#version 450

#include "rope_common.comp"

layout(binding = 0) buffer restrict readonly  tensorInA { float16_t inA[]; };
layout(binding = 1) buffer restrict readonly  tensorInB { int       inB[]; };
layout(binding = 2) buffer restrict writeonly tensorOut { float16_t out_[]; };

void main() {
    const uint i3 = gl_WorkGroupID.z;
    const uint i2 = gl_WorkGroupID.y;
    const uint i1 = gl_WorkGroupID.x;

    const bool is_neox = (pcs.mode & 2) != 0;

    float corr_dims[2];
    rope_yarn_corr_dims(pcs.n_dims, pcs.n_orig_ctx, pcs.freq_base, pcs.beta_fast, pcs.beta_slow, corr_dims);

    const float theta_scale = pow(pcs.freq_base, -2.0/pcs.n_dims);

    const int p = inB[pcs.inBOff + i2];

    float theta = float(p);

    if (!is_neox) {
        for (uint i0 = 0; i0 < pcs.ne0; i0 += 2) {
            float cos_theta, sin_theta;
            rope_yarn(theta, pcs.freq_scale, corr_dims, i0, pcs.ext_factor, pcs.attn_factor, cos_theta, sin_theta);

            theta *= theta_scale;

            const uint src      = uint((i3*pcs.nb03 + i2*pcs.nb02 + i1*pcs.nb01 + i0*pcs.nb00) / 2) + pcs.inAOff; // Based from in
            const uint dst_data = uint((i3*pcs.nb3  + i2*pcs.nb2  + i1*pcs.nb1  + i0*pcs.nb0)  / 2) + pcs.outOff; // Based from out_

            const float x0 = float(inA[src]);
            const float x1 = float(inA[src+1]);

            out_[dst_data]   = float16_t(x0*cos_theta - x1*sin_theta);
            out_[dst_data+1] = float16_t(x0*sin_theta + x1*cos_theta);
        }
    } else {
        const float inv_ndims = -1.f/pcs.n_dims;
        for (uint ic = 0; ic < pcs.n_dims; ic += 2) {
            const uint cur_rot = ic;

            float cos_theta, sin_theta;
            rope_yarn(theta, pcs.freq_scale, corr_dims, cur_rot, pcs.ext_factor, pcs.attn_factor, cos_theta, sin_theta);

            theta *= theta_scale;

            const uint i0 = ic/2;

            const uint src      = uint((i3*pcs.nb03 + i2*pcs.nb02 + i1*pcs.nb01 + i0*pcs.nb00) / 2) + pcs.inAOff; // Based from in
            const uint dst_data = uint((i3*pcs.nb3  + i2*pcs.nb2  + i1*pcs.nb1  + i0*pcs.nb0)  / 2) + pcs.outOff; // Based from out_

            const float x0 = float(inA[src]);
            const float x1 = float(inA[src+pcs.n_dims/2]);

            out_[dst_data]              = float16_t(x0*cos_theta - x1*sin_theta);
            out_[dst_data+pcs.n_dims/2] = float16_t(x0*sin_theta + x1*cos_theta);
        }

        for (uint ic = pcs.n_dims; ic < pcs.ne0; ic += 2) {
            const uint i0 = ic;

            const uint src      = uint((i3*pcs.nb03 + i2*pcs.nb02 + i1*pcs.nb01 + i0*pcs.nb00) / 2) + pcs.inAOff; // Based from in
            const uint dst_data = uint((i3*pcs.nb3  + i2*pcs.nb2  + i1*pcs.nb1  + i0*pcs.nb0)  / 2) + pcs.outOff; // Based from out_

            out_[dst_data + 0] = inA[src + 0];
            out_[dst_data + 1] = inA[src + 1];
        }
    }
}