llama.cpp/ggml/src/ggml-vulkan/vulkan-shaders/rms_norm.comp

#version 450

#include "generic_unary_head.comp"
#include "types.comp"

#extension GL_EXT_control_flow_attributes : enable
#define BLOCK_SIZE 512

layout(local_size_x = BLOCK_SIZE, local_size_y = 1, local_size_z = 1) in;

shared FLOAT_TYPE sum[BLOCK_SIZE];

void main() {
    const uint ncols     = p.ne00;
    const uint nrows     = gl_NumWorkGroups.x;
    const uint nchannels = gl_NumWorkGroups.y;

    const uint row       = gl_WorkGroupID.x;
    const uint channel   = gl_WorkGroupID.y;
    const uint samp      = gl_WorkGroupID.z;
    const uint tid       = gl_LocalInvocationID.x;

    const uint stride_row       = p.nb01;
    const uint stride_channel   = p.nb02;
    const uint stride_sample    = p.nb03;

    uint32_t a_offset = samp*stride_sample + channel*stride_channel + row*stride_row + get_aoffset();
    uint32_t d_offset = ((samp*nchannels + channel)*nrows + row)*ncols + get_doffset();

    sum[tid] = FLOAT_TYPE(0.0f); // partial sum for thread in warp

    [[unroll]] for (uint col = tid; col < ncols; col += BLOCK_SIZE) {
        const FLOAT_TYPE xi = FLOAT_TYPE(data_a[a_offset + col]);
        sum[tid] += xi * xi;
    }

    // sum up partial sums and write back result
    barrier();
    [[unroll]] for (int s = BLOCK_SIZE / 2; s > 0; s >>= 1) {
        if (tid < s) {
            sum[tid] += sum[tid + s];
        }
        barrier();
    }

    const FLOAT_TYPE mean = sum[0] / FLOAT_TYPE(ncols);
    const FLOAT_TYPE scale = inversesqrt(mean + FLOAT_TYPE(p.param1));

    [[unroll]] for (uint col = tid; col < ncols; col += BLOCK_SIZE) {
        data_d[d_offset + col] = D_TYPE(scale * FLOAT_TYPE(data_a[a_offset + col]));
    }
}