commit a842fef3579f5daf99ed33eb6940dc2763b8af3e
author Marat Dukhan <maratek@google.com> Mon Jul 19 21:07:40 2021 -0700
committer XNNPACK Team <xnnpack-github-robot@google.com> Mon Jul 19 21:08:36 2021 -0700
tree 34b560bfc733190f299e7bcf9cdeeb550394f0f7
parent e6a480568a54e8e4f440394c1e12dfe65b733b51

Rename zero_point_product parameter to bias in Quantized Add microkernels

PiperOrigin-RevId: 385704949

src/qs8-vaddc/gen/minmax-sse41-mul32-ld32-x32.c

diff --git a/src/qs8-vaddc/gen/minmax-sse41-mul32-ld32-x32.c b/src/qs8-vaddc/gen/minmax-sse41-mul32-ld32-x32.c
index d351768..c07928c 100644
--- a/src/qs8-vaddc/gen/minmax-sse41-mul32-ld32-x32.c
+++ b/src/qs8-vaddc/gen/minmax-sse41-mul32-ld32-x32.c
@@ -30,9 +30,9 @@
   const __m128i voutput_min = _mm_load_si128((const __m128i*) params->sse2.output_min);
   const __m128i voutput_max = _mm_load_si128((const __m128i*) params->sse2.output_max);
 
-  __m128i vzero_point_product = _mm_cvtsi32_si128(params->sse2.b_multiplier[0] * (int32_t) *input_b);
-  vzero_point_product = _mm_shuffle_epi32(vzero_point_product, _MM_SHUFFLE(0, 0, 0, 0));
-  vzero_point_product = _mm_add_epi32(vzero_point_product, _mm_load_si128((const __m128i*) params->sse2.zero_point_product));
+  __m128i vbias = _mm_cvtsi32_si128(params->sse2.b_multiplier[0] * (int32_t) *input_b);
+  vbias = _mm_shuffle_epi32(vbias, _MM_SHUFFLE(0, 0, 0, 0));
+  vbias = _mm_add_epi32(vbias, _mm_load_si128((const __m128i*) params->sse2.bias));
   for (; n >= 32 * sizeof(int8_t); n -= 32 * sizeof(int8_t)) {
     const __m128i va0123 = _mm_cvtepi8_epi32(_mm_loadu_si32(input_a));
     const __m128i va4567 = _mm_cvtepi8_epi32(_mm_loadu_si32(input_a + 4));
@@ -45,14 +45,14 @@
     input_a += 32;
     input_b += 32;
 
-    __m128i vacc0123 = _mm_add_epi32(vzero_point_product, _mm_mullo_epi32(va0123, va_multiplier));
-    __m128i vacc4567 = _mm_add_epi32(vzero_point_product, _mm_mullo_epi32(va4567, va_multiplier));
-    __m128i vacc89AB = _mm_add_epi32(vzero_point_product, _mm_mullo_epi32(va89AB, va_multiplier));
-    __m128i vaccCDEF = _mm_add_epi32(vzero_point_product, _mm_mullo_epi32(vaCDEF, va_multiplier));
-    __m128i vaccGHIJ = _mm_add_epi32(vzero_point_product, _mm_mullo_epi32(vaGHIJ, va_multiplier));
-    __m128i vaccKLMN = _mm_add_epi32(vzero_point_product, _mm_mullo_epi32(vaKLMN, va_multiplier));
-    __m128i vaccOPQR = _mm_add_epi32(vzero_point_product, _mm_mullo_epi32(vaOPQR, va_multiplier));
-    __m128i vaccSTUV = _mm_add_epi32(vzero_point_product, _mm_mullo_epi32(vaSTUV, va_multiplier));
+    __m128i vacc0123 = _mm_add_epi32(vbias, _mm_mullo_epi32(va0123, va_multiplier));
+    __m128i vacc4567 = _mm_add_epi32(vbias, _mm_mullo_epi32(va4567, va_multiplier));
+    __m128i vacc89AB = _mm_add_epi32(vbias, _mm_mullo_epi32(va89AB, va_multiplier));
+    __m128i vaccCDEF = _mm_add_epi32(vbias, _mm_mullo_epi32(vaCDEF, va_multiplier));
+    __m128i vaccGHIJ = _mm_add_epi32(vbias, _mm_mullo_epi32(vaGHIJ, va_multiplier));
+    __m128i vaccKLMN = _mm_add_epi32(vbias, _mm_mullo_epi32(vaKLMN, va_multiplier));
+    __m128i vaccOPQR = _mm_add_epi32(vbias, _mm_mullo_epi32(vaOPQR, va_multiplier));
+    __m128i vaccSTUV = _mm_add_epi32(vbias, _mm_mullo_epi32(vaSTUV, va_multiplier));
 
     const __m128i vrem0123 = _mm_add_epi32(_mm_and_si128(vacc0123, vremainder_mask), _mm_cmpgt_epi32(_mm_setzero_si128(), vacc0123));
     const __m128i vrem4567 = _mm_add_epi32(_mm_and_si128(vacc4567, vremainder_mask), _mm_cmpgt_epi32(_mm_setzero_si128(), vacc4567));
@@ -100,8 +100,8 @@
       const __m128i va4567 = _mm_cvtepi8_epi32(_mm_loadu_si32(input_a + 4));
       input_a += 8;
 
-      __m128i vacc0123 = _mm_add_epi32(vzero_point_product, _mm_mullo_epi32(va0123, va_multiplier));
-      __m128i vacc4567 = _mm_add_epi32(vzero_point_product, _mm_mullo_epi32(va4567, va_multiplier));
+      __m128i vacc0123 = _mm_add_epi32(vbias, _mm_mullo_epi32(va0123, va_multiplier));
+      __m128i vacc4567 = _mm_add_epi32(vbias, _mm_mullo_epi32(va4567, va_multiplier));
 
       const __m128i vrem0123 = _mm_add_epi32(_mm_and_si128(vacc0123, vremainder_mask), _mm_cmpgt_epi32(_mm_setzero_si128(), vacc0123));
       const __m128i vrem4567 = _mm_add_epi32(_mm_and_si128(vacc4567, vremainder_mask), _mm_cmpgt_epi32(_mm_setzero_si128(), vacc4567));