numkong 7.0.0 → 7.4.1

package/include/numkong/spatials/diamond.h

@@ -0,0 +1,82 @@
+/**
+ *  @brief SIMD-accelerated Batched Spatial Distances for Diamond Rapids.
+ *  @file include/numkong/spatials/diamond.h
+ *  @author Ash Vardanian
+ *  @date March 23, 2026
+ *
+ *  @sa include/numkong/spatials.h
+ *
+ *  Uses VCVTHF82PH/VCVTBF82PH + VDPPHPS via the batched dots/ infrastructure,
+ *  then derives angular and Euclidean distances.
+ */
+#ifndef NK_SPATIALS_DIAMOND_H
+#define NK_SPATIALS_DIAMOND_H
+
+#if NK_TARGET_X86_
+#if NK_TARGET_DIAMOND
+
+#include "numkong/dots/diamond.h"
+
+#if defined(__cplusplus)
+extern "C" {
+#endif
+
+#if defined(__clang__)
+#pragma clang attribute push(                                                                                    \
+    __attribute__((target("avx2,avx512f,avx512vl,avx512bw,avx512dq,avx512fp16,avx10.2-512,f16c,fma,bmi,bmi2"))), \
+    apply_to = function)
+#elif defined(__GNUC__)
+#pragma GCC push_options
+#pragma GCC target("avx2", "avx512f", "avx512vl", "avx512bw", "avx512dq", "avx512fp16", "avx10.2-512", "f16c", "fma", \
+                   "bmi", "bmi2")
+#endif
+
+nk_define_cross_normalized_packed_(angular, e4m3, diamond, e4m3, e4m3, f32, /*norm_value_type=*/f32, f32, nk_b128_vec_t,
+                                   nk_dots_packed_e4m3_diamond, nk_angular_through_f32_from_dot_haswell_,
+                                   nk_dots_reduce_sumsq_e4m3_, nk_load_b128_haswell_, nk_partial_load_b32x4_skylake_,
+                                   nk_store_b128_haswell_, nk_partial_store_b32x4_skylake_, 1)
+nk_define_cross_normalized_packed_(euclidean, e4m3, diamond, e4m3, e4m3, f32, /*norm_value_type=*/f32, f32,
+                                   nk_b128_vec_t, nk_dots_packed_e4m3_diamond,
+                                   nk_euclidean_through_f32_from_dot_haswell_, nk_dots_reduce_sumsq_e4m3_,
+                                   nk_load_b128_haswell_, nk_partial_load_b32x4_skylake_, nk_store_b128_haswell_,
+                                   nk_partial_store_b32x4_skylake_, 1)
+nk_define_cross_normalized_symmetric_(angular, e4m3, diamond, e4m3, f32, /*norm_value_type=*/f32, f32, nk_b128_vec_t,
+                                      nk_dots_symmetric_e4m3_diamond, nk_angular_through_f32_from_dot_haswell_,
+                                      nk_dots_reduce_sumsq_e4m3_, nk_load_b128_haswell_, nk_partial_load_b32x4_skylake_,
+                                      nk_store_b128_haswell_, nk_partial_store_b32x4_skylake_, 1)
+nk_define_cross_normalized_symmetric_(euclidean, e4m3, diamond, e4m3, f32, /*norm_value_type=*/f32, f32, nk_b128_vec_t,
+                                      nk_dots_symmetric_e4m3_diamond, nk_euclidean_through_f32_from_dot_haswell_,
+                                      nk_dots_reduce_sumsq_e4m3_, nk_load_b128_haswell_, nk_partial_load_b32x4_skylake_,
+                                      nk_store_b128_haswell_, nk_partial_store_b32x4_skylake_, 1)
+
+nk_define_cross_normalized_packed_(angular, e5m2, diamond, e5m2, e5m2, f32, /*norm_value_type=*/f32, f32, nk_b128_vec_t,
+                                   nk_dots_packed_e5m2_diamond, nk_angular_through_f32_from_dot_haswell_,
+                                   nk_dots_reduce_sumsq_e5m2_, nk_load_b128_haswell_, nk_partial_load_b32x4_skylake_,
+                                   nk_store_b128_haswell_, nk_partial_store_b32x4_skylake_, 1)
+nk_define_cross_normalized_packed_(euclidean, e5m2, diamond, e5m2, e5m2, f32, /*norm_value_type=*/f32, f32,
+                                   nk_b128_vec_t, nk_dots_packed_e5m2_diamond,
+                                   nk_euclidean_through_f32_from_dot_haswell_, nk_dots_reduce_sumsq_e5m2_,
+                                   nk_load_b128_haswell_, nk_partial_load_b32x4_skylake_, nk_store_b128_haswell_,
+                                   nk_partial_store_b32x4_skylake_, 1)
+nk_define_cross_normalized_symmetric_(angular, e5m2, diamond, e5m2, f32, /*norm_value_type=*/f32, f32, nk_b128_vec_t,
+                                      nk_dots_symmetric_e5m2_diamond, nk_angular_through_f32_from_dot_haswell_,
+                                      nk_dots_reduce_sumsq_e5m2_, nk_load_b128_haswell_, nk_partial_load_b32x4_skylake_,
+                                      nk_store_b128_haswell_, nk_partial_store_b32x4_skylake_, 1)
+nk_define_cross_normalized_symmetric_(euclidean, e5m2, diamond, e5m2, f32, /*norm_value_type=*/f32, f32, nk_b128_vec_t,
+                                      nk_dots_symmetric_e5m2_diamond, nk_euclidean_through_f32_from_dot_haswell_,
+                                      nk_dots_reduce_sumsq_e5m2_, nk_load_b128_haswell_, nk_partial_load_b32x4_skylake_,
+                                      nk_store_b128_haswell_, nk_partial_store_b32x4_skylake_, 1)
+
+#if defined(__clang__)
+#pragma clang attribute pop
+#elif defined(__GNUC__)
+#pragma GCC pop_options
+#endif
+
+#if defined(__cplusplus)
+} // extern "C"
+#endif
+
+#endif // NK_TARGET_DIAMOND
+#endif // NK_TARGET_X86_
+#endif // NK_SPATIALS_DIAMOND_H

package/include/numkong/spatials/haswell.h

@@ -78,11 +78,11 @@ nk_define_cross_normalized_symmetric_(euclidean, f16, haswell, f16, f32, /*norm_
                                       nk_dots_reduce_sumsq_f16_, nk_load_b128_haswell_, nk_partial_load_b32x4_haswell_,
                                       nk_store_b128_haswell_, nk_partial_store_b32x4_haswell_, 1)
 
-nk_define_cross_normalized_packed_(angular, bf16, haswell, bf16, f32, f32, /*norm_value_type=*/f32, f32, nk_b128_vec_t,
+nk_define_cross_normalized_packed_(angular, bf16, haswell, bf16, bf16, f32, /*norm_value_type=*/f32, f32, nk_b128_vec_t,
                                    nk_dots_packed_bf16_haswell, nk_angular_through_f32_from_dot_haswell_,
                                    nk_dots_reduce_sumsq_bf16_, nk_load_b128_haswell_, nk_partial_load_b32x4_haswell_,
                                    nk_store_b128_haswell_, nk_partial_store_b32x4_haswell_, 1)
-nk_define_cross_normalized_packed_(euclidean, bf16, haswell, bf16, f32, f32, /*norm_value_type=*/f32, f32,
+nk_define_cross_normalized_packed_(euclidean, bf16, haswell, bf16, bf16, f32, /*norm_value_type=*/f32, f32,
                                    nk_b128_vec_t, nk_dots_packed_bf16_haswell,
                                    nk_euclidean_through_f32_from_dot_haswell_, nk_dots_reduce_sumsq_bf16_,
                                    nk_load_b128_haswell_, nk_partial_load_b32x4_haswell_, nk_store_b128_haswell_,

package/include/numkong/spatials/loongsonasx.h

@@ -0,0 +1,153 @@
+/**
+ *  @brief Batched Spatial Distances for LoongArch LASX (256-bit).
+ *  @file include/numkong/spatials/loongsonasx.h
+ *  @author Ash Vardanian
+ *  @date March 23, 2026
+ *
+ *  @sa include/numkong/spatials.h
+ */
+#ifndef NK_SPATIALS_LOONGSONASX_H
+#define NK_SPATIALS_LOONGSONASX_H
+
+#if NK_TARGET_LOONGARCH_
+#if NK_TARGET_LOONGSONASX
+
+#include "numkong/spatial/loongsonasx.h"
+#include "numkong/spatial/serial.h"
+#include "numkong/dots/loongsonasx.h"
+
+#if defined(__cplusplus)
+extern "C" {
+#endif
+
+nk_define_cross_normalized_packed_(angular, f32, loongsonasx, f32, f32, f64, /*norm_value_type=*/f64, f64,
+                                   nk_b256_vec_t, nk_dots_packed_f32_loongsonasx,
+                                   nk_angular_through_f64_from_dot_loongsonasx_, nk_dots_reduce_sumsq_f32_,
+                                   nk_load_b256_loongsonasx_, nk_partial_load_b64x4_serial_, nk_store_b256_loongsonasx_,
+                                   nk_partial_store_b64x4_serial_, 1)
+nk_define_cross_normalized_packed_(euclidean, f32, loongsonasx, f32, f32, f64, /*norm_value_type=*/f64, f64,
+                                   nk_b256_vec_t, nk_dots_packed_f32_loongsonasx,
+                                   nk_euclidean_through_f64_from_dot_loongsonasx_, nk_dots_reduce_sumsq_f32_,
+                                   nk_load_b256_loongsonasx_, nk_partial_load_b64x4_serial_, nk_store_b256_loongsonasx_,
+                                   nk_partial_store_b64x4_serial_, 1)
+nk_define_cross_normalized_symmetric_(angular, f32, loongsonasx, f32, f64, /*norm_value_type=*/f64, f64, nk_b256_vec_t,
+                                      nk_dots_symmetric_f32_loongsonasx, nk_angular_through_f64_from_dot_loongsonasx_,
+                                      nk_dots_reduce_sumsq_f32_, nk_load_b256_loongsonasx_,
+                                      nk_partial_load_b64x4_serial_, nk_store_b256_loongsonasx_,
+                                      nk_partial_store_b64x4_serial_, 1)
+nk_define_cross_normalized_symmetric_(euclidean, f32, loongsonasx, f32, f64, /*norm_value_type=*/f64, f64,
+                                      nk_b256_vec_t, nk_dots_symmetric_f32_loongsonasx,
+                                      nk_euclidean_through_f64_from_dot_loongsonasx_, nk_dots_reduce_sumsq_f32_,
+                                      nk_load_b256_loongsonasx_, nk_partial_load_b64x4_serial_,
+                                      nk_store_b256_loongsonasx_, nk_partial_store_b64x4_serial_, 1)
+
+nk_define_cross_normalized_packed_(angular, f64, loongsonasx, f64, f64, f64, /*norm_value_type=*/f64, f64,
+                                   nk_b256_vec_t, nk_dots_packed_f64_loongsonasx,
+                                   nk_angular_through_f64_from_dot_loongsonasx_, nk_dots_reduce_sumsq_f64_,
+                                   nk_load_b256_loongsonasx_, nk_partial_load_b64x4_serial_, nk_store_b256_loongsonasx_,
+                                   nk_partial_store_b64x4_serial_, 1)
+nk_define_cross_normalized_packed_(euclidean, f64, loongsonasx, f64, f64, f64, /*norm_value_type=*/f64, f64,
+                                   nk_b256_vec_t, nk_dots_packed_f64_loongsonasx,
+                                   nk_euclidean_through_f64_from_dot_loongsonasx_, nk_dots_reduce_sumsq_f64_,
+                                   nk_load_b256_loongsonasx_, nk_partial_load_b64x4_serial_, nk_store_b256_loongsonasx_,
+                                   nk_partial_store_b64x4_serial_, 1)
+nk_define_cross_normalized_symmetric_(angular, f64, loongsonasx, f64, f64, /*norm_value_type=*/f64, f64, nk_b256_vec_t,
+                                      nk_dots_symmetric_f64_loongsonasx, nk_angular_through_f64_from_dot_loongsonasx_,
+                                      nk_dots_reduce_sumsq_f64_, nk_load_b256_loongsonasx_,
+                                      nk_partial_load_b64x4_serial_, nk_store_b256_loongsonasx_,
+                                      nk_partial_store_b64x4_serial_, 1)
+nk_define_cross_normalized_symmetric_(euclidean, f64, loongsonasx, f64, f64, /*norm_value_type=*/f64, f64,
+                                      nk_b256_vec_t, nk_dots_symmetric_f64_loongsonasx,
+                                      nk_euclidean_through_f64_from_dot_loongsonasx_, nk_dots_reduce_sumsq_f64_,
+                                      nk_load_b256_loongsonasx_, nk_partial_load_b64x4_serial_,
+                                      nk_store_b256_loongsonasx_, nk_partial_store_b64x4_serial_, 1)
+
+nk_define_cross_normalized_packed_(angular, i8, loongsonasx, i8, i8, i32, /*norm_value_type=*/u32, f32, nk_b128_vec_t,
+                                   nk_dots_packed_i8_loongsonasx, nk_angular_through_i32_from_dot_loongsonasx_,
+                                   nk_dots_reduce_sumsq_i8_, nk_load_b128_loongsonasx_, nk_partial_load_b32x4_serial_,
+                                   nk_store_b128_loongsonasx_, nk_partial_store_b32x4_serial_, 1)
+nk_define_cross_normalized_symmetric_(angular, i8, loongsonasx, i8, i32, /*norm_value_type=*/u32, f32, nk_b128_vec_t,
+                                      nk_dots_symmetric_i8_loongsonasx, nk_angular_through_i32_from_dot_loongsonasx_,
+                                      nk_dots_reduce_sumsq_i8_, nk_load_b128_loongsonasx_,
+                                      nk_partial_load_b32x4_serial_, nk_store_b128_loongsonasx_,
+                                      nk_partial_store_b32x4_serial_, 1)
+
+nk_define_cross_normalized_packed_(euclidean, i8, loongsonasx, i8, i8, i32, /*norm_value_type=*/u32, f32, nk_b128_vec_t,
+                                   nk_dots_packed_i8_loongsonasx, nk_euclidean_through_i32_from_dot_loongsonasx_,
+                                   nk_dots_reduce_sumsq_i8_, nk_load_b128_loongsonasx_, nk_partial_load_b32x4_serial_,
+                                   nk_store_b128_loongsonasx_, nk_partial_store_b32x4_serial_, 1)
+nk_define_cross_normalized_symmetric_(euclidean, i8, loongsonasx, i8, i32, /*norm_value_type=*/u32, f32, nk_b128_vec_t,
+                                      nk_dots_symmetric_i8_loongsonasx, nk_euclidean_through_i32_from_dot_loongsonasx_,
+                                      nk_dots_reduce_sumsq_i8_, nk_load_b128_loongsonasx_,
+                                      nk_partial_load_b32x4_serial_, nk_store_b128_loongsonasx_,
+                                      nk_partial_store_b32x4_serial_, 1)
+
+nk_define_cross_normalized_packed_(angular, u8, loongsonasx, u8, u8, u32, /*norm_value_type=*/u32, f32, nk_b128_vec_t,
+                                   nk_dots_packed_u8_loongsonasx, nk_angular_through_u32_from_dot_loongsonasx_,
+                                   nk_dots_reduce_sumsq_u8_, nk_load_b128_loongsonasx_, nk_partial_load_b32x4_serial_,
+                                   nk_store_b128_loongsonasx_, nk_partial_store_b32x4_serial_, 1)
+nk_define_cross_normalized_symmetric_(angular, u8, loongsonasx, u8, u32, /*norm_value_type=*/u32, f32, nk_b128_vec_t,
+                                      nk_dots_symmetric_u8_loongsonasx, nk_angular_through_u32_from_dot_loongsonasx_,
+                                      nk_dots_reduce_sumsq_u8_, nk_load_b128_loongsonasx_,
+                                      nk_partial_load_b32x4_serial_, nk_store_b128_loongsonasx_,
+                                      nk_partial_store_b32x4_serial_, 1)
+
+nk_define_cross_normalized_packed_(euclidean, u8, loongsonasx, u8, u8, u32, /*norm_value_type=*/u32, f32, nk_b128_vec_t,
+                                   nk_dots_packed_u8_loongsonasx, nk_euclidean_through_u32_from_dot_loongsonasx_,
+                                   nk_dots_reduce_sumsq_u8_, nk_load_b128_loongsonasx_, nk_partial_load_b32x4_serial_,
+                                   nk_store_b128_loongsonasx_, nk_partial_store_b32x4_serial_, 1)
+nk_define_cross_normalized_symmetric_(euclidean, u8, loongsonasx, u8, u32, /*norm_value_type=*/u32, f32, nk_b128_vec_t,
+                                      nk_dots_symmetric_u8_loongsonasx, nk_euclidean_through_u32_from_dot_loongsonasx_,
+                                      nk_dots_reduce_sumsq_u8_, nk_load_b128_loongsonasx_,
+                                      nk_partial_load_b32x4_serial_, nk_store_b128_loongsonasx_,
+                                      nk_partial_store_b32x4_serial_, 1)
+
+nk_define_cross_normalized_packed_(angular, bf16, loongsonasx, bf16, bf16, f32, /*norm_value_type=*/f32, f32,
+                                   nk_b128_vec_t, nk_dots_packed_bf16_loongsonasx,
+                                   nk_angular_through_f32_from_dot_loongsonasx_, nk_dots_reduce_sumsq_bf16_,
+                                   nk_load_b128_loongsonasx_, nk_partial_load_b32x4_serial_, nk_store_b128_loongsonasx_,
+                                   nk_partial_store_b32x4_serial_, 1)
+nk_define_cross_normalized_packed_(euclidean, bf16, loongsonasx, bf16, bf16, f32, /*norm_value_type=*/f32, f32,
+                                   nk_b128_vec_t, nk_dots_packed_bf16_loongsonasx,
+                                   nk_euclidean_through_f32_from_dot_loongsonasx_, nk_dots_reduce_sumsq_bf16_,
+                                   nk_load_b128_loongsonasx_, nk_partial_load_b32x4_serial_, nk_store_b128_loongsonasx_,
+                                   nk_partial_store_b32x4_serial_, 1)
+nk_define_cross_normalized_symmetric_(angular, bf16, loongsonasx, bf16, f32, /*norm_value_type=*/f32, f32,
+                                      nk_b128_vec_t, nk_dots_symmetric_bf16_loongsonasx,
+                                      nk_angular_through_f32_from_dot_loongsonasx_, nk_dots_reduce_sumsq_bf16_,
+                                      nk_load_b128_loongsonasx_, nk_partial_load_b32x4_serial_,
+                                      nk_store_b128_loongsonasx_, nk_partial_store_b32x4_serial_, 1)
+nk_define_cross_normalized_symmetric_(euclidean, bf16, loongsonasx, bf16, f32, /*norm_value_type=*/f32, f32,
+                                      nk_b128_vec_t, nk_dots_symmetric_bf16_loongsonasx,
+                                      nk_euclidean_through_f32_from_dot_loongsonasx_, nk_dots_reduce_sumsq_bf16_,
+                                      nk_load_b128_loongsonasx_, nk_partial_load_b32x4_serial_,
+                                      nk_store_b128_loongsonasx_, nk_partial_store_b32x4_serial_, 1)
+
+nk_define_cross_normalized_packed_(angular, f16, loongsonasx, f16, f32, f32, /*norm_value_type=*/f32, f32,
+                                   nk_b128_vec_t, nk_dots_packed_f16_loongsonasx,
+                                   nk_angular_through_f32_from_dot_loongsonasx_, nk_dots_reduce_sumsq_f16_,
+                                   nk_load_b128_loongsonasx_, nk_partial_load_b32x4_serial_, nk_store_b128_loongsonasx_,
+                                   nk_partial_store_b32x4_serial_, 1)
+nk_define_cross_normalized_packed_(euclidean, f16, loongsonasx, f16, f32, f32, /*norm_value_type=*/f32, f32,
+                                   nk_b128_vec_t, nk_dots_packed_f16_loongsonasx,
+                                   nk_euclidean_through_f32_from_dot_loongsonasx_, nk_dots_reduce_sumsq_f16_,
+                                   nk_load_b128_loongsonasx_, nk_partial_load_b32x4_serial_, nk_store_b128_loongsonasx_,
+                                   nk_partial_store_b32x4_serial_, 1)
+nk_define_cross_normalized_symmetric_(angular, f16, loongsonasx, f16, f32, /*norm_value_type=*/f32, f32, nk_b128_vec_t,
+                                      nk_dots_symmetric_f16_loongsonasx, nk_angular_through_f32_from_dot_loongsonasx_,
+                                      nk_dots_reduce_sumsq_f16_, nk_load_b128_loongsonasx_,
+                                      nk_partial_load_b32x4_serial_, nk_store_b128_loongsonasx_,
+                                      nk_partial_store_b32x4_serial_, 1)
+nk_define_cross_normalized_symmetric_(euclidean, f16, loongsonasx, f16, f32, /*norm_value_type=*/f32, f32,
+                                      nk_b128_vec_t, nk_dots_symmetric_f16_loongsonasx,
+                                      nk_euclidean_through_f32_from_dot_loongsonasx_, nk_dots_reduce_sumsq_f16_,
+                                      nk_load_b128_loongsonasx_, nk_partial_load_b32x4_serial_,
+                                      nk_store_b128_loongsonasx_, nk_partial_store_b32x4_serial_, 1)
+
+#if defined(__cplusplus)
+} // extern "C"
+#endif
+
+#endif // NK_TARGET_LOONGSONASX
+#endif // NK_TARGET_LOONGARCH_
+#endif // NK_SPATIALS_LOONGSONASX_H

package/include/numkong/spatials/neonfp8.h

@@ -0,0 +1,111 @@
+/**
+ *  @brief SIMD-accelerated Batched Spatial Distances for NEON FP8DOT4.
+ *  @file include/numkong/spatials/neonfp8.h
+ *  @author Ash Vardanian
+ *  @date March 23, 2026
+ *
+ *  @sa include/numkong/spatials.h
+ *
+ *  Uses FDOT (FEAT_FP8DOT4) for native FP8 dot products, then derives angular
+ *  and Euclidean distances via the batched dots/ infrastructure.
+ */
+#ifndef NK_SPATIALS_NEONFP8_H
+#define NK_SPATIALS_NEONFP8_H
+
+#if NK_TARGET_ARM_
+#if NK_TARGET_NEONFP8
+
+#include "numkong/dots/neonfp8.h"
+
+#if defined(__cplusplus)
+extern "C" {
+#endif
+
+#if defined(__clang__)
+#pragma clang attribute push(__attribute__((target("arch=armv8-a+simd+fp8dot4"))), apply_to = function)
+#elif defined(__GNUC__)
+#pragma GCC push_options
+#pragma GCC target("arch=armv8-a+simd+fp8dot4")
+#endif
+
+nk_define_cross_normalized_packed_(angular, e4m3, neonfp8, e4m3, e4m3, f32, f32, f32, nk_b128_vec_t,
+                                   nk_dots_packed_e4m3_neonfp8, nk_angular_through_f32_from_dot_neon_,
+                                   nk_dots_reduce_sumsq_e4m3_, nk_load_b128_neon_, nk_partial_load_b32x4_serial_,
+                                   nk_store_b128_neon_, nk_partial_store_b32x4_serial_, 1)
+nk_define_cross_normalized_packed_(euclidean, e4m3, neonfp8, e4m3, e4m3, f32, f32, f32, nk_b128_vec_t,
+                                   nk_dots_packed_e4m3_neonfp8, nk_euclidean_through_f32_from_dot_neon_,
+                                   nk_dots_reduce_sumsq_e4m3_, nk_load_b128_neon_, nk_partial_load_b32x4_serial_,
+                                   nk_store_b128_neon_, nk_partial_store_b32x4_serial_, 1)
+nk_define_cross_normalized_symmetric_(angular, e4m3, neonfp8, e4m3, f32, f32, f32, nk_b128_vec_t,
+                                      nk_dots_symmetric_e4m3_neonfp8, nk_angular_through_f32_from_dot_neon_,
+                                      nk_dots_reduce_sumsq_e4m3_, nk_load_b128_neon_, nk_partial_load_b32x4_serial_,
+                                      nk_store_b128_neon_, nk_partial_store_b32x4_serial_, 1)
+nk_define_cross_normalized_symmetric_(euclidean, e4m3, neonfp8, e4m3, f32, f32, f32, nk_b128_vec_t,
+                                      nk_dots_symmetric_e4m3_neonfp8, nk_euclidean_through_f32_from_dot_neon_,
+                                      nk_dots_reduce_sumsq_e4m3_, nk_load_b128_neon_, nk_partial_load_b32x4_serial_,
+                                      nk_store_b128_neon_, nk_partial_store_b32x4_serial_, 1)
+
+nk_define_cross_normalized_packed_(angular, e5m2, neonfp8, e5m2, e5m2, f32, f32, f32, nk_b128_vec_t,
+                                   nk_dots_packed_e5m2_neonfp8, nk_angular_through_f32_from_dot_neon_,
+                                   nk_dots_reduce_sumsq_e5m2_, nk_load_b128_neon_, nk_partial_load_b32x4_serial_,
+                                   nk_store_b128_neon_, nk_partial_store_b32x4_serial_, 1)
+nk_define_cross_normalized_packed_(euclidean, e5m2, neonfp8, e5m2, e5m2, f32, f32, f32, nk_b128_vec_t,
+                                   nk_dots_packed_e5m2_neonfp8, nk_euclidean_through_f32_from_dot_neon_,
+                                   nk_dots_reduce_sumsq_e5m2_, nk_load_b128_neon_, nk_partial_load_b32x4_serial_,
+                                   nk_store_b128_neon_, nk_partial_store_b32x4_serial_, 1)
+nk_define_cross_normalized_symmetric_(angular, e5m2, neonfp8, e5m2, f32, f32, f32, nk_b128_vec_t,
+                                      nk_dots_symmetric_e5m2_neonfp8, nk_angular_through_f32_from_dot_neon_,
+                                      nk_dots_reduce_sumsq_e5m2_, nk_load_b128_neon_, nk_partial_load_b32x4_serial_,
+                                      nk_store_b128_neon_, nk_partial_store_b32x4_serial_, 1)
+nk_define_cross_normalized_symmetric_(euclidean, e5m2, neonfp8, e5m2, f32, f32, f32, nk_b128_vec_t,
+                                      nk_dots_symmetric_e5m2_neonfp8, nk_euclidean_through_f32_from_dot_neon_,
+                                      nk_dots_reduce_sumsq_e5m2_, nk_load_b128_neon_, nk_partial_load_b32x4_serial_,
+                                      nk_store_b128_neon_, nk_partial_store_b32x4_serial_, 1)
+
+nk_define_cross_normalized_packed_(angular, e2m3, neonfp8, e2m3, e2m3, f32, f32, f32, nk_b128_vec_t,
+                                   nk_dots_packed_e2m3_neonfp8, nk_angular_through_f32_from_dot_neon_,
+                                   nk_dots_reduce_sumsq_e2m3_, nk_load_b128_neon_, nk_partial_load_b32x4_serial_,
+                                   nk_store_b128_neon_, nk_partial_store_b32x4_serial_, 1)
+nk_define_cross_normalized_packed_(euclidean, e2m3, neonfp8, e2m3, e2m3, f32, f32, f32, nk_b128_vec_t,
+                                   nk_dots_packed_e2m3_neonfp8, nk_euclidean_through_f32_from_dot_neon_,
+                                   nk_dots_reduce_sumsq_e2m3_, nk_load_b128_neon_, nk_partial_load_b32x4_serial_,
+                                   nk_store_b128_neon_, nk_partial_store_b32x4_serial_, 1)
+nk_define_cross_normalized_symmetric_(angular, e2m3, neonfp8, e2m3, f32, f32, f32, nk_b128_vec_t,
+                                      nk_dots_symmetric_e2m3_neonfp8, nk_angular_through_f32_from_dot_neon_,
+                                      nk_dots_reduce_sumsq_e2m3_, nk_load_b128_neon_, nk_partial_load_b32x4_serial_,
+                                      nk_store_b128_neon_, nk_partial_store_b32x4_serial_, 1)
+nk_define_cross_normalized_symmetric_(euclidean, e2m3, neonfp8, e2m3, f32, f32, f32, nk_b128_vec_t,
+                                      nk_dots_symmetric_e2m3_neonfp8, nk_euclidean_through_f32_from_dot_neon_,
+                                      nk_dots_reduce_sumsq_e2m3_, nk_load_b128_neon_, nk_partial_load_b32x4_serial_,
+                                      nk_store_b128_neon_, nk_partial_store_b32x4_serial_, 1)
+
+nk_define_cross_normalized_packed_(angular, e3m2, neonfp8, e3m2, e3m2, f32, f32, f32, nk_b128_vec_t,
+                                   nk_dots_packed_e3m2_neonfp8, nk_angular_through_f32_from_dot_neon_,
+                                   nk_dots_reduce_sumsq_e3m2_, nk_load_b128_neon_, nk_partial_load_b32x4_serial_,
+                                   nk_store_b128_neon_, nk_partial_store_b32x4_serial_, 1)
+nk_define_cross_normalized_packed_(euclidean, e3m2, neonfp8, e3m2, e3m2, f32, f32, f32, nk_b128_vec_t,
+                                   nk_dots_packed_e3m2_neonfp8, nk_euclidean_through_f32_from_dot_neon_,
+                                   nk_dots_reduce_sumsq_e3m2_, nk_load_b128_neon_, nk_partial_load_b32x4_serial_,
+                                   nk_store_b128_neon_, nk_partial_store_b32x4_serial_, 1)
+nk_define_cross_normalized_symmetric_(angular, e3m2, neonfp8, e3m2, f32, f32, f32, nk_b128_vec_t,
+                                      nk_dots_symmetric_e3m2_neonfp8, nk_angular_through_f32_from_dot_neon_,
+                                      nk_dots_reduce_sumsq_e3m2_, nk_load_b128_neon_, nk_partial_load_b32x4_serial_,
+                                      nk_store_b128_neon_, nk_partial_store_b32x4_serial_, 1)
+nk_define_cross_normalized_symmetric_(euclidean, e3m2, neonfp8, e3m2, f32, f32, f32, nk_b128_vec_t,
+                                      nk_dots_symmetric_e3m2_neonfp8, nk_euclidean_through_f32_from_dot_neon_,
+                                      nk_dots_reduce_sumsq_e3m2_, nk_load_b128_neon_, nk_partial_load_b32x4_serial_,
+                                      nk_store_b128_neon_, nk_partial_store_b32x4_serial_, 1)
+
+#if defined(__clang__)
+#pragma clang attribute pop
+#elif defined(__GNUC__)
+#pragma GCC pop_options
+#endif
+
+#if defined(__cplusplus)
+} // extern "C"
+#endif
+
+#endif // NK_TARGET_NEONFP8
+#endif // NK_TARGET_ARM_
+#endif // NK_SPATIALS_NEONFP8_H

package/include/numkong/spatials/neonsdot.h

@@ -94,6 +94,40 @@ nk_define_cross_normalized_symmetric_(euclidean, u4, neonsdot, u4x2, u32, /*norm
                                       nk_dots_reduce_sumsq_u4_, nk_load_b128_neon_, nk_partial_load_b32x4_serial_,
                                       nk_store_b128_neon_, nk_partial_store_b32x4_serial_, 2)
 
+nk_define_cross_normalized_packed_(angular, e2m3, neonsdot, e2m3, e2m3, f32, /*norm_value_type=*/f32, f32,
+                                   nk_b128_vec_t, nk_dots_packed_e2m3_neonsdot, nk_angular_through_f32_from_dot_neon_,
+                                   nk_dots_reduce_sumsq_e2m3_, nk_load_b128_neon_, nk_partial_load_b32x4_serial_,
+                                   nk_store_b128_neon_, nk_partial_store_b32x4_serial_, 1)
+nk_define_cross_normalized_packed_(euclidean, e2m3, neonsdot, e2m3, e2m3, f32, /*norm_value_type=*/f32, f32,
+                                   nk_b128_vec_t, nk_dots_packed_e2m3_neonsdot, nk_euclidean_through_f32_from_dot_neon_,
+                                   nk_dots_reduce_sumsq_e2m3_, nk_load_b128_neon_, nk_partial_load_b32x4_serial_,
+                                   nk_store_b128_neon_, nk_partial_store_b32x4_serial_, 1)
+nk_define_cross_normalized_symmetric_(angular, e2m3, neonsdot, e2m3, f32, /*norm_value_type=*/f32, f32, nk_b128_vec_t,
+                                      nk_dots_symmetric_e2m3_neonsdot, nk_angular_through_f32_from_dot_neon_,
+                                      nk_dots_reduce_sumsq_e2m3_, nk_load_b128_neon_, nk_partial_load_b32x4_serial_,
+                                      nk_store_b128_neon_, nk_partial_store_b32x4_serial_, 1)
+nk_define_cross_normalized_symmetric_(euclidean, e2m3, neonsdot, e2m3, f32, /*norm_value_type=*/f32, f32, nk_b128_vec_t,
+                                      nk_dots_symmetric_e2m3_neonsdot, nk_euclidean_through_f32_from_dot_neon_,
+                                      nk_dots_reduce_sumsq_e2m3_, nk_load_b128_neon_, nk_partial_load_b32x4_serial_,
+                                      nk_store_b128_neon_, nk_partial_store_b32x4_serial_, 1)
+
+nk_define_cross_normalized_packed_(angular, e3m2, neonsdot, e3m2, e3m2, f32, /*norm_value_type=*/f32, f32,
+                                   nk_b128_vec_t, nk_dots_packed_e3m2_neonsdot, nk_angular_through_f32_from_dot_neon_,
+                                   nk_dots_reduce_sumsq_e3m2_, nk_load_b128_neon_, nk_partial_load_b32x4_serial_,
+                                   nk_store_b128_neon_, nk_partial_store_b32x4_serial_, 1)
+nk_define_cross_normalized_packed_(euclidean, e3m2, neonsdot, e3m2, e3m2, f32, /*norm_value_type=*/f32, f32,
+                                   nk_b128_vec_t, nk_dots_packed_e3m2_neonsdot, nk_euclidean_through_f32_from_dot_neon_,
+                                   nk_dots_reduce_sumsq_e3m2_, nk_load_b128_neon_, nk_partial_load_b32x4_serial_,
+                                   nk_store_b128_neon_, nk_partial_store_b32x4_serial_, 1)
+nk_define_cross_normalized_symmetric_(angular, e3m2, neonsdot, e3m2, f32, /*norm_value_type=*/f32, f32, nk_b128_vec_t,
+                                      nk_dots_symmetric_e3m2_neonsdot, nk_angular_through_f32_from_dot_neon_,
+                                      nk_dots_reduce_sumsq_e3m2_, nk_load_b128_neon_, nk_partial_load_b32x4_serial_,
+                                      nk_store_b128_neon_, nk_partial_store_b32x4_serial_, 1)
+nk_define_cross_normalized_symmetric_(euclidean, e3m2, neonsdot, e3m2, f32, /*norm_value_type=*/f32, f32, nk_b128_vec_t,
+                                      nk_dots_symmetric_e3m2_neonsdot, nk_euclidean_through_f32_from_dot_neon_,
+                                      nk_dots_reduce_sumsq_e3m2_, nk_load_b128_neon_, nk_partial_load_b32x4_serial_,
+                                      nk_store_b128_neon_, nk_partial_store_b32x4_serial_, 1)
+
 #if defined(__clang__)
 #pragma clang attribute pop
 #elif defined(__GNUC__)

package/include/numkong/spatials/powervsx.h

@@ -0,0 +1,153 @@
+/**
+ *  @brief Batched Spatial Distances for Power VSX.
+ *  @file include/numkong/spatials/powervsx.h
+ *  @author Ash Vardanian
+ *  @date March 23, 2026
+ *
+ *  @sa include/numkong/spatials.h
+ */
+#ifndef NK_SPATIALS_POWERVSX_H
+#define NK_SPATIALS_POWERVSX_H
+
+#if NK_TARGET_POWER_
+#if NK_TARGET_POWERVSX
+
+#include "numkong/spatial/powervsx.h"
+#include "numkong/dots/powervsx.h"
+
+#if defined(__clang__)
+#pragma clang attribute push(__attribute__((target("power9-vector"))), apply_to = function)
+#elif defined(__GNUC__)
+#pragma GCC push_options
+#pragma GCC target("power9-vector")
+#endif
+
+#if defined(__cplusplus)
+extern "C" {
+#endif
+
+nk_define_cross_normalized_packed_(angular, f32, powervsx, f32, f32, f64, /*norm_value_type=*/f64, f64, nk_b256_vec_t,
+                                   nk_dots_packed_f32_powervsx, nk_angular_through_f64_from_dot_powervsx_,
+                                   nk_dots_reduce_sumsq_f32_, nk_load_b256_powervsx_, nk_partial_load_b64x4_powervsx_,
+                                   nk_store_b256_powervsx_, nk_partial_store_b64x4_powervsx_, 1)
+nk_define_cross_normalized_packed_(euclidean, f32, powervsx, f32, f32, f64, /*norm_value_type=*/f64, f64, nk_b256_vec_t,
+                                   nk_dots_packed_f32_powervsx, nk_euclidean_through_f64_from_dot_powervsx_,
+                                   nk_dots_reduce_sumsq_f32_, nk_load_b256_powervsx_, nk_partial_load_b64x4_powervsx_,
+                                   nk_store_b256_powervsx_, nk_partial_store_b64x4_powervsx_, 1)
+nk_define_cross_normalized_symmetric_(angular, f32, powervsx, f32, f64, /*norm_value_type=*/f64, f64, nk_b256_vec_t,
+                                      nk_dots_symmetric_f32_powervsx, nk_angular_through_f64_from_dot_powervsx_,
+                                      nk_dots_reduce_sumsq_f32_, nk_load_b256_powervsx_,
+                                      nk_partial_load_b64x4_powervsx_, nk_store_b256_powervsx_,
+                                      nk_partial_store_b64x4_powervsx_, 1)
+nk_define_cross_normalized_symmetric_(euclidean, f32, powervsx, f32, f64, /*norm_value_type=*/f64, f64, nk_b256_vec_t,
+                                      nk_dots_symmetric_f32_powervsx, nk_euclidean_through_f64_from_dot_powervsx_,
+                                      nk_dots_reduce_sumsq_f32_, nk_load_b256_powervsx_,
+                                      nk_partial_load_b64x4_powervsx_, nk_store_b256_powervsx_,
+                                      nk_partial_store_b64x4_powervsx_, 1)
+
+nk_define_cross_normalized_packed_(angular, bf16, powervsx, bf16, bf16, f32, /*norm_value_type=*/f32, f32,
+                                   nk_b128_vec_t, nk_dots_packed_bf16_powervsx,
+                                   nk_angular_through_f32_from_dot_powervsx_, nk_dots_reduce_sumsq_bf16_,
+                                   nk_load_b128_powervsx_, nk_partial_load_b32x4_powervsx_, nk_store_b128_powervsx_,
+                                   nk_partial_store_b32x4_powervsx_, 1)
+nk_define_cross_normalized_packed_(euclidean, bf16, powervsx, bf16, bf16, f32, /*norm_value_type=*/f32, f32,
+                                   nk_b128_vec_t, nk_dots_packed_bf16_powervsx,
+                                   nk_euclidean_through_f32_from_dot_powervsx_, nk_dots_reduce_sumsq_bf16_,
+                                   nk_load_b128_powervsx_, nk_partial_load_b32x4_powervsx_, nk_store_b128_powervsx_,
+                                   nk_partial_store_b32x4_powervsx_, 1)
+nk_define_cross_normalized_symmetric_(angular, bf16, powervsx, bf16, f32, /*norm_value_type=*/f32, f32, nk_b128_vec_t,
+                                      nk_dots_symmetric_bf16_powervsx, nk_angular_through_f32_from_dot_powervsx_,
+                                      nk_dots_reduce_sumsq_bf16_, nk_load_b128_powervsx_,
+                                      nk_partial_load_b32x4_powervsx_, nk_store_b128_powervsx_,
+                                      nk_partial_store_b32x4_powervsx_, 1)
+nk_define_cross_normalized_symmetric_(euclidean, bf16, powervsx, bf16, f32, /*norm_value_type=*/f32, f32, nk_b128_vec_t,
+                                      nk_dots_symmetric_bf16_powervsx, nk_euclidean_through_f32_from_dot_powervsx_,
+                                      nk_dots_reduce_sumsq_bf16_, nk_load_b128_powervsx_,
+                                      nk_partial_load_b32x4_powervsx_, nk_store_b128_powervsx_,
+                                      nk_partial_store_b32x4_powervsx_, 1)
+
+nk_define_cross_normalized_packed_(angular, f16, powervsx, f16, f16, f32, /*norm_value_type=*/f32, f32, nk_b128_vec_t,
+                                   nk_dots_packed_f16_powervsx, nk_angular_through_f32_from_dot_powervsx_,
+                                   nk_dots_reduce_sumsq_f16_, nk_load_b128_powervsx_, nk_partial_load_b32x4_powervsx_,
+                                   nk_store_b128_powervsx_, nk_partial_store_b32x4_powervsx_, 1)
+nk_define_cross_normalized_packed_(euclidean, f16, powervsx, f16, f16, f32, /*norm_value_type=*/f32, f32, nk_b128_vec_t,
+                                   nk_dots_packed_f16_powervsx, nk_euclidean_through_f32_from_dot_powervsx_,
+                                   nk_dots_reduce_sumsq_f16_, nk_load_b128_powervsx_, nk_partial_load_b32x4_powervsx_,
+                                   nk_store_b128_powervsx_, nk_partial_store_b32x4_powervsx_, 1)
+nk_define_cross_normalized_symmetric_(angular, f16, powervsx, f16, f32, /*norm_value_type=*/f32, f32, nk_b128_vec_t,
+                                      nk_dots_symmetric_f16_powervsx, nk_angular_through_f32_from_dot_powervsx_,
+                                      nk_dots_reduce_sumsq_f16_, nk_load_b128_powervsx_,
+                                      nk_partial_load_b32x4_powervsx_, nk_store_b128_powervsx_,
+                                      nk_partial_store_b32x4_powervsx_, 1)
+nk_define_cross_normalized_symmetric_(euclidean, f16, powervsx, f16, f32, /*norm_value_type=*/f32, f32, nk_b128_vec_t,
+                                      nk_dots_symmetric_f16_powervsx, nk_euclidean_through_f32_from_dot_powervsx_,
+                                      nk_dots_reduce_sumsq_f16_, nk_load_b128_powervsx_,
+                                      nk_partial_load_b32x4_powervsx_, nk_store_b128_powervsx_,
+                                      nk_partial_store_b32x4_powervsx_, 1)
+
+nk_define_cross_normalized_packed_(angular, i8, powervsx, i8, i8, i32, /*norm_value_type=*/u32, f32, nk_b128_vec_t,
+                                   nk_dots_packed_i8_powervsx, nk_angular_through_i32_from_dot_powervsx_,
+                                   nk_dots_reduce_sumsq_i8_, nk_load_b128_powervsx_, nk_partial_load_b32x4_powervsx_,
+                                   nk_store_b128_powervsx_, nk_partial_store_b32x4_powervsx_, 1)
+nk_define_cross_normalized_packed_(euclidean, i8, powervsx, i8, i8, i32, /*norm_value_type=*/u32, f32, nk_b128_vec_t,
+                                   nk_dots_packed_i8_powervsx, nk_euclidean_through_i32_from_dot_powervsx_,
+                                   nk_dots_reduce_sumsq_i8_, nk_load_b128_powervsx_, nk_partial_load_b32x4_powervsx_,
+                                   nk_store_b128_powervsx_, nk_partial_store_b32x4_powervsx_, 1)
+nk_define_cross_normalized_symmetric_(angular, i8, powervsx, i8, i32, /*norm_value_type=*/u32, f32, nk_b128_vec_t,
+                                      nk_dots_symmetric_i8_powervsx, nk_angular_through_i32_from_dot_powervsx_,
+                                      nk_dots_reduce_sumsq_i8_, nk_load_b128_powervsx_, nk_partial_load_b32x4_powervsx_,
+                                      nk_store_b128_powervsx_, nk_partial_store_b32x4_powervsx_, 1)
+nk_define_cross_normalized_symmetric_(euclidean, i8, powervsx, i8, i32, /*norm_value_type=*/u32, f32, nk_b128_vec_t,
+                                      nk_dots_symmetric_i8_powervsx, nk_euclidean_through_i32_from_dot_powervsx_,
+                                      nk_dots_reduce_sumsq_i8_, nk_load_b128_powervsx_, nk_partial_load_b32x4_powervsx_,
+                                      nk_store_b128_powervsx_, nk_partial_store_b32x4_powervsx_, 1)
+
+nk_define_cross_normalized_packed_(angular, u8, powervsx, u8, u8, u32, /*norm_value_type=*/u32, f32, nk_b128_vec_t,
+                                   nk_dots_packed_u8_powervsx, nk_angular_through_u32_from_dot_powervsx_,
+                                   nk_dots_reduce_sumsq_u8_, nk_load_b128_powervsx_, nk_partial_load_b32x4_powervsx_,
+                                   nk_store_b128_powervsx_, nk_partial_store_b32x4_powervsx_, 1)
+nk_define_cross_normalized_packed_(euclidean, u8, powervsx, u8, u8, u32, /*norm_value_type=*/u32, f32, nk_b128_vec_t,
+                                   nk_dots_packed_u8_powervsx, nk_euclidean_through_u32_from_dot_powervsx_,
+                                   nk_dots_reduce_sumsq_u8_, nk_load_b128_powervsx_, nk_partial_load_b32x4_powervsx_,
+                                   nk_store_b128_powervsx_, nk_partial_store_b32x4_powervsx_, 1)
+nk_define_cross_normalized_symmetric_(angular, u8, powervsx, u8, u32, /*norm_value_type=*/u32, f32, nk_b128_vec_t,
+                                      nk_dots_symmetric_u8_powervsx, nk_angular_through_u32_from_dot_powervsx_,
+                                      nk_dots_reduce_sumsq_u8_, nk_load_b128_powervsx_, nk_partial_load_b32x4_powervsx_,
+                                      nk_store_b128_powervsx_, nk_partial_store_b32x4_powervsx_, 1)
+nk_define_cross_normalized_symmetric_(euclidean, u8, powervsx, u8, u32, /*norm_value_type=*/u32, f32, nk_b128_vec_t,
+                                      nk_dots_symmetric_u8_powervsx, nk_euclidean_through_u32_from_dot_powervsx_,
+                                      nk_dots_reduce_sumsq_u8_, nk_load_b128_powervsx_, nk_partial_load_b32x4_powervsx_,
+                                      nk_store_b128_powervsx_, nk_partial_store_b32x4_powervsx_, 1)
+
+nk_define_cross_normalized_packed_(angular, f64, powervsx, f64, f64, f64, /*norm_value_type=*/f64, f64, nk_b256_vec_t,
+                                   nk_dots_packed_f64_powervsx, nk_angular_through_f64_from_dot_powervsx_,
+                                   nk_dots_reduce_sumsq_f64_, nk_load_b256_powervsx_, nk_partial_load_b64x4_powervsx_,
+                                   nk_store_b256_powervsx_, nk_partial_store_b64x4_powervsx_, 1)
+nk_define_cross_normalized_packed_(euclidean, f64, powervsx, f64, f64, f64, /*norm_value_type=*/f64, f64, nk_b256_vec_t,
+                                   nk_dots_packed_f64_powervsx, nk_euclidean_through_f64_from_dot_powervsx_,
+                                   nk_dots_reduce_sumsq_f64_, nk_load_b256_powervsx_, nk_partial_load_b64x4_powervsx_,
+                                   nk_store_b256_powervsx_, nk_partial_store_b64x4_powervsx_, 1)
+nk_define_cross_normalized_symmetric_(angular, f64, powervsx, f64, f64, /*norm_value_type=*/f64, f64, nk_b256_vec_t,
+                                      nk_dots_symmetric_f64_powervsx, nk_angular_through_f64_from_dot_powervsx_,
+                                      nk_dots_reduce_sumsq_f64_, nk_load_b256_powervsx_,
+                                      nk_partial_load_b64x4_powervsx_, nk_store_b256_powervsx_,
+                                      nk_partial_store_b64x4_powervsx_, 1)
+nk_define_cross_normalized_symmetric_(euclidean, f64, powervsx, f64, f64, /*norm_value_type=*/f64, f64, nk_b256_vec_t,
+                                      nk_dots_symmetric_f64_powervsx, nk_euclidean_through_f64_from_dot_powervsx_,
+                                      nk_dots_reduce_sumsq_f64_, nk_load_b256_powervsx_,
+                                      nk_partial_load_b64x4_powervsx_, nk_store_b256_powervsx_,
+                                      nk_partial_store_b64x4_powervsx_, 1)
+
+#if defined(__cplusplus)
+} // extern "C"
+#endif
+
+#if defined(__clang__)
+#pragma clang attribute pop
+#elif defined(__GNUC__)
+#pragma GCC pop_options
+#endif
+
+#endif // NK_TARGET_POWERVSX
+#endif // NK_TARGET_POWER_
+#endif // NK_SPATIALS_POWERVSX_H