commit 8a045659758ca7f02a1f8daa94b3fbe82802b725
author Marat Dukhan <maratek@google.com> Tue Jul 20 14:07:21 2021 -0700
committer XNNPACK Team <xnnpack-github-robot@google.com> Tue Jul 20 14:08:09 2021 -0700
tree 418f0fc9aade8d6b3378a3d1a8c61d42ff916771
parent c3f69fd3d543cdffbbf9e34630d3da66b22589fc

Use RNDNU requantization in QS8 VADD[C] microkernels

Rounding midpoints up (RNDNU) is more efficient than previously used RNDNA
(rounding midpoints away from zero) requantization

PiperOrigin-RevId: 385867165

src/qs8-vaddc/avx2-mul32-ld64.c.in

diff --git a/src/qs8-vaddc/avx2-mul32-ld64.c.in b/src/qs8-vaddc/avx2-mul32-ld64.c.in
index dd74f53..69d7d34 100644
--- a/src/qs8-vaddc/avx2-mul32-ld64.c.in
+++ b/src/qs8-vaddc/avx2-mul32-ld64.c.in
@@ -46,11 +46,7 @@
       __m256i vacc${ABC[N:N+8]} = _mm256_add_epi32(vbias, _mm256_mullo_epi32(va${ABC[N:N+8]}, va_multiplier));
 
     $for N in range(0, BATCH_TILE, 8):
-      const __m256i vadj${ABC[N:N+8]} = _mm256_srai_epi32(vacc${ABC[N:N+8]}, 31);
-      vacc${ABC[N:N+8]} = _mm256_add_epi32(vacc${ABC[N:N+8]}, vrounding);
-
-    $for N in range(0, BATCH_TILE, 8):
-      vacc${ABC[N:N+8]} = _mm256_sra_epi32(_mm256_add_epi32(vacc${ABC[N:N+8]}, vadj${ABC[N:N+8]}), vshift);
+      vacc${ABC[N:N+8]} = _mm256_sra_epi32(_mm256_add_epi32(vacc${ABC[N:N+8]}, vrounding), vshift);
 
     $for N in range(0, BATCH_TILE, 16):
       $if N + 8 < BATCH_TILE:
@@ -93,9 +89,7 @@
 
       __m256i vacc${ABC[0:8]} = _mm256_add_epi32(vbias, _mm256_mullo_epi32(va${ABC[0:8]}, va_multiplier));
 
-      const __m256i vadj${ABC[0:8]} = _mm256_srai_epi32(vacc${ABC[0:8]}, 31);
-      vacc${ABC[0:8]} = _mm256_add_epi32(vacc${ABC[0:8]}, vrounding);
-      vacc${ABC[0:8]} = _mm256_sra_epi32(_mm256_add_epi32(vacc${ABC[0:8]}, vadj${ABC[0:8]}), vshift);
+      vacc${ABC[0:8]} = _mm256_sra_epi32(_mm256_add_epi32(vacc${ABC[0:8]}, vrounding), vshift);
 
       $if BATCH_TILE > 8:
         __m128i vout${ABC[0:8]} = _mm_adds_epi16(_mm_packs_epi32(_mm256_castsi256_si128(vacc${ABC[0:8]}), _mm256_extracti128_si256(vacc${ABC[0:8]}, 1)), _mm256_castsi256_si128(voutput_zero_point));