usearch 2.15.1 → 2.15.3

package/simsimd/include/simsimd/types.h

@@ -35,7 +35,7 @@
 #define SIMSIMD_PUBLIC __attribute__((unused)) inline static
 #define SIMSIMD_INTERNAL __attribute__((always_inline)) inline static
 #else
-#define SIMSIMD_DYNAMIC inline static
+#define SIMSIMD_DYNAMIC
 #define SIMSIMD_PUBLIC inline static
 #define SIMSIMD_INTERNAL inline static
 #endif
@@ -98,6 +98,16 @@
 #define SIMSIMD_TARGET_SVE_BF16 SIMSIMD_TARGET_SVE
 #endif // !defined(SIMSIMD_TARGET_SVE_BF16)
 
+// Compiling for Arm: SIMSIMD_TARGET_SVE2
+#if !defined(SIMSIMD_TARGET_SVE2) || (SIMSIMD_TARGET_SVE2 && !SIMSIMD_TARGET_ARM)
+#if defined(__ARM_FEATURE_SVE)
+#define SIMSIMD_TARGET_SVE2 SIMSIMD_TARGET_ARM
+#else
+#undef SIMSIMD_TARGET_SVE2
+#define SIMSIMD_TARGET_SVE2 0
+#endif // defined(__ARM_FEATURE_SVE)
+#endif // !defined(SIMSIMD_TARGET_SVE2)
+
 // Compiling for x86: SIMSIMD_TARGET_HASWELL
 //
 // Starting with Ivy Bridge, Intel supports the `F16C` extensions for fast half-precision
@@ -114,7 +124,7 @@
 #endif // defined(__AVX2__)
 #endif // !defined(SIMSIMD_TARGET_HASWELL)
 
-// Compiling for x86: SIMSIMD_TARGET_SKYLAKE, SIMSIMD_TARGET_ICE, SIMSIMD_TARGET_SAPPHIRE
+// Compiling for x86: SIMSIMD_TARGET_SKYLAKE, SIMSIMD_TARGET_ICE, SIMSIMD_TARGET_GENOA, SIMSIMD_TARGET_SAPPHIRE
 //
 // To list all available macros for x86, take a recent compiler, like GCC 12 and run:
 //      gcc-12 -march=sapphirerapids -dM -E - < /dev/null | egrep "SSE|AVX" | sort
@@ -163,31 +173,37 @@
 #include <arm_neon.h>
 #endif
 
-#if SIMSIMD_TARGET_SVE
+#if SIMSIMD_TARGET_SVE || SIMSIMD_TARGET_SVE2
 #include <arm_sve.h>
 #endif
 
-#if SIMSIMD_TARGET_HASWELL || SIMSIMD_TARGET_SKYLAKE
+#if SIMSIMD_TARGET_HASWELL || SIMSIMD_TARGET_SKYLAKE || SIMSIMD_TARGET_ICE || SIMSIMD_TARGET_GENOA ||                  \
+    SIMSIMD_TARGET_SAPPHIRE
 #include <immintrin.h>
 #endif
 
 #endif
 
-#ifndef SIMSIMD_RSQRT
+#if !defined(SIMSIMD_SQRT)
 #include <math.h>
-#define SIMSIMD_RSQRT(x) (1 / sqrtf(x))
+#define SIMSIMD_SQRT(x) (sqrt(x))
 #endif
 
-#ifndef SIMSIMD_LOG
+#if !defined(SIMSIMD_RSQRT)
 #include <math.h>
-#define SIMSIMD_LOG(x) (logf(x))
+#define SIMSIMD_RSQRT(x) (1 / SIMSIMD_SQRT(x))
 #endif
 
-#ifndef SIMSIMD_F32_DIVISION_EPSILON
+#if !defined(SIMSIMD_LOG)
+#include <math.h>
+#define SIMSIMD_LOG(x) (log(x))
+#endif
+
+#if !defined(SIMSIMD_F32_DIVISION_EPSILON)
 #define SIMSIMD_F32_DIVISION_EPSILON (1e-7)
 #endif
 
-#ifndef SIMSIMD_F16_DIVISION_EPSILON
+#if !defined(SIMSIMD_F16_DIVISION_EPSILON)
 #define SIMSIMD_F16_DIVISION_EPSILON (1e-3)
 #endif
 
@@ -195,14 +211,20 @@
 extern "C" {
 #endif
 
-typedef int simsimd_i32_t;
-typedef float simsimd_f32_t;
-typedef double simsimd_f64_t;
-typedef signed char simsimd_i8_t;
 typedef unsigned char simsimd_b8_t;
-typedef long long simsimd_i64_t;
+
+typedef signed char simsimd_i8_t;
+typedef unsigned char simsimd_u8_t;
+typedef signed short simsimd_i16_t;
+typedef unsigned short simsimd_u16_t;
+typedef signed int simsimd_i32_t;
+typedef unsigned int simsimd_u32_t;
+typedef signed long long simsimd_i64_t;
 typedef unsigned long long simsimd_u64_t;
 
+typedef float simsimd_f32_t;
+typedef double simsimd_f64_t;
+
 typedef simsimd_u64_t simsimd_size_t;
 typedef simsimd_f64_t simsimd_distance_t;
 
@@ -215,17 +237,19 @@ typedef simsimd_f64_t simsimd_distance_t;
 #if !defined(SIMSIMD_NATIVE_F16) || SIMSIMD_NATIVE_F16
 #if (defined(__GNUC__) || defined(__clang__)) && (defined(__ARM_ARCH) || defined(__aarch64__)) &&                      \
     (defined(__ARM_FP16_FORMAT_IEEE))
-#if !defined(SIMSIMD_NATIVE_F16)
+#undef SIMSIMD_NATIVE_F16
 #define SIMSIMD_NATIVE_F16 1
-#endif
 typedef __fp16 simsimd_f16_t;
 #elif ((defined(__GNUC__) || defined(__clang__)) && (defined(__x86_64__) || defined(__i386__)) &&                      \
-       (defined(__SSE2__) || defined(__AVX512F__)))
+       (defined(__AVX512FP16__)))
 typedef _Float16 simsimd_f16_t;
-#if !defined(SIMSIMD_NATIVE_F16)
+#undef SIMSIMD_NATIVE_F16
 #define SIMSIMD_NATIVE_F16 1
+#else                                       // Unknown compiler or architecture
+#if defined(__GNUC__) || defined(__clang__) // Some compilers don't support warning pragmas
+#warning "Unknown compiler or architecture for float16."
 #endif
-#else // Unknown compiler or architecture
+#undef SIMSIMD_NATIVE_F16
 #define SIMSIMD_NATIVE_F16 0
 #endif // Unknown compiler or architecture
 #endif // !SIMSIMD_NATIVE_F16
@@ -242,23 +266,38 @@ typedef unsigned short simsimd_f16_t;
  *  - GCC or Clang on 64-bit x86: `_BFloat16`.
  *  - Default: `unsigned short`.
  *
+ *  The compilers have added `__bf16` support in compliance with the x86-64 psABI spec.
+ *  The motivation for this new special type is summed up as:
+ *
+ *      Currently `__bfloat16` is a typedef of short, which creates a problem where the
+ *      compiler does not raise any alarms if it is used to add, subtract, multiply or
+ *      divide, but the result of the calculation is actually meaningless.
+ *      To solve this problem, a real scalar type `__Bfloat16` needs to be introduced.
+ *      It is mainly used for intrinsics, not available for C standard operators.
+ *      `__Bfloat16` will also be used for movement like passing parameter, load and store,
+ *      vector initialization, vector shuffle, and etc. It creates a need for a
+ *      corresponding psABI.
+ *
  *  @warning Apple Clang has hard time with bf16.
  *  https://developer.apple.com/documentation/xcode/writing-arm64-code-for-apple-platforms
  *  https://forums.developer.apple.com/forums/thread/726201
+ *  https://www.phoronix.com/news/GCC-LLVM-bf16-BFloat16-Type
  */
 #if (defined(__GNUC__) || defined(__clang__)) && (defined(__ARM_ARCH) || defined(__aarch64__)) &&                      \
     (defined(__ARM_BF16_FORMAT_ALTERNATIVE))
-#if !defined(SIMSIMD_NATIVE_BF16)
+#undef SIMSIMD_NATIVE_BF16
 #define SIMSIMD_NATIVE_BF16 1
-#endif
-typedef __fp16 simsimd_bf16_t;
+typedef __bf16 simsimd_bf16_t;
 #elif ((defined(__GNUC__) || defined(__clang__)) && (defined(__x86_64__) || defined(__i386__)) &&                      \
-       (defined(__SSE2__) || defined(__AVX512F__)))
-typedef _Float16 simsimd_bf16_t;
-#if !defined(SIMSIMD_NATIVE_BF16)
+       (defined(__AVX512BF16__)))
+typedef __bfloat16 simsimd_bf16_t;
+#undef SIMSIMD_NATIVE_BF16
 #define SIMSIMD_NATIVE_BF16 1
+#else                                       // Unknown compiler or architecture
+#if defined(__GNUC__) || defined(__clang__) // Some compilers don't support warning pragmas
+#warning "Unknown compiler or architecture for bfloat16."
 #endif
-#else // Unknown compiler or architecture
+#undef SIMSIMD_NATIVE_BF16
 #define SIMSIMD_NATIVE_BF16 0
 #endif // Unknown compiler or architecture
 #endif // !SIMSIMD_NATIVE_BF16
@@ -286,17 +325,38 @@ typedef unsigned short simsimd_bf16_t;
 #endif
 #endif
 
-#define SIMSIMD_IDENTIFY(x) (x)
+/*
+ *  Let's make sure the sizes of the types are as expected.
+ *  In C the `_Static_assert` is only available with C 11 and later.
+ */
+#define SIMSIMD_STATIC_ASSERT(cond, msg) typedef char static_assertion_##msg[(cond) ? 1 : -1]
+SIMSIMD_STATIC_ASSERT(sizeof(simsimd_i8_t) == 1, simsimd_i8_t_must_be_1_byte);
+SIMSIMD_STATIC_ASSERT(sizeof(simsimd_u8_t) == 1, simsimd_u8_t_must_be_1_byte);
+SIMSIMD_STATIC_ASSERT(sizeof(simsimd_i16_t) == 2, simsimd_i16_t_must_be_2_bytes);
+SIMSIMD_STATIC_ASSERT(sizeof(simsimd_u16_t) == 2, simsimd_u16_t_must_be_2_bytes);
+SIMSIMD_STATIC_ASSERT(sizeof(simsimd_i32_t) == 4, simsimd_i32_t_must_be_4_bytes);
+SIMSIMD_STATIC_ASSERT(sizeof(simsimd_u32_t) == 4, simsimd_u32_t_must_be_4_bytes);
+SIMSIMD_STATIC_ASSERT(sizeof(simsimd_i64_t) == 8, simsimd_i64_t_must_be_8_bytes);
+SIMSIMD_STATIC_ASSERT(sizeof(simsimd_u64_t) == 8, simsimd_u64_t_must_be_8_bytes);
+SIMSIMD_STATIC_ASSERT(sizeof(simsimd_f32_t) == 4, simsimd_f32_t_must_be_4_bytes);
+SIMSIMD_STATIC_ASSERT(sizeof(simsimd_f64_t) == 8, simsimd_f64_t_must_be_8_bytes);
+SIMSIMD_STATIC_ASSERT(sizeof(simsimd_f16_t) == 2, simsimd_f16_t_must_be_2_bytes);
+SIMSIMD_STATIC_ASSERT(sizeof(simsimd_bf16_t) == 2, simsimd_bf16_t_must_be_2_bytes);
+
+#define SIMSIMD_DEREFERENCE(x) (*(x))
+#define SIMSIMD_EXPORT(x, y) *(y) = x
 
 /**
  *  @brief  Returns the value of the half-precision floating-point number,
  *          potentially decompressed into single-precision.
  */
-#ifndef SIMSIMD_UNCOMPRESS_F16
+#if !defined(SIMSIMD_F16_TO_F32)
 #if SIMSIMD_NATIVE_F16
-#define SIMSIMD_UNCOMPRESS_F16(x) (SIMSIMD_IDENTIFY(x))
+#define SIMSIMD_F16_TO_F32(x) (SIMSIMD_DEREFERENCE(x))
+#define SIMSIMD_F32_TO_F16(x, y) (SIMSIMD_EXPORT(x, y))
 #else
-#define SIMSIMD_UNCOMPRESS_F16(x) (simsimd_uncompress_f16(x))
+#define SIMSIMD_F16_TO_F32(x) (simsimd_f16_to_f32(x))
+#define SIMSIMD_F32_TO_F16(x, y) (simsimd_f32_to_f16(x, y))
 #endif
 #endif
 
@@ -304,22 +364,34 @@ typedef unsigned short simsimd_bf16_t;
  *  @brief  Returns the value of the half-precision brain floating-point number,
  *          potentially decompressed into single-precision.
  */
-#ifndef SIMSIMD_UNCOMPRESS_BF16
+#if !defined(SIMSIMD_BF16_TO_F32)
 #if SIMSIMD_NATIVE_BF16
-#define SIMSIMD_UNCOMPRESS_BF16(x) (SIMSIMD_IDENTIFY(x))
+#define SIMSIMD_BF16_TO_F32(x) (SIMSIMD_DEREFERENCE(x))
+#define SIMSIMD_F32_TO_BF16(x, y) (SIMSIMD_EXPORT(x, y))
 #else
-#define SIMSIMD_UNCOMPRESS_BF16(x) (simsimd_uncompress_bf16(x))
+#define SIMSIMD_BF16_TO_F32(x) (simsimd_bf16_to_f32(x))
+#define SIMSIMD_F32_TO_BF16(x, y) (simsimd_f32_to_bf16(x, y))
 #endif
 #endif
 
+/** @brief  Convenience type for half-precision floating-point type conversions. */
 typedef union {
     unsigned i;
     float f;
 } simsimd_f32i32_t;
 
 /**
- *  @brief  Computes `1/sqrt(x)` using the trick from Quake 3, replacing
- *          magic numbers with the ones suggested by Jan Kadlec.
+ *  @brief  Computes `1/sqrt(x)` using the trick from Quake 3,
+ *          replacing the magic numbers with the ones suggested by Jan Kadlec.
+ *
+ *  Subsequent additions by hardware manufacturers have made this algorithm redundant for the most part.
+ *  For example, on x86, Intel introduced the SSE instruction `rsqrtss` in 1999. In a 2009 benchmark on
+ *  the Intel Core 2, this instruction took 0.85ns per float compared to 3.54ns for the fast inverse
+ *  square root algorithm, and had less error. Carmack's Magic Number `rsqrt` had an average error
+ *  of 0.0990%, while SSE `rsqrtss` had 0.0094%, a 10x improvement.
+ *
+ *  https://web.archive.org/web/20210208132927/http://assemblyrequired.crashworks.org/timing-square-root/
+ *  https://stackoverflow.com/a/41460625/2766161
  */
 SIMSIMD_PUBLIC simsimd_f32_t simsimd_approximate_inverse_square_root(simsimd_f32_t number) {
     simsimd_f32i32_t conv;
@@ -351,20 +423,17 @@ SIMSIMD_PUBLIC simsimd_f32_t simsimd_approximate_log(simsimd_f32_t number) {
  *  https://gist.github.com/milhidaka/95863906fe828198f47991c813dbe233
  *  https://github.com/OpenCyphal/libcanard/blob/636795f4bc395f56af8d2c61d3757b5e762bb9e5/canard.c#L811-L834
  */
-SIMSIMD_PUBLIC simsimd_f32_t simsimd_uncompress_f16(unsigned short x) {
-    union float_or_unsigned_int_t {
-        float f;
-        unsigned int i;
-    };
+SIMSIMD_PUBLIC simsimd_f32_t simsimd_f16_to_f32(simsimd_f16_t const* x_ptr) {
+    unsigned short x = *(unsigned short const*)x_ptr;
     unsigned int exponent = (x & 0x7C00) >> 10;
     unsigned int mantissa = (x & 0x03FF) << 13;
-    union float_or_unsigned_int_t mantissa_union;
-    mantissa_union.f = (float)mantissa;
-    unsigned int v = (mantissa_union.i) >> 23;
-    union float_or_unsigned_int_t result_union;
-    result_union.i = (x & 0x8000) << 16 | (exponent != 0) * ((exponent + 112) << 23 | mantissa) |
-                     ((exponent == 0) & (mantissa != 0)) * ((v - 37) << 23 | ((mantissa << (150 - v)) & 0x007FE000));
-    return result_union.f;
+    simsimd_f32i32_t mantissa_conv;
+    mantissa_conv.f = (float)mantissa;
+    unsigned int v = (mantissa_conv.i) >> 23;
+    simsimd_f32i32_t conv;
+    conv.i = (x & 0x8000) << 16 | (exponent != 0) * ((exponent + 112) << 23 | mantissa) |
+             ((exponent == 0) & (mantissa != 0)) * ((v - 37) << 23 | ((mantissa << (150 - v)) & 0x007FE000));
+    return conv.f;
 }
 
 /**
@@ -376,20 +445,16 @@ SIMSIMD_PUBLIC simsimd_f32_t simsimd_uncompress_f16(unsigned short x) {
  *  https://gist.github.com/milhidaka/95863906fe828198f47991c813dbe233
  *  https://github.com/OpenCyphal/libcanard/blob/636795f4bc395f56af8d2c61d3757b5e762bb9e5/canard.c#L811-L834
  */
-SIMSIMD_PUBLIC unsigned short simsimd_compress_f16(simsimd_f32_t x) {
-    union float_or_unsigned_int_t {
-        float f;
-        unsigned int i;
-    } x_union;
-    x_union.f = x;
-
-    unsigned int b = x_union.i + 0x00001000;
+SIMSIMD_PUBLIC void simsimd_f32_to_f16(simsimd_f32_t x, simsimd_f16_t* result_ptr) {
+    simsimd_f32i32_t conv;
+    conv.f = x;
+    unsigned int b = conv.i + 0x00001000;
     unsigned int e = (b & 0x7F800000) >> 23;
     unsigned int m = b & 0x007FFFFF;
     unsigned short result = ((b & 0x80000000) >> 16) | (e > 112) * ((((e - 112) << 10) & 0x7C00) | (m >> 13)) |
                             ((e < 113) & (e > 101)) * ((((0x007FF000 + m) >> (125 - e)) + 1) >> 1) |
                             ((e > 143) * 0x7FFF);
-    return result;
+    *(unsigned short*)result_ptr = result;
 }
 
 /**
@@ -399,31 +464,36 @@ SIMSIMD_PUBLIC unsigned short simsimd_compress_f16(simsimd_f32_t x) {
  *  https://stackoverflow.com/questions/55253233/convert-fp32-to-bfloat16-in-c/55254307#55254307
  *  https://cloud.google.com/blog/products/ai-machine-learning/bfloat16-the-secret-to-high-performance-on-cloud-tpus
  */
-SIMSIMD_PUBLIC simsimd_f32_t simsimd_uncompress_bf16(unsigned short x) {
-    union float_or_unsigned_int_t {
-        float f;
-        unsigned int i;
-    };
-    union float_or_unsigned_int_t result_union;
-    result_union.i = x << 16; // Zero extends the mantissa
-    return result_union.f;
+SIMSIMD_PUBLIC simsimd_f32_t simsimd_bf16_to_f32(simsimd_bf16_t const* x_ptr) {
+    unsigned short x = *(unsigned short const*)x_ptr;
+    simsimd_f32i32_t conv;
+    conv.i = x << 16; // Zero extends the mantissa
+    return conv.f;
 }
 
 /**
  *  @brief  Compresses a `float` to a `bf16` representation.
+ *
+ *  https://stackoverflow.com/questions/55253233/convert-fp32-to-bfloat16-in-c/55254307#55254307
+ *  https://cloud.google.com/blog/products/ai-machine-learning/bfloat16-the-secret-to-high-performance-on-cloud-tpus
  */
-SIMSIMD_PUBLIC unsigned short simsimd_compress_bf16(simsimd_f32_t x) {
-    union float_or_unsigned_int_t {
-        float f;
-        unsigned int i;
-    };
-    union float_or_unsigned_int_t value;
-    value.f = x;
-    value.i >>= 16;
-    value.i &= 0xFFFF;
-    return (unsigned short)value.i;
+SIMSIMD_PUBLIC void simsimd_f32_to_bf16(simsimd_f32_t x, simsimd_bf16_t* result_ptr) {
+    simsimd_f32i32_t conv;
+    conv.f = x;
+    conv.i += 0x8000; // Rounding is optional
+    conv.i >>= 16;
+    // The top 16 bits will be zeroed out anyways
+    // conv.i &= 0xFFFF;
+    *(unsigned short*)result_ptr = (unsigned short)conv.i;
 }
 
+SIMSIMD_PUBLIC simsimd_u32_t simsimd_u32_rol(simsimd_u32_t x, int n) { return (x << n) | (x >> (32 - n)); }
+SIMSIMD_PUBLIC simsimd_u16_t simsimd_u16_rol(simsimd_u16_t x, int n) { return (x << n) | (x >> (16 - n)); }
+SIMSIMD_PUBLIC simsimd_u8_t simsimd_u8_rol(simsimd_u8_t x, int n) { return (x << n) | (x >> (8 - n)); }
+SIMSIMD_PUBLIC simsimd_u32_t simsimd_u32_ror(simsimd_u32_t x, int n) { return (x >> n) | (x << (32 - n)); }
+SIMSIMD_PUBLIC simsimd_u16_t simsimd_u16_ror(simsimd_u16_t x, int n) { return (x >> n) | (x << (16 - n)); }
+SIMSIMD_PUBLIC simsimd_u8_t simsimd_u8_ror(simsimd_u8_t x, int n) { return (x >> n) | (x << (8 - n)); }
+
 #ifdef __cplusplus
 } // extern "C"
 #endif

package/simsimd/javascript/simsimd.ts

@@ -83,20 +83,20 @@ export const jaccard = (a: Uint8Array, b: Uint8Array): number => {
 };
 
 /**
- * @brief Computes the kullbackleibler similarity coefficient between two vectors.
+ * @brief Computes the Kullback-Leibler divergence between two vectors.
  * @param {Float64Array|Float32Array} a - The first vector.
  * @param {Float64Array|Float32Array} b - The second vector.
- * @returns {number} The Jaccard similarity coefficient between vectors a and b.
+ * @returns {number} The Kullback-Leibler divergence between vectors a and b.
  */
 export const kullbackleibler = (a: Float64Array | Float32Array, b: Float64Array | Float32Array): number => {
   return compiled.kullbackleibler(a, b);
 };
 
 /**
- * @brief Computes the jensenshannon similarity coefficient between two vectors.
+ * @brief Computes the Jensen-Shannon divergence between two vectors.
  * @param {Float64Array|Float32Array} a - The first vector.
  * @param {Float64Array|Float32Array} b - The second vector.
- * @returns {number} The Jaccard similarity coefficient between vectors a and b.
+ * @returns {number} The Jensen-Shannon divergence between vectors a and b.
  */
 export const jensenshannon = (a: Float64Array | Float32Array, b: Float64Array | Float32Array): number => {
   return compiled.jensenshannon(a, b);

package/simsimd/package.json

@@ -1,6 +1,6 @@
 {
   "name": "simsimd",
-  "version": "5.0.0",
+  "version": "5.6.4",
   "description": "Fastest SIMD-Accelerated Vector Similarity Functions for x86 and Arm",
   "homepage": "https://github.com/ashvardanian/simsimd",
   "author": "Ash Vardanian",

package/simsimd/rust/lib.rs

@@ -54,24 +54,29 @@ type ComplexProduct = (f64, f64);
 extern "C" {
 
     fn simsimd_dot_f16(a: *const u16, b: *const u16, c: usize, d: *mut Distance);
+    fn simsimd_dot_bf16(a: *const u16, b: *const u16, c: usize, d: *mut Distance);
     fn simsimd_dot_f32(a: *const f32, b: *const f32, c: usize, d: *mut Distance);
     fn simsimd_dot_f64(a: *const f64, b: *const f64, c: usize, d: *mut Distance);
 
     fn simsimd_dot_f16c(a: *const u16, b: *const u16, c: usize, d: *mut Distance);
+    fn simsimd_dot_bf16c(a: *const u16, b: *const u16, c: usize, d: *mut Distance);
     fn simsimd_dot_f32c(a: *const f32, b: *const f32, c: usize, d: *mut Distance);
     fn simsimd_dot_f64c(a: *const f64, b: *const f64, c: usize, d: *mut Distance);
 
     fn simsimd_vdot_f16c(a: *const u16, b: *const u16, c: usize, d: *mut Distance);
+    fn simsimd_vdot_bf16c(a: *const u16, b: *const u16, c: usize, d: *mut Distance);
     fn simsimd_vdot_f32c(a: *const f32, b: *const f32, c: usize, d: *mut Distance);
     fn simsimd_vdot_f64c(a: *const f64, b: *const f64, c: usize, d: *mut Distance);
 
     fn simsimd_cos_i8(a: *const i8, b: *const i8, c: usize, d: *mut Distance);
     fn simsimd_cos_f16(a: *const u16, b: *const u16, c: usize, d: *mut Distance);
+    fn simsimd_cos_bf16(a: *const u16, b: *const u16, c: usize, d: *mut Distance);
     fn simsimd_cos_f32(a: *const f32, b: *const f32, c: usize, d: *mut Distance);
     fn simsimd_cos_f64(a: *const f64, b: *const f64, c: usize, d: *mut Distance);
 
     fn simsimd_l2sq_i8(a: *const i8, b: *const i8, c: usize, d: *mut Distance);
     fn simsimd_l2sq_f16(a: *const u16, b: *const u16, c: usize, d: *mut Distance);
+    fn simsimd_l2sq_bf16(a: *const u16, b: *const u16, c: usize, d: *mut Distance);
     fn simsimd_l2sq_f32(a: *const f32, b: *const f32, c: usize, d: *mut Distance);
     fn simsimd_l2sq_f64(a: *const f64, b: *const f64, c: usize, d: *mut Distance);
 
@@ -79,10 +84,12 @@ extern "C" {
     fn simsimd_jaccard_b8(a: *const u8, b: *const u8, c: usize, d: *mut Distance);
 
     fn simsimd_js_f16(a: *const u16, b: *const u16, c: usize, d: *mut Distance);
+    fn simsimd_js_bf16(a: *const u16, b: *const u16, c: usize, d: *mut Distance);
     fn simsimd_js_f32(a: *const f32, b: *const f32, c: usize, d: *mut Distance);
     fn simsimd_js_f64(a: *const f64, b: *const f64, c: usize, d: *mut Distance);
 
     fn simsimd_kl_f16(a: *const u16, b: *const u16, c: usize, d: *mut Distance);
+    fn simsimd_kl_bf16(a: *const u16, b: *const u16, c: usize, d: *mut Distance);
     fn simsimd_kl_f32(a: *const f32, b: *const f32, c: usize, d: *mut Distance);
     fn simsimd_kl_f64(a: *const f64, b: *const f64, c: usize, d: *mut Distance);
 
@@ -107,6 +114,12 @@ pub struct f16(u16);
 
 impl f16 {}
 
+/// A half-precision floating point number, called brain float.
+#[repr(transparent)]
+pub struct bf16(u16);
+
+impl bf16 {}
+
 /// The `capabilities` module provides functions for detecting the hardware features
 /// available on the current system.
 pub mod capabilities {
@@ -370,6 +383,50 @@ impl SpatialSimilarity for f16 {
     }
 }
 
+impl SpatialSimilarity for bf16 {
+    fn cos(a: &[Self], b: &[Self]) -> Option<Distance> {
+        if a.len() != b.len() {
+            return None;
+        }
+
+        // Explicitly cast `*const bf16` to `*const u16`
+        let a_ptr = a.as_ptr() as *const u16;
+        let b_ptr = b.as_ptr() as *const u16;
+        let mut distance_value: Distance = 0.0;
+        let distance_ptr: *mut Distance = &mut distance_value as *mut Distance;
+        unsafe { simsimd_cos_bf16(a_ptr, b_ptr, a.len(), distance_ptr) };
+        Some(distance_value)
+    }
+
+    fn dot(a: &[Self], b: &[Self]) -> Option<Distance> {
+        if a.len() != b.len() {
+            return None;
+        }
+
+        // Explicitly cast `*const bf16` to `*const u16`
+        let a_ptr = a.as_ptr() as *const u16;
+        let b_ptr = b.as_ptr() as *const u16;
+        let mut distance_value: Distance = 0.0;
+        let distance_ptr: *mut Distance = &mut distance_value as *mut Distance;
+        unsafe { simsimd_dot_bf16(a_ptr, b_ptr, a.len(), distance_ptr) };
+        Some(distance_value)
+    }
+
+    fn l2sq(a: &[Self], b: &[Self]) -> Option<Distance> {
+        if a.len() != b.len() {
+            return None;
+        }
+
+        // Explicitly cast `*const bf16` to `*const u16`
+        let a_ptr = a.as_ptr() as *const u16;
+        let b_ptr = b.as_ptr() as *const u16;
+        let mut distance_value: Distance = 0.0;
+        let distance_ptr: *mut Distance = &mut distance_value as *mut Distance;
+        unsafe { simsimd_l2sq_bf16(a_ptr, b_ptr, a.len(), distance_ptr) };
+        Some(distance_value)
+    }
+}
+
 impl SpatialSimilarity for f32 {
     fn cos(a: &[Self], b: &[Self]) -> Option<Distance> {
         if a.len() != b.len() {
@@ -464,6 +521,36 @@ impl ProbabilitySimilarity for f16 {
     }
 }
 
+impl ProbabilitySimilarity for bf16 {
+    fn jensenshannon(a: &[Self], b: &[Self]) -> Option<Distance> {
+        if a.len() != b.len() {
+            return None;
+        }
+
+        // Explicitly cast `*const bf16` to `*const u16`
+        let a_ptr = a.as_ptr() as *const u16;
+        let b_ptr = b.as_ptr() as *const u16;
+        let mut distance_value: Distance = 0.0;
+        let distance_ptr: *mut Distance = &mut distance_value as *mut Distance;
+        unsafe { simsimd_js_bf16(a_ptr, b_ptr, a.len(), distance_ptr) };
+        Some(distance_value)
+    }
+
+    fn kullbackleibler(a: &[Self], b: &[Self]) -> Option<Distance> {
+        if a.len() != b.len() {
+            return None;
+        }
+
+        // Explicitly cast `*const bf16` to `*const u16`
+        let a_ptr = a.as_ptr() as *const u16;
+        let b_ptr = b.as_ptr() as *const u16;
+        let mut distance_value: Distance = 0.0;
+        let distance_ptr: *mut Distance = &mut distance_value as *mut Distance;
+        unsafe { simsimd_kl_bf16(a_ptr, b_ptr, a.len(), distance_ptr) };
+        Some(distance_value)
+    }
+}
+
 impl ProbabilitySimilarity for f32 {
     fn jensenshannon(a: &[Self], b: &[Self]) -> Option<Distance> {
         if a.len() != b.len() {
@@ -536,6 +623,36 @@ impl ComplexProducts for f16 {
     }
 }
 
+impl ComplexProducts for bf16 {
+    fn dot(a: &[Self], b: &[Self]) -> Option<ComplexProduct> {
+        if a.len() != b.len() {
+            return None;
+        }
+        // Prepare the output array where the real and imaginary parts will be stored
+        let mut product: [Distance; 2] = [0.0, 0.0];
+        let product_ptr: *mut Distance = &mut product[0] as *mut _;
+        // Explicitly cast `*const bf16` to `*const u16`
+        let a_ptr = a.as_ptr() as *const u16;
+        let b_ptr = b.as_ptr() as *const u16;
+        unsafe { simsimd_dot_bf16c(a_ptr, b_ptr, a.len(), product_ptr) };
+        Some((product[0], product[1]))
+    }
+
+    fn vdot(a: &[Self], b: &[Self]) -> Option<ComplexProduct> {
+        if a.len() != b.len() {
+            return None;
+        }
+        // Prepare the output array where the real and imaginary parts will be stored
+        let mut product: [Distance; 2] = [0.0, 0.0];
+        let product_ptr: *mut Distance = &mut product[0] as *mut _;
+        // Explicitly cast `*const bf16` to `*const u16`
+        let a_ptr = a.as_ptr() as *const u16;
+        let b_ptr = b.as_ptr() as *const u16;
+        unsafe { simsimd_vdot_bf16c(a_ptr, b_ptr, a.len(), product_ptr) };
+        Some((product[0], product[1]))
+    }
+}
+
 impl ComplexProducts for f32 {
     fn dot(a: &[Self], b: &[Self]) -> Option<ComplexProduct> {
         if a.len() != b.len() {
@@ -583,6 +700,7 @@ impl ComplexProducts for f64 {
 #[cfg(test)]
 mod tests {
     use super::*;
+    use half::bf16 as HalfBF16;
     use half::f16 as HalfF16;
 
     #[test]
@@ -786,6 +904,26 @@ mod tests {
         }
     }
 
+    #[test]
+    fn test_cos_bf16_same() {
+        // Assuming these u16 values represent bf16 bit patterns, and they are identical
+        let a_u16: &[u16] = &[15360, 16384, 17408]; // Corresponding to some bf16 values
+        let b_u16: &[u16] = &[15360, 16384, 17408]; // Same as above for simplicity
+
+        // Reinterpret cast from &[u16] to &[bf16]
+        // SAFETY: This is safe as long as the representations are guaranteed to be identical,
+        // which they are for transparent structs wrapping the same type.
+        let a_bf16: &[bf16] =
+            unsafe { std::slice::from_raw_parts(a_u16.as_ptr() as *const bf16, a_u16.len()) };
+        let b_bf16: &[bf16] =
+            unsafe { std::slice::from_raw_parts(b_u16.as_ptr() as *const bf16, b_u16.len()) };
+
+        if let Some(result) = SpatialSimilarity::cosine(a_bf16, b_bf16) {
+            println!("The result of cos_bf16 is {:.8}", result);
+            assert_almost_equal(0.0, result, 0.01); // Example value, adjust according to actual expected value
+        }
+    }
+
     #[test]
     fn test_cos_f16_interop() {
         let a_half: Vec<HalfF16> = vec![1.0, 2.0, 3.0]
@@ -812,4 +950,31 @@ mod tests {
             assert_almost_equal(0.025, result, 0.01);
         }
     }
+
+    #[test]
+    fn test_cos_bf16_interop() {
+        let a_half: Vec<HalfBF16> = vec![1.0, 2.0, 3.0]
+            .iter()
+            .map(|&x| HalfBF16::from_f32(x))
+            .collect();
+        let b_half: Vec<HalfBF16> = vec![4.0, 5.0, 6.0]
+            .iter()
+            .map(|&x| HalfBF16::from_f32(x))
+            .collect();
+
+        // SAFETY: This is safe as long as the memory representations are guaranteed to be identical,
+        // which they are due to both being #[repr(transparent)] wrappers around u16.
+        let a_simsimd: &[bf16] =
+            unsafe { std::slice::from_raw_parts(a_half.as_ptr() as *const bf16, a_half.len()) };
+        let b_simsimd: &[bf16] =
+            unsafe { std::slice::from_raw_parts(b_half.as_ptr() as *const bf16, b_half.len()) };
+
+        // Use the reinterpret-casted slices with your SpatialSimilarity implementation
+        if let Some(result) = SpatialSimilarity::cosine(a_simsimd, b_simsimd) {
+            // Expected value might need adjustment depending on actual cosine functionality
+            // Assuming identical vectors yield cosine similarity of 1.0
+            println!("The result of cos_bf16 (interop) is {:.8}", result);
+            assert_almost_equal(0.025, result, 0.01);
+        }
+    }
 }