#version 450
#extension GL_EXT_shader_explicit_arithmetic_types : require

#include "mul_mat_vec_base.comp"

layout(local_size_x_id = 0, local_size_y = 1, local_size_z = 1) in;

layout (constant_id = 0) const uint BLOCK_SIZE = 32;

shared FLOAT_TYPE tmp[BLOCK_SIZE];

void main() {
    const uint row = gl_WorkGroupID.x + gl_NumWorkGroups.x * gl_WorkGroupID.z;

    if (row >= p.stride_d) {
        return;
    }

    uint a_offset, b_offset, d_offset;
    get_offsets(a_offset, b_offset, d_offset);

    const uint num_blocks_per_row = p.ncols / QUANT_K;
    const uint ib0 = a_offset / QUANT_K + row*num_blocks_per_row;

    // 16 threads are used to process each block
    const uint it_size = gl_WorkGroupSize.x/16;
    const uint tid = gl_LocalInvocationID.x;
    const uint itid = tid%16;  // 0...16
    const uint ix  = tid/16;

    const uint step = 8;

    const uint v_im = itid/step;                            // 0 or 1. 0 computes 0..., 1 computes 128...
    const uint v_in = itid - step*v_im;                     // 0...15 or 0...7

    const uint8_t m = uint8_t(1 << (4 * v_im));

    const uint l0 = 2*v_in;                                // 0...15
    const uint q_offset = 32*v_im + l0;
    const uint y_offset = 128*v_im + l0;

    FLOAT_TYPE temp = FLOAT_TYPE(0.0); // partial sum for thread in warp

    const uint s_shift = 4 * v_im;

    [[unroll]] for (uint i = ix; i < num_blocks_per_row; i += it_size) {
        const uint y_idx = i * QUANT_K + y_offset;

        const FLOAT_TYPE d = FLOAT_TYPE(data_a[ib0 + i].d);

        B_TYPE_VEC2 b0 = data_b_v2[(b_offset + y_idx) / 2 + 0];
        B_TYPE_VEC2 b16 = data_b_v2[(b_offset + y_idx) / 2 + 8];
        B_TYPE_VEC2 b32 = data_b_v2[(b_offset + y_idx) / 2 + 16];
        B_TYPE_VEC2 b48 = data_b_v2[(b_offset + y_idx) / 2 + 24];
        B_TYPE_VEC2 b64 = data_b_v2[(b_offset + y_idx) / 2 + 32];
        B_TYPE_VEC2 b80 = data_b_v2[(b_offset + y_idx) / 2 + 40];
        B_TYPE_VEC2 b96 = data_b_v2[(b_offset + y_idx) / 2 + 48];
        B_TYPE_VEC2 b112 = data_b_v2[(b_offset + y_idx) / 2 + 56];

        uint16_t s0_16 = data_a_packed16[ib0 + i].scales[0];
        uint16_t s2_16 = data_a_packed16[ib0 + i].scales[1];
        uint16_t s4_16 = data_a_packed16[ib0 + i].scales[2];
        uint16_t s6_16 = data_a_packed16[ib0 + i].scales[3];
        uint16_t s8_16 = data_a_packed16[ib0 + i].scales[4];
        uint16_t s10_16 = data_a_packed16[ib0 + i].scales[5];
        u8vec2 s0 = unpack8(s0_16);
        u8vec2 s2 = unpack8(s2_16);
        u8vec2 s4 = unpack8(s4_16);
        u8vec2 s6 = unpack8(s6_16);
        u8vec2 s8 = unpack8(s8_16);
        u8vec2 s10 = unpack8(s10_16);

        FLOAT_TYPE sum = FLOAT_TYPE(0.0);
        [[unroll]] for (int l = 0; l < 2; ++l) {
            sum = fma(FLOAT_TYPE(b0[l])   * FLOAT_TYPE(int8_t(((s0[0] >> s_shift) & 0xF) | ((s8[0]  >> (s_shift + 0) & 0x3) << 4)) - 32), FLOAT_TYPE(((data_a[ib0 + i].qs[q_offset + l   ]     ) & 3) - (((data_a[ib0 + i].hmask[l0 + l   ] & (m << 0)) != 0) ? 0 : 4)),
                  fma(FLOAT_TYPE(b32[l])  * FLOAT_TYPE(int8_t(((s2[0] >> s_shift) & 0xF) | ((s10[0] >> (s_shift + 0) & 0x3) << 4)) - 32), FLOAT_TYPE(((data_a[ib0 + i].qs[q_offset + l   ] >> 2) & 3) - (((data_a[ib0 + i].hmask[l0 + l   ] & (m << 1)) != 0) ? 0 : 4)),
                  fma(FLOAT_TYPE(b64[l])  * FLOAT_TYPE(int8_t(((s4[0] >> s_shift) & 0xF) | ((s8[0]  >> (s_shift + 2) & 0x3) << 4)) - 32), FLOAT_TYPE(((data_a[ib0 + i].qs[q_offset + l   ] >> 4) & 3) - (((data_a[ib0 + i].hmask[l0 + l   ] & (m << 2)) != 0) ? 0 : 4)),
                  fma(FLOAT_TYPE(b96[l])  * FLOAT_TYPE(int8_t(((s6[0] >> s_shift) & 0xF) | ((s10[0] >> (s_shift + 2) & 0x3) << 4)) - 32), FLOAT_TYPE(((data_a[ib0 + i].qs[q_offset + l   ] >> 6) & 3) - (((data_a[ib0 + i].hmask[l0 + l   ] & (m << 3)) != 0) ? 0 : 4)),
                  fma(FLOAT_TYPE(b16[l])  * FLOAT_TYPE(int8_t(((s0[1] >> s_shift) & 0xF) | ((s8[1]  >> (s_shift + 0) & 0x3) << 4)) - 32), FLOAT_TYPE(((data_a[ib0 + i].qs[q_offset + l+16]     ) & 3) - (((data_a[ib0 + i].hmask[l0 + l+16] & (m << 0)) != 0) ? 0 : 4)),
                  fma(FLOAT_TYPE(b48[l])  * FLOAT_TYPE(int8_t(((s2[1] >> s_shift) & 0xF) | ((s10[1] >> (s_shift + 0) & 0x3) << 4)) - 32), FLOAT_TYPE(((data_a[ib0 + i].qs[q_offset + l+16] >> 2) & 3) - (((data_a[ib0 + i].hmask[l0 + l+16] & (m << 1)) != 0) ? 0 : 4)),
                  fma(FLOAT_TYPE(b80[l])  * FLOAT_TYPE(int8_t(((s4[1] >> s_shift) & 0xF) | ((s8[1]  >> (s_shift + 2) & 0x3) << 4)) - 32), FLOAT_TYPE(((data_a[ib0 + i].qs[q_offset + l+16] >> 4) & 3) - (((data_a[ib0 + i].hmask[l0 + l+16] & (m << 2)) != 0) ? 0 : 4)),
                  fma(FLOAT_TYPE(b112[l]) * FLOAT_TYPE(int8_t(((s6[1] >> s_shift) & 0xF) | ((s10[1] >> (s_shift + 2) & 0x3) << 4)) - 32), FLOAT_TYPE(((data_a[ib0 + i].qs[q_offset + l+16] >> 6) & 3) - (((data_a[ib0 + i].hmask[l0 + l+16] & (m << 3)) != 0) ? 0 : 4)), sum))))))));
        }
        temp = fma(d, sum, temp);
    }

    tmp[gl_LocalInvocationID.x] = temp;

    // sum up partial sums and write back result
    barrier();
    [[unroll]] for (uint s = gl_WorkGroupSize.x/2; s > 0; s >>= 1) {
        if (tid < s) {
            tmp[tid] += tmp[tid + s];
        }
        barrier();
    }
    if (tid == 0) {
        data_d[d_offset + row] = D_TYPE(tmp[0]);
    }
}