reference/shaders-msl-no-opt/packing/matrix-multiply-unpacked-row-major-2.comp

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

using namespace metal;

typedef packed_float3 packed_rm_float3x3[3];

struct SSBO
{
    packed_rm_float3x3 m0;
    packed_rm_float3x3 m1;
    packed_float3 v0;
    packed_float3 v1;
};

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

kernel void main0(device SSBO& _11 [[buffer(0)]])
{
    _11.v0 = float3(_11.v1) * (float3x3(float3(_11.m1[0]), float3(_11.m1[1]), float3(_11.m1[2])) * float3x3(float3(_11.m0[0]), float3(_11.m0[1]), float3(_11.m0[2])));
    _11.v0 = (float3(_11.v1) * float3x3(float3(_11.m1[0]), float3(_11.m1[1]), float3(_11.m1[2]))) * float3x3(float3(_11.m0[0]), float3(_11.m0[1]), float3(_11.m0[2]));
    _11.v0 = float3x3(float3(_11.m1[0]), float3(_11.m1[1]), float3(_11.m1[2])) * (float3x3(float3(_11.m0[0]), float3(_11.m0[1]), float3(_11.m0[2])) * float3(_11.v1));
    _11.v0 = (float3x3(float3(_11.m1[0]), float3(_11.m1[1]), float3(_11.m1[2])) * float3x3(float3(_11.m0[0]), float3(_11.m0[1]), float3(_11.m0[2]))) * float3(_11.v1);
}