#version 450 core
#extension GL_KHR_memory_scope_semantics : enable
#extension GL_KHR_cooperative_matrix : enable
#extension GL_EXT_shader_explicit_arithmetic_types : enable
#extension GL_NV_cooperative_matrix2 : enable
#extension GL_NV_cooperative_matrix_decode_vector : enable
#extension GL_EXT_buffer_reference : enable

layout (local_size_x = 64, local_size_y = 1, local_size_z = 1) in;

buffer BufType {
   float16_t x[];
} Buf;

layout(buffer_reference, std430, buffer_reference_align = 4) buffer dwordBuf {
    uint32_t d;
};

float16_t decodeF16Scalar(const in dwordBuf b, const in uint32_t blockCoords[2], const in uint32_t coordInBlock[2])
{
    return float16_t(b.d);
}

f16vec2 decodeF16x2(const in dwordBuf b, const in uint32_t blockCoords[2], const in uint32_t coordInBlock[2])
{
    return unpackFloat2x16(b.d);
}

// Vector return with V == 3, not in {2, 4, 8}.
f16vec3 decodeF16x3_badLength(const in dwordBuf b, const in uint32_t blockCoords[2], const in uint32_t coordInBlock[2])
{
    return f16vec3(0);
}

// Vector return whose component type doesn't match the matrix type.
vec2 decodeF32x2_badType(const in dwordBuf b, const in uint32_t blockCoords[2], const in uint32_t coordInBlock[2])
{
    return vec2(0);
}

void main()
{
    coopmat<float16_t, gl_ScopeWorkgroup, 64, 32, gl_MatrixUseA> A;

    tensorLayoutNV<2> t  = createTensorLayoutNV(2);

    // Single-arg form must be scalar; a vector return is only legal in
    // the second function position now.
    coopMatLoadTensorNV(A, Buf.x, 0, t, decodeF16x2);

    // V == 3, not in {2, 4, 8}.
    coopMatLoadTensorNV(A, Buf.x, 0, t, decodeF16Scalar, decodeF16x3_badLength);

    // Vector component type doesn't match the matrix component type.
    coopMatLoadTensorNV(A, Buf.x, 0, t, decodeF16Scalar, decodeF32x2_badType);
}