Test/spv.coopmat2_decode_vector.comp

#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_EXT_long_vector : 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;

buffer BufType8 {
   uint8_t x[];
} Buf8;

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);
}

f16vec4 decodeF16x4(const in dwordBuf b, const in uint32_t blockCoords[2], const in uint32_t coordInBlock[2])
{
    return f16vec4(unpackFloat2x16(b.d), unpackFloat2x16(b.d));
}

uint8_t decodeU8Scalar(const in dwordBuf b, const in uint32_t blockCoords[2], const in uint32_t coordInBlock[2])
{
    return uint8_t(b.d);
}

u8vec4 decodeU8x4(const in dwordBuf b, const in uint32_t blockCoords[2], const in uint32_t coordInBlock[2])
{
    return u8vec4(uint8_t(b.d), uint8_t(b.d >> 8), uint8_t(b.d >> 16), uint8_t(b.d >> 24));
}

vector<uint8_t, 8> decodeU8x8(const in dwordBuf b, const in uint32_t blockCoords[2], const in uint32_t coordInBlock[2])
{
    return vector<uint8_t, 8>(uint8_t(b.d),       uint8_t(b.d >> 8),
                              uint8_t(b.d >> 16), uint8_t(b.d >> 24),
                              uint8_t(b.d),       uint8_t(b.d >> 8),
                              uint8_t(b.d >> 16), uint8_t(b.d >> 24));
}

void main()
{
    coopmat<float16_t, gl_ScopeWorkgroup, 64, 32, gl_MatrixUseA> A;
    coopmat<uint8_t,   gl_ScopeWorkgroup, 64, 32, gl_MatrixUseA> Au8;

    tensorLayoutNV<2> t  = createTensorLayoutNV(2);
    t = setTensorLayoutBlockSizeNV(t, 4, 8);
    t = setTensorLayoutDimensionNV(t, 256, 512);

    // Scalar-only decode (the SPV_NV_cooperative_matrix2 baseline). No
    // capability or extension from this extension is required when the
    // vector form is not used.
    coopMatLoadTensorNV(A,   Buf.x,  0, t, decodeF16Scalar);
    coopMatLoadTensorNV(Au8, Buf8.x, 0, t, decodeU8Scalar);

    // Scalar + vector decode pair: the implementation may invoke either
    // function per call site. V == 2 fp16.
    coopMatLoadTensorNV(A,   Buf.x,  0, t, decodeF16Scalar, decodeF16x2);

    // V == 4 fp16.
    coopMatLoadTensorNV(A,   Buf.x,  0, t, decodeF16Scalar, decodeF16x4);

    // V == 4 u8.
    coopMatLoadTensorNV(Au8, Buf8.x, 0, t, decodeU8Scalar, decodeU8x4);

    // V == 8 u8.
    coopMatLoadTensorNV(Au8, Buf8.x, 0, t, decodeU8Scalar, decodeU8x8);
}