AVX2 versions of QS8 VADD[C] microkernels

PiperOrigin-RevId: 330666026

5 files changed, 537 insertions, 0 deletions

diff --git a/src/qs8-vaddc/avx2-mul32-ld64.c.in b/src/qs8-vaddc/avx2-mul32-ld64.c.in
new file mode 100644
index 000000000..6b29b59eb
--- /dev/null
+++ b/src/qs8-vaddc/avx2-mul32-ld64.c.in
@@ -0,0 +1,145 @@
+// 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.
+
+$assert BATCH_TILE % 8 == 0
+$assert BATCH_TILE >= 8
+$ABC = "0123456789ABCDEFGHIJKLMNOPQRSTUVWXYZ"
+#include <assert.h>
+
+#include <immintrin.h>
+
+#include <xnnpack/intrinsics-polyfill.h>
+#include <xnnpack/vadd.h>
+
+
+void xnn_qs8_vaddc_minmax_ukernel__avx2_mul32_ld64_x${BATCH_TILE}(
+    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 __m256i vx_multiplier = _mm256_broadcastsi128_si256(_mm_load_si128((const __m128i*) params->sse2.x_multiplier));
+  const __m256i vremainder_mask = _mm256_broadcastsi128_si256(_mm_load_si128((const __m128i*) params->sse2.remainder_mask));
+  const __m256i vremainder_threshold = _mm256_broadcastsi128_si256(_mm_load_si128((const __m128i*) params->sse2.remainder_threshold));
+  const __m128i vshift = _mm_cvtsi32_si128((int) params->sse2.shift);
+  $if BATCH_TILE > 8:
+    const __m256i voutput_zero_point = _mm256_broadcastsi128_si256(_mm_load_si128((const __m128i*) params->sse2.output_zero_point));
+    const __m256i voutput_min = _mm256_broadcastsi128_si256(_mm_load_si128((const __m128i*) params->sse2.output_min));
+    const __m256i voutput_max = _mm256_broadcastsi128_si256(_mm_load_si128((const __m128i*) params->sse2.output_max));
+  $else:
+    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);
+
+  __m256i vzero_point_product = _mm256_broadcastsi128_si256(_mm_add_epi32(
+    _mm_broadcastd_epi32(_mm_cvtsi32_si128(params->sse2.y_multiplier[0] * (int32_t) *input_y)),
+    _mm_load_si128((const __m128i*) params->sse2.zero_point_product)));
+  for (; n >= ${BATCH_TILE} * sizeof(int8_t); n -= ${BATCH_TILE} * sizeof(int8_t)) {
+    const __m256i vx${ABC[0:8]} = _mm256_cvtepi8_epi32(_mm_loadl_epi64((const __m128i*) input_x));
+    $for N in range(8, BATCH_TILE, 8):
+      const __m256i vx${ABC[N:N+8]} = _mm256_cvtepi8_epi32(_mm_loadl_epi64((const __m128i*) (input_x + ${N})));
+    input_x += ${BATCH_TILE};
+
+    $for N in range(0, BATCH_TILE, 8):
+      __m256i vacc${ABC[N:N+8]} = _mm256_add_epi32(vzero_point_product, _mm256_mullo_epi32(vx${ABC[N:N+8]}, vx_multiplier));
+
+    $for N in range(0, BATCH_TILE, 8):
+      const __m256i vrem${ABC[N:N+8]} = _mm256_add_epi32(_mm256_and_si256(vacc${ABC[N:N+8]}, vremainder_mask), _mm256_srai_epi32(vacc${ABC[N:N+8]}, 31));
+
+    $for N in range(0, BATCH_TILE, 8):
+      vacc${ABC[N:N+8]} = _mm256_sub_epi32(_mm256_sra_epi32(vacc${ABC[N:N+8]}, vshift), _mm256_cmpgt_epi32(vrem${ABC[N:N+8]}, vremainder_threshold));
+
+    $for N in range(0, BATCH_TILE, 16):
+      $if N + 8 < BATCH_TILE:
+        __m256i vout${ABC[N:N+4]}${ABC[N+8:N+12]}${ABC[N+4:N+8]}${ABC[N+12:N+16]} = _mm256_adds_epi16(_mm256_packs_epi32(vacc${ABC[N:N+8]}, vacc${ABC[N+8:N+16]}), voutput_zero_point);
+      $elif BATCH_TILE > 8:
+        __m128i vout${ABC[N:N+8]} = _mm_adds_epi16(_mm_packs_epi32(_mm256_castsi256_si128(vacc${ABC[N:N+8]}), _mm256_extracti128_si256(vacc${ABC[N:N+8]}, 1)), _mm256_castsi256_si128(voutput_zero_point));
+      $else:
+        __m128i vout${ABC[N:N+8]} = _mm_adds_epi16(_mm_packs_epi32(_mm256_castsi256_si128(vacc${ABC[N:N+8]}), _mm256_extracti128_si256(vacc${ABC[N:N+8]}, 1)), voutput_zero_point);
+
+    $for N in range(0, BATCH_TILE, 16):
+      $if N + 8 < BATCH_TILE:
+        vout${ABC[N:N+4]}${ABC[N+8:N+12]}${ABC[N+4:N+8]}${ABC[N+12:N+16]} = _mm256_min_epi16(_mm256_max_epi16(vout${ABC[N:N+4]}${ABC[N+8:N+12]}${ABC[N+4:N+8]}${ABC[N+12:N+16]}, voutput_min), voutput_max);
+      $elif BATCH_TILE > 8:
+        vout${ABC[N:N+8]} = _mm_min_epi16(_mm_max_epi16(vout${ABC[N:N+8]}, _mm256_castsi256_si128(voutput_min)), _mm256_castsi256_si128(voutput_max));
+      $else:
+        vout${ABC[N:N+8]} = _mm_min_epi16(_mm_max_epi16(vout${ABC[N:N+8]}, voutput_min), voutput_max);
+
+    $for N in range(0, BATCH_TILE, 16):
+      $if N + 8 < BATCH_TILE:
+        __m128i vout${ABC[N:N+16]} = _mm_shuffle_epi32(_mm_packs_epi16(_mm256_castsi256_si128(vout${ABC[N:N+4]}${ABC[N+8:N+12]}${ABC[N+4:N+8]}${ABC[N+12:N+16]}), _mm256_extracti128_si256(vout${ABC[N:N+4]}${ABC[N+8:N+12]}${ABC[N+4:N+8]}${ABC[N+12:N+16]}, 1)), _MM_SHUFFLE(3, 1, 2, 0));
+      $else:
+        __m128i vout${ABC[N:N+8]}${ABC[N:N+8]} = _mm_packs_epi16(vout${ABC[N:N+8]}, vout${ABC[N:N+8]});
+
+    $if BATCH_TILE >= 16:
+      _mm_storeu_si128((__m128i*) output, vout${ABC[0:16]});
+    $else:
+      _mm_storel_epi64((__m128i*) output, vout${ABC[0:8]}${ABC[0:8]});
+    $for N in range(16, BATCH_TILE, 16):
+      $if N + 8 < BATCH_TILE:
+        _mm_storeu_si128((__m128i*) (output + ${N}), vout${ABC[N:N+16]});
+      $else:
+        _mm_storel_epi64((__m128i*) (output + ${N}), vout${ABC[N:N+8]}${ABC[N:N+8]});
+    output += ${BATCH_TILE};
+  }
+  if XNN_UNLIKELY(n != 0) {
+    ${"do " if BATCH_TILE > 8 else ""}{
+      const __m256i vx${ABC[0:8]} = _mm256_cvtepi8_epi32(_mm_loadl_epi64((const __m128i*) input_x));
+      $if BATCH_TILE > 8:
+        input_x += 8;
+
+      __m256i vacc${ABC[0:8]} = _mm256_add_epi32(vzero_point_product, _mm256_mullo_epi32(vx${ABC[0:8]}, vx_multiplier));
+
+      const __m256i vrem${ABC[0:8]} = _mm256_add_epi32(_mm256_and_si256(vacc${ABC[0:8]}, vremainder_mask), _mm256_srai_epi32(vacc${ABC[0:8]}, 31));
+
+      vacc${ABC[0:8]} = _mm256_sub_epi32(_mm256_sra_epi32(vacc${ABC[0:8]}, vshift), _mm256_cmpgt_epi32(vrem${ABC[0:8]}, vremainder_threshold));
+
+      $if BATCH_TILE > 8:
+        __m128i vout${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));
+        vout${ABC[0:8]} = _mm_min_epi16(_mm_max_epi16(vout${ABC[0:8]}, _mm256_castsi256_si128(voutput_min)), _mm256_castsi256_si128(voutput_max));
+      $else:
+        __m128i vout${ABC[0:8]} = _mm_adds_epi16(_mm_packs_epi32(_mm256_castsi256_si128(vacc${ABC[0:8]}), _mm256_extracti128_si256(vacc${ABC[0:8]}, 1)), voutput_zero_point);
+        vout${ABC[0:8]} = _mm_min_epi16(_mm_max_epi16(vout${ABC[0:8]}, voutput_min), voutput_max);
+      __m128i vout${ABC[0:8]}${ABC[0:8]} = _mm_packs_epi16(vout${ABC[0:8]}, vout${ABC[0:8]});
+
+      $if BATCH_TILE > 8:
+        if XNN_LIKELY(n >= (8 * sizeof(int8_t))) {
+          _mm_storel_epi64((__m128i*) output, vout${ABC[0:8]}${ABC[0:8]});
+          output += 8;
+          n -= 8 * sizeof(int8_t);
+        } else {
+          if (n & (4 * sizeof(int8_t))) {
+            *((uint32_t*) output) = (uint32_t) _mm_cvtsi128_si32(vout${ABC[0:8]}${ABC[0:8]});
+            vout${ABC[0:8]}${ABC[0:8]} = _mm_srli_epi64(vout${ABC[0:8]}${ABC[0:8]}, 32);
+            output += 4;
+          }
+          if (n & (2 * sizeof(int8_t))) {
+            *((uint16_t*) output) = (uint16_t) _mm_extract_epi16(vout${ABC[0:8]}${ABC[0:8]}, 0);
+            vout${ABC[0:8]}${ABC[0:8]} = _mm_srli_epi32(vout${ABC[0:8]}${ABC[0:8]}, 16);
+            output += 2;
+          }
+          if (n & (1 * sizeof(int8_t))) {
+            *output = (int8_t) _mm_extract_epi8(vout${ABC[0:8]}${ABC[0:8]}, 0);
+          }
+          n = 0;
+        }
+      $else:
+        if (n & (4 * sizeof(int8_t))) {
+          *((uint32_t*) output) = (uint32_t) _mm_cvtsi128_si32(vout${ABC[0:8]}${ABC[0:8]});
+          vout${ABC[0:8]}${ABC[0:8]} = _mm_srli_epi64(vout${ABC[0:8]}${ABC[0:8]}, 32);
+          output += 4;
+        }
+        if (n & (2 * sizeof(int8_t))) {
+          *((uint16_t*) output) = (uint16_t) _mm_extract_epi16(vout${ABC[0:8]}${ABC[0:8]}, 0);
+          vout${ABC[0:8]}${ABC[0:8]} = _mm_srli_epi32(vout${ABC[0:8]}${ABC[0:8]}, 16);
+          output += 2;
+        }
+        if (n & (1 * sizeof(int8_t))) {
+          *output = (int8_t) _mm_extract_epi8(vout${ABC[0:8]}${ABC[0:8]}, 0);
+        }
+    }${" while (n != 0);" if BATCH_TILE > 8 else ""}
+  }
+}
diff --git a/src/qs8-vaddc/gen/minmax-avx2-mul32-ld64-x16.c b/src/qs8-vaddc/gen/minmax-avx2-mul32-ld64-x16.c
new file mode 100644
index 000000000..94cf51a03
--- /dev/null
+++ b/src/qs8-vaddc/gen/minmax-avx2-mul32-ld64-x16.c
@@ -0,0 +1,96 @@
+// Auto-generated file. Do not edit!
+//   Template: src/qs8-vaddc/avx2-mul32-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 <immintrin.h>
+
+#include <xnnpack/intrinsics-polyfill.h>
+#include <xnnpack/vadd.h>
+
+
+void xnn_qs8_vaddc_minmax_ukernel__avx2_mul32_ld64_x16(
+    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 __m256i vx_multiplier = _mm256_broadcastsi128_si256(_mm_load_si128((const __m128i*) params->sse2.x_multiplier));
+  const __m256i vremainder_mask = _mm256_broadcastsi128_si256(_mm_load_si128((const __m128i*) params->sse2.remainder_mask));
+  const __m256i vremainder_threshold = _mm256_broadcastsi128_si256(_mm_load_si128((const __m128i*) params->sse2.remainder_threshold));
+  const __m128i vshift = _mm_cvtsi32_si128((int) params->sse2.shift);
+  const __m256i voutput_zero_point = _mm256_broadcastsi128_si256(_mm_load_si128((const __m128i*) params->sse2.output_zero_point));
+  const __m256i voutput_min = _mm256_broadcastsi128_si256(_mm_load_si128((const __m128i*) params->sse2.output_min));
+  const __m256i voutput_max = _mm256_broadcastsi128_si256(_mm_load_si128((const __m128i*) params->sse2.output_max));
+
+  __m256i vzero_point_product = _mm256_broadcastsi128_si256(_mm_add_epi32(
+    _mm_broadcastd_epi32(_mm_cvtsi32_si128(params->sse2.y_multiplier[0] * (int32_t) *input_y)),
+    _mm_load_si128((const __m128i*) params->sse2.zero_point_product)));
+  for (; n >= 16 * sizeof(int8_t); n -= 16 * sizeof(int8_t)) {
+    const __m256i vx01234567 = _mm256_cvtepi8_epi32(_mm_loadl_epi64((const __m128i*) input_x));
+    const __m256i vx89ABCDEF = _mm256_cvtepi8_epi32(_mm_loadl_epi64((const __m128i*) (input_x + 8)));
+    input_x += 16;
+
+    __m256i vacc01234567 = _mm256_add_epi32(vzero_point_product, _mm256_mullo_epi32(vx01234567, vx_multiplier));
+    __m256i vacc89ABCDEF = _mm256_add_epi32(vzero_point_product, _mm256_mullo_epi32(vx89ABCDEF, vx_multiplier));
+
+    const __m256i vrem01234567 = _mm256_add_epi32(_mm256_and_si256(vacc01234567, vremainder_mask), _mm256_srai_epi32(vacc01234567, 31));
+    const __m256i vrem89ABCDEF = _mm256_add_epi32(_mm256_and_si256(vacc89ABCDEF, vremainder_mask), _mm256_srai_epi32(vacc89ABCDEF, 31));
+
+    vacc01234567 = _mm256_sub_epi32(_mm256_sra_epi32(vacc01234567, vshift), _mm256_cmpgt_epi32(vrem01234567, vremainder_threshold));
+    vacc89ABCDEF = _mm256_sub_epi32(_mm256_sra_epi32(vacc89ABCDEF, vshift), _mm256_cmpgt_epi32(vrem89ABCDEF, vremainder_threshold));
+
+    __m256i vout012389AB4567CDEF = _mm256_adds_epi16(_mm256_packs_epi32(vacc01234567, vacc89ABCDEF), voutput_zero_point);
+
+    vout012389AB4567CDEF = _mm256_min_epi16(_mm256_max_epi16(vout012389AB4567CDEF, voutput_min), voutput_max);
+
+    __m128i vout0123456789ABCDEF = _mm_shuffle_epi32(_mm_packs_epi16(_mm256_castsi256_si128(vout012389AB4567CDEF), _mm256_extracti128_si256(vout012389AB4567CDEF, 1)), _MM_SHUFFLE(3, 1, 2, 0));
+
+    _mm_storeu_si128((__m128i*) output, vout0123456789ABCDEF);
+    output += 16;
+  }
+  if XNN_UNLIKELY(n != 0) {
+    do {
+      const __m256i vx01234567 = _mm256_cvtepi8_epi32(_mm_loadl_epi64((const __m128i*) input_x));
+      input_x += 8;
+
+      __m256i vacc01234567 = _mm256_add_epi32(vzero_point_product, _mm256_mullo_epi32(vx01234567, vx_multiplier));
+
+      const __m256i vrem01234567 = _mm256_add_epi32(_mm256_and_si256(vacc01234567, vremainder_mask), _mm256_srai_epi32(vacc01234567, 31));
+
+      vacc01234567 = _mm256_sub_epi32(_mm256_sra_epi32(vacc01234567, vshift), _mm256_cmpgt_epi32(vrem01234567, vremainder_threshold));
+
+      __m128i vout01234567 = _mm_adds_epi16(_mm_packs_epi32(_mm256_castsi256_si128(vacc01234567), _mm256_extracti128_si256(vacc01234567, 1)), _mm256_castsi256_si128(voutput_zero_point));
+      vout01234567 = _mm_min_epi16(_mm_max_epi16(vout01234567, _mm256_castsi256_si128(voutput_min)), _mm256_castsi256_si128(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 * sizeof(int8_t);
+      } 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);
+        }
+        n = 0;
+      }
+    } while (n != 0);
+  }
+}
diff --git a/src/qs8-vaddc/gen/minmax-avx2-mul32-ld64-x24.c b/src/qs8-vaddc/gen/minmax-avx2-mul32-ld64-x24.c
new file mode 100644
index 000000000..729963e8e
--- /dev/null
+++ b/src/qs8-vaddc/gen/minmax-avx2-mul32-ld64-x24.c
@@ -0,0 +1,104 @@
+// Auto-generated file. Do not edit!
+//   Template: src/qs8-vaddc/avx2-mul32-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 <immintrin.h>
+
+#include <xnnpack/intrinsics-polyfill.h>
+#include <xnnpack/vadd.h>
+
+
+void xnn_qs8_vaddc_minmax_ukernel__avx2_mul32_ld64_x24(
+    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 __m256i vx_multiplier = _mm256_broadcastsi128_si256(_mm_load_si128((const __m128i*) params->sse2.x_multiplier));
+  const __m256i vremainder_mask = _mm256_broadcastsi128_si256(_mm_load_si128((const __m128i*) params->sse2.remainder_mask));
+  const __m256i vremainder_threshold = _mm256_broadcastsi128_si256(_mm_load_si128((const __m128i*) params->sse2.remainder_threshold));
+  const __m128i vshift = _mm_cvtsi32_si128((int) params->sse2.shift);
+  const __m256i voutput_zero_point = _mm256_broadcastsi128_si256(_mm_load_si128((const __m128i*) params->sse2.output_zero_point));
+  const __m256i voutput_min = _mm256_broadcastsi128_si256(_mm_load_si128((const __m128i*) params->sse2.output_min));
+  const __m256i voutput_max = _mm256_broadcastsi128_si256(_mm_load_si128((const __m128i*) params->sse2.output_max));
+
+  __m256i vzero_point_product = _mm256_broadcastsi128_si256(_mm_add_epi32(
+    _mm_broadcastd_epi32(_mm_cvtsi32_si128(params->sse2.y_multiplier[0] * (int32_t) *input_y)),
+    _mm_load_si128((const __m128i*) params->sse2.zero_point_product)));
+  for (; n >= 24 * sizeof(int8_t); n -= 24 * sizeof(int8_t)) {
+    const __m256i vx01234567 = _mm256_cvtepi8_epi32(_mm_loadl_epi64((const __m128i*) input_x));
+    const __m256i vx89ABCDEF = _mm256_cvtepi8_epi32(_mm_loadl_epi64((const __m128i*) (input_x + 8)));
+    const __m256i vxGHIJKLMN = _mm256_cvtepi8_epi32(_mm_loadl_epi64((const __m128i*) (input_x + 16)));
+    input_x += 24;
+
+    __m256i vacc01234567 = _mm256_add_epi32(vzero_point_product, _mm256_mullo_epi32(vx01234567, vx_multiplier));
+    __m256i vacc89ABCDEF = _mm256_add_epi32(vzero_point_product, _mm256_mullo_epi32(vx89ABCDEF, vx_multiplier));
+    __m256i vaccGHIJKLMN = _mm256_add_epi32(vzero_point_product, _mm256_mullo_epi32(vxGHIJKLMN, vx_multiplier));
+
+    const __m256i vrem01234567 = _mm256_add_epi32(_mm256_and_si256(vacc01234567, vremainder_mask), _mm256_srai_epi32(vacc01234567, 31));
+    const __m256i vrem89ABCDEF = _mm256_add_epi32(_mm256_and_si256(vacc89ABCDEF, vremainder_mask), _mm256_srai_epi32(vacc89ABCDEF, 31));
+    const __m256i vremGHIJKLMN = _mm256_add_epi32(_mm256_and_si256(vaccGHIJKLMN, vremainder_mask), _mm256_srai_epi32(vaccGHIJKLMN, 31));
+
+    vacc01234567 = _mm256_sub_epi32(_mm256_sra_epi32(vacc01234567, vshift), _mm256_cmpgt_epi32(vrem01234567, vremainder_threshold));
+    vacc89ABCDEF = _mm256_sub_epi32(_mm256_sra_epi32(vacc89ABCDEF, vshift), _mm256_cmpgt_epi32(vrem89ABCDEF, vremainder_threshold));
+    vaccGHIJKLMN = _mm256_sub_epi32(_mm256_sra_epi32(vaccGHIJKLMN, vshift), _mm256_cmpgt_epi32(vremGHIJKLMN, vremainder_threshold));
+
+    __m256i vout012389AB4567CDEF = _mm256_adds_epi16(_mm256_packs_epi32(vacc01234567, vacc89ABCDEF), voutput_zero_point);
+    __m128i voutGHIJKLMN = _mm_adds_epi16(_mm_packs_epi32(_mm256_castsi256_si128(vaccGHIJKLMN), _mm256_extracti128_si256(vaccGHIJKLMN, 1)), _mm256_castsi256_si128(voutput_zero_point));
+
+    vout012389AB4567CDEF = _mm256_min_epi16(_mm256_max_epi16(vout012389AB4567CDEF, voutput_min), voutput_max);
+    voutGHIJKLMN = _mm_min_epi16(_mm_max_epi16(voutGHIJKLMN, _mm256_castsi256_si128(voutput_min)), _mm256_castsi256_si128(voutput_max));
+
+    __m128i vout0123456789ABCDEF = _mm_shuffle_epi32(_mm_packs_epi16(_mm256_castsi256_si128(vout012389AB4567CDEF), _mm256_extracti128_si256(vout012389AB4567CDEF, 1)), _MM_SHUFFLE(3, 1, 2, 0));
+    __m128i voutGHIJKLMNGHIJKLMN = _mm_packs_epi16(voutGHIJKLMN, voutGHIJKLMN);
+
+    _mm_storeu_si128((__m128i*) output, vout0123456789ABCDEF);
+    _mm_storel_epi64((__m128i*) (output + 16), voutGHIJKLMNGHIJKLMN);
+    output += 24;
+  }
+  if XNN_UNLIKELY(n != 0) {
+    do {
+      const __m256i vx01234567 = _mm256_cvtepi8_epi32(_mm_loadl_epi64((const __m128i*) input_x));
+      input_x += 8;
+
+      __m256i vacc01234567 = _mm256_add_epi32(vzero_point_product, _mm256_mullo_epi32(vx01234567, vx_multiplier));
+
+      const __m256i vrem01234567 = _mm256_add_epi32(_mm256_and_si256(vacc01234567, vremainder_mask), _mm256_srai_epi32(vacc01234567, 31));
+
+      vacc01234567 = _mm256_sub_epi32(_mm256_sra_epi32(vacc01234567, vshift), _mm256_cmpgt_epi32(vrem01234567, vremainder_threshold));
+
+      __m128i vout01234567 = _mm_adds_epi16(_mm_packs_epi32(_mm256_castsi256_si128(vacc01234567), _mm256_extracti128_si256(vacc01234567, 1)), _mm256_castsi256_si128(voutput_zero_point));
+      vout01234567 = _mm_min_epi16(_mm_max_epi16(vout01234567, _mm256_castsi256_si128(voutput_min)), _mm256_castsi256_si128(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 * sizeof(int8_t);
+      } 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);
+        }
+        n = 0;
+      }
+    } while (n != 0);
+  }
+}
diff --git a/src/qs8-vaddc/gen/minmax-avx2-mul32-ld64-x32.c b/src/qs8-vaddc/gen/minmax-avx2-mul32-ld64-x32.c
new file mode 100644
index 000000000..b2dab8f1f
--- /dev/null
+++ b/src/qs8-vaddc/gen/minmax-avx2-mul32-ld64-x32.c
@@ -0,0 +1,108 @@
+// Auto-generated file. Do not edit!
+//   Template: src/qs8-vaddc/avx2-mul32-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 <immintrin.h>
+
+#include <xnnpack/intrinsics-polyfill.h>
+#include <xnnpack/vadd.h>
+
+
+void xnn_qs8_vaddc_minmax_ukernel__avx2_mul32_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 __m256i vx_multiplier = _mm256_broadcastsi128_si256(_mm_load_si128((const __m128i*) params->sse2.x_multiplier));
+  const __m256i vremainder_mask = _mm256_broadcastsi128_si256(_mm_load_si128((const __m128i*) params->sse2.remainder_mask));
+  const __m256i vremainder_threshold = _mm256_broadcastsi128_si256(_mm_load_si128((const __m128i*) params->sse2.remainder_threshold));
+  const __m128i vshift = _mm_cvtsi32_si128((int) params->sse2.shift);
+  const __m256i voutput_zero_point = _mm256_broadcastsi128_si256(_mm_load_si128((const __m128i*) params->sse2.output_zero_point));
+  const __m256i voutput_min = _mm256_broadcastsi128_si256(_mm_load_si128((const __m128i*) params->sse2.output_min));
+  const __m256i voutput_max = _mm256_broadcastsi128_si256(_mm_load_si128((const __m128i*) params->sse2.output_max));
+
+  __m256i vzero_point_product = _mm256_broadcastsi128_si256(_mm_add_epi32(
+    _mm_broadcastd_epi32(_mm_cvtsi32_si128(params->sse2.y_multiplier[0] * (int32_t) *input_y)),
+    _mm_load_si128((const __m128i*) params->sse2.zero_point_product)));
+  for (; n >= 32 * sizeof(int8_t); n -= 32 * sizeof(int8_t)) {
+    const __m256i vx01234567 = _mm256_cvtepi8_epi32(_mm_loadl_epi64((const __m128i*) input_x));
+    const __m256i vx89ABCDEF = _mm256_cvtepi8_epi32(_mm_loadl_epi64((const __m128i*) (input_x + 8)));
+    const __m256i vxGHIJKLMN = _mm256_cvtepi8_epi32(_mm_loadl_epi64((const __m128i*) (input_x + 16)));
+    const __m256i vxOPQRSTUV = _mm256_cvtepi8_epi32(_mm_loadl_epi64((const __m128i*) (input_x + 24)));
+    input_x += 32;
+
+    __m256i vacc01234567 = _mm256_add_epi32(vzero_point_product, _mm256_mullo_epi32(vx01234567, vx_multiplier));
+    __m256i vacc89ABCDEF = _mm256_add_epi32(vzero_point_product, _mm256_mullo_epi32(vx89ABCDEF, vx_multiplier));
+    __m256i vaccGHIJKLMN = _mm256_add_epi32(vzero_point_product, _mm256_mullo_epi32(vxGHIJKLMN, vx_multiplier));
+    __m256i vaccOPQRSTUV = _mm256_add_epi32(vzero_point_product, _mm256_mullo_epi32(vxOPQRSTUV, vx_multiplier));
+
+    const __m256i vrem01234567 = _mm256_add_epi32(_mm256_and_si256(vacc01234567, vremainder_mask), _mm256_srai_epi32(vacc01234567, 31));
+    const __m256i vrem89ABCDEF = _mm256_add_epi32(_mm256_and_si256(vacc89ABCDEF, vremainder_mask), _mm256_srai_epi32(vacc89ABCDEF, 31));
+    const __m256i vremGHIJKLMN = _mm256_add_epi32(_mm256_and_si256(vaccGHIJKLMN, vremainder_mask), _mm256_srai_epi32(vaccGHIJKLMN, 31));
+    const __m256i vremOPQRSTUV = _mm256_add_epi32(_mm256_and_si256(vaccOPQRSTUV, vremainder_mask), _mm256_srai_epi32(vaccOPQRSTUV, 31));
+
+    vacc01234567 = _mm256_sub_epi32(_mm256_sra_epi32(vacc01234567, vshift), _mm256_cmpgt_epi32(vrem01234567, vremainder_threshold));
+    vacc89ABCDEF = _mm256_sub_epi32(_mm256_sra_epi32(vacc89ABCDEF, vshift), _mm256_cmpgt_epi32(vrem89ABCDEF, vremainder_threshold));
+    vaccGHIJKLMN = _mm256_sub_epi32(_mm256_sra_epi32(vaccGHIJKLMN, vshift), _mm256_cmpgt_epi32(vremGHIJKLMN, vremainder_threshold));
+    vaccOPQRSTUV = _mm256_sub_epi32(_mm256_sra_epi32(vaccOPQRSTUV, vshift), _mm256_cmpgt_epi32(vremOPQRSTUV, vremainder_threshold));
+
+    __m256i vout012389AB4567CDEF = _mm256_adds_epi16(_mm256_packs_epi32(vacc01234567, vacc89ABCDEF), voutput_zero_point);
+    __m256i voutGHIJOPQRKLMNSTUV = _mm256_adds_epi16(_mm256_packs_epi32(vaccGHIJKLMN, vaccOPQRSTUV), voutput_zero_point);
+
+    vout012389AB4567CDEF = _mm256_min_epi16(_mm256_max_epi16(vout012389AB4567CDEF, voutput_min), voutput_max);
+    voutGHIJOPQRKLMNSTUV = _mm256_min_epi16(_mm256_max_epi16(voutGHIJOPQRKLMNSTUV, voutput_min), voutput_max);
+
+    __m128i vout0123456789ABCDEF = _mm_shuffle_epi32(_mm_packs_epi16(_mm256_castsi256_si128(vout012389AB4567CDEF), _mm256_extracti128_si256(vout012389AB4567CDEF, 1)), _MM_SHUFFLE(3, 1, 2, 0));
+    __m128i voutGHIJKLMNOPQRSTUV = _mm_shuffle_epi32(_mm_packs_epi16(_mm256_castsi256_si128(voutGHIJOPQRKLMNSTUV), _mm256_extracti128_si256(voutGHIJOPQRKLMNSTUV, 1)), _MM_SHUFFLE(3, 1, 2, 0));
+
+    _mm_storeu_si128((__m128i*) output, vout0123456789ABCDEF);
+    _mm_storeu_si128((__m128i*) (output + 16), voutGHIJKLMNOPQRSTUV);
+    output += 32;
+  }
+  if XNN_UNLIKELY(n != 0) {
+    do {
+      const __m256i vx01234567 = _mm256_cvtepi8_epi32(_mm_loadl_epi64((const __m128i*) input_x));
+      input_x += 8;
+
+      __m256i vacc01234567 = _mm256_add_epi32(vzero_point_product, _mm256_mullo_epi32(vx01234567, vx_multiplier));
+
+      const __m256i vrem01234567 = _mm256_add_epi32(_mm256_and_si256(vacc01234567, vremainder_mask), _mm256_srai_epi32(vacc01234567, 31));
+
+      vacc01234567 = _mm256_sub_epi32(_mm256_sra_epi32(vacc01234567, vshift), _mm256_cmpgt_epi32(vrem01234567, vremainder_threshold));
+
+      __m128i vout01234567 = _mm_adds_epi16(_mm_packs_epi32(_mm256_castsi256_si128(vacc01234567), _mm256_extracti128_si256(vacc01234567, 1)), _mm256_castsi256_si128(voutput_zero_point));
+      vout01234567 = _mm_min_epi16(_mm_max_epi16(vout01234567, _mm256_castsi256_si128(voutput_min)), _mm256_castsi256_si128(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 * sizeof(int8_t);
+      } 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);
+        }
+        n = 0;
+      }
+    } while (n != 0);
+  }
+}
diff --git a/src/qs8-vaddc/gen/minmax-avx2-mul32-ld64-x8.c b/src/qs8-vaddc/gen/minmax-avx2-mul32-ld64-x8.c
new file mode 100644
index 000000000..127c635c1
--- /dev/null
+++ b/src/qs8-vaddc/gen/minmax-avx2-mul32-ld64-x8.c
@@ -0,0 +1,84 @@
+// Auto-generated file. Do not edit!
+//   Template: src/qs8-vaddc/avx2-mul32-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 <immintrin.h>
+
+#include <xnnpack/intrinsics-polyfill.h>
+#include <xnnpack/vadd.h>
+
+
+void xnn_qs8_vaddc_minmax_ukernel__avx2_mul32_ld64_x8(
+    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 __m256i vx_multiplier = _mm256_broadcastsi128_si256(_mm_load_si128((const __m128i*) params->sse2.x_multiplier));
+  const __m256i vremainder_mask = _mm256_broadcastsi128_si256(_mm_load_si128((const __m128i*) params->sse2.remainder_mask));
+  const __m256i vremainder_threshold = _mm256_broadcastsi128_si256(_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);
+
+  __m256i vzero_point_product = _mm256_broadcastsi128_si256(_mm_add_epi32(
+    _mm_broadcastd_epi32(_mm_cvtsi32_si128(params->sse2.y_multiplier[0] * (int32_t) *input_y)),
+    _mm_load_si128((const __m128i*) params->sse2.zero_point_product)));
+  for (; n >= 8 * sizeof(int8_t); n -= 8 * sizeof(int8_t)) {
+    const __m256i vx01234567 = _mm256_cvtepi8_epi32(_mm_loadl_epi64((const __m128i*) input_x));
+    input_x += 8;
+
+    __m256i vacc01234567 = _mm256_add_epi32(vzero_point_product, _mm256_mullo_epi32(vx01234567, vx_multiplier));
+
+    const __m256i vrem01234567 = _mm256_add_epi32(_mm256_and_si256(vacc01234567, vremainder_mask), _mm256_srai_epi32(vacc01234567, 31));
+
+    vacc01234567 = _mm256_sub_epi32(_mm256_sra_epi32(vacc01234567, vshift), _mm256_cmpgt_epi32(vrem01234567, vremainder_threshold));
+
+    __m128i vout01234567 = _mm_adds_epi16(_mm_packs_epi32(_mm256_castsi256_si128(vacc01234567), _mm256_extracti128_si256(vacc01234567, 1)), voutput_zero_point);
+
+    vout01234567 = _mm_min_epi16(_mm_max_epi16(vout01234567, voutput_min), voutput_max);
+
+    __m128i vout0123456701234567 = _mm_packs_epi16(vout01234567, vout01234567);
+
+    _mm_storel_epi64((__m128i*) output, vout0123456701234567);
+    output += 8;
+  }
+  if XNN_UNLIKELY(n != 0) {
+    {
+      const __m256i vx01234567 = _mm256_cvtepi8_epi32(_mm_loadl_epi64((const __m128i*) input_x));
+
+      __m256i vacc01234567 = _mm256_add_epi32(vzero_point_product, _mm256_mullo_epi32(vx01234567, vx_multiplier));
+
+      const __m256i vrem01234567 = _mm256_add_epi32(_mm256_and_si256(vacc01234567, vremainder_mask), _mm256_srai_epi32(vacc01234567, 31));
+
+      vacc01234567 = _mm256_sub_epi32(_mm256_sra_epi32(vacc01234567, vshift), _mm256_cmpgt_epi32(vrem01234567, vremainder_threshold));
+
+      __m128i vout01234567 = _mm_adds_epi16(_mm_packs_epi32(_mm256_castsi256_si128(vacc01234567), _mm256_extracti128_si256(vacc01234567, 1)), voutput_zero_point);
+      vout01234567 = _mm_min_epi16(_mm_max_epi16(vout01234567, voutput_min), voutput_max);
+      __m128i vout0123456701234567 = _mm_packs_epi16(vout01234567, vout01234567);
+
+      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);
+      }
+    }
+  }
+}