| Line | Branch | Exec | Source |
|---|---|---|---|
| 1 | /* | ||
| 2 | * Copyright (c) 2026 Tiger Data, Inc. | ||
| 3 | * Licensed under the PostgreSQL License. See LICENSE for details. | ||
| 4 | * | ||
| 5 | * vec16.h - Half-precision (float16) vector type and type dispatch | ||
| 6 | * | ||
| 7 | * Provides: | ||
| 8 | * - Platform-adaptive half type (F16C, _Float16, or uint16_t fallback) | ||
| 9 | * - Vec16 struct (binary-compatible with pgvector's HalfVector) | ||
| 10 | * - Vec32TypeOps vtable for type-generic k-means and quantization | ||
| 11 | * - Scalar and bulk conversion between half and float | ||
| 12 | * | ||
| 13 | * The vtable enables mixed-type distance computation (vec_type × float32 | ||
| 14 | * centroid) so that halfvec inputs read half the memory without bulk | ||
| 15 | * conversion. Full float32 conversion is only done where unavoidable | ||
| 16 | * (CBLAS sgemm, RaBitQ rotation). | ||
| 17 | */ | ||
| 18 | |||
| 19 | #ifndef VEC16_H | ||
| 20 | #define VEC16_H | ||
| 21 | |||
| 22 | #include <float.h> | ||
| 23 | #include <math.h> | ||
| 24 | #include <stdbool.h> | ||
| 25 | #include <stdint.h> | ||
| 26 | #include <string.h> | ||
| 27 | |||
| 28 | #include "core/types.h" | ||
| 29 | |||
| 30 | /* ---------------------------------------------------------------- | ||
| 31 | * Half type detection (independent flags, prefer _Float16) | ||
| 32 | * | ||
| 33 | * _Float16 is preferred when available because the compiler can | ||
| 34 | * auto-vectorize (float)h casts into bulk vcvtph2ps (8 at a time), | ||
| 35 | * whereas _cvtsh_ss is a scalar intrinsic (1 at a time). With | ||
| 36 | * -mf16c, the compiler emits F16C instructions for _Float16 casts. | ||
| 37 | * | ||
| 38 | * F16C intrinsics are the fallback when _Float16 is not supported | ||
| 39 | * but -mf16c is available (e.g. older compilers). | ||
| 40 | * ---------------------------------------------------------------- */ | ||
| 41 | |||
| 42 | #if defined(__FLT16_MAX__) && !defined(__FreeBSD__) && \ | ||
| 43 | (!defined(__i386__) || defined(__SSE2__)) | ||
| 44 | #define VS_FLT16_SUPPORT | ||
| 45 | #endif | ||
| 46 | |||
| 47 | /* The F16C vtable below is built on the x86-64 AVX2 helpers in | ||
| 48 | * algo/simd_utils.h; 32-bit x86 converts halves through the portable | ||
| 49 | * paths instead. */ | ||
| 50 | #if defined(__F16C__) && (defined(__x86_64__) || defined(_M_X64)) | ||
| 51 | #include <immintrin.h> | ||
| 52 | #define VS_F16C_SUPPORT | ||
| 53 | #endif | ||
| 54 | |||
| 55 | #ifdef VS_FLT16_SUPPORT | ||
| 56 | typedef _Float16 half; | ||
| 57 | #define VS_HALF_MAX FLT16_MAX | ||
| 58 | #else | ||
| 59 | typedef uint16_t half; | ||
| 60 | #define VS_HALF_MAX 65504 | ||
| 61 | #endif | ||
| 62 | |||
| 63 | /* ---------------------------------------------------------------- | ||
| 64 | * Scalar conversion (inline, 3-tier) | ||
| 65 | * ---------------------------------------------------------------- */ | ||
| 66 | |||
| 67 | VS_VTABLE_INLINE float | ||
| 68 | 3735490 | vs_half_to_float(half h) | |
| 69 | { | ||
| 70 | #if defined(VS_FLT16_SUPPORT) | ||
| 71 | 180479 | return (float)h; | |
| 72 | #elif defined(VS_F16C_SUPPORT) | ||
| 73 | return _cvtsh_ss(h); | ||
| 74 | #else | ||
| 75 | /* IEEE 754 bit manipulation fallback */ | ||
| 76 | uint16_t bits = h; | ||
| 77 | uint32_t sign = (uint32_t)(bits >> 15) << 31; | ||
| 78 | uint32_t exp = (bits >> 10) & 0x1F; | ||
| 79 | uint32_t frac = bits & 0x3FF; | ||
| 80 | uint32_t f32; | ||
| 81 | |||
| 82 | if (exp == 0) | ||
| 83 | { | ||
| 84 | if (frac == 0) | ||
| 85 | { | ||
| 86 | /* ±zero */ | ||
| 87 | f32 = sign; | ||
| 88 | } | ||
| 89 | else | ||
| 90 | { | ||
| 91 | /* subnormal: normalize */ | ||
| 92 | exp = 1; | ||
| 93 | while ((frac & 0x400) == 0) | ||
| 94 | { | ||
| 95 | frac <<= 1; | ||
| 96 | exp--; | ||
| 97 | } | ||
| 98 | frac &= 0x3FF; | ||
| 99 | f32 = sign | ((uint32_t)(exp + 127 - 15) << 23) | | ||
| 100 | ((uint32_t)frac << 13); | ||
| 101 | } | ||
| 102 | } | ||
| 103 | else if (exp == 31) | ||
| 104 | { | ||
| 105 | /* Inf or NaN */ | ||
| 106 | f32 = sign | 0x7F800000u | ((uint32_t)frac << 13); | ||
| 107 | } | ||
| 108 | else | ||
| 109 | { | ||
| 110 | /* normal */ | ||
| 111 | f32 = sign | ((uint32_t)(exp + 127 - 15) << 23) | | ||
| 112 | ((uint32_t)frac << 13); | ||
| 113 | } | ||
| 114 | |||
| 115 | float result; | ||
| 116 | memcpy(&result, &f32, sizeof(float)); | ||
| 117 | return result; | ||
| 118 | #endif | ||
| 119 | } | ||
| 120 | |||
| 121 | static inline half | ||
| 122 | 12533 | vs_float_to_half(float f) | |
| 123 | { | ||
| 124 | #if defined(VS_FLT16_SUPPORT) | ||
| 125 | 12533 | return (_Float16)f; | |
| 126 | #elif defined(VS_F16C_SUPPORT) | ||
| 127 | return _cvtss_sh(f, _MM_FROUND_TO_NEAREST_INT | _MM_FROUND_NO_EXC); | ||
| 128 | #else | ||
| 129 | /* IEEE 754 bit manipulation fallback */ | ||
| 130 | uint32_t bits; | ||
| 131 | memcpy(&bits, &f, sizeof(uint32_t)); | ||
| 132 | |||
| 133 | uint32_t sign = (bits >> 16) & 0x8000; | ||
| 134 | int32_t exp = ((bits >> 23) & 0xFF) - 127 + 15; | ||
| 135 | uint32_t frac = bits & 0x7FFFFF; | ||
| 136 | |||
| 137 | if (exp <= 0) | ||
| 138 | { | ||
| 139 | if (exp < -10) | ||
| 140 | { | ||
| 141 | /* Too small, flush to zero */ | ||
| 142 | return (half)(sign); | ||
| 143 | } | ||
| 144 | /* Subnormal in half precision */ | ||
| 145 | frac |= 0x800000; | ||
| 146 | uint32_t shift = (uint32_t)(1 - exp); | ||
| 147 | /* Round to nearest, ties to even */ | ||
| 148 | uint32_t round_bit = 1u << (shift + 12); | ||
| 149 | frac += round_bit >> 1; | ||
| 150 | if ((frac & round_bit) && (frac & (round_bit - 1)) == 0) | ||
| 151 | frac &= ~(round_bit >> 1); /* tie: round to even */ | ||
| 152 | return (half)(sign | (frac >> (shift + 13))); | ||
| 153 | } | ||
| 154 | else if (exp >= 31) | ||
| 155 | { | ||
| 156 | if (exp == 31 && frac != 0) | ||
| 157 | { | ||
| 158 | /* NaN: preserve some significand bits */ | ||
| 159 | return (half)(sign | 0x7E00 | (frac >> 13)); | ||
| 160 | } | ||
| 161 | /* Overflow: Inf */ | ||
| 162 | return (half)(sign | 0x7C00); | ||
| 163 | } | ||
| 164 | |||
| 165 | /* Round to nearest, ties to even */ | ||
| 166 | frac += 0x1000; /* round bit at position 12 */ | ||
| 167 | if (frac & 0x800000) | ||
| 168 | { | ||
| 169 | frac = 0; | ||
| 170 | exp++; | ||
| 171 | if (exp >= 31) | ||
| 172 | return (half)(sign | 0x7C00); /* overflow to Inf */ | ||
| 173 | } | ||
| 174 | |||
| 175 | return (half)(sign | ((uint32_t)exp << 10) | (frac >> 13)); | ||
| 176 | #endif | ||
| 177 | } | ||
| 178 | |||
| 179 | /* ---------------------------------------------------------------- | ||
| 180 | * Special value checks | ||
| 181 | * ---------------------------------------------------------------- */ | ||
| 182 | |||
| 183 | static inline bool | ||
| 184 | 8 | vs_half_is_nan(half h) | |
| 185 | { | ||
| 186 | #ifdef VS_FLT16_SUPPORT | ||
| 187 | 8 | return isnan(h); | |
| 188 | #else | ||
| 189 | uint16_t bits = h; | ||
| 190 | return ((bits & 0x7C00) == 0x7C00) && ((bits & 0x03FF) != 0); | ||
| 191 | #endif | ||
| 192 | } | ||
| 193 | |||
| 194 | static inline bool | ||
| 195 | 10 | vs_half_is_inf(half h) | |
| 196 | { | ||
| 197 | #ifdef VS_FLT16_SUPPORT | ||
| 198 |
2/2✓ Branch 0 taken 4 times.
✓ Branch 1 taken 6 times.
|
10 | return isinf(h); |
| 199 | #else | ||
| 200 | uint16_t bits = h; | ||
| 201 | return ((bits & 0x7FFF) == 0x7C00); | ||
| 202 | #endif | ||
| 203 | } | ||
| 204 | |||
| 205 | static inline bool | ||
| 206 | 10 | vs_half_is_zero(half h) | |
| 207 | { | ||
| 208 | #ifdef VS_FLT16_SUPPORT | ||
| 209 | 10 | return h == (_Float16)0; | |
| 210 | #else | ||
| 211 | return (h & 0x7FFF) == 0; | ||
| 212 | #endif | ||
| 213 | } | ||
| 214 | |||
| 215 | /* ---------------------------------------------------------------- | ||
| 216 | * Vec16 (binary-compatible with pgvector HalfVector) | ||
| 217 | * ---------------------------------------------------------------- */ | ||
| 218 | |||
| 219 | typedef struct Vec16 | ||
| 220 | { | ||
| 221 | int32_t vl_len_; /* varlena header / size in standalone mode */ | ||
| 222 | int16_t dim; /* number of dimensions */ | ||
| 223 | int16_t unused; /* reserved for future use, always zero */ | ||
| 224 | half x[]; /* flexible array member */ | ||
| 225 | } Vec16; | ||
| 226 | |||
| 227 | #define VEC16_SIZE(dim) (offsetof(Vec16, x) + sizeof(half) * (dim)) | ||
| 228 | #define VEC16_DIM(v) ((v)->dim) | ||
| 229 | #define VEC16_DATA(v) ((v)->x) | ||
| 230 | |||
| 231 | /* ---------------------------------------------------------------- | ||
| 232 | * Bulk conversion (SIMD-dispatched in vec16.c) | ||
| 233 | * ---------------------------------------------------------------- */ | ||
| 234 | |||
| 235 | void vs_half_to_float_array(const half *src, float *dst, uint32_t n); | ||
| 236 | void vs_float_to_half_array(const float *src, half *dst, uint32_t n); | ||
| 237 | |||
| 238 | /* ---------------------------------------------------------------- | ||
| 239 | * Vec16 lifecycle | ||
| 240 | * ---------------------------------------------------------------- */ | ||
| 241 | |||
| 242 | Vec16 *vec16_create(Dimension dim); | ||
| 243 | Vec16 *vec16_from_floats(const float *values, Dimension dim); | ||
| 244 | void vec16_free(Vec16 *v); | ||
| 245 | void vec16_set(Vec16 *v, const float *values); | ||
| 246 | |||
| 247 | /* Convert to Vec32Ref (requires caller-provided float32 buffer) */ | ||
| 248 | static inline Vec32Ref | ||
| 249 | 2 | Vec16ToRef(const Vec16 *hv, float *buffer) | |
| 250 | { | ||
| 251 | 2 | vs_half_to_float_array(hv->x, buffer, hv->dim); | |
| 252 | 2 | return (Vec32Ref){.data = buffer, .dim = (Dimension)hv->dim}; | |
| 253 | } | ||
| 254 | |||
| 255 | /* ---------------------------------------------------------------- | ||
| 256 | * Inline vtable for compile-time specialization | ||
| 257 | * | ||
| 258 | * These always_inline functions + static const vtable enable the | ||
| 259 | * compiler to inline through vtable function pointers when the | ||
| 260 | * pointer target is known at compile time. Used by k-means and | ||
| 261 | * other hot loops that dispatch once at the entry point. | ||
| 262 | * ---------------------------------------------------------------- */ | ||
| 263 | |||
| 264 | VS_VTABLE_INLINE float | ||
| 265 | 1007 | vs_f16_dot_product(const void *vec, const float *centroid, Dimension dim) | |
| 266 | { | ||
| 267 | 1007 | const half *v = (const half *)vec; | |
| 268 | 1007 | float sum = 0.0f; | |
| 269 |
2/4✓ Branch 0 taken 3523 times.
✓ Branch 1 taken 1007 times.
✗ Branch 2 not taken.
✗ Branch 3 not taken.
|
4530 | for (Dimension d = 0; d < dim; d++) |
| 270 | 3523 | sum += vs_half_to_float(v[d]) * centroid[d]; | |
| 271 | 1007 | return sum; | |
| 272 | } | ||
| 273 | |||
| 274 | VS_VTABLE_INLINE float | ||
| 275 | 146740 | vs_f16_l2_squared(const void *vec, const float *centroid, Dimension dim) | |
| 276 | { | ||
| 277 | 146740 | const half *v = (const half *)vec; | |
| 278 | 146740 | float sum = 0.0f; | |
| 279 |
2/2✓ Branch 0 taken 3594288 times.
✓ Branch 1 taken 146740 times.
|
3741028 | for (Dimension d = 0; d < dim; d++) |
| 280 | { | ||
| 281 | 3594288 | float diff = vs_half_to_float(v[d]) - centroid[d]; | |
| 282 | 3594288 | sum += diff * diff; | |
| 283 | } | ||
| 284 | 146740 | return sum; | |
| 285 | } | ||
| 286 | |||
| 287 | VS_VTABLE_INLINE float | ||
| 288 | 2004 | vs_f16_norm_sq(const void *vec, Dimension dim) | |
| 289 | { | ||
| 290 | 2004 | const half *v = (const half *)vec; | |
| 291 | 2004 | float sum = 0.0f; | |
| 292 |
2/2✓ Branch 0 taken 9060 times.
✓ Branch 1 taken 2004 times.
|
11064 | for (Dimension d = 0; d < dim; d++) |
| 293 | { | ||
| 294 | 9060 | float val = vs_half_to_float(v[d]); | |
| 295 | 9060 | sum += val * val; | |
| 296 | } | ||
| 297 |
0/2✗ Branch 0 not taken.
✗ Branch 1 not taken.
|
2004 | return sum; |
| 298 | } | ||
| 299 | |||
| 300 | VS_VTABLE_INLINE void | ||
| 301 | 13604 | vs_f16_sum_to_float(const void *vec, float *accum, Dimension dim) | |
| 302 | { | ||
| 303 | 13604 | const half *v = (const half *)vec; | |
| 304 |
2/2✓ Branch 0 taken 94662 times.
✓ Branch 1 taken 13604 times.
|
108266 | for (Dimension d = 0; d < dim; d++) |
| 305 | 94662 | accum[d] += vs_half_to_float(v[d]); | |
| 306 | 13604 | } | |
| 307 | |||
| 308 | VS_VTABLE_INLINE void | ||
| 309 | 46 | vs_f16_to_float_one(const void *src, float *dst, Dimension dim) | |
| 310 | { | ||
| 311 | 46 | vs_half_to_float_array((const half *)src, dst, dim); | |
| 312 | 46 | } | |
| 313 | |||
| 314 | VS_VTABLE_INLINE const float * | ||
| 315 | 13053 | vs_f16_to_float_block( | |
| 316 | const void *src, float *dst, uint32_t count, Dimension dim) | ||
| 317 | { | ||
| 318 | 13053 | vs_half_to_float_array((const half *)src, dst, (uint32_t)count * dim); | |
| 319 | 13053 | return dst; | |
| 320 | } | ||
| 321 | |||
| 322 | static const Vec32TypeOps vs_f16_type_ops = { | ||
| 323 | .name = "float16", | ||
| 324 | .element_size = sizeof(half), | ||
| 325 | .dot_product = vs_f16_dot_product, | ||
| 326 | .l2_squared = vs_f16_l2_squared, | ||
| 327 | .norm_sq = vs_f16_norm_sq, | ||
| 328 | .sum_to_float = vs_f16_sum_to_float, | ||
| 329 | .to_float_one = vs_f16_to_float_one, | ||
| 330 | .to_float_block = vs_f16_to_float_block, | ||
| 331 | }; | ||
| 332 | |||
| 333 | /* ---------------------------------------------------------------- | ||
| 334 | * F16C inline vtable for compile-time specialization | ||
| 335 | * | ||
| 336 | * Hand-written AVX2+FMA+F16C: loads 8 halfs via _mm256_cvtph_ps, | ||
| 337 | * accumulates with FMA, reduces via horizontal sum. Scalar tail | ||
| 338 | * for remainder. Used by k-means and RaBitQ wrappers that target | ||
| 339 | * AVX2 explicitly rather than relying on TARGET_CLONES | ||
| 340 | * auto-vectorization. | ||
| 341 | * ---------------------------------------------------------------- */ | ||
| 342 | |||
| 343 | #if defined(VS_F16C_SUPPORT) && !defined(VS_SIMD_NONE) | ||
| 344 | |||
| 345 | #include "algo/simd_utils.h" | ||
| 346 | |||
| 347 | VS_TARGET_F16C_AVX2 VS_VTABLE_INLINE float | ||
| 348 | 4 | vs_f16c_dot_product(const void *vec, const float *centroid, Dimension dim) | |
| 349 | { | ||
| 350 | 4 | const half *v = (const half *)vec; | |
| 351 | 4 | __m256 sum8 = _mm256_setzero_ps(); | |
| 352 | 4 | Dimension d = 0; | |
| 353 | |||
| 354 |
2/2✓ Branch 0 taken 32 times.
✓ Branch 1 taken 4 times.
|
36 | for (; d + 8 <= dim; d += 8) |
| 355 | { | ||
| 356 | 64 | __m128i h8 = _mm_loadu_si128((const __m128i *)(v + d)); | |
| 357 | 32 | __m256 fv = _mm256_cvtph_ps(h8); | |
| 358 | 64 | __m256 fc = _mm256_loadu_ps(centroid + d); | |
| 359 | 32 | sum8 = _mm256_fmadd_ps(fv, fc, sum8); | |
| 360 | } | ||
| 361 | |||
| 362 | 4 | float sum = vs_horizontal_sum_avx2(sum8); | |
| 363 | |||
| 364 |
2/2✓ Branch 0 taken 8 times.
✓ Branch 1 taken 4 times.
|
12 | for (; d < dim; d++) |
| 365 | 8 | sum += vs_half_to_float(v[d]) * centroid[d]; | |
| 366 | |||
| 367 | 4 | return sum; | |
| 368 | } | ||
| 369 | |||
| 370 | VS_TARGET_F16C_AVX2 VS_VTABLE_INLINE float | ||
| 371 | 4 | vs_f16c_l2_squared(const void *vec, const float *centroid, Dimension dim) | |
| 372 | { | ||
| 373 | 4 | const half *v = (const half *)vec; | |
| 374 | 4 | __m256 sum8 = _mm256_setzero_ps(); | |
| 375 | 4 | Dimension d = 0; | |
| 376 | |||
| 377 |
2/2✓ Branch 0 taken 32 times.
✓ Branch 1 taken 4 times.
|
36 | for (; d + 8 <= dim; d += 8) |
| 378 | { | ||
| 379 | 64 | __m128i h8 = _mm_loadu_si128((const __m128i *)(v + d)); | |
| 380 | 32 | __m256 fv = _mm256_cvtph_ps(h8); | |
| 381 | 64 | __m256 fc = _mm256_loadu_ps(centroid + d); | |
| 382 | 32 | __m256 diff = _mm256_sub_ps(fv, fc); | |
| 383 | 32 | sum8 = _mm256_fmadd_ps(diff, diff, sum8); | |
| 384 | } | ||
| 385 | |||
| 386 | 4 | float sum = vs_horizontal_sum_avx2(sum8); | |
| 387 | |||
| 388 |
2/2✓ Branch 0 taken 8 times.
✓ Branch 1 taken 4 times.
|
12 | for (; d < dim; d++) |
| 389 | { | ||
| 390 | 8 | float diff = vs_half_to_float(v[d]) - centroid[d]; | |
| 391 | 8 | sum += diff * diff; | |
| 392 | } | ||
| 393 | |||
| 394 | 4 | return sum; | |
| 395 | } | ||
| 396 | |||
| 397 | VS_TARGET_F16C_AVX2 VS_VTABLE_INLINE float | ||
| 398 | 4 | vs_f16c_norm_sq(const void *vec, Dimension dim) | |
| 399 | { | ||
| 400 | 4 | const half *v = (const half *)vec; | |
| 401 | 4 | __m256 sum8 = _mm256_setzero_ps(); | |
| 402 | 4 | Dimension d = 0; | |
| 403 | |||
| 404 |
2/2✓ Branch 0 taken 32 times.
✓ Branch 1 taken 4 times.
|
36 | for (; d + 8 <= dim; d += 8) |
| 405 | { | ||
| 406 | 64 | __m128i h8 = _mm_loadu_si128((const __m128i *)(v + d)); | |
| 407 | 32 | __m256 fv = _mm256_cvtph_ps(h8); | |
| 408 | 32 | sum8 = _mm256_fmadd_ps(fv, fv, sum8); | |
| 409 | } | ||
| 410 | |||
| 411 | 4 | float sum = vs_horizontal_sum_avx2(sum8); | |
| 412 | |||
| 413 |
2/2✓ Branch 0 taken 4 times.
✓ Branch 1 taken 4 times.
|
8 | for (; d < dim; d++) |
| 414 | { | ||
| 415 | 4 | float val = vs_half_to_float(v[d]); | |
| 416 | 4 | sum += val * val; | |
| 417 | } | ||
| 418 | |||
| 419 | 4 | return sum; | |
| 420 | } | ||
| 421 | |||
| 422 | VS_TARGET_F16C_AVX2 VS_VTABLE_INLINE void | ||
| 423 | 4 | vs_f16c_sum_to_float(const void *vec, float *accum, Dimension dim) | |
| 424 | { | ||
| 425 | 4 | const half *v = (const half *)vec; | |
| 426 | 4 | Dimension d = 0; | |
| 427 | |||
| 428 |
2/2✓ Branch 0 taken 32 times.
✓ Branch 1 taken 4 times.
|
36 | for (; d + 8 <= dim; d += 8) |
| 429 | { | ||
| 430 | 64 | __m128i h8 = _mm_loadu_si128((const __m128i *)(v + d)); | |
| 431 | 32 | __m256 fv = _mm256_cvtph_ps(h8); | |
| 432 | 64 | __m256 ac = _mm256_loadu_ps(accum + d); | |
| 433 | 32 | _mm256_storeu_ps(accum + d, _mm256_add_ps(ac, fv)); | |
| 434 | } | ||
| 435 | |||
| 436 |
2/2✓ Branch 0 taken 6 times.
✓ Branch 1 taken 4 times.
|
10 | for (; d < dim; d++) |
| 437 | 6 | accum[d] += vs_half_to_float(v[d]); | |
| 438 | 4 | } | |
| 439 | |||
| 440 | static const Vec32TypeOps vs_f16c_type_ops = { | ||
| 441 | .name = "float16-f16c", | ||
| 442 | .element_size = sizeof(half), | ||
| 443 | .dot_product = vs_f16c_dot_product, | ||
| 444 | .l2_squared = vs_f16c_l2_squared, | ||
| 445 | .norm_sq = vs_f16c_norm_sq, | ||
| 446 | .sum_to_float = vs_f16c_sum_to_float, | ||
| 447 | .to_float_one = vs_f16_to_float_one, | ||
| 448 | .to_float_block = vs_f16_to_float_block, | ||
| 449 | }; | ||
| 450 | |||
| 451 | #endif /* VS_F16C_SUPPORT && !VS_SIMD_NONE */ | ||
| 452 | |||
| 453 | #endif /* VEC16_H */ | ||
| 454 |