commit 9b474cfd4d5f237cb85274cd0c4ffde428385fec
author Marat Dukhan <maratek@google.com> Tue May 25 16:37:48 2021 -0700
committer XNNPACK Team <xnnpack-github-robot@google.com> Tue May 25 16:38:39 2021 -0700
tree 3637eb8607f9f139ac0c158853a06d00dcbed648
parent 0a5d7b4b8cc0b5cb9f8013c2bb56048e865493c8

Support FP32 requantization in AVX2 QS8 microkernels

PiperOrigin-RevId: 375822588

src/qs8-igemm/gen/4x4c2-minmax-gemmlowp-xop-ld128.c

diff --git a/src/qs8-igemm/gen/4x4c2-minmax-gemmlowp-xop-ld128.c b/src/qs8-igemm/gen/4x4c2-minmax-gemmlowp-xop-ld128.c
index 3eb4e00..61b2117 100644
--- a/src/qs8-igemm/gen/4x4c2-minmax-gemmlowp-xop-ld128.c
+++ b/src/qs8-igemm/gen/4x4c2-minmax-gemmlowp-xop-ld128.c
@@ -210,8 +210,8 @@
       p -= 4 * sizeof(void*);
     } while (p != 0);
 
-    const __m128i vmultiplier = _mm_load_si128((const __m128i*) params->sse4.multiplier);
-    const __m128i vrounding = _mm_load_si128((const __m128i*) params->sse4.rounding);
+    const __m128i vmultiplier = _mm_load_si128((const __m128i*) params->gemmlowp_sse4.multiplier);
+    const __m128i vrounding = _mm_load_si128((const __m128i*) params->gemmlowp_sse4.rounding);
 
     const __m128i vacc0x1133 = _mm_shuffle_epi32(vacc0x0123, _MM_SHUFFLE(3, 3, 1, 1));
     const __m128i vacc1x1133 = _mm_shuffle_epi32(vacc1x0123, _MM_SHUFFLE(3, 3, 1, 1));
@@ -242,7 +242,7 @@
     const __m128i vq31prod2x0123 = _mm_blend_epi16(vq31prod2x02, vq31prod2x13, 0xCC);
     const __m128i vq31prod3x0123 = _mm_blend_epi16(vq31prod3x02, vq31prod3x13, 0xCC);
 
-    const __m128i vremainder_mask = _mm_load_si128((const __m128i*) params->sse4.remainder_mask);
+    const __m128i vremainder_mask = _mm_load_si128((const __m128i*) params->gemmlowp_sse4.remainder_mask);
     const __m128i vrem0x0123 =
       _mm_add_epi32(_mm_and_si128(vq31prod0x0123, vremainder_mask), _mm_cmpgt_epi32(_mm_setzero_si128(), vq31prod0x0123));
     const __m128i vrem1x0123 =
@@ -252,8 +252,8 @@
     const __m128i vrem3x0123 =
       _mm_add_epi32(_mm_and_si128(vq31prod3x0123, vremainder_mask), _mm_cmpgt_epi32(_mm_setzero_si128(), vq31prod3x0123));
 
-    const __m128i vremainder_threshold = _mm_load_si128((const __m128i*) params->sse4.remainder_threshold);
-    const __m128i vshift = _mm_loadl_epi64((const __m128i*) params->sse4.shift);
+    const __m128i vremainder_threshold = _mm_load_si128((const __m128i*) params->gemmlowp_sse4.remainder_threshold);
+    const __m128i vshift = _mm_loadl_epi64((const __m128i*) params->gemmlowp_sse4.shift);
     vacc0x0123 =
       _mm_sub_epi32(_mm_sra_epi32(vq31prod0x0123, vshift), _mm_cmpgt_epi32(vrem0x0123, vremainder_threshold));
     vacc1x0123 =
@@ -263,15 +263,15 @@
     vacc3x0123 =
       _mm_sub_epi32(_mm_sra_epi32(vq31prod3x0123, vshift), _mm_cmpgt_epi32(vrem3x0123, vremainder_threshold));
 
-    const __m128i voutput_zero_point = _mm_load_si128((const __m128i*) params->sse4.output_zero_point);
+    const __m128i voutput_zero_point = _mm_load_si128((const __m128i*) params->gemmlowp_sse4.output_zero_point);
     __m128i vacc01x0123 = _mm_adds_epi16(_mm_packs_epi32(vacc0x0123, vacc1x0123), voutput_zero_point);
     __m128i vacc23x0123 = _mm_adds_epi16(_mm_packs_epi32(vacc2x0123, vacc3x0123), voutput_zero_point);
 
 
     __m128i vout = _mm_packs_epi16(vacc01x0123, vacc23x0123);
 
-    vout = _mm_max_epi8(vout, _mm_load_si128((const __m128i*) params->sse4.output_min));
-    vout = _mm_min_epi8(vout, _mm_load_si128((const __m128i*) params->sse4.output_max));
+    vout = _mm_max_epi8(vout, _mm_load_si128((const __m128i*) params->gemmlowp_sse4.output_min));
+    vout = _mm_min_epi8(vout, _mm_load_si128((const __m128i*) params->gemmlowp_sse4.output_max));
 
     if (nc >= 4) {
       *((uint32_t*) c3) = (uint32_t) _mm_extract_epi32(vout, 3);