Optimize QS8 VADD[C] microkernels for SSE4/AVX/XOP/AVX2

Use microkernel-specific parameters to remove some operations

PiperOrigin-RevId: 385914437
diff --git a/src/qs8-vadd/gen/minmax-avx-mul32-ld32-x32.c b/src/qs8-vadd/gen/minmax-avx-mul32-ld32-x32.c
index 5303bef..994ffb1 100644
--- a/src/qs8-vadd/gen/minmax-avx-mul32-ld32-x32.c
+++ b/src/qs8-vadd/gen/minmax-avx-mul32-ld32-x32.c
@@ -22,14 +22,14 @@
     int8_t* output,
     const union xnn_qs8_add_minmax_params params[restrict XNN_MIN_ELEMENTS(1)]) XNN_DISABLE_TSAN
 {
-  const __m128i vbias = _mm_load_si128((const __m128i*) params->sse2.bias);
-  const __m128i va_multiplier = _mm_load_si128((const __m128i*) params->sse2.a_multiplier);
-  const __m128i vb_multiplier = _mm_load_si128((const __m128i*) params->sse2.b_multiplier);
-  const __m128i vrounding = _mm_load_si128((const __m128i*) params->sse2.rounding);
-  const __m128i vshift = _mm_cvtsi32_si128((int) params->sse2.shift);
-  const __m128i voutput_zero_point = _mm_load_si128((const __m128i*) params->sse2.output_zero_point);
-  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);
+  const __m128i vbias = _mm_load_si128((const __m128i*) params->sse4_mul32.bias);
+  const __m128i va_multiplier = _mm_load_si128((const __m128i*) params->sse4_mul32.a_multiplier);
+  const __m128i vb_multiplier = _mm_load_si128((const __m128i*) params->sse4_mul32.b_multiplier);
+  const __m128i vrounding = _mm_load_si128((const __m128i*) params->sse4_mul32.rounding);
+  const __m128i vshift = _mm_loadu_si32(params->sse4_mul32.shift);
+  const __m128i voutput_zero_point = _mm_load_si128((const __m128i*) params->sse4_mul32.output_zero_point);
+  const __m128i voutput_min = _mm_load_si128((const __m128i*) params->sse4_mul32.output_min);
+  const __m128i voutput_max = _mm_load_si128((const __m128i*) params->sse4_mul32.output_max);
 
   for (; n >= 32 * sizeof(int8_t); n -= 32 * sizeof(int8_t)) {
     const __m128i va0123 = _mm_cvtepi8_epi32(_mm_loadu_si32(input_a));
@@ -78,23 +78,19 @@
     vaccOPQR = _mm_sra_epi32(_mm_add_epi32(vaccOPQR, vrounding), vshift);
     vaccSTUV = _mm_sra_epi32(_mm_add_epi32(vaccSTUV, vrounding), vshift);
 
-    __m128i vout01234567 = _mm_adds_epi16(_mm_packs_epi32(vacc0123, vacc4567), voutput_zero_point);
-    __m128i vout89ABCDEF = _mm_adds_epi16(_mm_packs_epi32(vacc89AB, vaccCDEF), voutput_zero_point);
-    __m128i voutGHIJKLMN = _mm_adds_epi16(_mm_packs_epi32(vaccGHIJ, vaccKLMN), voutput_zero_point);
-    __m128i voutOPQRSTUV = _mm_adds_epi16(_mm_packs_epi32(vaccOPQR, vaccSTUV), voutput_zero_point);
+    const __m128i vout01234567 = _mm_adds_epi16(_mm_packs_epi32(vacc0123, vacc4567), voutput_zero_point);
+    const __m128i vout89ABCDEF = _mm_adds_epi16(_mm_packs_epi32(vacc89AB, vaccCDEF), voutput_zero_point);
+    const __m128i voutGHIJKLMN = _mm_adds_epi16(_mm_packs_epi32(vaccGHIJ, vaccKLMN), voutput_zero_point);
+    const __m128i voutOPQRSTUV = _mm_adds_epi16(_mm_packs_epi32(vaccOPQR, vaccSTUV), voutput_zero_point);
 
-    vout01234567 = _mm_max_epi16(vout01234567, voutput_min);
-    vout89ABCDEF = _mm_max_epi16(vout89ABCDEF, voutput_min);
-    voutGHIJKLMN = _mm_max_epi16(voutGHIJKLMN, voutput_min);
-    voutOPQRSTUV = _mm_max_epi16(voutOPQRSTUV, voutput_min);
+    __m128i vout0123456789ABCDEF = _mm_packs_epi16(vout01234567, vout89ABCDEF);
+    __m128i voutGHIJKLMNOPQRSTUV = _mm_packs_epi16(voutGHIJKLMN, voutOPQRSTUV);
 
-    vout01234567 = _mm_min_epi16(vout01234567, voutput_max);
-    vout89ABCDEF = _mm_min_epi16(vout89ABCDEF, voutput_max);
-    voutGHIJKLMN = _mm_min_epi16(voutGHIJKLMN, voutput_max);
-    voutOPQRSTUV = _mm_min_epi16(voutOPQRSTUV, voutput_max);
+    vout0123456789ABCDEF = _mm_max_epi8(vout0123456789ABCDEF, voutput_min);
+    voutGHIJKLMNOPQRSTUV = _mm_max_epi8(voutGHIJKLMNOPQRSTUV, voutput_min);
 
-    const __m128i vout0123456789ABCDEF = _mm_packs_epi16(vout01234567, vout89ABCDEF);
-    const __m128i voutGHIJKLMNOPQRSTUV = _mm_packs_epi16(voutGHIJKLMN, voutOPQRSTUV);
+    vout0123456789ABCDEF = _mm_min_epi8(vout0123456789ABCDEF, voutput_max);
+    voutGHIJKLMNOPQRSTUV = _mm_min_epi8(voutGHIJKLMNOPQRSTUV, voutput_max);
 
     _mm_storeu_si128((__m128i*) output, vout0123456789ABCDEF);
     _mm_storeu_si128((__m128i*) (output + 16), voutGHIJKLMNOPQRSTUV);
@@ -118,11 +114,11 @@
       vacc0123 = _mm_sra_epi32(_mm_add_epi32(vacc0123, vrounding), vshift);
       vacc4567 = _mm_sra_epi32(_mm_add_epi32(vacc4567, vrounding), vshift);
 
-      __m128i vout01234567 = _mm_adds_epi16(_mm_packs_epi32(vacc0123, vacc4567), voutput_zero_point);
-      vout01234567 = _mm_max_epi16(vout01234567, voutput_min);
-      vout01234567 = _mm_min_epi16(vout01234567, voutput_max);
+      const __m128i vout01234567 = _mm_adds_epi16(_mm_packs_epi32(vacc0123, vacc4567), voutput_zero_point);
 
       __m128i vout0123456701234567 = _mm_packs_epi16(vout01234567, vout01234567);
+      vout0123456701234567 = _mm_max_epi8(vout0123456701234567, voutput_min);
+      vout0123456701234567 = _mm_min_epi8(vout0123456701234567, voutput_max);
 
       if XNN_LIKELY(n >= (8 * sizeof(int8_t))) {
         _mm_storel_epi64((__m128i*) output, vout0123456701234567);