#version 450

#include "common.comp"

#define BLOCKS_IN_QUANT QK4_0
#define SIZE_OF_BLOCK sizeof_block_q4_0
#define N_ROWS 4

#include "op_mul_mv_q_n_pre.comp"

// The q4_0 version of this function
float block_q_n_dot_y(uint block_index, uint yb, uint il) {
    vec2 acc = vec2(0.0, 0.0);
    const uint index = (block_index) * SIZE_OF_BLOCK + pcs.inAOff;
    float d = float(u8BufToFloat16(inA, index));
    float sumy = 0.0f;
    for (int i = 0; i < BLOCKS_IN_QUANT/4; i+=2) {
        const uint16_t b = u8BufToU16(inA, index + 2 + il + i);

        const float yl0 = inB[yb + i];
        const float yl1 = inB[yb + i + 1];
        const float yl8 = inB[yb + i + BLOCKS_IN_QUANT/2];
        const float yl9 = inB[yb + i + BLOCKS_IN_QUANT/2 + 1];

        sumy += yl0 + yl1 + yl8 + yl9;

        acc[0] += yl0 * (b & 0x000F) + yl1 / 256.f * (b & 0x0F00);
        acc[1] += yl8 / 16.f * (b & 0x00F0) + yl9 / 4096.f * (b & 0xF000);
    }
    return d * (sumy * -8.f + acc[0] + acc[1]);
}

#include "op_mul_mv_q_n.comp"