QS8 VADDC microkernels in SSE2 and SSE4.1 implementations
PiperOrigin-RevId: 325972893
diff --git a/src/qs8-vaddc/gen/minmax-sse41-mul16-ld64-x32.c b/src/qs8-vaddc/gen/minmax-sse41-mul16-ld64-x32.c
new file mode 100644
index 0000000..6d056b8
--- /dev/null
+++ b/src/qs8-vaddc/gen/minmax-sse41-mul16-ld64-x32.c
@@ -0,0 +1,163 @@
+// Auto-generated file. Do not edit!
+// Template: src/qs8-vaddc/sse-mul16-ld64.c.in
+// Generator: tools/xngen
+//
+// Copyright 2020 Google LLC
+//
+// This source code is licensed under the BSD-style license found in the
+// LICENSE file in the root directory of this source tree.
+
+#include <assert.h>
+
+#include <smmintrin.h>
+
+#include <xnnpack/vadd.h>
+
+
+void xnn_qs8_vaddc_minmax_ukernel__sse41_mul16_ld64_x32(
+ size_t n,
+ const int8_t* input_x,
+ const int8_t* input_y,
+ int8_t* output,
+ const union xnn_qs8_add_params params[restrict XNN_MIN_ELEMENTS(1)]) XNN_DISABLE_TSAN
+{
+ const __m128i vx_multiplier_lo = _mm_load_si128((const __m128i*) ¶ms->sse2.x_multiplier_lo);
+ const __m128i vx_multiplier_hi = _mm_load_si128((const __m128i*) ¶ms->sse2.x_multiplier_hi);
+ const __m128i vremainder_mask = _mm_load_si128((const __m128i*) params->sse2.remainder_mask);
+ const __m128i vremainder_threshold = _mm_load_si128((const __m128i*) params->sse2.remainder_threshold);
+ 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);
+
+ __m128i vzero_point_product = _mm_cvtsi32_si128(params->sse2.y_multiplier * (int32_t) *input_y);
+ 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*) ¶ms->sse2.zero_point_product));
+ for (; n >= 32 * sizeof(int8_t); n -= 32 * sizeof(int8_t)) {
+ const __m128i vx01234567 = _mm_cvtepi8_epi16(_mm_loadl_epi64((const __m128i*) input_x));
+ const __m128i vx89ABCDEF = _mm_cvtepi8_epi16(_mm_loadl_epi64((const __m128i*) (input_x + 8)));
+ const __m128i vxGHIJKLMN = _mm_cvtepi8_epi16(_mm_loadl_epi64((const __m128i*) (input_x + 16)));
+ const __m128i vxOPQRSTUV = _mm_cvtepi8_epi16(_mm_loadl_epi64((const __m128i*) (input_x + 24)));
+ input_x += 32;
+
+
+ __m128i vxprod01234567hi = _mm_mulhi_epu16(vx01234567, vx_multiplier_lo);
+ const __m128i vxprod01234567lo = _mm_mullo_epi16(vx01234567, vx_multiplier_lo);
+ __m128i vxprod89ABCDEFhi = _mm_mulhi_epu16(vx89ABCDEF, vx_multiplier_lo);
+ const __m128i vxprod89ABCDEFlo = _mm_mullo_epi16(vx89ABCDEF, vx_multiplier_lo);
+ __m128i vxprodGHIJKLMNhi = _mm_mulhi_epu16(vxGHIJKLMN, vx_multiplier_lo);
+ const __m128i vxprodGHIJKLMNlo = _mm_mullo_epi16(vxGHIJKLMN, vx_multiplier_lo);
+ __m128i vxprodOPQRSTUVhi = _mm_mulhi_epu16(vxOPQRSTUV, vx_multiplier_lo);
+ const __m128i vxprodOPQRSTUVlo = _mm_mullo_epi16(vxOPQRSTUV, vx_multiplier_lo);
+
+ vxprod01234567hi = _mm_add_epi16(vxprod01234567hi, _mm_mullo_epi16(vx01234567, vx_multiplier_hi));
+ vxprod89ABCDEFhi = _mm_add_epi16(vxprod89ABCDEFhi, _mm_mullo_epi16(vx89ABCDEF, vx_multiplier_hi));
+ vxprodGHIJKLMNhi = _mm_add_epi16(vxprodGHIJKLMNhi, _mm_mullo_epi16(vxGHIJKLMN, vx_multiplier_hi));
+ vxprodOPQRSTUVhi = _mm_add_epi16(vxprodOPQRSTUVhi, _mm_mullo_epi16(vxOPQRSTUV, vx_multiplier_hi));
+
+ vxprod01234567hi = _mm_sub_epi16(vxprod01234567hi, _mm_and_si128(_mm_srai_epi16(vx01234567, 15), vx_multiplier_lo));
+ vxprod89ABCDEFhi = _mm_sub_epi16(vxprod89ABCDEFhi, _mm_and_si128(_mm_srai_epi16(vx89ABCDEF, 15), vx_multiplier_lo));
+ vxprodGHIJKLMNhi = _mm_sub_epi16(vxprodGHIJKLMNhi, _mm_and_si128(_mm_srai_epi16(vxGHIJKLMN, 15), vx_multiplier_lo));
+ vxprodOPQRSTUVhi = _mm_sub_epi16(vxprodOPQRSTUVhi, _mm_and_si128(_mm_srai_epi16(vxOPQRSTUV, 15), vx_multiplier_lo));
+
+ __m128i vacc0123 = _mm_add_epi32(vzero_point_product, _mm_unpacklo_epi16(vxprod01234567lo, vxprod01234567hi));
+ __m128i vacc4567 = _mm_add_epi32(vzero_point_product, _mm_unpackhi_epi16(vxprod01234567lo, vxprod01234567hi));
+ __m128i vacc89AB = _mm_add_epi32(vzero_point_product, _mm_unpacklo_epi16(vxprod89ABCDEFlo, vxprod89ABCDEFhi));
+ __m128i vaccCDEF = _mm_add_epi32(vzero_point_product, _mm_unpackhi_epi16(vxprod89ABCDEFlo, vxprod89ABCDEFhi));
+ __m128i vaccGHIJ = _mm_add_epi32(vzero_point_product, _mm_unpacklo_epi16(vxprodGHIJKLMNlo, vxprodGHIJKLMNhi));
+ __m128i vaccKLMN = _mm_add_epi32(vzero_point_product, _mm_unpackhi_epi16(vxprodGHIJKLMNlo, vxprodGHIJKLMNhi));
+ __m128i vaccOPQR = _mm_add_epi32(vzero_point_product, _mm_unpacklo_epi16(vxprodOPQRSTUVlo, vxprodOPQRSTUVhi));
+ __m128i vaccSTUV = _mm_add_epi32(vzero_point_product, _mm_unpackhi_epi16(vxprodOPQRSTUVlo, vxprodOPQRSTUVhi));
+
+ 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));
+ const __m128i vrem89AB = _mm_add_epi32(_mm_and_si128(vacc89AB, vremainder_mask), _mm_cmpgt_epi32(_mm_setzero_si128(), vacc89AB));
+ const __m128i vremCDEF = _mm_add_epi32(_mm_and_si128(vaccCDEF, vremainder_mask), _mm_cmpgt_epi32(_mm_setzero_si128(), vaccCDEF));
+ const __m128i vremGHIJ = _mm_add_epi32(_mm_and_si128(vaccGHIJ, vremainder_mask), _mm_cmpgt_epi32(_mm_setzero_si128(), vaccGHIJ));
+ const __m128i vremKLMN = _mm_add_epi32(_mm_and_si128(vaccKLMN, vremainder_mask), _mm_cmpgt_epi32(_mm_setzero_si128(), vaccKLMN));
+ const __m128i vremOPQR = _mm_add_epi32(_mm_and_si128(vaccOPQR, vremainder_mask), _mm_cmpgt_epi32(_mm_setzero_si128(), vaccOPQR));
+ const __m128i vremSTUV = _mm_add_epi32(_mm_and_si128(vaccSTUV, vremainder_mask), _mm_cmpgt_epi32(_mm_setzero_si128(), vaccSTUV));
+
+ vacc0123 = _mm_sub_epi32(_mm_sra_epi32(vacc0123, vshift), _mm_cmpgt_epi32(vrem0123, vremainder_threshold));
+ vacc4567 = _mm_sub_epi32(_mm_sra_epi32(vacc4567, vshift), _mm_cmpgt_epi32(vrem4567, vremainder_threshold));
+ vacc89AB = _mm_sub_epi32(_mm_sra_epi32(vacc89AB, vshift), _mm_cmpgt_epi32(vrem89AB, vremainder_threshold));
+ vaccCDEF = _mm_sub_epi32(_mm_sra_epi32(vaccCDEF, vshift), _mm_cmpgt_epi32(vremCDEF, vremainder_threshold));
+ vaccGHIJ = _mm_sub_epi32(_mm_sra_epi32(vaccGHIJ, vshift), _mm_cmpgt_epi32(vremGHIJ, vremainder_threshold));
+ vaccKLMN = _mm_sub_epi32(_mm_sra_epi32(vaccKLMN, vshift), _mm_cmpgt_epi32(vremKLMN, vremainder_threshold));
+ vaccOPQR = _mm_sub_epi32(_mm_sra_epi32(vaccOPQR, vshift), _mm_cmpgt_epi32(vremOPQR, vremainder_threshold));
+ vaccSTUV = _mm_sub_epi32(_mm_sra_epi32(vaccSTUV, vshift), _mm_cmpgt_epi32(vremSTUV, vremainder_threshold));
+
+ __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);
+
+ 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);
+
+ 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);
+
+ const __m128i vout0123456789ABCDEF = _mm_packs_epi16(vout01234567, vout89ABCDEF);
+ const __m128i voutGHIJKLMNOPQRSTUV = _mm_packs_epi16(voutGHIJKLMN, voutOPQRSTUV);
+
+ _mm_storeu_si128((__m128i*) output, vout0123456789ABCDEF);
+ _mm_storeu_si128((__m128i*) (output + 16), voutGHIJKLMNOPQRSTUV);
+ output += 32;
+ }
+ if XNN_UNLIKELY(n != 0) {
+ do {
+ const __m128i vx01234567 = _mm_cvtepi8_epi16(_mm_loadl_epi64((const __m128i*) input_x));
+ input_x += 8;
+
+
+ __m128i vxprod01234567hi = _mm_mulhi_epu16(vx01234567, vx_multiplier_lo);
+ const __m128i vxprod01234567lo = _mm_mullo_epi16(vx01234567, vx_multiplier_lo);
+
+ vxprod01234567hi = _mm_add_epi16(vxprod01234567hi, _mm_mullo_epi16(vx01234567, vx_multiplier_hi));
+
+ vxprod01234567hi = _mm_sub_epi16(vxprod01234567hi, _mm_and_si128(_mm_srai_epi16(vx01234567, 15), vx_multiplier_lo));
+
+ __m128i vacc0123 = _mm_add_epi32(vzero_point_product, _mm_unpacklo_epi16(vxprod01234567lo, vxprod01234567hi));
+ __m128i vacc4567 = _mm_add_epi32(vzero_point_product, _mm_unpackhi_epi16(vxprod01234567lo, vxprod01234567hi));
+
+ 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));
+
+ vacc0123 = _mm_sub_epi32(_mm_sra_epi32(vacc0123, vshift), _mm_cmpgt_epi32(vrem0123, vremainder_threshold));
+ vacc4567 = _mm_sub_epi32(_mm_sra_epi32(vacc4567, vshift), _mm_cmpgt_epi32(vrem4567, vremainder_threshold));
+
+ __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);
+
+ __m128i vout0123456701234567 = _mm_packs_epi16(vout01234567, vout01234567);
+
+ if XNN_LIKELY(n >= (8 * sizeof(int8_t))) {
+ _mm_storel_epi64((__m128i*) output, vout0123456701234567);
+ output += 8;
+ n -= 8;
+ } else {
+ if (n & (4 * sizeof(int8_t))) {
+ *((uint32_t*) output) = (uint32_t) _mm_cvtsi128_si32(vout0123456701234567);
+ vout0123456701234567 = _mm_srli_epi64(vout0123456701234567, 32);
+ output += 4;
+ }
+ if (n & (2 * sizeof(int8_t))) {
+ *((uint16_t*) output) = (uint16_t) _mm_extract_epi16(vout0123456701234567, 0);
+ vout0123456701234567 = _mm_srli_epi32(vout0123456701234567, 16);
+ output += 2;
+ }
+ if (n & (1 * sizeof(int8_t))) {
+ *output = (int8_t) _mm_extract_epi8(vout0123456701234567, 0);
+ output += 1;
+ }
+ n = 0;
+ }
+ } while (n != 0);
+ }
+}