb7x0, va2x0);
+      int16x8_t vprod3x7 = vmull_s8(vb7x0, va3x0);
+      vprod0x7 = vmlal_s8(vprod0x7, vb7x1, va0x1);
+      vprod1x7 = vmlal_s8(vprod1x7, vb7x1, va1x1);
+      vprod2x7 = vmlal_s8(vprod2x7, vb7x1, va2x1);
+      vprod3x7 = vmlal_s8(vprod3x7, vb7x1, va3x1);
+      vacc0x7 = vpadalq_s16(vacc0x7, vprod0x7);
+      vacc1x7 = vpadalq_s16(vacc1x7, vprod1x7);
+      vacc2x7 = vpadalq_s16(vacc2x7, vprod2x7);
+      vacc3x7 = vpadalq_s16(vacc3x7, vprod3x7);
+
+      k -= 16 * sizeof(int8_t);
+    }
+    if (k >= 8 * sizeof(int8_t)) {
       const int8x8_t va0 = vld1_s8(a0); a0 += 8;
       const int8x8_t va1 = vld1_s8(a1); a1 += 8;
       const int8x8_t va2 = vld1_s8(a2); a2 += 8;
       const int8x8_t va3 = vld1_s8(a3); a3 += 8;
 
       const int8x8_t vb0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
-
       const int16x8_t vprod0x0 = vmull_s8(vb0, va0);
-      vacc0x0 = vpadalq_s16(vacc0x0, vprod0x0);
       const int16x8_t vprod1x0 = vmull_s8(vb0, va1);
-      vacc1x0 = vpadalq_s16(vacc1x0, vprod1x0);
       const int16x8_t vprod2x0 = vmull_s8(vb0, va2);
-      vacc2x0 = vpadalq_s16(vacc2x0, vprod2x0);
       const int16x8_t vprod3x0 = vmull_s8(vb0, va3);
+      vacc0x0 = vpadalq_s16(vacc0x0, vprod0x0);
+      vacc1x0 = vpadalq_s16(vacc1x0, vprod1x0);
+      vacc2x0 = vpadalq_s16(vacc2x0, vprod2x0);
       vacc3x0 = vpadalq_s16(vacc3x0, vprod3x0);
       const int8x8_t vb1 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
-
       const int16x8_t vprod0x1 = vmull_s8(vb1, va0);
-      vacc0x1 = vpadalq_s16(vacc0x1, vprod0x1);
       const int16x8_t vprod1x1 = vmull_s8(vb1, va1);
-      vacc1x1 = vpadalq_s16(vacc1x1, vprod1x1);
       const int16x8_t vprod2x1 = vmull_s8(vb1, va2);
-      vacc2x1 = vpadalq_s16(vacc2x1, vprod2x1);
       const int16x8_t vprod3x1 = vmull_s8(vb1, va3);
+      vacc0x1 = vpadalq_s16(vacc0x1, vprod0x1);
+      vacc1x1 = vpadalq_s16(vacc1x1, vprod1x1);
+      vacc2x1 = vpadalq_s16(vacc2x1, vprod2x1);
       vacc3x1 = vpadalq_s16(vacc3x1, vprod3x1);
       const int8x8_t vb2 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
-
       const int16x8_t vprod0x2 = vmull_s8(vb2, va0);
-      vacc0x2 = vpadalq_s16(vacc0x2, vprod0x2);
       const int16x8_t vprod1x2 = vmull_s8(vb2, va1);
-      vacc1x2 = vpadalq_s16(vacc1x2, vprod1x2);
       const int16x8_t vprod2x2 = vmull_s8(vb2, va2);
-      vacc2x2 = vpadalq_s16(vacc2x2, vprod2x2);
       const int16x8_t vprod3x2 = vmull_s8(vb2, va3);
+      vacc0x2 = vpadalq_s16(vacc0x2, vprod0x2);
+      vacc1x2 = vpadalq_s16(vacc1x2, vprod1x2);
+      vacc2x2 = vpadalq_s16(vacc2x2, vprod2x2);
       vacc3x2 = vpadalq_s16(vacc3x2, vprod3x2);
       const int8x8_t vb3 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
-
       const int16x8_t vprod0x3 = vmull_s8(vb3, va0);
-      vacc0x3 = vpadalq_s16(vacc0x3, vprod0x3);
       const int16x8_t vprod1x3 = vmull_s8(vb3, va1);
-      vacc1x3 = vpadalq_s16(vacc1x3, vprod1x3);
       const int16x8_t vprod2x3 = vmull_s8(vb3, va2);
-      vacc2x3 = vpadalq_s16(vacc2x3, vprod2x3);
       const int16x8_t vprod3x3 = vmull_s8(vb3, va3);
+      vacc0x3 = vpadalq_s16(vacc0x3, vprod0x3);
+      vacc1x3 = vpadalq_s16(vacc1x3, vprod1x3);
+      vacc2x3 = vpadalq_s16(vacc2x3, vprod2x3);
       vacc3x3 = vpadalq_s16(vacc3x3, vprod3x3);
       const int8x8_t vb4 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
-
       const int16x8_t vprod0x4 = vmull_s8(vb4, va0);
-      vacc0x4 = vpadalq_s16(vacc0x4, vprod0x4);
       const int16x8_t vprod1x4 = vmull_s8(vb4, va1);
-      vacc1x4 = vpadalq_s16(vacc1x4, vprod1x4);
       const int16x8_t vprod2x4 = vmull_s8(vb4, va2);
-      vacc2x4 = vpadalq_s16(vacc2x4, vprod2x4);
       const int16x8_t vprod3x4 = vmull_s8(vb4, va3);
+      vacc0x4 = vpadalq_s16(vacc0x4, vprod0x4);
+      vacc1x4 = vpadalq_s16(vacc1x4, vprod1x4);
+      vacc2x4 = vpadalq_s16(vacc2x4, vprod2x4);
       vacc3x4 = vpadalq_s16(vacc3x4, vprod3x4);
       const int8x8_t vb5 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
-
       const int16x8_t vprod0x5 = vmull_s8(vb5, va0);
-      vacc0x5 = vpadalq_s16(vacc0x5, vprod0x5);
       const int16x8_t vprod1x5 = vmull_s8(vb5, va1);
-      vacc1x5 = vpadalq_s16(vacc1x5, vprod1x5);
       const int16x8_t vprod2x5 = vmull_s8(vb5, va2);
-      vacc2x5 = vpadalq_s16(vacc2x5, vprod2x5);
       const int16x8_t vprod3x5 = vmull_s8(vb5, va3);
+      vacc0x5 = vpadalq_s16(vacc0x5, vprod0x5);
+      vacc1x5 = vpadalq_s16(vacc1x5, vprod1x5);
+      vacc2x5 = vpadalq_s16(vacc2x5, vprod2x5);
       vacc3x5 = vpadalq_s16(vacc3x5, vprod3x5);
       const int8x8_t vb6 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
-
       const int16x8_t vprod0x6 = vmull_s8(vb6, va0);
-      vacc0x6 = vpadalq_s16(vacc0x6, vprod0x6);
       const int16x8_t vprod1x6 = vmull_s8(vb6, va1);
-      vacc1x6 = vpadalq_s16(vacc1x6, vprod1x6);
       const int16x8_t vprod2x6 = vmull_s8(vb6, va2);
-      vacc2x6 = vpadalq_s16(vacc2x6, vprod2x6);
       const int16x8_t vprod3x6 = vmull_s8(vb6, va3);
+      vacc0x6 = vpadalq_s16(vacc0x6, vprod0x6);
+      vacc1x6 = vpadalq_s16(vacc1x6, vprod1x6);
+      vacc2x6 = vpadalq_s16(vacc2x6, vprod2x6);
       vacc3x6 = vpadalq_s16(vacc3x6, vprod3x6);
       const int8x8_t vb7 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
-
       const int16x8_t vprod0x7 = vmull_s8(vb7, va0);
-      vacc0x7 = vpadalq_s16(vacc0x7, vprod0x7);
       const int16x8_t vprod1x7 = vmull_s8(vb7, va1);
-      vacc1x7 = vpadalq_s16(vacc1x7, vprod1x7);
       const int16x8_t vprod2x7 = vmull_s8(vb7, va2);
-      vacc2x7 = vpadalq_s16(vacc2x7, vprod2x7);
       const int16x8_t vprod3x7 = vmull_s8(vb7, va3);
+      vacc0x7 = vpadalq_s16(vacc0x7, vprod0x7);
+      vacc1x7 = vpadalq_s16(vacc1x7, vprod1x7);
+      vacc2x7 = vpadalq_s16(vacc2x7, vprod2x7);
       vacc3x7 = vpadalq_s16(vacc3x7, vprod3x7);
 
       k -= 8 * sizeof(int8_t);
@@ -188,84 +306,76 @@
       const int8x8_t va3 = vld1_s8(a3); a3 += k;
 
       const int8x8_t vb0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
-
       const int16x8_t vprod0x0 = vmull_s8(vb0, va0);
-      vacc0x0 = vpadalq_s16(vacc0x0, vprod0x0);
       const int16x8_t vprod1x0 = vmull_s8(vb0, va1);
-      vacc1x0 = vpadalq_s16(vacc1x0, vprod1x0);
       const int16x8_t vprod2x0 = vmull_s8(vb0, va2);
-      vacc2x0 = vpadalq_s16(vacc2x0, vprod2x0);
       const int16x8_t vprod3x0 = vmull_s8(vb0, va3);
+      vacc0x0 = vpadalq_s16(vacc0x0, vprod0x0);
+      vacc1x0 = vpadalq_s16(vacc1x0, vprod1x0);
+      vacc2x0 = vpadalq_s16(vacc2x0, vprod2x0);
       vacc3x0 = vpadalq_s16(vacc3x0, vprod3x0);
       const int8x8_t vb1 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
-
       const int16x8_t vprod0x1 = vmull_s8(vb1, va0);
-      vacc0x1 = vpadalq_s16(vacc0x1, vprod0x1);
       const int16x8_t vprod1x1 = vmull_s8(vb1, va1);
-      vacc1x1 = vpadalq_s16(vacc1x1, vprod1x1);
       const int16x8_t vprod2x1 = vmull_s8(vb1, va2);
-      vacc2x1 = vpadalq_s16(vacc2x1, vprod2x1);
       const int16x8_t vprod3x1 = vmull_s8(vb1, va3);
+      vacc0x1 = vpadalq_s16(vacc0x1, vprod0x1);
+      vacc1x1 = vpadalq_s16(vacc1x1, vprod1x1);
+      vacc2x1 = vpadalq_s16(vacc2x1, vprod2x1);
       vacc3x1 = vpadalq_s16(vacc3x1, vprod3x1);
       const int8x8_t vb2 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
-
       const int16x8_t vprod0x2 = vmull_s8(vb2, va0);
-      vacc0x2 = vpadalq_s16(vacc0x2, vprod0x2);
       const int16x8_t vprod1x2 = vmull_s8(vb2, va1);
-      vacc1x2 = vpadalq_s16(vacc1x2, vprod1x2);
       const int16x8_t vprod2x2 = vmull_s8(vb2, va2);
-      vacc2x2 = vpadalq_s16(vacc2x2, vprod2x2);
       const int16x8_t vprod3x2 = vmull_s8(vb2, va3);
+      vacc0x2 = vpadalq_s16(vacc0x2, vprod0x2);
+      vacc1x2 = vpadalq_s16(vacc1x2, vprod1x2);
+      vacc2x2 = vpadalq_s16(vacc2x2, vprod2x2);
       vacc3x2 = vpadalq_s16(vacc3x2, vprod3x2);
       const int8x8_t vb3 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
-
       const int16x8_t vprod0x3 = vmull_s8(vb3, va0);
-      vacc0x3 = vpadalq_s16(vacc0x3, vprod0x3);
       const int16x8_t vprod1x3 = vmull_s8(vb3, va1);
-      vacc1x3 = vpadalq_s16(vacc1x3, vprod1x3);
       const int16x8_t vprod2x3 = vmull_s8(vb3, va2);
-      vacc2x3 = vpadalq_s16(vacc2x3, vprod2x3);
       const int16x8_t vprod3x3 = vmull_s8(vb3, va3);
+      vacc0x3 = vpadalq_s16(vacc0x3, vprod0x3);
+      vacc1x3 = vpadalq_s16(vacc1x3, vprod1x3);
+      vacc2x3 = vpadalq_s16(vacc2x3, vprod2x3);
       vacc3x3 = vpadalq_s16(vacc3x3, vprod3x3);
       const int8x8_t vb4 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
-
       const int16x8_t vprod0x4 = vmull_s8(vb4, va0);
-      vacc0x4 = vpadalq_s16(vacc0x4, vprod0x4);
       const int16x8_t vprod1x4 = vmull_s8(vb4, va1);
-      vacc1x4 = vpadalq_s16(vacc1x4, vprod1x4);
       const int16x8_t vprod2x4 = vmull_s8(vb4, va2);
-      vacc2x4 = vpadalq_s16(vacc2x4, vprod2x4);
       const int16x8_t vprod3x4 = vmull_s8(vb4, va3);
+      vacc0x4 = vpadalq_s16(vacc0x4, vprod0x4);
+      vacc1x4 = vpadalq_s16(vacc1x4, vprod1x4);
+      vacc2x4 = vpadalq_s16(vacc2x4, vprod2x4);
       vacc3x4 = vpadalq_s16(vacc3x4, vprod3x4);
       const int8x8_t vb5 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
-
       const int16x8_t vprod0x5 = vmull_s8(vb5, va0);
-      vacc0x5 = vpadalq_s16(vacc0x5, vprod0x5);
       const int16x8_t vprod1x5 = vmull_s8(vb5, va1);
-      vacc1x5 = vpadalq_s16(vacc1x5, vprod1x5);
       const int16x8_t vprod2x5 = vmull_s8(vb5, va2);
-      vacc2x5 = vpadalq_s16(vacc2x5, vprod2x5);
       const int16x8_t vprod3x5 = vmull_s8(vb5, va3);
+      vacc0x5 = vpadalq_s16(vacc0x5, vprod0x5);
+      vacc1x5 = vpadalq_s16(vacc1x5, vprod1x5);
+      vacc2x5 = vpadalq_s16(vacc2x5, vprod2x5);
       vacc3x5 = vpadalq_s16(vacc3x5, vprod3x5);
       const int8x8_t vb6 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
-
       const int16x8_t vprod0x6 = vmull_s8(vb6, va0);
-      vacc0x6 = vpadalq_s16(vacc0x6, vprod0x6);
       const int16x8_t vprod1x6 = vmull_s8(vb6, va1);
-      vacc1x6 = vpadalq_s16(vacc1x6, vprod1x6);
       const int16x8_t vprod2x6 = vmull_s8(vb6, va2);
-      vacc2x6 = vpadalq_s16(vacc2x6, vprod2x6);
       const int16x8_t vprod3x6 = vmull_s8(vb6, va3);
+      vacc0x6 = vpadalq_s16(vacc0x6, vprod0x6);
+      vacc1x6 = vpadalq_s16(vacc1x6, vprod1x6);
+      vacc2x6 = vpadalq_s16(vacc2x6, vprod2x6);
       vacc3x6 = vpadalq_s16(vacc3x6, vprod3x6);
       const int8x8_t vb7 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
-
       const int16x8_t vprod0x7 = vmull_s8(vb7, va0);
-      vacc0x7 = vpadalq_s16(vacc0x7, vprod0x7);
       const int16x8_t vprod1x7 = vmull_s8(vb7, va1);
-      vacc1x7 = vpadalq_s16(vacc1x7, vprod1x7);
       const int16x8_t vprod2x7 = vmull_s8(vb7, va2);
-      vacc2x7 = vpadalq_s16(vacc2x7, vprod2x7);
       const int16x8_t vprod3x7 = vmull_s8(vb7, va3);
+      vacc0x7 = vpadalq_s16(vacc0x7, vprod0x7);
+      vacc1x7 = vpadalq_s16(vacc1x7, vprod1x7);
+      vacc2x7 = vpadalq_s16(vacc2x7, vprod2x7);
       vacc3x7 = vpadalq_s16(vacc3x7, vprod3x7);
     }
 

diff --git a/test/gemm-microkernel-tester.h b/test/gemm-microkernel-tester.h
index 20ec6d4..c0ef22b 100644
--- a/test/gemm-microkernel-tester.h
+++ b/test/gemm-microkernel-tester.h

@@ -440,7 +440,7 @@
     auto rng = std::mt19937(random_device());
     auto i32rng = std::bind(std::uniform_int_distribution<int32_t>(-10000, 10000), rng);
     auto i8rng = std::bind(
-      std::uniform_int_distribution<int32_t>(std::numeric_limits<int8_t>::min(), std::numeric_limits<int8_t>::max()), rng);
+      std::uniform_int_distribution<int32_t>(-127, std::numeric_limits<int8_t>::max()), rng);
 
     std::vector<int8_t> a((m() - 1) * a_stride() + k() + XNN_EXTRA_BYTES / sizeof(int8_t));
     std::vector<int8_t> b(n() * k());

diff --git a/test/qs8-gemm-minmax.cc b/test/qs8-gemm-minmax.cc
index 2fee42d..06dd0d3 100644
--- a/test/qs8-gemm-minmax.cc
+++ b/test/qs8-gemm-minmax.cc

@@ -14615,7 +14615,7 @@
 
 
 #if XNN_ARCH_ARM || XNN_ARCH_ARM64
-  TEST(QS8_GEMM_MINMAX_1X8C8__NEON_MULL_PADAL, k_eq_8) {
+  TEST(QS8_GEMM_MINMAX_1X8C8__NEON_MULL_PADAL, k_eq_16) {
     TEST_REQUIRES_ARM_NEON;
     GemmMicrokernelTester()
       .mr(1)
@@ -14624,7 +14624,7 @@
       .sr(1)
       .m(1)
       .n(8)
-      .k(8)
+      .k(16)
       .Test(xnn_qs8_gemm_minmax_ukernel_1x8c8__neon_mull_padal);
   }
 
@@ -14637,12 +14637,12 @@
       .sr(1)
       .m(1)
       .n(8)
-      .k(8)
+      .k(16)
       .cn_stride(11)
       .Test(xnn_qs8_gemm_minmax_ukernel_1x8c8__neon_mull_padal);
   }
 
-  TEST(QS8_GEMM_MINMAX_1X8C8__NEON_MULL_PADAL, k_eq_8_strided_a) {
+  TEST(QS8_GEMM_MINMAX_1X8C8__NEON_MULL_PADAL, k_eq_16_strided_a) {
     TEST_REQUIRES_ARM_NEON;
     GemmMicrokernelTester()
       .mr(1)
@@ -14651,12 +14651,12 @@
       .sr(1)
       .m(1)
       .n(8)
-      .k(8)
-      .a_stride(11)
+      .k(16)
+      .a_stride(19)
       .Test(xnn_qs8_gemm_minmax_ukernel_1x8c8__neon_mull_padal);
   }
 
-  TEST(QS8_GEMM_MINMAX_1X8C8__NEON_MULL_PADAL, k_eq_8_subtile) {
+  TEST(QS8_GEMM_MINMAX_1X8C8__NEON_MULL_PADAL, k_eq_16_subtile) {
     TEST_REQUIRES_ARM_NEON;
     for (uint32_t m = 1; m <= 1; m++) {
       for (uint32_t n = 1; n <= 8; n++) {
@@ -14667,14 +14667,14 @@
           .sr(1)
           .m(m)
           .n(n)
-          .k(8)
+          .k(16)
           .iterations(1)
           .Test(xnn_qs8_gemm_minmax_ukernel_1x8c8__neon_mull_padal);
       }
     }
   }
 
-  TEST(QS8_GEMM_MINMAX_1X8C8__NEON_MULL_PADAL, k_eq_8_subtile_m) {
+  TEST(QS8_GEMM_MINMAX_1X8C8__NEON_MULL_PADAL, k_eq_16_subtile_m) {
     TEST_REQUIRES_ARM_NEON;
     for (uint32_t m = 1; m <= 1; m++) {
       GemmMicrokernelTester()
@@ -14684,13 +14684,13 @@
         .sr(1)
         .m(m)
         .n(8)
-        .k(8)
+        .k(16)
         .iterations(1)
         .Test(xnn_qs8_gemm_minmax_ukernel_1x8c8__neon_mull_padal);
     }
   }
 
-  TEST(QS8_GEMM_MINMAX_1X8C8__NEON_MULL_PADAL, k_eq_8_subtile_n) {
+  TEST(QS8_GEMM_MINMAX_1X8C8__NEON_MULL_PADAL, k_eq_16_subtile_n) {
     TEST_REQUIRES_ARM_NEON;
     for (uint32_t n = 1; n <= 8; n++) {
       GemmMicrokernelTester()
@@ -14700,15 +14700,15 @@
         .sr(1)
         .m(1)
         .n(n)
-        .k(8)
+        .k(16)
         .iterations(1)
         .Test(xnn_qs8_gemm_minmax_ukernel_1x8c8__neon_mull_padal);
     }
   }
 
-  TEST(QS8_GEMM_MINMAX_1X8C8__NEON_MULL_PADAL, k_lt_8) {
+  TEST(QS8_GEMM_MINMAX_1X8C8__NEON_MULL_PADAL, k_lt_16) {
     TEST_REQUIRES_ARM_NEON;
-    for (size_t k = 1; k < 8; k++) {
+    for (size_t k = 1; k < 16; k++) {
       GemmMicrokernelTester()
         .mr(1)
         .nr(8)
@@ -14721,60 +14721,9 @@
     }
   }
 
-  TEST(QS8_GEMM_MINMAX_1X8C8__NEON_MULL_PADAL, k_lt_8_strided_a) {
+  TEST(QS8_GEMM_MINMAX_1X8C8__NEON_MULL_PADAL, k_lt_16_strided_a) {
     TEST_REQUIRES_ARM_NEON;
-    for (size_t k = 1; k < 8; k++) {
-      GemmMicrokernelTester()
-        .mr(1)
-        .nr(8)
-        .kr(8)
-        .sr(1)
-        .m(1)
-        .n(8)
-        .k(k)
-        .a_stride(11)
-        .Test(xnn_qs8_gemm_minmax_ukernel_1x8c8__neon_mull_padal);
-    }
-  }
-
-  TEST(QS8_GEMM_MINMAX_1X8C8__NEON_MULL_PADAL, k_lt_8_subtile) {
-    TEST_REQUIRES_ARM_NEON;
-    for (size_t k = 1; k < 8; k++) {
-      for (uint32_t m = 1; m <= 1; m++) {
-        for (uint32_t n = 1; n <= 8; n++) {
-          GemmMicrokernelTester()
-            .mr(1)
-            .nr(8)
-            .kr(8)
-            .sr(1)
-            .m(m)
-            .n(n)
-            .k(k)
-            .iterations(1)
-            .Test(xnn_qs8_gemm_minmax_ukernel_1x8c8__neon_mull_padal);
-        }
-      }
-    }
-  }
-
-  TEST(QS8_GEMM_MINMAX_1X8C8__NEON_MULL_PADAL, k_gt_8) {
-    TEST_REQUIRES_ARM_NEON;
-    for (size_t k = 9; k < 16; k++) {
-      GemmMicrokernelTester()
-        .mr(1)
-        .nr(8)
-        .kr(8)
-        .sr(1)
-        .m(1)
-        .n(8)
-        .k(k)
-        .Test(xnn_qs8_gemm_minmax_ukernel_1x8c8__neon_mull_padal);
-    }
-  }
-
-  TEST(QS8_GEMM_MINMAX_1X8C8__NEON_MULL_PADAL, k_gt_8_strided_a) {
-    TEST_REQUIRES_ARM_NEON;
-    for (size_t k = 9; k < 16; k++) {
+    for (size_t k = 1; k < 16; k++) {
       GemmMicrokernelTester()
         .mr(1)
         .nr(8)
@@ -14788,9 +14737,9 @@
     }
   }
 
-  TEST(QS8_GEMM_MINMAX_1X8C8__NEON_MULL_PADAL, k_gt_8_subtile) {
+  TEST(QS8_GEMM_MINMAX_1X8C8__NEON_MULL_PADAL, k_lt_16_subtile) {
     TEST_REQUIRES_ARM_NEON;
-    for (size_t k = 9; k < 16; k++) {
+    for (size_t k = 1; k < 16; k++) {
       for (uint32_t m = 1; m <= 1; m++) {
         for (uint32_t n = 1; n <= 8; n++) {
           GemmMicrokernelTester()
@@ -14808,9 +14757,9 @@
     }
   }
 
-  TEST(QS8_GEMM_MINMAX_1X8C8__NEON_MULL_PADAL, k_div_8) {
+  TEST(QS8_GEMM_MINMAX_1X8C8__NEON_MULL_PADAL, k_gt_16) {
     TEST_REQUIRES_ARM_NEON;
-    for (size_t k = 16; k <= 80; k += 8) {
+    for (size_t k = 17; k < 32; k++) {
       GemmMicrokernelTester()
         .mr(1)
         .nr(8)
@@ -14823,9 +14772,9 @@
     }
   }
 
-  TEST(QS8_GEMM_MINMAX_1X8C8__NEON_MULL_PADAL, k_div_8_strided_a) {
+  TEST(QS8_GEMM_MINMAX_1X8C8__NEON_MULL_PADAL, k_gt_16_strided_a) {
     TEST_REQUIRES_ARM_NEON;
-    for (size_t k = 16; k <= 80; k += 8) {
+    for (size_t k = 17; k < 32; k++) {
       GemmMicrokernelTester()
         .mr(1)
         .nr(8)
@@ -14834,14 +14783,65 @@
         .m(1)
         .n(8)
         .k(k)
-        .a_stride(83)
+        .a_stride(37)
         .Test(xnn_qs8_gemm_minmax_ukernel_1x8c8__neon_mull_padal);
     }
   }
 
-  TEST(QS8_GEMM_MINMAX_1X8C8__NEON_MULL_PADAL, k_div_8_subtile) {
+  TEST(QS8_GEMM_MINMAX_1X8C8__NEON_MULL_PADAL, k_gt_16_subtile) {
     TEST_REQUIRES_ARM_NEON;
-    for (size_t k = 16; k <= 80; k += 8) {
+    for (size_t k = 17; k < 32; k++) {
+      for (uint32_t m = 1; m <= 1; m++) {
+        for (uint32_t n = 1; n <= 8; n++) {
+          GemmMicrokernelTester()
+            .mr(1)
+            .nr(8)
+            .kr(8)
+            .sr(1)
+            .m(m)
+            .n(n)
+            .k(k)
+            .iterations(1)
+            .Test(xnn_qs8_gemm_minmax_ukernel_1x8c8__neon_mull_padal);
+        }
+      }
+    }
+  }
+
+  TEST(QS8_GEMM_MINMAX_1X8C8__NEON_MULL_PADAL, k_div_16) {
+    TEST_REQUIRES_ARM_NEON;
+    for (size_t k = 32; k <= 160; k += 16) {
+      GemmMicrokernelTester()
+        .mr(1)
+        .nr(8)
+        .kr(8)
+        .sr(1)
+        .m(1)
+        .n(8)
+        .k(k)
+        .Test(xnn_qs8_gemm_minmax_ukernel_1x8c8__neon_mull_padal);
+    }
+  }
+
+  TEST(QS8_GEMM_MINMAX_1X8C8__NEON_MULL_PADAL, k_div_16_strided_a) {
+    TEST_REQUIRES_ARM_NEON;
+    for (size_t k = 32; k <= 160; k += 16) {
+      GemmMicrokernelTester()
+        .mr(1)
+        .nr(8)
+        .kr(8)
+        .sr(1)
+        .m(1)
+        .n(8)
+        .k(k)
+        .a_stride(163)
+        .Test(xnn_qs8_gemm_minmax_ukernel_1x8c8__neon_mull_padal);
+    }
+  }
+
+  TEST(QS8_GEMM_MINMAX_1X8C8__NEON_MULL_PADAL, k_div_16_subtile) {
+    TEST_REQUIRES_ARM_NEON;
+    for (size_t k = 32; k <= 160; k += 16) {
       for (uint32_t m = 1; m <= 1; m++) {
         for (uint32_t n = 1; n <= 8; n++) {
           GemmMicrokernelTester()
@@ -14862,7 +14862,7 @@
   TEST(QS8_GEMM_MINMAX_1X8C8__NEON_MULL_PADAL, n_gt_8) {
     TEST_REQUIRES_ARM_NEON;
     for (uint32_t n = 9; n < 16; n++) {
-      for (size_t k = 1; k <= 40; k += 9) {
+      for (size_t k = 1; k <= 80; k += 17) {
         GemmMicrokernelTester()
           .mr(1)
           .nr(8)
@@ -14879,7 +14879,7 @@
   TEST(QS8_GEMM_MINMAX_1X8C8__NEON_MULL_PADAL, n_gt_8_strided_cn) {
     TEST_REQUIRES_ARM_NEON;
     for (uint32_t n = 9; n < 16; n++) {
-      for (size_t k = 1; k <= 40; k += 9) {
+      for (size_t k = 1; k <= 80; k += 17) {
         GemmMicrokernelTester()
           .mr(1)
           .nr(8)
@@ -14897,7 +14897,7 @@
   TEST(QS8_GEMM_MINMAX_1X8C8__NEON_MULL_PADAL, n_gt_8_strided_a) {
     TEST_REQUIRES_ARM_NEON;
     for (uint32_t n = 9; n < 16; n++) {
-      for (size_t k = 1; k <= 40; k += 9) {
+      for (size_t k = 1; k <= 80; k += 17) {
         GemmMicrokernelTester()
           .mr(1)
           .nr(8)
@@ -14906,7 +14906,7 @@
           .m(1)
           .n(n)
           .k(k)
-          .a_stride(43)
+          .a_stride(83)
           .Test(xnn_qs8_gemm_minmax_ukernel_1x8c8__neon_mull_padal);
       }
     }
@@ -14915,7 +14915,7 @@
   TEST(QS8_GEMM_MINMAX_1X8C8__NEON_MULL_PADAL, n_gt_8_subtile) {
     TEST_REQUIRES_ARM_NEON;
     for (uint32_t n = 9; n < 16; n++) {
-      for (size_t k = 1; k <= 40; k += 9) {
+      for (size_t k = 1; k <= 80; k += 17) {
         for (uint32_t m = 1; m <= 1; m++) {
           GemmMicrokernelTester()
             .mr(1)
@@ -14935,7 +14935,7 @@
   TEST(QS8_GEMM_MINMAX_1X8C8__NEON_MULL_PADAL, n_div_8) {
     TEST_REQUIRES_ARM_NEON;
     for (uint32_t n = 16; n <= 24; n += 8) {
-      for (size_t k = 1; k <= 40; k += 9) {
+      for (size_t k = 1; k <= 80; k += 17) {
         GemmMicrokernelTester()
           .mr(1)
           .nr(8)
@@ -14952,7 +14952,7 @@
   TEST(QS8_GEMM_MINMAX_1X8C8__NEON_MULL_PADAL, n_div_8_strided_cn) {
     TEST_REQUIRES_ARM_NEON;
     for (uint32_t n = 16; n <= 24; n += 8) {
-      for (size_t k = 1; k <= 40; k += 9) {
+      for (size_t k = 1; k <= 80; k += 17) {
         GemmMicrokernelTester()
           .mr(1)
           .nr(8)
@@ -14970,7 +14970,7 @@
   TEST(QS8_GEMM_MINMAX_1X8C8__NEON_MULL_PADAL, n_div_8_strided_a) {
     TEST_REQUIRES_ARM_NEON;
     for (uint32_t n = 16; n <= 24; n += 8) {
-      for (size_t k = 1; k <= 40; k += 9) {
+      for (size_t k = 1; k <= 80; k += 17) {
         GemmMicrokernelTester()
           .mr(1)
           .nr(8)
@@ -14979,7 +14979,7 @@
           .m(1)
           .n(n)
           .k(k)
-          .a_stride(43)
+          .a_stride(83)
           .Test(xnn_qs8_gemm_minmax_ukernel_1x8c8__neon_mull_padal);
       }
     }
@@ -14988,7 +14988,7 @@
   TEST(QS8_GEMM_MINMAX_1X8C8__NEON_MULL_PADAL, n_div_8_subtile) {
     TEST_REQUIRES_ARM_NEON;
     for (uint32_t n = 16; n <= 24; n += 8) {
-      for (size_t k = 1; k <= 40; k += 9) {
+      for (size_t k = 1; k <= 80; k += 17) {
         for (uint32_t m = 1; m <= 1; m++) {
           GemmMicrokernelTester()
             .mr(1)
@@ -15007,7 +15007,7 @@
 
   TEST(QS8_GEMM_MINMAX_1X8C8__NEON_MULL_PADAL, strided_cm_subtile) {
     TEST_REQUIRES_ARM_NEON;
-    for (size_t k = 1; k <= 40; k += 9) {
+    for (size_t k = 1; k <= 80; k += 17) {
       for (uint32_t m = 1; m <= 1; m++) {
         for (uint32_t n = 1; n <= 8; n++) {
           GemmMicrokernelTester()
@@ -15035,7 +15035,7 @@
       .sr(1)
       .m(1)
       .n(8)
-      .k(8)
+      .k(16)
       .qmin(128)
       .Test(xnn_qs8_gemm_minmax_ukernel_1x8c8__neon_mull_padal);
   }
@@ -15049,7 +15049,7 @@
       .sr(1)
       .m(1)
       .n(8)
-      .k(8)
+      .k(16)
       .qmax(128)
       .Test(xnn_qs8_gemm_minmax_ukernel_1x8c8__neon_mull_padal);
   }
@@ -15063,7 +15063,7 @@
       .sr(1)
       .m(1)
       .n(8)
-      .k(8)
+      .k(16)
       .cm_stride(11)
       .Test(xnn_qs8_gemm_minmax_ukernel_1x8c8__neon_mull_padal);
   }
@@ -15071,7 +15071,7 @@
 
 
 #if XNN_ARCH_ARM || XNN_ARCH_ARM64
-  TEST(QS8_GEMM_MINMAX_2X8C8__NEON_MULL_PADAL, k_eq_8) {
+  TEST(QS8_GEMM_MINMAX_2X8C8__NEON_MULL_PADAL, k_eq_16) {
     TEST_REQUIRES_ARM_NEON;
     GemmMicrokernelTester()
       .mr(2)
@@ -15080,7 +15080,7 @@
       .sr(1)
       .m(2)
       .n(8)
-      .k(8)
+      .k(16)
       .Test(xnn_qs8_gemm_minmax_ukernel_2x8c8__neon_mull_padal);
   }
 
@@ -15093,12 +15093,12 @@
       .sr(1)
       .m(2)
       .n(8)
-      .k(8)
+      .k(16)
       .cn_stride(11)
       .Test(xnn_qs8_gemm_minmax_ukernel_2x8c8__neon_mull_padal);
   }
 
-  TEST(QS8_GEMM_MINMAX_2X8C8__NEON_MULL_PADAL, k_eq_8_strided_a) {
+  TEST(QS8_GEMM_MINMAX_2X8C8__NEON_MULL_PADAL, k_eq_16_strided_a) {
     TEST_REQUIRES_ARM_NEON;
     GemmMicrokernelTester()
       .mr(2)
@@ -15107,12 +15107,12 @@
       .sr(1)
       .m(2)
       .n(8)
-      .k(8)
-      .a_stride(11)
+      .k(16)
+      .a_stride(19)
       .Test(xnn_qs8_gemm_minmax_ukernel_2x8c8__neon_mull_padal);
   }
 
-  TEST(QS8_GEMM_MINMAX_2X8C8__NEON_MULL_PADAL, k_eq_8_subtile) {
+  TEST(QS8_GEMM_MINMAX_2X8C8__NEON_MULL_PADAL, k_eq_16_subtile) {
     TEST_REQUIRES_ARM_NEON;
     for (uint32_t m = 1; m <= 2; m++) {
       for (uint32_t n = 1; n <= 8; n++) {
@@ -15123,14 +15123,14 @@
           .sr(1)
           .m(m)
           .n(n)
-          .k(8)
+          .k(16)
           .iterations(1)
           .Test(xnn_qs8_gemm_minmax_ukernel_2x8c8__neon_mull_padal);
       }
     }
   }
 
-  TEST(QS8_GEMM_MINMAX_2X8C8__NEON_MULL_PADAL, k_eq_8_subtile_m) {
+  TEST(QS8_GEMM_MINMAX_2X8C8__NEON_MULL_PADAL, k_eq_16_subtile_m) {
     TEST_REQUIRES_ARM_NEON;
     for (uint32_t m = 1; m <= 2; m++) {
       GemmMicrokernelTester()
@@ -15140,13 +15140,13 @@
         .sr(1)
         .m(m)
         .n(8)
-        .k(8)
+        .k(16)
         .iterations(1)
         .Test(xnn_qs8_gemm_minmax_ukernel_2x8c8__neon_mull_padal);
     }
   }
 
-  TEST(QS8_GEMM_MINMAX_2X8C8__NEON_MULL_PADAL, k_eq_8_subtile_n) {
+  TEST(QS8_GEMM_MINMAX_2X8C8__NEON_MULL_PADAL, k_eq_16_subtile_n) {
     TEST_REQUIRES_ARM_NEON;
     for (uint32_t n = 1; n <= 8; n++) {
       GemmMicrokernelTester()
@@ -15156,15 +15156,15 @@
         .sr(1)
         .m(2)
         .n(n)
-        .k(8)
+        .k(16)
         .iterations(1)
         .Test(xnn_qs8_gemm_minmax_ukernel_2x8c8__neon_mull_padal);
     }
   }
 
-  TEST(QS8_GEMM_MINMAX_2X8C8__NEON_MULL_PADAL, k_lt_8) {
+  TEST(QS8_GEMM_MINMAX_2X8C8__NEON_MULL_PADAL, k_lt_16) {
     TEST_REQUIRES_ARM_NEON;
-    for (size_t k = 1; k < 8; k++) {
+    for (size_t k = 1; k < 16; k++) {
       GemmMicrokernelTester()
         .mr(2)
         .nr(8)
@@ -15177,60 +15177,9 @@
     }
   }
 
-  TEST(QS8_GEMM_MINMAX_2X8C8__NEON_MULL_PADAL, k_lt_8_strided_a) {
+  TEST(QS8_GEMM_MINMAX_2X8C8__NEON_MULL_PADAL, k_lt_16_strided_a) {
     TEST_REQUIRES_ARM_NEON;
-    for (size_t k = 1; k < 8; k++) {
-      GemmMicrokernelTester()
-        .mr(2)
-        .nr(8)
-        .kr(8)
-        .sr(1)
-        .m(2)
-        .n(8)
-        .k(k)
-        .a_stride(11)
-        .Test(xnn_qs8_gemm_minmax_ukernel_2x8c8__neon_mull_padal);
-    }
-  }
-
-  TEST(QS8_GEMM_MINMAX_2X8C8__NEON_MULL_PADAL, k_lt_8_subtile) {
-    TEST_REQUIRES_ARM_NEON;
-    for (size_t k = 1; k < 8; k++) {
-      for (uint32_t m = 1; m <= 2; m++) {
-        for (uint32_t n = 1; n <= 8; n++) {
-          GemmMicrokernelTester()
-            .mr(2)
-            .nr(8)
-            .kr(8)
-            .sr(1)
-            .m(m)
-            .n(n)
-            .k(k)
-            .iterations(1)
-            .Test(xnn_qs8_gemm_minmax_ukernel_2x8c8__neon_mull_padal);
-        }
-      }
-    }
-  }
-
-  TEST(QS8_GEMM_MINMAX_2X8C8__NEON_MULL_PADAL, k_gt_8) {
-    TEST_REQUIRES_ARM_NEON;
-    for (size_t k = 9; k < 16; k++) {
-      GemmMicrokernelTester()
-        .mr(2)
-        .nr(8)
-        .kr(8)
-        .sr(1)
-        .m(2)
-        .n(8)
-        .k(k)
-        .Test(xnn_qs8_gemm_minmax_ukernel_2x8c8__neon_mull_padal);
-    }
-  }
-
-  TEST(QS8_GEMM_MINMAX_2X8C8__NEON_MULL_PADAL, k_gt_8_strided_a) {
-    TEST_REQUIRES_ARM_NEON;
-    for (size_t k = 9; k < 16; k++) {
+    for (size_t k = 1; k < 16; k++) {
       GemmMicrokernelTester()
         .mr(2)
         .nr(8)
@@ -15244,9 +15193,9 @@
     }
   }
 
-  TEST(QS8_GEMM_MINMAX_2X8C8__NEON_MULL_PADAL, k_gt_8_subtile) {
+  TEST(QS8_GEMM_MINMAX_2X8C8__NEON_MULL_PADAL, k_lt_16_subtile) {
     TEST_REQUIRES_ARM_NEON;
-    for (size_t k = 9; k < 16; k++) {
+    for (size_t k = 1; k < 16; k++) {
       for (uint32_t m = 1; m <= 2; m++) {
         for (uint32_t n = 1; n <= 8; n++) {
           GemmMicrokernelTester()
@@ -15264,9 +15213,9 @@
     }
   }
 
-  TEST(QS8_GEMM_MINMAX_2X8C8__NEON_MULL_PADAL, k_div_8) {
+  TEST(QS8_GEMM_MINMAX_2X8C8__NEON_MULL_PADAL, k_gt_16) {
     TEST_REQUIRES_ARM_NEON;
-    for (size_t k = 16; k <= 80; k += 8) {
+    for (size_t k = 17; k < 32; k++) {
       GemmMicrokernelTester()
         .mr(2)
         .nr(8)
@@ -15279,9 +15228,9 @@
     }
   }
 
-  TEST(QS8_GEMM_MINMAX_2X8C8__NEON_MULL_PADAL, k_div_8_strided_a) {
+  TEST(QS8_GEMM_MINMAX_2X8C8__NEON_MULL_PADAL, k_gt_16_strided_a) {
     TEST_REQUIRES_ARM_NEON;
-    for (size_t k = 16; k <= 80; k += 8) {
+    for (size_t k = 17; k < 32; k++) {
       GemmMicrokernelTester()
         .mr(2)
         .nr(8)
@@ -15290,14 +15239,65 @@
         .m(2)
         .n(8)
         .k(k)
-        .a_stride(83)
+        .a_stride(37)
         .Test(xnn_qs8_gemm_minmax_ukernel_2x8c8__neon_mull_padal);
     }
   }
 
-  TEST(QS8_GEMM_MINMAX_2X8C8__NEON_MULL_PADAL, k_div_8_subtile) {
+  TEST(QS8_GEMM_MINMAX_2X8C8__NEON_MULL_PADAL, k_gt_16_subtile) {
     TEST_REQUIRES_ARM_NEON;
-    for (size_t k = 16; k <= 80; k += 8) {
+    for (size_t k = 17; k < 32; k++) {
+      for (uint32_t m = 1; m <= 2; m++) {
+        for (uint32_t n = 1; n <= 8; n++) {
+          GemmMicrokernelTester()
+            .mr(2)
+            .nr(8)
+            .kr(8)
+            .sr(1)
+            .m(m)
+            .n(n)
+            .k(k)
+            .iterations(1)
+            .Test(xnn_qs8_gemm_minmax_ukernel_2x8c8__neon_mull_padal);
+        }
+      }
+    }
+  }
+
+  TEST(QS8_GEMM_MINMAX_2X8C8__NEON_MULL_PADAL, k_div_16) {
+    TEST_REQUIRES_ARM_NEON;
+    for (size_t k = 32; k <= 160; k += 16) {
+      GemmMicrokernelTester()
+        .mr(2)
+        .nr(8)
+        .kr(8)
+        .sr(1)
+        .m(2)
+        .n(8)
+        .k(k)
+        .Test(xnn_qs8_gemm_minmax_ukernel_2x8c8__neon_mull_padal);
+    }
+  }
+
+  TEST(QS8_GEMM_MINMAX_2X8C8__NEON_MULL_PADAL, k_div_16_strided_a) {
+    TEST_REQUIRES_ARM_NEON;
+    for (size_t k = 32; k <= 160; k += 16) {
+      GemmMicrokernelTester()
+        .mr(2)
+        .nr(8)
+        .kr(8)
+        .sr(1)
+        .m(2)
+        .n(8)
+        .k(k)
+        .a_stride(163)
+        .Test(xnn_qs8_gemm_minmax_ukernel_2x8c8__neon_mull_padal);
+    }
+  }
+
+  TEST(QS8_GEMM_MINMAX_2X8C8__NEON_MULL_PADAL, k_div_16_subtile) {
+    TEST_REQUIRES_ARM_NEON;
+    for (size_t k = 32; k <= 160; k += 16) {
       for (uint32_t m = 1; m <= 2; m++) {
         for (uint32_t n = 1; n <= 8; n++) {
           GemmMicrokernelTester()
@@ -15318,7 +15318,7 @@
   TEST(QS8_GEMM_MINMAX_2X8C8__NEON_MULL_PADAL, n_gt_8) {
     TEST_REQUIRES_ARM_NEON;
     for (uint32_t n = 9; n < 16; n++) {
-      for (size_t k = 1; k <= 40; k += 9) {
+      for (size_t k = 1; k <= 80; k += 17) {
         GemmMicrokernelTester()
           .mr(2)
           .nr(8)
@@ -15335,7 +15335,7 @@
   TEST(QS8_GEMM_MINMAX_2X8C8__NEON_MULL_PADAL, n_gt_8_strided_cn) {
     TEST_REQUIRES_ARM_NEON;
     for (uint32_t n = 9; n < 16; n++) {
-      for (size_t k = 1; k <= 40; k += 9) {
+      for (size_t k = 1; k <= 80; k += 17) {
         GemmMicrokernelTester()
           .mr(2)
           .nr(8)
@@ -15353,7 +15353,7 @@
   TEST(QS8_GEMM_MINMAX_2X8C8__NEON_MULL_PADAL, n_gt_8_strided_a) {
     TEST_REQUIRES_ARM_NEON;
     for (uint32_t n = 9; n < 16; n++) {
-      for (size_t k = 1; k <= 40; k += 9) {
+      for (size_t k = 1; k <= 80; k += 17) {
         GemmMicrokernelTester()
           .mr(2)
           .nr(8)
@@ -15362,7 +15362,7 @@
           .m(2)
           .n(n)
           .k(k)
-          .a_stride(43)
+          .a_stride(83)
           .Test(xnn_qs8_gemm_minmax_ukernel_2x8c8__neon_mull_padal);
       }
     }
@@ -15371,7 +15371,7 @@
   TEST(QS8_GEMM_MINMAX_2X8C8__NEON_MULL_PADAL, n_gt_8_subtile) {
     TEST_REQUIRES_ARM_NEON;
     for (uint32_t n = 9; n < 16; n++) {
-      for (size_t k = 1; k <= 40; k += 9) {
+      for (size_t k = 1; k <= 80; k += 17) {
         for (uint32_t m = 1; m <= 2; m++) {
           GemmMicrokernelTester()
             .mr(2)
@@ -15391,7 +15391,7 @@
   TEST(QS8_GEMM_MINMAX_2X8C8__NEON_MULL_PADAL, n_div_8) {
     TEST_REQUIRES_ARM_NEON;
     for (uint32_t n = 16; n <= 24; n += 8) {
-      for (size_t k = 1; k <= 40; k += 9) {
+      for (size_t k = 1; k <= 80; k += 17) {
         GemmMicrokernelTester()
           .mr(2)
           .nr(8)
@@ -15408,7 +15408,7 @@
   TEST(QS8_GEMM_MINMAX_2X8C8__NEON_MULL_PADAL, n_div_8_strided_cn) {
     TEST_REQUIRES_ARM_NEON;
     for (uint32_t n = 16; n <= 24; n += 8) {
-      for (size_t k = 1; k <= 40; k += 9) {
+      for (size_t k = 1; k <= 80; k += 17) {
         GemmMicrokernelTester()
           .mr(2)
           .nr(8)
@@ -15426,7 +15426,7 @@
   TEST(QS8_GEMM_MINMAX_2X8C8__NEON_MULL_PADAL, n_div_8_strided_a) {
     TEST_REQUIRES_ARM_NEON;
     for (uint32_t n = 16; n <= 24; n += 8) {
-      for (size_t k = 1; k <= 40; k += 9) {
+      for (size_t k = 1; k <= 80; k += 17) {
         GemmMicrokernelTester()
           .mr(2)
           .nr(8)
@@ -15435,7 +15435,7 @@
           .m(2)
           .n(n)
           .k(k)
-          .a_stride(43)
+          .a_stride(83)
           .Test(xnn_qs8_gemm_minmax_ukernel_2x8c8__neon_mull_padal);
       }
     }
@@ -15444,7 +15444,7 @@
   TEST(QS8_GEMM_MINMAX_2X8C8__NEON_MULL_PADAL, n_div_8_subtile) {
     TEST_REQUIRES_ARM_NEON;
     for (uint32_t n = 16; n <= 24; n += 8) {
-      for (size_t k = 1; k <= 40; k += 9) {
+      for (size_t k = 1; k <= 80; k += 17) {
         for (uint32_t m = 1; m <= 2; m++) {
           GemmMicrokernelTester()
             .mr(2)
@@ -15463,7 +15463,7 @@
 
   TEST(QS8_GEMM_MINMAX_2X8C8__NEON_MULL_PADAL, strided_cm_subtile) {
     TEST_REQUIRES_ARM_NEON;
-    for (size_t k = 1; k <= 40; k += 9) {
+    for (size_t k = 1; k <= 80; k += 17) {
       for (uint32_t m = 1; m <= 2; m++) {
         for (uint32_t n = 1; n <= 8; n++) {
           GemmMicrokernelTester()
@@ -15491,7 +15491,7 @@
       .sr(1)
       .m(2)
       .n(8)
-      .k(8)
+      .k(16)
       .qmin(128)
       .Test(xnn_qs8_gemm_minmax_ukernel_2x8c8__neon_mull_padal);
   }
@@ -15505,7 +15505,7 @@
       .sr(1)
       .m(2)
       .n(8)
-      .k(8)
+      .k(16)
       .qmax(128)
       .Test(xnn_qs8_gemm_minmax_ukernel_2x8c8__neon_mull_padal);
   }
@@ -15519,7 +15519,7 @@
       .sr(1)
       .m(2)
       .n(8)
-      .k(8)
+      .k(16)
       .cm_stride(11)
       .Test(xnn_qs8_gemm_minmax_ukernel_2x8c8__neon_mull_padal);
   }
@@ -15527,7 +15527,7 @@
 
 
 #if XNN_ARCH_ARM || XNN_ARCH_ARM64
-  TEST(QS8_GEMM_MINMAX_3X8C8__NEON_MULL_PADAL, k_eq_8) {
+  TEST(QS8_GEMM_MINMAX_3X8C8__NEON_MULL_PADAL, k_eq_16) {
     TEST_REQUIRES_ARM_NEON;
     GemmMicrokernelTester()
       .mr(3)
@@ -15536,7 +15536,7 @@
       .sr(1)
       .m(3)
       .n(8)
-      .k(8)
+      .k(16)
       .Test(xnn_qs8_gemm_minmax_ukernel_3x8c8__neon_mull_padal);
   }
 
@@ -15549,12 +15549,12 @@
       .sr(1)
       .m(3)
       .n(8)
-      .k(8)
+      .k(16)
       .cn_stride(11)
       .Test(xnn_qs8_gemm_minmax_ukernel_3x8c8__neon_mull_padal);
   }
 
-  TEST(QS8_GEMM_MINMAX_3X8C8__NEON_MULL_PADAL, k_eq_8_strided_a) {
+  TEST(QS8_GEMM_MINMAX_3X8C8__NEON_MULL_PADAL, k_eq_16_strided_a) {
     TEST_REQUIRES_ARM_NEON;
     GemmMicrokernelTester()
       .mr(3)
@@ -15563,12 +15563,12 @@
       .sr(1)
       .m(3)
       .n(8)
-      .k(8)
-      .a_stride(11)
+      .k(16)
+      .a_stride(19)
       .Test(xnn_qs8_gemm_minmax_ukernel_3x8c8__neon_mull_padal);
   }
 
-  TEST(QS8_GEMM_MINMAX_3X8C8__NEON_MULL_PADAL, k_eq_8_subtile) {
+  TEST(QS8_GEMM_MINMAX_3X8C8__NEON_MULL_PADAL, k_eq_16_subtile) {
     TEST_REQUIRES_ARM_NEON;
     for (uint32_t m = 1; m <= 3; m++) {
       for (uint32_t n = 1; n <= 8; n++) {
@@ -15579,14 +15579,14 @@
           .sr(1)
           .m(m)
           .n(n)
-          .k(8)
+          .k(16)
           .iterations(1)
           .Test(xnn_qs8_gemm_minmax_ukernel_3x8c8__neon_mull_padal);
       }
     }
   }
 
-  TEST(QS8_GEMM_MINMAX_3X8C8__NEON_MULL_PADAL, k_eq_8_subtile_m) {
+  TEST(QS8_GEMM_MINMAX_3X8C8__NEON_MULL_PADAL, k_eq_16_subtile_m) {
     TEST_REQUIRES_ARM_NEON;
     for (uint32_t m = 1; m <= 3; m++) {
       GemmMicrokernelTester()
@@ -15596,13 +15596,13 @@
         .sr(1)
         .m(m)
         .n(8)
-        .k(8)
+        .k(16)
         .iterations(1)
         .Test(xnn_qs8_gemm_minmax_ukernel_3x8c8__neon_mull_padal);
     }
   }
 
-  TEST(QS8_GEMM_MINMAX_3X8C8__NEON_MULL_PADAL, k_eq_8_subtile_n) {
+  TEST(QS8_GEMM_MINMAX_3X8C8__NEON_MULL_PADAL, k_eq_16_subtile_n) {
     TEST_REQUIRES_ARM_NEON;
     for (uint32_t n = 1; n <= 8; n++) {
       GemmMicrokernelTester()
@@ -15612,15 +15612,15 @@
         .sr(1)
         .m(3)
         .n(n)
-        .k(8)
+        .k(16)
         .iterations(1)
         .Test(xnn_qs8_gemm_minmax_ukernel_3x8c8__neon_mull_padal);
     }
   }
 
-  TEST(QS8_GEMM_MINMAX_3X8C8__NEON_MULL_PADAL, k_lt_8) {
+  TEST(QS8_GEMM_MINMAX_3X8C8__NEON_MULL_PADAL, k_lt_16) {
     TEST_REQUIRES_ARM_NEON;
-    for (size_t k = 1; k < 8; k++) {
+    for (size_t k = 1; k < 16; k++) {
       GemmMicrokernelTester()
         .mr(3)
         .nr(8)
@@ -15633,60 +15633,9 @@
     }
   }
 
-  TEST(QS8_GEMM_MINMAX_3X8C8__NEON_MULL_PADAL, k_lt_8_strided_a) {
+  TEST(QS8_GEMM_MINMAX_3X8C8__NEON_MULL_PADAL, k_lt_16_strided_a) {
     TEST_REQUIRES_ARM_NEON;
-    for (size_t k = 1; k < 8; k++) {
-      GemmMicrokernelTester()
-        .mr(3)
-        .nr(8)
-        .kr(8)
-        .sr(1)
-        .m(3)
-        .n(8)
-        .k(k)
-        .a_stride(11)
-        .Test(xnn_qs8_gemm_minmax_ukernel_3x8c8__neon_mull_padal);
-    }
-  }
-
-  TEST(QS8_GEMM_MINMAX_3X8C8__NEON_MULL_PADAL, k_lt_8_subtile) {
-    TEST_REQUIRES_ARM_NEON;
-    for (size_t k = 1; k < 8; k++) {
-      for (uint32_t m = 1; m <= 3; m++) {
-        for (uint32_t n = 1; n <= 8; n++) {
-          GemmMicrokernelTester()
-            .mr(3)
-            .nr(8)
-            .kr(8)
-            .sr(1)
-            .m(m)
-            .n(n)
-            .k(k)
-            .iterations(1)
-            .Test(xnn_qs8_gemm_minmax_ukernel_3x8c8__neon_mull_padal);
-        }
-      }
-    }
-  }
-
-  TEST(QS8_GEMM_MINMAX_3X8C8__NEON_MULL_PADAL, k_gt_8) {
-    TEST_REQUIRES_ARM_NEON;
-    for (size_t k = 9; k < 16; k++) {
-      GemmMicrokernelTester()
-        .mr(3)
-        .nr(8)
-        .kr(8)
-        .sr(1)
-        .m(3)
-        .n(8)
-        .k(k)
-        .Test(xnn_qs8_gemm_minmax_ukernel_3x8c8__neon_mull_padal);
-    }
-  }
-
-  TEST(QS8_GEMM_MINMAX_3X8C8__NEON_MULL_PADAL, k_gt_8_strided_a) {
-    TEST_REQUIRES_ARM_NEON;
-    for (size_t k = 9; k < 16; k++) {
+    for (size_t k = 1; k < 16; k++) {
       GemmMicrokernelTester()
         .mr(3)
         .nr(8)
@@ -15700,9 +15649,9 @@
     }
   }
 
-  TEST(QS8_GEMM_MINMAX_3X8C8__NEON_MULL_PADAL, k_gt_8_subtile) {
+  TEST(QS8_GEMM_MINMAX_3X8C8__NEON_MULL_PADAL, k_lt_16_subtile) {
     TEST_REQUIRES_ARM_NEON;
-    for (size_t k = 9; k < 16; k++) {
+    for (size_t k = 1; k < 16; k++) {
       for (uint32_t m = 1; m <= 3; m++) {
         for (uint32_t n = 1; n <= 8; n++) {
           GemmMicrokernelTester()
@@ -15720,9 +15669,9 @@
     }
   }
 
-  TEST(QS8_GEMM_MINMAX_3X8C8__NEON_MULL_PADAL, k_div_8) {
+  TEST(QS8_GEMM_MINMAX_3X8C8__NEON_MULL_PADAL, k_gt_16) {
     TEST_REQUIRES_ARM_NEON;
-    for (size_t k = 16; k <= 80; k += 8) {
+    for (size_t k = 17; k < 32; k++) {
       GemmMicrokernelTester()
         .mr(3)
         .nr(8)
@@ -15735,9 +15684,9 @@
     }
   }
 
-  TEST(QS8_GEMM_MINMAX_3X8C8__NEON_MULL_PADAL, k_div_8_strided_a) {
+  TEST(QS8_GEMM_MINMAX_3X8C8__NEON_MULL_PADAL, k_gt_16_strided_a) {
     TEST_REQUIRES_ARM_NEON;
-    for (size_t k = 16; k <= 80; k += 8) {
+    for (size_t k = 17; k < 32; k++) {
       GemmMicrokernelTester()
         .mr(3)
         .nr(8)
@@ -15746,14 +15695,65 @@
         .m(3)
         .n(8)
         .k(k)
-        .a_stride(83)
+        .a_stride(37)
         .Test(xnn_qs8_gemm_minmax_ukernel_3x8c8__neon_mull_padal);
     }
   }
 
-  TEST(QS8_GEMM_MINMAX_3X8C8__NEON_MULL_PADAL, k_div_8_subtile) {
+  TEST(QS8_GEMM_MINMAX_3X8C8__NEON_MULL_PADAL, k_gt_16_subtile) {
     TEST_REQUIRES_ARM_NEON;
-    for (size_t k = 16; k <= 80; k += 8) {
+    for (size_t k = 17; k < 32; k++) {
+      for (uint32_t m = 1; m <= 3; m++) {
+        for (uint32_t n = 1; n <= 8; n++) {
+          GemmMicrokernelTester()
+            .mr(3)
+            .nr(8)
+            .kr(8)
+            .sr(1)
+            .m(m)
+            .n(n)
+            .k(k)
+            .iterations(1)
+            .Test(xnn_qs8_gemm_minmax_ukernel_3x8c8__neon_mull_padal);
+        }
+      }
+    }
+  }
+
+  TEST(QS8_GEMM_MINMAX_3X8C8__NEON_MULL_PADAL, k_div_16) {
+    TEST_REQUIRES_ARM_NEON;
+    for (size_t k = 32; k <= 160; k += 16) {
+      GemmMicrokernelTester()
+        .mr(3)
+        .nr(8)
+        .kr(8)
+        .sr(1)
+        .m(3)
+        .n(8)
+        .k(k)
+        .Test(xnn_qs8_gemm_minmax_ukernel_3x8c8__neon_mull_padal);
+    }
+  }
+
+  TEST(QS8_GEMM_MINMAX_3X8C8__NEON_MULL_PADAL, k_div_16_strided_a) {
+    TEST_REQUIRES_ARM_NEON;
+    for (size_t k = 32; k <= 160; k += 16) {
+      GemmMicrokernelTester()
+        .mr(3)
+        .nr(8)
+        .kr(8)
+        .sr(1)
+        .m(3)
+        .n(8)
+        .k(k)
+        .a_stride(163)
+        .Test(xnn_qs8_gemm_minmax_ukernel_3x8c8__neon_mull_padal);
+    }
+  }
+
+  TEST(QS8_GEMM_MINMAX_3X8C8__NEON_MULL_PADAL, k_div_16_subtile) {
+    TEST_REQUIRES_ARM_NEON;
+    for (size_t k = 32; k <= 160; k += 16) {
       for (uint32_t m = 1; m <= 3; m++) {
         for (uint32_t n = 1; n <= 8; n++) {
           GemmMicrokernelTester()
@@ -15774,7 +15774,7 @@
   TEST(QS8_GEMM_MINMAX_3X8C8__NEON_MULL_PADAL, n_gt_8) {
     TEST_REQUIRES_ARM_NEON;
     for (uint32_t n = 9; n < 16; n++) {
-      for (size_t k = 1; k <= 40; k += 9) {
+      for (size_t k = 1; k <= 80; k += 17) {
         GemmMicrokernelTester()
           .mr(3)
           .nr(8)
@@ -15791,7 +15791,7 @@
   TEST(QS8_GEMM_MINMAX_3X8C8__NEON_MULL_PADAL, n_gt_8_strided_cn) {
     TEST_REQUIRES_ARM_NEON;
     for (uint32_t n = 9; n < 16; n++) {
-      for (size_t k = 1; k <= 40; k += 9) {
+      for (size_t k = 1; k <= 80; k += 17) {
         GemmMicrokernelTester()
           .mr(3)
           .nr(8)
@@ -15809,7 +15809,7 @@
   TEST(QS8_GEMM_MINMAX_3X8C8__NEON_MULL_PADAL, n_gt_8_strided_a) {
     TEST_REQUIRES_ARM_NEON;
     for (uint32_t n = 9; n < 16; n++) {
-      for (size_t k = 1; k <= 40; k += 9) {
+      for (size_t k = 1; k <= 80; k += 17) {
         GemmMicrokernelTester()
           .mr(3)
           .nr(8)
@@ -15818,7 +15818,7 @@
           .m(3)
           .n(n)
           .k(k)
-          .a_stride(43)
+          .a_stride(83)
           .Test(xnn_qs8_gemm_minmax_ukernel_3x8c8__neon_mull_padal);
       }
     }
@@ -15827,7 +15827,7 @@
   TEST(QS8_GEMM_MINMAX_3X8C8__NEON_MULL_PADAL, n_gt_8_subtile) {
     TEST_REQUIRES_ARM_NEON;
     for (uint32_t n = 9; n < 16; n++) {
-      for (size_t k = 1; k <= 40; k += 9) {
+      for (size_t k = 1; k <= 80; k += 17) {
         for (uint32_t m = 1; m <= 3; m++) {
           GemmMicrokernelTester()
             .mr(3)
@@ -15847,7 +15847,7 @@
   TEST(QS8_GEMM_MINMAX_3X8C8__NEON_MULL_PADAL, n_div_8) {
     TEST_REQUIRES_ARM_NEON;
     for (uint32_t n = 16; n <= 24; n += 8) {
-      for (size_t k = 1; k <= 40; k += 9) {
+      for (size_t k = 1; k <= 80; k += 17) {
         GemmMicrokernelTester()
           .mr(3)
           .nr(8)
@@ -15864,7 +15864,7 @@
   TEST(QS8_GEMM_MINMAX_3X8C8__NEON_MULL_PADAL, n_div_8_strided_cn) {
     TEST_REQUIRES_ARM_NEON;
     for (uint32_t n = 16; n <= 24; n += 8) {
-      for (size_t k = 1; k <= 40; k += 9) {
+      for (size_t k = 1; k <= 80; k += 17) {
         GemmMicrokernelTester()
           .mr(3)
           .nr(8)
@@ -15882,7 +15882,7 @@
   TEST(QS8_GEMM_MINMAX_3X8C8__NEON_MULL_PADAL, n_div_8_strided_a) {
     TEST_REQUIRES_ARM_NEON;
     for (uint32_t n = 16; n <= 24; n += 8) {
-      for (size_t k = 1; k <= 40; k += 9) {
+      for (size_t k = 1; k <= 80; k += 17) {
         GemmMicrokernelTester()
           .mr(3)
           .nr(8)
@@ -15891,7 +15891,7 @@
           .m(3)
           .n(n)
           .k(k)
-          .a_stride(43)
+          .a_stride(83)
           .Test(xnn_qs8_gemm_minmax_ukernel_3x8c8__neon_mull_padal);
       }
     }
@@ -15900,7 +15900,7 @@
   TEST(QS8_GEMM_MINMAX_3X8C8__NEON_MULL_PADAL, n_div_8_subtile) {
     TEST_REQUIRES_ARM_NEON;
     for (uint32_t n = 16; n <= 24; n += 8) {
-      for (size_t k = 1; k <= 40; k += 9) {
+      for (size_t k = 1; k <= 80; k += 17) {
         for (uint32_t m = 1; m <= 3; m++) {
           GemmMicrokernelTester()
             .mr(3)
@@ -15919,7 +15919,7 @@
 
   TEST(QS8_GEMM_MINMAX_3X8C8__NEON_MULL_PADAL, strided_cm_subtile) {
     TEST_REQUIRES_ARM_NEON;
-    for (size_t k = 1; k <= 40; k += 9) {
+    for (size_t k = 1; k <= 80; k += 17) {
       for (uint32_t m = 1; m <= 3; m++) {
         for (uint32_t n = 1; n <= 8; n++) {
           GemmMicrokernelTester()
@@ -15947,7 +15947,7 @@
       .sr(1)
       .m(3)
       .n(8)
-      .k(8)
+      .k(16)
       .qmin(128)
       .Test(xnn_qs8_gemm_minmax_ukernel_3x8c8__neon_mull_padal);
   }
@@ -15961,7 +15961,7 @@
       .sr(1)
       .m(3)
       .n(8)
-      .k(8)
+      .k(16)
       .qmax(128)
       .Test(xnn_qs8_gemm_minmax_ukernel_3x8c8__neon_mull_padal);
   }
@@ -15975,7 +15975,7 @@
       .sr(1)
       .m(3)
       .n(8)
-      .k(8)
+      .k(16)
       .cm_stride(11)
       .Test(xnn_qs8_gemm_minmax_ukernel_3x8c8__neon_mull_padal);
   }
@@ -15983,7 +15983,7 @@
 
 
 #if XNN_ARCH_ARM || XNN_ARCH_ARM64
-  TEST(QS8_GEMM_MINMAX_4X8C8__NEON_MULL_PADAL, k_eq_8) {
+  TEST(QS8_GEMM_MINMAX_4X8C8__NEON_MULL_PADAL, k_eq_16) {
     TEST_REQUIRES_ARM_NEON;
     GemmMicrokernelTester()
       .mr(4)
@@ -15992,7 +15992,7 @@
       .sr(1)
       .m(4)
       .n(8)
-      .k(8)
+      .k(16)
       .Test(xnn_qs8_gemm_minmax_ukernel_4x8c8__neon_mull_padal);
   }
 
@@ -16005,12 +16005,12 @@
       .sr(1)
       .m(4)
       .n(8)
-      .k(8)
+      .k(16)
       .cn_stride(11)
       .Test(xnn_qs8_gemm_minmax_ukernel_4x8c8__neon_mull_padal);
   }
 
-  TEST(QS8_GEMM_MINMAX_4X8C8__NEON_MULL_PADAL, k_eq_8_strided_a) {
+  TEST(QS8_GEMM_MINMAX_4X8C8__NEON_MULL_PADAL, k_eq_16_strided_a) {
     TEST_REQUIRES_ARM_NEON;
     GemmMicrokernelTester()
       .mr(4)
@@ -16019,12 +16019,12 @@
       .sr(1)
       .m(4)
       .n(8)
-      .k(8)
-      .a_stride(11)
+      .k(16)
+      .a_stride(19)
       .Test(xnn_qs8_gemm_minmax_ukernel_4x8c8__neon_mull_padal);
   }
 
-  TEST(QS8_GEMM_MINMAX_4X8C8__NEON_MULL_PADAL, k_eq_8_subtile) {
+  TEST(QS8_GEMM_MINMAX_4X8C8__NEON_MULL_PADAL, k_eq_16_subtile) {
     TEST_REQUIRES_ARM_NEON;
     for (uint32_t m = 1; m <= 4; m++) {
       for (uint32_t n = 1; n <= 8; n++) {
@@ -16035,14 +16035,14 @@
           .sr(1)
           .m(m)
           .n(n)
-          .k(8)
+          .k(16)
           .iterations(1)
           .Test(xnn_qs8_gemm_minmax_ukernel_4x8c8__neon_mull_padal);
       }
     }
   }
 
-  TEST(QS8_GEMM_MINMAX_4X8C8__NEON_MULL_PADAL, k_eq_8_subtile_m) {
+  TEST(QS8_GEMM_MINMAX_4X8C8__NEON_MULL_PADAL, k_eq_16_subtile_m) {
     TEST_REQUIRES_ARM_NEON;
     for (uint32_t m = 1; m <= 4; m++) {
       GemmMicrokernelTester()
@@ -16052,13 +16052,13 @@
         .sr(1)
         .m(m)
         .n(8)
-        .k(8)
+        .k(16)
         .iterations(1)
         .Test(xnn_qs8_gemm_minmax_ukernel_4x8c8__neon_mull_padal);
     }
   }
 
-  TEST(QS8_GEMM_MINMAX_4X8C8__NEON_MULL_PADAL, k_eq_8_subtile_n) {
+  TEST(QS8_GEMM_MINMAX_4X8C8__NEON_MULL_PADAL, k_eq_16_subtile_n) {
     TEST_REQUIRES_ARM_NEON;
     for (uint32_t n = 1; n <= 8; n++) {
       GemmMicrokernelTester()
@@ -16068,15 +16068,15 @@
         .sr(1)
         .m(4)
         .n(n)
-        .k(8)
+        .k(16)
         .iterations(1)
         .Test(xnn_qs8_gemm_minmax_ukernel_4x8c8__neon_mull_padal);
     }
   }
 
-  TEST(QS8_GEMM_MINMAX_4X8C8__NEON_MULL_PADAL, k_lt_8) {
+  TEST(QS8_GEMM_MINMAX_4X8C8__NEON_MULL_PADAL, k_lt_16) {
     TEST_REQUIRES_ARM_NEON;
-    for (size_t k = 1; k < 8; k++) {
+    for (size_t k = 1; k < 16; k++) {
       GemmMicrokernelTester()
         .mr(4)
         .nr(8)
@@ -16089,60 +16089,9 @@
     }
   }
 
-  TEST(QS8_GEMM_MINMAX_4X8C8__NEON_MULL_PADAL, k_lt_8_strided_a) {
+  TEST(QS8_GEMM_MINMAX_4X8C8__NEON_MULL_PADAL, k_lt_16_strided_a) {
     TEST_REQUIRES_ARM_NEON;
-    for (size_t k = 1; k < 8; k++) {
-      GemmMicrokernelTester()
-        .mr(4)
-        .nr(8)
-        .kr(8)
-        .sr(1)
-        .m(4)
-        .n(8)
-        .k(k)
-        .a_stride(11)
-        .Test(xnn_qs8_gemm_minmax_ukernel_4x8c8__neon_mull_padal);
-    }
-  }
-
-  TEST(QS8_GEMM_MINMAX_4X8C8__NEON_MULL_PADAL, k_lt_8_subtile) {
-    TEST_REQUIRES_ARM_NEON;
-    for (size_t k = 1; k < 8; k++) {
-      for (uint32_t m = 1; m <= 4; m++) {
-        for (uint32_t n = 1; n <= 8; n++) {
-          GemmMicrokernelTester()
-            .mr(4)
-            .nr(8)
-            .kr(8)
-            .sr(1)
-            .m(m)
-            .n(n)
-            .k(k)
-            .iterations(1)
-            .Test(xnn_qs8_gemm_minmax_ukernel_4x8c8__neon_mull_padal);
-        }
-      }
-    }
-  }
-
-  TEST(QS8_GEMM_MINMAX_4X8C8__NEON_MULL_PADAL, k_gt_8) {
-    TEST_REQUIRES_ARM_NEON;
-    for (size_t k = 9; k < 16; k++) {
-      GemmMicrokernelTester()
-        .mr(4)
-        .nr(8)
-        .kr(8)
-        .sr(1)
-        .m(4)
-        .n(8)
-        .k(k)
-        .Test(xnn_qs8_gemm_minmax_ukernel_4x8c8__neon_mull_padal);
-    }
-  }
-
-  TEST(QS8_GEMM_MINMAX_4X8C8__NEON_MULL_PADAL, k_gt_8_strided_a) {
-    TEST_REQUIRES_ARM_NEON;
-    for (size_t k = 9; k < 16; k++) {
+    for (size_t k = 1; k < 16; k++) {
       GemmMicrokernelTester()
         .mr(4)
         .nr(8)
@@ -16156,9 +16105,9 @@
     }
   }
 
-  TEST(QS8_GEMM_MINMAX_4X8C8__NEON_MULL_PADAL, k_gt_8_subtile) {
+  TEST(QS8_GEMM_MINMAX_4X8C8__NEON_MULL_PADAL, k_lt_16_subtile) {
     TEST_REQUIRES_ARM_NEON;
-    for (size_t k = 9; k < 16; k++) {
+    for (size_t k = 1; k < 16; k++) {
       for (uint32_t m = 1; m <= 4; m++) {
         for (uint32_t n = 1; n <= 8; n++) {
           GemmMicrokernelTester()
@@ -16176,9 +16125,9 @@
     }
   }
 
-  TEST(QS8_GEMM_MINMAX_4X8C8__NEON_MULL_PADAL, k_div_8) {
+  TEST(QS8_GEMM_MINMAX_4X8C8__NEON_MULL_PADAL, k_gt_16) {
     TEST_REQUIRES_ARM_NEON;
-    for (size_t k = 16; k <= 80; k += 8) {
+    for (size_t k = 17; k < 32; k++) {
       GemmMicrokernelTester()
         .mr(4)
         .nr(8)
@@ -16191,9 +16140,9 @@
     }
   }
 
-  TEST(QS8_GEMM_MINMAX_4X8C8__NEON_MULL_PADAL, k_div_8_strided_a) {
+  TEST(QS8_GEMM_MINMAX_4X8C8__NEON_MULL_PADAL, k_gt_16_strided_a) {
     TEST_REQUIRES_ARM_NEON;
-    for (size_t k = 16; k <= 80; k += 8) {
+    for (size_t k = 17; k < 32; k++) {
       GemmMicrokernelTester()
         .mr(4)
         .nr(8)
@@ -16202,14 +16151,65 @@
         .m(4)
         .n(8)
         .k(k)
-        .a_stride(83)
+        .a_stride(37)
         .Test(xnn_qs8_gemm_minmax_ukernel_4x8c8__neon_mull_padal);
     }
   }
 
-  TEST(QS8_GEMM_MINMAX_4X8C8__NEON_MULL_PADAL, k_div_8_subtile) {
+  TEST(QS8_GEMM_MINMAX_4X8C8__NEON_MULL_PADAL, k_gt_16_subtile) {
     TEST_REQUIRES_ARM_NEON;
-    for (size_t k = 16; k <= 80; k += 8) {
+    for (size_t k = 17; k < 32; k++) {
+      for (uint32_t m = 1; m <= 4; m++) {
+        for (uint32_t n = 1; n <= 8; n++) {
+          GemmMicrokernelTester()
+            .mr(4)
+            .nr(8)
+            .kr(8)
+            .sr(1)
+            .m(m)
+            .n(n)
+            .k(k)
+            .iterations(1)
+            .Test(xnn_qs8_gemm_minmax_ukernel_4x8c8__neon_mull_padal);
+        }
+      }
+    }
+  }
+
+  TEST(QS8_GEMM_MINMAX_4X8C8__NEON_MULL_PADAL, k_div_16) {
+    TEST_REQUIRES_ARM_NEON;
+    for (size_t k = 32; k <= 160; k += 16) {
+      GemmMicrokernelTester()
+        .mr(4)
+        .nr(8)
+        .kr(8)
+        .sr(1)
+        .m(4)
+        .n(8)
+        .k(k)
+        .Test(xnn_qs8_gemm_minmax_ukernel_4x8c8__neon_mull_padal);
+    }
+  }
+
+  TEST(QS8_GEMM_MINMAX_4X8C8__NEON_MULL_PADAL, k_div_16_strided_a) {
+    TEST_REQUIRES_ARM_NEON;
+    for (size_t k = 32; k <= 160; k += 16) {
+      GemmMicrokernelTester()
+        .mr(4)
+        .nr(8)
+        .kr(8)
+        .sr(1)
+        .m(4)
+        .n(8)
+        .k(k)
+        .a_stride(163)
+        .Test(xnn_qs8_gemm_minmax_ukernel_4x8c8__neon_mull_padal);
+    }
+  }
+
+  TEST(QS8_GEMM_MINMAX_4X8C8__NEON_MULL_PADAL, k_div_16_subtile) {
+    TEST_REQUIRES_ARM_NEON;
+    for (size_t k = 32; k <= 160; k += 16) {
       for (uint32_t m = 1; m <= 4; m++) {
         for (uint32_t n = 1; n <= 8; n++) {
           GemmMicrokernelTester()
@@ -16230,7 +16230,7 @@
   TEST(QS8_GEMM_MINMAX_4X8C8__NEON_MULL_PADAL, n_gt_8) {
     TEST_REQUIRES_ARM_NEON;
     for (uint32_t n = 9; n < 16; n++) {
-      for (size_t k = 1; k <= 40; k += 9) {
+      for (size_t k = 1; k <= 80; k += 17) {
         GemmMicrokernelTester()
           .mr(4)
           .nr(8)
@@ -16247,7 +16247,7 @@
   TEST(QS8_GEMM_MINMAX_4X8C8__NEON_MULL_PADAL, n_gt_8_strided_cn) {
     TEST_REQUIRES_ARM_NEON;
     for (uint32_t n = 9; n < 16; n++) {
-      for (size_t k = 1; k <= 40; k += 9) {
+      for (size_t k = 1; k <= 80; k += 17) {
         GemmMicrokernelTester()
           .mr(4)
           .nr(8)
@@ -16265,7 +16265,7 @@
   TEST(QS8_GEMM_MINMAX_4X8C8__NEON_MULL_PADAL, n_gt_8_strided_a) {
     TEST_REQUIRES_ARM_NEON;
     for (uint32_t n = 9; n < 16; n++) {
-      for (size_t k = 1; k <= 40; k += 9) {
+      for (size_t k = 1; k <= 80; k += 17) {
         GemmMicrokernelTester()
           .mr(4)
           .nr(8)
@@ -16274,7 +16274,7 @@
           .m(4)
           .n(n)
           .k(k)
-          .a_stride(43)
+          .a_stride(83)
           .Test(xnn_qs8_gemm_minmax_ukernel_4x8c8__neon_mull_padal);
       }
     }
@@ -16283,7 +16283,7 @@
   TEST(QS8_GEMM_MINMAX_4X8C8__NEON_MULL_PADAL, n_gt_8_subtile) {
     TEST_REQUIRES_ARM_NEON;
     for (uint32_t n = 9; n < 16; n++) {
-      for (size_t k = 1; k <= 40; k += 9) {
+      for (size_t k = 1; k <= 80; k += 17) {
         for (uint32_t m = 1; m <= 4; m++) {
           GemmMicrokernelTester()
             .mr(4)
@@ -16303,7 +16303,7 @@
   TEST(QS8_GEMM_MINMAX_4X8C8__NEON_MULL_PADAL, n_div_8) {
     TEST_REQUIRES_ARM_NEON;
     for (uint32_t n = 16; n <= 24; n += 8) {
-      for (size_t k = 1; k <= 40; k += 9) {
+      for (size_t k = 1; k <= 80; k += 17) {
         GemmMicrokernelTester()
           .mr(4)
           .nr(8)
@@ -16320,7 +16320,7 @@
   TEST(QS8_GEMM_MINMAX_4X8C8__NEON_MULL_PADAL, n_div_8_strided_cn) {
     TEST_REQUIRES_ARM_NEON;
     for (uint32_t n = 16; n <= 24; n += 8) {
-      for (size_t k = 1; k <= 40; k += 9) {
+      for (size_t k = 1; k <= 80; k += 17) {
         GemmMicrokernelTester()
           .mr(4)
           .nr(8)
@@ -16338,7 +16338,7 @@
   TEST(QS8_GEMM_MINMAX_4X8C8__NEON_MULL_PADAL, n_div_8_strided_a) {
     TEST_REQUIRES_ARM_NEON;
     for (uint32_t n = 16; n <= 24; n += 8) {
-      for (size_t k = 1; k <= 40; k += 9) {
+      for (size_t k = 1; k <= 80; k += 17) {
         GemmMicrokernelTester()
           .mr(4)
           .nr(8)
@@ -16347,7 +16347,7 @@
           .m(4)
           .n(n)
           .k(k)
-          .a_stride(43)
+          .a_stride(83)
           .Test(xnn_qs8_gemm_minmax_ukernel_4x8c8__neon_mull_padal);
       }
     }
@@ -16356,7 +16356,7 @@
   TEST(QS8_GEMM_MINMAX_4X8C8__NEON_MULL_PADAL, n_div_8_subtile) {
     TEST_REQUIRES_ARM_NEON;
     for (uint32_t n = 16; n <= 24; n += 8) {
-      for (size_t k = 1; k <= 40; k += 9) {
+      for (size_t k = 1; k <= 80; k += 17) {
         for (uint32_t m = 1; m <= 4; m++) {
           GemmMicrokernelTester()
             .mr(4)
@@ -16375,7 +16375,7 @@
 
   TEST(QS8_GEMM_MINMAX_4X8C8__NEON_MULL_PADAL, strided_cm_subtile) {
     TEST_REQUIRES_ARM_NEON;
-    for (size_t k = 1; k <= 40; k += 9) {
+    for (size_t k = 1; k <= 80; k += 17) {
       for (uint32_t m = 1; m <= 4; m++) {
         for (uint32_t n = 1; n <= 8; n++) {
           GemmMicrokernelTester()
@@ -16403,7 +16403,7 @@
       .sr(1)
       .m(4)
       .n(8)
-      .k(8)
+      .k(16)
       .qmin(128)
       .Test(xnn_qs8_gemm_minmax_ukernel_4x8c8__neon_mull_padal);
   }
@@ -16417,7 +16417,7 @@
       .sr(1)
       .m(4)
       .n(8)
-      .k(8)
+      .k(16)
       .qmax(128)
       .Test(xnn_qs8_gemm_minmax_ukernel_4x8c8__neon_mull_padal);
   }
@@ -16431,7 +16431,7 @@
       .sr(1)
       .m(4)
       .n(8)
-      .k(8)
+      .k(16)
       .cm_stride(11)
       .Test(xnn_qs8_gemm_minmax_ukernel_4x8c8__neon_mull_padal);
   }
@@ -16439,7 +16439,7 @@
 
 
 #if XNN_ARCH_ARM || XNN_ARCH_ARM64
-  TEST(QS8_GEMM_MINMAX_1X16C8__NEON_MULL_PADAL, k_eq_8) {
+  TEST(QS8_GEMM_MINMAX_1X16C8__NEON_MULL_PADAL, k_eq_16) {
     TEST_REQUIRES_ARM_NEON;
     GemmMicrokernelTester()
       .mr(1)
@@ -16448,7 +16448,7 @@
       .sr(1)
       .m(1)
       .n(16)
-      .k(8)
+      .k(16)
       .Test(xnn_qs8_gemm_minmax_ukernel_1x16c8__neon_mull_padal);
   }
 
@@ -16461,12 +16461,12 @@
       .sr(1)
       .m(1)
       .n(16)
-      .k(8)
+      .k(16)
       .cn_stride(19)
       .Test(xnn_qs8_gemm_minmax_ukernel_1x16c8__neon_mull_padal);
   }
 
-  TEST(QS8_GEMM_MINMAX_1X16C8__NEON_MULL_PADAL, k_eq_8_strided_a) {
+  TEST(QS8_GEMM_MINMAX_1X16C8__NEON_MULL_PADAL, k_eq_16_strided_a) {
     TEST_REQUIRES_ARM_NEON;
     GemmMicrokernelTester()
       .mr(1)
@@ -16475,12 +16475,12 @@
       .sr(1)
       .m(1)
       .n(16)
-      .k(8)
-      .a_stride(11)
+      .k(16)
+      .a_stride(19)
       .Test(xnn_qs8_gemm_minmax_ukernel_1x16c8__neon_mull_padal);
   }
 
-  TEST(QS8_GEMM_MINMAX_1X16C8__NEON_MULL_PADAL, k_eq_8_subtile) {
+  TEST(QS8_GEMM_MINMAX_1X16C8__NEON_MULL_PADAL, k_eq_16_subtile) {
     TEST_REQUIRES_ARM_NEON;
     for (uint32_t m = 1; m <= 1; m++) {
       for (uint32_t n = 1; n <= 16; n++) {
@@ -16491,14 +16491,14 @@
           .sr(1)
           .m(m)
           .n(n)
-          .k(8)
+          .k(16)
           .iterations(1)
           .Test(xnn_qs8_gemm_minmax_ukernel_1x16c8__neon_mull_padal);
       }
     }
   }
 
-  TEST(QS8_GEMM_MINMAX_1X16C8__NEON_MULL_PADAL, k_eq_8_subtile_m) {
+  TEST(QS8_GEMM_MINMAX_1X16C8__NEON_MULL_PADAL, k_eq_16_subtile_m) {
     TEST_REQUIRES_ARM_NEON;
     for (uint32_t m = 1; m <= 1; m++) {
       GemmMicrokernelTester()
@@ -16508,13 +16508,13 @@
         .sr(1)
         .m(m)
         .n(16)
-        .k(8)
+        .k(16)
         .iterations(1)
         .Test(xnn_qs8_gemm_minmax_ukernel_1x16c8__neon_mull_padal);
     }
   }
 
-  TEST(QS8_GEMM_MINMAX_1X16C8__NEON_MULL_PADAL, k_eq_8_subtile_n) {
+  TEST(QS8_GEMM_MINMAX_1X16C8__NEON_MULL_PADAL, k_eq_16_subtile_n) {
     TEST_REQUIRES_ARM_NEON;
     for (uint32_t n = 1; n <= 16; n++) {
       GemmMicrokernelTester()
@@ -16524,15 +16524,15 @@
         .sr(1)
         .m(1)
         .n(n)
-        .k(8)
+        .k(16)
         .iterations(1)
         .Test(xnn_qs8_gemm_minmax_ukernel_1x16c8__neon_mull_padal);
     }
   }
 
-  TEST(QS8_GEMM_MINMAX_1X16C8__NEON_MULL_PADAL, k_lt_8) {
+  TEST(QS8_GEMM_MINMAX_1X16C8__NEON_MULL_PADAL, k_lt_16) {
     TEST_REQUIRES_ARM_NEON;
-    for (size_t k = 1; k < 8; k++) {
+    for (size_t k = 1; k < 16; k++) {
       GemmMicrokernelTester()
         .mr(1)
         .nr(16)
@@ -16545,60 +16545,9 @@
     }
   }
 
-  TEST(QS8_GEMM_MINMAX_1X16C8__NEON_MULL_PADAL, k_lt_8_strided_a) {
+  TEST(QS8_GEMM_MINMAX_1X16C8__NEON_MULL_PADAL, k_lt_16_strided_a) {
     TEST_REQUIRES_ARM_NEON;
-    for (size_t k = 1; k < 8; k++) {
-      GemmMicrokernelTester()
-        .mr(1)
-        .nr(16)
-        .kr(8)
-        .sr(1)
-        .m(1)
-        .n(16)
-        .k(k)
-        .a_stride(11)
-        .Test(xnn_qs8_gemm_minmax_ukernel_1x16c8__neon_mull_padal);
-    }
-  }
-
-  TEST(QS8_GEMM_MINMAX_1X16C8__NEON_MULL_PADAL, k_lt_8_subtile) {
-    TEST_REQUIRES_ARM_NEON;
-    for (size_t k = 1; k < 8; k++) {
-      for (uint32_t m = 1; m <= 1; m++) {
-        for (uint32_t n = 1; n <= 16; n++) {
-          GemmMicrokernelTester()
-            .mr(1)
-            .nr(16)
-            .kr(8)
-            .sr(1)
-            .m(m)
-            .n(n)
-            .k(k)
-            .iterations(1)
-            .Test(xnn_qs8_gemm_minmax_ukernel_1x16c8__neon_mull_padal);
-        }
-      }
-    }
-  }
-
-  TEST(QS8_GEMM_MINMAX_1X16C8__NEON_MULL_PADAL, k_gt_8) {
-    TEST_REQUIRES_ARM_NEON;
-    for (size_t k = 9; k < 16; k++) {
-      GemmMicrokernelTester()
-        .mr(1)
-        .nr(16)
-        .kr(8)
-        .sr(1)
-        .m(1)
-        .n(16)
-        .k(k)
-        .Test(xnn_qs8_gemm_minmax_ukernel_1x16c8__neon_mull_padal);
-    }
-  }
-
-  TEST(QS8_GEMM_MINMAX_1X16C8__NEON_MULL_PADAL, k_gt_8_strided_a) {
-    TEST_REQUIRES_ARM_NEON;
-    for (size_t k = 9; k < 16; k++) {
+    for (size_t k = 1; k < 16; k++) {
       GemmMicrokernelTester()
         .mr(1)
         .nr(16)
@@ -16612,9 +16561,9 @@
     }
   }
 
-  TEST(QS8_GEMM_MINMAX_1X16C8__NEON_MULL_PADAL, k_gt_8_subtile) {
+  TEST(QS8_GEMM_MINMAX_1X16C8__NEON_MULL_PADAL, k_lt_16_subtile) {
     TEST_REQUIRES_ARM_NEON;
-    for (size_t k = 9; k < 16; k++) {
+    for (size_t k = 1; k < 16; k++) {
       for (uint32_t m = 1; m <= 1; m++) {
         for (uint32_t n = 1; n <= 16; n++) {
           GemmMicrokernelTester()
@@ -16632,9 +16581,9 @@
     }
   }
 
-  TEST(QS8_GEMM_MINMAX_1X16C8__NEON_MULL_PADAL, k_div_8) {
+  TEST(QS8_GEMM_MINMAX_1X16C8__NEON_MULL_PADAL, k_gt_16) {
     TEST_REQUIRES_ARM_NEON;
-    for (size_t k = 16; k <= 80; k += 8) {
+    for (size_t k = 17; k < 32; k++) {
       GemmMicrokernelTester()
         .mr(1)
         .nr(16)
@@ -16647,9 +16596,9 @@
     }
   }
 
-  TEST(QS8_GEMM_MINMAX_1X16C8__NEON_MULL_PADAL, k_div_8_strided_a) {
+  TEST(QS8_GEMM_MINMAX_1X16C8__NEON_MULL_PADAL, k_gt_16_strided_a) {
     TEST_REQUIRES_ARM_NEON;
-    for (size_t k = 16; k <= 80; k += 8) {
+    for (size_t k = 17; k < 32; k++) {
       GemmMicrokernelTester()
         .mr(1)
         .nr(16)
@@ -16658,14 +16607,65 @@
         .m(1)
         .n(16)
         .k(k)
-        .a_stride(83)
+        .a_stride(37)
         .Test(xnn_qs8_gemm_minmax_ukernel_1x16c8__neon_mull_padal);
     }
   }
 
-  TEST(QS8_GEMM_MINMAX_1X16C8__NEON_MULL_PADAL, k_div_8_subtile) {
+  TEST(QS8_GEMM_MINMAX_1X16C8__NEON_MULL_PADAL, k_gt_16_subtile) {
     TEST_REQUIRES_ARM_NEON;
-    for (size_t k = 16; k <= 80; k += 8) {
+    for (size_t k = 17; k < 32; k++) {
+      for (uint32_t m = 1; m <= 1; m++) {
+        for (uint32_t n = 1; n <= 16; n++) {
+          GemmMicrokernelTester()
+            .mr(1)
+            .nr(16)
+            .kr(8)
+            .sr(1)
+            .m(m)
+            .n(n)
+            .k(k)
+            .iterations(1)
+            .Test(xnn_qs8_gemm_minmax_ukernel_1x16c8__neon_mull_padal);
+        }
+      }
+    }
+  }
+
+  TEST(QS8_GEMM_MINMAX_1X16C8__NEON_MULL_PADAL, k_div_16) {
+    TEST_REQUIRES_ARM_NEON;
+    for (size_t k = 32; k <= 160; k += 16) {
+      GemmMicrokernelTester()
+        .mr(1)
+        .nr(16)
+        .kr(8)
+        .sr(1)
+        .m(1)
+        .n(16)
+        .k(k)
+        .Test(xnn_qs8_gemm_minmax_ukernel_1x16c8__neon_mull_padal);
+    }
+  }
+
+  TEST(QS8_GEMM_MINMAX_1X16C8__NEON_MULL_PADAL, k_div_16_strided_a) {
+    TEST_REQUIRES_ARM_NEON;
+    for (size_t k = 32; k <= 160; k += 16) {
+      GemmMicrokernelTester()
+        .mr(1)
+        .nr(16)
+        .kr(8)
+        .sr(1)
+        .m(1)
+        .n(16)
+        .k(k)
+        .a_stride(163)
+        .Test(xnn_qs8_gemm_minmax_ukernel_1x16c8__neon_mull_padal);
+    }
+  }
+
+  TEST(QS8_GEMM_MINMAX_1X16C8__NEON_MULL_PADAL, k_div_16_subtile) {
+    TEST_REQUIRES_ARM_NEON;
+    for (size_t k = 32; k <= 160; k += 16) {
       for (uint32_t m = 1; m <= 1; m++) {
         for (uint32_t n = 1; n <= 16; n++) {
           GemmMicrokernelTester()
@@ -16686,7 +16686,7 @@
   TEST(QS8_GEMM_MINMAX_1X16C8__NEON_MULL_PADAL, n_gt_16) {
     TEST_REQUIRES_ARM_NEON;
     for (uint32_t n = 17; n < 32; n++) {
-      for (size_t k = 1; k <= 40; k += 9) {
+      for (size_t k = 1; k <= 80; k += 17) {
         GemmMicrokernelTester()
           .mr(1)
           .nr(16)
@@ -16703,7 +16703,7 @@
   TEST(QS8_GEMM_MINMAX_1X16C8__NEON_MULL_PADAL, n_gt_16_strided_cn) {
     TEST_REQUIRES_ARM_NEON;
     for (uint32_t n = 17; n < 32; n++) {
-      for (size_t k = 1; k <= 40; k += 9) {
+      for (size_t k = 1; k <= 80; k += 17) {
         GemmMicrokernelTester()
           .mr(1)
           .nr(16)
@@ -16721,7 +16721,7 @@
   TEST(QS8_GEMM_MINMAX_1X16C8__NEON_MULL_PADAL, n_gt_16_strided_a) {
     TEST_REQUIRES_ARM_NEON;
     for (uint32_t n = 17; n < 32; n++) {
-      for (size_t k = 1; k <= 40; k += 9) {
+      for (size_t k = 1; k <= 80; k += 17) {
         GemmMicrokernelTester()
           .mr(1)
           .nr(16)
@@ -16730,7 +16730,7 @@
           .m(1)
           .n(n)
           .k(k)
-          .a_stride(43)
+          .a_stride(83)
           .Test(xnn_qs8_gemm_minmax_ukernel_1x16c8__neon_mull_padal);
       }
     }
@@ -16739,7 +16739,7 @@
   TEST(QS8_GEMM_MINMAX_1X16C8__NEON_MULL_PADAL, n_gt_16_subtile) {
     TEST_REQUIRES_ARM_NEON;
     for (uint32_t n = 17; n < 32; n++) {
-      for (size_t k = 1; k <= 40; k += 9) {
+      for (size_t k = 1; k <= 80; k += 17) {
         for (uint32_t m = 1; m <= 1; m++) {
           GemmMicrokernelTester()
             .mr(1)
@@ -16759,7 +16759,7 @@
   TEST(QS8_GEMM_MINMAX_1X16C8__NEON_MULL_PADAL, n_div_16) {
     TEST_REQUIRES_ARM_NEON;
     for (uint32_t n = 32; n <= 48; n += 16) {
-      for (size_t k = 1; k <= 40; k += 9) {
+      for (size_t k = 1; k <= 80; k += 17) {
         GemmMicrokernelTester()
           .mr(1)
           .nr(16)
@@ -16776,7 +16776,7 @@
   TEST(QS8_GEMM_MINMAX_1X16C8__NEON_MULL_PADAL, n_div_16_strided_cn) {
     TEST_REQUIRES_ARM_NEON;
     for (uint32_t n = 32; n <= 48; n += 16) {
-      for (size_t k = 1; k <= 40; k += 9) {
+      for (size_t k = 1; k <= 80; k += 17) {
         GemmMicrokernelTester()
           .mr(1)
           .nr(16)
@@ -16794,7 +16794,7 @@
   TEST(QS8_GEMM_MINMAX_1X16C8__NEON_MULL_PADAL, n_div_16_strided_a) {
     TEST_REQUIRES_ARM_NEON;
     for (uint32_t n = 32; n <= 48; n += 16) {
-      for (size_t k = 1; k <= 40; k += 9) {
+      for (size_t k = 1; k <= 80; k += 17) {
         GemmMicrokernelTester()
           .mr(1)
           .nr(16)
@@ -16803,7 +16803,7 @@
           .m(1)
           .n(n)
           .k(k)
-          .a_stride(43)
+          .a_stride(83)
           .Test(xnn_qs8_gemm_minmax_ukernel_1x16c8__neon_mull_padal);
       }
     }
@@ -16812,7 +16812,7 @@
   TEST(QS8_GEMM_MINMAX_1X16C8__NEON_MULL_PADAL, n_div_16_subtile) {
     TEST_REQUIRES_ARM_NEON;
     for (uint32_t n = 32; n <= 48; n += 16) {
-      for (size_t k = 1; k <= 40; k += 9) {
+      for (size_t k = 1; k <= 80; k += 17) {
         for (uint32_t m = 1; m <= 1; m++) {
           GemmMicrokernelTester()
             .mr(1)
@@ -16831,7 +16831,7 @@
 
   TEST(QS8_GEMM_MINMAX_1X16C8__NEON_MULL_PADAL, strided_cm_subtile) {
     TEST_REQUIRES_ARM_NEON;
-    for (size_t k = 1; k <= 40; k += 9) {
+    for (size_t k = 1; k <= 80; k += 17) {
       for (uint32_t m = 1; m <= 1; m++) {
         for (uint32_t n = 1; n <= 16; n++) {
           GemmMicrokernelTester()
@@ -16859,7 +16859,7 @@
       .sr(1)
       .m(1)
       .n(16)
-      .k(8)
+      .k(16)
       .qmin(128)
       .Test(xnn_qs8_gemm_minmax_ukernel_1x16c8__neon_mull_padal);
   }
@@ -16873,7 +16873,7 @@
       .sr(1)
       .m(1)
       .n(16)
-      .k(8)
+      .k(16)
       .qmax(128)
       .Test(xnn_qs8_gemm_minmax_ukernel_1x16c8__neon_mull_padal);
   }
@@ -16887,7 +16887,7 @@
       .sr(1)
       .m(1)
       .n(16)
-      .k(8)
+      .k(16)
       .cm_stride(19)
       .Test(xnn_qs8_gemm_minmax_ukernel_1x16c8__neon_mull_padal);
   }
@@ -16895,7 +16895,7 @@
 
 
 #if XNN_ARCH_ARM || XNN_ARCH_ARM64
-  TEST(QS8_GEMM_MINMAX_2X16C8__NEON_MULL_PADAL, k_eq_8) {
+  TEST(QS8_GEMM_MINMAX_2X16C8__NEON_MULL_PADAL, k_eq_16) {
     TEST_REQUIRES_ARM_NEON;
     GemmMicrokernelTester()
       .mr(2)
@@ -16904,7 +16904,7 @@
       .sr(1)
       .m(2)
       .n(16)
-      .k(8)
+      .k(16)
       .Test(xnn_qs8_gemm_minmax_ukernel_2x16c8__neon_mull_padal);
   }
 
@@ -16917,12 +16917,12 @@
       .sr(1)
       .m(2)
       .n(16)
-      .k(8)
+      .k(16)
       .cn_stride(19)
       .Test(xnn_qs8_gemm_minmax_ukernel_2x16c8__neon_mull_padal);
   }
 
-  TEST(QS8_GEMM_MINMAX_2X16C8__NEON_MULL_PADAL, k_eq_8_strided_a) {
+  TEST(QS8_GEMM_MINMAX_2X16C8__NEON_MULL_PADAL, k_eq_16_strided_a) {
     TEST_REQUIRES_ARM_NEON;
     GemmMicrokernelTester()
       .mr(2)
@@ -16931,12 +16931,12 @@
       .sr(1)
       .m(2)
       .n(16)
-      .k(8)
-      .a_stride(11)
+      .k(16)
+      .a_stride(19)
       .Test(xnn_qs8_gemm_minmax_ukernel_2x16c8__neon_mull_padal);
   }
 
-  TEST(QS8_GEMM_MINMAX_2X16C8__NEON_MULL_PADAL, k_eq_8_subtile) {
+  TEST(QS8_GEMM_MINMAX_2X16C8__NEON_MULL_PADAL, k_eq_16_subtile) {
     TEST_REQUIRES_ARM_NEON;
     for (uint32_t m = 1; m <= 2; m++) {
       for (uint32_t n = 1; n <= 16; n++) {
@@ -16947,14 +16947,14 @@
           .sr(1)
           .m(m)
           .n(n)
-          .k(8)
+          .k(16)
           .iterations(1)
           .Test(xnn_qs8_gemm_minmax_ukernel_2x16c8__neon_mull_padal);
       }
     }
   }
 
-  TEST(QS8_GEMM_MINMAX_2X16C8__NEON_MULL_PADAL, k_eq_8_subtile_m) {
+  TEST(QS8_GEMM_MINMAX_2X16C8__NEON_MULL_PADAL, k_eq_16_subtile_m) {
     TEST_REQUIRES_ARM_NEON;
     for (uint32_t m = 1; m <= 2; m++) {
       GemmMicrokernelTester()
@@ -16964,13 +16964,13 @@
         .sr(1)
         .m(m)
         .n(16)
-        .k(8)
+        .k(16)
         .iterations(1)
         .Test(xnn_qs8_gemm_minmax_ukernel_2x16c8__neon_mull_padal);
     }
   }
 
-  TEST(QS8_GEMM_MINMAX_2X16C8__NEON_MULL_PADAL, k_eq_8_subtile_n) {
+  TEST(QS8_GEMM_MINMAX_2X16C8__NEON_MULL_PADAL, k_eq_16_subtile_n) {
     TEST_REQUIRES_ARM_NEON;
     for (uint32_t n = 1; n <= 16; n++) {
       GemmMicrokernelTester()
@@ -16980,15 +16980,15 @@
         .sr(1)
         .m(2)
         .n(n)
-        .k(8)
+        .k(16)
         .iterations(1)
         .Test(xnn_qs8_gemm_minmax_ukernel_2x16c8__neon_mull_padal);
     }
   }
 
-  TEST(QS8_GEMM_MINMAX_2X16C8__NEON_MULL_PADAL, k_lt_8) {
+  TEST(QS8_GEMM_MINMAX_2X16C8__NEON_MULL_PADAL, k_lt_16) {
     TEST_REQUIRES_ARM_NEON;
-    for (size_t k = 1; k < 8; k++) {
+    for (size_t k = 1; k < 16; k++) {
       GemmMicrokernelTester()
         .mr(2)
         .nr(16)
@@ -17001,60 +17001,9 @@
     }
   }
 
-  TEST(QS8_GEMM_MINMAX_2X16C8__NEON_MULL_PADAL, k_lt_8_strided_a) {
+  TEST(QS8_GEMM_MINMAX_2X16C8__NEON_MULL_PADAL, k_lt_16_strided_a) {
     TEST_REQUIRES_ARM_NEON;
-    for (size_t k = 1; k < 8; k++) {
-      GemmMicrokernelTester()
-        .mr(2)
-        .nr(16)
-        .kr(8)
-        .sr(1)
-        .m(2)
-        .n(16)
-        .k(k)
-        .a_stride(11)
-        .Test(xnn_qs8_gemm_minmax_ukernel_2x16c8__neon_mull_padal);
-    }
-  }
-
-  TEST(QS8_GEMM_MINMAX_2X16C8__NEON_MULL_PADAL, k_lt_8_subtile) {
-    TEST_REQUIRES_ARM_NEON;
-    for (size_t k = 1; k < 8; k++) {
-      for (uint32_t m = 1; m <= 2; m++) {
-        for (uint32_t n = 1; n <= 16; n++) {
-          GemmMicrokernelTester()
-            .mr(2)
-            .nr(16)
-            .kr(8)
-            .sr(1)
-            .m(m)
-            .n(n)
-            .k(k)
-            .iterations(1)
-            .Test(xnn_qs8_gemm_minmax_ukernel_2x16c8__neon_mull_padal);
-        }
-      }
-    }
-  }
-
-  TEST(QS8_GEMM_MINMAX_2X16C8__NEON_MULL_PADAL, k_gt_8) {
-    TEST_REQUIRES_ARM_NEON;
-    for (size_t k = 9; k < 16; k++) {
-      GemmMicrokernelTester()
-        .mr(2)
-        .nr(16)
-        .kr(8)
-        .sr(1)
-        .m(2)
-        .n(16)
-        .k(k)
-        .Test(xnn_qs8_gemm_minmax_ukernel_2x16c8__neon_mull_padal);
-    }
-  }
-
-  TEST(QS8_GEMM_MINMAX_2X16C8__NEON_MULL_PADAL, k_gt_8_strided_a) {
-    TEST_REQUIRES_ARM_NEON;
-    for (size_t k = 9; k < 16; k++) {
+    for (size_t k = 1; k < 16; k++) {
       GemmMicrokernelTester()
         .mr(2)
         .nr(16)
@@ -17068,9 +17017,9 @@
     }
   }
 
-  TEST(QS8_GEMM_MINMAX_2X16C8__NEON_MULL_PADAL, k_gt_8_subtile) {
+  TEST(QS8_GEMM_MINMAX_2X16C8__NEON_MULL_PADAL, k_lt_16_subtile) {
     TEST_REQUIRES_ARM_NEON;
-    for (size_t k = 9; k < 16; k++) {
+    for (size_t k = 1; k < 16; k++) {
       for (uint32_t m = 1; m <= 2; m++) {
         for (uint32_t n = 1; n <= 16; n++) {
           GemmMicrokernelTester()
@@ -17088,9 +17037,9 @@
     }
   }
 
-  TEST(QS8_GEMM_MINMAX_2X16C8__NEON_MULL_PADAL, k_div_8) {
+  TEST(QS8_GEMM_MINMAX_2X16C8__NEON_MULL_PADAL, k_gt_16) {
     TEST_REQUIRES_ARM_NEON;
-    for (size_t k = 16; k <= 80; k += 8) {
+    for (size_t k = 17; k < 32; k++) {
       GemmMicrokernelTester()
         .mr(2)
         .nr(16)
@@ -17103,9 +17052,9 @@
     }
   }
 
-  TEST(QS8_GEMM_MINMAX_2X16C8__NEON_MULL_PADAL, k_div_8_strided_a) {
+  TEST(QS8_GEMM_MINMAX_2X16C8__NEON_MULL_PADAL, k_gt_16_strided_a) {
     TEST_REQUIRES_ARM_NEON;
-    for (size_t k = 16; k <= 80; k += 8) {
+    for (size_t k = 17; k < 32; k++) {
       GemmMicrokernelTester()
         .mr(2)
         .nr(16)
@@ -17114,14 +17063,65 @@
         .m(2)
         .n(16)
         .k(k)
-        .a_stride(83)
+        .a_stride(37)
         .Test(xnn_qs8_gemm_minmax_ukernel_2x16c8__neon_mull_padal);
     }
   }
 
-  TEST(QS8_GEMM_MINMAX_2X16C8__NEON_MULL_PADAL, k_div_8_subtile) {
+  TEST(QS8_GEMM_MINMAX_2X16C8__NEON_MULL_PADAL, k_gt_16_subtile) {
     TEST_REQUIRES_ARM_NEON;
-    for (size_t k = 16; k <= 80; k += 8) {
+    for (size_t k = 17; k < 32; k++) {
+      for (uint32_t m = 1; m <= 2; m++) {
+        for (uint32_t n = 1; n <= 16; n++) {
+          GemmMicrokernelTester()
+            .mr(2)
+            .nr(16)
+            .kr(8)
+            .sr(1)
+            .m(m)
+            .n(n)
+            .k(k)
+            .iterations(1)
+            .Test(xnn_qs8_gemm_minmax_ukernel_2x16c8__neon_mull_padal);
+        }
+      }
+    }
+  }
+
+  TEST(QS8_GEMM_MINMAX_2X16C8__NEON_MULL_PADAL, k_div_16) {
+    TEST_REQUIRES_ARM_NEON;
+    for (size_t k = 32; k <= 160; k += 16) {
+      GemmMicrokernelTester()
+        .mr(2)
+        .nr(16)
+        .kr(8)
+        .sr(1)
+        .m(2)
+        .n(16)
+        .k(k)
+        .Test(xnn_qs8_gemm_minmax_ukernel_2x16c8__neon_mull_padal);
+    }
+  }
+
+  TEST(QS8_GEMM_MINMAX_2X16C8__NEON_MULL_PADAL, k_div_16_strided_a) {
+    TEST_REQUIRES_ARM_NEON;
+    for (size_t k = 32; k <= 160; k += 16) {
+      GemmMicrokernelTester()
+        .mr(2)
+        .nr(16)
+        .kr(8)
+        .sr(1)
+        .m(2)
+        .n(16)
+        .k(k)
+        .a_stride(163)
+        .Test(xnn_qs8_gemm_minmax_ukernel_2x16c8__neon_mull_padal);
+    }
+  }
+
+  TEST(QS8_GEMM_MINMAX_2X16C8__NEON_MULL_PADAL, k_div_16_subtile) {
+    TEST_REQUIRES_ARM_NEON;
+    for (size_t k = 32; k <= 160; k += 16) {
       for (uint32_t m = 1; m <= 2; m++) {
         for (uint32_t n = 1; n <= 16; n++) {
           GemmMicrokernelTester()
@@ -17142,7 +17142,7 @@
   TEST(QS8_GEMM_MINMAX_2X16C8__NEON_MULL_PADAL, n_gt_16) {
     TEST_REQUIRES_ARM_NEON;
     for (uint32_t n = 17; n < 32; n++) {
-      for (size_t k = 1; k <= 40; k += 9) {
+      for (size_t k = 1; k <= 80; k += 17) {
         GemmMicrokernelTester()
           .mr(2)
           .nr(16)
@@ -17159,7 +17159,7 @@
   TEST(QS8_GEMM_MINMAX_2X16C8__NEON_MULL_PADAL, n_gt_16_strided_cn) {
     TEST_REQUIRES_ARM_NEON;
     for (uint32_t n = 17; n < 32; n++) {
-      for (size_t k = 1; k <= 40; k += 9) {
+      for (size_t k = 1; k <= 80; k += 17) {
         GemmMicrokernelTester()
           .mr(2)
           .nr(16)
@@ -17177,7 +17177,7 @@
   TEST(QS8_GEMM_MINMAX_2X16C8__NEON_MULL_PADAL, n_gt_16_strided_a) {
     TEST_REQUIRES_ARM_NEON;
     for (uint32_t n = 17; n < 32; n++) {
-      for (size_t k = 1; k <= 40; k += 9) {
+      for (size_t k = 1; k <= 80; k += 17) {
         GemmMicrokernelTester()
           .mr(2)
           .nr(16)
@@ -17186,7 +17186,7 @@
           .m(2)
           .n(n)
           .k(k)
-          .a_stride(43)
+          .a_stride(83)
           .Test(xnn_qs8_gemm_minmax_ukernel_2x16c8__neon_mull_padal);
       }
     }
@@ -17195,7 +17195,7 @@
   TEST(QS8_GEMM_MINMAX_2X16C8__NEON_MULL_PADAL, n_gt_16_subtile) {
     TEST_REQUIRES_ARM_NEON;
     for (uint32_t n = 17; n < 32; n++) {
-      for (size_t k = 1; k <= 40; k += 9) {
+      for (size_t k = 1; k <= 80; k += 17) {
         for (uint32_t m = 1; m <= 2; m++) {
           GemmMicrokernelTester()
             .mr(2)
@@ -17215,7 +17215,7 @@
   TEST(QS8_GEMM_MINMAX_2X16C8__NEON_MULL_PADAL, n_div_16) {
     TEST_REQUIRES_ARM_NEON;
     for (uint32_t n = 32; n <= 48; n += 16) {
-      for (size_t k = 1; k <= 40; k += 9) {
+      for (size_t k = 1; k <= 80; k += 17) {
         GemmMicrokernelTester()
           .mr(2)
           .nr(16)
@@ -17232,7 +17232,7 @@
   TEST(QS8_GEMM_MINMAX_2X16C8__NEON_MULL_PADAL, n_div_16_strided_cn) {
     TEST_REQUIRES_ARM_NEON;
     for (uint32_t n = 32; n <= 48; n += 16) {
-      for (size_t k = 1; k <= 40; k += 9) {
+      for (size_t k = 1; k <= 80; k += 17) {
         GemmMicrokernelTester()
           .mr(2)
           .nr(16)
@@ -17250,7 +17250,7 @@
   TEST(QS8_GEMM_MINMAX_2X16C8__NEON_MULL_PADAL, n_div_16_strided_a) {
     TEST_REQUIRES_ARM_NEON;
     for (uint32_t n = 32; n <= 48; n += 16) {
-      for (size_t k = 1; k <= 40; k += 9) {
+      for (size_t k = 1; k <= 80; k += 17) {
         GemmMicrokernelTester()
           .mr(2)
           .nr(16)
@@ -17259,7 +17259,7 @@
           .m(2)
           .n(n)
           .k(k)
-          .a_stride(43)
+          .a_stride(83)
           .Test(xnn_qs8_gemm_minmax_ukernel_2x16c8__neon_mull_padal);
       }
     }
@@ -17268,7 +17268,7 @@
   TEST(QS8_GEMM_MINMAX_2X16C8__NEON_MULL_PADAL, n_div_16_subtile) {
     TEST_REQUIRES_ARM_NEON;
     for (uint32_t n = 32; n <= 48; n += 16) {
-      for (size_t k = 1; k <= 40; k += 9) {
+      for (size_t k = 1; k <= 80; k += 17) {
         for (uint32_t m = 1; m <= 2; m++) {
           GemmMicrokernelTester()
             .mr(2)
@@ -17287,7 +17287,7 @@
 
   TEST(QS8_GEMM_MINMAX_2X16C8__NEON_MULL_PADAL, strided_cm_subtile) {
     TEST_REQUIRES_ARM_NEON;
-    for (size_t k = 1; k <= 40; k += 9) {
+    for (size_t k = 1; k <= 80; k += 17) {
       for (uint32_t m = 1; m <= 2; m++) {
         for (uint32_t n = 1; n <= 16; n++) {
           GemmMicrokernelTester()
@@ -17315,7 +17315,7 @@
       .sr(1)
       .m(2)
       .n(16)
-      .k(8)
+      .k(16)
       .qmin(128)
       .Test(xnn_qs8_gemm_minmax_ukernel_2x16c8__neon_mull_padal);
   }
@@ -17329,7 +17329,7 @@
       .sr(1)
       .m(2)
       .n(16)
-      .k(8)
+      .k(16)
       .qmax(128)
       .Test(xnn_qs8_gemm_minmax_ukernel_2x16c8__neon_mull_padal);
   }
@@ -17343,7 +17343,7 @@
       .sr(1)
       .m(2)
       .n(16)
-      .k(8)
+      .k(16)
       .cm_stride(19)
       .Test(xnn_qs8_gemm_minmax_ukernel_2x16c8__neon_mull_padal);
   }
@@ -17351,7 +17351,7 @@
 
 
 #if XNN_ARCH_ARM || XNN_ARCH_ARM64
-  TEST(QS8_GEMM_MINMAX_3X16C8__NEON_MULL_PADAL, k_eq_8) {
+  TEST(QS8_GEMM_MINMAX_3X16C8__NEON_MULL_PADAL, k_eq_16) {
     TEST_REQUIRES_ARM_NEON;
     GemmMicrokernelTester()
       .mr(3)
@@ -17360,7 +17360,7 @@
       .sr(1)
       .m(3)
       .n(16)
-      .k(8)
+      .k(16)
       .Test(xnn_qs8_gemm_minmax_ukernel_3x16c8__neon_mull_padal);
   }
 
@@ -17373,12 +17373,12 @@
       .sr(1)
       .m(3)
       .n(16)
-      .k(8)
+      .k(16)
       .cn_stride(19)
       .Test(xnn_qs8_gemm_minmax_ukernel_3x16c8__neon_mull_padal);
   }
 
-  TEST(QS8_GEMM_MINMAX_3X16C8__NEON_MULL_PADAL, k_eq_8_strided_a) {
+  TEST(QS8_GEMM_MINMAX_3X16C8__NEON_MULL_PADAL, k_eq_16_strided_a) {
     TEST_REQUIRES_ARM_NEON;
     GemmMicrokernelTester()
       .mr(3)
@@ -17387,12 +17387,12 @@
       .sr(1)
       .m(3)
       .n(16)
-      .k(8)
-      .a_stride(11)
+      .k(16)
+      .a_stride(19)
       .Test(xnn_qs8_gemm_minmax_ukernel_3x16c8__neon_mull_padal);
   }
 
-  TEST(QS8_GEMM_MINMAX_3X16C8__NEON_MULL_PADAL, k_eq_8_subtile) {
+  TEST(QS8_GEMM_MINMAX_3X16C8__NEON_MULL_PADAL, k_eq_16_subtile) {
     TEST_REQUIRES_ARM_NEON;
     for (uint32_t m = 1; m <= 3; m++) {
       for (uint32_t n = 1; n <= 16; n++) {
@@ -17403,14 +17403,14 @@
           .sr(1)
           .m(m)
           .n(n)
-          .k(8)
+          .k(16)
           .iterations(1)
           .Test(xnn_qs8_gemm_minmax_ukernel_3x16c8__neon_mull_padal);
       }
     }
   }
 
-  TEST(QS8_GEMM_MINMAX_3X16C8__NEON_MULL_PADAL, k_eq_8_subtile_m) {
+  TEST(QS8_GEMM_MINMAX_3X16C8__NEON_MULL_PADAL, k_eq_16_subtile_m) {
     TEST_REQUIRES_ARM_NEON;
     for (uint32_t m = 1; m <= 3; m++) {
       GemmMicrokernelTester()
@@ -17420,13 +17420,13 @@
         .sr(1)
         .m(m)
         .n(16)
-        .k(8)
+        .k(16)
         .iterations(1)
         .Test(xnn_qs8_gemm_minmax_ukernel_3x16c8__neon_mull_padal);
     }
   }
 
-  TEST(QS8_GEMM_MINMAX_3X16C8__NEON_MULL_PADAL, k_eq_8_subtile_n) {
+  TEST(QS8_GEMM_MINMAX_3X16C8__NEON_MULL_PADAL, k_eq_16_subtile_n) {
     TEST_REQUIRES_ARM_NEON;
     for (uint32_t n = 1; n <= 16; n++) {
       GemmMicrokernelTester()
@@ -17436,15 +17436,15 @@
         .sr(1)
         .m(3)
         .n(n)
-        .k(8)
+        .k(16)
         .iterations(1)
         .Test(xnn_qs8_gemm_minmax_ukernel_3x16c8__neon_mull_padal);
     }
   }
 
-  TEST(QS8_GEMM_MINMAX_3X16C8__NEON_MULL_PADAL, k_lt_8) {
+  TEST(QS8_GEMM_MINMAX_3X16C8__NEON_MULL_PADAL, k_lt_16) {
     TEST_REQUIRES_ARM_NEON;
-    for (size_t k = 1; k < 8; k++) {
+    for (size_t k = 1; k < 16; k++) {
       GemmMicrokernelTester()
         .mr(3)
         .nr(16)
@@ -17457,60 +17457,9 @@
     }
   }
 
-  TEST(QS8_GEMM_MINMAX_3X16C8__NEON_MULL_PADAL, k_lt_8_strided_a) {
+  TEST(QS8_GEMM_MINMAX_3X16C8__NEON_MULL_PADAL, k_lt_16_strided_a) {
     TEST_REQUIRES_ARM_NEON;
-    for (size_t k = 1; k < 8; k++) {
-      GemmMicrokernelTester()
-        .mr(3)
-        .nr(16)
-        .kr(8)
-        .sr(1)
-        .m(3)
-        .n(16)
-        .k(k)
-        .a_stride(11)
-        .Test(xnn_qs8_gemm_minmax_ukernel_3x16c8__neon_mull_padal);
-    }
-  }
-
-  TEST(QS8_GEMM_MINMAX_3X16C8__NEON_MULL_PADAL, k_lt_8_subtile) {
-    TEST_REQUIRES_ARM_NEON;
-    for (size_t k = 1; k < 8; k++) {
-      for (uint32_t m = 1; m <= 3; m++) {
-        for (uint32_t n = 1; n <= 16; n++) {
-          GemmMicrokernelTester()
-            .mr(3)
-            .nr(16)
-            .kr(8)
-            .sr(1)
-            .m(m)
-            .n(n)
-            .k(k)
-            .iterations(1)
-            .Test(xnn_qs8_gemm_minmax_ukernel_3x16c8__neon_mull_padal);
-        }
-      }
-    }
-  }
-
-  TEST(QS8_GEMM_MINMAX_3X16C8__NEON_MULL_PADAL, k_gt_8) {
-    TEST_REQUIRES_ARM_NEON;
-    for (size_t k = 9; k < 16; k++) {
-      GemmMicrokernelTester()
-        .mr(3)
-        .nr(16)
-        .kr(8)
-        .sr(1)
-        .m(3)
-        .n(16)
-        .k(k)
-        .Test(xnn_qs8_gemm_minmax_ukernel_3x16c8__neon_mull_padal);
-    }
-  }
-
-  TEST(QS8_GEMM_MINMAX_3X16C8__NEON_MULL_PADAL, k_gt_8_strided_a) {
-    TEST_REQUIRES_ARM_NEON;
-    for (size_t k = 9; k < 16; k++) {
+    for (size_t k = 1; k < 16; k++) {
       GemmMicrokernelTester()
         .mr(3)
         .nr(16)
@@ -17524,9 +17473,9 @@
     }
   }
 
-  TEST(QS8_GEMM_MINMAX_3X16C8__NEON_MULL_PADAL, k_gt_8_subtile) {
+  TEST(QS8_GEMM_MINMAX_3X16C8__NEON_MULL_PADAL, k_lt_16_subtile) {
     TEST_REQUIRES_ARM_NEON;
-    for (size_t k = 9; k < 16; k++) {
+    for (size_t k = 1; k < 16; k++) {
       for (uint32_t m = 1; m <= 3; m++) {
         for (uint32_t n = 1; n <= 16; n++) {
           GemmMicrokernelTester()
@@ -17544,9 +17493,9 @@
     }
   }
 
-  TEST(QS8_GEMM_MINMAX_3X16C8__NEON_MULL_PADAL, k_div_8) {
+  TEST(QS8_GEMM_MINMAX_3X16C8__NEON_MULL_PADAL, k_gt_16) {
     TEST_REQUIRES_ARM_NEON;
-    for (size_t k = 16; k <= 80; k += 8) {
+    for (size_t k = 17; k < 32; k++) {
       GemmMicrokernelTester()
         .mr(3)
         .nr(16)
@@ -17559,9 +17508,9 @@
     }
   }
 
-  TEST(QS8_GEMM_MINMAX_3X16C8__NEON_MULL_PADAL, k_div_8_strided_a) {
+  TEST(QS8_GEMM_MINMAX_3X16C8__NEON_MULL_PADAL, k_gt_16_strided_a) {
     TEST_REQUIRES_ARM_NEON;
-    for (size_t k = 16; k <= 80; k += 8) {
+    for (size_t k = 17; k < 32; k++) {
       GemmMicrokernelTester()
         .mr(3)
         .nr(16)
@@ -17570,14 +17519,65 @@
         .m(3)
         .n(16)
         .k(k)
-        .a_stride(83)
+        .a_stride(37)
         .Test(xnn_qs8_gemm_minmax_ukernel_3x16c8__neon_mull_padal);
     }
   }
 
-  TEST(QS8_GEMM_MINMAX_3X16C8__NEON_MULL_PADAL, k_div_8_subtile) {
+  TEST(QS8_GEMM_MINMAX_3X16C8__NEON_MULL_PADAL, k_gt_16_subtile) {
     TEST_REQUIRES_ARM_NEON;
-    for (size_t k = 16; k <= 80; k += 8) {
+    for (size_t k = 17; k < 32; k++) {
+      for (uint32_t m = 1; m <= 3; m++) {
+        for (uint32_t n = 1; n <= 16; n++) {
+          GemmMicrokernelTester()
+            .mr(3)
+            .nr(16)
+            .kr(8)
+            .sr(1)
+            .m(m)
+            .n(n)
+            .k(k)
+            .iterations(1)
+            .Test(xnn_qs8_gemm_minmax_ukernel_3x16c8__neon_mull_padal);
+        }
+      }
+    }
+  }
+
+  TEST(QS8_GEMM_MINMAX_3X16C8__NEON_MULL_PADAL, k_div_16) {
+    TEST_REQUIRES_ARM_NEON;
+    for (size_t k = 32; k <= 160; k += 16) {
+      GemmMicrokernelTester()
+        .mr(3)
+        .nr(16)
+        .kr(8)
+        .sr(1)
+        .m(3)
+        .n(16)
+        .k(k)
+        .Test(xnn_qs8_gemm_minmax_ukernel_3x16c8__neon_mull_padal);
+    }
+  }
+
+  TEST(QS8_GEMM_MINMAX_3X16C8__NEON_MULL_PADAL, k_div_16_strided_a) {
+    TEST_REQUIRES_ARM_NEON;
+    for (size_t k = 32; k <= 160; k += 16) {
+      GemmMicrokernelTester()
+        .mr(3)
+        .nr(16)
+        .kr(8)
+        .sr(1)
+        .m(3)
+        .n(16)
+        .k(k)
+        .a_stride(163)
+        .Test(xnn_qs8_gemm_minmax_ukernel_3x16c8__neon_mull_padal);
+    }
+  }
+
+  TEST(QS8_GEMM_MINMAX_3X16C8__NEON_MULL_PADAL, k_div_16_subtile) {
+    TEST_REQUIRES_ARM_NEON;
+    for (size_t k = 32; k <= 160; k += 16) {
       for (uint32_t m = 1; m <= 3; m++) {
         for (uint32_t n = 1; n <= 16; n++) {
           GemmMicrokernelTester()
@@ -17598,7 +17598,7 @@
   TEST(QS8_GEMM_MINMAX_3X16C8__NEON_MULL_PADAL, n_gt_16) {
     TEST_REQUIRES_ARM_NEON;
     for (uint32_t n = 17; n < 32; n++) {
-      for (size_t k = 1; k <= 40; k += 9) {
+      for (size_t k = 1; k <= 80; k += 17) {
         GemmMicrokernelTester()
           .mr(3)
           .nr(16)
@@ -17615,7 +17615,7 @@
   TEST(QS8_GEMM_MINMAX_3X16C8__NEON_MULL_PADAL, n_gt_16_strided_cn) {
     TEST_REQUIRES_ARM_NEON;
     for (uint32_t n = 17; n < 32; n++) {
-      for (size_t k = 1; k <= 40; k += 9) {
+      for (size_t k = 1; k <= 80; k += 17) {
         GemmMicrokernelTester()
           .mr(3)
           .nr(16)
@@ -17633,7 +17633,7 @@
   TEST(QS8_GEMM_MINMAX_3X16C8__NEON_MULL_PADAL, n_gt_16_strided_a) {
     TEST_REQUIRES_ARM_NEON;
     for (uint32_t n = 17; n < 32; n++) {
-      for (size_t k = 1; k <= 40; k += 9) {
+      for (size_t k = 1; k <= 80; k += 17) {
         GemmMicrokernelTester()
           .mr(3)
           .nr(16)
@@ -17642,7 +17642,7 @@
           .m(3)
           .n(n)
           .k(k)
-          .a_stride(43)
+          .a_stride(83)
           .Test(xnn_qs8_gemm_minmax_ukernel_3x16c8__neon_mull_padal);
       }
     }
@@ -17651,7 +17651,7 @@
   TEST(QS8_GEMM_MINMAX_3X16C8__NEON_MULL_PADAL, n_gt_16_subtile) {
     TEST_REQUIRES_ARM_NEON;
     for (uint32_t n = 17; n < 32; n++) {
-      for (size_t k = 1; k <= 40; k += 9) {
+      for (size_t k = 1; k <= 80; k += 17) {
         for (uint32_t m = 1; m <= 3; m++) {
           GemmMicrokernelTester()
             .mr(3)
@@ -17671,7 +17671,7 @@
   TEST(QS8_GEMM_MINMAX_3X16C8__NEON_MULL_PADAL, n_div_16) {
     TEST_REQUIRES_ARM_NEON;
     for (uint32_t n = 32; n <= 48; n += 16) {
-      for (size_t k = 1; k <= 40; k += 9) {
+      for (size_t k = 1; k <= 80; k += 17) {
         GemmMicrokernelTester()
           .mr(3)
           .nr(16)
@@ -17688,7 +17688,7 @@
   TEST(QS8_GEMM_MINMAX_3X16C8__NEON_MULL_PADAL, n_div_16_strided_cn) {
     TEST_REQUIRES_ARM_NEON;
     for (uint32_t n = 32; n <= 48; n += 16) {
-      for (size_t k = 1; k <= 40; k += 9) {
+      for (size_t k = 1; k <= 80; k += 17) {
         GemmMicrokernelTester()
           .mr(3)
           .nr(16)
@@ -17706,7 +17706,7 @@
   TEST(QS8_GEMM_MINMAX_3X16C8__NEON_MULL_PADAL, n_div_16_strided_a) {
     TEST_REQUIRES_ARM_NEON;
     for (uint32_t n = 32; n <= 48; n += 16) {
-      for (size_t k = 1; k <= 40; k += 9) {
+      for (size_t k = 1; k <= 80; k += 17) {
         GemmMicrokernelTester()
           .mr(3)
           .nr(16)
@@ -17715,7 +17715,7 @@
           .m(3)
           .n(n)
           .k(k)
-          .a_stride(43)
+          .a_stride(83)
           .Test(xnn_qs8_gemm_minmax_ukernel_3x16c8__neon_mull_padal);
       }
     }
@@ -17724,7 +17724,7 @@
   TEST(QS8_GEMM_MINMAX_3X16C8__NEON_MULL_PADAL, n_div_16_subtile) {
     TEST_REQUIRES_ARM_NEON;
     for (uint32_t n = 32; n <= 48; n += 16) {
-      for (size_t k = 1; k <= 40; k += 9) {
+      for (size_t k = 1; k <= 80; k += 17) {
         for (uint32_t m = 1; m <= 3; m++) {
           GemmMicrokernelTester()
             .mr(3)
@@ -17743,7 +17743,7 @@
 
   TEST(QS8_GEMM_MINMAX_3X16C8__NEON_MULL_PADAL, strided_cm_subtile) {
     TEST_REQUIRES_ARM_NEON;
-    for (size_t k = 1; k <= 40; k += 9) {
+    for (size_t k = 1; k <= 80; k += 17) {
       for (uint32_t m = 1; m <= 3; m++) {
         for (uint32_t n = 1; n <= 16; n++) {
           GemmMicrokernelTester()
@@ -17771,7 +17771,7 @@
       .sr(1)
       .m(3)
       .n(16)
-      .k(8)
+      .k(16)
       .qmin(128)
       .Test(xnn_qs8_gemm_minmax_ukernel_3x16c8__neon_mull_padal);
   }
@@ -17785,7 +17785,7 @@
       .sr(1)
       .m(3)
       .n(16)
-      .k(8)
+      .k(16)
       .qmax(128)
       .Test(xnn_qs8_gemm_minmax_ukernel_3x16c8__neon_mull_padal);
   }
@@ -17799,7 +17799,7 @@
       .sr(1)
       .m(3)
       .n(16)
-      .k(8)
+      .k(16)
       .cm_stride(19)
       .Test(xnn_qs8_gemm_minmax_ukernel_3x16c8__neon_mull_padal);
   }
@@ -17807,7 +17807,7 @@
 
 
 #if XNN_ARCH_ARM || XNN_ARCH_ARM64
-  TEST(QS8_GEMM_MINMAX_4X16C8__NEON_MULL_PADAL, k_eq_8) {
+  TEST(QS8_GEMM_MINMAX_4X16C8__NEON_MULL_PADAL, k_eq_16) {
     TEST_REQUIRES_ARM_NEON;
     GemmMicrokernelTester()
       .mr(4)
@@ -17816,7 +17816,7 @@
       .sr(1)
       .m(4)
       .n(16)
-      .k(8)
+      .k(16)
       .Test(xnn_qs8_gemm_minmax_ukernel_4x16c8__neon_mull_padal);
   }
 
@@ -17829,12 +17829,12 @@
       .sr(1)
       .m(4)
       .n(16)
-      .k(8)
+      .k(16)
       .cn_stride(19)
       .Test(xnn_qs8_gemm_minmax_ukernel_4x16c8__neon_mull_padal);
   }
 
-  TEST(QS8_GEMM_MINMAX_4X16C8__NEON_MULL_PADAL, k_eq_8_strided_a) {
+  TEST(QS8_GEMM_MINMAX_4X16C8__NEON_MULL_PADAL, k_eq_16_strided_a) {
     TEST_REQUIRES_ARM_NEON;
     GemmMicrokernelTester()
       .mr(4)
@@ -17843,12 +17843,12 @@
       .sr(1)
       .m(4)
       .n(16)
-      .k(8)
-      .a_stride(11)
+      .k(16)
+      .a_stride(19)
       .Test(xnn_qs8_gemm_minmax_ukernel_4x16c8__neon_mull_padal);
   }
 
-  TEST(QS8_GEMM_MINMAX_4X16C8__NEON_MULL_PADAL, k_eq_8_subtile) {
+  TEST(QS8_GEMM_MINMAX_4X16C8__NEON_MULL_PADAL, k_eq_16_subtile) {
     TEST_REQUIRES_ARM_NEON;
     for (uint32_t m = 1; m <= 4; m++) {
       for (uint32_t n = 1; n <= 16; n++) {
@@ -17859,14 +17859,14 @@
           .sr(1)
           .m(m)
           .n(n)
-          .k(8)
+          .k(16)
           .iterations(1)
           .Test(xnn_qs8_gemm_minmax_ukernel_4x16c8__neon_mull_padal);
       }
     }
   }
 
-  TEST(QS8_GEMM_MINMAX_4X16C8__NEON_MULL_PADAL, k_eq_8_subtile_m) {
+  TEST(QS8_GEMM_MINMAX_4X16C8__NEON_MULL_PADAL, k_eq_16_subtile_m) {
     TEST_REQUIRES_ARM_NEON;
     for (uint32_t m = 1; m <= 4; m++) {
       GemmMicrokernelTester()
@@ -17876,13 +17876,13 @@
         .sr(1)
         .m(m)
         .n(16)
-        .k(8)
+        .k(16)
         .iterations(1)
         .Test(xnn_qs8_gemm_minmax_ukernel_4x16c8__neon_mull_padal);
     }
   }
 
-  TEST(QS8_GEMM_MINMAX_4X16C8__NEON_MULL_PADAL, k_eq_8_subtile_n) {
+  TEST(QS8_GEMM_MINMAX_4X16C8__NEON_MULL_PADAL, k_eq_16_subtile_n) {
     TEST_REQUIRES_ARM_NEON;
     for (uint32_t n = 1; n <= 16; n++) {
       GemmMicrokernelTester()
@@ -17892,15 +17892,15 @@
         .sr(1)
         .m(4)
         .n(n)
-        .k(8)
+        .k(16)
         .iterations(1)
         .Test(xnn_qs8_gemm_minmax_ukernel_4x16c8__neon_mull_padal);
     }
   }
 
-  TEST(QS8_GEMM_MINMAX_4X16C8__NEON_MULL_PADAL, k_lt_8) {
+  TEST(QS8_GEMM_MINMAX_4X16C8__NEON_MULL_PADAL, k_lt_16) {
     TEST_REQUIRES_ARM_NEON;
-    for (size_t k = 1; k < 8; k++) {
+    for (size_t k = 1; k < 16; k++) {
       GemmMicrokernelTester()
         .mr(4)
         .nr(16)
@@ -17913,60 +17913,9 @@
     }
   }
 
-  TEST(QS8_GEMM_MINMAX_4X16C8__NEON_MULL_PADAL, k_lt_8_strided_a) {
+  TEST(QS8_GEMM_MINMAX_4X16C8__NEON_MULL_PADAL, k_lt_16_strided_a) {
     TEST_REQUIRES_ARM_NEON;
-    for (size_t k = 1; k < 8; k++) {
-      GemmMicrokernelTester()
-        .mr(4)
-        .nr(16)
-        .kr(8)
-        .sr(1)
-        .m(4)
-        .n(16)
-        .k(k)
-        .a_stride(11)
-        .Test(xnn_qs8_gemm_minmax_ukernel_4x16c8__neon_mull_padal);
-    }
-  }
-
-  TEST(QS8_GEMM_MINMAX_4X16C8__NEON_MULL_PADAL, k_lt_8_subtile) {
-    TEST_REQUIRES_ARM_NEON;
-    for (size_t k = 1; k < 8; k++) {
-      for (uint32_t m = 1; m <= 4; m++) {
-        for (uint32_t n = 1; n <= 16; n++) {
-          GemmMicrokernelTester()
-            .mr(4)
-            .nr(16)
-            .kr(8)
-            .sr(1)
-            .m(m)
-            .n(n)
-            .k(k)
-            .iterations(1)
-            .Test(xnn_qs8_gemm_minmax_ukernel_4x16c8__neon_mull_padal);
-        }
-      }
-    }
-  }
-
-  TEST(QS8_GEMM_MINMAX_4X16C8__NEON_MULL_PADAL, k_gt_8) {
-    TEST_REQUIRES_ARM_NEON;
-    for (size_t k = 9; k < 16; k++) {
-      GemmMicrokernelTester()
-        .mr(4)
-        .nr(16)
-        .kr(8)
-        .sr(1)
-        .m(4)
-        .n(16)
-        .k(k)
-        .Test(xnn_qs8_gemm_minmax_ukernel_4x16c8__neon_mull_padal);
-    }
-  }
-
-  TEST(QS8_GEMM_MINMAX_4X16C8__NEON_MULL_PADAL, k_gt_8_strided_a) {
-    TEST_REQUIRES_ARM_NEON;
-    for (size_t k = 9; k < 16; k++) {
+    for (size_t k = 1; k < 16; k++) {
       GemmMicrokernelTester()
         .mr(4)
         .nr(16)
@@ -17980,9 +17929,9 @@
     }
   }
 
-  TEST(QS8_GEMM_MINMAX_4X16C8__NEON_MULL_PADAL, k_gt_8_subtile) {
+  TEST(QS8_GEMM_MINMAX_4X16C8__NEON_MULL_PADAL, k_lt_16_subtile) {
     TEST_REQUIRES_ARM_NEON;
-    for (size_t k = 9; k < 16; k++) {
+    for (size_t k = 1; k < 16; k++) {
       for (uint32_t m = 1; m <= 4; m++) {
         for (uint32_t n = 1; n <= 16; n++) {
           GemmMicrokernelTester()
@@ -18000,9 +17949,9 @@
     }
   }
 
-  TEST(QS8_GEMM_MINMAX_4X16C8__NEON_MULL_PADAL, k_div_8) {
+  TEST(QS8_GEMM_MINMAX_4X16C8__NEON_MULL_PADAL, k_gt_16) {
     TEST_REQUIRES_ARM_NEON;
-    for (size_t k = 16; k <= 80; k += 8) {
+    for (size_t k = 17; k < 32; k++) {
       GemmMicrokernelTester()
         .mr(4)
         .nr(16)
@@ -18015,9 +17964,9 @@
     }
   }
 
-  TEST(QS8_GEMM_MINMAX_4X16C8__NEON_MULL_PADAL, k_div_8_strided_a) {
+  TEST(QS8_GEMM_MINMAX_4X16C8__NEON_MULL_PADAL, k_gt_16_strided_a) {
     TEST_REQUIRES_ARM_NEON;
-    for (size_t k = 16; k <= 80; k += 8) {
+    for (size_t k = 17; k < 32; k++) {
       GemmMicrokernelTester()
         .mr(4)
         .nr(16)
@@ -18026,14 +17975,65 @@
         .m(4)
         .n(16)
         .k(k)
-        .a_stride(83)
+        .a_stride(37)
         .Test(xnn_qs8_gemm_minmax_ukernel_4x16c8__neon_mull_padal);
     }
   }
 
-  TEST(QS8_GEMM_MINMAX_4X16C8__NEON_MULL_PADAL, k_div_8_subtile) {
+  TEST(QS8_GEMM_MINMAX_4X16C8__NEON_MULL_PADAL, k_gt_16_subtile) {
     TEST_REQUIRES_ARM_NEON;
-    for (size_t k = 16; k <= 80; k += 8) {
+    for (size_t k = 17; k < 32; k++) {
+      for (uint32_t m = 1; m <= 4; m++) {
+        for (uint32_t n = 1; n <= 16; n++) {
+          GemmMicrokernelTester()
+            .mr(4)
+            .nr(16)
+            .kr(8)
+            .sr(1)
+            .m(m)
+            .n(n)
+            .k(k)
+            .iterations(1)
+            .Test(xnn_qs8_gemm_minmax_ukernel_4x16c8__neon_mull_padal);
+        }
+      }
+    }
+  }
+
+  TEST(QS8_GEMM_MINMAX_4X16C8__NEON_MULL_PADAL, k_div_16) {
+    TEST_REQUIRES_ARM_NEON;
+    for (size_t k = 32; k <= 160; k += 16) {
+      GemmMicrokernelTester()
+        .mr(4)
+        .nr(16)
+        .kr(8)
+        .sr(1)
+        .m(4)
+        .n(16)
+        .k(k)
+        .Test(xnn_qs8_gemm_minmax_ukernel_4x16c8__neon_mull_padal);
+    }
+  }
+
+  TEST(QS8_GEMM_MINMAX_4X16C8__NEON_MULL_PADAL, k_div_16_strided_a) {
+    TEST_REQUIRES_ARM_NEON;
+    for (size_t k = 32; k <= 160; k += 16) {
+      GemmMicrokernelTester()
+        .mr(4)
+        .nr(16)
+        .kr(8)
+        .sr(1)
+        .m(4)
+        .n(16)
+        .k(k)
+        .a_stride(163)
+        .Test(xnn_qs8_gemm_minmax_ukernel_4x16c8__neon_mull_padal);
+    }
+  }
+
+  TEST(QS8_GEMM_MINMAX_4X16C8__NEON_MULL_PADAL, k_div_16_subtile) {
+    TEST_REQUIRES_ARM_NEON;
+    for (size_t k = 32; k <= 160; k += 16) {
       for (uint32_t m = 1; m <= 4; m++) {
         for (uint32_t n = 1; n <= 16; n++) {
           GemmMicrokernelTester()
@@ -18054,7 +18054,7 @@
   TEST(QS8_GEMM_MINMAX_4X16C8__NEON_MULL_PADAL, n_gt_16) {
     TEST_REQUIRES_ARM_NEON;
     for (uint32_t n = 17; n < 32; n++) {
-      for (size_t k = 1; k <= 40; k += 9) {
+      for (size_t k = 1; k <= 80; k += 17) {
         GemmMicrokernelTester()
           .mr(4)
           .nr(16)
@@ -18071,7 +18071,7 @@
   TEST(QS8_GEMM_MINMAX_4X16C8__NEON_MULL_PADAL, n_gt_16_strided_cn) {
     TEST_REQUIRES_ARM_NEON;
     for (uint32_t n = 17; n < 32; n++) {
-      for (size_t k = 1; k <= 40; k += 9) {
+      for (size_t k = 1; k <= 80; k += 17) {
         GemmMicrokernelTester()
           .mr(4)
           .nr(16)
@@ -18089,7 +18089,7 @@
   TEST(QS8_GEMM_MINMAX_4X16C8__NEON_MULL_PADAL, n_gt_16_strided_a) {
     TEST_REQUIRES_ARM_NEON;
     for (uint32_t n = 17; n < 32; n++) {
-      for (size_t k = 1; k <= 40; k += 9) {
+      for (size_t k = 1; k <= 80; k += 17) {
         GemmMicrokernelTester()
           .mr(4)
           .nr(16)
@@ -18098,7 +18098,7 @@
           .m(4)
           .n(n)
           .k(k)
-          .a_stride(43)
+          .a_stride(83)
           .Test(xnn_qs8_gemm_minmax_ukernel_4x16c8__neon_mull_padal);
       }
     }
@@ -18107,7 +18107,7 @@
   TEST(QS8_GEMM_MINMAX_4X16C8__NEON_MULL_PADAL, n_gt_16_subtile) {
     TEST_REQUIRES_ARM_NEON;
     for (uint32_t n = 17; n < 32; n++) {
-      for (size_t k = 1; k <= 40; k += 9) {
+      for (size_t k = 1; k <= 80; k += 17) {
         for (uint32_t m = 1; m <= 4; m++) {
           GemmMicrokernelTester()
             .mr(4)
@@ -18127,7 +18127,7 @@
   TEST(QS8_GEMM_MINMAX_4X16C8__NEON_MULL_PADAL, n_div_16) {
     TEST_REQUIRES_ARM_NEON;
     for (uint32_t n = 32; n <= 48; n += 16) {
-      for (size_t k = 1; k <= 40; k += 9) {
+      for (size_t k = 1; k <= 80; k += 17) {
         GemmMicrokernelTester()
           .mr(4)
           .nr(16)
@@ -18144,7 +18144,7 @@
   TEST(QS8_GEMM_MINMAX_4X16C8__NEON_MULL_PADAL, n_div_16_strided_cn) {
     TEST_REQUIRES_ARM_NEON;
     for (uint32_t n = 32; n <= 48; n += 16) {
-      for (size_t k = 1; k <= 40; k += 9) {
+      for (size_t k = 1; k <= 80; k += 17) {
         GemmMicrokernelTester()
           .mr(4)
           .nr(16)
@@ -18162,7 +18162,7 @@
   TEST(QS8_GEMM_MINMAX_4X16C8__NEON_MULL_PADAL, n_div_16_strided_a) {
     TEST_REQUIRES_ARM_NEON;
     for (uint32_t n = 32; n <= 48; n += 16) {
-      for (size_t k = 1; k <= 40; k += 9) {
+      for (size_t k = 1; k <= 80; k += 17) {
         GemmMicrokernelTester()
           .mr(4)
           .nr(16)
@@ -18171,7 +18171,7 @@
           .m(4)
           .n(n)
           .k(k)
-          .a_stride(43)
+          .a_stride(83)
           .Test(xnn_qs8_gemm_minmax_ukernel_4x16c8__neon_mull_padal);
       }
     }
@@ -18180,7 +18180,7 @@
   TEST(QS8_GEMM_MINMAX_4X16C8__NEON_MULL_PADAL, n_div_16_subtile) {
     TEST_REQUIRES_ARM_NEON;
     for (uint32_t n = 32; n <= 48; n += 16) {
-      for (size_t k = 1; k <= 40; k += 9) {
+      for (size_t k = 1; k <= 80; k += 17) {
         for (uint32_t m = 1; m <= 4; m++) {
           GemmMicrokernelTester()
             .mr(4)
@@ -18199,7 +18199,7 @@
 
   TEST(QS8_GEMM_MINMAX_4X16C8__NEON_MULL_PADAL, strided_cm_subtile) {
     TEST_REQUIRES_ARM_NEON;
-    for (size_t k = 1; k <= 40; k += 9) {
+    for (size_t k = 1; k <= 80; k += 17) {
       for (uint32_t m = 1; m <= 4; m++) {
         for (uint32_t n = 1; n <= 16; n++) {
           GemmMicrokernelTester()
@@ -18227,7 +18227,7 @@
       .sr(1)
       .m(4)
       .n(16)
-      .k(8)
+      .k(16)
       .qmin(128)
       .Test(xnn_qs8_gemm_minmax_ukernel_4x16c8__neon_mull_padal);
   }
@@ -18241,7 +18241,7 @@
       .sr(1)
       .m(4)
       .n(16)
-      .k(8)
+      .k(16)
       .qmax(128)
       .Test(xnn_qs8_gemm_minmax_ukernel_4x16c8__neon_mull_padal);
   }
@@ -18255,7 +18255,7 @@
       .sr(1)
       .m(4)
       .n(16)
-      .k(8)
+      .k(16)
       .cm_stride(19)
       .Test(xnn_qs8_gemm_minmax_ukernel_4x16c8__neon_mull_padal);
   }

diff --git a/test/qs8-gemm-minmax.yaml b/test/qs8-gemm-minmax.yaml
index 78cabd2..ebcdf94 100644
--- a/test/qs8-gemm-minmax.yaml
+++ b/test/qs8-gemm-minmax.yaml

@@ -67,21 +67,21 @@
 - name: xnn_qs8_gemm_minmax_ukernel_4x16c2__neon_mlal_padal_dup
   k-block: 16
 - name: xnn_qs8_gemm_minmax_ukernel_1x8c8__neon_mull_padal
-  k-block: 8
+  k-block: 16
 - name: xnn_qs8_gemm_minmax_ukernel_2x8c8__neon_mull_padal
-  k-block: 8
+  k-block: 16
 - name: xnn_qs8_gemm_minmax_ukernel_3x8c8__neon_mull_padal
-  k-block: 8
+  k-block: 16
 - name: xnn_qs8_gemm_minmax_ukernel_4x8c8__neon_mull_padal
-  k-block: 8
+  k-block: 16
 - name: xnn_qs8_gemm_minmax_ukernel_1x16c8__neon_mull_padal
-  k-block: 8
+  k-block: 16
 - name: xnn_qs8_gemm_minmax_ukernel_2x16c8__neon_mull_padal
-  k-block: 8
+  k-block: 16
 - name: xnn_qs8_gemm_minmax_ukernel_3x16c8__neon_mull_padal
-  k-block: 8
+  k-block: 16
 - name: xnn_qs8_gemm_minmax_ukernel_4x16c8__neon_mull_padal
-  k-block: 8
+  k-block: 16
 - name: xnn_qs8_gemm_minmax_ukernel_1x8c4__neondot
   k-block: 8
 - name: xnn_qs8_gemm_minmax_ukernel_4x8c4__neondot
commit	c8532ae6ac9550ee44b126052ac39cc02ff58052	[log] [tgz]
author	Frank Barchard <fbarchard@google.com>	Wed Feb 10 20:01:39 2021 -0800
committer	XNNPACK Team <xnnpack-github-robot@google.com>	Wed Feb 10 20:02:26 2021 -0800
tree	11fec338f8f304e36fac5fdcc2f16be34637d9d9
parent	4c032f34f870ee35a10ba4c5e203de6cc2565865 [diff]