llama.cpp/ggml/src/ggml-kompute/kompute-shaders/op_rope_neox_f16.comp

#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 readonly  tensorInC { float     inC[]; };
layout(binding = 3) 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;

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

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

    float theta_base = float(inB[pcs.inBOff + i2]);
    float inv_ndims = -1.f/pcs.n_dims;

    float cos_theta;
    float sin_theta;

    for (uint i0 = 2*gl_LocalInvocationIndex; i0 < pcs.ne0; i0 += 2*gl_WorkGroupSize.x) {
        if (i0 < pcs.n_dims) {
            uint ic = i0/2;

            float theta = theta_base * pow(pcs.freq_base, inv_ndims*i0);

            const float freq_factor = pcs.has_freq_factors ? inC[pcs.inCOff + ic] : 1.0f;

            rope_yarn(theta/freq_factor, pcs.freq_scale, corr_dims, i0, pcs.ext_factor, pcs.attn_factor, cos_theta, sin_theta);

            const uint src      = uint((i3*pcs.nb03 + i2*pcs.nb02 + i1*pcs.nb01 + ic*pcs.nb00) / 2) + pcs.inAOff; // Based from in
            const uint dst_data = uint((i3*pcs.nb3  + i2*pcs.nb2  + i1*pcs.nb1  + ic*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);
        } else {
            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]   = inA[src];
            out_[dst_data+1] = inA[src+1];
        }
    }
}