% Many of these tests fail on MoltenVK because of a bug in Metal, which causes
% SIMD operations (i.e., wave operations for Apple) to be hoisted out of the
% "if"statement, thereby changing which threads (i.e., wave groups for Apple)
% appear as active. This was reported to Apple as FB15624583.

[require]
shader model >= 6.0
format r32-uint uav-load

[uav 0]
format r32-uint
size (buffer, 4)

0x0 0x1 0x2 0x13

[uav 1]
format r32-uint
size (buffer, 4)

0x100 0x200 0x400 0x800

[uav 2]
format r32-uint
size (buffer, 20)

0 0 0 0 0
0 0 0 0 0
0 0 0 0 0
0 0 0 0 0

[compute shader]
RWBuffer<uint> u0 : register(u0);
RWBuffer<uint> u1 : register(u1);
RWBuffer<uint> u2 : register(u2);

    [numthreads(4, 1, 1)]
void main(uint id : SV_GroupIndex)
{
    const unsigned int POS_COUNT = 5;
    unsigned int i, pos = 0;
    /* If SPV_KHR_subgroup_uniform_control_flow is not supported,
     * SPIR-V doesn't guarantee reconvergence after the loop. */
    for (i = 0; i < 2; ++i)
    {
        u2[POS_COUNT * id + pos++] = WaveActiveSum(u1[id]) + 16 * i + 1;
        if (u0[id] & (1 << i))
        {
            /* In many programming languages this statement could be
             * moved after the loop. Not here, because the set of
             * active invocations is supposed to depend on whether
             * we're still in the loop (i.e., before reconverging the
             * invocations that diverged during the loop) or not. */
            u2[POS_COUNT * id + pos++] = WaveActiveSum(u1[id]) + 16 * i + 2;
            break;
        }
        else
        {
            u2[POS_COUNT * id + pos++] = WaveActiveSum(u1[id]) + 16 * i + 3;
        }
    }
    u2[POS_COUNT * id + pos++] = WaveActiveSum(u1[id]) + 16 * i + 4;
}

[test]
dispatch 1 1 1
probe uav 2 (0) u32(0xf01)
bug(mvk) probe uav 2 (1) u32(0x503)
probe uav 2 (2) u32(0x511)
bug(mvk) probe uav 2 (3) u32(0x113)
probe uav 2 (4) u32(0xf24)

probe uav 2 (5) u32(0xf01)
bug(mvk) probe uav 2 (6) u32(0xa02)
probe uav 2 (7) u32(0xf04)
probe uav 2 (8) u32(0)
probe uav 2 (9) u32(0)

probe uav 2 (10) u32(0xf01)
bug(mvk) probe uav 2 (11) u32(0x503)
probe uav 2 (12) u32(0x511)
bug(mvk) probe uav 2 (13) u32(0x412)
probe uav 2 (14) u32(0xf14)

probe uav 2 (15) u32(0xf01)
bug(mvk) probe uav 2 (16) u32(0xa02)
probe uav 2 (17) u32(0xf04)
probe uav 2 (18) u32(0)
probe uav 2 (19) u32(0)

[uav 2]
format r32-uint
size (buffer, 20)

0 0 0 0 0
0 0 0 0 0
0 0 0 0 0
0 0 0 0 0

[compute shader]
RWBuffer<uint> u0 : register(u0);
RWBuffer<uint> u1 : register(u1);
RWBuffer<uint> u2 : register(u2);

    [numthreads(4, 1, 1)]
void main(uint id : SV_GroupIndex)
{
    const unsigned int POS_COUNT = 5;
    unsigned int i, pos = 0;
    /* If SPV_KHR_maximal_reconvergence is not supported, SPIR-V
     * doesn't guarantee reconvergence after the loop even if
     * SPV_KHR_subgroup_uniform_control_flow is supported, because the
     * flow wasn't fully convergent when entering the loop. */
    if (!(u0[id] & 0x10))
    {
        for (i = 0; i < 2; ++i)
        {
            u2[POS_COUNT * id + pos++] = WaveActiveSum(u1[id]) + 16 * i + 1;
            if (u0[id] & (1 << i))
            {
                u2[POS_COUNT * id + pos++] = WaveActiveSum(u1[id]) + 16 * i + 2;
                break;
            }
            else
            {
                u2[POS_COUNT * id + pos++] = WaveActiveSum(u1[id]) + 16 * i + 3;
            }
        }
        u2[POS_COUNT * id + pos++] = WaveActiveSum(u1[id]) + 16 * i + 4;
    }
}

[test]
dispatch 1 1 1
probe uav 2 (0) u32(0x701)
bug(mvk) probe uav 2 (1) u32(0x503)
probe uav 2 (2) u32(0x511)
bug(mvk) probe uav 2 (3) u32(0x113)
probe uav 2 (4) u32(0x724)

probe uav 2 (5) u32(0x701)
bug(mvk) probe uav 2 (6) u32(0x202)
probe uav 2 (7) u32(0x704)
probe uav 2 (8) u32(0)
probe uav 2 (9) u32(0)

probe uav 2 (10) u32(0x701)
bug(mvk) probe uav 2 (11) u32(0x503)
probe uav 2 (12) u32(0x511)
bug(mvk) probe uav 2 (13) u32(0x412)
probe uav 2 (14) u32(0x714)

probe uav 2 (15) u32(0)
probe uav 2 (16) u32(0)
probe uav 2 (17) u32(0)
probe uav 2 (18) u32(0)
probe uav 2 (19) u32(0)