neondot qs8 rsum use const for remainder masking

PiperOrigin-RevId: 635010053

Author: fbarchard

src/qs8-rsum/gen/qs8-rsum-minmax-fp32-neondot-u16.c

@@ -38,8 +38,8 @@ void xnn_qs8_rsum_minmax_fp32_ukernel__neondot_u16(
       vacc0 = vdotq_s32(vacc0, vt, vone);
     }
     if (XNN_UNLIKELY(batch != 0)) {
-      int8x16_t vt = vld1q_s8(input);
-      vone = vld1q_s8(&params->fp32_neon.mask_table[15 - batch]);
+      const int8x16_t vt = vld1q_s8(input);
+      const vone = vld1q_s8(&params->fp32_neon.mask_table[15 - batch]);
       vacc0 = vdotq_s32(vacc0, vt, vone);
     }
   }

src/qs8-rsum/gen/qs8-rsum-minmax-fp32-neondot-u32-acc2.c

@@ -41,8 +41,8 @@ void xnn_qs8_rsum_minmax_fp32_ukernel__neondot_u32_acc2(
       vacc0 = vdotq_s32(vacc0, vt, vone);
     }
     if (XNN_UNLIKELY(batch != 0)) {
-      int8x16_t vt = vld1q_s8(input);
-      vone = vld1q_s8(&params->fp32_neon.mask_table[15 - batch]);
+      const int8x16_t vt = vld1q_s8(input);
+      const vone = vld1q_s8(&params->fp32_neon.mask_table[15 - batch]);
       vacc0 = vdotq_s32(vacc0, vt, vone);
     }
   }

src/qs8-rsum/gen/qs8-rsum-minmax-fp32-neondot-u32.c

@@ -40,8 +40,8 @@ void xnn_qs8_rsum_minmax_fp32_ukernel__neondot_u32(
       vacc0 = vdotq_s32(vacc0, vt, vone);
     }
     if (XNN_UNLIKELY(batch != 0)) {
-      int8x16_t vt = vld1q_s8(input);
-      vone = vld1q_s8(&params->fp32_neon.mask_table[15 - batch]);
+      const int8x16_t vt = vld1q_s8(input);
+      const vone = vld1q_s8(&params->fp32_neon.mask_table[15 - batch]);
       vacc0 = vdotq_s32(vacc0, vt, vone);
     }
   }

src/qs8-rsum/gen/qs8-rsum-minmax-fp32-neondot-u64-acc2.c

@@ -45,8 +45,8 @@ void xnn_qs8_rsum_minmax_fp32_ukernel__neondot_u64_acc2(
       vacc0 = vdotq_s32(vacc0, vt, vone);
     }
     if (XNN_UNLIKELY(batch != 0)) {
-      int8x16_t vt = vld1q_s8(input);
-      vone = vld1q_s8(&params->fp32_neon.mask_table[15 - batch]);
+      const int8x16_t vt = vld1q_s8(input);
+      const vone = vld1q_s8(&params->fp32_neon.mask_table[15 - batch]);
       vacc0 = vdotq_s32(vacc0, vt, vone);
     }
   }

src/qs8-rsum/gen/qs8-rsum-minmax-fp32-neondot-u64-acc4.c

@@ -47,14 +47,14 @@ void xnn_qs8_rsum_minmax_fp32_ukernel__neondot_u64_acc4(
       vacc0 = vdotq_s32(vacc0, vt, vone);
     }
     if (XNN_UNLIKELY(batch != 0)) {
-      int8x16_t vt = vld1q_s8(input);
-      vone = vld1q_s8(&params->fp32_neon.mask_table[15 - batch]);
+      const int8x16_t vt = vld1q_s8(input);
+      const vone = vld1q_s8(&params->fp32_neon.mask_table[15 - batch]);
       vacc0 = vdotq_s32(vacc0, vt, vone);
     }
   }
   vacc0 = vaddq_s32(vacc0, vacc1);
+  vacc2 = vaddq_s32(vacc2, vacc3);
   vacc0 = vaddq_s32(vacc0, vacc2);
-  vacc0 = vaddq_s32(vacc0, vacc3);
   int32x2_t vacc_lo = vadd_s32(vget_low_s32(vacc0), vget_high_s32(vacc0));
   vacc_lo = vpadd_s32(vacc_lo, vacc_lo);
 

src/qs8-rsum/gen/qs8-rsum-minmax-fp32-neondot-u64.c

@@ -44,8 +44,8 @@ void xnn_qs8_rsum_minmax_fp32_ukernel__neondot_u64(
       vacc0 = vdotq_s32(vacc0, vt, vone);
     }
     if (XNN_UNLIKELY(batch != 0)) {
-      int8x16_t vt = vld1q_s8(input);
-      vone = vld1q_s8(&params->fp32_neon.mask_table[15 - batch]);
+      const int8x16_t vt = vld1q_s8(input);
+      const vone = vld1q_s8(&params->fp32_neon.mask_table[15 - batch]);
       vacc0 = vdotq_s32(vacc0, vt, vone);
     }
   }

src/qs8-rsum/neondot.c.in

@@ -45,8 +45,8 @@ void xnn_qs8_rsum_minmax_${REQUANTIZATION.lower()}_ukernel__neondot_u${CHANNEL_T
       vacc0 = vdotq_s32(vacc0, vt, vone);
     }
     if (XNN_UNLIKELY(batch != 0)) {
-      int8x16_t vt = vld1q_s8(input);
-      vone = vld1q_s8(&params->fp32_neon.mask_table[15 - batch]);
+      const int8x16_t vt = vld1q_s8(input);
+      const vone = vld1q_s8(&params->fp32_neon.mask_table[15 - batch]);
       vacc0 = vdotq_s32(vacc0, vt, vone);
     }
   }