#include <metal_stdlib>
#include <simd/simd.h>

using namespace metal;

struct Foo
{
    float4x4 m;
};

struct SSBO2
{
    Foo out_data[1];
};

struct SSBO
{
    Foo in_data[1];
};

constant uint3 gl_WorkGroupSize [[maybe_unused]] = uint3(1u);

kernel void main0(device SSBO2& _23 [[buffer(0)]], const device SSBO& _30 [[buffer(1)]], uint3 gl_GlobalInvocationID [[thread_position_in_grid]])
{
    _23.out_data[gl_GlobalInvocationID.x].m = _30.in_data[gl_GlobalInvocationID.x].m * _30.in_data[gl_GlobalInvocationID.x].m;
}