+
+#if defined (USE_AVX512)
+
+ [[maybe_unused]] const __m512i kOnes512 = _mm512_set1_epi16(1);
+
+ [[maybe_unused]] auto m512_hadd = [](__m512i sum, int bias) -> int {
+ return _mm512_reduce_add_epi32(sum) + bias;
+ };
+
+ // This function takes
+ // sum0 = [xmm0a, xmm0b, xmm0c, xmm0d]
+ // sum1 = [xmm1a, xmm1b, xmm1c, xmm1d]
+ // sum2 = [xmm2a, xmm2b, xmm2c, xmm2d]
+ // sum3 = [xmm3a, xmm3b, xmm3c, xmm3d]
+ // and returns
+ // ret = [
+ // reduce_add_epi32(xmm0a), reduce_add_epi32(xmm1a), reduce_add_epi32(xmm2a), reduce_add_epi32(xmm3a),
+ // reduce_add_epi32(xmm0b), reduce_add_epi32(xmm1b), reduce_add_epi32(xmm2b), reduce_add_epi32(xmm3b),
+ // reduce_add_epi32(xmm0c), reduce_add_epi32(xmm1c), reduce_add_epi32(xmm2c), reduce_add_epi32(xmm3c),
+ // reduce_add_epi32(xmm0d), reduce_add_epi32(xmm1d), reduce_add_epi32(xmm2d), reduce_add_epi32(xmm3d)
+ // ]
+ [[maybe_unused]] auto m512_hadd128x16_interleave = [](
+ __m512i sum0, __m512i sum1, __m512i sum2, __m512i sum3) -> __m512i {
+
+ __m512i sum01a = _mm512_unpacklo_epi32(sum0, sum1);
+ __m512i sum01b = _mm512_unpackhi_epi32(sum0, sum1);
+
+ __m512i sum23a = _mm512_unpacklo_epi32(sum2, sum3);
+ __m512i sum23b = _mm512_unpackhi_epi32(sum2, sum3);
+
+ __m512i sum01 = _mm512_add_epi32(sum01a, sum01b);
+ __m512i sum23 = _mm512_add_epi32(sum23a, sum23b);
+
+ __m512i sum0123a = _mm512_unpacklo_epi64(sum01, sum23);
+ __m512i sum0123b = _mm512_unpackhi_epi64(sum01, sum23);
+
+ return _mm512_add_epi32(sum0123a, sum0123b);
+ };
+
+ [[maybe_unused]] auto m512_haddx4 = [m512_hadd128x16_interleave](
+ __m512i sum0, __m512i sum1, __m512i sum2, __m512i sum3, __m128i bias) -> __m128i {
+
+ __m512i sum = m512_hadd128x16_interleave(sum0, sum1, sum2, sum3);
+
+ __m256i sum256lo = _mm512_castsi512_si256(sum);
+ __m256i sum256hi = _mm512_extracti64x4_epi64(sum, 1);
+
+ sum256lo = _mm256_add_epi32(sum256lo, sum256hi);
+
+ __m128i sum128lo = _mm256_castsi256_si128(sum256lo);
+ __m128i sum128hi = _mm256_extracti128_si256(sum256lo, 1);
+
+ return _mm_add_epi32(_mm_add_epi32(sum128lo, sum128hi), bias);
+ };
+
+ [[maybe_unused]] auto m512_haddx8 = [m512_hadd128x16_interleave](
+ __m512i sum0, __m512i sum1, __m512i sum2, __m512i sum3,
+ __m512i sum4, __m512i sum5, __m512i sum6, __m512i sum7, __m256i bias) -> __m256i {
+
+ __m512i suma = m512_hadd128x16_interleave(sum0, sum1, sum2, sum3);
+ __m512i sumb = m512_hadd128x16_interleave(sum4, sum5, sum6, sum7);
+
+ __m512i indices0 = _mm512_setr_epi64(0, 1, 8, 9, 4, 5, 12, 13);
+ __m512i indices1 = _mm512_setr_epi64(2, 3, 10, 11, 6, 7, 14, 15);
+ __m512i x = _mm512_add_epi32(
+ _mm512_permutex2var_epi64(suma, indices0, sumb),
+ _mm512_permutex2var_epi64(suma, indices1, sumb));
+
+ __m256i sum256lo = _mm512_castsi512_si256(x);
+ __m256i sum256hi = _mm512_extracti64x4_epi64(x, 1);
+
+ return _mm256_add_epi32(_mm256_add_epi32(sum256lo, sum256hi), bias);
+ };
+
+ [[maybe_unused]] auto m512_hadd256x8 =[m512_hadd128x16_interleave](
+ __m512i sum0, __m512i sum1, __m512i sum2, __m512i sum3, __m256i bias) -> __m256i {
+
+ __m512i sum = m512_hadd128x16_interleave(sum0, sum1, sum2, sum3);
+
+ __m512i indices = _mm512_setr_epi32(
+ 0, 4, 8, 12, 2, 6, 10, 14,
+ 1, 5, 9, 13, 3, 7, 11, 15);
+ sum = _mm512_permutexvar_epi32(indices, sum);
+
+ __m256i sum256lo = _mm512_castsi512_si256(sum);
+ __m256i sum256hi = _mm512_extracti64x4_epi64(sum, 1);
+
+ return _mm256_add_epi32(_mm256_hadd_epi32(sum256lo, sum256hi), bias);
+ };
+
+ [[maybe_unused]] auto m512_hadd256x16 = [m512_hadd128x16_interleave](
+ __m512i sum0, __m512i sum1, __m512i sum2, __m512i sum3,
+ __m512i sum4, __m512i sum5, __m512i sum6, __m512i sum7, __m512i bias) -> __m512i {
+
+ __m512i suma = m512_hadd128x16_interleave(sum0, sum1, sum2, sum3);
+ __m512i sumb = m512_hadd128x16_interleave(sum4, sum5, sum6, sum7);
+
+ __m512i indices0 = _mm512_setr_epi64(0, 1, 8, 9, 4, 5, 12, 13);
+ __m512i indices1 = _mm512_setr_epi64(2, 3, 10, 11, 6, 7, 14, 15);
+ __m512i x = _mm512_add_epi32(
+ _mm512_permutex2var_epi64(suma, indices0, sumb),
+ _mm512_permutex2var_epi64(suma, indices1, sumb));
+
+ __m512i indices = _mm512_setr_epi32(0, 8, 1, 9, 2, 10, 3, 11, 4, 12, 5, 13, 6, 14, 7, 15);
+ return _mm512_add_epi32(_mm512_permutexvar_epi32(indices, x), bias);
+ };
+
+ [[maybe_unused]] auto m512_add_dpbusd_epi32 = [=](__m512i& acc, __m512i a, __m512i b) {
+#if defined (USE_VNNI)
+ acc = _mm512_dpbusd_epi32(acc, a, b);
+#else
+ __m512i product0 = _mm512_maddubs_epi16(a, b);
+ product0 = _mm512_madd_epi16(product0, kOnes512);
+ acc = _mm512_add_epi32(acc, product0);
+#endif
+ };
+
+#endif
+#if defined (USE_AVX2)
+
+ [[maybe_unused]] const __m256i kOnes256 = _mm256_set1_epi16(1);
+
+ [[maybe_unused]] auto m256_hadd = [](__m256i sum, int bias) -> int {
+ __m128i sum128 = _mm_add_epi32(_mm256_castsi256_si128(sum), _mm256_extracti128_si256(sum, 1));
+ sum128 = _mm_add_epi32(sum128, _mm_shuffle_epi32(sum128, _MM_PERM_BADC));
+ sum128 = _mm_add_epi32(sum128, _mm_shuffle_epi32(sum128, _MM_PERM_CDAB));
+ return _mm_cvtsi128_si32(sum128) + bias;
+ };
+
+ [[maybe_unused]] auto m256_haddx4 = [](__m256i sum0, __m256i sum1, __m256i sum2, __m256i sum3, __m128i bias) -> __m128i {
+ sum0 = _mm256_hadd_epi32(sum0, sum1);
+ sum2 = _mm256_hadd_epi32(sum2, sum3);
+
+ sum0 = _mm256_hadd_epi32(sum0, sum2);
+
+ __m128i sum128lo = _mm256_castsi256_si128(sum0);
+ __m128i sum128hi = _mm256_extracti128_si256(sum0, 1);
+
+ return _mm_add_epi32(_mm_add_epi32(sum128lo, sum128hi), bias);
+ };
+
+ [[maybe_unused]] auto m256_add_dpbusd_epi32 = [=](__m256i& acc, __m256i a, __m256i b) {
+#if defined (USE_VNNI)
+ acc = _mm256_dpbusd_epi32(acc, a, b);
+#else
+ __m256i product0 = _mm256_maddubs_epi16(a, b);
+ product0 = _mm256_madd_epi16(product0, kOnes256);
+ acc = _mm256_add_epi32(acc, product0);
+#endif
+ };
+
+#endif
+
+#if defined (USE_SSSE3)
+
+ [[maybe_unused]] const __m128i kOnes128 = _mm_set1_epi16(1);
+
+ [[maybe_unused]] auto m128_hadd = [](__m128i sum, int bias) -> int {
+ sum = _mm_add_epi32(sum, _mm_shuffle_epi32(sum, 0x4E)); //_MM_PERM_BADC
+ sum = _mm_add_epi32(sum, _mm_shuffle_epi32(sum, 0xB1)); //_MM_PERM_CDAB
+ return _mm_cvtsi128_si32(sum) + bias;
+ };
+
+ [[maybe_unused]] auto m128_haddx4 = [](__m128i sum0, __m128i sum1, __m128i sum2, __m128i sum3, __m128i bias) -> __m128i {
+ sum0 = _mm_hadd_epi32(sum0, sum1);
+ sum2 = _mm_hadd_epi32(sum2, sum3);
+
+ sum0 = _mm_hadd_epi32(sum0, sum2);
+
+ return _mm_add_epi32(sum0, bias);
+ };
+
+ [[maybe_unused]] auto m128_add_dpbusd_epi32 = [=](__m128i& acc, __m128i a, __m128i b) {
+ __m128i product0 = _mm_maddubs_epi16(a, b);
+ product0 = _mm_madd_epi16(product0, kOnes128);
+ acc = _mm_add_epi32(acc, product0);
+ };
+
+#endif
+
+#if defined (USE_AVX512)
+
+ constexpr IndexType kNumChunks512 = kPaddedInputDimensions / (kSimdWidth * 2);
+ constexpr IndexType kNumChunks256 = kPaddedInputDimensions / kSimdWidth;
+