usearch 2.21.3 → 2.25.0

package/numkong/include/numkong/maxsim/haswell.h

@@ -0,0 +1,553 @@
+/**
+ *  @brief SIMD-accelerated MaxSim (angular distance late-interaction) for Haswell (AVX2).
+ *  @file include/numkong/maxsim/haswell.h
+ *  @author Ash Vardanian
+ *  @date February 28, 2026
+ *
+ *  @sa include/numkong/maxsim.h
+ *
+ *  Uses AVX2 VPMADDUBSW (u8×i8→i16) + VPMADDWD (i16→i32) for coarse i8 screening.
+ *  Quantization range [-79, 79] ensures no i16 saturation: worst pair sum = 2 × 207 × 79 = 32706 < 32767.
+ *  Bias correction via XOR-0x80 converts signed queries to unsigned, then subtracts 128 × sum_quantized.
+ *
+ *  4x4 register tiling: 4 queries × 4 documents = 16 YMM accumulators per depth loop.
+ *  Depth steps at 32 bytes (YMM width in bytes).
+ */
+#ifndef NK_MAXSIM_HASWELL_H
+#define NK_MAXSIM_HASWELL_H
+
+#if NK_TARGET_X8664_
+#if NK_TARGET_HASWELL
+
+#include "numkong/types.h"
+#include "numkong/maxsim/serial.h"   // `nk_maxsim_packed_header_t`
+#include "numkong/dot.h"             // `nk_dot_bf16`, `nk_dot_f32`, `nk_dot_f16`
+#include "numkong/cast/haswell.h"    // `nk_f16_to_f32_haswell`
+#include "numkong/spatial/haswell.h" // `nk_f32_sqrt_haswell`
+
+#if defined(__cplusplus)
+extern "C" {
+#endif
+
+#if defined(__clang__)
+#pragma clang attribute push(__attribute__((target("avx2,f16c,fma,bmi,bmi2"))), apply_to = function)
+#elif defined(__GNUC__)
+#pragma GCC push_options
+#pragma GCC target("avx2", "f16c", "fma", "bmi", "bmi2")
+#endif
+
+NK_PUBLIC nk_size_t nk_maxsim_packed_size_bf16_haswell(nk_size_t vector_count, nk_size_t depth) {
+    return nk_maxsim_packed_size_(vector_count, depth, sizeof(nk_bf16_t), 32);
+}
+
+NK_PUBLIC nk_size_t nk_maxsim_packed_size_f32_haswell(nk_size_t vector_count, nk_size_t depth) {
+    return nk_maxsim_packed_size_(vector_count, depth, sizeof(nk_f32_t), 32);
+}
+
+NK_PUBLIC nk_size_t nk_maxsim_packed_size_f16_haswell(nk_size_t vector_count, nk_size_t depth) {
+    return nk_maxsim_packed_size_(vector_count, depth, sizeof(nk_f16_t), 32);
+}
+
+NK_PUBLIC void nk_maxsim_pack_bf16_haswell( //
+    nk_bf16_t const *vectors, nk_size_t vector_count, nk_size_t depth, nk_size_t stride_in_bytes, void *packed) {
+
+    nk_size_t const element_bytes = sizeof(nk_bf16_t);
+    nk_size_t depth_i8_padded = nk_maxsim_packed_header_setup_(packed, vector_count, depth, 32, element_bytes);
+
+    nk_maxsim_packed_header_t const *header = (nk_maxsim_packed_header_t const *)packed;
+    nk_i8_t *quantized_i8 = (nk_i8_t *)((char *)packed + header->offset_i8_data);
+    nk_maxsim_vector_metadata_t *metadata = (nk_maxsim_vector_metadata_t *)((char *)packed + header->offset_metadata);
+    char *originals = (char *)packed + header->offset_original_data;
+    nk_size_t const original_stride = header->original_stride_bytes;
+
+    for (nk_size_t vector_index = 0; vector_index < vector_count; vector_index++) {
+        char const *source_row = (char const *)vectors + vector_index * stride_in_bytes;
+        nk_f32_t norm_sq;
+        nk_maxsim_quantize_vector_(source_row, element_bytes, depth, depth_i8_padded, 79.0f,
+                                   (nk_maxsim_to_f32_t)nk_bf16_to_f32_serial,
+                                   &quantized_i8[vector_index * depth_i8_padded], &metadata[vector_index], &norm_sq);
+        metadata[vector_index].inverse_norm_f32 = norm_sq > 0.0f ? (1.0f / nk_f32_sqrt_haswell(norm_sq)) : 0.0f;
+        char *destination_original = originals + vector_index * original_stride;
+        nk_copy_bytes_(destination_original, source_row, depth * element_bytes);
+        for (nk_size_t byte_index = depth * element_bytes; byte_index < original_stride; byte_index++)
+            destination_original[byte_index] = 0;
+    }
+}
+
+NK_PUBLIC void nk_maxsim_pack_f32_haswell( //
+    nk_f32_t const *vectors, nk_size_t vector_count, nk_size_t depth, nk_size_t stride_in_bytes, void *packed) {
+
+    nk_size_t const element_bytes = sizeof(nk_f32_t);
+    nk_size_t depth_i8_padded = nk_maxsim_packed_header_setup_(packed, vector_count, depth, 32, element_bytes);
+
+    nk_maxsim_packed_header_t const *header = (nk_maxsim_packed_header_t const *)packed;
+    nk_i8_t *quantized_i8 = (nk_i8_t *)((char *)packed + header->offset_i8_data);
+    nk_maxsim_vector_metadata_t *metadata = (nk_maxsim_vector_metadata_t *)((char *)packed + header->offset_metadata);
+    char *originals = (char *)packed + header->offset_original_data;
+    nk_size_t const original_stride = header->original_stride_bytes;
+
+    for (nk_size_t vector_index = 0; vector_index < vector_count; vector_index++) {
+        char const *source_row = (char const *)vectors + vector_index * stride_in_bytes;
+        nk_f32_t norm_sq;
+        nk_maxsim_quantize_vector_(source_row, element_bytes, depth, depth_i8_padded, 79.0f, nk_f32_to_f32_,
+                                   &quantized_i8[vector_index * depth_i8_padded], &metadata[vector_index], &norm_sq);
+        metadata[vector_index].inverse_norm_f32 = norm_sq > 0.0f ? (1.0f / nk_f32_sqrt_haswell(norm_sq)) : 0.0f;
+        char *destination_original = originals + vector_index * original_stride;
+        nk_copy_bytes_(destination_original, source_row, depth * element_bytes);
+        for (nk_size_t byte_index = depth * element_bytes; byte_index < original_stride; byte_index++)
+            destination_original[byte_index] = 0;
+    }
+}
+
+NK_PUBLIC void nk_maxsim_pack_f16_haswell( //
+    nk_f16_t const *vectors, nk_size_t vector_count, nk_size_t depth, nk_size_t stride_in_bytes, void *packed) {
+
+    nk_size_t const element_bytes = sizeof(nk_f16_t);
+    nk_size_t depth_i8_padded = nk_maxsim_packed_header_setup_(packed, vector_count, depth, 32, element_bytes);
+
+    nk_maxsim_packed_header_t const *header = (nk_maxsim_packed_header_t const *)packed;
+    nk_i8_t *quantized_i8 = (nk_i8_t *)((char *)packed + header->offset_i8_data);
+    nk_maxsim_vector_metadata_t *metadata = (nk_maxsim_vector_metadata_t *)((char *)packed + header->offset_metadata);
+    char *originals = (char *)packed + header->offset_original_data;
+    nk_size_t const original_stride = header->original_stride_bytes;
+
+    for (nk_size_t vector_index = 0; vector_index < vector_count; vector_index++) {
+        char const *source_row = (char const *)vectors + vector_index * stride_in_bytes;
+        nk_f32_t norm_sq;
+        nk_maxsim_quantize_vector_(source_row, element_bytes, depth, depth_i8_padded, 79.0f,
+                                   (nk_maxsim_to_f32_t)nk_f16_to_f32_haswell,
+                                   &quantized_i8[vector_index * depth_i8_padded], &metadata[vector_index], &norm_sq);
+        metadata[vector_index].inverse_norm_f32 = norm_sq > 0.0f ? (1.0f / nk_f32_sqrt_haswell(norm_sq)) : 0.0f;
+        char *destination_original = originals + vector_index * original_stride;
+        nk_copy_bytes_(destination_original, source_row, depth * element_bytes);
+        for (nk_size_t byte_index = depth * element_bytes; byte_index < original_stride; byte_index++)
+            destination_original[byte_index] = 0;
+    }
+}
+
+/** @brief Reduces 4 YMM i32x8 accumulators to a single __m128i with 4 horizontal sums. */
+NK_INTERNAL __m128i nk_maxsim_reduce_i32x8x4_haswell_(        //
+    __m256i accumulator_a_i32x8, __m256i accumulator_b_i32x8, //
+    __m256i accumulator_c_i32x8, __m256i accumulator_d_i32x8) {
+    // Step 1: 8 -> 4 (extract high 128-bit half and add to low half)
+    __m128i sum_a_i32x4 = _mm_add_epi32(_mm256_castsi256_si128(accumulator_a_i32x8),
+                                        _mm256_extracti128_si256(accumulator_a_i32x8, 1));
+    __m128i sum_b_i32x4 = _mm_add_epi32(_mm256_castsi256_si128(accumulator_b_i32x8),
+                                        _mm256_extracti128_si256(accumulator_b_i32x8, 1));
+    __m128i sum_c_i32x4 = _mm_add_epi32(_mm256_castsi256_si128(accumulator_c_i32x8),
+                                        _mm256_extracti128_si256(accumulator_c_i32x8, 1));
+    __m128i sum_d_i32x4 = _mm_add_epi32(_mm256_castsi256_si128(accumulator_d_i32x8),
+                                        _mm256_extracti128_si256(accumulator_d_i32x8, 1));
+    // Step 2: 4x4 transpose + reduce -> [sum_a, sum_b, sum_c, sum_d]
+    __m128i transpose_ab_low_i32x4 = _mm_unpacklo_epi32(sum_a_i32x4, sum_b_i32x4);
+    __m128i transpose_cd_low_i32x4 = _mm_unpacklo_epi32(sum_c_i32x4, sum_d_i32x4);
+    __m128i transpose_ab_high_i32x4 = _mm_unpackhi_epi32(sum_a_i32x4, sum_b_i32x4);
+    __m128i transpose_cd_high_i32x4 = _mm_unpackhi_epi32(sum_c_i32x4, sum_d_i32x4);
+    __m128i sum_lane_0_i32x4 = _mm_unpacklo_epi64(transpose_ab_low_i32x4, transpose_cd_low_i32x4);
+    __m128i sum_lane_1_i32x4 = _mm_unpackhi_epi64(transpose_ab_low_i32x4, transpose_cd_low_i32x4);
+    __m128i sum_lane_2_i32x4 = _mm_unpacklo_epi64(transpose_ab_high_i32x4, transpose_cd_high_i32x4);
+    __m128i sum_lane_3_i32x4 = _mm_unpackhi_epi64(transpose_ab_high_i32x4, transpose_cd_high_i32x4);
+    return _mm_add_epi32(_mm_add_epi32(sum_lane_0_i32x4, sum_lane_1_i32x4),
+                         _mm_add_epi32(sum_lane_2_i32x4, sum_lane_3_i32x4));
+}
+
+/**
+ *  @brief Factored coarse i8 argmax kernel for Haswell.
+ *  Uses AVX2 VPMADDUBSW (u8×i8→i16) + VPMADDWD (i16×1→i32) with XOR-0x80 bias.
+ *  4Q×4D register tiling with 16 YMM accumulators.
+ */
+NK_INTERNAL void nk_maxsim_coarse_argmax_haswell_(        //
+    nk_i8_t const *query_i8, nk_i8_t const *document_i8,  //
+    nk_maxsim_vector_metadata_t const *document_metadata, //
+    nk_size_t query_count, nk_size_t document_count,      //
+    nk_size_t depth_i8_padded, nk_u32_t *best_document_indices) {
+
+    __m256i const xor_mask_u8x32 = _mm256_set1_epi8((char)0x80);
+    __m256i const ones_i16x16 = _mm256_set1_epi16(1);
+
+    // Primary path: 4-query grouping
+    nk_size_t query_block_start_index = 0;
+    for (; query_block_start_index + 4 <= query_count; query_block_start_index += 4) {
+        __m128i running_max_i32x4 = _mm_set1_epi32(NK_I32_MIN);
+        __m128i running_argmax_i32x4 = _mm_setzero_si128();
+
+        // 4Q×4D document blocking
+        nk_size_t document_block_start_index = 0;
+        for (; document_block_start_index + 4 <= document_count; document_block_start_index += 4) {
+            __m256i accumulator_tiles_i32x8[4][4];
+            for (nk_size_t query_tile_index = 0; query_tile_index < 4; query_tile_index++)
+                for (nk_size_t document_tile_index = 0; document_tile_index < 4; document_tile_index++)
+                    accumulator_tiles_i32x8[query_tile_index][document_tile_index] = _mm256_setzero_si256();
+
+            for (nk_size_t depth_index = 0; depth_index < depth_i8_padded; depth_index += 32) {
+                __m256i query_biased_u8x32_0 = _mm256_xor_si256(
+                    _mm256_loadu_si256(
+                        (__m256i const *)(query_i8 + (query_block_start_index + 0) * depth_i8_padded + depth_index)),
+                    xor_mask_u8x32);
+                __m256i query_biased_u8x32_1 = _mm256_xor_si256(
+                    _mm256_loadu_si256(
+                        (__m256i const *)(query_i8 + (query_block_start_index + 1) * depth_i8_padded + depth_index)),
+                    xor_mask_u8x32);
+                __m256i query_biased_u8x32_2 = _mm256_xor_si256(
+                    _mm256_loadu_si256(
+                        (__m256i const *)(query_i8 + (query_block_start_index + 2) * depth_i8_padded + depth_index)),
+                    xor_mask_u8x32);
+                __m256i query_biased_u8x32_3 = _mm256_xor_si256(
+                    _mm256_loadu_si256(
+                        (__m256i const *)(query_i8 + (query_block_start_index + 3) * depth_i8_padded + depth_index)),
+                    xor_mask_u8x32);
+
+                __m256i document_i8x32, products_i16x16, products_i32x8;
+
+                // Document 0
+                document_i8x32 = _mm256_loadu_si256(
+                    (__m256i const *)(document_i8 + (document_block_start_index + 0) * depth_i8_padded + depth_index));
+                products_i16x16 = _mm256_maddubs_epi16(query_biased_u8x32_0, document_i8x32);
+                products_i32x8 = _mm256_madd_epi16(products_i16x16, ones_i16x16);
+                accumulator_tiles_i32x8[0][0] = _mm256_add_epi32(accumulator_tiles_i32x8[0][0], products_i32x8);
+                products_i16x16 = _mm256_maddubs_epi16(query_biased_u8x32_1, document_i8x32);
+                products_i32x8 = _mm256_madd_epi16(products_i16x16, ones_i16x16);
+                accumulator_tiles_i32x8[1][0] = _mm256_add_epi32(accumulator_tiles_i32x8[1][0], products_i32x8);
+                products_i16x16 = _mm256_maddubs_epi16(query_biased_u8x32_2, document_i8x32);
+                products_i32x8 = _mm256_madd_epi16(products_i16x16, ones_i16x16);
+                accumulator_tiles_i32x8[2][0] = _mm256_add_epi32(accumulator_tiles_i32x8[2][0], products_i32x8);
+                products_i16x16 = _mm256_maddubs_epi16(query_biased_u8x32_3, document_i8x32);
+                products_i32x8 = _mm256_madd_epi16(products_i16x16, ones_i16x16);
+                accumulator_tiles_i32x8[3][0] = _mm256_add_epi32(accumulator_tiles_i32x8[3][0], products_i32x8);
+
+                // Document 1
+                document_i8x32 = _mm256_loadu_si256(
+                    (__m256i const *)(document_i8 + (document_block_start_index + 1) * depth_i8_padded + depth_index));
+                products_i16x16 = _mm256_maddubs_epi16(query_biased_u8x32_0, document_i8x32);
+                products_i32x8 = _mm256_madd_epi16(products_i16x16, ones_i16x16);
+                accumulator_tiles_i32x8[0][1] = _mm256_add_epi32(accumulator_tiles_i32x8[0][1], products_i32x8);
+                products_i16x16 = _mm256_maddubs_epi16(query_biased_u8x32_1, document_i8x32);
+                products_i32x8 = _mm256_madd_epi16(products_i16x16, ones_i16x16);
+                accumulator_tiles_i32x8[1][1] = _mm256_add_epi32(accumulator_tiles_i32x8[1][1], products_i32x8);
+                products_i16x16 = _mm256_maddubs_epi16(query_biased_u8x32_2, document_i8x32);
+                products_i32x8 = _mm256_madd_epi16(products_i16x16, ones_i16x16);
+                accumulator_tiles_i32x8[2][1] = _mm256_add_epi32(accumulator_tiles_i32x8[2][1], products_i32x8);
+                products_i16x16 = _mm256_maddubs_epi16(query_biased_u8x32_3, document_i8x32);
+                products_i32x8 = _mm256_madd_epi16(products_i16x16, ones_i16x16);
+                accumulator_tiles_i32x8[3][1] = _mm256_add_epi32(accumulator_tiles_i32x8[3][1], products_i32x8);
+
+                // Document 2
+                document_i8x32 = _mm256_loadu_si256(
+                    (__m256i const *)(document_i8 + (document_block_start_index + 2) * depth_i8_padded + depth_index));
+                products_i16x16 = _mm256_maddubs_epi16(query_biased_u8x32_0, document_i8x32);
+                products_i32x8 = _mm256_madd_epi16(products_i16x16, ones_i16x16);
+                accumulator_tiles_i32x8[0][2] = _mm256_add_epi32(accumulator_tiles_i32x8[0][2], products_i32x8);
+                products_i16x16 = _mm256_maddubs_epi16(query_biased_u8x32_1, document_i8x32);
+                products_i32x8 = _mm256_madd_epi16(products_i16x16, ones_i16x16);
+                accumulator_tiles_i32x8[1][2] = _mm256_add_epi32(accumulator_tiles_i32x8[1][2], products_i32x8);
+                products_i16x16 = _mm256_maddubs_epi16(query_biased_u8x32_2, document_i8x32);
+                products_i32x8 = _mm256_madd_epi16(products_i16x16, ones_i16x16);
+                accumulator_tiles_i32x8[2][2] = _mm256_add_epi32(accumulator_tiles_i32x8[2][2], products_i32x8);
+                products_i16x16 = _mm256_maddubs_epi16(query_biased_u8x32_3, document_i8x32);
+                products_i32x8 = _mm256_madd_epi16(products_i16x16, ones_i16x16);
+                accumulator_tiles_i32x8[3][2] = _mm256_add_epi32(accumulator_tiles_i32x8[3][2], products_i32x8);
+
+                // Document 3
+                document_i8x32 = _mm256_loadu_si256(
+                    (__m256i const *)(document_i8 + (document_block_start_index + 3) * depth_i8_padded + depth_index));
+                products_i16x16 = _mm256_maddubs_epi16(query_biased_u8x32_0, document_i8x32);
+                products_i32x8 = _mm256_madd_epi16(products_i16x16, ones_i16x16);
+                accumulator_tiles_i32x8[0][3] = _mm256_add_epi32(accumulator_tiles_i32x8[0][3], products_i32x8);
+                products_i16x16 = _mm256_maddubs_epi16(query_biased_u8x32_1, document_i8x32);
+                products_i32x8 = _mm256_madd_epi16(products_i16x16, ones_i16x16);
+                accumulator_tiles_i32x8[1][3] = _mm256_add_epi32(accumulator_tiles_i32x8[1][3], products_i32x8);
+                products_i16x16 = _mm256_maddubs_epi16(query_biased_u8x32_2, document_i8x32);
+                products_i32x8 = _mm256_madd_epi16(products_i16x16, ones_i16x16);
+                accumulator_tiles_i32x8[2][3] = _mm256_add_epi32(accumulator_tiles_i32x8[2][3], products_i32x8);
+                products_i16x16 = _mm256_maddubs_epi16(query_biased_u8x32_3, document_i8x32);
+                products_i32x8 = _mm256_madd_epi16(products_i16x16, ones_i16x16);
+                accumulator_tiles_i32x8[3][3] = _mm256_add_epi32(accumulator_tiles_i32x8[3][3], products_i32x8);
+            }
+
+            // Reduce each query's 4 doc accumulators -> __m128i
+            __m128i query_0_coarse_dots_i32x4 = nk_maxsim_reduce_i32x8x4_haswell_(
+                accumulator_tiles_i32x8[0][0], accumulator_tiles_i32x8[0][1], accumulator_tiles_i32x8[0][2],
+                accumulator_tiles_i32x8[0][3]);
+            __m128i query_1_coarse_dots_i32x4 = nk_maxsim_reduce_i32x8x4_haswell_(
+                accumulator_tiles_i32x8[1][0], accumulator_tiles_i32x8[1][1], accumulator_tiles_i32x8[1][2],
+                accumulator_tiles_i32x8[1][3]);
+            __m128i query_2_coarse_dots_i32x4 = nk_maxsim_reduce_i32x8x4_haswell_(
+                accumulator_tiles_i32x8[2][0], accumulator_tiles_i32x8[2][1], accumulator_tiles_i32x8[2][2],
+                accumulator_tiles_i32x8[2][3]);
+            __m128i query_3_coarse_dots_i32x4 = nk_maxsim_reduce_i32x8x4_haswell_(
+                accumulator_tiles_i32x8[3][0], accumulator_tiles_i32x8[3][1], accumulator_tiles_i32x8[3][2],
+                accumulator_tiles_i32x8[3][3]);
+
+            // Bias correction: subtract 128 × sum_quantized for each document
+            __m128i bias_correction_i32x4 = _mm_set_epi32(
+                128 * document_metadata[document_block_start_index + 3].sum_i8_i32,
+                128 * document_metadata[document_block_start_index + 2].sum_i8_i32,
+                128 * document_metadata[document_block_start_index + 1].sum_i8_i32,
+                128 * document_metadata[document_block_start_index + 0].sum_i8_i32);
+            query_0_coarse_dots_i32x4 = _mm_sub_epi32(query_0_coarse_dots_i32x4, bias_correction_i32x4);
+            query_1_coarse_dots_i32x4 = _mm_sub_epi32(query_1_coarse_dots_i32x4, bias_correction_i32x4);
+            query_2_coarse_dots_i32x4 = _mm_sub_epi32(query_2_coarse_dots_i32x4, bias_correction_i32x4);
+            query_3_coarse_dots_i32x4 = _mm_sub_epi32(query_3_coarse_dots_i32x4, bias_correction_i32x4);
+
+            // 4x4 transpose: [query][doc] -> [doc][query] for vectorized argmax
+            __m128i transpose_queries_01_low_i32x4 = _mm_unpacklo_epi32(query_0_coarse_dots_i32x4,
+                                                                        query_1_coarse_dots_i32x4);
+            __m128i transpose_queries_23_low_i32x4 = _mm_unpacklo_epi32(query_2_coarse_dots_i32x4,
+                                                                        query_3_coarse_dots_i32x4);
+            __m128i transpose_queries_01_high_i32x4 = _mm_unpackhi_epi32(query_0_coarse_dots_i32x4,
+                                                                         query_1_coarse_dots_i32x4);
+            __m128i transpose_queries_23_high_i32x4 = _mm_unpackhi_epi32(query_2_coarse_dots_i32x4,
+                                                                         query_3_coarse_dots_i32x4);
+            __m128i document_0_dots_i32x4 = _mm_unpacklo_epi64(transpose_queries_01_low_i32x4,
+                                                               transpose_queries_23_low_i32x4);
+            __m128i document_1_dots_i32x4 = _mm_unpackhi_epi64(transpose_queries_01_low_i32x4,
+                                                               transpose_queries_23_low_i32x4);
+            __m128i document_2_dots_i32x4 = _mm_unpacklo_epi64(transpose_queries_01_high_i32x4,
+                                                               transpose_queries_23_high_i32x4);
+            __m128i document_3_dots_i32x4 = _mm_unpackhi_epi64(transpose_queries_01_high_i32x4,
+                                                               transpose_queries_23_high_i32x4);
+
+            // Branchless SIMD argmax
+            __m128i comparison_mask_i32x4, document_index_i32x4;
+
+            comparison_mask_i32x4 = _mm_cmpgt_epi32(document_0_dots_i32x4, running_max_i32x4);
+            document_index_i32x4 = _mm_set1_epi32((int)(document_block_start_index + 0));
+            running_max_i32x4 = _mm_blendv_epi8(running_max_i32x4, document_0_dots_i32x4, comparison_mask_i32x4);
+            running_argmax_i32x4 = _mm_blendv_epi8(running_argmax_i32x4, document_index_i32x4, comparison_mask_i32x4);
+
+            comparison_mask_i32x4 = _mm_cmpgt_epi32(document_1_dots_i32x4, running_max_i32x4);
+            document_index_i32x4 = _mm_set1_epi32((int)(document_block_start_index + 1));
+            running_max_i32x4 = _mm_blendv_epi8(running_max_i32x4, document_1_dots_i32x4, comparison_mask_i32x4);
+            running_argmax_i32x4 = _mm_blendv_epi8(running_argmax_i32x4, document_index_i32x4, comparison_mask_i32x4);
+
+            comparison_mask_i32x4 = _mm_cmpgt_epi32(document_2_dots_i32x4, running_max_i32x4);
+            document_index_i32x4 = _mm_set1_epi32((int)(document_block_start_index + 2));
+            running_max_i32x4 = _mm_blendv_epi8(running_max_i32x4, document_2_dots_i32x4, comparison_mask_i32x4);
+            running_argmax_i32x4 = _mm_blendv_epi8(running_argmax_i32x4, document_index_i32x4, comparison_mask_i32x4);
+
+            comparison_mask_i32x4 = _mm_cmpgt_epi32(document_3_dots_i32x4, running_max_i32x4);
+            document_index_i32x4 = _mm_set1_epi32((int)(document_block_start_index + 3));
+            running_max_i32x4 = _mm_blendv_epi8(running_max_i32x4, document_3_dots_i32x4, comparison_mask_i32x4);
+            running_argmax_i32x4 = _mm_blendv_epi8(running_argmax_i32x4, document_index_i32x4, comparison_mask_i32x4);
+        }
+
+        // Document tail: 4Q×1D
+        for (nk_size_t document_index = document_block_start_index; document_index < document_count; document_index++) {
+            nk_i8_t const *document_i8_row = document_i8 + document_index * depth_i8_padded;
+
+            __m256i accumulator_i32x8_0 = _mm256_setzero_si256();
+            __m256i accumulator_i32x8_1 = _mm256_setzero_si256();
+            __m256i accumulator_i32x8_2 = _mm256_setzero_si256();
+            __m256i accumulator_i32x8_3 = _mm256_setzero_si256();
+
+            for (nk_size_t depth_index = 0; depth_index < depth_i8_padded; depth_index += 32) {
+                __m256i document_i8x32 = _mm256_loadu_si256((__m256i const *)(document_i8_row + depth_index));
+                __m256i products_i16x16, products_i32x8;
+
+                products_i16x16 = _mm256_maddubs_epi16(
+                    _mm256_xor_si256(
+                        _mm256_loadu_si256((
+                            __m256i const *)(query_i8 + (query_block_start_index + 0) * depth_i8_padded + depth_index)),
+                        xor_mask_u8x32),
+                    document_i8x32);
+                products_i32x8 = _mm256_madd_epi16(products_i16x16, ones_i16x16);
+                accumulator_i32x8_0 = _mm256_add_epi32(accumulator_i32x8_0, products_i32x8);
+
+                products_i16x16 = _mm256_maddubs_epi16(
+                    _mm256_xor_si256(
+                        _mm256_loadu_si256((
+                            __m256i const *)(query_i8 + (query_block_start_index + 1) * depth_i8_padded + depth_index)),
+                        xor_mask_u8x32),
+                    document_i8x32);
+                products_i32x8 = _mm256_madd_epi16(products_i16x16, ones_i16x16);
+                accumulator_i32x8_1 = _mm256_add_epi32(accumulator_i32x8_1, products_i32x8);
+
+                products_i16x16 = _mm256_maddubs_epi16(
+                    _mm256_xor_si256(
+                        _mm256_loadu_si256((
+                            __m256i const *)(query_i8 + (query_block_start_index + 2) * depth_i8_padded + depth_index)),
+                        xor_mask_u8x32),
+                    document_i8x32);
+                products_i32x8 = _mm256_madd_epi16(products_i16x16, ones_i16x16);
+                accumulator_i32x8_2 = _mm256_add_epi32(accumulator_i32x8_2, products_i32x8);
+
+                products_i16x16 = _mm256_maddubs_epi16(
+                    _mm256_xor_si256(
+                        _mm256_loadu_si256((
+                            __m256i const *)(query_i8 + (query_block_start_index + 3) * depth_i8_padded + depth_index)),
+                        xor_mask_u8x32),
+                    document_i8x32);
+                products_i32x8 = _mm256_madd_epi16(products_i16x16, ones_i16x16);
+                accumulator_i32x8_3 = _mm256_add_epi32(accumulator_i32x8_3, products_i32x8);
+            }
+
+            __m128i reduced_i32x4 = nk_maxsim_reduce_i32x8x4_haswell_(accumulator_i32x8_0, accumulator_i32x8_1,
+                                                                      accumulator_i32x8_2, accumulator_i32x8_3);
+            nk_i32_t bias_correction_i32 = 128 * document_metadata[document_index].sum_i8_i32;
+            __m128i coarse_dots_i32x4 = _mm_sub_epi32(reduced_i32x4, _mm_set1_epi32(bias_correction_i32));
+
+            __m128i comparison_mask_i32x4 = _mm_cmpgt_epi32(coarse_dots_i32x4, running_max_i32x4);
+            __m128i document_index_i32x4 = _mm_set1_epi32((int)document_index);
+            running_max_i32x4 = _mm_blendv_epi8(running_max_i32x4, coarse_dots_i32x4, comparison_mask_i32x4);
+            running_argmax_i32x4 = _mm_blendv_epi8(running_argmax_i32x4, document_index_i32x4, comparison_mask_i32x4);
+        }
+
+        best_document_indices[query_block_start_index + 0] = (nk_u32_t)_mm_extract_epi32(running_argmax_i32x4, 0);
+        best_document_indices[query_block_start_index + 1] = (nk_u32_t)_mm_extract_epi32(running_argmax_i32x4, 1);
+        best_document_indices[query_block_start_index + 2] = (nk_u32_t)_mm_extract_epi32(running_argmax_i32x4, 2);
+        best_document_indices[query_block_start_index + 3] = (nk_u32_t)_mm_extract_epi32(running_argmax_i32x4, 3);
+    }
+
+    // Query tail: 1Q×1D
+    for (nk_size_t query_index = query_block_start_index; query_index < query_count; query_index++) {
+        nk_i8_t const *query_i8_row = query_i8 + query_index * depth_i8_padded;
+        nk_i32_t running_max_i32 = NK_I32_MIN;
+        nk_u32_t running_argmax_u32 = 0;
+
+        for (nk_size_t document_index = 0; document_index < document_count; document_index++) {
+            nk_i8_t const *document_i8_row = document_i8 + document_index * depth_i8_padded;
+            __m256i accumulator_i32x8 = _mm256_setzero_si256();
+
+            for (nk_size_t depth_index = 0; depth_index < depth_i8_padded; depth_index += 32) {
+                __m256i document_i8x32 = _mm256_loadu_si256((__m256i const *)(document_i8_row + depth_index));
+                __m256i query_biased_u8x32 = _mm256_xor_si256(
+                    _mm256_loadu_si256((__m256i const *)(query_i8_row + depth_index)), xor_mask_u8x32);
+                __m256i products_i16x16 = _mm256_maddubs_epi16(query_biased_u8x32, document_i8x32);
+                __m256i products_i32x8 = _mm256_madd_epi16(products_i16x16, ones_i16x16);
+                accumulator_i32x8 = _mm256_add_epi32(accumulator_i32x8, products_i32x8);
+            }
+
+            // Horizontal sum of 8 i32 lanes
+            __m128i sum_i32x4 = _mm_add_epi32(_mm256_castsi256_si128(accumulator_i32x8),
+                                              _mm256_extracti128_si256(accumulator_i32x8, 1));
+            sum_i32x4 = _mm_add_epi32(sum_i32x4, _mm_shuffle_epi32(sum_i32x4, 0x4E)); // 01001110
+            sum_i32x4 = _mm_add_epi32(sum_i32x4, _mm_shuffle_epi32(sum_i32x4, 0xB1)); // 10110001
+            nk_i32_t coarse_dot_i32 = _mm_extract_epi32(sum_i32x4, 0) -
+                                      128 * document_metadata[document_index].sum_i8_i32;
+
+            if (coarse_dot_i32 > running_max_i32) {
+                running_max_i32 = coarse_dot_i32;
+                running_argmax_u32 = (nk_u32_t)document_index;
+            }
+        }
+
+        best_document_indices[query_index] = running_argmax_u32;
+    }
+}
+
+NK_PUBLIC void nk_maxsim_packed_bf16_haswell( //
+    void const *query_packed, void const *document_packed, nk_size_t query_count, nk_size_t document_count,
+    nk_size_t depth, nk_f32_t *result) {
+
+    nk_maxsim_packed_regions_t regions = nk_maxsim_extract_packed_regions_(query_packed, document_packed);
+    nk_f64_t total_angular_distance = 0.0;
+
+    for (nk_size_t chunk_start = 0; chunk_start < query_count; chunk_start += 256) {
+        nk_size_t chunk_size = query_count - chunk_start < 256 ? query_count - chunk_start : 256;
+        nk_u32_t best_document_indices[256];
+
+        nk_maxsim_coarse_argmax_haswell_(regions.query_quantized + chunk_start * regions.depth_i8_padded,
+                                         regions.document_quantized, regions.document_metadata, chunk_size,
+                                         document_count, regions.depth_i8_padded, best_document_indices);
+
+        for (nk_size_t query_index = 0; query_index < chunk_size; query_index++) {
+            nk_u32_t best_document_index = best_document_indices[query_index];
+            nk_f32_t dot_result;
+            nk_dot_bf16((nk_bf16_t const *)(regions.query_originals +
+                                            (chunk_start + query_index) * regions.query_original_stride),
+                        (nk_bf16_t const *)(regions.document_originals +
+                                            best_document_index * regions.document_original_stride),
+                        depth, &dot_result);
+            nk_f32_t cosine = dot_result * regions.query_metadata[chunk_start + query_index].inverse_norm_f32 *
+                              regions.document_metadata[best_document_index].inverse_norm_f32;
+            nk_f32_t angular = 1.0f - cosine;
+            if (angular < 0.0f) angular = 0.0f;
+            total_angular_distance += (nk_f64_t)angular;
+        }
+    }
+
+    *result = (nk_f32_t)total_angular_distance;
+}
+
+NK_PUBLIC void nk_maxsim_packed_f32_haswell( //
+    void const *query_packed, void const *document_packed, nk_size_t query_count, nk_size_t document_count,
+    nk_size_t depth, nk_f64_t *result) {
+
+    nk_maxsim_packed_regions_t regions = nk_maxsim_extract_packed_regions_(query_packed, document_packed);
+    nk_f64_t total_angular_distance = 0.0;
+
+    for (nk_size_t chunk_start = 0; chunk_start < query_count; chunk_start += 256) {
+        nk_size_t chunk_size = query_count - chunk_start < 256 ? query_count - chunk_start : 256;
+        nk_u32_t best_document_indices[256];
+
+        nk_maxsim_coarse_argmax_haswell_(regions.query_quantized + chunk_start * regions.depth_i8_padded,
+                                         regions.document_quantized, regions.document_metadata, chunk_size,
+                                         document_count, regions.depth_i8_padded, best_document_indices);
+
+        for (nk_size_t query_index = 0; query_index < chunk_size; query_index++) {
+            nk_u32_t best_document_index = best_document_indices[query_index];
+            nk_f64_t dot_result;
+            nk_dot_f32(
+                (nk_f32_t const *)(regions.query_originals +
+                                   (chunk_start + query_index) * regions.query_original_stride),
+                (nk_f32_t const *)(regions.document_originals + best_document_index * regions.document_original_stride),
+                depth, &dot_result);
+            nk_f64_t cosine = dot_result *
+                              (nk_f64_t)regions.query_metadata[chunk_start + query_index].inverse_norm_f32 *
+                              (nk_f64_t)regions.document_metadata[best_document_index].inverse_norm_f32;
+            nk_f64_t angular = 1.0 - cosine;
+            if (angular < 0.0) angular = 0.0;
+            total_angular_distance += angular;
+        }
+    }
+
+    *result = total_angular_distance;
+}
+
+NK_PUBLIC void nk_maxsim_packed_f16_haswell( //
+    void const *query_packed, void const *document_packed, nk_size_t query_count, nk_size_t document_count,
+    nk_size_t depth, nk_f32_t *result) {
+
+    nk_maxsim_packed_regions_t regions = nk_maxsim_extract_packed_regions_(query_packed, document_packed);
+    nk_f64_t total_angular_distance = 0.0;
+
+    for (nk_size_t chunk_start = 0; chunk_start < query_count; chunk_start += 256) {
+        nk_size_t chunk_size = query_count - chunk_start < 256 ? query_count - chunk_start : 256;
+        nk_u32_t best_document_indices[256];
+
+        nk_maxsim_coarse_argmax_haswell_(regions.query_quantized + chunk_start * regions.depth_i8_padded,
+                                         regions.document_quantized, regions.document_metadata, chunk_size,
+                                         document_count, regions.depth_i8_padded, best_document_indices);
+
+        for (nk_size_t query_index = 0; query_index < chunk_size; query_index++) {
+            nk_u32_t best_document_index = best_document_indices[query_index];
+            nk_f32_t dot_result;
+            nk_dot_f16(
+                (nk_f16_t const *)(regions.query_originals +
+                                   (chunk_start + query_index) * regions.query_original_stride),
+                (nk_f16_t const *)(regions.document_originals + best_document_index * regions.document_original_stride),
+                depth, &dot_result);
+            nk_f32_t cosine = dot_result * regions.query_metadata[chunk_start + query_index].inverse_norm_f32 *
+                              regions.document_metadata[best_document_index].inverse_norm_f32;
+            nk_f32_t angular = 1.0f - cosine;
+            if (angular < 0.0f) angular = 0.0f;
+            total_angular_distance += (nk_f64_t)angular;
+        }
+    }
+
+    *result = (nk_f32_t)total_angular_distance;
+}
+
+#if defined(__clang__)
+#pragma clang attribute pop
+#elif defined(__GNUC__)
+#pragma GCC pop_options
+#endif
+
+#if defined(__cplusplus)
+} // extern "C"
+#endif
+
+#endif // NK_TARGET_HASWELL
+#endif // NK_TARGET_X8664_
+#endif // NK_MAXSIM_HASWELL_H