/*-*- mode:c;indent-tabs-mode:nil;c-basic-offset:4;tab-width:8;coding:utf-8 -*-│
│vi: set net ft=c ts=4 sts=4 sw=4 fenc=utf-8                                :vi│
╚──────────────────────────────────────────────────────────────────────────────╝
│                                                                              │
│  GGML                                                                        │
│  Copyright (c) 2023 Georgi Gerganov                                          │
│                                                                              │
│  Permission is hereby granted, free of charge, to any person obtaining       │
│  a copy of this software and associated documentation files (the             │
│  "Software"), to deal in the Software without restriction, including         │
│  without limitation the rights to use, copy, modify, merge, publish,         │
│  distribute, sublicense, and/or sell copies of the Software, and to          │
│  permit persons to whom the Software is furnished to do so, subject to       │
│  the following conditions:                                                   │
│                                                                              │
│  The above copyright notice and this permission notice shall be              │
│  included in all copies or substantial portions of the Software.             │
│                                                                              │
│  THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND,             │
│  EXPRESS OR IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF          │
│  MERCHANTABILITY, FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT.      │
│  IN NO EVENT SHALL THE AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY        │
│  CLAIM, DAMAGES OR OTHER LIABILITY, WHETHER IN AN ACTION OF CONTRACT,        │
│  TORT OR OTHERWISE, ARISING FROM, OUT OF OR IN CONNECTION WITH THE           │
│  SOFTWARE OR THE USE OR OTHER DEALINGS IN THE SOFTWARE.                      │
│                                                                              │
╚─────────────────────────────────────────────────────────────────────────────*/
#include "third_party/ggml/ggjt.v2.q4_0.h"
#include "libc/assert.h"
#include "libc/macros.internal.h"
#include "libc/math.h"
#include "third_party/ggml/ggjt.v2.internal.h"
#include "third_party/ggml/ggjt.v2.q8_0.h"

static_assert(sizeof(block_v2_q4_0) == sizeof(float) + V2_QK4_0 / 2,
              "wrong q4_0 block size/padding");

void dequantize_row_v2_q4_0(const void * restrict x_, float * restrict y, int k) {
    const block_v2_q4_0 * restrict x = x_;
    static const int qk = V2_QK4_0;

    assert(k % qk == 0);

    const int nb = k / qk;

    for (int i = 0; i < nb; i++) {
        const float d = x[i].d;

        for (int j = 0; j < qk/2; ++j) {
            const int x0 = (x[i].qs[j] & 0x0F) - 8;
            const int x1 = (x[i].qs[j] >>   4) - 8;

            y[i*qk + j + 0   ] = x0*d;
            y[i*qk + j + qk/2] = x1*d;
        }
    }
}

size_t ggml_quantize_v2_q4_0(const float * src, void * dst, int n, int k, int64_t * hist) {
    assert(k % V2_QK4_0 == 0);
    const int nb = k / V2_QK4_0;

    for (int b = 0; b < n; b += k) {
        block_v2_q4_0 * restrict y = (block_v2_q4_0 *) dst + b/V2_QK4_0;

        quantize_row_v2_q4_0_reference(src + b, y, k);

        for (int i = 0; i < nb; i++) {
            for (int j = 0; j < V2_QK4_0; j += 2) {
                const uint8_t vi0 = y[i].qs[j/2] & 0x0F;
                const uint8_t vi1 = y[i].qs[j/2] >> 4;

                hist[vi0]++;
                hist[vi1]++;
            }
        }
    }

    return (n/V2_QK4_0*sizeof(block_v2_q4_0));
}

void quantize_row_v2_q4_0(const float * restrict x, void * restrict y, int k) {
    quantize_row_v2_q4_0_reference(x, y, k);
}

// reference implementation for deterministic creation of model files
void quantize_row_v2_q4_0_reference(const float * restrict x, block_v2_q4_0 * restrict y, int k) {
    static const int qk = V2_QK4_0;

    assert(k % qk == 0);

    const int nb = k / qk;

    for (int i = 0; i < nb; i++) {
        float amax = 0.0f; // absolute max
        float max  = 0.0f;

        for (int j = 0; j < qk; j++) {
            const float v = x[i*qk + j];
            if (amax < fabsf(v)) {
                amax = fabsf(v);
                max  = v;
            }
        }

        const float d  = max / -8;
        const float id = d ? 1.0f/d : 0.0f;

        y[i].d = d;

        for (int j = 0; j < qk/2; ++j) {
            const float x0 = x[i*qk + 0    + j]*id;
            const float x1 = x[i*qk + qk/2 + j]*id;

            const uint8_t xi0 = MIN(15, (int8_t)(x0 + 8.5f));
            const uint8_t xi1 = MIN(15, (int8_t)(x1 + 8.5f));

            y[i].qs[j]  = xi0;
            y[i].qs[j] |= xi1 << 4;
        }
    }
}

void ggml_vec_dot_v2_q4_0_q8_0(const int n,
                               float * restrict s,
                               const void * restrict vx,
                               const void * restrict vy) {
    const int qk = V2_QK8_0;
    const int nb = n / qk;

    assert(n % qk == 0);
    assert(nb % 2 == 0);

    const block_v2_q4_0 * restrict x = vx;
    const block_v2_q8_0 * restrict y = vy;

#if defined(__ARM_NEON)
    float32x4_t sumv0 = vdupq_n_f32(0.0f);
    float32x4_t sumv1 = vdupq_n_f32(0.0f);

    for (int i = 0; i < nb; i += 2) {
        const block_v2_q4_0 * restrict x0 = &x[i + 0];
        const block_v2_q4_0 * restrict x1 = &x[i + 1];
        const block_v2_q8_0 * restrict y0 = &y[i + 0];
        const block_v2_q8_0 * restrict y1 = &y[i + 1];

        const uint8x16_t m4b   = vdupq_n_u8(0x0F);
        const int8x16_t  s8b   = vdupq_n_s8(0x8);

        const uint8x16_t v0_0 = vld1q_u8(x0->qs);
        const uint8x16_t v0_1 = vld1q_u8(x1->qs);

        // 4-bit -> 8-bit
        const int8x16_t v0_0l = vreinterpretq_s8_u8(vandq_u8  (v0_0, m4b));
        const int8x16_t v0_0h = vreinterpretq_s8_u8(vshrq_n_u8(v0_0, 4));
        const int8x16_t v0_1l = vreinterpretq_s8_u8(vandq_u8  (v0_1, m4b));
        const int8x16_t v0_1h = vreinterpretq_s8_u8(vshrq_n_u8(v0_1, 4));

        // sub 8
        const int8x16_t v0_0ls = vsubq_s8(v0_0l, s8b);
        const int8x16_t v0_0hs = vsubq_s8(v0_0h, s8b);
        const int8x16_t v0_1ls = vsubq_s8(v0_1l, s8b);
        const int8x16_t v0_1hs = vsubq_s8(v0_1h, s8b);

        // load y
        const int8x16_t v1_0l = vld1q_s8(y0->qs);
        const int8x16_t v1_0h = vld1q_s8(y0->qs + 16);
        const int8x16_t v1_1l = vld1q_s8(y1->qs);
        const int8x16_t v1_1h = vld1q_s8(y1->qs + 16);

#if defined(__ARM_FEATURE_DOTPROD)
        // dot product into int32x4_t
        const int32x4_t p_0 = vdotq_s32(vdotq_s32(vdupq_n_s32(0), v0_0ls, v1_0l), v0_0hs, v1_0h);
        const int32x4_t p_1 = vdotq_s32(vdotq_s32(vdupq_n_s32(0), v0_1ls, v1_1l), v0_1hs, v1_1h);

        sumv0 = vmlaq_n_f32(sumv0, vcvtq_f32_s32(p_0), x0->d*y0->d);
        sumv1 = vmlaq_n_f32(sumv1, vcvtq_f32_s32(p_1), x1->d*y1->d);
#else
        const int16x8_t pl0l = vmull_s8(vget_low_s8 (v0_0ls), vget_low_s8 (v1_0l));
        const int16x8_t pl0h = vmull_s8(vget_high_s8(v0_0ls), vget_high_s8(v1_0l));
        const int16x8_t ph0l = vmull_s8(vget_low_s8 (v0_0hs), vget_low_s8 (v1_0h));
        const int16x8_t ph0h = vmull_s8(vget_high_s8(v0_0hs), vget_high_s8(v1_0h));

        const int16x8_t pl1l = vmull_s8(vget_low_s8 (v0_1ls), vget_low_s8 (v1_1l));
        const int16x8_t pl1h = vmull_s8(vget_high_s8(v0_1ls), vget_high_s8(v1_1l));
        const int16x8_t ph1l = vmull_s8(vget_low_s8 (v0_1hs), vget_low_s8 (v1_1h));
        const int16x8_t ph1h = vmull_s8(vget_high_s8(v0_1hs), vget_high_s8(v1_1h));

        const int32x4_t pl0 = vaddq_s32(vpaddlq_s16(pl0l), vpaddlq_s16(pl0h));
        const int32x4_t ph0 = vaddq_s32(vpaddlq_s16(ph0l), vpaddlq_s16(ph0h));
        const int32x4_t pl1 = vaddq_s32(vpaddlq_s16(pl1l), vpaddlq_s16(pl1h));
        const int32x4_t ph1 = vaddq_s32(vpaddlq_s16(ph1l), vpaddlq_s16(ph1h));

        sumv0 = vmlaq_n_f32(sumv0, vcvtq_f32_s32(vaddq_s32(pl0, ph0)), x0->d*y0->d);
        sumv1 = vmlaq_n_f32(sumv1, vcvtq_f32_s32(vaddq_s32(pl1, ph1)), x1->d*y1->d);
#endif
    }

    *s = vaddvq_f32(sumv0) + vaddvq_f32(sumv1);
#elif defined(__AVX2__)
    // Initialize accumulator with zeros
    __m256 acc = _mm256_setzero_ps();

    // Main loop
    for (int i = 0; i < nb; ++i) {
        /* Compute combined scale for the block */
        const __m256 d = _mm256_mul_ps( _mm256_broadcast_ss( &x[i].d ), _mm256_broadcast_ss( &y[i].d ) );

        __m256i bx = bytes_from_nibbles_32(x[i].qs);

        // Now we have a vector with bytes in [ 0 .. 15 ] interval. Offset them into [ -8 .. +7 ] interval.
        const __m256i off = _mm256_set1_epi8( 8 );
        bx = _mm256_sub_epi8( bx, off );

        __m256i by = _mm256_loadu_si256((const __m256i *)y[i].qs);

        const __m256 q = mul_sum_i8_pairs_float(bx, by);

        /* Multiply q with scale and accumulate */
        acc = _mm256_fmadd_ps( d, q, acc );
    }

    *s = hsum_float_8(acc);
#elif defined(__AVX__)
    // Initialize accumulator with zeros
    __m256 acc = _mm256_setzero_ps();

    // Main loop
    for (int i = 0; i < nb; ++i) {
        // Compute combined scale for the block
        const __m256 d = _mm256_mul_ps( _mm256_broadcast_ss( &x[i].d ), _mm256_broadcast_ss( &y[i].d ) );

        const __m128i lowMask = _mm_set1_epi8(0xF);
        const __m128i off = _mm_set1_epi8(8);

        const __m128i tmp = _mm_loadu_si128((const __m128i *)x[i].qs);

        __m128i bx = _mm_and_si128(lowMask, tmp);
        __m128i by = _mm_loadu_si128((const __m128i *)y[i].qs);
        bx = _mm_sub_epi8(bx, off);
        const __m128i i32_0 = mul_sum_i8_pairs(bx, by);

        bx = _mm_and_si128(lowMask, _mm_srli_epi64(tmp, 4));
        by = _mm_loadu_si128((const __m128i *)(y[i].qs + 16));
        bx = _mm_sub_epi8(bx, off);
        const __m128i i32_1 = mul_sum_i8_pairs(bx, by);

        // Convert int32_t to float
        __m256 p = _mm256_cvtepi32_ps(_mm256_set_m128i(i32_0, i32_1));

        // Apply the scale, and accumulate
        acc = _mm256_add_ps(_mm256_mul_ps( d, p ), acc);
    }

    *s = hsum_float_8(acc);
#elif defined(__SSSE3__)
    // set constants
    const __m128i lowMask = _mm_set1_epi8(0xF);
    const __m128i off = _mm_set1_epi8(8);

    // Initialize accumulator with zeros
    __m128 acc_0 = _mm_setzero_ps();
    __m128 acc_1 = _mm_setzero_ps();
    __m128 acc_2 = _mm_setzero_ps();
    __m128 acc_3 = _mm_setzero_ps();

    // First round without accumulation
    {
        _mm_prefetch(&x[0] + sizeof(block_v2_q4_0), _MM_HINT_T0);
        _mm_prefetch(&y[0] + sizeof(block_v2_q8_0), _MM_HINT_T0);

        // Compute combined scale for the block 0 and 1
        const __m128 d_0_1 = _mm_mul_ps( _mm_set1_ps( x[0].d ), _mm_set1_ps( y[0].d ) );

        const __m128i tmp_0_1 = _mm_loadu_si128((const __m128i *)x[0].qs);

        __m128i bx_0 = _mm_and_si128(lowMask, tmp_0_1);
        __m128i by_0 = _mm_loadu_si128((const __m128i *)y[0].qs);
        bx_0 = _mm_sub_epi8(bx_0, off);
        const __m128i i32_0 = mul_sum_i8_pairs(bx_0, by_0);

        __m128i bx_1 = _mm_and_si128(lowMask, _mm_srli_epi64(tmp_0_1, 4));
        __m128i by_1 = _mm_loadu_si128((const __m128i *)(y[0].qs + 16));
        bx_1 = _mm_sub_epi8(bx_1, off);
        const __m128i i32_1 = mul_sum_i8_pairs(bx_1, by_1);

        _mm_prefetch(&x[1] + sizeof(block_v2_q4_0), _MM_HINT_T0);
        _mm_prefetch(&y[1] + sizeof(block_v2_q8_0), _MM_HINT_T0);

        // Compute combined scale for the block 2 and 3
        const __m128 d_2_3 = _mm_mul_ps( _mm_set1_ps( x[1].d ), _mm_set1_ps( y[1].d ) );

        const __m128i tmp_2_3 = _mm_loadu_si128((const __m128i *)x[1].qs);

        __m128i bx_2 = _mm_and_si128(lowMask, tmp_2_3);
        __m128i by_2 = _mm_loadu_si128((const __m128i *)y[1].qs);
        bx_2 = _mm_sub_epi8(bx_2, off);
        const __m128i i32_2 = mul_sum_i8_pairs(bx_2, by_2);

        __m128i bx_3 = _mm_and_si128(lowMask, _mm_srli_epi64(tmp_2_3, 4));
        __m128i by_3 = _mm_loadu_si128((const __m128i *)(y[1].qs + 16));
        bx_3 = _mm_sub_epi8(bx_3, off);
        const __m128i i32_3 = mul_sum_i8_pairs(bx_3, by_3);

        // Convert int32_t to float
        __m128 p0 = _mm_cvtepi32_ps(i32_0);
        __m128 p1 = _mm_cvtepi32_ps(i32_1);
        __m128 p2 = _mm_cvtepi32_ps(i32_2);
        __m128 p3 = _mm_cvtepi32_ps(i32_3);

        // Apply the scale
        acc_0 = _mm_mul_ps( d_0_1, p0 );
        acc_1 = _mm_mul_ps( d_0_1, p1 );
        acc_2 = _mm_mul_ps( d_2_3, p2 );
        acc_3 = _mm_mul_ps( d_2_3, p3 );
    }

    // Main loop
    for (int i = 2; i < nb; i+=2) {
        _mm_prefetch(&x[i] + sizeof(block_v2_q4_0), _MM_HINT_T0);
        _mm_prefetch(&y[i] + sizeof(block_v2_q8_0), _MM_HINT_T0);

        // Compute combined scale for the block 0 and 1
        const __m128 d_0_1 = _mm_mul_ps( _mm_set1_ps( x[i].d ), _mm_set1_ps( y[i].d ) );

        const __m128i tmp_0_1 = _mm_loadu_si128((const __m128i *)x[i].qs);

        __m128i bx_0 = _mm_and_si128(lowMask, tmp_0_1);
        __m128i by_0 = _mm_loadu_si128((const __m128i *)y[i].qs);
        bx_0 = _mm_sub_epi8(bx_0, off);
        const __m128i i32_0 = mul_sum_i8_pairs(bx_0, by_0);

        __m128i bx_1 = _mm_and_si128(lowMask, _mm_srli_epi64(tmp_0_1, 4));
        __m128i by_1 = _mm_loadu_si128((const __m128i *)(y[i].qs + 16));
        bx_1 = _mm_sub_epi8(bx_1, off);
        const __m128i i32_1 = mul_sum_i8_pairs(bx_1, by_1);

        _mm_prefetch(&x[i] + 2 * sizeof(block_v2_q4_0), _MM_HINT_T0);
        _mm_prefetch(&y[i] + 2 * sizeof(block_v2_q8_0), _MM_HINT_T0);

        // Compute combined scale for the block 2 and 3
        const __m128 d_2_3 = _mm_mul_ps( _mm_set1_ps( x[i + 1].d ), _mm_set1_ps( y[i + 1].d ) );

        const __m128i tmp_2_3 = _mm_loadu_si128((const __m128i *)x[i + 1].qs);

        __m128i bx_2 = _mm_and_si128(lowMask, tmp_2_3);
        __m128i by_2 = _mm_loadu_si128((const __m128i *)y[i + 1].qs);
        bx_2 = _mm_sub_epi8(bx_2, off);
        const __m128i i32_2 = mul_sum_i8_pairs(bx_2, by_2);

        __m128i bx_3 = _mm_and_si128(lowMask, _mm_srli_epi64(tmp_2_3, 4));
        __m128i by_3 = _mm_loadu_si128((const __m128i *)(y[i + 1].qs + 16));
        bx_3 = _mm_sub_epi8(bx_3, off);
        const __m128i i32_3 = mul_sum_i8_pairs(bx_3, by_3);

        // Convert int32_t to float
        __m128 p0 = _mm_cvtepi32_ps(i32_0);
        __m128 p1 = _mm_cvtepi32_ps(i32_1);
        __m128 p2 = _mm_cvtepi32_ps(i32_2);
        __m128 p3 = _mm_cvtepi32_ps(i32_3);

        // Apply the scale
        __m128 p0_d = _mm_mul_ps( d_0_1, p0 );
        __m128 p1_d = _mm_mul_ps( d_0_1, p1 );
        __m128 p2_d = _mm_mul_ps( d_2_3, p2 );
        __m128 p3_d = _mm_mul_ps( d_2_3, p3 );

        // Acummulate
        acc_0 = _mm_add_ps(p0_d, acc_0);
        acc_1 = _mm_add_ps(p1_d, acc_1);
        acc_2 = _mm_add_ps(p2_d, acc_2);
        acc_3 = _mm_add_ps(p3_d, acc_3);
    }

    *s = hsum_float_4x4(acc_0, acc_1, acc_2, acc_3);
#else
    // scalar
    float sumf = 0.0;

    for (int i = 0; i < nb; i++) {
        int sumi = 0;

        for (int j = 0; j < qk/2; ++j) {
            const int v0 = (x[i].qs[j] & 0x0F) - 8;
            const int v1 = (x[i].qs[j] >>   4) - 8;

            sumi += (v0 * y[i].qs[j]) + (v1 * y[i].qs[j + qk/2]);
        }

        sumf += (x[i].d*y[i].d)*sumi;
    }

    *s = sumf;
#endif
}