ggml/src/ggml-cpu-impl.h

#pragma once

// GGML CPU internal header

#include "ggml.h"
#include "ggml-impl.h"
#include <stdlib.h> // load `stdlib.h` before other headers to work around MinGW bug: https://sourceforge.net/p/mingw-w64/bugs/192/
//#include <stddef.h>
#include <stdbool.h>
#include <string.h> // memcpy
#include <math.h>   // fabsf


#ifdef __cplusplus
extern "C" {
#endif

#if defined(_MSC_VER)

#define m512bh(p) p
#define m512i(p) p

#else

#define m512bh(p) (__m512bh)(p)
#define m512i(p) (__m512i)(p)

#endif

/**
 * Converts brain16 to float32.
 *
 * The bfloat16 floating point format has the following structure:
 *
 *       ┌sign
 *       │
 *       │   ┌exponent
 *       │   │
 *       │   │      ┌mantissa
 *       │   │      │
 *       │┌──┴───┐┌─┴───┐
 *     0b0000000000000000 brain16
 *
 * Since bf16 has the same number of exponent bits as a 32bit float,
 * encoding and decoding numbers becomes relatively straightforward.
 *
 *       ┌sign
 *       │
 *       │   ┌exponent
 *       │   │
 *       │   │      ┌mantissa
 *       │   │      │
 *       │┌──┴───┐┌─┴───────────────────┐
 *     0b00000000000000000000000000000000 IEEE binary32
 *
 * For comparison, the standard fp16 format has fewer exponent bits.
 *
 *       ┌sign
 *       │
 *       │  ┌exponent
 *       │  │
 *       │  │    ┌mantissa
 *       │  │    │
 *       │┌─┴─┐┌─┴──────┐
 *     0b0000000000000000 IEEE binary16
 *
 * @see IEEE 754-2008
 */
static inline float ggml_compute_bf16_to_fp32(ggml_bf16_t h) {
    union {
        float f;
        uint32_t i;
    } u;
    u.i = (uint32_t)h.bits << 16;
    return u.f;
}

/**
 * Converts float32 to brain16.
 *
 * This is binary identical with Google Brain float conversion.
 * Floats shall round to nearest even, and NANs shall be quiet.
 * Subnormals aren't flushed to zero, except perhaps when used.
 * This code should vectorize nicely if using modern compilers.
 */
static inline ggml_bf16_t ggml_compute_fp32_to_bf16(float s) {
    ggml_bf16_t h;
    union {
        float f;
        uint32_t i;
    } u;
    u.f = s;
    if ((u.i & 0x7fffffff) > 0x7f800000) { /* nan */
        h.bits = (u.i >> 16) | 64; /* force to quiet */
        return h;
    }
    h.bits = (u.i + (0x7fff + ((u.i >> 16) & 1))) >> 16;
    return h;
}

#define GGML_FP32_TO_BF16(x) ggml_compute_fp32_to_bf16(x)
#define GGML_BF16_TO_FP32(x) ggml_compute_bf16_to_fp32(x)

// __FMA__ and __F16C__ are not defined in MSVC, however they are implied with AVX2/AVX512
#if defined(_MSC_VER) && (defined(__AVX2__) || defined(__AVX512F__))
#ifndef __FMA__
#define __FMA__
#endif
#ifndef __F16C__
#define __F16C__
#endif
#endif

// __SSE3__ and __SSSE3__ are not defined in MSVC, but SSE3/SSSE3 are present when AVX/AVX2/AVX512 are available
#if defined(_MSC_VER) && (defined(__AVX__) || defined(__AVX2__) || defined(__AVX512F__))
#ifndef __SSE3__
#define __SSE3__
#endif
#ifndef __SSSE3__
#define __SSSE3__
#endif
#endif

#if defined(__ARM_FEATURE_SVE)
#include <arm_sve.h>
#include <sys/prctl.h>
#endif

// 16-bit float
// on Arm, we use __fp16
// on x86, we use uint16_t
#if defined(__ARM_NEON)

// if YCM cannot find <arm_neon.h>, make a symbolic link to it, for example:
//
//   $ ln -sfn /Library/Developer/CommandLineTools/usr/lib/clang/13.1.6/include/arm_neon.h ./src/
//
#include <arm_neon.h>

#ifdef _MSC_VER

typedef uint16_t ggml_fp16_internal_t;

#define ggml_vld1q_u32(w,x,y,z) { ((w) + ((uint64_t)(x) << 32)), ((y) + ((uint64_t)(z) << 32)) }

#else

typedef __fp16 ggml_fp16_internal_t;

#define ggml_vld1q_u32(w,x,y,z) { (w), (x), (y), (z) }

#endif // _MSC_VER

#if !defined(__aarch64__)

// 32-bit ARM compatibility

// vaddlvq_s16
// vpaddq_s16
// vpaddq_s32
// vaddvq_s32
// vaddvq_f32
// vmaxvq_f32
// vcvtnq_s32_f32
// vzip1_u8
// vzip2_u8

inline static int32_t vaddlvq_s16(int16x8_t v) {
    int32x4_t v0 = vreinterpretq_s32_s64(vpaddlq_s32(vpaddlq_s16(v)));
    return vgetq_lane_s32(v0, 0) + vgetq_lane_s32(v0, 2);
}

inline static int16x8_t vpaddq_s16(int16x8_t a, int16x8_t b) {
    int16x4_t a0 = vpadd_s16(vget_low_s16(a), vget_high_s16(a));
    int16x4_t b0 = vpadd_s16(vget_low_s16(b), vget_high_s16(b));
    return vcombine_s16(a0, b0);
}

inline static int32x4_t vpaddq_s32(int32x4_t a, int32x4_t b) {
    int32x2_t a0 = vpadd_s32(vget_low_s32(a), vget_high_s32(a));
    int32x2_t b0 = vpadd_s32(vget_low_s32(b), vget_high_s32(b));
    return vcombine_s32(a0, b0);
}

inline static int32_t vaddvq_s32(int32x4_t v) {
    return vgetq_lane_s32(v, 0) + vgetq_lane_s32(v, 1) + vgetq_lane_s32(v, 2) + vgetq_lane_s32(v, 3);
}

inline static float vaddvq_f32(float32x4_t v) {
    return vgetq_lane_f32(v, 0) + vgetq_lane_f32(v, 1) + vgetq_lane_f32(v, 2) + vgetq_lane_f32(v, 3);
}

inline static float vmaxvq_f32(float32x4_t v) {
    return
        MAX(MAX(vgetq_lane_f32(v, 0), vgetq_lane_f32(v, 1)),
            MAX(vgetq_lane_f32(v, 2), vgetq_lane_f32(v, 3)));
}

inline static int32x4_t vcvtnq_s32_f32(float32x4_t v) {
    int32x4_t res;

    res[0] = roundf(vgetq_lane_f32(v, 0));
    res[1] = roundf(vgetq_lane_f32(v, 1));
    res[2] = roundf(vgetq_lane_f32(v, 2));
    res[3] = roundf(vgetq_lane_f32(v, 3));

    return res;
}

inline static uint8x8_t vzip1_u8(uint8x8_t a, uint8x8_t b) {
    uint8x8_t res;

    res[0] = a[0]; res[1] = b[0];
    res[2] = a[1]; res[3] = b[1];
    res[4] = a[2]; res[5] = b[2];
    res[6] = a[3]; res[7] = b[3];

    return res;
}

inline static uint8x8_t vzip2_u8(uint8x8_t a, uint8x8_t b) {
    uint8x8_t res;

    res[0] = a[4]; res[1] = b[4];
    res[2] = a[5]; res[3] = b[5];
    res[4] = a[6]; res[5] = b[6];
    res[6] = a[7]; res[7] = b[7];

    return res;
}

// vld1q_s16_x2
// vld1q_u8_x2
// vld1q_u8_x4
// vld1q_s8_x2
// vld1q_s8_x4
// TODO: double-check these work correctly

typedef struct ggml_int16x8x2_t {
    int16x8_t val[2];
} ggml_int16x8x2_t;

inline static ggml_int16x8x2_t ggml_vld1q_s16_x2(const int16_t * ptr) {
    ggml_int16x8x2_t res;

    res.val[0] = vld1q_s16(ptr + 0);
    res.val[1] = vld1q_s16(ptr + 8);

    return res;
}

typedef struct ggml_uint8x16x2_t {
    uint8x16_t val[2];
} ggml_uint8x16x2_t;

inline static ggml_uint8x16x2_t ggml_vld1q_u8_x2(const uint8_t * ptr) {
    ggml_uint8x16x2_t res;

    res.val[0] = vld1q_u8(ptr + 0);
    res.val[1] = vld1q_u8(ptr + 16);

    return res;
}

typedef struct ggml_uint8x16x4_t {
    uint8x16_t val[4];
} ggml_uint8x16x4_t;

inline static ggml_uint8x16x4_t ggml_vld1q_u8_x4(const uint8_t * ptr) {
    ggml_uint8x16x4_t res;

    res.val[0] = vld1q_u8(ptr + 0);
    res.val[1] = vld1q_u8(ptr + 16);
    res.val[2] = vld1q_u8(ptr + 32);
    res.val[3] = vld1q_u8(ptr + 48);

    return res;
}

typedef struct ggml_int8x16x2_t {
    int8x16_t val[2];
} ggml_int8x16x2_t;

inline static ggml_int8x16x2_t ggml_vld1q_s8_x2(const int8_t * ptr) {
    ggml_int8x16x2_t res;

    res.val[0] = vld1q_s8(ptr + 0);
    res.val[1] = vld1q_s8(ptr + 16);

    return res;
}

typedef struct ggml_int8x16x4_t {
    int8x16_t val[4];
} ggml_int8x16x4_t;

inline static ggml_int8x16x4_t ggml_vld1q_s8_x4(const int8_t * ptr) {
    ggml_int8x16x4_t res;

    res.val[0] = vld1q_s8(ptr + 0);
    res.val[1] = vld1q_s8(ptr + 16);
    res.val[2] = vld1q_s8(ptr + 32);
    res.val[3] = vld1q_s8(ptr + 48);

    return res;
}

// NOTE: not tested
inline static int8x16_t ggml_vqtbl1q_s8(int8x16_t a, uint8x16_t b) {
    int8x16_t res;

    res[ 0] = a[b[ 0]];
    res[ 1] = a[b[ 1]];
    res[ 2] = a[b[ 2]];
    res[ 3] = a[b[ 3]];
    res[ 4] = a[b[ 4]];
    res[ 5] = a[b[ 5]];
    res[ 6] = a[b[ 6]];
    res[ 7] = a[b[ 7]];
    res[ 8] = a[b[ 8]];
    res[ 9] = a[b[ 9]];
    res[10] = a[b[10]];
    res[11] = a[b[11]];
    res[12] = a[b[12]];
    res[13] = a[b[13]];
    res[14] = a[b[14]];
    res[15] = a[b[15]];

    return res;
}

// NOTE: not tested
inline static uint8x16_t ggml_vqtbl1q_u8(uint8x16_t a, uint8x16_t b) {
    uint8x16_t res;

    res[ 0] = a[b[ 0]];
    res[ 1] = a[b[ 1]];
    res[ 2] = a[b[ 2]];
    res[ 3] = a[b[ 3]];
    res[ 4] = a[b[ 4]];
    res[ 5] = a[b[ 5]];
    res[ 6] = a[b[ 6]];
    res[ 7] = a[b[ 7]];
    res[ 8] = a[b[ 8]];
    res[ 9] = a[b[ 9]];
    res[10] = a[b[10]];
    res[11] = a[b[11]];
    res[12] = a[b[12]];
    res[13] = a[b[13]];
    res[14] = a[b[14]];
    res[15] = a[b[15]];

    return res;
}

#else

#define ggml_int16x8x2_t  int16x8x2_t
#define ggml_uint8x16x2_t uint8x16x2_t
#define ggml_uint8x16x4_t uint8x16x4_t
#define ggml_int8x16x2_t  int8x16x2_t
#define ggml_int8x16x4_t  int8x16x4_t

#define ggml_vld1q_s16_x2 vld1q_s16_x2
#define ggml_vld1q_u8_x2  vld1q_u8_x2
#define ggml_vld1q_u8_x4  vld1q_u8_x4
#define ggml_vld1q_s8_x2  vld1q_s8_x2
#define ggml_vld1q_s8_x4  vld1q_s8_x4
#define ggml_vqtbl1q_s8   vqtbl1q_s8
#define ggml_vqtbl1q_u8   vqtbl1q_u8

#endif // !defined(__aarch64__)

#if !defined(__ARM_FEATURE_DOTPROD)

inline static int32x4_t ggml_vdotq_s32(int32x4_t acc, int8x16_t a, int8x16_t b) {
    const int16x8_t p0 = vmull_s8(vget_low_s8 (a), vget_low_s8 (b));
    const int16x8_t p1 = vmull_s8(vget_high_s8(a), vget_high_s8(b));

    return vaddq_s32(acc, vaddq_s32(vpaddlq_s16(p0), vpaddlq_s16(p1)));
}

#else

#define ggml_vdotq_s32(a, b, c) vdotq_s32(a, b, c)

#endif // !defined(__ARM_FEATURE_DOTPROD)

#endif // defined(__ARM_NEON)

#if defined(__ARM_NEON) && !defined(_MSC_VER)

#define GGML_COMPUTE_FP16_TO_FP32(x) ggml_compute_fp16_to_fp32(x)
#define GGML_COMPUTE_FP32_TO_FP16(x) ggml_compute_fp32_to_fp16(x)

#define GGML_FP16_TO_FP32(x) ggml_compute_fp16_to_fp32(x)

static inline float ggml_compute_fp16_to_fp32(ggml_fp16_t h) {
    ggml_fp16_internal_t tmp;
    memcpy(&tmp, &h, sizeof(ggml_fp16_t));
    return (float)tmp;
}

static inline ggml_fp16_t ggml_compute_fp32_to_fp16(float f) {
    ggml_fp16_t res;
    ggml_fp16_internal_t tmp = f;
    memcpy(&res, &tmp, sizeof(ggml_fp16_t));
    return res;
}

#else

#ifdef __wasm_simd128__
#include <wasm_simd128.h>
#else
#ifdef __POWER9_VECTOR__
#include <altivec.h>
#undef bool
#define bool _Bool
#else
#if defined(_MSC_VER) || defined(__MINGW32__)
#include <intrin.h>
#else
#if defined(__AVX__) || defined(__AVX2__) || defined(__AVX512F__) || defined(__SSSE3__) || defined(__SSE3__) || defined(__SSE__)
#if !defined(__riscv)
#include <immintrin.h>
#endif
#endif
#endif
#endif
#endif

#ifdef __riscv_v_intrinsic
#include <riscv_vector.h>
#endif

#if defined(__loongarch64)
#if defined(__loongarch_asx)
#include <lasxintrin.h>
#endif
#if defined(__loongarch_sx)
#include <lsxintrin.h>
#endif
#endif

#if defined(__loongarch_asx)

typedef union {
    int32_t i;
    float f;
} ft_union;

/* float type data load instructions */
static __m128 __lsx_vreplfr2vr_s(float val) {
    ft_union fi_tmpval = {.f = val};
    return (__m128)__lsx_vreplgr2vr_w(fi_tmpval.i);
}

static __m256 __lasx_xvreplfr2vr_s(float val) {
    ft_union fi_tmpval = {.f = val};
    return (__m256)__lasx_xvreplgr2vr_w(fi_tmpval.i);
}
#endif

#ifdef __F16C__

#ifdef _MSC_VER
#define GGML_COMPUTE_FP16_TO_FP32(x) _mm_cvtss_f32(_mm_cvtph_ps(_mm_cvtsi32_si128(x)))
#define GGML_COMPUTE_FP32_TO_FP16(x) _mm_extract_epi16(_mm_cvtps_ph(_mm_set_ss(x), 0), 0)
#else
#define GGML_COMPUTE_FP16_TO_FP32(x) _cvtsh_ss(x)
#define GGML_COMPUTE_FP32_TO_FP16(x) _cvtss_sh(x, 0)
#endif

#elif defined(__POWER9_VECTOR__)

#define GGML_COMPUTE_FP16_TO_FP32(x) ggml_compute_fp16_to_fp32(x)
#define GGML_COMPUTE_FP32_TO_FP16(x) ggml_compute_fp32_to_fp16(x)
/* the inline asm below is about 12% faster than the lookup method */
#define GGML_FP16_TO_FP32(x) GGML_COMPUTE_FP16_TO_FP32(x)
#define GGML_FP32_TO_FP16(x) GGML_COMPUTE_FP32_TO_FP16(x)

static inline float ggml_compute_fp16_to_fp32(ggml_fp16_t h) {
    register float f;
    register double d;
    __asm__(
        "mtfprd %0,%2\n"
        "xscvhpdp %0,%0\n"
        "frsp %1,%0\n" :
        /* temp */ "=d"(d),
        /* out */  "=f"(f):
        /* in */   "r"(h));
    return f;
}

static inline ggml_fp16_t ggml_compute_fp32_to_fp16(float f) {
    register double d;
    register ggml_fp16_t r;
    __asm__( /* xscvdphp can work on double or single precision */
        "xscvdphp %0,%2\n"
        "mffprd %1,%0\n" :
        /* temp */ "=d"(d),
        /* out */  "=r"(r):
        /* in */   "f"(f));
    return r;
}

#else

// FP16 <-> FP32
// ref: https://github.com/Maratyszcza/FP16

static inline float fp32_from_bits(uint32_t w) {
    union {
        uint32_t as_bits;
        float as_value;
    } fp32;
    fp32.as_bits = w;
    return fp32.as_value;
}

static inline uint32_t fp32_to_bits(float f) {
    union {
        float as_value;
        uint32_t as_bits;
    } fp32;
    fp32.as_value = f;
    return fp32.as_bits;
}

static inline float ggml_compute_fp16_to_fp32(ggml_fp16_t h) {
    const uint32_t w = (uint32_t) h << 16;
    const uint32_t sign = w & UINT32_C(0x80000000);
    const uint32_t two_w = w + w;

    const uint32_t exp_offset = UINT32_C(0xE0) << 23;
#if defined(__STDC_VERSION__) && (__STDC_VERSION__ >= 199901L) || defined(__GNUC__) && !defined(__STRICT_ANSI__)
    const float exp_scale = 0x1.0p-112f;
#else
    const float exp_scale = fp32_from_bits(UINT32_C(0x7800000));
#endif
    const float normalized_value = fp32_from_bits((two_w >> 4) + exp_offset) * exp_scale;

    const uint32_t magic_mask = UINT32_C(126) << 23;
    const float magic_bias = 0.5f;
    const float denormalized_value = fp32_from_bits((two_w >> 17) | magic_mask) - magic_bias;

    const uint32_t denormalized_cutoff = UINT32_C(1) << 27;
    const uint32_t result = sign |
        (two_w < denormalized_cutoff ? fp32_to_bits(denormalized_value) : fp32_to_bits(normalized_value));
    return fp32_from_bits(result);
}

static inline ggml_fp16_t ggml_compute_fp32_to_fp16(float f) {
#if defined(__STDC_VERSION__) && (__STDC_VERSION__ >= 199901L) || defined(__GNUC__) && !defined(__STRICT_ANSI__)
    const float scale_to_inf = 0x1.0p+112f;
    const float scale_to_zero = 0x1.0p-110f;
#else
    const float scale_to_inf = fp32_from_bits(UINT32_C(0x77800000));
    const float scale_to_zero = fp32_from_bits(UINT32_C(0x08800000));
#endif
    float base = (fabsf(f) * scale_to_inf) * scale_to_zero;

    const uint32_t w = fp32_to_bits(f);
    const uint32_t shl1_w = w + w;
    const uint32_t sign = w & UINT32_C(0x80000000);
    uint32_t bias = shl1_w & UINT32_C(0xFF000000);
    if (bias < UINT32_C(0x71000000)) {
        bias = UINT32_C(0x71000000);
    }

    base = fp32_from_bits((bias >> 1) + UINT32_C(0x07800000)) + base;
    const uint32_t bits = fp32_to_bits(base);
    const uint32_t exp_bits = (bits >> 13) & UINT32_C(0x00007C00);
    const uint32_t mantissa_bits = bits & UINT32_C(0x00000FFF);
    const uint32_t nonsign = exp_bits + mantissa_bits;
    return (sign >> 16) | (shl1_w > UINT32_C(0xFF000000) ? UINT16_C(0x7E00) : nonsign);
}

#define GGML_COMPUTE_FP16_TO_FP32(x) ggml_compute_fp16_to_fp32(x)
#define GGML_COMPUTE_FP32_TO_FP16(x) ggml_compute_fp32_to_fp16(x)

#endif // __F16C__

#endif // defined(__ARM_NEON) && (!defined(__MSC_VER)

#ifdef __ARM_FEATURE_SVE
#include <arm_sve.h>
#endif // __ARM_FEATURE_SVE

// precomputed f32 table for f16 (256 KB)
// defined in ggml.c, initialized in ggml_init()
extern float ggml_table_f32_f16[1 << 16];

// On ARM NEON, it's quicker to directly convert x -> x instead of calling into ggml_lookup_fp16_to_fp32,
// so we define GGML_FP16_TO_FP32 and GGML_FP32_TO_FP16 elsewhere for NEON.
// This is also true for POWER9.
#if !defined(GGML_FP16_TO_FP32)
inline static float ggml_lookup_fp16_to_fp32(ggml_fp16_t f) {
    uint16_t s;
    memcpy(&s, &f, sizeof(uint16_t));
    return ggml_table_f32_f16[s];
}

#define GGML_FP16_TO_FP32(x) ggml_lookup_fp16_to_fp32(x)
#endif

#if !defined(GGML_FP32_TO_FP16)
#define GGML_FP32_TO_FP16(x) GGML_COMPUTE_FP32_TO_FP16(x)
#endif

#ifdef __cplusplus
}
#endif