#version 450 core
#extension GL_KHR_memory_scope_semantics : enable
#extension GL_NV_cooperative_matrix : enable
#extension GL_EXT_shader_explicit_arithmetic_types_float16 : enable
#extension GL_EXT_buffer_reference : enable

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

const int X = 8;
layout(constant_id = 0) const int Y = 2;
const int Z = X*Y;

fcoopmatNV<16, gl_ScopeSubgroup, Z, 8> mC;
fcoopmatNV<16, gl_ScopeSubgroup, Z, 8> mC2[3];

int arr[mC.length()];
int arr2[mC2[1].length()];

layout(constant_id = 1) const float F = 3.0;

const fcoopmatNV<32, gl_ScopeSubgroup, Z, 8> mD = fcoopmatNV<32, gl_ScopeSubgroup, Z, 8>(0.0);
const fcoopmatNV<16, gl_ScopeSubgroup, 8, 8> mD2 = fcoopmatNV<16, gl_ScopeSubgroup, 8, 8>(1);

struct S { int a; int b; int c; };

const S s = S(12, 23, 34);

layout(set = 0, binding = 0, buffer_reference) coherent buffer Block {
    float y[1024*1024];
    float x[];
} block;

layout(set = 0, binding = 0) coherent buffer Block16 {
    float16_t y[1024*1024];
    float16_t x[];

    Block b;
} block16;

fcoopmatNV<16, gl_ScopeSubgroup, 8, 8> f16(fcoopmatNV<16, gl_ScopeSubgroup, 8, 8> m) { return -m; }
fcoopmatNV<32, gl_ScopeSubgroup, 8, 8> f32(fcoopmatNV<32, gl_ScopeSubgroup, 8, 8> m) { return -m; }

layout(constant_id = 2) const int SC = 1;
fcoopmatNV<16, gl_ScopeSubgroup, SC, SC> scm[SC][SC];

// sized for fcoopmatNV<16, gl_ScopeSubgroup, 16, 16>
shared uvec4 shmatrix[16*16*2/16];

void main()
{
    fcoopmatNV<32, gl_ScopeSubgroup, 16, (2>1?8:4)> m = fcoopmatNV<32, gl_ScopeSubgroup, 16, (2>1?8:4)>(0.0);

    m = m + m;
    m = m - m;
    m = -m;
    m = 2.0*m;
    m = m*2.0;

    fcoopmatNV<16, gl_ScopeSubgroup, 16, 8> m2 = fcoopmatNV<16, gl_ScopeSubgroup, 16, 8>(m);

    float x = m[1];
    m[0] = x;

    coopMatLoadNV(m, block.x, 16, 128, false);
    coopMatStoreNV(m, block.x, 16, 128, false);
    coopMatLoadNV(m2, block16.x, 16, 128, false);
    coopMatStoreNV(m2, block16.x, 16, 128, false);
    coopMatLoadNV(m, block16.b.x, 16, 128, false);
    coopMatStoreNV(m, block16.b.x, 16, 128, false);

    fcoopmatNV<16, gl_ScopeSubgroup, 16, 8> A;
    fcoopmatNV<16, gl_ScopeSubgroup, 8, 8> B;
    fcoopmatNV<32, gl_ScopeSubgroup, 16, 8> C;
    fcoopmatNV<32, gl_ScopeSubgroup, 16, 8> D;
    D = coopMatMulAddNV(A, B, C);

    int l = D.length();

    fcoopmatNV<16, gl_ScopeSubgroup, 8, 8> E;

    fcoopmatNV<16, gl_ScopeSubgroup, Z, Z> F = fcoopmatNV<16, gl_ScopeSubgroup, Z, Z>(0.0);

    fcoopmatNV<32, gl_ScopeSubgroup, 16, (2>1?8:4)> a[5];
    a[3][0] = 1.0;

    float md1 = mD[1];

    md1 += (m += m)[1234];

    mC2[1] = mC2[2];

    coopMatLoadNV(m, block.y, 16, 128, false);
    coopMatStoreNV(m, block.y, 16, 128, false);
    coopMatLoadNV(m2, block16.y, 16, 128, false);
    coopMatStoreNV(m2, block16.y, 16, 128, false);

    fcoopmatNV<16, gl_ScopeSubgroup, 8, 8> p1;
    fcoopmatNV<32, gl_ScopeSubgroup, 8, 8> p2;

    p1 = f16(p1);
    p2 = f32(p2);

    p1 = fcoopmatNV<16, gl_ScopeSubgroup, 8, 8>(0.0);
    p2 = fcoopmatNV<32, gl_ScopeSubgroup, 8, 8>(0.0);

    p1 /= p1;

    p1 *= float16_t(2.0);
    p2 *= 4.0;

    fcoopmatNV<16, gl_ScopeSubgroup, 16, 8> ms;
    coopMatLoadNV(ms, shmatrix, 1, 2, false);
    coopMatStoreNV(ms, shmatrix, 1, 2, false);

}