Reoptimize x86 requantization

PiperOrigin-RevId: 325031583
diff --git a/src/qs8-igemm/gen/1x4c8-minmax-ssse3-ld128.c b/src/qs8-igemm/gen/1x4c8-minmax-ssse3-ld128.c
index 66122ee..20f15c0 100644
--- a/src/qs8-igemm/gen/1x4c8-minmax-ssse3-ld128.c
+++ b/src/qs8-igemm/gen/1x4c8-minmax-ssse3-ld128.c
@@ -95,7 +95,7 @@
 
     const __m128i vabsacc0x0123 = _mm_abs_epi32(vacc0x0123);
 
-    const __m128i vabsacc0x1032 = _mm_shuffle_epi32(vabsacc0x0123, _MM_SHUFFLE(2, 3, 0, 1));
+    const __m128i vabsacc0x1133 = _mm_shuffle_epi32(vabsacc0x0123, _MM_SHUFFLE(3, 3, 1, 1));
 
     const __m128i vabsprod0x02 = _mm_mul_epu32(vabsacc0x0123, vmultiplier);
 
@@ -105,7 +105,7 @@
 
     const __m128i vq31prod0x02 = _mm_srli_epi64(_mm_add_epi64(vprod0x02, vrounding), 31);
 
-    const __m128i vabsprod0x13 = _mm_mul_epu32(vabsacc0x1032, vmultiplier);
+    const __m128i vabsprod0x13 = _mm_mul_epu32(vabsacc0x1133, vmultiplier);
 
     const __m128i vnmask0x13 = _mm_shuffle_epi32(vnmask0x0123, _MM_SHUFFLE(3, 3, 1, 1));